Print this page
5269 zfs: zpool import slow
While importing a pool all objsets are enumerated twice, once to check
the zil log chains and once to claim them. On pools with many datasets
this process might take a substantial amount of time.
Speed up the process by parallelizing it utilizing a taskq. The number
of parallel tasks is limited to 4 times the number of leaf vdevs.
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/zil.c
+++ new/usr/src/uts/common/fs/zfs/zil.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 24 */
25 25
26 26 /* Portions Copyright 2010 Robert Milkowski */
27 27
28 28 #include <sys/zfs_context.h>
29 29 #include <sys/spa.h>
30 30 #include <sys/dmu.h>
31 31 #include <sys/zap.h>
32 32 #include <sys/arc.h>
33 33 #include <sys/stat.h>
34 34 #include <sys/resource.h>
35 35 #include <sys/zil.h>
36 36 #include <sys/zil_impl.h>
37 37 #include <sys/dsl_dataset.h>
38 38 #include <sys/vdev_impl.h>
39 39 #include <sys/dmu_tx.h>
40 40 #include <sys/dsl_pool.h>
41 41
42 42 /*
43 43 * The zfs intent log (ZIL) saves transaction records of system calls
44 44 * that change the file system in memory with enough information
45 45 * to be able to replay them. These are stored in memory until
46 46 * either the DMU transaction group (txg) commits them to the stable pool
47 47 * and they can be discarded, or they are flushed to the stable log
48 48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
49 49 * requirement. In the event of a panic or power fail then those log
50 50 * records (transactions) are replayed.
51 51 *
52 52 * There is one ZIL per file system. Its on-disk (pool) format consists
53 53 * of 3 parts:
54 54 *
55 55 * - ZIL header
56 56 * - ZIL blocks
57 57 * - ZIL records
58 58 *
59 59 * A log record holds a system call transaction. Log blocks can
60 60 * hold many log records and the blocks are chained together.
61 61 * Each ZIL block contains a block pointer (blkptr_t) to the next
62 62 * ZIL block in the chain. The ZIL header points to the first
63 63 * block in the chain. Note there is not a fixed place in the pool
64 64 * to hold blocks. They are dynamically allocated and freed as
65 65 * needed from the blocks available. Figure X shows the ZIL structure:
66 66 */
67 67
68 68 /*
69 69 * Disable intent logging replay. This global ZIL switch affects all pools.
70 70 */
71 71 int zil_replay_disable = 0;
72 72
73 73 /*
74 74 * Tunable parameter for debugging or performance analysis. Setting
75 75 * zfs_nocacheflush will cause corruption on power loss if a volatile
76 76 * out-of-order write cache is enabled.
77 77 */
78 78 boolean_t zfs_nocacheflush = B_FALSE;
79 79
80 80 static kmem_cache_t *zil_lwb_cache;
81 81
82 82 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid);
83 83
84 84 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
85 85 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
86 86
87 87
88 88 /*
89 89 * ziltest is by and large an ugly hack, but very useful in
90 90 * checking replay without tedious work.
91 91 * When running ziltest we want to keep all itx's and so maintain
92 92 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
93 93 * We subtract TXG_CONCURRENT_STATES to allow for common code.
94 94 */
95 95 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
96 96
97 97 static int
98 98 zil_bp_compare(const void *x1, const void *x2)
99 99 {
100 100 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
101 101 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
102 102
103 103 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
104 104 return (-1);
105 105 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
106 106 return (1);
107 107
108 108 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
109 109 return (-1);
110 110 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
111 111 return (1);
112 112
113 113 return (0);
114 114 }
115 115
116 116 static void
117 117 zil_bp_tree_init(zilog_t *zilog)
118 118 {
119 119 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
120 120 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
121 121 }
122 122
123 123 static void
124 124 zil_bp_tree_fini(zilog_t *zilog)
125 125 {
126 126 avl_tree_t *t = &zilog->zl_bp_tree;
127 127 zil_bp_node_t *zn;
128 128 void *cookie = NULL;
129 129
130 130 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
131 131 kmem_free(zn, sizeof (zil_bp_node_t));
132 132
133 133 avl_destroy(t);
134 134 }
135 135
136 136 int
137 137 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
138 138 {
139 139 avl_tree_t *t = &zilog->zl_bp_tree;
140 140 const dva_t *dva;
141 141 zil_bp_node_t *zn;
142 142 avl_index_t where;
143 143
144 144 if (BP_IS_EMBEDDED(bp))
145 145 return (0);
146 146
147 147 dva = BP_IDENTITY(bp);
148 148
149 149 if (avl_find(t, dva, &where) != NULL)
150 150 return (SET_ERROR(EEXIST));
151 151
152 152 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
153 153 zn->zn_dva = *dva;
154 154 avl_insert(t, zn, where);
155 155
156 156 return (0);
157 157 }
158 158
159 159 static zil_header_t *
160 160 zil_header_in_syncing_context(zilog_t *zilog)
161 161 {
162 162 return ((zil_header_t *)zilog->zl_header);
163 163 }
164 164
165 165 static void
166 166 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
167 167 {
168 168 zio_cksum_t *zc = &bp->blk_cksum;
169 169
170 170 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
171 171 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
172 172 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
173 173 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
174 174 }
175 175
176 176 /*
177 177 * Read a log block and make sure it's valid.
178 178 */
179 179 static int
180 180 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
181 181 char **end)
182 182 {
183 183 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
184 184 uint32_t aflags = ARC_WAIT;
185 185 arc_buf_t *abuf = NULL;
186 186 zbookmark_t zb;
187 187 int error;
188 188
189 189 if (zilog->zl_header->zh_claim_txg == 0)
190 190 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
191 191
192 192 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
193 193 zio_flags |= ZIO_FLAG_SPECULATIVE;
194 194
195 195 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
196 196 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
197 197
198 198 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
199 199 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
200 200
201 201 if (error == 0) {
202 202 zio_cksum_t cksum = bp->blk_cksum;
203 203
204 204 /*
205 205 * Validate the checksummed log block.
206 206 *
207 207 * Sequence numbers should be... sequential. The checksum
208 208 * verifier for the next block should be bp's checksum plus 1.
209 209 *
210 210 * Also check the log chain linkage and size used.
211 211 */
212 212 cksum.zc_word[ZIL_ZC_SEQ]++;
213 213
214 214 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
215 215 zil_chain_t *zilc = abuf->b_data;
216 216 char *lr = (char *)(zilc + 1);
217 217 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
218 218
219 219 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
220 220 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
221 221 error = SET_ERROR(ECKSUM);
222 222 } else {
223 223 bcopy(lr, dst, len);
224 224 *end = (char *)dst + len;
225 225 *nbp = zilc->zc_next_blk;
226 226 }
227 227 } else {
228 228 char *lr = abuf->b_data;
229 229 uint64_t size = BP_GET_LSIZE(bp);
230 230 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
231 231
232 232 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
233 233 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
234 234 (zilc->zc_nused > (size - sizeof (*zilc)))) {
235 235 error = SET_ERROR(ECKSUM);
236 236 } else {
237 237 bcopy(lr, dst, zilc->zc_nused);
238 238 *end = (char *)dst + zilc->zc_nused;
239 239 *nbp = zilc->zc_next_blk;
240 240 }
241 241 }
242 242
243 243 VERIFY(arc_buf_remove_ref(abuf, &abuf));
244 244 }
245 245
246 246 return (error);
247 247 }
248 248
249 249 /*
250 250 * Read a TX_WRITE log data block.
251 251 */
252 252 static int
253 253 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
254 254 {
255 255 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
256 256 const blkptr_t *bp = &lr->lr_blkptr;
257 257 uint32_t aflags = ARC_WAIT;
258 258 arc_buf_t *abuf = NULL;
259 259 zbookmark_t zb;
260 260 int error;
261 261
262 262 if (BP_IS_HOLE(bp)) {
263 263 if (wbuf != NULL)
264 264 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
265 265 return (0);
266 266 }
267 267
268 268 if (zilog->zl_header->zh_claim_txg == 0)
269 269 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
270 270
271 271 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
272 272 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
273 273
274 274 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
275 275 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
276 276
277 277 if (error == 0) {
278 278 if (wbuf != NULL)
279 279 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
280 280 (void) arc_buf_remove_ref(abuf, &abuf);
281 281 }
282 282
283 283 return (error);
284 284 }
285 285
286 286 /*
287 287 * Parse the intent log, and call parse_func for each valid record within.
288 288 */
289 289 int
290 290 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
291 291 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
292 292 {
293 293 const zil_header_t *zh = zilog->zl_header;
294 294 boolean_t claimed = !!zh->zh_claim_txg;
295 295 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
296 296 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
297 297 uint64_t max_blk_seq = 0;
298 298 uint64_t max_lr_seq = 0;
299 299 uint64_t blk_count = 0;
300 300 uint64_t lr_count = 0;
301 301 blkptr_t blk, next_blk;
302 302 char *lrbuf, *lrp;
303 303 int error = 0;
304 304
305 305 /*
306 306 * Old logs didn't record the maximum zh_claim_lr_seq.
307 307 */
308 308 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
309 309 claim_lr_seq = UINT64_MAX;
310 310
311 311 /*
312 312 * Starting at the block pointed to by zh_log we read the log chain.
313 313 * For each block in the chain we strongly check that block to
314 314 * ensure its validity. We stop when an invalid block is found.
315 315 * For each block pointer in the chain we call parse_blk_func().
316 316 * For each record in each valid block we call parse_lr_func().
317 317 * If the log has been claimed, stop if we encounter a sequence
318 318 * number greater than the highest claimed sequence number.
319 319 */
320 320 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
321 321 zil_bp_tree_init(zilog);
322 322
323 323 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
324 324 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
325 325 int reclen;
326 326 char *end;
327 327
328 328 if (blk_seq > claim_blk_seq)
329 329 break;
330 330 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
331 331 break;
332 332 ASSERT3U(max_blk_seq, <, blk_seq);
333 333 max_blk_seq = blk_seq;
334 334 blk_count++;
335 335
336 336 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
337 337 break;
338 338
339 339 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
340 340 if (error != 0)
341 341 break;
342 342
343 343 for (lrp = lrbuf; lrp < end; lrp += reclen) {
344 344 lr_t *lr = (lr_t *)lrp;
345 345 reclen = lr->lrc_reclen;
346 346 ASSERT3U(reclen, >=, sizeof (lr_t));
347 347 if (lr->lrc_seq > claim_lr_seq)
348 348 goto done;
349 349 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
350 350 goto done;
351 351 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
352 352 max_lr_seq = lr->lrc_seq;
353 353 lr_count++;
354 354 }
355 355 }
356 356 done:
357 357 zilog->zl_parse_error = error;
358 358 zilog->zl_parse_blk_seq = max_blk_seq;
359 359 zilog->zl_parse_lr_seq = max_lr_seq;
360 360 zilog->zl_parse_blk_count = blk_count;
361 361 zilog->zl_parse_lr_count = lr_count;
362 362
363 363 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
364 364 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
365 365
366 366 zil_bp_tree_fini(zilog);
367 367 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
368 368
369 369 return (error);
370 370 }
371 371
372 372 static int
373 373 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
374 374 {
375 375 /*
376 376 * Claim log block if not already committed and not already claimed.
377 377 * If tx == NULL, just verify that the block is claimable.
378 378 */
379 379 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
380 380 zil_bp_tree_add(zilog, bp) != 0)
381 381 return (0);
382 382
383 383 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
384 384 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
385 385 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
386 386 }
387 387
388 388 static int
389 389 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
390 390 {
391 391 lr_write_t *lr = (lr_write_t *)lrc;
392 392 int error;
393 393
394 394 if (lrc->lrc_txtype != TX_WRITE)
395 395 return (0);
396 396
397 397 /*
398 398 * If the block is not readable, don't claim it. This can happen
399 399 * in normal operation when a log block is written to disk before
400 400 * some of the dmu_sync() blocks it points to. In this case, the
401 401 * transaction cannot have been committed to anyone (we would have
402 402 * waited for all writes to be stable first), so it is semantically
403 403 * correct to declare this the end of the log.
404 404 */
405 405 if (lr->lr_blkptr.blk_birth >= first_txg &&
406 406 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
407 407 return (error);
408 408 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
409 409 }
410 410
411 411 /* ARGSUSED */
412 412 static int
413 413 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
414 414 {
415 415 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
416 416
417 417 return (0);
418 418 }
419 419
420 420 static int
421 421 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
422 422 {
423 423 lr_write_t *lr = (lr_write_t *)lrc;
424 424 blkptr_t *bp = &lr->lr_blkptr;
425 425
426 426 /*
427 427 * If we previously claimed it, we need to free it.
428 428 */
429 429 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
430 430 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
431 431 !BP_IS_HOLE(bp))
432 432 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
433 433
434 434 return (0);
435 435 }
436 436
437 437 static lwb_t *
438 438 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
439 439 {
440 440 lwb_t *lwb;
441 441
442 442 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
443 443 lwb->lwb_zilog = zilog;
444 444 lwb->lwb_blk = *bp;
445 445 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
446 446 lwb->lwb_max_txg = txg;
447 447 lwb->lwb_zio = NULL;
448 448 lwb->lwb_tx = NULL;
449 449 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
450 450 lwb->lwb_nused = sizeof (zil_chain_t);
451 451 lwb->lwb_sz = BP_GET_LSIZE(bp);
452 452 } else {
453 453 lwb->lwb_nused = 0;
454 454 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
455 455 }
456 456
457 457 mutex_enter(&zilog->zl_lock);
458 458 list_insert_tail(&zilog->zl_lwb_list, lwb);
459 459 mutex_exit(&zilog->zl_lock);
460 460
461 461 return (lwb);
462 462 }
463 463
464 464 /*
465 465 * Called when we create in-memory log transactions so that we know
466 466 * to cleanup the itxs at the end of spa_sync().
467 467 */
468 468 void
469 469 zilog_dirty(zilog_t *zilog, uint64_t txg)
470 470 {
471 471 dsl_pool_t *dp = zilog->zl_dmu_pool;
472 472 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
473 473
474 474 if (dsl_dataset_is_snapshot(ds))
475 475 panic("dirtying snapshot!");
476 476
477 477 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
478 478 /* up the hold count until we can be written out */
479 479 dmu_buf_add_ref(ds->ds_dbuf, zilog);
480 480 }
481 481 }
482 482
483 483 boolean_t
484 484 zilog_is_dirty(zilog_t *zilog)
485 485 {
486 486 dsl_pool_t *dp = zilog->zl_dmu_pool;
487 487
488 488 for (int t = 0; t < TXG_SIZE; t++) {
489 489 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
490 490 return (B_TRUE);
491 491 }
492 492 return (B_FALSE);
493 493 }
494 494
495 495 /*
496 496 * Create an on-disk intent log.
497 497 */
498 498 static lwb_t *
499 499 zil_create(zilog_t *zilog)
500 500 {
501 501 const zil_header_t *zh = zilog->zl_header;
502 502 lwb_t *lwb = NULL;
503 503 uint64_t txg = 0;
504 504 dmu_tx_t *tx = NULL;
505 505 blkptr_t blk;
506 506 int error = 0;
507 507
508 508 /*
509 509 * Wait for any previous destroy to complete.
510 510 */
511 511 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
512 512
513 513 ASSERT(zh->zh_claim_txg == 0);
514 514 ASSERT(zh->zh_replay_seq == 0);
515 515
516 516 blk = zh->zh_log;
517 517
518 518 /*
519 519 * Allocate an initial log block if:
520 520 * - there isn't one already
521 521 * - the existing block is the wrong endianess
522 522 */
523 523 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
524 524 tx = dmu_tx_create(zilog->zl_os);
525 525 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
526 526 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
527 527 txg = dmu_tx_get_txg(tx);
528 528
529 529 if (!BP_IS_HOLE(&blk)) {
530 530 zio_free_zil(zilog->zl_spa, txg, &blk);
531 531 BP_ZERO(&blk);
532 532 }
533 533
534 534 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
535 535 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
536 536
537 537 if (error == 0)
538 538 zil_init_log_chain(zilog, &blk);
539 539 }
540 540
541 541 /*
542 542 * Allocate a log write buffer (lwb) for the first log block.
543 543 */
544 544 if (error == 0)
545 545 lwb = zil_alloc_lwb(zilog, &blk, txg);
546 546
547 547 /*
548 548 * If we just allocated the first log block, commit our transaction
549 549 * and wait for zil_sync() to stuff the block poiner into zh_log.
550 550 * (zh is part of the MOS, so we cannot modify it in open context.)
551 551 */
552 552 if (tx != NULL) {
553 553 dmu_tx_commit(tx);
554 554 txg_wait_synced(zilog->zl_dmu_pool, txg);
555 555 }
556 556
557 557 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
558 558
559 559 return (lwb);
560 560 }
561 561
562 562 /*
563 563 * In one tx, free all log blocks and clear the log header.
564 564 * If keep_first is set, then we're replaying a log with no content.
565 565 * We want to keep the first block, however, so that the first
566 566 * synchronous transaction doesn't require a txg_wait_synced()
567 567 * in zil_create(). We don't need to txg_wait_synced() here either
568 568 * when keep_first is set, because both zil_create() and zil_destroy()
569 569 * will wait for any in-progress destroys to complete.
570 570 */
571 571 void
572 572 zil_destroy(zilog_t *zilog, boolean_t keep_first)
573 573 {
574 574 const zil_header_t *zh = zilog->zl_header;
575 575 lwb_t *lwb;
576 576 dmu_tx_t *tx;
577 577 uint64_t txg;
578 578
579 579 /*
580 580 * Wait for any previous destroy to complete.
581 581 */
582 582 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
583 583
584 584 zilog->zl_old_header = *zh; /* debugging aid */
585 585
586 586 if (BP_IS_HOLE(&zh->zh_log))
587 587 return;
588 588
589 589 tx = dmu_tx_create(zilog->zl_os);
590 590 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
591 591 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
592 592 txg = dmu_tx_get_txg(tx);
593 593
594 594 mutex_enter(&zilog->zl_lock);
595 595
596 596 ASSERT3U(zilog->zl_destroy_txg, <, txg);
597 597 zilog->zl_destroy_txg = txg;
598 598 zilog->zl_keep_first = keep_first;
599 599
600 600 if (!list_is_empty(&zilog->zl_lwb_list)) {
601 601 ASSERT(zh->zh_claim_txg == 0);
602 602 VERIFY(!keep_first);
603 603 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
604 604 list_remove(&zilog->zl_lwb_list, lwb);
605 605 if (lwb->lwb_buf != NULL)
606 606 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
607 607 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
608 608 kmem_cache_free(zil_lwb_cache, lwb);
609 609 }
610 610 } else if (!keep_first) {
611 611 zil_destroy_sync(zilog, tx);
612 612 }
613 613 mutex_exit(&zilog->zl_lock);
614 614
615 615 dmu_tx_commit(tx);
616 616 }
|
↓ open down ↓ |
616 lines elided |
↑ open up ↑ |
617 617
618 618 void
619 619 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
620 620 {
621 621 ASSERT(list_is_empty(&zilog->zl_lwb_list));
622 622 (void) zil_parse(zilog, zil_free_log_block,
623 623 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
624 624 }
625 625
626 626 int
627 -zil_claim(const char *osname, void *txarg)
627 +zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
628 628 {
629 629 dmu_tx_t *tx = txarg;
630 630 uint64_t first_txg = dmu_tx_get_txg(tx);
631 631 zilog_t *zilog;
632 632 zil_header_t *zh;
633 633 objset_t *os;
634 634 int error;
635 635
636 - error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
636 + error = dmu_objset_own_obj(dp, ds->ds_object,
637 + DMU_OST_ANY, B_FALSE, FTAG, &os);
637 638 if (error != 0) {
638 - cmn_err(CE_WARN, "can't open objset for %s", osname);
639 + cmn_err(CE_WARN, "can't open objset %llu, error %d",
640 + (unsigned long long)ds->ds_object, error);
639 641 return (0);
640 642 }
641 643
642 644 zilog = dmu_objset_zil(os);
643 645 zh = zil_header_in_syncing_context(zilog);
644 646
645 647 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
646 648 if (!BP_IS_HOLE(&zh->zh_log))
647 649 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
648 650 BP_ZERO(&zh->zh_log);
649 651 dsl_dataset_dirty(dmu_objset_ds(os), tx);
650 652 dmu_objset_disown(os, FTAG);
651 653 return (0);
652 654 }
653 655
654 656 /*
655 657 * Claim all log blocks if we haven't already done so, and remember
656 658 * the highest claimed sequence number. This ensures that if we can
657 659 * read only part of the log now (e.g. due to a missing device),
658 660 * but we can read the entire log later, we will not try to replay
659 661 * or destroy beyond the last block we successfully claimed.
660 662 */
661 663 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
662 664 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
663 665 (void) zil_parse(zilog, zil_claim_log_block,
664 666 zil_claim_log_record, tx, first_txg);
665 667 zh->zh_claim_txg = first_txg;
666 668 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
667 669 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
668 670 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
669 671 zh->zh_flags |= ZIL_REPLAY_NEEDED;
670 672 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
671 673 dsl_dataset_dirty(dmu_objset_ds(os), tx);
672 674 }
673 675
674 676 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
|
↓ open down ↓ |
26 lines elided |
↑ open up ↑ |
675 677 dmu_objset_disown(os, FTAG);
676 678 return (0);
677 679 }
678 680
679 681 /*
680 682 * Check the log by walking the log chain.
681 683 * Checksum errors are ok as they indicate the end of the chain.
682 684 * Any other error (no device or read failure) returns an error.
683 685 */
684 686 int
685 -zil_check_log_chain(const char *osname, void *tx)
687 +zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
686 688 {
687 689 zilog_t *zilog;
688 690 objset_t *os;
689 691 blkptr_t *bp;
690 692 int error;
691 693
692 694 ASSERT(tx == NULL);
693 695
694 - error = dmu_objset_hold(osname, FTAG, &os);
696 + error = dmu_objset_from_ds(ds, &os);
695 697 if (error != 0) {
696 - cmn_err(CE_WARN, "can't open objset for %s", osname);
698 + cmn_err(CE_WARN, "can't open objset %llu, error %d",
699 + (unsigned long long)ds->ds_object, error);
697 700 return (0);
698 701 }
699 702
700 703 zilog = dmu_objset_zil(os);
701 704 bp = (blkptr_t *)&zilog->zl_header->zh_log;
702 705
703 706 /*
704 707 * Check the first block and determine if it's on a log device
705 708 * which may have been removed or faulted prior to loading this
706 709 * pool. If so, there's no point in checking the rest of the log
707 710 * as its content should have already been synced to the pool.
708 711 */
|
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
709 712 if (!BP_IS_HOLE(bp)) {
710 713 vdev_t *vd;
711 714 boolean_t valid = B_TRUE;
712 715
713 716 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
714 717 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
715 718 if (vd->vdev_islog && vdev_is_dead(vd))
716 719 valid = vdev_log_state_valid(vd);
717 720 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
718 721
719 - if (!valid) {
720 - dmu_objset_rele(os, FTAG);
722 + if (!valid)
721 723 return (0);
722 - }
723 724 }
724 725
725 726 /*
726 727 * Because tx == NULL, zil_claim_log_block() will not actually claim
727 728 * any blocks, but just determine whether it is possible to do so.
728 729 * In addition to checking the log chain, zil_claim_log_block()
729 730 * will invoke zio_claim() with a done func of spa_claim_notify(),
730 731 * which will update spa_max_claim_txg. See spa_load() for details.
731 732 */
732 733 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
733 734 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
734 735
735 - dmu_objset_rele(os, FTAG);
736 -
737 736 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
738 737 }
739 738
740 739 static int
741 740 zil_vdev_compare(const void *x1, const void *x2)
742 741 {
743 742 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
744 743 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
745 744
746 745 if (v1 < v2)
747 746 return (-1);
748 747 if (v1 > v2)
749 748 return (1);
750 749
751 750 return (0);
752 751 }
753 752
754 753 void
755 754 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
756 755 {
757 756 avl_tree_t *t = &zilog->zl_vdev_tree;
758 757 avl_index_t where;
759 758 zil_vdev_node_t *zv, zvsearch;
760 759 int ndvas = BP_GET_NDVAS(bp);
761 760 int i;
762 761
763 762 if (zfs_nocacheflush)
764 763 return;
765 764
766 765 ASSERT(zilog->zl_writer);
767 766
768 767 /*
769 768 * Even though we're zl_writer, we still need a lock because the
770 769 * zl_get_data() callbacks may have dmu_sync() done callbacks
771 770 * that will run concurrently.
772 771 */
773 772 mutex_enter(&zilog->zl_vdev_lock);
774 773 for (i = 0; i < ndvas; i++) {
775 774 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
776 775 if (avl_find(t, &zvsearch, &where) == NULL) {
777 776 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
778 777 zv->zv_vdev = zvsearch.zv_vdev;
779 778 avl_insert(t, zv, where);
780 779 }
781 780 }
782 781 mutex_exit(&zilog->zl_vdev_lock);
783 782 }
784 783
785 784 static void
786 785 zil_flush_vdevs(zilog_t *zilog)
787 786 {
788 787 spa_t *spa = zilog->zl_spa;
789 788 avl_tree_t *t = &zilog->zl_vdev_tree;
790 789 void *cookie = NULL;
791 790 zil_vdev_node_t *zv;
792 791 zio_t *zio;
793 792
794 793 ASSERT(zilog->zl_writer);
795 794
796 795 /*
797 796 * We don't need zl_vdev_lock here because we're the zl_writer,
798 797 * and all zl_get_data() callbacks are done.
799 798 */
800 799 if (avl_numnodes(t) == 0)
801 800 return;
802 801
803 802 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
804 803
805 804 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
806 805
807 806 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
808 807 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
809 808 if (vd != NULL)
810 809 zio_flush(zio, vd);
811 810 kmem_free(zv, sizeof (*zv));
812 811 }
813 812
814 813 /*
815 814 * Wait for all the flushes to complete. Not all devices actually
816 815 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
817 816 */
818 817 (void) zio_wait(zio);
819 818
820 819 spa_config_exit(spa, SCL_STATE, FTAG);
821 820 }
822 821
823 822 /*
824 823 * Function called when a log block write completes
825 824 */
826 825 static void
827 826 zil_lwb_write_done(zio_t *zio)
828 827 {
829 828 lwb_t *lwb = zio->io_private;
830 829 zilog_t *zilog = lwb->lwb_zilog;
831 830 dmu_tx_t *tx = lwb->lwb_tx;
832 831
833 832 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
834 833 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
835 834 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
836 835 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
837 836 ASSERT(!BP_IS_GANG(zio->io_bp));
838 837 ASSERT(!BP_IS_HOLE(zio->io_bp));
839 838 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
840 839
841 840 /*
842 841 * Ensure the lwb buffer pointer is cleared before releasing
843 842 * the txg. If we have had an allocation failure and
844 843 * the txg is waiting to sync then we want want zil_sync()
845 844 * to remove the lwb so that it's not picked up as the next new
846 845 * one in zil_commit_writer(). zil_sync() will only remove
847 846 * the lwb if lwb_buf is null.
848 847 */
849 848 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
850 849 mutex_enter(&zilog->zl_lock);
851 850 lwb->lwb_buf = NULL;
852 851 lwb->lwb_tx = NULL;
853 852 mutex_exit(&zilog->zl_lock);
854 853
855 854 /*
856 855 * Now that we've written this log block, we have a stable pointer
857 856 * to the next block in the chain, so it's OK to let the txg in
858 857 * which we allocated the next block sync.
859 858 */
860 859 dmu_tx_commit(tx);
861 860 }
862 861
863 862 /*
864 863 * Initialize the io for a log block.
865 864 */
866 865 static void
867 866 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
868 867 {
869 868 zbookmark_t zb;
870 869
871 870 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
872 871 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
873 872 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
874 873
875 874 if (zilog->zl_root_zio == NULL) {
876 875 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
877 876 ZIO_FLAG_CANFAIL);
878 877 }
879 878 if (lwb->lwb_zio == NULL) {
880 879 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
881 880 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
882 881 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
883 882 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
884 883 }
885 884 }
886 885
887 886 /*
888 887 * Define a limited set of intent log block sizes.
889 888 *
890 889 * These must be a multiple of 4KB. Note only the amount used (again
891 890 * aligned to 4KB) actually gets written. However, we can't always just
892 891 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
893 892 */
894 893 uint64_t zil_block_buckets[] = {
895 894 4096, /* non TX_WRITE */
896 895 8192+4096, /* data base */
897 896 32*1024 + 4096, /* NFS writes */
898 897 UINT64_MAX
899 898 };
900 899
901 900 /*
902 901 * Use the slog as long as the logbias is 'latency' and the current commit size
903 902 * is less than the limit or the total list size is less than 2X the limit.
904 903 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
905 904 */
906 905 uint64_t zil_slog_limit = 1024 * 1024;
907 906 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
908 907 (((zilog)->zl_cur_used < zil_slog_limit) || \
909 908 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
910 909
911 910 /*
912 911 * Start a log block write and advance to the next log block.
913 912 * Calls are serialized.
914 913 */
915 914 static lwb_t *
916 915 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
917 916 {
918 917 lwb_t *nlwb = NULL;
919 918 zil_chain_t *zilc;
920 919 spa_t *spa = zilog->zl_spa;
921 920 blkptr_t *bp;
922 921 dmu_tx_t *tx;
923 922 uint64_t txg;
924 923 uint64_t zil_blksz, wsz;
925 924 int i, error;
926 925
927 926 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
928 927 zilc = (zil_chain_t *)lwb->lwb_buf;
929 928 bp = &zilc->zc_next_blk;
930 929 } else {
931 930 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
932 931 bp = &zilc->zc_next_blk;
933 932 }
934 933
935 934 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
936 935
937 936 /*
938 937 * Allocate the next block and save its address in this block
939 938 * before writing it in order to establish the log chain.
940 939 * Note that if the allocation of nlwb synced before we wrote
941 940 * the block that points at it (lwb), we'd leak it if we crashed.
942 941 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
943 942 * We dirty the dataset to ensure that zil_sync() will be called
944 943 * to clean up in the event of allocation failure or I/O failure.
945 944 */
946 945 tx = dmu_tx_create(zilog->zl_os);
947 946 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
948 947 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
949 948 txg = dmu_tx_get_txg(tx);
950 949
951 950 lwb->lwb_tx = tx;
952 951
953 952 /*
954 953 * Log blocks are pre-allocated. Here we select the size of the next
955 954 * block, based on size used in the last block.
956 955 * - first find the smallest bucket that will fit the block from a
957 956 * limited set of block sizes. This is because it's faster to write
958 957 * blocks allocated from the same metaslab as they are adjacent or
959 958 * close.
960 959 * - next find the maximum from the new suggested size and an array of
961 960 * previous sizes. This lessens a picket fence effect of wrongly
962 961 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
963 962 * requests.
964 963 *
965 964 * Note we only write what is used, but we can't just allocate
966 965 * the maximum block size because we can exhaust the available
967 966 * pool log space.
968 967 */
969 968 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
970 969 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
971 970 continue;
972 971 zil_blksz = zil_block_buckets[i];
973 972 if (zil_blksz == UINT64_MAX)
974 973 zil_blksz = SPA_MAXBLOCKSIZE;
975 974 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
976 975 for (i = 0; i < ZIL_PREV_BLKS; i++)
977 976 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
978 977 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
979 978
980 979 BP_ZERO(bp);
981 980 /* pass the old blkptr in order to spread log blocks across devs */
982 981 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
983 982 USE_SLOG(zilog));
984 983 if (error == 0) {
985 984 ASSERT3U(bp->blk_birth, ==, txg);
986 985 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
987 986 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
988 987
989 988 /*
990 989 * Allocate a new log write buffer (lwb).
991 990 */
992 991 nlwb = zil_alloc_lwb(zilog, bp, txg);
993 992
994 993 /* Record the block for later vdev flushing */
995 994 zil_add_block(zilog, &lwb->lwb_blk);
996 995 }
997 996
998 997 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
999 998 /* For Slim ZIL only write what is used. */
1000 999 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1001 1000 ASSERT3U(wsz, <=, lwb->lwb_sz);
1002 1001 zio_shrink(lwb->lwb_zio, wsz);
1003 1002
1004 1003 } else {
1005 1004 wsz = lwb->lwb_sz;
1006 1005 }
1007 1006
1008 1007 zilc->zc_pad = 0;
1009 1008 zilc->zc_nused = lwb->lwb_nused;
1010 1009 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1011 1010
1012 1011 /*
1013 1012 * clear unused data for security
1014 1013 */
1015 1014 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1016 1015
1017 1016 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1018 1017
1019 1018 /*
1020 1019 * If there was an allocation failure then nlwb will be null which
1021 1020 * forces a txg_wait_synced().
1022 1021 */
1023 1022 return (nlwb);
1024 1023 }
1025 1024
1026 1025 static lwb_t *
1027 1026 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1028 1027 {
1029 1028 lr_t *lrc = &itx->itx_lr; /* common log record */
1030 1029 lr_write_t *lrw = (lr_write_t *)lrc;
1031 1030 char *lr_buf;
1032 1031 uint64_t txg = lrc->lrc_txg;
1033 1032 uint64_t reclen = lrc->lrc_reclen;
1034 1033 uint64_t dlen = 0;
1035 1034
1036 1035 if (lwb == NULL)
1037 1036 return (NULL);
1038 1037
1039 1038 ASSERT(lwb->lwb_buf != NULL);
1040 1039 ASSERT(zilog_is_dirty(zilog) ||
1041 1040 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1042 1041
1043 1042 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1044 1043 dlen = P2ROUNDUP_TYPED(
1045 1044 lrw->lr_length, sizeof (uint64_t), uint64_t);
1046 1045
1047 1046 zilog->zl_cur_used += (reclen + dlen);
1048 1047
1049 1048 zil_lwb_write_init(zilog, lwb);
1050 1049
1051 1050 /*
1052 1051 * If this record won't fit in the current log block, start a new one.
1053 1052 */
1054 1053 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1055 1054 lwb = zil_lwb_write_start(zilog, lwb);
1056 1055 if (lwb == NULL)
1057 1056 return (NULL);
1058 1057 zil_lwb_write_init(zilog, lwb);
1059 1058 ASSERT(LWB_EMPTY(lwb));
1060 1059 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1061 1060 txg_wait_synced(zilog->zl_dmu_pool, txg);
1062 1061 return (lwb);
1063 1062 }
1064 1063 }
1065 1064
1066 1065 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1067 1066 bcopy(lrc, lr_buf, reclen);
1068 1067 lrc = (lr_t *)lr_buf;
1069 1068 lrw = (lr_write_t *)lrc;
1070 1069
1071 1070 /*
1072 1071 * If it's a write, fetch the data or get its blkptr as appropriate.
1073 1072 */
1074 1073 if (lrc->lrc_txtype == TX_WRITE) {
1075 1074 if (txg > spa_freeze_txg(zilog->zl_spa))
1076 1075 txg_wait_synced(zilog->zl_dmu_pool, txg);
1077 1076 if (itx->itx_wr_state != WR_COPIED) {
1078 1077 char *dbuf;
1079 1078 int error;
1080 1079
1081 1080 if (dlen) {
1082 1081 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1083 1082 dbuf = lr_buf + reclen;
1084 1083 lrw->lr_common.lrc_reclen += dlen;
1085 1084 } else {
1086 1085 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1087 1086 dbuf = NULL;
1088 1087 }
1089 1088 error = zilog->zl_get_data(
1090 1089 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1091 1090 if (error == EIO) {
1092 1091 txg_wait_synced(zilog->zl_dmu_pool, txg);
1093 1092 return (lwb);
1094 1093 }
1095 1094 if (error != 0) {
1096 1095 ASSERT(error == ENOENT || error == EEXIST ||
1097 1096 error == EALREADY);
1098 1097 return (lwb);
1099 1098 }
1100 1099 }
1101 1100 }
1102 1101
1103 1102 /*
1104 1103 * We're actually making an entry, so update lrc_seq to be the
1105 1104 * log record sequence number. Note that this is generally not
1106 1105 * equal to the itx sequence number because not all transactions
1107 1106 * are synchronous, and sometimes spa_sync() gets there first.
1108 1107 */
1109 1108 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1110 1109 lwb->lwb_nused += reclen + dlen;
1111 1110 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1112 1111 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1113 1112 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1114 1113
1115 1114 return (lwb);
1116 1115 }
1117 1116
1118 1117 itx_t *
1119 1118 zil_itx_create(uint64_t txtype, size_t lrsize)
1120 1119 {
1121 1120 itx_t *itx;
1122 1121
1123 1122 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1124 1123
1125 1124 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1126 1125 itx->itx_lr.lrc_txtype = txtype;
1127 1126 itx->itx_lr.lrc_reclen = lrsize;
1128 1127 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1129 1128 itx->itx_lr.lrc_seq = 0; /* defensive */
1130 1129 itx->itx_sync = B_TRUE; /* default is synchronous */
1131 1130
1132 1131 return (itx);
1133 1132 }
1134 1133
1135 1134 void
1136 1135 zil_itx_destroy(itx_t *itx)
1137 1136 {
1138 1137 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1139 1138 }
1140 1139
1141 1140 /*
1142 1141 * Free up the sync and async itxs. The itxs_t has already been detached
1143 1142 * so no locks are needed.
1144 1143 */
1145 1144 static void
1146 1145 zil_itxg_clean(itxs_t *itxs)
1147 1146 {
1148 1147 itx_t *itx;
1149 1148 list_t *list;
1150 1149 avl_tree_t *t;
1151 1150 void *cookie;
1152 1151 itx_async_node_t *ian;
1153 1152
1154 1153 list = &itxs->i_sync_list;
1155 1154 while ((itx = list_head(list)) != NULL) {
1156 1155 list_remove(list, itx);
1157 1156 kmem_free(itx, offsetof(itx_t, itx_lr) +
1158 1157 itx->itx_lr.lrc_reclen);
1159 1158 }
1160 1159
1161 1160 cookie = NULL;
1162 1161 t = &itxs->i_async_tree;
1163 1162 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1164 1163 list = &ian->ia_list;
1165 1164 while ((itx = list_head(list)) != NULL) {
1166 1165 list_remove(list, itx);
1167 1166 kmem_free(itx, offsetof(itx_t, itx_lr) +
1168 1167 itx->itx_lr.lrc_reclen);
1169 1168 }
1170 1169 list_destroy(list);
1171 1170 kmem_free(ian, sizeof (itx_async_node_t));
1172 1171 }
1173 1172 avl_destroy(t);
1174 1173
1175 1174 kmem_free(itxs, sizeof (itxs_t));
1176 1175 }
1177 1176
1178 1177 static int
1179 1178 zil_aitx_compare(const void *x1, const void *x2)
1180 1179 {
1181 1180 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1182 1181 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1183 1182
1184 1183 if (o1 < o2)
1185 1184 return (-1);
1186 1185 if (o1 > o2)
1187 1186 return (1);
1188 1187
1189 1188 return (0);
1190 1189 }
1191 1190
1192 1191 /*
1193 1192 * Remove all async itx with the given oid.
1194 1193 */
1195 1194 static void
1196 1195 zil_remove_async(zilog_t *zilog, uint64_t oid)
1197 1196 {
1198 1197 uint64_t otxg, txg;
1199 1198 itx_async_node_t *ian;
1200 1199 avl_tree_t *t;
1201 1200 avl_index_t where;
1202 1201 list_t clean_list;
1203 1202 itx_t *itx;
1204 1203
1205 1204 ASSERT(oid != 0);
1206 1205 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1207 1206
1208 1207 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1209 1208 otxg = ZILTEST_TXG;
1210 1209 else
1211 1210 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1212 1211
1213 1212 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1214 1213 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1215 1214
1216 1215 mutex_enter(&itxg->itxg_lock);
1217 1216 if (itxg->itxg_txg != txg) {
1218 1217 mutex_exit(&itxg->itxg_lock);
1219 1218 continue;
1220 1219 }
1221 1220
1222 1221 /*
1223 1222 * Locate the object node and append its list.
1224 1223 */
1225 1224 t = &itxg->itxg_itxs->i_async_tree;
1226 1225 ian = avl_find(t, &oid, &where);
1227 1226 if (ian != NULL)
1228 1227 list_move_tail(&clean_list, &ian->ia_list);
1229 1228 mutex_exit(&itxg->itxg_lock);
1230 1229 }
1231 1230 while ((itx = list_head(&clean_list)) != NULL) {
1232 1231 list_remove(&clean_list, itx);
1233 1232 kmem_free(itx, offsetof(itx_t, itx_lr) +
1234 1233 itx->itx_lr.lrc_reclen);
1235 1234 }
1236 1235 list_destroy(&clean_list);
1237 1236 }
1238 1237
1239 1238 void
1240 1239 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1241 1240 {
1242 1241 uint64_t txg;
1243 1242 itxg_t *itxg;
1244 1243 itxs_t *itxs, *clean = NULL;
1245 1244
1246 1245 /*
1247 1246 * Object ids can be re-instantiated in the next txg so
1248 1247 * remove any async transactions to avoid future leaks.
1249 1248 * This can happen if a fsync occurs on the re-instantiated
1250 1249 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1251 1250 * the new file data and flushes a write record for the old object.
1252 1251 */
1253 1252 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1254 1253 zil_remove_async(zilog, itx->itx_oid);
1255 1254
1256 1255 /*
1257 1256 * Ensure the data of a renamed file is committed before the rename.
1258 1257 */
1259 1258 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1260 1259 zil_async_to_sync(zilog, itx->itx_oid);
1261 1260
1262 1261 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1263 1262 txg = ZILTEST_TXG;
1264 1263 else
1265 1264 txg = dmu_tx_get_txg(tx);
1266 1265
1267 1266 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1268 1267 mutex_enter(&itxg->itxg_lock);
1269 1268 itxs = itxg->itxg_itxs;
1270 1269 if (itxg->itxg_txg != txg) {
1271 1270 if (itxs != NULL) {
1272 1271 /*
1273 1272 * The zil_clean callback hasn't got around to cleaning
1274 1273 * this itxg. Save the itxs for release below.
1275 1274 * This should be rare.
1276 1275 */
1277 1276 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1278 1277 itxg->itxg_sod = 0;
1279 1278 clean = itxg->itxg_itxs;
1280 1279 }
1281 1280 ASSERT(itxg->itxg_sod == 0);
1282 1281 itxg->itxg_txg = txg;
1283 1282 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1284 1283
1285 1284 list_create(&itxs->i_sync_list, sizeof (itx_t),
1286 1285 offsetof(itx_t, itx_node));
1287 1286 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1288 1287 sizeof (itx_async_node_t),
1289 1288 offsetof(itx_async_node_t, ia_node));
1290 1289 }
1291 1290 if (itx->itx_sync) {
1292 1291 list_insert_tail(&itxs->i_sync_list, itx);
1293 1292 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1294 1293 itxg->itxg_sod += itx->itx_sod;
1295 1294 } else {
1296 1295 avl_tree_t *t = &itxs->i_async_tree;
1297 1296 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1298 1297 itx_async_node_t *ian;
1299 1298 avl_index_t where;
1300 1299
1301 1300 ian = avl_find(t, &foid, &where);
1302 1301 if (ian == NULL) {
1303 1302 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1304 1303 list_create(&ian->ia_list, sizeof (itx_t),
1305 1304 offsetof(itx_t, itx_node));
1306 1305 ian->ia_foid = foid;
1307 1306 avl_insert(t, ian, where);
1308 1307 }
1309 1308 list_insert_tail(&ian->ia_list, itx);
1310 1309 }
1311 1310
1312 1311 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1313 1312 zilog_dirty(zilog, txg);
1314 1313 mutex_exit(&itxg->itxg_lock);
1315 1314
1316 1315 /* Release the old itxs now we've dropped the lock */
1317 1316 if (clean != NULL)
1318 1317 zil_itxg_clean(clean);
1319 1318 }
1320 1319
1321 1320 /*
1322 1321 * If there are any in-memory intent log transactions which have now been
1323 1322 * synced then start up a taskq to free them. We should only do this after we
1324 1323 * have written out the uberblocks (i.e. txg has been comitted) so that
1325 1324 * don't inadvertently clean out in-memory log records that would be required
1326 1325 * by zil_commit().
1327 1326 */
1328 1327 void
1329 1328 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1330 1329 {
1331 1330 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1332 1331 itxs_t *clean_me;
1333 1332
1334 1333 mutex_enter(&itxg->itxg_lock);
1335 1334 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1336 1335 mutex_exit(&itxg->itxg_lock);
1337 1336 return;
1338 1337 }
1339 1338 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1340 1339 ASSERT(itxg->itxg_txg != 0);
1341 1340 ASSERT(zilog->zl_clean_taskq != NULL);
1342 1341 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1343 1342 itxg->itxg_sod = 0;
1344 1343 clean_me = itxg->itxg_itxs;
1345 1344 itxg->itxg_itxs = NULL;
1346 1345 itxg->itxg_txg = 0;
1347 1346 mutex_exit(&itxg->itxg_lock);
1348 1347 /*
1349 1348 * Preferably start a task queue to free up the old itxs but
1350 1349 * if taskq_dispatch can't allocate resources to do that then
1351 1350 * free it in-line. This should be rare. Note, using TQ_SLEEP
1352 1351 * created a bad performance problem.
1353 1352 */
1354 1353 if (taskq_dispatch(zilog->zl_clean_taskq,
1355 1354 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1356 1355 zil_itxg_clean(clean_me);
1357 1356 }
1358 1357
1359 1358 /*
1360 1359 * Get the list of itxs to commit into zl_itx_commit_list.
1361 1360 */
1362 1361 static void
1363 1362 zil_get_commit_list(zilog_t *zilog)
1364 1363 {
1365 1364 uint64_t otxg, txg;
1366 1365 list_t *commit_list = &zilog->zl_itx_commit_list;
1367 1366 uint64_t push_sod = 0;
1368 1367
1369 1368 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1370 1369 otxg = ZILTEST_TXG;
1371 1370 else
1372 1371 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1373 1372
1374 1373 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1375 1374 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1376 1375
1377 1376 mutex_enter(&itxg->itxg_lock);
1378 1377 if (itxg->itxg_txg != txg) {
1379 1378 mutex_exit(&itxg->itxg_lock);
1380 1379 continue;
1381 1380 }
1382 1381
1383 1382 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1384 1383 push_sod += itxg->itxg_sod;
1385 1384 itxg->itxg_sod = 0;
1386 1385
1387 1386 mutex_exit(&itxg->itxg_lock);
1388 1387 }
1389 1388 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1390 1389 }
1391 1390
1392 1391 /*
1393 1392 * Move the async itxs for a specified object to commit into sync lists.
1394 1393 */
1395 1394 static void
1396 1395 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1397 1396 {
1398 1397 uint64_t otxg, txg;
1399 1398 itx_async_node_t *ian;
1400 1399 avl_tree_t *t;
1401 1400 avl_index_t where;
1402 1401
1403 1402 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1404 1403 otxg = ZILTEST_TXG;
1405 1404 else
1406 1405 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1407 1406
1408 1407 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1409 1408 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1410 1409
1411 1410 mutex_enter(&itxg->itxg_lock);
1412 1411 if (itxg->itxg_txg != txg) {
1413 1412 mutex_exit(&itxg->itxg_lock);
1414 1413 continue;
1415 1414 }
1416 1415
1417 1416 /*
1418 1417 * If a foid is specified then find that node and append its
1419 1418 * list. Otherwise walk the tree appending all the lists
1420 1419 * to the sync list. We add to the end rather than the
1421 1420 * beginning to ensure the create has happened.
1422 1421 */
1423 1422 t = &itxg->itxg_itxs->i_async_tree;
1424 1423 if (foid != 0) {
1425 1424 ian = avl_find(t, &foid, &where);
1426 1425 if (ian != NULL) {
1427 1426 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1428 1427 &ian->ia_list);
1429 1428 }
1430 1429 } else {
1431 1430 void *cookie = NULL;
1432 1431
1433 1432 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1434 1433 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1435 1434 &ian->ia_list);
1436 1435 list_destroy(&ian->ia_list);
1437 1436 kmem_free(ian, sizeof (itx_async_node_t));
1438 1437 }
1439 1438 }
1440 1439 mutex_exit(&itxg->itxg_lock);
1441 1440 }
1442 1441 }
1443 1442
1444 1443 static void
1445 1444 zil_commit_writer(zilog_t *zilog)
1446 1445 {
1447 1446 uint64_t txg;
1448 1447 itx_t *itx;
1449 1448 lwb_t *lwb;
1450 1449 spa_t *spa = zilog->zl_spa;
1451 1450 int error = 0;
1452 1451
1453 1452 ASSERT(zilog->zl_root_zio == NULL);
1454 1453
1455 1454 mutex_exit(&zilog->zl_lock);
1456 1455
1457 1456 zil_get_commit_list(zilog);
1458 1457
1459 1458 /*
1460 1459 * Return if there's nothing to commit before we dirty the fs by
1461 1460 * calling zil_create().
1462 1461 */
1463 1462 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1464 1463 mutex_enter(&zilog->zl_lock);
1465 1464 return;
1466 1465 }
1467 1466
1468 1467 if (zilog->zl_suspend) {
1469 1468 lwb = NULL;
1470 1469 } else {
1471 1470 lwb = list_tail(&zilog->zl_lwb_list);
1472 1471 if (lwb == NULL)
1473 1472 lwb = zil_create(zilog);
1474 1473 }
1475 1474
1476 1475 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1477 1476 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1478 1477 txg = itx->itx_lr.lrc_txg;
1479 1478 ASSERT(txg);
1480 1479
1481 1480 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1482 1481 lwb = zil_lwb_commit(zilog, itx, lwb);
1483 1482 list_remove(&zilog->zl_itx_commit_list, itx);
1484 1483 kmem_free(itx, offsetof(itx_t, itx_lr)
1485 1484 + itx->itx_lr.lrc_reclen);
1486 1485 }
1487 1486 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1488 1487
1489 1488 /* write the last block out */
1490 1489 if (lwb != NULL && lwb->lwb_zio != NULL)
1491 1490 lwb = zil_lwb_write_start(zilog, lwb);
1492 1491
1493 1492 zilog->zl_cur_used = 0;
1494 1493
1495 1494 /*
1496 1495 * Wait if necessary for the log blocks to be on stable storage.
1497 1496 */
1498 1497 if (zilog->zl_root_zio) {
1499 1498 error = zio_wait(zilog->zl_root_zio);
1500 1499 zilog->zl_root_zio = NULL;
1501 1500 zil_flush_vdevs(zilog);
1502 1501 }
1503 1502
1504 1503 if (error || lwb == NULL)
1505 1504 txg_wait_synced(zilog->zl_dmu_pool, 0);
1506 1505
1507 1506 mutex_enter(&zilog->zl_lock);
1508 1507
1509 1508 /*
1510 1509 * Remember the highest committed log sequence number for ztest.
1511 1510 * We only update this value when all the log writes succeeded,
1512 1511 * because ztest wants to ASSERT that it got the whole log chain.
1513 1512 */
1514 1513 if (error == 0 && lwb != NULL)
1515 1514 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1516 1515 }
1517 1516
1518 1517 /*
1519 1518 * Commit zfs transactions to stable storage.
1520 1519 * If foid is 0 push out all transactions, otherwise push only those
1521 1520 * for that object or might reference that object.
1522 1521 *
1523 1522 * itxs are committed in batches. In a heavily stressed zil there will be
1524 1523 * a commit writer thread who is writing out a bunch of itxs to the log
1525 1524 * for a set of committing threads (cthreads) in the same batch as the writer.
1526 1525 * Those cthreads are all waiting on the same cv for that batch.
1527 1526 *
1528 1527 * There will also be a different and growing batch of threads that are
1529 1528 * waiting to commit (qthreads). When the committing batch completes
1530 1529 * a transition occurs such that the cthreads exit and the qthreads become
1531 1530 * cthreads. One of the new cthreads becomes the writer thread for the
1532 1531 * batch. Any new threads arriving become new qthreads.
1533 1532 *
1534 1533 * Only 2 condition variables are needed and there's no transition
1535 1534 * between the two cvs needed. They just flip-flop between qthreads
1536 1535 * and cthreads.
1537 1536 *
1538 1537 * Using this scheme we can efficiently wakeup up only those threads
1539 1538 * that have been committed.
1540 1539 */
1541 1540 void
1542 1541 zil_commit(zilog_t *zilog, uint64_t foid)
1543 1542 {
1544 1543 uint64_t mybatch;
1545 1544
1546 1545 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1547 1546 return;
1548 1547
1549 1548 /* move the async itxs for the foid to the sync queues */
1550 1549 zil_async_to_sync(zilog, foid);
1551 1550
1552 1551 mutex_enter(&zilog->zl_lock);
1553 1552 mybatch = zilog->zl_next_batch;
1554 1553 while (zilog->zl_writer) {
1555 1554 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1556 1555 if (mybatch <= zilog->zl_com_batch) {
1557 1556 mutex_exit(&zilog->zl_lock);
1558 1557 return;
1559 1558 }
1560 1559 }
1561 1560
1562 1561 zilog->zl_next_batch++;
1563 1562 zilog->zl_writer = B_TRUE;
1564 1563 zil_commit_writer(zilog);
1565 1564 zilog->zl_com_batch = mybatch;
1566 1565 zilog->zl_writer = B_FALSE;
1567 1566 mutex_exit(&zilog->zl_lock);
1568 1567
1569 1568 /* wake up one thread to become the next writer */
1570 1569 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1571 1570
1572 1571 /* wake up all threads waiting for this batch to be committed */
1573 1572 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1574 1573 }
1575 1574
1576 1575 /*
1577 1576 * Called in syncing context to free committed log blocks and update log header.
1578 1577 */
1579 1578 void
1580 1579 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1581 1580 {
1582 1581 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1583 1582 uint64_t txg = dmu_tx_get_txg(tx);
1584 1583 spa_t *spa = zilog->zl_spa;
1585 1584 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1586 1585 lwb_t *lwb;
1587 1586
1588 1587 /*
1589 1588 * We don't zero out zl_destroy_txg, so make sure we don't try
1590 1589 * to destroy it twice.
1591 1590 */
1592 1591 if (spa_sync_pass(spa) != 1)
1593 1592 return;
1594 1593
1595 1594 mutex_enter(&zilog->zl_lock);
1596 1595
1597 1596 ASSERT(zilog->zl_stop_sync == 0);
1598 1597
1599 1598 if (*replayed_seq != 0) {
1600 1599 ASSERT(zh->zh_replay_seq < *replayed_seq);
1601 1600 zh->zh_replay_seq = *replayed_seq;
1602 1601 *replayed_seq = 0;
1603 1602 }
1604 1603
1605 1604 if (zilog->zl_destroy_txg == txg) {
1606 1605 blkptr_t blk = zh->zh_log;
1607 1606
1608 1607 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1609 1608
1610 1609 bzero(zh, sizeof (zil_header_t));
1611 1610 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1612 1611
1613 1612 if (zilog->zl_keep_first) {
1614 1613 /*
1615 1614 * If this block was part of log chain that couldn't
1616 1615 * be claimed because a device was missing during
1617 1616 * zil_claim(), but that device later returns,
1618 1617 * then this block could erroneously appear valid.
1619 1618 * To guard against this, assign a new GUID to the new
1620 1619 * log chain so it doesn't matter what blk points to.
1621 1620 */
1622 1621 zil_init_log_chain(zilog, &blk);
1623 1622 zh->zh_log = blk;
1624 1623 }
1625 1624 }
1626 1625
1627 1626 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1628 1627 zh->zh_log = lwb->lwb_blk;
1629 1628 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1630 1629 break;
1631 1630 list_remove(&zilog->zl_lwb_list, lwb);
1632 1631 zio_free_zil(spa, txg, &lwb->lwb_blk);
1633 1632 kmem_cache_free(zil_lwb_cache, lwb);
1634 1633
1635 1634 /*
1636 1635 * If we don't have anything left in the lwb list then
1637 1636 * we've had an allocation failure and we need to zero
1638 1637 * out the zil_header blkptr so that we don't end
1639 1638 * up freeing the same block twice.
1640 1639 */
1641 1640 if (list_head(&zilog->zl_lwb_list) == NULL)
1642 1641 BP_ZERO(&zh->zh_log);
1643 1642 }
1644 1643 mutex_exit(&zilog->zl_lock);
1645 1644 }
1646 1645
1647 1646 void
1648 1647 zil_init(void)
1649 1648 {
1650 1649 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1651 1650 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1652 1651 }
1653 1652
1654 1653 void
1655 1654 zil_fini(void)
1656 1655 {
1657 1656 kmem_cache_destroy(zil_lwb_cache);
1658 1657 }
1659 1658
1660 1659 void
1661 1660 zil_set_sync(zilog_t *zilog, uint64_t sync)
1662 1661 {
1663 1662 zilog->zl_sync = sync;
1664 1663 }
1665 1664
1666 1665 void
1667 1666 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1668 1667 {
1669 1668 zilog->zl_logbias = logbias;
1670 1669 }
1671 1670
1672 1671 zilog_t *
1673 1672 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1674 1673 {
1675 1674 zilog_t *zilog;
1676 1675
1677 1676 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1678 1677
1679 1678 zilog->zl_header = zh_phys;
1680 1679 zilog->zl_os = os;
1681 1680 zilog->zl_spa = dmu_objset_spa(os);
1682 1681 zilog->zl_dmu_pool = dmu_objset_pool(os);
1683 1682 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1684 1683 zilog->zl_logbias = dmu_objset_logbias(os);
1685 1684 zilog->zl_sync = dmu_objset_syncprop(os);
1686 1685 zilog->zl_next_batch = 1;
1687 1686
1688 1687 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1689 1688
1690 1689 for (int i = 0; i < TXG_SIZE; i++) {
1691 1690 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1692 1691 MUTEX_DEFAULT, NULL);
1693 1692 }
1694 1693
1695 1694 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1696 1695 offsetof(lwb_t, lwb_node));
1697 1696
1698 1697 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1699 1698 offsetof(itx_t, itx_node));
1700 1699
1701 1700 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1702 1701
1703 1702 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1704 1703 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1705 1704
1706 1705 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1707 1706 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1708 1707 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1709 1708 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1710 1709
1711 1710 return (zilog);
1712 1711 }
1713 1712
1714 1713 void
1715 1714 zil_free(zilog_t *zilog)
1716 1715 {
1717 1716 zilog->zl_stop_sync = 1;
1718 1717
1719 1718 ASSERT0(zilog->zl_suspend);
1720 1719 ASSERT0(zilog->zl_suspending);
1721 1720
1722 1721 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1723 1722 list_destroy(&zilog->zl_lwb_list);
1724 1723
1725 1724 avl_destroy(&zilog->zl_vdev_tree);
1726 1725 mutex_destroy(&zilog->zl_vdev_lock);
1727 1726
1728 1727 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1729 1728 list_destroy(&zilog->zl_itx_commit_list);
1730 1729
1731 1730 for (int i = 0; i < TXG_SIZE; i++) {
1732 1731 /*
1733 1732 * It's possible for an itx to be generated that doesn't dirty
1734 1733 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1735 1734 * callback to remove the entry. We remove those here.
1736 1735 *
1737 1736 * Also free up the ziltest itxs.
1738 1737 */
1739 1738 if (zilog->zl_itxg[i].itxg_itxs)
1740 1739 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1741 1740 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1742 1741 }
1743 1742
1744 1743 mutex_destroy(&zilog->zl_lock);
1745 1744
1746 1745 cv_destroy(&zilog->zl_cv_writer);
1747 1746 cv_destroy(&zilog->zl_cv_suspend);
1748 1747 cv_destroy(&zilog->zl_cv_batch[0]);
1749 1748 cv_destroy(&zilog->zl_cv_batch[1]);
1750 1749
1751 1750 kmem_free(zilog, sizeof (zilog_t));
1752 1751 }
1753 1752
1754 1753 /*
1755 1754 * Open an intent log.
1756 1755 */
1757 1756 zilog_t *
1758 1757 zil_open(objset_t *os, zil_get_data_t *get_data)
1759 1758 {
1760 1759 zilog_t *zilog = dmu_objset_zil(os);
1761 1760
1762 1761 ASSERT(zilog->zl_clean_taskq == NULL);
1763 1762 ASSERT(zilog->zl_get_data == NULL);
1764 1763 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1765 1764
1766 1765 zilog->zl_get_data = get_data;
1767 1766 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1768 1767 2, 2, TASKQ_PREPOPULATE);
1769 1768
1770 1769 return (zilog);
1771 1770 }
1772 1771
1773 1772 /*
1774 1773 * Close an intent log.
1775 1774 */
1776 1775 void
1777 1776 zil_close(zilog_t *zilog)
1778 1777 {
1779 1778 lwb_t *lwb;
1780 1779 uint64_t txg = 0;
1781 1780
1782 1781 zil_commit(zilog, 0); /* commit all itx */
1783 1782
1784 1783 /*
1785 1784 * The lwb_max_txg for the stubby lwb will reflect the last activity
1786 1785 * for the zil. After a txg_wait_synced() on the txg we know all the
1787 1786 * callbacks have occurred that may clean the zil. Only then can we
1788 1787 * destroy the zl_clean_taskq.
1789 1788 */
1790 1789 mutex_enter(&zilog->zl_lock);
1791 1790 lwb = list_tail(&zilog->zl_lwb_list);
1792 1791 if (lwb != NULL)
1793 1792 txg = lwb->lwb_max_txg;
1794 1793 mutex_exit(&zilog->zl_lock);
1795 1794 if (txg)
1796 1795 txg_wait_synced(zilog->zl_dmu_pool, txg);
1797 1796 ASSERT(!zilog_is_dirty(zilog));
1798 1797
1799 1798 taskq_destroy(zilog->zl_clean_taskq);
1800 1799 zilog->zl_clean_taskq = NULL;
1801 1800 zilog->zl_get_data = NULL;
1802 1801
1803 1802 /*
1804 1803 * We should have only one LWB left on the list; remove it now.
1805 1804 */
1806 1805 mutex_enter(&zilog->zl_lock);
1807 1806 lwb = list_head(&zilog->zl_lwb_list);
1808 1807 if (lwb != NULL) {
1809 1808 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1810 1809 list_remove(&zilog->zl_lwb_list, lwb);
1811 1810 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1812 1811 kmem_cache_free(zil_lwb_cache, lwb);
1813 1812 }
1814 1813 mutex_exit(&zilog->zl_lock);
1815 1814 }
1816 1815
1817 1816 static char *suspend_tag = "zil suspending";
1818 1817
1819 1818 /*
1820 1819 * Suspend an intent log. While in suspended mode, we still honor
1821 1820 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1822 1821 * On old version pools, we suspend the log briefly when taking a
1823 1822 * snapshot so that it will have an empty intent log.
1824 1823 *
1825 1824 * Long holds are not really intended to be used the way we do here --
1826 1825 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1827 1826 * could fail. Therefore we take pains to only put a long hold if it is
1828 1827 * actually necessary. Fortunately, it will only be necessary if the
1829 1828 * objset is currently mounted (or the ZVOL equivalent). In that case it
1830 1829 * will already have a long hold, so we are not really making things any worse.
1831 1830 *
1832 1831 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1833 1832 * zvol_state_t), and use their mechanism to prevent their hold from being
1834 1833 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1835 1834 * very little gain.
1836 1835 *
1837 1836 * if cookiep == NULL, this does both the suspend & resume.
1838 1837 * Otherwise, it returns with the dataset "long held", and the cookie
1839 1838 * should be passed into zil_resume().
1840 1839 */
1841 1840 int
1842 1841 zil_suspend(const char *osname, void **cookiep)
1843 1842 {
1844 1843 objset_t *os;
1845 1844 zilog_t *zilog;
1846 1845 const zil_header_t *zh;
1847 1846 int error;
1848 1847
1849 1848 error = dmu_objset_hold(osname, suspend_tag, &os);
1850 1849 if (error != 0)
1851 1850 return (error);
1852 1851 zilog = dmu_objset_zil(os);
1853 1852
1854 1853 mutex_enter(&zilog->zl_lock);
1855 1854 zh = zilog->zl_header;
1856 1855
1857 1856 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1858 1857 mutex_exit(&zilog->zl_lock);
1859 1858 dmu_objset_rele(os, suspend_tag);
1860 1859 return (SET_ERROR(EBUSY));
1861 1860 }
1862 1861
1863 1862 /*
1864 1863 * Don't put a long hold in the cases where we can avoid it. This
1865 1864 * is when there is no cookie so we are doing a suspend & resume
1866 1865 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1867 1866 * for the suspend because it's already suspended, or there's no ZIL.
1868 1867 */
1869 1868 if (cookiep == NULL && !zilog->zl_suspending &&
1870 1869 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1871 1870 mutex_exit(&zilog->zl_lock);
1872 1871 dmu_objset_rele(os, suspend_tag);
1873 1872 return (0);
1874 1873 }
1875 1874
1876 1875 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1877 1876 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1878 1877
1879 1878 zilog->zl_suspend++;
1880 1879
1881 1880 if (zilog->zl_suspend > 1) {
1882 1881 /*
1883 1882 * Someone else is already suspending it.
1884 1883 * Just wait for them to finish.
1885 1884 */
1886 1885
1887 1886 while (zilog->zl_suspending)
1888 1887 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1889 1888 mutex_exit(&zilog->zl_lock);
1890 1889
1891 1890 if (cookiep == NULL)
1892 1891 zil_resume(os);
1893 1892 else
1894 1893 *cookiep = os;
1895 1894 return (0);
1896 1895 }
1897 1896
1898 1897 /*
1899 1898 * If there is no pointer to an on-disk block, this ZIL must not
1900 1899 * be active (e.g. filesystem not mounted), so there's nothing
1901 1900 * to clean up.
1902 1901 */
1903 1902 if (BP_IS_HOLE(&zh->zh_log)) {
1904 1903 ASSERT(cookiep != NULL); /* fast path already handled */
1905 1904
1906 1905 *cookiep = os;
1907 1906 mutex_exit(&zilog->zl_lock);
1908 1907 return (0);
1909 1908 }
1910 1909
1911 1910 zilog->zl_suspending = B_TRUE;
1912 1911 mutex_exit(&zilog->zl_lock);
1913 1912
1914 1913 zil_commit(zilog, 0);
1915 1914
1916 1915 zil_destroy(zilog, B_FALSE);
1917 1916
1918 1917 mutex_enter(&zilog->zl_lock);
1919 1918 zilog->zl_suspending = B_FALSE;
1920 1919 cv_broadcast(&zilog->zl_cv_suspend);
1921 1920 mutex_exit(&zilog->zl_lock);
1922 1921
1923 1922 if (cookiep == NULL)
1924 1923 zil_resume(os);
1925 1924 else
1926 1925 *cookiep = os;
1927 1926 return (0);
1928 1927 }
1929 1928
1930 1929 void
1931 1930 zil_resume(void *cookie)
1932 1931 {
1933 1932 objset_t *os = cookie;
1934 1933 zilog_t *zilog = dmu_objset_zil(os);
1935 1934
1936 1935 mutex_enter(&zilog->zl_lock);
1937 1936 ASSERT(zilog->zl_suspend != 0);
1938 1937 zilog->zl_suspend--;
1939 1938 mutex_exit(&zilog->zl_lock);
1940 1939 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1941 1940 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1942 1941 }
1943 1942
1944 1943 typedef struct zil_replay_arg {
1945 1944 zil_replay_func_t **zr_replay;
1946 1945 void *zr_arg;
1947 1946 boolean_t zr_byteswap;
1948 1947 char *zr_lr;
1949 1948 } zil_replay_arg_t;
1950 1949
1951 1950 static int
1952 1951 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1953 1952 {
1954 1953 char name[MAXNAMELEN];
1955 1954
1956 1955 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1957 1956
1958 1957 dmu_objset_name(zilog->zl_os, name);
1959 1958
1960 1959 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1961 1960 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1962 1961 (u_longlong_t)lr->lrc_seq,
1963 1962 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1964 1963 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1965 1964
1966 1965 return (error);
1967 1966 }
1968 1967
1969 1968 static int
1970 1969 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1971 1970 {
1972 1971 zil_replay_arg_t *zr = zra;
1973 1972 const zil_header_t *zh = zilog->zl_header;
1974 1973 uint64_t reclen = lr->lrc_reclen;
1975 1974 uint64_t txtype = lr->lrc_txtype;
1976 1975 int error = 0;
1977 1976
1978 1977 zilog->zl_replaying_seq = lr->lrc_seq;
1979 1978
1980 1979 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1981 1980 return (0);
1982 1981
1983 1982 if (lr->lrc_txg < claim_txg) /* already committed */
1984 1983 return (0);
1985 1984
1986 1985 /* Strip case-insensitive bit, still present in log record */
1987 1986 txtype &= ~TX_CI;
1988 1987
1989 1988 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1990 1989 return (zil_replay_error(zilog, lr, EINVAL));
1991 1990
1992 1991 /*
1993 1992 * If this record type can be logged out of order, the object
1994 1993 * (lr_foid) may no longer exist. That's legitimate, not an error.
1995 1994 */
1996 1995 if (TX_OOO(txtype)) {
1997 1996 error = dmu_object_info(zilog->zl_os,
1998 1997 ((lr_ooo_t *)lr)->lr_foid, NULL);
1999 1998 if (error == ENOENT || error == EEXIST)
2000 1999 return (0);
2001 2000 }
2002 2001
2003 2002 /*
2004 2003 * Make a copy of the data so we can revise and extend it.
2005 2004 */
2006 2005 bcopy(lr, zr->zr_lr, reclen);
2007 2006
2008 2007 /*
2009 2008 * If this is a TX_WRITE with a blkptr, suck in the data.
2010 2009 */
2011 2010 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2012 2011 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2013 2012 zr->zr_lr + reclen);
2014 2013 if (error != 0)
2015 2014 return (zil_replay_error(zilog, lr, error));
2016 2015 }
2017 2016
2018 2017 /*
2019 2018 * The log block containing this lr may have been byteswapped
2020 2019 * so that we can easily examine common fields like lrc_txtype.
2021 2020 * However, the log is a mix of different record types, and only the
2022 2021 * replay vectors know how to byteswap their records. Therefore, if
2023 2022 * the lr was byteswapped, undo it before invoking the replay vector.
2024 2023 */
2025 2024 if (zr->zr_byteswap)
2026 2025 byteswap_uint64_array(zr->zr_lr, reclen);
2027 2026
2028 2027 /*
2029 2028 * We must now do two things atomically: replay this log record,
2030 2029 * and update the log header sequence number to reflect the fact that
2031 2030 * we did so. At the end of each replay function the sequence number
2032 2031 * is updated if we are in replay mode.
2033 2032 */
2034 2033 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2035 2034 if (error != 0) {
2036 2035 /*
2037 2036 * The DMU's dnode layer doesn't see removes until the txg
2038 2037 * commits, so a subsequent claim can spuriously fail with
2039 2038 * EEXIST. So if we receive any error we try syncing out
2040 2039 * any removes then retry the transaction. Note that we
2041 2040 * specify B_FALSE for byteswap now, so we don't do it twice.
2042 2041 */
2043 2042 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2044 2043 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2045 2044 if (error != 0)
2046 2045 return (zil_replay_error(zilog, lr, error));
2047 2046 }
2048 2047 return (0);
2049 2048 }
2050 2049
2051 2050 /* ARGSUSED */
2052 2051 static int
2053 2052 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2054 2053 {
2055 2054 zilog->zl_replay_blks++;
2056 2055
2057 2056 return (0);
2058 2057 }
2059 2058
2060 2059 /*
2061 2060 * If this dataset has a non-empty intent log, replay it and destroy it.
2062 2061 */
2063 2062 void
2064 2063 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2065 2064 {
2066 2065 zilog_t *zilog = dmu_objset_zil(os);
2067 2066 const zil_header_t *zh = zilog->zl_header;
2068 2067 zil_replay_arg_t zr;
2069 2068
2070 2069 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2071 2070 zil_destroy(zilog, B_TRUE);
2072 2071 return;
2073 2072 }
2074 2073
2075 2074 zr.zr_replay = replay_func;
2076 2075 zr.zr_arg = arg;
2077 2076 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2078 2077 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2079 2078
2080 2079 /*
2081 2080 * Wait for in-progress removes to sync before starting replay.
2082 2081 */
2083 2082 txg_wait_synced(zilog->zl_dmu_pool, 0);
2084 2083
2085 2084 zilog->zl_replay = B_TRUE;
2086 2085 zilog->zl_replay_time = ddi_get_lbolt();
2087 2086 ASSERT(zilog->zl_replay_blks == 0);
2088 2087 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2089 2088 zh->zh_claim_txg);
2090 2089 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2091 2090
2092 2091 zil_destroy(zilog, B_FALSE);
2093 2092 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2094 2093 zilog->zl_replay = B_FALSE;
2095 2094 }
2096 2095
2097 2096 boolean_t
2098 2097 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2099 2098 {
2100 2099 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2101 2100 return (B_TRUE);
2102 2101
2103 2102 if (zilog->zl_replay) {
2104 2103 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2105 2104 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2106 2105 zilog->zl_replaying_seq;
2107 2106 return (B_TRUE);
2108 2107 }
2109 2108
2110 2109 return (B_FALSE);
2111 2110 }
2112 2111
2113 2112 /* ARGSUSED */
2114 2113 int
2115 2114 zil_vdev_offline(const char *osname, void *arg)
2116 2115 {
2117 2116 int error;
2118 2117
2119 2118 error = zil_suspend(osname, NULL);
2120 2119 if (error != 0)
2121 2120 return (SET_ERROR(EEXIST));
2122 2121 return (0);
2123 2122 }
|
↓ open down ↓ |
1377 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX