]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
1d3ba0bf | 9 | * or https://opensource.org/licenses/CDDL-1.0. |
34dc7c2f BB |
10 | * See the License for the specific language governing permissions |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
492f64e9 | 23 | * Copyright (c) 2011, 2018 by Delphix. All rights reserved. |
55922e73 | 24 | * Copyright (c) 2014 Integros [integros.com] |
2ffd89fc | 25 | * Copyright (c) 2018 Datto Inc. |
34dc7c2f BB |
26 | */ |
27 | ||
428870ff BB |
28 | /* Portions Copyright 2010 Robert Milkowski */ |
29 | ||
34dc7c2f BB |
30 | #include <sys/zfs_context.h> |
31 | #include <sys/spa.h> | |
d2734cce | 32 | #include <sys/spa_impl.h> |
34dc7c2f BB |
33 | #include <sys/dmu.h> |
34 | #include <sys/zap.h> | |
35 | #include <sys/arc.h> | |
36 | #include <sys/stat.h> | |
34dc7c2f BB |
37 | #include <sys/zil.h> |
38 | #include <sys/zil_impl.h> | |
39 | #include <sys/dsl_dataset.h> | |
572e2857 | 40 | #include <sys/vdev_impl.h> |
34dc7c2f | 41 | #include <sys/dmu_tx.h> |
428870ff | 42 | #include <sys/dsl_pool.h> |
920dd524 | 43 | #include <sys/metaslab.h> |
e5d1c27e | 44 | #include <sys/trace_zfs.h> |
a6255b7f | 45 | #include <sys/abd.h> |
67a1b037 | 46 | #include <sys/brt.h> |
fb087146 | 47 | #include <sys/wmsum.h> |
34dc7c2f BB |
48 | |
49 | /* | |
1ce23dca PS |
50 | * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system |
51 | * calls that change the file system. Each itx has enough information to | |
52 | * be able to replay them after a system crash, power loss, or | |
53 | * equivalent failure mode. These are stored in memory until either: | |
34dc7c2f | 54 | * |
1ce23dca PS |
55 | * 1. they are committed to the pool by the DMU transaction group |
56 | * (txg), at which point they can be discarded; or | |
57 | * 2. they are committed to the on-disk ZIL for the dataset being | |
58 | * modified (e.g. due to an fsync, O_DSYNC, or other synchronous | |
59 | * requirement). | |
34dc7c2f | 60 | * |
1ce23dca PS |
61 | * In the event of a crash or power loss, the itxs contained by each |
62 | * dataset's on-disk ZIL will be replayed when that dataset is first | |
e1cfd73f | 63 | * instantiated (e.g. if the dataset is a normal filesystem, when it is |
1ce23dca | 64 | * first mounted). |
34dc7c2f | 65 | * |
1ce23dca PS |
66 | * As hinted at above, there is one ZIL per dataset (both the in-memory |
67 | * representation, and the on-disk representation). The on-disk format | |
68 | * consists of 3 parts: | |
69 | * | |
70 | * - a single, per-dataset, ZIL header; which points to a chain of | |
71 | * - zero or more ZIL blocks; each of which contains | |
72 | * - zero or more ZIL records | |
73 | * | |
74 | * A ZIL record holds the information necessary to replay a single | |
75 | * system call transaction. A ZIL block can hold many ZIL records, and | |
76 | * the blocks are chained together, similarly to a singly linked list. | |
77 | * | |
78 | * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL | |
79 | * block in the chain, and the ZIL header points to the first block in | |
80 | * the chain. | |
81 | * | |
82 | * Note, there is not a fixed place in the pool to hold these ZIL | |
83 | * blocks; they are dynamically allocated and freed as needed from the | |
84 | * blocks available on the pool, though they can be preferentially | |
85 | * allocated from a dedicated "log" vdev. | |
34dc7c2f BB |
86 | */ |
87 | ||
1ce23dca PS |
88 | /* |
89 | * This controls the amount of time that a ZIL block (lwb) will remain | |
90 | * "open" when it isn't "full", and it has a thread waiting for it to be | |
91 | * committed to stable storage. Please refer to the zil_commit_waiter() | |
92 | * function (and the comments within it) for more details. | |
93 | */ | |
252f46be | 94 | static uint_t zfs_commit_timeout_pct = 10; |
0f740a4f | 95 | |
b6ad9671 ED |
96 | /* |
97 | * See zil.h for more information about these fields. | |
98 | */ | |
fb087146 | 99 | static zil_kstat_values_t zil_stats = { |
d1d7e268 MK |
100 | { "zil_commit_count", KSTAT_DATA_UINT64 }, |
101 | { "zil_commit_writer_count", KSTAT_DATA_UINT64 }, | |
102 | { "zil_itx_count", KSTAT_DATA_UINT64 }, | |
103 | { "zil_itx_indirect_count", KSTAT_DATA_UINT64 }, | |
104 | { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 }, | |
105 | { "zil_itx_copied_count", KSTAT_DATA_UINT64 }, | |
106 | { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 }, | |
107 | { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 }, | |
108 | { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 }, | |
109 | { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 }, | |
110 | { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 }, | |
b6fbe61f AM |
111 | { "zil_itx_metaslab_normal_write", KSTAT_DATA_UINT64 }, |
112 | { "zil_itx_metaslab_normal_alloc", KSTAT_DATA_UINT64 }, | |
d1d7e268 MK |
113 | { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 }, |
114 | { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 }, | |
b6fbe61f AM |
115 | { "zil_itx_metaslab_slog_write", KSTAT_DATA_UINT64 }, |
116 | { "zil_itx_metaslab_slog_alloc", KSTAT_DATA_UINT64 }, | |
b6ad9671 ED |
117 | }; |
118 | ||
fb087146 AH |
119 | static zil_sums_t zil_sums_global; |
120 | static kstat_t *zil_kstats_global; | |
b6ad9671 | 121 | |
34dc7c2f | 122 | /* |
d3cc8b15 | 123 | * Disable intent logging replay. This global ZIL switch affects all pools. |
34dc7c2f | 124 | */ |
d3cc8b15 | 125 | int zil_replay_disable = 0; |
34dc7c2f BB |
126 | |
127 | /* | |
53b1f5ea PS |
128 | * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to |
129 | * the disk(s) by the ZIL after an LWB write has completed. Setting this | |
130 | * will cause ZIL corruption on power loss if a volatile out-of-order | |
131 | * write cache is enabled. | |
34dc7c2f | 132 | */ |
18168da7 | 133 | static int zil_nocacheflush = 0; |
34dc7c2f | 134 | |
1b7c1e5c GDN |
135 | /* |
136 | * Limit SLOG write size per commit executed with synchronous priority. | |
137 | * Any writes above that will be executed with lower (asynchronous) priority | |
138 | * to limit potential SLOG device abuse by single active ZIL writer. | |
139 | */ | |
c0e58995 | 140 | static uint64_t zil_slog_bulk = 64 * 1024 * 1024; |
1b7c1e5c | 141 | |
34dc7c2f | 142 | static kmem_cache_t *zil_lwb_cache; |
1ce23dca | 143 | static kmem_cache_t *zil_zcw_cache; |
34dc7c2f | 144 | |
f63811f0 AM |
145 | static void zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx); |
146 | static itx_t *zil_itx_clone(itx_t *oitx); | |
147 | ||
34dc7c2f | 148 | static int |
428870ff | 149 | zil_bp_compare(const void *x1, const void *x2) |
34dc7c2f | 150 | { |
428870ff BB |
151 | const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; |
152 | const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; | |
34dc7c2f | 153 | |
ca577779 | 154 | int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2)); |
ee36c709 GN |
155 | if (likely(cmp)) |
156 | return (cmp); | |
34dc7c2f | 157 | |
ca577779 | 158 | return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2))); |
34dc7c2f BB |
159 | } |
160 | ||
161 | static void | |
428870ff | 162 | zil_bp_tree_init(zilog_t *zilog) |
34dc7c2f | 163 | { |
428870ff BB |
164 | avl_create(&zilog->zl_bp_tree, zil_bp_compare, |
165 | sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); | |
34dc7c2f BB |
166 | } |
167 | ||
168 | static void | |
428870ff | 169 | zil_bp_tree_fini(zilog_t *zilog) |
34dc7c2f | 170 | { |
428870ff BB |
171 | avl_tree_t *t = &zilog->zl_bp_tree; |
172 | zil_bp_node_t *zn; | |
34dc7c2f BB |
173 | void *cookie = NULL; |
174 | ||
175 | while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) | |
428870ff | 176 | kmem_free(zn, sizeof (zil_bp_node_t)); |
34dc7c2f BB |
177 | |
178 | avl_destroy(t); | |
179 | } | |
180 | ||
428870ff BB |
181 | int |
182 | zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) | |
34dc7c2f | 183 | { |
428870ff | 184 | avl_tree_t *t = &zilog->zl_bp_tree; |
9b67f605 | 185 | const dva_t *dva; |
428870ff | 186 | zil_bp_node_t *zn; |
34dc7c2f BB |
187 | avl_index_t where; |
188 | ||
9b67f605 MA |
189 | if (BP_IS_EMBEDDED(bp)) |
190 | return (0); | |
191 | ||
192 | dva = BP_IDENTITY(bp); | |
193 | ||
34dc7c2f | 194 | if (avl_find(t, dva, &where) != NULL) |
2e528b49 | 195 | return (SET_ERROR(EEXIST)); |
34dc7c2f | 196 | |
79c76d5b | 197 | zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); |
34dc7c2f BB |
198 | zn->zn_dva = *dva; |
199 | avl_insert(t, zn, where); | |
200 | ||
201 | return (0); | |
202 | } | |
203 | ||
204 | static zil_header_t * | |
205 | zil_header_in_syncing_context(zilog_t *zilog) | |
206 | { | |
207 | return ((zil_header_t *)zilog->zl_header); | |
208 | } | |
209 | ||
210 | static void | |
211 | zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) | |
212 | { | |
213 | zio_cksum_t *zc = &bp->blk_cksum; | |
214 | ||
29274c9f AM |
215 | (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0], |
216 | sizeof (zc->zc_word[ZIL_ZC_GUID_0])); | |
217 | (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1], | |
218 | sizeof (zc->zc_word[ZIL_ZC_GUID_1])); | |
34dc7c2f BB |
219 | zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); |
220 | zc->zc_word[ZIL_ZC_SEQ] = 1ULL; | |
221 | } | |
222 | ||
fb087146 AH |
223 | static int |
224 | zil_kstats_global_update(kstat_t *ksp, int rw) | |
225 | { | |
226 | zil_kstat_values_t *zs = ksp->ks_data; | |
227 | ASSERT3P(&zil_stats, ==, zs); | |
228 | ||
229 | if (rw == KSTAT_WRITE) { | |
230 | return (SET_ERROR(EACCES)); | |
231 | } | |
232 | ||
233 | zil_kstat_values_update(zs, &zil_sums_global); | |
234 | ||
235 | return (0); | |
236 | } | |
237 | ||
34dc7c2f | 238 | /* |
428870ff | 239 | * Read a log block and make sure it's valid. |
34dc7c2f BB |
240 | */ |
241 | static int | |
b5256303 | 242 | zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp, |
482da24e | 243 | blkptr_t *nbp, char **begin, char **end, arc_buf_t **abuf) |
34dc7c2f | 244 | { |
4938d01d | 245 | zio_flag_t zio_flags = ZIO_FLAG_CANFAIL; |
2a432414 | 246 | arc_flags_t aflags = ARC_FLAG_WAIT; |
5dbd68a3 | 247 | zbookmark_phys_t zb; |
34dc7c2f BB |
248 | int error; |
249 | ||
428870ff BB |
250 | if (zilog->zl_header->zh_claim_txg == 0) |
251 | zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; | |
34dc7c2f | 252 | |
428870ff BB |
253 | if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) |
254 | zio_flags |= ZIO_FLAG_SPECULATIVE; | |
34dc7c2f | 255 | |
b5256303 TC |
256 | if (!decrypt) |
257 | zio_flags |= ZIO_FLAG_RAW; | |
258 | ||
428870ff BB |
259 | SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], |
260 | ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
261 | ||
b5256303 | 262 | error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, |
482da24e | 263 | abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); |
34dc7c2f BB |
264 | |
265 | if (error == 0) { | |
34dc7c2f BB |
266 | zio_cksum_t cksum = bp->blk_cksum; |
267 | ||
268 | /* | |
b128c09f BB |
269 | * Validate the checksummed log block. |
270 | * | |
34dc7c2f BB |
271 | * Sequence numbers should be... sequential. The checksum |
272 | * verifier for the next block should be bp's checksum plus 1. | |
b128c09f BB |
273 | * |
274 | * Also check the log chain linkage and size used. | |
34dc7c2f BB |
275 | */ |
276 | cksum.zc_word[ZIL_ZC_SEQ]++; | |
277 | ||
482da24e | 278 | uint64_t size = BP_GET_LSIZE(bp); |
428870ff | 279 | if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { |
482da24e | 280 | zil_chain_t *zilc = (*abuf)->b_data; |
428870ff | 281 | char *lr = (char *)(zilc + 1); |
34dc7c2f | 282 | |
861166b0 | 283 | if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum, |
8e20e0ff | 284 | sizeof (cksum)) || |
482da24e AM |
285 | zilc->zc_nused < sizeof (*zilc) || |
286 | zilc->zc_nused > size) { | |
2e528b49 | 287 | error = SET_ERROR(ECKSUM); |
428870ff | 288 | } else { |
482da24e AM |
289 | *begin = lr; |
290 | *end = lr + zilc->zc_nused - sizeof (*zilc); | |
428870ff BB |
291 | *nbp = zilc->zc_next_blk; |
292 | } | |
293 | } else { | |
482da24e | 294 | char *lr = (*abuf)->b_data; |
428870ff BB |
295 | zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; |
296 | ||
861166b0 | 297 | if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum, |
8e20e0ff | 298 | sizeof (cksum)) || |
428870ff | 299 | (zilc->zc_nused > (size - sizeof (*zilc)))) { |
2e528b49 | 300 | error = SET_ERROR(ECKSUM); |
428870ff | 301 | } else { |
482da24e AM |
302 | *begin = lr; |
303 | *end = lr + zilc->zc_nused; | |
428870ff BB |
304 | *nbp = zilc->zc_next_blk; |
305 | } | |
34dc7c2f | 306 | } |
428870ff BB |
307 | } |
308 | ||
309 | return (error); | |
310 | } | |
311 | ||
312 | /* | |
313 | * Read a TX_WRITE log data block. | |
314 | */ | |
315 | static int | |
316 | zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) | |
317 | { | |
4938d01d | 318 | zio_flag_t zio_flags = ZIO_FLAG_CANFAIL; |
428870ff | 319 | const blkptr_t *bp = &lr->lr_blkptr; |
2a432414 | 320 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff | 321 | arc_buf_t *abuf = NULL; |
5dbd68a3 | 322 | zbookmark_phys_t zb; |
428870ff BB |
323 | int error; |
324 | ||
325 | if (BP_IS_HOLE(bp)) { | |
326 | if (wbuf != NULL) | |
861166b0 | 327 | memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length)); |
428870ff | 328 | return (0); |
34dc7c2f BB |
329 | } |
330 | ||
428870ff BB |
331 | if (zilog->zl_header->zh_claim_txg == 0) |
332 | zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; | |
333 | ||
b5256303 TC |
334 | /* |
335 | * If we are not using the resulting data, we are just checking that | |
336 | * it hasn't been corrupted so we don't need to waste CPU time | |
337 | * decompressing and decrypting it. | |
338 | */ | |
339 | if (wbuf == NULL) | |
340 | zio_flags |= ZIO_FLAG_RAW; | |
341 | ||
a6ccb36b | 342 | ASSERT3U(BP_GET_LSIZE(bp), !=, 0); |
428870ff BB |
343 | SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, |
344 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
345 | ||
294f6806 | 346 | error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, |
428870ff BB |
347 | ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); |
348 | ||
349 | if (error == 0) { | |
350 | if (wbuf != NULL) | |
861166b0 | 351 | memcpy(wbuf, abuf->b_data, arc_buf_size(abuf)); |
d3c2ae1c | 352 | arc_buf_destroy(abuf, &abuf); |
428870ff | 353 | } |
34dc7c2f BB |
354 | |
355 | return (error); | |
356 | } | |
357 | ||
fb087146 AH |
358 | void |
359 | zil_sums_init(zil_sums_t *zs) | |
360 | { | |
361 | wmsum_init(&zs->zil_commit_count, 0); | |
362 | wmsum_init(&zs->zil_commit_writer_count, 0); | |
363 | wmsum_init(&zs->zil_itx_count, 0); | |
364 | wmsum_init(&zs->zil_itx_indirect_count, 0); | |
365 | wmsum_init(&zs->zil_itx_indirect_bytes, 0); | |
366 | wmsum_init(&zs->zil_itx_copied_count, 0); | |
367 | wmsum_init(&zs->zil_itx_copied_bytes, 0); | |
368 | wmsum_init(&zs->zil_itx_needcopy_count, 0); | |
369 | wmsum_init(&zs->zil_itx_needcopy_bytes, 0); | |
370 | wmsum_init(&zs->zil_itx_metaslab_normal_count, 0); | |
371 | wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0); | |
b6fbe61f AM |
372 | wmsum_init(&zs->zil_itx_metaslab_normal_write, 0); |
373 | wmsum_init(&zs->zil_itx_metaslab_normal_alloc, 0); | |
fb087146 AH |
374 | wmsum_init(&zs->zil_itx_metaslab_slog_count, 0); |
375 | wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0); | |
b6fbe61f AM |
376 | wmsum_init(&zs->zil_itx_metaslab_slog_write, 0); |
377 | wmsum_init(&zs->zil_itx_metaslab_slog_alloc, 0); | |
fb087146 AH |
378 | } |
379 | ||
380 | void | |
381 | zil_sums_fini(zil_sums_t *zs) | |
382 | { | |
383 | wmsum_fini(&zs->zil_commit_count); | |
384 | wmsum_fini(&zs->zil_commit_writer_count); | |
385 | wmsum_fini(&zs->zil_itx_count); | |
386 | wmsum_fini(&zs->zil_itx_indirect_count); | |
387 | wmsum_fini(&zs->zil_itx_indirect_bytes); | |
388 | wmsum_fini(&zs->zil_itx_copied_count); | |
389 | wmsum_fini(&zs->zil_itx_copied_bytes); | |
390 | wmsum_fini(&zs->zil_itx_needcopy_count); | |
391 | wmsum_fini(&zs->zil_itx_needcopy_bytes); | |
392 | wmsum_fini(&zs->zil_itx_metaslab_normal_count); | |
393 | wmsum_fini(&zs->zil_itx_metaslab_normal_bytes); | |
b6fbe61f AM |
394 | wmsum_fini(&zs->zil_itx_metaslab_normal_write); |
395 | wmsum_fini(&zs->zil_itx_metaslab_normal_alloc); | |
fb087146 AH |
396 | wmsum_fini(&zs->zil_itx_metaslab_slog_count); |
397 | wmsum_fini(&zs->zil_itx_metaslab_slog_bytes); | |
b6fbe61f AM |
398 | wmsum_fini(&zs->zil_itx_metaslab_slog_write); |
399 | wmsum_fini(&zs->zil_itx_metaslab_slog_alloc); | |
fb087146 AH |
400 | } |
401 | ||
402 | void | |
403 | zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums) | |
404 | { | |
405 | zs->zil_commit_count.value.ui64 = | |
406 | wmsum_value(&zil_sums->zil_commit_count); | |
407 | zs->zil_commit_writer_count.value.ui64 = | |
408 | wmsum_value(&zil_sums->zil_commit_writer_count); | |
409 | zs->zil_itx_count.value.ui64 = | |
410 | wmsum_value(&zil_sums->zil_itx_count); | |
411 | zs->zil_itx_indirect_count.value.ui64 = | |
412 | wmsum_value(&zil_sums->zil_itx_indirect_count); | |
413 | zs->zil_itx_indirect_bytes.value.ui64 = | |
414 | wmsum_value(&zil_sums->zil_itx_indirect_bytes); | |
415 | zs->zil_itx_copied_count.value.ui64 = | |
416 | wmsum_value(&zil_sums->zil_itx_copied_count); | |
417 | zs->zil_itx_copied_bytes.value.ui64 = | |
418 | wmsum_value(&zil_sums->zil_itx_copied_bytes); | |
419 | zs->zil_itx_needcopy_count.value.ui64 = | |
420 | wmsum_value(&zil_sums->zil_itx_needcopy_count); | |
421 | zs->zil_itx_needcopy_bytes.value.ui64 = | |
422 | wmsum_value(&zil_sums->zil_itx_needcopy_bytes); | |
423 | zs->zil_itx_metaslab_normal_count.value.ui64 = | |
424 | wmsum_value(&zil_sums->zil_itx_metaslab_normal_count); | |
425 | zs->zil_itx_metaslab_normal_bytes.value.ui64 = | |
426 | wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes); | |
b6fbe61f AM |
427 | zs->zil_itx_metaslab_normal_write.value.ui64 = |
428 | wmsum_value(&zil_sums->zil_itx_metaslab_normal_write); | |
429 | zs->zil_itx_metaslab_normal_alloc.value.ui64 = | |
430 | wmsum_value(&zil_sums->zil_itx_metaslab_normal_alloc); | |
fb087146 AH |
431 | zs->zil_itx_metaslab_slog_count.value.ui64 = |
432 | wmsum_value(&zil_sums->zil_itx_metaslab_slog_count); | |
433 | zs->zil_itx_metaslab_slog_bytes.value.ui64 = | |
434 | wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes); | |
b6fbe61f AM |
435 | zs->zil_itx_metaslab_slog_write.value.ui64 = |
436 | wmsum_value(&zil_sums->zil_itx_metaslab_slog_write); | |
437 | zs->zil_itx_metaslab_slog_alloc.value.ui64 = | |
438 | wmsum_value(&zil_sums->zil_itx_metaslab_slog_alloc); | |
fb087146 AH |
439 | } |
440 | ||
34dc7c2f BB |
441 | /* |
442 | * Parse the intent log, and call parse_func for each valid record within. | |
34dc7c2f | 443 | */ |
428870ff | 444 | int |
34dc7c2f | 445 | zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, |
b5256303 TC |
446 | zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg, |
447 | boolean_t decrypt) | |
34dc7c2f BB |
448 | { |
449 | const zil_header_t *zh = zilog->zl_header; | |
428870ff BB |
450 | boolean_t claimed = !!zh->zh_claim_txg; |
451 | uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; | |
452 | uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; | |
453 | uint64_t max_blk_seq = 0; | |
454 | uint64_t max_lr_seq = 0; | |
455 | uint64_t blk_count = 0; | |
456 | uint64_t lr_count = 0; | |
861166b0 | 457 | blkptr_t blk, next_blk = {{{{0}}}}; |
428870ff | 458 | int error = 0; |
34dc7c2f | 459 | |
428870ff BB |
460 | /* |
461 | * Old logs didn't record the maximum zh_claim_lr_seq. | |
462 | */ | |
463 | if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) | |
464 | claim_lr_seq = UINT64_MAX; | |
34dc7c2f BB |
465 | |
466 | /* | |
467 | * Starting at the block pointed to by zh_log we read the log chain. | |
468 | * For each block in the chain we strongly check that block to | |
469 | * ensure its validity. We stop when an invalid block is found. | |
470 | * For each block pointer in the chain we call parse_blk_func(). | |
471 | * For each record in each valid block we call parse_lr_func(). | |
472 | * If the log has been claimed, stop if we encounter a sequence | |
473 | * number greater than the highest claimed sequence number. | |
474 | */ | |
428870ff | 475 | zil_bp_tree_init(zilog); |
34dc7c2f | 476 | |
428870ff BB |
477 | for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { |
478 | uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; | |
479 | int reclen; | |
482da24e AM |
480 | char *lrp, *end; |
481 | arc_buf_t *abuf = NULL; | |
34dc7c2f | 482 | |
428870ff BB |
483 | if (blk_seq > claim_blk_seq) |
484 | break; | |
b5256303 TC |
485 | |
486 | error = parse_blk_func(zilog, &blk, arg, txg); | |
487 | if (error != 0) | |
428870ff BB |
488 | break; |
489 | ASSERT3U(max_blk_seq, <, blk_seq); | |
490 | max_blk_seq = blk_seq; | |
491 | blk_count++; | |
34dc7c2f | 492 | |
428870ff BB |
493 | if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) |
494 | break; | |
34dc7c2f | 495 | |
b5256303 | 496 | error = zil_read_log_block(zilog, decrypt, &blk, &next_blk, |
482da24e | 497 | &lrp, &end, &abuf); |
748b9d5b | 498 | if (error != 0) { |
482da24e AM |
499 | if (abuf) |
500 | arc_buf_destroy(abuf, &abuf); | |
748b9d5b RM |
501 | if (claimed) { |
502 | char name[ZFS_MAX_DATASET_NAME_LEN]; | |
503 | ||
504 | dmu_objset_name(zilog->zl_os, name); | |
505 | ||
506 | cmn_err(CE_WARN, "ZFS read log block error %d, " | |
507 | "dataset %s, seq 0x%llx\n", error, name, | |
508 | (u_longlong_t)blk_seq); | |
509 | } | |
34dc7c2f | 510 | break; |
748b9d5b | 511 | } |
34dc7c2f | 512 | |
482da24e | 513 | for (; lrp < end; lrp += reclen) { |
34dc7c2f BB |
514 | lr_t *lr = (lr_t *)lrp; |
515 | reclen = lr->lrc_reclen; | |
516 | ASSERT3U(reclen, >=, sizeof (lr_t)); | |
2a27fd41 | 517 | ASSERT3U(reclen, <=, end - lrp); |
8e8acabd AM |
518 | if (lr->lrc_seq > claim_lr_seq) { |
519 | arc_buf_destroy(abuf, &abuf); | |
428870ff | 520 | goto done; |
8e8acabd | 521 | } |
b5256303 TC |
522 | |
523 | error = parse_lr_func(zilog, lr, arg, txg); | |
8e8acabd AM |
524 | if (error != 0) { |
525 | arc_buf_destroy(abuf, &abuf); | |
428870ff | 526 | goto done; |
8e8acabd | 527 | } |
428870ff BB |
528 | ASSERT3U(max_lr_seq, <, lr->lrc_seq); |
529 | max_lr_seq = lr->lrc_seq; | |
530 | lr_count++; | |
34dc7c2f | 531 | } |
482da24e | 532 | arc_buf_destroy(abuf, &abuf); |
34dc7c2f | 533 | } |
428870ff BB |
534 | done: |
535 | zilog->zl_parse_error = error; | |
536 | zilog->zl_parse_blk_seq = max_blk_seq; | |
537 | zilog->zl_parse_lr_seq = max_lr_seq; | |
538 | zilog->zl_parse_blk_count = blk_count; | |
539 | zilog->zl_parse_lr_count = lr_count; | |
540 | ||
428870ff | 541 | zil_bp_tree_fini(zilog); |
34dc7c2f | 542 | |
428870ff | 543 | return (error); |
34dc7c2f BB |
544 | } |
545 | ||
d2734cce | 546 | static int |
61868bb1 CS |
547 | zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, |
548 | uint64_t first_txg) | |
d2734cce | 549 | { |
14e4e3cb | 550 | (void) tx; |
d2734cce SD |
551 | ASSERT(!BP_IS_HOLE(bp)); |
552 | ||
553 | /* | |
554 | * As we call this function from the context of a rewind to a | |
555 | * checkpoint, each ZIL block whose txg is later than the txg | |
556 | * that we rewind to is invalid. Thus, we return -1 so | |
557 | * zil_parse() doesn't attempt to read it. | |
558 | */ | |
559 | if (bp->blk_birth >= first_txg) | |
560 | return (-1); | |
561 | ||
562 | if (zil_bp_tree_add(zilog, bp) != 0) | |
563 | return (0); | |
564 | ||
565 | zio_free(zilog->zl_spa, first_txg, bp); | |
566 | return (0); | |
567 | } | |
568 | ||
d2734cce | 569 | static int |
61868bb1 CS |
570 | zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, |
571 | uint64_t first_txg) | |
d2734cce | 572 | { |
14e4e3cb | 573 | (void) zilog, (void) lrc, (void) tx, (void) first_txg; |
d2734cce SD |
574 | return (0); |
575 | } | |
576 | ||
428870ff | 577 | static int |
61868bb1 CS |
578 | zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, |
579 | uint64_t first_txg) | |
34dc7c2f | 580 | { |
34dc7c2f BB |
581 | /* |
582 | * Claim log block if not already committed and not already claimed. | |
428870ff | 583 | * If tx == NULL, just verify that the block is claimable. |
34dc7c2f | 584 | */ |
b0bc7a84 MG |
585 | if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || |
586 | zil_bp_tree_add(zilog, bp) != 0) | |
428870ff BB |
587 | return (0); |
588 | ||
589 | return (zio_wait(zio_claim(NULL, zilog->zl_spa, | |
590 | tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, | |
591 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); | |
34dc7c2f BB |
592 | } |
593 | ||
428870ff | 594 | static int |
67a1b037 | 595 | zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg) |
34dc7c2f | 596 | { |
428870ff BB |
597 | lr_write_t *lr = (lr_write_t *)lrc; |
598 | int error; | |
599 | ||
2a27fd41 | 600 | ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr)); |
428870ff BB |
601 | |
602 | /* | |
603 | * If the block is not readable, don't claim it. This can happen | |
604 | * in normal operation when a log block is written to disk before | |
605 | * some of the dmu_sync() blocks it points to. In this case, the | |
606 | * transaction cannot have been committed to anyone (we would have | |
607 | * waited for all writes to be stable first), so it is semantically | |
608 | * correct to declare this the end of the log. | |
609 | */ | |
b5256303 TC |
610 | if (lr->lr_blkptr.blk_birth >= first_txg) { |
611 | error = zil_read_log_data(zilog, lr, NULL); | |
612 | if (error != 0) | |
613 | return (error); | |
614 | } | |
615 | ||
428870ff | 616 | return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); |
34dc7c2f BB |
617 | } |
618 | ||
67a1b037 | 619 | static int |
55b764e0 AM |
620 | zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx, |
621 | uint64_t first_txg) | |
67a1b037 PJD |
622 | { |
623 | const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc; | |
624 | const blkptr_t *bp; | |
55b764e0 | 625 | spa_t *spa = zilog->zl_spa; |
67a1b037 PJD |
626 | uint_t ii; |
627 | ||
2a27fd41 AM |
628 | ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr)); |
629 | ASSERT3U(lrc->lrc_reclen, >=, offsetof(lr_clone_range_t, | |
630 | lr_bps[lr->lr_nbps])); | |
67a1b037 PJD |
631 | |
632 | if (tx == NULL) { | |
633 | return (0); | |
634 | } | |
635 | ||
636 | /* | |
637 | * XXX: Do we need to byteswap lr? | |
638 | */ | |
639 | ||
67a1b037 PJD |
640 | for (ii = 0; ii < lr->lr_nbps; ii++) { |
641 | bp = &lr->lr_bps[ii]; | |
642 | ||
643 | /* | |
55b764e0 AM |
644 | * When data is embedded into the BP there is no need to create |
645 | * BRT entry as there is no data block. Just copy the BP as it | |
646 | * contains the data. | |
67a1b037 | 647 | */ |
55b764e0 AM |
648 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) |
649 | continue; | |
650 | ||
651 | /* | |
652 | * We can not handle block pointers from the future, since they | |
653 | * are not yet allocated. It should not normally happen, but | |
654 | * just in case lets be safe and just stop here now instead of | |
655 | * corrupting the pool. | |
656 | */ | |
657 | if (BP_PHYSICAL_BIRTH(bp) >= first_txg) | |
658 | return (SET_ERROR(ENOENT)); | |
659 | ||
660 | /* | |
661 | * Assert the block is really allocated before we reference it. | |
662 | */ | |
663 | metaslab_check_free(spa, bp); | |
664 | } | |
665 | ||
666 | for (ii = 0; ii < lr->lr_nbps; ii++) { | |
667 | bp = &lr->lr_bps[ii]; | |
668 | if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) | |
67a1b037 | 669 | brt_pending_add(spa, bp, tx); |
67a1b037 PJD |
670 | } |
671 | ||
672 | return (0); | |
673 | } | |
674 | ||
675 | static int | |
676 | zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, | |
677 | uint64_t first_txg) | |
678 | { | |
679 | ||
680 | switch (lrc->lrc_txtype) { | |
681 | case TX_WRITE: | |
682 | return (zil_claim_write(zilog, lrc, tx, first_txg)); | |
683 | case TX_CLONE_RANGE: | |
55b764e0 | 684 | return (zil_claim_clone_range(zilog, lrc, tx, first_txg)); |
67a1b037 PJD |
685 | default: |
686 | return (0); | |
687 | } | |
688 | } | |
689 | ||
428870ff | 690 | static int |
61868bb1 CS |
691 | zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, |
692 | uint64_t claim_txg) | |
34dc7c2f | 693 | { |
14e4e3cb AZ |
694 | (void) claim_txg; |
695 | ||
d2734cce | 696 | zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); |
428870ff BB |
697 | |
698 | return (0); | |
34dc7c2f BB |
699 | } |
700 | ||
428870ff | 701 | static int |
67a1b037 | 702 | zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg) |
34dc7c2f | 703 | { |
428870ff BB |
704 | lr_write_t *lr = (lr_write_t *)lrc; |
705 | blkptr_t *bp = &lr->lr_blkptr; | |
706 | ||
2a27fd41 | 707 | ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr)); |
67a1b037 | 708 | |
34dc7c2f BB |
709 | /* |
710 | * If we previously claimed it, we need to free it. | |
711 | */ | |
67a1b037 PJD |
712 | if (bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && |
713 | !BP_IS_HOLE(bp)) { | |
428870ff | 714 | zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); |
67a1b037 | 715 | } |
428870ff BB |
716 | |
717 | return (0); | |
718 | } | |
719 | ||
67a1b037 PJD |
720 | static int |
721 | zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx) | |
722 | { | |
723 | const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc; | |
724 | const blkptr_t *bp; | |
725 | spa_t *spa; | |
726 | uint_t ii; | |
727 | ||
2a27fd41 AM |
728 | ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr)); |
729 | ASSERT3U(lrc->lrc_reclen, >=, offsetof(lr_clone_range_t, | |
730 | lr_bps[lr->lr_nbps])); | |
67a1b037 PJD |
731 | |
732 | if (tx == NULL) { | |
733 | return (0); | |
734 | } | |
735 | ||
736 | spa = zilog->zl_spa; | |
737 | ||
738 | for (ii = 0; ii < lr->lr_nbps; ii++) { | |
739 | bp = &lr->lr_bps[ii]; | |
740 | ||
741 | if (!BP_IS_HOLE(bp)) { | |
742 | zio_free(spa, dmu_tx_get_txg(tx), bp); | |
743 | } | |
744 | } | |
745 | ||
746 | return (0); | |
747 | } | |
748 | ||
749 | static int | |
750 | zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, | |
751 | uint64_t claim_txg) | |
752 | { | |
753 | ||
754 | if (claim_txg == 0) { | |
755 | return (0); | |
756 | } | |
757 | ||
758 | switch (lrc->lrc_txtype) { | |
759 | case TX_WRITE: | |
760 | return (zil_free_write(zilog, lrc, tx, claim_txg)); | |
761 | case TX_CLONE_RANGE: | |
762 | return (zil_free_clone_range(zilog, lrc, tx)); | |
763 | default: | |
764 | return (0); | |
765 | } | |
766 | } | |
767 | ||
1ce23dca PS |
768 | static int |
769 | zil_lwb_vdev_compare(const void *x1, const void *x2) | |
770 | { | |
771 | const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; | |
772 | const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; | |
773 | ||
ca577779 | 774 | return (TREE_CMP(v1, v2)); |
1ce23dca PS |
775 | } |
776 | ||
eda3fcd5 AM |
777 | /* |
778 | * Allocate a new lwb. We may already have a block pointer for it, in which | |
779 | * case we get size and version from there. Or we may not yet, in which case | |
780 | * we choose them here and later make the block allocation match. | |
781 | */ | |
428870ff | 782 | static lwb_t * |
eda3fcd5 AM |
783 | zil_alloc_lwb(zilog_t *zilog, int sz, blkptr_t *bp, boolean_t slog, |
784 | uint64_t txg, lwb_state_t state) | |
428870ff BB |
785 | { |
786 | lwb_t *lwb; | |
787 | ||
79c76d5b | 788 | lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); |
428870ff | 789 | lwb->lwb_zilog = zilog; |
eda3fcd5 AM |
790 | if (bp) { |
791 | lwb->lwb_blk = *bp; | |
792 | lwb->lwb_slim = (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2); | |
793 | sz = BP_GET_LSIZE(bp); | |
794 | } else { | |
795 | BP_ZERO(&lwb->lwb_blk); | |
796 | lwb->lwb_slim = (spa_version(zilog->zl_spa) >= | |
797 | SPA_VERSION_SLIM_ZIL); | |
798 | } | |
1b7c1e5c | 799 | lwb->lwb_slog = slog; |
eda3fcd5 AM |
800 | lwb->lwb_error = 0; |
801 | if (lwb->lwb_slim) { | |
802 | lwb->lwb_nmax = sz; | |
f63811f0 | 803 | lwb->lwb_nused = lwb->lwb_nfilled = sizeof (zil_chain_t); |
f63811f0 | 804 | } else { |
eda3fcd5 | 805 | lwb->lwb_nmax = sz - sizeof (zil_chain_t); |
f63811f0 | 806 | lwb->lwb_nused = lwb->lwb_nfilled = 0; |
f63811f0 | 807 | } |
eda3fcd5 AM |
808 | lwb->lwb_sz = sz; |
809 | lwb->lwb_state = state; | |
810 | lwb->lwb_buf = zio_buf_alloc(sz); | |
811 | lwb->lwb_child_zio = NULL; | |
1ce23dca PS |
812 | lwb->lwb_write_zio = NULL; |
813 | lwb->lwb_root_zio = NULL; | |
1ce23dca | 814 | lwb->lwb_issued_timestamp = 0; |
152d6fda | 815 | lwb->lwb_issued_txg = 0; |
eda3fcd5 AM |
816 | lwb->lwb_alloc_txg = txg; |
817 | lwb->lwb_max_txg = 0; | |
428870ff BB |
818 | |
819 | mutex_enter(&zilog->zl_lock); | |
820 | list_insert_tail(&zilog->zl_lwb_list, lwb); | |
eda3fcd5 AM |
821 | if (state != LWB_STATE_NEW) |
822 | zilog->zl_last_lwb_opened = lwb; | |
428870ff BB |
823 | mutex_exit(&zilog->zl_lock); |
824 | ||
825 | return (lwb); | |
34dc7c2f BB |
826 | } |
827 | ||
1ce23dca PS |
828 | static void |
829 | zil_free_lwb(zilog_t *zilog, lwb_t *lwb) | |
830 | { | |
831 | ASSERT(MUTEX_HELD(&zilog->zl_lock)); | |
bbcf18c2 AM |
832 | ASSERT(lwb->lwb_state == LWB_STATE_NEW || |
833 | lwb->lwb_state == LWB_STATE_FLUSH_DONE); | |
eda3fcd5 | 834 | ASSERT3P(lwb->lwb_child_zio, ==, NULL); |
1ce23dca PS |
835 | ASSERT3P(lwb->lwb_write_zio, ==, NULL); |
836 | ASSERT3P(lwb->lwb_root_zio, ==, NULL); | |
eda3fcd5 | 837 | ASSERT3U(lwb->lwb_alloc_txg, <=, spa_syncing_txg(zilog->zl_spa)); |
2fe61a7e | 838 | ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa)); |
bbcf18c2 AM |
839 | VERIFY(list_is_empty(&lwb->lwb_itxs)); |
840 | VERIFY(list_is_empty(&lwb->lwb_waiters)); | |
841 | ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); | |
842 | ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); | |
1ce23dca PS |
843 | |
844 | /* | |
845 | * Clear the zilog's field to indicate this lwb is no longer | |
846 | * valid, and prevent use-after-free errors. | |
847 | */ | |
848 | if (zilog->zl_last_lwb_opened == lwb) | |
849 | zilog->zl_last_lwb_opened = NULL; | |
850 | ||
851 | kmem_cache_free(zil_lwb_cache, lwb); | |
852 | } | |
853 | ||
29809a6c MA |
854 | /* |
855 | * Called when we create in-memory log transactions so that we know | |
856 | * to cleanup the itxs at the end of spa_sync(). | |
857 | */ | |
65c7cc49 | 858 | static void |
29809a6c MA |
859 | zilog_dirty(zilog_t *zilog, uint64_t txg) |
860 | { | |
861 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
862 | dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); | |
863 | ||
1ce23dca PS |
864 | ASSERT(spa_writeable(zilog->zl_spa)); |
865 | ||
0c66c32d | 866 | if (ds->ds_is_snapshot) |
29809a6c MA |
867 | panic("dirtying snapshot!"); |
868 | ||
13fe0198 | 869 | if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { |
29809a6c MA |
870 | /* up the hold count until we can be written out */ |
871 | dmu_buf_add_ref(ds->ds_dbuf, zilog); | |
1ce23dca PS |
872 | |
873 | zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg); | |
29809a6c MA |
874 | } |
875 | } | |
876 | ||
55922e73 GW |
877 | /* |
878 | * Determine if the zil is dirty in the specified txg. Callers wanting to | |
879 | * ensure that the dirty state does not change must hold the itxg_lock for | |
880 | * the specified txg. Holding the lock will ensure that the zil cannot be | |
881 | * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current | |
882 | * state. | |
883 | */ | |
65c7cc49 | 884 | static boolean_t __maybe_unused |
55922e73 GW |
885 | zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) |
886 | { | |
887 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
888 | ||
889 | if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) | |
890 | return (B_TRUE); | |
891 | return (B_FALSE); | |
892 | } | |
893 | ||
894 | /* | |
895 | * Determine if the zil is dirty. The zil is considered dirty if it has | |
896 | * any pending itx records that have not been cleaned by zil_clean(). | |
897 | */ | |
65c7cc49 | 898 | static boolean_t |
29809a6c MA |
899 | zilog_is_dirty(zilog_t *zilog) |
900 | { | |
901 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
29809a6c | 902 | |
1c27024e | 903 | for (int t = 0; t < TXG_SIZE; t++) { |
29809a6c MA |
904 | if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) |
905 | return (B_TRUE); | |
906 | } | |
907 | return (B_FALSE); | |
908 | } | |
909 | ||
361a7e82 JP |
910 | /* |
911 | * Its called in zil_commit context (zil_process_commit_list()/zil_create()). | |
912 | * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled. | |
913 | * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every | |
914 | * zil_commit. | |
915 | */ | |
916 | static void | |
917 | zil_commit_activate_saxattr_feature(zilog_t *zilog) | |
918 | { | |
919 | dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); | |
920 | uint64_t txg = 0; | |
921 | dmu_tx_t *tx = NULL; | |
922 | ||
dbf6108b | 923 | if (spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) && |
361a7e82 | 924 | dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL && |
dbf6108b | 925 | !dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)) { |
361a7e82 JP |
926 | tx = dmu_tx_create(zilog->zl_os); |
927 | VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); | |
928 | dsl_dataset_dirty(ds, tx); | |
929 | txg = dmu_tx_get_txg(tx); | |
930 | ||
931 | mutex_enter(&ds->ds_lock); | |
932 | ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] = | |
933 | (void *)B_TRUE; | |
934 | mutex_exit(&ds->ds_lock); | |
935 | dmu_tx_commit(tx); | |
936 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
937 | } | |
938 | } | |
939 | ||
34dc7c2f BB |
940 | /* |
941 | * Create an on-disk intent log. | |
942 | */ | |
428870ff | 943 | static lwb_t * |
34dc7c2f BB |
944 | zil_create(zilog_t *zilog) |
945 | { | |
946 | const zil_header_t *zh = zilog->zl_header; | |
428870ff | 947 | lwb_t *lwb = NULL; |
34dc7c2f BB |
948 | uint64_t txg = 0; |
949 | dmu_tx_t *tx = NULL; | |
950 | blkptr_t blk; | |
951 | int error = 0; | |
1b7c1e5c | 952 | boolean_t slog = FALSE; |
361a7e82 JP |
953 | dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); |
954 | ||
34dc7c2f BB |
955 | |
956 | /* | |
957 | * Wait for any previous destroy to complete. | |
958 | */ | |
959 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
960 | ||
961 | ASSERT(zh->zh_claim_txg == 0); | |
962 | ASSERT(zh->zh_replay_seq == 0); | |
963 | ||
964 | blk = zh->zh_log; | |
965 | ||
966 | /* | |
428870ff BB |
967 | * Allocate an initial log block if: |
968 | * - there isn't one already | |
4e33ba4c | 969 | * - the existing block is the wrong endianness |
34dc7c2f | 970 | */ |
fb5f0bc8 | 971 | if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { |
34dc7c2f | 972 | tx = dmu_tx_create(zilog->zl_os); |
1ce23dca | 973 | VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); |
34dc7c2f BB |
974 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); |
975 | txg = dmu_tx_get_txg(tx); | |
976 | ||
fb5f0bc8 | 977 | if (!BP_IS_HOLE(&blk)) { |
d2734cce | 978 | zio_free(zilog->zl_spa, txg, &blk); |
fb5f0bc8 BB |
979 | BP_ZERO(&blk); |
980 | } | |
981 | ||
b5256303 | 982 | error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk, |
1b7c1e5c | 983 | ZIL_MIN_BLKSZ, &slog); |
34dc7c2f BB |
984 | if (error == 0) |
985 | zil_init_log_chain(zilog, &blk); | |
986 | } | |
987 | ||
988 | /* | |
1ce23dca | 989 | * Allocate a log write block (lwb) for the first log block. |
34dc7c2f | 990 | */ |
428870ff | 991 | if (error == 0) |
eda3fcd5 | 992 | lwb = zil_alloc_lwb(zilog, 0, &blk, slog, txg, LWB_STATE_NEW); |
34dc7c2f BB |
993 | |
994 | /* | |
995 | * If we just allocated the first log block, commit our transaction | |
2fe61a7e | 996 | * and wait for zil_sync() to stuff the block pointer into zh_log. |
34dc7c2f BB |
997 | * (zh is part of the MOS, so we cannot modify it in open context.) |
998 | */ | |
999 | if (tx != NULL) { | |
361a7e82 JP |
1000 | /* |
1001 | * If "zilsaxattr" feature is enabled on zpool, then activate | |
1002 | * it now when we're creating the ZIL chain. We can't wait with | |
1003 | * this until we write the first xattr log record because we | |
1004 | * need to wait for the feature activation to sync out. | |
1005 | */ | |
1006 | if (spa_feature_is_enabled(zilog->zl_spa, | |
1007 | SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) != | |
1008 | DMU_OST_ZVOL) { | |
1009 | mutex_enter(&ds->ds_lock); | |
1010 | ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] = | |
1011 | (void *)B_TRUE; | |
1012 | mutex_exit(&ds->ds_lock); | |
1013 | } | |
1014 | ||
34dc7c2f BB |
1015 | dmu_tx_commit(tx); |
1016 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
361a7e82 JP |
1017 | } else { |
1018 | /* | |
1019 | * This branch covers the case where we enable the feature on a | |
1020 | * zpool that has existing ZIL headers. | |
1021 | */ | |
1022 | zil_commit_activate_saxattr_feature(zilog); | |
34dc7c2f | 1023 | } |
361a7e82 JP |
1024 | IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) && |
1025 | dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL, | |
1026 | dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)); | |
34dc7c2f | 1027 | |
861166b0 | 1028 | ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); |
c04812f9 | 1029 | IMPLY(error == 0, lwb != NULL); |
428870ff BB |
1030 | |
1031 | return (lwb); | |
34dc7c2f BB |
1032 | } |
1033 | ||
1034 | /* | |
1ce23dca PS |
1035 | * In one tx, free all log blocks and clear the log header. If keep_first |
1036 | * is set, then we're replaying a log with no content. We want to keep the | |
1037 | * first block, however, so that the first synchronous transaction doesn't | |
1038 | * require a txg_wait_synced() in zil_create(). We don't need to | |
1039 | * txg_wait_synced() here either when keep_first is set, because both | |
1040 | * zil_create() and zil_destroy() will wait for any in-progress destroys | |
1041 | * to complete. | |
e197bb24 | 1042 | * Return B_TRUE if there were any entries to replay. |
34dc7c2f | 1043 | */ |
e197bb24 | 1044 | boolean_t |
34dc7c2f BB |
1045 | zil_destroy(zilog_t *zilog, boolean_t keep_first) |
1046 | { | |
1047 | const zil_header_t *zh = zilog->zl_header; | |
1048 | lwb_t *lwb; | |
1049 | dmu_tx_t *tx; | |
1050 | uint64_t txg; | |
1051 | ||
1052 | /* | |
1053 | * Wait for any previous destroy to complete. | |
1054 | */ | |
1055 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
1056 | ||
428870ff BB |
1057 | zilog->zl_old_header = *zh; /* debugging aid */ |
1058 | ||
34dc7c2f | 1059 | if (BP_IS_HOLE(&zh->zh_log)) |
e197bb24 | 1060 | return (B_FALSE); |
34dc7c2f BB |
1061 | |
1062 | tx = dmu_tx_create(zilog->zl_os); | |
1ce23dca | 1063 | VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); |
34dc7c2f BB |
1064 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); |
1065 | txg = dmu_tx_get_txg(tx); | |
1066 | ||
1067 | mutex_enter(&zilog->zl_lock); | |
1068 | ||
34dc7c2f BB |
1069 | ASSERT3U(zilog->zl_destroy_txg, <, txg); |
1070 | zilog->zl_destroy_txg = txg; | |
1071 | zilog->zl_keep_first = keep_first; | |
1072 | ||
1073 | if (!list_is_empty(&zilog->zl_lwb_list)) { | |
1074 | ASSERT(zh->zh_claim_txg == 0); | |
3e31d2b0 | 1075 | VERIFY(!keep_first); |
895e0313 | 1076 | while ((lwb = list_remove_head(&zilog->zl_lwb_list)) != NULL) { |
34dc7c2f BB |
1077 | if (lwb->lwb_buf != NULL) |
1078 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
eda3fcd5 AM |
1079 | if (!BP_IS_HOLE(&lwb->lwb_blk)) |
1080 | zio_free(zilog->zl_spa, txg, &lwb->lwb_blk); | |
1ce23dca | 1081 | zil_free_lwb(zilog, lwb); |
34dc7c2f | 1082 | } |
428870ff | 1083 | } else if (!keep_first) { |
29809a6c | 1084 | zil_destroy_sync(zilog, tx); |
34dc7c2f BB |
1085 | } |
1086 | mutex_exit(&zilog->zl_lock); | |
1087 | ||
1088 | dmu_tx_commit(tx); | |
e197bb24 AS |
1089 | |
1090 | return (B_TRUE); | |
34dc7c2f BB |
1091 | } |
1092 | ||
29809a6c MA |
1093 | void |
1094 | zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) | |
1095 | { | |
1096 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
1097 | (void) zil_parse(zilog, zil_free_log_block, | |
b5256303 | 1098 | zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE); |
29809a6c MA |
1099 | } |
1100 | ||
34dc7c2f | 1101 | int |
9c43027b | 1102 | zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) |
34dc7c2f BB |
1103 | { |
1104 | dmu_tx_t *tx = txarg; | |
34dc7c2f | 1105 | zilog_t *zilog; |
d2734cce | 1106 | uint64_t first_txg; |
34dc7c2f BB |
1107 | zil_header_t *zh; |
1108 | objset_t *os; | |
1109 | int error; | |
1110 | ||
9c43027b | 1111 | error = dmu_objset_own_obj(dp, ds->ds_object, |
b5256303 | 1112 | DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os); |
13fe0198 | 1113 | if (error != 0) { |
6d9036f3 MA |
1114 | /* |
1115 | * EBUSY indicates that the objset is inconsistent, in which | |
1116 | * case it can not have a ZIL. | |
1117 | */ | |
1118 | if (error != EBUSY) { | |
9c43027b AJ |
1119 | cmn_err(CE_WARN, "can't open objset for %llu, error %u", |
1120 | (unsigned long long)ds->ds_object, error); | |
6d9036f3 MA |
1121 | } |
1122 | ||
34dc7c2f BB |
1123 | return (0); |
1124 | } | |
1125 | ||
1126 | zilog = dmu_objset_zil(os); | |
1127 | zh = zil_header_in_syncing_context(zilog); | |
d2734cce SD |
1128 | ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa)); |
1129 | first_txg = spa_min_claim_txg(zilog->zl_spa); | |
34dc7c2f | 1130 | |
d2734cce SD |
1131 | /* |
1132 | * If the spa_log_state is not set to be cleared, check whether | |
1133 | * the current uberblock is a checkpoint one and if the current | |
1134 | * header has been claimed before moving on. | |
1135 | * | |
1136 | * If the current uberblock is a checkpointed uberblock then | |
1137 | * one of the following scenarios took place: | |
1138 | * | |
1139 | * 1] We are currently rewinding to the checkpoint of the pool. | |
1140 | * 2] We crashed in the middle of a checkpoint rewind but we | |
1141 | * did manage to write the checkpointed uberblock to the | |
1142 | * vdev labels, so when we tried to import the pool again | |
1143 | * the checkpointed uberblock was selected from the import | |
1144 | * procedure. | |
1145 | * | |
1146 | * In both cases we want to zero out all the ZIL blocks, except | |
1147 | * the ones that have been claimed at the time of the checkpoint | |
1148 | * (their zh_claim_txg != 0). The reason is that these blocks | |
1149 | * may be corrupted since we may have reused their locations on | |
1150 | * disk after we took the checkpoint. | |
1151 | * | |
1152 | * We could try to set spa_log_state to SPA_LOG_CLEAR earlier | |
1153 | * when we first figure out whether the current uberblock is | |
1154 | * checkpointed or not. Unfortunately, that would discard all | |
1155 | * the logs, including the ones that are claimed, and we would | |
1156 | * leak space. | |
1157 | */ | |
1158 | if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR || | |
1159 | (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && | |
1160 | zh->zh_claim_txg == 0)) { | |
1161 | if (!BP_IS_HOLE(&zh->zh_log)) { | |
1162 | (void) zil_parse(zilog, zil_clear_log_block, | |
1163 | zil_noop_log_record, tx, first_txg, B_FALSE); | |
1164 | } | |
9babb374 | 1165 | BP_ZERO(&zh->zh_log); |
b5256303 | 1166 | if (os->os_encrypted) |
1b66810b | 1167 | os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE; |
9babb374 | 1168 | dsl_dataset_dirty(dmu_objset_ds(os), tx); |
b5256303 | 1169 | dmu_objset_disown(os, B_FALSE, FTAG); |
428870ff | 1170 | return (0); |
9babb374 BB |
1171 | } |
1172 | ||
d2734cce SD |
1173 | /* |
1174 | * If we are not rewinding and opening the pool normally, then | |
1175 | * the min_claim_txg should be equal to the first txg of the pool. | |
1176 | */ | |
1177 | ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa)); | |
1178 | ||
34dc7c2f BB |
1179 | /* |
1180 | * Claim all log blocks if we haven't already done so, and remember | |
1181 | * the highest claimed sequence number. This ensures that if we can | |
1182 | * read only part of the log now (e.g. due to a missing device), | |
1183 | * but we can read the entire log later, we will not try to replay | |
1184 | * or destroy beyond the last block we successfully claimed. | |
1185 | */ | |
1186 | ASSERT3U(zh->zh_claim_txg, <=, first_txg); | |
1187 | if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { | |
428870ff | 1188 | (void) zil_parse(zilog, zil_claim_log_block, |
b5256303 | 1189 | zil_claim_log_record, tx, first_txg, B_FALSE); |
428870ff BB |
1190 | zh->zh_claim_txg = first_txg; |
1191 | zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; | |
1192 | zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; | |
1193 | if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) | |
1194 | zh->zh_flags |= ZIL_REPLAY_NEEDED; | |
1195 | zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; | |
d53bd7f5 | 1196 | if (os->os_encrypted) |
1b66810b | 1197 | os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE; |
34dc7c2f BB |
1198 | dsl_dataset_dirty(dmu_objset_ds(os), tx); |
1199 | } | |
1200 | ||
1201 | ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); | |
b5256303 | 1202 | dmu_objset_disown(os, B_FALSE, FTAG); |
34dc7c2f BB |
1203 | return (0); |
1204 | } | |
1205 | ||
b128c09f BB |
1206 | /* |
1207 | * Check the log by walking the log chain. | |
1208 | * Checksum errors are ok as they indicate the end of the chain. | |
1209 | * Any other error (no device or read failure) returns an error. | |
1210 | */ | |
b128c09f | 1211 | int |
9c43027b | 1212 | zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) |
b128c09f | 1213 | { |
14e4e3cb | 1214 | (void) dp; |
b128c09f | 1215 | zilog_t *zilog; |
b128c09f | 1216 | objset_t *os; |
572e2857 | 1217 | blkptr_t *bp; |
b128c09f BB |
1218 | int error; |
1219 | ||
428870ff BB |
1220 | ASSERT(tx == NULL); |
1221 | ||
9c43027b | 1222 | error = dmu_objset_from_ds(ds, &os); |
13fe0198 | 1223 | if (error != 0) { |
9c43027b AJ |
1224 | cmn_err(CE_WARN, "can't open objset %llu, error %d", |
1225 | (unsigned long long)ds->ds_object, error); | |
b128c09f BB |
1226 | return (0); |
1227 | } | |
1228 | ||
1229 | zilog = dmu_objset_zil(os); | |
572e2857 BB |
1230 | bp = (blkptr_t *)&zilog->zl_header->zh_log; |
1231 | ||
572e2857 BB |
1232 | if (!BP_IS_HOLE(bp)) { |
1233 | vdev_t *vd; | |
1234 | boolean_t valid = B_TRUE; | |
1235 | ||
d2734cce SD |
1236 | /* |
1237 | * Check the first block and determine if it's on a log device | |
1238 | * which may have been removed or faulted prior to loading this | |
1239 | * pool. If so, there's no point in checking the rest of the | |
1240 | * log as its content should have already been synced to the | |
1241 | * pool. | |
1242 | */ | |
572e2857 BB |
1243 | spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); |
1244 | vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); | |
1245 | if (vd->vdev_islog && vdev_is_dead(vd)) | |
1246 | valid = vdev_log_state_valid(vd); | |
1247 | spa_config_exit(os->os_spa, SCL_STATE, FTAG); | |
1248 | ||
9c43027b | 1249 | if (!valid) |
572e2857 | 1250 | return (0); |
d2734cce SD |
1251 | |
1252 | /* | |
1253 | * Check whether the current uberblock is checkpointed (e.g. | |
1254 | * we are rewinding) and whether the current header has been | |
1255 | * claimed or not. If it hasn't then skip verifying it. We | |
1256 | * do this because its ZIL blocks may be part of the pool's | |
1257 | * state before the rewind, which is no longer valid. | |
1258 | */ | |
1259 | zil_header_t *zh = zil_header_in_syncing_context(zilog); | |
1260 | if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && | |
1261 | zh->zh_claim_txg == 0) | |
1262 | return (0); | |
572e2857 | 1263 | } |
b128c09f | 1264 | |
428870ff BB |
1265 | /* |
1266 | * Because tx == NULL, zil_claim_log_block() will not actually claim | |
1267 | * any blocks, but just determine whether it is possible to do so. | |
1268 | * In addition to checking the log chain, zil_claim_log_block() | |
1269 | * will invoke zio_claim() with a done func of spa_claim_notify(), | |
1270 | * which will update spa_max_claim_txg. See spa_load() for details. | |
1271 | */ | |
1272 | error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, | |
d2734cce SD |
1273 | zilog->zl_header->zh_claim_txg ? -1ULL : |
1274 | spa_min_claim_txg(os->os_spa), B_FALSE); | |
428870ff | 1275 | |
428870ff | 1276 | return ((error == ECKSUM || error == ENOENT) ? 0 : error); |
b128c09f BB |
1277 | } |
1278 | ||
1ce23dca PS |
1279 | /* |
1280 | * When an itx is "skipped", this function is used to properly mark the | |
1281 | * waiter as "done, and signal any thread(s) waiting on it. An itx can | |
1282 | * be skipped (and not committed to an lwb) for a variety of reasons, | |
1283 | * one of them being that the itx was committed via spa_sync(), prior to | |
1284 | * it being committed to an lwb; this can happen if a thread calling | |
1285 | * zil_commit() is racing with spa_sync(). | |
1286 | */ | |
1287 | static void | |
1288 | zil_commit_waiter_skip(zil_commit_waiter_t *zcw) | |
34dc7c2f | 1289 | { |
1ce23dca PS |
1290 | mutex_enter(&zcw->zcw_lock); |
1291 | ASSERT3B(zcw->zcw_done, ==, B_FALSE); | |
1292 | zcw->zcw_done = B_TRUE; | |
1293 | cv_broadcast(&zcw->zcw_cv); | |
1294 | mutex_exit(&zcw->zcw_lock); | |
1295 | } | |
34dc7c2f | 1296 | |
1ce23dca PS |
1297 | /* |
1298 | * This function is used when the given waiter is to be linked into an | |
1299 | * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb. | |
1300 | * At this point, the waiter will no longer be referenced by the itx, | |
1301 | * and instead, will be referenced by the lwb. | |
1302 | */ | |
1303 | static void | |
1304 | zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb) | |
1305 | { | |
2fe61a7e PS |
1306 | /* |
1307 | * The lwb_waiters field of the lwb is protected by the zilog's | |
eda3fcd5 AM |
1308 | * zl_issuer_lock while the lwb is open and zl_lock otherwise. |
1309 | * zl_issuer_lock also protects leaving the open state. | |
1310 | * zcw_lwb setting is protected by zl_issuer_lock and state != | |
1311 | * flush_done, which transition is protected by zl_lock. | |
2fe61a7e | 1312 | */ |
eda3fcd5 AM |
1313 | ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_issuer_lock)); |
1314 | IMPLY(lwb->lwb_state != LWB_STATE_OPENED, | |
1315 | MUTEX_HELD(&lwb->lwb_zilog->zl_lock)); | |
1316 | ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW); | |
1317 | ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); | |
2fe61a7e | 1318 | |
1ce23dca | 1319 | ASSERT(!list_link_active(&zcw->zcw_node)); |
1ce23dca | 1320 | list_insert_tail(&lwb->lwb_waiters, zcw); |
eda3fcd5 | 1321 | ASSERT3P(zcw->zcw_lwb, ==, NULL); |
1ce23dca | 1322 | zcw->zcw_lwb = lwb; |
1ce23dca PS |
1323 | } |
1324 | ||
1325 | /* | |
1326 | * This function is used when zio_alloc_zil() fails to allocate a ZIL | |
1327 | * block, and the given waiter must be linked to the "nolwb waiters" | |
1328 | * list inside of zil_process_commit_list(). | |
1329 | */ | |
1330 | static void | |
1331 | zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb) | |
1332 | { | |
1ce23dca | 1333 | ASSERT(!list_link_active(&zcw->zcw_node)); |
1ce23dca | 1334 | list_insert_tail(nolwb, zcw); |
eda3fcd5 | 1335 | ASSERT3P(zcw->zcw_lwb, ==, NULL); |
34dc7c2f BB |
1336 | } |
1337 | ||
1338 | void | |
1ce23dca | 1339 | zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp) |
34dc7c2f | 1340 | { |
1ce23dca | 1341 | avl_tree_t *t = &lwb->lwb_vdev_tree; |
34dc7c2f BB |
1342 | avl_index_t where; |
1343 | zil_vdev_node_t *zv, zvsearch; | |
1344 | int ndvas = BP_GET_NDVAS(bp); | |
1345 | int i; | |
1346 | ||
bbcf18c2 AM |
1347 | ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); |
1348 | ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); | |
1349 | ||
53b1f5ea | 1350 | if (zil_nocacheflush) |
34dc7c2f BB |
1351 | return; |
1352 | ||
1ce23dca | 1353 | mutex_enter(&lwb->lwb_vdev_lock); |
34dc7c2f BB |
1354 | for (i = 0; i < ndvas; i++) { |
1355 | zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); | |
1356 | if (avl_find(t, &zvsearch, &where) == NULL) { | |
79c76d5b | 1357 | zv = kmem_alloc(sizeof (*zv), KM_SLEEP); |
34dc7c2f BB |
1358 | zv->zv_vdev = zvsearch.zv_vdev; |
1359 | avl_insert(t, zv, where); | |
1360 | } | |
1361 | } | |
1ce23dca | 1362 | mutex_exit(&lwb->lwb_vdev_lock); |
34dc7c2f BB |
1363 | } |
1364 | ||
900d09b2 PS |
1365 | static void |
1366 | zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb) | |
1367 | { | |
1368 | avl_tree_t *src = &lwb->lwb_vdev_tree; | |
1369 | avl_tree_t *dst = &nlwb->lwb_vdev_tree; | |
1370 | void *cookie = NULL; | |
1371 | zil_vdev_node_t *zv; | |
1372 | ||
1373 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); | |
1374 | ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE); | |
1375 | ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); | |
1376 | ||
1377 | /* | |
1378 | * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does | |
1379 | * not need the protection of lwb_vdev_lock (it will only be modified | |
1380 | * while holding zilog->zl_lock) as its writes and those of its | |
1381 | * children have all completed. The younger 'nlwb' may be waiting on | |
1382 | * future writes to additional vdevs. | |
1383 | */ | |
1384 | mutex_enter(&nlwb->lwb_vdev_lock); | |
1385 | /* | |
1386 | * Tear down the 'lwb' vdev tree, ensuring that entries which do not | |
1387 | * exist in 'nlwb' are moved to it, freeing any would-be duplicates. | |
1388 | */ | |
1389 | while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) { | |
1390 | avl_index_t where; | |
1391 | ||
1392 | if (avl_find(dst, zv, &where) == NULL) { | |
1393 | avl_insert(dst, zv, where); | |
1394 | } else { | |
1395 | kmem_free(zv, sizeof (*zv)); | |
1396 | } | |
1397 | } | |
1398 | mutex_exit(&nlwb->lwb_vdev_lock); | |
1399 | } | |
1400 | ||
1ce23dca PS |
1401 | void |
1402 | zil_lwb_add_txg(lwb_t *lwb, uint64_t txg) | |
1403 | { | |
1404 | lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); | |
1405 | } | |
1406 | ||
1407 | /* | |
900d09b2 | 1408 | * This function is a called after all vdevs associated with a given lwb |
1ce23dca | 1409 | * write have completed their DKIOCFLUSHWRITECACHE command; or as soon |
900d09b2 PS |
1410 | * as the lwb write completes, if "zil_nocacheflush" is set. Further, |
1411 | * all "previous" lwb's will have completed before this function is | |
1412 | * called; i.e. this function is called for all previous lwbs before | |
1413 | * it's called for "this" lwb (enforced via zio the dependencies | |
1414 | * configured in zil_lwb_set_zio_dependency()). | |
1ce23dca PS |
1415 | * |
1416 | * The intention is for this function to be called as soon as the | |
1417 | * contents of an lwb are considered "stable" on disk, and will survive | |
1418 | * any sudden loss of power. At this point, any threads waiting for the | |
1419 | * lwb to reach this state are signalled, and the "waiter" structures | |
1420 | * are marked "done". | |
1421 | */ | |
572e2857 | 1422 | static void |
1ce23dca | 1423 | zil_lwb_flush_vdevs_done(zio_t *zio) |
34dc7c2f | 1424 | { |
1ce23dca PS |
1425 | lwb_t *lwb = zio->io_private; |
1426 | zilog_t *zilog = lwb->lwb_zilog; | |
1ce23dca PS |
1427 | zil_commit_waiter_t *zcw; |
1428 | itx_t *itx; | |
1429 | ||
a604d324 GW |
1430 | spa_config_exit(zilog->zl_spa, SCL_STATE, lwb); |
1431 | ||
895e0313 | 1432 | hrtime_t t = gethrtime() - lwb->lwb_issued_timestamp; |
34dc7c2f | 1433 | |
1ce23dca | 1434 | mutex_enter(&zilog->zl_lock); |
34dc7c2f | 1435 | |
895e0313 | 1436 | zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 7 + t) / 8; |
34dc7c2f | 1437 | |
1ce23dca | 1438 | lwb->lwb_root_zio = NULL; |
900d09b2 | 1439 | |
b1b99e10 AM |
1440 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); |
1441 | lwb->lwb_state = LWB_STATE_FLUSH_DONE; | |
1442 | ||
1ce23dca PS |
1443 | if (zilog->zl_last_lwb_opened == lwb) { |
1444 | /* | |
1445 | * Remember the highest committed log sequence number | |
1446 | * for ztest. We only update this value when all the log | |
1447 | * writes succeeded, because ztest wants to ASSERT that | |
1448 | * it got the whole log chain. | |
1449 | */ | |
1450 | zilog->zl_commit_lr_seq = zilog->zl_lr_seq; | |
1451 | } | |
1452 | ||
b1b99e10 | 1453 | while ((itx = list_remove_head(&lwb->lwb_itxs)) != NULL) |
1ce23dca | 1454 | zil_itx_destroy(itx); |
1ce23dca | 1455 | |
b1b99e10 | 1456 | while ((zcw = list_remove_head(&lwb->lwb_waiters)) != NULL) { |
1ce23dca PS |
1457 | mutex_enter(&zcw->zcw_lock); |
1458 | ||
b1b99e10 | 1459 | ASSERT3P(zcw->zcw_lwb, ==, lwb); |
1ce23dca | 1460 | zcw->zcw_lwb = NULL; |
f82f0279 AK |
1461 | /* |
1462 | * We expect any ZIO errors from child ZIOs to have been | |
1463 | * propagated "up" to this specific LWB's root ZIO, in | |
1464 | * order for this error handling to work correctly. This | |
1465 | * includes ZIO errors from either this LWB's write or | |
1466 | * flush, as well as any errors from other dependent LWBs | |
1467 | * (e.g. a root LWB ZIO that might be a child of this LWB). | |
1468 | * | |
1469 | * With that said, it's important to note that LWB flush | |
1470 | * errors are not propagated up to the LWB root ZIO. | |
1471 | * This is incorrect behavior, and results in VDEV flush | |
1472 | * errors not being handled correctly here. See the | |
1473 | * comment above the call to "zio_flush" for details. | |
1474 | */ | |
1ce23dca PS |
1475 | |
1476 | zcw->zcw_zio_error = zio->io_error; | |
1477 | ||
1478 | ASSERT3B(zcw->zcw_done, ==, B_FALSE); | |
1479 | zcw->zcw_done = B_TRUE; | |
1480 | cv_broadcast(&zcw->zcw_cv); | |
1481 | ||
1482 | mutex_exit(&zcw->zcw_lock); | |
34dc7c2f | 1483 | } |
b1b99e10 AM |
1484 | |
1485 | uint64_t txg = lwb->lwb_issued_txg; | |
1486 | ||
1487 | /* Once we drop the lock, lwb may be freed by zil_sync(). */ | |
1488 | mutex_exit(&zilog->zl_lock); | |
34dc7c2f | 1489 | |
152d6fda | 1490 | mutex_enter(&zilog->zl_lwb_io_lock); |
152d6fda KJ |
1491 | ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0); |
1492 | zilog->zl_lwb_inflight[txg & TXG_MASK]--; | |
1493 | if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0) | |
1494 | cv_broadcast(&zilog->zl_lwb_io_cv); | |
1495 | mutex_exit(&zilog->zl_lwb_io_lock); | |
1496 | } | |
1497 | ||
1498 | /* | |
1499 | * Wait for the completion of all issued write/flush of that txg provided. | |
1500 | * It guarantees zil_lwb_flush_vdevs_done() is called and returned. | |
1501 | */ | |
1502 | static void | |
1503 | zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg) | |
1504 | { | |
1505 | ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa)); | |
1506 | ||
1507 | mutex_enter(&zilog->zl_lwb_io_lock); | |
1508 | while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0) | |
1509 | cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock); | |
1510 | mutex_exit(&zilog->zl_lwb_io_lock); | |
1511 | ||
1512 | #ifdef ZFS_DEBUG | |
1513 | mutex_enter(&zilog->zl_lock); | |
1514 | mutex_enter(&zilog->zl_lwb_io_lock); | |
1515 | lwb_t *lwb = list_head(&zilog->zl_lwb_list); | |
eda3fcd5 | 1516 | while (lwb != NULL) { |
152d6fda KJ |
1517 | if (lwb->lwb_issued_txg <= txg) { |
1518 | ASSERT(lwb->lwb_state != LWB_STATE_ISSUED); | |
1519 | ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE); | |
1520 | IMPLY(lwb->lwb_issued_txg > 0, | |
1521 | lwb->lwb_state == LWB_STATE_FLUSH_DONE); | |
1522 | } | |
7381ddf1 AM |
1523 | IMPLY(lwb->lwb_state == LWB_STATE_WRITE_DONE || |
1524 | lwb->lwb_state == LWB_STATE_FLUSH_DONE, | |
152d6fda KJ |
1525 | lwb->lwb_buf == NULL); |
1526 | lwb = list_next(&zilog->zl_lwb_list, lwb); | |
1527 | } | |
1528 | mutex_exit(&zilog->zl_lwb_io_lock); | |
1529 | mutex_exit(&zilog->zl_lock); | |
1530 | #endif | |
34dc7c2f BB |
1531 | } |
1532 | ||
1533 | /* | |
900d09b2 PS |
1534 | * This is called when an lwb's write zio completes. The callback's |
1535 | * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs | |
1536 | * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved | |
1537 | * in writing out this specific lwb's data, and in the case that cache | |
1538 | * flushes have been deferred, vdevs involved in writing the data for | |
1539 | * previous lwbs. The writes corresponding to all the vdevs in the | |
1540 | * lwb_vdev_tree will have completed by the time this is called, due to | |
1541 | * the zio dependencies configured in zil_lwb_set_zio_dependency(), | |
1542 | * which takes deferred flushes into account. The lwb will be "done" | |
1543 | * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio | |
1544 | * completion callback for the lwb's root zio. | |
34dc7c2f BB |
1545 | */ |
1546 | static void | |
1547 | zil_lwb_write_done(zio_t *zio) | |
1548 | { | |
1549 | lwb_t *lwb = zio->io_private; | |
1ce23dca | 1550 | spa_t *spa = zio->io_spa; |
34dc7c2f | 1551 | zilog_t *zilog = lwb->lwb_zilog; |
1ce23dca PS |
1552 | avl_tree_t *t = &lwb->lwb_vdev_tree; |
1553 | void *cookie = NULL; | |
1554 | zil_vdev_node_t *zv; | |
900d09b2 | 1555 | lwb_t *nlwb; |
1ce23dca | 1556 | |
a604d324 GW |
1557 | ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0); |
1558 | ||
e2af2acc | 1559 | abd_free(zio->io_abd); |
7381ddf1 AM |
1560 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); |
1561 | lwb->lwb_buf = NULL; | |
1ce23dca | 1562 | |
34dc7c2f | 1563 | mutex_enter(&zilog->zl_lock); |
900d09b2 PS |
1564 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED); |
1565 | lwb->lwb_state = LWB_STATE_WRITE_DONE; | |
eda3fcd5 | 1566 | lwb->lwb_child_zio = NULL; |
1ce23dca | 1567 | lwb->lwb_write_zio = NULL; |
90149552 AM |
1568 | |
1569 | /* | |
1570 | * If nlwb is not yet issued, zil_lwb_set_zio_dependency() is not | |
1571 | * called for it yet, and when it will be, it won't be able to make | |
1572 | * its write ZIO a parent this ZIO. In such case we can not defer | |
1573 | * our flushes or below may be a race between the done callbacks. | |
1574 | */ | |
900d09b2 | 1575 | nlwb = list_next(&zilog->zl_lwb_list, lwb); |
90149552 AM |
1576 | if (nlwb && nlwb->lwb_state != LWB_STATE_ISSUED) |
1577 | nlwb = NULL; | |
428870ff | 1578 | mutex_exit(&zilog->zl_lock); |
9babb374 | 1579 | |
1ce23dca PS |
1580 | if (avl_numnodes(t) == 0) |
1581 | return; | |
1582 | ||
9babb374 | 1583 | /* |
1ce23dca PS |
1584 | * If there was an IO error, we're not going to call zio_flush() |
1585 | * on these vdevs, so we simply empty the tree and free the | |
1586 | * nodes. We avoid calling zio_flush() since there isn't any | |
1587 | * good reason for doing so, after the lwb block failed to be | |
1588 | * written out. | |
f82f0279 AK |
1589 | * |
1590 | * Additionally, we don't perform any further error handling at | |
1591 | * this point (e.g. setting "zcw_zio_error" appropriately), as | |
1592 | * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus, | |
1593 | * we expect any error seen here, to have been propagated to | |
1594 | * that function). | |
9babb374 | 1595 | */ |
1ce23dca PS |
1596 | if (zio->io_error != 0) { |
1597 | while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) | |
1598 | kmem_free(zv, sizeof (*zv)); | |
1599 | return; | |
1600 | } | |
1601 | ||
900d09b2 PS |
1602 | /* |
1603 | * If this lwb does not have any threads waiting for it to | |
1604 | * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE | |
1605 | * command to the vdevs written to by "this" lwb, and instead | |
1606 | * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE | |
1607 | * command for those vdevs. Thus, we merge the vdev tree of | |
1608 | * "this" lwb with the vdev tree of the "next" lwb in the list, | |
1609 | * and assume the "next" lwb will handle flushing the vdevs (or | |
1610 | * deferring the flush(s) again). | |
1611 | * | |
1612 | * This is a useful performance optimization, especially for | |
1613 | * workloads with lots of async write activity and few sync | |
1614 | * write and/or fsync activity, as it has the potential to | |
1615 | * coalesce multiple flush commands to a vdev into one. | |
1616 | */ | |
895e0313 | 1617 | if (list_is_empty(&lwb->lwb_waiters) && nlwb != NULL) { |
900d09b2 PS |
1618 | zil_lwb_flush_defer(lwb, nlwb); |
1619 | ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); | |
1620 | return; | |
1621 | } | |
1622 | ||
1ce23dca PS |
1623 | while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { |
1624 | vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); | |
5a3bffab | 1625 | if (vd != NULL) { |
f82f0279 AK |
1626 | /* |
1627 | * The "ZIO_FLAG_DONT_PROPAGATE" is currently | |
1628 | * always used within "zio_flush". This means, | |
1629 | * any errors when flushing the vdev(s), will | |
1630 | * (unfortunately) not be handled correctly, | |
1631 | * since these "zio_flush" errors will not be | |
1632 | * propagated up to "zil_lwb_flush_vdevs_done". | |
1633 | */ | |
1ce23dca | 1634 | zio_flush(lwb->lwb_root_zio, vd); |
f82f0279 | 1635 | } |
1ce23dca PS |
1636 | kmem_free(zv, sizeof (*zv)); |
1637 | } | |
34dc7c2f BB |
1638 | } |
1639 | ||
eda3fcd5 AM |
1640 | /* |
1641 | * Build the zio dependency chain, which is used to preserve the ordering of | |
1642 | * lwb completions that is required by the semantics of the ZIL. Each new lwb | |
1643 | * zio becomes a parent of the previous lwb zio, such that the new lwb's zio | |
1644 | * cannot complete until the previous lwb's zio completes. | |
1645 | * | |
1646 | * This is required by the semantics of zil_commit(): the commit waiters | |
1647 | * attached to the lwbs will be woken in the lwb zio's completion callback, | |
1648 | * so this zio dependency graph ensures the waiters are woken in the correct | |
1649 | * order (the same order the lwbs were created). | |
1650 | */ | |
900d09b2 PS |
1651 | static void |
1652 | zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb) | |
1653 | { | |
900d09b2 PS |
1654 | ASSERT(MUTEX_HELD(&zilog->zl_lock)); |
1655 | ||
eda3fcd5 AM |
1656 | lwb_t *prev_lwb = list_prev(&zilog->zl_lwb_list, lwb); |
1657 | if (prev_lwb == NULL || | |
1658 | prev_lwb->lwb_state == LWB_STATE_FLUSH_DONE) | |
1659 | return; | |
1660 | ||
900d09b2 | 1661 | /* |
eda3fcd5 AM |
1662 | * If the previous lwb's write hasn't already completed, we also want |
1663 | * to order the completion of the lwb write zios (above, we only order | |
1664 | * the completion of the lwb root zios). This is required because of | |
1665 | * how we can defer the DKIOCFLUSHWRITECACHE commands for each lwb. | |
900d09b2 | 1666 | * |
eda3fcd5 AM |
1667 | * When the DKIOCFLUSHWRITECACHE commands are deferred, the previous |
1668 | * lwb will rely on this lwb to flush the vdevs written to by that | |
1669 | * previous lwb. Thus, we need to ensure this lwb doesn't issue the | |
1670 | * flush until after the previous lwb's write completes. We ensure | |
1671 | * this ordering by setting the zio parent/child relationship here. | |
1672 | * | |
1673 | * Without this relationship on the lwb's write zio, it's possible | |
1674 | * for this lwb's write to complete prior to the previous lwb's write | |
1675 | * completing; and thus, the vdevs for the previous lwb would be | |
1676 | * flushed prior to that lwb's data being written to those vdevs (the | |
1677 | * vdevs are flushed in the lwb write zio's completion handler, | |
1678 | * zil_lwb_write_done()). | |
1679 | */ | |
1680 | if (prev_lwb->lwb_state == LWB_STATE_ISSUED) { | |
1681 | ASSERT3P(prev_lwb->lwb_write_zio, !=, NULL); | |
1682 | zio_add_child(lwb->lwb_write_zio, prev_lwb->lwb_write_zio); | |
1683 | } else { | |
1684 | ASSERT3S(prev_lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); | |
900d09b2 | 1685 | } |
eda3fcd5 AM |
1686 | |
1687 | ASSERT3P(prev_lwb->lwb_root_zio, !=, NULL); | |
1688 | zio_add_child(lwb->lwb_root_zio, prev_lwb->lwb_root_zio); | |
900d09b2 PS |
1689 | } |
1690 | ||
1691 | ||
34dc7c2f | 1692 | /* |
1ce23dca | 1693 | * This function's purpose is to "open" an lwb such that it is ready to |
eda3fcd5 AM |
1694 | * accept new itxs being committed to it. This function is idempotent; if |
1695 | * the passed in lwb has already been opened, it is essentially a no-op. | |
34dc7c2f BB |
1696 | */ |
1697 | static void | |
1ce23dca | 1698 | zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb) |
34dc7c2f | 1699 | { |
1b2b0aca | 1700 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca | 1701 | |
eda3fcd5 AM |
1702 | if (lwb->lwb_state != LWB_STATE_NEW) { |
1703 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); | |
f63811f0 | 1704 | return; |
eda3fcd5 | 1705 | } |
1ce23dca | 1706 | |
b22bab25 | 1707 | mutex_enter(&zilog->zl_lock); |
f63811f0 | 1708 | lwb->lwb_state = LWB_STATE_OPENED; |
f63811f0 | 1709 | zilog->zl_last_lwb_opened = lwb; |
920dd524 | 1710 | mutex_exit(&zilog->zl_lock); |
34dc7c2f BB |
1711 | } |
1712 | ||
428870ff BB |
1713 | /* |
1714 | * Define a limited set of intent log block sizes. | |
d3cc8b15 | 1715 | * |
428870ff BB |
1716 | * These must be a multiple of 4KB. Note only the amount used (again |
1717 | * aligned to 4KB) actually gets written. However, we can't always just | |
f1512ee6 | 1718 | * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. |
428870ff | 1719 | */ |
18168da7 | 1720 | static const struct { |
f15d6a54 AM |
1721 | uint64_t limit; |
1722 | uint64_t blksz; | |
1723 | } zil_block_buckets[] = { | |
1724 | { 4096, 4096 }, /* non TX_WRITE */ | |
1725 | { 8192 + 4096, 8192 + 4096 }, /* database */ | |
1726 | { 32768 + 4096, 32768 + 4096 }, /* NFS writes */ | |
1727 | { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */ | |
f15d6a54 | 1728 | { UINT64_MAX, SPA_OLD_MAXBLOCKSIZE}, /* > 128KB writes */ |
428870ff BB |
1729 | }; |
1730 | ||
b8738257 MA |
1731 | /* |
1732 | * Maximum block size used by the ZIL. This is picked up when the ZIL is | |
1733 | * initialized. Otherwise this should not be used directly; see | |
1734 | * zl_max_block_size instead. | |
1735 | */ | |
fdc2d303 | 1736 | static uint_t zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE; |
b8738257 | 1737 | |
34dc7c2f | 1738 | /* |
f63811f0 AM |
1739 | * Close the log block for being issued and allocate the next one. |
1740 | * Has to be called under zl_issuer_lock to chain more lwbs. | |
34dc7c2f BB |
1741 | */ |
1742 | static lwb_t * | |
eda3fcd5 | 1743 | zil_lwb_write_close(zilog_t *zilog, lwb_t *lwb, lwb_state_t state) |
34dc7c2f | 1744 | { |
eda3fcd5 | 1745 | int i; |
428870ff | 1746 | |
1b2b0aca | 1747 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca | 1748 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); |
eda3fcd5 | 1749 | lwb->lwb_state = LWB_STATE_CLOSED; |
1ce23dca | 1750 | |
f63811f0 | 1751 | /* |
eda3fcd5 AM |
1752 | * If there was an allocation failure then returned NULL will trigger |
1753 | * zil_commit_writer_stall() at the caller. This is inherently racy, | |
1754 | * since allocation may not have happened yet. | |
34dc7c2f | 1755 | */ |
eda3fcd5 AM |
1756 | if (lwb->lwb_error != 0) |
1757 | return (NULL); | |
34dc7c2f BB |
1758 | |
1759 | /* | |
428870ff BB |
1760 | * Log blocks are pre-allocated. Here we select the size of the next |
1761 | * block, based on size used in the last block. | |
1762 | * - first find the smallest bucket that will fit the block from a | |
1763 | * limited set of block sizes. This is because it's faster to write | |
1764 | * blocks allocated from the same metaslab as they are adjacent or | |
1765 | * close. | |
1766 | * - next find the maximum from the new suggested size and an array of | |
1767 | * previous sizes. This lessens a picket fence effect of wrongly | |
2fe61a7e | 1768 | * guessing the size if we have a stream of say 2k, 64k, 2k, 64k |
428870ff BB |
1769 | * requests. |
1770 | * | |
1771 | * Note we only write what is used, but we can't just allocate | |
1772 | * the maximum block size because we can exhaust the available | |
1773 | * pool log space. | |
34dc7c2f | 1774 | */ |
eda3fcd5 | 1775 | uint64_t zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); |
f15d6a54 | 1776 | for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++) |
428870ff | 1777 | continue; |
f15d6a54 | 1778 | zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size); |
428870ff BB |
1779 | zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; |
1780 | for (i = 0; i < ZIL_PREV_BLKS; i++) | |
1781 | zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); | |
b6fbe61f AM |
1782 | DTRACE_PROBE3(zil__block__size, zilog_t *, zilog, |
1783 | uint64_t, zil_blksz, | |
1784 | uint64_t, zilog->zl_prev_blks[zilog->zl_prev_rotor]); | |
428870ff | 1785 | zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); |
34dc7c2f | 1786 | |
eda3fcd5 AM |
1787 | return (zil_alloc_lwb(zilog, zil_blksz, NULL, 0, 0, state)); |
1788 | } | |
34dc7c2f | 1789 | |
eda3fcd5 AM |
1790 | /* |
1791 | * Finalize previously closed block and issue the write zio. | |
1792 | */ | |
1793 | static void | |
1794 | zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb) | |
1795 | { | |
1796 | spa_t *spa = zilog->zl_spa; | |
1797 | zil_chain_t *zilc; | |
1798 | boolean_t slog; | |
1799 | zbookmark_phys_t zb; | |
1800 | zio_priority_t prio; | |
1801 | int error; | |
34dc7c2f | 1802 | |
eda3fcd5 | 1803 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED); |
f63811f0 | 1804 | |
eda3fcd5 AM |
1805 | /* Actually fill the lwb with the data. */ |
1806 | for (itx_t *itx = list_head(&lwb->lwb_itxs); itx; | |
1807 | itx = list_next(&lwb->lwb_itxs, itx)) | |
1808 | zil_lwb_commit(zilog, lwb, itx); | |
1809 | lwb->lwb_nused = lwb->lwb_nfilled; | |
2a27fd41 | 1810 | ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax); |
eda3fcd5 AM |
1811 | |
1812 | lwb->lwb_root_zio = zio_root(spa, zil_lwb_flush_vdevs_done, lwb, | |
1813 | ZIO_FLAG_CANFAIL); | |
f63811f0 | 1814 | |
2cb992a9 | 1815 | /* |
eda3fcd5 AM |
1816 | * The lwb is now ready to be issued, but it can be only if it already |
1817 | * got its block pointer allocated or the allocation has failed. | |
1818 | * Otherwise leave it as-is, relying on some other thread to issue it | |
1819 | * after allocating its block pointer via calling zil_lwb_write_issue() | |
1820 | * for the previous lwb(s) in the chain. | |
2cb992a9 | 1821 | */ |
eda3fcd5 AM |
1822 | mutex_enter(&zilog->zl_lock); |
1823 | lwb->lwb_state = LWB_STATE_READY; | |
1824 | if (BP_IS_HOLE(&lwb->lwb_blk) && lwb->lwb_error == 0) { | |
1825 | mutex_exit(&zilog->zl_lock); | |
1826 | return; | |
1827 | } | |
1828 | mutex_exit(&zilog->zl_lock); | |
1829 | ||
1830 | next_lwb: | |
1831 | if (lwb->lwb_slim) | |
1832 | zilc = (zil_chain_t *)lwb->lwb_buf; | |
1833 | else | |
1834 | zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_nmax); | |
1835 | int wsz = lwb->lwb_sz; | |
1836 | if (lwb->lwb_error == 0) { | |
1837 | abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, lwb->lwb_sz); | |
1838 | if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk) | |
1839 | prio = ZIO_PRIORITY_SYNC_WRITE; | |
1840 | else | |
1841 | prio = ZIO_PRIORITY_ASYNC_WRITE; | |
1842 | SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], | |
1843 | ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
1844 | lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
1845 | lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, spa, 0, | |
1846 | &lwb->lwb_blk, lwb_abd, lwb->lwb_sz, zil_lwb_write_done, | |
1847 | lwb, prio, ZIO_FLAG_CANFAIL, &zb); | |
1848 | zil_lwb_add_block(lwb, &lwb->lwb_blk); | |
1849 | ||
1850 | if (lwb->lwb_slim) { | |
1851 | /* For Slim ZIL only write what is used. */ | |
1852 | wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, | |
1853 | int); | |
1854 | ASSERT3S(wsz, <=, lwb->lwb_sz); | |
1855 | zio_shrink(lwb->lwb_write_zio, wsz); | |
1856 | wsz = lwb->lwb_write_zio->io_size; | |
2cb992a9 | 1857 | } |
eda3fcd5 AM |
1858 | memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused); |
1859 | zilc->zc_pad = 0; | |
1860 | zilc->zc_nused = lwb->lwb_nused; | |
1861 | zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; | |
2cb992a9 | 1862 | } else { |
eda3fcd5 AM |
1863 | /* |
1864 | * We can't write the lwb if there was an allocation failure, | |
1865 | * so create a null zio instead just to maintain dependencies. | |
1866 | */ | |
1867 | lwb->lwb_write_zio = zio_null(lwb->lwb_root_zio, spa, NULL, | |
1868 | zil_lwb_write_done, lwb, ZIO_FLAG_CANFAIL); | |
1869 | lwb->lwb_write_zio->io_error = lwb->lwb_error; | |
2cb992a9 | 1870 | } |
eda3fcd5 AM |
1871 | if (lwb->lwb_child_zio) |
1872 | zio_add_child(lwb->lwb_write_zio, lwb->lwb_child_zio); | |
2cb992a9 | 1873 | |
f63811f0 | 1874 | /* |
eda3fcd5 | 1875 | * Open transaction to allocate the next block pointer. |
f63811f0 | 1876 | */ |
eda3fcd5 AM |
1877 | dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); |
1878 | VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); | |
1879 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
1880 | uint64_t txg = dmu_tx_get_txg(tx); | |
f63811f0 | 1881 | |
eda3fcd5 AM |
1882 | /* |
1883 | * Allocate next the block pointer unless we are already in error. | |
1884 | */ | |
1885 | lwb_t *nlwb = list_next(&zilog->zl_lwb_list, lwb); | |
1886 | blkptr_t *bp = &zilc->zc_next_blk; | |
1887 | BP_ZERO(bp); | |
1888 | error = lwb->lwb_error; | |
1889 | if (error == 0) { | |
1890 | error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, nlwb->lwb_sz, | |
1891 | &slog); | |
1892 | } | |
1893 | if (error == 0) { | |
1894 | ASSERT3U(bp->blk_birth, ==, txg); | |
1895 | BP_SET_CHECKSUM(bp, nlwb->lwb_slim ? ZIO_CHECKSUM_ZILOG2 : | |
1896 | ZIO_CHECKSUM_ZILOG); | |
1897 | bp->blk_cksum = lwb->lwb_blk.blk_cksum; | |
1898 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; | |
f63811f0 AM |
1899 | } |
1900 | ||
eda3fcd5 AM |
1901 | /* |
1902 | * Reduce TXG open time by incrementing inflight counter and committing | |
1903 | * the transaciton. zil_sync() will wait for it to return to zero. | |
1904 | */ | |
1905 | mutex_enter(&zilog->zl_lwb_io_lock); | |
1906 | lwb->lwb_issued_txg = txg; | |
1907 | zilog->zl_lwb_inflight[txg & TXG_MASK]++; | |
1908 | zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg); | |
1909 | mutex_exit(&zilog->zl_lwb_io_lock); | |
1910 | dmu_tx_commit(tx); | |
34dc7c2f | 1911 | |
eda3fcd5 | 1912 | spa_config_enter(spa, SCL_STATE, lwb, RW_READER); |
34dc7c2f BB |
1913 | |
1914 | /* | |
eda3fcd5 AM |
1915 | * We've completed all potentially blocking operations. Update the |
1916 | * nlwb and allow it proceed without possible lock order reversals. | |
34dc7c2f | 1917 | */ |
eda3fcd5 AM |
1918 | mutex_enter(&zilog->zl_lock); |
1919 | zil_lwb_set_zio_dependency(zilog, lwb); | |
1920 | lwb->lwb_state = LWB_STATE_ISSUED; | |
1921 | ||
1922 | if (nlwb) { | |
1923 | nlwb->lwb_blk = *bp; | |
1924 | nlwb->lwb_error = error; | |
1925 | nlwb->lwb_slog = slog; | |
1926 | nlwb->lwb_alloc_txg = txg; | |
1927 | if (nlwb->lwb_state != LWB_STATE_READY) | |
1928 | nlwb = NULL; | |
1929 | } | |
1930 | mutex_exit(&zilog->zl_lock); | |
34dc7c2f | 1931 | |
f63811f0 AM |
1932 | if (lwb->lwb_slog) { |
1933 | ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count); | |
1934 | ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes, | |
1935 | lwb->lwb_nused); | |
b6fbe61f AM |
1936 | ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_write, |
1937 | wsz); | |
1938 | ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_alloc, | |
1939 | BP_GET_LSIZE(&lwb->lwb_blk)); | |
f63811f0 AM |
1940 | } else { |
1941 | ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count); | |
1942 | ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes, | |
1943 | lwb->lwb_nused); | |
b6fbe61f AM |
1944 | ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_write, |
1945 | wsz); | |
1946 | ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_alloc, | |
1947 | BP_GET_LSIZE(&lwb->lwb_blk)); | |
f63811f0 | 1948 | } |
1ce23dca | 1949 | lwb->lwb_issued_timestamp = gethrtime(); |
eda3fcd5 AM |
1950 | if (lwb->lwb_child_zio) |
1951 | zio_nowait(lwb->lwb_child_zio); | |
9da6b604 AM |
1952 | zio_nowait(lwb->lwb_write_zio); |
1953 | zio_nowait(lwb->lwb_root_zio); | |
eda3fcd5 AM |
1954 | |
1955 | /* | |
1956 | * If nlwb was ready when we gave it the block pointer, | |
1957 | * it is on us to issue it and possibly following ones. | |
1958 | */ | |
1959 | lwb = nlwb; | |
1960 | if (lwb) | |
1961 | goto next_lwb; | |
34dc7c2f BB |
1962 | } |
1963 | ||
b8738257 | 1964 | /* |
67a1b037 | 1965 | * Maximum amount of data that can be put into single log block. |
b8738257 MA |
1966 | */ |
1967 | uint64_t | |
67a1b037 | 1968 | zil_max_log_data(zilog_t *zilog, size_t hdrsize) |
b8738257 | 1969 | { |
67a1b037 | 1970 | return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize); |
b8738257 MA |
1971 | } |
1972 | ||
1973 | /* | |
1974 | * Maximum amount of log space we agree to waste to reduce number of | |
66b81b34 | 1975 | * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~6%). |
b8738257 MA |
1976 | */ |
1977 | static inline uint64_t | |
1978 | zil_max_waste_space(zilog_t *zilog) | |
1979 | { | |
66b81b34 | 1980 | return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 16); |
b8738257 MA |
1981 | } |
1982 | ||
1983 | /* | |
1984 | * Maximum amount of write data for WR_COPIED. For correctness, consumers | |
1985 | * must fall back to WR_NEED_COPY if we can't fit the entire record into one | |
1986 | * maximum sized log block, because each WR_COPIED record must fit in a | |
66b81b34 AM |
1987 | * single log block. Below that it is a tradeoff of additional memory copy |
1988 | * and possibly worse log space efficiency vs additional range lock/unlock. | |
b8738257 | 1989 | */ |
66b81b34 AM |
1990 | static uint_t zil_maxcopied = 7680; |
1991 | ||
b8738257 MA |
1992 | uint64_t |
1993 | zil_max_copied_data(zilog_t *zilog) | |
1994 | { | |
66b81b34 AM |
1995 | uint64_t max_data = zil_max_log_data(zilog, sizeof (lr_write_t)); |
1996 | return (MIN(max_data, zil_maxcopied)); | |
b8738257 MA |
1997 | } |
1998 | ||
f63811f0 AM |
1999 | /* |
2000 | * Estimate space needed in the lwb for the itx. Allocate more lwbs or | |
2001 | * split the itx as needed, but don't touch the actual transaction data. | |
2002 | * Has to be called under zl_issuer_lock to call zil_lwb_write_close() | |
2003 | * to chain more lwbs. | |
2004 | */ | |
34dc7c2f | 2005 | static lwb_t * |
f63811f0 | 2006 | zil_lwb_assign(zilog_t *zilog, lwb_t *lwb, itx_t *itx, list_t *ilwbs) |
34dc7c2f | 2007 | { |
f63811f0 AM |
2008 | itx_t *citx; |
2009 | lr_t *lr, *clr; | |
2010 | lr_write_t *lrw; | |
2011 | uint64_t dlen, dnow, lwb_sp, reclen, max_log_data; | |
34dc7c2f | 2012 | |
1b2b0aca | 2013 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca PS |
2014 | ASSERT3P(lwb, !=, NULL); |
2015 | ASSERT3P(lwb->lwb_buf, !=, NULL); | |
2016 | ||
2017 | zil_lwb_write_open(zilog, lwb); | |
428870ff | 2018 | |
f63811f0 AM |
2019 | lr = &itx->itx_lr; |
2020 | lrw = (lr_write_t *)lr; | |
1ce23dca PS |
2021 | |
2022 | /* | |
2023 | * A commit itx doesn't represent any on-disk state; instead | |
2024 | * it's simply used as a place holder on the commit list, and | |
2025 | * provides a mechanism for attaching a "commit waiter" onto the | |
2026 | * correct lwb (such that the waiter can be signalled upon | |
2027 | * completion of that lwb). Thus, we don't process this itx's | |
2028 | * log record if it's a commit itx (these itx's don't have log | |
2029 | * records), and instead link the itx's waiter onto the lwb's | |
2030 | * list of waiters. | |
2031 | * | |
2032 | * For more details, see the comment above zil_commit(). | |
2033 | */ | |
f63811f0 | 2034 | if (lr->lrc_txtype == TX_COMMIT) { |
1ce23dca | 2035 | zil_commit_waiter_link_lwb(itx->itx_private, lwb); |
f63811f0 | 2036 | list_insert_tail(&lwb->lwb_itxs, itx); |
1ce23dca PS |
2037 | return (lwb); |
2038 | } | |
34dc7c2f | 2039 | |
2a27fd41 | 2040 | reclen = lr->lrc_reclen; |
f63811f0 | 2041 | if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { |
2a27fd41 | 2042 | ASSERT3U(reclen, ==, sizeof (lr_write_t)); |
34dc7c2f | 2043 | dlen = P2ROUNDUP_TYPED( |
428870ff | 2044 | lrw->lr_length, sizeof (uint64_t), uint64_t); |
1b7c1e5c | 2045 | } else { |
2a27fd41 | 2046 | ASSERT3U(reclen, >=, sizeof (lr_t)); |
f63811f0 | 2047 | dlen = 0; |
1b7c1e5c | 2048 | } |
2a27fd41 | 2049 | ASSERT3U(reclen, <=, zil_max_log_data(zilog, 0)); |
34dc7c2f BB |
2050 | zilog->zl_cur_used += (reclen + dlen); |
2051 | ||
1b7c1e5c | 2052 | cont: |
34dc7c2f BB |
2053 | /* |
2054 | * If this record won't fit in the current log block, start a new one. | |
1b7c1e5c | 2055 | * For WR_NEED_COPY optimize layout for minimal number of chunks. |
34dc7c2f | 2056 | */ |
eda3fcd5 | 2057 | lwb_sp = lwb->lwb_nmax - lwb->lwb_nused; |
67a1b037 | 2058 | max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t)); |
1b7c1e5c | 2059 | if (reclen > lwb_sp || (reclen + dlen > lwb_sp && |
b8738257 MA |
2060 | lwb_sp < zil_max_waste_space(zilog) && |
2061 | (dlen % max_log_data == 0 || | |
2062 | lwb_sp < reclen + dlen % max_log_data))) { | |
eda3fcd5 AM |
2063 | list_insert_tail(ilwbs, lwb); |
2064 | lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_OPENED); | |
34dc7c2f BB |
2065 | if (lwb == NULL) |
2066 | return (NULL); | |
eda3fcd5 | 2067 | lwb_sp = lwb->lwb_nmax - lwb->lwb_nused; |
34dc7c2f BB |
2068 | } |
2069 | ||
2a27fd41 AM |
2070 | /* |
2071 | * There must be enough space in the log block to hold reclen. | |
2072 | * For WR_COPIED, we need to fit the whole record in one block, | |
2073 | * and reclen is the write record header size + the data size. | |
2074 | * For WR_NEED_COPY, we can create multiple records, splitting | |
2075 | * the data into multiple blocks, so we only need to fit one | |
2076 | * word of data per block; in this case reclen is just the header | |
2077 | * size (no data). | |
2078 | */ | |
2079 | ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp); | |
2080 | ||
1b7c1e5c | 2081 | dnow = MIN(dlen, lwb_sp - reclen); |
f63811f0 AM |
2082 | if (dlen > dnow) { |
2083 | ASSERT3U(lr->lrc_txtype, ==, TX_WRITE); | |
2084 | ASSERT3U(itx->itx_wr_state, ==, WR_NEED_COPY); | |
2085 | citx = zil_itx_clone(itx); | |
2086 | clr = &citx->itx_lr; | |
2087 | lr_write_t *clrw = (lr_write_t *)clr; | |
2088 | clrw->lr_length = dnow; | |
2089 | lrw->lr_offset += dnow; | |
2090 | lrw->lr_length -= dnow; | |
2091 | } else { | |
2092 | citx = itx; | |
2093 | clr = lr; | |
2094 | } | |
2095 | ||
2096 | /* | |
2097 | * We're actually making an entry, so update lrc_seq to be the | |
2098 | * log record sequence number. Note that this is generally not | |
2099 | * equal to the itx sequence number because not all transactions | |
2100 | * are synchronous, and sometimes spa_sync() gets there first. | |
2101 | */ | |
2102 | clr->lrc_seq = ++zilog->zl_lr_seq; | |
2103 | ||
2104 | lwb->lwb_nused += reclen + dnow; | |
eda3fcd5 | 2105 | ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax); |
f63811f0 AM |
2106 | ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); |
2107 | ||
2108 | zil_lwb_add_txg(lwb, lr->lrc_txg); | |
2109 | list_insert_tail(&lwb->lwb_itxs, citx); | |
2110 | ||
2111 | dlen -= dnow; | |
2112 | if (dlen > 0) { | |
2113 | zilog->zl_cur_used += reclen; | |
2114 | goto cont; | |
2115 | } | |
2116 | ||
eda3fcd5 AM |
2117 | if (lr->lrc_txtype == TX_WRITE && |
2118 | lr->lrc_txg > spa_freeze_txg(zilog->zl_spa)) | |
2119 | txg_wait_synced(zilog->zl_dmu_pool, lr->lrc_txg); | |
f63811f0 AM |
2120 | |
2121 | return (lwb); | |
2122 | } | |
2123 | ||
2124 | /* | |
2125 | * Fill the actual transaction data into the lwb, following zil_lwb_assign(). | |
2126 | * Does not require locking. | |
2127 | */ | |
2128 | static void | |
2129 | zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx) | |
2130 | { | |
2131 | lr_t *lr, *lrb; | |
2132 | lr_write_t *lrw, *lrwb; | |
2133 | char *lr_buf; | |
2134 | uint64_t dlen, reclen; | |
2135 | ||
2136 | lr = &itx->itx_lr; | |
2137 | lrw = (lr_write_t *)lr; | |
2138 | ||
2139 | if (lr->lrc_txtype == TX_COMMIT) | |
2140 | return; | |
2141 | ||
2142 | if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { | |
2143 | dlen = P2ROUNDUP_TYPED( | |
2144 | lrw->lr_length, sizeof (uint64_t), uint64_t); | |
2145 | } else { | |
2146 | dlen = 0; | |
2147 | } | |
2148 | reclen = lr->lrc_reclen; | |
2149 | ASSERT3U(reclen + dlen, <=, lwb->lwb_nused - lwb->lwb_nfilled); | |
2150 | ||
2151 | lr_buf = lwb->lwb_buf + lwb->lwb_nfilled; | |
2152 | memcpy(lr_buf, lr, reclen); | |
2153 | lrb = (lr_t *)lr_buf; /* Like lr, but inside lwb. */ | |
2154 | lrwb = (lr_write_t *)lrb; /* Like lrw, but inside lwb. */ | |
34dc7c2f | 2155 | |
fb087146 | 2156 | ZIL_STAT_BUMP(zilog, zil_itx_count); |
b6ad9671 | 2157 | |
34dc7c2f BB |
2158 | /* |
2159 | * If it's a write, fetch the data or get its blkptr as appropriate. | |
2160 | */ | |
f63811f0 | 2161 | if (lr->lrc_txtype == TX_WRITE) { |
b6ad9671 | 2162 | if (itx->itx_wr_state == WR_COPIED) { |
fb087146 AH |
2163 | ZIL_STAT_BUMP(zilog, zil_itx_copied_count); |
2164 | ZIL_STAT_INCR(zilog, zil_itx_copied_bytes, | |
2165 | lrw->lr_length); | |
b6ad9671 | 2166 | } else { |
34dc7c2f BB |
2167 | char *dbuf; |
2168 | int error; | |
2169 | ||
1b7c1e5c | 2170 | if (itx->itx_wr_state == WR_NEED_COPY) { |
428870ff | 2171 | dbuf = lr_buf + reclen; |
f63811f0 | 2172 | lrb->lrc_reclen += dlen; |
fb087146 AH |
2173 | ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count); |
2174 | ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes, | |
f63811f0 | 2175 | dlen); |
34dc7c2f | 2176 | } else { |
1ce23dca | 2177 | ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT); |
34dc7c2f | 2178 | dbuf = NULL; |
fb087146 AH |
2179 | ZIL_STAT_BUMP(zilog, zil_itx_indirect_count); |
2180 | ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes, | |
d1d7e268 | 2181 | lrw->lr_length); |
eda3fcd5 | 2182 | if (lwb->lwb_child_zio == NULL) { |
3afdc97d AM |
2183 | lwb->lwb_child_zio = zio_null(NULL, |
2184 | zilog->zl_spa, NULL, NULL, NULL, | |
eda3fcd5 AM |
2185 | ZIO_FLAG_CANFAIL); |
2186 | } | |
34dc7c2f | 2187 | } |
1ce23dca PS |
2188 | |
2189 | /* | |
eda3fcd5 AM |
2190 | * The "lwb_child_zio" we pass in will become a child of |
2191 | * "lwb_write_zio", when one is created, so one will be | |
2192 | * a parent of any zio's created by the "zl_get_data". | |
2193 | * This way "lwb_write_zio" will first wait for children | |
2194 | * block pointers before own writing, and then for their | |
2195 | * writing completion before the vdev cache flushing. | |
1ce23dca PS |
2196 | */ |
2197 | error = zilog->zl_get_data(itx->itx_private, | |
296a4a36 | 2198 | itx->itx_gen, lrwb, dbuf, lwb, |
eda3fcd5 | 2199 | lwb->lwb_child_zio); |
f63811f0 | 2200 | if (dbuf != NULL && error == 0) { |
3a185275 | 2201 | /* Zero any padding bytes in the last block. */ |
f63811f0 AM |
2202 | memset((char *)dbuf + lrwb->lr_length, 0, |
2203 | dlen - lrwb->lr_length); | |
2204 | } | |
1ce23dca | 2205 | |
3a7c3511 RY |
2206 | /* |
2207 | * Typically, the only return values we should see from | |
2208 | * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or | |
2209 | * EALREADY. However, it is also possible to see other | |
2210 | * error values such as ENOSPC or EINVAL from | |
2211 | * dmu_read() -> dnode_hold() -> dnode_hold_impl() or | |
2212 | * ENXIO as well as a multitude of others from the | |
2213 | * block layer through dmu_buf_hold() -> dbuf_read() | |
2214 | * -> zio_wait(), as well as through dmu_read() -> | |
2215 | * dnode_hold() -> dnode_hold_impl() -> dbuf_read() -> | |
2216 | * zio_wait(). When these errors happen, we can assume | |
2217 | * that neither an immediate write nor an indirect | |
2218 | * write occurred, so we need to fall back to | |
2219 | * txg_wait_synced(). This is unusual, so we print to | |
2220 | * dmesg whenever one of these errors occurs. | |
2221 | */ | |
2222 | switch (error) { | |
2223 | case 0: | |
2224 | break; | |
2225 | default: | |
2226 | cmn_err(CE_WARN, "zil_lwb_commit() received " | |
2227 | "unexpected error %d from ->zl_get_data()" | |
2228 | ". Falling back to txg_wait_synced().", | |
2229 | error); | |
2230 | zfs_fallthrough; | |
2231 | case EIO: | |
eda3fcd5 AM |
2232 | txg_wait_synced(zilog->zl_dmu_pool, |
2233 | lr->lrc_txg); | |
3a7c3511 RY |
2234 | zfs_fallthrough; |
2235 | case ENOENT: | |
2236 | zfs_fallthrough; | |
2237 | case EEXIST: | |
2238 | zfs_fallthrough; | |
2239 | case EALREADY: | |
f63811f0 | 2240 | return; |
34dc7c2f BB |
2241 | } |
2242 | } | |
2243 | } | |
2244 | ||
f63811f0 AM |
2245 | lwb->lwb_nfilled += reclen + dlen; |
2246 | ASSERT3S(lwb->lwb_nfilled, <=, lwb->lwb_nused); | |
2247 | ASSERT0(P2PHASE(lwb->lwb_nfilled, sizeof (uint64_t))); | |
34dc7c2f BB |
2248 | } |
2249 | ||
2250 | itx_t * | |
58714c28 | 2251 | zil_itx_create(uint64_t txtype, size_t olrsize) |
34dc7c2f | 2252 | { |
58714c28 | 2253 | size_t itxsize, lrsize; |
34dc7c2f BB |
2254 | itx_t *itx; |
2255 | ||
2a27fd41 | 2256 | ASSERT3U(olrsize, >=, sizeof (lr_t)); |
58714c28 | 2257 | lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t); |
2a27fd41 | 2258 | ASSERT3U(lrsize, >=, olrsize); |
72841b9f | 2259 | itxsize = offsetof(itx_t, itx_lr) + lrsize; |
34dc7c2f | 2260 | |
72841b9f | 2261 | itx = zio_data_buf_alloc(itxsize); |
34dc7c2f BB |
2262 | itx->itx_lr.lrc_txtype = txtype; |
2263 | itx->itx_lr.lrc_reclen = lrsize; | |
34dc7c2f | 2264 | itx->itx_lr.lrc_seq = 0; /* defensive */ |
861166b0 | 2265 | memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize); |
572e2857 | 2266 | itx->itx_sync = B_TRUE; /* default is synchronous */ |
119a394a ED |
2267 | itx->itx_callback = NULL; |
2268 | itx->itx_callback_data = NULL; | |
72841b9f | 2269 | itx->itx_size = itxsize; |
34dc7c2f BB |
2270 | |
2271 | return (itx); | |
2272 | } | |
2273 | ||
f63811f0 AM |
2274 | static itx_t * |
2275 | zil_itx_clone(itx_t *oitx) | |
2276 | { | |
2a27fd41 AM |
2277 | ASSERT3U(oitx->itx_size, >=, sizeof (itx_t)); |
2278 | ASSERT3U(oitx->itx_size, ==, | |
2279 | offsetof(itx_t, itx_lr) + oitx->itx_lr.lrc_reclen); | |
2280 | ||
f63811f0 AM |
2281 | itx_t *itx = zio_data_buf_alloc(oitx->itx_size); |
2282 | memcpy(itx, oitx, oitx->itx_size); | |
2283 | itx->itx_callback = NULL; | |
2284 | itx->itx_callback_data = NULL; | |
2285 | return (itx); | |
2286 | } | |
2287 | ||
428870ff BB |
2288 | void |
2289 | zil_itx_destroy(itx_t *itx) | |
2290 | { | |
2a27fd41 AM |
2291 | ASSERT3U(itx->itx_size, >=, sizeof (itx_t)); |
2292 | ASSERT3U(itx->itx_lr.lrc_reclen, ==, | |
2293 | itx->itx_size - offsetof(itx_t, itx_lr)); | |
1ce23dca PS |
2294 | IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL); |
2295 | IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); | |
2296 | ||
2297 | if (itx->itx_callback != NULL) | |
2298 | itx->itx_callback(itx->itx_callback_data); | |
2299 | ||
72841b9f | 2300 | zio_data_buf_free(itx, itx->itx_size); |
428870ff BB |
2301 | } |
2302 | ||
572e2857 BB |
2303 | /* |
2304 | * Free up the sync and async itxs. The itxs_t has already been detached | |
2305 | * so no locks are needed. | |
2306 | */ | |
2307 | static void | |
23c13c7e | 2308 | zil_itxg_clean(void *arg) |
34dc7c2f | 2309 | { |
572e2857 BB |
2310 | itx_t *itx; |
2311 | list_t *list; | |
2312 | avl_tree_t *t; | |
2313 | void *cookie; | |
23c13c7e | 2314 | itxs_t *itxs = arg; |
572e2857 BB |
2315 | itx_async_node_t *ian; |
2316 | ||
2317 | list = &itxs->i_sync_list; | |
895e0313 | 2318 | while ((itx = list_remove_head(list)) != NULL) { |
1ce23dca PS |
2319 | /* |
2320 | * In the general case, commit itxs will not be found | |
2321 | * here, as they'll be committed to an lwb via | |
f63811f0 | 2322 | * zil_lwb_assign(), and free'd in that function. Having |
1ce23dca PS |
2323 | * said that, it is still possible for commit itxs to be |
2324 | * found here, due to the following race: | |
2325 | * | |
2326 | * - a thread calls zil_commit() which assigns the | |
2327 | * commit itx to a per-txg i_sync_list | |
2328 | * - zil_itxg_clean() is called (e.g. via spa_sync()) | |
2329 | * while the waiter is still on the i_sync_list | |
2330 | * | |
2331 | * There's nothing to prevent syncing the txg while the | |
2332 | * waiter is on the i_sync_list. This normally doesn't | |
2333 | * happen because spa_sync() is slower than zil_commit(), | |
2334 | * but if zil_commit() calls txg_wait_synced() (e.g. | |
2335 | * because zil_create() or zil_commit_writer_stall() is | |
2336 | * called) we will hit this case. | |
2337 | */ | |
2338 | if (itx->itx_lr.lrc_txtype == TX_COMMIT) | |
2339 | zil_commit_waiter_skip(itx->itx_private); | |
2340 | ||
19ea3d25 | 2341 | zil_itx_destroy(itx); |
572e2857 | 2342 | } |
34dc7c2f | 2343 | |
572e2857 BB |
2344 | cookie = NULL; |
2345 | t = &itxs->i_async_tree; | |
2346 | while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { | |
2347 | list = &ian->ia_list; | |
895e0313 | 2348 | while ((itx = list_remove_head(list)) != NULL) { |
1ce23dca PS |
2349 | /* commit itxs should never be on the async lists. */ |
2350 | ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); | |
19ea3d25 | 2351 | zil_itx_destroy(itx); |
572e2857 BB |
2352 | } |
2353 | list_destroy(list); | |
2354 | kmem_free(ian, sizeof (itx_async_node_t)); | |
2355 | } | |
2356 | avl_destroy(t); | |
34dc7c2f | 2357 | |
572e2857 BB |
2358 | kmem_free(itxs, sizeof (itxs_t)); |
2359 | } | |
34dc7c2f | 2360 | |
572e2857 BB |
2361 | static int |
2362 | zil_aitx_compare(const void *x1, const void *x2) | |
2363 | { | |
2364 | const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; | |
2365 | const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; | |
2366 | ||
ca577779 | 2367 | return (TREE_CMP(o1, o2)); |
34dc7c2f BB |
2368 | } |
2369 | ||
2370 | /* | |
572e2857 | 2371 | * Remove all async itx with the given oid. |
34dc7c2f | 2372 | */ |
8e556c5e | 2373 | void |
572e2857 | 2374 | zil_remove_async(zilog_t *zilog, uint64_t oid) |
34dc7c2f | 2375 | { |
572e2857 | 2376 | uint64_t otxg, txg; |
2a27fd41 | 2377 | itx_async_node_t *ian, ian_search; |
572e2857 BB |
2378 | avl_tree_t *t; |
2379 | avl_index_t where; | |
34dc7c2f BB |
2380 | list_t clean_list; |
2381 | itx_t *itx; | |
2382 | ||
572e2857 | 2383 | ASSERT(oid != 0); |
34dc7c2f BB |
2384 | list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); |
2385 | ||
572e2857 BB |
2386 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ |
2387 | otxg = ZILTEST_TXG; | |
2388 | else | |
2389 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
34dc7c2f | 2390 | |
572e2857 BB |
2391 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { |
2392 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
2393 | ||
2394 | mutex_enter(&itxg->itxg_lock); | |
2395 | if (itxg->itxg_txg != txg) { | |
2396 | mutex_exit(&itxg->itxg_lock); | |
2397 | continue; | |
2398 | } | |
34dc7c2f | 2399 | |
572e2857 BB |
2400 | /* |
2401 | * Locate the object node and append its list. | |
2402 | */ | |
2403 | t = &itxg->itxg_itxs->i_async_tree; | |
2a27fd41 AM |
2404 | ian_search.ia_foid = oid; |
2405 | ian = avl_find(t, &ian_search, &where); | |
572e2857 BB |
2406 | if (ian != NULL) |
2407 | list_move_tail(&clean_list, &ian->ia_list); | |
2408 | mutex_exit(&itxg->itxg_lock); | |
2409 | } | |
895e0313 | 2410 | while ((itx = list_remove_head(&clean_list)) != NULL) { |
1ce23dca PS |
2411 | /* commit itxs should never be on the async lists. */ |
2412 | ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); | |
19ea3d25 | 2413 | zil_itx_destroy(itx); |
34dc7c2f BB |
2414 | } |
2415 | list_destroy(&clean_list); | |
2416 | } | |
2417 | ||
572e2857 BB |
2418 | void |
2419 | zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) | |
2420 | { | |
2421 | uint64_t txg; | |
2422 | itxg_t *itxg; | |
2423 | itxs_t *itxs, *clean = NULL; | |
2424 | ||
572e2857 BB |
2425 | /* |
2426 | * Ensure the data of a renamed file is committed before the rename. | |
2427 | */ | |
2428 | if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) | |
2429 | zil_async_to_sync(zilog, itx->itx_oid); | |
2430 | ||
29809a6c | 2431 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) |
572e2857 BB |
2432 | txg = ZILTEST_TXG; |
2433 | else | |
2434 | txg = dmu_tx_get_txg(tx); | |
2435 | ||
2436 | itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
2437 | mutex_enter(&itxg->itxg_lock); | |
2438 | itxs = itxg->itxg_itxs; | |
2439 | if (itxg->itxg_txg != txg) { | |
2440 | if (itxs != NULL) { | |
2441 | /* | |
2442 | * The zil_clean callback hasn't got around to cleaning | |
2443 | * this itxg. Save the itxs for release below. | |
2444 | * This should be rare. | |
2445 | */ | |
55922e73 | 2446 | zfs_dbgmsg("zil_itx_assign: missed itx cleanup for " |
8e739b2c | 2447 | "txg %llu", (u_longlong_t)itxg->itxg_txg); |
572e2857 BB |
2448 | clean = itxg->itxg_itxs; |
2449 | } | |
572e2857 | 2450 | itxg->itxg_txg = txg; |
d1d7e268 | 2451 | itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), |
79c76d5b | 2452 | KM_SLEEP); |
572e2857 BB |
2453 | |
2454 | list_create(&itxs->i_sync_list, sizeof (itx_t), | |
2455 | offsetof(itx_t, itx_node)); | |
2456 | avl_create(&itxs->i_async_tree, zil_aitx_compare, | |
2457 | sizeof (itx_async_node_t), | |
2458 | offsetof(itx_async_node_t, ia_node)); | |
2459 | } | |
2460 | if (itx->itx_sync) { | |
2461 | list_insert_tail(&itxs->i_sync_list, itx); | |
572e2857 BB |
2462 | } else { |
2463 | avl_tree_t *t = &itxs->i_async_tree; | |
50c957f7 NB |
2464 | uint64_t foid = |
2465 | LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid); | |
572e2857 BB |
2466 | itx_async_node_t *ian; |
2467 | avl_index_t where; | |
2468 | ||
2469 | ian = avl_find(t, &foid, &where); | |
2470 | if (ian == NULL) { | |
d1d7e268 | 2471 | ian = kmem_alloc(sizeof (itx_async_node_t), |
79c76d5b | 2472 | KM_SLEEP); |
572e2857 BB |
2473 | list_create(&ian->ia_list, sizeof (itx_t), |
2474 | offsetof(itx_t, itx_node)); | |
2475 | ian->ia_foid = foid; | |
2476 | avl_insert(t, ian, where); | |
2477 | } | |
2478 | list_insert_tail(&ian->ia_list, itx); | |
2479 | } | |
2480 | ||
2481 | itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); | |
1ce23dca PS |
2482 | |
2483 | /* | |
2484 | * We don't want to dirty the ZIL using ZILTEST_TXG, because | |
2485 | * zil_clean() will never be called using ZILTEST_TXG. Thus, we | |
2486 | * need to be careful to always dirty the ZIL using the "real" | |
2487 | * TXG (not itxg_txg) even when the SPA is frozen. | |
2488 | */ | |
2489 | zilog_dirty(zilog, dmu_tx_get_txg(tx)); | |
572e2857 BB |
2490 | mutex_exit(&itxg->itxg_lock); |
2491 | ||
2492 | /* Release the old itxs now we've dropped the lock */ | |
2493 | if (clean != NULL) | |
2494 | zil_itxg_clean(clean); | |
2495 | } | |
2496 | ||
34dc7c2f BB |
2497 | /* |
2498 | * If there are any in-memory intent log transactions which have now been | |
29809a6c | 2499 | * synced then start up a taskq to free them. We should only do this after we |
e1cfd73f | 2500 | * have written out the uberblocks (i.e. txg has been committed) so that |
29809a6c MA |
2501 | * don't inadvertently clean out in-memory log records that would be required |
2502 | * by zil_commit(). | |
34dc7c2f BB |
2503 | */ |
2504 | void | |
572e2857 | 2505 | zil_clean(zilog_t *zilog, uint64_t synced_txg) |
34dc7c2f | 2506 | { |
572e2857 BB |
2507 | itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; |
2508 | itxs_t *clean_me; | |
34dc7c2f | 2509 | |
1ce23dca PS |
2510 | ASSERT3U(synced_txg, <, ZILTEST_TXG); |
2511 | ||
572e2857 BB |
2512 | mutex_enter(&itxg->itxg_lock); |
2513 | if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { | |
2514 | mutex_exit(&itxg->itxg_lock); | |
2515 | return; | |
2516 | } | |
2517 | ASSERT3U(itxg->itxg_txg, <=, synced_txg); | |
a032ac4b | 2518 | ASSERT3U(itxg->itxg_txg, !=, 0); |
572e2857 BB |
2519 | clean_me = itxg->itxg_itxs; |
2520 | itxg->itxg_itxs = NULL; | |
2521 | itxg->itxg_txg = 0; | |
2522 | mutex_exit(&itxg->itxg_lock); | |
2523 | /* | |
2524 | * Preferably start a task queue to free up the old itxs but | |
2525 | * if taskq_dispatch can't allocate resources to do that then | |
2526 | * free it in-line. This should be rare. Note, using TQ_SLEEP | |
2527 | * created a bad performance problem. | |
2528 | */ | |
a032ac4b BB |
2529 | ASSERT3P(zilog->zl_dmu_pool, !=, NULL); |
2530 | ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL); | |
2531 | taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq, | |
23c13c7e | 2532 | zil_itxg_clean, clean_me, TQ_NOSLEEP); |
a032ac4b | 2533 | if (id == TASKQID_INVALID) |
572e2857 BB |
2534 | zil_itxg_clean(clean_me); |
2535 | } | |
2536 | ||
2537 | /* | |
1ce23dca PS |
2538 | * This function will traverse the queue of itxs that need to be |
2539 | * committed, and move them onto the ZIL's zl_itx_commit_list. | |
572e2857 | 2540 | */ |
233425a1 | 2541 | static uint64_t |
572e2857 BB |
2542 | zil_get_commit_list(zilog_t *zilog) |
2543 | { | |
233425a1 | 2544 | uint64_t otxg, txg, wtxg = 0; |
572e2857 | 2545 | list_t *commit_list = &zilog->zl_itx_commit_list; |
572e2857 | 2546 | |
1b2b0aca | 2547 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca | 2548 | |
572e2857 BB |
2549 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ |
2550 | otxg = ZILTEST_TXG; | |
2551 | else | |
2552 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
2553 | ||
55922e73 GW |
2554 | /* |
2555 | * This is inherently racy, since there is nothing to prevent | |
2556 | * the last synced txg from changing. That's okay since we'll | |
2557 | * only commit things in the future. | |
2558 | */ | |
572e2857 BB |
2559 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { |
2560 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
2561 | ||
2562 | mutex_enter(&itxg->itxg_lock); | |
2563 | if (itxg->itxg_txg != txg) { | |
2564 | mutex_exit(&itxg->itxg_lock); | |
2565 | continue; | |
2566 | } | |
2567 | ||
55922e73 GW |
2568 | /* |
2569 | * If we're adding itx records to the zl_itx_commit_list, | |
2570 | * then the zil better be dirty in this "txg". We can assert | |
2571 | * that here since we're holding the itxg_lock which will | |
2572 | * prevent spa_sync from cleaning it. Once we add the itxs | |
2573 | * to the zl_itx_commit_list we must commit it to disk even | |
2574 | * if it's unnecessary (i.e. the txg was synced). | |
2575 | */ | |
2576 | ASSERT(zilog_is_dirty_in_txg(zilog, txg) || | |
2577 | spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); | |
233425a1 AM |
2578 | list_t *sync_list = &itxg->itxg_itxs->i_sync_list; |
2579 | if (unlikely(zilog->zl_suspend > 0)) { | |
2580 | /* | |
2581 | * ZIL was just suspended, but we lost the race. | |
2582 | * Allow all earlier itxs to be committed, but ask | |
2583 | * caller to do txg_wait_synced(txg) for any new. | |
2584 | */ | |
2585 | if (!list_is_empty(sync_list)) | |
2586 | wtxg = MAX(wtxg, txg); | |
2587 | } else { | |
2588 | list_move_tail(commit_list, sync_list); | |
2589 | } | |
572e2857 BB |
2590 | |
2591 | mutex_exit(&itxg->itxg_lock); | |
2592 | } | |
233425a1 | 2593 | return (wtxg); |
572e2857 BB |
2594 | } |
2595 | ||
2596 | /* | |
2597 | * Move the async itxs for a specified object to commit into sync lists. | |
2598 | */ | |
eedb3a62 | 2599 | void |
572e2857 BB |
2600 | zil_async_to_sync(zilog_t *zilog, uint64_t foid) |
2601 | { | |
2602 | uint64_t otxg, txg; | |
2a27fd41 | 2603 | itx_async_node_t *ian, ian_search; |
572e2857 BB |
2604 | avl_tree_t *t; |
2605 | avl_index_t where; | |
2606 | ||
2607 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ | |
2608 | otxg = ZILTEST_TXG; | |
2609 | else | |
2610 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
2611 | ||
55922e73 GW |
2612 | /* |
2613 | * This is inherently racy, since there is nothing to prevent | |
2614 | * the last synced txg from changing. | |
2615 | */ | |
572e2857 BB |
2616 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { |
2617 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
2618 | ||
2619 | mutex_enter(&itxg->itxg_lock); | |
2620 | if (itxg->itxg_txg != txg) { | |
2621 | mutex_exit(&itxg->itxg_lock); | |
2622 | continue; | |
2623 | } | |
2624 | ||
2625 | /* | |
2626 | * If a foid is specified then find that node and append its | |
2627 | * list. Otherwise walk the tree appending all the lists | |
2628 | * to the sync list. We add to the end rather than the | |
2629 | * beginning to ensure the create has happened. | |
2630 | */ | |
2631 | t = &itxg->itxg_itxs->i_async_tree; | |
2632 | if (foid != 0) { | |
2a27fd41 AM |
2633 | ian_search.ia_foid = foid; |
2634 | ian = avl_find(t, &ian_search, &where); | |
572e2857 BB |
2635 | if (ian != NULL) { |
2636 | list_move_tail(&itxg->itxg_itxs->i_sync_list, | |
2637 | &ian->ia_list); | |
2638 | } | |
2639 | } else { | |
2640 | void *cookie = NULL; | |
2641 | ||
2642 | while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { | |
2643 | list_move_tail(&itxg->itxg_itxs->i_sync_list, | |
2644 | &ian->ia_list); | |
2645 | list_destroy(&ian->ia_list); | |
2646 | kmem_free(ian, sizeof (itx_async_node_t)); | |
2647 | } | |
2648 | } | |
2649 | mutex_exit(&itxg->itxg_lock); | |
34dc7c2f | 2650 | } |
34dc7c2f BB |
2651 | } |
2652 | ||
1ce23dca PS |
2653 | /* |
2654 | * This function will prune commit itxs that are at the head of the | |
2655 | * commit list (it won't prune past the first non-commit itx), and | |
2656 | * either: a) attach them to the last lwb that's still pending | |
2657 | * completion, or b) skip them altogether. | |
2658 | * | |
2659 | * This is used as a performance optimization to prevent commit itxs | |
2660 | * from generating new lwbs when it's unnecessary to do so. | |
2661 | */ | |
b128c09f | 2662 | static void |
1ce23dca | 2663 | zil_prune_commit_list(zilog_t *zilog) |
34dc7c2f | 2664 | { |
572e2857 | 2665 | itx_t *itx; |
34dc7c2f | 2666 | |
1b2b0aca | 2667 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
572e2857 | 2668 | |
1ce23dca PS |
2669 | while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { |
2670 | lr_t *lrc = &itx->itx_lr; | |
2671 | if (lrc->lrc_txtype != TX_COMMIT) | |
2672 | break; | |
572e2857 | 2673 | |
1ce23dca PS |
2674 | mutex_enter(&zilog->zl_lock); |
2675 | ||
2676 | lwb_t *last_lwb = zilog->zl_last_lwb_opened; | |
900d09b2 PS |
2677 | if (last_lwb == NULL || |
2678 | last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) { | |
1ce23dca PS |
2679 | /* |
2680 | * All of the itxs this waiter was waiting on | |
2681 | * must have already completed (or there were | |
2682 | * never any itx's for it to wait on), so it's | |
2683 | * safe to skip this waiter and mark it done. | |
2684 | */ | |
2685 | zil_commit_waiter_skip(itx->itx_private); | |
2686 | } else { | |
2687 | zil_commit_waiter_link_lwb(itx->itx_private, last_lwb); | |
1ce23dca PS |
2688 | } |
2689 | ||
2690 | mutex_exit(&zilog->zl_lock); | |
2691 | ||
2692 | list_remove(&zilog->zl_itx_commit_list, itx); | |
2693 | zil_itx_destroy(itx); | |
2694 | } | |
2695 | ||
2696 | IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); | |
2697 | } | |
2698 | ||
2699 | static void | |
2700 | zil_commit_writer_stall(zilog_t *zilog) | |
2701 | { | |
2702 | /* | |
2703 | * When zio_alloc_zil() fails to allocate the next lwb block on | |
2704 | * disk, we must call txg_wait_synced() to ensure all of the | |
2705 | * lwbs in the zilog's zl_lwb_list are synced and then freed (in | |
2706 | * zil_sync()), such that any subsequent ZIL writer (i.e. a call | |
2707 | * to zil_process_commit_list()) will have to call zil_create(), | |
2708 | * and start a new ZIL chain. | |
2709 | * | |
2710 | * Since zil_alloc_zil() failed, the lwb that was previously | |
2711 | * issued does not have a pointer to the "next" lwb on disk. | |
2712 | * Thus, if another ZIL writer thread was to allocate the "next" | |
2713 | * on-disk lwb, that block could be leaked in the event of a | |
2714 | * crash (because the previous lwb on-disk would not point to | |
2715 | * it). | |
2716 | * | |
1b2b0aca | 2717 | * We must hold the zilog's zl_issuer_lock while we do this, to |
1ce23dca PS |
2718 | * ensure no new threads enter zil_process_commit_list() until |
2719 | * all lwb's in the zl_lwb_list have been synced and freed | |
2720 | * (which is achieved via the txg_wait_synced() call). | |
2721 | */ | |
1b2b0aca | 2722 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca | 2723 | txg_wait_synced(zilog->zl_dmu_pool, 0); |
895e0313 | 2724 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
1ce23dca PS |
2725 | } |
2726 | ||
252f46be AM |
2727 | static void |
2728 | zil_burst_done(zilog_t *zilog) | |
2729 | { | |
2730 | if (!list_is_empty(&zilog->zl_itx_commit_list) || | |
2731 | zilog->zl_cur_used == 0) | |
2732 | return; | |
2733 | ||
2734 | if (zilog->zl_parallel) | |
2735 | zilog->zl_parallel--; | |
2736 | ||
2737 | zilog->zl_cur_used = 0; | |
2738 | } | |
2739 | ||
1ce23dca PS |
2740 | /* |
2741 | * This function will traverse the commit list, creating new lwbs as | |
2742 | * needed, and committing the itxs from the commit list to these newly | |
2743 | * created lwbs. Additionally, as a new lwb is created, the previous | |
2744 | * lwb will be issued to the zio layer to be written to disk. | |
2745 | */ | |
2746 | static void | |
f63811f0 | 2747 | zil_process_commit_list(zilog_t *zilog, zil_commit_waiter_t *zcw, list_t *ilwbs) |
1ce23dca PS |
2748 | { |
2749 | spa_t *spa = zilog->zl_spa; | |
2750 | list_t nolwb_itxs; | |
2751 | list_t nolwb_waiters; | |
0f740a4f | 2752 | lwb_t *lwb, *plwb; |
1ce23dca PS |
2753 | itx_t *itx; |
2754 | ||
1b2b0aca | 2755 | ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); |
572e2857 BB |
2756 | |
2757 | /* | |
2758 | * Return if there's nothing to commit before we dirty the fs by | |
2759 | * calling zil_create(). | |
2760 | */ | |
895e0313 | 2761 | if (list_is_empty(&zilog->zl_itx_commit_list)) |
572e2857 | 2762 | return; |
34dc7c2f | 2763 | |
1ce23dca PS |
2764 | list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); |
2765 | list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t), | |
2766 | offsetof(zil_commit_waiter_t, zcw_node)); | |
2767 | ||
2768 | lwb = list_tail(&zilog->zl_lwb_list); | |
2769 | if (lwb == NULL) { | |
2770 | lwb = zil_create(zilog); | |
34dc7c2f | 2771 | } else { |
361a7e82 JP |
2772 | /* |
2773 | * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will | |
2774 | * have already been created (zl_lwb_list not empty). | |
2775 | */ | |
2776 | zil_commit_activate_saxattr_feature(zilog); | |
eda3fcd5 AM |
2777 | ASSERT(lwb->lwb_state == LWB_STATE_NEW || |
2778 | lwb->lwb_state == LWB_STATE_OPENED); | |
252f46be AM |
2779 | |
2780 | /* | |
2781 | * If the lwb is still opened, it means the workload is really | |
2782 | * multi-threaded and we won the chance of write aggregation. | |
2783 | * If it is not opened yet, but previous lwb is still not | |
2784 | * flushed, it still means the workload is multi-threaded, but | |
2785 | * there was too much time between the commits to aggregate, so | |
2786 | * we try aggregation next times, but without too much hopes. | |
2787 | */ | |
2788 | if (lwb->lwb_state == LWB_STATE_OPENED) { | |
2789 | zilog->zl_parallel = ZIL_BURSTS; | |
2790 | } else if ((plwb = list_prev(&zilog->zl_lwb_list, lwb)) | |
2791 | != NULL && plwb->lwb_state != LWB_STATE_FLUSH_DONE) { | |
2792 | zilog->zl_parallel = MAX(zilog->zl_parallel, | |
2793 | ZIL_BURSTS / 2); | |
2794 | } | |
34dc7c2f BB |
2795 | } |
2796 | ||
895e0313 | 2797 | while ((itx = list_remove_head(&zilog->zl_itx_commit_list)) != NULL) { |
1ce23dca PS |
2798 | lr_t *lrc = &itx->itx_lr; |
2799 | uint64_t txg = lrc->lrc_txg; | |
2800 | ||
55922e73 | 2801 | ASSERT3U(txg, !=, 0); |
34dc7c2f | 2802 | |
1ce23dca PS |
2803 | if (lrc->lrc_txtype == TX_COMMIT) { |
2804 | DTRACE_PROBE2(zil__process__commit__itx, | |
2805 | zilog_t *, zilog, itx_t *, itx); | |
2806 | } else { | |
2807 | DTRACE_PROBE2(zil__process__normal__itx, | |
2808 | zilog_t *, zilog, itx_t *, itx); | |
2809 | } | |
2810 | ||
1ce23dca PS |
2811 | boolean_t synced = txg <= spa_last_synced_txg(spa); |
2812 | boolean_t frozen = txg > spa_freeze_txg(spa); | |
2813 | ||
2fe61a7e PS |
2814 | /* |
2815 | * If the txg of this itx has already been synced out, then | |
2816 | * we don't need to commit this itx to an lwb. This is | |
2817 | * because the data of this itx will have already been | |
2818 | * written to the main pool. This is inherently racy, and | |
2819 | * it's still ok to commit an itx whose txg has already | |
2820 | * been synced; this will result in a write that's | |
2821 | * unnecessary, but will do no harm. | |
2822 | * | |
2823 | * With that said, we always want to commit TX_COMMIT itxs | |
2824 | * to an lwb, regardless of whether or not that itx's txg | |
2825 | * has been synced out. We do this to ensure any OPENED lwb | |
2826 | * will always have at least one zil_commit_waiter_t linked | |
2827 | * to the lwb. | |
2828 | * | |
2829 | * As a counter-example, if we skipped TX_COMMIT itx's | |
2830 | * whose txg had already been synced, the following | |
2831 | * situation could occur if we happened to be racing with | |
2832 | * spa_sync: | |
2833 | * | |
2834 | * 1. We commit a non-TX_COMMIT itx to an lwb, where the | |
2835 | * itx's txg is 10 and the last synced txg is 9. | |
2836 | * 2. spa_sync finishes syncing out txg 10. | |
2837 | * 3. We move to the next itx in the list, it's a TX_COMMIT | |
2838 | * whose txg is 10, so we skip it rather than committing | |
2839 | * it to the lwb used in (1). | |
2840 | * | |
2841 | * If the itx that is skipped in (3) is the last TX_COMMIT | |
2842 | * itx in the commit list, than it's possible for the lwb | |
2843 | * used in (1) to remain in the OPENED state indefinitely. | |
2844 | * | |
2845 | * To prevent the above scenario from occurring, ensuring | |
2846 | * that once an lwb is OPENED it will transition to ISSUED | |
2847 | * and eventually DONE, we always commit TX_COMMIT itx's to | |
2848 | * an lwb here, even if that itx's txg has already been | |
2849 | * synced. | |
2850 | * | |
2851 | * Finally, if the pool is frozen, we _always_ commit the | |
2852 | * itx. The point of freezing the pool is to prevent data | |
2853 | * from being written to the main pool via spa_sync, and | |
2854 | * instead rely solely on the ZIL to persistently store the | |
2855 | * data; i.e. when the pool is frozen, the last synced txg | |
2856 | * value can't be trusted. | |
2857 | */ | |
2858 | if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) { | |
1ce23dca | 2859 | if (lwb != NULL) { |
f63811f0 AM |
2860 | lwb = zil_lwb_assign(zilog, lwb, itx, ilwbs); |
2861 | if (lwb == NULL) { | |
1ce23dca | 2862 | list_insert_tail(&nolwb_itxs, itx); |
f63811f0 AM |
2863 | } else if ((zcw->zcw_lwb != NULL && |
2864 | zcw->zcw_lwb != lwb) || zcw->zcw_done) { | |
2865 | /* | |
2866 | * Our lwb is done, leave the rest of | |
2867 | * itx list to somebody else who care. | |
2868 | */ | |
252f46be | 2869 | zilog->zl_parallel = ZIL_BURSTS; |
f63811f0 AM |
2870 | break; |
2871 | } | |
1ce23dca PS |
2872 | } else { |
2873 | if (lrc->lrc_txtype == TX_COMMIT) { | |
2874 | zil_commit_waiter_link_nolwb( | |
2875 | itx->itx_private, &nolwb_waiters); | |
2876 | } | |
1ce23dca PS |
2877 | list_insert_tail(&nolwb_itxs, itx); |
2878 | } | |
2879 | } else { | |
2fe61a7e | 2880 | ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT); |
1ce23dca PS |
2881 | zil_itx_destroy(itx); |
2882 | } | |
34dc7c2f | 2883 | } |
34dc7c2f | 2884 | |
1ce23dca PS |
2885 | if (lwb == NULL) { |
2886 | /* | |
2887 | * This indicates zio_alloc_zil() failed to allocate the | |
2888 | * "next" lwb on-disk. When this happens, we must stall | |
2889 | * the ZIL write pipeline; see the comment within | |
2890 | * zil_commit_writer_stall() for more details. | |
2891 | */ | |
f63811f0 AM |
2892 | while ((lwb = list_remove_head(ilwbs)) != NULL) |
2893 | zil_lwb_write_issue(zilog, lwb); | |
1ce23dca | 2894 | zil_commit_writer_stall(zilog); |
34dc7c2f | 2895 | |
1ce23dca PS |
2896 | /* |
2897 | * Additionally, we have to signal and mark the "nolwb" | |
2898 | * waiters as "done" here, since without an lwb, we | |
2899 | * can't do this via zil_lwb_flush_vdevs_done() like | |
2900 | * normal. | |
2901 | */ | |
2902 | zil_commit_waiter_t *zcw; | |
895e0313 | 2903 | while ((zcw = list_remove_head(&nolwb_waiters)) != NULL) |
1ce23dca | 2904 | zil_commit_waiter_skip(zcw); |
1ce23dca PS |
2905 | |
2906 | /* | |
2907 | * And finally, we have to destroy the itx's that | |
2908 | * couldn't be committed to an lwb; this will also call | |
2909 | * the itx's callback if one exists for the itx. | |
2910 | */ | |
895e0313 | 2911 | while ((itx = list_remove_head(&nolwb_itxs)) != NULL) |
1ce23dca | 2912 | zil_itx_destroy(itx); |
1ce23dca PS |
2913 | } else { |
2914 | ASSERT(list_is_empty(&nolwb_waiters)); | |
2915 | ASSERT3P(lwb, !=, NULL); | |
eda3fcd5 AM |
2916 | ASSERT(lwb->lwb_state == LWB_STATE_NEW || |
2917 | lwb->lwb_state == LWB_STATE_OPENED); | |
1ce23dca PS |
2918 | |
2919 | /* | |
2920 | * At this point, the ZIL block pointed at by the "lwb" | |
eda3fcd5 | 2921 | * variable is in "new" or "opened" state. |
1ce23dca | 2922 | * |
eda3fcd5 AM |
2923 | * If it's "new", then no itxs have been committed to it, so |
2924 | * there's no point in issuing its zio (i.e. it's "empty"). | |
1ce23dca | 2925 | * |
eda3fcd5 | 2926 | * If it's "opened", then it contains one or more itxs that |
2fe61a7e PS |
2927 | * eventually need to be committed to stable storage. In |
2928 | * this case we intentionally do not issue the lwb's zio | |
2929 | * to disk yet, and instead rely on one of the following | |
2930 | * two mechanisms for issuing the zio: | |
1ce23dca | 2931 | * |
eda3fcd5 AM |
2932 | * 1. Ideally, there will be more ZIL activity occurring on |
2933 | * the system, such that this function will be immediately | |
2934 | * called again by different thread and this lwb will be | |
2935 | * closed by zil_lwb_assign(). This way, the lwb will be | |
2936 | * "full" when it is issued to disk, and we'll make use of | |
2937 | * the lwb's size the best we can. | |
1ce23dca | 2938 | * |
2fe61a7e | 2939 | * 2. If there isn't sufficient ZIL activity occurring on |
eda3fcd5 AM |
2940 | * the system, zil_commit_waiter() will close it and issue |
2941 | * the zio. If this occurs, the lwb is not guaranteed | |
1ce23dca PS |
2942 | * to be "full" by the time its zio is issued, and means |
2943 | * the size of the lwb was "too large" given the amount | |
2fe61a7e | 2944 | * of ZIL activity occurring on the system at that time. |
1ce23dca PS |
2945 | * |
2946 | * We do this for a couple of reasons: | |
2947 | * | |
2948 | * 1. To try and reduce the number of IOPs needed to | |
2949 | * write the same number of itxs. If an lwb has space | |
2fe61a7e | 2950 | * available in its buffer for more itxs, and more itxs |
1ce23dca PS |
2951 | * will be committed relatively soon (relative to the |
2952 | * latency of performing a write), then it's beneficial | |
2953 | * to wait for these "next" itxs. This way, more itxs | |
2954 | * can be committed to stable storage with fewer writes. | |
2955 | * | |
2956 | * 2. To try and use the largest lwb block size that the | |
2957 | * incoming rate of itxs can support. Again, this is to | |
2958 | * try and pack as many itxs into as few lwbs as | |
2959 | * possible, without significantly impacting the latency | |
2960 | * of each individual itx. | |
2961 | */ | |
252f46be AM |
2962 | if (lwb->lwb_state == LWB_STATE_OPENED && !zilog->zl_parallel) { |
2963 | list_insert_tail(ilwbs, lwb); | |
2964 | lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW); | |
2965 | zil_burst_done(zilog); | |
2966 | if (lwb == NULL) { | |
2967 | while ((lwb = list_remove_head(ilwbs)) != NULL) | |
2968 | zil_lwb_write_issue(zilog, lwb); | |
2969 | zil_commit_writer_stall(zilog); | |
0f740a4f AM |
2970 | } |
2971 | } | |
1ce23dca PS |
2972 | } |
2973 | } | |
2974 | ||
2975 | /* | |
2976 | * This function is responsible for ensuring the passed in commit waiter | |
2977 | * (and associated commit itx) is committed to an lwb. If the waiter is | |
2978 | * not already committed to an lwb, all itxs in the zilog's queue of | |
2979 | * itxs will be processed. The assumption is the passed in waiter's | |
2980 | * commit itx will found in the queue just like the other non-commit | |
2981 | * itxs, such that when the entire queue is processed, the waiter will | |
2fe61a7e | 2982 | * have been committed to an lwb. |
1ce23dca PS |
2983 | * |
2984 | * The lwb associated with the passed in waiter is not guaranteed to | |
2985 | * have been issued by the time this function completes. If the lwb is | |
2986 | * not issued, we rely on future calls to zil_commit_writer() to issue | |
2987 | * the lwb, or the timeout mechanism found in zil_commit_waiter(). | |
2988 | */ | |
233425a1 | 2989 | static uint64_t |
1ce23dca PS |
2990 | zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw) |
2991 | { | |
f63811f0 AM |
2992 | list_t ilwbs; |
2993 | lwb_t *lwb; | |
233425a1 | 2994 | uint64_t wtxg = 0; |
f63811f0 | 2995 | |
1ce23dca PS |
2996 | ASSERT(!MUTEX_HELD(&zilog->zl_lock)); |
2997 | ASSERT(spa_writeable(zilog->zl_spa)); | |
1ce23dca | 2998 | |
f63811f0 | 2999 | list_create(&ilwbs, sizeof (lwb_t), offsetof(lwb_t, lwb_issue_node)); |
1b2b0aca | 3000 | mutex_enter(&zilog->zl_issuer_lock); |
1ce23dca PS |
3001 | |
3002 | if (zcw->zcw_lwb != NULL || zcw->zcw_done) { | |
3003 | /* | |
3004 | * It's possible that, while we were waiting to acquire | |
1b2b0aca | 3005 | * the "zl_issuer_lock", another thread committed this |
1ce23dca PS |
3006 | * waiter to an lwb. If that occurs, we bail out early, |
3007 | * without processing any of the zilog's queue of itxs. | |
3008 | * | |
3009 | * On certain workloads and system configurations, the | |
1b2b0aca | 3010 | * "zl_issuer_lock" can become highly contended. In an |
1ce23dca PS |
3011 | * attempt to reduce this contention, we immediately drop |
3012 | * the lock if the waiter has already been processed. | |
3013 | * | |
3014 | * We've measured this optimization to reduce CPU spent | |
3015 | * contending on this lock by up to 5%, using a system | |
3016 | * with 32 CPUs, low latency storage (~50 usec writes), | |
3017 | * and 1024 threads performing sync writes. | |
3018 | */ | |
3019 | goto out; | |
3020 | } | |
3021 | ||
fb087146 | 3022 | ZIL_STAT_BUMP(zilog, zil_commit_writer_count); |
1ce23dca | 3023 | |
233425a1 | 3024 | wtxg = zil_get_commit_list(zilog); |
1ce23dca | 3025 | zil_prune_commit_list(zilog); |
f63811f0 | 3026 | zil_process_commit_list(zilog, zcw, &ilwbs); |
1ce23dca PS |
3027 | |
3028 | out: | |
1b2b0aca | 3029 | mutex_exit(&zilog->zl_issuer_lock); |
f63811f0 AM |
3030 | while ((lwb = list_remove_head(&ilwbs)) != NULL) |
3031 | zil_lwb_write_issue(zilog, lwb); | |
3032 | list_destroy(&ilwbs); | |
233425a1 | 3033 | return (wtxg); |
1ce23dca PS |
3034 | } |
3035 | ||
3036 | static void | |
3037 | zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw) | |
3038 | { | |
1b2b0aca | 3039 | ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca PS |
3040 | ASSERT(MUTEX_HELD(&zcw->zcw_lock)); |
3041 | ASSERT3B(zcw->zcw_done, ==, B_FALSE); | |
3042 | ||
3043 | lwb_t *lwb = zcw->zcw_lwb; | |
3044 | ASSERT3P(lwb, !=, NULL); | |
eda3fcd5 | 3045 | ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW); |
34dc7c2f BB |
3046 | |
3047 | /* | |
1ce23dca PS |
3048 | * If the lwb has already been issued by another thread, we can |
3049 | * immediately return since there's no work to be done (the | |
3050 | * point of this function is to issue the lwb). Additionally, we | |
1b2b0aca | 3051 | * do this prior to acquiring the zl_issuer_lock, to avoid |
1ce23dca | 3052 | * acquiring it when it's not necessary to do so. |
34dc7c2f | 3053 | */ |
eda3fcd5 | 3054 | if (lwb->lwb_state != LWB_STATE_OPENED) |
1ce23dca | 3055 | return; |
34dc7c2f | 3056 | |
1ce23dca | 3057 | /* |
f63811f0 | 3058 | * In order to call zil_lwb_write_close() we must hold the |
1b2b0aca | 3059 | * zilog's "zl_issuer_lock". We can't simply acquire that lock, |
1ce23dca | 3060 | * since we're already holding the commit waiter's "zcw_lock", |
2fe61a7e | 3061 | * and those two locks are acquired in the opposite order |
1ce23dca PS |
3062 | * elsewhere. |
3063 | */ | |
3064 | mutex_exit(&zcw->zcw_lock); | |
1b2b0aca | 3065 | mutex_enter(&zilog->zl_issuer_lock); |
1ce23dca | 3066 | mutex_enter(&zcw->zcw_lock); |
34dc7c2f | 3067 | |
1ce23dca PS |
3068 | /* |
3069 | * Since we just dropped and re-acquired the commit waiter's | |
3070 | * lock, we have to re-check to see if the waiter was marked | |
3071 | * "done" during that process. If the waiter was marked "done", | |
3072 | * the "lwb" pointer is no longer valid (it can be free'd after | |
3073 | * the waiter is marked "done"), so without this check we could | |
3074 | * wind up with a use-after-free error below. | |
3075 | */ | |
f63811f0 | 3076 | if (zcw->zcw_done) { |
eda3fcd5 AM |
3077 | mutex_exit(&zilog->zl_issuer_lock); |
3078 | return; | |
f63811f0 | 3079 | } |
119a394a | 3080 | |
1ce23dca PS |
3081 | ASSERT3P(lwb, ==, zcw->zcw_lwb); |
3082 | ||
3083 | /* | |
2fe61a7e PS |
3084 | * We've already checked this above, but since we hadn't acquired |
3085 | * the zilog's zl_issuer_lock, we have to perform this check a | |
3086 | * second time while holding the lock. | |
3087 | * | |
3088 | * We don't need to hold the zl_lock since the lwb cannot transition | |
eda3fcd5 AM |
3089 | * from OPENED to CLOSED while we hold the zl_issuer_lock. The lwb |
3090 | * _can_ transition from CLOSED to DONE, but it's OK to race with | |
2fe61a7e | 3091 | * that transition since we treat the lwb the same, whether it's in |
eda3fcd5 | 3092 | * the CLOSED, ISSUED or DONE states. |
2fe61a7e PS |
3093 | * |
3094 | * The important thing, is we treat the lwb differently depending on | |
eda3fcd5 AM |
3095 | * if it's OPENED or CLOSED, and block any other threads that might |
3096 | * attempt to close/issue this lwb. For that reason we hold the | |
2fe61a7e | 3097 | * zl_issuer_lock when checking the lwb_state; we must not call |
eda3fcd5 | 3098 | * zil_lwb_write_close() if the lwb had already been closed/issued. |
2fe61a7e PS |
3099 | * |
3100 | * See the comment above the lwb_state_t structure definition for | |
3101 | * more details on the lwb states, and locking requirements. | |
1ce23dca | 3102 | */ |
eda3fcd5 AM |
3103 | if (lwb->lwb_state != LWB_STATE_OPENED) { |
3104 | mutex_exit(&zilog->zl_issuer_lock); | |
3105 | return; | |
f63811f0 | 3106 | } |
1ce23dca | 3107 | |
eda3fcd5 AM |
3108 | /* |
3109 | * We do not need zcw_lock once we hold zl_issuer_lock and know lwb | |
3110 | * is still open. But we have to drop it to avoid a deadlock in case | |
3111 | * callback of zio issued by zil_lwb_write_issue() try to get it, | |
3112 | * while zil_lwb_write_issue() is blocked on attempt to issue next | |
3113 | * lwb it found in LWB_STATE_READY state. | |
3114 | */ | |
3115 | mutex_exit(&zcw->zcw_lock); | |
1ce23dca PS |
3116 | |
3117 | /* | |
3118 | * As described in the comments above zil_commit_waiter() and | |
3119 | * zil_process_commit_list(), we need to issue this lwb's zio | |
3120 | * since we've reached the commit waiter's timeout and it still | |
3121 | * hasn't been issued. | |
3122 | */ | |
eda3fcd5 | 3123 | lwb_t *nlwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW); |
1ce23dca | 3124 | |
eda3fcd5 | 3125 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED); |
1ce23dca | 3126 | |
252f46be | 3127 | zil_burst_done(zilog); |
1ce23dca PS |
3128 | |
3129 | if (nlwb == NULL) { | |
3130 | /* | |
f63811f0 | 3131 | * When zil_lwb_write_close() returns NULL, this |
1ce23dca PS |
3132 | * indicates zio_alloc_zil() failed to allocate the |
3133 | * "next" lwb on-disk. When this occurs, the ZIL write | |
3134 | * pipeline must be stalled; see the comment within the | |
3135 | * zil_commit_writer_stall() function for more details. | |
1ce23dca | 3136 | */ |
f63811f0 | 3137 | zil_lwb_write_issue(zilog, lwb); |
1ce23dca | 3138 | zil_commit_writer_stall(zilog); |
eda3fcd5 AM |
3139 | mutex_exit(&zilog->zl_issuer_lock); |
3140 | } else { | |
3141 | mutex_exit(&zilog->zl_issuer_lock); | |
f63811f0 | 3142 | zil_lwb_write_issue(zilog, lwb); |
eda3fcd5 AM |
3143 | } |
3144 | mutex_enter(&zcw->zcw_lock); | |
1ce23dca PS |
3145 | } |
3146 | ||
3147 | /* | |
3148 | * This function is responsible for performing the following two tasks: | |
3149 | * | |
3150 | * 1. its primary responsibility is to block until the given "commit | |
3151 | * waiter" is considered "done". | |
3152 | * | |
3153 | * 2. its secondary responsibility is to issue the zio for the lwb that | |
3154 | * the given "commit waiter" is waiting on, if this function has | |
3155 | * waited "long enough" and the lwb is still in the "open" state. | |
3156 | * | |
3157 | * Given a sufficient amount of itxs being generated and written using | |
f63811f0 | 3158 | * the ZIL, the lwb's zio will be issued via the zil_lwb_assign() |
1ce23dca PS |
3159 | * function. If this does not occur, this secondary responsibility will |
3160 | * ensure the lwb is issued even if there is not other synchronous | |
3161 | * activity on the system. | |
3162 | * | |
3163 | * For more details, see zil_process_commit_list(); more specifically, | |
3164 | * the comment at the bottom of that function. | |
3165 | */ | |
3166 | static void | |
3167 | zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw) | |
3168 | { | |
3169 | ASSERT(!MUTEX_HELD(&zilog->zl_lock)); | |
1b2b0aca | 3170 | ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); |
1ce23dca | 3171 | ASSERT(spa_writeable(zilog->zl_spa)); |
1ce23dca PS |
3172 | |
3173 | mutex_enter(&zcw->zcw_lock); | |
428870ff BB |
3174 | |
3175 | /* | |
1ce23dca PS |
3176 | * The timeout is scaled based on the lwb latency to avoid |
3177 | * significantly impacting the latency of each individual itx. | |
3178 | * For more details, see the comment at the bottom of the | |
3179 | * zil_process_commit_list() function. | |
428870ff | 3180 | */ |
1ce23dca PS |
3181 | int pct = MAX(zfs_commit_timeout_pct, 1); |
3182 | hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100; | |
3183 | hrtime_t wakeup = gethrtime() + sleep; | |
3184 | boolean_t timedout = B_FALSE; | |
3185 | ||
3186 | while (!zcw->zcw_done) { | |
3187 | ASSERT(MUTEX_HELD(&zcw->zcw_lock)); | |
3188 | ||
3189 | lwb_t *lwb = zcw->zcw_lwb; | |
3190 | ||
3191 | /* | |
3192 | * Usually, the waiter will have a non-NULL lwb field here, | |
3193 | * but it's possible for it to be NULL as a result of | |
3194 | * zil_commit() racing with spa_sync(). | |
3195 | * | |
3196 | * When zil_clean() is called, it's possible for the itxg | |
3197 | * list (which may be cleaned via a taskq) to contain | |
3198 | * commit itxs. When this occurs, the commit waiters linked | |
3199 | * off of these commit itxs will not be committed to an | |
3200 | * lwb. Additionally, these commit waiters will not be | |
3201 | * marked done until zil_commit_waiter_skip() is called via | |
3202 | * zil_itxg_clean(). | |
3203 | * | |
3204 | * Thus, it's possible for this commit waiter (i.e. the | |
3205 | * "zcw" variable) to be found in this "in between" state; | |
3206 | * where it's "zcw_lwb" field is NULL, and it hasn't yet | |
3207 | * been skipped, so it's "zcw_done" field is still B_FALSE. | |
3208 | */ | |
eda3fcd5 | 3209 | IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_NEW); |
1ce23dca PS |
3210 | |
3211 | if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) { | |
3212 | ASSERT3B(timedout, ==, B_FALSE); | |
3213 | ||
3214 | /* | |
3215 | * If the lwb hasn't been issued yet, then we | |
3216 | * need to wait with a timeout, in case this | |
3217 | * function needs to issue the lwb after the | |
3218 | * timeout is reached; responsibility (2) from | |
3219 | * the comment above this function. | |
3220 | */ | |
8056a756 | 3221 | int rc = cv_timedwait_hires(&zcw->zcw_cv, |
1ce23dca PS |
3222 | &zcw->zcw_lock, wakeup, USEC2NSEC(1), |
3223 | CALLOUT_FLAG_ABSOLUTE); | |
3224 | ||
8056a756 | 3225 | if (rc != -1 || zcw->zcw_done) |
1ce23dca PS |
3226 | continue; |
3227 | ||
3228 | timedout = B_TRUE; | |
3229 | zil_commit_waiter_timeout(zilog, zcw); | |
3230 | ||
3231 | if (!zcw->zcw_done) { | |
3232 | /* | |
3233 | * If the commit waiter has already been | |
3234 | * marked "done", it's possible for the | |
3235 | * waiter's lwb structure to have already | |
3236 | * been freed. Thus, we can only reliably | |
3237 | * make these assertions if the waiter | |
3238 | * isn't done. | |
3239 | */ | |
3240 | ASSERT3P(lwb, ==, zcw->zcw_lwb); | |
3241 | ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED); | |
3242 | } | |
3243 | } else { | |
3244 | /* | |
3245 | * If the lwb isn't open, then it must have already | |
3246 | * been issued. In that case, there's no need to | |
3247 | * use a timeout when waiting for the lwb to | |
3248 | * complete. | |
3249 | * | |
3250 | * Additionally, if the lwb is NULL, the waiter | |
2fe61a7e | 3251 | * will soon be signaled and marked done via |
1ce23dca PS |
3252 | * zil_clean() and zil_itxg_clean(), so no timeout |
3253 | * is required. | |
3254 | */ | |
3255 | ||
3256 | IMPLY(lwb != NULL, | |
eda3fcd5 AM |
3257 | lwb->lwb_state == LWB_STATE_CLOSED || |
3258 | lwb->lwb_state == LWB_STATE_READY || | |
1ce23dca | 3259 | lwb->lwb_state == LWB_STATE_ISSUED || |
900d09b2 PS |
3260 | lwb->lwb_state == LWB_STATE_WRITE_DONE || |
3261 | lwb->lwb_state == LWB_STATE_FLUSH_DONE); | |
1ce23dca PS |
3262 | cv_wait(&zcw->zcw_cv, &zcw->zcw_lock); |
3263 | } | |
3264 | } | |
3265 | ||
3266 | mutex_exit(&zcw->zcw_lock); | |
3267 | } | |
3268 | ||
3269 | static zil_commit_waiter_t * | |
3270 | zil_alloc_commit_waiter(void) | |
3271 | { | |
3272 | zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP); | |
3273 | ||
3274 | cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL); | |
3275 | mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL); | |
3276 | list_link_init(&zcw->zcw_node); | |
3277 | zcw->zcw_lwb = NULL; | |
3278 | zcw->zcw_done = B_FALSE; | |
3279 | zcw->zcw_zio_error = 0; | |
3280 | ||
3281 | return (zcw); | |
3282 | } | |
3283 | ||
3284 | static void | |
3285 | zil_free_commit_waiter(zil_commit_waiter_t *zcw) | |
3286 | { | |
3287 | ASSERT(!list_link_active(&zcw->zcw_node)); | |
3288 | ASSERT3P(zcw->zcw_lwb, ==, NULL); | |
3289 | ASSERT3B(zcw->zcw_done, ==, B_TRUE); | |
3290 | mutex_destroy(&zcw->zcw_lock); | |
3291 | cv_destroy(&zcw->zcw_cv); | |
3292 | kmem_cache_free(zil_zcw_cache, zcw); | |
34dc7c2f BB |
3293 | } |
3294 | ||
3295 | /* | |
1ce23dca PS |
3296 | * This function is used to create a TX_COMMIT itx and assign it. This |
3297 | * way, it will be linked into the ZIL's list of synchronous itxs, and | |
3298 | * then later committed to an lwb (or skipped) when | |
3299 | * zil_process_commit_list() is called. | |
3300 | */ | |
3301 | static void | |
3302 | zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw) | |
3303 | { | |
3304 | dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); | |
2fd1c304 AM |
3305 | |
3306 | /* | |
3307 | * Since we are not going to create any new dirty data, and we | |
3308 | * can even help with clearing the existing dirty data, we | |
3309 | * should not be subject to the dirty data based delays. We | |
3310 | * use TXG_NOTHROTTLE to bypass the delay mechanism. | |
3311 | */ | |
3312 | VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); | |
1ce23dca PS |
3313 | |
3314 | itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t)); | |
3315 | itx->itx_sync = B_TRUE; | |
3316 | itx->itx_private = zcw; | |
3317 | ||
3318 | zil_itx_assign(zilog, itx, tx); | |
3319 | ||
3320 | dmu_tx_commit(tx); | |
3321 | } | |
3322 | ||
3323 | /* | |
3324 | * Commit ZFS Intent Log transactions (itxs) to stable storage. | |
3325 | * | |
3326 | * When writing ZIL transactions to the on-disk representation of the | |
3327 | * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple | |
3328 | * itxs can be committed to a single lwb. Once a lwb is written and | |
3329 | * committed to stable storage (i.e. the lwb is written, and vdevs have | |
3330 | * been flushed), each itx that was committed to that lwb is also | |
3331 | * considered to be committed to stable storage. | |
3332 | * | |
3333 | * When an itx is committed to an lwb, the log record (lr_t) contained | |
3334 | * by the itx is copied into the lwb's zio buffer, and once this buffer | |
3335 | * is written to disk, it becomes an on-disk ZIL block. | |
3336 | * | |
3337 | * As itxs are generated, they're inserted into the ZIL's queue of | |
3338 | * uncommitted itxs. The semantics of zil_commit() are such that it will | |
3339 | * block until all itxs that were in the queue when it was called, are | |
3340 | * committed to stable storage. | |
3341 | * | |
3342 | * If "foid" is zero, this means all "synchronous" and "asynchronous" | |
3343 | * itxs, for all objects in the dataset, will be committed to stable | |
3344 | * storage prior to zil_commit() returning. If "foid" is non-zero, all | |
3345 | * "synchronous" itxs for all objects, but only "asynchronous" itxs | |
3346 | * that correspond to the foid passed in, will be committed to stable | |
3347 | * storage prior to zil_commit() returning. | |
3348 | * | |
3349 | * Generally speaking, when zil_commit() is called, the consumer doesn't | |
3350 | * actually care about _all_ of the uncommitted itxs. Instead, they're | |
3351 | * simply trying to waiting for a specific itx to be committed to disk, | |
3352 | * but the interface(s) for interacting with the ZIL don't allow such | |
3353 | * fine-grained communication. A better interface would allow a consumer | |
3354 | * to create and assign an itx, and then pass a reference to this itx to | |
3355 | * zil_commit(); such that zil_commit() would return as soon as that | |
3356 | * specific itx was committed to disk (instead of waiting for _all_ | |
3357 | * itxs to be committed). | |
3358 | * | |
3359 | * When a thread calls zil_commit() a special "commit itx" will be | |
3360 | * generated, along with a corresponding "waiter" for this commit itx. | |
3361 | * zil_commit() will wait on this waiter's CV, such that when the waiter | |
2fe61a7e | 3362 | * is marked done, and signaled, zil_commit() will return. |
1ce23dca PS |
3363 | * |
3364 | * This commit itx is inserted into the queue of uncommitted itxs. This | |
3365 | * provides an easy mechanism for determining which itxs were in the | |
3366 | * queue prior to zil_commit() having been called, and which itxs were | |
3367 | * added after zil_commit() was called. | |
3368 | * | |
2310dba9 | 3369 | * The commit itx is special; it doesn't have any on-disk representation. |
1ce23dca PS |
3370 | * When a commit itx is "committed" to an lwb, the waiter associated |
3371 | * with it is linked onto the lwb's list of waiters. Then, when that lwb | |
2fe61a7e | 3372 | * completes, each waiter on the lwb's list is marked done and signaled |
1ce23dca PS |
3373 | * -- allowing the thread waiting on the waiter to return from zil_commit(). |
3374 | * | |
3375 | * It's important to point out a few critical factors that allow us | |
3376 | * to make use of the commit itxs, commit waiters, per-lwb lists of | |
3377 | * commit waiters, and zio completion callbacks like we're doing: | |
572e2857 | 3378 | * |
1ce23dca | 3379 | * 1. The list of waiters for each lwb is traversed, and each commit |
2fe61a7e | 3380 | * waiter is marked "done" and signaled, in the zio completion |
1ce23dca | 3381 | * callback of the lwb's zio[*]. |
572e2857 | 3382 | * |
2fe61a7e | 3383 | * * Actually, the waiters are signaled in the zio completion |
1ce23dca PS |
3384 | * callback of the root zio for the DKIOCFLUSHWRITECACHE commands |
3385 | * that are sent to the vdevs upon completion of the lwb zio. | |
572e2857 | 3386 | * |
1ce23dca PS |
3387 | * 2. When the itxs are inserted into the ZIL's queue of uncommitted |
3388 | * itxs, the order in which they are inserted is preserved[*]; as | |
3389 | * itxs are added to the queue, they are added to the tail of | |
3390 | * in-memory linked lists. | |
572e2857 | 3391 | * |
1ce23dca PS |
3392 | * When committing the itxs to lwbs (to be written to disk), they |
3393 | * are committed in the same order in which the itxs were added to | |
3394 | * the uncommitted queue's linked list(s); i.e. the linked list of | |
3395 | * itxs to commit is traversed from head to tail, and each itx is | |
3396 | * committed to an lwb in that order. | |
3397 | * | |
3398 | * * To clarify: | |
3399 | * | |
3400 | * - the order of "sync" itxs is preserved w.r.t. other | |
3401 | * "sync" itxs, regardless of the corresponding objects. | |
3402 | * - the order of "async" itxs is preserved w.r.t. other | |
3403 | * "async" itxs corresponding to the same object. | |
3404 | * - the order of "async" itxs is *not* preserved w.r.t. other | |
3405 | * "async" itxs corresponding to different objects. | |
3406 | * - the order of "sync" itxs w.r.t. "async" itxs (or vice | |
3407 | * versa) is *not* preserved, even for itxs that correspond | |
3408 | * to the same object. | |
3409 | * | |
3410 | * For more details, see: zil_itx_assign(), zil_async_to_sync(), | |
3411 | * zil_get_commit_list(), and zil_process_commit_list(). | |
3412 | * | |
3413 | * 3. The lwbs represent a linked list of blocks on disk. Thus, any | |
3414 | * lwb cannot be considered committed to stable storage, until its | |
3415 | * "previous" lwb is also committed to stable storage. This fact, | |
3416 | * coupled with the fact described above, means that itxs are | |
3417 | * committed in (roughly) the order in which they were generated. | |
3418 | * This is essential because itxs are dependent on prior itxs. | |
3419 | * Thus, we *must not* deem an itx as being committed to stable | |
3420 | * storage, until *all* prior itxs have also been committed to | |
3421 | * stable storage. | |
3422 | * | |
3423 | * To enforce this ordering of lwb zio's, while still leveraging as | |
3424 | * much of the underlying storage performance as possible, we rely | |
3425 | * on two fundamental concepts: | |
3426 | * | |
3427 | * 1. The creation and issuance of lwb zio's is protected by | |
1b2b0aca | 3428 | * the zilog's "zl_issuer_lock", which ensures only a single |
1ce23dca PS |
3429 | * thread is creating and/or issuing lwb's at a time |
3430 | * 2. The "previous" lwb is a child of the "current" lwb | |
2fe61a7e | 3431 | * (leveraging the zio parent-child dependency graph) |
1ce23dca PS |
3432 | * |
3433 | * By relying on this parent-child zio relationship, we can have | |
3434 | * many lwb zio's concurrently issued to the underlying storage, | |
3435 | * but the order in which they complete will be the same order in | |
3436 | * which they were created. | |
34dc7c2f BB |
3437 | */ |
3438 | void | |
572e2857 | 3439 | zil_commit(zilog_t *zilog, uint64_t foid) |
34dc7c2f | 3440 | { |
1ce23dca PS |
3441 | /* |
3442 | * We should never attempt to call zil_commit on a snapshot for | |
3443 | * a couple of reasons: | |
3444 | * | |
3445 | * 1. A snapshot may never be modified, thus it cannot have any | |
3446 | * in-flight itxs that would have modified the dataset. | |
3447 | * | |
3448 | * 2. By design, when zil_commit() is called, a commit itx will | |
3449 | * be assigned to this zilog; as a result, the zilog will be | |
3450 | * dirtied. We must not dirty the zilog of a snapshot; there's | |
3451 | * checks in the code that enforce this invariant, and will | |
3452 | * cause a panic if it's not upheld. | |
3453 | */ | |
3454 | ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE); | |
34dc7c2f | 3455 | |
572e2857 BB |
3456 | if (zilog->zl_sync == ZFS_SYNC_DISABLED) |
3457 | return; | |
34dc7c2f | 3458 | |
1ce23dca PS |
3459 | if (!spa_writeable(zilog->zl_spa)) { |
3460 | /* | |
3461 | * If the SPA is not writable, there should never be any | |
3462 | * pending itxs waiting to be committed to disk. If that | |
3463 | * weren't true, we'd skip writing those itxs out, and | |
2fe61a7e | 3464 | * would break the semantics of zil_commit(); thus, we're |
1ce23dca PS |
3465 | * verifying that truth before we return to the caller. |
3466 | */ | |
3467 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
3468 | ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); | |
3469 | for (int i = 0; i < TXG_SIZE; i++) | |
3470 | ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL); | |
3471 | return; | |
3472 | } | |
3473 | ||
3474 | /* | |
3475 | * If the ZIL is suspended, we don't want to dirty it by calling | |
3476 | * zil_commit_itx_assign() below, nor can we write out | |
3477 | * lwbs like would be done in zil_commit_write(). Thus, we | |
3478 | * simply rely on txg_wait_synced() to maintain the necessary | |
3479 | * semantics, and avoid calling those functions altogether. | |
3480 | */ | |
3481 | if (zilog->zl_suspend > 0) { | |
3482 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
3483 | return; | |
3484 | } | |
3485 | ||
2fe61a7e PS |
3486 | zil_commit_impl(zilog, foid); |
3487 | } | |
3488 | ||
3489 | void | |
3490 | zil_commit_impl(zilog_t *zilog, uint64_t foid) | |
3491 | { | |
fb087146 | 3492 | ZIL_STAT_BUMP(zilog, zil_commit_count); |
b6ad9671 | 3493 | |
1ce23dca PS |
3494 | /* |
3495 | * Move the "async" itxs for the specified foid to the "sync" | |
3496 | * queues, such that they will be later committed (or skipped) | |
3497 | * to an lwb when zil_process_commit_list() is called. | |
3498 | * | |
3499 | * Since these "async" itxs must be committed prior to this | |
3500 | * call to zil_commit returning, we must perform this operation | |
3501 | * before we call zil_commit_itx_assign(). | |
3502 | */ | |
572e2857 | 3503 | zil_async_to_sync(zilog, foid); |
34dc7c2f | 3504 | |
1ce23dca PS |
3505 | /* |
3506 | * We allocate a new "waiter" structure which will initially be | |
3507 | * linked to the commit itx using the itx's "itx_private" field. | |
3508 | * Since the commit itx doesn't represent any on-disk state, | |
3509 | * when it's committed to an lwb, rather than copying the its | |
3510 | * lr_t into the lwb's buffer, the commit itx's "waiter" will be | |
3511 | * added to the lwb's list of waiters. Then, when the lwb is | |
3512 | * committed to stable storage, each waiter in the lwb's list of | |
3513 | * waiters will be marked "done", and signalled. | |
3514 | * | |
3515 | * We must create the waiter and assign the commit itx prior to | |
3516 | * calling zil_commit_writer(), or else our specific commit itx | |
3517 | * is not guaranteed to be committed to an lwb prior to calling | |
3518 | * zil_commit_waiter(). | |
3519 | */ | |
3520 | zil_commit_waiter_t *zcw = zil_alloc_commit_waiter(); | |
3521 | zil_commit_itx_assign(zilog, zcw); | |
428870ff | 3522 | |
233425a1 | 3523 | uint64_t wtxg = zil_commit_writer(zilog, zcw); |
1ce23dca | 3524 | zil_commit_waiter(zilog, zcw); |
428870ff | 3525 | |
1ce23dca PS |
3526 | if (zcw->zcw_zio_error != 0) { |
3527 | /* | |
3528 | * If there was an error writing out the ZIL blocks that | |
3529 | * this thread is waiting on, then we fallback to | |
3530 | * relying on spa_sync() to write out the data this | |
3531 | * thread is waiting on. Obviously this has performance | |
3532 | * implications, but the expectation is for this to be | |
3533 | * an exceptional case, and shouldn't occur often. | |
3534 | */ | |
3535 | DTRACE_PROBE2(zil__commit__io__error, | |
3536 | zilog_t *, zilog, zil_commit_waiter_t *, zcw); | |
3537 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
233425a1 AM |
3538 | } else if (wtxg != 0) { |
3539 | txg_wait_synced(zilog->zl_dmu_pool, wtxg); | |
1ce23dca | 3540 | } |
8c0712fd | 3541 | |
1ce23dca | 3542 | zil_free_commit_waiter(zcw); |
428870ff BB |
3543 | } |
3544 | ||
34dc7c2f BB |
3545 | /* |
3546 | * Called in syncing context to free committed log blocks and update log header. | |
3547 | */ | |
3548 | void | |
3549 | zil_sync(zilog_t *zilog, dmu_tx_t *tx) | |
3550 | { | |
3551 | zil_header_t *zh = zil_header_in_syncing_context(zilog); | |
3552 | uint64_t txg = dmu_tx_get_txg(tx); | |
3553 | spa_t *spa = zilog->zl_spa; | |
428870ff | 3554 | uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; |
34dc7c2f BB |
3555 | lwb_t *lwb; |
3556 | ||
9babb374 BB |
3557 | /* |
3558 | * We don't zero out zl_destroy_txg, so make sure we don't try | |
3559 | * to destroy it twice. | |
3560 | */ | |
3561 | if (spa_sync_pass(spa) != 1) | |
3562 | return; | |
3563 | ||
152d6fda KJ |
3564 | zil_lwb_flush_wait_all(zilog, txg); |
3565 | ||
34dc7c2f BB |
3566 | mutex_enter(&zilog->zl_lock); |
3567 | ||
3568 | ASSERT(zilog->zl_stop_sync == 0); | |
3569 | ||
428870ff BB |
3570 | if (*replayed_seq != 0) { |
3571 | ASSERT(zh->zh_replay_seq < *replayed_seq); | |
3572 | zh->zh_replay_seq = *replayed_seq; | |
3573 | *replayed_seq = 0; | |
3574 | } | |
34dc7c2f BB |
3575 | |
3576 | if (zilog->zl_destroy_txg == txg) { | |
3577 | blkptr_t blk = zh->zh_log; | |
361a7e82 | 3578 | dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); |
34dc7c2f | 3579 | |
895e0313 | 3580 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
34dc7c2f | 3581 | |
861166b0 AZ |
3582 | memset(zh, 0, sizeof (zil_header_t)); |
3583 | memset(zilog->zl_replayed_seq, 0, | |
3584 | sizeof (zilog->zl_replayed_seq)); | |
34dc7c2f BB |
3585 | |
3586 | if (zilog->zl_keep_first) { | |
3587 | /* | |
3588 | * If this block was part of log chain that couldn't | |
3589 | * be claimed because a device was missing during | |
3590 | * zil_claim(), but that device later returns, | |
3591 | * then this block could erroneously appear valid. | |
3592 | * To guard against this, assign a new GUID to the new | |
3593 | * log chain so it doesn't matter what blk points to. | |
3594 | */ | |
3595 | zil_init_log_chain(zilog, &blk); | |
3596 | zh->zh_log = blk; | |
361a7e82 JP |
3597 | } else { |
3598 | /* | |
3599 | * A destroyed ZIL chain can't contain any TX_SETSAXATTR | |
3600 | * records. So, deactivate the feature for this dataset. | |
3601 | * We activate it again when we start a new ZIL chain. | |
3602 | */ | |
3603 | if (dsl_dataset_feature_is_active(ds, | |
3604 | SPA_FEATURE_ZILSAXATTR)) | |
3605 | dsl_dataset_deactivate_feature(ds, | |
3606 | SPA_FEATURE_ZILSAXATTR, tx); | |
34dc7c2f BB |
3607 | } |
3608 | } | |
3609 | ||
9babb374 | 3610 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { |
34dc7c2f | 3611 | zh->zh_log = lwb->lwb_blk; |
7381ddf1 | 3612 | if (lwb->lwb_state != LWB_STATE_FLUSH_DONE || |
eda3fcd5 | 3613 | lwb->lwb_alloc_txg > txg || lwb->lwb_max_txg > txg) |
34dc7c2f BB |
3614 | break; |
3615 | list_remove(&zilog->zl_lwb_list, lwb); | |
eda3fcd5 AM |
3616 | if (!BP_IS_HOLE(&lwb->lwb_blk)) |
3617 | zio_free(spa, txg, &lwb->lwb_blk); | |
1ce23dca | 3618 | zil_free_lwb(zilog, lwb); |
34dc7c2f BB |
3619 | |
3620 | /* | |
3621 | * If we don't have anything left in the lwb list then | |
3622 | * we've had an allocation failure and we need to zero | |
3623 | * out the zil_header blkptr so that we don't end | |
3624 | * up freeing the same block twice. | |
3625 | */ | |
895e0313 | 3626 | if (list_is_empty(&zilog->zl_lwb_list)) |
34dc7c2f BB |
3627 | BP_ZERO(&zh->zh_log); |
3628 | } | |
920dd524 | 3629 | |
34dc7c2f BB |
3630 | mutex_exit(&zilog->zl_lock); |
3631 | } | |
3632 | ||
1ce23dca PS |
3633 | static int |
3634 | zil_lwb_cons(void *vbuf, void *unused, int kmflag) | |
3635 | { | |
14e4e3cb | 3636 | (void) unused, (void) kmflag; |
1ce23dca PS |
3637 | lwb_t *lwb = vbuf; |
3638 | list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); | |
3639 | list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t), | |
3640 | offsetof(zil_commit_waiter_t, zcw_node)); | |
3641 | avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare, | |
3642 | sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); | |
3643 | mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL); | |
3644 | return (0); | |
3645 | } | |
3646 | ||
1ce23dca PS |
3647 | static void |
3648 | zil_lwb_dest(void *vbuf, void *unused) | |
3649 | { | |
14e4e3cb | 3650 | (void) unused; |
1ce23dca PS |
3651 | lwb_t *lwb = vbuf; |
3652 | mutex_destroy(&lwb->lwb_vdev_lock); | |
3653 | avl_destroy(&lwb->lwb_vdev_tree); | |
3654 | list_destroy(&lwb->lwb_waiters); | |
3655 | list_destroy(&lwb->lwb_itxs); | |
3656 | } | |
3657 | ||
34dc7c2f BB |
3658 | void |
3659 | zil_init(void) | |
3660 | { | |
3661 | zil_lwb_cache = kmem_cache_create("zil_lwb_cache", | |
1ce23dca PS |
3662 | sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0); |
3663 | ||
3664 | zil_zcw_cache = kmem_cache_create("zil_zcw_cache", | |
3665 | sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
b6ad9671 | 3666 | |
fb087146 AH |
3667 | zil_sums_init(&zil_sums_global); |
3668 | zil_kstats_global = kstat_create("zfs", 0, "zil", "misc", | |
d1d7e268 | 3669 | KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t), |
b6ad9671 ED |
3670 | KSTAT_FLAG_VIRTUAL); |
3671 | ||
fb087146 AH |
3672 | if (zil_kstats_global != NULL) { |
3673 | zil_kstats_global->ks_data = &zil_stats; | |
3674 | zil_kstats_global->ks_update = zil_kstats_global_update; | |
3675 | zil_kstats_global->ks_private = NULL; | |
3676 | kstat_install(zil_kstats_global); | |
b6ad9671 | 3677 | } |
34dc7c2f BB |
3678 | } |
3679 | ||
3680 | void | |
3681 | zil_fini(void) | |
3682 | { | |
1ce23dca | 3683 | kmem_cache_destroy(zil_zcw_cache); |
34dc7c2f | 3684 | kmem_cache_destroy(zil_lwb_cache); |
b6ad9671 | 3685 | |
fb087146 AH |
3686 | if (zil_kstats_global != NULL) { |
3687 | kstat_delete(zil_kstats_global); | |
3688 | zil_kstats_global = NULL; | |
b6ad9671 | 3689 | } |
fb087146 AH |
3690 | |
3691 | zil_sums_fini(&zil_sums_global); | |
34dc7c2f BB |
3692 | } |
3693 | ||
428870ff BB |
3694 | void |
3695 | zil_set_sync(zilog_t *zilog, uint64_t sync) | |
3696 | { | |
3697 | zilog->zl_sync = sync; | |
3698 | } | |
3699 | ||
3700 | void | |
3701 | zil_set_logbias(zilog_t *zilog, uint64_t logbias) | |
3702 | { | |
3703 | zilog->zl_logbias = logbias; | |
3704 | } | |
3705 | ||
34dc7c2f BB |
3706 | zilog_t * |
3707 | zil_alloc(objset_t *os, zil_header_t *zh_phys) | |
3708 | { | |
3709 | zilog_t *zilog; | |
3710 | ||
79c76d5b | 3711 | zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); |
34dc7c2f BB |
3712 | |
3713 | zilog->zl_header = zh_phys; | |
3714 | zilog->zl_os = os; | |
3715 | zilog->zl_spa = dmu_objset_spa(os); | |
3716 | zilog->zl_dmu_pool = dmu_objset_pool(os); | |
3717 | zilog->zl_destroy_txg = TXG_INITIAL - 1; | |
428870ff BB |
3718 | zilog->zl_logbias = dmu_objset_logbias(os); |
3719 | zilog->zl_sync = dmu_objset_syncprop(os); | |
1ce23dca PS |
3720 | zilog->zl_dirty_max_txg = 0; |
3721 | zilog->zl_last_lwb_opened = NULL; | |
3722 | zilog->zl_last_lwb_latency = 0; | |
b8738257 | 3723 | zilog->zl_max_block_size = zil_maxblocksize; |
34dc7c2f BB |
3724 | |
3725 | mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); | |
1b2b0aca | 3726 | mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL); |
152d6fda | 3727 | mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f | 3728 | |
1c27024e | 3729 | for (int i = 0; i < TXG_SIZE; i++) { |
572e2857 BB |
3730 | mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, |
3731 | MUTEX_DEFAULT, NULL); | |
3732 | } | |
34dc7c2f BB |
3733 | |
3734 | list_create(&zilog->zl_lwb_list, sizeof (lwb_t), | |
3735 | offsetof(lwb_t, lwb_node)); | |
3736 | ||
572e2857 BB |
3737 | list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), |
3738 | offsetof(itx_t, itx_node)); | |
3739 | ||
34dc7c2f | 3740 | cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); |
152d6fda | 3741 | cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL); |
34dc7c2f BB |
3742 | |
3743 | return (zilog); | |
3744 | } | |
3745 | ||
3746 | void | |
3747 | zil_free(zilog_t *zilog) | |
3748 | { | |
d6320ddb | 3749 | int i; |
34dc7c2f BB |
3750 | |
3751 | zilog->zl_stop_sync = 1; | |
3752 | ||
13fe0198 MA |
3753 | ASSERT0(zilog->zl_suspend); |
3754 | ASSERT0(zilog->zl_suspending); | |
3755 | ||
3e31d2b0 | 3756 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
34dc7c2f BB |
3757 | list_destroy(&zilog->zl_lwb_list); |
3758 | ||
572e2857 BB |
3759 | ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); |
3760 | list_destroy(&zilog->zl_itx_commit_list); | |
3761 | ||
d6320ddb | 3762 | for (i = 0; i < TXG_SIZE; i++) { |
572e2857 BB |
3763 | /* |
3764 | * It's possible for an itx to be generated that doesn't dirty | |
3765 | * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() | |
3766 | * callback to remove the entry. We remove those here. | |
3767 | * | |
3768 | * Also free up the ziltest itxs. | |
3769 | */ | |
3770 | if (zilog->zl_itxg[i].itxg_itxs) | |
3771 | zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); | |
3772 | mutex_destroy(&zilog->zl_itxg[i].itxg_lock); | |
3773 | } | |
3774 | ||
1b2b0aca | 3775 | mutex_destroy(&zilog->zl_issuer_lock); |
34dc7c2f | 3776 | mutex_destroy(&zilog->zl_lock); |
152d6fda | 3777 | mutex_destroy(&zilog->zl_lwb_io_lock); |
34dc7c2f | 3778 | |
34dc7c2f | 3779 | cv_destroy(&zilog->zl_cv_suspend); |
152d6fda | 3780 | cv_destroy(&zilog->zl_lwb_io_cv); |
34dc7c2f BB |
3781 | |
3782 | kmem_free(zilog, sizeof (zilog_t)); | |
3783 | } | |
3784 | ||
34dc7c2f BB |
3785 | /* |
3786 | * Open an intent log. | |
3787 | */ | |
3788 | zilog_t * | |
fb087146 | 3789 | zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums) |
34dc7c2f BB |
3790 | { |
3791 | zilog_t *zilog = dmu_objset_zil(os); | |
3792 | ||
1ce23dca PS |
3793 | ASSERT3P(zilog->zl_get_data, ==, NULL); |
3794 | ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); | |
3e31d2b0 ES |
3795 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
3796 | ||
34dc7c2f | 3797 | zilog->zl_get_data = get_data; |
fb087146 | 3798 | zilog->zl_sums = zil_sums; |
34dc7c2f BB |
3799 | |
3800 | return (zilog); | |
3801 | } | |
3802 | ||
3803 | /* | |
3804 | * Close an intent log. | |
3805 | */ | |
3806 | void | |
3807 | zil_close(zilog_t *zilog) | |
3808 | { | |
3e31d2b0 | 3809 | lwb_t *lwb; |
1ce23dca | 3810 | uint64_t txg; |
572e2857 | 3811 | |
1ce23dca PS |
3812 | if (!dmu_objset_is_snapshot(zilog->zl_os)) { |
3813 | zil_commit(zilog, 0); | |
3814 | } else { | |
895e0313 | 3815 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
1ce23dca PS |
3816 | ASSERT0(zilog->zl_dirty_max_txg); |
3817 | ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE); | |
3818 | } | |
572e2857 | 3819 | |
572e2857 | 3820 | mutex_enter(&zilog->zl_lock); |
eda3fcd5 | 3821 | txg = zilog->zl_dirty_max_txg; |
3e31d2b0 | 3822 | lwb = list_tail(&zilog->zl_lwb_list); |
eda3fcd5 AM |
3823 | if (lwb != NULL) { |
3824 | txg = MAX(txg, lwb->lwb_alloc_txg); | |
3825 | txg = MAX(txg, lwb->lwb_max_txg); | |
3826 | } | |
572e2857 | 3827 | mutex_exit(&zilog->zl_lock); |
1ce23dca PS |
3828 | |
3829 | /* | |
152d6fda KJ |
3830 | * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends |
3831 | * on the time when the dmu_tx transaction is assigned in | |
eda3fcd5 | 3832 | * zil_lwb_write_issue(). |
152d6fda KJ |
3833 | */ |
3834 | mutex_enter(&zilog->zl_lwb_io_lock); | |
3835 | txg = MAX(zilog->zl_lwb_max_issued_txg, txg); | |
3836 | mutex_exit(&zilog->zl_lwb_io_lock); | |
3837 | ||
3838 | /* | |
3839 | * We need to use txg_wait_synced() to wait until that txg is synced. | |
3840 | * zil_sync() will guarantee all lwbs up to that txg have been | |
3841 | * written out, flushed, and cleaned. | |
1ce23dca PS |
3842 | */ |
3843 | if (txg != 0) | |
34dc7c2f | 3844 | txg_wait_synced(zilog->zl_dmu_pool, txg); |
55922e73 GW |
3845 | |
3846 | if (zilog_is_dirty(zilog)) | |
8e739b2c RE |
3847 | zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog, |
3848 | (u_longlong_t)txg); | |
50c957f7 | 3849 | if (txg < spa_freeze_txg(zilog->zl_spa)) |
55922e73 | 3850 | VERIFY(!zilog_is_dirty(zilog)); |
34dc7c2f | 3851 | |
34dc7c2f | 3852 | zilog->zl_get_data = NULL; |
3e31d2b0 ES |
3853 | |
3854 | /* | |
1ce23dca | 3855 | * We should have only one lwb left on the list; remove it now. |
3e31d2b0 ES |
3856 | */ |
3857 | mutex_enter(&zilog->zl_lock); | |
895e0313 | 3858 | lwb = list_remove_head(&zilog->zl_lwb_list); |
3e31d2b0 | 3859 | if (lwb != NULL) { |
895e0313 | 3860 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
eda3fcd5 | 3861 | ASSERT3S(lwb->lwb_state, ==, LWB_STATE_NEW); |
3e31d2b0 | 3862 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); |
1ce23dca | 3863 | zil_free_lwb(zilog, lwb); |
3e31d2b0 ES |
3864 | } |
3865 | mutex_exit(&zilog->zl_lock); | |
34dc7c2f BB |
3866 | } |
3867 | ||
a926aab9 | 3868 | static const char *suspend_tag = "zil suspending"; |
13fe0198 | 3869 | |
34dc7c2f BB |
3870 | /* |
3871 | * Suspend an intent log. While in suspended mode, we still honor | |
3872 | * synchronous semantics, but we rely on txg_wait_synced() to do it. | |
13fe0198 MA |
3873 | * On old version pools, we suspend the log briefly when taking a |
3874 | * snapshot so that it will have an empty intent log. | |
3875 | * | |
3876 | * Long holds are not really intended to be used the way we do here -- | |
3877 | * held for such a short time. A concurrent caller of dsl_dataset_long_held() | |
3878 | * could fail. Therefore we take pains to only put a long hold if it is | |
3879 | * actually necessary. Fortunately, it will only be necessary if the | |
3880 | * objset is currently mounted (or the ZVOL equivalent). In that case it | |
3881 | * will already have a long hold, so we are not really making things any worse. | |
3882 | * | |
3883 | * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or | |
3884 | * zvol_state_t), and use their mechanism to prevent their hold from being | |
3885 | * dropped (e.g. VFS_HOLD()). However, that would be even more pain for | |
3886 | * very little gain. | |
3887 | * | |
3888 | * if cookiep == NULL, this does both the suspend & resume. | |
3889 | * Otherwise, it returns with the dataset "long held", and the cookie | |
3890 | * should be passed into zil_resume(). | |
34dc7c2f BB |
3891 | */ |
3892 | int | |
13fe0198 | 3893 | zil_suspend(const char *osname, void **cookiep) |
34dc7c2f | 3894 | { |
13fe0198 MA |
3895 | objset_t *os; |
3896 | zilog_t *zilog; | |
3897 | const zil_header_t *zh; | |
3898 | int error; | |
3899 | ||
3900 | error = dmu_objset_hold(osname, suspend_tag, &os); | |
3901 | if (error != 0) | |
3902 | return (error); | |
3903 | zilog = dmu_objset_zil(os); | |
34dc7c2f BB |
3904 | |
3905 | mutex_enter(&zilog->zl_lock); | |
13fe0198 MA |
3906 | zh = zilog->zl_header; |
3907 | ||
9babb374 | 3908 | if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ |
34dc7c2f | 3909 | mutex_exit(&zilog->zl_lock); |
13fe0198 | 3910 | dmu_objset_rele(os, suspend_tag); |
2e528b49 | 3911 | return (SET_ERROR(EBUSY)); |
34dc7c2f | 3912 | } |
13fe0198 MA |
3913 | |
3914 | /* | |
3915 | * Don't put a long hold in the cases where we can avoid it. This | |
3916 | * is when there is no cookie so we are doing a suspend & resume | |
3917 | * (i.e. called from zil_vdev_offline()), and there's nothing to do | |
3918 | * for the suspend because it's already suspended, or there's no ZIL. | |
3919 | */ | |
3920 | if (cookiep == NULL && !zilog->zl_suspending && | |
3921 | (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { | |
3922 | mutex_exit(&zilog->zl_lock); | |
3923 | dmu_objset_rele(os, suspend_tag); | |
3924 | return (0); | |
3925 | } | |
3926 | ||
3927 | dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); | |
3928 | dsl_pool_rele(dmu_objset_pool(os), suspend_tag); | |
3929 | ||
3930 | zilog->zl_suspend++; | |
3931 | ||
3932 | if (zilog->zl_suspend > 1) { | |
34dc7c2f | 3933 | /* |
13fe0198 | 3934 | * Someone else is already suspending it. |
34dc7c2f BB |
3935 | * Just wait for them to finish. |
3936 | */ | |
13fe0198 | 3937 | |
34dc7c2f BB |
3938 | while (zilog->zl_suspending) |
3939 | cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); | |
34dc7c2f | 3940 | mutex_exit(&zilog->zl_lock); |
13fe0198 MA |
3941 | |
3942 | if (cookiep == NULL) | |
3943 | zil_resume(os); | |
3944 | else | |
3945 | *cookiep = os; | |
3946 | return (0); | |
3947 | } | |
3948 | ||
3949 | /* | |
3950 | * If there is no pointer to an on-disk block, this ZIL must not | |
3951 | * be active (e.g. filesystem not mounted), so there's nothing | |
3952 | * to clean up. | |
3953 | */ | |
3954 | if (BP_IS_HOLE(&zh->zh_log)) { | |
3955 | ASSERT(cookiep != NULL); /* fast path already handled */ | |
3956 | ||
3957 | *cookiep = os; | |
3958 | mutex_exit(&zilog->zl_lock); | |
34dc7c2f BB |
3959 | return (0); |
3960 | } | |
13fe0198 | 3961 | |
4807c0ba TC |
3962 | /* |
3963 | * The ZIL has work to do. Ensure that the associated encryption | |
3964 | * key will remain mapped while we are committing the log by | |
3965 | * grabbing a reference to it. If the key isn't loaded we have no | |
3966 | * choice but to return an error until the wrapping key is loaded. | |
3967 | */ | |
52ce99dd TC |
3968 | if (os->os_encrypted && |
3969 | dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) { | |
4807c0ba TC |
3970 | zilog->zl_suspend--; |
3971 | mutex_exit(&zilog->zl_lock); | |
3972 | dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); | |
3973 | dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); | |
2ffd89fc | 3974 | return (SET_ERROR(EACCES)); |
4807c0ba TC |
3975 | } |
3976 | ||
34dc7c2f BB |
3977 | zilog->zl_suspending = B_TRUE; |
3978 | mutex_exit(&zilog->zl_lock); | |
3979 | ||
2fe61a7e PS |
3980 | /* |
3981 | * We need to use zil_commit_impl to ensure we wait for all | |
eda3fcd5 | 3982 | * LWB_STATE_OPENED, _CLOSED and _READY lwbs to be committed |
2fe61a7e PS |
3983 | * to disk before proceeding. If we used zil_commit instead, it |
3984 | * would just call txg_wait_synced(), because zl_suspend is set. | |
3985 | * txg_wait_synced() doesn't wait for these lwb's to be | |
900d09b2 | 3986 | * LWB_STATE_FLUSH_DONE before returning. |
2fe61a7e PS |
3987 | */ |
3988 | zil_commit_impl(zilog, 0); | |
3989 | ||
3990 | /* | |
900d09b2 PS |
3991 | * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we |
3992 | * use txg_wait_synced() to ensure the data from the zilog has | |
3993 | * migrated to the main pool before calling zil_destroy(). | |
2fe61a7e PS |
3994 | */ |
3995 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
34dc7c2f BB |
3996 | |
3997 | zil_destroy(zilog, B_FALSE); | |
3998 | ||
3999 | mutex_enter(&zilog->zl_lock); | |
4000 | zilog->zl_suspending = B_FALSE; | |
4001 | cv_broadcast(&zilog->zl_cv_suspend); | |
4002 | mutex_exit(&zilog->zl_lock); | |
4003 | ||
52ce99dd TC |
4004 | if (os->os_encrypted) |
4005 | dsl_dataset_remove_key_mapping(dmu_objset_ds(os)); | |
4807c0ba | 4006 | |
13fe0198 MA |
4007 | if (cookiep == NULL) |
4008 | zil_resume(os); | |
4009 | else | |
4010 | *cookiep = os; | |
34dc7c2f BB |
4011 | return (0); |
4012 | } | |
4013 | ||
4014 | void | |
13fe0198 | 4015 | zil_resume(void *cookie) |
34dc7c2f | 4016 | { |
13fe0198 MA |
4017 | objset_t *os = cookie; |
4018 | zilog_t *zilog = dmu_objset_zil(os); | |
4019 | ||
34dc7c2f BB |
4020 | mutex_enter(&zilog->zl_lock); |
4021 | ASSERT(zilog->zl_suspend != 0); | |
4022 | zilog->zl_suspend--; | |
4023 | mutex_exit(&zilog->zl_lock); | |
13fe0198 MA |
4024 | dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); |
4025 | dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); | |
34dc7c2f BB |
4026 | } |
4027 | ||
4028 | typedef struct zil_replay_arg { | |
18168da7 | 4029 | zil_replay_func_t *const *zr_replay; |
34dc7c2f | 4030 | void *zr_arg; |
34dc7c2f | 4031 | boolean_t zr_byteswap; |
428870ff | 4032 | char *zr_lr; |
34dc7c2f BB |
4033 | } zil_replay_arg_t; |
4034 | ||
428870ff | 4035 | static int |
61868bb1 | 4036 | zil_replay_error(zilog_t *zilog, const lr_t *lr, int error) |
428870ff | 4037 | { |
eca7b760 | 4038 | char name[ZFS_MAX_DATASET_NAME_LEN]; |
428870ff BB |
4039 | |
4040 | zilog->zl_replaying_seq--; /* didn't actually replay this one */ | |
4041 | ||
4042 | dmu_objset_name(zilog->zl_os, name); | |
4043 | ||
4044 | cmn_err(CE_WARN, "ZFS replay transaction error %d, " | |
4045 | "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, | |
4046 | (u_longlong_t)lr->lrc_seq, | |
4047 | (u_longlong_t)(lr->lrc_txtype & ~TX_CI), | |
4048 | (lr->lrc_txtype & TX_CI) ? "CI" : ""); | |
4049 | ||
4050 | return (error); | |
4051 | } | |
4052 | ||
4053 | static int | |
61868bb1 CS |
4054 | zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra, |
4055 | uint64_t claim_txg) | |
34dc7c2f BB |
4056 | { |
4057 | zil_replay_arg_t *zr = zra; | |
4058 | const zil_header_t *zh = zilog->zl_header; | |
4059 | uint64_t reclen = lr->lrc_reclen; | |
4060 | uint64_t txtype = lr->lrc_txtype; | |
428870ff | 4061 | int error = 0; |
34dc7c2f | 4062 | |
428870ff | 4063 | zilog->zl_replaying_seq = lr->lrc_seq; |
34dc7c2f BB |
4064 | |
4065 | if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ | |
428870ff BB |
4066 | return (0); |
4067 | ||
4068 | if (lr->lrc_txg < claim_txg) /* already committed */ | |
4069 | return (0); | |
34dc7c2f BB |
4070 | |
4071 | /* Strip case-insensitive bit, still present in log record */ | |
4072 | txtype &= ~TX_CI; | |
4073 | ||
428870ff BB |
4074 | if (txtype == 0 || txtype >= TX_MAX_TYPE) |
4075 | return (zil_replay_error(zilog, lr, EINVAL)); | |
4076 | ||
4077 | /* | |
4078 | * If this record type can be logged out of order, the object | |
4079 | * (lr_foid) may no longer exist. That's legitimate, not an error. | |
4080 | */ | |
4081 | if (TX_OOO(txtype)) { | |
4082 | error = dmu_object_info(zilog->zl_os, | |
50c957f7 | 4083 | LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL); |
428870ff BB |
4084 | if (error == ENOENT || error == EEXIST) |
4085 | return (0); | |
fb5f0bc8 BB |
4086 | } |
4087 | ||
34dc7c2f BB |
4088 | /* |
4089 | * Make a copy of the data so we can revise and extend it. | |
4090 | */ | |
861166b0 | 4091 | memcpy(zr->zr_lr, lr, reclen); |
428870ff BB |
4092 | |
4093 | /* | |
4094 | * If this is a TX_WRITE with a blkptr, suck in the data. | |
4095 | */ | |
4096 | if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { | |
4097 | error = zil_read_log_data(zilog, (lr_write_t *)lr, | |
4098 | zr->zr_lr + reclen); | |
13fe0198 | 4099 | if (error != 0) |
428870ff BB |
4100 | return (zil_replay_error(zilog, lr, error)); |
4101 | } | |
34dc7c2f BB |
4102 | |
4103 | /* | |
4104 | * The log block containing this lr may have been byteswapped | |
4105 | * so that we can easily examine common fields like lrc_txtype. | |
428870ff | 4106 | * However, the log is a mix of different record types, and only the |
34dc7c2f BB |
4107 | * replay vectors know how to byteswap their records. Therefore, if |
4108 | * the lr was byteswapped, undo it before invoking the replay vector. | |
4109 | */ | |
4110 | if (zr->zr_byteswap) | |
428870ff | 4111 | byteswap_uint64_array(zr->zr_lr, reclen); |
34dc7c2f BB |
4112 | |
4113 | /* | |
4114 | * We must now do two things atomically: replay this log record, | |
fb5f0bc8 BB |
4115 | * and update the log header sequence number to reflect the fact that |
4116 | * we did so. At the end of each replay function the sequence number | |
4117 | * is updated if we are in replay mode. | |
34dc7c2f | 4118 | */ |
428870ff | 4119 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); |
13fe0198 | 4120 | if (error != 0) { |
34dc7c2f BB |
4121 | /* |
4122 | * The DMU's dnode layer doesn't see removes until the txg | |
4123 | * commits, so a subsequent claim can spuriously fail with | |
fb5f0bc8 | 4124 | * EEXIST. So if we receive any error we try syncing out |
428870ff BB |
4125 | * any removes then retry the transaction. Note that we |
4126 | * specify B_FALSE for byteswap now, so we don't do it twice. | |
34dc7c2f | 4127 | */ |
428870ff BB |
4128 | txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); |
4129 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); | |
13fe0198 | 4130 | if (error != 0) |
428870ff | 4131 | return (zil_replay_error(zilog, lr, error)); |
34dc7c2f | 4132 | } |
428870ff | 4133 | return (0); |
34dc7c2f BB |
4134 | } |
4135 | ||
428870ff | 4136 | static int |
61868bb1 | 4137 | zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg) |
34dc7c2f | 4138 | { |
14e4e3cb AZ |
4139 | (void) bp, (void) arg, (void) claim_txg; |
4140 | ||
34dc7c2f | 4141 | zilog->zl_replay_blks++; |
428870ff BB |
4142 | |
4143 | return (0); | |
34dc7c2f BB |
4144 | } |
4145 | ||
4146 | /* | |
4147 | * If this dataset has a non-empty intent log, replay it and destroy it. | |
e197bb24 | 4148 | * Return B_TRUE if there were any entries to replay. |
34dc7c2f | 4149 | */ |
e197bb24 | 4150 | boolean_t |
18168da7 AZ |
4151 | zil_replay(objset_t *os, void *arg, |
4152 | zil_replay_func_t *const replay_func[TX_MAX_TYPE]) | |
34dc7c2f BB |
4153 | { |
4154 | zilog_t *zilog = dmu_objset_zil(os); | |
4155 | const zil_header_t *zh = zilog->zl_header; | |
4156 | zil_replay_arg_t zr; | |
4157 | ||
9babb374 | 4158 | if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { |
e197bb24 | 4159 | return (zil_destroy(zilog, B_TRUE)); |
34dc7c2f BB |
4160 | } |
4161 | ||
34dc7c2f BB |
4162 | zr.zr_replay = replay_func; |
4163 | zr.zr_arg = arg; | |
34dc7c2f | 4164 | zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); |
79c76d5b | 4165 | zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); |
34dc7c2f BB |
4166 | |
4167 | /* | |
4168 | * Wait for in-progress removes to sync before starting replay. | |
4169 | */ | |
4170 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
4171 | ||
fb5f0bc8 | 4172 | zilog->zl_replay = B_TRUE; |
428870ff | 4173 | zilog->zl_replay_time = ddi_get_lbolt(); |
34dc7c2f BB |
4174 | ASSERT(zilog->zl_replay_blks == 0); |
4175 | (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, | |
b5256303 | 4176 | zh->zh_claim_txg, B_TRUE); |
00b46022 | 4177 | vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); |
34dc7c2f BB |
4178 | |
4179 | zil_destroy(zilog, B_FALSE); | |
4180 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
fb5f0bc8 | 4181 | zilog->zl_replay = B_FALSE; |
e197bb24 AS |
4182 | |
4183 | return (B_TRUE); | |
34dc7c2f BB |
4184 | } |
4185 | ||
428870ff BB |
4186 | boolean_t |
4187 | zil_replaying(zilog_t *zilog, dmu_tx_t *tx) | |
34dc7c2f | 4188 | { |
428870ff BB |
4189 | if (zilog->zl_sync == ZFS_SYNC_DISABLED) |
4190 | return (B_TRUE); | |
34dc7c2f | 4191 | |
428870ff BB |
4192 | if (zilog->zl_replay) { |
4193 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
4194 | zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = | |
4195 | zilog->zl_replaying_seq; | |
4196 | return (B_TRUE); | |
34dc7c2f BB |
4197 | } |
4198 | ||
428870ff | 4199 | return (B_FALSE); |
34dc7c2f | 4200 | } |
9babb374 | 4201 | |
9babb374 | 4202 | int |
a1d477c2 | 4203 | zil_reset(const char *osname, void *arg) |
9babb374 | 4204 | { |
14e4e3cb | 4205 | (void) arg; |
9babb374 | 4206 | |
14e4e3cb | 4207 | int error = zil_suspend(osname, NULL); |
2ffd89fc PZ |
4208 | /* EACCES means crypto key not loaded */ |
4209 | if ((error == EACCES) || (error == EBUSY)) | |
4210 | return (SET_ERROR(error)); | |
13fe0198 | 4211 | if (error != 0) |
2e528b49 | 4212 | return (SET_ERROR(EEXIST)); |
13fe0198 | 4213 | return (0); |
9babb374 | 4214 | } |
c409e464 | 4215 | |
0f699108 AZ |
4216 | EXPORT_SYMBOL(zil_alloc); |
4217 | EXPORT_SYMBOL(zil_free); | |
4218 | EXPORT_SYMBOL(zil_open); | |
4219 | EXPORT_SYMBOL(zil_close); | |
4220 | EXPORT_SYMBOL(zil_replay); | |
4221 | EXPORT_SYMBOL(zil_replaying); | |
4222 | EXPORT_SYMBOL(zil_destroy); | |
4223 | EXPORT_SYMBOL(zil_destroy_sync); | |
4224 | EXPORT_SYMBOL(zil_itx_create); | |
4225 | EXPORT_SYMBOL(zil_itx_destroy); | |
4226 | EXPORT_SYMBOL(zil_itx_assign); | |
4227 | EXPORT_SYMBOL(zil_commit); | |
0f699108 AZ |
4228 | EXPORT_SYMBOL(zil_claim); |
4229 | EXPORT_SYMBOL(zil_check_log_chain); | |
4230 | EXPORT_SYMBOL(zil_sync); | |
4231 | EXPORT_SYMBOL(zil_clean); | |
4232 | EXPORT_SYMBOL(zil_suspend); | |
4233 | EXPORT_SYMBOL(zil_resume); | |
1ce23dca | 4234 | EXPORT_SYMBOL(zil_lwb_add_block); |
0f699108 AZ |
4235 | EXPORT_SYMBOL(zil_bp_tree_add); |
4236 | EXPORT_SYMBOL(zil_set_sync); | |
4237 | EXPORT_SYMBOL(zil_set_logbias); | |
fb087146 AH |
4238 | EXPORT_SYMBOL(zil_sums_init); |
4239 | EXPORT_SYMBOL(zil_sums_fini); | |
4240 | EXPORT_SYMBOL(zil_kstat_values_update); | |
0f699108 | 4241 | |
fdc2d303 | 4242 | ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW, |
03fdcb9a | 4243 | "ZIL block open timeout percentage"); |
2fe61a7e | 4244 | |
03fdcb9a MM |
4245 | ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW, |
4246 | "Disable intent logging replay"); | |
c409e464 | 4247 | |
03fdcb9a MM |
4248 | ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW, |
4249 | "Disable ZIL cache flushes"); | |
ee191e80 | 4250 | |
ab8d9c17 | 4251 | ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, U64, ZMOD_RW, |
03fdcb9a | 4252 | "Limit in bytes slog sync writes per commit"); |
b8738257 | 4253 | |
fdc2d303 | 4254 | ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW, |
03fdcb9a | 4255 | "Limit in bytes of ZIL log block size"); |
66b81b34 AM |
4256 | |
4257 | ZFS_MODULE_PARAM(zfs_zil, zil_, maxcopied, UINT, ZMOD_RW, | |
4258 | "Limit in bytes WR_COPIED size"); |