]>
Commit | Line | Data |
---|---|---|
70e083d2 TG |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. | |
23 | * Copyright (c) 2011, 2014 by Delphix. All rights reserved. | |
24 | */ | |
25 | ||
26 | /* Portions Copyright 2010 Robert Milkowski */ | |
27 | ||
28 | #include <sys/zfs_context.h> | |
29 | #include <sys/spa.h> | |
30 | #include <sys/dmu.h> | |
31 | #include <sys/zap.h> | |
32 | #include <sys/arc.h> | |
33 | #include <sys/stat.h> | |
34 | #include <sys/resource.h> | |
35 | #include <sys/zil.h> | |
36 | #include <sys/zil_impl.h> | |
37 | #include <sys/dsl_dataset.h> | |
38 | #include <sys/vdev_impl.h> | |
39 | #include <sys/dmu_tx.h> | |
40 | #include <sys/dsl_pool.h> | |
41 | #include <sys/metaslab.h> | |
42 | #include <sys/trace_zil.h> | |
43 | ||
44 | /* | |
45 | * The zfs intent log (ZIL) saves transaction records of system calls | |
46 | * that change the file system in memory with enough information | |
47 | * to be able to replay them. These are stored in memory until | |
48 | * either the DMU transaction group (txg) commits them to the stable pool | |
49 | * and they can be discarded, or they are flushed to the stable log | |
50 | * (also in the pool) due to a fsync, O_DSYNC or other synchronous | |
51 | * requirement. In the event of a panic or power fail then those log | |
52 | * records (transactions) are replayed. | |
53 | * | |
54 | * There is one ZIL per file system. Its on-disk (pool) format consists | |
55 | * of 3 parts: | |
56 | * | |
57 | * - ZIL header | |
58 | * - ZIL blocks | |
59 | * - ZIL records | |
60 | * | |
61 | * A log record holds a system call transaction. Log blocks can | |
62 | * hold many log records and the blocks are chained together. | |
63 | * Each ZIL block contains a block pointer (blkptr_t) to the next | |
64 | * ZIL block in the chain. The ZIL header points to the first | |
65 | * block in the chain. Note there is not a fixed place in the pool | |
66 | * to hold blocks. They are dynamically allocated and freed as | |
67 | * needed from the blocks available. Figure X shows the ZIL structure: | |
68 | */ | |
69 | ||
70 | /* | |
71 | * See zil.h for more information about these fields. | |
72 | */ | |
73 | zil_stats_t zil_stats = { | |
74 | { "zil_commit_count", KSTAT_DATA_UINT64 }, | |
75 | { "zil_commit_writer_count", KSTAT_DATA_UINT64 }, | |
76 | { "zil_itx_count", KSTAT_DATA_UINT64 }, | |
77 | { "zil_itx_indirect_count", KSTAT_DATA_UINT64 }, | |
78 | { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 }, | |
79 | { "zil_itx_copied_count", KSTAT_DATA_UINT64 }, | |
80 | { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 }, | |
81 | { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 }, | |
82 | { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 }, | |
83 | { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 }, | |
84 | { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 }, | |
85 | { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 }, | |
86 | { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 }, | |
87 | }; | |
88 | ||
89 | static kstat_t *zil_ksp; | |
90 | ||
91 | /* | |
92 | * Disable intent logging replay. This global ZIL switch affects all pools. | |
93 | */ | |
94 | int zil_replay_disable = 0; | |
95 | ||
96 | /* | |
97 | * Tunable parameter for debugging or performance analysis. Setting | |
98 | * zfs_nocacheflush will cause corruption on power loss if a volatile | |
99 | * out-of-order write cache is enabled. | |
100 | */ | |
101 | int zfs_nocacheflush = 0; | |
102 | ||
103 | static kmem_cache_t *zil_lwb_cache; | |
104 | ||
105 | static void zil_async_to_sync(zilog_t *zilog, uint64_t foid); | |
106 | ||
107 | #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ | |
108 | sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) | |
109 | ||
110 | ||
111 | /* | |
112 | * ziltest is by and large an ugly hack, but very useful in | |
113 | * checking replay without tedious work. | |
114 | * When running ziltest we want to keep all itx's and so maintain | |
115 | * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG | |
116 | * We subtract TXG_CONCURRENT_STATES to allow for common code. | |
117 | */ | |
118 | #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) | |
119 | ||
120 | static int | |
121 | zil_bp_compare(const void *x1, const void *x2) | |
122 | { | |
123 | const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; | |
124 | const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; | |
125 | ||
126 | if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) | |
127 | return (-1); | |
128 | if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) | |
129 | return (1); | |
130 | ||
131 | if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) | |
132 | return (-1); | |
133 | if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) | |
134 | return (1); | |
135 | ||
136 | return (0); | |
137 | } | |
138 | ||
139 | static void | |
140 | zil_bp_tree_init(zilog_t *zilog) | |
141 | { | |
142 | avl_create(&zilog->zl_bp_tree, zil_bp_compare, | |
143 | sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); | |
144 | } | |
145 | ||
146 | static void | |
147 | zil_bp_tree_fini(zilog_t *zilog) | |
148 | { | |
149 | avl_tree_t *t = &zilog->zl_bp_tree; | |
150 | zil_bp_node_t *zn; | |
151 | void *cookie = NULL; | |
152 | ||
153 | while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) | |
154 | kmem_free(zn, sizeof (zil_bp_node_t)); | |
155 | ||
156 | avl_destroy(t); | |
157 | } | |
158 | ||
159 | int | |
160 | zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) | |
161 | { | |
162 | avl_tree_t *t = &zilog->zl_bp_tree; | |
163 | const dva_t *dva; | |
164 | zil_bp_node_t *zn; | |
165 | avl_index_t where; | |
166 | ||
167 | if (BP_IS_EMBEDDED(bp)) | |
168 | return (0); | |
169 | ||
170 | dva = BP_IDENTITY(bp); | |
171 | ||
172 | if (avl_find(t, dva, &where) != NULL) | |
173 | return (SET_ERROR(EEXIST)); | |
174 | ||
175 | zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); | |
176 | zn->zn_dva = *dva; | |
177 | avl_insert(t, zn, where); | |
178 | ||
179 | return (0); | |
180 | } | |
181 | ||
182 | static zil_header_t * | |
183 | zil_header_in_syncing_context(zilog_t *zilog) | |
184 | { | |
185 | return ((zil_header_t *)zilog->zl_header); | |
186 | } | |
187 | ||
188 | static void | |
189 | zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) | |
190 | { | |
191 | zio_cksum_t *zc = &bp->blk_cksum; | |
192 | ||
193 | zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); | |
194 | zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); | |
195 | zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); | |
196 | zc->zc_word[ZIL_ZC_SEQ] = 1ULL; | |
197 | } | |
198 | ||
199 | /* | |
200 | * Read a log block and make sure it's valid. | |
201 | */ | |
202 | static int | |
203 | zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, | |
204 | char **end) | |
205 | { | |
206 | enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; | |
207 | arc_flags_t aflags = ARC_FLAG_WAIT; | |
208 | arc_buf_t *abuf = NULL; | |
209 | zbookmark_phys_t zb; | |
210 | int error; | |
211 | ||
212 | if (zilog->zl_header->zh_claim_txg == 0) | |
213 | zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; | |
214 | ||
215 | if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) | |
216 | zio_flags |= ZIO_FLAG_SPECULATIVE; | |
217 | ||
218 | SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], | |
219 | ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
220 | ||
221 | error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, | |
222 | ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); | |
223 | ||
224 | if (error == 0) { | |
225 | zio_cksum_t cksum = bp->blk_cksum; | |
226 | ||
227 | /* | |
228 | * Validate the checksummed log block. | |
229 | * | |
230 | * Sequence numbers should be... sequential. The checksum | |
231 | * verifier for the next block should be bp's checksum plus 1. | |
232 | * | |
233 | * Also check the log chain linkage and size used. | |
234 | */ | |
235 | cksum.zc_word[ZIL_ZC_SEQ]++; | |
236 | ||
237 | if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { | |
238 | zil_chain_t *zilc = abuf->b_data; | |
239 | char *lr = (char *)(zilc + 1); | |
240 | uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); | |
241 | ||
242 | if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, | |
243 | sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { | |
244 | error = SET_ERROR(ECKSUM); | |
245 | } else { | |
246 | ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); | |
247 | bcopy(lr, dst, len); | |
248 | *end = (char *)dst + len; | |
249 | *nbp = zilc->zc_next_blk; | |
250 | } | |
251 | } else { | |
252 | char *lr = abuf->b_data; | |
253 | uint64_t size = BP_GET_LSIZE(bp); | |
254 | zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; | |
255 | ||
256 | if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, | |
257 | sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || | |
258 | (zilc->zc_nused > (size - sizeof (*zilc)))) { | |
259 | error = SET_ERROR(ECKSUM); | |
260 | } else { | |
261 | ASSERT3U(zilc->zc_nused, <=, | |
262 | SPA_OLD_MAXBLOCKSIZE); | |
263 | bcopy(lr, dst, zilc->zc_nused); | |
264 | *end = (char *)dst + zilc->zc_nused; | |
265 | *nbp = zilc->zc_next_blk; | |
266 | } | |
267 | } | |
268 | ||
269 | VERIFY(arc_buf_remove_ref(abuf, &abuf)); | |
270 | } | |
271 | ||
272 | return (error); | |
273 | } | |
274 | ||
275 | /* | |
276 | * Read a TX_WRITE log data block. | |
277 | */ | |
278 | static int | |
279 | zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) | |
280 | { | |
281 | enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; | |
282 | const blkptr_t *bp = &lr->lr_blkptr; | |
283 | arc_flags_t aflags = ARC_FLAG_WAIT; | |
284 | arc_buf_t *abuf = NULL; | |
285 | zbookmark_phys_t zb; | |
286 | int error; | |
287 | ||
288 | if (BP_IS_HOLE(bp)) { | |
289 | if (wbuf != NULL) | |
290 | bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); | |
291 | return (0); | |
292 | } | |
293 | ||
294 | if (zilog->zl_header->zh_claim_txg == 0) | |
295 | zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; | |
296 | ||
297 | SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, | |
298 | ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); | |
299 | ||
300 | error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, | |
301 | ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); | |
302 | ||
303 | if (error == 0) { | |
304 | if (wbuf != NULL) | |
305 | bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); | |
306 | (void) arc_buf_remove_ref(abuf, &abuf); | |
307 | } | |
308 | ||
309 | return (error); | |
310 | } | |
311 | ||
312 | /* | |
313 | * Parse the intent log, and call parse_func for each valid record within. | |
314 | */ | |
315 | int | |
316 | zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, | |
317 | zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) | |
318 | { | |
319 | const zil_header_t *zh = zilog->zl_header; | |
320 | boolean_t claimed = !!zh->zh_claim_txg; | |
321 | uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; | |
322 | uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; | |
323 | uint64_t max_blk_seq = 0; | |
324 | uint64_t max_lr_seq = 0; | |
325 | uint64_t blk_count = 0; | |
326 | uint64_t lr_count = 0; | |
327 | blkptr_t blk, next_blk; | |
328 | char *lrbuf, *lrp; | |
329 | int error = 0; | |
330 | ||
331 | bzero(&next_blk, sizeof (blkptr_t)); | |
332 | ||
333 | /* | |
334 | * Old logs didn't record the maximum zh_claim_lr_seq. | |
335 | */ | |
336 | if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) | |
337 | claim_lr_seq = UINT64_MAX; | |
338 | ||
339 | /* | |
340 | * Starting at the block pointed to by zh_log we read the log chain. | |
341 | * For each block in the chain we strongly check that block to | |
342 | * ensure its validity. We stop when an invalid block is found. | |
343 | * For each block pointer in the chain we call parse_blk_func(). | |
344 | * For each record in each valid block we call parse_lr_func(). | |
345 | * If the log has been claimed, stop if we encounter a sequence | |
346 | * number greater than the highest claimed sequence number. | |
347 | */ | |
348 | lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); | |
349 | zil_bp_tree_init(zilog); | |
350 | ||
351 | for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { | |
352 | uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; | |
353 | int reclen; | |
354 | char *end = NULL; | |
355 | ||
356 | if (blk_seq > claim_blk_seq) | |
357 | break; | |
358 | if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) | |
359 | break; | |
360 | ASSERT3U(max_blk_seq, <, blk_seq); | |
361 | max_blk_seq = blk_seq; | |
362 | blk_count++; | |
363 | ||
364 | if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) | |
365 | break; | |
366 | ||
367 | error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); | |
368 | if (error != 0) | |
369 | break; | |
370 | ||
371 | for (lrp = lrbuf; lrp < end; lrp += reclen) { | |
372 | lr_t *lr = (lr_t *)lrp; | |
373 | reclen = lr->lrc_reclen; | |
374 | ASSERT3U(reclen, >=, sizeof (lr_t)); | |
375 | if (lr->lrc_seq > claim_lr_seq) | |
376 | goto done; | |
377 | if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) | |
378 | goto done; | |
379 | ASSERT3U(max_lr_seq, <, lr->lrc_seq); | |
380 | max_lr_seq = lr->lrc_seq; | |
381 | lr_count++; | |
382 | } | |
383 | } | |
384 | done: | |
385 | zilog->zl_parse_error = error; | |
386 | zilog->zl_parse_blk_seq = max_blk_seq; | |
387 | zilog->zl_parse_lr_seq = max_lr_seq; | |
388 | zilog->zl_parse_blk_count = blk_count; | |
389 | zilog->zl_parse_lr_count = lr_count; | |
390 | ||
391 | ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || | |
392 | (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); | |
393 | ||
394 | zil_bp_tree_fini(zilog); | |
395 | zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); | |
396 | ||
397 | return (error); | |
398 | } | |
399 | ||
400 | static int | |
401 | zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) | |
402 | { | |
403 | /* | |
404 | * Claim log block if not already committed and not already claimed. | |
405 | * If tx == NULL, just verify that the block is claimable. | |
406 | */ | |
407 | if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || | |
408 | zil_bp_tree_add(zilog, bp) != 0) | |
409 | return (0); | |
410 | ||
411 | return (zio_wait(zio_claim(NULL, zilog->zl_spa, | |
412 | tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, | |
413 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); | |
414 | } | |
415 | ||
416 | static int | |
417 | zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) | |
418 | { | |
419 | lr_write_t *lr = (lr_write_t *)lrc; | |
420 | int error; | |
421 | ||
422 | if (lrc->lrc_txtype != TX_WRITE) | |
423 | return (0); | |
424 | ||
425 | /* | |
426 | * If the block is not readable, don't claim it. This can happen | |
427 | * in normal operation when a log block is written to disk before | |
428 | * some of the dmu_sync() blocks it points to. In this case, the | |
429 | * transaction cannot have been committed to anyone (we would have | |
430 | * waited for all writes to be stable first), so it is semantically | |
431 | * correct to declare this the end of the log. | |
432 | */ | |
433 | if (lr->lr_blkptr.blk_birth >= first_txg && | |
434 | (error = zil_read_log_data(zilog, lr, NULL)) != 0) | |
435 | return (error); | |
436 | return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); | |
437 | } | |
438 | ||
439 | /* ARGSUSED */ | |
440 | static int | |
441 | zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) | |
442 | { | |
443 | zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); | |
444 | ||
445 | return (0); | |
446 | } | |
447 | ||
448 | static int | |
449 | zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) | |
450 | { | |
451 | lr_write_t *lr = (lr_write_t *)lrc; | |
452 | blkptr_t *bp = &lr->lr_blkptr; | |
453 | ||
454 | /* | |
455 | * If we previously claimed it, we need to free it. | |
456 | */ | |
457 | if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && | |
458 | bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && | |
459 | !BP_IS_HOLE(bp)) | |
460 | zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); | |
461 | ||
462 | return (0); | |
463 | } | |
464 | ||
465 | static lwb_t * | |
466 | zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg, boolean_t fastwrite) | |
467 | { | |
468 | lwb_t *lwb; | |
469 | ||
470 | lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); | |
471 | lwb->lwb_zilog = zilog; | |
472 | lwb->lwb_blk = *bp; | |
473 | lwb->lwb_fastwrite = fastwrite; | |
474 | lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); | |
475 | lwb->lwb_max_txg = txg; | |
476 | lwb->lwb_zio = NULL; | |
477 | lwb->lwb_tx = NULL; | |
478 | if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { | |
479 | lwb->lwb_nused = sizeof (zil_chain_t); | |
480 | lwb->lwb_sz = BP_GET_LSIZE(bp); | |
481 | } else { | |
482 | lwb->lwb_nused = 0; | |
483 | lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); | |
484 | } | |
485 | ||
486 | mutex_enter(&zilog->zl_lock); | |
487 | list_insert_tail(&zilog->zl_lwb_list, lwb); | |
488 | mutex_exit(&zilog->zl_lock); | |
489 | ||
490 | return (lwb); | |
491 | } | |
492 | ||
493 | /* | |
494 | * Called when we create in-memory log transactions so that we know | |
495 | * to cleanup the itxs at the end of spa_sync(). | |
496 | */ | |
497 | void | |
498 | zilog_dirty(zilog_t *zilog, uint64_t txg) | |
499 | { | |
500 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
501 | dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); | |
502 | ||
503 | if (ds->ds_is_snapshot) | |
504 | panic("dirtying snapshot!"); | |
505 | ||
506 | if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { | |
507 | /* up the hold count until we can be written out */ | |
508 | dmu_buf_add_ref(ds->ds_dbuf, zilog); | |
509 | } | |
510 | } | |
511 | ||
512 | boolean_t | |
513 | zilog_is_dirty(zilog_t *zilog) | |
514 | { | |
515 | dsl_pool_t *dp = zilog->zl_dmu_pool; | |
516 | int t; | |
517 | ||
518 | for (t = 0; t < TXG_SIZE; t++) { | |
519 | if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) | |
520 | return (B_TRUE); | |
521 | } | |
522 | return (B_FALSE); | |
523 | } | |
524 | ||
525 | /* | |
526 | * Create an on-disk intent log. | |
527 | */ | |
528 | static lwb_t * | |
529 | zil_create(zilog_t *zilog) | |
530 | { | |
531 | const zil_header_t *zh = zilog->zl_header; | |
532 | lwb_t *lwb = NULL; | |
533 | uint64_t txg = 0; | |
534 | dmu_tx_t *tx = NULL; | |
535 | blkptr_t blk; | |
536 | int error = 0; | |
537 | boolean_t fastwrite = FALSE; | |
538 | ||
539 | /* | |
540 | * Wait for any previous destroy to complete. | |
541 | */ | |
542 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
543 | ||
544 | ASSERT(zh->zh_claim_txg == 0); | |
545 | ASSERT(zh->zh_replay_seq == 0); | |
546 | ||
547 | blk = zh->zh_log; | |
548 | ||
549 | /* | |
550 | * Allocate an initial log block if: | |
551 | * - there isn't one already | |
552 | * - the existing block is the wrong endianess | |
553 | */ | |
554 | if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { | |
555 | tx = dmu_tx_create(zilog->zl_os); | |
556 | VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); | |
557 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
558 | txg = dmu_tx_get_txg(tx); | |
559 | ||
560 | if (!BP_IS_HOLE(&blk)) { | |
561 | zio_free_zil(zilog->zl_spa, txg, &blk); | |
562 | BP_ZERO(&blk); | |
563 | } | |
564 | ||
565 | error = zio_alloc_zil(zilog->zl_spa, txg, &blk, | |
566 | ZIL_MIN_BLKSZ, B_TRUE); | |
567 | fastwrite = TRUE; | |
568 | ||
569 | if (error == 0) | |
570 | zil_init_log_chain(zilog, &blk); | |
571 | } | |
572 | ||
573 | /* | |
574 | * Allocate a log write buffer (lwb) for the first log block. | |
575 | */ | |
576 | if (error == 0) | |
577 | lwb = zil_alloc_lwb(zilog, &blk, txg, fastwrite); | |
578 | ||
579 | /* | |
580 | * If we just allocated the first log block, commit our transaction | |
581 | * and wait for zil_sync() to stuff the block poiner into zh_log. | |
582 | * (zh is part of the MOS, so we cannot modify it in open context.) | |
583 | */ | |
584 | if (tx != NULL) { | |
585 | dmu_tx_commit(tx); | |
586 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
587 | } | |
588 | ||
589 | ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); | |
590 | ||
591 | return (lwb); | |
592 | } | |
593 | ||
594 | /* | |
595 | * In one tx, free all log blocks and clear the log header. | |
596 | * If keep_first is set, then we're replaying a log with no content. | |
597 | * We want to keep the first block, however, so that the first | |
598 | * synchronous transaction doesn't require a txg_wait_synced() | |
599 | * in zil_create(). We don't need to txg_wait_synced() here either | |
600 | * when keep_first is set, because both zil_create() and zil_destroy() | |
601 | * will wait for any in-progress destroys to complete. | |
602 | */ | |
603 | void | |
604 | zil_destroy(zilog_t *zilog, boolean_t keep_first) | |
605 | { | |
606 | const zil_header_t *zh = zilog->zl_header; | |
607 | lwb_t *lwb; | |
608 | dmu_tx_t *tx; | |
609 | uint64_t txg; | |
610 | ||
611 | /* | |
612 | * Wait for any previous destroy to complete. | |
613 | */ | |
614 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
615 | ||
616 | zilog->zl_old_header = *zh; /* debugging aid */ | |
617 | ||
618 | if (BP_IS_HOLE(&zh->zh_log)) | |
619 | return; | |
620 | ||
621 | tx = dmu_tx_create(zilog->zl_os); | |
622 | VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); | |
623 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
624 | txg = dmu_tx_get_txg(tx); | |
625 | ||
626 | mutex_enter(&zilog->zl_lock); | |
627 | ||
628 | ASSERT3U(zilog->zl_destroy_txg, <, txg); | |
629 | zilog->zl_destroy_txg = txg; | |
630 | zilog->zl_keep_first = keep_first; | |
631 | ||
632 | if (!list_is_empty(&zilog->zl_lwb_list)) { | |
633 | ASSERT(zh->zh_claim_txg == 0); | |
634 | VERIFY(!keep_first); | |
635 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { | |
636 | ASSERT(lwb->lwb_zio == NULL); | |
637 | if (lwb->lwb_fastwrite) | |
638 | metaslab_fastwrite_unmark(zilog->zl_spa, | |
639 | &lwb->lwb_blk); | |
640 | list_remove(&zilog->zl_lwb_list, lwb); | |
641 | if (lwb->lwb_buf != NULL) | |
642 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
643 | zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); | |
644 | kmem_cache_free(zil_lwb_cache, lwb); | |
645 | } | |
646 | } else if (!keep_first) { | |
647 | zil_destroy_sync(zilog, tx); | |
648 | } | |
649 | mutex_exit(&zilog->zl_lock); | |
650 | ||
651 | dmu_tx_commit(tx); | |
652 | } | |
653 | ||
654 | void | |
655 | zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) | |
656 | { | |
657 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
658 | (void) zil_parse(zilog, zil_free_log_block, | |
659 | zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); | |
660 | } | |
661 | ||
662 | int | |
663 | zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) | |
664 | { | |
665 | dmu_tx_t *tx = txarg; | |
666 | uint64_t first_txg = dmu_tx_get_txg(tx); | |
667 | zilog_t *zilog; | |
668 | zil_header_t *zh; | |
669 | objset_t *os; | |
670 | int error; | |
671 | ||
672 | error = dmu_objset_own_obj(dp, ds->ds_object, | |
673 | DMU_OST_ANY, B_FALSE, FTAG, &os); | |
674 | if (error != 0) { | |
675 | /* | |
676 | * EBUSY indicates that the objset is inconsistent, in which | |
677 | * case it can not have a ZIL. | |
678 | */ | |
679 | if (error != EBUSY) { | |
680 | cmn_err(CE_WARN, "can't open objset for %llu, error %u", | |
681 | (unsigned long long)ds->ds_object, error); | |
682 | } | |
683 | ||
684 | return (0); | |
685 | } | |
686 | ||
687 | zilog = dmu_objset_zil(os); | |
688 | zh = zil_header_in_syncing_context(zilog); | |
689 | ||
690 | if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { | |
691 | if (!BP_IS_HOLE(&zh->zh_log)) | |
692 | zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); | |
693 | BP_ZERO(&zh->zh_log); | |
694 | dsl_dataset_dirty(dmu_objset_ds(os), tx); | |
695 | dmu_objset_disown(os, FTAG); | |
696 | return (0); | |
697 | } | |
698 | ||
699 | /* | |
700 | * Claim all log blocks if we haven't already done so, and remember | |
701 | * the highest claimed sequence number. This ensures that if we can | |
702 | * read only part of the log now (e.g. due to a missing device), | |
703 | * but we can read the entire log later, we will not try to replay | |
704 | * or destroy beyond the last block we successfully claimed. | |
705 | */ | |
706 | ASSERT3U(zh->zh_claim_txg, <=, first_txg); | |
707 | if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { | |
708 | (void) zil_parse(zilog, zil_claim_log_block, | |
709 | zil_claim_log_record, tx, first_txg); | |
710 | zh->zh_claim_txg = first_txg; | |
711 | zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; | |
712 | zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; | |
713 | if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) | |
714 | zh->zh_flags |= ZIL_REPLAY_NEEDED; | |
715 | zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; | |
716 | dsl_dataset_dirty(dmu_objset_ds(os), tx); | |
717 | } | |
718 | ||
719 | ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); | |
720 | dmu_objset_disown(os, FTAG); | |
721 | return (0); | |
722 | } | |
723 | ||
724 | /* | |
725 | * Check the log by walking the log chain. | |
726 | * Checksum errors are ok as they indicate the end of the chain. | |
727 | * Any other error (no device or read failure) returns an error. | |
728 | */ | |
729 | /* ARGSUSED */ | |
730 | int | |
731 | zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) | |
732 | { | |
733 | zilog_t *zilog; | |
734 | objset_t *os; | |
735 | blkptr_t *bp; | |
736 | int error; | |
737 | ||
738 | ASSERT(tx == NULL); | |
739 | ||
740 | error = dmu_objset_from_ds(ds, &os); | |
741 | if (error != 0) { | |
742 | cmn_err(CE_WARN, "can't open objset %llu, error %d", | |
743 | (unsigned long long)ds->ds_object, error); | |
744 | return (0); | |
745 | } | |
746 | ||
747 | zilog = dmu_objset_zil(os); | |
748 | bp = (blkptr_t *)&zilog->zl_header->zh_log; | |
749 | ||
750 | /* | |
751 | * Check the first block and determine if it's on a log device | |
752 | * which may have been removed or faulted prior to loading this | |
753 | * pool. If so, there's no point in checking the rest of the log | |
754 | * as its content should have already been synced to the pool. | |
755 | */ | |
756 | if (!BP_IS_HOLE(bp)) { | |
757 | vdev_t *vd; | |
758 | boolean_t valid = B_TRUE; | |
759 | ||
760 | spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); | |
761 | vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); | |
762 | if (vd->vdev_islog && vdev_is_dead(vd)) | |
763 | valid = vdev_log_state_valid(vd); | |
764 | spa_config_exit(os->os_spa, SCL_STATE, FTAG); | |
765 | ||
766 | if (!valid) | |
767 | return (0); | |
768 | } | |
769 | ||
770 | /* | |
771 | * Because tx == NULL, zil_claim_log_block() will not actually claim | |
772 | * any blocks, but just determine whether it is possible to do so. | |
773 | * In addition to checking the log chain, zil_claim_log_block() | |
774 | * will invoke zio_claim() with a done func of spa_claim_notify(), | |
775 | * which will update spa_max_claim_txg. See spa_load() for details. | |
776 | */ | |
777 | error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, | |
778 | zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); | |
779 | ||
780 | return ((error == ECKSUM || error == ENOENT) ? 0 : error); | |
781 | } | |
782 | ||
783 | static int | |
784 | zil_vdev_compare(const void *x1, const void *x2) | |
785 | { | |
786 | const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; | |
787 | const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; | |
788 | ||
789 | if (v1 < v2) | |
790 | return (-1); | |
791 | if (v1 > v2) | |
792 | return (1); | |
793 | ||
794 | return (0); | |
795 | } | |
796 | ||
797 | void | |
798 | zil_add_block(zilog_t *zilog, const blkptr_t *bp) | |
799 | { | |
800 | avl_tree_t *t = &zilog->zl_vdev_tree; | |
801 | avl_index_t where; | |
802 | zil_vdev_node_t *zv, zvsearch; | |
803 | int ndvas = BP_GET_NDVAS(bp); | |
804 | int i; | |
805 | ||
806 | if (zfs_nocacheflush) | |
807 | return; | |
808 | ||
809 | ASSERT(zilog->zl_writer); | |
810 | ||
811 | /* | |
812 | * Even though we're zl_writer, we still need a lock because the | |
813 | * zl_get_data() callbacks may have dmu_sync() done callbacks | |
814 | * that will run concurrently. | |
815 | */ | |
816 | mutex_enter(&zilog->zl_vdev_lock); | |
817 | for (i = 0; i < ndvas; i++) { | |
818 | zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); | |
819 | if (avl_find(t, &zvsearch, &where) == NULL) { | |
820 | zv = kmem_alloc(sizeof (*zv), KM_SLEEP); | |
821 | zv->zv_vdev = zvsearch.zv_vdev; | |
822 | avl_insert(t, zv, where); | |
823 | } | |
824 | } | |
825 | mutex_exit(&zilog->zl_vdev_lock); | |
826 | } | |
827 | ||
828 | static void | |
829 | zil_flush_vdevs(zilog_t *zilog) | |
830 | { | |
831 | spa_t *spa = zilog->zl_spa; | |
832 | avl_tree_t *t = &zilog->zl_vdev_tree; | |
833 | void *cookie = NULL; | |
834 | zil_vdev_node_t *zv; | |
835 | zio_t *zio; | |
836 | ||
837 | ASSERT(zilog->zl_writer); | |
838 | ||
839 | /* | |
840 | * We don't need zl_vdev_lock here because we're the zl_writer, | |
841 | * and all zl_get_data() callbacks are done. | |
842 | */ | |
843 | if (avl_numnodes(t) == 0) | |
844 | return; | |
845 | ||
846 | spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); | |
847 | ||
848 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); | |
849 | ||
850 | while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { | |
851 | vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); | |
852 | if (vd != NULL) | |
853 | zio_flush(zio, vd); | |
854 | kmem_free(zv, sizeof (*zv)); | |
855 | } | |
856 | ||
857 | /* | |
858 | * Wait for all the flushes to complete. Not all devices actually | |
859 | * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. | |
860 | */ | |
861 | (void) zio_wait(zio); | |
862 | ||
863 | spa_config_exit(spa, SCL_STATE, FTAG); | |
864 | } | |
865 | ||
866 | /* | |
867 | * Function called when a log block write completes | |
868 | */ | |
869 | static void | |
870 | zil_lwb_write_done(zio_t *zio) | |
871 | { | |
872 | lwb_t *lwb = zio->io_private; | |
873 | zilog_t *zilog = lwb->lwb_zilog; | |
874 | dmu_tx_t *tx = lwb->lwb_tx; | |
875 | ||
876 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
877 | ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); | |
878 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
879 | ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); | |
880 | ASSERT(!BP_IS_GANG(zio->io_bp)); | |
881 | ASSERT(!BP_IS_HOLE(zio->io_bp)); | |
882 | ASSERT(BP_GET_FILL(zio->io_bp) == 0); | |
883 | ||
884 | /* | |
885 | * Ensure the lwb buffer pointer is cleared before releasing | |
886 | * the txg. If we have had an allocation failure and | |
887 | * the txg is waiting to sync then we want want zil_sync() | |
888 | * to remove the lwb so that it's not picked up as the next new | |
889 | * one in zil_commit_writer(). zil_sync() will only remove | |
890 | * the lwb if lwb_buf is null. | |
891 | */ | |
892 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
893 | mutex_enter(&zilog->zl_lock); | |
894 | lwb->lwb_zio = NULL; | |
895 | lwb->lwb_fastwrite = FALSE; | |
896 | lwb->lwb_buf = NULL; | |
897 | lwb->lwb_tx = NULL; | |
898 | mutex_exit(&zilog->zl_lock); | |
899 | ||
900 | /* | |
901 | * Now that we've written this log block, we have a stable pointer | |
902 | * to the next block in the chain, so it's OK to let the txg in | |
903 | * which we allocated the next block sync. | |
904 | */ | |
905 | dmu_tx_commit(tx); | |
906 | } | |
907 | ||
908 | /* | |
909 | * Initialize the io for a log block. | |
910 | */ | |
911 | static void | |
912 | zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) | |
913 | { | |
914 | zbookmark_phys_t zb; | |
915 | ||
916 | SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], | |
917 | ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, | |
918 | lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); | |
919 | ||
920 | if (zilog->zl_root_zio == NULL) { | |
921 | zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, | |
922 | ZIO_FLAG_CANFAIL); | |
923 | } | |
924 | ||
925 | /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */ | |
926 | mutex_enter(&zilog->zl_lock); | |
927 | if (lwb->lwb_zio == NULL) { | |
928 | if (!lwb->lwb_fastwrite) { | |
929 | metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk); | |
930 | lwb->lwb_fastwrite = 1; | |
931 | } | |
932 | lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, | |
933 | 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), | |
934 | zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, | |
935 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | | |
936 | ZIO_FLAG_FASTWRITE, &zb); | |
937 | } | |
938 | mutex_exit(&zilog->zl_lock); | |
939 | } | |
940 | ||
941 | /* | |
942 | * Define a limited set of intent log block sizes. | |
943 | * | |
944 | * These must be a multiple of 4KB. Note only the amount used (again | |
945 | * aligned to 4KB) actually gets written. However, we can't always just | |
946 | * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. | |
947 | */ | |
948 | uint64_t zil_block_buckets[] = { | |
949 | 4096, /* non TX_WRITE */ | |
950 | 8192+4096, /* data base */ | |
951 | 32*1024 + 4096, /* NFS writes */ | |
952 | UINT64_MAX | |
953 | }; | |
954 | ||
955 | /* | |
956 | * Use the slog as long as the current commit size is less than the | |
957 | * limit or the total list size is less than 2X the limit. Limit | |
958 | * checking is disabled by setting zil_slog_limit to UINT64_MAX. | |
959 | */ | |
960 | unsigned long zil_slog_limit = 1024 * 1024; | |
961 | #define USE_SLOG(zilog) (((zilog)->zl_cur_used < zil_slog_limit) || \ | |
962 | ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))) | |
963 | ||
964 | /* | |
965 | * Start a log block write and advance to the next log block. | |
966 | * Calls are serialized. | |
967 | */ | |
968 | static lwb_t * | |
969 | zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) | |
970 | { | |
971 | lwb_t *nlwb = NULL; | |
972 | zil_chain_t *zilc; | |
973 | spa_t *spa = zilog->zl_spa; | |
974 | blkptr_t *bp; | |
975 | dmu_tx_t *tx; | |
976 | uint64_t txg; | |
977 | uint64_t zil_blksz, wsz; | |
978 | int i, error; | |
979 | boolean_t use_slog; | |
980 | ||
981 | if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { | |
982 | zilc = (zil_chain_t *)lwb->lwb_buf; | |
983 | bp = &zilc->zc_next_blk; | |
984 | } else { | |
985 | zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); | |
986 | bp = &zilc->zc_next_blk; | |
987 | } | |
988 | ||
989 | ASSERT(lwb->lwb_nused <= lwb->lwb_sz); | |
990 | ||
991 | /* | |
992 | * Allocate the next block and save its address in this block | |
993 | * before writing it in order to establish the log chain. | |
994 | * Note that if the allocation of nlwb synced before we wrote | |
995 | * the block that points at it (lwb), we'd leak it if we crashed. | |
996 | * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). | |
997 | * We dirty the dataset to ensure that zil_sync() will be called | |
998 | * to clean up in the event of allocation failure or I/O failure. | |
999 | */ | |
1000 | tx = dmu_tx_create(zilog->zl_os); | |
1001 | VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); | |
1002 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
1003 | txg = dmu_tx_get_txg(tx); | |
1004 | ||
1005 | lwb->lwb_tx = tx; | |
1006 | ||
1007 | /* | |
1008 | * Log blocks are pre-allocated. Here we select the size of the next | |
1009 | * block, based on size used in the last block. | |
1010 | * - first find the smallest bucket that will fit the block from a | |
1011 | * limited set of block sizes. This is because it's faster to write | |
1012 | * blocks allocated from the same metaslab as they are adjacent or | |
1013 | * close. | |
1014 | * - next find the maximum from the new suggested size and an array of | |
1015 | * previous sizes. This lessens a picket fence effect of wrongly | |
1016 | * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k | |
1017 | * requests. | |
1018 | * | |
1019 | * Note we only write what is used, but we can't just allocate | |
1020 | * the maximum block size because we can exhaust the available | |
1021 | * pool log space. | |
1022 | */ | |
1023 | zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); | |
1024 | for (i = 0; zil_blksz > zil_block_buckets[i]; i++) | |
1025 | continue; | |
1026 | zil_blksz = zil_block_buckets[i]; | |
1027 | if (zil_blksz == UINT64_MAX) | |
1028 | zil_blksz = SPA_OLD_MAXBLOCKSIZE; | |
1029 | zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; | |
1030 | for (i = 0; i < ZIL_PREV_BLKS; i++) | |
1031 | zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); | |
1032 | zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); | |
1033 | ||
1034 | BP_ZERO(bp); | |
1035 | use_slog = USE_SLOG(zilog); | |
1036 | error = zio_alloc_zil(spa, txg, bp, zil_blksz, | |
1037 | USE_SLOG(zilog)); | |
1038 | if (use_slog) { | |
1039 | ZIL_STAT_BUMP(zil_itx_metaslab_slog_count); | |
1040 | ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused); | |
1041 | } else { | |
1042 | ZIL_STAT_BUMP(zil_itx_metaslab_normal_count); | |
1043 | ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused); | |
1044 | } | |
1045 | if (error == 0) { | |
1046 | ASSERT3U(bp->blk_birth, ==, txg); | |
1047 | bp->blk_cksum = lwb->lwb_blk.blk_cksum; | |
1048 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; | |
1049 | ||
1050 | /* | |
1051 | * Allocate a new log write buffer (lwb). | |
1052 | */ | |
1053 | nlwb = zil_alloc_lwb(zilog, bp, txg, TRUE); | |
1054 | ||
1055 | /* Record the block for later vdev flushing */ | |
1056 | zil_add_block(zilog, &lwb->lwb_blk); | |
1057 | } | |
1058 | ||
1059 | if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { | |
1060 | /* For Slim ZIL only write what is used. */ | |
1061 | wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); | |
1062 | ASSERT3U(wsz, <=, lwb->lwb_sz); | |
1063 | zio_shrink(lwb->lwb_zio, wsz); | |
1064 | ||
1065 | } else { | |
1066 | wsz = lwb->lwb_sz; | |
1067 | } | |
1068 | ||
1069 | zilc->zc_pad = 0; | |
1070 | zilc->zc_nused = lwb->lwb_nused; | |
1071 | zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; | |
1072 | ||
1073 | /* | |
1074 | * clear unused data for security | |
1075 | */ | |
1076 | bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); | |
1077 | ||
1078 | zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ | |
1079 | ||
1080 | /* | |
1081 | * If there was an allocation failure then nlwb will be null which | |
1082 | * forces a txg_wait_synced(). | |
1083 | */ | |
1084 | return (nlwb); | |
1085 | } | |
1086 | ||
1087 | static lwb_t * | |
1088 | zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) | |
1089 | { | |
1090 | lr_t *lrc = &itx->itx_lr; /* common log record */ | |
1091 | lr_write_t *lrw = (lr_write_t *)lrc; | |
1092 | char *lr_buf; | |
1093 | uint64_t txg = lrc->lrc_txg; | |
1094 | uint64_t reclen = lrc->lrc_reclen; | |
1095 | uint64_t dlen = 0; | |
1096 | ||
1097 | if (lwb == NULL) | |
1098 | return (NULL); | |
1099 | ||
1100 | ASSERT(lwb->lwb_buf != NULL); | |
1101 | ASSERT(zilog_is_dirty(zilog) || | |
1102 | spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); | |
1103 | ||
1104 | if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) | |
1105 | dlen = P2ROUNDUP_TYPED( | |
1106 | lrw->lr_length, sizeof (uint64_t), uint64_t); | |
1107 | ||
1108 | zilog->zl_cur_used += (reclen + dlen); | |
1109 | ||
1110 | zil_lwb_write_init(zilog, lwb); | |
1111 | ||
1112 | /* | |
1113 | * If this record won't fit in the current log block, start a new one. | |
1114 | */ | |
1115 | if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { | |
1116 | lwb = zil_lwb_write_start(zilog, lwb); | |
1117 | if (lwb == NULL) | |
1118 | return (NULL); | |
1119 | zil_lwb_write_init(zilog, lwb); | |
1120 | ASSERT(LWB_EMPTY(lwb)); | |
1121 | if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { | |
1122 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
1123 | return (lwb); | |
1124 | } | |
1125 | } | |
1126 | ||
1127 | lr_buf = lwb->lwb_buf + lwb->lwb_nused; | |
1128 | bcopy(lrc, lr_buf, reclen); | |
1129 | lrc = (lr_t *)lr_buf; | |
1130 | lrw = (lr_write_t *)lrc; | |
1131 | ||
1132 | ZIL_STAT_BUMP(zil_itx_count); | |
1133 | ||
1134 | /* | |
1135 | * If it's a write, fetch the data or get its blkptr as appropriate. | |
1136 | */ | |
1137 | if (lrc->lrc_txtype == TX_WRITE) { | |
1138 | if (txg > spa_freeze_txg(zilog->zl_spa)) | |
1139 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
1140 | if (itx->itx_wr_state == WR_COPIED) { | |
1141 | ZIL_STAT_BUMP(zil_itx_copied_count); | |
1142 | ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length); | |
1143 | } else { | |
1144 | char *dbuf; | |
1145 | int error; | |
1146 | ||
1147 | if (dlen) { | |
1148 | ASSERT(itx->itx_wr_state == WR_NEED_COPY); | |
1149 | dbuf = lr_buf + reclen; | |
1150 | lrw->lr_common.lrc_reclen += dlen; | |
1151 | ZIL_STAT_BUMP(zil_itx_needcopy_count); | |
1152 | ZIL_STAT_INCR(zil_itx_needcopy_bytes, | |
1153 | lrw->lr_length); | |
1154 | } else { | |
1155 | ASSERT(itx->itx_wr_state == WR_INDIRECT); | |
1156 | dbuf = NULL; | |
1157 | ZIL_STAT_BUMP(zil_itx_indirect_count); | |
1158 | ZIL_STAT_INCR(zil_itx_indirect_bytes, | |
1159 | lrw->lr_length); | |
1160 | } | |
1161 | error = zilog->zl_get_data( | |
1162 | itx->itx_private, lrw, dbuf, lwb->lwb_zio); | |
1163 | if (error == EIO) { | |
1164 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
1165 | return (lwb); | |
1166 | } | |
1167 | if (error != 0) { | |
1168 | ASSERT(error == ENOENT || error == EEXIST || | |
1169 | error == EALREADY); | |
1170 | return (lwb); | |
1171 | } | |
1172 | } | |
1173 | } | |
1174 | ||
1175 | /* | |
1176 | * We're actually making an entry, so update lrc_seq to be the | |
1177 | * log record sequence number. Note that this is generally not | |
1178 | * equal to the itx sequence number because not all transactions | |
1179 | * are synchronous, and sometimes spa_sync() gets there first. | |
1180 | */ | |
1181 | lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ | |
1182 | lwb->lwb_nused += reclen + dlen; | |
1183 | lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); | |
1184 | ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); | |
1185 | ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); | |
1186 | ||
1187 | return (lwb); | |
1188 | } | |
1189 | ||
1190 | itx_t * | |
1191 | zil_itx_create(uint64_t txtype, size_t lrsize) | |
1192 | { | |
1193 | itx_t *itx; | |
1194 | ||
1195 | lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); | |
1196 | ||
1197 | itx = zio_data_buf_alloc(offsetof(itx_t, itx_lr) + lrsize); | |
1198 | itx->itx_lr.lrc_txtype = txtype; | |
1199 | itx->itx_lr.lrc_reclen = lrsize; | |
1200 | itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ | |
1201 | itx->itx_lr.lrc_seq = 0; /* defensive */ | |
1202 | itx->itx_sync = B_TRUE; /* default is synchronous */ | |
1203 | itx->itx_callback = NULL; | |
1204 | itx->itx_callback_data = NULL; | |
1205 | ||
1206 | return (itx); | |
1207 | } | |
1208 | ||
1209 | void | |
1210 | zil_itx_destroy(itx_t *itx) | |
1211 | { | |
1212 | zio_data_buf_free(itx, offsetof(itx_t, itx_lr)+itx->itx_lr.lrc_reclen); | |
1213 | } | |
1214 | ||
1215 | /* | |
1216 | * Free up the sync and async itxs. The itxs_t has already been detached | |
1217 | * so no locks are needed. | |
1218 | */ | |
1219 | static void | |
1220 | zil_itxg_clean(itxs_t *itxs) | |
1221 | { | |
1222 | itx_t *itx; | |
1223 | list_t *list; | |
1224 | avl_tree_t *t; | |
1225 | void *cookie; | |
1226 | itx_async_node_t *ian; | |
1227 | ||
1228 | list = &itxs->i_sync_list; | |
1229 | while ((itx = list_head(list)) != NULL) { | |
1230 | if (itx->itx_callback != NULL) | |
1231 | itx->itx_callback(itx->itx_callback_data); | |
1232 | list_remove(list, itx); | |
1233 | zil_itx_destroy(itx); | |
1234 | } | |
1235 | ||
1236 | cookie = NULL; | |
1237 | t = &itxs->i_async_tree; | |
1238 | while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { | |
1239 | list = &ian->ia_list; | |
1240 | while ((itx = list_head(list)) != NULL) { | |
1241 | if (itx->itx_callback != NULL) | |
1242 | itx->itx_callback(itx->itx_callback_data); | |
1243 | list_remove(list, itx); | |
1244 | zil_itx_destroy(itx); | |
1245 | } | |
1246 | list_destroy(list); | |
1247 | kmem_free(ian, sizeof (itx_async_node_t)); | |
1248 | } | |
1249 | avl_destroy(t); | |
1250 | ||
1251 | kmem_free(itxs, sizeof (itxs_t)); | |
1252 | } | |
1253 | ||
1254 | static int | |
1255 | zil_aitx_compare(const void *x1, const void *x2) | |
1256 | { | |
1257 | const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; | |
1258 | const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; | |
1259 | ||
1260 | if (o1 < o2) | |
1261 | return (-1); | |
1262 | if (o1 > o2) | |
1263 | return (1); | |
1264 | ||
1265 | return (0); | |
1266 | } | |
1267 | ||
1268 | /* | |
1269 | * Remove all async itx with the given oid. | |
1270 | */ | |
1271 | static void | |
1272 | zil_remove_async(zilog_t *zilog, uint64_t oid) | |
1273 | { | |
1274 | uint64_t otxg, txg; | |
1275 | itx_async_node_t *ian; | |
1276 | avl_tree_t *t; | |
1277 | avl_index_t where; | |
1278 | list_t clean_list; | |
1279 | itx_t *itx; | |
1280 | ||
1281 | ASSERT(oid != 0); | |
1282 | list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); | |
1283 | ||
1284 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ | |
1285 | otxg = ZILTEST_TXG; | |
1286 | else | |
1287 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
1288 | ||
1289 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { | |
1290 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1291 | ||
1292 | mutex_enter(&itxg->itxg_lock); | |
1293 | if (itxg->itxg_txg != txg) { | |
1294 | mutex_exit(&itxg->itxg_lock); | |
1295 | continue; | |
1296 | } | |
1297 | ||
1298 | /* | |
1299 | * Locate the object node and append its list. | |
1300 | */ | |
1301 | t = &itxg->itxg_itxs->i_async_tree; | |
1302 | ian = avl_find(t, &oid, &where); | |
1303 | if (ian != NULL) | |
1304 | list_move_tail(&clean_list, &ian->ia_list); | |
1305 | mutex_exit(&itxg->itxg_lock); | |
1306 | } | |
1307 | while ((itx = list_head(&clean_list)) != NULL) { | |
1308 | if (itx->itx_callback != NULL) | |
1309 | itx->itx_callback(itx->itx_callback_data); | |
1310 | list_remove(&clean_list, itx); | |
1311 | zil_itx_destroy(itx); | |
1312 | } | |
1313 | list_destroy(&clean_list); | |
1314 | } | |
1315 | ||
1316 | void | |
1317 | zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) | |
1318 | { | |
1319 | uint64_t txg; | |
1320 | itxg_t *itxg; | |
1321 | itxs_t *itxs, *clean = NULL; | |
1322 | ||
1323 | /* | |
1324 | * Object ids can be re-instantiated in the next txg so | |
1325 | * remove any async transactions to avoid future leaks. | |
1326 | * This can happen if a fsync occurs on the re-instantiated | |
1327 | * object for a WR_INDIRECT or WR_NEED_COPY write, which gets | |
1328 | * the new file data and flushes a write record for the old object. | |
1329 | */ | |
1330 | if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) | |
1331 | zil_remove_async(zilog, itx->itx_oid); | |
1332 | ||
1333 | /* | |
1334 | * Ensure the data of a renamed file is committed before the rename. | |
1335 | */ | |
1336 | if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) | |
1337 | zil_async_to_sync(zilog, itx->itx_oid); | |
1338 | ||
1339 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) | |
1340 | txg = ZILTEST_TXG; | |
1341 | else | |
1342 | txg = dmu_tx_get_txg(tx); | |
1343 | ||
1344 | itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1345 | mutex_enter(&itxg->itxg_lock); | |
1346 | itxs = itxg->itxg_itxs; | |
1347 | if (itxg->itxg_txg != txg) { | |
1348 | if (itxs != NULL) { | |
1349 | /* | |
1350 | * The zil_clean callback hasn't got around to cleaning | |
1351 | * this itxg. Save the itxs for release below. | |
1352 | * This should be rare. | |
1353 | */ | |
1354 | atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); | |
1355 | itxg->itxg_sod = 0; | |
1356 | clean = itxg->itxg_itxs; | |
1357 | } | |
1358 | ASSERT(itxg->itxg_sod == 0); | |
1359 | itxg->itxg_txg = txg; | |
1360 | itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), | |
1361 | KM_SLEEP); | |
1362 | ||
1363 | list_create(&itxs->i_sync_list, sizeof (itx_t), | |
1364 | offsetof(itx_t, itx_node)); | |
1365 | avl_create(&itxs->i_async_tree, zil_aitx_compare, | |
1366 | sizeof (itx_async_node_t), | |
1367 | offsetof(itx_async_node_t, ia_node)); | |
1368 | } | |
1369 | if (itx->itx_sync) { | |
1370 | list_insert_tail(&itxs->i_sync_list, itx); | |
1371 | atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); | |
1372 | itxg->itxg_sod += itx->itx_sod; | |
1373 | } else { | |
1374 | avl_tree_t *t = &itxs->i_async_tree; | |
1375 | uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; | |
1376 | itx_async_node_t *ian; | |
1377 | avl_index_t where; | |
1378 | ||
1379 | ian = avl_find(t, &foid, &where); | |
1380 | if (ian == NULL) { | |
1381 | ian = kmem_alloc(sizeof (itx_async_node_t), | |
1382 | KM_SLEEP); | |
1383 | list_create(&ian->ia_list, sizeof (itx_t), | |
1384 | offsetof(itx_t, itx_node)); | |
1385 | ian->ia_foid = foid; | |
1386 | avl_insert(t, ian, where); | |
1387 | } | |
1388 | list_insert_tail(&ian->ia_list, itx); | |
1389 | } | |
1390 | ||
1391 | itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); | |
1392 | zilog_dirty(zilog, txg); | |
1393 | mutex_exit(&itxg->itxg_lock); | |
1394 | ||
1395 | /* Release the old itxs now we've dropped the lock */ | |
1396 | if (clean != NULL) | |
1397 | zil_itxg_clean(clean); | |
1398 | } | |
1399 | ||
1400 | /* | |
1401 | * If there are any in-memory intent log transactions which have now been | |
1402 | * synced then start up a taskq to free them. We should only do this after we | |
1403 | * have written out the uberblocks (i.e. txg has been comitted) so that | |
1404 | * don't inadvertently clean out in-memory log records that would be required | |
1405 | * by zil_commit(). | |
1406 | */ | |
1407 | void | |
1408 | zil_clean(zilog_t *zilog, uint64_t synced_txg) | |
1409 | { | |
1410 | itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; | |
1411 | itxs_t *clean_me; | |
1412 | ||
1413 | mutex_enter(&itxg->itxg_lock); | |
1414 | if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { | |
1415 | mutex_exit(&itxg->itxg_lock); | |
1416 | return; | |
1417 | } | |
1418 | ASSERT3U(itxg->itxg_txg, <=, synced_txg); | |
1419 | ASSERT(itxg->itxg_txg != 0); | |
1420 | ASSERT(zilog->zl_clean_taskq != NULL); | |
1421 | atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); | |
1422 | itxg->itxg_sod = 0; | |
1423 | clean_me = itxg->itxg_itxs; | |
1424 | itxg->itxg_itxs = NULL; | |
1425 | itxg->itxg_txg = 0; | |
1426 | mutex_exit(&itxg->itxg_lock); | |
1427 | /* | |
1428 | * Preferably start a task queue to free up the old itxs but | |
1429 | * if taskq_dispatch can't allocate resources to do that then | |
1430 | * free it in-line. This should be rare. Note, using TQ_SLEEP | |
1431 | * created a bad performance problem. | |
1432 | */ | |
1433 | if (taskq_dispatch(zilog->zl_clean_taskq, | |
1434 | (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) | |
1435 | zil_itxg_clean(clean_me); | |
1436 | } | |
1437 | ||
1438 | /* | |
1439 | * Get the list of itxs to commit into zl_itx_commit_list. | |
1440 | */ | |
1441 | static void | |
1442 | zil_get_commit_list(zilog_t *zilog) | |
1443 | { | |
1444 | uint64_t otxg, txg; | |
1445 | list_t *commit_list = &zilog->zl_itx_commit_list; | |
1446 | uint64_t push_sod = 0; | |
1447 | ||
1448 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ | |
1449 | otxg = ZILTEST_TXG; | |
1450 | else | |
1451 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
1452 | ||
1453 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { | |
1454 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1455 | ||
1456 | mutex_enter(&itxg->itxg_lock); | |
1457 | if (itxg->itxg_txg != txg) { | |
1458 | mutex_exit(&itxg->itxg_lock); | |
1459 | continue; | |
1460 | } | |
1461 | ||
1462 | list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); | |
1463 | push_sod += itxg->itxg_sod; | |
1464 | itxg->itxg_sod = 0; | |
1465 | ||
1466 | mutex_exit(&itxg->itxg_lock); | |
1467 | } | |
1468 | atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); | |
1469 | } | |
1470 | ||
1471 | /* | |
1472 | * Move the async itxs for a specified object to commit into sync lists. | |
1473 | */ | |
1474 | static void | |
1475 | zil_async_to_sync(zilog_t *zilog, uint64_t foid) | |
1476 | { | |
1477 | uint64_t otxg, txg; | |
1478 | itx_async_node_t *ian; | |
1479 | avl_tree_t *t; | |
1480 | avl_index_t where; | |
1481 | ||
1482 | if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ | |
1483 | otxg = ZILTEST_TXG; | |
1484 | else | |
1485 | otxg = spa_last_synced_txg(zilog->zl_spa) + 1; | |
1486 | ||
1487 | for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { | |
1488 | itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; | |
1489 | ||
1490 | mutex_enter(&itxg->itxg_lock); | |
1491 | if (itxg->itxg_txg != txg) { | |
1492 | mutex_exit(&itxg->itxg_lock); | |
1493 | continue; | |
1494 | } | |
1495 | ||
1496 | /* | |
1497 | * If a foid is specified then find that node and append its | |
1498 | * list. Otherwise walk the tree appending all the lists | |
1499 | * to the sync list. We add to the end rather than the | |
1500 | * beginning to ensure the create has happened. | |
1501 | */ | |
1502 | t = &itxg->itxg_itxs->i_async_tree; | |
1503 | if (foid != 0) { | |
1504 | ian = avl_find(t, &foid, &where); | |
1505 | if (ian != NULL) { | |
1506 | list_move_tail(&itxg->itxg_itxs->i_sync_list, | |
1507 | &ian->ia_list); | |
1508 | } | |
1509 | } else { | |
1510 | void *cookie = NULL; | |
1511 | ||
1512 | while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { | |
1513 | list_move_tail(&itxg->itxg_itxs->i_sync_list, | |
1514 | &ian->ia_list); | |
1515 | list_destroy(&ian->ia_list); | |
1516 | kmem_free(ian, sizeof (itx_async_node_t)); | |
1517 | } | |
1518 | } | |
1519 | mutex_exit(&itxg->itxg_lock); | |
1520 | } | |
1521 | } | |
1522 | ||
1523 | static void | |
1524 | zil_commit_writer(zilog_t *zilog) | |
1525 | { | |
1526 | uint64_t txg; | |
1527 | itx_t *itx; | |
1528 | lwb_t *lwb; | |
1529 | spa_t *spa = zilog->zl_spa; | |
1530 | int error = 0; | |
1531 | ||
1532 | ASSERT(zilog->zl_root_zio == NULL); | |
1533 | ||
1534 | mutex_exit(&zilog->zl_lock); | |
1535 | ||
1536 | zil_get_commit_list(zilog); | |
1537 | ||
1538 | /* | |
1539 | * Return if there's nothing to commit before we dirty the fs by | |
1540 | * calling zil_create(). | |
1541 | */ | |
1542 | if (list_head(&zilog->zl_itx_commit_list) == NULL) { | |
1543 | mutex_enter(&zilog->zl_lock); | |
1544 | return; | |
1545 | } | |
1546 | ||
1547 | if (zilog->zl_suspend) { | |
1548 | lwb = NULL; | |
1549 | } else { | |
1550 | lwb = list_tail(&zilog->zl_lwb_list); | |
1551 | if (lwb == NULL) | |
1552 | lwb = zil_create(zilog); | |
1553 | } | |
1554 | ||
1555 | DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); | |
1556 | for (itx = list_head(&zilog->zl_itx_commit_list); itx != NULL; | |
1557 | itx = list_next(&zilog->zl_itx_commit_list, itx)) { | |
1558 | txg = itx->itx_lr.lrc_txg; | |
1559 | ASSERT(txg); | |
1560 | ||
1561 | if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) | |
1562 | lwb = zil_lwb_commit(zilog, itx, lwb); | |
1563 | } | |
1564 | DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); | |
1565 | ||
1566 | /* write the last block out */ | |
1567 | if (lwb != NULL && lwb->lwb_zio != NULL) | |
1568 | lwb = zil_lwb_write_start(zilog, lwb); | |
1569 | ||
1570 | zilog->zl_cur_used = 0; | |
1571 | ||
1572 | /* | |
1573 | * Wait if necessary for the log blocks to be on stable storage. | |
1574 | */ | |
1575 | if (zilog->zl_root_zio) { | |
1576 | error = zio_wait(zilog->zl_root_zio); | |
1577 | zilog->zl_root_zio = NULL; | |
1578 | zil_flush_vdevs(zilog); | |
1579 | } | |
1580 | ||
1581 | if (error || lwb == NULL) | |
1582 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
1583 | ||
1584 | while ((itx = list_head(&zilog->zl_itx_commit_list))) { | |
1585 | txg = itx->itx_lr.lrc_txg; | |
1586 | ASSERT(txg); | |
1587 | ||
1588 | if (itx->itx_callback != NULL) | |
1589 | itx->itx_callback(itx->itx_callback_data); | |
1590 | list_remove(&zilog->zl_itx_commit_list, itx); | |
1591 | zil_itx_destroy(itx); | |
1592 | } | |
1593 | ||
1594 | mutex_enter(&zilog->zl_lock); | |
1595 | ||
1596 | /* | |
1597 | * Remember the highest committed log sequence number for ztest. | |
1598 | * We only update this value when all the log writes succeeded, | |
1599 | * because ztest wants to ASSERT that it got the whole log chain. | |
1600 | */ | |
1601 | if (error == 0 && lwb != NULL) | |
1602 | zilog->zl_commit_lr_seq = zilog->zl_lr_seq; | |
1603 | } | |
1604 | ||
1605 | /* | |
1606 | * Commit zfs transactions to stable storage. | |
1607 | * If foid is 0 push out all transactions, otherwise push only those | |
1608 | * for that object or might reference that object. | |
1609 | * | |
1610 | * itxs are committed in batches. In a heavily stressed zil there will be | |
1611 | * a commit writer thread who is writing out a bunch of itxs to the log | |
1612 | * for a set of committing threads (cthreads) in the same batch as the writer. | |
1613 | * Those cthreads are all waiting on the same cv for that batch. | |
1614 | * | |
1615 | * There will also be a different and growing batch of threads that are | |
1616 | * waiting to commit (qthreads). When the committing batch completes | |
1617 | * a transition occurs such that the cthreads exit and the qthreads become | |
1618 | * cthreads. One of the new cthreads becomes the writer thread for the | |
1619 | * batch. Any new threads arriving become new qthreads. | |
1620 | * | |
1621 | * Only 2 condition variables are needed and there's no transition | |
1622 | * between the two cvs needed. They just flip-flop between qthreads | |
1623 | * and cthreads. | |
1624 | * | |
1625 | * Using this scheme we can efficiently wakeup up only those threads | |
1626 | * that have been committed. | |
1627 | */ | |
1628 | void | |
1629 | zil_commit(zilog_t *zilog, uint64_t foid) | |
1630 | { | |
1631 | uint64_t mybatch; | |
1632 | ||
1633 | if (zilog->zl_sync == ZFS_SYNC_DISABLED) | |
1634 | return; | |
1635 | ||
1636 | ZIL_STAT_BUMP(zil_commit_count); | |
1637 | ||
1638 | /* move the async itxs for the foid to the sync queues */ | |
1639 | zil_async_to_sync(zilog, foid); | |
1640 | ||
1641 | mutex_enter(&zilog->zl_lock); | |
1642 | mybatch = zilog->zl_next_batch; | |
1643 | while (zilog->zl_writer) { | |
1644 | cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); | |
1645 | if (mybatch <= zilog->zl_com_batch) { | |
1646 | mutex_exit(&zilog->zl_lock); | |
1647 | return; | |
1648 | } | |
1649 | } | |
1650 | ||
1651 | zilog->zl_next_batch++; | |
1652 | zilog->zl_writer = B_TRUE; | |
1653 | ZIL_STAT_BUMP(zil_commit_writer_count); | |
1654 | zil_commit_writer(zilog); | |
1655 | zilog->zl_com_batch = mybatch; | |
1656 | zilog->zl_writer = B_FALSE; | |
1657 | ||
1658 | /* wake up one thread to become the next writer */ | |
1659 | cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); | |
1660 | ||
1661 | /* wake up all threads waiting for this batch to be committed */ | |
1662 | cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); | |
1663 | ||
1664 | mutex_exit(&zilog->zl_lock); | |
1665 | } | |
1666 | ||
1667 | /* | |
1668 | * Called in syncing context to free committed log blocks and update log header. | |
1669 | */ | |
1670 | void | |
1671 | zil_sync(zilog_t *zilog, dmu_tx_t *tx) | |
1672 | { | |
1673 | zil_header_t *zh = zil_header_in_syncing_context(zilog); | |
1674 | uint64_t txg = dmu_tx_get_txg(tx); | |
1675 | spa_t *spa = zilog->zl_spa; | |
1676 | uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; | |
1677 | lwb_t *lwb; | |
1678 | ||
1679 | /* | |
1680 | * We don't zero out zl_destroy_txg, so make sure we don't try | |
1681 | * to destroy it twice. | |
1682 | */ | |
1683 | if (spa_sync_pass(spa) != 1) | |
1684 | return; | |
1685 | ||
1686 | mutex_enter(&zilog->zl_lock); | |
1687 | ||
1688 | ASSERT(zilog->zl_stop_sync == 0); | |
1689 | ||
1690 | if (*replayed_seq != 0) { | |
1691 | ASSERT(zh->zh_replay_seq < *replayed_seq); | |
1692 | zh->zh_replay_seq = *replayed_seq; | |
1693 | *replayed_seq = 0; | |
1694 | } | |
1695 | ||
1696 | if (zilog->zl_destroy_txg == txg) { | |
1697 | blkptr_t blk = zh->zh_log; | |
1698 | ||
1699 | ASSERT(list_head(&zilog->zl_lwb_list) == NULL); | |
1700 | ||
1701 | bzero(zh, sizeof (zil_header_t)); | |
1702 | bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); | |
1703 | ||
1704 | if (zilog->zl_keep_first) { | |
1705 | /* | |
1706 | * If this block was part of log chain that couldn't | |
1707 | * be claimed because a device was missing during | |
1708 | * zil_claim(), but that device later returns, | |
1709 | * then this block could erroneously appear valid. | |
1710 | * To guard against this, assign a new GUID to the new | |
1711 | * log chain so it doesn't matter what blk points to. | |
1712 | */ | |
1713 | zil_init_log_chain(zilog, &blk); | |
1714 | zh->zh_log = blk; | |
1715 | } | |
1716 | } | |
1717 | ||
1718 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { | |
1719 | zh->zh_log = lwb->lwb_blk; | |
1720 | if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) | |
1721 | break; | |
1722 | ||
1723 | ASSERT(lwb->lwb_zio == NULL); | |
1724 | ||
1725 | list_remove(&zilog->zl_lwb_list, lwb); | |
1726 | zio_free_zil(spa, txg, &lwb->lwb_blk); | |
1727 | kmem_cache_free(zil_lwb_cache, lwb); | |
1728 | ||
1729 | /* | |
1730 | * If we don't have anything left in the lwb list then | |
1731 | * we've had an allocation failure and we need to zero | |
1732 | * out the zil_header blkptr so that we don't end | |
1733 | * up freeing the same block twice. | |
1734 | */ | |
1735 | if (list_head(&zilog->zl_lwb_list) == NULL) | |
1736 | BP_ZERO(&zh->zh_log); | |
1737 | } | |
1738 | ||
1739 | /* | |
1740 | * Remove fastwrite on any blocks that have been pre-allocated for | |
1741 | * the next commit. This prevents fastwrite counter pollution by | |
1742 | * unused, long-lived LWBs. | |
1743 | */ | |
1744 | for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) { | |
1745 | if (lwb->lwb_fastwrite && !lwb->lwb_zio) { | |
1746 | metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); | |
1747 | lwb->lwb_fastwrite = 0; | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | mutex_exit(&zilog->zl_lock); | |
1752 | } | |
1753 | ||
1754 | void | |
1755 | zil_init(void) | |
1756 | { | |
1757 | zil_lwb_cache = kmem_cache_create("zil_lwb_cache", | |
1758 | sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
1759 | ||
1760 | zil_ksp = kstat_create("zfs", 0, "zil", "misc", | |
1761 | KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t), | |
1762 | KSTAT_FLAG_VIRTUAL); | |
1763 | ||
1764 | if (zil_ksp != NULL) { | |
1765 | zil_ksp->ks_data = &zil_stats; | |
1766 | kstat_install(zil_ksp); | |
1767 | } | |
1768 | } | |
1769 | ||
1770 | void | |
1771 | zil_fini(void) | |
1772 | { | |
1773 | kmem_cache_destroy(zil_lwb_cache); | |
1774 | ||
1775 | if (zil_ksp != NULL) { | |
1776 | kstat_delete(zil_ksp); | |
1777 | zil_ksp = NULL; | |
1778 | } | |
1779 | } | |
1780 | ||
1781 | void | |
1782 | zil_set_sync(zilog_t *zilog, uint64_t sync) | |
1783 | { | |
1784 | zilog->zl_sync = sync; | |
1785 | } | |
1786 | ||
1787 | void | |
1788 | zil_set_logbias(zilog_t *zilog, uint64_t logbias) | |
1789 | { | |
1790 | zilog->zl_logbias = logbias; | |
1791 | } | |
1792 | ||
1793 | zilog_t * | |
1794 | zil_alloc(objset_t *os, zil_header_t *zh_phys) | |
1795 | { | |
1796 | zilog_t *zilog; | |
1797 | int i; | |
1798 | ||
1799 | zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); | |
1800 | ||
1801 | zilog->zl_header = zh_phys; | |
1802 | zilog->zl_os = os; | |
1803 | zilog->zl_spa = dmu_objset_spa(os); | |
1804 | zilog->zl_dmu_pool = dmu_objset_pool(os); | |
1805 | zilog->zl_destroy_txg = TXG_INITIAL - 1; | |
1806 | zilog->zl_logbias = dmu_objset_logbias(os); | |
1807 | zilog->zl_sync = dmu_objset_syncprop(os); | |
1808 | zilog->zl_next_batch = 1; | |
1809 | ||
1810 | mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); | |
1811 | ||
1812 | for (i = 0; i < TXG_SIZE; i++) { | |
1813 | mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, | |
1814 | MUTEX_DEFAULT, NULL); | |
1815 | } | |
1816 | ||
1817 | list_create(&zilog->zl_lwb_list, sizeof (lwb_t), | |
1818 | offsetof(lwb_t, lwb_node)); | |
1819 | ||
1820 | list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), | |
1821 | offsetof(itx_t, itx_node)); | |
1822 | ||
1823 | mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); | |
1824 | ||
1825 | avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, | |
1826 | sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); | |
1827 | ||
1828 | cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); | |
1829 | cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); | |
1830 | cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); | |
1831 | cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); | |
1832 | ||
1833 | return (zilog); | |
1834 | } | |
1835 | ||
1836 | void | |
1837 | zil_free(zilog_t *zilog) | |
1838 | { | |
1839 | int i; | |
1840 | ||
1841 | zilog->zl_stop_sync = 1; | |
1842 | ||
1843 | ASSERT0(zilog->zl_suspend); | |
1844 | ASSERT0(zilog->zl_suspending); | |
1845 | ||
1846 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
1847 | list_destroy(&zilog->zl_lwb_list); | |
1848 | ||
1849 | avl_destroy(&zilog->zl_vdev_tree); | |
1850 | mutex_destroy(&zilog->zl_vdev_lock); | |
1851 | ||
1852 | ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); | |
1853 | list_destroy(&zilog->zl_itx_commit_list); | |
1854 | ||
1855 | for (i = 0; i < TXG_SIZE; i++) { | |
1856 | /* | |
1857 | * It's possible for an itx to be generated that doesn't dirty | |
1858 | * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() | |
1859 | * callback to remove the entry. We remove those here. | |
1860 | * | |
1861 | * Also free up the ziltest itxs. | |
1862 | */ | |
1863 | if (zilog->zl_itxg[i].itxg_itxs) | |
1864 | zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); | |
1865 | mutex_destroy(&zilog->zl_itxg[i].itxg_lock); | |
1866 | } | |
1867 | ||
1868 | mutex_destroy(&zilog->zl_lock); | |
1869 | ||
1870 | cv_destroy(&zilog->zl_cv_writer); | |
1871 | cv_destroy(&zilog->zl_cv_suspend); | |
1872 | cv_destroy(&zilog->zl_cv_batch[0]); | |
1873 | cv_destroy(&zilog->zl_cv_batch[1]); | |
1874 | ||
1875 | kmem_free(zilog, sizeof (zilog_t)); | |
1876 | } | |
1877 | ||
1878 | /* | |
1879 | * Open an intent log. | |
1880 | */ | |
1881 | zilog_t * | |
1882 | zil_open(objset_t *os, zil_get_data_t *get_data) | |
1883 | { | |
1884 | zilog_t *zilog = dmu_objset_zil(os); | |
1885 | ||
1886 | ASSERT(zilog->zl_clean_taskq == NULL); | |
1887 | ASSERT(zilog->zl_get_data == NULL); | |
1888 | ASSERT(list_is_empty(&zilog->zl_lwb_list)); | |
1889 | ||
1890 | zilog->zl_get_data = get_data; | |
1891 | zilog->zl_clean_taskq = taskq_create("zil_clean", 1, defclsyspri, | |
1892 | 2, 2, TASKQ_PREPOPULATE); | |
1893 | ||
1894 | return (zilog); | |
1895 | } | |
1896 | ||
1897 | /* | |
1898 | * Close an intent log. | |
1899 | */ | |
1900 | void | |
1901 | zil_close(zilog_t *zilog) | |
1902 | { | |
1903 | lwb_t *lwb; | |
1904 | uint64_t txg = 0; | |
1905 | ||
1906 | zil_commit(zilog, 0); /* commit all itx */ | |
1907 | ||
1908 | /* | |
1909 | * The lwb_max_txg for the stubby lwb will reflect the last activity | |
1910 | * for the zil. After a txg_wait_synced() on the txg we know all the | |
1911 | * callbacks have occurred that may clean the zil. Only then can we | |
1912 | * destroy the zl_clean_taskq. | |
1913 | */ | |
1914 | mutex_enter(&zilog->zl_lock); | |
1915 | lwb = list_tail(&zilog->zl_lwb_list); | |
1916 | if (lwb != NULL) | |
1917 | txg = lwb->lwb_max_txg; | |
1918 | mutex_exit(&zilog->zl_lock); | |
1919 | if (txg) | |
1920 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
1921 | ASSERT(!zilog_is_dirty(zilog)); | |
1922 | ||
1923 | taskq_destroy(zilog->zl_clean_taskq); | |
1924 | zilog->zl_clean_taskq = NULL; | |
1925 | zilog->zl_get_data = NULL; | |
1926 | ||
1927 | /* | |
1928 | * We should have only one LWB left on the list; remove it now. | |
1929 | */ | |
1930 | mutex_enter(&zilog->zl_lock); | |
1931 | lwb = list_head(&zilog->zl_lwb_list); | |
1932 | if (lwb != NULL) { | |
1933 | ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); | |
1934 | ASSERT(lwb->lwb_zio == NULL); | |
1935 | if (lwb->lwb_fastwrite) | |
1936 | metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); | |
1937 | list_remove(&zilog->zl_lwb_list, lwb); | |
1938 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
1939 | kmem_cache_free(zil_lwb_cache, lwb); | |
1940 | } | |
1941 | mutex_exit(&zilog->zl_lock); | |
1942 | } | |
1943 | ||
1944 | static char *suspend_tag = "zil suspending"; | |
1945 | ||
1946 | /* | |
1947 | * Suspend an intent log. While in suspended mode, we still honor | |
1948 | * synchronous semantics, but we rely on txg_wait_synced() to do it. | |
1949 | * On old version pools, we suspend the log briefly when taking a | |
1950 | * snapshot so that it will have an empty intent log. | |
1951 | * | |
1952 | * Long holds are not really intended to be used the way we do here -- | |
1953 | * held for such a short time. A concurrent caller of dsl_dataset_long_held() | |
1954 | * could fail. Therefore we take pains to only put a long hold if it is | |
1955 | * actually necessary. Fortunately, it will only be necessary if the | |
1956 | * objset is currently mounted (or the ZVOL equivalent). In that case it | |
1957 | * will already have a long hold, so we are not really making things any worse. | |
1958 | * | |
1959 | * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or | |
1960 | * zvol_state_t), and use their mechanism to prevent their hold from being | |
1961 | * dropped (e.g. VFS_HOLD()). However, that would be even more pain for | |
1962 | * very little gain. | |
1963 | * | |
1964 | * if cookiep == NULL, this does both the suspend & resume. | |
1965 | * Otherwise, it returns with the dataset "long held", and the cookie | |
1966 | * should be passed into zil_resume(). | |
1967 | */ | |
1968 | int | |
1969 | zil_suspend(const char *osname, void **cookiep) | |
1970 | { | |
1971 | objset_t *os; | |
1972 | zilog_t *zilog; | |
1973 | const zil_header_t *zh; | |
1974 | int error; | |
1975 | ||
1976 | error = dmu_objset_hold(osname, suspend_tag, &os); | |
1977 | if (error != 0) | |
1978 | return (error); | |
1979 | zilog = dmu_objset_zil(os); | |
1980 | ||
1981 | mutex_enter(&zilog->zl_lock); | |
1982 | zh = zilog->zl_header; | |
1983 | ||
1984 | if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ | |
1985 | mutex_exit(&zilog->zl_lock); | |
1986 | dmu_objset_rele(os, suspend_tag); | |
1987 | return (SET_ERROR(EBUSY)); | |
1988 | } | |
1989 | ||
1990 | /* | |
1991 | * Don't put a long hold in the cases where we can avoid it. This | |
1992 | * is when there is no cookie so we are doing a suspend & resume | |
1993 | * (i.e. called from zil_vdev_offline()), and there's nothing to do | |
1994 | * for the suspend because it's already suspended, or there's no ZIL. | |
1995 | */ | |
1996 | if (cookiep == NULL && !zilog->zl_suspending && | |
1997 | (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { | |
1998 | mutex_exit(&zilog->zl_lock); | |
1999 | dmu_objset_rele(os, suspend_tag); | |
2000 | return (0); | |
2001 | } | |
2002 | ||
2003 | dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); | |
2004 | dsl_pool_rele(dmu_objset_pool(os), suspend_tag); | |
2005 | ||
2006 | zilog->zl_suspend++; | |
2007 | ||
2008 | if (zilog->zl_suspend > 1) { | |
2009 | /* | |
2010 | * Someone else is already suspending it. | |
2011 | * Just wait for them to finish. | |
2012 | */ | |
2013 | ||
2014 | while (zilog->zl_suspending) | |
2015 | cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); | |
2016 | mutex_exit(&zilog->zl_lock); | |
2017 | ||
2018 | if (cookiep == NULL) | |
2019 | zil_resume(os); | |
2020 | else | |
2021 | *cookiep = os; | |
2022 | return (0); | |
2023 | } | |
2024 | ||
2025 | /* | |
2026 | * If there is no pointer to an on-disk block, this ZIL must not | |
2027 | * be active (e.g. filesystem not mounted), so there's nothing | |
2028 | * to clean up. | |
2029 | */ | |
2030 | if (BP_IS_HOLE(&zh->zh_log)) { | |
2031 | ASSERT(cookiep != NULL); /* fast path already handled */ | |
2032 | ||
2033 | *cookiep = os; | |
2034 | mutex_exit(&zilog->zl_lock); | |
2035 | return (0); | |
2036 | } | |
2037 | ||
2038 | zilog->zl_suspending = B_TRUE; | |
2039 | mutex_exit(&zilog->zl_lock); | |
2040 | ||
2041 | zil_commit(zilog, 0); | |
2042 | ||
2043 | zil_destroy(zilog, B_FALSE); | |
2044 | ||
2045 | mutex_enter(&zilog->zl_lock); | |
2046 | zilog->zl_suspending = B_FALSE; | |
2047 | cv_broadcast(&zilog->zl_cv_suspend); | |
2048 | mutex_exit(&zilog->zl_lock); | |
2049 | ||
2050 | if (cookiep == NULL) | |
2051 | zil_resume(os); | |
2052 | else | |
2053 | *cookiep = os; | |
2054 | return (0); | |
2055 | } | |
2056 | ||
2057 | void | |
2058 | zil_resume(void *cookie) | |
2059 | { | |
2060 | objset_t *os = cookie; | |
2061 | zilog_t *zilog = dmu_objset_zil(os); | |
2062 | ||
2063 | mutex_enter(&zilog->zl_lock); | |
2064 | ASSERT(zilog->zl_suspend != 0); | |
2065 | zilog->zl_suspend--; | |
2066 | mutex_exit(&zilog->zl_lock); | |
2067 | dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); | |
2068 | dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); | |
2069 | } | |
2070 | ||
2071 | typedef struct zil_replay_arg { | |
2072 | zil_replay_func_t *zr_replay; | |
2073 | void *zr_arg; | |
2074 | boolean_t zr_byteswap; | |
2075 | char *zr_lr; | |
2076 | } zil_replay_arg_t; | |
2077 | ||
2078 | static int | |
2079 | zil_replay_error(zilog_t *zilog, lr_t *lr, int error) | |
2080 | { | |
2081 | char name[MAXNAMELEN]; | |
2082 | ||
2083 | zilog->zl_replaying_seq--; /* didn't actually replay this one */ | |
2084 | ||
2085 | dmu_objset_name(zilog->zl_os, name); | |
2086 | ||
2087 | cmn_err(CE_WARN, "ZFS replay transaction error %d, " | |
2088 | "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, | |
2089 | (u_longlong_t)lr->lrc_seq, | |
2090 | (u_longlong_t)(lr->lrc_txtype & ~TX_CI), | |
2091 | (lr->lrc_txtype & TX_CI) ? "CI" : ""); | |
2092 | ||
2093 | return (error); | |
2094 | } | |
2095 | ||
2096 | static int | |
2097 | zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) | |
2098 | { | |
2099 | zil_replay_arg_t *zr = zra; | |
2100 | const zil_header_t *zh = zilog->zl_header; | |
2101 | uint64_t reclen = lr->lrc_reclen; | |
2102 | uint64_t txtype = lr->lrc_txtype; | |
2103 | int error = 0; | |
2104 | ||
2105 | zilog->zl_replaying_seq = lr->lrc_seq; | |
2106 | ||
2107 | if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ | |
2108 | return (0); | |
2109 | ||
2110 | if (lr->lrc_txg < claim_txg) /* already committed */ | |
2111 | return (0); | |
2112 | ||
2113 | /* Strip case-insensitive bit, still present in log record */ | |
2114 | txtype &= ~TX_CI; | |
2115 | ||
2116 | if (txtype == 0 || txtype >= TX_MAX_TYPE) | |
2117 | return (zil_replay_error(zilog, lr, EINVAL)); | |
2118 | ||
2119 | /* | |
2120 | * If this record type can be logged out of order, the object | |
2121 | * (lr_foid) may no longer exist. That's legitimate, not an error. | |
2122 | */ | |
2123 | if (TX_OOO(txtype)) { | |
2124 | error = dmu_object_info(zilog->zl_os, | |
2125 | ((lr_ooo_t *)lr)->lr_foid, NULL); | |
2126 | if (error == ENOENT || error == EEXIST) | |
2127 | return (0); | |
2128 | } | |
2129 | ||
2130 | /* | |
2131 | * Make a copy of the data so we can revise and extend it. | |
2132 | */ | |
2133 | bcopy(lr, zr->zr_lr, reclen); | |
2134 | ||
2135 | /* | |
2136 | * If this is a TX_WRITE with a blkptr, suck in the data. | |
2137 | */ | |
2138 | if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { | |
2139 | error = zil_read_log_data(zilog, (lr_write_t *)lr, | |
2140 | zr->zr_lr + reclen); | |
2141 | if (error != 0) | |
2142 | return (zil_replay_error(zilog, lr, error)); | |
2143 | } | |
2144 | ||
2145 | /* | |
2146 | * The log block containing this lr may have been byteswapped | |
2147 | * so that we can easily examine common fields like lrc_txtype. | |
2148 | * However, the log is a mix of different record types, and only the | |
2149 | * replay vectors know how to byteswap their records. Therefore, if | |
2150 | * the lr was byteswapped, undo it before invoking the replay vector. | |
2151 | */ | |
2152 | if (zr->zr_byteswap) | |
2153 | byteswap_uint64_array(zr->zr_lr, reclen); | |
2154 | ||
2155 | /* | |
2156 | * We must now do two things atomically: replay this log record, | |
2157 | * and update the log header sequence number to reflect the fact that | |
2158 | * we did so. At the end of each replay function the sequence number | |
2159 | * is updated if we are in replay mode. | |
2160 | */ | |
2161 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); | |
2162 | if (error != 0) { | |
2163 | /* | |
2164 | * The DMU's dnode layer doesn't see removes until the txg | |
2165 | * commits, so a subsequent claim can spuriously fail with | |
2166 | * EEXIST. So if we receive any error we try syncing out | |
2167 | * any removes then retry the transaction. Note that we | |
2168 | * specify B_FALSE for byteswap now, so we don't do it twice. | |
2169 | */ | |
2170 | txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); | |
2171 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); | |
2172 | if (error != 0) | |
2173 | return (zil_replay_error(zilog, lr, error)); | |
2174 | } | |
2175 | return (0); | |
2176 | } | |
2177 | ||
2178 | /* ARGSUSED */ | |
2179 | static int | |
2180 | zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) | |
2181 | { | |
2182 | zilog->zl_replay_blks++; | |
2183 | ||
2184 | return (0); | |
2185 | } | |
2186 | ||
2187 | /* | |
2188 | * If this dataset has a non-empty intent log, replay it and destroy it. | |
2189 | */ | |
2190 | void | |
2191 | zil_replay(objset_t *os, void *arg, zil_replay_func_t replay_func[TX_MAX_TYPE]) | |
2192 | { | |
2193 | zilog_t *zilog = dmu_objset_zil(os); | |
2194 | const zil_header_t *zh = zilog->zl_header; | |
2195 | zil_replay_arg_t zr; | |
2196 | ||
2197 | if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { | |
2198 | zil_destroy(zilog, B_TRUE); | |
2199 | return; | |
2200 | } | |
2201 | ||
2202 | zr.zr_replay = replay_func; | |
2203 | zr.zr_arg = arg; | |
2204 | zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); | |
2205 | zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); | |
2206 | ||
2207 | /* | |
2208 | * Wait for in-progress removes to sync before starting replay. | |
2209 | */ | |
2210 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
2211 | ||
2212 | zilog->zl_replay = B_TRUE; | |
2213 | zilog->zl_replay_time = ddi_get_lbolt(); | |
2214 | ASSERT(zilog->zl_replay_blks == 0); | |
2215 | (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, | |
2216 | zh->zh_claim_txg); | |
2217 | vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); | |
2218 | ||
2219 | zil_destroy(zilog, B_FALSE); | |
2220 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
2221 | zilog->zl_replay = B_FALSE; | |
2222 | } | |
2223 | ||
2224 | boolean_t | |
2225 | zil_replaying(zilog_t *zilog, dmu_tx_t *tx) | |
2226 | { | |
2227 | if (zilog->zl_sync == ZFS_SYNC_DISABLED) | |
2228 | return (B_TRUE); | |
2229 | ||
2230 | if (zilog->zl_replay) { | |
2231 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
2232 | zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = | |
2233 | zilog->zl_replaying_seq; | |
2234 | return (B_TRUE); | |
2235 | } | |
2236 | ||
2237 | return (B_FALSE); | |
2238 | } | |
2239 | ||
2240 | /* ARGSUSED */ | |
2241 | int | |
2242 | zil_vdev_offline(const char *osname, void *arg) | |
2243 | { | |
2244 | int error; | |
2245 | ||
2246 | error = zil_suspend(osname, NULL); | |
2247 | if (error != 0) | |
2248 | return (SET_ERROR(EEXIST)); | |
2249 | return (0); | |
2250 | } | |
2251 | ||
2252 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
2253 | EXPORT_SYMBOL(zil_alloc); | |
2254 | EXPORT_SYMBOL(zil_free); | |
2255 | EXPORT_SYMBOL(zil_open); | |
2256 | EXPORT_SYMBOL(zil_close); | |
2257 | EXPORT_SYMBOL(zil_replay); | |
2258 | EXPORT_SYMBOL(zil_replaying); | |
2259 | EXPORT_SYMBOL(zil_destroy); | |
2260 | EXPORT_SYMBOL(zil_destroy_sync); | |
2261 | EXPORT_SYMBOL(zil_itx_create); | |
2262 | EXPORT_SYMBOL(zil_itx_destroy); | |
2263 | EXPORT_SYMBOL(zil_itx_assign); | |
2264 | EXPORT_SYMBOL(zil_commit); | |
2265 | EXPORT_SYMBOL(zil_vdev_offline); | |
2266 | EXPORT_SYMBOL(zil_claim); | |
2267 | EXPORT_SYMBOL(zil_check_log_chain); | |
2268 | EXPORT_SYMBOL(zil_sync); | |
2269 | EXPORT_SYMBOL(zil_clean); | |
2270 | EXPORT_SYMBOL(zil_suspend); | |
2271 | EXPORT_SYMBOL(zil_resume); | |
2272 | EXPORT_SYMBOL(zil_add_block); | |
2273 | EXPORT_SYMBOL(zil_bp_tree_add); | |
2274 | EXPORT_SYMBOL(zil_set_sync); | |
2275 | EXPORT_SYMBOL(zil_set_logbias); | |
2276 | ||
2277 | module_param(zil_replay_disable, int, 0644); | |
2278 | MODULE_PARM_DESC(zil_replay_disable, "Disable intent logging replay"); | |
2279 | ||
2280 | module_param(zfs_nocacheflush, int, 0644); | |
2281 | MODULE_PARM_DESC(zfs_nocacheflush, "Disable cache flushes"); | |
2282 | ||
2283 | module_param(zil_slog_limit, ulong, 0644); | |
2284 | MODULE_PARM_DESC(zil_slog_limit, "Max commit bytes to separate log device"); | |
2285 | #endif |