Print this page
4045 zfs write throttle & i/o scheduler performance work
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
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 = BP_IDENTITY(bp);
141 141 zil_bp_node_t *zn;
142 142 avl_index_t where;
143 143
144 144 if (avl_find(t, dva, &where) != NULL)
145 145 return (SET_ERROR(EEXIST));
146 146
147 147 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
148 148 zn->zn_dva = *dva;
149 149 avl_insert(t, zn, where);
150 150
151 151 return (0);
152 152 }
153 153
154 154 static zil_header_t *
155 155 zil_header_in_syncing_context(zilog_t *zilog)
156 156 {
157 157 return ((zil_header_t *)zilog->zl_header);
158 158 }
159 159
160 160 static void
161 161 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
162 162 {
163 163 zio_cksum_t *zc = &bp->blk_cksum;
164 164
165 165 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
166 166 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
167 167 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
168 168 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
169 169 }
170 170
171 171 /*
172 172 * Read a log block and make sure it's valid.
173 173 */
174 174 static int
175 175 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
176 176 char **end)
177 177 {
178 178 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
179 179 uint32_t aflags = ARC_WAIT;
180 180 arc_buf_t *abuf = NULL;
181 181 zbookmark_t zb;
182 182 int error;
183 183
184 184 if (zilog->zl_header->zh_claim_txg == 0)
185 185 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
186 186
187 187 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
188 188 zio_flags |= ZIO_FLAG_SPECULATIVE;
189 189
190 190 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
191 191 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
192 192
193 193 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
194 194 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
195 195
196 196 if (error == 0) {
197 197 zio_cksum_t cksum = bp->blk_cksum;
198 198
199 199 /*
200 200 * Validate the checksummed log block.
201 201 *
202 202 * Sequence numbers should be... sequential. The checksum
203 203 * verifier for the next block should be bp's checksum plus 1.
204 204 *
205 205 * Also check the log chain linkage and size used.
206 206 */
207 207 cksum.zc_word[ZIL_ZC_SEQ]++;
208 208
209 209 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
210 210 zil_chain_t *zilc = abuf->b_data;
211 211 char *lr = (char *)(zilc + 1);
212 212 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
213 213
214 214 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
215 215 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
216 216 error = SET_ERROR(ECKSUM);
217 217 } else {
218 218 bcopy(lr, dst, len);
219 219 *end = (char *)dst + len;
220 220 *nbp = zilc->zc_next_blk;
221 221 }
222 222 } else {
223 223 char *lr = abuf->b_data;
224 224 uint64_t size = BP_GET_LSIZE(bp);
225 225 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
226 226
227 227 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
228 228 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
229 229 (zilc->zc_nused > (size - sizeof (*zilc)))) {
230 230 error = SET_ERROR(ECKSUM);
231 231 } else {
232 232 bcopy(lr, dst, zilc->zc_nused);
233 233 *end = (char *)dst + zilc->zc_nused;
234 234 *nbp = zilc->zc_next_blk;
235 235 }
236 236 }
237 237
238 238 VERIFY(arc_buf_remove_ref(abuf, &abuf));
239 239 }
240 240
241 241 return (error);
242 242 }
243 243
244 244 /*
245 245 * Read a TX_WRITE log data block.
246 246 */
247 247 static int
248 248 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
249 249 {
250 250 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
251 251 const blkptr_t *bp = &lr->lr_blkptr;
252 252 uint32_t aflags = ARC_WAIT;
253 253 arc_buf_t *abuf = NULL;
254 254 zbookmark_t zb;
255 255 int error;
256 256
257 257 if (BP_IS_HOLE(bp)) {
258 258 if (wbuf != NULL)
259 259 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
260 260 return (0);
261 261 }
262 262
263 263 if (zilog->zl_header->zh_claim_txg == 0)
264 264 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
265 265
266 266 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
267 267 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
268 268
269 269 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
270 270 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
271 271
272 272 if (error == 0) {
273 273 if (wbuf != NULL)
274 274 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
275 275 (void) arc_buf_remove_ref(abuf, &abuf);
276 276 }
277 277
278 278 return (error);
279 279 }
280 280
281 281 /*
282 282 * Parse the intent log, and call parse_func for each valid record within.
283 283 */
284 284 int
285 285 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
286 286 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
287 287 {
288 288 const zil_header_t *zh = zilog->zl_header;
289 289 boolean_t claimed = !!zh->zh_claim_txg;
290 290 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
291 291 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
292 292 uint64_t max_blk_seq = 0;
293 293 uint64_t max_lr_seq = 0;
294 294 uint64_t blk_count = 0;
295 295 uint64_t lr_count = 0;
296 296 blkptr_t blk, next_blk;
297 297 char *lrbuf, *lrp;
298 298 int error = 0;
299 299
300 300 /*
301 301 * Old logs didn't record the maximum zh_claim_lr_seq.
302 302 */
303 303 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
304 304 claim_lr_seq = UINT64_MAX;
305 305
306 306 /*
307 307 * Starting at the block pointed to by zh_log we read the log chain.
308 308 * For each block in the chain we strongly check that block to
309 309 * ensure its validity. We stop when an invalid block is found.
310 310 * For each block pointer in the chain we call parse_blk_func().
311 311 * For each record in each valid block we call parse_lr_func().
312 312 * If the log has been claimed, stop if we encounter a sequence
313 313 * number greater than the highest claimed sequence number.
314 314 */
315 315 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
316 316 zil_bp_tree_init(zilog);
317 317
318 318 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
319 319 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
320 320 int reclen;
321 321 char *end;
322 322
323 323 if (blk_seq > claim_blk_seq)
324 324 break;
325 325 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
326 326 break;
327 327 ASSERT3U(max_blk_seq, <, blk_seq);
328 328 max_blk_seq = blk_seq;
329 329 blk_count++;
330 330
331 331 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
332 332 break;
333 333
334 334 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
335 335 if (error != 0)
336 336 break;
337 337
338 338 for (lrp = lrbuf; lrp < end; lrp += reclen) {
339 339 lr_t *lr = (lr_t *)lrp;
340 340 reclen = lr->lrc_reclen;
341 341 ASSERT3U(reclen, >=, sizeof (lr_t));
342 342 if (lr->lrc_seq > claim_lr_seq)
343 343 goto done;
344 344 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
345 345 goto done;
346 346 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
347 347 max_lr_seq = lr->lrc_seq;
348 348 lr_count++;
349 349 }
350 350 }
351 351 done:
352 352 zilog->zl_parse_error = error;
353 353 zilog->zl_parse_blk_seq = max_blk_seq;
354 354 zilog->zl_parse_lr_seq = max_lr_seq;
355 355 zilog->zl_parse_blk_count = blk_count;
356 356 zilog->zl_parse_lr_count = lr_count;
357 357
358 358 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
359 359 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
360 360
361 361 zil_bp_tree_fini(zilog);
362 362 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
363 363
364 364 return (error);
365 365 }
366 366
367 367 static int
368 368 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
369 369 {
370 370 /*
371 371 * Claim log block if not already committed and not already claimed.
372 372 * If tx == NULL, just verify that the block is claimable.
373 373 */
374 374 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
375 375 return (0);
376 376
377 377 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
378 378 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
379 379 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
380 380 }
381 381
382 382 static int
383 383 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
384 384 {
385 385 lr_write_t *lr = (lr_write_t *)lrc;
386 386 int error;
387 387
388 388 if (lrc->lrc_txtype != TX_WRITE)
389 389 return (0);
390 390
391 391 /*
392 392 * If the block is not readable, don't claim it. This can happen
393 393 * in normal operation when a log block is written to disk before
394 394 * some of the dmu_sync() blocks it points to. In this case, the
395 395 * transaction cannot have been committed to anyone (we would have
396 396 * waited for all writes to be stable first), so it is semantically
397 397 * correct to declare this the end of the log.
398 398 */
399 399 if (lr->lr_blkptr.blk_birth >= first_txg &&
400 400 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
401 401 return (error);
402 402 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
403 403 }
404 404
405 405 /* ARGSUSED */
406 406 static int
407 407 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
408 408 {
409 409 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
410 410
411 411 return (0);
412 412 }
413 413
414 414 static int
415 415 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
416 416 {
417 417 lr_write_t *lr = (lr_write_t *)lrc;
418 418 blkptr_t *bp = &lr->lr_blkptr;
419 419
420 420 /*
421 421 * If we previously claimed it, we need to free it.
422 422 */
423 423 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
424 424 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
425 425 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
426 426
427 427 return (0);
428 428 }
429 429
430 430 static lwb_t *
431 431 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
432 432 {
433 433 lwb_t *lwb;
434 434
435 435 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
436 436 lwb->lwb_zilog = zilog;
437 437 lwb->lwb_blk = *bp;
438 438 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
439 439 lwb->lwb_max_txg = txg;
440 440 lwb->lwb_zio = NULL;
441 441 lwb->lwb_tx = NULL;
442 442 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
443 443 lwb->lwb_nused = sizeof (zil_chain_t);
444 444 lwb->lwb_sz = BP_GET_LSIZE(bp);
445 445 } else {
446 446 lwb->lwb_nused = 0;
447 447 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
448 448 }
449 449
450 450 mutex_enter(&zilog->zl_lock);
451 451 list_insert_tail(&zilog->zl_lwb_list, lwb);
452 452 mutex_exit(&zilog->zl_lock);
453 453
454 454 return (lwb);
455 455 }
456 456
457 457 /*
458 458 * Called when we create in-memory log transactions so that we know
459 459 * to cleanup the itxs at the end of spa_sync().
460 460 */
461 461 void
462 462 zilog_dirty(zilog_t *zilog, uint64_t txg)
463 463 {
464 464 dsl_pool_t *dp = zilog->zl_dmu_pool;
465 465 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
466 466
467 467 if (dsl_dataset_is_snapshot(ds))
468 468 panic("dirtying snapshot!");
469 469
470 470 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
471 471 /* up the hold count until we can be written out */
472 472 dmu_buf_add_ref(ds->ds_dbuf, zilog);
473 473 }
474 474 }
475 475
476 476 boolean_t
477 477 zilog_is_dirty(zilog_t *zilog)
478 478 {
479 479 dsl_pool_t *dp = zilog->zl_dmu_pool;
480 480
481 481 for (int t = 0; t < TXG_SIZE; t++) {
482 482 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
483 483 return (B_TRUE);
484 484 }
485 485 return (B_FALSE);
486 486 }
487 487
488 488 /*
489 489 * Create an on-disk intent log.
490 490 */
491 491 static lwb_t *
492 492 zil_create(zilog_t *zilog)
493 493 {
494 494 const zil_header_t *zh = zilog->zl_header;
495 495 lwb_t *lwb = NULL;
496 496 uint64_t txg = 0;
497 497 dmu_tx_t *tx = NULL;
498 498 blkptr_t blk;
499 499 int error = 0;
500 500
501 501 /*
502 502 * Wait for any previous destroy to complete.
503 503 */
504 504 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
505 505
506 506 ASSERT(zh->zh_claim_txg == 0);
507 507 ASSERT(zh->zh_replay_seq == 0);
508 508
509 509 blk = zh->zh_log;
510 510
511 511 /*
512 512 * Allocate an initial log block if:
513 513 * - there isn't one already
514 514 * - the existing block is the wrong endianess
515 515 */
516 516 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
517 517 tx = dmu_tx_create(zilog->zl_os);
518 518 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
519 519 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
520 520 txg = dmu_tx_get_txg(tx);
521 521
522 522 if (!BP_IS_HOLE(&blk)) {
523 523 zio_free_zil(zilog->zl_spa, txg, &blk);
524 524 BP_ZERO(&blk);
525 525 }
526 526
527 527 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
528 528 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
529 529
530 530 if (error == 0)
531 531 zil_init_log_chain(zilog, &blk);
532 532 }
533 533
534 534 /*
535 535 * Allocate a log write buffer (lwb) for the first log block.
536 536 */
537 537 if (error == 0)
538 538 lwb = zil_alloc_lwb(zilog, &blk, txg);
539 539
540 540 /*
541 541 * If we just allocated the first log block, commit our transaction
542 542 * and wait for zil_sync() to stuff the block poiner into zh_log.
543 543 * (zh is part of the MOS, so we cannot modify it in open context.)
544 544 */
545 545 if (tx != NULL) {
546 546 dmu_tx_commit(tx);
547 547 txg_wait_synced(zilog->zl_dmu_pool, txg);
548 548 }
549 549
550 550 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
551 551
552 552 return (lwb);
553 553 }
554 554
555 555 /*
556 556 * In one tx, free all log blocks and clear the log header.
557 557 * If keep_first is set, then we're replaying a log with no content.
558 558 * We want to keep the first block, however, so that the first
559 559 * synchronous transaction doesn't require a txg_wait_synced()
560 560 * in zil_create(). We don't need to txg_wait_synced() here either
561 561 * when keep_first is set, because both zil_create() and zil_destroy()
562 562 * will wait for any in-progress destroys to complete.
563 563 */
564 564 void
565 565 zil_destroy(zilog_t *zilog, boolean_t keep_first)
566 566 {
567 567 const zil_header_t *zh = zilog->zl_header;
568 568 lwb_t *lwb;
569 569 dmu_tx_t *tx;
570 570 uint64_t txg;
571 571
572 572 /*
573 573 * Wait for any previous destroy to complete.
574 574 */
575 575 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
576 576
577 577 zilog->zl_old_header = *zh; /* debugging aid */
578 578
579 579 if (BP_IS_HOLE(&zh->zh_log))
580 580 return;
581 581
582 582 tx = dmu_tx_create(zilog->zl_os);
583 583 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
584 584 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
585 585 txg = dmu_tx_get_txg(tx);
586 586
587 587 mutex_enter(&zilog->zl_lock);
588 588
589 589 ASSERT3U(zilog->zl_destroy_txg, <, txg);
590 590 zilog->zl_destroy_txg = txg;
591 591 zilog->zl_keep_first = keep_first;
592 592
593 593 if (!list_is_empty(&zilog->zl_lwb_list)) {
594 594 ASSERT(zh->zh_claim_txg == 0);
595 595 VERIFY(!keep_first);
596 596 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
597 597 list_remove(&zilog->zl_lwb_list, lwb);
598 598 if (lwb->lwb_buf != NULL)
599 599 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
600 600 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
601 601 kmem_cache_free(zil_lwb_cache, lwb);
602 602 }
603 603 } else if (!keep_first) {
604 604 zil_destroy_sync(zilog, tx);
605 605 }
606 606 mutex_exit(&zilog->zl_lock);
607 607
608 608 dmu_tx_commit(tx);
609 609 }
610 610
611 611 void
612 612 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
613 613 {
614 614 ASSERT(list_is_empty(&zilog->zl_lwb_list));
615 615 (void) zil_parse(zilog, zil_free_log_block,
616 616 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
617 617 }
618 618
619 619 int
620 620 zil_claim(const char *osname, void *txarg)
621 621 {
622 622 dmu_tx_t *tx = txarg;
623 623 uint64_t first_txg = dmu_tx_get_txg(tx);
624 624 zilog_t *zilog;
625 625 zil_header_t *zh;
626 626 objset_t *os;
627 627 int error;
628 628
629 629 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
630 630 if (error != 0) {
631 631 cmn_err(CE_WARN, "can't open objset for %s", osname);
632 632 return (0);
633 633 }
634 634
635 635 zilog = dmu_objset_zil(os);
636 636 zh = zil_header_in_syncing_context(zilog);
637 637
638 638 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
639 639 if (!BP_IS_HOLE(&zh->zh_log))
640 640 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
641 641 BP_ZERO(&zh->zh_log);
642 642 dsl_dataset_dirty(dmu_objset_ds(os), tx);
643 643 dmu_objset_disown(os, FTAG);
644 644 return (0);
645 645 }
646 646
647 647 /*
648 648 * Claim all log blocks if we haven't already done so, and remember
649 649 * the highest claimed sequence number. This ensures that if we can
650 650 * read only part of the log now (e.g. due to a missing device),
651 651 * but we can read the entire log later, we will not try to replay
652 652 * or destroy beyond the last block we successfully claimed.
653 653 */
654 654 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
655 655 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
656 656 (void) zil_parse(zilog, zil_claim_log_block,
657 657 zil_claim_log_record, tx, first_txg);
658 658 zh->zh_claim_txg = first_txg;
659 659 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
660 660 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
661 661 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
662 662 zh->zh_flags |= ZIL_REPLAY_NEEDED;
663 663 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
664 664 dsl_dataset_dirty(dmu_objset_ds(os), tx);
665 665 }
666 666
667 667 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
668 668 dmu_objset_disown(os, FTAG);
669 669 return (0);
670 670 }
671 671
672 672 /*
673 673 * Check the log by walking the log chain.
674 674 * Checksum errors are ok as they indicate the end of the chain.
675 675 * Any other error (no device or read failure) returns an error.
676 676 */
677 677 int
678 678 zil_check_log_chain(const char *osname, void *tx)
679 679 {
680 680 zilog_t *zilog;
681 681 objset_t *os;
682 682 blkptr_t *bp;
683 683 int error;
684 684
685 685 ASSERT(tx == NULL);
686 686
687 687 error = dmu_objset_hold(osname, FTAG, &os);
688 688 if (error != 0) {
689 689 cmn_err(CE_WARN, "can't open objset for %s", osname);
690 690 return (0);
691 691 }
692 692
693 693 zilog = dmu_objset_zil(os);
694 694 bp = (blkptr_t *)&zilog->zl_header->zh_log;
695 695
696 696 /*
697 697 * Check the first block and determine if it's on a log device
698 698 * which may have been removed or faulted prior to loading this
699 699 * pool. If so, there's no point in checking the rest of the log
700 700 * as its content should have already been synced to the pool.
701 701 */
702 702 if (!BP_IS_HOLE(bp)) {
703 703 vdev_t *vd;
704 704 boolean_t valid = B_TRUE;
705 705
706 706 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
707 707 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
708 708 if (vd->vdev_islog && vdev_is_dead(vd))
709 709 valid = vdev_log_state_valid(vd);
710 710 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
711 711
712 712 if (!valid) {
713 713 dmu_objset_rele(os, FTAG);
714 714 return (0);
715 715 }
716 716 }
717 717
718 718 /*
719 719 * Because tx == NULL, zil_claim_log_block() will not actually claim
720 720 * any blocks, but just determine whether it is possible to do so.
721 721 * In addition to checking the log chain, zil_claim_log_block()
722 722 * will invoke zio_claim() with a done func of spa_claim_notify(),
723 723 * which will update spa_max_claim_txg. See spa_load() for details.
724 724 */
725 725 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
726 726 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
727 727
728 728 dmu_objset_rele(os, FTAG);
729 729
730 730 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
731 731 }
732 732
733 733 static int
734 734 zil_vdev_compare(const void *x1, const void *x2)
735 735 {
736 736 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
737 737 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
738 738
739 739 if (v1 < v2)
740 740 return (-1);
741 741 if (v1 > v2)
742 742 return (1);
743 743
744 744 return (0);
745 745 }
746 746
747 747 void
748 748 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
749 749 {
750 750 avl_tree_t *t = &zilog->zl_vdev_tree;
751 751 avl_index_t where;
752 752 zil_vdev_node_t *zv, zvsearch;
753 753 int ndvas = BP_GET_NDVAS(bp);
754 754 int i;
755 755
756 756 if (zfs_nocacheflush)
757 757 return;
758 758
759 759 ASSERT(zilog->zl_writer);
760 760
761 761 /*
762 762 * Even though we're zl_writer, we still need a lock because the
763 763 * zl_get_data() callbacks may have dmu_sync() done callbacks
764 764 * that will run concurrently.
765 765 */
766 766 mutex_enter(&zilog->zl_vdev_lock);
767 767 for (i = 0; i < ndvas; i++) {
768 768 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
769 769 if (avl_find(t, &zvsearch, &where) == NULL) {
770 770 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
771 771 zv->zv_vdev = zvsearch.zv_vdev;
772 772 avl_insert(t, zv, where);
773 773 }
774 774 }
775 775 mutex_exit(&zilog->zl_vdev_lock);
776 776 }
777 777
778 778 static void
779 779 zil_flush_vdevs(zilog_t *zilog)
780 780 {
781 781 spa_t *spa = zilog->zl_spa;
782 782 avl_tree_t *t = &zilog->zl_vdev_tree;
783 783 void *cookie = NULL;
784 784 zil_vdev_node_t *zv;
785 785 zio_t *zio;
786 786
787 787 ASSERT(zilog->zl_writer);
788 788
789 789 /*
790 790 * We don't need zl_vdev_lock here because we're the zl_writer,
791 791 * and all zl_get_data() callbacks are done.
792 792 */
793 793 if (avl_numnodes(t) == 0)
794 794 return;
795 795
796 796 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
797 797
798 798 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
799 799
800 800 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
801 801 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
802 802 if (vd != NULL)
803 803 zio_flush(zio, vd);
804 804 kmem_free(zv, sizeof (*zv));
805 805 }
806 806
807 807 /*
808 808 * Wait for all the flushes to complete. Not all devices actually
809 809 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
810 810 */
811 811 (void) zio_wait(zio);
812 812
813 813 spa_config_exit(spa, SCL_STATE, FTAG);
814 814 }
815 815
816 816 /*
817 817 * Function called when a log block write completes
818 818 */
819 819 static void
820 820 zil_lwb_write_done(zio_t *zio)
821 821 {
822 822 lwb_t *lwb = zio->io_private;
823 823 zilog_t *zilog = lwb->lwb_zilog;
824 824 dmu_tx_t *tx = lwb->lwb_tx;
825 825
826 826 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
827 827 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
828 828 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
829 829 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
830 830 ASSERT(!BP_IS_GANG(zio->io_bp));
831 831 ASSERT(!BP_IS_HOLE(zio->io_bp));
832 832 ASSERT(zio->io_bp->blk_fill == 0);
833 833
834 834 /*
835 835 * Ensure the lwb buffer pointer is cleared before releasing
836 836 * the txg. If we have had an allocation failure and
837 837 * the txg is waiting to sync then we want want zil_sync()
838 838 * to remove the lwb so that it's not picked up as the next new
839 839 * one in zil_commit_writer(). zil_sync() will only remove
840 840 * the lwb if lwb_buf is null.
841 841 */
842 842 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
843 843 mutex_enter(&zilog->zl_lock);
844 844 lwb->lwb_buf = NULL;
845 845 lwb->lwb_tx = NULL;
846 846 mutex_exit(&zilog->zl_lock);
847 847
848 848 /*
849 849 * Now that we've written this log block, we have a stable pointer
850 850 * to the next block in the chain, so it's OK to let the txg in
851 851 * which we allocated the next block sync.
852 852 */
853 853 dmu_tx_commit(tx);
854 854 }
855 855
856 856 /*
857 857 * Initialize the io for a log block.
858 858 */
859 859 static void
860 860 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
861 861 {
862 862 zbookmark_t zb;
863 863
864 864 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
↓ open down ↓ |
864 lines elided |
↑ open up ↑ |
865 865 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
866 866 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
867 867
868 868 if (zilog->zl_root_zio == NULL) {
869 869 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
870 870 ZIO_FLAG_CANFAIL);
871 871 }
872 872 if (lwb->lwb_zio == NULL) {
873 873 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
874 874 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
875 - zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
875 + zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
876 876 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
877 877 }
878 878 }
879 879
880 880 /*
881 881 * Define a limited set of intent log block sizes.
882 882 *
883 883 * These must be a multiple of 4KB. Note only the amount used (again
884 884 * aligned to 4KB) actually gets written. However, we can't always just
885 885 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
886 886 */
887 887 uint64_t zil_block_buckets[] = {
888 888 4096, /* non TX_WRITE */
889 889 8192+4096, /* data base */
890 890 32*1024 + 4096, /* NFS writes */
891 891 UINT64_MAX
892 892 };
893 893
894 894 /*
895 895 * Use the slog as long as the logbias is 'latency' and the current commit size
896 896 * is less than the limit or the total list size is less than 2X the limit.
897 897 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
898 898 */
899 899 uint64_t zil_slog_limit = 1024 * 1024;
900 900 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
901 901 (((zilog)->zl_cur_used < zil_slog_limit) || \
902 902 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
903 903
904 904 /*
905 905 * Start a log block write and advance to the next log block.
906 906 * Calls are serialized.
907 907 */
908 908 static lwb_t *
909 909 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
910 910 {
911 911 lwb_t *nlwb = NULL;
912 912 zil_chain_t *zilc;
913 913 spa_t *spa = zilog->zl_spa;
914 914 blkptr_t *bp;
915 915 dmu_tx_t *tx;
916 916 uint64_t txg;
917 917 uint64_t zil_blksz, wsz;
918 918 int i, error;
919 919
920 920 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
921 921 zilc = (zil_chain_t *)lwb->lwb_buf;
922 922 bp = &zilc->zc_next_blk;
923 923 } else {
924 924 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
925 925 bp = &zilc->zc_next_blk;
926 926 }
927 927
928 928 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
929 929
930 930 /*
931 931 * Allocate the next block and save its address in this block
932 932 * before writing it in order to establish the log chain.
933 933 * Note that if the allocation of nlwb synced before we wrote
934 934 * the block that points at it (lwb), we'd leak it if we crashed.
935 935 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
936 936 * We dirty the dataset to ensure that zil_sync() will be called
937 937 * to clean up in the event of allocation failure or I/O failure.
938 938 */
939 939 tx = dmu_tx_create(zilog->zl_os);
940 940 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
941 941 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
942 942 txg = dmu_tx_get_txg(tx);
943 943
944 944 lwb->lwb_tx = tx;
945 945
946 946 /*
947 947 * Log blocks are pre-allocated. Here we select the size of the next
948 948 * block, based on size used in the last block.
949 949 * - first find the smallest bucket that will fit the block from a
950 950 * limited set of block sizes. This is because it's faster to write
951 951 * blocks allocated from the same metaslab as they are adjacent or
952 952 * close.
953 953 * - next find the maximum from the new suggested size and an array of
954 954 * previous sizes. This lessens a picket fence effect of wrongly
955 955 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
956 956 * requests.
957 957 *
958 958 * Note we only write what is used, but we can't just allocate
959 959 * the maximum block size because we can exhaust the available
960 960 * pool log space.
961 961 */
962 962 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
963 963 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
964 964 continue;
965 965 zil_blksz = zil_block_buckets[i];
966 966 if (zil_blksz == UINT64_MAX)
967 967 zil_blksz = SPA_MAXBLOCKSIZE;
968 968 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
969 969 for (i = 0; i < ZIL_PREV_BLKS; i++)
970 970 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
971 971 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
972 972
973 973 BP_ZERO(bp);
974 974 /* pass the old blkptr in order to spread log blocks across devs */
975 975 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
976 976 USE_SLOG(zilog));
977 977 if (error == 0) {
978 978 ASSERT3U(bp->blk_birth, ==, txg);
979 979 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
980 980 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
981 981
982 982 /*
983 983 * Allocate a new log write buffer (lwb).
984 984 */
985 985 nlwb = zil_alloc_lwb(zilog, bp, txg);
986 986
987 987 /* Record the block for later vdev flushing */
988 988 zil_add_block(zilog, &lwb->lwb_blk);
989 989 }
990 990
991 991 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
992 992 /* For Slim ZIL only write what is used. */
993 993 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
994 994 ASSERT3U(wsz, <=, lwb->lwb_sz);
995 995 zio_shrink(lwb->lwb_zio, wsz);
996 996
997 997 } else {
998 998 wsz = lwb->lwb_sz;
999 999 }
1000 1000
1001 1001 zilc->zc_pad = 0;
1002 1002 zilc->zc_nused = lwb->lwb_nused;
1003 1003 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1004 1004
1005 1005 /*
1006 1006 * clear unused data for security
1007 1007 */
1008 1008 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1009 1009
1010 1010 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1011 1011
1012 1012 /*
1013 1013 * If there was an allocation failure then nlwb will be null which
1014 1014 * forces a txg_wait_synced().
1015 1015 */
1016 1016 return (nlwb);
1017 1017 }
1018 1018
1019 1019 static lwb_t *
1020 1020 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1021 1021 {
1022 1022 lr_t *lrc = &itx->itx_lr; /* common log record */
1023 1023 lr_write_t *lrw = (lr_write_t *)lrc;
1024 1024 char *lr_buf;
1025 1025 uint64_t txg = lrc->lrc_txg;
1026 1026 uint64_t reclen = lrc->lrc_reclen;
1027 1027 uint64_t dlen = 0;
1028 1028
1029 1029 if (lwb == NULL)
1030 1030 return (NULL);
1031 1031
1032 1032 ASSERT(lwb->lwb_buf != NULL);
1033 1033 ASSERT(zilog_is_dirty(zilog) ||
1034 1034 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1035 1035
1036 1036 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1037 1037 dlen = P2ROUNDUP_TYPED(
1038 1038 lrw->lr_length, sizeof (uint64_t), uint64_t);
1039 1039
1040 1040 zilog->zl_cur_used += (reclen + dlen);
1041 1041
1042 1042 zil_lwb_write_init(zilog, lwb);
1043 1043
1044 1044 /*
1045 1045 * If this record won't fit in the current log block, start a new one.
1046 1046 */
1047 1047 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1048 1048 lwb = zil_lwb_write_start(zilog, lwb);
1049 1049 if (lwb == NULL)
1050 1050 return (NULL);
1051 1051 zil_lwb_write_init(zilog, lwb);
1052 1052 ASSERT(LWB_EMPTY(lwb));
1053 1053 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1054 1054 txg_wait_synced(zilog->zl_dmu_pool, txg);
1055 1055 return (lwb);
1056 1056 }
1057 1057 }
1058 1058
1059 1059 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1060 1060 bcopy(lrc, lr_buf, reclen);
1061 1061 lrc = (lr_t *)lr_buf;
1062 1062 lrw = (lr_write_t *)lrc;
1063 1063
1064 1064 /*
1065 1065 * If it's a write, fetch the data or get its blkptr as appropriate.
1066 1066 */
1067 1067 if (lrc->lrc_txtype == TX_WRITE) {
1068 1068 if (txg > spa_freeze_txg(zilog->zl_spa))
1069 1069 txg_wait_synced(zilog->zl_dmu_pool, txg);
1070 1070 if (itx->itx_wr_state != WR_COPIED) {
1071 1071 char *dbuf;
1072 1072 int error;
1073 1073
1074 1074 if (dlen) {
1075 1075 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1076 1076 dbuf = lr_buf + reclen;
1077 1077 lrw->lr_common.lrc_reclen += dlen;
1078 1078 } else {
1079 1079 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1080 1080 dbuf = NULL;
1081 1081 }
1082 1082 error = zilog->zl_get_data(
1083 1083 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1084 1084 if (error == EIO) {
1085 1085 txg_wait_synced(zilog->zl_dmu_pool, txg);
1086 1086 return (lwb);
1087 1087 }
1088 1088 if (error != 0) {
1089 1089 ASSERT(error == ENOENT || error == EEXIST ||
1090 1090 error == EALREADY);
1091 1091 return (lwb);
1092 1092 }
1093 1093 }
1094 1094 }
1095 1095
1096 1096 /*
1097 1097 * We're actually making an entry, so update lrc_seq to be the
1098 1098 * log record sequence number. Note that this is generally not
1099 1099 * equal to the itx sequence number because not all transactions
1100 1100 * are synchronous, and sometimes spa_sync() gets there first.
1101 1101 */
1102 1102 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1103 1103 lwb->lwb_nused += reclen + dlen;
1104 1104 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1105 1105 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1106 1106 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1107 1107
1108 1108 return (lwb);
1109 1109 }
1110 1110
1111 1111 itx_t *
1112 1112 zil_itx_create(uint64_t txtype, size_t lrsize)
1113 1113 {
1114 1114 itx_t *itx;
1115 1115
1116 1116 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1117 1117
1118 1118 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1119 1119 itx->itx_lr.lrc_txtype = txtype;
1120 1120 itx->itx_lr.lrc_reclen = lrsize;
1121 1121 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1122 1122 itx->itx_lr.lrc_seq = 0; /* defensive */
1123 1123 itx->itx_sync = B_TRUE; /* default is synchronous */
1124 1124
1125 1125 return (itx);
1126 1126 }
1127 1127
1128 1128 void
1129 1129 zil_itx_destroy(itx_t *itx)
1130 1130 {
1131 1131 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1132 1132 }
1133 1133
1134 1134 /*
1135 1135 * Free up the sync and async itxs. The itxs_t has already been detached
1136 1136 * so no locks are needed.
1137 1137 */
1138 1138 static void
1139 1139 zil_itxg_clean(itxs_t *itxs)
1140 1140 {
1141 1141 itx_t *itx;
1142 1142 list_t *list;
1143 1143 avl_tree_t *t;
1144 1144 void *cookie;
1145 1145 itx_async_node_t *ian;
1146 1146
1147 1147 list = &itxs->i_sync_list;
1148 1148 while ((itx = list_head(list)) != NULL) {
1149 1149 list_remove(list, itx);
1150 1150 kmem_free(itx, offsetof(itx_t, itx_lr) +
1151 1151 itx->itx_lr.lrc_reclen);
1152 1152 }
1153 1153
1154 1154 cookie = NULL;
1155 1155 t = &itxs->i_async_tree;
1156 1156 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1157 1157 list = &ian->ia_list;
1158 1158 while ((itx = list_head(list)) != NULL) {
1159 1159 list_remove(list, itx);
1160 1160 kmem_free(itx, offsetof(itx_t, itx_lr) +
1161 1161 itx->itx_lr.lrc_reclen);
1162 1162 }
1163 1163 list_destroy(list);
1164 1164 kmem_free(ian, sizeof (itx_async_node_t));
1165 1165 }
1166 1166 avl_destroy(t);
1167 1167
1168 1168 kmem_free(itxs, sizeof (itxs_t));
1169 1169 }
1170 1170
1171 1171 static int
1172 1172 zil_aitx_compare(const void *x1, const void *x2)
1173 1173 {
1174 1174 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1175 1175 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1176 1176
1177 1177 if (o1 < o2)
1178 1178 return (-1);
1179 1179 if (o1 > o2)
1180 1180 return (1);
1181 1181
1182 1182 return (0);
1183 1183 }
1184 1184
1185 1185 /*
1186 1186 * Remove all async itx with the given oid.
1187 1187 */
1188 1188 static void
1189 1189 zil_remove_async(zilog_t *zilog, uint64_t oid)
1190 1190 {
1191 1191 uint64_t otxg, txg;
1192 1192 itx_async_node_t *ian;
1193 1193 avl_tree_t *t;
1194 1194 avl_index_t where;
1195 1195 list_t clean_list;
1196 1196 itx_t *itx;
1197 1197
1198 1198 ASSERT(oid != 0);
1199 1199 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1200 1200
1201 1201 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1202 1202 otxg = ZILTEST_TXG;
1203 1203 else
1204 1204 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1205 1205
1206 1206 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1207 1207 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1208 1208
1209 1209 mutex_enter(&itxg->itxg_lock);
1210 1210 if (itxg->itxg_txg != txg) {
1211 1211 mutex_exit(&itxg->itxg_lock);
1212 1212 continue;
1213 1213 }
1214 1214
1215 1215 /*
1216 1216 * Locate the object node and append its list.
1217 1217 */
1218 1218 t = &itxg->itxg_itxs->i_async_tree;
1219 1219 ian = avl_find(t, &oid, &where);
1220 1220 if (ian != NULL)
1221 1221 list_move_tail(&clean_list, &ian->ia_list);
1222 1222 mutex_exit(&itxg->itxg_lock);
1223 1223 }
1224 1224 while ((itx = list_head(&clean_list)) != NULL) {
1225 1225 list_remove(&clean_list, itx);
1226 1226 kmem_free(itx, offsetof(itx_t, itx_lr) +
1227 1227 itx->itx_lr.lrc_reclen);
1228 1228 }
1229 1229 list_destroy(&clean_list);
1230 1230 }
1231 1231
1232 1232 void
1233 1233 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1234 1234 {
1235 1235 uint64_t txg;
1236 1236 itxg_t *itxg;
1237 1237 itxs_t *itxs, *clean = NULL;
1238 1238
1239 1239 /*
1240 1240 * Object ids can be re-instantiated in the next txg so
1241 1241 * remove any async transactions to avoid future leaks.
1242 1242 * This can happen if a fsync occurs on the re-instantiated
1243 1243 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1244 1244 * the new file data and flushes a write record for the old object.
1245 1245 */
1246 1246 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1247 1247 zil_remove_async(zilog, itx->itx_oid);
1248 1248
1249 1249 /*
1250 1250 * Ensure the data of a renamed file is committed before the rename.
1251 1251 */
1252 1252 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1253 1253 zil_async_to_sync(zilog, itx->itx_oid);
1254 1254
1255 1255 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1256 1256 txg = ZILTEST_TXG;
1257 1257 else
1258 1258 txg = dmu_tx_get_txg(tx);
1259 1259
1260 1260 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1261 1261 mutex_enter(&itxg->itxg_lock);
1262 1262 itxs = itxg->itxg_itxs;
1263 1263 if (itxg->itxg_txg != txg) {
1264 1264 if (itxs != NULL) {
1265 1265 /*
1266 1266 * The zil_clean callback hasn't got around to cleaning
1267 1267 * this itxg. Save the itxs for release below.
1268 1268 * This should be rare.
1269 1269 */
1270 1270 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1271 1271 itxg->itxg_sod = 0;
1272 1272 clean = itxg->itxg_itxs;
1273 1273 }
1274 1274 ASSERT(itxg->itxg_sod == 0);
1275 1275 itxg->itxg_txg = txg;
1276 1276 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1277 1277
1278 1278 list_create(&itxs->i_sync_list, sizeof (itx_t),
1279 1279 offsetof(itx_t, itx_node));
1280 1280 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1281 1281 sizeof (itx_async_node_t),
1282 1282 offsetof(itx_async_node_t, ia_node));
1283 1283 }
1284 1284 if (itx->itx_sync) {
1285 1285 list_insert_tail(&itxs->i_sync_list, itx);
1286 1286 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1287 1287 itxg->itxg_sod += itx->itx_sod;
1288 1288 } else {
1289 1289 avl_tree_t *t = &itxs->i_async_tree;
1290 1290 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1291 1291 itx_async_node_t *ian;
1292 1292 avl_index_t where;
1293 1293
1294 1294 ian = avl_find(t, &foid, &where);
1295 1295 if (ian == NULL) {
1296 1296 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1297 1297 list_create(&ian->ia_list, sizeof (itx_t),
1298 1298 offsetof(itx_t, itx_node));
1299 1299 ian->ia_foid = foid;
1300 1300 avl_insert(t, ian, where);
1301 1301 }
1302 1302 list_insert_tail(&ian->ia_list, itx);
1303 1303 }
1304 1304
1305 1305 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1306 1306 zilog_dirty(zilog, txg);
1307 1307 mutex_exit(&itxg->itxg_lock);
1308 1308
1309 1309 /* Release the old itxs now we've dropped the lock */
1310 1310 if (clean != NULL)
1311 1311 zil_itxg_clean(clean);
1312 1312 }
1313 1313
1314 1314 /*
1315 1315 * If there are any in-memory intent log transactions which have now been
1316 1316 * synced then start up a taskq to free them. We should only do this after we
1317 1317 * have written out the uberblocks (i.e. txg has been comitted) so that
1318 1318 * don't inadvertently clean out in-memory log records that would be required
1319 1319 * by zil_commit().
1320 1320 */
1321 1321 void
1322 1322 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1323 1323 {
1324 1324 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1325 1325 itxs_t *clean_me;
1326 1326
1327 1327 mutex_enter(&itxg->itxg_lock);
1328 1328 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1329 1329 mutex_exit(&itxg->itxg_lock);
1330 1330 return;
1331 1331 }
1332 1332 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1333 1333 ASSERT(itxg->itxg_txg != 0);
1334 1334 ASSERT(zilog->zl_clean_taskq != NULL);
1335 1335 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1336 1336 itxg->itxg_sod = 0;
1337 1337 clean_me = itxg->itxg_itxs;
1338 1338 itxg->itxg_itxs = NULL;
1339 1339 itxg->itxg_txg = 0;
1340 1340 mutex_exit(&itxg->itxg_lock);
1341 1341 /*
1342 1342 * Preferably start a task queue to free up the old itxs but
1343 1343 * if taskq_dispatch can't allocate resources to do that then
1344 1344 * free it in-line. This should be rare. Note, using TQ_SLEEP
1345 1345 * created a bad performance problem.
1346 1346 */
1347 1347 if (taskq_dispatch(zilog->zl_clean_taskq,
1348 1348 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1349 1349 zil_itxg_clean(clean_me);
1350 1350 }
1351 1351
1352 1352 /*
1353 1353 * Get the list of itxs to commit into zl_itx_commit_list.
1354 1354 */
1355 1355 static void
1356 1356 zil_get_commit_list(zilog_t *zilog)
1357 1357 {
1358 1358 uint64_t otxg, txg;
1359 1359 list_t *commit_list = &zilog->zl_itx_commit_list;
1360 1360 uint64_t push_sod = 0;
1361 1361
1362 1362 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1363 1363 otxg = ZILTEST_TXG;
1364 1364 else
1365 1365 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1366 1366
1367 1367 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1368 1368 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1369 1369
1370 1370 mutex_enter(&itxg->itxg_lock);
1371 1371 if (itxg->itxg_txg != txg) {
1372 1372 mutex_exit(&itxg->itxg_lock);
1373 1373 continue;
1374 1374 }
1375 1375
1376 1376 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1377 1377 push_sod += itxg->itxg_sod;
1378 1378 itxg->itxg_sod = 0;
1379 1379
1380 1380 mutex_exit(&itxg->itxg_lock);
1381 1381 }
1382 1382 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1383 1383 }
1384 1384
1385 1385 /*
1386 1386 * Move the async itxs for a specified object to commit into sync lists.
1387 1387 */
1388 1388 static void
1389 1389 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1390 1390 {
1391 1391 uint64_t otxg, txg;
1392 1392 itx_async_node_t *ian;
1393 1393 avl_tree_t *t;
1394 1394 avl_index_t where;
1395 1395
1396 1396 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1397 1397 otxg = ZILTEST_TXG;
1398 1398 else
1399 1399 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1400 1400
1401 1401 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1402 1402 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1403 1403
1404 1404 mutex_enter(&itxg->itxg_lock);
1405 1405 if (itxg->itxg_txg != txg) {
1406 1406 mutex_exit(&itxg->itxg_lock);
1407 1407 continue;
1408 1408 }
1409 1409
1410 1410 /*
1411 1411 * If a foid is specified then find that node and append its
1412 1412 * list. Otherwise walk the tree appending all the lists
1413 1413 * to the sync list. We add to the end rather than the
1414 1414 * beginning to ensure the create has happened.
1415 1415 */
1416 1416 t = &itxg->itxg_itxs->i_async_tree;
1417 1417 if (foid != 0) {
1418 1418 ian = avl_find(t, &foid, &where);
1419 1419 if (ian != NULL) {
1420 1420 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1421 1421 &ian->ia_list);
1422 1422 }
1423 1423 } else {
1424 1424 void *cookie = NULL;
1425 1425
1426 1426 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1427 1427 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1428 1428 &ian->ia_list);
1429 1429 list_destroy(&ian->ia_list);
1430 1430 kmem_free(ian, sizeof (itx_async_node_t));
1431 1431 }
1432 1432 }
1433 1433 mutex_exit(&itxg->itxg_lock);
1434 1434 }
1435 1435 }
1436 1436
1437 1437 static void
1438 1438 zil_commit_writer(zilog_t *zilog)
1439 1439 {
1440 1440 uint64_t txg;
1441 1441 itx_t *itx;
1442 1442 lwb_t *lwb;
1443 1443 spa_t *spa = zilog->zl_spa;
1444 1444 int error = 0;
1445 1445
1446 1446 ASSERT(zilog->zl_root_zio == NULL);
1447 1447
1448 1448 mutex_exit(&zilog->zl_lock);
1449 1449
1450 1450 zil_get_commit_list(zilog);
1451 1451
1452 1452 /*
1453 1453 * Return if there's nothing to commit before we dirty the fs by
1454 1454 * calling zil_create().
1455 1455 */
1456 1456 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1457 1457 mutex_enter(&zilog->zl_lock);
1458 1458 return;
1459 1459 }
1460 1460
1461 1461 if (zilog->zl_suspend) {
1462 1462 lwb = NULL;
1463 1463 } else {
1464 1464 lwb = list_tail(&zilog->zl_lwb_list);
1465 1465 if (lwb == NULL)
1466 1466 lwb = zil_create(zilog);
1467 1467 }
1468 1468
1469 1469 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1470 1470 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1471 1471 txg = itx->itx_lr.lrc_txg;
1472 1472 ASSERT(txg);
1473 1473
1474 1474 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1475 1475 lwb = zil_lwb_commit(zilog, itx, lwb);
1476 1476 list_remove(&zilog->zl_itx_commit_list, itx);
1477 1477 kmem_free(itx, offsetof(itx_t, itx_lr)
1478 1478 + itx->itx_lr.lrc_reclen);
1479 1479 }
1480 1480 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1481 1481
1482 1482 /* write the last block out */
1483 1483 if (lwb != NULL && lwb->lwb_zio != NULL)
1484 1484 lwb = zil_lwb_write_start(zilog, lwb);
1485 1485
1486 1486 zilog->zl_cur_used = 0;
1487 1487
1488 1488 /*
1489 1489 * Wait if necessary for the log blocks to be on stable storage.
1490 1490 */
1491 1491 if (zilog->zl_root_zio) {
1492 1492 error = zio_wait(zilog->zl_root_zio);
1493 1493 zilog->zl_root_zio = NULL;
1494 1494 zil_flush_vdevs(zilog);
1495 1495 }
1496 1496
1497 1497 if (error || lwb == NULL)
1498 1498 txg_wait_synced(zilog->zl_dmu_pool, 0);
1499 1499
1500 1500 mutex_enter(&zilog->zl_lock);
1501 1501
1502 1502 /*
1503 1503 * Remember the highest committed log sequence number for ztest.
1504 1504 * We only update this value when all the log writes succeeded,
1505 1505 * because ztest wants to ASSERT that it got the whole log chain.
1506 1506 */
1507 1507 if (error == 0 && lwb != NULL)
1508 1508 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1509 1509 }
1510 1510
1511 1511 /*
1512 1512 * Commit zfs transactions to stable storage.
1513 1513 * If foid is 0 push out all transactions, otherwise push only those
1514 1514 * for that object or might reference that object.
1515 1515 *
1516 1516 * itxs are committed in batches. In a heavily stressed zil there will be
1517 1517 * a commit writer thread who is writing out a bunch of itxs to the log
1518 1518 * for a set of committing threads (cthreads) in the same batch as the writer.
1519 1519 * Those cthreads are all waiting on the same cv for that batch.
1520 1520 *
1521 1521 * There will also be a different and growing batch of threads that are
1522 1522 * waiting to commit (qthreads). When the committing batch completes
1523 1523 * a transition occurs such that the cthreads exit and the qthreads become
1524 1524 * cthreads. One of the new cthreads becomes the writer thread for the
1525 1525 * batch. Any new threads arriving become new qthreads.
1526 1526 *
1527 1527 * Only 2 condition variables are needed and there's no transition
1528 1528 * between the two cvs needed. They just flip-flop between qthreads
1529 1529 * and cthreads.
1530 1530 *
1531 1531 * Using this scheme we can efficiently wakeup up only those threads
1532 1532 * that have been committed.
1533 1533 */
1534 1534 void
1535 1535 zil_commit(zilog_t *zilog, uint64_t foid)
1536 1536 {
1537 1537 uint64_t mybatch;
1538 1538
1539 1539 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1540 1540 return;
1541 1541
1542 1542 /* move the async itxs for the foid to the sync queues */
1543 1543 zil_async_to_sync(zilog, foid);
1544 1544
1545 1545 mutex_enter(&zilog->zl_lock);
1546 1546 mybatch = zilog->zl_next_batch;
1547 1547 while (zilog->zl_writer) {
1548 1548 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1549 1549 if (mybatch <= zilog->zl_com_batch) {
1550 1550 mutex_exit(&zilog->zl_lock);
1551 1551 return;
1552 1552 }
1553 1553 }
1554 1554
1555 1555 zilog->zl_next_batch++;
1556 1556 zilog->zl_writer = B_TRUE;
1557 1557 zil_commit_writer(zilog);
1558 1558 zilog->zl_com_batch = mybatch;
1559 1559 zilog->zl_writer = B_FALSE;
1560 1560 mutex_exit(&zilog->zl_lock);
1561 1561
1562 1562 /* wake up one thread to become the next writer */
1563 1563 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1564 1564
1565 1565 /* wake up all threads waiting for this batch to be committed */
1566 1566 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1567 1567 }
1568 1568
1569 1569 /*
1570 1570 * Called in syncing context to free committed log blocks and update log header.
1571 1571 */
1572 1572 void
1573 1573 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1574 1574 {
1575 1575 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1576 1576 uint64_t txg = dmu_tx_get_txg(tx);
1577 1577 spa_t *spa = zilog->zl_spa;
1578 1578 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1579 1579 lwb_t *lwb;
1580 1580
1581 1581 /*
1582 1582 * We don't zero out zl_destroy_txg, so make sure we don't try
1583 1583 * to destroy it twice.
1584 1584 */
1585 1585 if (spa_sync_pass(spa) != 1)
1586 1586 return;
1587 1587
1588 1588 mutex_enter(&zilog->zl_lock);
1589 1589
1590 1590 ASSERT(zilog->zl_stop_sync == 0);
1591 1591
1592 1592 if (*replayed_seq != 0) {
1593 1593 ASSERT(zh->zh_replay_seq < *replayed_seq);
1594 1594 zh->zh_replay_seq = *replayed_seq;
1595 1595 *replayed_seq = 0;
1596 1596 }
1597 1597
1598 1598 if (zilog->zl_destroy_txg == txg) {
1599 1599 blkptr_t blk = zh->zh_log;
1600 1600
1601 1601 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1602 1602
1603 1603 bzero(zh, sizeof (zil_header_t));
1604 1604 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1605 1605
1606 1606 if (zilog->zl_keep_first) {
1607 1607 /*
1608 1608 * If this block was part of log chain that couldn't
1609 1609 * be claimed because a device was missing during
1610 1610 * zil_claim(), but that device later returns,
1611 1611 * then this block could erroneously appear valid.
1612 1612 * To guard against this, assign a new GUID to the new
1613 1613 * log chain so it doesn't matter what blk points to.
1614 1614 */
1615 1615 zil_init_log_chain(zilog, &blk);
1616 1616 zh->zh_log = blk;
1617 1617 }
1618 1618 }
1619 1619
1620 1620 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1621 1621 zh->zh_log = lwb->lwb_blk;
1622 1622 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1623 1623 break;
1624 1624 list_remove(&zilog->zl_lwb_list, lwb);
1625 1625 zio_free_zil(spa, txg, &lwb->lwb_blk);
1626 1626 kmem_cache_free(zil_lwb_cache, lwb);
1627 1627
1628 1628 /*
1629 1629 * If we don't have anything left in the lwb list then
1630 1630 * we've had an allocation failure and we need to zero
1631 1631 * out the zil_header blkptr so that we don't end
1632 1632 * up freeing the same block twice.
1633 1633 */
1634 1634 if (list_head(&zilog->zl_lwb_list) == NULL)
1635 1635 BP_ZERO(&zh->zh_log);
1636 1636 }
1637 1637 mutex_exit(&zilog->zl_lock);
1638 1638 }
1639 1639
1640 1640 void
1641 1641 zil_init(void)
1642 1642 {
1643 1643 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1644 1644 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1645 1645 }
1646 1646
1647 1647 void
1648 1648 zil_fini(void)
1649 1649 {
1650 1650 kmem_cache_destroy(zil_lwb_cache);
1651 1651 }
1652 1652
1653 1653 void
1654 1654 zil_set_sync(zilog_t *zilog, uint64_t sync)
1655 1655 {
1656 1656 zilog->zl_sync = sync;
1657 1657 }
1658 1658
1659 1659 void
1660 1660 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1661 1661 {
1662 1662 zilog->zl_logbias = logbias;
1663 1663 }
1664 1664
1665 1665 zilog_t *
1666 1666 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1667 1667 {
1668 1668 zilog_t *zilog;
1669 1669
1670 1670 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1671 1671
1672 1672 zilog->zl_header = zh_phys;
1673 1673 zilog->zl_os = os;
1674 1674 zilog->zl_spa = dmu_objset_spa(os);
1675 1675 zilog->zl_dmu_pool = dmu_objset_pool(os);
1676 1676 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1677 1677 zilog->zl_logbias = dmu_objset_logbias(os);
1678 1678 zilog->zl_sync = dmu_objset_syncprop(os);
1679 1679 zilog->zl_next_batch = 1;
1680 1680
1681 1681 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1682 1682
1683 1683 for (int i = 0; i < TXG_SIZE; i++) {
1684 1684 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1685 1685 MUTEX_DEFAULT, NULL);
1686 1686 }
1687 1687
1688 1688 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1689 1689 offsetof(lwb_t, lwb_node));
1690 1690
1691 1691 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1692 1692 offsetof(itx_t, itx_node));
1693 1693
1694 1694 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1695 1695
1696 1696 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1697 1697 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1698 1698
1699 1699 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1700 1700 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1701 1701 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1702 1702 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1703 1703
1704 1704 return (zilog);
1705 1705 }
1706 1706
1707 1707 void
1708 1708 zil_free(zilog_t *zilog)
1709 1709 {
1710 1710 zilog->zl_stop_sync = 1;
1711 1711
1712 1712 ASSERT0(zilog->zl_suspend);
1713 1713 ASSERT0(zilog->zl_suspending);
1714 1714
1715 1715 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1716 1716 list_destroy(&zilog->zl_lwb_list);
1717 1717
1718 1718 avl_destroy(&zilog->zl_vdev_tree);
1719 1719 mutex_destroy(&zilog->zl_vdev_lock);
1720 1720
1721 1721 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1722 1722 list_destroy(&zilog->zl_itx_commit_list);
1723 1723
1724 1724 for (int i = 0; i < TXG_SIZE; i++) {
1725 1725 /*
1726 1726 * It's possible for an itx to be generated that doesn't dirty
1727 1727 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1728 1728 * callback to remove the entry. We remove those here.
1729 1729 *
1730 1730 * Also free up the ziltest itxs.
1731 1731 */
1732 1732 if (zilog->zl_itxg[i].itxg_itxs)
1733 1733 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1734 1734 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1735 1735 }
1736 1736
1737 1737 mutex_destroy(&zilog->zl_lock);
1738 1738
1739 1739 cv_destroy(&zilog->zl_cv_writer);
1740 1740 cv_destroy(&zilog->zl_cv_suspend);
1741 1741 cv_destroy(&zilog->zl_cv_batch[0]);
1742 1742 cv_destroy(&zilog->zl_cv_batch[1]);
1743 1743
1744 1744 kmem_free(zilog, sizeof (zilog_t));
1745 1745 }
1746 1746
1747 1747 /*
1748 1748 * Open an intent log.
1749 1749 */
1750 1750 zilog_t *
1751 1751 zil_open(objset_t *os, zil_get_data_t *get_data)
1752 1752 {
1753 1753 zilog_t *zilog = dmu_objset_zil(os);
1754 1754
1755 1755 ASSERT(zilog->zl_clean_taskq == NULL);
1756 1756 ASSERT(zilog->zl_get_data == NULL);
1757 1757 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1758 1758
1759 1759 zilog->zl_get_data = get_data;
1760 1760 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1761 1761 2, 2, TASKQ_PREPOPULATE);
1762 1762
1763 1763 return (zilog);
1764 1764 }
1765 1765
1766 1766 /*
1767 1767 * Close an intent log.
1768 1768 */
1769 1769 void
1770 1770 zil_close(zilog_t *zilog)
1771 1771 {
1772 1772 lwb_t *lwb;
1773 1773 uint64_t txg = 0;
1774 1774
1775 1775 zil_commit(zilog, 0); /* commit all itx */
1776 1776
1777 1777 /*
1778 1778 * The lwb_max_txg for the stubby lwb will reflect the last activity
1779 1779 * for the zil. After a txg_wait_synced() on the txg we know all the
1780 1780 * callbacks have occurred that may clean the zil. Only then can we
1781 1781 * destroy the zl_clean_taskq.
1782 1782 */
1783 1783 mutex_enter(&zilog->zl_lock);
1784 1784 lwb = list_tail(&zilog->zl_lwb_list);
1785 1785 if (lwb != NULL)
1786 1786 txg = lwb->lwb_max_txg;
1787 1787 mutex_exit(&zilog->zl_lock);
1788 1788 if (txg)
1789 1789 txg_wait_synced(zilog->zl_dmu_pool, txg);
1790 1790 ASSERT(!zilog_is_dirty(zilog));
1791 1791
1792 1792 taskq_destroy(zilog->zl_clean_taskq);
1793 1793 zilog->zl_clean_taskq = NULL;
1794 1794 zilog->zl_get_data = NULL;
1795 1795
1796 1796 /*
1797 1797 * We should have only one LWB left on the list; remove it now.
1798 1798 */
1799 1799 mutex_enter(&zilog->zl_lock);
1800 1800 lwb = list_head(&zilog->zl_lwb_list);
1801 1801 if (lwb != NULL) {
1802 1802 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1803 1803 list_remove(&zilog->zl_lwb_list, lwb);
1804 1804 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1805 1805 kmem_cache_free(zil_lwb_cache, lwb);
1806 1806 }
1807 1807 mutex_exit(&zilog->zl_lock);
1808 1808 }
1809 1809
1810 1810 static char *suspend_tag = "zil suspending";
1811 1811
1812 1812 /*
1813 1813 * Suspend an intent log. While in suspended mode, we still honor
1814 1814 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1815 1815 * On old version pools, we suspend the log briefly when taking a
1816 1816 * snapshot so that it will have an empty intent log.
1817 1817 *
1818 1818 * Long holds are not really intended to be used the way we do here --
1819 1819 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1820 1820 * could fail. Therefore we take pains to only put a long hold if it is
1821 1821 * actually necessary. Fortunately, it will only be necessary if the
1822 1822 * objset is currently mounted (or the ZVOL equivalent). In that case it
1823 1823 * will already have a long hold, so we are not really making things any worse.
1824 1824 *
1825 1825 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1826 1826 * zvol_state_t), and use their mechanism to prevent their hold from being
1827 1827 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1828 1828 * very little gain.
1829 1829 *
1830 1830 * if cookiep == NULL, this does both the suspend & resume.
1831 1831 * Otherwise, it returns with the dataset "long held", and the cookie
1832 1832 * should be passed into zil_resume().
1833 1833 */
1834 1834 int
1835 1835 zil_suspend(const char *osname, void **cookiep)
1836 1836 {
1837 1837 objset_t *os;
1838 1838 zilog_t *zilog;
1839 1839 const zil_header_t *zh;
1840 1840 int error;
1841 1841
1842 1842 error = dmu_objset_hold(osname, suspend_tag, &os);
1843 1843 if (error != 0)
1844 1844 return (error);
1845 1845 zilog = dmu_objset_zil(os);
1846 1846
1847 1847 mutex_enter(&zilog->zl_lock);
1848 1848 zh = zilog->zl_header;
1849 1849
1850 1850 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1851 1851 mutex_exit(&zilog->zl_lock);
1852 1852 dmu_objset_rele(os, suspend_tag);
1853 1853 return (SET_ERROR(EBUSY));
1854 1854 }
1855 1855
1856 1856 /*
1857 1857 * Don't put a long hold in the cases where we can avoid it. This
1858 1858 * is when there is no cookie so we are doing a suspend & resume
1859 1859 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1860 1860 * for the suspend because it's already suspended, or there's no ZIL.
1861 1861 */
1862 1862 if (cookiep == NULL && !zilog->zl_suspending &&
1863 1863 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1864 1864 mutex_exit(&zilog->zl_lock);
1865 1865 dmu_objset_rele(os, suspend_tag);
1866 1866 return (0);
1867 1867 }
1868 1868
1869 1869 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1870 1870 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1871 1871
1872 1872 zilog->zl_suspend++;
1873 1873
1874 1874 if (zilog->zl_suspend > 1) {
1875 1875 /*
1876 1876 * Someone else is already suspending it.
1877 1877 * Just wait for them to finish.
1878 1878 */
1879 1879
1880 1880 while (zilog->zl_suspending)
1881 1881 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1882 1882 mutex_exit(&zilog->zl_lock);
1883 1883
1884 1884 if (cookiep == NULL)
1885 1885 zil_resume(os);
1886 1886 else
1887 1887 *cookiep = os;
1888 1888 return (0);
1889 1889 }
1890 1890
1891 1891 /*
1892 1892 * If there is no pointer to an on-disk block, this ZIL must not
1893 1893 * be active (e.g. filesystem not mounted), so there's nothing
1894 1894 * to clean up.
1895 1895 */
1896 1896 if (BP_IS_HOLE(&zh->zh_log)) {
1897 1897 ASSERT(cookiep != NULL); /* fast path already handled */
1898 1898
1899 1899 *cookiep = os;
1900 1900 mutex_exit(&zilog->zl_lock);
1901 1901 return (0);
1902 1902 }
1903 1903
1904 1904 zilog->zl_suspending = B_TRUE;
1905 1905 mutex_exit(&zilog->zl_lock);
1906 1906
1907 1907 zil_commit(zilog, 0);
1908 1908
1909 1909 zil_destroy(zilog, B_FALSE);
1910 1910
1911 1911 mutex_enter(&zilog->zl_lock);
1912 1912 zilog->zl_suspending = B_FALSE;
1913 1913 cv_broadcast(&zilog->zl_cv_suspend);
1914 1914 mutex_exit(&zilog->zl_lock);
1915 1915
1916 1916 if (cookiep == NULL)
1917 1917 zil_resume(os);
1918 1918 else
1919 1919 *cookiep = os;
1920 1920 return (0);
1921 1921 }
1922 1922
1923 1923 void
1924 1924 zil_resume(void *cookie)
1925 1925 {
1926 1926 objset_t *os = cookie;
1927 1927 zilog_t *zilog = dmu_objset_zil(os);
1928 1928
1929 1929 mutex_enter(&zilog->zl_lock);
1930 1930 ASSERT(zilog->zl_suspend != 0);
1931 1931 zilog->zl_suspend--;
1932 1932 mutex_exit(&zilog->zl_lock);
1933 1933 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1934 1934 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1935 1935 }
1936 1936
1937 1937 typedef struct zil_replay_arg {
1938 1938 zil_replay_func_t **zr_replay;
1939 1939 void *zr_arg;
1940 1940 boolean_t zr_byteswap;
1941 1941 char *zr_lr;
1942 1942 } zil_replay_arg_t;
1943 1943
1944 1944 static int
1945 1945 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1946 1946 {
1947 1947 char name[MAXNAMELEN];
1948 1948
1949 1949 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1950 1950
1951 1951 dmu_objset_name(zilog->zl_os, name);
1952 1952
1953 1953 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1954 1954 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1955 1955 (u_longlong_t)lr->lrc_seq,
1956 1956 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1957 1957 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1958 1958
1959 1959 return (error);
1960 1960 }
1961 1961
1962 1962 static int
1963 1963 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1964 1964 {
1965 1965 zil_replay_arg_t *zr = zra;
1966 1966 const zil_header_t *zh = zilog->zl_header;
1967 1967 uint64_t reclen = lr->lrc_reclen;
1968 1968 uint64_t txtype = lr->lrc_txtype;
1969 1969 int error = 0;
1970 1970
1971 1971 zilog->zl_replaying_seq = lr->lrc_seq;
1972 1972
1973 1973 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1974 1974 return (0);
1975 1975
1976 1976 if (lr->lrc_txg < claim_txg) /* already committed */
1977 1977 return (0);
1978 1978
1979 1979 /* Strip case-insensitive bit, still present in log record */
1980 1980 txtype &= ~TX_CI;
1981 1981
1982 1982 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1983 1983 return (zil_replay_error(zilog, lr, EINVAL));
1984 1984
1985 1985 /*
1986 1986 * If this record type can be logged out of order, the object
1987 1987 * (lr_foid) may no longer exist. That's legitimate, not an error.
1988 1988 */
1989 1989 if (TX_OOO(txtype)) {
1990 1990 error = dmu_object_info(zilog->zl_os,
1991 1991 ((lr_ooo_t *)lr)->lr_foid, NULL);
1992 1992 if (error == ENOENT || error == EEXIST)
1993 1993 return (0);
1994 1994 }
1995 1995
1996 1996 /*
1997 1997 * Make a copy of the data so we can revise and extend it.
1998 1998 */
1999 1999 bcopy(lr, zr->zr_lr, reclen);
2000 2000
2001 2001 /*
2002 2002 * If this is a TX_WRITE with a blkptr, suck in the data.
2003 2003 */
2004 2004 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2005 2005 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2006 2006 zr->zr_lr + reclen);
2007 2007 if (error != 0)
2008 2008 return (zil_replay_error(zilog, lr, error));
2009 2009 }
2010 2010
2011 2011 /*
2012 2012 * The log block containing this lr may have been byteswapped
2013 2013 * so that we can easily examine common fields like lrc_txtype.
2014 2014 * However, the log is a mix of different record types, and only the
2015 2015 * replay vectors know how to byteswap their records. Therefore, if
2016 2016 * the lr was byteswapped, undo it before invoking the replay vector.
2017 2017 */
2018 2018 if (zr->zr_byteswap)
2019 2019 byteswap_uint64_array(zr->zr_lr, reclen);
2020 2020
2021 2021 /*
2022 2022 * We must now do two things atomically: replay this log record,
2023 2023 * and update the log header sequence number to reflect the fact that
2024 2024 * we did so. At the end of each replay function the sequence number
2025 2025 * is updated if we are in replay mode.
2026 2026 */
2027 2027 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2028 2028 if (error != 0) {
2029 2029 /*
2030 2030 * The DMU's dnode layer doesn't see removes until the txg
2031 2031 * commits, so a subsequent claim can spuriously fail with
2032 2032 * EEXIST. So if we receive any error we try syncing out
2033 2033 * any removes then retry the transaction. Note that we
2034 2034 * specify B_FALSE for byteswap now, so we don't do it twice.
2035 2035 */
2036 2036 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2037 2037 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2038 2038 if (error != 0)
2039 2039 return (zil_replay_error(zilog, lr, error));
2040 2040 }
2041 2041 return (0);
2042 2042 }
2043 2043
2044 2044 /* ARGSUSED */
2045 2045 static int
2046 2046 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2047 2047 {
2048 2048 zilog->zl_replay_blks++;
2049 2049
2050 2050 return (0);
2051 2051 }
2052 2052
2053 2053 /*
2054 2054 * If this dataset has a non-empty intent log, replay it and destroy it.
2055 2055 */
2056 2056 void
2057 2057 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2058 2058 {
2059 2059 zilog_t *zilog = dmu_objset_zil(os);
2060 2060 const zil_header_t *zh = zilog->zl_header;
2061 2061 zil_replay_arg_t zr;
2062 2062
2063 2063 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2064 2064 zil_destroy(zilog, B_TRUE);
2065 2065 return;
2066 2066 }
2067 2067
2068 2068 zr.zr_replay = replay_func;
2069 2069 zr.zr_arg = arg;
2070 2070 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2071 2071 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2072 2072
2073 2073 /*
2074 2074 * Wait for in-progress removes to sync before starting replay.
2075 2075 */
2076 2076 txg_wait_synced(zilog->zl_dmu_pool, 0);
2077 2077
2078 2078 zilog->zl_replay = B_TRUE;
2079 2079 zilog->zl_replay_time = ddi_get_lbolt();
2080 2080 ASSERT(zilog->zl_replay_blks == 0);
2081 2081 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2082 2082 zh->zh_claim_txg);
2083 2083 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2084 2084
2085 2085 zil_destroy(zilog, B_FALSE);
2086 2086 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2087 2087 zilog->zl_replay = B_FALSE;
2088 2088 }
2089 2089
2090 2090 boolean_t
2091 2091 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2092 2092 {
2093 2093 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2094 2094 return (B_TRUE);
2095 2095
2096 2096 if (zilog->zl_replay) {
2097 2097 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2098 2098 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2099 2099 zilog->zl_replaying_seq;
2100 2100 return (B_TRUE);
2101 2101 }
2102 2102
2103 2103 return (B_FALSE);
2104 2104 }
2105 2105
2106 2106 /* ARGSUSED */
2107 2107 int
2108 2108 zil_vdev_offline(const char *osname, void *arg)
2109 2109 {
2110 2110 int error;
2111 2111
2112 2112 error = zil_suspend(osname, NULL);
2113 2113 if (error != 0)
2114 2114 return (SET_ERROR(EEXIST));
2115 2115 return (0);
2116 2116 }
↓ open down ↓ |
1231 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX