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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2016 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 */
26
27 /* Portions Copyright 2010 Robert Milkowski */
28
29 #include <sys/zfs_context.h>
30 #include <sys/spa.h>
31 #include <sys/dmu.h>
32 #include <sys/zap.h>
33 #include <sys/arc.h>
34 #include <sys/stat.h>
35 #include <sys/resource.h>
36 #include <sys/zil.h>
37 #include <sys/zil_impl.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/dmu_tx.h>
41 #include <sys/dsl_pool.h>
42
43 /*
44 * The zfs intent log (ZIL) saves transaction records of system calls
45 * that change the file system in memory with enough information
46 * to be able to replay them. These are stored in memory until
47 * either the DMU transaction group (txg) commits them to the stable pool
48 * and they can be discarded, or they are flushed to the stable log
49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous
50 * requirement. In the event of a panic or power fail then those log
51 * records (transactions) are replayed.
52 *
53 * There is one ZIL per file system. Its on-disk (pool) format consists
54 * of 3 parts:
55 *
56 * - ZIL header
57 * - ZIL blocks
58 * - ZIL records
59 *
60 * A log record holds a system call transaction. Log blocks can
61 * hold many log records and the blocks are chained together.
62 * Each ZIL block contains a block pointer (blkptr_t) to the next
63 * ZIL block in the chain. The ZIL header points to the first
64 * block in the chain. Note there is not a fixed place in the pool
65 * to hold blocks. They are dynamically allocated and freed as
66 * needed from the blocks available. Figure X shows the ZIL structure:
67 */
68
69 /*
70 * Disable intent logging replay. This global ZIL switch affects all pools.
71 */
72 int zil_replay_disable = 0;
73 SYSCTL_DECL(_vfs_zfs);
74 SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN,
75 &zil_replay_disable, 0, "Disable intent logging replay");
76
77 /*
78 * Tunable parameter for debugging or performance analysis. Setting
79 * zfs_nocacheflush will cause corruption on power loss if a volatile
80 * out-of-order write cache is enabled.
81 */
82 boolean_t zfs_nocacheflush = B_FALSE;
83 SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN,
84 &zfs_nocacheflush, 0, "Disable cache flush");
85 boolean_t zfs_trim_enabled = B_TRUE;
86 SYSCTL_DECL(_vfs_zfs_trim);
87 SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0,
88 "Enable ZFS TRIM");
89
90 /*
91 * Limit SLOG write size per commit executed with synchronous priority.
92 * Any writes above that executed with lower (asynchronous) priority to
93 * limit potential SLOG device abuse by single active ZIL writer.
94 */
95 uint64_t zil_slog_limit = 768 * 1024;
96 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_limit, CTLFLAG_RWTUN,
97 &zil_slog_limit, 0, "Maximal SLOG commit size with sync priority");
98
99 static kmem_cache_t *zil_lwb_cache;
100
101 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \
102 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused))
103
104
105 /*
106 * ziltest is by and large an ugly hack, but very useful in
107 * checking replay without tedious work.
108 * When running ziltest we want to keep all itx's and so maintain
109 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG
110 * We subtract TXG_CONCURRENT_STATES to allow for common code.
111 */
112 #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES)
113
114 static int
zil_bp_compare(const void * x1,const void * x2)115 zil_bp_compare(const void *x1, const void *x2)
116 {
117 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
118 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
119
120 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2))
121 return (-1);
122 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2))
123 return (1);
124
125 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2))
126 return (-1);
127 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2))
128 return (1);
129
130 return (0);
131 }
132
133 static void
zil_bp_tree_init(zilog_t * zilog)134 zil_bp_tree_init(zilog_t *zilog)
135 {
136 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
137 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
138 }
139
140 static void
zil_bp_tree_fini(zilog_t * zilog)141 zil_bp_tree_fini(zilog_t *zilog)
142 {
143 avl_tree_t *t = &zilog->zl_bp_tree;
144 zil_bp_node_t *zn;
145 void *cookie = NULL;
146
147 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
148 kmem_free(zn, sizeof (zil_bp_node_t));
149
150 avl_destroy(t);
151 }
152
153 int
zil_bp_tree_add(zilog_t * zilog,const blkptr_t * bp)154 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
155 {
156 avl_tree_t *t = &zilog->zl_bp_tree;
157 const dva_t *dva;
158 zil_bp_node_t *zn;
159 avl_index_t where;
160
161 if (BP_IS_EMBEDDED(bp))
162 return (0);
163
164 dva = BP_IDENTITY(bp);
165
166 if (avl_find(t, dva, &where) != NULL)
167 return (SET_ERROR(EEXIST));
168
169 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
170 zn->zn_dva = *dva;
171 avl_insert(t, zn, where);
172
173 return (0);
174 }
175
176 static zil_header_t *
zil_header_in_syncing_context(zilog_t * zilog)177 zil_header_in_syncing_context(zilog_t *zilog)
178 {
179 return ((zil_header_t *)zilog->zl_header);
180 }
181
182 static void
zil_init_log_chain(zilog_t * zilog,blkptr_t * bp)183 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
184 {
185 zio_cksum_t *zc = &bp->blk_cksum;
186
187 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL);
188 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL);
189 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
190 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
191 }
192
193 /*
194 * Read a log block and make sure it's valid.
195 */
196 static int
zil_read_log_block(zilog_t * zilog,const blkptr_t * bp,blkptr_t * nbp,void * dst,char ** end)197 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
198 char **end)
199 {
200 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
201 arc_flags_t aflags = ARC_FLAG_WAIT;
202 arc_buf_t *abuf = NULL;
203 zbookmark_phys_t zb;
204 int error;
205
206 if (zilog->zl_header->zh_claim_txg == 0)
207 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
208
209 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
210 zio_flags |= ZIO_FLAG_SPECULATIVE;
211
212 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
213 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
214
215 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
216 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
217
218 if (error == 0) {
219 zio_cksum_t cksum = bp->blk_cksum;
220
221 /*
222 * Validate the checksummed log block.
223 *
224 * Sequence numbers should be... sequential. The checksum
225 * verifier for the next block should be bp's checksum plus 1.
226 *
227 * Also check the log chain linkage and size used.
228 */
229 cksum.zc_word[ZIL_ZC_SEQ]++;
230
231 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
232 zil_chain_t *zilc = abuf->b_data;
233 char *lr = (char *)(zilc + 1);
234 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t);
235
236 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
237 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
238 error = SET_ERROR(ECKSUM);
239 } else {
240 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
241 bcopy(lr, dst, len);
242 *end = (char *)dst + len;
243 *nbp = zilc->zc_next_blk;
244 }
245 } else {
246 char *lr = abuf->b_data;
247 uint64_t size = BP_GET_LSIZE(bp);
248 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
249
250 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
251 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
252 (zilc->zc_nused > (size - sizeof (*zilc)))) {
253 error = SET_ERROR(ECKSUM);
254 } else {
255 ASSERT3U(zilc->zc_nused, <=,
256 SPA_OLD_MAXBLOCKSIZE);
257 bcopy(lr, dst, zilc->zc_nused);
258 *end = (char *)dst + zilc->zc_nused;
259 *nbp = zilc->zc_next_blk;
260 }
261 }
262
263 arc_buf_destroy(abuf, &abuf);
264 }
265
266 return (error);
267 }
268
269 /*
270 * Read a TX_WRITE log data block.
271 */
272 static int
zil_read_log_data(zilog_t * zilog,const lr_write_t * lr,void * wbuf)273 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
274 {
275 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL;
276 const blkptr_t *bp = &lr->lr_blkptr;
277 arc_flags_t aflags = ARC_FLAG_WAIT;
278 arc_buf_t *abuf = NULL;
279 zbookmark_phys_t zb;
280 int error;
281
282 if (BP_IS_HOLE(bp)) {
283 if (wbuf != NULL)
284 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length));
285 return (0);
286 }
287
288 if (zilog->zl_header->zh_claim_txg == 0)
289 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
290
291 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
292 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
293
294 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
295 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
296
297 if (error == 0) {
298 if (wbuf != NULL)
299 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf));
300 arc_buf_destroy(abuf, &abuf);
301 }
302
303 return (error);
304 }
305
306 /*
307 * Parse the intent log, and call parse_func for each valid record within.
308 */
309 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)310 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
311 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg)
312 {
313 const zil_header_t *zh = zilog->zl_header;
314 boolean_t claimed = !!zh->zh_claim_txg;
315 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
316 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
317 uint64_t max_blk_seq = 0;
318 uint64_t max_lr_seq = 0;
319 uint64_t blk_count = 0;
320 uint64_t lr_count = 0;
321 blkptr_t blk, next_blk;
322 char *lrbuf, *lrp;
323 int error = 0;
324
325 /*
326 * Old logs didn't record the maximum zh_claim_lr_seq.
327 */
328 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
329 claim_lr_seq = UINT64_MAX;
330
331 /*
332 * Starting at the block pointed to by zh_log we read the log chain.
333 * For each block in the chain we strongly check that block to
334 * ensure its validity. We stop when an invalid block is found.
335 * For each block pointer in the chain we call parse_blk_func().
336 * For each record in each valid block we call parse_lr_func().
337 * If the log has been claimed, stop if we encounter a sequence
338 * number greater than the highest claimed sequence number.
339 */
340 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
341 zil_bp_tree_init(zilog);
342
343 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
344 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
345 int reclen;
346 char *end;
347
348 if (blk_seq > claim_blk_seq)
349 break;
350 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0)
351 break;
352 ASSERT3U(max_blk_seq, <, blk_seq);
353 max_blk_seq = blk_seq;
354 blk_count++;
355
356 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
357 break;
358
359 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end);
360 if (error != 0)
361 break;
362
363 for (lrp = lrbuf; lrp < end; lrp += reclen) {
364 lr_t *lr = (lr_t *)lrp;
365 reclen = lr->lrc_reclen;
366 ASSERT3U(reclen, >=, sizeof (lr_t));
367 if (lr->lrc_seq > claim_lr_seq)
368 goto done;
369 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0)
370 goto done;
371 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
372 max_lr_seq = lr->lrc_seq;
373 lr_count++;
374 }
375 }
376 done:
377 zilog->zl_parse_error = error;
378 zilog->zl_parse_blk_seq = max_blk_seq;
379 zilog->zl_parse_lr_seq = max_lr_seq;
380 zilog->zl_parse_blk_count = blk_count;
381 zilog->zl_parse_lr_count = lr_count;
382
383 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) ||
384 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
385
386 zil_bp_tree_fini(zilog);
387 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
388
389 return (error);
390 }
391
392 static int
zil_claim_log_block(zilog_t * zilog,blkptr_t * bp,void * tx,uint64_t first_txg)393 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg)
394 {
395 /*
396 * Claim log block if not already committed and not already claimed.
397 * If tx == NULL, just verify that the block is claimable.
398 */
399 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
400 zil_bp_tree_add(zilog, bp) != 0)
401 return (0);
402
403 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
404 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
405 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
406 }
407
408 static int
zil_claim_log_record(zilog_t * zilog,lr_t * lrc,void * tx,uint64_t first_txg)409 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg)
410 {
411 lr_write_t *lr = (lr_write_t *)lrc;
412 int error;
413
414 if (lrc->lrc_txtype != TX_WRITE)
415 return (0);
416
417 /*
418 * If the block is not readable, don't claim it. This can happen
419 * in normal operation when a log block is written to disk before
420 * some of the dmu_sync() blocks it points to. In this case, the
421 * transaction cannot have been committed to anyone (we would have
422 * waited for all writes to be stable first), so it is semantically
423 * correct to declare this the end of the log.
424 */
425 if (lr->lr_blkptr.blk_birth >= first_txg &&
426 (error = zil_read_log_data(zilog, lr, NULL)) != 0)
427 return (error);
428 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
429 }
430
431 /* ARGSUSED */
432 static int
zil_free_log_block(zilog_t * zilog,blkptr_t * bp,void * tx,uint64_t claim_txg)433 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg)
434 {
435 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
436
437 return (0);
438 }
439
440 static int
zil_free_log_record(zilog_t * zilog,lr_t * lrc,void * tx,uint64_t claim_txg)441 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg)
442 {
443 lr_write_t *lr = (lr_write_t *)lrc;
444 blkptr_t *bp = &lr->lr_blkptr;
445
446 /*
447 * If we previously claimed it, we need to free it.
448 */
449 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE &&
450 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
451 !BP_IS_HOLE(bp))
452 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
453
454 return (0);
455 }
456
457 static lwb_t *
zil_alloc_lwb(zilog_t * zilog,blkptr_t * bp,boolean_t slog,uint64_t txg)458 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg)
459 {
460 lwb_t *lwb;
461
462 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
463 lwb->lwb_zilog = zilog;
464 lwb->lwb_blk = *bp;
465 lwb->lwb_slog = slog;
466 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
467 lwb->lwb_max_txg = txg;
468 lwb->lwb_zio = NULL;
469 lwb->lwb_tx = NULL;
470 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
471 lwb->lwb_nused = sizeof (zil_chain_t);
472 lwb->lwb_sz = BP_GET_LSIZE(bp);
473 } else {
474 lwb->lwb_nused = 0;
475 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
476 }
477
478 mutex_enter(&zilog->zl_lock);
479 list_insert_tail(&zilog->zl_lwb_list, lwb);
480 mutex_exit(&zilog->zl_lock);
481
482 return (lwb);
483 }
484
485 /*
486 * Called when we create in-memory log transactions so that we know
487 * to cleanup the itxs at the end of spa_sync().
488 */
489 void
zilog_dirty(zilog_t * zilog,uint64_t txg)490 zilog_dirty(zilog_t *zilog, uint64_t txg)
491 {
492 dsl_pool_t *dp = zilog->zl_dmu_pool;
493 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
494
495 if (ds->ds_is_snapshot)
496 panic("dirtying snapshot!");
497
498 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
499 /* up the hold count until we can be written out */
500 dmu_buf_add_ref(ds->ds_dbuf, zilog);
501 }
502 }
503
504 /*
505 * Determine if the zil is dirty in the specified txg. Callers wanting to
506 * ensure that the dirty state does not change must hold the itxg_lock for
507 * the specified txg. Holding the lock will ensure that the zil cannot be
508 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
509 * state.
510 */
511 boolean_t
zilog_is_dirty_in_txg(zilog_t * zilog,uint64_t txg)512 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
513 {
514 dsl_pool_t *dp = zilog->zl_dmu_pool;
515
516 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
517 return (B_TRUE);
518 return (B_FALSE);
519 }
520
521 /*
522 * Determine if the zil is dirty. The zil is considered dirty if it has
523 * any pending itx records that have not been cleaned by zil_clean().
524 */
525 boolean_t
zilog_is_dirty(zilog_t * zilog)526 zilog_is_dirty(zilog_t *zilog)
527 {
528 dsl_pool_t *dp = zilog->zl_dmu_pool;
529
530 for (int t = 0; t < TXG_SIZE; t++) {
531 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
532 return (B_TRUE);
533 }
534 return (B_FALSE);
535 }
536
537 /*
538 * Create an on-disk intent log.
539 */
540 static lwb_t *
zil_create(zilog_t * zilog)541 zil_create(zilog_t *zilog)
542 {
543 const zil_header_t *zh = zilog->zl_header;
544 lwb_t *lwb = NULL;
545 uint64_t txg = 0;
546 dmu_tx_t *tx = NULL;
547 blkptr_t blk;
548 int error = 0;
549 boolean_t slog = FALSE;
550
551 /*
552 * Wait for any previous destroy to complete.
553 */
554 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
555
556 ASSERT(zh->zh_claim_txg == 0);
557 ASSERT(zh->zh_replay_seq == 0);
558
559 blk = zh->zh_log;
560
561 /*
562 * Allocate an initial log block if:
563 * - there isn't one already
564 * - the existing block is the wrong endianess
565 */
566 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
567 tx = dmu_tx_create(zilog->zl_os);
568 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
569 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
570 txg = dmu_tx_get_txg(tx);
571
572 if (!BP_IS_HOLE(&blk)) {
573 zio_free_zil(zilog->zl_spa, txg, &blk);
574 BP_ZERO(&blk);
575 }
576
577 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL,
578 ZIL_MIN_BLKSZ, &slog);
579
580 if (error == 0)
581 zil_init_log_chain(zilog, &blk);
582 }
583
584 /*
585 * Allocate a log write buffer (lwb) for the first log block.
586 */
587 if (error == 0)
588 lwb = zil_alloc_lwb(zilog, &blk, slog, txg);
589
590 /*
591 * If we just allocated the first log block, commit our transaction
592 * and wait for zil_sync() to stuff the block poiner into zh_log.
593 * (zh is part of the MOS, so we cannot modify it in open context.)
594 */
595 if (tx != NULL) {
596 dmu_tx_commit(tx);
597 txg_wait_synced(zilog->zl_dmu_pool, txg);
598 }
599
600 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
601
602 return (lwb);
603 }
604
605 /*
606 * In one tx, free all log blocks and clear the log header.
607 * If keep_first is set, then we're replaying a log with no content.
608 * We want to keep the first block, however, so that the first
609 * synchronous transaction doesn't require a txg_wait_synced()
610 * in zil_create(). We don't need to txg_wait_synced() here either
611 * when keep_first is set, because both zil_create() and zil_destroy()
612 * will wait for any in-progress destroys to complete.
613 */
614 void
zil_destroy(zilog_t * zilog,boolean_t keep_first)615 zil_destroy(zilog_t *zilog, boolean_t keep_first)
616 {
617 const zil_header_t *zh = zilog->zl_header;
618 lwb_t *lwb;
619 dmu_tx_t *tx;
620 uint64_t txg;
621
622 /*
623 * Wait for any previous destroy to complete.
624 */
625 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
626
627 zilog->zl_old_header = *zh; /* debugging aid */
628
629 if (BP_IS_HOLE(&zh->zh_log))
630 return;
631
632 tx = dmu_tx_create(zilog->zl_os);
633 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
634 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
635 txg = dmu_tx_get_txg(tx);
636
637 mutex_enter(&zilog->zl_lock);
638
639 ASSERT3U(zilog->zl_destroy_txg, <, txg);
640 zilog->zl_destroy_txg = txg;
641 zilog->zl_keep_first = keep_first;
642
643 if (!list_is_empty(&zilog->zl_lwb_list)) {
644 ASSERT(zh->zh_claim_txg == 0);
645 VERIFY(!keep_first);
646 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
647 list_remove(&zilog->zl_lwb_list, lwb);
648 if (lwb->lwb_buf != NULL)
649 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
650 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk);
651 kmem_cache_free(zil_lwb_cache, lwb);
652 }
653 } else if (!keep_first) {
654 zil_destroy_sync(zilog, tx);
655 }
656 mutex_exit(&zilog->zl_lock);
657
658 dmu_tx_commit(tx);
659 }
660
661 void
zil_destroy_sync(zilog_t * zilog,dmu_tx_t * tx)662 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
663 {
664 ASSERT(list_is_empty(&zilog->zl_lwb_list));
665 (void) zil_parse(zilog, zil_free_log_block,
666 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg);
667 }
668
669 int
zil_claim(dsl_pool_t * dp,dsl_dataset_t * ds,void * txarg)670 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
671 {
672 dmu_tx_t *tx = txarg;
673 uint64_t first_txg = dmu_tx_get_txg(tx);
674 zilog_t *zilog;
675 zil_header_t *zh;
676 objset_t *os;
677 int error;
678
679 error = dmu_objset_own_obj(dp, ds->ds_object,
680 DMU_OST_ANY, B_FALSE, FTAG, &os);
681 if (error != 0) {
682 /*
683 * EBUSY indicates that the objset is inconsistent, in which
684 * case it can not have a ZIL.
685 */
686 if (error != EBUSY) {
687 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
688 (unsigned long long)ds->ds_object, error);
689 }
690 return (0);
691 }
692
693 zilog = dmu_objset_zil(os);
694 zh = zil_header_in_syncing_context(zilog);
695
696 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) {
697 if (!BP_IS_HOLE(&zh->zh_log))
698 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log);
699 BP_ZERO(&zh->zh_log);
700 dsl_dataset_dirty(dmu_objset_ds(os), tx);
701 dmu_objset_disown(os, FTAG);
702 return (0);
703 }
704
705 /*
706 * Claim all log blocks if we haven't already done so, and remember
707 * the highest claimed sequence number. This ensures that if we can
708 * read only part of the log now (e.g. due to a missing device),
709 * but we can read the entire log later, we will not try to replay
710 * or destroy beyond the last block we successfully claimed.
711 */
712 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
713 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
714 (void) zil_parse(zilog, zil_claim_log_block,
715 zil_claim_log_record, tx, first_txg);
716 zh->zh_claim_txg = first_txg;
717 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
718 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
719 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
720 zh->zh_flags |= ZIL_REPLAY_NEEDED;
721 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
722 dsl_dataset_dirty(dmu_objset_ds(os), tx);
723 }
724
725 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
726 dmu_objset_disown(os, FTAG);
727 return (0);
728 }
729
730 /*
731 * Check the log by walking the log chain.
732 * Checksum errors are ok as they indicate the end of the chain.
733 * Any other error (no device or read failure) returns an error.
734 */
735 /* ARGSUSED */
736 int
zil_check_log_chain(dsl_pool_t * dp,dsl_dataset_t * ds,void * tx)737 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
738 {
739 zilog_t *zilog;
740 objset_t *os;
741 blkptr_t *bp;
742 int error;
743
744 ASSERT(tx == NULL);
745
746 error = dmu_objset_from_ds(ds, &os);
747 if (error != 0) {
748 cmn_err(CE_WARN, "can't open objset %llu, error %d",
749 (unsigned long long)ds->ds_object, error);
750 return (0);
751 }
752
753 zilog = dmu_objset_zil(os);
754 bp = (blkptr_t *)&zilog->zl_header->zh_log;
755
756 /*
757 * Check the first block and determine if it's on a log device
758 * which may have been removed or faulted prior to loading this
759 * pool. If so, there's no point in checking the rest of the log
760 * as its content should have already been synced to the pool.
761 */
762 if (!BP_IS_HOLE(bp)) {
763 vdev_t *vd;
764 boolean_t valid = B_TRUE;
765
766 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
767 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
768 if (vd->vdev_islog && vdev_is_dead(vd))
769 valid = vdev_log_state_valid(vd);
770 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
771
772 if (!valid)
773 return (0);
774 }
775
776 /*
777 * Because tx == NULL, zil_claim_log_block() will not actually claim
778 * any blocks, but just determine whether it is possible to do so.
779 * In addition to checking the log chain, zil_claim_log_block()
780 * will invoke zio_claim() with a done func of spa_claim_notify(),
781 * which will update spa_max_claim_txg. See spa_load() for details.
782 */
783 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
784 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa));
785
786 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
787 }
788
789 static int
zil_vdev_compare(const void * x1,const void * x2)790 zil_vdev_compare(const void *x1, const void *x2)
791 {
792 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
793 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
794
795 if (v1 < v2)
796 return (-1);
797 if (v1 > v2)
798 return (1);
799
800 return (0);
801 }
802
803 void
zil_add_block(zilog_t * zilog,const blkptr_t * bp)804 zil_add_block(zilog_t *zilog, const blkptr_t *bp)
805 {
806 avl_tree_t *t = &zilog->zl_vdev_tree;
807 avl_index_t where;
808 zil_vdev_node_t *zv, zvsearch;
809 int ndvas = BP_GET_NDVAS(bp);
810 int i;
811
812 if (zfs_nocacheflush)
813 return;
814
815 ASSERT(zilog->zl_writer);
816
817 /*
818 * Even though we're zl_writer, we still need a lock because the
819 * zl_get_data() callbacks may have dmu_sync() done callbacks
820 * that will run concurrently.
821 */
822 mutex_enter(&zilog->zl_vdev_lock);
823 for (i = 0; i < ndvas; i++) {
824 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
825 if (avl_find(t, &zvsearch, &where) == NULL) {
826 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
827 zv->zv_vdev = zvsearch.zv_vdev;
828 avl_insert(t, zv, where);
829 }
830 }
831 mutex_exit(&zilog->zl_vdev_lock);
832 }
833
834 static void
zil_flush_vdevs(zilog_t * zilog)835 zil_flush_vdevs(zilog_t *zilog)
836 {
837 spa_t *spa = zilog->zl_spa;
838 avl_tree_t *t = &zilog->zl_vdev_tree;
839 void *cookie = NULL;
840 zil_vdev_node_t *zv;
841 zio_t *zio = NULL;
842
843 ASSERT(zilog->zl_writer);
844
845 /*
846 * We don't need zl_vdev_lock here because we're the zl_writer,
847 * and all zl_get_data() callbacks are done.
848 */
849 if (avl_numnodes(t) == 0)
850 return;
851
852 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
853
854 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
855 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
856 if (vd != NULL && !vd->vdev_nowritecache) {
857 if (zio == NULL)
858 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
859 zio_flush(zio, vd);
860 }
861 kmem_free(zv, sizeof (*zv));
862 }
863
864 /*
865 * Wait for all the flushes to complete. Not all devices actually
866 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails.
867 */
868 if (zio)
869 (void) zio_wait(zio);
870
871 spa_config_exit(spa, SCL_STATE, FTAG);
872 }
873
874 /*
875 * Function called when a log block write completes
876 */
877 static void
zil_lwb_write_done(zio_t * zio)878 zil_lwb_write_done(zio_t *zio)
879 {
880 lwb_t *lwb = zio->io_private;
881 zilog_t *zilog = lwb->lwb_zilog;
882 dmu_tx_t *tx = lwb->lwb_tx;
883
884 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
885 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
886 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
887 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
888 ASSERT(!BP_IS_GANG(zio->io_bp));
889 ASSERT(!BP_IS_HOLE(zio->io_bp));
890 ASSERT(BP_GET_FILL(zio->io_bp) == 0);
891
892 /*
893 * Ensure the lwb buffer pointer is cleared before releasing
894 * the txg. If we have had an allocation failure and
895 * the txg is waiting to sync then we want want zil_sync()
896 * to remove the lwb so that it's not picked up as the next new
897 * one in zil_commit_writer(). zil_sync() will only remove
898 * the lwb if lwb_buf is null.
899 */
900 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
901 mutex_enter(&zilog->zl_lock);
902 lwb->lwb_buf = NULL;
903 lwb->lwb_tx = NULL;
904 mutex_exit(&zilog->zl_lock);
905
906 /*
907 * Now that we've written this log block, we have a stable pointer
908 * to the next block in the chain, so it's OK to let the txg in
909 * which we allocated the next block sync.
910 */
911 dmu_tx_commit(tx);
912 }
913
914 /*
915 * Initialize the io for a log block.
916 */
917 static void
zil_lwb_write_init(zilog_t * zilog,lwb_t * lwb)918 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
919 {
920 zbookmark_phys_t zb;
921 zio_priority_t prio;
922
923 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
924 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
925 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
926
927 if (zilog->zl_root_zio == NULL) {
928 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL,
929 ZIO_FLAG_CANFAIL);
930 }
931 if (lwb->lwb_zio == NULL) {
932 if (zilog->zl_cur_used <= zil_slog_limit || !lwb->lwb_slog)
933 prio = ZIO_PRIORITY_SYNC_WRITE;
934 else
935 prio = ZIO_PRIORITY_ASYNC_WRITE;
936 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa,
937 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk),
938 zil_lwb_write_done, lwb, prio,
939 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb);
940 }
941 }
942
943 /*
944 * Define a limited set of intent log block sizes.
945 *
946 * These must be a multiple of 4KB. Note only the amount used (again
947 * aligned to 4KB) actually gets written. However, we can't always just
948 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
949 */
950 uint64_t zil_block_buckets[] = {
951 4096, /* non TX_WRITE */
952 8192+4096, /* data base */
953 32*1024 + 4096, /* NFS writes */
954 UINT64_MAX
955 };
956
957 /*
958 * Start a log block write and advance to the next log block.
959 * Calls are serialized.
960 */
961 static lwb_t *
zil_lwb_write_start(zilog_t * zilog,lwb_t * lwb,boolean_t last)962 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb, boolean_t last)
963 {
964 lwb_t *nlwb = NULL;
965 zil_chain_t *zilc;
966 spa_t *spa = zilog->zl_spa;
967 blkptr_t *bp;
968 dmu_tx_t *tx;
969 uint64_t txg;
970 uint64_t zil_blksz, wsz;
971 int i, error;
972 boolean_t slog;
973
974 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
975 zilc = (zil_chain_t *)lwb->lwb_buf;
976 bp = &zilc->zc_next_blk;
977 } else {
978 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
979 bp = &zilc->zc_next_blk;
980 }
981
982 ASSERT(lwb->lwb_nused <= lwb->lwb_sz);
983
984 /*
985 * Allocate the next block and save its address in this block
986 * before writing it in order to establish the log chain.
987 * Note that if the allocation of nlwb synced before we wrote
988 * the block that points at it (lwb), we'd leak it if we crashed.
989 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done().
990 * We dirty the dataset to ensure that zil_sync() will be called
991 * to clean up in the event of allocation failure or I/O failure.
992 */
993 tx = dmu_tx_create(zilog->zl_os);
994 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
995 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
996 txg = dmu_tx_get_txg(tx);
997
998 lwb->lwb_tx = tx;
999
1000 /*
1001 * Log blocks are pre-allocated. Here we select the size of the next
1002 * block, based on size used in the last block.
1003 * - first find the smallest bucket that will fit the block from a
1004 * limited set of block sizes. This is because it's faster to write
1005 * blocks allocated from the same metaslab as they are adjacent or
1006 * close.
1007 * - next find the maximum from the new suggested size and an array of
1008 * previous sizes. This lessens a picket fence effect of wrongly
1009 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k
1010 * requests.
1011 *
1012 * Note we only write what is used, but we can't just allocate
1013 * the maximum block size because we can exhaust the available
1014 * pool log space.
1015 */
1016 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1017 for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
1018 continue;
1019 zil_blksz = zil_block_buckets[i];
1020 if (zil_blksz == UINT64_MAX)
1021 zil_blksz = SPA_OLD_MAXBLOCKSIZE;
1022 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1023 for (i = 0; i < ZIL_PREV_BLKS; i++)
1024 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1025 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1026
1027 BP_ZERO(bp);
1028 /* pass the old blkptr in order to spread log blocks across devs */
1029 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog);
1030 if (error == 0) {
1031 ASSERT3U(bp->blk_birth, ==, txg);
1032 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1033 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1034
1035 /*
1036 * Allocate a new log write buffer (lwb).
1037 */
1038 nlwb = zil_alloc_lwb(zilog, bp, slog, txg);
1039
1040 /* Record the block for later vdev flushing */
1041 zil_add_block(zilog, &lwb->lwb_blk);
1042 }
1043
1044 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1045 /* For Slim ZIL only write what is used. */
1046 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t);
1047 ASSERT3U(wsz, <=, lwb->lwb_sz);
1048 zio_shrink(lwb->lwb_zio, wsz);
1049
1050 } else {
1051 wsz = lwb->lwb_sz;
1052 }
1053
1054 zilc->zc_pad = 0;
1055 zilc->zc_nused = lwb->lwb_nused;
1056 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1057
1058 /*
1059 * clear unused data for security
1060 */
1061 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused);
1062
1063 if (last)
1064 lwb->lwb_zio->io_pipeline &= ~ZIO_STAGE_ISSUE_ASYNC;
1065 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */
1066
1067 /*
1068 * If there was an allocation failure then nlwb will be null which
1069 * forces a txg_wait_synced().
1070 */
1071 return (nlwb);
1072 }
1073
1074 static lwb_t *
zil_lwb_commit(zilog_t * zilog,itx_t * itx,lwb_t * lwb)1075 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
1076 {
1077 lr_t *lrcb, *lrc = &itx->itx_lr; /* common log record */
1078 lr_write_t *lrwb, *lrw = (lr_write_t *)lrc;
1079 char *lr_buf;
1080 uint64_t txg = lrc->lrc_txg;
1081 uint64_t reclen = lrc->lrc_reclen;
1082 uint64_t dlen = 0;
1083 uint64_t dnow, lwb_sp;
1084
1085 if (lwb == NULL)
1086 return (NULL);
1087
1088 ASSERT(lwb->lwb_buf != NULL);
1089
1090 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY)
1091 dlen = P2ROUNDUP_TYPED(
1092 lrw->lr_length, sizeof (uint64_t), uint64_t);
1093
1094 zilog->zl_cur_used += (reclen + dlen);
1095
1096 zil_lwb_write_init(zilog, lwb);
1097
1098 cont:
1099 /*
1100 * If this record won't fit in the current log block, start a new one.
1101 * For WR_NEED_COPY optimize layout for minimal number of chunks, but
1102 * try to keep wasted space withing reasonable range (12%).
1103 */
1104 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1105 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
1106 lwb_sp < ZIL_MAX_LOG_DATA / 8 && (dlen % ZIL_MAX_LOG_DATA == 0 ||
1107 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
1108 lwb = zil_lwb_write_start(zilog, lwb, B_FALSE);
1109 if (lwb == NULL)
1110 return (NULL);
1111 zil_lwb_write_init(zilog, lwb);
1112 ASSERT(LWB_EMPTY(lwb));
1113 lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
1114 ASSERT3U(reclen + MIN(dlen, sizeof(uint64_t)), <=, lwb_sp);
1115 }
1116
1117 dnow = MIN(dlen, lwb_sp - reclen);
1118 lr_buf = lwb->lwb_buf + lwb->lwb_nused;
1119 bcopy(lrc, lr_buf, reclen);
1120 lrcb = (lr_t *)lr_buf;
1121 lrwb = (lr_write_t *)lrcb;
1122
1123 /*
1124 * If it's a write, fetch the data or get its blkptr as appropriate.
1125 */
1126 if (lrc->lrc_txtype == TX_WRITE) {
1127 if (txg > spa_freeze_txg(zilog->zl_spa))
1128 txg_wait_synced(zilog->zl_dmu_pool, txg);
1129 if (itx->itx_wr_state != WR_COPIED) {
1130 char *dbuf;
1131 int error;
1132
1133 if (itx->itx_wr_state == WR_NEED_COPY) {
1134 dbuf = lr_buf + reclen;
1135 lrcb->lrc_reclen += dnow;
1136 if (lrwb->lr_length > dnow)
1137 lrwb->lr_length = dnow;
1138 lrw->lr_offset += dnow;
1139 lrw->lr_length -= dnow;
1140 } else {
1141 ASSERT(itx->itx_wr_state == WR_INDIRECT);
1142 dbuf = NULL;
1143 }
1144 error = zilog->zl_get_data(
1145 itx->itx_private, lrwb, dbuf, lwb->lwb_zio);
1146 if (error == EIO) {
1147 txg_wait_synced(zilog->zl_dmu_pool, txg);
1148 return (lwb);
1149 }
1150 if (error != 0) {
1151 ASSERT(error == ENOENT || error == EEXIST ||
1152 error == EALREADY);
1153 return (lwb);
1154 }
1155 }
1156 }
1157
1158 /*
1159 * We're actually making an entry, so update lrc_seq to be the
1160 * log record sequence number. Note that this is generally not
1161 * equal to the itx sequence number because not all transactions
1162 * are synchronous, and sometimes spa_sync() gets there first.
1163 */
1164 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */
1165 lwb->lwb_nused += reclen + dnow;
1166 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1167 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
1168 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
1169
1170 dlen -= dnow;
1171 if (dlen > 0) {
1172 zilog->zl_cur_used += reclen;
1173 goto cont;
1174 }
1175
1176 return (lwb);
1177 }
1178
1179 itx_t *
zil_itx_create(uint64_t txtype,size_t lrsize)1180 zil_itx_create(uint64_t txtype, size_t lrsize)
1181 {
1182 itx_t *itx;
1183
1184 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t);
1185
1186 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP);
1187 itx->itx_lr.lrc_txtype = txtype;
1188 itx->itx_lr.lrc_reclen = lrsize;
1189 itx->itx_lr.lrc_seq = 0; /* defensive */
1190 itx->itx_sync = B_TRUE; /* default is synchronous */
1191
1192 return (itx);
1193 }
1194
1195 void
zil_itx_destroy(itx_t * itx)1196 zil_itx_destroy(itx_t *itx)
1197 {
1198 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen);
1199 }
1200
1201 /*
1202 * Free up the sync and async itxs. The itxs_t has already been detached
1203 * so no locks are needed.
1204 */
1205 static void
zil_itxg_clean(itxs_t * itxs)1206 zil_itxg_clean(itxs_t *itxs)
1207 {
1208 itx_t *itx;
1209 list_t *list;
1210 avl_tree_t *t;
1211 void *cookie;
1212 itx_async_node_t *ian;
1213
1214 list = &itxs->i_sync_list;
1215 while ((itx = list_head(list)) != NULL) {
1216 list_remove(list, itx);
1217 kmem_free(itx, offsetof(itx_t, itx_lr) +
1218 itx->itx_lr.lrc_reclen);
1219 }
1220
1221 cookie = NULL;
1222 t = &itxs->i_async_tree;
1223 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1224 list = &ian->ia_list;
1225 while ((itx = list_head(list)) != NULL) {
1226 list_remove(list, itx);
1227 kmem_free(itx, offsetof(itx_t, itx_lr) +
1228 itx->itx_lr.lrc_reclen);
1229 }
1230 list_destroy(list);
1231 kmem_free(ian, sizeof (itx_async_node_t));
1232 }
1233 avl_destroy(t);
1234
1235 kmem_free(itxs, sizeof (itxs_t));
1236 }
1237
1238 static int
zil_aitx_compare(const void * x1,const void * x2)1239 zil_aitx_compare(const void *x1, const void *x2)
1240 {
1241 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
1242 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
1243
1244 if (o1 < o2)
1245 return (-1);
1246 if (o1 > o2)
1247 return (1);
1248
1249 return (0);
1250 }
1251
1252 /*
1253 * Remove all async itx with the given oid.
1254 */
1255 static void
zil_remove_async(zilog_t * zilog,uint64_t oid)1256 zil_remove_async(zilog_t *zilog, uint64_t oid)
1257 {
1258 uint64_t otxg, txg;
1259 itx_async_node_t *ian;
1260 avl_tree_t *t;
1261 avl_index_t where;
1262 list_t clean_list;
1263 itx_t *itx;
1264
1265 ASSERT(oid != 0);
1266 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
1267
1268 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1269 otxg = ZILTEST_TXG;
1270 else
1271 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1272
1273 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1274 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1275
1276 mutex_enter(&itxg->itxg_lock);
1277 if (itxg->itxg_txg != txg) {
1278 mutex_exit(&itxg->itxg_lock);
1279 continue;
1280 }
1281
1282 /*
1283 * Locate the object node and append its list.
1284 */
1285 t = &itxg->itxg_itxs->i_async_tree;
1286 ian = avl_find(t, &oid, &where);
1287 if (ian != NULL)
1288 list_move_tail(&clean_list, &ian->ia_list);
1289 mutex_exit(&itxg->itxg_lock);
1290 }
1291 while ((itx = list_head(&clean_list)) != NULL) {
1292 list_remove(&clean_list, itx);
1293 kmem_free(itx, offsetof(itx_t, itx_lr) +
1294 itx->itx_lr.lrc_reclen);
1295 }
1296 list_destroy(&clean_list);
1297 }
1298
1299 void
zil_itx_assign(zilog_t * zilog,itx_t * itx,dmu_tx_t * tx)1300 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
1301 {
1302 uint64_t txg;
1303 itxg_t *itxg;
1304 itxs_t *itxs, *clean = NULL;
1305
1306 /*
1307 * Object ids can be re-instantiated in the next txg so
1308 * remove any async transactions to avoid future leaks.
1309 * This can happen if a fsync occurs on the re-instantiated
1310 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets
1311 * the new file data and flushes a write record for the old object.
1312 */
1313 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE)
1314 zil_remove_async(zilog, itx->itx_oid);
1315
1316 /*
1317 * Ensure the data of a renamed file is committed before the rename.
1318 */
1319 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
1320 zil_async_to_sync(zilog, itx->itx_oid);
1321
1322 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
1323 txg = ZILTEST_TXG;
1324 else
1325 txg = dmu_tx_get_txg(tx);
1326
1327 itxg = &zilog->zl_itxg[txg & TXG_MASK];
1328 mutex_enter(&itxg->itxg_lock);
1329 itxs = itxg->itxg_itxs;
1330 if (itxg->itxg_txg != txg) {
1331 if (itxs != NULL) {
1332 /*
1333 * The zil_clean callback hasn't got around to cleaning
1334 * this itxg. Save the itxs for release below.
1335 * This should be rare.
1336 */
1337 clean = itxg->itxg_itxs;
1338 }
1339 itxg->itxg_txg = txg;
1340 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP);
1341
1342 list_create(&itxs->i_sync_list, sizeof (itx_t),
1343 offsetof(itx_t, itx_node));
1344 avl_create(&itxs->i_async_tree, zil_aitx_compare,
1345 sizeof (itx_async_node_t),
1346 offsetof(itx_async_node_t, ia_node));
1347 }
1348 if (itx->itx_sync) {
1349 list_insert_tail(&itxs->i_sync_list, itx);
1350 } else {
1351 avl_tree_t *t = &itxs->i_async_tree;
1352 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid;
1353 itx_async_node_t *ian;
1354 avl_index_t where;
1355
1356 ian = avl_find(t, &foid, &where);
1357 if (ian == NULL) {
1358 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP);
1359 list_create(&ian->ia_list, sizeof (itx_t),
1360 offsetof(itx_t, itx_node));
1361 ian->ia_foid = foid;
1362 avl_insert(t, ian, where);
1363 }
1364 list_insert_tail(&ian->ia_list, itx);
1365 }
1366
1367 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
1368 zilog_dirty(zilog, txg);
1369 mutex_exit(&itxg->itxg_lock);
1370
1371 /* Release the old itxs now we've dropped the lock */
1372 if (clean != NULL)
1373 zil_itxg_clean(clean);
1374 }
1375
1376 /*
1377 * If there are any in-memory intent log transactions which have now been
1378 * synced then start up a taskq to free them. We should only do this after we
1379 * have written out the uberblocks (i.e. txg has been comitted) so that
1380 * don't inadvertently clean out in-memory log records that would be required
1381 * by zil_commit().
1382 */
1383 void
zil_clean(zilog_t * zilog,uint64_t synced_txg)1384 zil_clean(zilog_t *zilog, uint64_t synced_txg)
1385 {
1386 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
1387 itxs_t *clean_me;
1388
1389 mutex_enter(&itxg->itxg_lock);
1390 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
1391 mutex_exit(&itxg->itxg_lock);
1392 return;
1393 }
1394 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
1395 ASSERT(itxg->itxg_txg != 0);
1396 ASSERT(zilog->zl_clean_taskq != NULL);
1397 clean_me = itxg->itxg_itxs;
1398 itxg->itxg_itxs = NULL;
1399 itxg->itxg_txg = 0;
1400 mutex_exit(&itxg->itxg_lock);
1401 /*
1402 * Preferably start a task queue to free up the old itxs but
1403 * if taskq_dispatch can't allocate resources to do that then
1404 * free it in-line. This should be rare. Note, using TQ_SLEEP
1405 * created a bad performance problem.
1406 */
1407 if (taskq_dispatch(zilog->zl_clean_taskq,
1408 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
1409 zil_itxg_clean(clean_me);
1410 }
1411
1412 /*
1413 * Get the list of itxs to commit into zl_itx_commit_list.
1414 */
1415 static void
zil_get_commit_list(zilog_t * zilog)1416 zil_get_commit_list(zilog_t *zilog)
1417 {
1418 uint64_t otxg, txg;
1419 list_t *commit_list = &zilog->zl_itx_commit_list;
1420
1421 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1422 otxg = ZILTEST_TXG;
1423 else
1424 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1425
1426 /*
1427 * This is inherently racy, since there is nothing to prevent
1428 * the last synced txg from changing. That's okay since we'll
1429 * only commit things in the future.
1430 */
1431 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1432 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1433
1434 mutex_enter(&itxg->itxg_lock);
1435 if (itxg->itxg_txg != txg) {
1436 mutex_exit(&itxg->itxg_lock);
1437 continue;
1438 }
1439
1440 /*
1441 * If we're adding itx records to the zl_itx_commit_list,
1442 * then the zil better be dirty in this "txg". We can assert
1443 * that here since we're holding the itxg_lock which will
1444 * prevent spa_sync from cleaning it. Once we add the itxs
1445 * to the zl_itx_commit_list we must commit it to disk even
1446 * if it's unnecessary (i.e. the txg was synced).
1447 */
1448 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
1449 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
1450 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list);
1451
1452 mutex_exit(&itxg->itxg_lock);
1453 }
1454 }
1455
1456 /*
1457 * Move the async itxs for a specified object to commit into sync lists.
1458 */
1459 void
zil_async_to_sync(zilog_t * zilog,uint64_t foid)1460 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
1461 {
1462 uint64_t otxg, txg;
1463 itx_async_node_t *ian;
1464 avl_tree_t *t;
1465 avl_index_t where;
1466
1467 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
1468 otxg = ZILTEST_TXG;
1469 else
1470 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
1471
1472 /*
1473 * This is inherently racy, since there is nothing to prevent
1474 * the last synced txg from changing.
1475 */
1476 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
1477 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
1478
1479 mutex_enter(&itxg->itxg_lock);
1480 if (itxg->itxg_txg != txg) {
1481 mutex_exit(&itxg->itxg_lock);
1482 continue;
1483 }
1484
1485 /*
1486 * If a foid is specified then find that node and append its
1487 * list. Otherwise walk the tree appending all the lists
1488 * to the sync list. We add to the end rather than the
1489 * beginning to ensure the create has happened.
1490 */
1491 t = &itxg->itxg_itxs->i_async_tree;
1492 if (foid != 0) {
1493 ian = avl_find(t, &foid, &where);
1494 if (ian != NULL) {
1495 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1496 &ian->ia_list);
1497 }
1498 } else {
1499 void *cookie = NULL;
1500
1501 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
1502 list_move_tail(&itxg->itxg_itxs->i_sync_list,
1503 &ian->ia_list);
1504 list_destroy(&ian->ia_list);
1505 kmem_free(ian, sizeof (itx_async_node_t));
1506 }
1507 }
1508 mutex_exit(&itxg->itxg_lock);
1509 }
1510 }
1511
1512 static void
zil_commit_writer(zilog_t * zilog)1513 zil_commit_writer(zilog_t *zilog)
1514 {
1515 uint64_t txg;
1516 itx_t *itx;
1517 lwb_t *lwb;
1518 spa_t *spa = zilog->zl_spa;
1519 int error = 0;
1520
1521 ASSERT(zilog->zl_root_zio == NULL);
1522
1523 mutex_exit(&zilog->zl_lock);
1524
1525 zil_get_commit_list(zilog);
1526
1527 /*
1528 * Return if there's nothing to commit before we dirty the fs by
1529 * calling zil_create().
1530 */
1531 if (list_head(&zilog->zl_itx_commit_list) == NULL) {
1532 mutex_enter(&zilog->zl_lock);
1533 return;
1534 }
1535
1536 if (zilog->zl_suspend) {
1537 lwb = NULL;
1538 } else {
1539 lwb = list_tail(&zilog->zl_lwb_list);
1540 if (lwb == NULL)
1541 lwb = zil_create(zilog);
1542 }
1543
1544 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog);
1545 while (itx = list_head(&zilog->zl_itx_commit_list)) {
1546 txg = itx->itx_lr.lrc_txg;
1547 ASSERT3U(txg, !=, 0);
1548
1549 /*
1550 * This is inherently racy and may result in us writing
1551 * out a log block for a txg that was just synced. This is
1552 * ok since we'll end cleaning up that log block the next
1553 * time we call zil_sync().
1554 */
1555 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa))
1556 lwb = zil_lwb_commit(zilog, itx, lwb);
1557 list_remove(&zilog->zl_itx_commit_list, itx);
1558 kmem_free(itx, offsetof(itx_t, itx_lr)
1559 + itx->itx_lr.lrc_reclen);
1560 }
1561 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
1562
1563 /* write the last block out */
1564 if (lwb != NULL && lwb->lwb_zio != NULL)
1565 lwb = zil_lwb_write_start(zilog, lwb, B_TRUE);
1566
1567 zilog->zl_cur_used = 0;
1568
1569 /*
1570 * Wait if necessary for the log blocks to be on stable storage.
1571 */
1572 if (zilog->zl_root_zio) {
1573 error = zio_wait(zilog->zl_root_zio);
1574 zilog->zl_root_zio = NULL;
1575 zil_flush_vdevs(zilog);
1576 }
1577
1578 if (error || lwb == NULL)
1579 txg_wait_synced(zilog->zl_dmu_pool, 0);
1580
1581 mutex_enter(&zilog->zl_lock);
1582
1583 /*
1584 * Remember the highest committed log sequence number for ztest.
1585 * We only update this value when all the log writes succeeded,
1586 * because ztest wants to ASSERT that it got the whole log chain.
1587 */
1588 if (error == 0 && lwb != NULL)
1589 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1590 }
1591
1592 /*
1593 * Commit zfs transactions to stable storage.
1594 * If foid is 0 push out all transactions, otherwise push only those
1595 * for that object or might reference that object.
1596 *
1597 * itxs are committed in batches. In a heavily stressed zil there will be
1598 * a commit writer thread who is writing out a bunch of itxs to the log
1599 * for a set of committing threads (cthreads) in the same batch as the writer.
1600 * Those cthreads are all waiting on the same cv for that batch.
1601 *
1602 * There will also be a different and growing batch of threads that are
1603 * waiting to commit (qthreads). When the committing batch completes
1604 * a transition occurs such that the cthreads exit and the qthreads become
1605 * cthreads. One of the new cthreads becomes the writer thread for the
1606 * batch. Any new threads arriving become new qthreads.
1607 *
1608 * Only 2 condition variables are needed and there's no transition
1609 * between the two cvs needed. They just flip-flop between qthreads
1610 * and cthreads.
1611 *
1612 * Using this scheme we can efficiently wakeup up only those threads
1613 * that have been committed.
1614 */
1615 void
zil_commit(zilog_t * zilog,uint64_t foid)1616 zil_commit(zilog_t *zilog, uint64_t foid)
1617 {
1618 uint64_t mybatch;
1619
1620 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
1621 return;
1622
1623 /* move the async itxs for the foid to the sync queues */
1624 zil_async_to_sync(zilog, foid);
1625
1626 mutex_enter(&zilog->zl_lock);
1627 mybatch = zilog->zl_next_batch;
1628 while (zilog->zl_writer) {
1629 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock);
1630 if (mybatch <= zilog->zl_com_batch) {
1631 mutex_exit(&zilog->zl_lock);
1632 return;
1633 }
1634 }
1635
1636 zilog->zl_next_batch++;
1637 zilog->zl_writer = B_TRUE;
1638 zil_commit_writer(zilog);
1639 zilog->zl_com_batch = mybatch;
1640 zilog->zl_writer = B_FALSE;
1641 mutex_exit(&zilog->zl_lock);
1642
1643 /* wake up one thread to become the next writer */
1644 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]);
1645
1646 /* wake up all threads waiting for this batch to be committed */
1647 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]);
1648 }
1649
1650 /*
1651 * Called in syncing context to free committed log blocks and update log header.
1652 */
1653 void
zil_sync(zilog_t * zilog,dmu_tx_t * tx)1654 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
1655 {
1656 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1657 uint64_t txg = dmu_tx_get_txg(tx);
1658 spa_t *spa = zilog->zl_spa;
1659 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
1660 lwb_t *lwb;
1661
1662 /*
1663 * We don't zero out zl_destroy_txg, so make sure we don't try
1664 * to destroy it twice.
1665 */
1666 if (spa_sync_pass(spa) != 1)
1667 return;
1668
1669 mutex_enter(&zilog->zl_lock);
1670
1671 ASSERT(zilog->zl_stop_sync == 0);
1672
1673 if (*replayed_seq != 0) {
1674 ASSERT(zh->zh_replay_seq < *replayed_seq);
1675 zh->zh_replay_seq = *replayed_seq;
1676 *replayed_seq = 0;
1677 }
1678
1679 if (zilog->zl_destroy_txg == txg) {
1680 blkptr_t blk = zh->zh_log;
1681
1682 ASSERT(list_head(&zilog->zl_lwb_list) == NULL);
1683
1684 bzero(zh, sizeof (zil_header_t));
1685 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
1686
1687 if (zilog->zl_keep_first) {
1688 /*
1689 * If this block was part of log chain that couldn't
1690 * be claimed because a device was missing during
1691 * zil_claim(), but that device later returns,
1692 * then this block could erroneously appear valid.
1693 * To guard against this, assign a new GUID to the new
1694 * log chain so it doesn't matter what blk points to.
1695 */
1696 zil_init_log_chain(zilog, &blk);
1697 zh->zh_log = blk;
1698 }
1699 }
1700
1701 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
1702 zh->zh_log = lwb->lwb_blk;
1703 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg)
1704 break;
1705 list_remove(&zilog->zl_lwb_list, lwb);
1706 zio_free_zil(spa, txg, &lwb->lwb_blk);
1707 kmem_cache_free(zil_lwb_cache, lwb);
1708
1709 /*
1710 * If we don't have anything left in the lwb list then
1711 * we've had an allocation failure and we need to zero
1712 * out the zil_header blkptr so that we don't end
1713 * up freeing the same block twice.
1714 */
1715 if (list_head(&zilog->zl_lwb_list) == NULL)
1716 BP_ZERO(&zh->zh_log);
1717 }
1718 mutex_exit(&zilog->zl_lock);
1719 }
1720
1721 void
zil_init(void)1722 zil_init(void)
1723 {
1724 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
1725 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0);
1726 }
1727
1728 void
zil_fini(void)1729 zil_fini(void)
1730 {
1731 kmem_cache_destroy(zil_lwb_cache);
1732 }
1733
1734 void
zil_set_sync(zilog_t * zilog,uint64_t sync)1735 zil_set_sync(zilog_t *zilog, uint64_t sync)
1736 {
1737 zilog->zl_sync = sync;
1738 }
1739
1740 void
zil_set_logbias(zilog_t * zilog,uint64_t logbias)1741 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
1742 {
1743 zilog->zl_logbias = logbias;
1744 }
1745
1746 zilog_t *
zil_alloc(objset_t * os,zil_header_t * zh_phys)1747 zil_alloc(objset_t *os, zil_header_t *zh_phys)
1748 {
1749 zilog_t *zilog;
1750
1751 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
1752
1753 zilog->zl_header = zh_phys;
1754 zilog->zl_os = os;
1755 zilog->zl_spa = dmu_objset_spa(os);
1756 zilog->zl_dmu_pool = dmu_objset_pool(os);
1757 zilog->zl_destroy_txg = TXG_INITIAL - 1;
1758 zilog->zl_logbias = dmu_objset_logbias(os);
1759 zilog->zl_sync = dmu_objset_syncprop(os);
1760 zilog->zl_next_batch = 1;
1761
1762 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
1763
1764 for (int i = 0; i < TXG_SIZE; i++) {
1765 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
1766 MUTEX_DEFAULT, NULL);
1767 }
1768
1769 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
1770 offsetof(lwb_t, lwb_node));
1771
1772 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
1773 offsetof(itx_t, itx_node));
1774
1775 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
1776
1777 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare,
1778 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
1779
1780 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL);
1781 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
1782 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL);
1783 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL);
1784
1785 return (zilog);
1786 }
1787
1788 void
zil_free(zilog_t * zilog)1789 zil_free(zilog_t *zilog)
1790 {
1791 zilog->zl_stop_sync = 1;
1792
1793 ASSERT0(zilog->zl_suspend);
1794 ASSERT0(zilog->zl_suspending);
1795
1796 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1797 list_destroy(&zilog->zl_lwb_list);
1798
1799 avl_destroy(&zilog->zl_vdev_tree);
1800 mutex_destroy(&zilog->zl_vdev_lock);
1801
1802 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
1803 list_destroy(&zilog->zl_itx_commit_list);
1804
1805 for (int i = 0; i < TXG_SIZE; i++) {
1806 /*
1807 * It's possible for an itx to be generated that doesn't dirty
1808 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
1809 * callback to remove the entry. We remove those here.
1810 *
1811 * Also free up the ziltest itxs.
1812 */
1813 if (zilog->zl_itxg[i].itxg_itxs)
1814 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
1815 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
1816 }
1817
1818 mutex_destroy(&zilog->zl_lock);
1819
1820 cv_destroy(&zilog->zl_cv_writer);
1821 cv_destroy(&zilog->zl_cv_suspend);
1822 cv_destroy(&zilog->zl_cv_batch[0]);
1823 cv_destroy(&zilog->zl_cv_batch[1]);
1824
1825 kmem_free(zilog, sizeof (zilog_t));
1826 }
1827
1828 /*
1829 * Open an intent log.
1830 */
1831 zilog_t *
zil_open(objset_t * os,zil_get_data_t * get_data)1832 zil_open(objset_t *os, zil_get_data_t *get_data)
1833 {
1834 zilog_t *zilog = dmu_objset_zil(os);
1835
1836 ASSERT(zilog->zl_clean_taskq == NULL);
1837 ASSERT(zilog->zl_get_data == NULL);
1838 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1839
1840 zilog->zl_get_data = get_data;
1841 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
1842 2, 2, TASKQ_PREPOPULATE);
1843
1844 return (zilog);
1845 }
1846
1847 /*
1848 * Close an intent log.
1849 */
1850 void
zil_close(zilog_t * zilog)1851 zil_close(zilog_t *zilog)
1852 {
1853 lwb_t *lwb;
1854 uint64_t txg = 0;
1855
1856 zil_commit(zilog, 0); /* commit all itx */
1857
1858 /*
1859 * The lwb_max_txg for the stubby lwb will reflect the last activity
1860 * for the zil. After a txg_wait_synced() on the txg we know all the
1861 * callbacks have occurred that may clean the zil. Only then can we
1862 * destroy the zl_clean_taskq.
1863 */
1864 mutex_enter(&zilog->zl_lock);
1865 lwb = list_tail(&zilog->zl_lwb_list);
1866 if (lwb != NULL)
1867 txg = lwb->lwb_max_txg;
1868 mutex_exit(&zilog->zl_lock);
1869 if (txg)
1870 txg_wait_synced(zilog->zl_dmu_pool, txg);
1871
1872 if (zilog_is_dirty(zilog))
1873 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
1874 VERIFY(!zilog_is_dirty(zilog));
1875
1876 taskq_destroy(zilog->zl_clean_taskq);
1877 zilog->zl_clean_taskq = NULL;
1878 zilog->zl_get_data = NULL;
1879
1880 /*
1881 * We should have only one LWB left on the list; remove it now.
1882 */
1883 mutex_enter(&zilog->zl_lock);
1884 lwb = list_head(&zilog->zl_lwb_list);
1885 if (lwb != NULL) {
1886 ASSERT(lwb == list_tail(&zilog->zl_lwb_list));
1887 list_remove(&zilog->zl_lwb_list, lwb);
1888 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1889 kmem_cache_free(zil_lwb_cache, lwb);
1890 }
1891 mutex_exit(&zilog->zl_lock);
1892 }
1893
1894 static char *suspend_tag = "zil suspending";
1895
1896 /*
1897 * Suspend an intent log. While in suspended mode, we still honor
1898 * synchronous semantics, but we rely on txg_wait_synced() to do it.
1899 * On old version pools, we suspend the log briefly when taking a
1900 * snapshot so that it will have an empty intent log.
1901 *
1902 * Long holds are not really intended to be used the way we do here --
1903 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
1904 * could fail. Therefore we take pains to only put a long hold if it is
1905 * actually necessary. Fortunately, it will only be necessary if the
1906 * objset is currently mounted (or the ZVOL equivalent). In that case it
1907 * will already have a long hold, so we are not really making things any worse.
1908 *
1909 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
1910 * zvol_state_t), and use their mechanism to prevent their hold from being
1911 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
1912 * very little gain.
1913 *
1914 * if cookiep == NULL, this does both the suspend & resume.
1915 * Otherwise, it returns with the dataset "long held", and the cookie
1916 * should be passed into zil_resume().
1917 */
1918 int
zil_suspend(const char * osname,void ** cookiep)1919 zil_suspend(const char *osname, void **cookiep)
1920 {
1921 objset_t *os;
1922 zilog_t *zilog;
1923 const zil_header_t *zh;
1924 int error;
1925
1926 error = dmu_objset_hold(osname, suspend_tag, &os);
1927 if (error != 0)
1928 return (error);
1929 zilog = dmu_objset_zil(os);
1930
1931 mutex_enter(&zilog->zl_lock);
1932 zh = zilog->zl_header;
1933
1934 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
1935 mutex_exit(&zilog->zl_lock);
1936 dmu_objset_rele(os, suspend_tag);
1937 return (SET_ERROR(EBUSY));
1938 }
1939
1940 /*
1941 * Don't put a long hold in the cases where we can avoid it. This
1942 * is when there is no cookie so we are doing a suspend & resume
1943 * (i.e. called from zil_vdev_offline()), and there's nothing to do
1944 * for the suspend because it's already suspended, or there's no ZIL.
1945 */
1946 if (cookiep == NULL && !zilog->zl_suspending &&
1947 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
1948 mutex_exit(&zilog->zl_lock);
1949 dmu_objset_rele(os, suspend_tag);
1950 return (0);
1951 }
1952
1953 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
1954 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
1955
1956 zilog->zl_suspend++;
1957
1958 if (zilog->zl_suspend > 1) {
1959 /*
1960 * Someone else is already suspending it.
1961 * Just wait for them to finish.
1962 */
1963
1964 while (zilog->zl_suspending)
1965 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
1966 mutex_exit(&zilog->zl_lock);
1967
1968 if (cookiep == NULL)
1969 zil_resume(os);
1970 else
1971 *cookiep = os;
1972 return (0);
1973 }
1974
1975 /*
1976 * If there is no pointer to an on-disk block, this ZIL must not
1977 * be active (e.g. filesystem not mounted), so there's nothing
1978 * to clean up.
1979 */
1980 if (BP_IS_HOLE(&zh->zh_log)) {
1981 ASSERT(cookiep != NULL); /* fast path already handled */
1982
1983 *cookiep = os;
1984 mutex_exit(&zilog->zl_lock);
1985 return (0);
1986 }
1987
1988 zilog->zl_suspending = B_TRUE;
1989 mutex_exit(&zilog->zl_lock);
1990
1991 zil_commit(zilog, 0);
1992
1993 zil_destroy(zilog, B_FALSE);
1994
1995 mutex_enter(&zilog->zl_lock);
1996 zilog->zl_suspending = B_FALSE;
1997 cv_broadcast(&zilog->zl_cv_suspend);
1998 mutex_exit(&zilog->zl_lock);
1999
2000 if (cookiep == NULL)
2001 zil_resume(os);
2002 else
2003 *cookiep = os;
2004 return (0);
2005 }
2006
2007 void
zil_resume(void * cookie)2008 zil_resume(void *cookie)
2009 {
2010 objset_t *os = cookie;
2011 zilog_t *zilog = dmu_objset_zil(os);
2012
2013 mutex_enter(&zilog->zl_lock);
2014 ASSERT(zilog->zl_suspend != 0);
2015 zilog->zl_suspend--;
2016 mutex_exit(&zilog->zl_lock);
2017 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
2018 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
2019 }
2020
2021 typedef struct zil_replay_arg {
2022 zil_replay_func_t **zr_replay;
2023 void *zr_arg;
2024 boolean_t zr_byteswap;
2025 char *zr_lr;
2026 } zil_replay_arg_t;
2027
2028 static int
zil_replay_error(zilog_t * zilog,lr_t * lr,int error)2029 zil_replay_error(zilog_t *zilog, lr_t *lr, int error)
2030 {
2031 char name[ZFS_MAX_DATASET_NAME_LEN];
2032
2033 zilog->zl_replaying_seq--; /* didn't actually replay this one */
2034
2035 dmu_objset_name(zilog->zl_os, name);
2036
2037 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
2038 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
2039 (u_longlong_t)lr->lrc_seq,
2040 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
2041 (lr->lrc_txtype & TX_CI) ? "CI" : "");
2042
2043 return (error);
2044 }
2045
2046 static int
zil_replay_log_record(zilog_t * zilog,lr_t * lr,void * zra,uint64_t claim_txg)2047 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg)
2048 {
2049 zil_replay_arg_t *zr = zra;
2050 const zil_header_t *zh = zilog->zl_header;
2051 uint64_t reclen = lr->lrc_reclen;
2052 uint64_t txtype = lr->lrc_txtype;
2053 int error = 0;
2054
2055 zilog->zl_replaying_seq = lr->lrc_seq;
2056
2057 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
2058 return (0);
2059
2060 if (lr->lrc_txg < claim_txg) /* already committed */
2061 return (0);
2062
2063 /* Strip case-insensitive bit, still present in log record */
2064 txtype &= ~TX_CI;
2065
2066 if (txtype == 0 || txtype >= TX_MAX_TYPE)
2067 return (zil_replay_error(zilog, lr, EINVAL));
2068
2069 /*
2070 * If this record type can be logged out of order, the object
2071 * (lr_foid) may no longer exist. That's legitimate, not an error.
2072 */
2073 if (TX_OOO(txtype)) {
2074 error = dmu_object_info(zilog->zl_os,
2075 ((lr_ooo_t *)lr)->lr_foid, NULL);
2076 if (error == ENOENT || error == EEXIST)
2077 return (0);
2078 }
2079
2080 /*
2081 * Make a copy of the data so we can revise and extend it.
2082 */
2083 bcopy(lr, zr->zr_lr, reclen);
2084
2085 /*
2086 * If this is a TX_WRITE with a blkptr, suck in the data.
2087 */
2088 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
2089 error = zil_read_log_data(zilog, (lr_write_t *)lr,
2090 zr->zr_lr + reclen);
2091 if (error != 0)
2092 return (zil_replay_error(zilog, lr, error));
2093 }
2094
2095 /*
2096 * The log block containing this lr may have been byteswapped
2097 * so that we can easily examine common fields like lrc_txtype.
2098 * However, the log is a mix of different record types, and only the
2099 * replay vectors know how to byteswap their records. Therefore, if
2100 * the lr was byteswapped, undo it before invoking the replay vector.
2101 */
2102 if (zr->zr_byteswap)
2103 byteswap_uint64_array(zr->zr_lr, reclen);
2104
2105 /*
2106 * We must now do two things atomically: replay this log record,
2107 * and update the log header sequence number to reflect the fact that
2108 * we did so. At the end of each replay function the sequence number
2109 * is updated if we are in replay mode.
2110 */
2111 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
2112 if (error != 0) {
2113 /*
2114 * The DMU's dnode layer doesn't see removes until the txg
2115 * commits, so a subsequent claim can spuriously fail with
2116 * EEXIST. So if we receive any error we try syncing out
2117 * any removes then retry the transaction. Note that we
2118 * specify B_FALSE for byteswap now, so we don't do it twice.
2119 */
2120 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
2121 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
2122 if (error != 0)
2123 return (zil_replay_error(zilog, lr, error));
2124 }
2125 return (0);
2126 }
2127
2128 /* ARGSUSED */
2129 static int
zil_incr_blks(zilog_t * zilog,blkptr_t * bp,void * arg,uint64_t claim_txg)2130 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
2131 {
2132 zilog->zl_replay_blks++;
2133
2134 return (0);
2135 }
2136
2137 /*
2138 * If this dataset has a non-empty intent log, replay it and destroy it.
2139 */
2140 void
zil_replay(objset_t * os,void * arg,zil_replay_func_t * replay_func[TX_MAX_TYPE])2141 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
2142 {
2143 zilog_t *zilog = dmu_objset_zil(os);
2144 const zil_header_t *zh = zilog->zl_header;
2145 zil_replay_arg_t zr;
2146
2147 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
2148 zil_destroy(zilog, B_TRUE);
2149 return;
2150 }
2151
2152 zr.zr_replay = replay_func;
2153 zr.zr_arg = arg;
2154 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
2155 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
2156
2157 /*
2158 * Wait for in-progress removes to sync before starting replay.
2159 */
2160 txg_wait_synced(zilog->zl_dmu_pool, 0);
2161
2162 zilog->zl_replay = B_TRUE;
2163 zilog->zl_replay_time = ddi_get_lbolt();
2164 ASSERT(zilog->zl_replay_blks == 0);
2165 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
2166 zh->zh_claim_txg);
2167 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
2168
2169 zil_destroy(zilog, B_FALSE);
2170 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
2171 zilog->zl_replay = B_FALSE;
2172 }
2173
2174 boolean_t
zil_replaying(zilog_t * zilog,dmu_tx_t * tx)2175 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
2176 {
2177 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
2178 return (B_TRUE);
2179
2180 if (zilog->zl_replay) {
2181 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
2182 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
2183 zilog->zl_replaying_seq;
2184 return (B_TRUE);
2185 }
2186
2187 return (B_FALSE);
2188 }
2189
2190 /* ARGSUSED */
2191 int
zil_vdev_offline(const char * osname,void * arg)2192 zil_vdev_offline(const char *osname, void *arg)
2193 {
2194 int error;
2195
2196 error = zil_suspend(osname, NULL);
2197 if (error != 0)
2198 return (SET_ERROR(EEXIST));
2199 return (0);
2200 }
2201