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