Print this page
5269 zfs: zpool import slow
PORTING: this code relies on the property of taskq_wait to wait
until no more tasks are queued and no more tasks are active. As
we always queue new tasks from within other tasks, task_wait
reliably waits for the full recursion to finish, even though we
enqueue new tasks after taskq_wait has been called.
On platforms other than illumos, taskq_wait may not have this
property.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Reviewed by: George Wilson <george.wilson@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) 2011, 2014 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 arc_flags_t aflags = ARC_FLAG_WAIT;
185 185 arc_buf_t *abuf = NULL;
186 186 zbookmark_phys_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 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
224 224 bcopy(lr, dst, len);
225 225 *end = (char *)dst + len;
226 226 *nbp = zilc->zc_next_blk;
227 227 }
228 228 } else {
229 229 char *lr = abuf->b_data;
230 230 uint64_t size = BP_GET_LSIZE(bp);
231 231 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
232 232
233 233 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
234 234 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
235 235 (zilc->zc_nused > (size - sizeof (*zilc)))) {
236 236 error = SET_ERROR(ECKSUM);
237 237 } else {
238 238 ASSERT3U(zilc->zc_nused, <=,
239 239 SPA_OLD_MAXBLOCKSIZE);
240 240 bcopy(lr, dst, zilc->zc_nused);
241 241 *end = (char *)dst + zilc->zc_nused;
242 242 *nbp = zilc->zc_next_blk;
243 243 }
244 244 }
245 245
246 246 VERIFY(arc_buf_remove_ref(abuf, &abuf));
247 247 }
248 248
249 249 return (error);
250 250 }
251 251
252 252 /*
253 253 * Read a TX_WRITE log data block.
254 254 */
255 255 static int
256 256 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
257 257 {
258 258 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
259 259 const blkptr_t *bp = &lr->lr_blkptr;
260 260 arc_flags_t aflags = ARC_FLAG_WAIT;
261 261 arc_buf_t *abuf = NULL;
262 262 zbookmark_phys_t zb;
263 263 int error;
264 264
265 265 if (BP_IS_HOLE(bp)) {
266 266 if (wbuf != NULL)
267 267 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
268 268 return (0);
269 269 }
270 270
271 271 if (zilog->zl_header->zh_claim_txg == 0)
272 272 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
273 273
274 274 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
275 275 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
276 276
277 277 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
278 278 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
279 279
280 280 if (error == 0) {
281 281 if (wbuf != NULL)
282 282 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
283 283 (void) arc_buf_remove_ref(abuf, &abuf);
284 284 }
285 285
286 286 return (error);
287 287 }
288 288
289 289 /*
290 290 * Parse the intent log, and call parse_func for each valid record within.
291 291 */
292 292 int
293 293 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
294 294 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
295 295 {
296 296 const zil_header_t *zh = zilog->zl_header;
297 297 boolean_t claimed = !!zh->zh_claim_txg;
298 298 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
299 299 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
300 300 uint64_t max_blk_seq = 0;
301 301 uint64_t max_lr_seq = 0;
302 302 uint64_t blk_count = 0;
303 303 uint64_t lr_count = 0;
304 304 blkptr_t blk, next_blk;
305 305 char *lrbuf, *lrp;
306 306 int error = 0;
307 307
308 308 /*
309 309 * Old logs didn't record the maximum zh_claim_lr_seq.
310 310 */
311 311 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
312 312 claim_lr_seq = UINT64_MAX;
313 313
314 314 /*
315 315 * Starting at the block pointed to by zh_log we read the log chain.
316 316 * For each block in the chain we strongly check that block to
317 317 * ensure its validity. We stop when an invalid block is found.
318 318 * For each block pointer in the chain we call parse_blk_func().
319 319 * For each record in each valid block we call parse_lr_func().
320 320 * If the log has been claimed, stop if we encounter a sequence
321 321 * number greater than the highest claimed sequence number.
322 322 */
323 323 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
324 324 zil_bp_tree_init(zilog);
325 325
326 326 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
327 327 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
328 328 int reclen;
329 329 char *end;
330 330
331 331 if (blk_seq > claim_blk_seq)
332 332 break;
333 333 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
334 334 break;
335 335 ASSERT3U(max_blk_seq, <, blk_seq);
336 336 max_blk_seq = blk_seq;
337 337 blk_count++;
338 338
339 339 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
340 340 break;
341 341
342 342 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
343 343 if (error != 0)
344 344 break;
345 345
346 346 for (lrp = lrbuf; lrp < end; lrp += reclen) {
347 347 lr_t *lr = (lr_t *)lrp;
348 348 reclen = lr->lrc_reclen;
349 349 ASSERT3U(reclen, >=, sizeof (lr_t));
350 350 if (lr->lrc_seq > claim_lr_seq)
351 351 goto done;
352 352 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
353 353 goto done;
354 354 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
355 355 max_lr_seq = lr->lrc_seq;
356 356 lr_count++;
357 357 }
358 358 }
359 359 done:
360 360 zilog->zl_parse_error = error;
361 361 zilog->zl_parse_blk_seq = max_blk_seq;
362 362 zilog->zl_parse_lr_seq = max_lr_seq;
363 363 zilog->zl_parse_blk_count = blk_count;
364 364 zilog->zl_parse_lr_count = lr_count;
365 365
366 366 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
367 367 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
368 368
369 369 zil_bp_tree_fini(zilog);
370 370 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
371 371
372 372 return (error);
373 373 }
374 374
375 375 static int
376 376 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
377 377 {
378 378 /*
379 379 * Claim log block if not already committed and not already claimed.
380 380 * If tx == NULL, just verify that the block is claimable.
381 381 */
382 382 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
383 383 zil_bp_tree_add(zilog, bp) != 0)
384 384 return (0);
385 385
386 386 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
387 387 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
388 388 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
389 389 }
390 390
391 391 static int
392 392 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
393 393 {
394 394 lr_write_t *lr = (lr_write_t *)lrc;
395 395 int error;
396 396
397 397 if (lrc->lrc_txtype != TX_WRITE)
398 398 return (0);
399 399
400 400 /*
401 401 * If the block is not readable, don't claim it. This can happen
402 402 * in normal operation when a log block is written to disk before
403 403 * some of the dmu_sync() blocks it points to. In this case, the
404 404 * transaction cannot have been committed to anyone (we would have
405 405 * waited for all writes to be stable first), so it is semantically
406 406 * correct to declare this the end of the log.
407 407 */
408 408 if (lr->lr_blkptr.blk_birth >= first_txg &&
409 409 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
410 410 return (error);
411 411 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
412 412 }
413 413
414 414 /* ARGSUSED */
415 415 static int
416 416 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
417 417 {
418 418 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
419 419
420 420 return (0);
421 421 }
422 422
423 423 static int
424 424 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
425 425 {
426 426 lr_write_t *lr = (lr_write_t *)lrc;
427 427 blkptr_t *bp = &lr->lr_blkptr;
428 428
429 429 /*
430 430 * If we previously claimed it, we need to free it.
431 431 */
432 432 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
433 433 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
434 434 !BP_IS_HOLE(bp))
435 435 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
436 436
437 437 return (0);
438 438 }
439 439
440 440 static lwb_t *
441 441 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
442 442 {
443 443 lwb_t *lwb;
444 444
445 445 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
446 446 lwb->lwb_zilog = zilog;
447 447 lwb->lwb_blk = *bp;
448 448 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
449 449 lwb->lwb_max_txg = txg;
450 450 lwb->lwb_zio = NULL;
451 451 lwb->lwb_tx = NULL;
452 452 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
453 453 lwb->lwb_nused = sizeof (zil_chain_t);
454 454 lwb->lwb_sz = BP_GET_LSIZE(bp);
455 455 } else {
456 456 lwb->lwb_nused = 0;
457 457 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
458 458 }
459 459
460 460 mutex_enter(&zilog->zl_lock);
461 461 list_insert_tail(&zilog->zl_lwb_list, lwb);
462 462 mutex_exit(&zilog->zl_lock);
463 463
464 464 return (lwb);
465 465 }
466 466
467 467 /*
468 468 * Called when we create in-memory log transactions so that we know
469 469 * to cleanup the itxs at the end of spa_sync().
470 470 */
471 471 void
472 472 zilog_dirty(zilog_t *zilog, uint64_t txg)
473 473 {
474 474 dsl_pool_t *dp = zilog->zl_dmu_pool;
475 475 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
476 476
477 477 if (ds->ds_is_snapshot)
478 478 panic("dirtying snapshot!");
479 479
480 480 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
481 481 /* up the hold count until we can be written out */
482 482 dmu_buf_add_ref(ds->ds_dbuf, zilog);
483 483 }
484 484 }
485 485
486 486 boolean_t
487 487 zilog_is_dirty(zilog_t *zilog)
488 488 {
489 489 dsl_pool_t *dp = zilog->zl_dmu_pool;
490 490
491 491 for (int t = 0; t < TXG_SIZE; t++) {
492 492 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
493 493 return (B_TRUE);
494 494 }
495 495 return (B_FALSE);
496 496 }
497 497
498 498 /*
499 499 * Create an on-disk intent log.
500 500 */
501 501 static lwb_t *
502 502 zil_create(zilog_t *zilog)
503 503 {
504 504 const zil_header_t *zh = zilog->zl_header;
505 505 lwb_t *lwb = NULL;
506 506 uint64_t txg = 0;
507 507 dmu_tx_t *tx = NULL;
508 508 blkptr_t blk;
509 509 int error = 0;
510 510
511 511 /*
512 512 * Wait for any previous destroy to complete.
513 513 */
514 514 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
515 515
516 516 ASSERT(zh->zh_claim_txg == 0);
517 517 ASSERT(zh->zh_replay_seq == 0);
518 518
519 519 blk = zh->zh_log;
520 520
521 521 /*
522 522 * Allocate an initial log block if:
523 523 * - there isn't one already
524 524 * - the existing block is the wrong endianess
525 525 */
526 526 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
527 527 tx = dmu_tx_create(zilog->zl_os);
528 528 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
529 529 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
530 530 txg = dmu_tx_get_txg(tx);
531 531
532 532 if (!BP_IS_HOLE(&blk)) {
533 533 zio_free_zil(zilog->zl_spa, txg, &blk);
534 534 BP_ZERO(&blk);
535 535 }
536 536
537 537 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
538 538 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
539 539
540 540 if (error == 0)
541 541 zil_init_log_chain(zilog, &blk);
542 542 }
543 543
544 544 /*
545 545 * Allocate a log write buffer (lwb) for the first log block.
546 546 */
547 547 if (error == 0)
548 548 lwb = zil_alloc_lwb(zilog, &blk, txg);
549 549
550 550 /*
551 551 * If we just allocated the first log block, commit our transaction
552 552 * and wait for zil_sync() to stuff the block poiner into zh_log.
553 553 * (zh is part of the MOS, so we cannot modify it in open context.)
554 554 */
555 555 if (tx != NULL) {
556 556 dmu_tx_commit(tx);
557 557 txg_wait_synced(zilog->zl_dmu_pool, txg);
558 558 }
559 559
560 560 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
561 561
562 562 return (lwb);
563 563 }
564 564
565 565 /*
566 566 * In one tx, free all log blocks and clear the log header.
567 567 * If keep_first is set, then we're replaying a log with no content.
568 568 * We want to keep the first block, however, so that the first
569 569 * synchronous transaction doesn't require a txg_wait_synced()
570 570 * in zil_create(). We don't need to txg_wait_synced() here either
571 571 * when keep_first is set, because both zil_create() and zil_destroy()
572 572 * will wait for any in-progress destroys to complete.
573 573 */
574 574 void
575 575 zil_destroy(zilog_t *zilog, boolean_t keep_first)
576 576 {
577 577 const zil_header_t *zh = zilog->zl_header;
578 578 lwb_t *lwb;
579 579 dmu_tx_t *tx;
580 580 uint64_t txg;
581 581
582 582 /*
583 583 * Wait for any previous destroy to complete.
584 584 */
585 585 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
586 586
587 587 zilog->zl_old_header = *zh; /* debugging aid */
588 588
589 589 if (BP_IS_HOLE(&zh->zh_log))
590 590 return;
591 591
592 592 tx = dmu_tx_create(zilog->zl_os);
593 593 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
594 594 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
595 595 txg = dmu_tx_get_txg(tx);
596 596
597 597 mutex_enter(&zilog->zl_lock);
598 598
599 599 ASSERT3U(zilog->zl_destroy_txg, <, txg);
600 600 zilog->zl_destroy_txg = txg;
601 601 zilog->zl_keep_first = keep_first;
602 602
603 603 if (!list_is_empty(&zilog->zl_lwb_list)) {
604 604 ASSERT(zh->zh_claim_txg == 0);
605 605 VERIFY(!keep_first);
606 606 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
607 607 list_remove(&zilog->zl_lwb_list, lwb);
608 608 if (lwb->lwb_buf != NULL)
609 609 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
610 610 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
611 611 kmem_cache_free(zil_lwb_cache, lwb);
612 612 }
613 613 } else if (!keep_first) {
614 614 zil_destroy_sync(zilog, tx);
615 615 }
616 616 mutex_exit(&zilog->zl_lock);
617 617
618 618 dmu_tx_commit(tx);
619 619 }
|
↓ open down ↓ |
619 lines elided |
↑ open up ↑ |
620 620
621 621 void
622 622 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
623 623 {
624 624 ASSERT(list_is_empty(&zilog->zl_lwb_list));
625 625 (void) zil_parse(zilog, zil_free_log_block,
626 626 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
627 627 }
628 628
629 629 int
630 -zil_claim(const char *osname, void *txarg)
630 +zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
631 631 {
632 632 dmu_tx_t *tx = txarg;
633 633 uint64_t first_txg = dmu_tx_get_txg(tx);
634 634 zilog_t *zilog;
635 635 zil_header_t *zh;
636 636 objset_t *os;
637 637 int error;
638 638
639 - error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os);
639 + error = dmu_objset_own_obj(dp, ds->ds_object,
640 + DMU_OST_ANY, B_FALSE, FTAG, &os);
640 641 if (error != 0) {
641 642 /*
642 643 * EBUSY indicates that the objset is inconsistent, in which
643 644 * case it can not have a ZIL.
644 645 */
645 646 if (error != EBUSY) {
646 - cmn_err(CE_WARN, "can't open objset for %s, error %u",
647 - osname, error);
647 + cmn_err(CE_WARN, "can't open objset for %llu, error %u",
648 + (unsigned long long)ds->ds_object, error);
648 649 }
649 650 return (0);
650 651 }
651 652
652 653 zilog = dmu_objset_zil(os);
653 654 zh = zil_header_in_syncing_context(zilog);
654 655
655 656 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
656 657 if (!BP_IS_HOLE(&zh->zh_log))
657 658 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
658 659 BP_ZERO(&zh->zh_log);
659 660 dsl_dataset_dirty(dmu_objset_ds(os), tx);
660 661 dmu_objset_disown(os, FTAG);
661 662 return (0);
662 663 }
663 664
664 665 /*
665 666 * Claim all log blocks if we haven't already done so, and remember
666 667 * the highest claimed sequence number. This ensures that if we can
667 668 * read only part of the log now (e.g. due to a missing device),
668 669 * but we can read the entire log later, we will not try to replay
669 670 * or destroy beyond the last block we successfully claimed.
670 671 */
671 672 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
672 673 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
673 674 (void) zil_parse(zilog, zil_claim_log_block,
674 675 zil_claim_log_record, tx, first_txg);
675 676 zh->zh_claim_txg = first_txg;
676 677 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
677 678 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
678 679 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
679 680 zh->zh_flags |= ZIL_REPLAY_NEEDED;
680 681 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
681 682 dsl_dataset_dirty(dmu_objset_ds(os), tx);
682 683 }
683 684
|
↓ open down ↓ |
26 lines elided |
↑ open up ↑ |
684 685 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
685 686 dmu_objset_disown(os, FTAG);
686 687 return (0);
687 688 }
688 689
689 690 /*
690 691 * Check the log by walking the log chain.
691 692 * Checksum errors are ok as they indicate the end of the chain.
692 693 * Any other error (no device or read failure) returns an error.
693 694 */
695 +/* ARGSUSED */
694 696 int
695 -zil_check_log_chain(const char *osname, void *tx)
697 +zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
696 698 {
697 699 zilog_t *zilog;
698 700 objset_t *os;
699 701 blkptr_t *bp;
700 702 int error;
701 703
702 704 ASSERT(tx == NULL);
703 705
704 - error = dmu_objset_hold(osname, FTAG, &os);
706 + error = dmu_objset_from_ds(ds, &os);
705 707 if (error != 0) {
706 - cmn_err(CE_WARN, "can't open objset for %s", osname);
708 + cmn_err(CE_WARN, "can't open objset %llu, error %d",
709 + (unsigned long long)ds->ds_object, error);
707 710 return (0);
708 711 }
709 712
710 713 zilog = dmu_objset_zil(os);
711 714 bp = (blkptr_t *)&zilog->zl_header->zh_log;
712 715
713 716 /*
714 717 * Check the first block and determine if it's on a log device
715 718 * which may have been removed or faulted prior to loading this
716 719 * pool. If so, there's no point in checking the rest of the log
717 720 * as its content should have already been synced to the pool.
718 721 */
|
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
719 722 if (!BP_IS_HOLE(bp)) {
720 723 vdev_t *vd;
721 724 boolean_t valid = B_TRUE;
722 725
723 726 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
724 727 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
725 728 if (vd->vdev_islog && vdev_is_dead(vd))
726 729 valid = vdev_log_state_valid(vd);
727 730 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
728 731
729 - if (!valid) {
730 - dmu_objset_rele(os, FTAG);
732 + if (!valid)
731 733 return (0);
732 - }
733 734 }
734 735
735 736 /*
736 737 * Because tx == NULL, zil_claim_log_block() will not actually claim
737 738 * any blocks, but just determine whether it is possible to do so.
738 739 * In addition to checking the log chain, zil_claim_log_block()
739 740 * will invoke zio_claim() with a done func of spa_claim_notify(),
740 741 * which will update spa_max_claim_txg. See spa_load() for details.
741 742 */
742 743 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
743 744 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
744 745
745 - dmu_objset_rele(os, FTAG);
746 -
747 746 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
748 747 }
749 748
750 749 static int
751 750 zil_vdev_compare(const void *x1, const void *x2)
752 751 {
753 752 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
754 753 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
755 754
756 755 if (v1 < v2)
757 756 return (-1);
758 757 if (v1 > v2)
759 758 return (1);
760 759
761 760 return (0);
762 761 }
763 762
764 763 void
765 764 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
766 765 {
767 766 avl_tree_t *t = &zilog->zl_vdev_tree;
768 767 avl_index_t where;
769 768 zil_vdev_node_t *zv, zvsearch;
770 769 int ndvas = BP_GET_NDVAS(bp);
771 770 int i;
772 771
773 772 if (zfs_nocacheflush)
774 773 return;
775 774
776 775 ASSERT(zilog->zl_writer);
777 776
778 777 /*
779 778 * Even though we're zl_writer, we still need a lock because the
780 779 * zl_get_data() callbacks may have dmu_sync() done callbacks
781 780 * that will run concurrently.
782 781 */
783 782 mutex_enter(&zilog->zl_vdev_lock);
784 783 for (i = 0; i < ndvas; i++) {
785 784 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
786 785 if (avl_find(t, &zvsearch, &where) == NULL) {
787 786 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
788 787 zv->zv_vdev = zvsearch.zv_vdev;
789 788 avl_insert(t, zv, where);
790 789 }
791 790 }
792 791 mutex_exit(&zilog->zl_vdev_lock);
793 792 }
794 793
795 794 static void
796 795 zil_flush_vdevs(zilog_t *zilog)
797 796 {
798 797 spa_t *spa = zilog->zl_spa;
799 798 avl_tree_t *t = &zilog->zl_vdev_tree;
800 799 void *cookie = NULL;
801 800 zil_vdev_node_t *zv;
802 801 zio_t *zio;
803 802
804 803 ASSERT(zilog->zl_writer);
805 804
806 805 /*
807 806 * We don't need zl_vdev_lock here because we're the zl_writer,
808 807 * and all zl_get_data() callbacks are done.
809 808 */
810 809 if (avl_numnodes(t) == 0)
811 810 return;
812 811
813 812 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
814 813
815 814 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
816 815
817 816 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
818 817 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
819 818 if (vd != NULL)
820 819 zio_flush(zio, vd);
821 820 kmem_free(zv, sizeof (*zv));
822 821 }
823 822
824 823 /*
825 824 * Wait for all the flushes to complete. Not all devices actually
826 825 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
827 826 */
828 827 (void) zio_wait(zio);
829 828
830 829 spa_config_exit(spa, SCL_STATE, FTAG);
831 830 }
832 831
833 832 /*
834 833 * Function called when a log block write completes
835 834 */
836 835 static void
837 836 zil_lwb_write_done(zio_t *zio)
838 837 {
839 838 lwb_t *lwb = zio->io_private;
840 839 zilog_t *zilog = lwb->lwb_zilog;
841 840 dmu_tx_t *tx = lwb->lwb_tx;
842 841
843 842 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
844 843 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
845 844 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
846 845 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
847 846 ASSERT(!BP_IS_GANG(zio->io_bp));
848 847 ASSERT(!BP_IS_HOLE(zio->io_bp));
849 848 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
850 849
851 850 /*
852 851 * Ensure the lwb buffer pointer is cleared before releasing
853 852 * the txg. If we have had an allocation failure and
854 853 * the txg is waiting to sync then we want want zil_sync()
855 854 * to remove the lwb so that it's not picked up as the next new
856 855 * one in zil_commit_writer(). zil_sync() will only remove
857 856 * the lwb if lwb_buf is null.
858 857 */
859 858 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
860 859 mutex_enter(&zilog->zl_lock);
861 860 lwb->lwb_buf = NULL;
862 861 lwb->lwb_tx = NULL;
863 862 mutex_exit(&zilog->zl_lock);
864 863
865 864 /*
866 865 * Now that we've written this log block, we have a stable pointer
867 866 * to the next block in the chain, so it's OK to let the txg in
868 867 * which we allocated the next block sync.
869 868 */
870 869 dmu_tx_commit(tx);
871 870 }
872 871
873 872 /*
874 873 * Initialize the io for a log block.
875 874 */
876 875 static void
877 876 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
878 877 {
879 878 zbookmark_phys_t zb;
880 879
881 880 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
882 881 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
883 882 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
884 883
885 884 if (zilog->zl_root_zio == NULL) {
886 885 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
887 886 ZIO_FLAG_CANFAIL);
888 887 }
889 888 if (lwb->lwb_zio == NULL) {
890 889 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
891 890 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
892 891 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE,
893 892 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
894 893 }
895 894 }
896 895
897 896 /*
898 897 * Define a limited set of intent log block sizes.
899 898 *
900 899 * These must be a multiple of 4KB. Note only the amount used (again
901 900 * aligned to 4KB) actually gets written. However, we can't always just
902 901 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
903 902 */
904 903 uint64_t zil_block_buckets[] = {
905 904 4096, /* non TX_WRITE */
906 905 8192+4096, /* data base */
907 906 32*1024 + 4096, /* NFS writes */
908 907 UINT64_MAX
909 908 };
910 909
911 910 /*
912 911 * Use the slog as long as the logbias is 'latency' and the current commit size
913 912 * is less than the limit or the total list size is less than 2X the limit.
914 913 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
915 914 */
916 915 uint64_t zil_slog_limit = 1024 * 1024;
917 916 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
918 917 (((zilog)->zl_cur_used < zil_slog_limit) || \
919 918 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
920 919
921 920 /*
922 921 * Start a log block write and advance to the next log block.
923 922 * Calls are serialized.
924 923 */
925 924 static lwb_t *
926 925 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
927 926 {
928 927 lwb_t *nlwb = NULL;
929 928 zil_chain_t *zilc;
930 929 spa_t *spa = zilog->zl_spa;
931 930 blkptr_t *bp;
932 931 dmu_tx_t *tx;
933 932 uint64_t txg;
934 933 uint64_t zil_blksz, wsz;
935 934 int i, error;
936 935
937 936 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
938 937 zilc = (zil_chain_t *)lwb->lwb_buf;
939 938 bp = &zilc->zc_next_blk;
940 939 } else {
941 940 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
942 941 bp = &zilc->zc_next_blk;
943 942 }
944 943
945 944 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
946 945
947 946 /*
948 947 * Allocate the next block and save its address in this block
949 948 * before writing it in order to establish the log chain.
950 949 * Note that if the allocation of nlwb synced before we wrote
951 950 * the block that points at it (lwb), we'd leak it if we crashed.
952 951 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
953 952 * We dirty the dataset to ensure that zil_sync() will be called
954 953 * to clean up in the event of allocation failure or I/O failure.
955 954 */
956 955 tx = dmu_tx_create(zilog->zl_os);
957 956 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
958 957 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
959 958 txg = dmu_tx_get_txg(tx);
960 959
961 960 lwb->lwb_tx = tx;
962 961
963 962 /*
964 963 * Log blocks are pre-allocated. Here we select the size of the next
965 964 * block, based on size used in the last block.
966 965 * - first find the smallest bucket that will fit the block from a
967 966 * limited set of block sizes. This is because it's faster to write
968 967 * blocks allocated from the same metaslab as they are adjacent or
969 968 * close.
970 969 * - next find the maximum from the new suggested size and an array of
971 970 * previous sizes. This lessens a picket fence effect of wrongly
972 971 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
973 972 * requests.
974 973 *
975 974 * Note we only write what is used, but we can't just allocate
976 975 * the maximum block size because we can exhaust the available
977 976 * pool log space.
978 977 */
979 978 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
980 979 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
981 980 continue;
982 981 zil_blksz = zil_block_buckets[i];
983 982 if (zil_blksz == UINT64_MAX)
984 983 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
985 984 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
986 985 for (i = 0; i < ZIL_PREV_BLKS; i++)
987 986 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
988 987 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
989 988
990 989 BP_ZERO(bp);
991 990 /* pass the old blkptr in order to spread log blocks across devs */
992 991 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
993 992 USE_SLOG(zilog));
994 993 if (error == 0) {
995 994 ASSERT3U(bp->blk_birth, ==, txg);
996 995 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
997 996 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
998 997
999 998 /*
1000 999 * Allocate a new log write buffer (lwb).
1001 1000 */
1002 1001 nlwb = zil_alloc_lwb(zilog, bp, txg);
1003 1002
1004 1003 /* Record the block for later vdev flushing */
1005 1004 zil_add_block(zilog, &lwb->lwb_blk);
1006 1005 }
1007 1006
1008 1007 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1009 1008 /* For Slim ZIL only write what is used. */
1010 1009 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1011 1010 ASSERT3U(wsz, <=, lwb->lwb_sz);
1012 1011 zio_shrink(lwb->lwb_zio, wsz);
1013 1012
1014 1013 } else {
1015 1014 wsz = lwb->lwb_sz;
1016 1015 }
1017 1016
1018 1017 zilc->zc_pad = 0;
1019 1018 zilc->zc_nused = lwb->lwb_nused;
1020 1019 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1021 1020
1022 1021 /*
1023 1022 * clear unused data for security
1024 1023 */
1025 1024 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1026 1025
1027 1026 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1028 1027
1029 1028 /*
1030 1029 * If there was an allocation failure then nlwb will be null which
1031 1030 * forces a txg_wait_synced().
1032 1031 */
1033 1032 return (nlwb);
1034 1033 }
1035 1034
1036 1035 static lwb_t *
1037 1036 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1038 1037 {
1039 1038 lr_t *lrc = &itx->itx_lr; /* common log record */
1040 1039 lr_write_t *lrw = (lr_write_t *)lrc;
1041 1040 char *lr_buf;
1042 1041 uint64_t txg = lrc->lrc_txg;
1043 1042 uint64_t reclen = lrc->lrc_reclen;
1044 1043 uint64_t dlen = 0;
1045 1044
1046 1045 if (lwb == NULL)
1047 1046 return (NULL);
1048 1047
1049 1048 ASSERT(lwb->lwb_buf != NULL);
1050 1049 ASSERT(zilog_is_dirty(zilog) ||
1051 1050 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1052 1051
1053 1052 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1054 1053 dlen = P2ROUNDUP_TYPED(
1055 1054 lrw->lr_length, sizeof (uint64_t), uint64_t);
1056 1055
1057 1056 zilog->zl_cur_used += (reclen + dlen);
1058 1057
1059 1058 zil_lwb_write_init(zilog, lwb);
1060 1059
1061 1060 /*
1062 1061 * If this record won't fit in the current log block, start a new one.
1063 1062 */
1064 1063 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1065 1064 lwb = zil_lwb_write_start(zilog, lwb);
1066 1065 if (lwb == NULL)
1067 1066 return (NULL);
1068 1067 zil_lwb_write_init(zilog, lwb);
1069 1068 ASSERT(LWB_EMPTY(lwb));
1070 1069 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
1071 1070 txg_wait_synced(zilog->zl_dmu_pool, txg);
1072 1071 return (lwb);
1073 1072 }
1074 1073 }
1075 1074
1076 1075 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1077 1076 bcopy(lrc, lr_buf, reclen);
1078 1077 lrc = (lr_t *)lr_buf;
1079 1078 lrw = (lr_write_t *)lrc;
1080 1079
1081 1080 /*
1082 1081 * If it's a write, fetch the data or get its blkptr as appropriate.
1083 1082 */
1084 1083 if (lrc->lrc_txtype == TX_WRITE) {
1085 1084 if (txg > spa_freeze_txg(zilog->zl_spa))
1086 1085 txg_wait_synced(zilog->zl_dmu_pool, txg);
1087 1086 if (itx->itx_wr_state != WR_COPIED) {
1088 1087 char *dbuf;
1089 1088 int error;
1090 1089
1091 1090 if (dlen) {
1092 1091 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
1093 1092 dbuf = lr_buf + reclen;
1094 1093 lrw->lr_common.lrc_reclen += dlen;
1095 1094 } else {
1096 1095 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1097 1096 dbuf = NULL;
1098 1097 }
1099 1098 error = zilog->zl_get_data(
1100 1099 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1101 1100 if (error == EIO) {
1102 1101 txg_wait_synced(zilog->zl_dmu_pool, txg);
1103 1102 return (lwb);
1104 1103 }
1105 1104 if (error != 0) {
1106 1105 ASSERT(error == ENOENT || error == EEXIST ||
1107 1106 error == EALREADY);
1108 1107 return (lwb);
1109 1108 }
1110 1109 }
1111 1110 }
1112 1111
1113 1112 /*
1114 1113 * We're actually making an entry, so update lrc_seq to be the
1115 1114 * log record sequence number. Note that this is generally not
1116 1115 * equal to the itx sequence number because not all transactions
1117 1116 * are synchronous, and sometimes spa_sync() gets there first.
1118 1117 */
1119 1118 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1120 1119 lwb->lwb_nused += reclen + dlen;
1121 1120 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1122 1121 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1123 1122 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1124 1123
1125 1124 return (lwb);
1126 1125 }
1127 1126
1128 1127 itx_t *
1129 1128 zil_itx_create(uint64_t txtype, size_t lrsize)
1130 1129 {
1131 1130 itx_t *itx;
1132 1131
1133 1132 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1134 1133
1135 1134 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1136 1135 itx->itx_lr.lrc_txtype = txtype;
1137 1136 itx->itx_lr.lrc_reclen = lrsize;
1138 1137 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1139 1138 itx->itx_lr.lrc_seq = 0; /* defensive */
1140 1139 itx->itx_sync = B_TRUE; /* default is synchronous */
1141 1140
1142 1141 return (itx);
1143 1142 }
1144 1143
1145 1144 void
1146 1145 zil_itx_destroy(itx_t *itx)
1147 1146 {
1148 1147 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1149 1148 }
1150 1149
1151 1150 /*
1152 1151 * Free up the sync and async itxs. The itxs_t has already been detached
1153 1152 * so no locks are needed.
1154 1153 */
1155 1154 static void
1156 1155 zil_itxg_clean(itxs_t *itxs)
1157 1156 {
1158 1157 itx_t *itx;
1159 1158 list_t *list;
1160 1159 avl_tree_t *t;
1161 1160 void *cookie;
1162 1161 itx_async_node_t *ian;
1163 1162
1164 1163 list = &itxs->i_sync_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
1171 1170 cookie = NULL;
1172 1171 t = &itxs->i_async_tree;
1173 1172 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1174 1173 list = &ian->ia_list;
1175 1174 while ((itx = list_head(list)) != NULL) {
1176 1175 list_remove(list, itx);
1177 1176 kmem_free(itx, offsetof(itx_t, itx_lr) +
1178 1177 itx->itx_lr.lrc_reclen);
1179 1178 }
1180 1179 list_destroy(list);
1181 1180 kmem_free(ian, sizeof (itx_async_node_t));
1182 1181 }
1183 1182 avl_destroy(t);
1184 1183
1185 1184 kmem_free(itxs, sizeof (itxs_t));
1186 1185 }
1187 1186
1188 1187 static int
1189 1188 zil_aitx_compare(const void *x1, const void *x2)
1190 1189 {
1191 1190 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1192 1191 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1193 1192
1194 1193 if (o1 < o2)
1195 1194 return (-1);
1196 1195 if (o1 > o2)
1197 1196 return (1);
1198 1197
1199 1198 return (0);
1200 1199 }
1201 1200
1202 1201 /*
1203 1202 * Remove all async itx with the given oid.
1204 1203 */
1205 1204 static void
1206 1205 zil_remove_async(zilog_t *zilog, uint64_t oid)
1207 1206 {
1208 1207 uint64_t otxg, txg;
1209 1208 itx_async_node_t *ian;
1210 1209 avl_tree_t *t;
1211 1210 avl_index_t where;
1212 1211 list_t clean_list;
1213 1212 itx_t *itx;
1214 1213
1215 1214 ASSERT(oid != 0);
1216 1215 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1217 1216
1218 1217 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1219 1218 otxg = ZILTEST_TXG;
1220 1219 else
1221 1220 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1222 1221
1223 1222 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1224 1223 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1225 1224
1226 1225 mutex_enter(&itxg->itxg_lock);
1227 1226 if (itxg->itxg_txg != txg) {
1228 1227 mutex_exit(&itxg->itxg_lock);
1229 1228 continue;
1230 1229 }
1231 1230
1232 1231 /*
1233 1232 * Locate the object node and append its list.
1234 1233 */
1235 1234 t = &itxg->itxg_itxs->i_async_tree;
1236 1235 ian = avl_find(t, &oid, &where);
1237 1236 if (ian != NULL)
1238 1237 list_move_tail(&clean_list, &ian->ia_list);
1239 1238 mutex_exit(&itxg->itxg_lock);
1240 1239 }
1241 1240 while ((itx = list_head(&clean_list)) != NULL) {
1242 1241 list_remove(&clean_list, itx);
1243 1242 kmem_free(itx, offsetof(itx_t, itx_lr) +
1244 1243 itx->itx_lr.lrc_reclen);
1245 1244 }
1246 1245 list_destroy(&clean_list);
1247 1246 }
1248 1247
1249 1248 void
1250 1249 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1251 1250 {
1252 1251 uint64_t txg;
1253 1252 itxg_t *itxg;
1254 1253 itxs_t *itxs, *clean = NULL;
1255 1254
1256 1255 /*
1257 1256 * Object ids can be re-instantiated in the next txg so
1258 1257 * remove any async transactions to avoid future leaks.
1259 1258 * This can happen if a fsync occurs on the re-instantiated
1260 1259 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1261 1260 * the new file data and flushes a write record for the old object.
1262 1261 */
1263 1262 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1264 1263 zil_remove_async(zilog, itx->itx_oid);
1265 1264
1266 1265 /*
1267 1266 * Ensure the data of a renamed file is committed before the rename.
1268 1267 */
1269 1268 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1270 1269 zil_async_to_sync(zilog, itx->itx_oid);
1271 1270
1272 1271 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1273 1272 txg = ZILTEST_TXG;
1274 1273 else
1275 1274 txg = dmu_tx_get_txg(tx);
1276 1275
1277 1276 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1278 1277 mutex_enter(&itxg->itxg_lock);
1279 1278 itxs = itxg->itxg_itxs;
1280 1279 if (itxg->itxg_txg != txg) {
1281 1280 if (itxs != NULL) {
1282 1281 /*
1283 1282 * The zil_clean callback hasn't got around to cleaning
1284 1283 * this itxg. Save the itxs for release below.
1285 1284 * This should be rare.
1286 1285 */
1287 1286 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1288 1287 itxg->itxg_sod = 0;
1289 1288 clean = itxg->itxg_itxs;
1290 1289 }
1291 1290 ASSERT(itxg->itxg_sod == 0);
1292 1291 itxg->itxg_txg = txg;
1293 1292 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1294 1293
1295 1294 list_create(&itxs->i_sync_list, sizeof (itx_t),
1296 1295 offsetof(itx_t, itx_node));
1297 1296 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1298 1297 sizeof (itx_async_node_t),
1299 1298 offsetof(itx_async_node_t, ia_node));
1300 1299 }
1301 1300 if (itx->itx_sync) {
1302 1301 list_insert_tail(&itxs->i_sync_list, itx);
1303 1302 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod);
1304 1303 itxg->itxg_sod += itx->itx_sod;
1305 1304 } else {
1306 1305 avl_tree_t *t = &itxs->i_async_tree;
1307 1306 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1308 1307 itx_async_node_t *ian;
1309 1308 avl_index_t where;
1310 1309
1311 1310 ian = avl_find(t, &foid, &where);
1312 1311 if (ian == NULL) {
1313 1312 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1314 1313 list_create(&ian->ia_list, sizeof (itx_t),
1315 1314 offsetof(itx_t, itx_node));
1316 1315 ian->ia_foid = foid;
1317 1316 avl_insert(t, ian, where);
1318 1317 }
1319 1318 list_insert_tail(&ian->ia_list, itx);
1320 1319 }
1321 1320
1322 1321 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1323 1322 zilog_dirty(zilog, txg);
1324 1323 mutex_exit(&itxg->itxg_lock);
1325 1324
1326 1325 /* Release the old itxs now we've dropped the lock */
1327 1326 if (clean != NULL)
1328 1327 zil_itxg_clean(clean);
1329 1328 }
1330 1329
1331 1330 /*
1332 1331 * If there are any in-memory intent log transactions which have now been
1333 1332 * synced then start up a taskq to free them. We should only do this after we
1334 1333 * have written out the uberblocks (i.e. txg has been comitted) so that
1335 1334 * don't inadvertently clean out in-memory log records that would be required
1336 1335 * by zil_commit().
1337 1336 */
1338 1337 void
1339 1338 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1340 1339 {
1341 1340 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1342 1341 itxs_t *clean_me;
1343 1342
1344 1343 mutex_enter(&itxg->itxg_lock);
1345 1344 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1346 1345 mutex_exit(&itxg->itxg_lock);
1347 1346 return;
1348 1347 }
1349 1348 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1350 1349 ASSERT(itxg->itxg_txg != 0);
1351 1350 ASSERT(zilog->zl_clean_taskq != NULL);
1352 1351 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod);
1353 1352 itxg->itxg_sod = 0;
1354 1353 clean_me = itxg->itxg_itxs;
1355 1354 itxg->itxg_itxs = NULL;
1356 1355 itxg->itxg_txg = 0;
1357 1356 mutex_exit(&itxg->itxg_lock);
1358 1357 /*
1359 1358 * Preferably start a task queue to free up the old itxs but
1360 1359 * if taskq_dispatch can't allocate resources to do that then
1361 1360 * free it in-line. This should be rare. Note, using TQ_SLEEP
1362 1361 * created a bad performance problem.
1363 1362 */
1364 1363 if (taskq_dispatch(zilog->zl_clean_taskq,
1365 1364 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL)
1366 1365 zil_itxg_clean(clean_me);
1367 1366 }
1368 1367
1369 1368 /*
1370 1369 * Get the list of itxs to commit into zl_itx_commit_list.
1371 1370 */
1372 1371 static void
1373 1372 zil_get_commit_list(zilog_t *zilog)
1374 1373 {
1375 1374 uint64_t otxg, txg;
1376 1375 list_t *commit_list = &zilog->zl_itx_commit_list;
1377 1376 uint64_t push_sod = 0;
1378 1377
1379 1378 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1380 1379 otxg = ZILTEST_TXG;
1381 1380 else
1382 1381 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1383 1382
1384 1383 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1385 1384 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1386 1385
1387 1386 mutex_enter(&itxg->itxg_lock);
1388 1387 if (itxg->itxg_txg != txg) {
1389 1388 mutex_exit(&itxg->itxg_lock);
1390 1389 continue;
1391 1390 }
1392 1391
1393 1392 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1394 1393 push_sod += itxg->itxg_sod;
1395 1394 itxg->itxg_sod = 0;
1396 1395
1397 1396 mutex_exit(&itxg->itxg_lock);
1398 1397 }
1399 1398 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod);
1400 1399 }
1401 1400
1402 1401 /*
1403 1402 * Move the async itxs for a specified object to commit into sync lists.
1404 1403 */
1405 1404 static void
1406 1405 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1407 1406 {
1408 1407 uint64_t otxg, txg;
1409 1408 itx_async_node_t *ian;
1410 1409 avl_tree_t *t;
1411 1410 avl_index_t where;
1412 1411
1413 1412 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1414 1413 otxg = ZILTEST_TXG;
1415 1414 else
1416 1415 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1417 1416
1418 1417 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1419 1418 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1420 1419
1421 1420 mutex_enter(&itxg->itxg_lock);
1422 1421 if (itxg->itxg_txg != txg) {
1423 1422 mutex_exit(&itxg->itxg_lock);
1424 1423 continue;
1425 1424 }
1426 1425
1427 1426 /*
1428 1427 * If a foid is specified then find that node and append its
1429 1428 * list. Otherwise walk the tree appending all the lists
1430 1429 * to the sync list. We add to the end rather than the
1431 1430 * beginning to ensure the create has happened.
1432 1431 */
1433 1432 t = &itxg->itxg_itxs->i_async_tree;
1434 1433 if (foid != 0) {
1435 1434 ian = avl_find(t, &foid, &where);
1436 1435 if (ian != NULL) {
1437 1436 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1438 1437 &ian->ia_list);
1439 1438 }
1440 1439 } else {
1441 1440 void *cookie = NULL;
1442 1441
1443 1442 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1444 1443 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1445 1444 &ian->ia_list);
1446 1445 list_destroy(&ian->ia_list);
1447 1446 kmem_free(ian, sizeof (itx_async_node_t));
1448 1447 }
1449 1448 }
1450 1449 mutex_exit(&itxg->itxg_lock);
1451 1450 }
1452 1451 }
1453 1452
1454 1453 static void
1455 1454 zil_commit_writer(zilog_t *zilog)
1456 1455 {
1457 1456 uint64_t txg;
1458 1457 itx_t *itx;
1459 1458 lwb_t *lwb;
1460 1459 spa_t *spa = zilog->zl_spa;
1461 1460 int error = 0;
1462 1461
1463 1462 ASSERT(zilog->zl_root_zio == NULL);
1464 1463
1465 1464 mutex_exit(&zilog->zl_lock);
1466 1465
1467 1466 zil_get_commit_list(zilog);
1468 1467
1469 1468 /*
1470 1469 * Return if there's nothing to commit before we dirty the fs by
1471 1470 * calling zil_create().
1472 1471 */
1473 1472 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1474 1473 mutex_enter(&zilog->zl_lock);
1475 1474 return;
1476 1475 }
1477 1476
1478 1477 if (zilog->zl_suspend) {
1479 1478 lwb = NULL;
1480 1479 } else {
1481 1480 lwb = list_tail(&zilog->zl_lwb_list);
1482 1481 if (lwb == NULL)
1483 1482 lwb = zil_create(zilog);
1484 1483 }
1485 1484
1486 1485 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1487 1486 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1488 1487 txg = itx->itx_lr.lrc_txg;
1489 1488 ASSERT(txg);
1490 1489
1491 1490 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1492 1491 lwb = zil_lwb_commit(zilog, itx, lwb);
1493 1492 list_remove(&zilog->zl_itx_commit_list, itx);
1494 1493 kmem_free(itx, offsetof(itx_t, itx_lr)
1495 1494 + itx->itx_lr.lrc_reclen);
1496 1495 }
1497 1496 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1498 1497
1499 1498 /* write the last block out */
1500 1499 if (lwb != NULL && lwb->lwb_zio != NULL)
1501 1500 lwb = zil_lwb_write_start(zilog, lwb);
1502 1501
1503 1502 zilog->zl_cur_used = 0;
1504 1503
1505 1504 /*
1506 1505 * Wait if necessary for the log blocks to be on stable storage.
1507 1506 */
1508 1507 if (zilog->zl_root_zio) {
1509 1508 error = zio_wait(zilog->zl_root_zio);
1510 1509 zilog->zl_root_zio = NULL;
1511 1510 zil_flush_vdevs(zilog);
1512 1511 }
1513 1512
1514 1513 if (error || lwb == NULL)
1515 1514 txg_wait_synced(zilog->zl_dmu_pool, 0);
1516 1515
1517 1516 mutex_enter(&zilog->zl_lock);
1518 1517
1519 1518 /*
1520 1519 * Remember the highest committed log sequence number for ztest.
1521 1520 * We only update this value when all the log writes succeeded,
1522 1521 * because ztest wants to ASSERT that it got the whole log chain.
1523 1522 */
1524 1523 if (error == 0 && lwb != NULL)
1525 1524 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1526 1525 }
1527 1526
1528 1527 /*
1529 1528 * Commit zfs transactions to stable storage.
1530 1529 * If foid is 0 push out all transactions, otherwise push only those
1531 1530 * for that object or might reference that object.
1532 1531 *
1533 1532 * itxs are committed in batches. In a heavily stressed zil there will be
1534 1533 * a commit writer thread who is writing out a bunch of itxs to the log
1535 1534 * for a set of committing threads (cthreads) in the same batch as the writer.
1536 1535 * Those cthreads are all waiting on the same cv for that batch.
1537 1536 *
1538 1537 * There will also be a different and growing batch of threads that are
1539 1538 * waiting to commit (qthreads). When the committing batch completes
1540 1539 * a transition occurs such that the cthreads exit and the qthreads become
1541 1540 * cthreads. One of the new cthreads becomes the writer thread for the
1542 1541 * batch. Any new threads arriving become new qthreads.
1543 1542 *
1544 1543 * Only 2 condition variables are needed and there's no transition
1545 1544 * between the two cvs needed. They just flip-flop between qthreads
1546 1545 * and cthreads.
1547 1546 *
1548 1547 * Using this scheme we can efficiently wakeup up only those threads
1549 1548 * that have been committed.
1550 1549 */
1551 1550 void
1552 1551 zil_commit(zilog_t *zilog, uint64_t foid)
1553 1552 {
1554 1553 uint64_t mybatch;
1555 1554
1556 1555 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1557 1556 return;
1558 1557
1559 1558 /* move the async itxs for the foid to the sync queues */
1560 1559 zil_async_to_sync(zilog, foid);
1561 1560
1562 1561 mutex_enter(&zilog->zl_lock);
1563 1562 mybatch = zilog->zl_next_batch;
1564 1563 while (zilog->zl_writer) {
1565 1564 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1566 1565 if (mybatch <= zilog->zl_com_batch) {
1567 1566 mutex_exit(&zilog->zl_lock);
1568 1567 return;
1569 1568 }
1570 1569 }
1571 1570
1572 1571 zilog->zl_next_batch++;
1573 1572 zilog->zl_writer = B_TRUE;
1574 1573 zil_commit_writer(zilog);
1575 1574 zilog->zl_com_batch = mybatch;
1576 1575 zilog->zl_writer = B_FALSE;
1577 1576 mutex_exit(&zilog->zl_lock);
1578 1577
1579 1578 /* wake up one thread to become the next writer */
1580 1579 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1581 1580
1582 1581 /* wake up all threads waiting for this batch to be committed */
1583 1582 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1584 1583 }
1585 1584
1586 1585 /*
1587 1586 * Called in syncing context to free committed log blocks and update log header.
1588 1587 */
1589 1588 void
1590 1589 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1591 1590 {
1592 1591 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1593 1592 uint64_t txg = dmu_tx_get_txg(tx);
1594 1593 spa_t *spa = zilog->zl_spa;
1595 1594 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1596 1595 lwb_t *lwb;
1597 1596
1598 1597 /*
1599 1598 * We don't zero out zl_destroy_txg, so make sure we don't try
1600 1599 * to destroy it twice.
1601 1600 */
1602 1601 if (spa_sync_pass(spa) != 1)
1603 1602 return;
1604 1603
1605 1604 mutex_enter(&zilog->zl_lock);
1606 1605
1607 1606 ASSERT(zilog->zl_stop_sync == 0);
1608 1607
1609 1608 if (*replayed_seq != 0) {
1610 1609 ASSERT(zh->zh_replay_seq < *replayed_seq);
1611 1610 zh->zh_replay_seq = *replayed_seq;
1612 1611 *replayed_seq = 0;
1613 1612 }
1614 1613
1615 1614 if (zilog->zl_destroy_txg == txg) {
1616 1615 blkptr_t blk = zh->zh_log;
1617 1616
1618 1617 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1619 1618
1620 1619 bzero(zh, sizeof (zil_header_t));
1621 1620 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1622 1621
1623 1622 if (zilog->zl_keep_first) {
1624 1623 /*
1625 1624 * If this block was part of log chain that couldn't
1626 1625 * be claimed because a device was missing during
1627 1626 * zil_claim(), but that device later returns,
1628 1627 * then this block could erroneously appear valid.
1629 1628 * To guard against this, assign a new GUID to the new
1630 1629 * log chain so it doesn't matter what blk points to.
1631 1630 */
1632 1631 zil_init_log_chain(zilog, &blk);
1633 1632 zh->zh_log = blk;
1634 1633 }
1635 1634 }
1636 1635
1637 1636 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1638 1637 zh->zh_log = lwb->lwb_blk;
1639 1638 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1640 1639 break;
1641 1640 list_remove(&zilog->zl_lwb_list, lwb);
1642 1641 zio_free_zil(spa, txg, &lwb->lwb_blk);
1643 1642 kmem_cache_free(zil_lwb_cache, lwb);
1644 1643
1645 1644 /*
1646 1645 * If we don't have anything left in the lwb list then
1647 1646 * we've had an allocation failure and we need to zero
1648 1647 * out the zil_header blkptr so that we don't end
1649 1648 * up freeing the same block twice.
1650 1649 */
1651 1650 if (list_head(&zilog->zl_lwb_list) == NULL)
1652 1651 BP_ZERO(&zh->zh_log);
1653 1652 }
1654 1653 mutex_exit(&zilog->zl_lock);
1655 1654 }
1656 1655
1657 1656 void
1658 1657 zil_init(void)
1659 1658 {
1660 1659 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1661 1660 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1662 1661 }
1663 1662
1664 1663 void
1665 1664 zil_fini(void)
1666 1665 {
1667 1666 kmem_cache_destroy(zil_lwb_cache);
1668 1667 }
1669 1668
1670 1669 void
1671 1670 zil_set_sync(zilog_t *zilog, uint64_t sync)
1672 1671 {
1673 1672 zilog->zl_sync = sync;
1674 1673 }
1675 1674
1676 1675 void
1677 1676 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1678 1677 {
1679 1678 zilog->zl_logbias = logbias;
1680 1679 }
1681 1680
1682 1681 zilog_t *
1683 1682 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1684 1683 {
1685 1684 zilog_t *zilog;
1686 1685
1687 1686 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1688 1687
1689 1688 zilog->zl_header = zh_phys;
1690 1689 zilog->zl_os = os;
1691 1690 zilog->zl_spa = dmu_objset_spa(os);
1692 1691 zilog->zl_dmu_pool = dmu_objset_pool(os);
1693 1692 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1694 1693 zilog->zl_logbias = dmu_objset_logbias(os);
1695 1694 zilog->zl_sync = dmu_objset_syncprop(os);
1696 1695 zilog->zl_next_batch = 1;
1697 1696
1698 1697 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1699 1698
1700 1699 for (int i = 0; i < TXG_SIZE; i++) {
1701 1700 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1702 1701 MUTEX_DEFAULT, NULL);
1703 1702 }
1704 1703
1705 1704 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1706 1705 offsetof(lwb_t, lwb_node));
1707 1706
1708 1707 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1709 1708 offsetof(itx_t, itx_node));
1710 1709
1711 1710 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1712 1711
1713 1712 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1714 1713 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1715 1714
1716 1715 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1717 1716 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1718 1717 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1719 1718 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1720 1719
1721 1720 return (zilog);
1722 1721 }
1723 1722
1724 1723 void
1725 1724 zil_free(zilog_t *zilog)
1726 1725 {
1727 1726 zilog->zl_stop_sync = 1;
1728 1727
1729 1728 ASSERT0(zilog->zl_suspend);
1730 1729 ASSERT0(zilog->zl_suspending);
1731 1730
1732 1731 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1733 1732 list_destroy(&zilog->zl_lwb_list);
1734 1733
1735 1734 avl_destroy(&zilog->zl_vdev_tree);
1736 1735 mutex_destroy(&zilog->zl_vdev_lock);
1737 1736
1738 1737 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1739 1738 list_destroy(&zilog->zl_itx_commit_list);
1740 1739
1741 1740 for (int i = 0; i < TXG_SIZE; i++) {
1742 1741 /*
1743 1742 * It's possible for an itx to be generated that doesn't dirty
1744 1743 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1745 1744 * callback to remove the entry. We remove those here.
1746 1745 *
1747 1746 * Also free up the ziltest itxs.
1748 1747 */
1749 1748 if (zilog->zl_itxg[i].itxg_itxs)
1750 1749 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1751 1750 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1752 1751 }
1753 1752
1754 1753 mutex_destroy(&zilog->zl_lock);
1755 1754
1756 1755 cv_destroy(&zilog->zl_cv_writer);
1757 1756 cv_destroy(&zilog->zl_cv_suspend);
1758 1757 cv_destroy(&zilog->zl_cv_batch[0]);
1759 1758 cv_destroy(&zilog->zl_cv_batch[1]);
1760 1759
1761 1760 kmem_free(zilog, sizeof (zilog_t));
1762 1761 }
1763 1762
1764 1763 /*
1765 1764 * Open an intent log.
1766 1765 */
1767 1766 zilog_t *
1768 1767 zil_open(objset_t *os, zil_get_data_t *get_data)
1769 1768 {
1770 1769 zilog_t *zilog = dmu_objset_zil(os);
1771 1770
1772 1771 ASSERT(zilog->zl_clean_taskq == NULL);
1773 1772 ASSERT(zilog->zl_get_data == NULL);
1774 1773 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1775 1774
1776 1775 zilog->zl_get_data = get_data;
1777 1776 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1778 1777 2, 2, TASKQ_PREPOPULATE);
1779 1778
1780 1779 return (zilog);
1781 1780 }
1782 1781
1783 1782 /*
1784 1783 * Close an intent log.
1785 1784 */
1786 1785 void
1787 1786 zil_close(zilog_t *zilog)
1788 1787 {
1789 1788 lwb_t *lwb;
1790 1789 uint64_t txg = 0;
1791 1790
1792 1791 zil_commit(zilog, 0); /* commit all itx */
1793 1792
1794 1793 /*
1795 1794 * The lwb_max_txg for the stubby lwb will reflect the last activity
1796 1795 * for the zil. After a txg_wait_synced() on the txg we know all the
1797 1796 * callbacks have occurred that may clean the zil. Only then can we
1798 1797 * destroy the zl_clean_taskq.
1799 1798 */
1800 1799 mutex_enter(&zilog->zl_lock);
1801 1800 lwb = list_tail(&zilog->zl_lwb_list);
1802 1801 if (lwb != NULL)
1803 1802 txg = lwb->lwb_max_txg;
1804 1803 mutex_exit(&zilog->zl_lock);
1805 1804 if (txg)
1806 1805 txg_wait_synced(zilog->zl_dmu_pool, txg);
1807 1806 ASSERT(!zilog_is_dirty(zilog));
1808 1807
1809 1808 taskq_destroy(zilog->zl_clean_taskq);
1810 1809 zilog->zl_clean_taskq = NULL;
1811 1810 zilog->zl_get_data = NULL;
1812 1811
1813 1812 /*
1814 1813 * We should have only one LWB left on the list; remove it now.
1815 1814 */
1816 1815 mutex_enter(&zilog->zl_lock);
1817 1816 lwb = list_head(&zilog->zl_lwb_list);
1818 1817 if (lwb != NULL) {
1819 1818 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1820 1819 list_remove(&zilog->zl_lwb_list, lwb);
1821 1820 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1822 1821 kmem_cache_free(zil_lwb_cache, lwb);
1823 1822 }
1824 1823 mutex_exit(&zilog->zl_lock);
1825 1824 }
1826 1825
1827 1826 static char *suspend_tag = "zil suspending";
1828 1827
1829 1828 /*
1830 1829 * Suspend an intent log. While in suspended mode, we still honor
1831 1830 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1832 1831 * On old version pools, we suspend the log briefly when taking a
1833 1832 * snapshot so that it will have an empty intent log.
1834 1833 *
1835 1834 * Long holds are not really intended to be used the way we do here --
1836 1835 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1837 1836 * could fail. Therefore we take pains to only put a long hold if it is
1838 1837 * actually necessary. Fortunately, it will only be necessary if the
1839 1838 * objset is currently mounted (or the ZVOL equivalent). In that case it
1840 1839 * will already have a long hold, so we are not really making things any worse.
1841 1840 *
1842 1841 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1843 1842 * zvol_state_t), and use their mechanism to prevent their hold from being
1844 1843 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1845 1844 * very little gain.
1846 1845 *
1847 1846 * if cookiep == NULL, this does both the suspend & resume.
1848 1847 * Otherwise, it returns with the dataset "long held", and the cookie
1849 1848 * should be passed into zil_resume().
1850 1849 */
1851 1850 int
1852 1851 zil_suspend(const char *osname, void **cookiep)
1853 1852 {
1854 1853 objset_t *os;
1855 1854 zilog_t *zilog;
1856 1855 const zil_header_t *zh;
1857 1856 int error;
1858 1857
1859 1858 error = dmu_objset_hold(osname, suspend_tag, &os);
1860 1859 if (error != 0)
1861 1860 return (error);
1862 1861 zilog = dmu_objset_zil(os);
1863 1862
1864 1863 mutex_enter(&zilog->zl_lock);
1865 1864 zh = zilog->zl_header;
1866 1865
1867 1866 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1868 1867 mutex_exit(&zilog->zl_lock);
1869 1868 dmu_objset_rele(os, suspend_tag);
1870 1869 return (SET_ERROR(EBUSY));
1871 1870 }
1872 1871
1873 1872 /*
1874 1873 * Don't put a long hold in the cases where we can avoid it. This
1875 1874 * is when there is no cookie so we are doing a suspend & resume
1876 1875 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1877 1876 * for the suspend because it's already suspended, or there's no ZIL.
1878 1877 */
1879 1878 if (cookiep == NULL && !zilog->zl_suspending &&
1880 1879 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1881 1880 mutex_exit(&zilog->zl_lock);
1882 1881 dmu_objset_rele(os, suspend_tag);
1883 1882 return (0);
1884 1883 }
1885 1884
1886 1885 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1887 1886 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1888 1887
1889 1888 zilog->zl_suspend++;
1890 1889
1891 1890 if (zilog->zl_suspend > 1) {
1892 1891 /*
1893 1892 * Someone else is already suspending it.
1894 1893 * Just wait for them to finish.
1895 1894 */
1896 1895
1897 1896 while (zilog->zl_suspending)
1898 1897 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1899 1898 mutex_exit(&zilog->zl_lock);
1900 1899
1901 1900 if (cookiep == NULL)
1902 1901 zil_resume(os);
1903 1902 else
1904 1903 *cookiep = os;
1905 1904 return (0);
1906 1905 }
1907 1906
1908 1907 /*
1909 1908 * If there is no pointer to an on-disk block, this ZIL must not
1910 1909 * be active (e.g. filesystem not mounted), so there's nothing
1911 1910 * to clean up.
1912 1911 */
1913 1912 if (BP_IS_HOLE(&zh->zh_log)) {
1914 1913 ASSERT(cookiep != NULL); /* fast path already handled */
1915 1914
1916 1915 *cookiep = os;
1917 1916 mutex_exit(&zilog->zl_lock);
1918 1917 return (0);
1919 1918 }
1920 1919
1921 1920 zilog->zl_suspending = B_TRUE;
1922 1921 mutex_exit(&zilog->zl_lock);
1923 1922
1924 1923 zil_commit(zilog, 0);
1925 1924
1926 1925 zil_destroy(zilog, B_FALSE);
1927 1926
1928 1927 mutex_enter(&zilog->zl_lock);
1929 1928 zilog->zl_suspending = B_FALSE;
1930 1929 cv_broadcast(&zilog->zl_cv_suspend);
1931 1930 mutex_exit(&zilog->zl_lock);
1932 1931
1933 1932 if (cookiep == NULL)
1934 1933 zil_resume(os);
1935 1934 else
1936 1935 *cookiep = os;
1937 1936 return (0);
1938 1937 }
1939 1938
1940 1939 void
1941 1940 zil_resume(void *cookie)
1942 1941 {
1943 1942 objset_t *os = cookie;
1944 1943 zilog_t *zilog = dmu_objset_zil(os);
1945 1944
1946 1945 mutex_enter(&zilog->zl_lock);
1947 1946 ASSERT(zilog->zl_suspend != 0);
1948 1947 zilog->zl_suspend--;
1949 1948 mutex_exit(&zilog->zl_lock);
1950 1949 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
1951 1950 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
1952 1951 }
1953 1952
1954 1953 typedef struct zil_replay_arg {
1955 1954 zil_replay_func_t **zr_replay;
1956 1955 void *zr_arg;
1957 1956 boolean_t zr_byteswap;
1958 1957 char *zr_lr;
1959 1958 } zil_replay_arg_t;
1960 1959
1961 1960 static int
1962 1961 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1963 1962 {
1964 1963 char name[MAXNAMELEN];
1965 1964
1966 1965 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1967 1966
1968 1967 dmu_objset_name(zilog->zl_os, name);
1969 1968
1970 1969 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1971 1970 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1972 1971 (u_longlong_t)lr->lrc_seq,
1973 1972 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1974 1973 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1975 1974
1976 1975 return (error);
1977 1976 }
1978 1977
1979 1978 static int
1980 1979 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1981 1980 {
1982 1981 zil_replay_arg_t *zr = zra;
1983 1982 const zil_header_t *zh = zilog->zl_header;
1984 1983 uint64_t reclen = lr->lrc_reclen;
1985 1984 uint64_t txtype = lr->lrc_txtype;
1986 1985 int error = 0;
1987 1986
1988 1987 zilog->zl_replaying_seq = lr->lrc_seq;
1989 1988
1990 1989 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1991 1990 return (0);
1992 1991
1993 1992 if (lr->lrc_txg < claim_txg) /* already committed */
1994 1993 return (0);
1995 1994
1996 1995 /* Strip case-insensitive bit, still present in log record */
1997 1996 txtype &= ~TX_CI;
1998 1997
1999 1998 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2000 1999 return (zil_replay_error(zilog, lr, EINVAL));
2001 2000
2002 2001 /*
2003 2002 * If this record type can be logged out of order, the object
2004 2003 * (lr_foid) may no longer exist. That's legitimate, not an error.
2005 2004 */
2006 2005 if (TX_OOO(txtype)) {
2007 2006 error = dmu_object_info(zilog->zl_os,
2008 2007 ((lr_ooo_t *)lr)->lr_foid, NULL);
2009 2008 if (error == ENOENT || error == EEXIST)
2010 2009 return (0);
2011 2010 }
2012 2011
2013 2012 /*
2014 2013 * Make a copy of the data so we can revise and extend it.
2015 2014 */
2016 2015 bcopy(lr, zr->zr_lr, reclen);
2017 2016
2018 2017 /*
2019 2018 * If this is a TX_WRITE with a blkptr, suck in the data.
2020 2019 */
2021 2020 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2022 2021 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2023 2022 zr->zr_lr + reclen);
2024 2023 if (error != 0)
2025 2024 return (zil_replay_error(zilog, lr, error));
2026 2025 }
2027 2026
2028 2027 /*
2029 2028 * The log block containing this lr may have been byteswapped
2030 2029 * so that we can easily examine common fields like lrc_txtype.
2031 2030 * However, the log is a mix of different record types, and only the
2032 2031 * replay vectors know how to byteswap their records. Therefore, if
2033 2032 * the lr was byteswapped, undo it before invoking the replay vector.
2034 2033 */
2035 2034 if (zr->zr_byteswap)
2036 2035 byteswap_uint64_array(zr->zr_lr, reclen);
2037 2036
2038 2037 /*
2039 2038 * We must now do two things atomically: replay this log record,
2040 2039 * and update the log header sequence number to reflect the fact that
2041 2040 * we did so. At the end of each replay function the sequence number
2042 2041 * is updated if we are in replay mode.
2043 2042 */
2044 2043 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2045 2044 if (error != 0) {
2046 2045 /*
2047 2046 * The DMU's dnode layer doesn't see removes until the txg
2048 2047 * commits, so a subsequent claim can spuriously fail with
2049 2048 * EEXIST. So if we receive any error we try syncing out
2050 2049 * any removes then retry the transaction. Note that we
2051 2050 * specify B_FALSE for byteswap now, so we don't do it twice.
2052 2051 */
2053 2052 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2054 2053 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2055 2054 if (error != 0)
2056 2055 return (zil_replay_error(zilog, lr, error));
2057 2056 }
2058 2057 return (0);
2059 2058 }
2060 2059
2061 2060 /* ARGSUSED */
2062 2061 static int
2063 2062 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2064 2063 {
2065 2064 zilog->zl_replay_blks++;
2066 2065
2067 2066 return (0);
2068 2067 }
2069 2068
2070 2069 /*
2071 2070 * If this dataset has a non-empty intent log, replay it and destroy it.
2072 2071 */
2073 2072 void
2074 2073 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2075 2074 {
2076 2075 zilog_t *zilog = dmu_objset_zil(os);
2077 2076 const zil_header_t *zh = zilog->zl_header;
2078 2077 zil_replay_arg_t zr;
2079 2078
2080 2079 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2081 2080 zil_destroy(zilog, B_TRUE);
2082 2081 return;
2083 2082 }
2084 2083
2085 2084 zr.zr_replay = replay_func;
2086 2085 zr.zr_arg = arg;
2087 2086 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2088 2087 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2089 2088
2090 2089 /*
2091 2090 * Wait for in-progress removes to sync before starting replay.
2092 2091 */
2093 2092 txg_wait_synced(zilog->zl_dmu_pool, 0);
2094 2093
2095 2094 zilog->zl_replay = B_TRUE;
2096 2095 zilog->zl_replay_time = ddi_get_lbolt();
2097 2096 ASSERT(zilog->zl_replay_blks == 0);
2098 2097 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2099 2098 zh->zh_claim_txg);
2100 2099 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2101 2100
2102 2101 zil_destroy(zilog, B_FALSE);
2103 2102 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2104 2103 zilog->zl_replay = B_FALSE;
2105 2104 }
2106 2105
2107 2106 boolean_t
2108 2107 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2109 2108 {
2110 2109 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2111 2110 return (B_TRUE);
2112 2111
2113 2112 if (zilog->zl_replay) {
2114 2113 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2115 2114 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2116 2115 zilog->zl_replaying_seq;
2117 2116 return (B_TRUE);
2118 2117 }
2119 2118
2120 2119 return (B_FALSE);
2121 2120 }
2122 2121
2123 2122 /* ARGSUSED */
2124 2123 int
2125 2124 zil_vdev_offline(const char *osname, void *arg)
2126 2125 {
2127 2126 int error;
2128 2127
2129 2128 error = zil_suspend(osname, NULL);
2130 2129 if (error != 0)
2131 2130 return (SET_ERROR(EEXIST));
2132 2131 return (0);
2133 2132 }
|
↓ open down ↓ |
1377 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX