1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 #include <sys/zfs_context.h>
27 #include <sys/spa.h>
28 #include <sys/dmu.h>
29 #include <sys/zap.h>
30 #include <sys/arc.h>
31 #include <sys/stat.h>
32 #include <sys/resource.h>
33 #include <sys/zil.h>
34 #include <sys/zil_impl.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/vdev.h>
37 #include <sys/dmu_tx.h>
38
39 /*
40 * The zfs intent log (ZIL) saves transaction records of system calls
41 * that change the file system in memory with enough information
42 * to be able to replay them. These are stored in memory until
43 * either the DMU transaction group (txg) commits them to the stable pool
44 * and they can be discarded, or they are flushed to the stable log
45 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
46 * requirement. In the event of a panic or power fail then those log
47 * records (transactions) are replayed.
48 *
49 * There is one ZIL per file system. Its on-disk (pool) format consists
50 * of 3 parts:
51 *
52 * - ZIL header
53 * - ZIL blocks
54 * - ZIL records
55 *
56 * A log record holds a system call transaction. Log blocks can
57 * hold many log records and the blocks are chained together.
58 * Each ZIL block contains a block pointer (blkptr_t) to the next
59 * ZIL block in the chain. The ZIL header points to the first
60 * block in the chain. Note there is not a fixed place in the pool
61 * to hold blocks. They are dynamically allocated and freed as
62 * needed from the blocks available. Figure X shows the ZIL structure:
63 */
64
65 /*
66 * This global ZIL switch affects all pools
67 */
68 int zil_disable = 0; /* disable intent logging */
69
70 /*
71 * Tunable parameter for debugging or performance analysis. Setting
72 * zfs_nocacheflush will cause corruption on power loss if a volatile
73 * out-of-order write cache is enabled.
74 */
75 boolean_t zfs_nocacheflush = B_FALSE;
76
77 static kmem_cache_t *zil_lwb_cache;
78
79 static boolean_t zil_empty(zilog_t *zilog);
80
81 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
82 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
83
84
85 static int
zil_bp_compare(const void * x1,const void * x2)86 zil_bp_compare(const void *x1, const void *x2)
87 {
88 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
89 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
90
91 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
92 return (-1);
93 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
94 return (1);
95
96 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
97 return (-1);
98 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
99 return (1);
100
101 return (0);
102 }
103
104 static void
zil_bp_tree_init(zilog_t * zilog)105 zil_bp_tree_init(zilog_t *zilog)
106 {
107 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
108 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
109 }
110
111 static void
zil_bp_tree_fini(zilog_t * zilog)112 zil_bp_tree_fini(zilog_t *zilog)
113 {
114 avl_tree_t *t = &zilog->zl_bp_tree;
115 zil_bp_node_t *zn;
116 void *cookie = NULL;
117
118 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
119 kmem_free(zn, sizeof (zil_bp_node_t));
120
121 avl_destroy(t);
122 }
123
124 int
zil_bp_tree_add(zilog_t * zilog,const blkptr_t * bp)125 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
126 {
127 avl_tree_t *t = &zilog->zl_bp_tree;
128 const dva_t *dva = BP_IDENTITY(bp);
129 zil_bp_node_t *zn;
130 avl_index_t where;
131
132 if (avl_find(t, dva, &where) != NULL)
133 return (EEXIST);
134
135 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
136 zn->zn_dva = *dva;
137 avl_insert(t, zn, where);
138
139 return (0);
140 }
141
142 static zil_header_t *
zil_header_in_syncing_context(zilog_t * zilog)143 zil_header_in_syncing_context(zilog_t *zilog)
144 {
145 return ((zil_header_t *)zilog->zl_header);
146 }
147
148 static void
zil_init_log_chain(zilog_t * zilog,blkptr_t * bp)149 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
150 {
151 zio_cksum_t *zc = &bp->blk_cksum;
152
153 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
154 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
155 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
156 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
157 }
158
159 /*
160 * Read a log block and make sure it's valid.
161 */
162 static int
zil_read_log_block(zilog_t * zilog,const blkptr_t * bp,blkptr_t * nbp,void * dst,char ** end)163 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
164 char **end)
165 {
166 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
167 uint32_t aflags = ARC_WAIT;
168 arc_buf_t *abuf = NULL;
169 zbookmark_t zb;
170 int error;
171
172 if (zilog->zl_header->zh_claim_txg == 0)
173 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
174
175 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
176 zio_flags |= ZIO_FLAG_SPECULATIVE;
177
178 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
179 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
180
181 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
182 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
183
184 if (error == 0) {
185 zio_cksum_t cksum = bp->blk_cksum;
186
187 /*
188 * Validate the checksummed log block.
189 *
190 * Sequence numbers should be... sequential. The checksum
191 * verifier for the next block should be bp's checksum plus 1.
192 *
193 * Also check the log chain linkage and size used.
194 */
195 cksum.zc_word[ZIL_ZC_SEQ]++;
196
197 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
198 zil_chain_t *zilc = abuf->b_data;
199 char *lr = (char *)(zilc + 1);
200 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
201
202 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
203 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
204 error = ECKSUM;
205 } else {
206 bcopy(lr, dst, len);
207 *end = (char *)dst + len;
208 *nbp = zilc->zc_next_blk;
209 }
210 } else {
211 char *lr = abuf->b_data;
212 uint64_t size = BP_GET_LSIZE(bp);
213 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
214
215 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
216 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
217 (zilc->zc_nused > (size - sizeof (*zilc)))) {
218 error = ECKSUM;
219 } else {
220 bcopy(lr, dst, zilc->zc_nused);
221 *end = (char *)dst + zilc->zc_nused;
222 *nbp = zilc->zc_next_blk;
223 }
224 }
225
226 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1);
227 }
228
229 return (error);
230 }
231
232 /*
233 * Read a TX_WRITE log data block.
234 */
235 static int
zil_read_log_data(zilog_t * zilog,const lr_write_t * lr,void * wbuf)236 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
237 {
238 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
239 const blkptr_t *bp = &lr->lr_blkptr;
240 uint32_t aflags = ARC_WAIT;
241 arc_buf_t *abuf = NULL;
242 zbookmark_t zb;
243 int error;
244
245 if (BP_IS_HOLE(bp)) {
246 if (wbuf != NULL)
247 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
248 return (0);
249 }
250
251 if (zilog->zl_header->zh_claim_txg == 0)
252 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
253
254 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
255 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
256
257 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
258 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
259
260 if (error == 0) {
261 if (wbuf != NULL)
262 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
263 (void) arc_buf_remove_ref(abuf, &abuf);
264 }
265
266 return (error);
267 }
268
269 /*
270 * Parse the intent log, and call parse_func for each valid record within.
271 */
272 int
zil_parse(zilog_t * zilog,zil_parse_blk_func_t * parse_blk_func,zil_parse_lr_func_t * parse_lr_func,void * arg,uint64_t txg)273 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
274 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
275 {
276 const zil_header_t *zh = zilog->zl_header;
277 boolean_t claimed = !!zh->zh_claim_txg;
278 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
279 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
280 uint64_t max_blk_seq = 0;
281 uint64_t max_lr_seq = 0;
282 uint64_t blk_count = 0;
283 uint64_t lr_count = 0;
284 blkptr_t blk, next_blk;
285 char *lrbuf, *lrp;
286 int error = 0;
287
288 /*
289 * Old logs didn't record the maximum zh_claim_lr_seq.
290 */
291 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
292 claim_lr_seq = UINT64_MAX;
293
294 /*
295 * Starting at the block pointed to by zh_log we read the log chain.
296 * For each block in the chain we strongly check that block to
297 * ensure its validity. We stop when an invalid block is found.
298 * For each block pointer in the chain we call parse_blk_func().
299 * For each record in each valid block we call parse_lr_func().
300 * If the log has been claimed, stop if we encounter a sequence
301 * number greater than the highest claimed sequence number.
302 */
303 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
304 zil_bp_tree_init(zilog);
305
306 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
307 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
308 int reclen;
309 char *end;
310
311 if (blk_seq > claim_blk_seq)
312 break;
313 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
314 break;
315 ASSERT3U(max_blk_seq, <, blk_seq);
316 max_blk_seq = blk_seq;
317 blk_count++;
318
319 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
320 break;
321
322 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
323 if (error)
324 break;
325
326 for (lrp = lrbuf; lrp < end; lrp += reclen) {
327 lr_t *lr = (lr_t *)lrp;
328 reclen = lr->lrc_reclen;
329 ASSERT3U(reclen, >=, sizeof (lr_t));
330 if (lr->lrc_seq > claim_lr_seq)
331 goto done;
332 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
333 goto done;
334 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
335 max_lr_seq = lr->lrc_seq;
336 lr_count++;
337 }
338 }
339 done:
340 zilog->zl_parse_error = error;
341 zilog->zl_parse_blk_seq = max_blk_seq;
342 zilog->zl_parse_lr_seq = max_lr_seq;
343 zilog->zl_parse_blk_count = blk_count;
344 zilog->zl_parse_lr_count = lr_count;
345
346 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
347 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
348
349 zil_bp_tree_fini(zilog);
350 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
351
352 return (error);
353 }
354
355 static int
zil_claim_log_block(zilog_t * zilog,blkptr_t * bp,void * tx,uint64_t first_txg)356 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
357 {
358 /*
359 * Claim log block if not already committed and not already claimed.
360 * If tx == NULL, just verify that the block is claimable.
361 */
362 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0)
363 return (0);
364
365 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
366 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
367 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
368 }
369
370 static int
zil_claim_log_record(zilog_t * zilog,lr_t * lrc,void * tx,uint64_t first_txg)371 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
372 {
373 lr_write_t *lr = (lr_write_t *)lrc;
374 int error;
375
376 if (lrc->lrc_txtype != TX_WRITE)
377 return (0);
378
379 /*
380 * If the block is not readable, don't claim it. This can happen
381 * in normal operation when a log block is written to disk before
382 * some of the dmu_sync() blocks it points to. In this case, the
383 * transaction cannot have been committed to anyone (we would have
384 * waited for all writes to be stable first), so it is semantically
385 * correct to declare this the end of the log.
386 */
387 if (lr->lr_blkptr.blk_birth >= first_txg &&
388 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
389 return (error);
390 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
391 }
392
393 /* ARGSUSED */
394 static int
zil_free_log_block(zilog_t * zilog,blkptr_t * bp,void * tx,uint64_t claim_txg)395 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
396 {
397 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
398
399 return (0);
400 }
401
402 static int
zil_free_log_record(zilog_t * zilog,lr_t * lrc,void * tx,uint64_t claim_txg)403 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
404 {
405 lr_write_t *lr = (lr_write_t *)lrc;
406 blkptr_t *bp = &lr->lr_blkptr;
407
408 /*
409 * If we previously claimed it, we need to free it.
410 */
411 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
412 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0)
413 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
414
415 return (0);
416 }
417
418 static lwb_t *
zil_alloc_lwb(zilog_t * zilog,blkptr_t * bp,uint64_t txg)419 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg)
420 {
421 lwb_t *lwb;
422
423 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
424 lwb->lwb_zilog = zilog;
425 lwb->lwb_blk = *bp;
426 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
427 lwb->lwb_max_txg = txg;
428 lwb->lwb_zio = NULL;
429 lwb->lwb_tx = NULL;
430 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
431 lwb->lwb_nused = sizeof (zil_chain_t);
432 lwb->lwb_sz = BP_GET_LSIZE(bp);
433 } else {
434 lwb->lwb_nused = 0;
435 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
436 }
437
438 mutex_enter(&zilog->zl_lock);
439 list_insert_tail(&zilog->zl_lwb_list, lwb);
440 mutex_exit(&zilog->zl_lock);
441
442 return (lwb);
443 }
444
445 /*
446 * Create an on-disk intent log.
447 */
448 static lwb_t *
zil_create(zilog_t * zilog)449 zil_create(zilog_t *zilog)
450 {
451 const zil_header_t *zh = zilog->zl_header;
452 lwb_t *lwb = NULL;
453 uint64_t txg = 0;
454 dmu_tx_t *tx = NULL;
455 blkptr_t blk;
456 int error = 0;
457
458 /*
459 * Wait for any previous destroy to complete.
460 */
461 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
462
463 ASSERT(zh->zh_claim_txg == 0);
464 ASSERT(zh->zh_replay_seq == 0);
465
466 blk = zh->zh_log;
467
468 /*
469 * Allocate an initial log block if:
470 * - there isn't one already
471 * - the existing block is the wrong endianess
472 */
473 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
474 tx = dmu_tx_create(zilog->zl_os);
475 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
476 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
477 txg = dmu_tx_get_txg(tx);
478
479 if (!BP_IS_HOLE(&blk)) {
480 zio_free_zil(zilog->zl_spa, txg, &blk);
481 BP_ZERO(&blk);
482 }
483
484 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
485 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY);
486
487 if (error == 0)
488 zil_init_log_chain(zilog, &blk);
489 }
490
491 /*
492 * Allocate a log write buffer (lwb) for the first log block.
493 */
494 if (error == 0)
495 lwb = zil_alloc_lwb(zilog, &blk, txg);
496
497 /*
498 * If we just allocated the first log block, commit our transaction
499 * and wait for zil_sync() to stuff the block poiner into zh_log.
500 * (zh is part of the MOS, so we cannot modify it in open context.)
501 */
502 if (tx != NULL) {
503 dmu_tx_commit(tx);
504 txg_wait_synced(zilog->zl_dmu_pool, txg);
505 }
506
507 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
508
509 return (lwb);
510 }
511
512 /*
513 * In one tx, free all log blocks and clear the log header.
514 * If keep_first is set, then we're replaying a log with no content.
515 * We want to keep the first block, however, so that the first
516 * synchronous transaction doesn't require a txg_wait_synced()
517 * in zil_create(). We don't need to txg_wait_synced() here either
518 * when keep_first is set, because both zil_create() and zil_destroy()
519 * will wait for any in-progress destroys to complete.
520 */
521 void
zil_destroy(zilog_t * zilog,boolean_t keep_first)522 zil_destroy(zilog_t *zilog, boolean_t keep_first)
523 {
524 const zil_header_t *zh = zilog->zl_header;
525 lwb_t *lwb;
526 dmu_tx_t *tx;
527 uint64_t txg;
528
529 /*
530 * Wait for any previous destroy to complete.
531 */
532 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
533
534 zilog->zl_old_header = *zh; /* debugging aid */
535
536 if (BP_IS_HOLE(&zh->zh_log))
537 return;
538
539 tx = dmu_tx_create(zilog->zl_os);
540 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
541 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
542 txg = dmu_tx_get_txg(tx);
543
544 mutex_enter(&zilog->zl_lock);
545
546 ASSERT3U(zilog->zl_destroy_txg, <, txg);
547 zilog->zl_destroy_txg = txg;
548 zilog->zl_keep_first = keep_first;
549
550 if (!list_is_empty(&zilog->zl_lwb_list)) {
551 ASSERT(zh->zh_claim_txg == 0);
552 ASSERT(!keep_first);
553 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
554 list_remove(&zilog->zl_lwb_list, lwb);
555 if (lwb->lwb_buf != NULL)
556 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
557 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
558 kmem_cache_free(zil_lwb_cache, lwb);
559 }
560 } else if (!keep_first) {
561 (void) zil_parse(zilog, zil_free_log_block,
562 zil_free_log_record, tx, zh->zh_claim_txg);
563 }
564 mutex_exit(&zilog->zl_lock);
565
566 dmu_tx_commit(tx);
567 }
568
569 int
zil_claim(const char * osname,void * txarg)570 zil_claim(const char *osname, void *txarg)
571 {
572 dmu_tx_t *tx = txarg;
573 uint64_t first_txg = dmu_tx_get_txg(tx);
574 zilog_t *zilog;
575 zil_header_t *zh;
576 objset_t *os;
577 int error;
578
579 error = dmu_objset_hold(osname, FTAG, &os);
580 if (error) {
581 cmn_err(CE_WARN, "can't open objset for %s", osname);
582 return (0);
583 }
584
585 zilog = dmu_objset_zil(os);
586 zh = zil_header_in_syncing_context(zilog);
587
588 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
589 if (!BP_IS_HOLE(&zh->zh_log))
590 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
591 BP_ZERO(&zh->zh_log);
592 dsl_dataset_dirty(dmu_objset_ds(os), tx);
593 dmu_objset_rele(os, FTAG);
594 return (0);
595 }
596
597 /*
598 * Claim all log blocks if we haven't already done so, and remember
599 * the highest claimed sequence number. This ensures that if we can
600 * read only part of the log now (e.g. due to a missing device),
601 * but we can read the entire log later, we will not try to replay
602 * or destroy beyond the last block we successfully claimed.
603 */
604 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
605 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
606 (void) zil_parse(zilog, zil_claim_log_block,
607 zil_claim_log_record, tx, first_txg);
608 zh->zh_claim_txg = first_txg;
609 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
610 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
611 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
612 zh->zh_flags |= ZIL_REPLAY_NEEDED;
613 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
614 dsl_dataset_dirty(dmu_objset_ds(os), tx);
615 }
616
617 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
618 dmu_objset_rele(os, FTAG);
619 return (0);
620 }
621
622 /*
623 * Check the log by walking the log chain.
624 * Checksum errors are ok as they indicate the end of the chain.
625 * Any other error (no device or read failure) returns an error.
626 */
627 int
zil_check_log_chain(const char * osname,void * tx)628 zil_check_log_chain(const char *osname, void *tx)
629 {
630 zilog_t *zilog;
631 objset_t *os;
632 int error;
633
634 ASSERT(tx == NULL);
635
636 error = dmu_objset_hold(osname, FTAG, &os);
637 if (error) {
638 cmn_err(CE_WARN, "can't open objset for %s", osname);
639 return (0);
640 }
641
642 zilog = dmu_objset_zil(os);
643
644 /*
645 * Because tx == NULL, zil_claim_log_block() will not actually claim
646 * any blocks, but just determine whether it is possible to do so.
647 * In addition to checking the log chain, zil_claim_log_block()
648 * will invoke zio_claim() with a done func of spa_claim_notify(),
649 * which will update spa_max_claim_txg. See spa_load() for details.
650 */
651 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
652 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
653
654 dmu_objset_rele(os, FTAG);
655
656 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
657 }
658
659 static int
zil_vdev_compare(const void * x1,const void * x2)660 zil_vdev_compare(const void *x1, const void *x2)
661 {
662 uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
663 uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
664
665 if (v1 < v2)
666 return (-1);
667 if (v1 > v2)
668 return (1);
669
670 return (0);
671 }
672
673 void
zil_add_block(zilog_t * zilog,const blkptr_t * bp)674 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
675 {
676 avl_tree_t *t = &zilog->zl_vdev_tree;
677 avl_index_t where;
678 zil_vdev_node_t *zv, zvsearch;
679 int ndvas = BP_GET_NDVAS(bp);
680 int i;
681
682 if (zfs_nocacheflush)
683 return;
684
685 ASSERT(zilog->zl_writer);
686
687 /*
688 * Even though we're zl_writer, we still need a lock because the
689 * zl_get_data() callbacks may have dmu_sync() done callbacks
690 * that will run concurrently.
691 */
692 mutex_enter(&zilog->zl_vdev_lock);
693 for (i = 0; i < ndvas; i++) {
694 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
695 if (avl_find(t, &zvsearch, &where) == NULL) {
696 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
697 zv->zv_vdev = zvsearch.zv_vdev;
698 avl_insert(t, zv, where);
699 }
700 }
701 mutex_exit(&zilog->zl_vdev_lock);
702 }
703
704 void
zil_flush_vdevs(zilog_t * zilog)705 zil_flush_vdevs(zilog_t *zilog)
706 {
707 spa_t *spa = zilog->zl_spa;
708 avl_tree_t *t = &zilog->zl_vdev_tree;
709 void *cookie = NULL;
710 zil_vdev_node_t *zv;
711 zio_t *zio;
712
713 ASSERT(zilog->zl_writer);
714
715 /*
716 * We don't need zl_vdev_lock here because we're the zl_writer,
717 * and all zl_get_data() callbacks are done.
718 */
719 if (avl_numnodes(t) == 0)
720 return;
721
722 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
723
724 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
725
726 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
727 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
728 if (vd != NULL)
729 zio_flush(zio, vd);
730 kmem_free(zv, sizeof (*zv));
731 }
732
733 /*
734 * Wait for all the flushes to complete. Not all devices actually
735 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
736 */
737 (void) zio_wait(zio);
738
739 spa_config_exit(spa, SCL_STATE, FTAG);
740 }
741
742 /*
743 * Function called when a log block write completes
744 */
745 static void
zil_lwb_write_done(zio_t * zio)746 zil_lwb_write_done(zio_t *zio)
747 {
748 lwb_t *lwb = zio->io_private;
749 zilog_t *zilog = lwb->lwb_zilog;
750 dmu_tx_t *tx = lwb->lwb_tx;
751
752 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
753 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
754 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
755 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
756 ASSERT(!BP_IS_GANG(zio->io_bp));
757 ASSERT(!BP_IS_HOLE(zio->io_bp));
758 ASSERT(zio->io_bp->blk_fill == 0);
759
760 /*
761 * Ensure the lwb buffer pointer is cleared before releasing
762 * the txg. If we have had an allocation failure and
763 * the txg is waiting to sync then we want want zil_sync()
764 * to remove the lwb so that it's not picked up as the next new
765 * one in zil_commit_writer(). zil_sync() will only remove
766 * the lwb if lwb_buf is null.
767 */
768 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
769 mutex_enter(&zilog->zl_lock);
770 lwb->lwb_buf = NULL;
771 lwb->lwb_tx = NULL;
772 mutex_exit(&zilog->zl_lock);
773
774 /*
775 * Now that we've written this log block, we have a stable pointer
776 * to the next block in the chain, so it's OK to let the txg in
777 * which we allocated the next block sync.
778 */
779 dmu_tx_commit(tx);
780 }
781
782 /*
783 * Initialize the io for a log block.
784 */
785 static void
zil_lwb_write_init(zilog_t * zilog,lwb_t * lwb)786 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
787 {
788 zbookmark_t zb;
789
790 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
791 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
792 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
793
794 if (zilog->zl_root_zio == NULL) {
795 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
796 ZIO_FLAG_CANFAIL);
797 }
798 if (lwb->lwb_zio == NULL) {
799 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
800 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
801 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE,
802 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
803 }
804 }
805
806 /*
807 * Define a limited set of intent log block sizes.
808 * These must be a multiple of 4KB. Note only the amount used (again
809 * aligned to 4KB) actually gets written. However, we can't always just
810 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
811 */
812 uint64_t zil_block_buckets[] = {
813 4096, /* non TX_WRITE */
814 8192+4096, /* data base */
815 32*1024 + 4096, /* NFS writes */
816 UINT64_MAX
817 };
818
819 /*
820 * Use the slog as long as the logbias is 'latency' and the current commit size
821 * is less than the limit or the total list size is less than 2X the limit.
822 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX.
823 */
824 uint64_t zil_slog_limit = 1024 * 1024;
825 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \
826 (((zilog)->zl_cur_used < zil_slog_limit) || \
827 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1))))
828
829 /*
830 * Start a log block write and advance to the next log block.
831 * Calls are serialized.
832 */
833 static lwb_t *
zil_lwb_write_start(zilog_t * zilog,lwb_t * lwb)834 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
835 {
836 lwb_t *nlwb = NULL;
837 zil_chain_t *zilc;
838 spa_t *spa = zilog->zl_spa;
839 blkptr_t *bp;
840 dmu_tx_t *tx;
841 uint64_t txg;
842 uint64_t zil_blksz;
843 int i, error;
844
845 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
846 zilc = (zil_chain_t *)lwb->lwb_buf;
847 bp = &zilc->zc_next_blk;
848 } else {
849 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
850 bp = &zilc->zc_next_blk;
851 }
852
853 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
854
855 /*
856 * Allocate the next block and save its address in this block
857 * before writing it in order to establish the log chain.
858 * Note that if the allocation of nlwb synced before we wrote
859 * the block that points at it (lwb), we'd leak it if we crashed.
860 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
861 * We dirty the dataset to ensure that zil_sync() will be called
862 * to clean up in the event of allocation failure or I/O failure.
863 */
864 tx = dmu_tx_create(zilog->zl_os);
865 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
866 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
867 txg = dmu_tx_get_txg(tx);
868
869 lwb->lwb_tx = tx;
870
871 /*
872 * Log blocks are pre-allocated. Here we select the size of the next
873 * block, based on size used in the last block.
874 * - first find the smallest bucket that will fit the block from a
875 * limited set of block sizes. This is because it's faster to write
876 * blocks allocated from the same metaslab as they are adjacent or
877 * close.
878 * - next find the maximum from the new suggested size and an array of
879 * previous sizes. This lessens a picket fence effect of wrongly
880 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
881 * requests.
882 *
883 * Note we only write what is used, but we can't just allocate
884 * the maximum block size because we can exhaust the available
885 * pool log space.
886 */
887 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
888 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
889 continue;
890 zil_blksz = zil_block_buckets[i];
891 if (zil_blksz == UINT64_MAX)
892 zil_blksz = SPA_MAXBLOCKSIZE;
893 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
894 for (i = 0; i < ZIL_PREV_BLKS; i++)
895 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
896 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
897
898 BP_ZERO(bp);
899 /* pass the old blkptr in order to spread log blocks across devs */
900 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz,
901 USE_SLOG(zilog));
902 if (!error) {
903 ASSERT3U(bp->blk_birth, ==, txg);
904 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
905 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
906
907 /*
908 * Allocate a new log write buffer (lwb).
909 */
910 nlwb = zil_alloc_lwb(zilog, bp, txg);
911
912 /* Record the block for later vdev flushing */
913 zil_add_block(zilog, &lwb->lwb_blk);
914 }
915
916 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
917 uint64_t len;
918
919 /* For Slim ZIL only write what is used. */
920 len = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
921 ASSERT3U(len, <=, lwb->lwb_sz);
922 zio_shrink(lwb->lwb_zio, len);
923
924 }
925 zilc->zc_pad = 0;
926 zilc->zc_nused = lwb->lwb_nused;
927 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
928
929 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
930
931 /*
932 * If there was an allocation failure then nlwb will be null which
933 * forces a txg_wait_synced().
934 */
935 return (nlwb);
936 }
937
938 static lwb_t *
zil_lwb_commit(zilog_t * zilog,itx_t * itx,lwb_t * lwb)939 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
940 {
941 lr_t *lrc = &itx->itx_lr; /* common log record */
942 lr_write_t *lrw = (lr_write_t *)lrc;
943 char *lr_buf;
944 uint64_t txg = lrc->lrc_txg;
945 uint64_t reclen = lrc->lrc_reclen;
946 uint64_t dlen = 0;
947
948 if (lwb == NULL)
949 return (NULL);
950
951 ASSERT(lwb->lwb_buf != NULL);
952
953 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
954 dlen = P2ROUNDUP_TYPED(
955 lrw->lr_length, sizeof (uint64_t), uint64_t);
956
957 zilog->zl_cur_used += (reclen + dlen);
958
959 zil_lwb_write_init(zilog, lwb);
960
961 /*
962 * If this record won't fit in the current log block, start a new one.
963 */
964 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
965 lwb = zil_lwb_write_start(zilog, lwb);
966 if (lwb == NULL)
967 return (NULL);
968 zil_lwb_write_init(zilog, lwb);
969 ASSERT(LWB_EMPTY(lwb));
970 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) {
971 txg_wait_synced(zilog->zl_dmu_pool, txg);
972 return (lwb);
973 }
974 }
975
976 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
977 bcopy(lrc, lr_buf, reclen);
978 lrc = (lr_t *)lr_buf;
979 lrw = (lr_write_t *)lrc;
980
981 /*
982 * If it's a write, fetch the data or get its blkptr as appropriate.
983 */
984 if (lrc->lrc_txtype == TX_WRITE) {
985 if (txg > spa_freeze_txg(zilog->zl_spa))
986 txg_wait_synced(zilog->zl_dmu_pool, txg);
987 if (itx->itx_wr_state != WR_COPIED) {
988 char *dbuf;
989 int error;
990
991 if (dlen) {
992 ASSERT(itx->itx_wr_state == WR_NEED_COPY);
993 dbuf = lr_buf + reclen;
994 lrw->lr_common.lrc_reclen += dlen;
995 } else {
996 ASSERT(itx->itx_wr_state == WR_INDIRECT);
997 dbuf = NULL;
998 }
999 error = zilog->zl_get_data(
1000 itx->itx_private, lrw, dbuf, lwb->lwb_zio);
1001 if (error == EIO) {
1002 txg_wait_synced(zilog->zl_dmu_pool, txg);
1003 return (lwb);
1004 }
1005 if (error) {
1006 ASSERT(error == ENOENT || error == EEXIST ||
1007 error == EALREADY);
1008 return (lwb);
1009 }
1010 }
1011 }
1012
1013 /*
1014 * We're actually making an entry, so update lrc_seq to be the
1015 * log record sequence number. Note that this is generally not
1016 * equal to the itx sequence number because not all transactions
1017 * are synchronous, and sometimes spa_sync() gets there first.
1018 */
1019 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1020 lwb->lwb_nused += reclen + dlen;
1021 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1022 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1023 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0);
1024
1025 return (lwb);
1026 }
1027
1028 itx_t *
zil_itx_create(uint64_t txtype,size_t lrsize)1029 zil_itx_create(uint64_t txtype, size_t lrsize)
1030 {
1031 itx_t *itx;
1032
1033 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1034
1035 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1036 itx->itx_lr.lrc_txtype = txtype;
1037 itx->itx_lr.lrc_reclen = lrsize;
1038 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */
1039 itx->itx_lr.lrc_seq = 0; /* defensive */
1040
1041 return (itx);
1042 }
1043
1044 void
zil_itx_destroy(itx_t * itx)1045 zil_itx_destroy(itx_t *itx)
1046 {
1047 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1048 }
1049
1050 uint64_t
zil_itx_assign(zilog_t * zilog,itx_t * itx,dmu_tx_t * tx)1051 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1052 {
1053 uint64_t seq;
1054
1055 ASSERT(itx->itx_lr.lrc_seq == 0);
1056 ASSERT(!zilog->zl_replay);
1057
1058 mutex_enter(&zilog->zl_lock);
1059 list_insert_tail(&zilog->zl_itx_list, itx);
1060 zilog->zl_itx_list_sz += itx->itx_sod;
1061 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1062 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq;
1063 mutex_exit(&zilog->zl_lock);
1064
1065 return (seq);
1066 }
1067
1068 /*
1069 * Free up all in-memory intent log transactions that have now been synced.
1070 */
1071 static void
zil_itx_clean(zilog_t * zilog)1072 zil_itx_clean(zilog_t *zilog)
1073 {
1074 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa);
1075 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa);
1076 list_t clean_list;
1077 itx_t *itx;
1078
1079 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1080
1081 mutex_enter(&zilog->zl_lock);
1082 /* wait for a log writer to finish walking list */
1083 while (zilog->zl_writer) {
1084 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1085 }
1086
1087 /*
1088 * Move the sync'd log transactions to a separate list so we can call
1089 * kmem_free without holding the zl_lock.
1090 *
1091 * There is no need to set zl_writer as we don't drop zl_lock here
1092 */
1093 while ((itx = list_head(&zilog->zl_itx_list)) != NULL &&
1094 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) {
1095 list_remove(&zilog->zl_itx_list, itx);
1096 zilog->zl_itx_list_sz -= itx->itx_sod;
1097 list_insert_tail(&clean_list, itx);
1098 }
1099 cv_broadcast(&zilog->zl_cv_writer);
1100 mutex_exit(&zilog->zl_lock);
1101
1102 /* destroy sync'd log transactions */
1103 while ((itx = list_head(&clean_list)) != NULL) {
1104 list_remove(&clean_list, itx);
1105 zil_itx_destroy(itx);
1106 }
1107 list_destroy(&clean_list);
1108 }
1109
1110 /*
1111 * If there are any in-memory intent log transactions which have now been
1112 * synced then start up a taskq to free them.
1113 */
1114 void
zil_clean(zilog_t * zilog)1115 zil_clean(zilog_t *zilog)
1116 {
1117 itx_t *itx;
1118
1119 mutex_enter(&zilog->zl_lock);
1120 itx = list_head(&zilog->zl_itx_list);
1121 if ((itx != NULL) &&
1122 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) {
1123 (void) taskq_dispatch(zilog->zl_clean_taskq,
1124 (task_func_t *)zil_itx_clean, zilog, TQ_NOSLEEP);
1125 }
1126 mutex_exit(&zilog->zl_lock);
1127 }
1128
1129 static void
zil_commit_writer(zilog_t * zilog,uint64_t seq,uint64_t foid)1130 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid)
1131 {
1132 uint64_t txg;
1133 uint64_t commit_seq = 0;
1134 itx_t *itx, *itx_next;
1135 lwb_t *lwb;
1136 spa_t *spa;
1137 int error = 0;
1138
1139 zilog->zl_writer = B_TRUE;
1140 ASSERT(zilog->zl_root_zio == NULL);
1141 spa = zilog->zl_spa;
1142
1143 if (zilog->zl_suspend) {
1144 lwb = NULL;
1145 } else {
1146 lwb = list_tail(&zilog->zl_lwb_list);
1147 if (lwb == NULL) {
1148 /*
1149 * Return if there's nothing to flush before we
1150 * dirty the fs by calling zil_create()
1151 */
1152 if (list_is_empty(&zilog->zl_itx_list)) {
1153 zilog->zl_writer = B_FALSE;
1154 return;
1155 }
1156 mutex_exit(&zilog->zl_lock);
1157 lwb = zil_create(zilog);
1158 mutex_enter(&zilog->zl_lock);
1159 }
1160 }
1161 ASSERT(lwb == NULL || lwb->lwb_zio == NULL);
1162
1163 /* Loop through in-memory log transactions filling log blocks. */
1164 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1165
1166 for (itx = list_head(&zilog->zl_itx_list); itx; itx = itx_next) {
1167 /*
1168 * Save the next pointer. Even though we drop zl_lock below,
1169 * all threads that can remove itx list entries (other writers
1170 * and zil_itx_clean()) can't do so until they have zl_writer.
1171 */
1172 itx_next = list_next(&zilog->zl_itx_list, itx);
1173
1174 /*
1175 * Determine whether to push this itx.
1176 * Push all transactions related to specified foid and
1177 * all other transactions except those that can be logged
1178 * out of order (TX_WRITE, TX_TRUNCATE, TX_SETATTR, TX_ACL)
1179 * for all other files.
1180 *
1181 * If foid == 0 (meaning "push all foids") or
1182 * itx->itx_sync is set (meaning O_[D]SYNC), push regardless.
1183 */
1184 if (foid != 0 && !itx->itx_sync &&
1185 TX_OOO(itx->itx_lr.lrc_txtype) &&
1186 ((lr_ooo_t *)&itx->itx_lr)->lr_foid != foid)
1187 continue; /* skip this record */
1188
1189 if ((itx->itx_lr.lrc_seq > seq) &&
1190 ((lwb == NULL) || (LWB_EMPTY(lwb)) ||
1191 (lwb->lwb_nused + itx->itx_sod > lwb->lwb_sz)))
1192 break;
1193
1194 list_remove(&zilog->zl_itx_list, itx);
1195 zilog->zl_itx_list_sz -= itx->itx_sod;
1196
1197 mutex_exit(&zilog->zl_lock);
1198
1199 txg = itx->itx_lr.lrc_txg;
1200 ASSERT(txg);
1201
1202 if (txg > spa_last_synced_txg(spa) ||
1203 txg > spa_freeze_txg(spa))
1204 lwb = zil_lwb_commit(zilog, itx, lwb);
1205
1206 zil_itx_destroy(itx);
1207
1208 mutex_enter(&zilog->zl_lock);
1209 }
1210 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1211 /* determine commit sequence number */
1212 itx = list_head(&zilog->zl_itx_list);
1213 if (itx)
1214 commit_seq = itx->itx_lr.lrc_seq - 1;
1215 else
1216 commit_seq = zilog->zl_itx_seq;
1217 mutex_exit(&zilog->zl_lock);
1218
1219 /* write the last block out */
1220 if (lwb != NULL && lwb->lwb_zio != NULL)
1221 lwb = zil_lwb_write_start(zilog, lwb);
1222
1223 zilog->zl_prev_used = zilog->zl_cur_used;
1224 zilog->zl_cur_used = 0;
1225
1226 /*
1227 * Wait if necessary for the log blocks to be on stable storage.
1228 */
1229 if (zilog->zl_root_zio) {
1230 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog);
1231 error = zio_wait(zilog->zl_root_zio);
1232 zilog->zl_root_zio = NULL;
1233 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog);
1234 zil_flush_vdevs(zilog);
1235 }
1236
1237 if (error || lwb == NULL)
1238 txg_wait_synced(zilog->zl_dmu_pool, 0);
1239
1240 mutex_enter(&zilog->zl_lock);
1241 zilog->zl_writer = B_FALSE;
1242
1243 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq);
1244 zilog->zl_commit_seq = commit_seq;
1245
1246 /*
1247 * Remember the highest committed log sequence number for ztest.
1248 * We only update this value when all the log writes succeeded,
1249 * because ztest wants to ASSERT that it got the whole log chain.
1250 */
1251 if (error == 0 && lwb != NULL)
1252 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1253 }
1254
1255 /*
1256 * Push zfs transactions to stable storage up to the supplied sequence number.
1257 * If foid is 0 push out all transactions, otherwise push only those
1258 * for that file or might have been used to create that file.
1259 */
1260 void
zil_commit(zilog_t * zilog,uint64_t seq,uint64_t foid)1261 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid)
1262 {
1263 if (zilog == NULL || seq == 0)
1264 return;
1265
1266 mutex_enter(&zilog->zl_lock);
1267
1268 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */
1269
1270 while (zilog->zl_writer) {
1271 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1272 if (seq <= zilog->zl_commit_seq) {
1273 mutex_exit(&zilog->zl_lock);
1274 return;
1275 }
1276 }
1277 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */
1278 /* wake up others waiting on the commit */
1279 cv_broadcast(&zilog->zl_cv_writer);
1280 mutex_exit(&zilog->zl_lock);
1281 }
1282
1283 /*
1284 * Report whether all transactions are committed.
1285 */
1286 static boolean_t
zil_is_committed(zilog_t * zilog)1287 zil_is_committed(zilog_t *zilog)
1288 {
1289 lwb_t *lwb;
1290 boolean_t committed;
1291
1292 mutex_enter(&zilog->zl_lock);
1293
1294 while (zilog->zl_writer)
1295 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1296
1297 if (!list_is_empty(&zilog->zl_itx_list))
1298 committed = B_FALSE; /* unpushed transactions */
1299 else if ((lwb = list_head(&zilog->zl_lwb_list)) == NULL)
1300 committed = B_TRUE; /* intent log never used */
1301 else if (list_next(&zilog->zl_lwb_list, lwb) != NULL)
1302 committed = B_FALSE; /* zil_sync() not done yet */
1303 else
1304 committed = B_TRUE; /* everything synced */
1305
1306 mutex_exit(&zilog->zl_lock);
1307 return (committed);
1308 }
1309
1310 /*
1311 * Called in syncing context to free committed log blocks and update log header.
1312 */
1313 void
zil_sync(zilog_t * zilog,dmu_tx_t * tx)1314 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1315 {
1316 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1317 uint64_t txg = dmu_tx_get_txg(tx);
1318 spa_t *spa = zilog->zl_spa;
1319 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1320 lwb_t *lwb;
1321
1322 /*
1323 * We don't zero out zl_destroy_txg, so make sure we don't try
1324 * to destroy it twice.
1325 */
1326 if (spa_sync_pass(spa) != 1)
1327 return;
1328
1329 mutex_enter(&zilog->zl_lock);
1330
1331 ASSERT(zilog->zl_stop_sync == 0);
1332
1333 if (*replayed_seq != 0) {
1334 ASSERT(zh->zh_replay_seq < *replayed_seq);
1335 zh->zh_replay_seq = *replayed_seq;
1336 *replayed_seq = 0;
1337 }
1338
1339 if (zilog->zl_destroy_txg == txg) {
1340 blkptr_t blk = zh->zh_log;
1341
1342 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1343
1344 bzero(zh, sizeof (zil_header_t));
1345 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1346
1347 if (zilog->zl_keep_first) {
1348 /*
1349 * If this block was part of log chain that couldn't
1350 * be claimed because a device was missing during
1351 * zil_claim(), but that device later returns,
1352 * then this block could erroneously appear valid.
1353 * To guard against this, assign a new GUID to the new
1354 * log chain so it doesn't matter what blk points to.
1355 */
1356 zil_init_log_chain(zilog, &blk);
1357 zh->zh_log = blk;
1358 }
1359 }
1360
1361 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1362 zh->zh_log = lwb->lwb_blk;
1363 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1364 break;
1365 list_remove(&zilog->zl_lwb_list, lwb);
1366 zio_free_zil(spa, txg, &lwb->lwb_blk);
1367 kmem_cache_free(zil_lwb_cache, lwb);
1368
1369 /*
1370 * If we don't have anything left in the lwb list then
1371 * we've had an allocation failure and we need to zero
1372 * out the zil_header blkptr so that we don't end
1373 * up freeing the same block twice.
1374 */
1375 if (list_head(&zilog->zl_lwb_list) == NULL)
1376 BP_ZERO(&zh->zh_log);
1377 }
1378 mutex_exit(&zilog->zl_lock);
1379 }
1380
1381 void
zil_init(void)1382 zil_init(void)
1383 {
1384 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1385 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1386 }
1387
1388 void
zil_fini(void)1389 zil_fini(void)
1390 {
1391 kmem_cache_destroy(zil_lwb_cache);
1392 }
1393
1394 void
zil_set_logbias(zilog_t * zilog,uint64_t logbias)1395 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1396 {
1397 zilog->zl_logbias = logbias;
1398 }
1399
1400 zilog_t *
zil_alloc(objset_t * os,zil_header_t * zh_phys)1401 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1402 {
1403 zilog_t *zilog;
1404
1405 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1406
1407 zilog->zl_header = zh_phys;
1408 zilog->zl_os = os;
1409 zilog->zl_spa = dmu_objset_spa(os);
1410 zilog->zl_dmu_pool = dmu_objset_pool(os);
1411 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1412 zilog->zl_logbias = dmu_objset_logbias(os);
1413
1414 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1415
1416 list_create(&zilog->zl_itx_list, sizeof (itx_t),
1417 offsetof(itx_t, itx_node));
1418
1419 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1420 offsetof(lwb_t, lwb_node));
1421
1422 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1423
1424 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1425 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1426
1427 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1428 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1429
1430 return (zilog);
1431 }
1432
1433 void
zil_free(zilog_t * zilog)1434 zil_free(zilog_t *zilog)
1435 {
1436 lwb_t *lwb;
1437
1438 zilog->zl_stop_sync = 1;
1439
1440 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1441 list_remove(&zilog->zl_lwb_list, lwb);
1442 if (lwb->lwb_buf != NULL)
1443 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1444 kmem_cache_free(zil_lwb_cache, lwb);
1445 }
1446 list_destroy(&zilog->zl_lwb_list);
1447
1448 avl_destroy(&zilog->zl_vdev_tree);
1449 mutex_destroy(&zilog->zl_vdev_lock);
1450
1451 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1452 list_destroy(&zilog->zl_itx_list);
1453 mutex_destroy(&zilog->zl_lock);
1454
1455 cv_destroy(&zilog->zl_cv_writer);
1456 cv_destroy(&zilog->zl_cv_suspend);
1457
1458 kmem_free(zilog, sizeof (zilog_t));
1459 }
1460
1461 /*
1462 * Open an intent log.
1463 */
1464 zilog_t *
zil_open(objset_t * os,zil_get_data_t * get_data)1465 zil_open(objset_t *os, zil_get_data_t *get_data)
1466 {
1467 zilog_t *zilog = dmu_objset_zil(os);
1468
1469 zilog->zl_get_data = get_data;
1470 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1471 2, 2, TASKQ_PREPOPULATE);
1472
1473 return (zilog);
1474 }
1475
1476 /*
1477 * Close an intent log.
1478 */
1479 void
zil_close(zilog_t * zilog)1480 zil_close(zilog_t *zilog)
1481 {
1482 /*
1483 * If the log isn't already committed, mark the objset dirty
1484 * (so zil_sync() will be called) and wait for that txg to sync.
1485 */
1486 if (!zil_is_committed(zilog)) {
1487 uint64_t txg;
1488 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1489 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
1490 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1491 txg = dmu_tx_get_txg(tx);
1492 dmu_tx_commit(tx);
1493 txg_wait_synced(zilog->zl_dmu_pool, txg);
1494 }
1495
1496 taskq_destroy(zilog->zl_clean_taskq);
1497 zilog->zl_clean_taskq = NULL;
1498 zilog->zl_get_data = NULL;
1499
1500 zil_itx_clean(zilog);
1501 ASSERT(list_head(&zilog->zl_itx_list) == NULL);
1502 }
1503
1504 /*
1505 * Suspend an intent log. While in suspended mode, we still honor
1506 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1507 * We suspend the log briefly when taking a snapshot so that the snapshot
1508 * contains all the data it's supposed to, and has an empty intent log.
1509 */
1510 int
zil_suspend(zilog_t * zilog)1511 zil_suspend(zilog_t *zilog)
1512 {
1513 const zil_header_t *zh = zilog->zl_header;
1514
1515 mutex_enter(&zilog->zl_lock);
1516 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1517 mutex_exit(&zilog->zl_lock);
1518 return (EBUSY);
1519 }
1520 if (zilog->zl_suspend++ != 0) {
1521 /*
1522 * Someone else already began a suspend.
1523 * Just wait for them to finish.
1524 */
1525 while (zilog->zl_suspending)
1526 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1527 mutex_exit(&zilog->zl_lock);
1528 return (0);
1529 }
1530 zilog->zl_suspending = B_TRUE;
1531 mutex_exit(&zilog->zl_lock);
1532
1533 zil_commit(zilog, UINT64_MAX, 0);
1534
1535 /*
1536 * Wait for any in-flight log writes to complete.
1537 */
1538 mutex_enter(&zilog->zl_lock);
1539 while (zilog->zl_writer)
1540 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock);
1541 mutex_exit(&zilog->zl_lock);
1542
1543 zil_destroy(zilog, B_FALSE);
1544
1545 mutex_enter(&zilog->zl_lock);
1546 zilog->zl_suspending = B_FALSE;
1547 cv_broadcast(&zilog->zl_cv_suspend);
1548 mutex_exit(&zilog->zl_lock);
1549
1550 return (0);
1551 }
1552
1553 void
zil_resume(zilog_t * zilog)1554 zil_resume(zilog_t *zilog)
1555 {
1556 mutex_enter(&zilog->zl_lock);
1557 ASSERT(zilog->zl_suspend != 0);
1558 zilog->zl_suspend--;
1559 mutex_exit(&zilog->zl_lock);
1560 }
1561
1562 typedef struct zil_replay_arg {
1563 zil_replay_func_t **zr_replay;
1564 void *zr_arg;
1565 boolean_t zr_byteswap;
1566 char *zr_lr;
1567 } zil_replay_arg_t;
1568
1569 static int
zil_replay_error(zilog_t * zilog,lr_t * lr,int error)1570 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
1571 {
1572 char name[MAXNAMELEN];
1573
1574 zilog->zl_replaying_seq--; /* didn't actually replay this one */
1575
1576 dmu_objset_name(zilog->zl_os, name);
1577
1578 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
1579 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
1580 (u_longlong_t)lr->lrc_seq,
1581 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
1582 (lr->lrc_txtype & TX_CI) ? "CI" : "");
1583
1584 return (error);
1585 }
1586
1587 static int
zil_replay_log_record(zilog_t * zilog,lr_t * lr,void * zra,uint64_t claim_txg)1588 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
1589 {
1590 zil_replay_arg_t *zr = zra;
1591 const zil_header_t *zh = zilog->zl_header;
1592 uint64_t reclen = lr->lrc_reclen;
1593 uint64_t txtype = lr->lrc_txtype;
1594 int error = 0;
1595
1596 zilog->zl_replaying_seq = lr->lrc_seq;
1597
1598 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
1599 return (0);
1600
1601 if (lr->lrc_txg < claim_txg) /* already committed */
1602 return (0);
1603
1604 /* Strip case-insensitive bit, still present in log record */
1605 txtype &= ~TX_CI;
1606
1607 if (txtype == 0 || txtype >= TX_MAX_TYPE)
1608 return (zil_replay_error(zilog, lr, EINVAL));
1609
1610 /*
1611 * If this record type can be logged out of order, the object
1612 * (lr_foid) may no longer exist. That's legitimate, not an error.
1613 */
1614 if (TX_OOO(txtype)) {
1615 error = dmu_object_info(zilog->zl_os,
1616 ((lr_ooo_t *)lr)->lr_foid, NULL);
1617 if (error == ENOENT || error == EEXIST)
1618 return (0);
1619 }
1620
1621 /*
1622 * Make a copy of the data so we can revise and extend it.
1623 */
1624 bcopy(lr, zr->zr_lr, reclen);
1625
1626 /*
1627 * If this is a TX_WRITE with a blkptr, suck in the data.
1628 */
1629 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
1630 error = zil_read_log_data(zilog, (lr_write_t *)lr,
1631 zr->zr_lr + reclen);
1632 if (error)
1633 return (zil_replay_error(zilog, lr, error));
1634 }
1635
1636 /*
1637 * The log block containing this lr may have been byteswapped
1638 * so that we can easily examine common fields like lrc_txtype.
1639 * However, the log is a mix of different record types, and only the
1640 * replay vectors know how to byteswap their records. Therefore, if
1641 * the lr was byteswapped, undo it before invoking the replay vector.
1642 */
1643 if (zr->zr_byteswap)
1644 byteswap_uint64_array(zr->zr_lr, reclen);
1645
1646 /*
1647 * We must now do two things atomically: replay this log record,
1648 * and update the log header sequence number to reflect the fact that
1649 * we did so. At the end of each replay function the sequence number
1650 * is updated if we are in replay mode.
1651 */
1652 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
1653 if (error) {
1654 /*
1655 * The DMU's dnode layer doesn't see removes until the txg
1656 * commits, so a subsequent claim can spuriously fail with
1657 * EEXIST. So if we receive any error we try syncing out
1658 * any removes then retry the transaction. Note that we
1659 * specify B_FALSE for byteswap now, so we don't do it twice.
1660 */
1661 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
1662 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
1663 if (error)
1664 return (zil_replay_error(zilog, lr, error));
1665 }
1666 return (0);
1667 }
1668
1669 /* ARGSUSED */
1670 static int
zil_incr_blks(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)1671 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1672 {
1673 zilog->zl_replay_blks++;
1674
1675 return (0);
1676 }
1677
1678 /*
1679 * If this dataset has a non-empty intent log, replay it and destroy it.
1680 */
1681 void
zil_replay(objset_t * os,void * arg,zil_replay_func_t * replay_func[TX_MAX_TYPE])1682 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
1683 {
1684 zilog_t *zilog = dmu_objset_zil(os);
1685 const zil_header_t *zh = zilog->zl_header;
1686 zil_replay_arg_t zr;
1687
1688 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
1689 zil_destroy(zilog, B_TRUE);
1690 return;
1691 }
1692
1693 zr.zr_replay = replay_func;
1694 zr.zr_arg = arg;
1695 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
1696 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
1697
1698 /*
1699 * Wait for in-progress removes to sync before starting replay.
1700 */
1701 txg_wait_synced(zilog->zl_dmu_pool, 0);
1702
1703 zilog->zl_replay = B_TRUE;
1704 zilog->zl_replay_time = ddi_get_lbolt();
1705 ASSERT(zilog->zl_replay_blks == 0);
1706 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
1707 zh->zh_claim_txg);
1708 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
1709
1710 zil_destroy(zilog, B_FALSE);
1711 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1712 zilog->zl_replay = B_FALSE;
1713 }
1714
1715 boolean_t
zil_replaying(zilog_t * zilog,dmu_tx_t * tx)1716 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
1717 {
1718 if (zilog == NULL)
1719 return (B_TRUE);
1720
1721 if (zilog->zl_replay) {
1722 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1723 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
1724 zilog->zl_replaying_seq;
1725 return (B_TRUE);
1726 }
1727
1728 return (B_FALSE);
1729 }
1730
1731 /* ARGSUSED */
1732 int
zil_vdev_offline(const char * osname,void * arg)1733 zil_vdev_offline(const char *osname, void *arg)
1734 {
1735 objset_t *os;
1736 zilog_t *zilog;
1737 int error;
1738
1739 error = dmu_objset_hold(osname, FTAG, &os);
1740 if (error)
1741 return (error);
1742
1743 zilog = dmu_objset_zil(os);
1744 if (zil_suspend(zilog) != 0)
1745 error = EEXIST;
1746 else
1747 zil_resume(zilog);
1748 dmu_objset_rele(os, FTAG);
1749 return (error);
1750 }
1751