]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
9babb374 | 22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. |
34dc7c2f BB |
23 | * Use is subject to license terms. |
24 | */ | |
25 | ||
34dc7c2f BB |
26 | #include <sys/zfs_context.h> |
27 | #include <sys/spa.h> | |
9babb374 | 28 | #include <sys/spa_impl.h> |
34dc7c2f BB |
29 | #include <sys/dmu.h> |
30 | #include <sys/zap.h> | |
31 | #include <sys/arc.h> | |
32 | #include <sys/stat.h> | |
33 | #include <sys/resource.h> | |
34 | #include <sys/zil.h> | |
35 | #include <sys/zil_impl.h> | |
36 | #include <sys/dsl_dataset.h> | |
37 | #include <sys/vdev.h> | |
38 | #include <sys/dmu_tx.h> | |
39 | ||
40 | /* | |
41 | * The zfs intent log (ZIL) saves transaction records of system calls | |
42 | * that change the file system in memory with enough information | |
43 | * to be able to replay them. These are stored in memory until | |
44 | * either the DMU transaction group (txg) commits them to the stable pool | |
45 | * and they can be discarded, or they are flushed to the stable log | |
46 | * (also in the pool) due to a fsync, O_DSYNC or other synchronous | |
47 | * requirement. In the event of a panic or power fail then those log | |
48 | * records (transactions) are replayed. | |
49 | * | |
50 | * There is one ZIL per file system. Its on-disk (pool) format consists | |
51 | * of 3 parts: | |
52 | * | |
53 | * - ZIL header | |
54 | * - ZIL blocks | |
55 | * - ZIL records | |
56 | * | |
57 | * A log record holds a system call transaction. Log blocks can | |
58 | * hold many log records and the blocks are chained together. | |
59 | * Each ZIL block contains a block pointer (blkptr_t) to the next | |
60 | * ZIL block in the chain. The ZIL header points to the first | |
61 | * block in the chain. Note there is not a fixed place in the pool | |
62 | * to hold blocks. They are dynamically allocated and freed as | |
63 | * needed from the blocks available. Figure X shows the ZIL structure: | |
64 | */ | |
65 | ||
66 | /* | |
67 | * This global ZIL switch affects all pools | |
68 | */ | |
69 | int zil_disable = 0; /* disable intent logging */ | |
70 | ||
71 | /* | |
72 | * Tunable parameter for debugging or performance analysis. Setting | |
73 | * zfs_nocacheflush will cause corruption on power loss if a volatile | |
74 | * out-of-order write cache is enabled. | |
75 | */ | |
76 | boolean_t zfs_nocacheflush = B_FALSE; | |
77 | ||
78 | static kmem_cache_t *zil_lwb_cache; | |
79 | ||
80 | static int | |
81 | zil_dva_compare(const void *x1, const void *x2) | |
82 | { | |
83 | const dva_t *dva1 = x1; | |
84 | const dva_t *dva2 = x2; | |
85 | ||
86 | if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) | |
87 | return (-1); | |
88 | if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) | |
89 | return (1); | |
90 | ||
91 | if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) | |
92 | return (-1); | |
93 | if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) | |
94 | return (1); | |
95 | ||
96 | return (0); | |
97 | } | |
98 | ||
99 | static void | |
100 | zil_dva_tree_init(avl_tree_t *t) | |
101 | { | |
102 | avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t), | |
103 | offsetof(zil_dva_node_t, zn_node)); | |
104 | } | |
105 | ||
106 | static void | |
107 | zil_dva_tree_fini(avl_tree_t *t) | |
108 | { | |
109 | zil_dva_node_t *zn; | |
110 | void *cookie = NULL; | |
111 | ||
112 | while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) | |
113 | kmem_free(zn, sizeof (zil_dva_node_t)); | |
114 | ||
115 | avl_destroy(t); | |
116 | } | |
117 | ||
118 | static int | |
119 | zil_dva_tree_add(avl_tree_t *t, dva_t *dva) | |
120 | { | |
121 | zil_dva_node_t *zn; | |
122 | avl_index_t where; | |
123 | ||
124 | if (avl_find(t, dva, &where) != NULL) | |
125 | return (EEXIST); | |
126 | ||
127 | zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP); | |
128 | zn->zn_dva = *dva; | |
129 | avl_insert(t, zn, where); | |
130 | ||
131 | return (0); | |
132 | } | |
133 | ||
134 | static zil_header_t * | |
135 | zil_header_in_syncing_context(zilog_t *zilog) | |
136 | { | |
137 | return ((zil_header_t *)zilog->zl_header); | |
138 | } | |
139 | ||
140 | static void | |
141 | zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) | |
142 | { | |
143 | zio_cksum_t *zc = &bp->blk_cksum; | |
144 | ||
145 | zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); | |
146 | zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); | |
147 | zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); | |
148 | zc->zc_word[ZIL_ZC_SEQ] = 1ULL; | |
149 | } | |
150 | ||
151 | /* | |
152 | * Read a log block, make sure it's valid, and byteswap it if necessary. | |
153 | */ | |
154 | static int | |
155 | zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp) | |
156 | { | |
157 | blkptr_t blk = *bp; | |
158 | zbookmark_t zb; | |
159 | uint32_t aflags = ARC_WAIT; | |
160 | int error; | |
161 | ||
162 | zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET]; | |
163 | zb.zb_object = 0; | |
164 | zb.zb_level = -1; | |
165 | zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ]; | |
166 | ||
167 | *abufpp = NULL; | |
168 | ||
b128c09f BB |
169 | /* |
170 | * We shouldn't be doing any scrubbing while we're doing log | |
171 | * replay, it's OK to not lock. | |
172 | */ | |
173 | error = arc_read_nolock(NULL, zilog->zl_spa, &blk, | |
34dc7c2f BB |
174 | arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL | |
175 | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb); | |
176 | ||
177 | if (error == 0) { | |
178 | char *data = (*abufpp)->b_data; | |
179 | uint64_t blksz = BP_GET_LSIZE(bp); | |
180 | zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1; | |
181 | zio_cksum_t cksum = bp->blk_cksum; | |
182 | ||
183 | /* | |
b128c09f BB |
184 | * Validate the checksummed log block. |
185 | * | |
34dc7c2f BB |
186 | * Sequence numbers should be... sequential. The checksum |
187 | * verifier for the next block should be bp's checksum plus 1. | |
b128c09f BB |
188 | * |
189 | * Also check the log chain linkage and size used. | |
34dc7c2f BB |
190 | */ |
191 | cksum.zc_word[ZIL_ZC_SEQ]++; | |
192 | ||
b128c09f BB |
193 | if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, |
194 | sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) || | |
195 | (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))) { | |
196 | error = ECKSUM; | |
197 | } | |
34dc7c2f BB |
198 | |
199 | if (error) { | |
200 | VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1); | |
201 | *abufpp = NULL; | |
202 | } | |
203 | } | |
204 | ||
205 | dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid); | |
206 | ||
207 | return (error); | |
208 | } | |
209 | ||
210 | /* | |
211 | * Parse the intent log, and call parse_func for each valid record within. | |
212 | * Return the highest sequence number. | |
213 | */ | |
214 | uint64_t | |
215 | zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, | |
216 | zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) | |
217 | { | |
218 | const zil_header_t *zh = zilog->zl_header; | |
219 | uint64_t claim_seq = zh->zh_claim_seq; | |
220 | uint64_t seq = 0; | |
221 | uint64_t max_seq = 0; | |
222 | blkptr_t blk = zh->zh_log; | |
223 | arc_buf_t *abuf; | |
224 | char *lrbuf, *lrp; | |
225 | zil_trailer_t *ztp; | |
226 | int reclen, error; | |
227 | ||
228 | if (BP_IS_HOLE(&blk)) | |
229 | return (max_seq); | |
230 | ||
231 | /* | |
232 | * Starting at the block pointed to by zh_log we read the log chain. | |
233 | * For each block in the chain we strongly check that block to | |
234 | * ensure its validity. We stop when an invalid block is found. | |
235 | * For each block pointer in the chain we call parse_blk_func(). | |
236 | * For each record in each valid block we call parse_lr_func(). | |
237 | * If the log has been claimed, stop if we encounter a sequence | |
238 | * number greater than the highest claimed sequence number. | |
239 | */ | |
240 | zil_dva_tree_init(&zilog->zl_dva_tree); | |
241 | for (;;) { | |
242 | seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; | |
243 | ||
244 | if (claim_seq != 0 && seq > claim_seq) | |
245 | break; | |
246 | ||
247 | ASSERT(max_seq < seq); | |
248 | max_seq = seq; | |
249 | ||
250 | error = zil_read_log_block(zilog, &blk, &abuf); | |
251 | ||
252 | if (parse_blk_func != NULL) | |
253 | parse_blk_func(zilog, &blk, arg, txg); | |
254 | ||
255 | if (error) | |
256 | break; | |
257 | ||
258 | lrbuf = abuf->b_data; | |
259 | ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; | |
260 | blk = ztp->zit_next_blk; | |
261 | ||
262 | if (parse_lr_func == NULL) { | |
263 | VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); | |
264 | continue; | |
265 | } | |
266 | ||
267 | for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { | |
268 | lr_t *lr = (lr_t *)lrp; | |
269 | reclen = lr->lrc_reclen; | |
270 | ASSERT3U(reclen, >=, sizeof (lr_t)); | |
271 | parse_lr_func(zilog, lr, arg, txg); | |
272 | } | |
273 | VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); | |
274 | } | |
275 | zil_dva_tree_fini(&zilog->zl_dva_tree); | |
276 | ||
277 | return (max_seq); | |
278 | } | |
279 | ||
280 | /* ARGSUSED */ | |
281 | static void | |
282 | zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) | |
283 | { | |
284 | spa_t *spa = zilog->zl_spa; | |
285 | int err; | |
286 | ||
287 | /* | |
288 | * Claim log block if not already committed and not already claimed. | |
289 | */ | |
290 | if (bp->blk_birth >= first_txg && | |
291 | zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) { | |
b128c09f BB |
292 | err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL, |
293 | ZIO_FLAG_MUSTSUCCEED)); | |
34dc7c2f BB |
294 | ASSERT(err == 0); |
295 | } | |
296 | } | |
297 | ||
298 | static void | |
299 | zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) | |
300 | { | |
301 | if (lrc->lrc_txtype == TX_WRITE) { | |
302 | lr_write_t *lr = (lr_write_t *)lrc; | |
303 | zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg); | |
304 | } | |
305 | } | |
306 | ||
307 | /* ARGSUSED */ | |
308 | static void | |
309 | zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) | |
310 | { | |
311 | zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx)); | |
312 | } | |
313 | ||
314 | static void | |
315 | zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) | |
316 | { | |
317 | /* | |
318 | * If we previously claimed it, we need to free it. | |
319 | */ | |
320 | if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) { | |
321 | lr_write_t *lr = (lr_write_t *)lrc; | |
322 | blkptr_t *bp = &lr->lr_blkptr; | |
323 | if (bp->blk_birth >= claim_txg && | |
324 | !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) { | |
325 | (void) arc_free(NULL, zilog->zl_spa, | |
326 | dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT); | |
327 | } | |
328 | } | |
329 | } | |
330 | ||
331 | /* | |
332 | * Create an on-disk intent log. | |
333 | */ | |
334 | static void | |
335 | zil_create(zilog_t *zilog) | |
336 | { | |
337 | const zil_header_t *zh = zilog->zl_header; | |
338 | lwb_t *lwb; | |
339 | uint64_t txg = 0; | |
340 | dmu_tx_t *tx = NULL; | |
341 | blkptr_t blk; | |
342 | int error = 0; | |
343 | ||
344 | /* | |
345 | * Wait for any previous destroy to complete. | |
346 | */ | |
347 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
348 | ||
349 | ASSERT(zh->zh_claim_txg == 0); | |
350 | ASSERT(zh->zh_replay_seq == 0); | |
351 | ||
352 | blk = zh->zh_log; | |
353 | ||
354 | /* | |
fb5f0bc8 BB |
355 | * If we don't already have an initial log block or we have one |
356 | * but it's the wrong endianness then allocate one. | |
34dc7c2f | 357 | */ |
fb5f0bc8 | 358 | if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { |
34dc7c2f BB |
359 | tx = dmu_tx_create(zilog->zl_os); |
360 | (void) dmu_tx_assign(tx, TXG_WAIT); | |
361 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
362 | txg = dmu_tx_get_txg(tx); | |
363 | ||
fb5f0bc8 BB |
364 | if (!BP_IS_HOLE(&blk)) { |
365 | zio_free_blk(zilog->zl_spa, &blk, txg); | |
366 | BP_ZERO(&blk); | |
367 | } | |
368 | ||
34dc7c2f BB |
369 | error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk, |
370 | NULL, txg); | |
371 | ||
372 | if (error == 0) | |
373 | zil_init_log_chain(zilog, &blk); | |
374 | } | |
375 | ||
376 | /* | |
377 | * Allocate a log write buffer (lwb) for the first log block. | |
378 | */ | |
379 | if (error == 0) { | |
380 | lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); | |
381 | lwb->lwb_zilog = zilog; | |
382 | lwb->lwb_blk = blk; | |
383 | lwb->lwb_nused = 0; | |
384 | lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); | |
385 | lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); | |
386 | lwb->lwb_max_txg = txg; | |
387 | lwb->lwb_zio = NULL; | |
388 | ||
389 | mutex_enter(&zilog->zl_lock); | |
390 | list_insert_tail(&zilog->zl_lwb_list, lwb); | |
391 | mutex_exit(&zilog->zl_lock); | |
392 | } | |
393 | ||
394 | /* | |
395 | * If we just allocated the first log block, commit our transaction | |
396 | * and wait for zil_sync() to stuff the block poiner into zh_log. | |
397 | * (zh is part of the MOS, so we cannot modify it in open context.) | |
398 | */ | |
399 | if (tx != NULL) { | |
400 | dmu_tx_commit(tx); | |
401 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
402 | } | |
403 | ||
404 | ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); | |
405 | } | |
406 | ||
407 | /* | |
408 | * In one tx, free all log blocks and clear the log header. | |
409 | * If keep_first is set, then we're replaying a log with no content. | |
410 | * We want to keep the first block, however, so that the first | |
411 | * synchronous transaction doesn't require a txg_wait_synced() | |
412 | * in zil_create(). We don't need to txg_wait_synced() here either | |
413 | * when keep_first is set, because both zil_create() and zil_destroy() | |
414 | * will wait for any in-progress destroys to complete. | |
415 | */ | |
416 | void | |
417 | zil_destroy(zilog_t *zilog, boolean_t keep_first) | |
418 | { | |
419 | const zil_header_t *zh = zilog->zl_header; | |
420 | lwb_t *lwb; | |
421 | dmu_tx_t *tx; | |
422 | uint64_t txg; | |
423 | ||
424 | /* | |
425 | * Wait for any previous destroy to complete. | |
426 | */ | |
427 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
428 | ||
429 | if (BP_IS_HOLE(&zh->zh_log)) | |
430 | return; | |
431 | ||
432 | tx = dmu_tx_create(zilog->zl_os); | |
433 | (void) dmu_tx_assign(tx, TXG_WAIT); | |
434 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
435 | txg = dmu_tx_get_txg(tx); | |
436 | ||
437 | mutex_enter(&zilog->zl_lock); | |
438 | ||
439 | /* | |
440 | * It is possible for the ZIL to get the previously mounted zilog | |
441 | * structure of the same dataset if quickly remounted and the dbuf | |
442 | * eviction has not completed. In this case we can see a non | |
443 | * empty lwb list and keep_first will be set. We fix this by | |
444 | * clearing the keep_first. This will be slower but it's very rare. | |
445 | */ | |
446 | if (!list_is_empty(&zilog->zl_lwb_list) && keep_first) | |
447 | keep_first = B_FALSE; | |
448 | ||
449 | ASSERT3U(zilog->zl_destroy_txg, <, txg); | |
450 | zilog->zl_destroy_txg = txg; | |
451 | zilog->zl_keep_first = keep_first; | |
452 | ||
453 | if (!list_is_empty(&zilog->zl_lwb_list)) { | |
454 | ASSERT(zh->zh_claim_txg == 0); | |
455 | ASSERT(!keep_first); | |
456 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { | |
457 | list_remove(&zilog->zl_lwb_list, lwb); | |
458 | if (lwb->lwb_buf != NULL) | |
459 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
460 | zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg); | |
461 | kmem_cache_free(zil_lwb_cache, lwb); | |
462 | } | |
463 | } else { | |
464 | if (!keep_first) { | |
465 | (void) zil_parse(zilog, zil_free_log_block, | |
466 | zil_free_log_record, tx, zh->zh_claim_txg); | |
467 | } | |
468 | } | |
469 | mutex_exit(&zilog->zl_lock); | |
470 | ||
471 | dmu_tx_commit(tx); | |
472 | } | |
473 | ||
474 | /* | |
9babb374 | 475 | * return true if the initial log block is not valid |
34dc7c2f | 476 | */ |
9babb374 BB |
477 | static boolean_t |
478 | zil_empty(zilog_t *zilog) | |
34dc7c2f BB |
479 | { |
480 | const zil_header_t *zh = zilog->zl_header; | |
9babb374 | 481 | arc_buf_t *abuf = NULL; |
34dc7c2f BB |
482 | |
483 | if (BP_IS_HOLE(&zh->zh_log)) | |
9babb374 | 484 | return (B_TRUE); |
34dc7c2f | 485 | |
9babb374 BB |
486 | if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0) |
487 | return (B_TRUE); | |
34dc7c2f | 488 | |
9babb374 BB |
489 | VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); |
490 | return (B_FALSE); | |
34dc7c2f BB |
491 | } |
492 | ||
493 | int | |
494 | zil_claim(char *osname, void *txarg) | |
495 | { | |
496 | dmu_tx_t *tx = txarg; | |
497 | uint64_t first_txg = dmu_tx_get_txg(tx); | |
498 | zilog_t *zilog; | |
499 | zil_header_t *zh; | |
500 | objset_t *os; | |
501 | int error; | |
502 | ||
b128c09f | 503 | error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); |
34dc7c2f | 504 | if (error) { |
b128c09f | 505 | cmn_err(CE_WARN, "can't open objset for %s", osname); |
34dc7c2f BB |
506 | return (0); |
507 | } | |
508 | ||
509 | zilog = dmu_objset_zil(os); | |
510 | zh = zil_header_in_syncing_context(zilog); | |
511 | ||
9babb374 BB |
512 | if (zilog->zl_spa->spa_log_state == SPA_LOG_CLEAR) { |
513 | if (!BP_IS_HOLE(&zh->zh_log)) | |
514 | zio_free_blk(zilog->zl_spa, &zh->zh_log, first_txg); | |
515 | BP_ZERO(&zh->zh_log); | |
516 | dsl_dataset_dirty(dmu_objset_ds(os), tx); | |
517 | } | |
518 | ||
519 | /* | |
520 | * Record here whether the zil has any records to replay. | |
521 | * If the header block pointer is null or the block points | |
522 | * to the stubby then we know there are no valid log records. | |
523 | * We use the header to store this state as the the zilog gets | |
524 | * freed later in dmu_objset_close(). | |
525 | * The flags (and the rest of the header fields) are cleared in | |
526 | * zil_sync() as a result of a zil_destroy(), after replaying the log. | |
527 | * | |
528 | * Note, the intent log can be empty but still need the | |
529 | * stubby to be claimed. | |
530 | */ | |
531 | if (!zil_empty(zilog)) { | |
532 | zh->zh_flags |= ZIL_REPLAY_NEEDED; | |
533 | dsl_dataset_dirty(dmu_objset_ds(os), tx); | |
534 | } | |
535 | ||
34dc7c2f BB |
536 | /* |
537 | * Claim all log blocks if we haven't already done so, and remember | |
538 | * the highest claimed sequence number. This ensures that if we can | |
539 | * read only part of the log now (e.g. due to a missing device), | |
540 | * but we can read the entire log later, we will not try to replay | |
541 | * or destroy beyond the last block we successfully claimed. | |
542 | */ | |
543 | ASSERT3U(zh->zh_claim_txg, <=, first_txg); | |
544 | if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { | |
545 | zh->zh_claim_txg = first_txg; | |
546 | zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block, | |
547 | zil_claim_log_record, tx, first_txg); | |
548 | dsl_dataset_dirty(dmu_objset_ds(os), tx); | |
549 | } | |
550 | ||
551 | ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); | |
552 | dmu_objset_close(os); | |
553 | return (0); | |
554 | } | |
555 | ||
b128c09f BB |
556 | /* |
557 | * Check the log by walking the log chain. | |
558 | * Checksum errors are ok as they indicate the end of the chain. | |
559 | * Any other error (no device or read failure) returns an error. | |
560 | */ | |
561 | /* ARGSUSED */ | |
562 | int | |
563 | zil_check_log_chain(char *osname, void *txarg) | |
564 | { | |
565 | zilog_t *zilog; | |
566 | zil_header_t *zh; | |
567 | blkptr_t blk; | |
568 | arc_buf_t *abuf; | |
569 | objset_t *os; | |
570 | char *lrbuf; | |
571 | zil_trailer_t *ztp; | |
572 | int error; | |
573 | ||
574 | error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); | |
575 | if (error) { | |
576 | cmn_err(CE_WARN, "can't open objset for %s", osname); | |
577 | return (0); | |
578 | } | |
579 | ||
580 | zilog = dmu_objset_zil(os); | |
581 | zh = zil_header_in_syncing_context(zilog); | |
582 | blk = zh->zh_log; | |
583 | if (BP_IS_HOLE(&blk)) { | |
584 | dmu_objset_close(os); | |
585 | return (0); /* no chain */ | |
586 | } | |
587 | ||
588 | for (;;) { | |
589 | error = zil_read_log_block(zilog, &blk, &abuf); | |
590 | if (error) | |
591 | break; | |
592 | lrbuf = abuf->b_data; | |
593 | ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; | |
594 | blk = ztp->zit_next_blk; | |
595 | VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); | |
596 | } | |
597 | dmu_objset_close(os); | |
598 | if (error == ECKSUM) | |
599 | return (0); /* normal end of chain */ | |
600 | return (error); | |
601 | } | |
602 | ||
34dc7c2f BB |
603 | static int |
604 | zil_vdev_compare(const void *x1, const void *x2) | |
605 | { | |
606 | uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; | |
607 | uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; | |
608 | ||
609 | if (v1 < v2) | |
610 | return (-1); | |
611 | if (v1 > v2) | |
612 | return (1); | |
613 | ||
614 | return (0); | |
615 | } | |
616 | ||
617 | void | |
618 | zil_add_block(zilog_t *zilog, blkptr_t *bp) | |
619 | { | |
620 | avl_tree_t *t = &zilog->zl_vdev_tree; | |
621 | avl_index_t where; | |
622 | zil_vdev_node_t *zv, zvsearch; | |
623 | int ndvas = BP_GET_NDVAS(bp); | |
624 | int i; | |
625 | ||
626 | if (zfs_nocacheflush) | |
627 | return; | |
628 | ||
629 | ASSERT(zilog->zl_writer); | |
630 | ||
631 | /* | |
632 | * Even though we're zl_writer, we still need a lock because the | |
633 | * zl_get_data() callbacks may have dmu_sync() done callbacks | |
634 | * that will run concurrently. | |
635 | */ | |
636 | mutex_enter(&zilog->zl_vdev_lock); | |
637 | for (i = 0; i < ndvas; i++) { | |
638 | zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); | |
639 | if (avl_find(t, &zvsearch, &where) == NULL) { | |
640 | zv = kmem_alloc(sizeof (*zv), KM_SLEEP); | |
641 | zv->zv_vdev = zvsearch.zv_vdev; | |
642 | avl_insert(t, zv, where); | |
643 | } | |
644 | } | |
645 | mutex_exit(&zilog->zl_vdev_lock); | |
646 | } | |
647 | ||
648 | void | |
649 | zil_flush_vdevs(zilog_t *zilog) | |
650 | { | |
651 | spa_t *spa = zilog->zl_spa; | |
652 | avl_tree_t *t = &zilog->zl_vdev_tree; | |
653 | void *cookie = NULL; | |
654 | zil_vdev_node_t *zv; | |
655 | zio_t *zio; | |
656 | ||
657 | ASSERT(zilog->zl_writer); | |
658 | ||
659 | /* | |
660 | * We don't need zl_vdev_lock here because we're the zl_writer, | |
661 | * and all zl_get_data() callbacks are done. | |
662 | */ | |
663 | if (avl_numnodes(t) == 0) | |
664 | return; | |
665 | ||
b128c09f | 666 | spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); |
34dc7c2f | 667 | |
b128c09f | 668 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
34dc7c2f BB |
669 | |
670 | while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { | |
671 | vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); | |
672 | if (vd != NULL) | |
673 | zio_flush(zio, vd); | |
674 | kmem_free(zv, sizeof (*zv)); | |
675 | } | |
676 | ||
677 | /* | |
678 | * Wait for all the flushes to complete. Not all devices actually | |
679 | * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. | |
680 | */ | |
681 | (void) zio_wait(zio); | |
682 | ||
b128c09f | 683 | spa_config_exit(spa, SCL_STATE, FTAG); |
34dc7c2f BB |
684 | } |
685 | ||
686 | /* | |
687 | * Function called when a log block write completes | |
688 | */ | |
689 | static void | |
690 | zil_lwb_write_done(zio_t *zio) | |
691 | { | |
692 | lwb_t *lwb = zio->io_private; | |
693 | zilog_t *zilog = lwb->lwb_zilog; | |
694 | ||
b128c09f BB |
695 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); |
696 | ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG); | |
697 | ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); | |
698 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
699 | ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); | |
700 | ASSERT(!BP_IS_GANG(zio->io_bp)); | |
701 | ASSERT(!BP_IS_HOLE(zio->io_bp)); | |
702 | ASSERT(zio->io_bp->blk_fill == 0); | |
703 | ||
34dc7c2f | 704 | /* |
9babb374 BB |
705 | * Ensure the lwb buffer pointer is cleared before releasing |
706 | * the txg. If we have had an allocation failure and | |
707 | * the txg is waiting to sync then we want want zil_sync() | |
708 | * to remove the lwb so that it's not picked up as the next new | |
709 | * one in zil_commit_writer(). zil_sync() will only remove | |
710 | * the lwb if lwb_buf is null. | |
34dc7c2f | 711 | */ |
34dc7c2f BB |
712 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); |
713 | mutex_enter(&zilog->zl_lock); | |
714 | lwb->lwb_buf = NULL; | |
715 | if (zio->io_error) | |
716 | zilog->zl_log_error = B_TRUE; | |
717 | mutex_exit(&zilog->zl_lock); | |
9babb374 BB |
718 | |
719 | /* | |
720 | * Now that we've written this log block, we have a stable pointer | |
721 | * to the next block in the chain, so it's OK to let the txg in | |
722 | * which we allocated the next block sync. | |
723 | */ | |
724 | txg_rele_to_sync(&lwb->lwb_txgh); | |
34dc7c2f BB |
725 | } |
726 | ||
727 | /* | |
728 | * Initialize the io for a log block. | |
34dc7c2f BB |
729 | */ |
730 | static void | |
731 | zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) | |
732 | { | |
733 | zbookmark_t zb; | |
734 | ||
735 | zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET]; | |
736 | zb.zb_object = 0; | |
737 | zb.zb_level = -1; | |
738 | zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; | |
739 | ||
740 | if (zilog->zl_root_zio == NULL) { | |
741 | zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, | |
742 | ZIO_FLAG_CANFAIL); | |
743 | } | |
744 | if (lwb->lwb_zio == NULL) { | |
745 | lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, | |
9babb374 BB |
746 | 0, &lwb->lwb_blk, lwb->lwb_buf, lwb->lwb_sz, |
747 | zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, | |
748 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb); | |
34dc7c2f BB |
749 | } |
750 | } | |
751 | ||
752 | /* | |
753 | * Start a log block write and advance to the next log block. | |
754 | * Calls are serialized. | |
755 | */ | |
756 | static lwb_t * | |
757 | zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) | |
758 | { | |
759 | lwb_t *nlwb; | |
760 | zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; | |
761 | spa_t *spa = zilog->zl_spa; | |
762 | blkptr_t *bp = &ztp->zit_next_blk; | |
763 | uint64_t txg; | |
764 | uint64_t zil_blksz; | |
765 | int error; | |
766 | ||
767 | ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); | |
768 | ||
769 | /* | |
770 | * Allocate the next block and save its address in this block | |
771 | * before writing it in order to establish the log chain. | |
772 | * Note that if the allocation of nlwb synced before we wrote | |
773 | * the block that points at it (lwb), we'd leak it if we crashed. | |
774 | * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done(). | |
775 | */ | |
776 | txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh); | |
777 | txg_rele_to_quiesce(&lwb->lwb_txgh); | |
778 | ||
779 | /* | |
780 | * Pick a ZIL blocksize. We request a size that is the | |
781 | * maximum of the previous used size, the current used size and | |
782 | * the amount waiting in the queue. | |
783 | */ | |
784 | zil_blksz = MAX(zilog->zl_prev_used, | |
785 | zilog->zl_cur_used + sizeof (*ztp)); | |
786 | zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); | |
787 | zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); | |
788 | if (zil_blksz > ZIL_MAX_BLKSZ) | |
789 | zil_blksz = ZIL_MAX_BLKSZ; | |
790 | ||
791 | BP_ZERO(bp); | |
792 | /* pass the old blkptr in order to spread log blocks across devs */ | |
793 | error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg); | |
794 | if (error) { | |
795 | dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg); | |
796 | ||
797 | /* | |
798 | * We dirty the dataset to ensure that zil_sync() will | |
799 | * be called to remove this lwb from our zl_lwb_list. | |
800 | * Failing to do so, may leave an lwb with a NULL lwb_buf | |
801 | * hanging around on the zl_lwb_list. | |
802 | */ | |
803 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
804 | dmu_tx_commit(tx); | |
805 | ||
806 | /* | |
807 | * Since we've just experienced an allocation failure so we | |
808 | * terminate the current lwb and send it on its way. | |
809 | */ | |
810 | ztp->zit_pad = 0; | |
811 | ztp->zit_nused = lwb->lwb_nused; | |
812 | ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; | |
813 | zio_nowait(lwb->lwb_zio); | |
814 | ||
815 | /* | |
816 | * By returning NULL the caller will call tx_wait_synced() | |
817 | */ | |
818 | return (NULL); | |
819 | } | |
820 | ||
821 | ASSERT3U(bp->blk_birth, ==, txg); | |
822 | ztp->zit_pad = 0; | |
823 | ztp->zit_nused = lwb->lwb_nused; | |
824 | ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; | |
825 | bp->blk_cksum = lwb->lwb_blk.blk_cksum; | |
826 | bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; | |
827 | ||
828 | /* | |
829 | * Allocate a new log write buffer (lwb). | |
830 | */ | |
831 | nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); | |
832 | ||
833 | nlwb->lwb_zilog = zilog; | |
834 | nlwb->lwb_blk = *bp; | |
835 | nlwb->lwb_nused = 0; | |
836 | nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); | |
837 | nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); | |
838 | nlwb->lwb_max_txg = txg; | |
839 | nlwb->lwb_zio = NULL; | |
840 | ||
841 | /* | |
842 | * Put new lwb at the end of the log chain | |
843 | */ | |
844 | mutex_enter(&zilog->zl_lock); | |
845 | list_insert_tail(&zilog->zl_lwb_list, nlwb); | |
846 | mutex_exit(&zilog->zl_lock); | |
847 | ||
848 | /* Record the block for later vdev flushing */ | |
849 | zil_add_block(zilog, &lwb->lwb_blk); | |
850 | ||
851 | /* | |
852 | * kick off the write for the old log block | |
853 | */ | |
854 | dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg); | |
855 | ASSERT(lwb->lwb_zio); | |
856 | zio_nowait(lwb->lwb_zio); | |
857 | ||
858 | return (nlwb); | |
859 | } | |
860 | ||
861 | static lwb_t * | |
862 | zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) | |
863 | { | |
864 | lr_t *lrc = &itx->itx_lr; /* common log record */ | |
865 | lr_write_t *lr = (lr_write_t *)lrc; | |
866 | uint64_t txg = lrc->lrc_txg; | |
867 | uint64_t reclen = lrc->lrc_reclen; | |
868 | uint64_t dlen; | |
869 | ||
870 | if (lwb == NULL) | |
871 | return (NULL); | |
872 | ASSERT(lwb->lwb_buf != NULL); | |
873 | ||
874 | if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) | |
875 | dlen = P2ROUNDUP_TYPED( | |
876 | lr->lr_length, sizeof (uint64_t), uint64_t); | |
877 | else | |
878 | dlen = 0; | |
879 | ||
880 | zilog->zl_cur_used += (reclen + dlen); | |
881 | ||
882 | zil_lwb_write_init(zilog, lwb); | |
883 | ||
884 | /* | |
885 | * If this record won't fit in the current log block, start a new one. | |
886 | */ | |
887 | if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { | |
888 | lwb = zil_lwb_write_start(zilog, lwb); | |
889 | if (lwb == NULL) | |
890 | return (NULL); | |
891 | zil_lwb_write_init(zilog, lwb); | |
892 | ASSERT(lwb->lwb_nused == 0); | |
893 | if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { | |
894 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
895 | return (lwb); | |
896 | } | |
897 | } | |
898 | ||
899 | /* | |
900 | * Update the lrc_seq, to be log record sequence number. See zil.h | |
901 | * Then copy the record to the log buffer. | |
902 | */ | |
903 | lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ | |
904 | bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen); | |
905 | ||
906 | /* | |
907 | * If it's a write, fetch the data or get its blkptr as appropriate. | |
908 | */ | |
909 | if (lrc->lrc_txtype == TX_WRITE) { | |
910 | if (txg > spa_freeze_txg(zilog->zl_spa)) | |
911 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
912 | if (itx->itx_wr_state != WR_COPIED) { | |
913 | char *dbuf; | |
914 | int error; | |
915 | ||
916 | /* alignment is guaranteed */ | |
917 | lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused); | |
918 | if (dlen) { | |
919 | ASSERT(itx->itx_wr_state == WR_NEED_COPY); | |
920 | dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen; | |
921 | lr->lr_common.lrc_reclen += dlen; | |
922 | } else { | |
923 | ASSERT(itx->itx_wr_state == WR_INDIRECT); | |
924 | dbuf = NULL; | |
925 | } | |
926 | error = zilog->zl_get_data( | |
927 | itx->itx_private, lr, dbuf, lwb->lwb_zio); | |
928 | if (error) { | |
929 | ASSERT(error == ENOENT || error == EEXIST || | |
930 | error == EALREADY); | |
931 | return (lwb); | |
932 | } | |
933 | } | |
934 | } | |
935 | ||
936 | lwb->lwb_nused += reclen + dlen; | |
937 | lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); | |
938 | ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); | |
939 | ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); | |
940 | ||
941 | return (lwb); | |
942 | } | |
943 | ||
944 | itx_t * | |
945 | zil_itx_create(uint64_t txtype, size_t lrsize) | |
946 | { | |
947 | itx_t *itx; | |
948 | ||
949 | lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); | |
950 | ||
951 | itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); | |
952 | itx->itx_lr.lrc_txtype = txtype; | |
953 | itx->itx_lr.lrc_reclen = lrsize; | |
954 | itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ | |
955 | itx->itx_lr.lrc_seq = 0; /* defensive */ | |
956 | ||
957 | return (itx); | |
958 | } | |
959 | ||
960 | uint64_t | |
961 | zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) | |
962 | { | |
963 | uint64_t seq; | |
964 | ||
965 | ASSERT(itx->itx_lr.lrc_seq == 0); | |
966 | ||
967 | mutex_enter(&zilog->zl_lock); | |
968 | list_insert_tail(&zilog->zl_itx_list, itx); | |
969 | zilog->zl_itx_list_sz += itx->itx_sod; | |
970 | itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); | |
971 | itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; | |
972 | mutex_exit(&zilog->zl_lock); | |
973 | ||
974 | return (seq); | |
975 | } | |
976 | ||
977 | /* | |
978 | * Free up all in-memory intent log transactions that have now been synced. | |
979 | */ | |
980 | static void | |
981 | zil_itx_clean(zilog_t *zilog) | |
982 | { | |
983 | uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); | |
984 | uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); | |
985 | list_t clean_list; | |
986 | itx_t *itx; | |
987 | ||
988 | list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); | |
989 | ||
990 | mutex_enter(&zilog->zl_lock); | |
991 | /* wait for a log writer to finish walking list */ | |
992 | while (zilog->zl_writer) { | |
993 | cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); | |
994 | } | |
995 | ||
996 | /* | |
997 | * Move the sync'd log transactions to a separate list so we can call | |
998 | * kmem_free without holding the zl_lock. | |
999 | * | |
1000 | * There is no need to set zl_writer as we don't drop zl_lock here | |
1001 | */ | |
1002 | while ((itx = list_head(&zilog->zl_itx_list)) != NULL && | |
1003 | itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { | |
1004 | list_remove(&zilog->zl_itx_list, itx); | |
1005 | zilog->zl_itx_list_sz -= itx->itx_sod; | |
1006 | list_insert_tail(&clean_list, itx); | |
1007 | } | |
1008 | cv_broadcast(&zilog->zl_cv_writer); | |
1009 | mutex_exit(&zilog->zl_lock); | |
1010 | ||
1011 | /* destroy sync'd log transactions */ | |
1012 | while ((itx = list_head(&clean_list)) != NULL) { | |
1013 | list_remove(&clean_list, itx); | |
1014 | kmem_free(itx, offsetof(itx_t, itx_lr) | |
1015 | + itx->itx_lr.lrc_reclen); | |
1016 | } | |
1017 | list_destroy(&clean_list); | |
1018 | } | |
1019 | ||
1020 | /* | |
1021 | * If there are any in-memory intent log transactions which have now been | |
1022 | * synced then start up a taskq to free them. | |
1023 | */ | |
1024 | void | |
1025 | zil_clean(zilog_t *zilog) | |
1026 | { | |
1027 | itx_t *itx; | |
1028 | ||
1029 | mutex_enter(&zilog->zl_lock); | |
1030 | itx = list_head(&zilog->zl_itx_list); | |
1031 | if ((itx != NULL) && | |
1032 | (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { | |
1033 | (void) taskq_dispatch(zilog->zl_clean_taskq, | |
9babb374 | 1034 | (task_func_t *)zil_itx_clean, zilog, TQ_SLEEP); |
34dc7c2f BB |
1035 | } |
1036 | mutex_exit(&zilog->zl_lock); | |
1037 | } | |
1038 | ||
b128c09f | 1039 | static void |
34dc7c2f BB |
1040 | zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) |
1041 | { | |
1042 | uint64_t txg; | |
1043 | uint64_t commit_seq = 0; | |
1044 | itx_t *itx, *itx_next = (itx_t *)-1; | |
1045 | lwb_t *lwb; | |
1046 | spa_t *spa; | |
1047 | ||
1048 | zilog->zl_writer = B_TRUE; | |
b128c09f | 1049 | ASSERT(zilog->zl_root_zio == NULL); |
34dc7c2f BB |
1050 | spa = zilog->zl_spa; |
1051 | ||
1052 | if (zilog->zl_suspend) { | |
1053 | lwb = NULL; | |
1054 | } else { | |
1055 | lwb = list_tail(&zilog->zl_lwb_list); | |
1056 | if (lwb == NULL) { | |
1057 | /* | |
1058 | * Return if there's nothing to flush before we | |
1059 | * dirty the fs by calling zil_create() | |
1060 | */ | |
1061 | if (list_is_empty(&zilog->zl_itx_list)) { | |
1062 | zilog->zl_writer = B_FALSE; | |
1063 | return; | |
1064 | } | |
1065 | mutex_exit(&zilog->zl_lock); | |
1066 | zil_create(zilog); | |
1067 | mutex_enter(&zilog->zl_lock); | |
1068 | lwb = list_tail(&zilog->zl_lwb_list); | |
1069 | } | |
1070 | } | |
1071 | ||
1072 | /* Loop through in-memory log transactions filling log blocks. */ | |
1073 | DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); | |
1074 | for (;;) { | |
1075 | /* | |
1076 | * Find the next itx to push: | |
1077 | * Push all transactions related to specified foid and all | |
1078 | * other transactions except TX_WRITE, TX_TRUNCATE, | |
1079 | * TX_SETATTR and TX_ACL for all other files. | |
1080 | */ | |
1081 | if (itx_next != (itx_t *)-1) | |
1082 | itx = itx_next; | |
1083 | else | |
1084 | itx = list_head(&zilog->zl_itx_list); | |
1085 | for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) { | |
1086 | if (foid == 0) /* push all foids? */ | |
1087 | break; | |
1088 | if (itx->itx_sync) /* push all O_[D]SYNC */ | |
1089 | break; | |
1090 | switch (itx->itx_lr.lrc_txtype) { | |
1091 | case TX_SETATTR: | |
1092 | case TX_WRITE: | |
1093 | case TX_TRUNCATE: | |
1094 | case TX_ACL: | |
1095 | /* lr_foid is same offset for these records */ | |
1096 | if (((lr_write_t *)&itx->itx_lr)->lr_foid | |
1097 | != foid) { | |
1098 | continue; /* skip this record */ | |
1099 | } | |
1100 | } | |
1101 | break; | |
1102 | } | |
1103 | if (itx == NULL) | |
1104 | break; | |
1105 | ||
1106 | if ((itx->itx_lr.lrc_seq > seq) && | |
1107 | ((lwb == NULL) || (lwb->lwb_nused == 0) || | |
1108 | (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) { | |
1109 | break; | |
1110 | } | |
1111 | ||
1112 | /* | |
1113 | * Save the next pointer. Even though we soon drop | |
1114 | * zl_lock all threads that may change the list | |
1115 | * (another writer or zil_itx_clean) can't do so until | |
1116 | * they have zl_writer. | |
1117 | */ | |
1118 | itx_next = list_next(&zilog->zl_itx_list, itx); | |
1119 | list_remove(&zilog->zl_itx_list, itx); | |
1120 | zilog->zl_itx_list_sz -= itx->itx_sod; | |
1121 | mutex_exit(&zilog->zl_lock); | |
1122 | txg = itx->itx_lr.lrc_txg; | |
1123 | ASSERT(txg); | |
1124 | ||
1125 | if (txg > spa_last_synced_txg(spa) || | |
1126 | txg > spa_freeze_txg(spa)) | |
1127 | lwb = zil_lwb_commit(zilog, itx, lwb); | |
1128 | kmem_free(itx, offsetof(itx_t, itx_lr) | |
1129 | + itx->itx_lr.lrc_reclen); | |
1130 | mutex_enter(&zilog->zl_lock); | |
1131 | } | |
1132 | DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); | |
1133 | /* determine commit sequence number */ | |
1134 | itx = list_head(&zilog->zl_itx_list); | |
1135 | if (itx) | |
1136 | commit_seq = itx->itx_lr.lrc_seq; | |
1137 | else | |
1138 | commit_seq = zilog->zl_itx_seq; | |
1139 | mutex_exit(&zilog->zl_lock); | |
1140 | ||
1141 | /* write the last block out */ | |
1142 | if (lwb != NULL && lwb->lwb_zio != NULL) | |
1143 | lwb = zil_lwb_write_start(zilog, lwb); | |
1144 | ||
1145 | zilog->zl_prev_used = zilog->zl_cur_used; | |
1146 | zilog->zl_cur_used = 0; | |
1147 | ||
1148 | /* | |
1149 | * Wait if necessary for the log blocks to be on stable storage. | |
1150 | */ | |
1151 | if (zilog->zl_root_zio) { | |
1152 | DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); | |
1153 | (void) zio_wait(zilog->zl_root_zio); | |
b128c09f | 1154 | zilog->zl_root_zio = NULL; |
34dc7c2f BB |
1155 | DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); |
1156 | zil_flush_vdevs(zilog); | |
1157 | } | |
1158 | ||
1159 | if (zilog->zl_log_error || lwb == NULL) { | |
1160 | zilog->zl_log_error = 0; | |
1161 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
1162 | } | |
1163 | ||
1164 | mutex_enter(&zilog->zl_lock); | |
1165 | zilog->zl_writer = B_FALSE; | |
1166 | ||
1167 | ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); | |
1168 | zilog->zl_commit_seq = commit_seq; | |
1169 | } | |
1170 | ||
1171 | /* | |
1172 | * Push zfs transactions to stable storage up to the supplied sequence number. | |
1173 | * If foid is 0 push out all transactions, otherwise push only those | |
1174 | * for that file or might have been used to create that file. | |
1175 | */ | |
1176 | void | |
1177 | zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) | |
1178 | { | |
1179 | if (zilog == NULL || seq == 0) | |
1180 | return; | |
1181 | ||
1182 | mutex_enter(&zilog->zl_lock); | |
1183 | ||
1184 | seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ | |
1185 | ||
1186 | while (zilog->zl_writer) { | |
1187 | cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); | |
1188 | if (seq < zilog->zl_commit_seq) { | |
1189 | mutex_exit(&zilog->zl_lock); | |
1190 | return; | |
1191 | } | |
1192 | } | |
1193 | zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ | |
1194 | /* wake up others waiting on the commit */ | |
1195 | cv_broadcast(&zilog->zl_cv_writer); | |
1196 | mutex_exit(&zilog->zl_lock); | |
1197 | } | |
1198 | ||
1199 | /* | |
1200 | * Called in syncing context to free committed log blocks and update log header. | |
1201 | */ | |
1202 | void | |
1203 | zil_sync(zilog_t *zilog, dmu_tx_t *tx) | |
1204 | { | |
1205 | zil_header_t *zh = zil_header_in_syncing_context(zilog); | |
1206 | uint64_t txg = dmu_tx_get_txg(tx); | |
1207 | spa_t *spa = zilog->zl_spa; | |
1208 | lwb_t *lwb; | |
1209 | ||
9babb374 BB |
1210 | /* |
1211 | * We don't zero out zl_destroy_txg, so make sure we don't try | |
1212 | * to destroy it twice. | |
1213 | */ | |
1214 | if (spa_sync_pass(spa) != 1) | |
1215 | return; | |
1216 | ||
34dc7c2f BB |
1217 | mutex_enter(&zilog->zl_lock); |
1218 | ||
1219 | ASSERT(zilog->zl_stop_sync == 0); | |
1220 | ||
fb5f0bc8 | 1221 | zh->zh_replay_seq = zilog->zl_replayed_seq[txg & TXG_MASK]; |
34dc7c2f BB |
1222 | |
1223 | if (zilog->zl_destroy_txg == txg) { | |
1224 | blkptr_t blk = zh->zh_log; | |
1225 | ||
1226 | ASSERT(list_head(&zilog->zl_lwb_list) == NULL); | |
34dc7c2f BB |
1227 | |
1228 | bzero(zh, sizeof (zil_header_t)); | |
fb5f0bc8 | 1229 | bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); |
34dc7c2f BB |
1230 | |
1231 | if (zilog->zl_keep_first) { | |
1232 | /* | |
1233 | * If this block was part of log chain that couldn't | |
1234 | * be claimed because a device was missing during | |
1235 | * zil_claim(), but that device later returns, | |
1236 | * then this block could erroneously appear valid. | |
1237 | * To guard against this, assign a new GUID to the new | |
1238 | * log chain so it doesn't matter what blk points to. | |
1239 | */ | |
1240 | zil_init_log_chain(zilog, &blk); | |
1241 | zh->zh_log = blk; | |
1242 | } | |
1243 | } | |
1244 | ||
9babb374 | 1245 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { |
34dc7c2f BB |
1246 | zh->zh_log = lwb->lwb_blk; |
1247 | if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) | |
1248 | break; | |
1249 | list_remove(&zilog->zl_lwb_list, lwb); | |
1250 | zio_free_blk(spa, &lwb->lwb_blk, txg); | |
1251 | kmem_cache_free(zil_lwb_cache, lwb); | |
1252 | ||
1253 | /* | |
1254 | * If we don't have anything left in the lwb list then | |
1255 | * we've had an allocation failure and we need to zero | |
1256 | * out the zil_header blkptr so that we don't end | |
1257 | * up freeing the same block twice. | |
1258 | */ | |
1259 | if (list_head(&zilog->zl_lwb_list) == NULL) | |
1260 | BP_ZERO(&zh->zh_log); | |
1261 | } | |
1262 | mutex_exit(&zilog->zl_lock); | |
1263 | } | |
1264 | ||
1265 | void | |
1266 | zil_init(void) | |
1267 | { | |
1268 | zil_lwb_cache = kmem_cache_create("zil_lwb_cache", | |
1269 | sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
1270 | } | |
1271 | ||
1272 | void | |
1273 | zil_fini(void) | |
1274 | { | |
1275 | kmem_cache_destroy(zil_lwb_cache); | |
1276 | } | |
1277 | ||
1278 | zilog_t * | |
1279 | zil_alloc(objset_t *os, zil_header_t *zh_phys) | |
1280 | { | |
1281 | zilog_t *zilog; | |
1282 | ||
1283 | zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); | |
1284 | ||
1285 | zilog->zl_header = zh_phys; | |
1286 | zilog->zl_os = os; | |
1287 | zilog->zl_spa = dmu_objset_spa(os); | |
1288 | zilog->zl_dmu_pool = dmu_objset_pool(os); | |
1289 | zilog->zl_destroy_txg = TXG_INITIAL - 1; | |
1290 | ||
1291 | mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); | |
1292 | ||
1293 | list_create(&zilog->zl_itx_list, sizeof (itx_t), | |
1294 | offsetof(itx_t, itx_node)); | |
1295 | ||
1296 | list_create(&zilog->zl_lwb_list, sizeof (lwb_t), | |
1297 | offsetof(lwb_t, lwb_node)); | |
1298 | ||
1299 | mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); | |
1300 | ||
1301 | avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, | |
1302 | sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); | |
1303 | ||
1304 | cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); | |
1305 | cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); | |
1306 | ||
1307 | return (zilog); | |
1308 | } | |
1309 | ||
1310 | void | |
1311 | zil_free(zilog_t *zilog) | |
1312 | { | |
1313 | lwb_t *lwb; | |
1314 | ||
1315 | zilog->zl_stop_sync = 1; | |
1316 | ||
1317 | while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { | |
1318 | list_remove(&zilog->zl_lwb_list, lwb); | |
1319 | if (lwb->lwb_buf != NULL) | |
1320 | zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); | |
1321 | kmem_cache_free(zil_lwb_cache, lwb); | |
1322 | } | |
1323 | list_destroy(&zilog->zl_lwb_list); | |
1324 | ||
1325 | avl_destroy(&zilog->zl_vdev_tree); | |
1326 | mutex_destroy(&zilog->zl_vdev_lock); | |
1327 | ||
1328 | ASSERT(list_head(&zilog->zl_itx_list) == NULL); | |
1329 | list_destroy(&zilog->zl_itx_list); | |
1330 | mutex_destroy(&zilog->zl_lock); | |
1331 | ||
1332 | cv_destroy(&zilog->zl_cv_writer); | |
1333 | cv_destroy(&zilog->zl_cv_suspend); | |
1334 | ||
1335 | kmem_free(zilog, sizeof (zilog_t)); | |
1336 | } | |
1337 | ||
34dc7c2f BB |
1338 | /* |
1339 | * Open an intent log. | |
1340 | */ | |
1341 | zilog_t * | |
1342 | zil_open(objset_t *os, zil_get_data_t *get_data) | |
1343 | { | |
1344 | zilog_t *zilog = dmu_objset_zil(os); | |
1345 | ||
1346 | zilog->zl_get_data = get_data; | |
1347 | zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, | |
1348 | 2, 2, TASKQ_PREPOPULATE); | |
1349 | ||
1350 | return (zilog); | |
1351 | } | |
1352 | ||
1353 | /* | |
1354 | * Close an intent log. | |
1355 | */ | |
1356 | void | |
1357 | zil_close(zilog_t *zilog) | |
1358 | { | |
1359 | /* | |
1360 | * If the log isn't already committed, mark the objset dirty | |
1361 | * (so zil_sync() will be called) and wait for that txg to sync. | |
1362 | */ | |
1363 | if (!zil_is_committed(zilog)) { | |
1364 | uint64_t txg; | |
1365 | dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); | |
1366 | (void) dmu_tx_assign(tx, TXG_WAIT); | |
1367 | dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); | |
1368 | txg = dmu_tx_get_txg(tx); | |
1369 | dmu_tx_commit(tx); | |
1370 | txg_wait_synced(zilog->zl_dmu_pool, txg); | |
1371 | } | |
1372 | ||
1373 | taskq_destroy(zilog->zl_clean_taskq); | |
1374 | zilog->zl_clean_taskq = NULL; | |
1375 | zilog->zl_get_data = NULL; | |
1376 | ||
1377 | zil_itx_clean(zilog); | |
1378 | ASSERT(list_head(&zilog->zl_itx_list) == NULL); | |
1379 | } | |
1380 | ||
1381 | /* | |
1382 | * Suspend an intent log. While in suspended mode, we still honor | |
1383 | * synchronous semantics, but we rely on txg_wait_synced() to do it. | |
1384 | * We suspend the log briefly when taking a snapshot so that the snapshot | |
1385 | * contains all the data it's supposed to, and has an empty intent log. | |
1386 | */ | |
1387 | int | |
1388 | zil_suspend(zilog_t *zilog) | |
1389 | { | |
1390 | const zil_header_t *zh = zilog->zl_header; | |
1391 | ||
1392 | mutex_enter(&zilog->zl_lock); | |
9babb374 | 1393 | if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ |
34dc7c2f BB |
1394 | mutex_exit(&zilog->zl_lock); |
1395 | return (EBUSY); | |
1396 | } | |
1397 | if (zilog->zl_suspend++ != 0) { | |
1398 | /* | |
1399 | * Someone else already began a suspend. | |
1400 | * Just wait for them to finish. | |
1401 | */ | |
1402 | while (zilog->zl_suspending) | |
1403 | cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); | |
34dc7c2f BB |
1404 | mutex_exit(&zilog->zl_lock); |
1405 | return (0); | |
1406 | } | |
1407 | zilog->zl_suspending = B_TRUE; | |
1408 | mutex_exit(&zilog->zl_lock); | |
1409 | ||
1410 | zil_commit(zilog, UINT64_MAX, 0); | |
1411 | ||
1412 | /* | |
1413 | * Wait for any in-flight log writes to complete. | |
1414 | */ | |
1415 | mutex_enter(&zilog->zl_lock); | |
1416 | while (zilog->zl_writer) | |
1417 | cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); | |
1418 | mutex_exit(&zilog->zl_lock); | |
1419 | ||
1420 | zil_destroy(zilog, B_FALSE); | |
1421 | ||
1422 | mutex_enter(&zilog->zl_lock); | |
1423 | zilog->zl_suspending = B_FALSE; | |
1424 | cv_broadcast(&zilog->zl_cv_suspend); | |
1425 | mutex_exit(&zilog->zl_lock); | |
1426 | ||
1427 | return (0); | |
1428 | } | |
1429 | ||
1430 | void | |
1431 | zil_resume(zilog_t *zilog) | |
1432 | { | |
1433 | mutex_enter(&zilog->zl_lock); | |
1434 | ASSERT(zilog->zl_suspend != 0); | |
1435 | zilog->zl_suspend--; | |
1436 | mutex_exit(&zilog->zl_lock); | |
1437 | } | |
1438 | ||
1439 | typedef struct zil_replay_arg { | |
1440 | objset_t *zr_os; | |
1441 | zil_replay_func_t **zr_replay; | |
1442 | void *zr_arg; | |
34dc7c2f BB |
1443 | boolean_t zr_byteswap; |
1444 | char *zr_lrbuf; | |
1445 | } zil_replay_arg_t; | |
1446 | ||
1447 | static void | |
1448 | zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) | |
1449 | { | |
1450 | zil_replay_arg_t *zr = zra; | |
1451 | const zil_header_t *zh = zilog->zl_header; | |
1452 | uint64_t reclen = lr->lrc_reclen; | |
1453 | uint64_t txtype = lr->lrc_txtype; | |
1454 | char *name; | |
fb5f0bc8 | 1455 | int pass, error; |
34dc7c2f | 1456 | |
fb5f0bc8 | 1457 | if (!zilog->zl_replay) /* giving up */ |
34dc7c2f BB |
1458 | return; |
1459 | ||
1460 | if (lr->lrc_txg < claim_txg) /* already committed */ | |
1461 | return; | |
1462 | ||
1463 | if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ | |
1464 | return; | |
1465 | ||
1466 | /* Strip case-insensitive bit, still present in log record */ | |
1467 | txtype &= ~TX_CI; | |
1468 | ||
fb5f0bc8 BB |
1469 | if (txtype == 0 || txtype >= TX_MAX_TYPE) { |
1470 | error = EINVAL; | |
1471 | goto bad; | |
1472 | } | |
1473 | ||
34dc7c2f BB |
1474 | /* |
1475 | * Make a copy of the data so we can revise and extend it. | |
1476 | */ | |
1477 | bcopy(lr, zr->zr_lrbuf, reclen); | |
1478 | ||
1479 | /* | |
1480 | * The log block containing this lr may have been byteswapped | |
1481 | * so that we can easily examine common fields like lrc_txtype. | |
1482 | * However, the log is a mix of different data types, and only the | |
1483 | * replay vectors know how to byteswap their records. Therefore, if | |
1484 | * the lr was byteswapped, undo it before invoking the replay vector. | |
1485 | */ | |
1486 | if (zr->zr_byteswap) | |
1487 | byteswap_uint64_array(zr->zr_lrbuf, reclen); | |
1488 | ||
1489 | /* | |
1490 | * If this is a TX_WRITE with a blkptr, suck in the data. | |
1491 | */ | |
1492 | if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { | |
1493 | lr_write_t *lrw = (lr_write_t *)lr; | |
1494 | blkptr_t *wbp = &lrw->lr_blkptr; | |
1495 | uint64_t wlen = lrw->lr_length; | |
1496 | char *wbuf = zr->zr_lrbuf + reclen; | |
1497 | ||
1498 | if (BP_IS_HOLE(wbp)) { /* compressed to a hole */ | |
1499 | bzero(wbuf, wlen); | |
1500 | } else { | |
1501 | /* | |
1502 | * A subsequent write may have overwritten this block, | |
1503 | * in which case wbp may have been been freed and | |
1504 | * reallocated, and our read of wbp may fail with a | |
1505 | * checksum error. We can safely ignore this because | |
1506 | * the later write will provide the correct data. | |
1507 | */ | |
1508 | zbookmark_t zb; | |
1509 | ||
1510 | zb.zb_objset = dmu_objset_id(zilog->zl_os); | |
1511 | zb.zb_object = lrw->lr_foid; | |
1512 | zb.zb_level = -1; | |
1513 | zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp); | |
1514 | ||
1515 | (void) zio_wait(zio_read(NULL, zilog->zl_spa, | |
1516 | wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL, | |
1517 | ZIO_PRIORITY_SYNC_READ, | |
1518 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb)); | |
1519 | (void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen); | |
1520 | } | |
1521 | } | |
1522 | ||
1523 | /* | |
1524 | * We must now do two things atomically: replay this log record, | |
fb5f0bc8 BB |
1525 | * and update the log header sequence number to reflect the fact that |
1526 | * we did so. At the end of each replay function the sequence number | |
1527 | * is updated if we are in replay mode. | |
34dc7c2f | 1528 | */ |
fb5f0bc8 BB |
1529 | for (pass = 1; pass <= 2; pass++) { |
1530 | zilog->zl_replaying_seq = lr->lrc_seq; | |
1531 | /* Only byteswap (if needed) on the 1st pass. */ | |
1532 | error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf, | |
1533 | zr->zr_byteswap && pass == 1); | |
34dc7c2f BB |
1534 | |
1535 | if (!error) | |
1536 | return; | |
1537 | ||
1538 | /* | |
1539 | * The DMU's dnode layer doesn't see removes until the txg | |
1540 | * commits, so a subsequent claim can spuriously fail with | |
fb5f0bc8 BB |
1541 | * EEXIST. So if we receive any error we try syncing out |
1542 | * any removes then retry the transaction. | |
34dc7c2f | 1543 | */ |
fb5f0bc8 | 1544 | if (pass == 1) |
34dc7c2f | 1545 | txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); |
34dc7c2f BB |
1546 | } |
1547 | ||
b128c09f | 1548 | bad: |
fb5f0bc8 | 1549 | ASSERT(error); |
34dc7c2f BB |
1550 | name = kmem_alloc(MAXNAMELEN, KM_SLEEP); |
1551 | dmu_objset_name(zr->zr_os, name); | |
1552 | cmn_err(CE_WARN, "ZFS replay transaction error %d, " | |
1553 | "dataset %s, seq 0x%llx, txtype %llu %s\n", | |
1554 | error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype, | |
1555 | (lr->lrc_txtype & TX_CI) ? "CI" : ""); | |
fb5f0bc8 | 1556 | zilog->zl_replay = B_FALSE; |
34dc7c2f BB |
1557 | kmem_free(name, MAXNAMELEN); |
1558 | } | |
1559 | ||
1560 | /* ARGSUSED */ | |
1561 | static void | |
1562 | zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) | |
1563 | { | |
1564 | zilog->zl_replay_blks++; | |
1565 | } | |
1566 | ||
1567 | /* | |
1568 | * If this dataset has a non-empty intent log, replay it and destroy it. | |
1569 | */ | |
1570 | void | |
fb5f0bc8 | 1571 | zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) |
34dc7c2f BB |
1572 | { |
1573 | zilog_t *zilog = dmu_objset_zil(os); | |
1574 | const zil_header_t *zh = zilog->zl_header; | |
1575 | zil_replay_arg_t zr; | |
1576 | ||
9babb374 | 1577 | if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { |
34dc7c2f BB |
1578 | zil_destroy(zilog, B_TRUE); |
1579 | return; | |
1580 | } | |
1581 | ||
1582 | zr.zr_os = os; | |
1583 | zr.zr_replay = replay_func; | |
1584 | zr.zr_arg = arg; | |
34dc7c2f BB |
1585 | zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); |
1586 | zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); | |
1587 | ||
1588 | /* | |
1589 | * Wait for in-progress removes to sync before starting replay. | |
1590 | */ | |
1591 | txg_wait_synced(zilog->zl_dmu_pool, 0); | |
1592 | ||
fb5f0bc8 | 1593 | zilog->zl_replay = B_TRUE; |
34dc7c2f BB |
1594 | zilog->zl_replay_time = lbolt; |
1595 | ASSERT(zilog->zl_replay_blks == 0); | |
1596 | (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, | |
1597 | zh->zh_claim_txg); | |
1598 | kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE); | |
1599 | ||
1600 | zil_destroy(zilog, B_FALSE); | |
1601 | txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); | |
fb5f0bc8 | 1602 | zilog->zl_replay = B_FALSE; |
34dc7c2f BB |
1603 | } |
1604 | ||
1605 | /* | |
1606 | * Report whether all transactions are committed | |
1607 | */ | |
1608 | int | |
1609 | zil_is_committed(zilog_t *zilog) | |
1610 | { | |
1611 | lwb_t *lwb; | |
1612 | int ret; | |
1613 | ||
1614 | mutex_enter(&zilog->zl_lock); | |
1615 | while (zilog->zl_writer) | |
1616 | cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); | |
1617 | ||
1618 | /* recent unpushed intent log transactions? */ | |
1619 | if (!list_is_empty(&zilog->zl_itx_list)) { | |
1620 | ret = B_FALSE; | |
1621 | goto out; | |
1622 | } | |
1623 | ||
1624 | /* intent log never used? */ | |
1625 | lwb = list_head(&zilog->zl_lwb_list); | |
1626 | if (lwb == NULL) { | |
1627 | ret = B_TRUE; | |
1628 | goto out; | |
1629 | } | |
1630 | ||
1631 | /* | |
1632 | * more than 1 log buffer means zil_sync() hasn't yet freed | |
1633 | * entries after a txg has committed | |
1634 | */ | |
1635 | if (list_next(&zilog->zl_lwb_list, lwb)) { | |
1636 | ret = B_FALSE; | |
1637 | goto out; | |
1638 | } | |
1639 | ||
1640 | ASSERT(zil_empty(zilog)); | |
1641 | ret = B_TRUE; | |
1642 | out: | |
1643 | cv_broadcast(&zilog->zl_cv_writer); | |
1644 | mutex_exit(&zilog->zl_lock); | |
1645 | return (ret); | |
1646 | } | |
9babb374 BB |
1647 | |
1648 | /* ARGSUSED */ | |
1649 | int | |
1650 | zil_vdev_offline(char *osname, void *arg) | |
1651 | { | |
1652 | objset_t *os; | |
1653 | zilog_t *zilog; | |
1654 | int error; | |
1655 | ||
1656 | error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); | |
1657 | if (error) | |
1658 | return (error); | |
1659 | ||
1660 | zilog = dmu_objset_zil(os); | |
1661 | if (zil_suspend(zilog) != 0) | |
1662 | error = EEXIST; | |
1663 | else | |
1664 | zil_resume(zilog); | |
1665 | dmu_objset_close(os); | |
1666 | return (error); | |
1667 | } |