xref: /freebsd/sys/contrib/openzfs/module/zfs/zil.c (revision 22b267e823719ef22790ed548fb11da8aa81d1ae)
1eda14cbcSMatt Macy /*
2eda14cbcSMatt Macy  * CDDL HEADER START
3eda14cbcSMatt Macy  *
4eda14cbcSMatt Macy  * The contents of this file are subject to the terms of the
5eda14cbcSMatt Macy  * Common Development and Distribution License (the "License").
6eda14cbcSMatt Macy  * You may not use this file except in compliance with the License.
7eda14cbcSMatt Macy  *
8eda14cbcSMatt Macy  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9271171e0SMartin Matuska  * or https://opensource.org/licenses/CDDL-1.0.
10eda14cbcSMatt Macy  * See the License for the specific language governing permissions
11eda14cbcSMatt Macy  * and limitations under the License.
12eda14cbcSMatt Macy  *
13eda14cbcSMatt Macy  * When distributing Covered Code, include this CDDL HEADER in each
14eda14cbcSMatt Macy  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15eda14cbcSMatt Macy  * If applicable, add the following below this CDDL HEADER, with the
16eda14cbcSMatt Macy  * fields enclosed by brackets "[]" replaced with your own identifying
17eda14cbcSMatt Macy  * information: Portions Copyright [yyyy] [name of copyright owner]
18eda14cbcSMatt Macy  *
19eda14cbcSMatt Macy  * CDDL HEADER END
20eda14cbcSMatt Macy  */
21eda14cbcSMatt Macy /*
22eda14cbcSMatt Macy  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23eda14cbcSMatt Macy  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24eda14cbcSMatt Macy  * Copyright (c) 2014 Integros [integros.com]
25eda14cbcSMatt Macy  * Copyright (c) 2018 Datto Inc.
26eda14cbcSMatt Macy  */
27eda14cbcSMatt Macy 
28eda14cbcSMatt Macy /* Portions Copyright 2010 Robert Milkowski */
29eda14cbcSMatt Macy 
30eda14cbcSMatt Macy #include <sys/zfs_context.h>
31eda14cbcSMatt Macy #include <sys/spa.h>
32eda14cbcSMatt Macy #include <sys/spa_impl.h>
33eda14cbcSMatt Macy #include <sys/dmu.h>
34eda14cbcSMatt Macy #include <sys/zap.h>
35eda14cbcSMatt Macy #include <sys/arc.h>
36eda14cbcSMatt Macy #include <sys/stat.h>
37eda14cbcSMatt Macy #include <sys/zil.h>
38eda14cbcSMatt Macy #include <sys/zil_impl.h>
39eda14cbcSMatt Macy #include <sys/dsl_dataset.h>
40eda14cbcSMatt Macy #include <sys/vdev_impl.h>
41eda14cbcSMatt Macy #include <sys/dmu_tx.h>
42eda14cbcSMatt Macy #include <sys/dsl_pool.h>
43eda14cbcSMatt Macy #include <sys/metaslab.h>
44eda14cbcSMatt Macy #include <sys/trace_zfs.h>
45eda14cbcSMatt Macy #include <sys/abd.h>
462a58b312SMartin Matuska #include <sys/brt.h>
47271171e0SMartin Matuska #include <sys/wmsum.h>
48eda14cbcSMatt Macy 
49eda14cbcSMatt Macy /*
50eda14cbcSMatt Macy  * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
51eda14cbcSMatt Macy  * calls that change the file system. Each itx has enough information to
52eda14cbcSMatt Macy  * be able to replay them after a system crash, power loss, or
53eda14cbcSMatt Macy  * equivalent failure mode. These are stored in memory until either:
54eda14cbcSMatt Macy  *
55eda14cbcSMatt Macy  *   1. they are committed to the pool by the DMU transaction group
56eda14cbcSMatt Macy  *      (txg), at which point they can be discarded; or
57eda14cbcSMatt Macy  *   2. they are committed to the on-disk ZIL for the dataset being
58eda14cbcSMatt Macy  *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
59eda14cbcSMatt Macy  *      requirement).
60eda14cbcSMatt Macy  *
61eda14cbcSMatt Macy  * In the event of a crash or power loss, the itxs contained by each
62eda14cbcSMatt Macy  * dataset's on-disk ZIL will be replayed when that dataset is first
63eda14cbcSMatt Macy  * instantiated (e.g. if the dataset is a normal filesystem, when it is
64eda14cbcSMatt Macy  * first mounted).
65eda14cbcSMatt Macy  *
66eda14cbcSMatt Macy  * As hinted at above, there is one ZIL per dataset (both the in-memory
67eda14cbcSMatt Macy  * representation, and the on-disk representation). The on-disk format
68eda14cbcSMatt Macy  * consists of 3 parts:
69eda14cbcSMatt Macy  *
70eda14cbcSMatt Macy  * 	- a single, per-dataset, ZIL header; which points to a chain of
71eda14cbcSMatt Macy  * 	- zero or more ZIL blocks; each of which contains
72eda14cbcSMatt Macy  * 	- zero or more ZIL records
73eda14cbcSMatt Macy  *
74eda14cbcSMatt Macy  * A ZIL record holds the information necessary to replay a single
75eda14cbcSMatt Macy  * system call transaction. A ZIL block can hold many ZIL records, and
76eda14cbcSMatt Macy  * the blocks are chained together, similarly to a singly linked list.
77eda14cbcSMatt Macy  *
78eda14cbcSMatt Macy  * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
79eda14cbcSMatt Macy  * block in the chain, and the ZIL header points to the first block in
80eda14cbcSMatt Macy  * the chain.
81eda14cbcSMatt Macy  *
82eda14cbcSMatt Macy  * Note, there is not a fixed place in the pool to hold these ZIL
83eda14cbcSMatt Macy  * blocks; they are dynamically allocated and freed as needed from the
84eda14cbcSMatt Macy  * blocks available on the pool, though they can be preferentially
85eda14cbcSMatt Macy  * allocated from a dedicated "log" vdev.
86eda14cbcSMatt Macy  */
87eda14cbcSMatt Macy 
88eda14cbcSMatt Macy /*
89eda14cbcSMatt Macy  * This controls the amount of time that a ZIL block (lwb) will remain
90eda14cbcSMatt Macy  * "open" when it isn't "full", and it has a thread waiting for it to be
91eda14cbcSMatt Macy  * committed to stable storage. Please refer to the zil_commit_waiter()
92eda14cbcSMatt Macy  * function (and the comments within it) for more details.
93eda14cbcSMatt Macy  */
94be181ee2SMartin Matuska static uint_t zfs_commit_timeout_pct = 5;
95eda14cbcSMatt Macy 
96eda14cbcSMatt Macy /*
9715f0b8c3SMartin Matuska  * Minimal time we care to delay commit waiting for more ZIL records.
9815f0b8c3SMartin Matuska  * At least FreeBSD kernel can't sleep for less than 2us at its best.
9915f0b8c3SMartin Matuska  * So requests to sleep for less then 5us is a waste of CPU time with
10015f0b8c3SMartin Matuska  * a risk of significant log latency increase due to oversleep.
10115f0b8c3SMartin Matuska  */
10215f0b8c3SMartin Matuska static uint64_t zil_min_commit_timeout = 5000;
10315f0b8c3SMartin Matuska 
10415f0b8c3SMartin Matuska /*
105eda14cbcSMatt Macy  * See zil.h for more information about these fields.
106eda14cbcSMatt Macy  */
107271171e0SMartin Matuska static zil_kstat_values_t zil_stats = {
108eda14cbcSMatt Macy 	{ "zil_commit_count",			KSTAT_DATA_UINT64 },
109eda14cbcSMatt Macy 	{ "zil_commit_writer_count",		KSTAT_DATA_UINT64 },
110eda14cbcSMatt Macy 	{ "zil_itx_count",			KSTAT_DATA_UINT64 },
111eda14cbcSMatt Macy 	{ "zil_itx_indirect_count",		KSTAT_DATA_UINT64 },
112eda14cbcSMatt Macy 	{ "zil_itx_indirect_bytes",		KSTAT_DATA_UINT64 },
113eda14cbcSMatt Macy 	{ "zil_itx_copied_count",		KSTAT_DATA_UINT64 },
114eda14cbcSMatt Macy 	{ "zil_itx_copied_bytes",		KSTAT_DATA_UINT64 },
115eda14cbcSMatt Macy 	{ "zil_itx_needcopy_count",		KSTAT_DATA_UINT64 },
116eda14cbcSMatt Macy 	{ "zil_itx_needcopy_bytes",		KSTAT_DATA_UINT64 },
117eda14cbcSMatt Macy 	{ "zil_itx_metaslab_normal_count",	KSTAT_DATA_UINT64 },
118eda14cbcSMatt Macy 	{ "zil_itx_metaslab_normal_bytes",	KSTAT_DATA_UINT64 },
1194e8d558cSMartin Matuska 	{ "zil_itx_metaslab_normal_write",	KSTAT_DATA_UINT64 },
1204e8d558cSMartin Matuska 	{ "zil_itx_metaslab_normal_alloc",	KSTAT_DATA_UINT64 },
121eda14cbcSMatt Macy 	{ "zil_itx_metaslab_slog_count",	KSTAT_DATA_UINT64 },
122eda14cbcSMatt Macy 	{ "zil_itx_metaslab_slog_bytes",	KSTAT_DATA_UINT64 },
1234e8d558cSMartin Matuska 	{ "zil_itx_metaslab_slog_write",	KSTAT_DATA_UINT64 },
1244e8d558cSMartin Matuska 	{ "zil_itx_metaslab_slog_alloc",	KSTAT_DATA_UINT64 },
125eda14cbcSMatt Macy };
126eda14cbcSMatt Macy 
127271171e0SMartin Matuska static zil_sums_t zil_sums_global;
128271171e0SMartin Matuska static kstat_t *zil_kstats_global;
129eda14cbcSMatt Macy 
130eda14cbcSMatt Macy /*
131eda14cbcSMatt Macy  * Disable intent logging replay.  This global ZIL switch affects all pools.
132eda14cbcSMatt Macy  */
133eda14cbcSMatt Macy int zil_replay_disable = 0;
134eda14cbcSMatt Macy 
135eda14cbcSMatt Macy /*
136eda14cbcSMatt Macy  * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
137eda14cbcSMatt Macy  * the disk(s) by the ZIL after an LWB write has completed. Setting this
138eda14cbcSMatt Macy  * will cause ZIL corruption on power loss if a volatile out-of-order
139eda14cbcSMatt Macy  * write cache is enabled.
140eda14cbcSMatt Macy  */
141e92ffd9bSMartin Matuska static int zil_nocacheflush = 0;
142eda14cbcSMatt Macy 
143eda14cbcSMatt Macy /*
144eda14cbcSMatt Macy  * Limit SLOG write size per commit executed with synchronous priority.
145eda14cbcSMatt Macy  * Any writes above that will be executed with lower (asynchronous) priority
146eda14cbcSMatt Macy  * to limit potential SLOG device abuse by single active ZIL writer.
147eda14cbcSMatt Macy  */
148*22b267e8SMartin Matuska static uint64_t zil_slog_bulk = 64 * 1024 * 1024;
149eda14cbcSMatt Macy 
150eda14cbcSMatt Macy static kmem_cache_t *zil_lwb_cache;
151eda14cbcSMatt Macy static kmem_cache_t *zil_zcw_cache;
152eda14cbcSMatt Macy 
1534e8d558cSMartin Matuska static void zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx);
1544e8d558cSMartin Matuska static itx_t *zil_itx_clone(itx_t *oitx);
1554e8d558cSMartin Matuska 
156eda14cbcSMatt Macy static int
157eda14cbcSMatt Macy zil_bp_compare(const void *x1, const void *x2)
158eda14cbcSMatt Macy {
159eda14cbcSMatt Macy 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
160eda14cbcSMatt Macy 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
161eda14cbcSMatt Macy 
162eda14cbcSMatt Macy 	int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
163eda14cbcSMatt Macy 	if (likely(cmp))
164eda14cbcSMatt Macy 		return (cmp);
165eda14cbcSMatt Macy 
166eda14cbcSMatt Macy 	return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
167eda14cbcSMatt Macy }
168eda14cbcSMatt Macy 
169eda14cbcSMatt Macy static void
170eda14cbcSMatt Macy zil_bp_tree_init(zilog_t *zilog)
171eda14cbcSMatt Macy {
172eda14cbcSMatt Macy 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
173eda14cbcSMatt Macy 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
174eda14cbcSMatt Macy }
175eda14cbcSMatt Macy 
176eda14cbcSMatt Macy static void
177eda14cbcSMatt Macy zil_bp_tree_fini(zilog_t *zilog)
178eda14cbcSMatt Macy {
179eda14cbcSMatt Macy 	avl_tree_t *t = &zilog->zl_bp_tree;
180eda14cbcSMatt Macy 	zil_bp_node_t *zn;
181eda14cbcSMatt Macy 	void *cookie = NULL;
182eda14cbcSMatt Macy 
183eda14cbcSMatt Macy 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
184eda14cbcSMatt Macy 		kmem_free(zn, sizeof (zil_bp_node_t));
185eda14cbcSMatt Macy 
186eda14cbcSMatt Macy 	avl_destroy(t);
187eda14cbcSMatt Macy }
188eda14cbcSMatt Macy 
189eda14cbcSMatt Macy int
190eda14cbcSMatt Macy zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
191eda14cbcSMatt Macy {
192eda14cbcSMatt Macy 	avl_tree_t *t = &zilog->zl_bp_tree;
193eda14cbcSMatt Macy 	const dva_t *dva;
194eda14cbcSMatt Macy 	zil_bp_node_t *zn;
195eda14cbcSMatt Macy 	avl_index_t where;
196eda14cbcSMatt Macy 
197eda14cbcSMatt Macy 	if (BP_IS_EMBEDDED(bp))
198eda14cbcSMatt Macy 		return (0);
199eda14cbcSMatt Macy 
200eda14cbcSMatt Macy 	dva = BP_IDENTITY(bp);
201eda14cbcSMatt Macy 
202eda14cbcSMatt Macy 	if (avl_find(t, dva, &where) != NULL)
203eda14cbcSMatt Macy 		return (SET_ERROR(EEXIST));
204eda14cbcSMatt Macy 
205eda14cbcSMatt Macy 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
206eda14cbcSMatt Macy 	zn->zn_dva = *dva;
207eda14cbcSMatt Macy 	avl_insert(t, zn, where);
208eda14cbcSMatt Macy 
209eda14cbcSMatt Macy 	return (0);
210eda14cbcSMatt Macy }
211eda14cbcSMatt Macy 
212eda14cbcSMatt Macy static zil_header_t *
213eda14cbcSMatt Macy zil_header_in_syncing_context(zilog_t *zilog)
214eda14cbcSMatt Macy {
215eda14cbcSMatt Macy 	return ((zil_header_t *)zilog->zl_header);
216eda14cbcSMatt Macy }
217eda14cbcSMatt Macy 
218eda14cbcSMatt Macy static void
219eda14cbcSMatt Macy zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
220eda14cbcSMatt Macy {
221eda14cbcSMatt Macy 	zio_cksum_t *zc = &bp->blk_cksum;
222eda14cbcSMatt Macy 
22333b8c039SMartin Matuska 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
22433b8c039SMartin Matuska 	    sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
22533b8c039SMartin Matuska 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
22633b8c039SMartin Matuska 	    sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
227eda14cbcSMatt Macy 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
228eda14cbcSMatt Macy 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
229eda14cbcSMatt Macy }
230eda14cbcSMatt Macy 
231271171e0SMartin Matuska static int
232271171e0SMartin Matuska zil_kstats_global_update(kstat_t *ksp, int rw)
233271171e0SMartin Matuska {
234271171e0SMartin Matuska 	zil_kstat_values_t *zs = ksp->ks_data;
235271171e0SMartin Matuska 	ASSERT3P(&zil_stats, ==, zs);
236271171e0SMartin Matuska 
237271171e0SMartin Matuska 	if (rw == KSTAT_WRITE) {
238271171e0SMartin Matuska 		return (SET_ERROR(EACCES));
239271171e0SMartin Matuska 	}
240271171e0SMartin Matuska 
241271171e0SMartin Matuska 	zil_kstat_values_update(zs, &zil_sums_global);
242271171e0SMartin Matuska 
243271171e0SMartin Matuska 	return (0);
244271171e0SMartin Matuska }
245271171e0SMartin Matuska 
246eda14cbcSMatt Macy /*
247eda14cbcSMatt Macy  * Read a log block and make sure it's valid.
248eda14cbcSMatt Macy  */
249eda14cbcSMatt Macy static int
250eda14cbcSMatt Macy zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
2514e8d558cSMartin Matuska     blkptr_t *nbp, char **begin, char **end, arc_buf_t **abuf)
252eda14cbcSMatt Macy {
253dbd5678dSMartin Matuska 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
254eda14cbcSMatt Macy 	arc_flags_t aflags = ARC_FLAG_WAIT;
255eda14cbcSMatt Macy 	zbookmark_phys_t zb;
256eda14cbcSMatt Macy 	int error;
257eda14cbcSMatt Macy 
258eda14cbcSMatt Macy 	if (zilog->zl_header->zh_claim_txg == 0)
259eda14cbcSMatt Macy 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
260eda14cbcSMatt Macy 
261eda14cbcSMatt Macy 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
262eda14cbcSMatt Macy 		zio_flags |= ZIO_FLAG_SPECULATIVE;
263eda14cbcSMatt Macy 
264eda14cbcSMatt Macy 	if (!decrypt)
265eda14cbcSMatt Macy 		zio_flags |= ZIO_FLAG_RAW;
266eda14cbcSMatt Macy 
267eda14cbcSMatt Macy 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
268eda14cbcSMatt Macy 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
269eda14cbcSMatt Macy 
270eda14cbcSMatt Macy 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
2714e8d558cSMartin Matuska 	    abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
272eda14cbcSMatt Macy 
273eda14cbcSMatt Macy 	if (error == 0) {
274eda14cbcSMatt Macy 		zio_cksum_t cksum = bp->blk_cksum;
275eda14cbcSMatt Macy 
276eda14cbcSMatt Macy 		/*
277eda14cbcSMatt Macy 		 * Validate the checksummed log block.
278eda14cbcSMatt Macy 		 *
279eda14cbcSMatt Macy 		 * Sequence numbers should be... sequential.  The checksum
280eda14cbcSMatt Macy 		 * verifier for the next block should be bp's checksum plus 1.
281eda14cbcSMatt Macy 		 *
282eda14cbcSMatt Macy 		 * Also check the log chain linkage and size used.
283eda14cbcSMatt Macy 		 */
284eda14cbcSMatt Macy 		cksum.zc_word[ZIL_ZC_SEQ]++;
285eda14cbcSMatt Macy 
2864e8d558cSMartin Matuska 		uint64_t size = BP_GET_LSIZE(bp);
287eda14cbcSMatt Macy 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
2884e8d558cSMartin Matuska 			zil_chain_t *zilc = (*abuf)->b_data;
289eda14cbcSMatt Macy 			char *lr = (char *)(zilc + 1);
290eda14cbcSMatt Macy 
291da5137abSMartin Matuska 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
292315ee00fSMartin Matuska 			    sizeof (cksum)) ||
2934e8d558cSMartin Matuska 			    zilc->zc_nused < sizeof (*zilc) ||
2944e8d558cSMartin Matuska 			    zilc->zc_nused > size) {
295eda14cbcSMatt Macy 				error = SET_ERROR(ECKSUM);
296eda14cbcSMatt Macy 			} else {
2974e8d558cSMartin Matuska 				*begin = lr;
2984e8d558cSMartin Matuska 				*end = lr + zilc->zc_nused - sizeof (*zilc);
299eda14cbcSMatt Macy 				*nbp = zilc->zc_next_blk;
300eda14cbcSMatt Macy 			}
301eda14cbcSMatt Macy 		} else {
3024e8d558cSMartin Matuska 			char *lr = (*abuf)->b_data;
303eda14cbcSMatt Macy 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
304eda14cbcSMatt Macy 
305da5137abSMartin Matuska 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
306315ee00fSMartin Matuska 			    sizeof (cksum)) ||
307eda14cbcSMatt Macy 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
308eda14cbcSMatt Macy 				error = SET_ERROR(ECKSUM);
309eda14cbcSMatt Macy 			} else {
3104e8d558cSMartin Matuska 				*begin = lr;
3114e8d558cSMartin Matuska 				*end = lr + zilc->zc_nused;
312eda14cbcSMatt Macy 				*nbp = zilc->zc_next_blk;
313eda14cbcSMatt Macy 			}
314eda14cbcSMatt Macy 		}
315eda14cbcSMatt Macy 	}
316eda14cbcSMatt Macy 
317eda14cbcSMatt Macy 	return (error);
318eda14cbcSMatt Macy }
319eda14cbcSMatt Macy 
320eda14cbcSMatt Macy /*
321eda14cbcSMatt Macy  * Read a TX_WRITE log data block.
322eda14cbcSMatt Macy  */
323eda14cbcSMatt Macy static int
324eda14cbcSMatt Macy zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
325eda14cbcSMatt Macy {
326dbd5678dSMartin Matuska 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
327eda14cbcSMatt Macy 	const blkptr_t *bp = &lr->lr_blkptr;
328eda14cbcSMatt Macy 	arc_flags_t aflags = ARC_FLAG_WAIT;
329eda14cbcSMatt Macy 	arc_buf_t *abuf = NULL;
330eda14cbcSMatt Macy 	zbookmark_phys_t zb;
331eda14cbcSMatt Macy 	int error;
332eda14cbcSMatt Macy 
333eda14cbcSMatt Macy 	if (BP_IS_HOLE(bp)) {
334eda14cbcSMatt Macy 		if (wbuf != NULL)
335da5137abSMartin Matuska 			memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length));
336eda14cbcSMatt Macy 		return (0);
337eda14cbcSMatt Macy 	}
338eda14cbcSMatt Macy 
339eda14cbcSMatt Macy 	if (zilog->zl_header->zh_claim_txg == 0)
340eda14cbcSMatt Macy 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
341eda14cbcSMatt Macy 
342eda14cbcSMatt Macy 	/*
343eda14cbcSMatt Macy 	 * If we are not using the resulting data, we are just checking that
344eda14cbcSMatt Macy 	 * it hasn't been corrupted so we don't need to waste CPU time
345eda14cbcSMatt Macy 	 * decompressing and decrypting it.
346eda14cbcSMatt Macy 	 */
347eda14cbcSMatt Macy 	if (wbuf == NULL)
348eda14cbcSMatt Macy 		zio_flags |= ZIO_FLAG_RAW;
349eda14cbcSMatt Macy 
350dbd5678dSMartin Matuska 	ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
351eda14cbcSMatt Macy 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
352eda14cbcSMatt Macy 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
353eda14cbcSMatt Macy 
354eda14cbcSMatt Macy 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
355eda14cbcSMatt Macy 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
356eda14cbcSMatt Macy 
357eda14cbcSMatt Macy 	if (error == 0) {
358eda14cbcSMatt Macy 		if (wbuf != NULL)
359da5137abSMartin Matuska 			memcpy(wbuf, abuf->b_data, arc_buf_size(abuf));
360eda14cbcSMatt Macy 		arc_buf_destroy(abuf, &abuf);
361eda14cbcSMatt Macy 	}
362eda14cbcSMatt Macy 
363eda14cbcSMatt Macy 	return (error);
364eda14cbcSMatt Macy }
365eda14cbcSMatt Macy 
366271171e0SMartin Matuska void
367271171e0SMartin Matuska zil_sums_init(zil_sums_t *zs)
368271171e0SMartin Matuska {
369271171e0SMartin Matuska 	wmsum_init(&zs->zil_commit_count, 0);
370271171e0SMartin Matuska 	wmsum_init(&zs->zil_commit_writer_count, 0);
371271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_count, 0);
372271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_indirect_count, 0);
373271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_indirect_bytes, 0);
374271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_copied_count, 0);
375271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_copied_bytes, 0);
376271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_needcopy_count, 0);
377271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_needcopy_bytes, 0);
378271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_normal_count, 0);
379271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0);
3804e8d558cSMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_normal_write, 0);
3814e8d558cSMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_normal_alloc, 0);
382271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_slog_count, 0);
383271171e0SMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0);
3844e8d558cSMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_slog_write, 0);
3854e8d558cSMartin Matuska 	wmsum_init(&zs->zil_itx_metaslab_slog_alloc, 0);
386271171e0SMartin Matuska }
387271171e0SMartin Matuska 
388271171e0SMartin Matuska void
389271171e0SMartin Matuska zil_sums_fini(zil_sums_t *zs)
390271171e0SMartin Matuska {
391271171e0SMartin Matuska 	wmsum_fini(&zs->zil_commit_count);
392271171e0SMartin Matuska 	wmsum_fini(&zs->zil_commit_writer_count);
393271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_count);
394271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_indirect_count);
395271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_indirect_bytes);
396271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_copied_count);
397271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_copied_bytes);
398271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_needcopy_count);
399271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_needcopy_bytes);
400271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_normal_count);
401271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_normal_bytes);
4024e8d558cSMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_normal_write);
4034e8d558cSMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_normal_alloc);
404271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_slog_count);
405271171e0SMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_slog_bytes);
4064e8d558cSMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_slog_write);
4074e8d558cSMartin Matuska 	wmsum_fini(&zs->zil_itx_metaslab_slog_alloc);
408271171e0SMartin Matuska }
409271171e0SMartin Matuska 
410271171e0SMartin Matuska void
411271171e0SMartin Matuska zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums)
412271171e0SMartin Matuska {
413271171e0SMartin Matuska 	zs->zil_commit_count.value.ui64 =
414271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_commit_count);
415271171e0SMartin Matuska 	zs->zil_commit_writer_count.value.ui64 =
416271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_commit_writer_count);
417271171e0SMartin Matuska 	zs->zil_itx_count.value.ui64 =
418271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_count);
419271171e0SMartin Matuska 	zs->zil_itx_indirect_count.value.ui64 =
420271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_indirect_count);
421271171e0SMartin Matuska 	zs->zil_itx_indirect_bytes.value.ui64 =
422271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_indirect_bytes);
423271171e0SMartin Matuska 	zs->zil_itx_copied_count.value.ui64 =
424271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_copied_count);
425271171e0SMartin Matuska 	zs->zil_itx_copied_bytes.value.ui64 =
426271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_copied_bytes);
427271171e0SMartin Matuska 	zs->zil_itx_needcopy_count.value.ui64 =
428271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_needcopy_count);
429271171e0SMartin Matuska 	zs->zil_itx_needcopy_bytes.value.ui64 =
430271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_needcopy_bytes);
431271171e0SMartin Matuska 	zs->zil_itx_metaslab_normal_count.value.ui64 =
432271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_count);
433271171e0SMartin Matuska 	zs->zil_itx_metaslab_normal_bytes.value.ui64 =
434271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes);
4354e8d558cSMartin Matuska 	zs->zil_itx_metaslab_normal_write.value.ui64 =
4364e8d558cSMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_write);
4374e8d558cSMartin Matuska 	zs->zil_itx_metaslab_normal_alloc.value.ui64 =
4384e8d558cSMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_alloc);
439271171e0SMartin Matuska 	zs->zil_itx_metaslab_slog_count.value.ui64 =
440271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_count);
441271171e0SMartin Matuska 	zs->zil_itx_metaslab_slog_bytes.value.ui64 =
442271171e0SMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes);
4434e8d558cSMartin Matuska 	zs->zil_itx_metaslab_slog_write.value.ui64 =
4444e8d558cSMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_write);
4454e8d558cSMartin Matuska 	zs->zil_itx_metaslab_slog_alloc.value.ui64 =
4464e8d558cSMartin Matuska 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_alloc);
447271171e0SMartin Matuska }
448271171e0SMartin Matuska 
449eda14cbcSMatt Macy /*
450eda14cbcSMatt Macy  * Parse the intent log, and call parse_func for each valid record within.
451eda14cbcSMatt Macy  */
452eda14cbcSMatt Macy int
453eda14cbcSMatt Macy zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
454eda14cbcSMatt Macy     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
455eda14cbcSMatt Macy     boolean_t decrypt)
456eda14cbcSMatt Macy {
457eda14cbcSMatt Macy 	const zil_header_t *zh = zilog->zl_header;
458eda14cbcSMatt Macy 	boolean_t claimed = !!zh->zh_claim_txg;
459eda14cbcSMatt Macy 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
460eda14cbcSMatt Macy 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
461eda14cbcSMatt Macy 	uint64_t max_blk_seq = 0;
462eda14cbcSMatt Macy 	uint64_t max_lr_seq = 0;
463eda14cbcSMatt Macy 	uint64_t blk_count = 0;
464eda14cbcSMatt Macy 	uint64_t lr_count = 0;
465da5137abSMartin Matuska 	blkptr_t blk, next_blk = {{{{0}}}};
466eda14cbcSMatt Macy 	int error = 0;
467eda14cbcSMatt Macy 
468eda14cbcSMatt Macy 	/*
469eda14cbcSMatt Macy 	 * Old logs didn't record the maximum zh_claim_lr_seq.
470eda14cbcSMatt Macy 	 */
471eda14cbcSMatt Macy 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
472eda14cbcSMatt Macy 		claim_lr_seq = UINT64_MAX;
473eda14cbcSMatt Macy 
474eda14cbcSMatt Macy 	/*
475eda14cbcSMatt Macy 	 * Starting at the block pointed to by zh_log we read the log chain.
476eda14cbcSMatt Macy 	 * For each block in the chain we strongly check that block to
477eda14cbcSMatt Macy 	 * ensure its validity.  We stop when an invalid block is found.
478eda14cbcSMatt Macy 	 * For each block pointer in the chain we call parse_blk_func().
479eda14cbcSMatt Macy 	 * For each record in each valid block we call parse_lr_func().
480eda14cbcSMatt Macy 	 * If the log has been claimed, stop if we encounter a sequence
481eda14cbcSMatt Macy 	 * number greater than the highest claimed sequence number.
482eda14cbcSMatt Macy 	 */
483eda14cbcSMatt Macy 	zil_bp_tree_init(zilog);
484eda14cbcSMatt Macy 
485eda14cbcSMatt Macy 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
486eda14cbcSMatt Macy 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
487eda14cbcSMatt Macy 		int reclen;
4884e8d558cSMartin Matuska 		char *lrp, *end;
4894e8d558cSMartin Matuska 		arc_buf_t *abuf = NULL;
490eda14cbcSMatt Macy 
491eda14cbcSMatt Macy 		if (blk_seq > claim_blk_seq)
492eda14cbcSMatt Macy 			break;
493eda14cbcSMatt Macy 
494eda14cbcSMatt Macy 		error = parse_blk_func(zilog, &blk, arg, txg);
495eda14cbcSMatt Macy 		if (error != 0)
496eda14cbcSMatt Macy 			break;
497eda14cbcSMatt Macy 		ASSERT3U(max_blk_seq, <, blk_seq);
498eda14cbcSMatt Macy 		max_blk_seq = blk_seq;
499eda14cbcSMatt Macy 		blk_count++;
500eda14cbcSMatt Macy 
501eda14cbcSMatt Macy 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
502eda14cbcSMatt Macy 			break;
503eda14cbcSMatt Macy 
504eda14cbcSMatt Macy 		error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
5054e8d558cSMartin Matuska 		    &lrp, &end, &abuf);
506dbd5678dSMartin Matuska 		if (error != 0) {
5074e8d558cSMartin Matuska 			if (abuf)
5084e8d558cSMartin Matuska 				arc_buf_destroy(abuf, &abuf);
509dbd5678dSMartin Matuska 			if (claimed) {
510dbd5678dSMartin Matuska 				char name[ZFS_MAX_DATASET_NAME_LEN];
511dbd5678dSMartin Matuska 
512dbd5678dSMartin Matuska 				dmu_objset_name(zilog->zl_os, name);
513dbd5678dSMartin Matuska 
514dbd5678dSMartin Matuska 				cmn_err(CE_WARN, "ZFS read log block error %d, "
515dbd5678dSMartin Matuska 				    "dataset %s, seq 0x%llx\n", error, name,
516dbd5678dSMartin Matuska 				    (u_longlong_t)blk_seq);
517dbd5678dSMartin Matuska 			}
518eda14cbcSMatt Macy 			break;
519dbd5678dSMartin Matuska 		}
520eda14cbcSMatt Macy 
5214e8d558cSMartin Matuska 		for (; lrp < end; lrp += reclen) {
522eda14cbcSMatt Macy 			lr_t *lr = (lr_t *)lrp;
523eda14cbcSMatt Macy 			reclen = lr->lrc_reclen;
524eda14cbcSMatt Macy 			ASSERT3U(reclen, >=, sizeof (lr_t));
525f190c36bSMartin Matuska 			if (lr->lrc_seq > claim_lr_seq) {
526f190c36bSMartin Matuska 				arc_buf_destroy(abuf, &abuf);
527eda14cbcSMatt Macy 				goto done;
528f190c36bSMartin Matuska 			}
529eda14cbcSMatt Macy 
530eda14cbcSMatt Macy 			error = parse_lr_func(zilog, lr, arg, txg);
531f190c36bSMartin Matuska 			if (error != 0) {
532f190c36bSMartin Matuska 				arc_buf_destroy(abuf, &abuf);
533eda14cbcSMatt Macy 				goto done;
534f190c36bSMartin Matuska 			}
535eda14cbcSMatt Macy 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
536eda14cbcSMatt Macy 			max_lr_seq = lr->lrc_seq;
537eda14cbcSMatt Macy 			lr_count++;
538eda14cbcSMatt Macy 		}
5394e8d558cSMartin Matuska 		arc_buf_destroy(abuf, &abuf);
540eda14cbcSMatt Macy 	}
541eda14cbcSMatt Macy done:
542eda14cbcSMatt Macy 	zilog->zl_parse_error = error;
543eda14cbcSMatt Macy 	zilog->zl_parse_blk_seq = max_blk_seq;
544eda14cbcSMatt Macy 	zilog->zl_parse_lr_seq = max_lr_seq;
545eda14cbcSMatt Macy 	zilog->zl_parse_blk_count = blk_count;
546eda14cbcSMatt Macy 	zilog->zl_parse_lr_count = lr_count;
547eda14cbcSMatt Macy 
548eda14cbcSMatt Macy 	zil_bp_tree_fini(zilog);
549eda14cbcSMatt Macy 
550eda14cbcSMatt Macy 	return (error);
551eda14cbcSMatt Macy }
552eda14cbcSMatt Macy 
553eda14cbcSMatt Macy static int
554180f8225SMatt Macy zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
555180f8225SMatt Macy     uint64_t first_txg)
556eda14cbcSMatt Macy {
557e92ffd9bSMartin Matuska 	(void) tx;
558eda14cbcSMatt Macy 	ASSERT(!BP_IS_HOLE(bp));
559eda14cbcSMatt Macy 
560eda14cbcSMatt Macy 	/*
561eda14cbcSMatt Macy 	 * As we call this function from the context of a rewind to a
562eda14cbcSMatt Macy 	 * checkpoint, each ZIL block whose txg is later than the txg
563eda14cbcSMatt Macy 	 * that we rewind to is invalid. Thus, we return -1 so
564eda14cbcSMatt Macy 	 * zil_parse() doesn't attempt to read it.
565eda14cbcSMatt Macy 	 */
566eda14cbcSMatt Macy 	if (bp->blk_birth >= first_txg)
567eda14cbcSMatt Macy 		return (-1);
568eda14cbcSMatt Macy 
569eda14cbcSMatt Macy 	if (zil_bp_tree_add(zilog, bp) != 0)
570eda14cbcSMatt Macy 		return (0);
571eda14cbcSMatt Macy 
572eda14cbcSMatt Macy 	zio_free(zilog->zl_spa, first_txg, bp);
573eda14cbcSMatt Macy 	return (0);
574eda14cbcSMatt Macy }
575eda14cbcSMatt Macy 
576eda14cbcSMatt Macy static int
577180f8225SMatt Macy zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
578180f8225SMatt Macy     uint64_t first_txg)
579eda14cbcSMatt Macy {
580e92ffd9bSMartin Matuska 	(void) zilog, (void) lrc, (void) tx, (void) first_txg;
581eda14cbcSMatt Macy 	return (0);
582eda14cbcSMatt Macy }
583eda14cbcSMatt Macy 
584eda14cbcSMatt Macy static int
585180f8225SMatt Macy zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
586180f8225SMatt Macy     uint64_t first_txg)
587eda14cbcSMatt Macy {
588eda14cbcSMatt Macy 	/*
589eda14cbcSMatt Macy 	 * Claim log block if not already committed and not already claimed.
590eda14cbcSMatt Macy 	 * If tx == NULL, just verify that the block is claimable.
591eda14cbcSMatt Macy 	 */
592eda14cbcSMatt Macy 	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
593eda14cbcSMatt Macy 	    zil_bp_tree_add(zilog, bp) != 0)
594eda14cbcSMatt Macy 		return (0);
595eda14cbcSMatt Macy 
596eda14cbcSMatt Macy 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
597eda14cbcSMatt Macy 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
598eda14cbcSMatt Macy 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
599eda14cbcSMatt Macy }
600eda14cbcSMatt Macy 
601eda14cbcSMatt Macy static int
6022a58b312SMartin Matuska zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg)
603eda14cbcSMatt Macy {
604eda14cbcSMatt Macy 	lr_write_t *lr = (lr_write_t *)lrc;
605eda14cbcSMatt Macy 	int error;
606eda14cbcSMatt Macy 
6072a58b312SMartin Matuska 	ASSERT(lrc->lrc_txtype == TX_WRITE);
608eda14cbcSMatt Macy 
609eda14cbcSMatt Macy 	/*
610eda14cbcSMatt Macy 	 * If the block is not readable, don't claim it.  This can happen
611eda14cbcSMatt Macy 	 * in normal operation when a log block is written to disk before
612eda14cbcSMatt Macy 	 * some of the dmu_sync() blocks it points to.  In this case, the
613eda14cbcSMatt Macy 	 * transaction cannot have been committed to anyone (we would have
614eda14cbcSMatt Macy 	 * waited for all writes to be stable first), so it is semantically
615eda14cbcSMatt Macy 	 * correct to declare this the end of the log.
616eda14cbcSMatt Macy 	 */
617eda14cbcSMatt Macy 	if (lr->lr_blkptr.blk_birth >= first_txg) {
618eda14cbcSMatt Macy 		error = zil_read_log_data(zilog, lr, NULL);
619eda14cbcSMatt Macy 		if (error != 0)
620eda14cbcSMatt Macy 			return (error);
621eda14cbcSMatt Macy 	}
622eda14cbcSMatt Macy 
623eda14cbcSMatt Macy 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
624eda14cbcSMatt Macy }
625eda14cbcSMatt Macy 
626eda14cbcSMatt Macy static int
6272a58b312SMartin Matuska zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
6282a58b312SMartin Matuska {
6292a58b312SMartin Matuska 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
6302a58b312SMartin Matuska 	const blkptr_t *bp;
6312a58b312SMartin Matuska 	spa_t *spa;
6322a58b312SMartin Matuska 	uint_t ii;
6332a58b312SMartin Matuska 
6342a58b312SMartin Matuska 	ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE);
6352a58b312SMartin Matuska 
6362a58b312SMartin Matuska 	if (tx == NULL) {
6372a58b312SMartin Matuska 		return (0);
6382a58b312SMartin Matuska 	}
6392a58b312SMartin Matuska 
6402a58b312SMartin Matuska 	/*
6412a58b312SMartin Matuska 	 * XXX: Do we need to byteswap lr?
6422a58b312SMartin Matuska 	 */
6432a58b312SMartin Matuska 
6442a58b312SMartin Matuska 	spa = zilog->zl_spa;
6452a58b312SMartin Matuska 
6462a58b312SMartin Matuska 	for (ii = 0; ii < lr->lr_nbps; ii++) {
6472a58b312SMartin Matuska 		bp = &lr->lr_bps[ii];
6482a58b312SMartin Matuska 
6492a58b312SMartin Matuska 		/*
6502a58b312SMartin Matuska 		 * When data in embedded into BP there is no need to create
6512a58b312SMartin Matuska 		 * BRT entry as there is no data block. Just copy the BP as
6522a58b312SMartin Matuska 		 * it contains the data.
6532a58b312SMartin Matuska 		 */
6542a58b312SMartin Matuska 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
6552a58b312SMartin Matuska 			brt_pending_add(spa, bp, tx);
6562a58b312SMartin Matuska 		}
6572a58b312SMartin Matuska 	}
6582a58b312SMartin Matuska 
6592a58b312SMartin Matuska 	return (0);
6602a58b312SMartin Matuska }
6612a58b312SMartin Matuska 
6622a58b312SMartin Matuska static int
6632a58b312SMartin Matuska zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
6642a58b312SMartin Matuska     uint64_t first_txg)
6652a58b312SMartin Matuska {
6662a58b312SMartin Matuska 
6672a58b312SMartin Matuska 	switch (lrc->lrc_txtype) {
6682a58b312SMartin Matuska 	case TX_WRITE:
6692a58b312SMartin Matuska 		return (zil_claim_write(zilog, lrc, tx, first_txg));
6702a58b312SMartin Matuska 	case TX_CLONE_RANGE:
6712a58b312SMartin Matuska 		return (zil_claim_clone_range(zilog, lrc, tx));
6722a58b312SMartin Matuska 	default:
6732a58b312SMartin Matuska 		return (0);
6742a58b312SMartin Matuska 	}
6752a58b312SMartin Matuska }
6762a58b312SMartin Matuska 
6772a58b312SMartin Matuska static int
678180f8225SMatt Macy zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
679180f8225SMatt Macy     uint64_t claim_txg)
680eda14cbcSMatt Macy {
681e92ffd9bSMartin Matuska 	(void) claim_txg;
682e92ffd9bSMartin Matuska 
683eda14cbcSMatt Macy 	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
684eda14cbcSMatt Macy 
685eda14cbcSMatt Macy 	return (0);
686eda14cbcSMatt Macy }
687eda14cbcSMatt Macy 
688eda14cbcSMatt Macy static int
6892a58b312SMartin Matuska zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg)
690eda14cbcSMatt Macy {
691eda14cbcSMatt Macy 	lr_write_t *lr = (lr_write_t *)lrc;
692eda14cbcSMatt Macy 	blkptr_t *bp = &lr->lr_blkptr;
693eda14cbcSMatt Macy 
6942a58b312SMartin Matuska 	ASSERT(lrc->lrc_txtype == TX_WRITE);
6952a58b312SMartin Matuska 
696eda14cbcSMatt Macy 	/*
697eda14cbcSMatt Macy 	 * If we previously claimed it, we need to free it.
698eda14cbcSMatt Macy 	 */
6992a58b312SMartin Matuska 	if (bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
7002a58b312SMartin Matuska 	    !BP_IS_HOLE(bp)) {
701eda14cbcSMatt Macy 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
7022a58b312SMartin Matuska 	}
703eda14cbcSMatt Macy 
704eda14cbcSMatt Macy 	return (0);
705eda14cbcSMatt Macy }
706eda14cbcSMatt Macy 
707eda14cbcSMatt Macy static int
7082a58b312SMartin Matuska zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
7092a58b312SMartin Matuska {
7102a58b312SMartin Matuska 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
7112a58b312SMartin Matuska 	const blkptr_t *bp;
7122a58b312SMartin Matuska 	spa_t *spa;
7132a58b312SMartin Matuska 	uint_t ii;
7142a58b312SMartin Matuska 
7152a58b312SMartin Matuska 	ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE);
7162a58b312SMartin Matuska 
7172a58b312SMartin Matuska 	if (tx == NULL) {
7182a58b312SMartin Matuska 		return (0);
7192a58b312SMartin Matuska 	}
7202a58b312SMartin Matuska 
7212a58b312SMartin Matuska 	spa = zilog->zl_spa;
7222a58b312SMartin Matuska 
7232a58b312SMartin Matuska 	for (ii = 0; ii < lr->lr_nbps; ii++) {
7242a58b312SMartin Matuska 		bp = &lr->lr_bps[ii];
7252a58b312SMartin Matuska 
7262a58b312SMartin Matuska 		if (!BP_IS_HOLE(bp)) {
7272a58b312SMartin Matuska 			zio_free(spa, dmu_tx_get_txg(tx), bp);
7282a58b312SMartin Matuska 		}
7292a58b312SMartin Matuska 	}
7302a58b312SMartin Matuska 
7312a58b312SMartin Matuska 	return (0);
7322a58b312SMartin Matuska }
7332a58b312SMartin Matuska 
7342a58b312SMartin Matuska static int
7352a58b312SMartin Matuska zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
7362a58b312SMartin Matuska     uint64_t claim_txg)
7372a58b312SMartin Matuska {
7382a58b312SMartin Matuska 
7392a58b312SMartin Matuska 	if (claim_txg == 0) {
7402a58b312SMartin Matuska 		return (0);
7412a58b312SMartin Matuska 	}
7422a58b312SMartin Matuska 
7432a58b312SMartin Matuska 	switch (lrc->lrc_txtype) {
7442a58b312SMartin Matuska 	case TX_WRITE:
7452a58b312SMartin Matuska 		return (zil_free_write(zilog, lrc, tx, claim_txg));
7462a58b312SMartin Matuska 	case TX_CLONE_RANGE:
7472a58b312SMartin Matuska 		return (zil_free_clone_range(zilog, lrc, tx));
7482a58b312SMartin Matuska 	default:
7492a58b312SMartin Matuska 		return (0);
7502a58b312SMartin Matuska 	}
7512a58b312SMartin Matuska }
7522a58b312SMartin Matuska 
7532a58b312SMartin Matuska static int
754eda14cbcSMatt Macy zil_lwb_vdev_compare(const void *x1, const void *x2)
755eda14cbcSMatt Macy {
756eda14cbcSMatt Macy 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
757eda14cbcSMatt Macy 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
758eda14cbcSMatt Macy 
759eda14cbcSMatt Macy 	return (TREE_CMP(v1, v2));
760eda14cbcSMatt Macy }
761eda14cbcSMatt Macy 
762315ee00fSMartin Matuska /*
763315ee00fSMartin Matuska  * Allocate a new lwb.  We may already have a block pointer for it, in which
764315ee00fSMartin Matuska  * case we get size and version from there.  Or we may not yet, in which case
765315ee00fSMartin Matuska  * we choose them here and later make the block allocation match.
766315ee00fSMartin Matuska  */
767eda14cbcSMatt Macy static lwb_t *
768315ee00fSMartin Matuska zil_alloc_lwb(zilog_t *zilog, int sz, blkptr_t *bp, boolean_t slog,
769315ee00fSMartin Matuska     uint64_t txg, lwb_state_t state)
770eda14cbcSMatt Macy {
771eda14cbcSMatt Macy 	lwb_t *lwb;
772eda14cbcSMatt Macy 
773eda14cbcSMatt Macy 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
774eda14cbcSMatt Macy 	lwb->lwb_zilog = zilog;
775315ee00fSMartin Matuska 	if (bp) {
776eda14cbcSMatt Macy 		lwb->lwb_blk = *bp;
777315ee00fSMartin Matuska 		lwb->lwb_slim = (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2);
778315ee00fSMartin Matuska 		sz = BP_GET_LSIZE(bp);
7794e8d558cSMartin Matuska 	} else {
780315ee00fSMartin Matuska 		BP_ZERO(&lwb->lwb_blk);
781315ee00fSMartin Matuska 		lwb->lwb_slim = (spa_version(zilog->zl_spa) >=
782315ee00fSMartin Matuska 		    SPA_VERSION_SLIM_ZIL);
7834e8d558cSMartin Matuska 	}
784315ee00fSMartin Matuska 	lwb->lwb_slog = slog;
785315ee00fSMartin Matuska 	lwb->lwb_error = 0;
786315ee00fSMartin Matuska 	if (lwb->lwb_slim) {
787315ee00fSMartin Matuska 		lwb->lwb_nmax = sz;
788315ee00fSMartin Matuska 		lwb->lwb_nused = lwb->lwb_nfilled = sizeof (zil_chain_t);
789315ee00fSMartin Matuska 	} else {
790315ee00fSMartin Matuska 		lwb->lwb_nmax = sz - sizeof (zil_chain_t);
791315ee00fSMartin Matuska 		lwb->lwb_nused = lwb->lwb_nfilled = 0;
792315ee00fSMartin Matuska 	}
793315ee00fSMartin Matuska 	lwb->lwb_sz = sz;
794315ee00fSMartin Matuska 	lwb->lwb_state = state;
795315ee00fSMartin Matuska 	lwb->lwb_buf = zio_buf_alloc(sz);
796315ee00fSMartin Matuska 	lwb->lwb_child_zio = NULL;
797eda14cbcSMatt Macy 	lwb->lwb_write_zio = NULL;
798eda14cbcSMatt Macy 	lwb->lwb_root_zio = NULL;
799eda14cbcSMatt Macy 	lwb->lwb_issued_timestamp = 0;
800e3aa18adSMartin Matuska 	lwb->lwb_issued_txg = 0;
801315ee00fSMartin Matuska 	lwb->lwb_alloc_txg = txg;
802315ee00fSMartin Matuska 	lwb->lwb_max_txg = 0;
803eda14cbcSMatt Macy 
804eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
805eda14cbcSMatt Macy 	list_insert_tail(&zilog->zl_lwb_list, lwb);
806315ee00fSMartin Matuska 	if (state != LWB_STATE_NEW)
807315ee00fSMartin Matuska 		zilog->zl_last_lwb_opened = lwb;
808eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
809eda14cbcSMatt Macy 
810eda14cbcSMatt Macy 	return (lwb);
811eda14cbcSMatt Macy }
812eda14cbcSMatt Macy 
813eda14cbcSMatt Macy static void
814eda14cbcSMatt Macy zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
815eda14cbcSMatt Macy {
816eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
8172ad756a6SMartin Matuska 	ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
8182ad756a6SMartin Matuska 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
819315ee00fSMartin Matuska 	ASSERT3P(lwb->lwb_child_zio, ==, NULL);
820eda14cbcSMatt Macy 	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
821eda14cbcSMatt Macy 	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
822315ee00fSMartin Matuska 	ASSERT3U(lwb->lwb_alloc_txg, <=, spa_syncing_txg(zilog->zl_spa));
823eda14cbcSMatt Macy 	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
8242ad756a6SMartin Matuska 	VERIFY(list_is_empty(&lwb->lwb_itxs));
8252ad756a6SMartin Matuska 	VERIFY(list_is_empty(&lwb->lwb_waiters));
8262ad756a6SMartin Matuska 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
8272ad756a6SMartin Matuska 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
828eda14cbcSMatt Macy 
829eda14cbcSMatt Macy 	/*
830eda14cbcSMatt Macy 	 * Clear the zilog's field to indicate this lwb is no longer
831eda14cbcSMatt Macy 	 * valid, and prevent use-after-free errors.
832eda14cbcSMatt Macy 	 */
833eda14cbcSMatt Macy 	if (zilog->zl_last_lwb_opened == lwb)
834eda14cbcSMatt Macy 		zilog->zl_last_lwb_opened = NULL;
835eda14cbcSMatt Macy 
836eda14cbcSMatt Macy 	kmem_cache_free(zil_lwb_cache, lwb);
837eda14cbcSMatt Macy }
838eda14cbcSMatt Macy 
839eda14cbcSMatt Macy /*
840eda14cbcSMatt Macy  * Called when we create in-memory log transactions so that we know
841eda14cbcSMatt Macy  * to cleanup the itxs at the end of spa_sync().
842eda14cbcSMatt Macy  */
843eda14cbcSMatt Macy static void
844eda14cbcSMatt Macy zilog_dirty(zilog_t *zilog, uint64_t txg)
845eda14cbcSMatt Macy {
846eda14cbcSMatt Macy 	dsl_pool_t *dp = zilog->zl_dmu_pool;
847eda14cbcSMatt Macy 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
848eda14cbcSMatt Macy 
849eda14cbcSMatt Macy 	ASSERT(spa_writeable(zilog->zl_spa));
850eda14cbcSMatt Macy 
851eda14cbcSMatt Macy 	if (ds->ds_is_snapshot)
852eda14cbcSMatt Macy 		panic("dirtying snapshot!");
853eda14cbcSMatt Macy 
854eda14cbcSMatt Macy 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
855eda14cbcSMatt Macy 		/* up the hold count until we can be written out */
856eda14cbcSMatt Macy 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
857eda14cbcSMatt Macy 
858eda14cbcSMatt Macy 		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
859eda14cbcSMatt Macy 	}
860eda14cbcSMatt Macy }
861eda14cbcSMatt Macy 
862eda14cbcSMatt Macy /*
863eda14cbcSMatt Macy  * Determine if the zil is dirty in the specified txg. Callers wanting to
864eda14cbcSMatt Macy  * ensure that the dirty state does not change must hold the itxg_lock for
865eda14cbcSMatt Macy  * the specified txg. Holding the lock will ensure that the zil cannot be
866eda14cbcSMatt Macy  * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
867eda14cbcSMatt Macy  * state.
868eda14cbcSMatt Macy  */
869eda14cbcSMatt Macy static boolean_t __maybe_unused
870eda14cbcSMatt Macy zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
871eda14cbcSMatt Macy {
872eda14cbcSMatt Macy 	dsl_pool_t *dp = zilog->zl_dmu_pool;
873eda14cbcSMatt Macy 
874eda14cbcSMatt Macy 	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
875eda14cbcSMatt Macy 		return (B_TRUE);
876eda14cbcSMatt Macy 	return (B_FALSE);
877eda14cbcSMatt Macy }
878eda14cbcSMatt Macy 
879eda14cbcSMatt Macy /*
880eda14cbcSMatt Macy  * Determine if the zil is dirty. The zil is considered dirty if it has
881eda14cbcSMatt Macy  * any pending itx records that have not been cleaned by zil_clean().
882eda14cbcSMatt Macy  */
883eda14cbcSMatt Macy static boolean_t
884eda14cbcSMatt Macy zilog_is_dirty(zilog_t *zilog)
885eda14cbcSMatt Macy {
886eda14cbcSMatt Macy 	dsl_pool_t *dp = zilog->zl_dmu_pool;
887eda14cbcSMatt Macy 
888eda14cbcSMatt Macy 	for (int t = 0; t < TXG_SIZE; t++) {
889eda14cbcSMatt Macy 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
890eda14cbcSMatt Macy 			return (B_TRUE);
891eda14cbcSMatt Macy 	}
892eda14cbcSMatt Macy 	return (B_FALSE);
893eda14cbcSMatt Macy }
894eda14cbcSMatt Macy 
895eda14cbcSMatt Macy /*
896c03c5b1cSMartin Matuska  * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
897c03c5b1cSMartin Matuska  * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
898c03c5b1cSMartin Matuska  * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
899c03c5b1cSMartin Matuska  * zil_commit.
900c03c5b1cSMartin Matuska  */
901c03c5b1cSMartin Matuska static void
902c03c5b1cSMartin Matuska zil_commit_activate_saxattr_feature(zilog_t *zilog)
903c03c5b1cSMartin Matuska {
904c03c5b1cSMartin Matuska 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
905c03c5b1cSMartin Matuska 	uint64_t txg = 0;
906c03c5b1cSMartin Matuska 	dmu_tx_t *tx = NULL;
907c03c5b1cSMartin Matuska 
908dbd5678dSMartin Matuska 	if (spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
909c03c5b1cSMartin Matuska 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL &&
910dbd5678dSMartin Matuska 	    !dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)) {
911c03c5b1cSMartin Matuska 		tx = dmu_tx_create(zilog->zl_os);
912c03c5b1cSMartin Matuska 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
913c03c5b1cSMartin Matuska 		dsl_dataset_dirty(ds, tx);
914c03c5b1cSMartin Matuska 		txg = dmu_tx_get_txg(tx);
915c03c5b1cSMartin Matuska 
916c03c5b1cSMartin Matuska 		mutex_enter(&ds->ds_lock);
917c03c5b1cSMartin Matuska 		ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
918c03c5b1cSMartin Matuska 		    (void *)B_TRUE;
919c03c5b1cSMartin Matuska 		mutex_exit(&ds->ds_lock);
920c03c5b1cSMartin Matuska 		dmu_tx_commit(tx);
921c03c5b1cSMartin Matuska 		txg_wait_synced(zilog->zl_dmu_pool, txg);
922c03c5b1cSMartin Matuska 	}
923c03c5b1cSMartin Matuska }
924c03c5b1cSMartin Matuska 
925c03c5b1cSMartin Matuska /*
926eda14cbcSMatt Macy  * Create an on-disk intent log.
927eda14cbcSMatt Macy  */
928eda14cbcSMatt Macy static lwb_t *
929eda14cbcSMatt Macy zil_create(zilog_t *zilog)
930eda14cbcSMatt Macy {
931eda14cbcSMatt Macy 	const zil_header_t *zh = zilog->zl_header;
932eda14cbcSMatt Macy 	lwb_t *lwb = NULL;
933eda14cbcSMatt Macy 	uint64_t txg = 0;
934eda14cbcSMatt Macy 	dmu_tx_t *tx = NULL;
935eda14cbcSMatt Macy 	blkptr_t blk;
936eda14cbcSMatt Macy 	int error = 0;
937eda14cbcSMatt Macy 	boolean_t slog = FALSE;
938c03c5b1cSMartin Matuska 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
939c03c5b1cSMartin Matuska 
940eda14cbcSMatt Macy 
941eda14cbcSMatt Macy 	/*
942eda14cbcSMatt Macy 	 * Wait for any previous destroy to complete.
943eda14cbcSMatt Macy 	 */
944eda14cbcSMatt Macy 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
945eda14cbcSMatt Macy 
946eda14cbcSMatt Macy 	ASSERT(zh->zh_claim_txg == 0);
947eda14cbcSMatt Macy 	ASSERT(zh->zh_replay_seq == 0);
948eda14cbcSMatt Macy 
949eda14cbcSMatt Macy 	blk = zh->zh_log;
950eda14cbcSMatt Macy 
951eda14cbcSMatt Macy 	/*
952eda14cbcSMatt Macy 	 * Allocate an initial log block if:
953eda14cbcSMatt Macy 	 *    - there isn't one already
954eda14cbcSMatt Macy 	 *    - the existing block is the wrong endianness
955eda14cbcSMatt Macy 	 */
956eda14cbcSMatt Macy 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
957eda14cbcSMatt Macy 		tx = dmu_tx_create(zilog->zl_os);
958eda14cbcSMatt Macy 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
959eda14cbcSMatt Macy 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
960eda14cbcSMatt Macy 		txg = dmu_tx_get_txg(tx);
961eda14cbcSMatt Macy 
962eda14cbcSMatt Macy 		if (!BP_IS_HOLE(&blk)) {
963eda14cbcSMatt Macy 			zio_free(zilog->zl_spa, txg, &blk);
964eda14cbcSMatt Macy 			BP_ZERO(&blk);
965eda14cbcSMatt Macy 		}
966eda14cbcSMatt Macy 
967eda14cbcSMatt Macy 		error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
968eda14cbcSMatt Macy 		    ZIL_MIN_BLKSZ, &slog);
969eda14cbcSMatt Macy 		if (error == 0)
970eda14cbcSMatt Macy 			zil_init_log_chain(zilog, &blk);
971eda14cbcSMatt Macy 	}
972eda14cbcSMatt Macy 
973eda14cbcSMatt Macy 	/*
974eda14cbcSMatt Macy 	 * Allocate a log write block (lwb) for the first log block.
975eda14cbcSMatt Macy 	 */
976eda14cbcSMatt Macy 	if (error == 0)
977315ee00fSMartin Matuska 		lwb = zil_alloc_lwb(zilog, 0, &blk, slog, txg, LWB_STATE_NEW);
978eda14cbcSMatt Macy 
979eda14cbcSMatt Macy 	/*
980eda14cbcSMatt Macy 	 * If we just allocated the first log block, commit our transaction
981eda14cbcSMatt Macy 	 * and wait for zil_sync() to stuff the block pointer into zh_log.
982eda14cbcSMatt Macy 	 * (zh is part of the MOS, so we cannot modify it in open context.)
983eda14cbcSMatt Macy 	 */
984eda14cbcSMatt Macy 	if (tx != NULL) {
985c03c5b1cSMartin Matuska 		/*
986c03c5b1cSMartin Matuska 		 * If "zilsaxattr" feature is enabled on zpool, then activate
987c03c5b1cSMartin Matuska 		 * it now when we're creating the ZIL chain. We can't wait with
988c03c5b1cSMartin Matuska 		 * this until we write the first xattr log record because we
989c03c5b1cSMartin Matuska 		 * need to wait for the feature activation to sync out.
990c03c5b1cSMartin Matuska 		 */
991c03c5b1cSMartin Matuska 		if (spa_feature_is_enabled(zilog->zl_spa,
992c03c5b1cSMartin Matuska 		    SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) !=
993c03c5b1cSMartin Matuska 		    DMU_OST_ZVOL) {
994c03c5b1cSMartin Matuska 			mutex_enter(&ds->ds_lock);
995c03c5b1cSMartin Matuska 			ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
996c03c5b1cSMartin Matuska 			    (void *)B_TRUE;
997c03c5b1cSMartin Matuska 			mutex_exit(&ds->ds_lock);
998c03c5b1cSMartin Matuska 		}
999c03c5b1cSMartin Matuska 
1000eda14cbcSMatt Macy 		dmu_tx_commit(tx);
1001eda14cbcSMatt Macy 		txg_wait_synced(zilog->zl_dmu_pool, txg);
1002c03c5b1cSMartin Matuska 	} else {
1003c03c5b1cSMartin Matuska 		/*
1004c03c5b1cSMartin Matuska 		 * This branch covers the case where we enable the feature on a
1005c03c5b1cSMartin Matuska 		 * zpool that has existing ZIL headers.
1006c03c5b1cSMartin Matuska 		 */
1007c03c5b1cSMartin Matuska 		zil_commit_activate_saxattr_feature(zilog);
1008eda14cbcSMatt Macy 	}
1009c03c5b1cSMartin Matuska 	IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
1010c03c5b1cSMartin Matuska 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL,
1011c03c5b1cSMartin Matuska 	    dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR));
1012eda14cbcSMatt Macy 
1013da5137abSMartin Matuska 	ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
1014eda14cbcSMatt Macy 	IMPLY(error == 0, lwb != NULL);
1015eda14cbcSMatt Macy 
1016eda14cbcSMatt Macy 	return (lwb);
1017eda14cbcSMatt Macy }
1018eda14cbcSMatt Macy 
1019eda14cbcSMatt Macy /*
1020eda14cbcSMatt Macy  * In one tx, free all log blocks and clear the log header. If keep_first
1021eda14cbcSMatt Macy  * is set, then we're replaying a log with no content. We want to keep the
1022eda14cbcSMatt Macy  * first block, however, so that the first synchronous transaction doesn't
1023eda14cbcSMatt Macy  * require a txg_wait_synced() in zil_create(). We don't need to
1024eda14cbcSMatt Macy  * txg_wait_synced() here either when keep_first is set, because both
1025eda14cbcSMatt Macy  * zil_create() and zil_destroy() will wait for any in-progress destroys
1026eda14cbcSMatt Macy  * to complete.
1027dbd5678dSMartin Matuska  * Return B_TRUE if there were any entries to replay.
1028eda14cbcSMatt Macy  */
1029dbd5678dSMartin Matuska boolean_t
1030eda14cbcSMatt Macy zil_destroy(zilog_t *zilog, boolean_t keep_first)
1031eda14cbcSMatt Macy {
1032eda14cbcSMatt Macy 	const zil_header_t *zh = zilog->zl_header;
1033eda14cbcSMatt Macy 	lwb_t *lwb;
1034eda14cbcSMatt Macy 	dmu_tx_t *tx;
1035eda14cbcSMatt Macy 	uint64_t txg;
1036eda14cbcSMatt Macy 
1037eda14cbcSMatt Macy 	/*
1038eda14cbcSMatt Macy 	 * Wait for any previous destroy to complete.
1039eda14cbcSMatt Macy 	 */
1040eda14cbcSMatt Macy 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1041eda14cbcSMatt Macy 
1042eda14cbcSMatt Macy 	zilog->zl_old_header = *zh;		/* debugging aid */
1043eda14cbcSMatt Macy 
1044eda14cbcSMatt Macy 	if (BP_IS_HOLE(&zh->zh_log))
1045dbd5678dSMartin Matuska 		return (B_FALSE);
1046eda14cbcSMatt Macy 
1047eda14cbcSMatt Macy 	tx = dmu_tx_create(zilog->zl_os);
1048eda14cbcSMatt Macy 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1049eda14cbcSMatt Macy 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1050eda14cbcSMatt Macy 	txg = dmu_tx_get_txg(tx);
1051eda14cbcSMatt Macy 
1052eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
1053eda14cbcSMatt Macy 
1054eda14cbcSMatt Macy 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
1055eda14cbcSMatt Macy 	zilog->zl_destroy_txg = txg;
1056eda14cbcSMatt Macy 	zilog->zl_keep_first = keep_first;
1057eda14cbcSMatt Macy 
1058eda14cbcSMatt Macy 	if (!list_is_empty(&zilog->zl_lwb_list)) {
1059eda14cbcSMatt Macy 		ASSERT(zh->zh_claim_txg == 0);
1060eda14cbcSMatt Macy 		VERIFY(!keep_first);
1061c0a83fe0SMartin Matuska 		while ((lwb = list_remove_head(&zilog->zl_lwb_list)) != NULL) {
1062eda14cbcSMatt Macy 			if (lwb->lwb_buf != NULL)
1063eda14cbcSMatt Macy 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1064315ee00fSMartin Matuska 			if (!BP_IS_HOLE(&lwb->lwb_blk))
1065eda14cbcSMatt Macy 				zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
1066eda14cbcSMatt Macy 			zil_free_lwb(zilog, lwb);
1067eda14cbcSMatt Macy 		}
1068eda14cbcSMatt Macy 	} else if (!keep_first) {
1069eda14cbcSMatt Macy 		zil_destroy_sync(zilog, tx);
1070eda14cbcSMatt Macy 	}
1071eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
1072eda14cbcSMatt Macy 
1073eda14cbcSMatt Macy 	dmu_tx_commit(tx);
1074dbd5678dSMartin Matuska 
1075dbd5678dSMartin Matuska 	return (B_TRUE);
1076eda14cbcSMatt Macy }
1077eda14cbcSMatt Macy 
1078eda14cbcSMatt Macy void
1079eda14cbcSMatt Macy zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
1080eda14cbcSMatt Macy {
1081eda14cbcSMatt Macy 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1082eda14cbcSMatt Macy 	(void) zil_parse(zilog, zil_free_log_block,
1083eda14cbcSMatt Macy 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
1084eda14cbcSMatt Macy }
1085eda14cbcSMatt Macy 
1086eda14cbcSMatt Macy int
1087eda14cbcSMatt Macy zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
1088eda14cbcSMatt Macy {
1089eda14cbcSMatt Macy 	dmu_tx_t *tx = txarg;
1090eda14cbcSMatt Macy 	zilog_t *zilog;
1091eda14cbcSMatt Macy 	uint64_t first_txg;
1092eda14cbcSMatt Macy 	zil_header_t *zh;
1093eda14cbcSMatt Macy 	objset_t *os;
1094eda14cbcSMatt Macy 	int error;
1095eda14cbcSMatt Macy 
1096eda14cbcSMatt Macy 	error = dmu_objset_own_obj(dp, ds->ds_object,
1097eda14cbcSMatt Macy 	    DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
1098eda14cbcSMatt Macy 	if (error != 0) {
1099eda14cbcSMatt Macy 		/*
1100eda14cbcSMatt Macy 		 * EBUSY indicates that the objset is inconsistent, in which
1101eda14cbcSMatt Macy 		 * case it can not have a ZIL.
1102eda14cbcSMatt Macy 		 */
1103eda14cbcSMatt Macy 		if (error != EBUSY) {
1104eda14cbcSMatt Macy 			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
1105eda14cbcSMatt Macy 			    (unsigned long long)ds->ds_object, error);
1106eda14cbcSMatt Macy 		}
1107eda14cbcSMatt Macy 
1108eda14cbcSMatt Macy 		return (0);
1109eda14cbcSMatt Macy 	}
1110eda14cbcSMatt Macy 
1111eda14cbcSMatt Macy 	zilog = dmu_objset_zil(os);
1112eda14cbcSMatt Macy 	zh = zil_header_in_syncing_context(zilog);
1113eda14cbcSMatt Macy 	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
1114eda14cbcSMatt Macy 	first_txg = spa_min_claim_txg(zilog->zl_spa);
1115eda14cbcSMatt Macy 
1116eda14cbcSMatt Macy 	/*
1117eda14cbcSMatt Macy 	 * If the spa_log_state is not set to be cleared, check whether
1118eda14cbcSMatt Macy 	 * the current uberblock is a checkpoint one and if the current
1119eda14cbcSMatt Macy 	 * header has been claimed before moving on.
1120eda14cbcSMatt Macy 	 *
1121eda14cbcSMatt Macy 	 * If the current uberblock is a checkpointed uberblock then
1122eda14cbcSMatt Macy 	 * one of the following scenarios took place:
1123eda14cbcSMatt Macy 	 *
1124eda14cbcSMatt Macy 	 * 1] We are currently rewinding to the checkpoint of the pool.
1125eda14cbcSMatt Macy 	 * 2] We crashed in the middle of a checkpoint rewind but we
1126eda14cbcSMatt Macy 	 *    did manage to write the checkpointed uberblock to the
1127eda14cbcSMatt Macy 	 *    vdev labels, so when we tried to import the pool again
1128eda14cbcSMatt Macy 	 *    the checkpointed uberblock was selected from the import
1129eda14cbcSMatt Macy 	 *    procedure.
1130eda14cbcSMatt Macy 	 *
1131eda14cbcSMatt Macy 	 * In both cases we want to zero out all the ZIL blocks, except
1132eda14cbcSMatt Macy 	 * the ones that have been claimed at the time of the checkpoint
1133eda14cbcSMatt Macy 	 * (their zh_claim_txg != 0). The reason is that these blocks
1134eda14cbcSMatt Macy 	 * may be corrupted since we may have reused their locations on
1135eda14cbcSMatt Macy 	 * disk after we took the checkpoint.
1136eda14cbcSMatt Macy 	 *
1137eda14cbcSMatt Macy 	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1138eda14cbcSMatt Macy 	 * when we first figure out whether the current uberblock is
1139eda14cbcSMatt Macy 	 * checkpointed or not. Unfortunately, that would discard all
1140eda14cbcSMatt Macy 	 * the logs, including the ones that are claimed, and we would
1141eda14cbcSMatt Macy 	 * leak space.
1142eda14cbcSMatt Macy 	 */
1143eda14cbcSMatt Macy 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
1144eda14cbcSMatt Macy 	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1145eda14cbcSMatt Macy 	    zh->zh_claim_txg == 0)) {
1146eda14cbcSMatt Macy 		if (!BP_IS_HOLE(&zh->zh_log)) {
1147eda14cbcSMatt Macy 			(void) zil_parse(zilog, zil_clear_log_block,
1148eda14cbcSMatt Macy 			    zil_noop_log_record, tx, first_txg, B_FALSE);
1149eda14cbcSMatt Macy 		}
1150eda14cbcSMatt Macy 		BP_ZERO(&zh->zh_log);
1151eda14cbcSMatt Macy 		if (os->os_encrypted)
1152eda14cbcSMatt Macy 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1153eda14cbcSMatt Macy 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1154eda14cbcSMatt Macy 		dmu_objset_disown(os, B_FALSE, FTAG);
1155eda14cbcSMatt Macy 		return (0);
1156eda14cbcSMatt Macy 	}
1157eda14cbcSMatt Macy 
1158eda14cbcSMatt Macy 	/*
1159eda14cbcSMatt Macy 	 * If we are not rewinding and opening the pool normally, then
1160eda14cbcSMatt Macy 	 * the min_claim_txg should be equal to the first txg of the pool.
1161eda14cbcSMatt Macy 	 */
1162eda14cbcSMatt Macy 	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
1163eda14cbcSMatt Macy 
1164eda14cbcSMatt Macy 	/*
1165eda14cbcSMatt Macy 	 * Claim all log blocks if we haven't already done so, and remember
1166eda14cbcSMatt Macy 	 * the highest claimed sequence number.  This ensures that if we can
1167eda14cbcSMatt Macy 	 * read only part of the log now (e.g. due to a missing device),
1168eda14cbcSMatt Macy 	 * but we can read the entire log later, we will not try to replay
1169eda14cbcSMatt Macy 	 * or destroy beyond the last block we successfully claimed.
1170eda14cbcSMatt Macy 	 */
1171eda14cbcSMatt Macy 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
1172eda14cbcSMatt Macy 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
1173eda14cbcSMatt Macy 		(void) zil_parse(zilog, zil_claim_log_block,
1174eda14cbcSMatt Macy 		    zil_claim_log_record, tx, first_txg, B_FALSE);
1175eda14cbcSMatt Macy 		zh->zh_claim_txg = first_txg;
1176eda14cbcSMatt Macy 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
1177eda14cbcSMatt Macy 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
1178eda14cbcSMatt Macy 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
1179eda14cbcSMatt Macy 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
1180eda14cbcSMatt Macy 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
1181eda14cbcSMatt Macy 		if (os->os_encrypted)
1182eda14cbcSMatt Macy 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1183eda14cbcSMatt Macy 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1184eda14cbcSMatt Macy 	}
1185eda14cbcSMatt Macy 
1186eda14cbcSMatt Macy 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
1187eda14cbcSMatt Macy 	dmu_objset_disown(os, B_FALSE, FTAG);
1188eda14cbcSMatt Macy 	return (0);
1189eda14cbcSMatt Macy }
1190eda14cbcSMatt Macy 
1191eda14cbcSMatt Macy /*
1192eda14cbcSMatt Macy  * Check the log by walking the log chain.
1193eda14cbcSMatt Macy  * Checksum errors are ok as they indicate the end of the chain.
1194eda14cbcSMatt Macy  * Any other error (no device or read failure) returns an error.
1195eda14cbcSMatt Macy  */
1196eda14cbcSMatt Macy int
1197eda14cbcSMatt Macy zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
1198eda14cbcSMatt Macy {
1199e92ffd9bSMartin Matuska 	(void) dp;
1200eda14cbcSMatt Macy 	zilog_t *zilog;
1201eda14cbcSMatt Macy 	objset_t *os;
1202eda14cbcSMatt Macy 	blkptr_t *bp;
1203eda14cbcSMatt Macy 	int error;
1204eda14cbcSMatt Macy 
1205eda14cbcSMatt Macy 	ASSERT(tx == NULL);
1206eda14cbcSMatt Macy 
1207eda14cbcSMatt Macy 	error = dmu_objset_from_ds(ds, &os);
1208eda14cbcSMatt Macy 	if (error != 0) {
1209eda14cbcSMatt Macy 		cmn_err(CE_WARN, "can't open objset %llu, error %d",
1210eda14cbcSMatt Macy 		    (unsigned long long)ds->ds_object, error);
1211eda14cbcSMatt Macy 		return (0);
1212eda14cbcSMatt Macy 	}
1213eda14cbcSMatt Macy 
1214eda14cbcSMatt Macy 	zilog = dmu_objset_zil(os);
1215eda14cbcSMatt Macy 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
1216eda14cbcSMatt Macy 
1217eda14cbcSMatt Macy 	if (!BP_IS_HOLE(bp)) {
1218eda14cbcSMatt Macy 		vdev_t *vd;
1219eda14cbcSMatt Macy 		boolean_t valid = B_TRUE;
1220eda14cbcSMatt Macy 
1221eda14cbcSMatt Macy 		/*
1222eda14cbcSMatt Macy 		 * Check the first block and determine if it's on a log device
1223eda14cbcSMatt Macy 		 * which may have been removed or faulted prior to loading this
1224eda14cbcSMatt Macy 		 * pool.  If so, there's no point in checking the rest of the
1225eda14cbcSMatt Macy 		 * log as its content should have already been synced to the
1226eda14cbcSMatt Macy 		 * pool.
1227eda14cbcSMatt Macy 		 */
1228eda14cbcSMatt Macy 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
1229eda14cbcSMatt Macy 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
1230eda14cbcSMatt Macy 		if (vd->vdev_islog && vdev_is_dead(vd))
1231eda14cbcSMatt Macy 			valid = vdev_log_state_valid(vd);
1232eda14cbcSMatt Macy 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
1233eda14cbcSMatt Macy 
1234eda14cbcSMatt Macy 		if (!valid)
1235eda14cbcSMatt Macy 			return (0);
1236eda14cbcSMatt Macy 
1237eda14cbcSMatt Macy 		/*
1238eda14cbcSMatt Macy 		 * Check whether the current uberblock is checkpointed (e.g.
1239eda14cbcSMatt Macy 		 * we are rewinding) and whether the current header has been
1240eda14cbcSMatt Macy 		 * claimed or not. If it hasn't then skip verifying it. We
1241eda14cbcSMatt Macy 		 * do this because its ZIL blocks may be part of the pool's
1242eda14cbcSMatt Macy 		 * state before the rewind, which is no longer valid.
1243eda14cbcSMatt Macy 		 */
1244eda14cbcSMatt Macy 		zil_header_t *zh = zil_header_in_syncing_context(zilog);
1245eda14cbcSMatt Macy 		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1246eda14cbcSMatt Macy 		    zh->zh_claim_txg == 0)
1247eda14cbcSMatt Macy 			return (0);
1248eda14cbcSMatt Macy 	}
1249eda14cbcSMatt Macy 
1250eda14cbcSMatt Macy 	/*
1251eda14cbcSMatt Macy 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
1252eda14cbcSMatt Macy 	 * any blocks, but just determine whether it is possible to do so.
1253eda14cbcSMatt Macy 	 * In addition to checking the log chain, zil_claim_log_block()
1254eda14cbcSMatt Macy 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
1255eda14cbcSMatt Macy 	 * which will update spa_max_claim_txg.  See spa_load() for details.
1256eda14cbcSMatt Macy 	 */
1257eda14cbcSMatt Macy 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
1258eda14cbcSMatt Macy 	    zilog->zl_header->zh_claim_txg ? -1ULL :
1259eda14cbcSMatt Macy 	    spa_min_claim_txg(os->os_spa), B_FALSE);
1260eda14cbcSMatt Macy 
1261eda14cbcSMatt Macy 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
1262eda14cbcSMatt Macy }
1263eda14cbcSMatt Macy 
1264eda14cbcSMatt Macy /*
1265eda14cbcSMatt Macy  * When an itx is "skipped", this function is used to properly mark the
1266eda14cbcSMatt Macy  * waiter as "done, and signal any thread(s) waiting on it. An itx can
1267eda14cbcSMatt Macy  * be skipped (and not committed to an lwb) for a variety of reasons,
1268eda14cbcSMatt Macy  * one of them being that the itx was committed via spa_sync(), prior to
1269eda14cbcSMatt Macy  * it being committed to an lwb; this can happen if a thread calling
1270eda14cbcSMatt Macy  * zil_commit() is racing with spa_sync().
1271eda14cbcSMatt Macy  */
1272eda14cbcSMatt Macy static void
1273eda14cbcSMatt Macy zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
1274eda14cbcSMatt Macy {
1275eda14cbcSMatt Macy 	mutex_enter(&zcw->zcw_lock);
1276eda14cbcSMatt Macy 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1277eda14cbcSMatt Macy 	zcw->zcw_done = B_TRUE;
1278eda14cbcSMatt Macy 	cv_broadcast(&zcw->zcw_cv);
1279eda14cbcSMatt Macy 	mutex_exit(&zcw->zcw_lock);
1280eda14cbcSMatt Macy }
1281eda14cbcSMatt Macy 
1282eda14cbcSMatt Macy /*
1283eda14cbcSMatt Macy  * This function is used when the given waiter is to be linked into an
1284eda14cbcSMatt Macy  * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1285eda14cbcSMatt Macy  * At this point, the waiter will no longer be referenced by the itx,
1286eda14cbcSMatt Macy  * and instead, will be referenced by the lwb.
1287eda14cbcSMatt Macy  */
1288eda14cbcSMatt Macy static void
1289eda14cbcSMatt Macy zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1290eda14cbcSMatt Macy {
1291eda14cbcSMatt Macy 	/*
1292eda14cbcSMatt Macy 	 * The lwb_waiters field of the lwb is protected by the zilog's
1293315ee00fSMartin Matuska 	 * zl_issuer_lock while the lwb is open and zl_lock otherwise.
1294315ee00fSMartin Matuska 	 * zl_issuer_lock also protects leaving the open state.
1295315ee00fSMartin Matuska 	 * zcw_lwb setting is protected by zl_issuer_lock and state !=
1296315ee00fSMartin Matuska 	 * flush_done, which transition is protected by zl_lock.
1297eda14cbcSMatt Macy 	 */
1298315ee00fSMartin Matuska 	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_issuer_lock));
1299315ee00fSMartin Matuska 	IMPLY(lwb->lwb_state != LWB_STATE_OPENED,
1300315ee00fSMartin Matuska 	    MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1301315ee00fSMartin Matuska 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW);
1302315ee00fSMartin Matuska 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1303eda14cbcSMatt Macy 
1304eda14cbcSMatt Macy 	ASSERT(!list_link_active(&zcw->zcw_node));
1305eda14cbcSMatt Macy 	list_insert_tail(&lwb->lwb_waiters, zcw);
1306315ee00fSMartin Matuska 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1307eda14cbcSMatt Macy 	zcw->zcw_lwb = lwb;
1308eda14cbcSMatt Macy }
1309eda14cbcSMatt Macy 
1310eda14cbcSMatt Macy /*
1311eda14cbcSMatt Macy  * This function is used when zio_alloc_zil() fails to allocate a ZIL
1312eda14cbcSMatt Macy  * block, and the given waiter must be linked to the "nolwb waiters"
1313eda14cbcSMatt Macy  * list inside of zil_process_commit_list().
1314eda14cbcSMatt Macy  */
1315eda14cbcSMatt Macy static void
1316eda14cbcSMatt Macy zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1317eda14cbcSMatt Macy {
1318eda14cbcSMatt Macy 	ASSERT(!list_link_active(&zcw->zcw_node));
1319eda14cbcSMatt Macy 	list_insert_tail(nolwb, zcw);
1320315ee00fSMartin Matuska 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1321eda14cbcSMatt Macy }
1322eda14cbcSMatt Macy 
1323eda14cbcSMatt Macy void
1324eda14cbcSMatt Macy zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1325eda14cbcSMatt Macy {
1326eda14cbcSMatt Macy 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1327eda14cbcSMatt Macy 	avl_index_t where;
1328eda14cbcSMatt Macy 	zil_vdev_node_t *zv, zvsearch;
1329eda14cbcSMatt Macy 	int ndvas = BP_GET_NDVAS(bp);
1330eda14cbcSMatt Macy 	int i;
1331eda14cbcSMatt Macy 
13322ad756a6SMartin Matuska 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
13332ad756a6SMartin Matuska 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
13342ad756a6SMartin Matuska 
1335eda14cbcSMatt Macy 	if (zil_nocacheflush)
1336eda14cbcSMatt Macy 		return;
1337eda14cbcSMatt Macy 
1338eda14cbcSMatt Macy 	mutex_enter(&lwb->lwb_vdev_lock);
1339eda14cbcSMatt Macy 	for (i = 0; i < ndvas; i++) {
1340eda14cbcSMatt Macy 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1341eda14cbcSMatt Macy 		if (avl_find(t, &zvsearch, &where) == NULL) {
1342eda14cbcSMatt Macy 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1343eda14cbcSMatt Macy 			zv->zv_vdev = zvsearch.zv_vdev;
1344eda14cbcSMatt Macy 			avl_insert(t, zv, where);
1345eda14cbcSMatt Macy 		}
1346eda14cbcSMatt Macy 	}
1347eda14cbcSMatt Macy 	mutex_exit(&lwb->lwb_vdev_lock);
1348eda14cbcSMatt Macy }
1349eda14cbcSMatt Macy 
1350eda14cbcSMatt Macy static void
1351eda14cbcSMatt Macy zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1352eda14cbcSMatt Macy {
1353eda14cbcSMatt Macy 	avl_tree_t *src = &lwb->lwb_vdev_tree;
1354eda14cbcSMatt Macy 	avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1355eda14cbcSMatt Macy 	void *cookie = NULL;
1356eda14cbcSMatt Macy 	zil_vdev_node_t *zv;
1357eda14cbcSMatt Macy 
1358eda14cbcSMatt Macy 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1359eda14cbcSMatt Macy 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1360eda14cbcSMatt Macy 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1361eda14cbcSMatt Macy 
1362eda14cbcSMatt Macy 	/*
1363eda14cbcSMatt Macy 	 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1364eda14cbcSMatt Macy 	 * not need the protection of lwb_vdev_lock (it will only be modified
1365eda14cbcSMatt Macy 	 * while holding zilog->zl_lock) as its writes and those of its
1366eda14cbcSMatt Macy 	 * children have all completed.  The younger 'nlwb' may be waiting on
1367eda14cbcSMatt Macy 	 * future writes to additional vdevs.
1368eda14cbcSMatt Macy 	 */
1369eda14cbcSMatt Macy 	mutex_enter(&nlwb->lwb_vdev_lock);
1370eda14cbcSMatt Macy 	/*
1371eda14cbcSMatt Macy 	 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1372eda14cbcSMatt Macy 	 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1373eda14cbcSMatt Macy 	 */
1374eda14cbcSMatt Macy 	while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1375eda14cbcSMatt Macy 		avl_index_t where;
1376eda14cbcSMatt Macy 
1377eda14cbcSMatt Macy 		if (avl_find(dst, zv, &where) == NULL) {
1378eda14cbcSMatt Macy 			avl_insert(dst, zv, where);
1379eda14cbcSMatt Macy 		} else {
1380eda14cbcSMatt Macy 			kmem_free(zv, sizeof (*zv));
1381eda14cbcSMatt Macy 		}
1382eda14cbcSMatt Macy 	}
1383eda14cbcSMatt Macy 	mutex_exit(&nlwb->lwb_vdev_lock);
1384eda14cbcSMatt Macy }
1385eda14cbcSMatt Macy 
1386eda14cbcSMatt Macy void
1387eda14cbcSMatt Macy zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1388eda14cbcSMatt Macy {
1389eda14cbcSMatt Macy 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1390eda14cbcSMatt Macy }
1391eda14cbcSMatt Macy 
1392eda14cbcSMatt Macy /*
1393eda14cbcSMatt Macy  * This function is a called after all vdevs associated with a given lwb
1394eda14cbcSMatt Macy  * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1395eda14cbcSMatt Macy  * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1396eda14cbcSMatt Macy  * all "previous" lwb's will have completed before this function is
1397eda14cbcSMatt Macy  * called; i.e. this function is called for all previous lwbs before
1398eda14cbcSMatt Macy  * it's called for "this" lwb (enforced via zio the dependencies
1399eda14cbcSMatt Macy  * configured in zil_lwb_set_zio_dependency()).
1400eda14cbcSMatt Macy  *
1401eda14cbcSMatt Macy  * The intention is for this function to be called as soon as the
1402eda14cbcSMatt Macy  * contents of an lwb are considered "stable" on disk, and will survive
1403eda14cbcSMatt Macy  * any sudden loss of power. At this point, any threads waiting for the
1404eda14cbcSMatt Macy  * lwb to reach this state are signalled, and the "waiter" structures
1405eda14cbcSMatt Macy  * are marked "done".
1406eda14cbcSMatt Macy  */
1407eda14cbcSMatt Macy static void
1408eda14cbcSMatt Macy zil_lwb_flush_vdevs_done(zio_t *zio)
1409eda14cbcSMatt Macy {
1410eda14cbcSMatt Macy 	lwb_t *lwb = zio->io_private;
1411eda14cbcSMatt Macy 	zilog_t *zilog = lwb->lwb_zilog;
1412eda14cbcSMatt Macy 	zil_commit_waiter_t *zcw;
1413eda14cbcSMatt Macy 	itx_t *itx;
1414eda14cbcSMatt Macy 
1415eda14cbcSMatt Macy 	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1416eda14cbcSMatt Macy 
1417c0a83fe0SMartin Matuska 	hrtime_t t = gethrtime() - lwb->lwb_issued_timestamp;
1418eda14cbcSMatt Macy 
1419eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
1420eda14cbcSMatt Macy 
1421c0a83fe0SMartin Matuska 	zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 7 + t) / 8;
1422eda14cbcSMatt Macy 
1423eda14cbcSMatt Macy 	lwb->lwb_root_zio = NULL;
1424eda14cbcSMatt Macy 
14252ad756a6SMartin Matuska 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
14262ad756a6SMartin Matuska 	lwb->lwb_state = LWB_STATE_FLUSH_DONE;
14272ad756a6SMartin Matuska 
1428eda14cbcSMatt Macy 	if (zilog->zl_last_lwb_opened == lwb) {
1429eda14cbcSMatt Macy 		/*
1430eda14cbcSMatt Macy 		 * Remember the highest committed log sequence number
1431eda14cbcSMatt Macy 		 * for ztest. We only update this value when all the log
1432eda14cbcSMatt Macy 		 * writes succeeded, because ztest wants to ASSERT that
1433eda14cbcSMatt Macy 		 * it got the whole log chain.
1434eda14cbcSMatt Macy 		 */
1435eda14cbcSMatt Macy 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1436eda14cbcSMatt Macy 	}
1437eda14cbcSMatt Macy 
14382ad756a6SMartin Matuska 	while ((itx = list_remove_head(&lwb->lwb_itxs)) != NULL)
14394e8d558cSMartin Matuska 		zil_itx_destroy(itx);
14404e8d558cSMartin Matuska 
14412ad756a6SMartin Matuska 	while ((zcw = list_remove_head(&lwb->lwb_waiters)) != NULL) {
1442eda14cbcSMatt Macy 		mutex_enter(&zcw->zcw_lock);
1443eda14cbcSMatt Macy 
14442ad756a6SMartin Matuska 		ASSERT3P(zcw->zcw_lwb, ==, lwb);
1445eda14cbcSMatt Macy 		zcw->zcw_lwb = NULL;
144653b70c86SMartin Matuska 		/*
144753b70c86SMartin Matuska 		 * We expect any ZIO errors from child ZIOs to have been
144853b70c86SMartin Matuska 		 * propagated "up" to this specific LWB's root ZIO, in
144953b70c86SMartin Matuska 		 * order for this error handling to work correctly. This
145053b70c86SMartin Matuska 		 * includes ZIO errors from either this LWB's write or
145153b70c86SMartin Matuska 		 * flush, as well as any errors from other dependent LWBs
145253b70c86SMartin Matuska 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
145353b70c86SMartin Matuska 		 *
145453b70c86SMartin Matuska 		 * With that said, it's important to note that LWB flush
145553b70c86SMartin Matuska 		 * errors are not propagated up to the LWB root ZIO.
145653b70c86SMartin Matuska 		 * This is incorrect behavior, and results in VDEV flush
145753b70c86SMartin Matuska 		 * errors not being handled correctly here. See the
145853b70c86SMartin Matuska 		 * comment above the call to "zio_flush" for details.
145953b70c86SMartin Matuska 		 */
1460eda14cbcSMatt Macy 
1461eda14cbcSMatt Macy 		zcw->zcw_zio_error = zio->io_error;
1462eda14cbcSMatt Macy 
1463eda14cbcSMatt Macy 		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1464eda14cbcSMatt Macy 		zcw->zcw_done = B_TRUE;
1465eda14cbcSMatt Macy 		cv_broadcast(&zcw->zcw_cv);
1466eda14cbcSMatt Macy 
1467eda14cbcSMatt Macy 		mutex_exit(&zcw->zcw_lock);
1468eda14cbcSMatt Macy 	}
14692ad756a6SMartin Matuska 
14702ad756a6SMartin Matuska 	uint64_t txg = lwb->lwb_issued_txg;
14712ad756a6SMartin Matuska 
14722ad756a6SMartin Matuska 	/* Once we drop the lock, lwb may be freed by zil_sync(). */
14732ad756a6SMartin Matuska 	mutex_exit(&zilog->zl_lock);
1474eda14cbcSMatt Macy 
1475e3aa18adSMartin Matuska 	mutex_enter(&zilog->zl_lwb_io_lock);
1476e3aa18adSMartin Matuska 	ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0);
1477e3aa18adSMartin Matuska 	zilog->zl_lwb_inflight[txg & TXG_MASK]--;
1478e3aa18adSMartin Matuska 	if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0)
1479e3aa18adSMartin Matuska 		cv_broadcast(&zilog->zl_lwb_io_cv);
1480e3aa18adSMartin Matuska 	mutex_exit(&zilog->zl_lwb_io_lock);
1481e3aa18adSMartin Matuska }
1482e3aa18adSMartin Matuska 
1483eda14cbcSMatt Macy /*
1484e3aa18adSMartin Matuska  * Wait for the completion of all issued write/flush of that txg provided.
1485e3aa18adSMartin Matuska  * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1486eda14cbcSMatt Macy  */
1487e3aa18adSMartin Matuska static void
1488e3aa18adSMartin Matuska zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg)
1489e3aa18adSMartin Matuska {
1490e3aa18adSMartin Matuska 	ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa));
1491e3aa18adSMartin Matuska 
1492e3aa18adSMartin Matuska 	mutex_enter(&zilog->zl_lwb_io_lock);
1493e3aa18adSMartin Matuska 	while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0)
1494e3aa18adSMartin Matuska 		cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock);
1495e3aa18adSMartin Matuska 	mutex_exit(&zilog->zl_lwb_io_lock);
1496e3aa18adSMartin Matuska 
1497e3aa18adSMartin Matuska #ifdef ZFS_DEBUG
1498e3aa18adSMartin Matuska 	mutex_enter(&zilog->zl_lock);
1499e3aa18adSMartin Matuska 	mutex_enter(&zilog->zl_lwb_io_lock);
1500e3aa18adSMartin Matuska 	lwb_t *lwb = list_head(&zilog->zl_lwb_list);
1501315ee00fSMartin Matuska 	while (lwb != NULL) {
1502e3aa18adSMartin Matuska 		if (lwb->lwb_issued_txg <= txg) {
1503e3aa18adSMartin Matuska 			ASSERT(lwb->lwb_state != LWB_STATE_ISSUED);
1504e3aa18adSMartin Matuska 			ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE);
1505e3aa18adSMartin Matuska 			IMPLY(lwb->lwb_issued_txg > 0,
1506e3aa18adSMartin Matuska 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
1507e3aa18adSMartin Matuska 		}
1508c0a83fe0SMartin Matuska 		IMPLY(lwb->lwb_state == LWB_STATE_WRITE_DONE ||
1509c0a83fe0SMartin Matuska 		    lwb->lwb_state == LWB_STATE_FLUSH_DONE,
1510e3aa18adSMartin Matuska 		    lwb->lwb_buf == NULL);
1511e3aa18adSMartin Matuska 		lwb = list_next(&zilog->zl_lwb_list, lwb);
1512e3aa18adSMartin Matuska 	}
1513e3aa18adSMartin Matuska 	mutex_exit(&zilog->zl_lwb_io_lock);
1514e3aa18adSMartin Matuska 	mutex_exit(&zilog->zl_lock);
1515e3aa18adSMartin Matuska #endif
1516eda14cbcSMatt Macy }
1517eda14cbcSMatt Macy 
1518eda14cbcSMatt Macy /*
1519eda14cbcSMatt Macy  * This is called when an lwb's write zio completes. The callback's
1520eda14cbcSMatt Macy  * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1521eda14cbcSMatt Macy  * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1522eda14cbcSMatt Macy  * in writing out this specific lwb's data, and in the case that cache
1523eda14cbcSMatt Macy  * flushes have been deferred, vdevs involved in writing the data for
1524eda14cbcSMatt Macy  * previous lwbs. The writes corresponding to all the vdevs in the
1525eda14cbcSMatt Macy  * lwb_vdev_tree will have completed by the time this is called, due to
1526eda14cbcSMatt Macy  * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1527eda14cbcSMatt Macy  * which takes deferred flushes into account. The lwb will be "done"
1528eda14cbcSMatt Macy  * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1529eda14cbcSMatt Macy  * completion callback for the lwb's root zio.
1530eda14cbcSMatt Macy  */
1531eda14cbcSMatt Macy static void
1532eda14cbcSMatt Macy zil_lwb_write_done(zio_t *zio)
1533eda14cbcSMatt Macy {
1534eda14cbcSMatt Macy 	lwb_t *lwb = zio->io_private;
1535eda14cbcSMatt Macy 	spa_t *spa = zio->io_spa;
1536eda14cbcSMatt Macy 	zilog_t *zilog = lwb->lwb_zilog;
1537eda14cbcSMatt Macy 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1538eda14cbcSMatt Macy 	void *cookie = NULL;
1539eda14cbcSMatt Macy 	zil_vdev_node_t *zv;
1540eda14cbcSMatt Macy 	lwb_t *nlwb;
1541eda14cbcSMatt Macy 
1542eda14cbcSMatt Macy 	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1543eda14cbcSMatt Macy 
1544184c1b94SMartin Matuska 	abd_free(zio->io_abd);
1545c0a83fe0SMartin Matuska 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1546c0a83fe0SMartin Matuska 	lwb->lwb_buf = NULL;
1547eda14cbcSMatt Macy 
1548eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
1549eda14cbcSMatt Macy 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1550eda14cbcSMatt Macy 	lwb->lwb_state = LWB_STATE_WRITE_DONE;
1551315ee00fSMartin Matuska 	lwb->lwb_child_zio = NULL;
1552eda14cbcSMatt Macy 	lwb->lwb_write_zio = NULL;
1553cbfe9975SMartin Matuska 
1554cbfe9975SMartin Matuska 	/*
1555cbfe9975SMartin Matuska 	 * If nlwb is not yet issued, zil_lwb_set_zio_dependency() is not
1556cbfe9975SMartin Matuska 	 * called for it yet, and when it will be, it won't be able to make
1557cbfe9975SMartin Matuska 	 * its write ZIO a parent this ZIO.  In such case we can not defer
1558cbfe9975SMartin Matuska 	 * our flushes or below may be a race between the done callbacks.
1559cbfe9975SMartin Matuska 	 */
1560eda14cbcSMatt Macy 	nlwb = list_next(&zilog->zl_lwb_list, lwb);
1561cbfe9975SMartin Matuska 	if (nlwb && nlwb->lwb_state != LWB_STATE_ISSUED)
1562cbfe9975SMartin Matuska 		nlwb = NULL;
1563eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
1564eda14cbcSMatt Macy 
1565eda14cbcSMatt Macy 	if (avl_numnodes(t) == 0)
1566eda14cbcSMatt Macy 		return;
1567eda14cbcSMatt Macy 
1568eda14cbcSMatt Macy 	/*
1569eda14cbcSMatt Macy 	 * If there was an IO error, we're not going to call zio_flush()
1570eda14cbcSMatt Macy 	 * on these vdevs, so we simply empty the tree and free the
1571eda14cbcSMatt Macy 	 * nodes. We avoid calling zio_flush() since there isn't any
1572eda14cbcSMatt Macy 	 * good reason for doing so, after the lwb block failed to be
1573eda14cbcSMatt Macy 	 * written out.
157453b70c86SMartin Matuska 	 *
157553b70c86SMartin Matuska 	 * Additionally, we don't perform any further error handling at
157653b70c86SMartin Matuska 	 * this point (e.g. setting "zcw_zio_error" appropriately), as
157753b70c86SMartin Matuska 	 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
157853b70c86SMartin Matuska 	 * we expect any error seen here, to have been propagated to
157953b70c86SMartin Matuska 	 * that function).
1580eda14cbcSMatt Macy 	 */
1581eda14cbcSMatt Macy 	if (zio->io_error != 0) {
1582eda14cbcSMatt Macy 		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1583eda14cbcSMatt Macy 			kmem_free(zv, sizeof (*zv));
1584eda14cbcSMatt Macy 		return;
1585eda14cbcSMatt Macy 	}
1586eda14cbcSMatt Macy 
1587eda14cbcSMatt Macy 	/*
1588eda14cbcSMatt Macy 	 * If this lwb does not have any threads waiting for it to
1589eda14cbcSMatt Macy 	 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1590eda14cbcSMatt Macy 	 * command to the vdevs written to by "this" lwb, and instead
1591eda14cbcSMatt Macy 	 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1592eda14cbcSMatt Macy 	 * command for those vdevs. Thus, we merge the vdev tree of
1593eda14cbcSMatt Macy 	 * "this" lwb with the vdev tree of the "next" lwb in the list,
1594eda14cbcSMatt Macy 	 * and assume the "next" lwb will handle flushing the vdevs (or
1595eda14cbcSMatt Macy 	 * deferring the flush(s) again).
1596eda14cbcSMatt Macy 	 *
1597eda14cbcSMatt Macy 	 * This is a useful performance optimization, especially for
1598eda14cbcSMatt Macy 	 * workloads with lots of async write activity and few sync
1599eda14cbcSMatt Macy 	 * write and/or fsync activity, as it has the potential to
1600eda14cbcSMatt Macy 	 * coalesce multiple flush commands to a vdev into one.
1601eda14cbcSMatt Macy 	 */
1602c0a83fe0SMartin Matuska 	if (list_is_empty(&lwb->lwb_waiters) && nlwb != NULL) {
1603eda14cbcSMatt Macy 		zil_lwb_flush_defer(lwb, nlwb);
1604eda14cbcSMatt Macy 		ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1605eda14cbcSMatt Macy 		return;
1606eda14cbcSMatt Macy 	}
1607eda14cbcSMatt Macy 
1608eda14cbcSMatt Macy 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1609eda14cbcSMatt Macy 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1610c0a83fe0SMartin Matuska 		if (vd != NULL && !vd->vdev_nowritecache) {
161153b70c86SMartin Matuska 			/*
161253b70c86SMartin Matuska 			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
161353b70c86SMartin Matuska 			 * always used within "zio_flush". This means,
161453b70c86SMartin Matuska 			 * any errors when flushing the vdev(s), will
161553b70c86SMartin Matuska 			 * (unfortunately) not be handled correctly,
161653b70c86SMartin Matuska 			 * since these "zio_flush" errors will not be
161753b70c86SMartin Matuska 			 * propagated up to "zil_lwb_flush_vdevs_done".
161853b70c86SMartin Matuska 			 */
1619eda14cbcSMatt Macy 			zio_flush(lwb->lwb_root_zio, vd);
162053b70c86SMartin Matuska 		}
1621eda14cbcSMatt Macy 		kmem_free(zv, sizeof (*zv));
1622eda14cbcSMatt Macy 	}
1623eda14cbcSMatt Macy }
1624eda14cbcSMatt Macy 
1625315ee00fSMartin Matuska /*
1626315ee00fSMartin Matuska  * Build the zio dependency chain, which is used to preserve the ordering of
1627315ee00fSMartin Matuska  * lwb completions that is required by the semantics of the ZIL. Each new lwb
1628315ee00fSMartin Matuska  * zio becomes a parent of the previous lwb zio, such that the new lwb's zio
1629315ee00fSMartin Matuska  * cannot complete until the previous lwb's zio completes.
1630315ee00fSMartin Matuska  *
1631315ee00fSMartin Matuska  * This is required by the semantics of zil_commit(): the commit waiters
1632315ee00fSMartin Matuska  * attached to the lwbs will be woken in the lwb zio's completion callback,
1633315ee00fSMartin Matuska  * so this zio dependency graph ensures the waiters are woken in the correct
1634315ee00fSMartin Matuska  * order (the same order the lwbs were created).
1635315ee00fSMartin Matuska  */
1636eda14cbcSMatt Macy static void
1637eda14cbcSMatt Macy zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1638eda14cbcSMatt Macy {
1639eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
1640eda14cbcSMatt Macy 
1641315ee00fSMartin Matuska 	lwb_t *prev_lwb = list_prev(&zilog->zl_lwb_list, lwb);
1642315ee00fSMartin Matuska 	if (prev_lwb == NULL ||
1643315ee00fSMartin Matuska 	    prev_lwb->lwb_state == LWB_STATE_FLUSH_DONE)
1644315ee00fSMartin Matuska 		return;
1645eda14cbcSMatt Macy 
1646eda14cbcSMatt Macy 	/*
1647315ee00fSMartin Matuska 	 * If the previous lwb's write hasn't already completed, we also want
1648315ee00fSMartin Matuska 	 * to order the completion of the lwb write zios (above, we only order
1649315ee00fSMartin Matuska 	 * the completion of the lwb root zios). This is required because of
1650315ee00fSMartin Matuska 	 * how we can defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1651eda14cbcSMatt Macy 	 *
1652315ee00fSMartin Matuska 	 * When the DKIOCFLUSHWRITECACHE commands are deferred, the previous
1653315ee00fSMartin Matuska 	 * lwb will rely on this lwb to flush the vdevs written to by that
1654315ee00fSMartin Matuska 	 * previous lwb. Thus, we need to ensure this lwb doesn't issue the
1655315ee00fSMartin Matuska 	 * flush until after the previous lwb's write completes. We ensure
1656315ee00fSMartin Matuska 	 * this ordering by setting the zio parent/child relationship here.
1657eda14cbcSMatt Macy 	 *
1658315ee00fSMartin Matuska 	 * Without this relationship on the lwb's write zio, it's possible
1659315ee00fSMartin Matuska 	 * for this lwb's write to complete prior to the previous lwb's write
1660315ee00fSMartin Matuska 	 * completing; and thus, the vdevs for the previous lwb would be
1661315ee00fSMartin Matuska 	 * flushed prior to that lwb's data being written to those vdevs (the
1662315ee00fSMartin Matuska 	 * vdevs are flushed in the lwb write zio's completion handler,
1663315ee00fSMartin Matuska 	 * zil_lwb_write_done()).
1664eda14cbcSMatt Macy 	 */
1665315ee00fSMartin Matuska 	if (prev_lwb->lwb_state == LWB_STATE_ISSUED) {
1666315ee00fSMartin Matuska 		ASSERT3P(prev_lwb->lwb_write_zio, !=, NULL);
1667315ee00fSMartin Matuska 		zio_add_child(lwb->lwb_write_zio, prev_lwb->lwb_write_zio);
1668315ee00fSMartin Matuska 	} else {
1669315ee00fSMartin Matuska 		ASSERT3S(prev_lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1670315ee00fSMartin Matuska 	}
1671eda14cbcSMatt Macy 
1672315ee00fSMartin Matuska 	ASSERT3P(prev_lwb->lwb_root_zio, !=, NULL);
1673315ee00fSMartin Matuska 	zio_add_child(lwb->lwb_root_zio, prev_lwb->lwb_root_zio);
1674eda14cbcSMatt Macy }
1675eda14cbcSMatt Macy 
1676eda14cbcSMatt Macy 
1677eda14cbcSMatt Macy /*
1678eda14cbcSMatt Macy  * This function's purpose is to "open" an lwb such that it is ready to
1679315ee00fSMartin Matuska  * accept new itxs being committed to it. This function is idempotent; if
1680315ee00fSMartin Matuska  * the passed in lwb has already been opened, it is essentially a no-op.
1681eda14cbcSMatt Macy  */
1682eda14cbcSMatt Macy static void
1683eda14cbcSMatt Macy zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1684eda14cbcSMatt Macy {
1685eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1686eda14cbcSMatt Macy 
1687315ee00fSMartin Matuska 	if (lwb->lwb_state != LWB_STATE_NEW) {
1688315ee00fSMartin Matuska 		ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
16894e8d558cSMartin Matuska 		return;
16904e8d558cSMartin Matuska 	}
16914e8d558cSMartin Matuska 
1692315ee00fSMartin Matuska 	mutex_enter(&zilog->zl_lock);
1693eda14cbcSMatt Macy 	lwb->lwb_state = LWB_STATE_OPENED;
1694eda14cbcSMatt Macy 	zilog->zl_last_lwb_opened = lwb;
1695eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
1696eda14cbcSMatt Macy }
1697eda14cbcSMatt Macy 
1698eda14cbcSMatt Macy /*
1699eda14cbcSMatt Macy  * Define a limited set of intent log block sizes.
1700eda14cbcSMatt Macy  *
1701eda14cbcSMatt Macy  * These must be a multiple of 4KB. Note only the amount used (again
1702eda14cbcSMatt Macy  * aligned to 4KB) actually gets written. However, we can't always just
1703eda14cbcSMatt Macy  * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1704eda14cbcSMatt Macy  */
1705e92ffd9bSMartin Matuska static const struct {
1706eda14cbcSMatt Macy 	uint64_t	limit;
1707eda14cbcSMatt Macy 	uint64_t	blksz;
1708eda14cbcSMatt Macy } zil_block_buckets[] = {
1709eda14cbcSMatt Macy 	{ 4096,		4096 },			/* non TX_WRITE */
1710eda14cbcSMatt Macy 	{ 8192 + 4096,	8192 + 4096 },		/* database */
1711eda14cbcSMatt Macy 	{ 32768 + 4096,	32768 + 4096 },		/* NFS writes */
1712eda14cbcSMatt Macy 	{ 65536 + 4096,	65536 + 4096 },		/* 64KB writes */
1713eda14cbcSMatt Macy 	{ 131072,	131072 },		/* < 128KB writes */
1714eda14cbcSMatt Macy 	{ 131072 +4096,	65536 + 4096 },		/* 128KB writes */
1715eda14cbcSMatt Macy 	{ UINT64_MAX,	SPA_OLD_MAXBLOCKSIZE},	/* > 128KB writes */
1716eda14cbcSMatt Macy };
1717eda14cbcSMatt Macy 
1718eda14cbcSMatt Macy /*
1719eda14cbcSMatt Macy  * Maximum block size used by the ZIL.  This is picked up when the ZIL is
1720eda14cbcSMatt Macy  * initialized.  Otherwise this should not be used directly; see
1721eda14cbcSMatt Macy  * zl_max_block_size instead.
1722eda14cbcSMatt Macy  */
1723be181ee2SMartin Matuska static uint_t zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1724eda14cbcSMatt Macy 
1725eda14cbcSMatt Macy /*
17264e8d558cSMartin Matuska  * Close the log block for being issued and allocate the next one.
17274e8d558cSMartin Matuska  * Has to be called under zl_issuer_lock to chain more lwbs.
1728eda14cbcSMatt Macy  */
1729eda14cbcSMatt Macy static lwb_t *
1730315ee00fSMartin Matuska zil_lwb_write_close(zilog_t *zilog, lwb_t *lwb, lwb_state_t state)
1731eda14cbcSMatt Macy {
1732315ee00fSMartin Matuska 	int i;
1733eda14cbcSMatt Macy 
1734eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1735eda14cbcSMatt Macy 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1736315ee00fSMartin Matuska 	lwb->lwb_state = LWB_STATE_CLOSED;
1737eda14cbcSMatt Macy 
17384e8d558cSMartin Matuska 	/*
1739315ee00fSMartin Matuska 	 * If there was an allocation failure then returned NULL will trigger
1740315ee00fSMartin Matuska 	 * zil_commit_writer_stall() at the caller.  This is inherently racy,
1741315ee00fSMartin Matuska 	 * since allocation may not have happened yet.
17424e8d558cSMartin Matuska 	 */
1743315ee00fSMartin Matuska 	if (lwb->lwb_error != 0)
1744315ee00fSMartin Matuska 		return (NULL);
1745eda14cbcSMatt Macy 
1746eda14cbcSMatt Macy 	/*
1747eda14cbcSMatt Macy 	 * Log blocks are pre-allocated. Here we select the size of the next
1748eda14cbcSMatt Macy 	 * block, based on size used in the last block.
1749eda14cbcSMatt Macy 	 * - first find the smallest bucket that will fit the block from a
1750eda14cbcSMatt Macy 	 *   limited set of block sizes. This is because it's faster to write
1751eda14cbcSMatt Macy 	 *   blocks allocated from the same metaslab as they are adjacent or
1752eda14cbcSMatt Macy 	 *   close.
1753eda14cbcSMatt Macy 	 * - next find the maximum from the new suggested size and an array of
1754eda14cbcSMatt Macy 	 *   previous sizes. This lessens a picket fence effect of wrongly
1755eda14cbcSMatt Macy 	 *   guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1756eda14cbcSMatt Macy 	 *   requests.
1757eda14cbcSMatt Macy 	 *
1758eda14cbcSMatt Macy 	 * Note we only write what is used, but we can't just allocate
1759eda14cbcSMatt Macy 	 * the maximum block size because we can exhaust the available
1760eda14cbcSMatt Macy 	 * pool log space.
1761eda14cbcSMatt Macy 	 */
1762315ee00fSMartin Matuska 	uint64_t zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1763eda14cbcSMatt Macy 	for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1764eda14cbcSMatt Macy 		continue;
1765eda14cbcSMatt Macy 	zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1766eda14cbcSMatt Macy 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1767eda14cbcSMatt Macy 	for (i = 0; i < ZIL_PREV_BLKS; i++)
1768eda14cbcSMatt Macy 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
17694e8d558cSMartin Matuska 	DTRACE_PROBE3(zil__block__size, zilog_t *, zilog,
17704e8d558cSMartin Matuska 	    uint64_t, zil_blksz,
17714e8d558cSMartin Matuska 	    uint64_t, zilog->zl_prev_blks[zilog->zl_prev_rotor]);
1772eda14cbcSMatt Macy 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1773eda14cbcSMatt Macy 
1774315ee00fSMartin Matuska 	return (zil_alloc_lwb(zilog, zil_blksz, NULL, 0, 0, state));
17754e8d558cSMartin Matuska }
17764e8d558cSMartin Matuska 
17774e8d558cSMartin Matuska /*
17784e8d558cSMartin Matuska  * Finalize previously closed block and issue the write zio.
17794e8d558cSMartin Matuska  */
17804e8d558cSMartin Matuska static void
17814e8d558cSMartin Matuska zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
17824e8d558cSMartin Matuska {
1783315ee00fSMartin Matuska 	spa_t *spa = zilog->zl_spa;
17844e8d558cSMartin Matuska 	zil_chain_t *zilc;
1785315ee00fSMartin Matuska 	boolean_t slog;
1786315ee00fSMartin Matuska 	zbookmark_phys_t zb;
1787315ee00fSMartin Matuska 	zio_priority_t prio;
1788315ee00fSMartin Matuska 	int error;
17894e8d558cSMartin Matuska 
1790315ee00fSMartin Matuska 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED);
1791315ee00fSMartin Matuska 
1792315ee00fSMartin Matuska 	/* Actually fill the lwb with the data. */
17934e8d558cSMartin Matuska 	for (itx_t *itx = list_head(&lwb->lwb_itxs); itx;
17944e8d558cSMartin Matuska 	    itx = list_next(&lwb->lwb_itxs, itx))
17954e8d558cSMartin Matuska 		zil_lwb_commit(zilog, lwb, itx);
17964e8d558cSMartin Matuska 	lwb->lwb_nused = lwb->lwb_nfilled;
17974e8d558cSMartin Matuska 
1798315ee00fSMartin Matuska 	lwb->lwb_root_zio = zio_root(spa, zil_lwb_flush_vdevs_done, lwb,
1799315ee00fSMartin Matuska 	    ZIO_FLAG_CANFAIL);
1800315ee00fSMartin Matuska 
1801315ee00fSMartin Matuska 	/*
1802315ee00fSMartin Matuska 	 * The lwb is now ready to be issued, but it can be only if it already
1803315ee00fSMartin Matuska 	 * got its block pointer allocated or the allocation has failed.
1804315ee00fSMartin Matuska 	 * Otherwise leave it as-is, relying on some other thread to issue it
1805315ee00fSMartin Matuska 	 * after allocating its block pointer via calling zil_lwb_write_issue()
1806315ee00fSMartin Matuska 	 * for the previous lwb(s) in the chain.
1807315ee00fSMartin Matuska 	 */
1808315ee00fSMartin Matuska 	mutex_enter(&zilog->zl_lock);
1809315ee00fSMartin Matuska 	lwb->lwb_state = LWB_STATE_READY;
1810315ee00fSMartin Matuska 	if (BP_IS_HOLE(&lwb->lwb_blk) && lwb->lwb_error == 0) {
1811315ee00fSMartin Matuska 		mutex_exit(&zilog->zl_lock);
1812315ee00fSMartin Matuska 		return;
1813315ee00fSMartin Matuska 	}
1814315ee00fSMartin Matuska 	mutex_exit(&zilog->zl_lock);
1815315ee00fSMartin Matuska 
1816315ee00fSMartin Matuska next_lwb:
1817315ee00fSMartin Matuska 	if (lwb->lwb_slim)
1818315ee00fSMartin Matuska 		zilc = (zil_chain_t *)lwb->lwb_buf;
1819315ee00fSMartin Matuska 	else
1820315ee00fSMartin Matuska 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_nmax);
1821315ee00fSMartin Matuska 	int wsz = lwb->lwb_sz;
1822315ee00fSMartin Matuska 	if (lwb->lwb_error == 0) {
1823315ee00fSMartin Matuska 		abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, lwb->lwb_sz);
1824315ee00fSMartin Matuska 		if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1825315ee00fSMartin Matuska 			prio = ZIO_PRIORITY_SYNC_WRITE;
1826315ee00fSMartin Matuska 		else
1827315ee00fSMartin Matuska 			prio = ZIO_PRIORITY_ASYNC_WRITE;
1828315ee00fSMartin Matuska 		SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1829315ee00fSMartin Matuska 		    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1830315ee00fSMartin Matuska 		    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1831315ee00fSMartin Matuska 		lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, spa, 0,
1832315ee00fSMartin Matuska 		    &lwb->lwb_blk, lwb_abd, lwb->lwb_sz, zil_lwb_write_done,
1833315ee00fSMartin Matuska 		    lwb, prio, ZIO_FLAG_CANFAIL, &zb);
1834315ee00fSMartin Matuska 		zil_lwb_add_block(lwb, &lwb->lwb_blk);
1835315ee00fSMartin Matuska 
1836315ee00fSMartin Matuska 		if (lwb->lwb_slim) {
1837eda14cbcSMatt Macy 			/* For Slim ZIL only write what is used. */
1838315ee00fSMartin Matuska 			wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ,
1839315ee00fSMartin Matuska 			    int);
18404e8d558cSMartin Matuska 			ASSERT3S(wsz, <=, lwb->lwb_sz);
1841eda14cbcSMatt Macy 			zio_shrink(lwb->lwb_write_zio, wsz);
1842e639e0d2SMartin Matuska 			wsz = lwb->lwb_write_zio->io_size;
1843eda14cbcSMatt Macy 		}
1844315ee00fSMartin Matuska 		memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused);
1845eda14cbcSMatt Macy 		zilc->zc_pad = 0;
1846eda14cbcSMatt Macy 		zilc->zc_nused = lwb->lwb_nused;
1847eda14cbcSMatt Macy 		zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1848315ee00fSMartin Matuska 	} else {
1849315ee00fSMartin Matuska 		/*
1850315ee00fSMartin Matuska 		 * We can't write the lwb if there was an allocation failure,
1851315ee00fSMartin Matuska 		 * so create a null zio instead just to maintain dependencies.
1852315ee00fSMartin Matuska 		 */
1853315ee00fSMartin Matuska 		lwb->lwb_write_zio = zio_null(lwb->lwb_root_zio, spa, NULL,
1854315ee00fSMartin Matuska 		    zil_lwb_write_done, lwb, ZIO_FLAG_CANFAIL);
1855315ee00fSMartin Matuska 		lwb->lwb_write_zio->io_error = lwb->lwb_error;
1856315ee00fSMartin Matuska 	}
1857315ee00fSMartin Matuska 	if (lwb->lwb_child_zio)
1858315ee00fSMartin Matuska 		zio_add_child(lwb->lwb_write_zio, lwb->lwb_child_zio);
1859eda14cbcSMatt Macy 
1860eda14cbcSMatt Macy 	/*
1861315ee00fSMartin Matuska 	 * Open transaction to allocate the next block pointer.
1862eda14cbcSMatt Macy 	 */
1863315ee00fSMartin Matuska 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1864315ee00fSMartin Matuska 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1865315ee00fSMartin Matuska 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1866315ee00fSMartin Matuska 	uint64_t txg = dmu_tx_get_txg(tx);
1867315ee00fSMartin Matuska 
1868315ee00fSMartin Matuska 	/*
1869315ee00fSMartin Matuska 	 * Allocate next the block pointer unless we are already in error.
1870315ee00fSMartin Matuska 	 */
1871315ee00fSMartin Matuska 	lwb_t *nlwb = list_next(&zilog->zl_lwb_list, lwb);
1872315ee00fSMartin Matuska 	blkptr_t *bp = &zilc->zc_next_blk;
1873315ee00fSMartin Matuska 	BP_ZERO(bp);
1874315ee00fSMartin Matuska 	error = lwb->lwb_error;
1875315ee00fSMartin Matuska 	if (error == 0) {
1876315ee00fSMartin Matuska 		error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, nlwb->lwb_sz,
1877315ee00fSMartin Matuska 		    &slog);
1878315ee00fSMartin Matuska 	}
1879315ee00fSMartin Matuska 	if (error == 0) {
1880315ee00fSMartin Matuska 		ASSERT3U(bp->blk_birth, ==, txg);
1881315ee00fSMartin Matuska 		BP_SET_CHECKSUM(bp, nlwb->lwb_slim ? ZIO_CHECKSUM_ZILOG2 :
1882315ee00fSMartin Matuska 		    ZIO_CHECKSUM_ZILOG);
1883315ee00fSMartin Matuska 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1884315ee00fSMartin Matuska 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1885315ee00fSMartin Matuska 	}
1886315ee00fSMartin Matuska 
1887315ee00fSMartin Matuska 	/*
1888315ee00fSMartin Matuska 	 * Reduce TXG open time by incrementing inflight counter and committing
1889315ee00fSMartin Matuska 	 * the transaciton.  zil_sync() will wait for it to return to zero.
1890315ee00fSMartin Matuska 	 */
1891315ee00fSMartin Matuska 	mutex_enter(&zilog->zl_lwb_io_lock);
1892315ee00fSMartin Matuska 	lwb->lwb_issued_txg = txg;
1893315ee00fSMartin Matuska 	zilog->zl_lwb_inflight[txg & TXG_MASK]++;
1894315ee00fSMartin Matuska 	zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg);
1895315ee00fSMartin Matuska 	mutex_exit(&zilog->zl_lwb_io_lock);
1896315ee00fSMartin Matuska 	dmu_tx_commit(tx);
1897315ee00fSMartin Matuska 
1898315ee00fSMartin Matuska 	spa_config_enter(spa, SCL_STATE, lwb, RW_READER);
1899315ee00fSMartin Matuska 
1900315ee00fSMartin Matuska 	/*
1901315ee00fSMartin Matuska 	 * We've completed all potentially blocking operations.  Update the
1902315ee00fSMartin Matuska 	 * nlwb and allow it proceed without possible lock order reversals.
1903315ee00fSMartin Matuska 	 */
1904315ee00fSMartin Matuska 	mutex_enter(&zilog->zl_lock);
1905315ee00fSMartin Matuska 	zil_lwb_set_zio_dependency(zilog, lwb);
1906315ee00fSMartin Matuska 	lwb->lwb_state = LWB_STATE_ISSUED;
1907315ee00fSMartin Matuska 
1908315ee00fSMartin Matuska 	if (nlwb) {
1909315ee00fSMartin Matuska 		nlwb->lwb_blk = *bp;
1910315ee00fSMartin Matuska 		nlwb->lwb_error = error;
1911315ee00fSMartin Matuska 		nlwb->lwb_slog = slog;
1912315ee00fSMartin Matuska 		nlwb->lwb_alloc_txg = txg;
1913315ee00fSMartin Matuska 		if (nlwb->lwb_state != LWB_STATE_READY)
1914315ee00fSMartin Matuska 			nlwb = NULL;
1915315ee00fSMartin Matuska 	}
1916315ee00fSMartin Matuska 	mutex_exit(&zilog->zl_lock);
1917eda14cbcSMatt Macy 
19184e8d558cSMartin Matuska 	if (lwb->lwb_slog) {
19194e8d558cSMartin Matuska 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count);
19204e8d558cSMartin Matuska 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes,
19214e8d558cSMartin Matuska 		    lwb->lwb_nused);
19224e8d558cSMartin Matuska 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_write,
19234e8d558cSMartin Matuska 		    wsz);
19244e8d558cSMartin Matuska 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_alloc,
19254e8d558cSMartin Matuska 		    BP_GET_LSIZE(&lwb->lwb_blk));
19264e8d558cSMartin Matuska 	} else {
19274e8d558cSMartin Matuska 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count);
19284e8d558cSMartin Matuska 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes,
19294e8d558cSMartin Matuska 		    lwb->lwb_nused);
19304e8d558cSMartin Matuska 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_write,
19314e8d558cSMartin Matuska 		    wsz);
19324e8d558cSMartin Matuska 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_alloc,
19334e8d558cSMartin Matuska 		    BP_GET_LSIZE(&lwb->lwb_blk));
19344e8d558cSMartin Matuska 	}
1935eda14cbcSMatt Macy 	lwb->lwb_issued_timestamp = gethrtime();
1936315ee00fSMartin Matuska 	if (lwb->lwb_child_zio)
1937315ee00fSMartin Matuska 		zio_nowait(lwb->lwb_child_zio);
19382ad756a6SMartin Matuska 	zio_nowait(lwb->lwb_write_zio);
19392ad756a6SMartin Matuska 	zio_nowait(lwb->lwb_root_zio);
1940315ee00fSMartin Matuska 
1941315ee00fSMartin Matuska 	/*
1942315ee00fSMartin Matuska 	 * If nlwb was ready when we gave it the block pointer,
1943315ee00fSMartin Matuska 	 * it is on us to issue it and possibly following ones.
1944315ee00fSMartin Matuska 	 */
1945315ee00fSMartin Matuska 	lwb = nlwb;
1946315ee00fSMartin Matuska 	if (lwb)
1947315ee00fSMartin Matuska 		goto next_lwb;
1948eda14cbcSMatt Macy }
1949eda14cbcSMatt Macy 
1950eda14cbcSMatt Macy /*
19512a58b312SMartin Matuska  * Maximum amount of data that can be put into single log block.
1952eda14cbcSMatt Macy  */
1953eda14cbcSMatt Macy uint64_t
19542a58b312SMartin Matuska zil_max_log_data(zilog_t *zilog, size_t hdrsize)
1955eda14cbcSMatt Macy {
19562a58b312SMartin Matuska 	return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize);
1957eda14cbcSMatt Macy }
1958eda14cbcSMatt Macy 
1959eda14cbcSMatt Macy /*
1960eda14cbcSMatt Macy  * Maximum amount of log space we agree to waste to reduce number of
1961b2526e8bSMartin Matuska  * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~6%).
1962eda14cbcSMatt Macy  */
1963eda14cbcSMatt Macy static inline uint64_t
1964eda14cbcSMatt Macy zil_max_waste_space(zilog_t *zilog)
1965eda14cbcSMatt Macy {
1966b2526e8bSMartin Matuska 	return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 16);
1967eda14cbcSMatt Macy }
1968eda14cbcSMatt Macy 
1969eda14cbcSMatt Macy /*
1970eda14cbcSMatt Macy  * Maximum amount of write data for WR_COPIED.  For correctness, consumers
1971eda14cbcSMatt Macy  * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1972eda14cbcSMatt Macy  * maximum sized log block, because each WR_COPIED record must fit in a
1973b2526e8bSMartin Matuska  * single log block.  Below that it is a tradeoff of additional memory copy
1974b2526e8bSMartin Matuska  * and possibly worse log space efficiency vs additional range lock/unlock.
1975eda14cbcSMatt Macy  */
1976b2526e8bSMartin Matuska static uint_t zil_maxcopied = 7680;
1977b2526e8bSMartin Matuska 
1978eda14cbcSMatt Macy uint64_t
1979eda14cbcSMatt Macy zil_max_copied_data(zilog_t *zilog)
1980eda14cbcSMatt Macy {
1981b2526e8bSMartin Matuska 	uint64_t max_data = zil_max_log_data(zilog, sizeof (lr_write_t));
1982b2526e8bSMartin Matuska 	return (MIN(max_data, zil_maxcopied));
1983eda14cbcSMatt Macy }
1984eda14cbcSMatt Macy 
19854e8d558cSMartin Matuska /*
19864e8d558cSMartin Matuska  * Estimate space needed in the lwb for the itx.  Allocate more lwbs or
19874e8d558cSMartin Matuska  * split the itx as needed, but don't touch the actual transaction data.
19884e8d558cSMartin Matuska  * Has to be called under zl_issuer_lock to call zil_lwb_write_close()
19894e8d558cSMartin Matuska  * to chain more lwbs.
19904e8d558cSMartin Matuska  */
1991eda14cbcSMatt Macy static lwb_t *
19924e8d558cSMartin Matuska zil_lwb_assign(zilog_t *zilog, lwb_t *lwb, itx_t *itx, list_t *ilwbs)
1993eda14cbcSMatt Macy {
19944e8d558cSMartin Matuska 	itx_t *citx;
19954e8d558cSMartin Matuska 	lr_t *lr, *clr;
19964e8d558cSMartin Matuska 	lr_write_t *lrw;
19974e8d558cSMartin Matuska 	uint64_t dlen, dnow, lwb_sp, reclen, max_log_data;
1998eda14cbcSMatt Macy 
1999eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2000eda14cbcSMatt Macy 	ASSERT3P(lwb, !=, NULL);
2001eda14cbcSMatt Macy 	ASSERT3P(lwb->lwb_buf, !=, NULL);
2002eda14cbcSMatt Macy 
2003eda14cbcSMatt Macy 	zil_lwb_write_open(zilog, lwb);
2004eda14cbcSMatt Macy 
20054e8d558cSMartin Matuska 	lr = &itx->itx_lr;
20064e8d558cSMartin Matuska 	lrw = (lr_write_t *)lr;
2007eda14cbcSMatt Macy 
2008eda14cbcSMatt Macy 	/*
2009eda14cbcSMatt Macy 	 * A commit itx doesn't represent any on-disk state; instead
2010eda14cbcSMatt Macy 	 * it's simply used as a place holder on the commit list, and
2011eda14cbcSMatt Macy 	 * provides a mechanism for attaching a "commit waiter" onto the
2012eda14cbcSMatt Macy 	 * correct lwb (such that the waiter can be signalled upon
2013eda14cbcSMatt Macy 	 * completion of that lwb). Thus, we don't process this itx's
2014eda14cbcSMatt Macy 	 * log record if it's a commit itx (these itx's don't have log
2015eda14cbcSMatt Macy 	 * records), and instead link the itx's waiter onto the lwb's
2016eda14cbcSMatt Macy 	 * list of waiters.
2017eda14cbcSMatt Macy 	 *
2018eda14cbcSMatt Macy 	 * For more details, see the comment above zil_commit().
2019eda14cbcSMatt Macy 	 */
20204e8d558cSMartin Matuska 	if (lr->lrc_txtype == TX_COMMIT) {
2021eda14cbcSMatt Macy 		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
20224e8d558cSMartin Matuska 		list_insert_tail(&lwb->lwb_itxs, itx);
2023eda14cbcSMatt Macy 		return (lwb);
2024eda14cbcSMatt Macy 	}
2025eda14cbcSMatt Macy 
20264e8d558cSMartin Matuska 	if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
2027eda14cbcSMatt Macy 		dlen = P2ROUNDUP_TYPED(
2028eda14cbcSMatt Macy 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
2029eda14cbcSMatt Macy 	} else {
20304e8d558cSMartin Matuska 		dlen = 0;
2031eda14cbcSMatt Macy 	}
20324e8d558cSMartin Matuska 	reclen = lr->lrc_reclen;
2033eda14cbcSMatt Macy 	zilog->zl_cur_used += (reclen + dlen);
2034eda14cbcSMatt Macy 
2035eda14cbcSMatt Macy cont:
2036eda14cbcSMatt Macy 	/*
2037eda14cbcSMatt Macy 	 * If this record won't fit in the current log block, start a new one.
2038eda14cbcSMatt Macy 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
2039eda14cbcSMatt Macy 	 */
2040315ee00fSMartin Matuska 	lwb_sp = lwb->lwb_nmax - lwb->lwb_nused;
20412a58b312SMartin Matuska 	max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t));
2042eda14cbcSMatt Macy 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
2043eda14cbcSMatt Macy 	    lwb_sp < zil_max_waste_space(zilog) &&
2044eda14cbcSMatt Macy 	    (dlen % max_log_data == 0 ||
2045eda14cbcSMatt Macy 	    lwb_sp < reclen + dlen % max_log_data))) {
2046315ee00fSMartin Matuska 		list_insert_tail(ilwbs, lwb);
2047315ee00fSMartin Matuska 		lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_OPENED);
2048eda14cbcSMatt Macy 		if (lwb == NULL)
2049eda14cbcSMatt Macy 			return (NULL);
2050315ee00fSMartin Matuska 		lwb_sp = lwb->lwb_nmax - lwb->lwb_nused;
2051eda14cbcSMatt Macy 
2052eda14cbcSMatt Macy 		/*
2053eda14cbcSMatt Macy 		 * There must be enough space in the new, empty log block to
2054eda14cbcSMatt Macy 		 * hold reclen.  For WR_COPIED, we need to fit the whole
2055eda14cbcSMatt Macy 		 * record in one block, and reclen is the header size + the
2056eda14cbcSMatt Macy 		 * data size. For WR_NEED_COPY, we can create multiple
2057eda14cbcSMatt Macy 		 * records, splitting the data into multiple blocks, so we
2058eda14cbcSMatt Macy 		 * only need to fit one word of data per block; in this case
2059eda14cbcSMatt Macy 		 * reclen is just the header size (no data).
2060eda14cbcSMatt Macy 		 */
2061eda14cbcSMatt Macy 		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
2062eda14cbcSMatt Macy 	}
2063eda14cbcSMatt Macy 
2064eda14cbcSMatt Macy 	dnow = MIN(dlen, lwb_sp - reclen);
20654e8d558cSMartin Matuska 	if (dlen > dnow) {
20664e8d558cSMartin Matuska 		ASSERT3U(lr->lrc_txtype, ==, TX_WRITE);
20674e8d558cSMartin Matuska 		ASSERT3U(itx->itx_wr_state, ==, WR_NEED_COPY);
20684e8d558cSMartin Matuska 		citx = zil_itx_clone(itx);
20694e8d558cSMartin Matuska 		clr = &citx->itx_lr;
20704e8d558cSMartin Matuska 		lr_write_t *clrw = (lr_write_t *)clr;
20714e8d558cSMartin Matuska 		clrw->lr_length = dnow;
20724e8d558cSMartin Matuska 		lrw->lr_offset += dnow;
20734e8d558cSMartin Matuska 		lrw->lr_length -= dnow;
20744e8d558cSMartin Matuska 	} else {
20754e8d558cSMartin Matuska 		citx = itx;
20764e8d558cSMartin Matuska 		clr = lr;
20774e8d558cSMartin Matuska 	}
20784e8d558cSMartin Matuska 
20794e8d558cSMartin Matuska 	/*
20804e8d558cSMartin Matuska 	 * We're actually making an entry, so update lrc_seq to be the
20814e8d558cSMartin Matuska 	 * log record sequence number.  Note that this is generally not
20824e8d558cSMartin Matuska 	 * equal to the itx sequence number because not all transactions
20834e8d558cSMartin Matuska 	 * are synchronous, and sometimes spa_sync() gets there first.
20844e8d558cSMartin Matuska 	 */
20854e8d558cSMartin Matuska 	clr->lrc_seq = ++zilog->zl_lr_seq;
20864e8d558cSMartin Matuska 
20874e8d558cSMartin Matuska 	lwb->lwb_nused += reclen + dnow;
2088315ee00fSMartin Matuska 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax);
20894e8d558cSMartin Matuska 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
20904e8d558cSMartin Matuska 
20914e8d558cSMartin Matuska 	zil_lwb_add_txg(lwb, lr->lrc_txg);
20924e8d558cSMartin Matuska 	list_insert_tail(&lwb->lwb_itxs, citx);
20934e8d558cSMartin Matuska 
20944e8d558cSMartin Matuska 	dlen -= dnow;
20954e8d558cSMartin Matuska 	if (dlen > 0) {
20964e8d558cSMartin Matuska 		zilog->zl_cur_used += reclen;
20974e8d558cSMartin Matuska 		goto cont;
20984e8d558cSMartin Matuska 	}
20994e8d558cSMartin Matuska 
2100315ee00fSMartin Matuska 	if (lr->lrc_txtype == TX_WRITE &&
2101315ee00fSMartin Matuska 	    lr->lrc_txg > spa_freeze_txg(zilog->zl_spa))
21024e8d558cSMartin Matuska 		txg_wait_synced(zilog->zl_dmu_pool, lr->lrc_txg);
21034e8d558cSMartin Matuska 
21044e8d558cSMartin Matuska 	return (lwb);
21054e8d558cSMartin Matuska }
21064e8d558cSMartin Matuska 
21074e8d558cSMartin Matuska /*
21084e8d558cSMartin Matuska  * Fill the actual transaction data into the lwb, following zil_lwb_assign().
21094e8d558cSMartin Matuska  * Does not require locking.
21104e8d558cSMartin Matuska  */
21114e8d558cSMartin Matuska static void
21124e8d558cSMartin Matuska zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx)
21134e8d558cSMartin Matuska {
21144e8d558cSMartin Matuska 	lr_t *lr, *lrb;
21154e8d558cSMartin Matuska 	lr_write_t *lrw, *lrwb;
21164e8d558cSMartin Matuska 	char *lr_buf;
21174e8d558cSMartin Matuska 	uint64_t dlen, reclen;
21184e8d558cSMartin Matuska 
21194e8d558cSMartin Matuska 	lr = &itx->itx_lr;
21204e8d558cSMartin Matuska 	lrw = (lr_write_t *)lr;
21214e8d558cSMartin Matuska 
21224e8d558cSMartin Matuska 	if (lr->lrc_txtype == TX_COMMIT)
21234e8d558cSMartin Matuska 		return;
21244e8d558cSMartin Matuska 
21254e8d558cSMartin Matuska 	if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
21264e8d558cSMartin Matuska 		dlen = P2ROUNDUP_TYPED(
21274e8d558cSMartin Matuska 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
21284e8d558cSMartin Matuska 	} else {
21294e8d558cSMartin Matuska 		dlen = 0;
21304e8d558cSMartin Matuska 	}
21314e8d558cSMartin Matuska 	reclen = lr->lrc_reclen;
21324e8d558cSMartin Matuska 	ASSERT3U(reclen + dlen, <=, lwb->lwb_nused - lwb->lwb_nfilled);
21334e8d558cSMartin Matuska 
21344e8d558cSMartin Matuska 	lr_buf = lwb->lwb_buf + lwb->lwb_nfilled;
21354e8d558cSMartin Matuska 	memcpy(lr_buf, lr, reclen);
21364e8d558cSMartin Matuska 	lrb = (lr_t *)lr_buf;		/* Like lr, but inside lwb. */
21374e8d558cSMartin Matuska 	lrwb = (lr_write_t *)lrb;	/* Like lrw, but inside lwb. */
2138eda14cbcSMatt Macy 
2139271171e0SMartin Matuska 	ZIL_STAT_BUMP(zilog, zil_itx_count);
2140eda14cbcSMatt Macy 
2141eda14cbcSMatt Macy 	/*
2142eda14cbcSMatt Macy 	 * If it's a write, fetch the data or get its blkptr as appropriate.
2143eda14cbcSMatt Macy 	 */
21444e8d558cSMartin Matuska 	if (lr->lrc_txtype == TX_WRITE) {
2145eda14cbcSMatt Macy 		if (itx->itx_wr_state == WR_COPIED) {
2146271171e0SMartin Matuska 			ZIL_STAT_BUMP(zilog, zil_itx_copied_count);
2147271171e0SMartin Matuska 			ZIL_STAT_INCR(zilog, zil_itx_copied_bytes,
2148271171e0SMartin Matuska 			    lrw->lr_length);
2149eda14cbcSMatt Macy 		} else {
2150eda14cbcSMatt Macy 			char *dbuf;
2151eda14cbcSMatt Macy 			int error;
2152eda14cbcSMatt Macy 
2153eda14cbcSMatt Macy 			if (itx->itx_wr_state == WR_NEED_COPY) {
2154eda14cbcSMatt Macy 				dbuf = lr_buf + reclen;
21554e8d558cSMartin Matuska 				lrb->lrc_reclen += dlen;
2156271171e0SMartin Matuska 				ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count);
2157271171e0SMartin Matuska 				ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes,
21584e8d558cSMartin Matuska 				    dlen);
2159eda14cbcSMatt Macy 			} else {
2160eda14cbcSMatt Macy 				ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
2161eda14cbcSMatt Macy 				dbuf = NULL;
2162271171e0SMartin Matuska 				ZIL_STAT_BUMP(zilog, zil_itx_indirect_count);
2163271171e0SMartin Matuska 				ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes,
2164eda14cbcSMatt Macy 				    lrw->lr_length);
2165315ee00fSMartin Matuska 				if (lwb->lwb_child_zio == NULL) {
2166315ee00fSMartin Matuska 					lwb->lwb_child_zio = zio_root(
2167315ee00fSMartin Matuska 					    zilog->zl_spa, NULL, NULL,
2168315ee00fSMartin Matuska 					    ZIO_FLAG_CANFAIL);
2169315ee00fSMartin Matuska 				}
2170eda14cbcSMatt Macy 			}
2171eda14cbcSMatt Macy 
2172eda14cbcSMatt Macy 			/*
2173315ee00fSMartin Matuska 			 * The "lwb_child_zio" we pass in will become a child of
2174315ee00fSMartin Matuska 			 * "lwb_write_zio", when one is created, so one will be
2175315ee00fSMartin Matuska 			 * a parent of any zio's created by the "zl_get_data".
2176315ee00fSMartin Matuska 			 * This way "lwb_write_zio" will first wait for children
2177315ee00fSMartin Matuska 			 * block pointers before own writing, and then for their
2178315ee00fSMartin Matuska 			 * writing completion before the vdev cache flushing.
2179eda14cbcSMatt Macy 			 */
2180eda14cbcSMatt Macy 			error = zilog->zl_get_data(itx->itx_private,
2181f9693befSMartin Matuska 			    itx->itx_gen, lrwb, dbuf, lwb,
2182315ee00fSMartin Matuska 			    lwb->lwb_child_zio);
21834e8d558cSMartin Matuska 			if (dbuf != NULL && error == 0) {
21841f88aa09SMartin Matuska 				/* Zero any padding bytes in the last block. */
21854e8d558cSMartin Matuska 				memset((char *)dbuf + lrwb->lr_length, 0,
21864e8d558cSMartin Matuska 				    dlen - lrwb->lr_length);
21874e8d558cSMartin Matuska 			}
2188eda14cbcSMatt Macy 
21892a58b312SMartin Matuska 			/*
21902a58b312SMartin Matuska 			 * Typically, the only return values we should see from
21912a58b312SMartin Matuska 			 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
21922a58b312SMartin Matuska 			 *  EALREADY. However, it is also possible to see other
21932a58b312SMartin Matuska 			 *  error values such as ENOSPC or EINVAL from
21942a58b312SMartin Matuska 			 *  dmu_read() -> dnode_hold() -> dnode_hold_impl() or
21952a58b312SMartin Matuska 			 *  ENXIO as well as a multitude of others from the
21962a58b312SMartin Matuska 			 *  block layer through dmu_buf_hold() -> dbuf_read()
21972a58b312SMartin Matuska 			 *  -> zio_wait(), as well as through dmu_read() ->
21982a58b312SMartin Matuska 			 *  dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
21992a58b312SMartin Matuska 			 *  zio_wait(). When these errors happen, we can assume
22002a58b312SMartin Matuska 			 *  that neither an immediate write nor an indirect
22012a58b312SMartin Matuska 			 *  write occurred, so we need to fall back to
22022a58b312SMartin Matuska 			 *  txg_wait_synced(). This is unusual, so we print to
22032a58b312SMartin Matuska 			 *  dmesg whenever one of these errors occurs.
22042a58b312SMartin Matuska 			 */
22052a58b312SMartin Matuska 			switch (error) {
22062a58b312SMartin Matuska 			case 0:
22072a58b312SMartin Matuska 				break;
22082a58b312SMartin Matuska 			default:
22092a58b312SMartin Matuska 				cmn_err(CE_WARN, "zil_lwb_commit() received "
22102a58b312SMartin Matuska 				    "unexpected error %d from ->zl_get_data()"
22112a58b312SMartin Matuska 				    ". Falling back to txg_wait_synced().",
22122a58b312SMartin Matuska 				    error);
22132a58b312SMartin Matuska 				zfs_fallthrough;
22142a58b312SMartin Matuska 			case EIO:
22154e8d558cSMartin Matuska 				txg_wait_synced(zilog->zl_dmu_pool,
22164e8d558cSMartin Matuska 				    lr->lrc_txg);
22172a58b312SMartin Matuska 				zfs_fallthrough;
22182a58b312SMartin Matuska 			case ENOENT:
22192a58b312SMartin Matuska 				zfs_fallthrough;
22202a58b312SMartin Matuska 			case EEXIST:
22212a58b312SMartin Matuska 				zfs_fallthrough;
22222a58b312SMartin Matuska 			case EALREADY:
22234e8d558cSMartin Matuska 				return;
2224eda14cbcSMatt Macy 			}
2225eda14cbcSMatt Macy 		}
2226eda14cbcSMatt Macy 	}
2227eda14cbcSMatt Macy 
22284e8d558cSMartin Matuska 	lwb->lwb_nfilled += reclen + dlen;
22294e8d558cSMartin Matuska 	ASSERT3S(lwb->lwb_nfilled, <=, lwb->lwb_nused);
22304e8d558cSMartin Matuska 	ASSERT0(P2PHASE(lwb->lwb_nfilled, sizeof (uint64_t)));
2231eda14cbcSMatt Macy }
2232eda14cbcSMatt Macy 
2233eda14cbcSMatt Macy itx_t *
22341f88aa09SMartin Matuska zil_itx_create(uint64_t txtype, size_t olrsize)
2235eda14cbcSMatt Macy {
22361f88aa09SMartin Matuska 	size_t itxsize, lrsize;
2237eda14cbcSMatt Macy 	itx_t *itx;
2238eda14cbcSMatt Macy 
22391f88aa09SMartin Matuska 	lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
2240eda14cbcSMatt Macy 	itxsize = offsetof(itx_t, itx_lr) + lrsize;
2241eda14cbcSMatt Macy 
2242eda14cbcSMatt Macy 	itx = zio_data_buf_alloc(itxsize);
2243eda14cbcSMatt Macy 	itx->itx_lr.lrc_txtype = txtype;
2244eda14cbcSMatt Macy 	itx->itx_lr.lrc_reclen = lrsize;
2245eda14cbcSMatt Macy 	itx->itx_lr.lrc_seq = 0;	/* defensive */
2246da5137abSMartin Matuska 	memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize);
2247eda14cbcSMatt Macy 	itx->itx_sync = B_TRUE;		/* default is synchronous */
2248eda14cbcSMatt Macy 	itx->itx_callback = NULL;
2249eda14cbcSMatt Macy 	itx->itx_callback_data = NULL;
2250eda14cbcSMatt Macy 	itx->itx_size = itxsize;
2251eda14cbcSMatt Macy 
2252eda14cbcSMatt Macy 	return (itx);
2253eda14cbcSMatt Macy }
2254eda14cbcSMatt Macy 
22554e8d558cSMartin Matuska static itx_t *
22564e8d558cSMartin Matuska zil_itx_clone(itx_t *oitx)
22574e8d558cSMartin Matuska {
22584e8d558cSMartin Matuska 	itx_t *itx = zio_data_buf_alloc(oitx->itx_size);
22594e8d558cSMartin Matuska 	memcpy(itx, oitx, oitx->itx_size);
22604e8d558cSMartin Matuska 	itx->itx_callback = NULL;
22614e8d558cSMartin Matuska 	itx->itx_callback_data = NULL;
22624e8d558cSMartin Matuska 	return (itx);
22634e8d558cSMartin Matuska }
22644e8d558cSMartin Matuska 
2265eda14cbcSMatt Macy void
2266eda14cbcSMatt Macy zil_itx_destroy(itx_t *itx)
2267eda14cbcSMatt Macy {
2268eda14cbcSMatt Macy 	IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
2269eda14cbcSMatt Macy 	IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2270eda14cbcSMatt Macy 
2271eda14cbcSMatt Macy 	if (itx->itx_callback != NULL)
2272eda14cbcSMatt Macy 		itx->itx_callback(itx->itx_callback_data);
2273eda14cbcSMatt Macy 
2274eda14cbcSMatt Macy 	zio_data_buf_free(itx, itx->itx_size);
2275eda14cbcSMatt Macy }
2276eda14cbcSMatt Macy 
2277eda14cbcSMatt Macy /*
2278eda14cbcSMatt Macy  * Free up the sync and async itxs. The itxs_t has already been detached
2279eda14cbcSMatt Macy  * so no locks are needed.
2280eda14cbcSMatt Macy  */
2281eda14cbcSMatt Macy static void
22823f9d360cSMartin Matuska zil_itxg_clean(void *arg)
2283eda14cbcSMatt Macy {
2284eda14cbcSMatt Macy 	itx_t *itx;
2285eda14cbcSMatt Macy 	list_t *list;
2286eda14cbcSMatt Macy 	avl_tree_t *t;
2287eda14cbcSMatt Macy 	void *cookie;
22883f9d360cSMartin Matuska 	itxs_t *itxs = arg;
2289eda14cbcSMatt Macy 	itx_async_node_t *ian;
2290eda14cbcSMatt Macy 
2291eda14cbcSMatt Macy 	list = &itxs->i_sync_list;
2292c0a83fe0SMartin Matuska 	while ((itx = list_remove_head(list)) != NULL) {
2293eda14cbcSMatt Macy 		/*
2294eda14cbcSMatt Macy 		 * In the general case, commit itxs will not be found
2295eda14cbcSMatt Macy 		 * here, as they'll be committed to an lwb via
22964e8d558cSMartin Matuska 		 * zil_lwb_assign(), and free'd in that function. Having
2297eda14cbcSMatt Macy 		 * said that, it is still possible for commit itxs to be
2298eda14cbcSMatt Macy 		 * found here, due to the following race:
2299eda14cbcSMatt Macy 		 *
2300eda14cbcSMatt Macy 		 *	- a thread calls zil_commit() which assigns the
2301eda14cbcSMatt Macy 		 *	  commit itx to a per-txg i_sync_list
2302eda14cbcSMatt Macy 		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
2303eda14cbcSMatt Macy 		 *	  while the waiter is still on the i_sync_list
2304eda14cbcSMatt Macy 		 *
2305eda14cbcSMatt Macy 		 * There's nothing to prevent syncing the txg while the
2306eda14cbcSMatt Macy 		 * waiter is on the i_sync_list. This normally doesn't
2307eda14cbcSMatt Macy 		 * happen because spa_sync() is slower than zil_commit(),
2308eda14cbcSMatt Macy 		 * but if zil_commit() calls txg_wait_synced() (e.g.
2309eda14cbcSMatt Macy 		 * because zil_create() or zil_commit_writer_stall() is
2310eda14cbcSMatt Macy 		 * called) we will hit this case.
2311eda14cbcSMatt Macy 		 */
2312eda14cbcSMatt Macy 		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
2313eda14cbcSMatt Macy 			zil_commit_waiter_skip(itx->itx_private);
2314eda14cbcSMatt Macy 
2315eda14cbcSMatt Macy 		zil_itx_destroy(itx);
2316eda14cbcSMatt Macy 	}
2317eda14cbcSMatt Macy 
2318eda14cbcSMatt Macy 	cookie = NULL;
2319eda14cbcSMatt Macy 	t = &itxs->i_async_tree;
2320eda14cbcSMatt Macy 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2321eda14cbcSMatt Macy 		list = &ian->ia_list;
2322c0a83fe0SMartin Matuska 		while ((itx = list_remove_head(list)) != NULL) {
2323eda14cbcSMatt Macy 			/* commit itxs should never be on the async lists. */
2324eda14cbcSMatt Macy 			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2325eda14cbcSMatt Macy 			zil_itx_destroy(itx);
2326eda14cbcSMatt Macy 		}
2327eda14cbcSMatt Macy 		list_destroy(list);
2328eda14cbcSMatt Macy 		kmem_free(ian, sizeof (itx_async_node_t));
2329eda14cbcSMatt Macy 	}
2330eda14cbcSMatt Macy 	avl_destroy(t);
2331eda14cbcSMatt Macy 
2332eda14cbcSMatt Macy 	kmem_free(itxs, sizeof (itxs_t));
2333eda14cbcSMatt Macy }
2334eda14cbcSMatt Macy 
2335eda14cbcSMatt Macy static int
2336eda14cbcSMatt Macy zil_aitx_compare(const void *x1, const void *x2)
2337eda14cbcSMatt Macy {
2338eda14cbcSMatt Macy 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
2339eda14cbcSMatt Macy 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
2340eda14cbcSMatt Macy 
2341eda14cbcSMatt Macy 	return (TREE_CMP(o1, o2));
2342eda14cbcSMatt Macy }
2343eda14cbcSMatt Macy 
2344eda14cbcSMatt Macy /*
2345eda14cbcSMatt Macy  * Remove all async itx with the given oid.
2346eda14cbcSMatt Macy  */
2347eda14cbcSMatt Macy void
2348eda14cbcSMatt Macy zil_remove_async(zilog_t *zilog, uint64_t oid)
2349eda14cbcSMatt Macy {
2350eda14cbcSMatt Macy 	uint64_t otxg, txg;
2351eda14cbcSMatt Macy 	itx_async_node_t *ian;
2352eda14cbcSMatt Macy 	avl_tree_t *t;
2353eda14cbcSMatt Macy 	avl_index_t where;
2354eda14cbcSMatt Macy 	list_t clean_list;
2355eda14cbcSMatt Macy 	itx_t *itx;
2356eda14cbcSMatt Macy 
2357eda14cbcSMatt Macy 	ASSERT(oid != 0);
2358eda14cbcSMatt Macy 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
2359eda14cbcSMatt Macy 
2360eda14cbcSMatt Macy 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2361eda14cbcSMatt Macy 		otxg = ZILTEST_TXG;
2362eda14cbcSMatt Macy 	else
2363eda14cbcSMatt Macy 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2364eda14cbcSMatt Macy 
2365eda14cbcSMatt Macy 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2366eda14cbcSMatt Macy 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2367eda14cbcSMatt Macy 
2368eda14cbcSMatt Macy 		mutex_enter(&itxg->itxg_lock);
2369eda14cbcSMatt Macy 		if (itxg->itxg_txg != txg) {
2370eda14cbcSMatt Macy 			mutex_exit(&itxg->itxg_lock);
2371eda14cbcSMatt Macy 			continue;
2372eda14cbcSMatt Macy 		}
2373eda14cbcSMatt Macy 
2374eda14cbcSMatt Macy 		/*
2375eda14cbcSMatt Macy 		 * Locate the object node and append its list.
2376eda14cbcSMatt Macy 		 */
2377eda14cbcSMatt Macy 		t = &itxg->itxg_itxs->i_async_tree;
2378eda14cbcSMatt Macy 		ian = avl_find(t, &oid, &where);
2379eda14cbcSMatt Macy 		if (ian != NULL)
2380eda14cbcSMatt Macy 			list_move_tail(&clean_list, &ian->ia_list);
2381eda14cbcSMatt Macy 		mutex_exit(&itxg->itxg_lock);
2382eda14cbcSMatt Macy 	}
2383c0a83fe0SMartin Matuska 	while ((itx = list_remove_head(&clean_list)) != NULL) {
2384eda14cbcSMatt Macy 		/* commit itxs should never be on the async lists. */
2385eda14cbcSMatt Macy 		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2386eda14cbcSMatt Macy 		zil_itx_destroy(itx);
2387eda14cbcSMatt Macy 	}
2388eda14cbcSMatt Macy 	list_destroy(&clean_list);
2389eda14cbcSMatt Macy }
2390eda14cbcSMatt Macy 
2391eda14cbcSMatt Macy void
2392eda14cbcSMatt Macy zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
2393eda14cbcSMatt Macy {
2394eda14cbcSMatt Macy 	uint64_t txg;
2395eda14cbcSMatt Macy 	itxg_t *itxg;
2396eda14cbcSMatt Macy 	itxs_t *itxs, *clean = NULL;
2397eda14cbcSMatt Macy 
2398eda14cbcSMatt Macy 	/*
2399eda14cbcSMatt Macy 	 * Ensure the data of a renamed file is committed before the rename.
2400eda14cbcSMatt Macy 	 */
2401eda14cbcSMatt Macy 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
2402eda14cbcSMatt Macy 		zil_async_to_sync(zilog, itx->itx_oid);
2403eda14cbcSMatt Macy 
2404eda14cbcSMatt Macy 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
2405eda14cbcSMatt Macy 		txg = ZILTEST_TXG;
2406eda14cbcSMatt Macy 	else
2407eda14cbcSMatt Macy 		txg = dmu_tx_get_txg(tx);
2408eda14cbcSMatt Macy 
2409eda14cbcSMatt Macy 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
2410eda14cbcSMatt Macy 	mutex_enter(&itxg->itxg_lock);
2411eda14cbcSMatt Macy 	itxs = itxg->itxg_itxs;
2412eda14cbcSMatt Macy 	if (itxg->itxg_txg != txg) {
2413eda14cbcSMatt Macy 		if (itxs != NULL) {
2414eda14cbcSMatt Macy 			/*
2415eda14cbcSMatt Macy 			 * The zil_clean callback hasn't got around to cleaning
2416eda14cbcSMatt Macy 			 * this itxg. Save the itxs for release below.
2417eda14cbcSMatt Macy 			 * This should be rare.
2418eda14cbcSMatt Macy 			 */
2419eda14cbcSMatt Macy 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
242033b8c039SMartin Matuska 			    "txg %llu", (u_longlong_t)itxg->itxg_txg);
2421eda14cbcSMatt Macy 			clean = itxg->itxg_itxs;
2422eda14cbcSMatt Macy 		}
2423eda14cbcSMatt Macy 		itxg->itxg_txg = txg;
2424eda14cbcSMatt Macy 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2425eda14cbcSMatt Macy 		    KM_SLEEP);
2426eda14cbcSMatt Macy 
2427eda14cbcSMatt Macy 		list_create(&itxs->i_sync_list, sizeof (itx_t),
2428eda14cbcSMatt Macy 		    offsetof(itx_t, itx_node));
2429eda14cbcSMatt Macy 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
2430eda14cbcSMatt Macy 		    sizeof (itx_async_node_t),
2431eda14cbcSMatt Macy 		    offsetof(itx_async_node_t, ia_node));
2432eda14cbcSMatt Macy 	}
2433eda14cbcSMatt Macy 	if (itx->itx_sync) {
2434eda14cbcSMatt Macy 		list_insert_tail(&itxs->i_sync_list, itx);
2435eda14cbcSMatt Macy 	} else {
2436eda14cbcSMatt Macy 		avl_tree_t *t = &itxs->i_async_tree;
2437eda14cbcSMatt Macy 		uint64_t foid =
2438eda14cbcSMatt Macy 		    LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2439eda14cbcSMatt Macy 		itx_async_node_t *ian;
2440eda14cbcSMatt Macy 		avl_index_t where;
2441eda14cbcSMatt Macy 
2442eda14cbcSMatt Macy 		ian = avl_find(t, &foid, &where);
2443eda14cbcSMatt Macy 		if (ian == NULL) {
2444eda14cbcSMatt Macy 			ian = kmem_alloc(sizeof (itx_async_node_t),
2445eda14cbcSMatt Macy 			    KM_SLEEP);
2446eda14cbcSMatt Macy 			list_create(&ian->ia_list, sizeof (itx_t),
2447eda14cbcSMatt Macy 			    offsetof(itx_t, itx_node));
2448eda14cbcSMatt Macy 			ian->ia_foid = foid;
2449eda14cbcSMatt Macy 			avl_insert(t, ian, where);
2450eda14cbcSMatt Macy 		}
2451eda14cbcSMatt Macy 		list_insert_tail(&ian->ia_list, itx);
2452eda14cbcSMatt Macy 	}
2453eda14cbcSMatt Macy 
2454eda14cbcSMatt Macy 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2455eda14cbcSMatt Macy 
2456eda14cbcSMatt Macy 	/*
2457eda14cbcSMatt Macy 	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2458eda14cbcSMatt Macy 	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2459eda14cbcSMatt Macy 	 * need to be careful to always dirty the ZIL using the "real"
2460eda14cbcSMatt Macy 	 * TXG (not itxg_txg) even when the SPA is frozen.
2461eda14cbcSMatt Macy 	 */
2462eda14cbcSMatt Macy 	zilog_dirty(zilog, dmu_tx_get_txg(tx));
2463eda14cbcSMatt Macy 	mutex_exit(&itxg->itxg_lock);
2464eda14cbcSMatt Macy 
2465eda14cbcSMatt Macy 	/* Release the old itxs now we've dropped the lock */
2466eda14cbcSMatt Macy 	if (clean != NULL)
2467eda14cbcSMatt Macy 		zil_itxg_clean(clean);
2468eda14cbcSMatt Macy }
2469eda14cbcSMatt Macy 
2470eda14cbcSMatt Macy /*
2471eda14cbcSMatt Macy  * If there are any in-memory intent log transactions which have now been
2472eda14cbcSMatt Macy  * synced then start up a taskq to free them. We should only do this after we
2473eda14cbcSMatt Macy  * have written out the uberblocks (i.e. txg has been committed) so that
2474eda14cbcSMatt Macy  * don't inadvertently clean out in-memory log records that would be required
2475eda14cbcSMatt Macy  * by zil_commit().
2476eda14cbcSMatt Macy  */
2477eda14cbcSMatt Macy void
2478eda14cbcSMatt Macy zil_clean(zilog_t *zilog, uint64_t synced_txg)
2479eda14cbcSMatt Macy {
2480eda14cbcSMatt Macy 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2481eda14cbcSMatt Macy 	itxs_t *clean_me;
2482eda14cbcSMatt Macy 
2483eda14cbcSMatt Macy 	ASSERT3U(synced_txg, <, ZILTEST_TXG);
2484eda14cbcSMatt Macy 
2485eda14cbcSMatt Macy 	mutex_enter(&itxg->itxg_lock);
2486eda14cbcSMatt Macy 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2487eda14cbcSMatt Macy 		mutex_exit(&itxg->itxg_lock);
2488eda14cbcSMatt Macy 		return;
2489eda14cbcSMatt Macy 	}
2490eda14cbcSMatt Macy 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2491eda14cbcSMatt Macy 	ASSERT3U(itxg->itxg_txg, !=, 0);
2492eda14cbcSMatt Macy 	clean_me = itxg->itxg_itxs;
2493eda14cbcSMatt Macy 	itxg->itxg_itxs = NULL;
2494eda14cbcSMatt Macy 	itxg->itxg_txg = 0;
2495eda14cbcSMatt Macy 	mutex_exit(&itxg->itxg_lock);
2496eda14cbcSMatt Macy 	/*
2497eda14cbcSMatt Macy 	 * Preferably start a task queue to free up the old itxs but
2498eda14cbcSMatt Macy 	 * if taskq_dispatch can't allocate resources to do that then
2499eda14cbcSMatt Macy 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
2500eda14cbcSMatt Macy 	 * created a bad performance problem.
2501eda14cbcSMatt Macy 	 */
2502eda14cbcSMatt Macy 	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2503eda14cbcSMatt Macy 	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2504eda14cbcSMatt Macy 	taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
25053f9d360cSMartin Matuska 	    zil_itxg_clean, clean_me, TQ_NOSLEEP);
2506eda14cbcSMatt Macy 	if (id == TASKQID_INVALID)
2507eda14cbcSMatt Macy 		zil_itxg_clean(clean_me);
2508eda14cbcSMatt Macy }
2509eda14cbcSMatt Macy 
2510eda14cbcSMatt Macy /*
2511eda14cbcSMatt Macy  * This function will traverse the queue of itxs that need to be
2512eda14cbcSMatt Macy  * committed, and move them onto the ZIL's zl_itx_commit_list.
2513eda14cbcSMatt Macy  */
25140a97523dSMartin Matuska static uint64_t
2515eda14cbcSMatt Macy zil_get_commit_list(zilog_t *zilog)
2516eda14cbcSMatt Macy {
25170a97523dSMartin Matuska 	uint64_t otxg, txg, wtxg = 0;
2518eda14cbcSMatt Macy 	list_t *commit_list = &zilog->zl_itx_commit_list;
2519eda14cbcSMatt Macy 
2520eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2521eda14cbcSMatt Macy 
2522eda14cbcSMatt Macy 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2523eda14cbcSMatt Macy 		otxg = ZILTEST_TXG;
2524eda14cbcSMatt Macy 	else
2525eda14cbcSMatt Macy 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2526eda14cbcSMatt Macy 
2527eda14cbcSMatt Macy 	/*
2528eda14cbcSMatt Macy 	 * This is inherently racy, since there is nothing to prevent
2529eda14cbcSMatt Macy 	 * the last synced txg from changing. That's okay since we'll
2530eda14cbcSMatt Macy 	 * only commit things in the future.
2531eda14cbcSMatt Macy 	 */
2532eda14cbcSMatt Macy 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2533eda14cbcSMatt Macy 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2534eda14cbcSMatt Macy 
2535eda14cbcSMatt Macy 		mutex_enter(&itxg->itxg_lock);
2536eda14cbcSMatt Macy 		if (itxg->itxg_txg != txg) {
2537eda14cbcSMatt Macy 			mutex_exit(&itxg->itxg_lock);
2538eda14cbcSMatt Macy 			continue;
2539eda14cbcSMatt Macy 		}
2540eda14cbcSMatt Macy 
2541eda14cbcSMatt Macy 		/*
2542eda14cbcSMatt Macy 		 * If we're adding itx records to the zl_itx_commit_list,
2543eda14cbcSMatt Macy 		 * then the zil better be dirty in this "txg". We can assert
2544eda14cbcSMatt Macy 		 * that here since we're holding the itxg_lock which will
2545eda14cbcSMatt Macy 		 * prevent spa_sync from cleaning it. Once we add the itxs
2546eda14cbcSMatt Macy 		 * to the zl_itx_commit_list we must commit it to disk even
2547eda14cbcSMatt Macy 		 * if it's unnecessary (i.e. the txg was synced).
2548eda14cbcSMatt Macy 		 */
2549eda14cbcSMatt Macy 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2550eda14cbcSMatt Macy 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
25510a97523dSMartin Matuska 		list_t *sync_list = &itxg->itxg_itxs->i_sync_list;
25520a97523dSMartin Matuska 		if (unlikely(zilog->zl_suspend > 0)) {
25530a97523dSMartin Matuska 			/*
25540a97523dSMartin Matuska 			 * ZIL was just suspended, but we lost the race.
25550a97523dSMartin Matuska 			 * Allow all earlier itxs to be committed, but ask
25560a97523dSMartin Matuska 			 * caller to do txg_wait_synced(txg) for any new.
25570a97523dSMartin Matuska 			 */
25580a97523dSMartin Matuska 			if (!list_is_empty(sync_list))
25590a97523dSMartin Matuska 				wtxg = MAX(wtxg, txg);
25600a97523dSMartin Matuska 		} else {
25610a97523dSMartin Matuska 			list_move_tail(commit_list, sync_list);
25620a97523dSMartin Matuska 		}
2563eda14cbcSMatt Macy 
2564eda14cbcSMatt Macy 		mutex_exit(&itxg->itxg_lock);
2565eda14cbcSMatt Macy 	}
25660a97523dSMartin Matuska 	return (wtxg);
2567eda14cbcSMatt Macy }
2568eda14cbcSMatt Macy 
2569eda14cbcSMatt Macy /*
2570eda14cbcSMatt Macy  * Move the async itxs for a specified object to commit into sync lists.
2571eda14cbcSMatt Macy  */
2572eda14cbcSMatt Macy void
2573eda14cbcSMatt Macy zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2574eda14cbcSMatt Macy {
2575eda14cbcSMatt Macy 	uint64_t otxg, txg;
2576eda14cbcSMatt Macy 	itx_async_node_t *ian;
2577eda14cbcSMatt Macy 	avl_tree_t *t;
2578eda14cbcSMatt Macy 	avl_index_t where;
2579eda14cbcSMatt Macy 
2580eda14cbcSMatt Macy 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2581eda14cbcSMatt Macy 		otxg = ZILTEST_TXG;
2582eda14cbcSMatt Macy 	else
2583eda14cbcSMatt Macy 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2584eda14cbcSMatt Macy 
2585eda14cbcSMatt Macy 	/*
2586eda14cbcSMatt Macy 	 * This is inherently racy, since there is nothing to prevent
2587eda14cbcSMatt Macy 	 * the last synced txg from changing.
2588eda14cbcSMatt Macy 	 */
2589eda14cbcSMatt Macy 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2590eda14cbcSMatt Macy 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2591eda14cbcSMatt Macy 
2592eda14cbcSMatt Macy 		mutex_enter(&itxg->itxg_lock);
2593eda14cbcSMatt Macy 		if (itxg->itxg_txg != txg) {
2594eda14cbcSMatt Macy 			mutex_exit(&itxg->itxg_lock);
2595eda14cbcSMatt Macy 			continue;
2596eda14cbcSMatt Macy 		}
2597eda14cbcSMatt Macy 
2598eda14cbcSMatt Macy 		/*
2599eda14cbcSMatt Macy 		 * If a foid is specified then find that node and append its
2600eda14cbcSMatt Macy 		 * list. Otherwise walk the tree appending all the lists
2601eda14cbcSMatt Macy 		 * to the sync list. We add to the end rather than the
2602eda14cbcSMatt Macy 		 * beginning to ensure the create has happened.
2603eda14cbcSMatt Macy 		 */
2604eda14cbcSMatt Macy 		t = &itxg->itxg_itxs->i_async_tree;
2605eda14cbcSMatt Macy 		if (foid != 0) {
2606eda14cbcSMatt Macy 			ian = avl_find(t, &foid, &where);
2607eda14cbcSMatt Macy 			if (ian != NULL) {
2608eda14cbcSMatt Macy 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2609eda14cbcSMatt Macy 				    &ian->ia_list);
2610eda14cbcSMatt Macy 			}
2611eda14cbcSMatt Macy 		} else {
2612eda14cbcSMatt Macy 			void *cookie = NULL;
2613eda14cbcSMatt Macy 
2614eda14cbcSMatt Macy 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2615eda14cbcSMatt Macy 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2616eda14cbcSMatt Macy 				    &ian->ia_list);
2617eda14cbcSMatt Macy 				list_destroy(&ian->ia_list);
2618eda14cbcSMatt Macy 				kmem_free(ian, sizeof (itx_async_node_t));
2619eda14cbcSMatt Macy 			}
2620eda14cbcSMatt Macy 		}
2621eda14cbcSMatt Macy 		mutex_exit(&itxg->itxg_lock);
2622eda14cbcSMatt Macy 	}
2623eda14cbcSMatt Macy }
2624eda14cbcSMatt Macy 
2625eda14cbcSMatt Macy /*
2626eda14cbcSMatt Macy  * This function will prune commit itxs that are at the head of the
2627eda14cbcSMatt Macy  * commit list (it won't prune past the first non-commit itx), and
2628eda14cbcSMatt Macy  * either: a) attach them to the last lwb that's still pending
2629eda14cbcSMatt Macy  * completion, or b) skip them altogether.
2630eda14cbcSMatt Macy  *
2631eda14cbcSMatt Macy  * This is used as a performance optimization to prevent commit itxs
2632eda14cbcSMatt Macy  * from generating new lwbs when it's unnecessary to do so.
2633eda14cbcSMatt Macy  */
2634eda14cbcSMatt Macy static void
2635eda14cbcSMatt Macy zil_prune_commit_list(zilog_t *zilog)
2636eda14cbcSMatt Macy {
2637eda14cbcSMatt Macy 	itx_t *itx;
2638eda14cbcSMatt Macy 
2639eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2640eda14cbcSMatt Macy 
2641eda14cbcSMatt Macy 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2642eda14cbcSMatt Macy 		lr_t *lrc = &itx->itx_lr;
2643eda14cbcSMatt Macy 		if (lrc->lrc_txtype != TX_COMMIT)
2644eda14cbcSMatt Macy 			break;
2645eda14cbcSMatt Macy 
2646eda14cbcSMatt Macy 		mutex_enter(&zilog->zl_lock);
2647eda14cbcSMatt Macy 
2648eda14cbcSMatt Macy 		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2649eda14cbcSMatt Macy 		if (last_lwb == NULL ||
2650eda14cbcSMatt Macy 		    last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2651eda14cbcSMatt Macy 			/*
2652eda14cbcSMatt Macy 			 * All of the itxs this waiter was waiting on
2653eda14cbcSMatt Macy 			 * must have already completed (or there were
2654eda14cbcSMatt Macy 			 * never any itx's for it to wait on), so it's
2655eda14cbcSMatt Macy 			 * safe to skip this waiter and mark it done.
2656eda14cbcSMatt Macy 			 */
2657eda14cbcSMatt Macy 			zil_commit_waiter_skip(itx->itx_private);
2658eda14cbcSMatt Macy 		} else {
2659eda14cbcSMatt Macy 			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2660eda14cbcSMatt Macy 		}
2661eda14cbcSMatt Macy 
2662eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_lock);
2663eda14cbcSMatt Macy 
2664eda14cbcSMatt Macy 		list_remove(&zilog->zl_itx_commit_list, itx);
2665eda14cbcSMatt Macy 		zil_itx_destroy(itx);
2666eda14cbcSMatt Macy 	}
2667eda14cbcSMatt Macy 
2668eda14cbcSMatt Macy 	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2669eda14cbcSMatt Macy }
2670eda14cbcSMatt Macy 
2671eda14cbcSMatt Macy static void
2672eda14cbcSMatt Macy zil_commit_writer_stall(zilog_t *zilog)
2673eda14cbcSMatt Macy {
2674eda14cbcSMatt Macy 	/*
2675eda14cbcSMatt Macy 	 * When zio_alloc_zil() fails to allocate the next lwb block on
2676eda14cbcSMatt Macy 	 * disk, we must call txg_wait_synced() to ensure all of the
2677eda14cbcSMatt Macy 	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2678eda14cbcSMatt Macy 	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2679eda14cbcSMatt Macy 	 * to zil_process_commit_list()) will have to call zil_create(),
2680eda14cbcSMatt Macy 	 * and start a new ZIL chain.
2681eda14cbcSMatt Macy 	 *
2682eda14cbcSMatt Macy 	 * Since zil_alloc_zil() failed, the lwb that was previously
2683eda14cbcSMatt Macy 	 * issued does not have a pointer to the "next" lwb on disk.
2684eda14cbcSMatt Macy 	 * Thus, if another ZIL writer thread was to allocate the "next"
2685eda14cbcSMatt Macy 	 * on-disk lwb, that block could be leaked in the event of a
2686eda14cbcSMatt Macy 	 * crash (because the previous lwb on-disk would not point to
2687eda14cbcSMatt Macy 	 * it).
2688eda14cbcSMatt Macy 	 *
2689eda14cbcSMatt Macy 	 * We must hold the zilog's zl_issuer_lock while we do this, to
2690eda14cbcSMatt Macy 	 * ensure no new threads enter zil_process_commit_list() until
2691eda14cbcSMatt Macy 	 * all lwb's in the zl_lwb_list have been synced and freed
2692eda14cbcSMatt Macy 	 * (which is achieved via the txg_wait_synced() call).
2693eda14cbcSMatt Macy 	 */
2694eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2695eda14cbcSMatt Macy 	txg_wait_synced(zilog->zl_dmu_pool, 0);
2696c0a83fe0SMartin Matuska 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2697eda14cbcSMatt Macy }
2698eda14cbcSMatt Macy 
2699eda14cbcSMatt Macy /*
2700eda14cbcSMatt Macy  * This function will traverse the commit list, creating new lwbs as
2701eda14cbcSMatt Macy  * needed, and committing the itxs from the commit list to these newly
2702eda14cbcSMatt Macy  * created lwbs. Additionally, as a new lwb is created, the previous
2703eda14cbcSMatt Macy  * lwb will be issued to the zio layer to be written to disk.
2704eda14cbcSMatt Macy  */
2705eda14cbcSMatt Macy static void
27064e8d558cSMartin Matuska zil_process_commit_list(zilog_t *zilog, zil_commit_waiter_t *zcw, list_t *ilwbs)
2707eda14cbcSMatt Macy {
2708eda14cbcSMatt Macy 	spa_t *spa = zilog->zl_spa;
2709eda14cbcSMatt Macy 	list_t nolwb_itxs;
2710eda14cbcSMatt Macy 	list_t nolwb_waiters;
271115f0b8c3SMartin Matuska 	lwb_t *lwb, *plwb;
2712eda14cbcSMatt Macy 	itx_t *itx;
271315f0b8c3SMartin Matuska 	boolean_t first = B_TRUE;
2714eda14cbcSMatt Macy 
2715eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2716eda14cbcSMatt Macy 
2717eda14cbcSMatt Macy 	/*
2718eda14cbcSMatt Macy 	 * Return if there's nothing to commit before we dirty the fs by
2719eda14cbcSMatt Macy 	 * calling zil_create().
2720eda14cbcSMatt Macy 	 */
2721c0a83fe0SMartin Matuska 	if (list_is_empty(&zilog->zl_itx_commit_list))
2722eda14cbcSMatt Macy 		return;
2723eda14cbcSMatt Macy 
2724eda14cbcSMatt Macy 	list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2725eda14cbcSMatt Macy 	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2726eda14cbcSMatt Macy 	    offsetof(zil_commit_waiter_t, zcw_node));
2727eda14cbcSMatt Macy 
2728eda14cbcSMatt Macy 	lwb = list_tail(&zilog->zl_lwb_list);
2729eda14cbcSMatt Macy 	if (lwb == NULL) {
2730eda14cbcSMatt Macy 		lwb = zil_create(zilog);
2731eda14cbcSMatt Macy 	} else {
2732c03c5b1cSMartin Matuska 		/*
2733c03c5b1cSMartin Matuska 		 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2734c03c5b1cSMartin Matuska 		 * have already been created (zl_lwb_list not empty).
2735c03c5b1cSMartin Matuska 		 */
2736c03c5b1cSMartin Matuska 		zil_commit_activate_saxattr_feature(zilog);
2737315ee00fSMartin Matuska 		ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
2738315ee00fSMartin Matuska 		    lwb->lwb_state == LWB_STATE_OPENED);
2739315ee00fSMartin Matuska 		first = (lwb->lwb_state == LWB_STATE_NEW) &&
274015f0b8c3SMartin Matuska 		    ((plwb = list_prev(&zilog->zl_lwb_list, lwb)) == NULL ||
274115f0b8c3SMartin Matuska 		    plwb->lwb_state == LWB_STATE_FLUSH_DONE);
2742eda14cbcSMatt Macy 	}
2743eda14cbcSMatt Macy 
2744c0a83fe0SMartin Matuska 	while ((itx = list_remove_head(&zilog->zl_itx_commit_list)) != NULL) {
2745eda14cbcSMatt Macy 		lr_t *lrc = &itx->itx_lr;
2746eda14cbcSMatt Macy 		uint64_t txg = lrc->lrc_txg;
2747eda14cbcSMatt Macy 
2748eda14cbcSMatt Macy 		ASSERT3U(txg, !=, 0);
2749eda14cbcSMatt Macy 
2750eda14cbcSMatt Macy 		if (lrc->lrc_txtype == TX_COMMIT) {
2751eda14cbcSMatt Macy 			DTRACE_PROBE2(zil__process__commit__itx,
2752eda14cbcSMatt Macy 			    zilog_t *, zilog, itx_t *, itx);
2753eda14cbcSMatt Macy 		} else {
2754eda14cbcSMatt Macy 			DTRACE_PROBE2(zil__process__normal__itx,
2755eda14cbcSMatt Macy 			    zilog_t *, zilog, itx_t *, itx);
2756eda14cbcSMatt Macy 		}
2757eda14cbcSMatt Macy 
2758eda14cbcSMatt Macy 		boolean_t synced = txg <= spa_last_synced_txg(spa);
2759eda14cbcSMatt Macy 		boolean_t frozen = txg > spa_freeze_txg(spa);
2760eda14cbcSMatt Macy 
2761eda14cbcSMatt Macy 		/*
2762eda14cbcSMatt Macy 		 * If the txg of this itx has already been synced out, then
2763eda14cbcSMatt Macy 		 * we don't need to commit this itx to an lwb. This is
2764eda14cbcSMatt Macy 		 * because the data of this itx will have already been
2765eda14cbcSMatt Macy 		 * written to the main pool. This is inherently racy, and
2766eda14cbcSMatt Macy 		 * it's still ok to commit an itx whose txg has already
2767eda14cbcSMatt Macy 		 * been synced; this will result in a write that's
2768eda14cbcSMatt Macy 		 * unnecessary, but will do no harm.
2769eda14cbcSMatt Macy 		 *
2770eda14cbcSMatt Macy 		 * With that said, we always want to commit TX_COMMIT itxs
2771eda14cbcSMatt Macy 		 * to an lwb, regardless of whether or not that itx's txg
2772eda14cbcSMatt Macy 		 * has been synced out. We do this to ensure any OPENED lwb
2773eda14cbcSMatt Macy 		 * will always have at least one zil_commit_waiter_t linked
2774eda14cbcSMatt Macy 		 * to the lwb.
2775eda14cbcSMatt Macy 		 *
2776eda14cbcSMatt Macy 		 * As a counter-example, if we skipped TX_COMMIT itx's
2777eda14cbcSMatt Macy 		 * whose txg had already been synced, the following
2778eda14cbcSMatt Macy 		 * situation could occur if we happened to be racing with
2779eda14cbcSMatt Macy 		 * spa_sync:
2780eda14cbcSMatt Macy 		 *
2781eda14cbcSMatt Macy 		 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2782eda14cbcSMatt Macy 		 *    itx's txg is 10 and the last synced txg is 9.
2783eda14cbcSMatt Macy 		 * 2. spa_sync finishes syncing out txg 10.
2784eda14cbcSMatt Macy 		 * 3. We move to the next itx in the list, it's a TX_COMMIT
2785eda14cbcSMatt Macy 		 *    whose txg is 10, so we skip it rather than committing
2786eda14cbcSMatt Macy 		 *    it to the lwb used in (1).
2787eda14cbcSMatt Macy 		 *
2788eda14cbcSMatt Macy 		 * If the itx that is skipped in (3) is the last TX_COMMIT
2789eda14cbcSMatt Macy 		 * itx in the commit list, than it's possible for the lwb
2790eda14cbcSMatt Macy 		 * used in (1) to remain in the OPENED state indefinitely.
2791eda14cbcSMatt Macy 		 *
2792eda14cbcSMatt Macy 		 * To prevent the above scenario from occurring, ensuring
2793eda14cbcSMatt Macy 		 * that once an lwb is OPENED it will transition to ISSUED
2794eda14cbcSMatt Macy 		 * and eventually DONE, we always commit TX_COMMIT itx's to
2795eda14cbcSMatt Macy 		 * an lwb here, even if that itx's txg has already been
2796eda14cbcSMatt Macy 		 * synced.
2797eda14cbcSMatt Macy 		 *
2798eda14cbcSMatt Macy 		 * Finally, if the pool is frozen, we _always_ commit the
2799eda14cbcSMatt Macy 		 * itx.  The point of freezing the pool is to prevent data
2800eda14cbcSMatt Macy 		 * from being written to the main pool via spa_sync, and
2801eda14cbcSMatt Macy 		 * instead rely solely on the ZIL to persistently store the
2802eda14cbcSMatt Macy 		 * data; i.e.  when the pool is frozen, the last synced txg
2803eda14cbcSMatt Macy 		 * value can't be trusted.
2804eda14cbcSMatt Macy 		 */
2805eda14cbcSMatt Macy 		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2806eda14cbcSMatt Macy 			if (lwb != NULL) {
28074e8d558cSMartin Matuska 				lwb = zil_lwb_assign(zilog, lwb, itx, ilwbs);
28084e8d558cSMartin Matuska 				if (lwb == NULL) {
2809eda14cbcSMatt Macy 					list_insert_tail(&nolwb_itxs, itx);
28104e8d558cSMartin Matuska 				} else if ((zcw->zcw_lwb != NULL &&
28114e8d558cSMartin Matuska 				    zcw->zcw_lwb != lwb) || zcw->zcw_done) {
28124e8d558cSMartin Matuska 					/*
28134e8d558cSMartin Matuska 					 * Our lwb is done, leave the rest of
28144e8d558cSMartin Matuska 					 * itx list to somebody else who care.
28154e8d558cSMartin Matuska 					 */
28164e8d558cSMartin Matuska 					first = B_FALSE;
28174e8d558cSMartin Matuska 					break;
28184e8d558cSMartin Matuska 				}
2819eda14cbcSMatt Macy 			} else {
2820eda14cbcSMatt Macy 				if (lrc->lrc_txtype == TX_COMMIT) {
2821eda14cbcSMatt Macy 					zil_commit_waiter_link_nolwb(
2822eda14cbcSMatt Macy 					    itx->itx_private, &nolwb_waiters);
2823eda14cbcSMatt Macy 				}
2824eda14cbcSMatt Macy 				list_insert_tail(&nolwb_itxs, itx);
2825eda14cbcSMatt Macy 			}
2826eda14cbcSMatt Macy 		} else {
2827eda14cbcSMatt Macy 			ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2828eda14cbcSMatt Macy 			zil_itx_destroy(itx);
2829eda14cbcSMatt Macy 		}
2830eda14cbcSMatt Macy 	}
2831eda14cbcSMatt Macy 
2832eda14cbcSMatt Macy 	if (lwb == NULL) {
2833eda14cbcSMatt Macy 		/*
2834eda14cbcSMatt Macy 		 * This indicates zio_alloc_zil() failed to allocate the
2835eda14cbcSMatt Macy 		 * "next" lwb on-disk. When this happens, we must stall
2836eda14cbcSMatt Macy 		 * the ZIL write pipeline; see the comment within
2837eda14cbcSMatt Macy 		 * zil_commit_writer_stall() for more details.
2838eda14cbcSMatt Macy 		 */
28394e8d558cSMartin Matuska 		while ((lwb = list_remove_head(ilwbs)) != NULL)
28404e8d558cSMartin Matuska 			zil_lwb_write_issue(zilog, lwb);
2841eda14cbcSMatt Macy 		zil_commit_writer_stall(zilog);
2842eda14cbcSMatt Macy 
2843eda14cbcSMatt Macy 		/*
2844eda14cbcSMatt Macy 		 * Additionally, we have to signal and mark the "nolwb"
2845eda14cbcSMatt Macy 		 * waiters as "done" here, since without an lwb, we
2846eda14cbcSMatt Macy 		 * can't do this via zil_lwb_flush_vdevs_done() like
2847eda14cbcSMatt Macy 		 * normal.
2848eda14cbcSMatt Macy 		 */
2849eda14cbcSMatt Macy 		zil_commit_waiter_t *zcw;
2850c0a83fe0SMartin Matuska 		while ((zcw = list_remove_head(&nolwb_waiters)) != NULL)
2851eda14cbcSMatt Macy 			zil_commit_waiter_skip(zcw);
2852eda14cbcSMatt Macy 
2853eda14cbcSMatt Macy 		/*
2854eda14cbcSMatt Macy 		 * And finally, we have to destroy the itx's that
2855eda14cbcSMatt Macy 		 * couldn't be committed to an lwb; this will also call
2856eda14cbcSMatt Macy 		 * the itx's callback if one exists for the itx.
2857eda14cbcSMatt Macy 		 */
2858c0a83fe0SMartin Matuska 		while ((itx = list_remove_head(&nolwb_itxs)) != NULL)
2859eda14cbcSMatt Macy 			zil_itx_destroy(itx);
2860eda14cbcSMatt Macy 	} else {
2861eda14cbcSMatt Macy 		ASSERT(list_is_empty(&nolwb_waiters));
2862eda14cbcSMatt Macy 		ASSERT3P(lwb, !=, NULL);
2863315ee00fSMartin Matuska 		ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
2864315ee00fSMartin Matuska 		    lwb->lwb_state == LWB_STATE_OPENED);
2865eda14cbcSMatt Macy 
2866eda14cbcSMatt Macy 		/*
2867eda14cbcSMatt Macy 		 * At this point, the ZIL block pointed at by the "lwb"
2868315ee00fSMartin Matuska 		 * variable is in "new" or "opened" state.
2869eda14cbcSMatt Macy 		 *
2870315ee00fSMartin Matuska 		 * If it's "new", then no itxs have been committed to it, so
2871315ee00fSMartin Matuska 		 * there's no point in issuing its zio (i.e. it's "empty").
2872eda14cbcSMatt Macy 		 *
2873315ee00fSMartin Matuska 		 * If it's "opened", then it contains one or more itxs that
2874eda14cbcSMatt Macy 		 * eventually need to be committed to stable storage. In
2875eda14cbcSMatt Macy 		 * this case we intentionally do not issue the lwb's zio
2876eda14cbcSMatt Macy 		 * to disk yet, and instead rely on one of the following
2877eda14cbcSMatt Macy 		 * two mechanisms for issuing the zio:
2878eda14cbcSMatt Macy 		 *
2879315ee00fSMartin Matuska 		 * 1. Ideally, there will be more ZIL activity occurring on
2880315ee00fSMartin Matuska 		 * the system, such that this function will be immediately
2881315ee00fSMartin Matuska 		 * called again by different thread and this lwb will be
2882315ee00fSMartin Matuska 		 * closed by zil_lwb_assign().  This way, the lwb will be
2883315ee00fSMartin Matuska 		 * "full" when it is issued to disk, and we'll make use of
2884315ee00fSMartin Matuska 		 * the lwb's size the best we can.
2885eda14cbcSMatt Macy 		 *
2886eda14cbcSMatt Macy 		 * 2. If there isn't sufficient ZIL activity occurring on
2887315ee00fSMartin Matuska 		 * the system, zil_commit_waiter() will close it and issue
2888315ee00fSMartin Matuska 		 * the zio.  If this occurs, the lwb is not guaranteed
2889eda14cbcSMatt Macy 		 * to be "full" by the time its zio is issued, and means
2890eda14cbcSMatt Macy 		 * the size of the lwb was "too large" given the amount
2891eda14cbcSMatt Macy 		 * of ZIL activity occurring on the system at that time.
2892eda14cbcSMatt Macy 		 *
2893eda14cbcSMatt Macy 		 * We do this for a couple of reasons:
2894eda14cbcSMatt Macy 		 *
2895eda14cbcSMatt Macy 		 * 1. To try and reduce the number of IOPs needed to
2896eda14cbcSMatt Macy 		 * write the same number of itxs. If an lwb has space
2897eda14cbcSMatt Macy 		 * available in its buffer for more itxs, and more itxs
2898eda14cbcSMatt Macy 		 * will be committed relatively soon (relative to the
2899eda14cbcSMatt Macy 		 * latency of performing a write), then it's beneficial
2900eda14cbcSMatt Macy 		 * to wait for these "next" itxs. This way, more itxs
2901eda14cbcSMatt Macy 		 * can be committed to stable storage with fewer writes.
2902eda14cbcSMatt Macy 		 *
2903eda14cbcSMatt Macy 		 * 2. To try and use the largest lwb block size that the
2904eda14cbcSMatt Macy 		 * incoming rate of itxs can support. Again, this is to
2905eda14cbcSMatt Macy 		 * try and pack as many itxs into as few lwbs as
2906eda14cbcSMatt Macy 		 * possible, without significantly impacting the latency
2907eda14cbcSMatt Macy 		 * of each individual itx.
290815f0b8c3SMartin Matuska 		 *
290915f0b8c3SMartin Matuska 		 * If we had no already running or open LWBs, it can be
291015f0b8c3SMartin Matuska 		 * the workload is single-threaded.  And if the ZIL write
291115f0b8c3SMartin Matuska 		 * latency is very small or if the LWB is almost full, it
291215f0b8c3SMartin Matuska 		 * may be cheaper to bypass the delay.
2913eda14cbcSMatt Macy 		 */
291415f0b8c3SMartin Matuska 		if (lwb->lwb_state == LWB_STATE_OPENED && first) {
291515f0b8c3SMartin Matuska 			hrtime_t sleep = zilog->zl_last_lwb_latency *
291615f0b8c3SMartin Matuska 			    zfs_commit_timeout_pct / 100;
291715f0b8c3SMartin Matuska 			if (sleep < zil_min_commit_timeout ||
2918315ee00fSMartin Matuska 			    lwb->lwb_nmax - lwb->lwb_nused <
2919315ee00fSMartin Matuska 			    lwb->lwb_nmax / 8) {
2920315ee00fSMartin Matuska 				list_insert_tail(ilwbs, lwb);
2921315ee00fSMartin Matuska 				lwb = zil_lwb_write_close(zilog, lwb,
2922315ee00fSMartin Matuska 				    LWB_STATE_NEW);
292315f0b8c3SMartin Matuska 				zilog->zl_cur_used = 0;
29244e8d558cSMartin Matuska 				if (lwb == NULL) {
29254e8d558cSMartin Matuska 					while ((lwb = list_remove_head(ilwbs))
29264e8d558cSMartin Matuska 					    != NULL)
29274e8d558cSMartin Matuska 						zil_lwb_write_issue(zilog, lwb);
292815f0b8c3SMartin Matuska 					zil_commit_writer_stall(zilog);
292915f0b8c3SMartin Matuska 				}
293015f0b8c3SMartin Matuska 			}
2931eda14cbcSMatt Macy 		}
2932eda14cbcSMatt Macy 	}
29334e8d558cSMartin Matuska }
2934eda14cbcSMatt Macy 
2935eda14cbcSMatt Macy /*
2936eda14cbcSMatt Macy  * This function is responsible for ensuring the passed in commit waiter
2937eda14cbcSMatt Macy  * (and associated commit itx) is committed to an lwb. If the waiter is
2938eda14cbcSMatt Macy  * not already committed to an lwb, all itxs in the zilog's queue of
2939eda14cbcSMatt Macy  * itxs will be processed. The assumption is the passed in waiter's
2940eda14cbcSMatt Macy  * commit itx will found in the queue just like the other non-commit
2941eda14cbcSMatt Macy  * itxs, such that when the entire queue is processed, the waiter will
2942eda14cbcSMatt Macy  * have been committed to an lwb.
2943eda14cbcSMatt Macy  *
2944eda14cbcSMatt Macy  * The lwb associated with the passed in waiter is not guaranteed to
2945eda14cbcSMatt Macy  * have been issued by the time this function completes. If the lwb is
2946eda14cbcSMatt Macy  * not issued, we rely on future calls to zil_commit_writer() to issue
2947eda14cbcSMatt Macy  * the lwb, or the timeout mechanism found in zil_commit_waiter().
2948eda14cbcSMatt Macy  */
29490a97523dSMartin Matuska static uint64_t
2950eda14cbcSMatt Macy zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2951eda14cbcSMatt Macy {
29524e8d558cSMartin Matuska 	list_t ilwbs;
29534e8d558cSMartin Matuska 	lwb_t *lwb;
29540a97523dSMartin Matuska 	uint64_t wtxg = 0;
29554e8d558cSMartin Matuska 
2956eda14cbcSMatt Macy 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2957eda14cbcSMatt Macy 	ASSERT(spa_writeable(zilog->zl_spa));
2958eda14cbcSMatt Macy 
29594e8d558cSMartin Matuska 	list_create(&ilwbs, sizeof (lwb_t), offsetof(lwb_t, lwb_issue_node));
2960eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_issuer_lock);
2961eda14cbcSMatt Macy 
2962eda14cbcSMatt Macy 	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2963eda14cbcSMatt Macy 		/*
2964eda14cbcSMatt Macy 		 * It's possible that, while we were waiting to acquire
2965eda14cbcSMatt Macy 		 * the "zl_issuer_lock", another thread committed this
2966eda14cbcSMatt Macy 		 * waiter to an lwb. If that occurs, we bail out early,
2967eda14cbcSMatt Macy 		 * without processing any of the zilog's queue of itxs.
2968eda14cbcSMatt Macy 		 *
2969eda14cbcSMatt Macy 		 * On certain workloads and system configurations, the
2970eda14cbcSMatt Macy 		 * "zl_issuer_lock" can become highly contended. In an
2971eda14cbcSMatt Macy 		 * attempt to reduce this contention, we immediately drop
2972eda14cbcSMatt Macy 		 * the lock if the waiter has already been processed.
2973eda14cbcSMatt Macy 		 *
2974eda14cbcSMatt Macy 		 * We've measured this optimization to reduce CPU spent
2975eda14cbcSMatt Macy 		 * contending on this lock by up to 5%, using a system
2976eda14cbcSMatt Macy 		 * with 32 CPUs, low latency storage (~50 usec writes),
2977eda14cbcSMatt Macy 		 * and 1024 threads performing sync writes.
2978eda14cbcSMatt Macy 		 */
2979eda14cbcSMatt Macy 		goto out;
2980eda14cbcSMatt Macy 	}
2981eda14cbcSMatt Macy 
2982271171e0SMartin Matuska 	ZIL_STAT_BUMP(zilog, zil_commit_writer_count);
2983eda14cbcSMatt Macy 
29840a97523dSMartin Matuska 	wtxg = zil_get_commit_list(zilog);
2985eda14cbcSMatt Macy 	zil_prune_commit_list(zilog);
29864e8d558cSMartin Matuska 	zil_process_commit_list(zilog, zcw, &ilwbs);
2987eda14cbcSMatt Macy 
2988eda14cbcSMatt Macy out:
2989eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_issuer_lock);
29904e8d558cSMartin Matuska 	while ((lwb = list_remove_head(&ilwbs)) != NULL)
29914e8d558cSMartin Matuska 		zil_lwb_write_issue(zilog, lwb);
29924e8d558cSMartin Matuska 	list_destroy(&ilwbs);
29930a97523dSMartin Matuska 	return (wtxg);
2994eda14cbcSMatt Macy }
2995eda14cbcSMatt Macy 
2996eda14cbcSMatt Macy static void
2997eda14cbcSMatt Macy zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
2998eda14cbcSMatt Macy {
2999eda14cbcSMatt Macy 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3000eda14cbcSMatt Macy 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3001eda14cbcSMatt Macy 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
3002eda14cbcSMatt Macy 
3003eda14cbcSMatt Macy 	lwb_t *lwb = zcw->zcw_lwb;
3004eda14cbcSMatt Macy 	ASSERT3P(lwb, !=, NULL);
3005315ee00fSMartin Matuska 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW);
3006eda14cbcSMatt Macy 
3007eda14cbcSMatt Macy 	/*
3008eda14cbcSMatt Macy 	 * If the lwb has already been issued by another thread, we can
3009eda14cbcSMatt Macy 	 * immediately return since there's no work to be done (the
3010eda14cbcSMatt Macy 	 * point of this function is to issue the lwb). Additionally, we
3011eda14cbcSMatt Macy 	 * do this prior to acquiring the zl_issuer_lock, to avoid
3012eda14cbcSMatt Macy 	 * acquiring it when it's not necessary to do so.
3013eda14cbcSMatt Macy 	 */
3014315ee00fSMartin Matuska 	if (lwb->lwb_state != LWB_STATE_OPENED)
3015eda14cbcSMatt Macy 		return;
3016eda14cbcSMatt Macy 
3017eda14cbcSMatt Macy 	/*
30184e8d558cSMartin Matuska 	 * In order to call zil_lwb_write_close() we must hold the
3019eda14cbcSMatt Macy 	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
3020eda14cbcSMatt Macy 	 * since we're already holding the commit waiter's "zcw_lock",
3021eda14cbcSMatt Macy 	 * and those two locks are acquired in the opposite order
3022eda14cbcSMatt Macy 	 * elsewhere.
3023eda14cbcSMatt Macy 	 */
3024eda14cbcSMatt Macy 	mutex_exit(&zcw->zcw_lock);
3025eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_issuer_lock);
3026eda14cbcSMatt Macy 	mutex_enter(&zcw->zcw_lock);
3027eda14cbcSMatt Macy 
3028eda14cbcSMatt Macy 	/*
3029eda14cbcSMatt Macy 	 * Since we just dropped and re-acquired the commit waiter's
3030eda14cbcSMatt Macy 	 * lock, we have to re-check to see if the waiter was marked
3031eda14cbcSMatt Macy 	 * "done" during that process. If the waiter was marked "done",
3032eda14cbcSMatt Macy 	 * the "lwb" pointer is no longer valid (it can be free'd after
3033eda14cbcSMatt Macy 	 * the waiter is marked "done"), so without this check we could
3034eda14cbcSMatt Macy 	 * wind up with a use-after-free error below.
3035eda14cbcSMatt Macy 	 */
30364e8d558cSMartin Matuska 	if (zcw->zcw_done) {
3037315ee00fSMartin Matuska 		mutex_exit(&zilog->zl_issuer_lock);
3038315ee00fSMartin Matuska 		return;
30394e8d558cSMartin Matuska 	}
3040eda14cbcSMatt Macy 
3041eda14cbcSMatt Macy 	ASSERT3P(lwb, ==, zcw->zcw_lwb);
3042eda14cbcSMatt Macy 
3043eda14cbcSMatt Macy 	/*
3044eda14cbcSMatt Macy 	 * We've already checked this above, but since we hadn't acquired
3045eda14cbcSMatt Macy 	 * the zilog's zl_issuer_lock, we have to perform this check a
3046eda14cbcSMatt Macy 	 * second time while holding the lock.
3047eda14cbcSMatt Macy 	 *
3048eda14cbcSMatt Macy 	 * We don't need to hold the zl_lock since the lwb cannot transition
3049315ee00fSMartin Matuska 	 * from OPENED to CLOSED while we hold the zl_issuer_lock. The lwb
3050315ee00fSMartin Matuska 	 * _can_ transition from CLOSED to DONE, but it's OK to race with
3051eda14cbcSMatt Macy 	 * that transition since we treat the lwb the same, whether it's in
3052315ee00fSMartin Matuska 	 * the CLOSED, ISSUED or DONE states.
3053eda14cbcSMatt Macy 	 *
3054eda14cbcSMatt Macy 	 * The important thing, is we treat the lwb differently depending on
3055315ee00fSMartin Matuska 	 * if it's OPENED or CLOSED, and block any other threads that might
3056315ee00fSMartin Matuska 	 * attempt to close/issue this lwb. For that reason we hold the
3057eda14cbcSMatt Macy 	 * zl_issuer_lock when checking the lwb_state; we must not call
3058315ee00fSMartin Matuska 	 * zil_lwb_write_close() if the lwb had already been closed/issued.
3059eda14cbcSMatt Macy 	 *
3060eda14cbcSMatt Macy 	 * See the comment above the lwb_state_t structure definition for
3061eda14cbcSMatt Macy 	 * more details on the lwb states, and locking requirements.
3062eda14cbcSMatt Macy 	 */
3063315ee00fSMartin Matuska 	if (lwb->lwb_state != LWB_STATE_OPENED) {
3064315ee00fSMartin Matuska 		mutex_exit(&zilog->zl_issuer_lock);
3065315ee00fSMartin Matuska 		return;
30664e8d558cSMartin Matuska 	}
3067eda14cbcSMatt Macy 
3068315ee00fSMartin Matuska 	/*
3069315ee00fSMartin Matuska 	 * We do not need zcw_lock once we hold zl_issuer_lock and know lwb
3070315ee00fSMartin Matuska 	 * is still open.  But we have to drop it to avoid a deadlock in case
3071315ee00fSMartin Matuska 	 * callback of zio issued by zil_lwb_write_issue() try to get it,
3072315ee00fSMartin Matuska 	 * while zil_lwb_write_issue() is blocked on attempt to issue next
3073315ee00fSMartin Matuska 	 * lwb it found in LWB_STATE_READY state.
3074315ee00fSMartin Matuska 	 */
3075315ee00fSMartin Matuska 	mutex_exit(&zcw->zcw_lock);
3076eda14cbcSMatt Macy 
3077eda14cbcSMatt Macy 	/*
3078eda14cbcSMatt Macy 	 * As described in the comments above zil_commit_waiter() and
3079eda14cbcSMatt Macy 	 * zil_process_commit_list(), we need to issue this lwb's zio
3080eda14cbcSMatt Macy 	 * since we've reached the commit waiter's timeout and it still
3081eda14cbcSMatt Macy 	 * hasn't been issued.
3082eda14cbcSMatt Macy 	 */
3083315ee00fSMartin Matuska 	lwb_t *nlwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW);
3084eda14cbcSMatt Macy 
3085315ee00fSMartin Matuska 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED);
3086eda14cbcSMatt Macy 
3087eda14cbcSMatt Macy 	/*
3088eda14cbcSMatt Macy 	 * Since the lwb's zio hadn't been issued by the time this thread
3089eda14cbcSMatt Macy 	 * reached its timeout, we reset the zilog's "zl_cur_used" field
3090eda14cbcSMatt Macy 	 * to influence the zil block size selection algorithm.
3091eda14cbcSMatt Macy 	 *
3092eda14cbcSMatt Macy 	 * By having to issue the lwb's zio here, it means the size of the
3093eda14cbcSMatt Macy 	 * lwb was too large, given the incoming throughput of itxs.  By
3094eda14cbcSMatt Macy 	 * setting "zl_cur_used" to zero, we communicate this fact to the
3095eda14cbcSMatt Macy 	 * block size selection algorithm, so it can take this information
3096eda14cbcSMatt Macy 	 * into account, and potentially select a smaller size for the
3097eda14cbcSMatt Macy 	 * next lwb block that is allocated.
3098eda14cbcSMatt Macy 	 */
3099eda14cbcSMatt Macy 	zilog->zl_cur_used = 0;
3100eda14cbcSMatt Macy 
3101eda14cbcSMatt Macy 	if (nlwb == NULL) {
3102eda14cbcSMatt Macy 		/*
31034e8d558cSMartin Matuska 		 * When zil_lwb_write_close() returns NULL, this
3104eda14cbcSMatt Macy 		 * indicates zio_alloc_zil() failed to allocate the
3105eda14cbcSMatt Macy 		 * "next" lwb on-disk. When this occurs, the ZIL write
3106eda14cbcSMatt Macy 		 * pipeline must be stalled; see the comment within the
3107eda14cbcSMatt Macy 		 * zil_commit_writer_stall() function for more details.
3108eda14cbcSMatt Macy 		 */
31094e8d558cSMartin Matuska 		zil_lwb_write_issue(zilog, lwb);
3110eda14cbcSMatt Macy 		zil_commit_writer_stall(zilog);
3111eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_issuer_lock);
3112315ee00fSMartin Matuska 	} else {
3113315ee00fSMartin Matuska 		mutex_exit(&zilog->zl_issuer_lock);
31144e8d558cSMartin Matuska 		zil_lwb_write_issue(zilog, lwb);
3115315ee00fSMartin Matuska 	}
3116315ee00fSMartin Matuska 	mutex_enter(&zcw->zcw_lock);
3117eda14cbcSMatt Macy }
3118eda14cbcSMatt Macy 
3119eda14cbcSMatt Macy /*
3120eda14cbcSMatt Macy  * This function is responsible for performing the following two tasks:
3121eda14cbcSMatt Macy  *
3122eda14cbcSMatt Macy  * 1. its primary responsibility is to block until the given "commit
3123eda14cbcSMatt Macy  *    waiter" is considered "done".
3124eda14cbcSMatt Macy  *
3125eda14cbcSMatt Macy  * 2. its secondary responsibility is to issue the zio for the lwb that
3126eda14cbcSMatt Macy  *    the given "commit waiter" is waiting on, if this function has
3127eda14cbcSMatt Macy  *    waited "long enough" and the lwb is still in the "open" state.
3128eda14cbcSMatt Macy  *
3129eda14cbcSMatt Macy  * Given a sufficient amount of itxs being generated and written using
31304e8d558cSMartin Matuska  * the ZIL, the lwb's zio will be issued via the zil_lwb_assign()
3131eda14cbcSMatt Macy  * function. If this does not occur, this secondary responsibility will
3132eda14cbcSMatt Macy  * ensure the lwb is issued even if there is not other synchronous
3133eda14cbcSMatt Macy  * activity on the system.
3134eda14cbcSMatt Macy  *
3135eda14cbcSMatt Macy  * For more details, see zil_process_commit_list(); more specifically,
3136eda14cbcSMatt Macy  * the comment at the bottom of that function.
3137eda14cbcSMatt Macy  */
3138eda14cbcSMatt Macy static void
3139eda14cbcSMatt Macy zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
3140eda14cbcSMatt Macy {
3141eda14cbcSMatt Macy 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3142eda14cbcSMatt Macy 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3143eda14cbcSMatt Macy 	ASSERT(spa_writeable(zilog->zl_spa));
3144eda14cbcSMatt Macy 
3145eda14cbcSMatt Macy 	mutex_enter(&zcw->zcw_lock);
3146eda14cbcSMatt Macy 
3147eda14cbcSMatt Macy 	/*
3148eda14cbcSMatt Macy 	 * The timeout is scaled based on the lwb latency to avoid
3149eda14cbcSMatt Macy 	 * significantly impacting the latency of each individual itx.
3150eda14cbcSMatt Macy 	 * For more details, see the comment at the bottom of the
3151eda14cbcSMatt Macy 	 * zil_process_commit_list() function.
3152eda14cbcSMatt Macy 	 */
3153eda14cbcSMatt Macy 	int pct = MAX(zfs_commit_timeout_pct, 1);
3154eda14cbcSMatt Macy 	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
3155eda14cbcSMatt Macy 	hrtime_t wakeup = gethrtime() + sleep;
3156eda14cbcSMatt Macy 	boolean_t timedout = B_FALSE;
3157eda14cbcSMatt Macy 
3158eda14cbcSMatt Macy 	while (!zcw->zcw_done) {
3159eda14cbcSMatt Macy 		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3160eda14cbcSMatt Macy 
3161eda14cbcSMatt Macy 		lwb_t *lwb = zcw->zcw_lwb;
3162eda14cbcSMatt Macy 
3163eda14cbcSMatt Macy 		/*
3164eda14cbcSMatt Macy 		 * Usually, the waiter will have a non-NULL lwb field here,
3165eda14cbcSMatt Macy 		 * but it's possible for it to be NULL as a result of
3166eda14cbcSMatt Macy 		 * zil_commit() racing with spa_sync().
3167eda14cbcSMatt Macy 		 *
3168eda14cbcSMatt Macy 		 * When zil_clean() is called, it's possible for the itxg
3169eda14cbcSMatt Macy 		 * list (which may be cleaned via a taskq) to contain
3170eda14cbcSMatt Macy 		 * commit itxs. When this occurs, the commit waiters linked
3171eda14cbcSMatt Macy 		 * off of these commit itxs will not be committed to an
3172eda14cbcSMatt Macy 		 * lwb.  Additionally, these commit waiters will not be
3173eda14cbcSMatt Macy 		 * marked done until zil_commit_waiter_skip() is called via
3174eda14cbcSMatt Macy 		 * zil_itxg_clean().
3175eda14cbcSMatt Macy 		 *
3176eda14cbcSMatt Macy 		 * Thus, it's possible for this commit waiter (i.e. the
3177eda14cbcSMatt Macy 		 * "zcw" variable) to be found in this "in between" state;
3178eda14cbcSMatt Macy 		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3179eda14cbcSMatt Macy 		 * been skipped, so it's "zcw_done" field is still B_FALSE.
3180eda14cbcSMatt Macy 		 */
3181315ee00fSMartin Matuska 		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_NEW);
3182eda14cbcSMatt Macy 
3183eda14cbcSMatt Macy 		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
3184eda14cbcSMatt Macy 			ASSERT3B(timedout, ==, B_FALSE);
3185eda14cbcSMatt Macy 
3186eda14cbcSMatt Macy 			/*
3187eda14cbcSMatt Macy 			 * If the lwb hasn't been issued yet, then we
3188eda14cbcSMatt Macy 			 * need to wait with a timeout, in case this
3189eda14cbcSMatt Macy 			 * function needs to issue the lwb after the
3190eda14cbcSMatt Macy 			 * timeout is reached; responsibility (2) from
3191eda14cbcSMatt Macy 			 * the comment above this function.
3192eda14cbcSMatt Macy 			 */
3193eda14cbcSMatt Macy 			int rc = cv_timedwait_hires(&zcw->zcw_cv,
3194eda14cbcSMatt Macy 			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
3195eda14cbcSMatt Macy 			    CALLOUT_FLAG_ABSOLUTE);
3196eda14cbcSMatt Macy 
3197eda14cbcSMatt Macy 			if (rc != -1 || zcw->zcw_done)
3198eda14cbcSMatt Macy 				continue;
3199eda14cbcSMatt Macy 
3200eda14cbcSMatt Macy 			timedout = B_TRUE;
3201eda14cbcSMatt Macy 			zil_commit_waiter_timeout(zilog, zcw);
3202eda14cbcSMatt Macy 
3203eda14cbcSMatt Macy 			if (!zcw->zcw_done) {
3204eda14cbcSMatt Macy 				/*
3205eda14cbcSMatt Macy 				 * If the commit waiter has already been
3206eda14cbcSMatt Macy 				 * marked "done", it's possible for the
3207eda14cbcSMatt Macy 				 * waiter's lwb structure to have already
3208eda14cbcSMatt Macy 				 * been freed.  Thus, we can only reliably
3209eda14cbcSMatt Macy 				 * make these assertions if the waiter
3210eda14cbcSMatt Macy 				 * isn't done.
3211eda14cbcSMatt Macy 				 */
3212eda14cbcSMatt Macy 				ASSERT3P(lwb, ==, zcw->zcw_lwb);
3213eda14cbcSMatt Macy 				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
3214eda14cbcSMatt Macy 			}
3215eda14cbcSMatt Macy 		} else {
3216eda14cbcSMatt Macy 			/*
3217eda14cbcSMatt Macy 			 * If the lwb isn't open, then it must have already
3218eda14cbcSMatt Macy 			 * been issued. In that case, there's no need to
3219eda14cbcSMatt Macy 			 * use a timeout when waiting for the lwb to
3220eda14cbcSMatt Macy 			 * complete.
3221eda14cbcSMatt Macy 			 *
3222eda14cbcSMatt Macy 			 * Additionally, if the lwb is NULL, the waiter
3223eda14cbcSMatt Macy 			 * will soon be signaled and marked done via
3224eda14cbcSMatt Macy 			 * zil_clean() and zil_itxg_clean(), so no timeout
3225eda14cbcSMatt Macy 			 * is required.
3226eda14cbcSMatt Macy 			 */
3227eda14cbcSMatt Macy 
3228eda14cbcSMatt Macy 			IMPLY(lwb != NULL,
3229315ee00fSMartin Matuska 			    lwb->lwb_state == LWB_STATE_CLOSED ||
3230315ee00fSMartin Matuska 			    lwb->lwb_state == LWB_STATE_READY ||
3231eda14cbcSMatt Macy 			    lwb->lwb_state == LWB_STATE_ISSUED ||
3232eda14cbcSMatt Macy 			    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3233eda14cbcSMatt Macy 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
3234eda14cbcSMatt Macy 			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
3235eda14cbcSMatt Macy 		}
3236eda14cbcSMatt Macy 	}
3237eda14cbcSMatt Macy 
3238eda14cbcSMatt Macy 	mutex_exit(&zcw->zcw_lock);
3239eda14cbcSMatt Macy }
3240eda14cbcSMatt Macy 
3241eda14cbcSMatt Macy static zil_commit_waiter_t *
3242eda14cbcSMatt Macy zil_alloc_commit_waiter(void)
3243eda14cbcSMatt Macy {
3244eda14cbcSMatt Macy 	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
3245eda14cbcSMatt Macy 
3246eda14cbcSMatt Macy 	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
3247eda14cbcSMatt Macy 	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
3248eda14cbcSMatt Macy 	list_link_init(&zcw->zcw_node);
3249eda14cbcSMatt Macy 	zcw->zcw_lwb = NULL;
3250eda14cbcSMatt Macy 	zcw->zcw_done = B_FALSE;
3251eda14cbcSMatt Macy 	zcw->zcw_zio_error = 0;
3252eda14cbcSMatt Macy 
3253eda14cbcSMatt Macy 	return (zcw);
3254eda14cbcSMatt Macy }
3255eda14cbcSMatt Macy 
3256eda14cbcSMatt Macy static void
3257eda14cbcSMatt Macy zil_free_commit_waiter(zil_commit_waiter_t *zcw)
3258eda14cbcSMatt Macy {
3259eda14cbcSMatt Macy 	ASSERT(!list_link_active(&zcw->zcw_node));
3260eda14cbcSMatt Macy 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
3261eda14cbcSMatt Macy 	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
3262eda14cbcSMatt Macy 	mutex_destroy(&zcw->zcw_lock);
3263eda14cbcSMatt Macy 	cv_destroy(&zcw->zcw_cv);
3264eda14cbcSMatt Macy 	kmem_cache_free(zil_zcw_cache, zcw);
3265eda14cbcSMatt Macy }
3266eda14cbcSMatt Macy 
3267eda14cbcSMatt Macy /*
3268eda14cbcSMatt Macy  * This function is used to create a TX_COMMIT itx and assign it. This
3269eda14cbcSMatt Macy  * way, it will be linked into the ZIL's list of synchronous itxs, and
3270eda14cbcSMatt Macy  * then later committed to an lwb (or skipped) when
3271eda14cbcSMatt Macy  * zil_process_commit_list() is called.
3272eda14cbcSMatt Macy  */
3273eda14cbcSMatt Macy static void
3274eda14cbcSMatt Macy zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
3275eda14cbcSMatt Macy {
3276eda14cbcSMatt Macy 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
3277d411c1d6SMartin Matuska 
3278d411c1d6SMartin Matuska 	/*
3279d411c1d6SMartin Matuska 	 * Since we are not going to create any new dirty data, and we
3280d411c1d6SMartin Matuska 	 * can even help with clearing the existing dirty data, we
3281d411c1d6SMartin Matuska 	 * should not be subject to the dirty data based delays. We
3282d411c1d6SMartin Matuska 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
3283d411c1d6SMartin Matuska 	 */
3284d411c1d6SMartin Matuska 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
3285eda14cbcSMatt Macy 
3286eda14cbcSMatt Macy 	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
3287eda14cbcSMatt Macy 	itx->itx_sync = B_TRUE;
3288eda14cbcSMatt Macy 	itx->itx_private = zcw;
3289eda14cbcSMatt Macy 
3290eda14cbcSMatt Macy 	zil_itx_assign(zilog, itx, tx);
3291eda14cbcSMatt Macy 
3292eda14cbcSMatt Macy 	dmu_tx_commit(tx);
3293eda14cbcSMatt Macy }
3294eda14cbcSMatt Macy 
3295eda14cbcSMatt Macy /*
3296eda14cbcSMatt Macy  * Commit ZFS Intent Log transactions (itxs) to stable storage.
3297eda14cbcSMatt Macy  *
3298eda14cbcSMatt Macy  * When writing ZIL transactions to the on-disk representation of the
3299eda14cbcSMatt Macy  * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3300eda14cbcSMatt Macy  * itxs can be committed to a single lwb. Once a lwb is written and
3301eda14cbcSMatt Macy  * committed to stable storage (i.e. the lwb is written, and vdevs have
3302eda14cbcSMatt Macy  * been flushed), each itx that was committed to that lwb is also
3303eda14cbcSMatt Macy  * considered to be committed to stable storage.
3304eda14cbcSMatt Macy  *
3305eda14cbcSMatt Macy  * When an itx is committed to an lwb, the log record (lr_t) contained
3306eda14cbcSMatt Macy  * by the itx is copied into the lwb's zio buffer, and once this buffer
3307eda14cbcSMatt Macy  * is written to disk, it becomes an on-disk ZIL block.
3308eda14cbcSMatt Macy  *
3309eda14cbcSMatt Macy  * As itxs are generated, they're inserted into the ZIL's queue of
3310eda14cbcSMatt Macy  * uncommitted itxs. The semantics of zil_commit() are such that it will
3311eda14cbcSMatt Macy  * block until all itxs that were in the queue when it was called, are
3312eda14cbcSMatt Macy  * committed to stable storage.
3313eda14cbcSMatt Macy  *
3314eda14cbcSMatt Macy  * If "foid" is zero, this means all "synchronous" and "asynchronous"
3315eda14cbcSMatt Macy  * itxs, for all objects in the dataset, will be committed to stable
3316eda14cbcSMatt Macy  * storage prior to zil_commit() returning. If "foid" is non-zero, all
3317eda14cbcSMatt Macy  * "synchronous" itxs for all objects, but only "asynchronous" itxs
3318eda14cbcSMatt Macy  * that correspond to the foid passed in, will be committed to stable
3319eda14cbcSMatt Macy  * storage prior to zil_commit() returning.
3320eda14cbcSMatt Macy  *
3321eda14cbcSMatt Macy  * Generally speaking, when zil_commit() is called, the consumer doesn't
3322eda14cbcSMatt Macy  * actually care about _all_ of the uncommitted itxs. Instead, they're
3323eda14cbcSMatt Macy  * simply trying to waiting for a specific itx to be committed to disk,
3324eda14cbcSMatt Macy  * but the interface(s) for interacting with the ZIL don't allow such
3325eda14cbcSMatt Macy  * fine-grained communication. A better interface would allow a consumer
3326eda14cbcSMatt Macy  * to create and assign an itx, and then pass a reference to this itx to
3327eda14cbcSMatt Macy  * zil_commit(); such that zil_commit() would return as soon as that
3328eda14cbcSMatt Macy  * specific itx was committed to disk (instead of waiting for _all_
3329eda14cbcSMatt Macy  * itxs to be committed).
3330eda14cbcSMatt Macy  *
3331eda14cbcSMatt Macy  * When a thread calls zil_commit() a special "commit itx" will be
3332eda14cbcSMatt Macy  * generated, along with a corresponding "waiter" for this commit itx.
3333eda14cbcSMatt Macy  * zil_commit() will wait on this waiter's CV, such that when the waiter
3334eda14cbcSMatt Macy  * is marked done, and signaled, zil_commit() will return.
3335eda14cbcSMatt Macy  *
3336eda14cbcSMatt Macy  * This commit itx is inserted into the queue of uncommitted itxs. This
3337eda14cbcSMatt Macy  * provides an easy mechanism for determining which itxs were in the
3338eda14cbcSMatt Macy  * queue prior to zil_commit() having been called, and which itxs were
3339eda14cbcSMatt Macy  * added after zil_commit() was called.
3340eda14cbcSMatt Macy  *
3341e3aa18adSMartin Matuska  * The commit itx is special; it doesn't have any on-disk representation.
3342eda14cbcSMatt Macy  * When a commit itx is "committed" to an lwb, the waiter associated
3343eda14cbcSMatt Macy  * with it is linked onto the lwb's list of waiters. Then, when that lwb
3344eda14cbcSMatt Macy  * completes, each waiter on the lwb's list is marked done and signaled
3345eda14cbcSMatt Macy  * -- allowing the thread waiting on the waiter to return from zil_commit().
3346eda14cbcSMatt Macy  *
3347eda14cbcSMatt Macy  * It's important to point out a few critical factors that allow us
3348eda14cbcSMatt Macy  * to make use of the commit itxs, commit waiters, per-lwb lists of
3349eda14cbcSMatt Macy  * commit waiters, and zio completion callbacks like we're doing:
3350eda14cbcSMatt Macy  *
3351eda14cbcSMatt Macy  *   1. The list of waiters for each lwb is traversed, and each commit
3352eda14cbcSMatt Macy  *      waiter is marked "done" and signaled, in the zio completion
3353eda14cbcSMatt Macy  *      callback of the lwb's zio[*].
3354eda14cbcSMatt Macy  *
3355eda14cbcSMatt Macy  *      * Actually, the waiters are signaled in the zio completion
3356eda14cbcSMatt Macy  *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3357eda14cbcSMatt Macy  *        that are sent to the vdevs upon completion of the lwb zio.
3358eda14cbcSMatt Macy  *
3359eda14cbcSMatt Macy  *   2. When the itxs are inserted into the ZIL's queue of uncommitted
3360eda14cbcSMatt Macy  *      itxs, the order in which they are inserted is preserved[*]; as
3361eda14cbcSMatt Macy  *      itxs are added to the queue, they are added to the tail of
3362eda14cbcSMatt Macy  *      in-memory linked lists.
3363eda14cbcSMatt Macy  *
3364eda14cbcSMatt Macy  *      When committing the itxs to lwbs (to be written to disk), they
3365eda14cbcSMatt Macy  *      are committed in the same order in which the itxs were added to
3366eda14cbcSMatt Macy  *      the uncommitted queue's linked list(s); i.e. the linked list of
3367eda14cbcSMatt Macy  *      itxs to commit is traversed from head to tail, and each itx is
3368eda14cbcSMatt Macy  *      committed to an lwb in that order.
3369eda14cbcSMatt Macy  *
3370eda14cbcSMatt Macy  *      * To clarify:
3371eda14cbcSMatt Macy  *
3372eda14cbcSMatt Macy  *        - the order of "sync" itxs is preserved w.r.t. other
3373eda14cbcSMatt Macy  *          "sync" itxs, regardless of the corresponding objects.
3374eda14cbcSMatt Macy  *        - the order of "async" itxs is preserved w.r.t. other
3375eda14cbcSMatt Macy  *          "async" itxs corresponding to the same object.
3376eda14cbcSMatt Macy  *        - the order of "async" itxs is *not* preserved w.r.t. other
3377eda14cbcSMatt Macy  *          "async" itxs corresponding to different objects.
3378eda14cbcSMatt Macy  *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
3379eda14cbcSMatt Macy  *          versa) is *not* preserved, even for itxs that correspond
3380eda14cbcSMatt Macy  *          to the same object.
3381eda14cbcSMatt Macy  *
3382eda14cbcSMatt Macy  *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
3383eda14cbcSMatt Macy  *      zil_get_commit_list(), and zil_process_commit_list().
3384eda14cbcSMatt Macy  *
3385eda14cbcSMatt Macy  *   3. The lwbs represent a linked list of blocks on disk. Thus, any
3386eda14cbcSMatt Macy  *      lwb cannot be considered committed to stable storage, until its
3387eda14cbcSMatt Macy  *      "previous" lwb is also committed to stable storage. This fact,
3388eda14cbcSMatt Macy  *      coupled with the fact described above, means that itxs are
3389eda14cbcSMatt Macy  *      committed in (roughly) the order in which they were generated.
3390eda14cbcSMatt Macy  *      This is essential because itxs are dependent on prior itxs.
3391eda14cbcSMatt Macy  *      Thus, we *must not* deem an itx as being committed to stable
3392eda14cbcSMatt Macy  *      storage, until *all* prior itxs have also been committed to
3393eda14cbcSMatt Macy  *      stable storage.
3394eda14cbcSMatt Macy  *
3395eda14cbcSMatt Macy  *      To enforce this ordering of lwb zio's, while still leveraging as
3396eda14cbcSMatt Macy  *      much of the underlying storage performance as possible, we rely
3397eda14cbcSMatt Macy  *      on two fundamental concepts:
3398eda14cbcSMatt Macy  *
3399eda14cbcSMatt Macy  *          1. The creation and issuance of lwb zio's is protected by
3400eda14cbcSMatt Macy  *             the zilog's "zl_issuer_lock", which ensures only a single
3401eda14cbcSMatt Macy  *             thread is creating and/or issuing lwb's at a time
3402eda14cbcSMatt Macy  *          2. The "previous" lwb is a child of the "current" lwb
3403eda14cbcSMatt Macy  *             (leveraging the zio parent-child dependency graph)
3404eda14cbcSMatt Macy  *
3405eda14cbcSMatt Macy  *      By relying on this parent-child zio relationship, we can have
3406eda14cbcSMatt Macy  *      many lwb zio's concurrently issued to the underlying storage,
3407eda14cbcSMatt Macy  *      but the order in which they complete will be the same order in
3408eda14cbcSMatt Macy  *      which they were created.
3409eda14cbcSMatt Macy  */
3410eda14cbcSMatt Macy void
3411eda14cbcSMatt Macy zil_commit(zilog_t *zilog, uint64_t foid)
3412eda14cbcSMatt Macy {
3413eda14cbcSMatt Macy 	/*
3414eda14cbcSMatt Macy 	 * We should never attempt to call zil_commit on a snapshot for
3415eda14cbcSMatt Macy 	 * a couple of reasons:
3416eda14cbcSMatt Macy 	 *
3417eda14cbcSMatt Macy 	 * 1. A snapshot may never be modified, thus it cannot have any
3418eda14cbcSMatt Macy 	 *    in-flight itxs that would have modified the dataset.
3419eda14cbcSMatt Macy 	 *
3420eda14cbcSMatt Macy 	 * 2. By design, when zil_commit() is called, a commit itx will
3421eda14cbcSMatt Macy 	 *    be assigned to this zilog; as a result, the zilog will be
3422eda14cbcSMatt Macy 	 *    dirtied. We must not dirty the zilog of a snapshot; there's
3423eda14cbcSMatt Macy 	 *    checks in the code that enforce this invariant, and will
3424eda14cbcSMatt Macy 	 *    cause a panic if it's not upheld.
3425eda14cbcSMatt Macy 	 */
3426eda14cbcSMatt Macy 	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
3427eda14cbcSMatt Macy 
3428eda14cbcSMatt Macy 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3429eda14cbcSMatt Macy 		return;
3430eda14cbcSMatt Macy 
3431eda14cbcSMatt Macy 	if (!spa_writeable(zilog->zl_spa)) {
3432eda14cbcSMatt Macy 		/*
3433eda14cbcSMatt Macy 		 * If the SPA is not writable, there should never be any
3434eda14cbcSMatt Macy 		 * pending itxs waiting to be committed to disk. If that
3435eda14cbcSMatt Macy 		 * weren't true, we'd skip writing those itxs out, and
3436eda14cbcSMatt Macy 		 * would break the semantics of zil_commit(); thus, we're
3437eda14cbcSMatt Macy 		 * verifying that truth before we return to the caller.
3438eda14cbcSMatt Macy 		 */
3439eda14cbcSMatt Macy 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3440eda14cbcSMatt Macy 		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3441eda14cbcSMatt Macy 		for (int i = 0; i < TXG_SIZE; i++)
3442eda14cbcSMatt Macy 			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
3443eda14cbcSMatt Macy 		return;
3444eda14cbcSMatt Macy 	}
3445eda14cbcSMatt Macy 
3446eda14cbcSMatt Macy 	/*
3447eda14cbcSMatt Macy 	 * If the ZIL is suspended, we don't want to dirty it by calling
3448eda14cbcSMatt Macy 	 * zil_commit_itx_assign() below, nor can we write out
3449eda14cbcSMatt Macy 	 * lwbs like would be done in zil_commit_write(). Thus, we
3450eda14cbcSMatt Macy 	 * simply rely on txg_wait_synced() to maintain the necessary
3451eda14cbcSMatt Macy 	 * semantics, and avoid calling those functions altogether.
3452eda14cbcSMatt Macy 	 */
3453eda14cbcSMatt Macy 	if (zilog->zl_suspend > 0) {
3454eda14cbcSMatt Macy 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3455eda14cbcSMatt Macy 		return;
3456eda14cbcSMatt Macy 	}
3457eda14cbcSMatt Macy 
3458eda14cbcSMatt Macy 	zil_commit_impl(zilog, foid);
3459eda14cbcSMatt Macy }
3460eda14cbcSMatt Macy 
3461eda14cbcSMatt Macy void
3462eda14cbcSMatt Macy zil_commit_impl(zilog_t *zilog, uint64_t foid)
3463eda14cbcSMatt Macy {
3464271171e0SMartin Matuska 	ZIL_STAT_BUMP(zilog, zil_commit_count);
3465eda14cbcSMatt Macy 
3466eda14cbcSMatt Macy 	/*
3467eda14cbcSMatt Macy 	 * Move the "async" itxs for the specified foid to the "sync"
3468eda14cbcSMatt Macy 	 * queues, such that they will be later committed (or skipped)
3469eda14cbcSMatt Macy 	 * to an lwb when zil_process_commit_list() is called.
3470eda14cbcSMatt Macy 	 *
3471eda14cbcSMatt Macy 	 * Since these "async" itxs must be committed prior to this
3472eda14cbcSMatt Macy 	 * call to zil_commit returning, we must perform this operation
3473eda14cbcSMatt Macy 	 * before we call zil_commit_itx_assign().
3474eda14cbcSMatt Macy 	 */
3475eda14cbcSMatt Macy 	zil_async_to_sync(zilog, foid);
3476eda14cbcSMatt Macy 
3477eda14cbcSMatt Macy 	/*
3478eda14cbcSMatt Macy 	 * We allocate a new "waiter" structure which will initially be
3479eda14cbcSMatt Macy 	 * linked to the commit itx using the itx's "itx_private" field.
3480eda14cbcSMatt Macy 	 * Since the commit itx doesn't represent any on-disk state,
3481eda14cbcSMatt Macy 	 * when it's committed to an lwb, rather than copying the its
3482eda14cbcSMatt Macy 	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3483eda14cbcSMatt Macy 	 * added to the lwb's list of waiters. Then, when the lwb is
3484eda14cbcSMatt Macy 	 * committed to stable storage, each waiter in the lwb's list of
3485eda14cbcSMatt Macy 	 * waiters will be marked "done", and signalled.
3486eda14cbcSMatt Macy 	 *
3487eda14cbcSMatt Macy 	 * We must create the waiter and assign the commit itx prior to
3488eda14cbcSMatt Macy 	 * calling zil_commit_writer(), or else our specific commit itx
3489eda14cbcSMatt Macy 	 * is not guaranteed to be committed to an lwb prior to calling
3490eda14cbcSMatt Macy 	 * zil_commit_waiter().
3491eda14cbcSMatt Macy 	 */
3492eda14cbcSMatt Macy 	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
3493eda14cbcSMatt Macy 	zil_commit_itx_assign(zilog, zcw);
3494eda14cbcSMatt Macy 
34950a97523dSMartin Matuska 	uint64_t wtxg = zil_commit_writer(zilog, zcw);
3496eda14cbcSMatt Macy 	zil_commit_waiter(zilog, zcw);
3497eda14cbcSMatt Macy 
3498eda14cbcSMatt Macy 	if (zcw->zcw_zio_error != 0) {
3499eda14cbcSMatt Macy 		/*
3500eda14cbcSMatt Macy 		 * If there was an error writing out the ZIL blocks that
3501eda14cbcSMatt Macy 		 * this thread is waiting on, then we fallback to
3502eda14cbcSMatt Macy 		 * relying on spa_sync() to write out the data this
3503eda14cbcSMatt Macy 		 * thread is waiting on. Obviously this has performance
3504eda14cbcSMatt Macy 		 * implications, but the expectation is for this to be
3505eda14cbcSMatt Macy 		 * an exceptional case, and shouldn't occur often.
3506eda14cbcSMatt Macy 		 */
3507eda14cbcSMatt Macy 		DTRACE_PROBE2(zil__commit__io__error,
3508eda14cbcSMatt Macy 		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3509eda14cbcSMatt Macy 		txg_wait_synced(zilog->zl_dmu_pool, 0);
35100a97523dSMartin Matuska 	} else if (wtxg != 0) {
35110a97523dSMartin Matuska 		txg_wait_synced(zilog->zl_dmu_pool, wtxg);
3512eda14cbcSMatt Macy 	}
3513eda14cbcSMatt Macy 
3514eda14cbcSMatt Macy 	zil_free_commit_waiter(zcw);
3515eda14cbcSMatt Macy }
3516eda14cbcSMatt Macy 
3517eda14cbcSMatt Macy /*
3518eda14cbcSMatt Macy  * Called in syncing context to free committed log blocks and update log header.
3519eda14cbcSMatt Macy  */
3520eda14cbcSMatt Macy void
3521eda14cbcSMatt Macy zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3522eda14cbcSMatt Macy {
3523eda14cbcSMatt Macy 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
3524eda14cbcSMatt Macy 	uint64_t txg = dmu_tx_get_txg(tx);
3525eda14cbcSMatt Macy 	spa_t *spa = zilog->zl_spa;
3526eda14cbcSMatt Macy 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3527eda14cbcSMatt Macy 	lwb_t *lwb;
3528eda14cbcSMatt Macy 
3529eda14cbcSMatt Macy 	/*
3530eda14cbcSMatt Macy 	 * We don't zero out zl_destroy_txg, so make sure we don't try
3531eda14cbcSMatt Macy 	 * to destroy it twice.
3532eda14cbcSMatt Macy 	 */
3533eda14cbcSMatt Macy 	if (spa_sync_pass(spa) != 1)
3534eda14cbcSMatt Macy 		return;
3535eda14cbcSMatt Macy 
3536e3aa18adSMartin Matuska 	zil_lwb_flush_wait_all(zilog, txg);
3537e3aa18adSMartin Matuska 
3538eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
3539eda14cbcSMatt Macy 
3540eda14cbcSMatt Macy 	ASSERT(zilog->zl_stop_sync == 0);
3541eda14cbcSMatt Macy 
3542eda14cbcSMatt Macy 	if (*replayed_seq != 0) {
3543eda14cbcSMatt Macy 		ASSERT(zh->zh_replay_seq < *replayed_seq);
3544eda14cbcSMatt Macy 		zh->zh_replay_seq = *replayed_seq;
3545eda14cbcSMatt Macy 		*replayed_seq = 0;
3546eda14cbcSMatt Macy 	}
3547eda14cbcSMatt Macy 
3548eda14cbcSMatt Macy 	if (zilog->zl_destroy_txg == txg) {
3549eda14cbcSMatt Macy 		blkptr_t blk = zh->zh_log;
3550c03c5b1cSMartin Matuska 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
3551eda14cbcSMatt Macy 
3552c0a83fe0SMartin Matuska 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3553eda14cbcSMatt Macy 
3554da5137abSMartin Matuska 		memset(zh, 0, sizeof (zil_header_t));
3555da5137abSMartin Matuska 		memset(zilog->zl_replayed_seq, 0,
3556da5137abSMartin Matuska 		    sizeof (zilog->zl_replayed_seq));
3557eda14cbcSMatt Macy 
3558eda14cbcSMatt Macy 		if (zilog->zl_keep_first) {
3559eda14cbcSMatt Macy 			/*
3560eda14cbcSMatt Macy 			 * If this block was part of log chain that couldn't
3561eda14cbcSMatt Macy 			 * be claimed because a device was missing during
3562eda14cbcSMatt Macy 			 * zil_claim(), but that device later returns,
3563eda14cbcSMatt Macy 			 * then this block could erroneously appear valid.
3564eda14cbcSMatt Macy 			 * To guard against this, assign a new GUID to the new
3565eda14cbcSMatt Macy 			 * log chain so it doesn't matter what blk points to.
3566eda14cbcSMatt Macy 			 */
3567eda14cbcSMatt Macy 			zil_init_log_chain(zilog, &blk);
3568eda14cbcSMatt Macy 			zh->zh_log = blk;
3569c03c5b1cSMartin Matuska 		} else {
3570c03c5b1cSMartin Matuska 			/*
3571c03c5b1cSMartin Matuska 			 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3572c03c5b1cSMartin Matuska 			 * records. So, deactivate the feature for this dataset.
3573c03c5b1cSMartin Matuska 			 * We activate it again when we start a new ZIL chain.
3574c03c5b1cSMartin Matuska 			 */
3575c03c5b1cSMartin Matuska 			if (dsl_dataset_feature_is_active(ds,
3576c03c5b1cSMartin Matuska 			    SPA_FEATURE_ZILSAXATTR))
3577c03c5b1cSMartin Matuska 				dsl_dataset_deactivate_feature(ds,
3578c03c5b1cSMartin Matuska 				    SPA_FEATURE_ZILSAXATTR, tx);
3579eda14cbcSMatt Macy 		}
3580eda14cbcSMatt Macy 	}
3581eda14cbcSMatt Macy 
3582eda14cbcSMatt Macy 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3583eda14cbcSMatt Macy 		zh->zh_log = lwb->lwb_blk;
3584c0a83fe0SMartin Matuska 		if (lwb->lwb_state != LWB_STATE_FLUSH_DONE ||
3585315ee00fSMartin Matuska 		    lwb->lwb_alloc_txg > txg || lwb->lwb_max_txg > txg)
3586eda14cbcSMatt Macy 			break;
3587eda14cbcSMatt Macy 		list_remove(&zilog->zl_lwb_list, lwb);
3588315ee00fSMartin Matuska 		if (!BP_IS_HOLE(&lwb->lwb_blk))
3589eda14cbcSMatt Macy 			zio_free(spa, txg, &lwb->lwb_blk);
3590eda14cbcSMatt Macy 		zil_free_lwb(zilog, lwb);
3591eda14cbcSMatt Macy 
3592eda14cbcSMatt Macy 		/*
3593eda14cbcSMatt Macy 		 * If we don't have anything left in the lwb list then
3594eda14cbcSMatt Macy 		 * we've had an allocation failure and we need to zero
3595eda14cbcSMatt Macy 		 * out the zil_header blkptr so that we don't end
3596eda14cbcSMatt Macy 		 * up freeing the same block twice.
3597eda14cbcSMatt Macy 		 */
3598c0a83fe0SMartin Matuska 		if (list_is_empty(&zilog->zl_lwb_list))
3599eda14cbcSMatt Macy 			BP_ZERO(&zh->zh_log);
3600eda14cbcSMatt Macy 	}
3601eda14cbcSMatt Macy 
3602eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
3603eda14cbcSMatt Macy }
3604eda14cbcSMatt Macy 
3605eda14cbcSMatt Macy static int
3606eda14cbcSMatt Macy zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3607eda14cbcSMatt Macy {
3608e92ffd9bSMartin Matuska 	(void) unused, (void) kmflag;
3609eda14cbcSMatt Macy 	lwb_t *lwb = vbuf;
3610eda14cbcSMatt Macy 	list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3611eda14cbcSMatt Macy 	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3612eda14cbcSMatt Macy 	    offsetof(zil_commit_waiter_t, zcw_node));
3613eda14cbcSMatt Macy 	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3614eda14cbcSMatt Macy 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3615eda14cbcSMatt Macy 	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3616eda14cbcSMatt Macy 	return (0);
3617eda14cbcSMatt Macy }
3618eda14cbcSMatt Macy 
3619eda14cbcSMatt Macy static void
3620eda14cbcSMatt Macy zil_lwb_dest(void *vbuf, void *unused)
3621eda14cbcSMatt Macy {
3622e92ffd9bSMartin Matuska 	(void) unused;
3623eda14cbcSMatt Macy 	lwb_t *lwb = vbuf;
3624eda14cbcSMatt Macy 	mutex_destroy(&lwb->lwb_vdev_lock);
3625eda14cbcSMatt Macy 	avl_destroy(&lwb->lwb_vdev_tree);
3626eda14cbcSMatt Macy 	list_destroy(&lwb->lwb_waiters);
3627eda14cbcSMatt Macy 	list_destroy(&lwb->lwb_itxs);
3628eda14cbcSMatt Macy }
3629eda14cbcSMatt Macy 
3630eda14cbcSMatt Macy void
3631eda14cbcSMatt Macy zil_init(void)
3632eda14cbcSMatt Macy {
3633eda14cbcSMatt Macy 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3634eda14cbcSMatt Macy 	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3635eda14cbcSMatt Macy 
3636eda14cbcSMatt Macy 	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3637eda14cbcSMatt Macy 	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3638eda14cbcSMatt Macy 
3639271171e0SMartin Matuska 	zil_sums_init(&zil_sums_global);
3640271171e0SMartin Matuska 	zil_kstats_global = kstat_create("zfs", 0, "zil", "misc",
3641eda14cbcSMatt Macy 	    KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3642eda14cbcSMatt Macy 	    KSTAT_FLAG_VIRTUAL);
3643eda14cbcSMatt Macy 
3644271171e0SMartin Matuska 	if (zil_kstats_global != NULL) {
3645271171e0SMartin Matuska 		zil_kstats_global->ks_data = &zil_stats;
3646271171e0SMartin Matuska 		zil_kstats_global->ks_update = zil_kstats_global_update;
3647271171e0SMartin Matuska 		zil_kstats_global->ks_private = NULL;
3648271171e0SMartin Matuska 		kstat_install(zil_kstats_global);
3649eda14cbcSMatt Macy 	}
3650eda14cbcSMatt Macy }
3651eda14cbcSMatt Macy 
3652eda14cbcSMatt Macy void
3653eda14cbcSMatt Macy zil_fini(void)
3654eda14cbcSMatt Macy {
3655eda14cbcSMatt Macy 	kmem_cache_destroy(zil_zcw_cache);
3656eda14cbcSMatt Macy 	kmem_cache_destroy(zil_lwb_cache);
3657eda14cbcSMatt Macy 
3658271171e0SMartin Matuska 	if (zil_kstats_global != NULL) {
3659271171e0SMartin Matuska 		kstat_delete(zil_kstats_global);
3660271171e0SMartin Matuska 		zil_kstats_global = NULL;
3661eda14cbcSMatt Macy 	}
3662271171e0SMartin Matuska 
3663271171e0SMartin Matuska 	zil_sums_fini(&zil_sums_global);
3664eda14cbcSMatt Macy }
3665eda14cbcSMatt Macy 
3666eda14cbcSMatt Macy void
3667eda14cbcSMatt Macy zil_set_sync(zilog_t *zilog, uint64_t sync)
3668eda14cbcSMatt Macy {
3669eda14cbcSMatt Macy 	zilog->zl_sync = sync;
3670eda14cbcSMatt Macy }
3671eda14cbcSMatt Macy 
3672eda14cbcSMatt Macy void
3673eda14cbcSMatt Macy zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3674eda14cbcSMatt Macy {
3675eda14cbcSMatt Macy 	zilog->zl_logbias = logbias;
3676eda14cbcSMatt Macy }
3677eda14cbcSMatt Macy 
3678eda14cbcSMatt Macy zilog_t *
3679eda14cbcSMatt Macy zil_alloc(objset_t *os, zil_header_t *zh_phys)
3680eda14cbcSMatt Macy {
3681eda14cbcSMatt Macy 	zilog_t *zilog;
3682eda14cbcSMatt Macy 
3683eda14cbcSMatt Macy 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3684eda14cbcSMatt Macy 
3685eda14cbcSMatt Macy 	zilog->zl_header = zh_phys;
3686eda14cbcSMatt Macy 	zilog->zl_os = os;
3687eda14cbcSMatt Macy 	zilog->zl_spa = dmu_objset_spa(os);
3688eda14cbcSMatt Macy 	zilog->zl_dmu_pool = dmu_objset_pool(os);
3689eda14cbcSMatt Macy 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
3690eda14cbcSMatt Macy 	zilog->zl_logbias = dmu_objset_logbias(os);
3691eda14cbcSMatt Macy 	zilog->zl_sync = dmu_objset_syncprop(os);
3692eda14cbcSMatt Macy 	zilog->zl_dirty_max_txg = 0;
3693eda14cbcSMatt Macy 	zilog->zl_last_lwb_opened = NULL;
3694eda14cbcSMatt Macy 	zilog->zl_last_lwb_latency = 0;
3695eda14cbcSMatt Macy 	zilog->zl_max_block_size = zil_maxblocksize;
3696eda14cbcSMatt Macy 
3697eda14cbcSMatt Macy 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3698eda14cbcSMatt Macy 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3699e3aa18adSMartin Matuska 	mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL);
3700eda14cbcSMatt Macy 
3701eda14cbcSMatt Macy 	for (int i = 0; i < TXG_SIZE; i++) {
3702eda14cbcSMatt Macy 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3703eda14cbcSMatt Macy 		    MUTEX_DEFAULT, NULL);
3704eda14cbcSMatt Macy 	}
3705eda14cbcSMatt Macy 
3706eda14cbcSMatt Macy 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3707eda14cbcSMatt Macy 	    offsetof(lwb_t, lwb_node));
3708eda14cbcSMatt Macy 
3709eda14cbcSMatt Macy 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3710eda14cbcSMatt Macy 	    offsetof(itx_t, itx_node));
3711eda14cbcSMatt Macy 
3712eda14cbcSMatt Macy 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3713e3aa18adSMartin Matuska 	cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL);
3714eda14cbcSMatt Macy 
3715eda14cbcSMatt Macy 	return (zilog);
3716eda14cbcSMatt Macy }
3717eda14cbcSMatt Macy 
3718eda14cbcSMatt Macy void
3719eda14cbcSMatt Macy zil_free(zilog_t *zilog)
3720eda14cbcSMatt Macy {
3721eda14cbcSMatt Macy 	int i;
3722eda14cbcSMatt Macy 
3723eda14cbcSMatt Macy 	zilog->zl_stop_sync = 1;
3724eda14cbcSMatt Macy 
3725eda14cbcSMatt Macy 	ASSERT0(zilog->zl_suspend);
3726eda14cbcSMatt Macy 	ASSERT0(zilog->zl_suspending);
3727eda14cbcSMatt Macy 
3728eda14cbcSMatt Macy 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3729eda14cbcSMatt Macy 	list_destroy(&zilog->zl_lwb_list);
3730eda14cbcSMatt Macy 
3731eda14cbcSMatt Macy 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3732eda14cbcSMatt Macy 	list_destroy(&zilog->zl_itx_commit_list);
3733eda14cbcSMatt Macy 
3734eda14cbcSMatt Macy 	for (i = 0; i < TXG_SIZE; i++) {
3735eda14cbcSMatt Macy 		/*
3736eda14cbcSMatt Macy 		 * It's possible for an itx to be generated that doesn't dirty
3737eda14cbcSMatt Macy 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3738eda14cbcSMatt Macy 		 * callback to remove the entry. We remove those here.
3739eda14cbcSMatt Macy 		 *
3740eda14cbcSMatt Macy 		 * Also free up the ziltest itxs.
3741eda14cbcSMatt Macy 		 */
3742eda14cbcSMatt Macy 		if (zilog->zl_itxg[i].itxg_itxs)
3743eda14cbcSMatt Macy 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3744eda14cbcSMatt Macy 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3745eda14cbcSMatt Macy 	}
3746eda14cbcSMatt Macy 
3747eda14cbcSMatt Macy 	mutex_destroy(&zilog->zl_issuer_lock);
3748eda14cbcSMatt Macy 	mutex_destroy(&zilog->zl_lock);
3749e3aa18adSMartin Matuska 	mutex_destroy(&zilog->zl_lwb_io_lock);
3750eda14cbcSMatt Macy 
3751eda14cbcSMatt Macy 	cv_destroy(&zilog->zl_cv_suspend);
3752e3aa18adSMartin Matuska 	cv_destroy(&zilog->zl_lwb_io_cv);
3753eda14cbcSMatt Macy 
3754eda14cbcSMatt Macy 	kmem_free(zilog, sizeof (zilog_t));
3755eda14cbcSMatt Macy }
3756eda14cbcSMatt Macy 
3757eda14cbcSMatt Macy /*
3758eda14cbcSMatt Macy  * Open an intent log.
3759eda14cbcSMatt Macy  */
3760eda14cbcSMatt Macy zilog_t *
3761271171e0SMartin Matuska zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums)
3762eda14cbcSMatt Macy {
3763eda14cbcSMatt Macy 	zilog_t *zilog = dmu_objset_zil(os);
3764eda14cbcSMatt Macy 
3765eda14cbcSMatt Macy 	ASSERT3P(zilog->zl_get_data, ==, NULL);
3766eda14cbcSMatt Macy 	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3767eda14cbcSMatt Macy 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3768eda14cbcSMatt Macy 
3769eda14cbcSMatt Macy 	zilog->zl_get_data = get_data;
3770271171e0SMartin Matuska 	zilog->zl_sums = zil_sums;
3771eda14cbcSMatt Macy 
3772eda14cbcSMatt Macy 	return (zilog);
3773eda14cbcSMatt Macy }
3774eda14cbcSMatt Macy 
3775eda14cbcSMatt Macy /*
3776eda14cbcSMatt Macy  * Close an intent log.
3777eda14cbcSMatt Macy  */
3778eda14cbcSMatt Macy void
3779eda14cbcSMatt Macy zil_close(zilog_t *zilog)
3780eda14cbcSMatt Macy {
3781eda14cbcSMatt Macy 	lwb_t *lwb;
3782eda14cbcSMatt Macy 	uint64_t txg;
3783eda14cbcSMatt Macy 
3784eda14cbcSMatt Macy 	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3785eda14cbcSMatt Macy 		zil_commit(zilog, 0);
3786eda14cbcSMatt Macy 	} else {
3787c0a83fe0SMartin Matuska 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3788eda14cbcSMatt Macy 		ASSERT0(zilog->zl_dirty_max_txg);
3789eda14cbcSMatt Macy 		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3790eda14cbcSMatt Macy 	}
3791eda14cbcSMatt Macy 
3792eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
3793eda14cbcSMatt Macy 	txg = zilog->zl_dirty_max_txg;
3794315ee00fSMartin Matuska 	lwb = list_tail(&zilog->zl_lwb_list);
3795315ee00fSMartin Matuska 	if (lwb != NULL) {
3796315ee00fSMartin Matuska 		txg = MAX(txg, lwb->lwb_alloc_txg);
3797315ee00fSMartin Matuska 		txg = MAX(txg, lwb->lwb_max_txg);
3798315ee00fSMartin Matuska 	}
3799eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
3800eda14cbcSMatt Macy 
3801eda14cbcSMatt Macy 	/*
3802e3aa18adSMartin Matuska 	 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3803e3aa18adSMartin Matuska 	 * on the time when the dmu_tx transaction is assigned in
3804315ee00fSMartin Matuska 	 * zil_lwb_write_issue().
3805e3aa18adSMartin Matuska 	 */
3806e3aa18adSMartin Matuska 	mutex_enter(&zilog->zl_lwb_io_lock);
3807e3aa18adSMartin Matuska 	txg = MAX(zilog->zl_lwb_max_issued_txg, txg);
3808e3aa18adSMartin Matuska 	mutex_exit(&zilog->zl_lwb_io_lock);
3809e3aa18adSMartin Matuska 
3810e3aa18adSMartin Matuska 	/*
3811e3aa18adSMartin Matuska 	 * We need to use txg_wait_synced() to wait until that txg is synced.
3812e3aa18adSMartin Matuska 	 * zil_sync() will guarantee all lwbs up to that txg have been
3813e3aa18adSMartin Matuska 	 * written out, flushed, and cleaned.
3814eda14cbcSMatt Macy 	 */
3815eda14cbcSMatt Macy 	if (txg != 0)
3816eda14cbcSMatt Macy 		txg_wait_synced(zilog->zl_dmu_pool, txg);
3817eda14cbcSMatt Macy 
3818eda14cbcSMatt Macy 	if (zilog_is_dirty(zilog))
381933b8c039SMartin Matuska 		zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
382033b8c039SMartin Matuska 		    (u_longlong_t)txg);
3821eda14cbcSMatt Macy 	if (txg < spa_freeze_txg(zilog->zl_spa))
3822eda14cbcSMatt Macy 		VERIFY(!zilog_is_dirty(zilog));
3823eda14cbcSMatt Macy 
3824eda14cbcSMatt Macy 	zilog->zl_get_data = NULL;
3825eda14cbcSMatt Macy 
3826eda14cbcSMatt Macy 	/*
3827eda14cbcSMatt Macy 	 * We should have only one lwb left on the list; remove it now.
3828eda14cbcSMatt Macy 	 */
3829eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
3830c0a83fe0SMartin Matuska 	lwb = list_remove_head(&zilog->zl_lwb_list);
3831eda14cbcSMatt Macy 	if (lwb != NULL) {
3832c0a83fe0SMartin Matuska 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3833315ee00fSMartin Matuska 		ASSERT3S(lwb->lwb_state, ==, LWB_STATE_NEW);
3834eda14cbcSMatt Macy 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3835eda14cbcSMatt Macy 		zil_free_lwb(zilog, lwb);
3836eda14cbcSMatt Macy 	}
3837eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
3838eda14cbcSMatt Macy }
3839eda14cbcSMatt Macy 
3840a0b956f5SMartin Matuska static const char *suspend_tag = "zil suspending";
3841eda14cbcSMatt Macy 
3842eda14cbcSMatt Macy /*
3843eda14cbcSMatt Macy  * Suspend an intent log.  While in suspended mode, we still honor
3844eda14cbcSMatt Macy  * synchronous semantics, but we rely on txg_wait_synced() to do it.
3845eda14cbcSMatt Macy  * On old version pools, we suspend the log briefly when taking a
3846eda14cbcSMatt Macy  * snapshot so that it will have an empty intent log.
3847eda14cbcSMatt Macy  *
3848eda14cbcSMatt Macy  * Long holds are not really intended to be used the way we do here --
3849eda14cbcSMatt Macy  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
3850eda14cbcSMatt Macy  * could fail.  Therefore we take pains to only put a long hold if it is
3851eda14cbcSMatt Macy  * actually necessary.  Fortunately, it will only be necessary if the
3852eda14cbcSMatt Macy  * objset is currently mounted (or the ZVOL equivalent).  In that case it
3853eda14cbcSMatt Macy  * will already have a long hold, so we are not really making things any worse.
3854eda14cbcSMatt Macy  *
3855eda14cbcSMatt Macy  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3856eda14cbcSMatt Macy  * zvol_state_t), and use their mechanism to prevent their hold from being
3857eda14cbcSMatt Macy  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
3858eda14cbcSMatt Macy  * very little gain.
3859eda14cbcSMatt Macy  *
3860eda14cbcSMatt Macy  * if cookiep == NULL, this does both the suspend & resume.
3861eda14cbcSMatt Macy  * Otherwise, it returns with the dataset "long held", and the cookie
3862eda14cbcSMatt Macy  * should be passed into zil_resume().
3863eda14cbcSMatt Macy  */
3864eda14cbcSMatt Macy int
3865eda14cbcSMatt Macy zil_suspend(const char *osname, void **cookiep)
3866eda14cbcSMatt Macy {
3867eda14cbcSMatt Macy 	objset_t *os;
3868eda14cbcSMatt Macy 	zilog_t *zilog;
3869eda14cbcSMatt Macy 	const zil_header_t *zh;
3870eda14cbcSMatt Macy 	int error;
3871eda14cbcSMatt Macy 
3872eda14cbcSMatt Macy 	error = dmu_objset_hold(osname, suspend_tag, &os);
3873eda14cbcSMatt Macy 	if (error != 0)
3874eda14cbcSMatt Macy 		return (error);
3875eda14cbcSMatt Macy 	zilog = dmu_objset_zil(os);
3876eda14cbcSMatt Macy 
3877eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
3878eda14cbcSMatt Macy 	zh = zilog->zl_header;
3879eda14cbcSMatt Macy 
3880eda14cbcSMatt Macy 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
3881eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_lock);
3882eda14cbcSMatt Macy 		dmu_objset_rele(os, suspend_tag);
3883eda14cbcSMatt Macy 		return (SET_ERROR(EBUSY));
3884eda14cbcSMatt Macy 	}
3885eda14cbcSMatt Macy 
3886eda14cbcSMatt Macy 	/*
3887eda14cbcSMatt Macy 	 * Don't put a long hold in the cases where we can avoid it.  This
3888eda14cbcSMatt Macy 	 * is when there is no cookie so we are doing a suspend & resume
3889eda14cbcSMatt Macy 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3890eda14cbcSMatt Macy 	 * for the suspend because it's already suspended, or there's no ZIL.
3891eda14cbcSMatt Macy 	 */
3892eda14cbcSMatt Macy 	if (cookiep == NULL && !zilog->zl_suspending &&
3893eda14cbcSMatt Macy 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3894eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_lock);
3895eda14cbcSMatt Macy 		dmu_objset_rele(os, suspend_tag);
3896eda14cbcSMatt Macy 		return (0);
3897eda14cbcSMatt Macy 	}
3898eda14cbcSMatt Macy 
3899eda14cbcSMatt Macy 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3900eda14cbcSMatt Macy 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3901eda14cbcSMatt Macy 
3902eda14cbcSMatt Macy 	zilog->zl_suspend++;
3903eda14cbcSMatt Macy 
3904eda14cbcSMatt Macy 	if (zilog->zl_suspend > 1) {
3905eda14cbcSMatt Macy 		/*
3906eda14cbcSMatt Macy 		 * Someone else is already suspending it.
3907eda14cbcSMatt Macy 		 * Just wait for them to finish.
3908eda14cbcSMatt Macy 		 */
3909eda14cbcSMatt Macy 
3910eda14cbcSMatt Macy 		while (zilog->zl_suspending)
3911eda14cbcSMatt Macy 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3912eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_lock);
3913eda14cbcSMatt Macy 
3914eda14cbcSMatt Macy 		if (cookiep == NULL)
3915eda14cbcSMatt Macy 			zil_resume(os);
3916eda14cbcSMatt Macy 		else
3917eda14cbcSMatt Macy 			*cookiep = os;
3918eda14cbcSMatt Macy 		return (0);
3919eda14cbcSMatt Macy 	}
3920eda14cbcSMatt Macy 
3921eda14cbcSMatt Macy 	/*
3922eda14cbcSMatt Macy 	 * If there is no pointer to an on-disk block, this ZIL must not
3923eda14cbcSMatt Macy 	 * be active (e.g. filesystem not mounted), so there's nothing
3924eda14cbcSMatt Macy 	 * to clean up.
3925eda14cbcSMatt Macy 	 */
3926eda14cbcSMatt Macy 	if (BP_IS_HOLE(&zh->zh_log)) {
3927eda14cbcSMatt Macy 		ASSERT(cookiep != NULL); /* fast path already handled */
3928eda14cbcSMatt Macy 
3929eda14cbcSMatt Macy 		*cookiep = os;
3930eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_lock);
3931eda14cbcSMatt Macy 		return (0);
3932eda14cbcSMatt Macy 	}
3933eda14cbcSMatt Macy 
3934eda14cbcSMatt Macy 	/*
3935eda14cbcSMatt Macy 	 * The ZIL has work to do. Ensure that the associated encryption
3936eda14cbcSMatt Macy 	 * key will remain mapped while we are committing the log by
3937eda14cbcSMatt Macy 	 * grabbing a reference to it. If the key isn't loaded we have no
3938eda14cbcSMatt Macy 	 * choice but to return an error until the wrapping key is loaded.
3939eda14cbcSMatt Macy 	 */
3940eda14cbcSMatt Macy 	if (os->os_encrypted &&
3941eda14cbcSMatt Macy 	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
3942eda14cbcSMatt Macy 		zilog->zl_suspend--;
3943eda14cbcSMatt Macy 		mutex_exit(&zilog->zl_lock);
3944eda14cbcSMatt Macy 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3945eda14cbcSMatt Macy 		dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3946eda14cbcSMatt Macy 		return (SET_ERROR(EACCES));
3947eda14cbcSMatt Macy 	}
3948eda14cbcSMatt Macy 
3949eda14cbcSMatt Macy 	zilog->zl_suspending = B_TRUE;
3950eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
3951eda14cbcSMatt Macy 
3952eda14cbcSMatt Macy 	/*
3953eda14cbcSMatt Macy 	 * We need to use zil_commit_impl to ensure we wait for all
3954315ee00fSMartin Matuska 	 * LWB_STATE_OPENED, _CLOSED and _READY lwbs to be committed
3955eda14cbcSMatt Macy 	 * to disk before proceeding. If we used zil_commit instead, it
3956eda14cbcSMatt Macy 	 * would just call txg_wait_synced(), because zl_suspend is set.
3957eda14cbcSMatt Macy 	 * txg_wait_synced() doesn't wait for these lwb's to be
3958eda14cbcSMatt Macy 	 * LWB_STATE_FLUSH_DONE before returning.
3959eda14cbcSMatt Macy 	 */
3960eda14cbcSMatt Macy 	zil_commit_impl(zilog, 0);
3961eda14cbcSMatt Macy 
3962eda14cbcSMatt Macy 	/*
3963eda14cbcSMatt Macy 	 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
3964eda14cbcSMatt Macy 	 * use txg_wait_synced() to ensure the data from the zilog has
3965eda14cbcSMatt Macy 	 * migrated to the main pool before calling zil_destroy().
3966eda14cbcSMatt Macy 	 */
3967eda14cbcSMatt Macy 	txg_wait_synced(zilog->zl_dmu_pool, 0);
3968eda14cbcSMatt Macy 
3969eda14cbcSMatt Macy 	zil_destroy(zilog, B_FALSE);
3970eda14cbcSMatt Macy 
3971eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
3972eda14cbcSMatt Macy 	zilog->zl_suspending = B_FALSE;
3973eda14cbcSMatt Macy 	cv_broadcast(&zilog->zl_cv_suspend);
3974eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
3975eda14cbcSMatt Macy 
3976eda14cbcSMatt Macy 	if (os->os_encrypted)
3977eda14cbcSMatt Macy 		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
3978eda14cbcSMatt Macy 
3979eda14cbcSMatt Macy 	if (cookiep == NULL)
3980eda14cbcSMatt Macy 		zil_resume(os);
3981eda14cbcSMatt Macy 	else
3982eda14cbcSMatt Macy 		*cookiep = os;
3983eda14cbcSMatt Macy 	return (0);
3984eda14cbcSMatt Macy }
3985eda14cbcSMatt Macy 
3986eda14cbcSMatt Macy void
3987eda14cbcSMatt Macy zil_resume(void *cookie)
3988eda14cbcSMatt Macy {
3989eda14cbcSMatt Macy 	objset_t *os = cookie;
3990eda14cbcSMatt Macy 	zilog_t *zilog = dmu_objset_zil(os);
3991eda14cbcSMatt Macy 
3992eda14cbcSMatt Macy 	mutex_enter(&zilog->zl_lock);
3993eda14cbcSMatt Macy 	ASSERT(zilog->zl_suspend != 0);
3994eda14cbcSMatt Macy 	zilog->zl_suspend--;
3995eda14cbcSMatt Macy 	mutex_exit(&zilog->zl_lock);
3996eda14cbcSMatt Macy 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3997eda14cbcSMatt Macy 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3998eda14cbcSMatt Macy }
3999eda14cbcSMatt Macy 
4000eda14cbcSMatt Macy typedef struct zil_replay_arg {
4001e92ffd9bSMartin Matuska 	zil_replay_func_t *const *zr_replay;
4002eda14cbcSMatt Macy 	void		*zr_arg;
4003eda14cbcSMatt Macy 	boolean_t	zr_byteswap;
4004eda14cbcSMatt Macy 	char		*zr_lr;
4005eda14cbcSMatt Macy } zil_replay_arg_t;
4006eda14cbcSMatt Macy 
4007eda14cbcSMatt Macy static int
4008180f8225SMatt Macy zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
4009eda14cbcSMatt Macy {
4010eda14cbcSMatt Macy 	char name[ZFS_MAX_DATASET_NAME_LEN];
4011eda14cbcSMatt Macy 
4012eda14cbcSMatt Macy 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
4013eda14cbcSMatt Macy 
4014eda14cbcSMatt Macy 	dmu_objset_name(zilog->zl_os, name);
4015eda14cbcSMatt Macy 
4016eda14cbcSMatt Macy 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
4017eda14cbcSMatt Macy 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
4018eda14cbcSMatt Macy 	    (u_longlong_t)lr->lrc_seq,
4019eda14cbcSMatt Macy 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
4020eda14cbcSMatt Macy 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
4021eda14cbcSMatt Macy 
4022eda14cbcSMatt Macy 	return (error);
4023eda14cbcSMatt Macy }
4024eda14cbcSMatt Macy 
4025eda14cbcSMatt Macy static int
4026180f8225SMatt Macy zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
4027180f8225SMatt Macy     uint64_t claim_txg)
4028eda14cbcSMatt Macy {
4029eda14cbcSMatt Macy 	zil_replay_arg_t *zr = zra;
4030eda14cbcSMatt Macy 	const zil_header_t *zh = zilog->zl_header;
4031eda14cbcSMatt Macy 	uint64_t reclen = lr->lrc_reclen;
4032eda14cbcSMatt Macy 	uint64_t txtype = lr->lrc_txtype;
4033eda14cbcSMatt Macy 	int error = 0;
4034eda14cbcSMatt Macy 
4035eda14cbcSMatt Macy 	zilog->zl_replaying_seq = lr->lrc_seq;
4036eda14cbcSMatt Macy 
4037eda14cbcSMatt Macy 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
4038eda14cbcSMatt Macy 		return (0);
4039eda14cbcSMatt Macy 
4040eda14cbcSMatt Macy 	if (lr->lrc_txg < claim_txg)		/* already committed */
4041eda14cbcSMatt Macy 		return (0);
4042eda14cbcSMatt Macy 
4043eda14cbcSMatt Macy 	/* Strip case-insensitive bit, still present in log record */
4044eda14cbcSMatt Macy 	txtype &= ~TX_CI;
4045eda14cbcSMatt Macy 
4046eda14cbcSMatt Macy 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
4047eda14cbcSMatt Macy 		return (zil_replay_error(zilog, lr, EINVAL));
4048eda14cbcSMatt Macy 
4049eda14cbcSMatt Macy 	/*
4050eda14cbcSMatt Macy 	 * If this record type can be logged out of order, the object
4051eda14cbcSMatt Macy 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
4052eda14cbcSMatt Macy 	 */
4053eda14cbcSMatt Macy 	if (TX_OOO(txtype)) {
4054eda14cbcSMatt Macy 		error = dmu_object_info(zilog->zl_os,
4055eda14cbcSMatt Macy 		    LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
4056eda14cbcSMatt Macy 		if (error == ENOENT || error == EEXIST)
4057eda14cbcSMatt Macy 			return (0);
4058eda14cbcSMatt Macy 	}
4059eda14cbcSMatt Macy 
4060eda14cbcSMatt Macy 	/*
4061eda14cbcSMatt Macy 	 * Make a copy of the data so we can revise and extend it.
4062eda14cbcSMatt Macy 	 */
4063da5137abSMartin Matuska 	memcpy(zr->zr_lr, lr, reclen);
4064eda14cbcSMatt Macy 
4065eda14cbcSMatt Macy 	/*
4066eda14cbcSMatt Macy 	 * If this is a TX_WRITE with a blkptr, suck in the data.
4067eda14cbcSMatt Macy 	 */
4068eda14cbcSMatt Macy 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
4069eda14cbcSMatt Macy 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
4070eda14cbcSMatt Macy 		    zr->zr_lr + reclen);
4071eda14cbcSMatt Macy 		if (error != 0)
4072eda14cbcSMatt Macy 			return (zil_replay_error(zilog, lr, error));
4073eda14cbcSMatt Macy 	}
4074eda14cbcSMatt Macy 
4075eda14cbcSMatt Macy 	/*
4076eda14cbcSMatt Macy 	 * The log block containing this lr may have been byteswapped
4077eda14cbcSMatt Macy 	 * so that we can easily examine common fields like lrc_txtype.
4078eda14cbcSMatt Macy 	 * However, the log is a mix of different record types, and only the
4079eda14cbcSMatt Macy 	 * replay vectors know how to byteswap their records.  Therefore, if
4080eda14cbcSMatt Macy 	 * the lr was byteswapped, undo it before invoking the replay vector.
4081eda14cbcSMatt Macy 	 */
4082eda14cbcSMatt Macy 	if (zr->zr_byteswap)
4083eda14cbcSMatt Macy 		byteswap_uint64_array(zr->zr_lr, reclen);
4084eda14cbcSMatt Macy 
4085eda14cbcSMatt Macy 	/*
4086eda14cbcSMatt Macy 	 * We must now do two things atomically: replay this log record,
4087eda14cbcSMatt Macy 	 * and update the log header sequence number to reflect the fact that
4088eda14cbcSMatt Macy 	 * we did so. At the end of each replay function the sequence number
4089eda14cbcSMatt Macy 	 * is updated if we are in replay mode.
4090eda14cbcSMatt Macy 	 */
4091eda14cbcSMatt Macy 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
4092eda14cbcSMatt Macy 	if (error != 0) {
4093eda14cbcSMatt Macy 		/*
4094eda14cbcSMatt Macy 		 * The DMU's dnode layer doesn't see removes until the txg
4095eda14cbcSMatt Macy 		 * commits, so a subsequent claim can spuriously fail with
4096eda14cbcSMatt Macy 		 * EEXIST. So if we receive any error we try syncing out
4097eda14cbcSMatt Macy 		 * any removes then retry the transaction.  Note that we
4098eda14cbcSMatt Macy 		 * specify B_FALSE for byteswap now, so we don't do it twice.
4099eda14cbcSMatt Macy 		 */
4100eda14cbcSMatt Macy 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
4101eda14cbcSMatt Macy 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
4102eda14cbcSMatt Macy 		if (error != 0)
4103eda14cbcSMatt Macy 			return (zil_replay_error(zilog, lr, error));
4104eda14cbcSMatt Macy 	}
4105eda14cbcSMatt Macy 	return (0);
4106eda14cbcSMatt Macy }
4107eda14cbcSMatt Macy 
4108eda14cbcSMatt Macy static int
4109180f8225SMatt Macy zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
4110eda14cbcSMatt Macy {
4111e92ffd9bSMartin Matuska 	(void) bp, (void) arg, (void) claim_txg;
4112e92ffd9bSMartin Matuska 
4113eda14cbcSMatt Macy 	zilog->zl_replay_blks++;
4114eda14cbcSMatt Macy 
4115eda14cbcSMatt Macy 	return (0);
4116eda14cbcSMatt Macy }
4117eda14cbcSMatt Macy 
4118eda14cbcSMatt Macy /*
4119eda14cbcSMatt Macy  * If this dataset has a non-empty intent log, replay it and destroy it.
4120dbd5678dSMartin Matuska  * Return B_TRUE if there were any entries to replay.
4121eda14cbcSMatt Macy  */
4122dbd5678dSMartin Matuska boolean_t
4123e92ffd9bSMartin Matuska zil_replay(objset_t *os, void *arg,
4124e92ffd9bSMartin Matuska     zil_replay_func_t *const replay_func[TX_MAX_TYPE])
4125eda14cbcSMatt Macy {
4126eda14cbcSMatt Macy 	zilog_t *zilog = dmu_objset_zil(os);
4127eda14cbcSMatt Macy 	const zil_header_t *zh = zilog->zl_header;
4128eda14cbcSMatt Macy 	zil_replay_arg_t zr;
4129eda14cbcSMatt Macy 
4130eda14cbcSMatt Macy 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
4131dbd5678dSMartin Matuska 		return (zil_destroy(zilog, B_TRUE));
4132eda14cbcSMatt Macy 	}
4133eda14cbcSMatt Macy 
4134eda14cbcSMatt Macy 	zr.zr_replay = replay_func;
4135eda14cbcSMatt Macy 	zr.zr_arg = arg;
4136eda14cbcSMatt Macy 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
4137eda14cbcSMatt Macy 	zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
4138eda14cbcSMatt Macy 
4139eda14cbcSMatt Macy 	/*
4140eda14cbcSMatt Macy 	 * Wait for in-progress removes to sync before starting replay.
4141eda14cbcSMatt Macy 	 */
4142eda14cbcSMatt Macy 	txg_wait_synced(zilog->zl_dmu_pool, 0);
4143eda14cbcSMatt Macy 
4144eda14cbcSMatt Macy 	zilog->zl_replay = B_TRUE;
4145eda14cbcSMatt Macy 	zilog->zl_replay_time = ddi_get_lbolt();
4146eda14cbcSMatt Macy 	ASSERT(zilog->zl_replay_blks == 0);
4147eda14cbcSMatt Macy 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
4148eda14cbcSMatt Macy 	    zh->zh_claim_txg, B_TRUE);
4149eda14cbcSMatt Macy 	vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
4150eda14cbcSMatt Macy 
4151eda14cbcSMatt Macy 	zil_destroy(zilog, B_FALSE);
4152eda14cbcSMatt Macy 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
4153eda14cbcSMatt Macy 	zilog->zl_replay = B_FALSE;
4154dbd5678dSMartin Matuska 
4155dbd5678dSMartin Matuska 	return (B_TRUE);
4156eda14cbcSMatt Macy }
4157eda14cbcSMatt Macy 
4158eda14cbcSMatt Macy boolean_t
4159eda14cbcSMatt Macy zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
4160eda14cbcSMatt Macy {
4161eda14cbcSMatt Macy 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
4162eda14cbcSMatt Macy 		return (B_TRUE);
4163eda14cbcSMatt Macy 
4164eda14cbcSMatt Macy 	if (zilog->zl_replay) {
4165eda14cbcSMatt Macy 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
4166eda14cbcSMatt Macy 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
4167eda14cbcSMatt Macy 		    zilog->zl_replaying_seq;
4168eda14cbcSMatt Macy 		return (B_TRUE);
4169eda14cbcSMatt Macy 	}
4170eda14cbcSMatt Macy 
4171eda14cbcSMatt Macy 	return (B_FALSE);
4172eda14cbcSMatt Macy }
4173eda14cbcSMatt Macy 
4174eda14cbcSMatt Macy int
4175eda14cbcSMatt Macy zil_reset(const char *osname, void *arg)
4176eda14cbcSMatt Macy {
4177e92ffd9bSMartin Matuska 	(void) arg;
4178eda14cbcSMatt Macy 
4179e92ffd9bSMartin Matuska 	int error = zil_suspend(osname, NULL);
4180eda14cbcSMatt Macy 	/* EACCES means crypto key not loaded */
4181eda14cbcSMatt Macy 	if ((error == EACCES) || (error == EBUSY))
4182eda14cbcSMatt Macy 		return (SET_ERROR(error));
4183eda14cbcSMatt Macy 	if (error != 0)
4184eda14cbcSMatt Macy 		return (SET_ERROR(EEXIST));
4185eda14cbcSMatt Macy 	return (0);
4186eda14cbcSMatt Macy }
4187eda14cbcSMatt Macy 
4188eda14cbcSMatt Macy EXPORT_SYMBOL(zil_alloc);
4189eda14cbcSMatt Macy EXPORT_SYMBOL(zil_free);
4190eda14cbcSMatt Macy EXPORT_SYMBOL(zil_open);
4191eda14cbcSMatt Macy EXPORT_SYMBOL(zil_close);
4192eda14cbcSMatt Macy EXPORT_SYMBOL(zil_replay);
4193eda14cbcSMatt Macy EXPORT_SYMBOL(zil_replaying);
4194eda14cbcSMatt Macy EXPORT_SYMBOL(zil_destroy);
4195eda14cbcSMatt Macy EXPORT_SYMBOL(zil_destroy_sync);
4196eda14cbcSMatt Macy EXPORT_SYMBOL(zil_itx_create);
4197eda14cbcSMatt Macy EXPORT_SYMBOL(zil_itx_destroy);
4198eda14cbcSMatt Macy EXPORT_SYMBOL(zil_itx_assign);
4199eda14cbcSMatt Macy EXPORT_SYMBOL(zil_commit);
4200eda14cbcSMatt Macy EXPORT_SYMBOL(zil_claim);
4201eda14cbcSMatt Macy EXPORT_SYMBOL(zil_check_log_chain);
4202eda14cbcSMatt Macy EXPORT_SYMBOL(zil_sync);
4203eda14cbcSMatt Macy EXPORT_SYMBOL(zil_clean);
4204eda14cbcSMatt Macy EXPORT_SYMBOL(zil_suspend);
4205eda14cbcSMatt Macy EXPORT_SYMBOL(zil_resume);
4206eda14cbcSMatt Macy EXPORT_SYMBOL(zil_lwb_add_block);
4207eda14cbcSMatt Macy EXPORT_SYMBOL(zil_bp_tree_add);
4208eda14cbcSMatt Macy EXPORT_SYMBOL(zil_set_sync);
4209eda14cbcSMatt Macy EXPORT_SYMBOL(zil_set_logbias);
4210271171e0SMartin Matuska EXPORT_SYMBOL(zil_sums_init);
4211271171e0SMartin Matuska EXPORT_SYMBOL(zil_sums_fini);
4212271171e0SMartin Matuska EXPORT_SYMBOL(zil_kstat_values_update);
4213eda14cbcSMatt Macy 
4214be181ee2SMartin Matuska ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW,
4215eda14cbcSMatt Macy 	"ZIL block open timeout percentage");
4216eda14cbcSMatt Macy 
421715f0b8c3SMartin Matuska ZFS_MODULE_PARAM(zfs_zil, zil_, min_commit_timeout, U64, ZMOD_RW,
421815f0b8c3SMartin Matuska 	"Minimum delay we care for ZIL block commit");
421915f0b8c3SMartin Matuska 
4220eda14cbcSMatt Macy ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
4221eda14cbcSMatt Macy 	"Disable intent logging replay");
4222eda14cbcSMatt Macy 
4223eda14cbcSMatt Macy ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
4224eda14cbcSMatt Macy 	"Disable ZIL cache flushes");
4225eda14cbcSMatt Macy 
4226dbd5678dSMartin Matuska ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, U64, ZMOD_RW,
4227eda14cbcSMatt Macy 	"Limit in bytes slog sync writes per commit");
4228eda14cbcSMatt Macy 
4229be181ee2SMartin Matuska ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW,
4230eda14cbcSMatt Macy 	"Limit in bytes of ZIL log block size");
4231b2526e8bSMartin Matuska 
4232b2526e8bSMartin Matuska ZFS_MODULE_PARAM(zfs_zil, zil_, maxcopied, UINT, ZMOD_RW,
4233b2526e8bSMartin Matuska 	"Limit in bytes WR_COPIED size");
4234