]>
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 | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
2e528b49 | 23 | * Copyright (c) 2013 by Delphix. All rights reserved. |
34dc7c2f BB |
24 | */ |
25 | ||
428870ff BB |
26 | /* Portions Copyright 2010 Robert Milkowski */ |
27 | ||
34dc7c2f BB |
28 | #include <sys/zfs_context.h> |
29 | #include <sys/spa.h> | |
30 | #include <sys/dmu.h> | |
31 | #include <sys/zap.h> | |
32 | #include <sys/arc.h> | |
33 | #include <sys/stat.h> | |
34 | #include <sys/resource.h> | |
35 | #include <sys/zil.h> | |
36 | #include <sys/zil_impl.h> | |
37 | #include <sys/dsl_dataset.h> | |
572e2857 | 38 | #include <sys/vdev_impl.h> |
34dc7c2f | 39 | #include <sys/dmu_tx.h> |
428870ff | 40 | #include <sys/dsl_pool.h> |
920dd524 | 41 | #include <sys/metaslab.h> |
34dc7c2f BB |
42 | |
43 | /* | |
44 | * The zfs intent log (ZIL) saves transaction records of system calls | |
45 | * that change the file system in memory with enough information | |
46 | * to be able to replay them. These are stored in memory until | |
47 | * either the DMU transaction group (txg) commits them to the stable pool | |
48 | * and they can be discarded, or they are flushed to the stable log | |
49 | * (also in the pool) due to a fsync, O_DSYNC or other synchronous | |
50 | * requirement. In the event of a panic or power fail then those log | |
51 | * records (transactions) are replayed. | |
52 | * | |
53 | * There is one ZIL per file system. Its on-disk (pool) format consists | |
54 | * of 3 parts: | |
55 | * | |
56 | * - ZIL header | |
57 | * - ZIL blocks | |
58 | * - ZIL records | |
59 | * | |
60 | * A log record holds a system call transaction. Log blocks can | |
61 | * hold many log records and the blocks are chained together. | |
62 | * Each ZIL block contains a block pointer (blkptr_t) to the next | |
63 | * ZIL block in the chain. The ZIL header points to the first | |
64 | * block in the chain. Note there is not a fixed place in the pool | |
65 | * to hold blocks. They are dynamically allocated and freed as | |
66 | * needed from the blocks available. Figure X shows the ZIL structure: | |
67 | */ | |
68 | ||
b6ad9671 ED |
69 | /* |
70 | * See zil.h for more information about these fields. | |
71 | */ | |
72 | zil_stats_t zil_stats = { | |
d1d7e268 MK |
73 | { "zil_commit_count", KSTAT_DATA_UINT64 }, |
74 | { "zil_commit_writer_count", KSTAT_DATA_UINT64 }, | |
75 | { "zil_itx_count", KSTAT_DATA_UINT64 }, | |
76 | { "zil_itx_indirect_count", KSTAT_DATA_UINT64 }, | |
77 | { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 }, | |
78 | { "zil_itx_copied_count", KSTAT_DATA_UINT64 }, | |
79 | { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 }, | |
80 | { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 }, | |
81 | { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 }, | |
82 | { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 }, | |
83 | { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 }, | |
84 | { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 }, | |
85 | { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 }, | |
b6ad9671 ED |
86 | }; |
87 | ||
88 | static kstat_t *zil_ksp; | |
89 | ||
34dc7c2f | 90 | /* |
d3cc8b15 | 91 | * Disable intent logging replay. This global ZIL switch affects all pools. |
34dc7c2f | 92 | */ |
d3cc8b15 | 93 | int zil_replay_disable = 0; |
34dc7c2f BB |
94 | |
95 | /* | |
96 | * Tunable parameter for debugging or performance analysis. Setting | |
97 | * zfs_nocacheflush will cause corruption on power loss if a volatile | |
98 | * out-of-order write cache is enabled. | |
99 | */ | |
c409e464 | 100 | int zfs_nocacheflush = 0; |
34dc7c2f BB |
101 | |
102 | static kmem_cache_t *zil_lwb_cache; | |
103 | ||
572e2857 | 104 | static void zil_async_to_sync(zilog_t *zilog, uint64_t foid); |
428870ff BB |
105 | |
106 | #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ | |
107 | sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) | |
108 | ||
109 | ||
572e2857 BB |
110 | /* |
111 | * ziltest is by and large an ugly hack, but very useful in | |
112 | * checking replay without tedious work. | |
113 | * When running ziltest we want to keep all itx's and so maintain | |
114 | * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG | |
115 | * We subtract TXG_CONCURRENT_STATES to allow for common code. | |
116 | */ | |
117 | #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) | |
118 | ||
34dc7c2f | 119 | static int |
428870ff | 120 | zil_bp_compare(const void *x1, const void *x2) |
34dc7c2f | 121 | { |
428870ff BB |
122 | const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; |
123 | const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; | |
34dc7c2f BB |
124 | |
125 | if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) | |
126 | return (-1); | |
127 | if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) | |
128 | return (1); | |
129 | ||
130 | if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) | |
131 | return (-1); | |
132 | if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) | |
133 | return (1); | |
134 | ||
135 | return (0); | |
136 | } | |
137 | ||
138 | static void | |
428870ff | 139 | zil_bp_tree_init(zilog_t *zilog) |
34dc7c2f | 140 | { |
428870ff BB |
141 | avl_create(&zilog->zl_bp_tree, zil_bp_compare, |
142 | sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); | |
34dc7c2f BB |
143 | } |
144 | ||
145 | static void | |
428870ff | 146 | zil_bp_tree_fini(zilog_t *zilog) |
34dc7c2f | 147 | { |
428870ff BB |
148 | avl_tree_t *t = &zilog->zl_bp_tree; |
149 | zil_bp_node_t *zn; | |
34dc7c2f BB |
150 | void *cookie = NULL; |
151 | ||
152 | while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) | |
428870ff | 153 | kmem_free(zn, sizeof (zil_bp_node_t)); |
34dc7c2f BB |
154 | |
155 | avl_destroy(t); | |
156 | } | |
157 | ||
428870ff BB |
158 | int |
159 | zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) | |
34dc7c2f | 160 | { |
428870ff | 161 | avl_tree_t *t = &zilog->zl_bp_tree; |
9b67f605 | 162 | const dva_t *dva; |
428870ff | 163 | zil_bp_node_t *zn; |
34dc7c2f BB |
164 | avl_index_t where; |
165 | ||
9b67f605 MA |
166 | if (BP_IS_EMBEDDED(bp)) |
167 | return (0); | |
168 | ||
169 | dva = BP_IDENTITY(bp); | |
170 | ||
34dc7c2f | 171 | if (avl_find(t, dva, &where) != NULL) |
2e528b49 | 172 | return (SET_ERROR(EEXIST)); |
34dc7c2f | 173 | |
b8d06fca | 174 | zn = kmem_alloc(sizeof (zil_bp_node_t), KM_PUSHPAGE); |
34dc7c2f BB |
175 | zn->zn_dva = *dva; |
176 | avl_insert(t, zn, where); | |
177 | ||
178 | return (0); | |
179 | } | |
180 | ||
181 | static zil_header_t * | |
182 | zil_header_in_syncing_context(zilog_t *zilog) | |
183 | { | |
184 | return ((zil_header_t *)zilog->zl_header); | |
185 | } | |
186 | ||
187 | static void | |
188 | zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) | |
189 | { | |
190 | zio_cksum_t *zc = &bp->blk_cksum; | |
191 | ||
192 | zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); | |
193 | zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); | |
194 | zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); | |
195 | zc->zc_word[ZIL_ZC_SEQ] = 1ULL; | |
196 | } | |
197 | ||
198 | /* | |
428870ff | 199 | * Read a log block and make sure it's valid. |
34dc7c2f BB |
200 | */ |
201 | static int | |
428870ff BB |
202 | zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, |
203 | char **end) | |
34dc7c2f | 204 | { |
428870ff | 205 | enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; |
34dc7c2f | 206 | uint32_t aflags = ARC_WAIT; |
428870ff BB |
207 | arc_buf_t *abuf = NULL; |
208 | zbookmark_t zb; | |
34dc7c2f BB |
209 | int error; |
210 | ||
428870ff BB |
211 | if (zilog->zl_header->zh_claim_txg == 0) |
212 | zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; | |
34dc7c2f | 213 | |
428870ff BB |
214 | if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) |
215 | zio_flags |= ZIO_FLAG_SPECULATIVE; | |
34dc7c2f | 216 | |
428870ff BB |
217 | SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], |
218 | ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
219 | ||
294f6806 | 220 | error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, |
428870ff | 221 | ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); |
34dc7c2f BB |
222 | |
223 | if (error == 0) { | |
34dc7c2f BB |
224 | zio_cksum_t cksum = bp->blk_cksum; |
225 | ||
226 | /* | |
b128c09f BB |
227 | * Validate the checksummed log block. |
228 | * | |
34dc7c2f BB |
229 | * Sequence numbers should be... sequential. The checksum |
230 | * verifier for the next block should be bp's checksum plus 1. | |
b128c09f BB |
231 | * |
232 | * Also check the log chain linkage and size used. | |
34dc7c2f BB |
233 | */ |
234 | cksum.zc_word[ZIL_ZC_SEQ]++; | |
235 | ||
428870ff BB |
236 | if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { |
237 | zil_chain_t *zilc = abuf->b_data; | |
238 | char *lr = (char *)(zilc + 1); | |
239 | uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); | |
34dc7c2f | 240 | |
428870ff BB |
241 | if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, |
242 | sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { | |
2e528b49 | 243 | error = SET_ERROR(ECKSUM); |
428870ff BB |
244 | } else { |
245 | bcopy(lr, dst, len); | |
246 | *end = (char *)dst + len; | |
247 | *nbp = zilc->zc_next_blk; | |
248 | } | |
249 | } else { | |
250 | char *lr = abuf->b_data; | |
251 | uint64_t size = BP_GET_LSIZE(bp); | |
252 | zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; | |
253 | ||
254 | if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, | |
255 | sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || | |
256 | (zilc->zc_nused > (size - sizeof (*zilc)))) { | |
2e528b49 | 257 | error = SET_ERROR(ECKSUM); |
428870ff BB |
258 | } else { |
259 | bcopy(lr, dst, zilc->zc_nused); | |
260 | *end = (char *)dst + zilc->zc_nused; | |
261 | *nbp = zilc->zc_next_blk; | |
262 | } | |
34dc7c2f | 263 | } |
428870ff | 264 | |
13fe0198 | 265 | VERIFY(arc_buf_remove_ref(abuf, &abuf)); |
428870ff BB |
266 | } |
267 | ||
268 | return (error); | |
269 | } | |
270 | ||
271 | /* | |
272 | * Read a TX_WRITE log data block. | |
273 | */ | |
274 | static int | |
275 | zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) | |
276 | { | |
277 | enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; | |
278 | const blkptr_t *bp = &lr->lr_blkptr; | |
279 | uint32_t aflags = ARC_WAIT; | |
280 | arc_buf_t *abuf = NULL; | |
281 | zbookmark_t zb; | |
282 | int error; | |
283 | ||
284 | if (BP_IS_HOLE(bp)) { | |
285 | if (wbuf != NULL) | |
286 | bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); | |
287 | return (0); | |
34dc7c2f BB |
288 | } |
289 | ||
428870ff BB |
290 | if (zilog->zl_header->zh_claim_txg == 0) |
291 | zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; | |
292 | ||
293 | SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, | |
294 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
295 | ||
294f6806 | 296 | error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, |
428870ff BB |
297 | ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); |
298 | ||
299 | if (error == 0) { | |
300 | if (wbuf != NULL) | |
301 | bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); | |
302 | (void) arc_buf_remove_ref(abuf, &abuf); | |
303 | } | |
34dc7c2f BB |
304 | |
305 | return (error); | |
306 | } | |
307 | ||
308 | /* | |
309 | * Parse the intent log, and call parse_func for each valid record within. | |
34dc7c2f | 310 | */ |
428870ff | 311 | int |
34dc7c2f BB |
312 | zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, |
313 | zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) | |
314 | { | |
315 | const zil_header_t *zh = zilog->zl_header; | |
428870ff BB |
316 | boolean_t claimed = !!zh->zh_claim_txg; |
317 | uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; | |
318 | uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; | |
319 | uint64_t max_blk_seq = 0; | |
320 | uint64_t max_lr_seq = 0; | |
321 | uint64_t blk_count = 0; | |
322 | uint64_t lr_count = 0; | |
323 | blkptr_t blk, next_blk; | |
34dc7c2f | 324 | char *lrbuf, *lrp; |
428870ff | 325 | int error = 0; |
34dc7c2f | 326 | |
d1d7e268 | 327 | bzero(&next_blk, sizeof (blkptr_t)); |
d4ed6673 | 328 | |
428870ff BB |
329 | /* |
330 | * Old logs didn't record the maximum zh_claim_lr_seq. | |
331 | */ | |
332 | if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) | |
333 | claim_lr_seq = UINT64_MAX; | |
34dc7c2f BB |
334 | |
335 | /* | |
336 | * Starting at the block pointed to by zh_log we read the log chain. | |
337 | * For each block in the chain we strongly check that block to | |
338 | * ensure its validity. We stop when an invalid block is found. | |
339 | * For each block pointer in the chain we call parse_blk_func(). | |
340 | * For each record in each valid block we call parse_lr_func(). | |
341 | * If the log has been claimed, stop if we encounter a sequence | |
342 | * number greater than the highest claimed sequence number. | |
343 | */ | |
428870ff BB |
344 | lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); |
345 | zil_bp_tree_init(zilog); | |
34dc7c2f | 346 | |
428870ff BB |
347 | for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { |
348 | uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; | |
349 | int reclen; | |
d4ed6673 | 350 | char *end = NULL; |
34dc7c2f | 351 | |
428870ff BB |
352 | if (blk_seq > claim_blk_seq) |
353 | break; | |
354 | if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) | |
355 | break; | |
356 | ASSERT3U(max_blk_seq, <, blk_seq); | |
357 | max_blk_seq = blk_seq; | |
358 | blk_count++; | |
34dc7c2f | 359 | |
428870ff BB |
360 | if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) |
361 | break; | |
34dc7c2f | 362 | |
428870ff | 363 | error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); |
13fe0198 | 364 | if (error != 0) |
34dc7c2f BB |
365 | break; |
366 | ||
428870ff | 367 | for (lrp = lrbuf; lrp < end; lrp += reclen) { |
34dc7c2f BB |
368 | lr_t *lr = (lr_t *)lrp; |
369 | reclen = lr->lrc_reclen; | |
370 | ASSERT3U(reclen, >=, sizeof (lr_t)); | |
428870ff BB |
371 | if (lr->lrc_seq > claim_lr_seq) |
372 | goto done; | |
373 | if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) | |
374 | goto done; | |
375 | ASSERT3U(max_lr_seq, <, lr->lrc_seq); | |
376 | max_lr_seq = lr->lrc_seq; | |
377 | lr_count++; | |
34dc7c2f | 378 | } |
34dc7c2f | 379 | } |
428870ff BB |
380 | done: |
381 | zilog->zl_parse_error = error; | |
382 | zilog->zl_parse_blk_seq = max_blk_seq; | |
383 | zilog->zl_parse_lr_seq = max_lr_seq; | |
384 | zilog->zl_parse_blk_count = blk_count; | |
385 | zilog->zl_parse_lr_count = lr_count; | |
386 | ||
387 | ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || | |
388 | (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); | |
389 | ||
390 | zil_bp_tree_fini(zilog); | |
391 | zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); | |
34dc7c2f | 392 | |
428870ff | 393 | return (error); |
34dc7c2f BB |
394 | } |
395 | ||
428870ff | 396 | static int |
34dc7c2f BB |
397 | zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) |
398 | { | |
34dc7c2f BB |
399 | /* |
400 | * Claim log block if not already committed and not already claimed. | |
428870ff | 401 | * If tx == NULL, just verify that the block is claimable. |
34dc7c2f | 402 | */ |
b0bc7a84 MG |
403 | if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || |
404 | zil_bp_tree_add(zilog, bp) != 0) | |
428870ff BB |
405 | return (0); |
406 | ||
407 | return (zio_wait(zio_claim(NULL, zilog->zl_spa, | |
408 | tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, | |
409 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); | |
34dc7c2f BB |
410 | } |
411 | ||
428870ff | 412 | static int |
34dc7c2f BB |
413 | zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) |
414 | { | |
428870ff BB |
415 | lr_write_t *lr = (lr_write_t *)lrc; |
416 | int error; | |
417 | ||
418 | if (lrc->lrc_txtype != TX_WRITE) | |
419 | return (0); | |
420 | ||
421 | /* | |
422 | * If the block is not readable, don't claim it. This can happen | |
423 | * in normal operation when a log block is written to disk before | |
424 | * some of the dmu_sync() blocks it points to. In this case, the | |
425 | * transaction cannot have been committed to anyone (we would have | |
426 | * waited for all writes to be stable first), so it is semantically | |
427 | * correct to declare this the end of the log. | |
428 | */ | |
429 | if (lr->lr_blkptr.blk_birth >= first_txg && | |
430 | (error = zil_read_log_data(zilog, lr, NULL)) != 0) | |
431 | return (error); | |
432 | return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); | |
34dc7c2f BB |
433 | } |
434 | ||
435 | /* ARGSUSED */ | |
428870ff | 436 | static int |
34dc7c2f BB |
437 | zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) |
438 | { | |
428870ff BB |
439 | zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); |
440 | ||
441 | return (0); | |
34dc7c2f BB |
442 | } |
443 | ||
428870ff | 444 | static int |
34dc7c2f BB |
445 | zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) |
446 | { | |
428870ff BB |
447 | lr_write_t *lr = (lr_write_t *)lrc; |
448 | blkptr_t *bp = &lr->lr_blkptr; | |
449 | ||
34dc7c2f BB |
450 | /* |
451 | * If we previously claimed it, we need to free it. | |
452 | */ | |
428870ff | 453 | if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && |
b0bc7a84 MG |
454 | bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && |
455 | !BP_IS_HOLE(bp)) | |
428870ff BB |
456 | zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); |
457 | ||
458 | return (0); | |
459 | } | |
460 | ||
461 | static lwb_t * | |
920dd524 | 462 | zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg, boolean_t fastwrite) |
428870ff BB |
463 | { |
464 | lwb_t *lwb; | |
465 | ||
b8d06fca | 466 | lwb = kmem_cache_alloc(zil_lwb_cache, KM_PUSHPAGE); |
428870ff BB |
467 | lwb->lwb_zilog = zilog; |
468 | lwb->lwb_blk = *bp; | |
920dd524 | 469 | lwb->lwb_fastwrite = fastwrite; |
428870ff BB |
470 | lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); |
471 | lwb->lwb_max_txg = txg; | |
472 | lwb->lwb_zio = NULL; | |
473 | lwb->lwb_tx = NULL; | |
474 | if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { | |
475 | lwb->lwb_nused = sizeof (zil_chain_t); | |
476 | lwb->lwb_sz = BP_GET_LSIZE(bp); | |
477 | } else { | |
478 | lwb->lwb_nused = 0; | |
479 | lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); | |
34dc7c2f | 480 | } |
428870ff BB |
481 | |
482 | mutex_enter(&zilog->zl_lock); | |
483 | list_insert_tail(&zilog->zl_lwb_list, lwb); | |
484 | mutex_exit(&zilog->zl_lock); | |
485 | ||
486 | return (lwb); | |
34dc7c2f BB |
487 | } |
488 | ||
29809a6c MA |
489 | /* |
490 | * Called when we create in-memory log transactions so that we know | |
491 | * to cleanup the itxs at the end of spa_sync(). | |
492 | */ | |
493 | void | |
494 | zilog_dirty(zilog_t *zilog, uint64_t txg) | |
495 | { | |
496 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
497 | dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); | |
498 | ||
499 | if (dsl_dataset_is_snapshot(ds)) | |
500 | panic("dirtying snapshot!"); | |
501 | ||
13fe0198 | 502 | if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { |
29809a6c MA |
503 | /* up the hold count until we can be written out */ |
504 | dmu_buf_add_ref(ds->ds_dbuf, zilog); | |
505 | } | |
506 | } | |
507 | ||
508 | boolean_t | |
509 | zilog_is_dirty(zilog_t *zilog) | |
510 | { | |
511 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
512 | int t; | |
513 | ||
514 | for (t = 0; t < TXG_SIZE; t++) { | |
515 | if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) | |
516 | return (B_TRUE); | |
517 | } | |
518 | return (B_FALSE); | |
519 | } | |
520 | ||
34dc7c2f BB |
521 | /* |
522 | * Create an on-disk intent log. | |
523 | */ | |
428870ff | 524 | static lwb_t * |
34dc7c2f BB |
525 | zil_create(zilog_t *zilog) |
526 | { | |
527 | const zil_header_t *zh = zilog->zl_header; | |
428870ff | 528 | lwb_t *lwb = NULL; |
34dc7c2f BB |
529 | uint64_t txg = 0; |
530 | dmu_tx_t *tx = NULL; | |
531 | blkptr_t blk; | |
532 | int error = 0; | |
920dd524 | 533 | boolean_t fastwrite = FALSE; |
34dc7c2f BB |
534 | |
535 | /* | |
536 | * Wait for any previous destroy to complete. | |
537 | */ | |
538 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
539 | ||
540 | ASSERT(zh->zh_claim_txg == 0); | |
541 | ASSERT(zh->zh_replay_seq == 0); | |
542 | ||
543 | blk = zh->zh_log; | |
544 | ||
545 | /* | |
428870ff BB |
546 | * Allocate an initial log block if: |
547 | * - there isn't one already | |
548 | * - the existing block is the wrong endianess | |
34dc7c2f | 549 | */ |
fb5f0bc8 | 550 | if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { |
34dc7c2f | 551 | tx = dmu_tx_create(zilog->zl_os); |
428870ff | 552 | VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); |
34dc7c2f BB |
553 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); |
554 | txg = dmu_tx_get_txg(tx); | |
555 | ||
fb5f0bc8 | 556 | if (!BP_IS_HOLE(&blk)) { |
428870ff | 557 | zio_free_zil(zilog->zl_spa, txg, &blk); |
fb5f0bc8 BB |
558 | BP_ZERO(&blk); |
559 | } | |
560 | ||
920dd524 | 561 | error = zio_alloc_zil(zilog->zl_spa, txg, &blk, |
5d7a86d1 | 562 | ZIL_MIN_BLKSZ, B_TRUE); |
920dd524 | 563 | fastwrite = TRUE; |
34dc7c2f BB |
564 | |
565 | if (error == 0) | |
566 | zil_init_log_chain(zilog, &blk); | |
567 | } | |
568 | ||
569 | /* | |
570 | * Allocate a log write buffer (lwb) for the first log block. | |
571 | */ | |
428870ff | 572 | if (error == 0) |
920dd524 | 573 | lwb = zil_alloc_lwb(zilog, &blk, txg, fastwrite); |
34dc7c2f BB |
574 | |
575 | /* | |
576 | * If we just allocated the first log block, commit our transaction | |
577 | * and wait for zil_sync() to stuff the block poiner into zh_log. | |
578 | * (zh is part of the MOS, so we cannot modify it in open context.) | |
579 | */ | |
580 | if (tx != NULL) { | |
581 | dmu_tx_commit(tx); | |
582 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
583 | } | |
584 | ||
585 | ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); | |
428870ff BB |
586 | |
587 | return (lwb); | |
34dc7c2f BB |
588 | } |
589 | ||
590 | /* | |
591 | * In one tx, free all log blocks and clear the log header. | |
592 | * If keep_first is set, then we're replaying a log with no content. | |
593 | * We want to keep the first block, however, so that the first | |
594 | * synchronous transaction doesn't require a txg_wait_synced() | |
595 | * in zil_create(). We don't need to txg_wait_synced() here either | |
596 | * when keep_first is set, because both zil_create() and zil_destroy() | |
597 | * will wait for any in-progress destroys to complete. | |
598 | */ | |
599 | void | |
600 | zil_destroy(zilog_t *zilog, boolean_t keep_first) | |
601 | { | |
602 | const zil_header_t *zh = zilog->zl_header; | |
603 | lwb_t *lwb; | |
604 | dmu_tx_t *tx; | |
605 | uint64_t txg; | |
606 | ||
607 | /* | |
608 | * Wait for any previous destroy to complete. | |
609 | */ | |
610 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
611 | ||
428870ff BB |
612 | zilog->zl_old_header = *zh; /* debugging aid */ |
613 | ||
34dc7c2f BB |
614 | if (BP_IS_HOLE(&zh->zh_log)) |
615 | return; | |
616 | ||
617 | tx = dmu_tx_create(zilog->zl_os); | |
428870ff | 618 | VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); |
34dc7c2f BB |
619 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); |
620 | txg = dmu_tx_get_txg(tx); | |
621 | ||
622 | mutex_enter(&zilog->zl_lock); | |
623 | ||
34dc7c2f BB |
624 | ASSERT3U(zilog->zl_destroy_txg, <, txg); |
625 | zilog->zl_destroy_txg = txg; | |
626 | zilog->zl_keep_first = keep_first; | |
627 | ||
628 | if (!list_is_empty(&zilog->zl_lwb_list)) { | |
629 | ASSERT(zh->zh_claim_txg == 0); | |
3e31d2b0 | 630 | VERIFY(!keep_first); |
34dc7c2f | 631 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { |
920dd524 ED |
632 | ASSERT(lwb->lwb_zio == NULL); |
633 | if (lwb->lwb_fastwrite) | |
634 | metaslab_fastwrite_unmark(zilog->zl_spa, | |
635 | &lwb->lwb_blk); | |
34dc7c2f BB |
636 | list_remove(&zilog->zl_lwb_list, lwb); |
637 | if (lwb->lwb_buf != NULL) | |
638 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
428870ff | 639 | zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); |
34dc7c2f BB |
640 | kmem_cache_free(zil_lwb_cache, lwb); |
641 | } | |
428870ff | 642 | } else if (!keep_first) { |
29809a6c | 643 | zil_destroy_sync(zilog, tx); |
34dc7c2f BB |
644 | } |
645 | mutex_exit(&zilog->zl_lock); | |
646 | ||
647 | dmu_tx_commit(tx); | |
648 | } | |
649 | ||
29809a6c MA |
650 | void |
651 | zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) | |
652 | { | |
653 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
654 | (void) zil_parse(zilog, zil_free_log_block, | |
655 | zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); | |
656 | } | |
657 | ||
34dc7c2f | 658 | int |
428870ff | 659 | zil_claim(const char *osname, void *txarg) |
34dc7c2f BB |
660 | { |
661 | dmu_tx_t *tx = txarg; | |
662 | uint64_t first_txg = dmu_tx_get_txg(tx); | |
663 | zilog_t *zilog; | |
664 | zil_header_t *zh; | |
665 | objset_t *os; | |
666 | int error; | |
667 | ||
13fe0198 MA |
668 | error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); |
669 | if (error != 0) { | |
b128c09f | 670 | cmn_err(CE_WARN, "can't open objset for %s", osname); |
34dc7c2f BB |
671 | return (0); |
672 | } | |
673 | ||
674 | zilog = dmu_objset_zil(os); | |
675 | zh = zil_header_in_syncing_context(zilog); | |
676 | ||
428870ff | 677 | if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { |
9babb374 | 678 | if (!BP_IS_HOLE(&zh->zh_log)) |
428870ff | 679 | zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); |
9babb374 BB |
680 | BP_ZERO(&zh->zh_log); |
681 | dsl_dataset_dirty(dmu_objset_ds(os), tx); | |
13fe0198 | 682 | dmu_objset_disown(os, FTAG); |
428870ff | 683 | return (0); |
9babb374 BB |
684 | } |
685 | ||
34dc7c2f BB |
686 | /* |
687 | * Claim all log blocks if we haven't already done so, and remember | |
688 | * the highest claimed sequence number. This ensures that if we can | |
689 | * read only part of the log now (e.g. due to a missing device), | |
690 | * but we can read the entire log later, we will not try to replay | |
691 | * or destroy beyond the last block we successfully claimed. | |
692 | */ | |
693 | ASSERT3U(zh->zh_claim_txg, <=, first_txg); | |
694 | if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { | |
428870ff | 695 | (void) zil_parse(zilog, zil_claim_log_block, |
34dc7c2f | 696 | zil_claim_log_record, tx, first_txg); |
428870ff BB |
697 | zh->zh_claim_txg = first_txg; |
698 | zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; | |
699 | zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; | |
700 | if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) | |
701 | zh->zh_flags |= ZIL_REPLAY_NEEDED; | |
702 | zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; | |
34dc7c2f BB |
703 | dsl_dataset_dirty(dmu_objset_ds(os), tx); |
704 | } | |
705 | ||
706 | ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); | |
13fe0198 | 707 | dmu_objset_disown(os, FTAG); |
34dc7c2f BB |
708 | return (0); |
709 | } | |
710 | ||
b128c09f BB |
711 | /* |
712 | * Check the log by walking the log chain. | |
713 | * Checksum errors are ok as they indicate the end of the chain. | |
714 | * Any other error (no device or read failure) returns an error. | |
715 | */ | |
b128c09f | 716 | int |
428870ff | 717 | zil_check_log_chain(const char *osname, void *tx) |
b128c09f BB |
718 | { |
719 | zilog_t *zilog; | |
b128c09f | 720 | objset_t *os; |
572e2857 | 721 | blkptr_t *bp; |
b128c09f BB |
722 | int error; |
723 | ||
428870ff BB |
724 | ASSERT(tx == NULL); |
725 | ||
726 | error = dmu_objset_hold(osname, FTAG, &os); | |
13fe0198 | 727 | if (error != 0) { |
b128c09f BB |
728 | cmn_err(CE_WARN, "can't open objset for %s", osname); |
729 | return (0); | |
730 | } | |
731 | ||
732 | zilog = dmu_objset_zil(os); | |
572e2857 BB |
733 | bp = (blkptr_t *)&zilog->zl_header->zh_log; |
734 | ||
735 | /* | |
736 | * Check the first block and determine if it's on a log device | |
737 | * which may have been removed or faulted prior to loading this | |
738 | * pool. If so, there's no point in checking the rest of the log | |
739 | * as its content should have already been synced to the pool. | |
740 | */ | |
741 | if (!BP_IS_HOLE(bp)) { | |
742 | vdev_t *vd; | |
743 | boolean_t valid = B_TRUE; | |
744 | ||
745 | spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); | |
746 | vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); | |
747 | if (vd->vdev_islog && vdev_is_dead(vd)) | |
748 | valid = vdev_log_state_valid(vd); | |
749 | spa_config_exit(os->os_spa, SCL_STATE, FTAG); | |
750 | ||
751 | if (!valid) { | |
752 | dmu_objset_rele(os, FTAG); | |
753 | return (0); | |
754 | } | |
755 | } | |
b128c09f | 756 | |
428870ff BB |
757 | /* |
758 | * Because tx == NULL, zil_claim_log_block() will not actually claim | |
759 | * any blocks, but just determine whether it is possible to do so. | |
760 | * In addition to checking the log chain, zil_claim_log_block() | |
761 | * will invoke zio_claim() with a done func of spa_claim_notify(), | |
762 | * which will update spa_max_claim_txg. See spa_load() for details. | |
763 | */ | |
764 | error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, | |
765 | zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); | |
766 | ||
767 | dmu_objset_rele(os, FTAG); | |
768 | ||
769 | return ((error == ECKSUM || error == ENOENT) ? 0 : error); | |
b128c09f BB |
770 | } |
771 | ||
34dc7c2f BB |
772 | static int |
773 | zil_vdev_compare(const void *x1, const void *x2) | |
774 | { | |
572e2857 BB |
775 | const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; |
776 | const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; | |
34dc7c2f BB |
777 | |
778 | if (v1 < v2) | |
779 | return (-1); | |
780 | if (v1 > v2) | |
781 | return (1); | |
782 | ||
783 | return (0); | |
784 | } | |
785 | ||
786 | void | |
428870ff | 787 | zil_add_block(zilog_t *zilog, const blkptr_t *bp) |
34dc7c2f BB |
788 | { |
789 | avl_tree_t *t = &zilog->zl_vdev_tree; | |
790 | avl_index_t where; | |
791 | zil_vdev_node_t *zv, zvsearch; | |
792 | int ndvas = BP_GET_NDVAS(bp); | |
793 | int i; | |
794 | ||
795 | if (zfs_nocacheflush) | |
796 | return; | |
797 | ||
798 | ASSERT(zilog->zl_writer); | |
799 | ||
800 | /* | |
801 | * Even though we're zl_writer, we still need a lock because the | |
802 | * zl_get_data() callbacks may have dmu_sync() done callbacks | |
803 | * that will run concurrently. | |
804 | */ | |
805 | mutex_enter(&zilog->zl_vdev_lock); | |
806 | for (i = 0; i < ndvas; i++) { | |
807 | zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); | |
808 | if (avl_find(t, &zvsearch, &where) == NULL) { | |
b8d06fca | 809 | zv = kmem_alloc(sizeof (*zv), KM_PUSHPAGE); |
34dc7c2f BB |
810 | zv->zv_vdev = zvsearch.zv_vdev; |
811 | avl_insert(t, zv, where); | |
812 | } | |
813 | } | |
814 | mutex_exit(&zilog->zl_vdev_lock); | |
815 | } | |
816 | ||
572e2857 | 817 | static void |
34dc7c2f BB |
818 | zil_flush_vdevs(zilog_t *zilog) |
819 | { | |
820 | spa_t *spa = zilog->zl_spa; | |
821 | avl_tree_t *t = &zilog->zl_vdev_tree; | |
822 | void *cookie = NULL; | |
823 | zil_vdev_node_t *zv; | |
824 | zio_t *zio; | |
825 | ||
826 | ASSERT(zilog->zl_writer); | |
827 | ||
828 | /* | |
829 | * We don't need zl_vdev_lock here because we're the zl_writer, | |
830 | * and all zl_get_data() callbacks are done. | |
831 | */ | |
832 | if (avl_numnodes(t) == 0) | |
833 | return; | |
834 | ||
b128c09f | 835 | spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); |
34dc7c2f | 836 | |
b128c09f | 837 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
34dc7c2f BB |
838 | |
839 | while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { | |
840 | vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); | |
841 | if (vd != NULL) | |
842 | zio_flush(zio, vd); | |
843 | kmem_free(zv, sizeof (*zv)); | |
844 | } | |
845 | ||
846 | /* | |
847 | * Wait for all the flushes to complete. Not all devices actually | |
848 | * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. | |
849 | */ | |
850 | (void) zio_wait(zio); | |
851 | ||
b128c09f | 852 | spa_config_exit(spa, SCL_STATE, FTAG); |
34dc7c2f BB |
853 | } |
854 | ||
855 | /* | |
856 | * Function called when a log block write completes | |
857 | */ | |
858 | static void | |
859 | zil_lwb_write_done(zio_t *zio) | |
860 | { | |
861 | lwb_t *lwb = zio->io_private; | |
862 | zilog_t *zilog = lwb->lwb_zilog; | |
428870ff | 863 | dmu_tx_t *tx = lwb->lwb_tx; |
34dc7c2f | 864 | |
b128c09f | 865 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); |
b128c09f BB |
866 | ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); |
867 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
868 | ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); | |
869 | ASSERT(!BP_IS_GANG(zio->io_bp)); | |
870 | ASSERT(!BP_IS_HOLE(zio->io_bp)); | |
9b67f605 | 871 | ASSERT(BP_GET_FILL(zio->io_bp) == 0); |
b128c09f | 872 | |
34dc7c2f | 873 | /* |
9babb374 BB |
874 | * Ensure the lwb buffer pointer is cleared before releasing |
875 | * the txg. If we have had an allocation failure and | |
876 | * the txg is waiting to sync then we want want zil_sync() | |
877 | * to remove the lwb so that it's not picked up as the next new | |
878 | * one in zil_commit_writer(). zil_sync() will only remove | |
879 | * the lwb if lwb_buf is null. | |
34dc7c2f | 880 | */ |
34dc7c2f BB |
881 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); |
882 | mutex_enter(&zilog->zl_lock); | |
920dd524 ED |
883 | lwb->lwb_zio = NULL; |
884 | lwb->lwb_fastwrite = FALSE; | |
34dc7c2f | 885 | lwb->lwb_buf = NULL; |
428870ff BB |
886 | lwb->lwb_tx = NULL; |
887 | mutex_exit(&zilog->zl_lock); | |
9babb374 BB |
888 | |
889 | /* | |
890 | * Now that we've written this log block, we have a stable pointer | |
891 | * to the next block in the chain, so it's OK to let the txg in | |
428870ff | 892 | * which we allocated the next block sync. |
9babb374 | 893 | */ |
428870ff | 894 | dmu_tx_commit(tx); |
34dc7c2f BB |
895 | } |
896 | ||
897 | /* | |
898 | * Initialize the io for a log block. | |
34dc7c2f BB |
899 | */ |
900 | static void | |
901 | zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) | |
902 | { | |
903 | zbookmark_t zb; | |
904 | ||
428870ff BB |
905 | SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], |
906 | ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
907 | lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
34dc7c2f BB |
908 | |
909 | if (zilog->zl_root_zio == NULL) { | |
910 | zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, | |
911 | ZIO_FLAG_CANFAIL); | |
912 | } | |
920dd524 ED |
913 | |
914 | /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */ | |
915 | mutex_enter(&zilog->zl_lock); | |
34dc7c2f | 916 | if (lwb->lwb_zio == NULL) { |
920dd524 ED |
917 | if (!lwb->lwb_fastwrite) { |
918 | metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk); | |
919 | lwb->lwb_fastwrite = 1; | |
920 | } | |
34dc7c2f | 921 | lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, |
428870ff | 922 | 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), |
e8b96c60 | 923 | zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, |
920dd524 ED |
924 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | |
925 | ZIO_FLAG_FASTWRITE, &zb); | |
34dc7c2f | 926 | } |
920dd524 | 927 | mutex_exit(&zilog->zl_lock); |
34dc7c2f BB |
928 | } |
929 | ||
428870ff BB |
930 | /* |
931 | * Define a limited set of intent log block sizes. | |
d3cc8b15 | 932 | * |
428870ff BB |
933 | * These must be a multiple of 4KB. Note only the amount used (again |
934 | * aligned to 4KB) actually gets written. However, we can't always just | |
935 | * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. | |
936 | */ | |
937 | uint64_t zil_block_buckets[] = { | |
938 | 4096, /* non TX_WRITE */ | |
939 | 8192+4096, /* data base */ | |
940 | 32*1024 + 4096, /* NFS writes */ | |
941 | UINT64_MAX | |
942 | }; | |
943 | ||
944 | /* | |
5d7a86d1 ED |
945 | * Use the slog as long as the current commit size is less than the |
946 | * limit or the total list size is less than 2X the limit. Limit | |
947 | * checking is disabled by setting zil_slog_limit to UINT64_MAX. | |
428870ff | 948 | */ |
ee191e80 | 949 | unsigned long zil_slog_limit = 1024 * 1024; |
5d7a86d1 ED |
950 | #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \ |
951 | ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))) | |
428870ff | 952 | |
34dc7c2f BB |
953 | /* |
954 | * Start a log block write and advance to the next log block. | |
955 | * Calls are serialized. | |
956 | */ | |
957 | static lwb_t * | |
958 | zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) | |
959 | { | |
428870ff BB |
960 | lwb_t *nlwb = NULL; |
961 | zil_chain_t *zilc; | |
34dc7c2f | 962 | spa_t *spa = zilog->zl_spa; |
428870ff BB |
963 | blkptr_t *bp; |
964 | dmu_tx_t *tx; | |
34dc7c2f | 965 | uint64_t txg; |
428870ff BB |
966 | uint64_t zil_blksz, wsz; |
967 | int i, error; | |
b6ad9671 | 968 | boolean_t use_slog; |
428870ff BB |
969 | |
970 | if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { | |
971 | zilc = (zil_chain_t *)lwb->lwb_buf; | |
972 | bp = &zilc->zc_next_blk; | |
973 | } else { | |
974 | zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); | |
975 | bp = &zilc->zc_next_blk; | |
976 | } | |
34dc7c2f | 977 | |
428870ff | 978 | ASSERT(lwb->lwb_nused <= lwb->lwb_sz); |
34dc7c2f BB |
979 | |
980 | /* | |
981 | * Allocate the next block and save its address in this block | |
982 | * before writing it in order to establish the log chain. | |
983 | * Note that if the allocation of nlwb synced before we wrote | |
984 | * the block that points at it (lwb), we'd leak it if we crashed. | |
428870ff BB |
985 | * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). |
986 | * We dirty the dataset to ensure that zil_sync() will be called | |
987 | * to clean up in the event of allocation failure or I/O failure. | |
34dc7c2f | 988 | */ |
428870ff BB |
989 | tx = dmu_tx_create(zilog->zl_os); |
990 | VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); | |
991 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
992 | txg = dmu_tx_get_txg(tx); | |
993 | ||
994 | lwb->lwb_tx = tx; | |
34dc7c2f BB |
995 | |
996 | /* | |
428870ff BB |
997 | * Log blocks are pre-allocated. Here we select the size of the next |
998 | * block, based on size used in the last block. | |
999 | * - first find the smallest bucket that will fit the block from a | |
1000 | * limited set of block sizes. This is because it's faster to write | |
1001 | * blocks allocated from the same metaslab as they are adjacent or | |
1002 | * close. | |
1003 | * - next find the maximum from the new suggested size and an array of | |
1004 | * previous sizes. This lessens a picket fence effect of wrongly | |
1005 | * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k | |
1006 | * requests. | |
1007 | * | |
1008 | * Note we only write what is used, but we can't just allocate | |
1009 | * the maximum block size because we can exhaust the available | |
1010 | * pool log space. | |
34dc7c2f | 1011 | */ |
428870ff BB |
1012 | zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); |
1013 | for (i = 0; zil_blksz > zil_block_buckets[i]; i++) | |
1014 | continue; | |
1015 | zil_blksz = zil_block_buckets[i]; | |
1016 | if (zil_blksz == UINT64_MAX) | |
1017 | zil_blksz = SPA_MAXBLOCKSIZE; | |
1018 | zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; | |
1019 | for (i = 0; i < ZIL_PREV_BLKS; i++) | |
1020 | zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); | |
1021 | zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); | |
34dc7c2f BB |
1022 | |
1023 | BP_ZERO(bp); | |
b6ad9671 | 1024 | use_slog = USE_SLOG(zilog); |
13fe0198 MA |
1025 | error = zio_alloc_zil(spa, txg, bp, zil_blksz, |
1026 | USE_SLOG(zilog)); | |
d1d7e268 | 1027 | if (use_slog) { |
b6ad9671 ED |
1028 | ZIL_STAT_BUMP(zil_itx_metaslab_slog_count); |
1029 | ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused); | |
d1d7e268 | 1030 | } else { |
b6ad9671 ED |
1031 | ZIL_STAT_BUMP(zil_itx_metaslab_normal_count); |
1032 | ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused); | |
1033 | } | |
13fe0198 | 1034 | if (error == 0) { |
428870ff BB |
1035 | ASSERT3U(bp->blk_birth, ==, txg); |
1036 | bp->blk_cksum = lwb->lwb_blk.blk_cksum; | |
1037 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; | |
34dc7c2f BB |
1038 | |
1039 | /* | |
428870ff | 1040 | * Allocate a new log write buffer (lwb). |
34dc7c2f | 1041 | */ |
920dd524 | 1042 | nlwb = zil_alloc_lwb(zilog, bp, txg, TRUE); |
34dc7c2f | 1043 | |
428870ff BB |
1044 | /* Record the block for later vdev flushing */ |
1045 | zil_add_block(zilog, &lwb->lwb_blk); | |
34dc7c2f BB |
1046 | } |
1047 | ||
428870ff BB |
1048 | if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { |
1049 | /* For Slim ZIL only write what is used. */ | |
1050 | wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); | |
1051 | ASSERT3U(wsz, <=, lwb->lwb_sz); | |
1052 | zio_shrink(lwb->lwb_zio, wsz); | |
34dc7c2f | 1053 | |
428870ff BB |
1054 | } else { |
1055 | wsz = lwb->lwb_sz; | |
1056 | } | |
34dc7c2f | 1057 | |
428870ff BB |
1058 | zilc->zc_pad = 0; |
1059 | zilc->zc_nused = lwb->lwb_nused; | |
1060 | zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; | |
34dc7c2f BB |
1061 | |
1062 | /* | |
428870ff | 1063 | * clear unused data for security |
34dc7c2f | 1064 | */ |
428870ff | 1065 | bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); |
34dc7c2f | 1066 | |
428870ff | 1067 | zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ |
34dc7c2f BB |
1068 | |
1069 | /* | |
428870ff BB |
1070 | * If there was an allocation failure then nlwb will be null which |
1071 | * forces a txg_wait_synced(). | |
34dc7c2f | 1072 | */ |
34dc7c2f BB |
1073 | return (nlwb); |
1074 | } | |
1075 | ||
1076 | static lwb_t * | |
1077 | zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) | |
1078 | { | |
1079 | lr_t *lrc = &itx->itx_lr; /* common log record */ | |
428870ff BB |
1080 | lr_write_t *lrw = (lr_write_t *)lrc; |
1081 | char *lr_buf; | |
34dc7c2f BB |
1082 | uint64_t txg = lrc->lrc_txg; |
1083 | uint64_t reclen = lrc->lrc_reclen; | |
428870ff | 1084 | uint64_t dlen = 0; |
34dc7c2f BB |
1085 | |
1086 | if (lwb == NULL) | |
1087 | return (NULL); | |
428870ff | 1088 | |
34dc7c2f | 1089 | ASSERT(lwb->lwb_buf != NULL); |
29809a6c MA |
1090 | ASSERT(zilog_is_dirty(zilog) || |
1091 | spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); | |
34dc7c2f BB |
1092 | |
1093 | if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) | |
1094 | dlen = P2ROUNDUP_TYPED( | |
428870ff | 1095 | lrw->lr_length, sizeof (uint64_t), uint64_t); |
34dc7c2f BB |
1096 | |
1097 | zilog->zl_cur_used += (reclen + dlen); | |
1098 | ||
1099 | zil_lwb_write_init(zilog, lwb); | |
1100 | ||
1101 | /* | |
1102 | * If this record won't fit in the current log block, start a new one. | |
1103 | */ | |
428870ff | 1104 | if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { |
34dc7c2f BB |
1105 | lwb = zil_lwb_write_start(zilog, lwb); |
1106 | if (lwb == NULL) | |
1107 | return (NULL); | |
1108 | zil_lwb_write_init(zilog, lwb); | |
428870ff BB |
1109 | ASSERT(LWB_EMPTY(lwb)); |
1110 | if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { | |
34dc7c2f BB |
1111 | txg_wait_synced(zilog->zl_dmu_pool, txg); |
1112 | return (lwb); | |
1113 | } | |
1114 | } | |
1115 | ||
428870ff BB |
1116 | lr_buf = lwb->lwb_buf + lwb->lwb_nused; |
1117 | bcopy(lrc, lr_buf, reclen); | |
1118 | lrc = (lr_t *)lr_buf; | |
1119 | lrw = (lr_write_t *)lrc; | |
34dc7c2f | 1120 | |
b6ad9671 ED |
1121 | ZIL_STAT_BUMP(zil_itx_count); |
1122 | ||
34dc7c2f BB |
1123 | /* |
1124 | * If it's a write, fetch the data or get its blkptr as appropriate. | |
1125 | */ | |
1126 | if (lrc->lrc_txtype == TX_WRITE) { | |
1127 | if (txg > spa_freeze_txg(zilog->zl_spa)) | |
1128 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
b6ad9671 ED |
1129 | if (itx->itx_wr_state == WR_COPIED) { |
1130 | ZIL_STAT_BUMP(zil_itx_copied_count); | |
1131 | ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length); | |
1132 | } else { | |
34dc7c2f BB |
1133 | char *dbuf; |
1134 | int error; | |
1135 | ||
34dc7c2f BB |
1136 | if (dlen) { |
1137 | ASSERT(itx->itx_wr_state == WR_NEED_COPY); | |
428870ff BB |
1138 | dbuf = lr_buf + reclen; |
1139 | lrw->lr_common.lrc_reclen += dlen; | |
b6ad9671 | 1140 | ZIL_STAT_BUMP(zil_itx_needcopy_count); |
d1d7e268 MK |
1141 | ZIL_STAT_INCR(zil_itx_needcopy_bytes, |
1142 | lrw->lr_length); | |
34dc7c2f BB |
1143 | } else { |
1144 | ASSERT(itx->itx_wr_state == WR_INDIRECT); | |
1145 | dbuf = NULL; | |
b6ad9671 | 1146 | ZIL_STAT_BUMP(zil_itx_indirect_count); |
d1d7e268 MK |
1147 | ZIL_STAT_INCR(zil_itx_indirect_bytes, |
1148 | lrw->lr_length); | |
34dc7c2f BB |
1149 | } |
1150 | error = zilog->zl_get_data( | |
428870ff | 1151 | itx->itx_private, lrw, dbuf, lwb->lwb_zio); |
45d1cae3 BB |
1152 | if (error == EIO) { |
1153 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
1154 | return (lwb); | |
1155 | } | |
13fe0198 | 1156 | if (error != 0) { |
34dc7c2f BB |
1157 | ASSERT(error == ENOENT || error == EEXIST || |
1158 | error == EALREADY); | |
1159 | return (lwb); | |
1160 | } | |
1161 | } | |
1162 | } | |
1163 | ||
428870ff BB |
1164 | /* |
1165 | * We're actually making an entry, so update lrc_seq to be the | |
1166 | * log record sequence number. Note that this is generally not | |
1167 | * equal to the itx sequence number because not all transactions | |
1168 | * are synchronous, and sometimes spa_sync() gets there first. | |
1169 | */ | |
1170 | lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ | |
34dc7c2f BB |
1171 | lwb->lwb_nused += reclen + dlen; |
1172 | lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); | |
428870ff | 1173 | ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); |
c99c9001 | 1174 | ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); |
34dc7c2f BB |
1175 | |
1176 | return (lwb); | |
1177 | } | |
1178 | ||
1179 | itx_t * | |
1180 | zil_itx_create(uint64_t txtype, size_t lrsize) | |
1181 | { | |
1182 | itx_t *itx; | |
1183 | ||
1184 | lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); | |
1185 | ||
701b1f81 BB |
1186 | itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, |
1187 | KM_PUSHPAGE | KM_NODEBUG); | |
34dc7c2f BB |
1188 | itx->itx_lr.lrc_txtype = txtype; |
1189 | itx->itx_lr.lrc_reclen = lrsize; | |
1190 | itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ | |
1191 | itx->itx_lr.lrc_seq = 0; /* defensive */ | |
572e2857 | 1192 | itx->itx_sync = B_TRUE; /* default is synchronous */ |
119a394a ED |
1193 | itx->itx_callback = NULL; |
1194 | itx->itx_callback_data = NULL; | |
34dc7c2f BB |
1195 | |
1196 | return (itx); | |
1197 | } | |
1198 | ||
428870ff BB |
1199 | void |
1200 | zil_itx_destroy(itx_t *itx) | |
1201 | { | |
1202 | kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); | |
1203 | } | |
1204 | ||
572e2857 BB |
1205 | /* |
1206 | * Free up the sync and async itxs. The itxs_t has already been detached | |
1207 | * so no locks are needed. | |
1208 | */ | |
1209 | static void | |
1210 | zil_itxg_clean(itxs_t *itxs) | |
34dc7c2f | 1211 | { |
572e2857 BB |
1212 | itx_t *itx; |
1213 | list_t *list; | |
1214 | avl_tree_t *t; | |
1215 | void *cookie; | |
1216 | itx_async_node_t *ian; | |
1217 | ||
1218 | list = &itxs->i_sync_list; | |
1219 | while ((itx = list_head(list)) != NULL) { | |
119a394a ED |
1220 | if (itx->itx_callback != NULL) |
1221 | itx->itx_callback(itx->itx_callback_data); | |
572e2857 BB |
1222 | list_remove(list, itx); |
1223 | kmem_free(itx, offsetof(itx_t, itx_lr) + | |
1224 | itx->itx_lr.lrc_reclen); | |
1225 | } | |
34dc7c2f | 1226 | |
572e2857 BB |
1227 | cookie = NULL; |
1228 | t = &itxs->i_async_tree; | |
1229 | while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { | |
1230 | list = &ian->ia_list; | |
1231 | while ((itx = list_head(list)) != NULL) { | |
119a394a ED |
1232 | if (itx->itx_callback != NULL) |
1233 | itx->itx_callback(itx->itx_callback_data); | |
572e2857 BB |
1234 | list_remove(list, itx); |
1235 | kmem_free(itx, offsetof(itx_t, itx_lr) + | |
1236 | itx->itx_lr.lrc_reclen); | |
1237 | } | |
1238 | list_destroy(list); | |
1239 | kmem_free(ian, sizeof (itx_async_node_t)); | |
1240 | } | |
1241 | avl_destroy(t); | |
34dc7c2f | 1242 | |
572e2857 BB |
1243 | kmem_free(itxs, sizeof (itxs_t)); |
1244 | } | |
34dc7c2f | 1245 | |
572e2857 BB |
1246 | static int |
1247 | zil_aitx_compare(const void *x1, const void *x2) | |
1248 | { | |
1249 | const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; | |
1250 | const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; | |
1251 | ||
1252 | if (o1 < o2) | |
1253 | return (-1); | |
1254 | if (o1 > o2) | |
1255 | return (1); | |
1256 | ||
1257 | return (0); | |
34dc7c2f BB |
1258 | } |
1259 | ||
1260 | /* | |
572e2857 | 1261 | * Remove all async itx with the given oid. |
34dc7c2f BB |
1262 | */ |
1263 | static void | |
572e2857 | 1264 | zil_remove_async(zilog_t *zilog, uint64_t oid) |
34dc7c2f | 1265 | { |
572e2857 BB |
1266 | uint64_t otxg, txg; |
1267 | itx_async_node_t *ian; | |
1268 | avl_tree_t *t; | |
1269 | avl_index_t where; | |
34dc7c2f BB |
1270 | list_t clean_list; |
1271 | itx_t *itx; | |
1272 | ||
572e2857 | 1273 | ASSERT(oid != 0); |
34dc7c2f BB |
1274 | list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); |
1275 | ||
572e2857 BB |
1276 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ |
1277 | otxg = ZILTEST_TXG; | |
1278 | else | |
1279 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
34dc7c2f | 1280 | |
572e2857 BB |
1281 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { |
1282 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1283 | ||
1284 | mutex_enter(&itxg->itxg_lock); | |
1285 | if (itxg->itxg_txg != txg) { | |
1286 | mutex_exit(&itxg->itxg_lock); | |
1287 | continue; | |
1288 | } | |
34dc7c2f | 1289 | |
572e2857 BB |
1290 | /* |
1291 | * Locate the object node and append its list. | |
1292 | */ | |
1293 | t = &itxg->itxg_itxs->i_async_tree; | |
1294 | ian = avl_find(t, &oid, &where); | |
1295 | if (ian != NULL) | |
1296 | list_move_tail(&clean_list, &ian->ia_list); | |
1297 | mutex_exit(&itxg->itxg_lock); | |
1298 | } | |
34dc7c2f | 1299 | while ((itx = list_head(&clean_list)) != NULL) { |
119a394a ED |
1300 | if (itx->itx_callback != NULL) |
1301 | itx->itx_callback(itx->itx_callback_data); | |
34dc7c2f | 1302 | list_remove(&clean_list, itx); |
572e2857 BB |
1303 | kmem_free(itx, offsetof(itx_t, itx_lr) + |
1304 | itx->itx_lr.lrc_reclen); | |
34dc7c2f BB |
1305 | } |
1306 | list_destroy(&clean_list); | |
1307 | } | |
1308 | ||
572e2857 BB |
1309 | void |
1310 | zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) | |
1311 | { | |
1312 | uint64_t txg; | |
1313 | itxg_t *itxg; | |
1314 | itxs_t *itxs, *clean = NULL; | |
1315 | ||
1316 | /* | |
1317 | * Object ids can be re-instantiated in the next txg so | |
1318 | * remove any async transactions to avoid future leaks. | |
1319 | * This can happen if a fsync occurs on the re-instantiated | |
1320 | * object for a WR_INDIRECT or WR_NEED_COPY write, which gets | |
1321 | * the new file data and flushes a write record for the old object. | |
1322 | */ | |
1323 | if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) | |
1324 | zil_remove_async(zilog, itx->itx_oid); | |
1325 | ||
1326 | /* | |
1327 | * Ensure the data of a renamed file is committed before the rename. | |
1328 | */ | |
1329 | if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) | |
1330 | zil_async_to_sync(zilog, itx->itx_oid); | |
1331 | ||
29809a6c | 1332 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) |
572e2857 BB |
1333 | txg = ZILTEST_TXG; |
1334 | else | |
1335 | txg = dmu_tx_get_txg(tx); | |
1336 | ||
1337 | itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1338 | mutex_enter(&itxg->itxg_lock); | |
1339 | itxs = itxg->itxg_itxs; | |
1340 | if (itxg->itxg_txg != txg) { | |
1341 | if (itxs != NULL) { | |
1342 | /* | |
1343 | * The zil_clean callback hasn't got around to cleaning | |
1344 | * this itxg. Save the itxs for release below. | |
1345 | * This should be rare. | |
1346 | */ | |
1347 | atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); | |
1348 | itxg->itxg_sod = 0; | |
1349 | clean = itxg->itxg_itxs; | |
1350 | } | |
1351 | ASSERT(itxg->itxg_sod == 0); | |
1352 | itxg->itxg_txg = txg; | |
d1d7e268 MK |
1353 | itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), |
1354 | KM_PUSHPAGE); | |
572e2857 BB |
1355 | |
1356 | list_create(&itxs->i_sync_list, sizeof (itx_t), | |
1357 | offsetof(itx_t, itx_node)); | |
1358 | avl_create(&itxs->i_async_tree, zil_aitx_compare, | |
1359 | sizeof (itx_async_node_t), | |
1360 | offsetof(itx_async_node_t, ia_node)); | |
1361 | } | |
1362 | if (itx->itx_sync) { | |
1363 | list_insert_tail(&itxs->i_sync_list, itx); | |
1364 | atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); | |
1365 | itxg->itxg_sod += itx->itx_sod; | |
1366 | } else { | |
1367 | avl_tree_t *t = &itxs->i_async_tree; | |
1368 | uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; | |
1369 | itx_async_node_t *ian; | |
1370 | avl_index_t where; | |
1371 | ||
1372 | ian = avl_find(t, &foid, &where); | |
1373 | if (ian == NULL) { | |
d1d7e268 MK |
1374 | ian = kmem_alloc(sizeof (itx_async_node_t), |
1375 | KM_PUSHPAGE); | |
572e2857 BB |
1376 | list_create(&ian->ia_list, sizeof (itx_t), |
1377 | offsetof(itx_t, itx_node)); | |
1378 | ian->ia_foid = foid; | |
1379 | avl_insert(t, ian, where); | |
1380 | } | |
1381 | list_insert_tail(&ian->ia_list, itx); | |
1382 | } | |
1383 | ||
1384 | itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); | |
29809a6c | 1385 | zilog_dirty(zilog, txg); |
572e2857 BB |
1386 | mutex_exit(&itxg->itxg_lock); |
1387 | ||
1388 | /* Release the old itxs now we've dropped the lock */ | |
1389 | if (clean != NULL) | |
1390 | zil_itxg_clean(clean); | |
1391 | } | |
1392 | ||
34dc7c2f BB |
1393 | /* |
1394 | * If there are any in-memory intent log transactions which have now been | |
29809a6c MA |
1395 | * synced then start up a taskq to free them. We should only do this after we |
1396 | * have written out the uberblocks (i.e. txg has been comitted) so that | |
1397 | * don't inadvertently clean out in-memory log records that would be required | |
1398 | * by zil_commit(). | |
34dc7c2f BB |
1399 | */ |
1400 | void | |
572e2857 | 1401 | zil_clean(zilog_t *zilog, uint64_t synced_txg) |
34dc7c2f | 1402 | { |
572e2857 BB |
1403 | itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; |
1404 | itxs_t *clean_me; | |
34dc7c2f | 1405 | |
572e2857 BB |
1406 | mutex_enter(&itxg->itxg_lock); |
1407 | if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { | |
1408 | mutex_exit(&itxg->itxg_lock); | |
1409 | return; | |
1410 | } | |
1411 | ASSERT3U(itxg->itxg_txg, <=, synced_txg); | |
1412 | ASSERT(itxg->itxg_txg != 0); | |
1413 | ASSERT(zilog->zl_clean_taskq != NULL); | |
1414 | atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); | |
1415 | itxg->itxg_sod = 0; | |
1416 | clean_me = itxg->itxg_itxs; | |
1417 | itxg->itxg_itxs = NULL; | |
1418 | itxg->itxg_txg = 0; | |
1419 | mutex_exit(&itxg->itxg_lock); | |
1420 | /* | |
1421 | * Preferably start a task queue to free up the old itxs but | |
1422 | * if taskq_dispatch can't allocate resources to do that then | |
1423 | * free it in-line. This should be rare. Note, using TQ_SLEEP | |
1424 | * created a bad performance problem. | |
1425 | */ | |
1426 | if (taskq_dispatch(zilog->zl_clean_taskq, | |
b8864a23 | 1427 | (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) |
572e2857 BB |
1428 | zil_itxg_clean(clean_me); |
1429 | } | |
1430 | ||
1431 | /* | |
1432 | * Get the list of itxs to commit into zl_itx_commit_list. | |
1433 | */ | |
1434 | static void | |
1435 | zil_get_commit_list(zilog_t *zilog) | |
1436 | { | |
1437 | uint64_t otxg, txg; | |
1438 | list_t *commit_list = &zilog->zl_itx_commit_list; | |
1439 | uint64_t push_sod = 0; | |
1440 | ||
1441 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ | |
1442 | otxg = ZILTEST_TXG; | |
1443 | else | |
1444 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
1445 | ||
1446 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { | |
1447 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1448 | ||
1449 | mutex_enter(&itxg->itxg_lock); | |
1450 | if (itxg->itxg_txg != txg) { | |
1451 | mutex_exit(&itxg->itxg_lock); | |
1452 | continue; | |
1453 | } | |
1454 | ||
1455 | list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); | |
1456 | push_sod += itxg->itxg_sod; | |
1457 | itxg->itxg_sod = 0; | |
1458 | ||
1459 | mutex_exit(&itxg->itxg_lock); | |
1460 | } | |
1461 | atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); | |
1462 | } | |
1463 | ||
1464 | /* | |
1465 | * Move the async itxs for a specified object to commit into sync lists. | |
1466 | */ | |
1467 | static void | |
1468 | zil_async_to_sync(zilog_t *zilog, uint64_t foid) | |
1469 | { | |
1470 | uint64_t otxg, txg; | |
1471 | itx_async_node_t *ian; | |
1472 | avl_tree_t *t; | |
1473 | avl_index_t where; | |
1474 | ||
1475 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ | |
1476 | otxg = ZILTEST_TXG; | |
1477 | else | |
1478 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
1479 | ||
1480 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { | |
1481 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1482 | ||
1483 | mutex_enter(&itxg->itxg_lock); | |
1484 | if (itxg->itxg_txg != txg) { | |
1485 | mutex_exit(&itxg->itxg_lock); | |
1486 | continue; | |
1487 | } | |
1488 | ||
1489 | /* | |
1490 | * If a foid is specified then find that node and append its | |
1491 | * list. Otherwise walk the tree appending all the lists | |
1492 | * to the sync list. We add to the end rather than the | |
1493 | * beginning to ensure the create has happened. | |
1494 | */ | |
1495 | t = &itxg->itxg_itxs->i_async_tree; | |
1496 | if (foid != 0) { | |
1497 | ian = avl_find(t, &foid, &where); | |
1498 | if (ian != NULL) { | |
1499 | list_move_tail(&itxg->itxg_itxs->i_sync_list, | |
1500 | &ian->ia_list); | |
1501 | } | |
1502 | } else { | |
1503 | void *cookie = NULL; | |
1504 | ||
1505 | while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { | |
1506 | list_move_tail(&itxg->itxg_itxs->i_sync_list, | |
1507 | &ian->ia_list); | |
1508 | list_destroy(&ian->ia_list); | |
1509 | kmem_free(ian, sizeof (itx_async_node_t)); | |
1510 | } | |
1511 | } | |
1512 | mutex_exit(&itxg->itxg_lock); | |
34dc7c2f | 1513 | } |
34dc7c2f BB |
1514 | } |
1515 | ||
b128c09f | 1516 | static void |
572e2857 | 1517 | zil_commit_writer(zilog_t *zilog) |
34dc7c2f BB |
1518 | { |
1519 | uint64_t txg; | |
572e2857 | 1520 | itx_t *itx; |
34dc7c2f | 1521 | lwb_t *lwb; |
572e2857 | 1522 | spa_t *spa = zilog->zl_spa; |
428870ff | 1523 | int error = 0; |
34dc7c2f | 1524 | |
b128c09f | 1525 | ASSERT(zilog->zl_root_zio == NULL); |
572e2857 BB |
1526 | |
1527 | mutex_exit(&zilog->zl_lock); | |
1528 | ||
1529 | zil_get_commit_list(zilog); | |
1530 | ||
1531 | /* | |
1532 | * Return if there's nothing to commit before we dirty the fs by | |
1533 | * calling zil_create(). | |
1534 | */ | |
1535 | if (list_head(&zilog->zl_itx_commit_list) == NULL) { | |
1536 | mutex_enter(&zilog->zl_lock); | |
1537 | return; | |
1538 | } | |
34dc7c2f BB |
1539 | |
1540 | if (zilog->zl_suspend) { | |
1541 | lwb = NULL; | |
1542 | } else { | |
1543 | lwb = list_tail(&zilog->zl_lwb_list); | |
572e2857 | 1544 | if (lwb == NULL) |
428870ff | 1545 | lwb = zil_create(zilog); |
34dc7c2f BB |
1546 | } |
1547 | ||
34dc7c2f | 1548 | DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); |
119a394a | 1549 | for (itx = list_head(&zilog->zl_itx_commit_list); itx != NULL; |
d1d7e268 | 1550 | itx = list_next(&zilog->zl_itx_commit_list, itx)) { |
34dc7c2f BB |
1551 | txg = itx->itx_lr.lrc_txg; |
1552 | ASSERT(txg); | |
1553 | ||
572e2857 | 1554 | if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) |
34dc7c2f | 1555 | lwb = zil_lwb_commit(zilog, itx, lwb); |
34dc7c2f BB |
1556 | } |
1557 | DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); | |
34dc7c2f BB |
1558 | |
1559 | /* write the last block out */ | |
1560 | if (lwb != NULL && lwb->lwb_zio != NULL) | |
1561 | lwb = zil_lwb_write_start(zilog, lwb); | |
1562 | ||
34dc7c2f BB |
1563 | zilog->zl_cur_used = 0; |
1564 | ||
1565 | /* | |
1566 | * Wait if necessary for the log blocks to be on stable storage. | |
1567 | */ | |
1568 | if (zilog->zl_root_zio) { | |
428870ff | 1569 | error = zio_wait(zilog->zl_root_zio); |
b128c09f | 1570 | zilog->zl_root_zio = NULL; |
34dc7c2f BB |
1571 | zil_flush_vdevs(zilog); |
1572 | } | |
1573 | ||
428870ff | 1574 | if (error || lwb == NULL) |
34dc7c2f | 1575 | txg_wait_synced(zilog->zl_dmu_pool, 0); |
34dc7c2f | 1576 | |
119a394a ED |
1577 | while ((itx = list_head(&zilog->zl_itx_commit_list))) { |
1578 | txg = itx->itx_lr.lrc_txg; | |
1579 | ASSERT(txg); | |
1580 | ||
1581 | if (itx->itx_callback != NULL) | |
1582 | itx->itx_callback(itx->itx_callback_data); | |
1583 | list_remove(&zilog->zl_itx_commit_list, itx); | |
1584 | kmem_free(itx, offsetof(itx_t, itx_lr) | |
1585 | + itx->itx_lr.lrc_reclen); | |
1586 | } | |
1587 | ||
34dc7c2f | 1588 | mutex_enter(&zilog->zl_lock); |
428870ff BB |
1589 | |
1590 | /* | |
1591 | * Remember the highest committed log sequence number for ztest. | |
1592 | * We only update this value when all the log writes succeeded, | |
1593 | * because ztest wants to ASSERT that it got the whole log chain. | |
1594 | */ | |
1595 | if (error == 0 && lwb != NULL) | |
1596 | zilog->zl_commit_lr_seq = zilog->zl_lr_seq; | |
34dc7c2f BB |
1597 | } |
1598 | ||
1599 | /* | |
572e2857 | 1600 | * Commit zfs transactions to stable storage. |
34dc7c2f | 1601 | * If foid is 0 push out all transactions, otherwise push only those |
572e2857 BB |
1602 | * for that object or might reference that object. |
1603 | * | |
1604 | * itxs are committed in batches. In a heavily stressed zil there will be | |
1605 | * a commit writer thread who is writing out a bunch of itxs to the log | |
1606 | * for a set of committing threads (cthreads) in the same batch as the writer. | |
1607 | * Those cthreads are all waiting on the same cv for that batch. | |
1608 | * | |
1609 | * There will also be a different and growing batch of threads that are | |
1610 | * waiting to commit (qthreads). When the committing batch completes | |
1611 | * a transition occurs such that the cthreads exit and the qthreads become | |
1612 | * cthreads. One of the new cthreads becomes the writer thread for the | |
1613 | * batch. Any new threads arriving become new qthreads. | |
1614 | * | |
1615 | * Only 2 condition variables are needed and there's no transition | |
1616 | * between the two cvs needed. They just flip-flop between qthreads | |
1617 | * and cthreads. | |
1618 | * | |
1619 | * Using this scheme we can efficiently wakeup up only those threads | |
1620 | * that have been committed. | |
34dc7c2f BB |
1621 | */ |
1622 | void | |
572e2857 | 1623 | zil_commit(zilog_t *zilog, uint64_t foid) |
34dc7c2f | 1624 | { |
572e2857 | 1625 | uint64_t mybatch; |
34dc7c2f | 1626 | |
572e2857 BB |
1627 | if (zilog->zl_sync == ZFS_SYNC_DISABLED) |
1628 | return; | |
34dc7c2f | 1629 | |
b6ad9671 ED |
1630 | ZIL_STAT_BUMP(zil_commit_count); |
1631 | ||
572e2857 BB |
1632 | /* move the async itxs for the foid to the sync queues */ |
1633 | zil_async_to_sync(zilog, foid); | |
34dc7c2f | 1634 | |
572e2857 BB |
1635 | mutex_enter(&zilog->zl_lock); |
1636 | mybatch = zilog->zl_next_batch; | |
34dc7c2f | 1637 | while (zilog->zl_writer) { |
572e2857 BB |
1638 | cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); |
1639 | if (mybatch <= zilog->zl_com_batch) { | |
34dc7c2f BB |
1640 | mutex_exit(&zilog->zl_lock); |
1641 | return; | |
1642 | } | |
1643 | } | |
428870ff | 1644 | |
572e2857 BB |
1645 | zilog->zl_next_batch++; |
1646 | zilog->zl_writer = B_TRUE; | |
b6ad9671 | 1647 | ZIL_STAT_BUMP(zil_commit_writer_count); |
572e2857 BB |
1648 | zil_commit_writer(zilog); |
1649 | zilog->zl_com_batch = mybatch; | |
1650 | zilog->zl_writer = B_FALSE; | |
428870ff | 1651 | |
572e2857 BB |
1652 | /* wake up one thread to become the next writer */ |
1653 | cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); | |
428870ff | 1654 | |
572e2857 BB |
1655 | /* wake up all threads waiting for this batch to be committed */ |
1656 | cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); | |
8c0712fd BB |
1657 | |
1658 | mutex_exit(&zilog->zl_lock); | |
428870ff BB |
1659 | } |
1660 | ||
34dc7c2f BB |
1661 | /* |
1662 | * Called in syncing context to free committed log blocks and update log header. | |
1663 | */ | |
1664 | void | |
1665 | zil_sync(zilog_t *zilog, dmu_tx_t *tx) | |
1666 | { | |
1667 | zil_header_t *zh = zil_header_in_syncing_context(zilog); | |
1668 | uint64_t txg = dmu_tx_get_txg(tx); | |
1669 | spa_t *spa = zilog->zl_spa; | |
428870ff | 1670 | uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; |
34dc7c2f BB |
1671 | lwb_t *lwb; |
1672 | ||
9babb374 BB |
1673 | /* |
1674 | * We don't zero out zl_destroy_txg, so make sure we don't try | |
1675 | * to destroy it twice. | |
1676 | */ | |
1677 | if (spa_sync_pass(spa) != 1) | |
1678 | return; | |
1679 | ||
34dc7c2f BB |
1680 | mutex_enter(&zilog->zl_lock); |
1681 | ||
1682 | ASSERT(zilog->zl_stop_sync == 0); | |
1683 | ||
428870ff BB |
1684 | if (*replayed_seq != 0) { |
1685 | ASSERT(zh->zh_replay_seq < *replayed_seq); | |
1686 | zh->zh_replay_seq = *replayed_seq; | |
1687 | *replayed_seq = 0; | |
1688 | } | |
34dc7c2f BB |
1689 | |
1690 | if (zilog->zl_destroy_txg == txg) { | |
1691 | blkptr_t blk = zh->zh_log; | |
1692 | ||
1693 | ASSERT(list_head(&zilog->zl_lwb_list) == NULL); | |
34dc7c2f BB |
1694 | |
1695 | bzero(zh, sizeof (zil_header_t)); | |
fb5f0bc8 | 1696 | bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); |
34dc7c2f BB |
1697 | |
1698 | if (zilog->zl_keep_first) { | |
1699 | /* | |
1700 | * If this block was part of log chain that couldn't | |
1701 | * be claimed because a device was missing during | |
1702 | * zil_claim(), but that device later returns, | |
1703 | * then this block could erroneously appear valid. | |
1704 | * To guard against this, assign a new GUID to the new | |
1705 | * log chain so it doesn't matter what blk points to. | |
1706 | */ | |
1707 | zil_init_log_chain(zilog, &blk); | |
1708 | zh->zh_log = blk; | |
1709 | } | |
1710 | } | |
1711 | ||
9babb374 | 1712 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { |
34dc7c2f BB |
1713 | zh->zh_log = lwb->lwb_blk; |
1714 | if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) | |
1715 | break; | |
920dd524 ED |
1716 | |
1717 | ASSERT(lwb->lwb_zio == NULL); | |
1718 | ||
34dc7c2f | 1719 | list_remove(&zilog->zl_lwb_list, lwb); |
428870ff | 1720 | zio_free_zil(spa, txg, &lwb->lwb_blk); |
34dc7c2f BB |
1721 | kmem_cache_free(zil_lwb_cache, lwb); |
1722 | ||
1723 | /* | |
1724 | * If we don't have anything left in the lwb list then | |
1725 | * we've had an allocation failure and we need to zero | |
1726 | * out the zil_header blkptr so that we don't end | |
1727 | * up freeing the same block twice. | |
1728 | */ | |
1729 | if (list_head(&zilog->zl_lwb_list) == NULL) | |
1730 | BP_ZERO(&zh->zh_log); | |
1731 | } | |
920dd524 ED |
1732 | |
1733 | /* | |
1734 | * Remove fastwrite on any blocks that have been pre-allocated for | |
1735 | * the next commit. This prevents fastwrite counter pollution by | |
1736 | * unused, long-lived LWBs. | |
1737 | */ | |
1738 | for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) { | |
1739 | if (lwb->lwb_fastwrite && !lwb->lwb_zio) { | |
1740 | metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); | |
1741 | lwb->lwb_fastwrite = 0; | |
1742 | } | |
1743 | } | |
1744 | ||
34dc7c2f BB |
1745 | mutex_exit(&zilog->zl_lock); |
1746 | } | |
1747 | ||
1748 | void | |
1749 | zil_init(void) | |
1750 | { | |
1751 | zil_lwb_cache = kmem_cache_create("zil_lwb_cache", | |
1752 | sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
b6ad9671 ED |
1753 | |
1754 | zil_ksp = kstat_create("zfs", 0, "zil", "misc", | |
d1d7e268 | 1755 | KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t), |
b6ad9671 ED |
1756 | KSTAT_FLAG_VIRTUAL); |
1757 | ||
1758 | if (zil_ksp != NULL) { | |
1759 | zil_ksp->ks_data = &zil_stats; | |
1760 | kstat_install(zil_ksp); | |
1761 | } | |
34dc7c2f BB |
1762 | } |
1763 | ||
1764 | void | |
1765 | zil_fini(void) | |
1766 | { | |
1767 | kmem_cache_destroy(zil_lwb_cache); | |
b6ad9671 ED |
1768 | |
1769 | if (zil_ksp != NULL) { | |
1770 | kstat_delete(zil_ksp); | |
1771 | zil_ksp = NULL; | |
1772 | } | |
34dc7c2f BB |
1773 | } |
1774 | ||
428870ff BB |
1775 | void |
1776 | zil_set_sync(zilog_t *zilog, uint64_t sync) | |
1777 | { | |
1778 | zilog->zl_sync = sync; | |
1779 | } | |
1780 | ||
1781 | void | |
1782 | zil_set_logbias(zilog_t *zilog, uint64_t logbias) | |
1783 | { | |
1784 | zilog->zl_logbias = logbias; | |
1785 | } | |
1786 | ||
34dc7c2f BB |
1787 | zilog_t * |
1788 | zil_alloc(objset_t *os, zil_header_t *zh_phys) | |
1789 | { | |
1790 | zilog_t *zilog; | |
d6320ddb | 1791 | int i; |
34dc7c2f | 1792 | |
b8d06fca | 1793 | zilog = kmem_zalloc(sizeof (zilog_t), KM_PUSHPAGE); |
34dc7c2f BB |
1794 | |
1795 | zilog->zl_header = zh_phys; | |
1796 | zilog->zl_os = os; | |
1797 | zilog->zl_spa = dmu_objset_spa(os); | |
1798 | zilog->zl_dmu_pool = dmu_objset_pool(os); | |
1799 | zilog->zl_destroy_txg = TXG_INITIAL - 1; | |
428870ff BB |
1800 | zilog->zl_logbias = dmu_objset_logbias(os); |
1801 | zilog->zl_sync = dmu_objset_syncprop(os); | |
572e2857 | 1802 | zilog->zl_next_batch = 1; |
34dc7c2f BB |
1803 | |
1804 | mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); | |
1805 | ||
d6320ddb | 1806 | for (i = 0; i < TXG_SIZE; i++) { |
572e2857 BB |
1807 | mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, |
1808 | MUTEX_DEFAULT, NULL); | |
1809 | } | |
34dc7c2f BB |
1810 | |
1811 | list_create(&zilog->zl_lwb_list, sizeof (lwb_t), | |
1812 | offsetof(lwb_t, lwb_node)); | |
1813 | ||
572e2857 BB |
1814 | list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), |
1815 | offsetof(itx_t, itx_node)); | |
1816 | ||
34dc7c2f BB |
1817 | mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); |
1818 | ||
1819 | avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, | |
1820 | sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); | |
1821 | ||
1822 | cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); | |
1823 | cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); | |
572e2857 BB |
1824 | cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); |
1825 | cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); | |
34dc7c2f BB |
1826 | |
1827 | return (zilog); | |
1828 | } | |
1829 | ||
1830 | void | |
1831 | zil_free(zilog_t *zilog) | |
1832 | { | |
d6320ddb | 1833 | int i; |
34dc7c2f BB |
1834 | |
1835 | zilog->zl_stop_sync = 1; | |
1836 | ||
13fe0198 MA |
1837 | ASSERT0(zilog->zl_suspend); |
1838 | ASSERT0(zilog->zl_suspending); | |
1839 | ||
3e31d2b0 | 1840 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); |
34dc7c2f BB |
1841 | list_destroy(&zilog->zl_lwb_list); |
1842 | ||
1843 | avl_destroy(&zilog->zl_vdev_tree); | |
1844 | mutex_destroy(&zilog->zl_vdev_lock); | |
1845 | ||
572e2857 BB |
1846 | ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); |
1847 | list_destroy(&zilog->zl_itx_commit_list); | |
1848 | ||
d6320ddb | 1849 | for (i = 0; i < TXG_SIZE; i++) { |
572e2857 BB |
1850 | /* |
1851 | * It's possible for an itx to be generated that doesn't dirty | |
1852 | * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() | |
1853 | * callback to remove the entry. We remove those here. | |
1854 | * | |
1855 | * Also free up the ziltest itxs. | |
1856 | */ | |
1857 | if (zilog->zl_itxg[i].itxg_itxs) | |
1858 | zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); | |
1859 | mutex_destroy(&zilog->zl_itxg[i].itxg_lock); | |
1860 | } | |
1861 | ||
34dc7c2f BB |
1862 | mutex_destroy(&zilog->zl_lock); |
1863 | ||
1864 | cv_destroy(&zilog->zl_cv_writer); | |
1865 | cv_destroy(&zilog->zl_cv_suspend); | |
572e2857 BB |
1866 | cv_destroy(&zilog->zl_cv_batch[0]); |
1867 | cv_destroy(&zilog->zl_cv_batch[1]); | |
34dc7c2f BB |
1868 | |
1869 | kmem_free(zilog, sizeof (zilog_t)); | |
1870 | } | |
1871 | ||
34dc7c2f BB |
1872 | /* |
1873 | * Open an intent log. | |
1874 | */ | |
1875 | zilog_t * | |
1876 | zil_open(objset_t *os, zil_get_data_t *get_data) | |
1877 | { | |
1878 | zilog_t *zilog = dmu_objset_zil(os); | |
1879 | ||
3e31d2b0 ES |
1880 | ASSERT(zilog->zl_clean_taskq == NULL); |
1881 | ASSERT(zilog->zl_get_data == NULL); | |
1882 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
1883 | ||
34dc7c2f BB |
1884 | zilog->zl_get_data = get_data; |
1885 | zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, | |
1886 | 2, 2, TASKQ_PREPOPULATE); | |
1887 | ||
1888 | return (zilog); | |
1889 | } | |
1890 | ||
1891 | /* | |
1892 | * Close an intent log. | |
1893 | */ | |
1894 | void | |
1895 | zil_close(zilog_t *zilog) | |
1896 | { | |
3e31d2b0 | 1897 | lwb_t *lwb; |
572e2857 BB |
1898 | uint64_t txg = 0; |
1899 | ||
1900 | zil_commit(zilog, 0); /* commit all itx */ | |
1901 | ||
34dc7c2f | 1902 | /* |
572e2857 BB |
1903 | * The lwb_max_txg for the stubby lwb will reflect the last activity |
1904 | * for the zil. After a txg_wait_synced() on the txg we know all the | |
1905 | * callbacks have occurred that may clean the zil. Only then can we | |
1906 | * destroy the zl_clean_taskq. | |
34dc7c2f | 1907 | */ |
572e2857 | 1908 | mutex_enter(&zilog->zl_lock); |
3e31d2b0 ES |
1909 | lwb = list_tail(&zilog->zl_lwb_list); |
1910 | if (lwb != NULL) | |
1911 | txg = lwb->lwb_max_txg; | |
572e2857 BB |
1912 | mutex_exit(&zilog->zl_lock); |
1913 | if (txg) | |
34dc7c2f | 1914 | txg_wait_synced(zilog->zl_dmu_pool, txg); |
29809a6c | 1915 | ASSERT(!zilog_is_dirty(zilog)); |
34dc7c2f BB |
1916 | |
1917 | taskq_destroy(zilog->zl_clean_taskq); | |
1918 | zilog->zl_clean_taskq = NULL; | |
1919 | zilog->zl_get_data = NULL; | |
3e31d2b0 ES |
1920 | |
1921 | /* | |
1922 | * We should have only one LWB left on the list; remove it now. | |
1923 | */ | |
1924 | mutex_enter(&zilog->zl_lock); | |
1925 | lwb = list_head(&zilog->zl_lwb_list); | |
1926 | if (lwb != NULL) { | |
1927 | ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); | |
920dd524 ED |
1928 | ASSERT(lwb->lwb_zio == NULL); |
1929 | if (lwb->lwb_fastwrite) | |
1930 | metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); | |
3e31d2b0 ES |
1931 | list_remove(&zilog->zl_lwb_list, lwb); |
1932 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
1933 | kmem_cache_free(zil_lwb_cache, lwb); | |
1934 | } | |
1935 | mutex_exit(&zilog->zl_lock); | |
34dc7c2f BB |
1936 | } |
1937 | ||
13fe0198 MA |
1938 | static char *suspend_tag = "zil suspending"; |
1939 | ||
34dc7c2f BB |
1940 | /* |
1941 | * Suspend an intent log. While in suspended mode, we still honor | |
1942 | * synchronous semantics, but we rely on txg_wait_synced() to do it. | |
13fe0198 MA |
1943 | * On old version pools, we suspend the log briefly when taking a |
1944 | * snapshot so that it will have an empty intent log. | |
1945 | * | |
1946 | * Long holds are not really intended to be used the way we do here -- | |
1947 | * held for such a short time. A concurrent caller of dsl_dataset_long_held() | |
1948 | * could fail. Therefore we take pains to only put a long hold if it is | |
1949 | * actually necessary. Fortunately, it will only be necessary if the | |
1950 | * objset is currently mounted (or the ZVOL equivalent). In that case it | |
1951 | * will already have a long hold, so we are not really making things any worse. | |
1952 | * | |
1953 | * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or | |
1954 | * zvol_state_t), and use their mechanism to prevent their hold from being | |
1955 | * dropped (e.g. VFS_HOLD()). However, that would be even more pain for | |
1956 | * very little gain. | |
1957 | * | |
1958 | * if cookiep == NULL, this does both the suspend & resume. | |
1959 | * Otherwise, it returns with the dataset "long held", and the cookie | |
1960 | * should be passed into zil_resume(). | |
34dc7c2f BB |
1961 | */ |
1962 | int | |
13fe0198 | 1963 | zil_suspend(const char *osname, void **cookiep) |
34dc7c2f | 1964 | { |
13fe0198 MA |
1965 | objset_t *os; |
1966 | zilog_t *zilog; | |
1967 | const zil_header_t *zh; | |
1968 | int error; | |
1969 | ||
1970 | error = dmu_objset_hold(osname, suspend_tag, &os); | |
1971 | if (error != 0) | |
1972 | return (error); | |
1973 | zilog = dmu_objset_zil(os); | |
34dc7c2f BB |
1974 | |
1975 | mutex_enter(&zilog->zl_lock); | |
13fe0198 MA |
1976 | zh = zilog->zl_header; |
1977 | ||
9babb374 | 1978 | if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ |
34dc7c2f | 1979 | mutex_exit(&zilog->zl_lock); |
13fe0198 | 1980 | dmu_objset_rele(os, suspend_tag); |
2e528b49 | 1981 | return (SET_ERROR(EBUSY)); |
34dc7c2f | 1982 | } |
13fe0198 MA |
1983 | |
1984 | /* | |
1985 | * Don't put a long hold in the cases where we can avoid it. This | |
1986 | * is when there is no cookie so we are doing a suspend & resume | |
1987 | * (i.e. called from zil_vdev_offline()), and there's nothing to do | |
1988 | * for the suspend because it's already suspended, or there's no ZIL. | |
1989 | */ | |
1990 | if (cookiep == NULL && !zilog->zl_suspending && | |
1991 | (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { | |
1992 | mutex_exit(&zilog->zl_lock); | |
1993 | dmu_objset_rele(os, suspend_tag); | |
1994 | return (0); | |
1995 | } | |
1996 | ||
1997 | dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); | |
1998 | dsl_pool_rele(dmu_objset_pool(os), suspend_tag); | |
1999 | ||
2000 | zilog->zl_suspend++; | |
2001 | ||
2002 | if (zilog->zl_suspend > 1) { | |
34dc7c2f | 2003 | /* |
13fe0198 | 2004 | * Someone else is already suspending it. |
34dc7c2f BB |
2005 | * Just wait for them to finish. |
2006 | */ | |
13fe0198 | 2007 | |
34dc7c2f BB |
2008 | while (zilog->zl_suspending) |
2009 | cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); | |
34dc7c2f | 2010 | mutex_exit(&zilog->zl_lock); |
13fe0198 MA |
2011 | |
2012 | if (cookiep == NULL) | |
2013 | zil_resume(os); | |
2014 | else | |
2015 | *cookiep = os; | |
2016 | return (0); | |
2017 | } | |
2018 | ||
2019 | /* | |
2020 | * If there is no pointer to an on-disk block, this ZIL must not | |
2021 | * be active (e.g. filesystem not mounted), so there's nothing | |
2022 | * to clean up. | |
2023 | */ | |
2024 | if (BP_IS_HOLE(&zh->zh_log)) { | |
2025 | ASSERT(cookiep != NULL); /* fast path already handled */ | |
2026 | ||
2027 | *cookiep = os; | |
2028 | mutex_exit(&zilog->zl_lock); | |
34dc7c2f BB |
2029 | return (0); |
2030 | } | |
13fe0198 | 2031 | |
34dc7c2f BB |
2032 | zilog->zl_suspending = B_TRUE; |
2033 | mutex_exit(&zilog->zl_lock); | |
2034 | ||
572e2857 | 2035 | zil_commit(zilog, 0); |
34dc7c2f BB |
2036 | |
2037 | zil_destroy(zilog, B_FALSE); | |
2038 | ||
2039 | mutex_enter(&zilog->zl_lock); | |
2040 | zilog->zl_suspending = B_FALSE; | |
2041 | cv_broadcast(&zilog->zl_cv_suspend); | |
2042 | mutex_exit(&zilog->zl_lock); | |
2043 | ||
13fe0198 MA |
2044 | if (cookiep == NULL) |
2045 | zil_resume(os); | |
2046 | else | |
2047 | *cookiep = os; | |
34dc7c2f BB |
2048 | return (0); |
2049 | } | |
2050 | ||
2051 | void | |
13fe0198 | 2052 | zil_resume(void *cookie) |
34dc7c2f | 2053 | { |
13fe0198 MA |
2054 | objset_t *os = cookie; |
2055 | zilog_t *zilog = dmu_objset_zil(os); | |
2056 | ||
34dc7c2f BB |
2057 | mutex_enter(&zilog->zl_lock); |
2058 | ASSERT(zilog->zl_suspend != 0); | |
2059 | zilog->zl_suspend--; | |
2060 | mutex_exit(&zilog->zl_lock); | |
13fe0198 MA |
2061 | dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); |
2062 | dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); | |
34dc7c2f BB |
2063 | } |
2064 | ||
2065 | typedef struct zil_replay_arg { | |
b01615d5 | 2066 | zil_replay_func_t *zr_replay; |
34dc7c2f | 2067 | void *zr_arg; |
34dc7c2f | 2068 | boolean_t zr_byteswap; |
428870ff | 2069 | char *zr_lr; |
34dc7c2f BB |
2070 | } zil_replay_arg_t; |
2071 | ||
428870ff BB |
2072 | static int |
2073 | zil_replay_error(zilog_t *zilog, lr_t *lr, int error) | |
2074 | { | |
2075 | char name[MAXNAMELEN]; | |
2076 | ||
2077 | zilog->zl_replaying_seq--; /* didn't actually replay this one */ | |
2078 | ||
2079 | dmu_objset_name(zilog->zl_os, name); | |
2080 | ||
2081 | cmn_err(CE_WARN, "ZFS replay transaction error %d, " | |
2082 | "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, | |
2083 | (u_longlong_t)lr->lrc_seq, | |
2084 | (u_longlong_t)(lr->lrc_txtype & ~TX_CI), | |
2085 | (lr->lrc_txtype & TX_CI) ? "CI" : ""); | |
2086 | ||
2087 | return (error); | |
2088 | } | |
2089 | ||
2090 | static int | |
34dc7c2f BB |
2091 | zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) |
2092 | { | |
2093 | zil_replay_arg_t *zr = zra; | |
2094 | const zil_header_t *zh = zilog->zl_header; | |
2095 | uint64_t reclen = lr->lrc_reclen; | |
2096 | uint64_t txtype = lr->lrc_txtype; | |
428870ff | 2097 | int error = 0; |
34dc7c2f | 2098 | |
428870ff | 2099 | zilog->zl_replaying_seq = lr->lrc_seq; |
34dc7c2f BB |
2100 | |
2101 | if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ | |
428870ff BB |
2102 | return (0); |
2103 | ||
2104 | if (lr->lrc_txg < claim_txg) /* already committed */ | |
2105 | return (0); | |
34dc7c2f BB |
2106 | |
2107 | /* Strip case-insensitive bit, still present in log record */ | |
2108 | txtype &= ~TX_CI; | |
2109 | ||
428870ff BB |
2110 | if (txtype == 0 || txtype >= TX_MAX_TYPE) |
2111 | return (zil_replay_error(zilog, lr, EINVAL)); | |
2112 | ||
2113 | /* | |
2114 | * If this record type can be logged out of order, the object | |
2115 | * (lr_foid) may no longer exist. That's legitimate, not an error. | |
2116 | */ | |
2117 | if (TX_OOO(txtype)) { | |
2118 | error = dmu_object_info(zilog->zl_os, | |
2119 | ((lr_ooo_t *)lr)->lr_foid, NULL); | |
2120 | if (error == ENOENT || error == EEXIST) | |
2121 | return (0); | |
fb5f0bc8 BB |
2122 | } |
2123 | ||
34dc7c2f BB |
2124 | /* |
2125 | * Make a copy of the data so we can revise and extend it. | |
2126 | */ | |
428870ff BB |
2127 | bcopy(lr, zr->zr_lr, reclen); |
2128 | ||
2129 | /* | |
2130 | * If this is a TX_WRITE with a blkptr, suck in the data. | |
2131 | */ | |
2132 | if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { | |
2133 | error = zil_read_log_data(zilog, (lr_write_t *)lr, | |
2134 | zr->zr_lr + reclen); | |
13fe0198 | 2135 | if (error != 0) |
428870ff BB |
2136 | return (zil_replay_error(zilog, lr, error)); |
2137 | } | |
34dc7c2f BB |
2138 | |
2139 | /* | |
2140 | * The log block containing this lr may have been byteswapped | |
2141 | * so that we can easily examine common fields like lrc_txtype. | |
428870ff | 2142 | * However, the log is a mix of different record types, and only the |
34dc7c2f BB |
2143 | * replay vectors know how to byteswap their records. Therefore, if |
2144 | * the lr was byteswapped, undo it before invoking the replay vector. | |
2145 | */ | |
2146 | if (zr->zr_byteswap) | |
428870ff | 2147 | byteswap_uint64_array(zr->zr_lr, reclen); |
34dc7c2f BB |
2148 | |
2149 | /* | |
2150 | * We must now do two things atomically: replay this log record, | |
fb5f0bc8 BB |
2151 | * and update the log header sequence number to reflect the fact that |
2152 | * we did so. At the end of each replay function the sequence number | |
2153 | * is updated if we are in replay mode. | |
34dc7c2f | 2154 | */ |
428870ff | 2155 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); |
13fe0198 | 2156 | if (error != 0) { |
34dc7c2f BB |
2157 | /* |
2158 | * The DMU's dnode layer doesn't see removes until the txg | |
2159 | * commits, so a subsequent claim can spuriously fail with | |
fb5f0bc8 | 2160 | * EEXIST. So if we receive any error we try syncing out |
428870ff BB |
2161 | * any removes then retry the transaction. Note that we |
2162 | * specify B_FALSE for byteswap now, so we don't do it twice. | |
34dc7c2f | 2163 | */ |
428870ff BB |
2164 | txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); |
2165 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); | |
13fe0198 | 2166 | if (error != 0) |
428870ff | 2167 | return (zil_replay_error(zilog, lr, error)); |
34dc7c2f | 2168 | } |
428870ff | 2169 | return (0); |
34dc7c2f BB |
2170 | } |
2171 | ||
2172 | /* ARGSUSED */ | |
428870ff | 2173 | static int |
34dc7c2f BB |
2174 | zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) |
2175 | { | |
2176 | zilog->zl_replay_blks++; | |
428870ff BB |
2177 | |
2178 | return (0); | |
34dc7c2f BB |
2179 | } |
2180 | ||
2181 | /* | |
2182 | * If this dataset has a non-empty intent log, replay it and destroy it. | |
2183 | */ | |
2184 | void | |
b01615d5 | 2185 | zil_replay(objset_t *os, void *arg, zil_replay_func_t replay_func[TX_MAX_TYPE]) |
34dc7c2f BB |
2186 | { |
2187 | zilog_t *zilog = dmu_objset_zil(os); | |
2188 | const zil_header_t *zh = zilog->zl_header; | |
2189 | zil_replay_arg_t zr; | |
2190 | ||
9babb374 | 2191 | if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { |
34dc7c2f BB |
2192 | zil_destroy(zilog, B_TRUE); |
2193 | return; | |
2194 | } | |
2195 | ||
34dc7c2f BB |
2196 | zr.zr_replay = replay_func; |
2197 | zr.zr_arg = arg; | |
34dc7c2f | 2198 | zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); |
b8d06fca | 2199 | zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_PUSHPAGE); |
34dc7c2f BB |
2200 | |
2201 | /* | |
2202 | * Wait for in-progress removes to sync before starting replay. | |
2203 | */ | |
2204 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
2205 | ||
fb5f0bc8 | 2206 | zilog->zl_replay = B_TRUE; |
428870ff | 2207 | zilog->zl_replay_time = ddi_get_lbolt(); |
34dc7c2f BB |
2208 | ASSERT(zilog->zl_replay_blks == 0); |
2209 | (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, | |
2210 | zh->zh_claim_txg); | |
00b46022 | 2211 | vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); |
34dc7c2f BB |
2212 | |
2213 | zil_destroy(zilog, B_FALSE); | |
2214 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
fb5f0bc8 | 2215 | zilog->zl_replay = B_FALSE; |
34dc7c2f BB |
2216 | } |
2217 | ||
428870ff BB |
2218 | boolean_t |
2219 | zil_replaying(zilog_t *zilog, dmu_tx_t *tx) | |
34dc7c2f | 2220 | { |
428870ff BB |
2221 | if (zilog->zl_sync == ZFS_SYNC_DISABLED) |
2222 | return (B_TRUE); | |
34dc7c2f | 2223 | |
428870ff BB |
2224 | if (zilog->zl_replay) { |
2225 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
2226 | zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = | |
2227 | zilog->zl_replaying_seq; | |
2228 | return (B_TRUE); | |
34dc7c2f BB |
2229 | } |
2230 | ||
428870ff | 2231 | return (B_FALSE); |
34dc7c2f | 2232 | } |
9babb374 BB |
2233 | |
2234 | /* ARGSUSED */ | |
2235 | int | |
428870ff | 2236 | zil_vdev_offline(const char *osname, void *arg) |
9babb374 | 2237 | { |
9babb374 BB |
2238 | int error; |
2239 | ||
13fe0198 MA |
2240 | error = zil_suspend(osname, NULL); |
2241 | if (error != 0) | |
2e528b49 | 2242 | return (SET_ERROR(EEXIST)); |
13fe0198 | 2243 | return (0); |
9babb374 | 2244 | } |
c409e464 BB |
2245 | |
2246 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
2247 | module_param(zil_replay_disable, int, 0644); | |
2248 | MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay"); | |
2249 | ||
2250 | module_param(zfs_nocacheflush, int, 0644); | |
2251 | MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes"); | |
ee191e80 ED |
2252 | |
2253 | module_param(zil_slog_limit, ulong, 0644); | |
2254 | MODULE_PARM_DESC(zil_slog_limit, "Max commit bytes to separate log device"); | |
c409e464 | 2255 | #endif |