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 https://opensource.org/licenses/CDDL-1.0.
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, 2018 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Integros [integros.com]
25 * Copyright (c) 2018 Datto Inc.
26 */
27
28 /* Portions Copyright 2010 Robert Milkowski */
29
30 #include <sys/zfs_context.h>
31 #include <sys/spa.h>
32 #include <sys/spa_impl.h>
33 #include <sys/dmu.h>
34 #include <sys/zap.h>
35 #include <sys/arc.h>
36 #include <sys/stat.h>
37 #include <sys/zil.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
43 #include <sys/metaslab.h>
44 #include <sys/trace_zfs.h>
45 #include <sys/abd.h>
46 #include <sys/brt.h>
47 #include <sys/wmsum.h>
48
49 /*
50 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
51 * calls that change the file system. Each itx has enough information to
52 * be able to replay them after a system crash, power loss, or
53 * equivalent failure mode. These are stored in memory until either:
54 *
55 * 1. they are committed to the pool by the DMU transaction group
56 * (txg), at which point they can be discarded; or
57 * 2. they are committed to the on-disk ZIL for the dataset being
58 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous
59 * requirement).
60 *
61 * In the event of a crash or power loss, the itxs contained by each
62 * dataset's on-disk ZIL will be replayed when that dataset is first
63 * instantiated (e.g. if the dataset is a normal filesystem, when it is
64 * first mounted).
65 *
66 * As hinted at above, there is one ZIL per dataset (both the in-memory
67 * representation, and the on-disk representation). The on-disk format
68 * consists of 3 parts:
69 *
70 * - a single, per-dataset, ZIL header; which points to a chain of
71 * - zero or more ZIL blocks; each of which contains
72 * - zero or more ZIL records
73 *
74 * A ZIL record holds the information necessary to replay a single
75 * system call transaction. A ZIL block can hold many ZIL records, and
76 * the blocks are chained together, similarly to a singly linked list.
77 *
78 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
79 * block in the chain, and the ZIL header points to the first block in
80 * the chain.
81 *
82 * Note, there is not a fixed place in the pool to hold these ZIL
83 * blocks; they are dynamically allocated and freed as needed from the
84 * blocks available on the pool, though they can be preferentially
85 * allocated from a dedicated "log" vdev.
86 */
87
88 /*
89 * This controls the amount of time that a ZIL block (lwb) will remain
90 * "open" when it isn't "full", and it has a thread waiting for it to be
91 * committed to stable storage. Please refer to the zil_commit_waiter()
92 * function (and the comments within it) for more details.
93 */
94 static uint_t zfs_commit_timeout_pct = 10;
95
96 /*
97 * See zil.h for more information about these fields.
98 */
99 static zil_kstat_values_t zil_stats = {
100 { "zil_commit_count", KSTAT_DATA_UINT64 },
101 { "zil_commit_writer_count", KSTAT_DATA_UINT64 },
102 { "zil_itx_count", KSTAT_DATA_UINT64 },
103 { "zil_itx_indirect_count", KSTAT_DATA_UINT64 },
104 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 },
105 { "zil_itx_copied_count", KSTAT_DATA_UINT64 },
106 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 },
107 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 },
108 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 },
109 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 },
110 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 },
111 { "zil_itx_metaslab_normal_write", KSTAT_DATA_UINT64 },
112 { "zil_itx_metaslab_normal_alloc", KSTAT_DATA_UINT64 },
113 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 },
114 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 },
115 { "zil_itx_metaslab_slog_write", KSTAT_DATA_UINT64 },
116 { "zil_itx_metaslab_slog_alloc", KSTAT_DATA_UINT64 },
117 };
118
119 static zil_sums_t zil_sums_global;
120 static kstat_t *zil_kstats_global;
121
122 /*
123 * Disable intent logging replay. This global ZIL switch affects all pools.
124 */
125 int zil_replay_disable = 0;
126
127 /*
128 * Disable the flush commands that are normally sent to the disk(s) by the ZIL
129 * after an LWB write has completed. Setting this will cause ZIL corruption on
130 * power loss if a volatile out-of-order write cache is enabled.
131 */
132 static int zil_nocacheflush = 0;
133
134 /*
135 * Limit SLOG write size per commit executed with synchronous priority.
136 * Any writes above that will be executed with lower (asynchronous) priority
137 * to limit potential SLOG device abuse by single active ZIL writer.
138 */
139 static uint64_t zil_slog_bulk = 64 * 1024 * 1024;
140
141 static kmem_cache_t *zil_lwb_cache;
142 static kmem_cache_t *zil_zcw_cache;
143
144 static void zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx);
145 static itx_t *zil_itx_clone(itx_t *oitx);
146 static uint64_t zil_max_waste_space(zilog_t *zilog);
147
148 static int
zil_bp_compare(const void * x1,const void * x2)149 zil_bp_compare(const void *x1, const void *x2)
150 {
151 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
152 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
153
154 int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
155 if (likely(cmp))
156 return (cmp);
157
158 return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
159 }
160
161 static void
zil_bp_tree_init(zilog_t * zilog)162 zil_bp_tree_init(zilog_t *zilog)
163 {
164 avl_create(&zilog->zl_bp_tree, zil_bp_compare,
165 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
166 }
167
168 static void
zil_bp_tree_fini(zilog_t * zilog)169 zil_bp_tree_fini(zilog_t *zilog)
170 {
171 avl_tree_t *t = &zilog->zl_bp_tree;
172 zil_bp_node_t *zn;
173 void *cookie = NULL;
174
175 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
176 kmem_free(zn, sizeof (zil_bp_node_t));
177
178 avl_destroy(t);
179 }
180
181 int
zil_bp_tree_add(zilog_t * zilog,const blkptr_t * bp)182 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
183 {
184 avl_tree_t *t = &zilog->zl_bp_tree;
185 const dva_t *dva;
186 zil_bp_node_t *zn;
187 avl_index_t where;
188
189 if (BP_IS_EMBEDDED(bp))
190 return (0);
191
192 dva = BP_IDENTITY(bp);
193
194 if (avl_find(t, dva, &where) != NULL)
195 return (SET_ERROR(EEXIST));
196
197 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
198 zn->zn_dva = *dva;
199 avl_insert(t, zn, where);
200
201 return (0);
202 }
203
204 static zil_header_t *
zil_header_in_syncing_context(zilog_t * zilog)205 zil_header_in_syncing_context(zilog_t *zilog)
206 {
207 return ((zil_header_t *)zilog->zl_header);
208 }
209
210 static void
zil_init_log_chain(zilog_t * zilog,blkptr_t * bp)211 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
212 {
213 zio_cksum_t *zc = &bp->blk_cksum;
214
215 (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
216 sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
217 (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
218 sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
219 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
220 zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
221 }
222
223 static int
zil_kstats_global_update(kstat_t * ksp,int rw)224 zil_kstats_global_update(kstat_t *ksp, int rw)
225 {
226 zil_kstat_values_t *zs = ksp->ks_data;
227 ASSERT3P(&zil_stats, ==, zs);
228
229 if (rw == KSTAT_WRITE) {
230 return (SET_ERROR(EACCES));
231 }
232
233 zil_kstat_values_update(zs, &zil_sums_global);
234
235 return (0);
236 }
237
238 /*
239 * Read a log block and make sure it's valid.
240 */
241 static int
zil_read_log_block(zilog_t * zilog,boolean_t decrypt,const blkptr_t * bp,blkptr_t * nbp,char ** begin,char ** end,arc_buf_t ** abuf)242 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
243 blkptr_t *nbp, char **begin, char **end, arc_buf_t **abuf)
244 {
245 zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
246 arc_flags_t aflags = ARC_FLAG_WAIT;
247 zbookmark_phys_t zb;
248 int error;
249
250 if (zilog->zl_header->zh_claim_txg == 0)
251 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
252
253 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
254 zio_flags |= ZIO_FLAG_SPECULATIVE;
255
256 if (!decrypt)
257 zio_flags |= ZIO_FLAG_RAW;
258
259 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
260 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
261
262 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
263 abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
264
265 if (error == 0) {
266 zio_cksum_t cksum = bp->blk_cksum;
267
268 /*
269 * Validate the checksummed log block.
270 *
271 * Sequence numbers should be... sequential. The checksum
272 * verifier for the next block should be bp's checksum plus 1.
273 *
274 * Also check the log chain linkage and size used.
275 */
276 cksum.zc_word[ZIL_ZC_SEQ]++;
277
278 uint64_t size = BP_GET_LSIZE(bp);
279 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
280 zil_chain_t *zilc = (*abuf)->b_data;
281 char *lr = (char *)(zilc + 1);
282
283 if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
284 sizeof (cksum)) ||
285 zilc->zc_nused < sizeof (*zilc) ||
286 zilc->zc_nused > size) {
287 error = SET_ERROR(ECKSUM);
288 } else {
289 *begin = lr;
290 *end = lr + zilc->zc_nused - sizeof (*zilc);
291 *nbp = zilc->zc_next_blk;
292 }
293 } else {
294 char *lr = (*abuf)->b_data;
295 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
296
297 if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
298 sizeof (cksum)) ||
299 (zilc->zc_nused > (size - sizeof (*zilc)))) {
300 error = SET_ERROR(ECKSUM);
301 } else {
302 *begin = lr;
303 *end = lr + zilc->zc_nused;
304 *nbp = zilc->zc_next_blk;
305 }
306 }
307 }
308
309 return (error);
310 }
311
312 /*
313 * Read a TX_WRITE log data block.
314 */
315 static int
zil_read_log_data(zilog_t * zilog,const lr_write_t * lr,void * wbuf)316 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
317 {
318 zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
319 const blkptr_t *bp = &lr->lr_blkptr;
320 arc_flags_t aflags = ARC_FLAG_WAIT;
321 arc_buf_t *abuf = NULL;
322 zbookmark_phys_t zb;
323 int error;
324
325 if (BP_IS_HOLE(bp)) {
326 if (wbuf != NULL)
327 memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length));
328 return (0);
329 }
330
331 if (zilog->zl_header->zh_claim_txg == 0)
332 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
333
334 /*
335 * If we are not using the resulting data, we are just checking that
336 * it hasn't been corrupted so we don't need to waste CPU time
337 * decompressing and decrypting it.
338 */
339 if (wbuf == NULL)
340 zio_flags |= ZIO_FLAG_RAW;
341
342 ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
343 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
344 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
345
346 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
347 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
348
349 if (error == 0) {
350 if (wbuf != NULL)
351 memcpy(wbuf, abuf->b_data, arc_buf_size(abuf));
352 arc_buf_destroy(abuf, &abuf);
353 }
354
355 return (error);
356 }
357
358 void
zil_sums_init(zil_sums_t * zs)359 zil_sums_init(zil_sums_t *zs)
360 {
361 wmsum_init(&zs->zil_commit_count, 0);
362 wmsum_init(&zs->zil_commit_writer_count, 0);
363 wmsum_init(&zs->zil_itx_count, 0);
364 wmsum_init(&zs->zil_itx_indirect_count, 0);
365 wmsum_init(&zs->zil_itx_indirect_bytes, 0);
366 wmsum_init(&zs->zil_itx_copied_count, 0);
367 wmsum_init(&zs->zil_itx_copied_bytes, 0);
368 wmsum_init(&zs->zil_itx_needcopy_count, 0);
369 wmsum_init(&zs->zil_itx_needcopy_bytes, 0);
370 wmsum_init(&zs->zil_itx_metaslab_normal_count, 0);
371 wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0);
372 wmsum_init(&zs->zil_itx_metaslab_normal_write, 0);
373 wmsum_init(&zs->zil_itx_metaslab_normal_alloc, 0);
374 wmsum_init(&zs->zil_itx_metaslab_slog_count, 0);
375 wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0);
376 wmsum_init(&zs->zil_itx_metaslab_slog_write, 0);
377 wmsum_init(&zs->zil_itx_metaslab_slog_alloc, 0);
378 }
379
380 void
zil_sums_fini(zil_sums_t * zs)381 zil_sums_fini(zil_sums_t *zs)
382 {
383 wmsum_fini(&zs->zil_commit_count);
384 wmsum_fini(&zs->zil_commit_writer_count);
385 wmsum_fini(&zs->zil_itx_count);
386 wmsum_fini(&zs->zil_itx_indirect_count);
387 wmsum_fini(&zs->zil_itx_indirect_bytes);
388 wmsum_fini(&zs->zil_itx_copied_count);
389 wmsum_fini(&zs->zil_itx_copied_bytes);
390 wmsum_fini(&zs->zil_itx_needcopy_count);
391 wmsum_fini(&zs->zil_itx_needcopy_bytes);
392 wmsum_fini(&zs->zil_itx_metaslab_normal_count);
393 wmsum_fini(&zs->zil_itx_metaslab_normal_bytes);
394 wmsum_fini(&zs->zil_itx_metaslab_normal_write);
395 wmsum_fini(&zs->zil_itx_metaslab_normal_alloc);
396 wmsum_fini(&zs->zil_itx_metaslab_slog_count);
397 wmsum_fini(&zs->zil_itx_metaslab_slog_bytes);
398 wmsum_fini(&zs->zil_itx_metaslab_slog_write);
399 wmsum_fini(&zs->zil_itx_metaslab_slog_alloc);
400 }
401
402 void
zil_kstat_values_update(zil_kstat_values_t * zs,zil_sums_t * zil_sums)403 zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums)
404 {
405 zs->zil_commit_count.value.ui64 =
406 wmsum_value(&zil_sums->zil_commit_count);
407 zs->zil_commit_writer_count.value.ui64 =
408 wmsum_value(&zil_sums->zil_commit_writer_count);
409 zs->zil_itx_count.value.ui64 =
410 wmsum_value(&zil_sums->zil_itx_count);
411 zs->zil_itx_indirect_count.value.ui64 =
412 wmsum_value(&zil_sums->zil_itx_indirect_count);
413 zs->zil_itx_indirect_bytes.value.ui64 =
414 wmsum_value(&zil_sums->zil_itx_indirect_bytes);
415 zs->zil_itx_copied_count.value.ui64 =
416 wmsum_value(&zil_sums->zil_itx_copied_count);
417 zs->zil_itx_copied_bytes.value.ui64 =
418 wmsum_value(&zil_sums->zil_itx_copied_bytes);
419 zs->zil_itx_needcopy_count.value.ui64 =
420 wmsum_value(&zil_sums->zil_itx_needcopy_count);
421 zs->zil_itx_needcopy_bytes.value.ui64 =
422 wmsum_value(&zil_sums->zil_itx_needcopy_bytes);
423 zs->zil_itx_metaslab_normal_count.value.ui64 =
424 wmsum_value(&zil_sums->zil_itx_metaslab_normal_count);
425 zs->zil_itx_metaslab_normal_bytes.value.ui64 =
426 wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes);
427 zs->zil_itx_metaslab_normal_write.value.ui64 =
428 wmsum_value(&zil_sums->zil_itx_metaslab_normal_write);
429 zs->zil_itx_metaslab_normal_alloc.value.ui64 =
430 wmsum_value(&zil_sums->zil_itx_metaslab_normal_alloc);
431 zs->zil_itx_metaslab_slog_count.value.ui64 =
432 wmsum_value(&zil_sums->zil_itx_metaslab_slog_count);
433 zs->zil_itx_metaslab_slog_bytes.value.ui64 =
434 wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes);
435 zs->zil_itx_metaslab_slog_write.value.ui64 =
436 wmsum_value(&zil_sums->zil_itx_metaslab_slog_write);
437 zs->zil_itx_metaslab_slog_alloc.value.ui64 =
438 wmsum_value(&zil_sums->zil_itx_metaslab_slog_alloc);
439 }
440
441 /*
442 * Parse the intent log, and call parse_func for each valid record within.
443 */
444 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,boolean_t decrypt)445 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
446 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
447 boolean_t decrypt)
448 {
449 const zil_header_t *zh = zilog->zl_header;
450 boolean_t claimed = !!zh->zh_claim_txg;
451 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
452 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
453 uint64_t max_blk_seq = 0;
454 uint64_t max_lr_seq = 0;
455 uint64_t blk_count = 0;
456 uint64_t lr_count = 0;
457 blkptr_t blk, next_blk = {{{{0}}}};
458 int error = 0;
459
460 /*
461 * Old logs didn't record the maximum zh_claim_lr_seq.
462 */
463 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
464 claim_lr_seq = UINT64_MAX;
465
466 /*
467 * Starting at the block pointed to by zh_log we read the log chain.
468 * For each block in the chain we strongly check that block to
469 * ensure its validity. We stop when an invalid block is found.
470 * For each block pointer in the chain we call parse_blk_func().
471 * For each record in each valid block we call parse_lr_func().
472 * If the log has been claimed, stop if we encounter a sequence
473 * number greater than the highest claimed sequence number.
474 */
475 zil_bp_tree_init(zilog);
476
477 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
478 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
479 int reclen;
480 char *lrp, *end;
481 arc_buf_t *abuf = NULL;
482
483 if (blk_seq > claim_blk_seq)
484 break;
485
486 error = parse_blk_func(zilog, &blk, arg, txg);
487 if (error != 0)
488 break;
489 ASSERT3U(max_blk_seq, <, blk_seq);
490 max_blk_seq = blk_seq;
491 blk_count++;
492
493 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
494 break;
495
496 error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
497 &lrp, &end, &abuf);
498 if (error != 0) {
499 if (abuf)
500 arc_buf_destroy(abuf, &abuf);
501 if (claimed) {
502 char name[ZFS_MAX_DATASET_NAME_LEN];
503
504 dmu_objset_name(zilog->zl_os, name);
505
506 cmn_err(CE_WARN, "ZFS read log block error %d, "
507 "dataset %s, seq 0x%llx\n", error, name,
508 (u_longlong_t)blk_seq);
509 }
510 break;
511 }
512
513 for (; lrp < end; lrp += reclen) {
514 lr_t *lr = (lr_t *)lrp;
515 reclen = lr->lrc_reclen;
516 ASSERT3U(reclen, >=, sizeof (lr_t));
517 ASSERT3U(reclen, <=, end - lrp);
518 if (lr->lrc_seq > claim_lr_seq) {
519 arc_buf_destroy(abuf, &abuf);
520 goto done;
521 }
522
523 error = parse_lr_func(zilog, lr, arg, txg);
524 if (error != 0) {
525 arc_buf_destroy(abuf, &abuf);
526 goto done;
527 }
528 ASSERT3U(max_lr_seq, <, lr->lrc_seq);
529 max_lr_seq = lr->lrc_seq;
530 lr_count++;
531 }
532 arc_buf_destroy(abuf, &abuf);
533 }
534 done:
535 zilog->zl_parse_error = error;
536 zilog->zl_parse_blk_seq = max_blk_seq;
537 zilog->zl_parse_lr_seq = max_lr_seq;
538 zilog->zl_parse_blk_count = blk_count;
539 zilog->zl_parse_lr_count = lr_count;
540
541 zil_bp_tree_fini(zilog);
542
543 return (error);
544 }
545
546 static int
zil_clear_log_block(zilog_t * zilog,const blkptr_t * bp,void * tx,uint64_t first_txg)547 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
548 uint64_t first_txg)
549 {
550 (void) tx;
551 ASSERT(!BP_IS_HOLE(bp));
552
553 /*
554 * As we call this function from the context of a rewind to a
555 * checkpoint, each ZIL block whose txg is later than the txg
556 * that we rewind to is invalid. Thus, we return -1 so
557 * zil_parse() doesn't attempt to read it.
558 */
559 if (BP_GET_LOGICAL_BIRTH(bp) >= first_txg)
560 return (-1);
561
562 if (zil_bp_tree_add(zilog, bp) != 0)
563 return (0);
564
565 zio_free(zilog->zl_spa, first_txg, bp);
566 return (0);
567 }
568
569 static int
zil_noop_log_record(zilog_t * zilog,const lr_t * lrc,void * tx,uint64_t first_txg)570 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
571 uint64_t first_txg)
572 {
573 (void) zilog, (void) lrc, (void) tx, (void) first_txg;
574 return (0);
575 }
576
577 static int
zil_claim_log_block(zilog_t * zilog,const blkptr_t * bp,void * tx,uint64_t first_txg)578 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
579 uint64_t first_txg)
580 {
581 /*
582 * Claim log block if not already committed and not already claimed.
583 * If tx == NULL, just verify that the block is claimable.
584 */
585 if (BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) < first_txg ||
586 zil_bp_tree_add(zilog, bp) != 0)
587 return (0);
588
589 return (zio_wait(zio_claim(NULL, zilog->zl_spa,
590 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
591 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
592 }
593
594 static int
zil_claim_write(zilog_t * zilog,const lr_t * lrc,void * tx,uint64_t first_txg)595 zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg)
596 {
597 lr_write_t *lr = (lr_write_t *)lrc;
598 int error;
599
600 ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
601
602 /*
603 * If the block is not readable, don't claim it. This can happen
604 * in normal operation when a log block is written to disk before
605 * some of the dmu_sync() blocks it points to. In this case, the
606 * transaction cannot have been committed to anyone (we would have
607 * waited for all writes to be stable first), so it is semantically
608 * correct to declare this the end of the log.
609 */
610 if (BP_GET_LOGICAL_BIRTH(&lr->lr_blkptr) >= first_txg) {
611 error = zil_read_log_data(zilog, lr, NULL);
612 if (error != 0)
613 return (error);
614 }
615
616 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
617 }
618
619 static int
zil_claim_clone_range(zilog_t * zilog,const lr_t * lrc,void * tx,uint64_t first_txg)620 zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx,
621 uint64_t first_txg)
622 {
623 const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
624 const blkptr_t *bp;
625 spa_t *spa = zilog->zl_spa;
626 uint_t ii;
627
628 ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
629 ASSERT3U(lrc->lrc_reclen, >=, offsetof(lr_clone_range_t,
630 lr_bps[lr->lr_nbps]));
631
632 if (tx == NULL) {
633 return (0);
634 }
635
636 /*
637 * XXX: Do we need to byteswap lr?
638 */
639
640 for (ii = 0; ii < lr->lr_nbps; ii++) {
641 bp = &lr->lr_bps[ii];
642
643 /*
644 * When data is embedded into the BP there is no need to create
645 * BRT entry as there is no data block. Just copy the BP as it
646 * contains the data.
647 */
648 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
649 continue;
650
651 /*
652 * We can not handle block pointers from the future, since they
653 * are not yet allocated. It should not normally happen, but
654 * just in case lets be safe and just stop here now instead of
655 * corrupting the pool.
656 */
657 if (BP_GET_BIRTH(bp) >= first_txg)
658 return (SET_ERROR(ENOENT));
659
660 /*
661 * Assert the block is really allocated before we reference it.
662 */
663 metaslab_check_free(spa, bp);
664 }
665
666 for (ii = 0; ii < lr->lr_nbps; ii++) {
667 bp = &lr->lr_bps[ii];
668 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp))
669 brt_pending_add(spa, bp, tx);
670 }
671
672 return (0);
673 }
674
675 static int
zil_claim_log_record(zilog_t * zilog,const lr_t * lrc,void * tx,uint64_t first_txg)676 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
677 uint64_t first_txg)
678 {
679
680 switch (lrc->lrc_txtype) {
681 case TX_WRITE:
682 return (zil_claim_write(zilog, lrc, tx, first_txg));
683 case TX_CLONE_RANGE:
684 return (zil_claim_clone_range(zilog, lrc, tx, first_txg));
685 default:
686 return (0);
687 }
688 }
689
690 static int
zil_free_log_block(zilog_t * zilog,const blkptr_t * bp,void * tx,uint64_t claim_txg)691 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
692 uint64_t claim_txg)
693 {
694 (void) claim_txg;
695
696 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
697
698 return (0);
699 }
700
701 static int
zil_free_write(zilog_t * zilog,const lr_t * lrc,void * tx,uint64_t claim_txg)702 zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg)
703 {
704 lr_write_t *lr = (lr_write_t *)lrc;
705 blkptr_t *bp = &lr->lr_blkptr;
706
707 ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
708
709 /*
710 * If we previously claimed it, we need to free it.
711 */
712 if (BP_GET_LOGICAL_BIRTH(bp) >= claim_txg &&
713 zil_bp_tree_add(zilog, bp) == 0 && !BP_IS_HOLE(bp)) {
714 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
715 }
716
717 return (0);
718 }
719
720 static int
zil_free_clone_range(zilog_t * zilog,const lr_t * lrc,void * tx)721 zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
722 {
723 const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
724 const blkptr_t *bp;
725 spa_t *spa;
726 uint_t ii;
727
728 ASSERT3U(lrc->lrc_reclen, >=, sizeof (*lr));
729 ASSERT3U(lrc->lrc_reclen, >=, offsetof(lr_clone_range_t,
730 lr_bps[lr->lr_nbps]));
731
732 if (tx == NULL) {
733 return (0);
734 }
735
736 spa = zilog->zl_spa;
737
738 for (ii = 0; ii < lr->lr_nbps; ii++) {
739 bp = &lr->lr_bps[ii];
740
741 if (!BP_IS_HOLE(bp)) {
742 zio_free(spa, dmu_tx_get_txg(tx), bp);
743 }
744 }
745
746 return (0);
747 }
748
749 static int
zil_free_log_record(zilog_t * zilog,const lr_t * lrc,void * tx,uint64_t claim_txg)750 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
751 uint64_t claim_txg)
752 {
753
754 if (claim_txg == 0) {
755 return (0);
756 }
757
758 switch (lrc->lrc_txtype) {
759 case TX_WRITE:
760 return (zil_free_write(zilog, lrc, tx, claim_txg));
761 case TX_CLONE_RANGE:
762 return (zil_free_clone_range(zilog, lrc, tx));
763 default:
764 return (0);
765 }
766 }
767
768 static int
zil_lwb_vdev_compare(const void * x1,const void * x2)769 zil_lwb_vdev_compare(const void *x1, const void *x2)
770 {
771 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
772 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
773
774 return (TREE_CMP(v1, v2));
775 }
776
777 /*
778 * Allocate a new lwb. We may already have a block pointer for it, in which
779 * case we get size and version from there. Or we may not yet, in which case
780 * we choose them here and later make the block allocation match.
781 */
782 static lwb_t *
zil_alloc_lwb(zilog_t * zilog,int sz,blkptr_t * bp,boolean_t slog,uint64_t txg,lwb_state_t state)783 zil_alloc_lwb(zilog_t *zilog, int sz, blkptr_t *bp, boolean_t slog,
784 uint64_t txg, lwb_state_t state)
785 {
786 lwb_t *lwb;
787
788 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
789 lwb->lwb_zilog = zilog;
790 if (bp) {
791 lwb->lwb_blk = *bp;
792 lwb->lwb_slim = (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2);
793 sz = BP_GET_LSIZE(bp);
794 } else {
795 BP_ZERO(&lwb->lwb_blk);
796 lwb->lwb_slim = (spa_version(zilog->zl_spa) >=
797 SPA_VERSION_SLIM_ZIL);
798 }
799 lwb->lwb_slog = slog;
800 lwb->lwb_error = 0;
801 if (lwb->lwb_slim) {
802 lwb->lwb_nmax = sz;
803 lwb->lwb_nused = lwb->lwb_nfilled = sizeof (zil_chain_t);
804 } else {
805 lwb->lwb_nmax = sz - sizeof (zil_chain_t);
806 lwb->lwb_nused = lwb->lwb_nfilled = 0;
807 }
808 lwb->lwb_sz = sz;
809 lwb->lwb_state = state;
810 lwb->lwb_buf = zio_buf_alloc(sz);
811 lwb->lwb_child_zio = NULL;
812 lwb->lwb_write_zio = NULL;
813 lwb->lwb_root_zio = NULL;
814 lwb->lwb_issued_timestamp = 0;
815 lwb->lwb_issued_txg = 0;
816 lwb->lwb_alloc_txg = txg;
817 lwb->lwb_max_txg = 0;
818
819 mutex_enter(&zilog->zl_lock);
820 list_insert_tail(&zilog->zl_lwb_list, lwb);
821 if (state != LWB_STATE_NEW)
822 zilog->zl_last_lwb_opened = lwb;
823 mutex_exit(&zilog->zl_lock);
824
825 return (lwb);
826 }
827
828 static void
zil_free_lwb(zilog_t * zilog,lwb_t * lwb)829 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
830 {
831 ASSERT(MUTEX_HELD(&zilog->zl_lock));
832 ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
833 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
834 ASSERT3P(lwb->lwb_child_zio, ==, NULL);
835 ASSERT3P(lwb->lwb_write_zio, ==, NULL);
836 ASSERT3P(lwb->lwb_root_zio, ==, NULL);
837 ASSERT3U(lwb->lwb_alloc_txg, <=, spa_syncing_txg(zilog->zl_spa));
838 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
839 VERIFY(list_is_empty(&lwb->lwb_itxs));
840 VERIFY(list_is_empty(&lwb->lwb_waiters));
841 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
842 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
843
844 /*
845 * Clear the zilog's field to indicate this lwb is no longer
846 * valid, and prevent use-after-free errors.
847 */
848 if (zilog->zl_last_lwb_opened == lwb)
849 zilog->zl_last_lwb_opened = NULL;
850
851 kmem_cache_free(zil_lwb_cache, lwb);
852 }
853
854 /*
855 * Called when we create in-memory log transactions so that we know
856 * to cleanup the itxs at the end of spa_sync().
857 */
858 static void
zilog_dirty(zilog_t * zilog,uint64_t txg)859 zilog_dirty(zilog_t *zilog, uint64_t txg)
860 {
861 dsl_pool_t *dp = zilog->zl_dmu_pool;
862 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
863
864 ASSERT(spa_writeable(zilog->zl_spa));
865
866 if (ds->ds_is_snapshot)
867 panic("dirtying snapshot!");
868
869 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
870 /* up the hold count until we can be written out */
871 dmu_buf_add_ref(ds->ds_dbuf, zilog);
872
873 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
874 }
875 }
876
877 /*
878 * Determine if the zil is dirty in the specified txg. Callers wanting to
879 * ensure that the dirty state does not change must hold the itxg_lock for
880 * the specified txg. Holding the lock will ensure that the zil cannot be
881 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
882 * state.
883 */
884 static boolean_t __maybe_unused
zilog_is_dirty_in_txg(zilog_t * zilog,uint64_t txg)885 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
886 {
887 dsl_pool_t *dp = zilog->zl_dmu_pool;
888
889 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
890 return (B_TRUE);
891 return (B_FALSE);
892 }
893
894 /*
895 * Determine if the zil is dirty. The zil is considered dirty if it has
896 * any pending itx records that have not been cleaned by zil_clean().
897 */
898 static boolean_t
zilog_is_dirty(zilog_t * zilog)899 zilog_is_dirty(zilog_t *zilog)
900 {
901 dsl_pool_t *dp = zilog->zl_dmu_pool;
902
903 for (int t = 0; t < TXG_SIZE; t++) {
904 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
905 return (B_TRUE);
906 }
907 return (B_FALSE);
908 }
909
910 /*
911 * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
912 * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
913 * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
914 * zil_commit.
915 */
916 static void
zil_commit_activate_saxattr_feature(zilog_t * zilog)917 zil_commit_activate_saxattr_feature(zilog_t *zilog)
918 {
919 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
920 uint64_t txg = 0;
921 dmu_tx_t *tx = NULL;
922
923 if (spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
924 dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL &&
925 !dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)) {
926 tx = dmu_tx_create(zilog->zl_os);
927 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
928 dsl_dataset_dirty(ds, tx);
929 txg = dmu_tx_get_txg(tx);
930
931 mutex_enter(&ds->ds_lock);
932 ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
933 (void *)B_TRUE;
934 mutex_exit(&ds->ds_lock);
935 dmu_tx_commit(tx);
936 txg_wait_synced(zilog->zl_dmu_pool, txg);
937 }
938 }
939
940 /*
941 * Create an on-disk intent log.
942 */
943 static lwb_t *
zil_create(zilog_t * zilog)944 zil_create(zilog_t *zilog)
945 {
946 const zil_header_t *zh = zilog->zl_header;
947 lwb_t *lwb = NULL;
948 uint64_t txg = 0;
949 dmu_tx_t *tx = NULL;
950 blkptr_t blk;
951 int error = 0;
952 boolean_t slog = FALSE;
953 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
954
955
956 /*
957 * Wait for any previous destroy to complete.
958 */
959 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
960
961 ASSERT(zh->zh_claim_txg == 0);
962 ASSERT(zh->zh_replay_seq == 0);
963
964 blk = zh->zh_log;
965
966 /*
967 * Allocate an initial log block if:
968 * - there isn't one already
969 * - the existing block is the wrong endianness
970 */
971 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
972 tx = dmu_tx_create(zilog->zl_os);
973 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
974 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
975 txg = dmu_tx_get_txg(tx);
976
977 if (!BP_IS_HOLE(&blk)) {
978 zio_free(zilog->zl_spa, txg, &blk);
979 BP_ZERO(&blk);
980 }
981
982 error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
983 ZIL_MIN_BLKSZ, &slog);
984 if (error == 0)
985 zil_init_log_chain(zilog, &blk);
986 }
987
988 /*
989 * Allocate a log write block (lwb) for the first log block.
990 */
991 if (error == 0)
992 lwb = zil_alloc_lwb(zilog, 0, &blk, slog, txg, LWB_STATE_NEW);
993
994 /*
995 * If we just allocated the first log block, commit our transaction
996 * and wait for zil_sync() to stuff the block pointer into zh_log.
997 * (zh is part of the MOS, so we cannot modify it in open context.)
998 */
999 if (tx != NULL) {
1000 /*
1001 * If "zilsaxattr" feature is enabled on zpool, then activate
1002 * it now when we're creating the ZIL chain. We can't wait with
1003 * this until we write the first xattr log record because we
1004 * need to wait for the feature activation to sync out.
1005 */
1006 if (spa_feature_is_enabled(zilog->zl_spa,
1007 SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) !=
1008 DMU_OST_ZVOL) {
1009 mutex_enter(&ds->ds_lock);
1010 ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
1011 (void *)B_TRUE;
1012 mutex_exit(&ds->ds_lock);
1013 }
1014
1015 dmu_tx_commit(tx);
1016 txg_wait_synced(zilog->zl_dmu_pool, txg);
1017 } else {
1018 /*
1019 * This branch covers the case where we enable the feature on a
1020 * zpool that has existing ZIL headers.
1021 */
1022 zil_commit_activate_saxattr_feature(zilog);
1023 }
1024 IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
1025 dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL,
1026 dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR));
1027
1028 ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
1029 IMPLY(error == 0, lwb != NULL);
1030
1031 return (lwb);
1032 }
1033
1034 /*
1035 * In one tx, free all log blocks and clear the log header. If keep_first
1036 * is set, then we're replaying a log with no content. We want to keep the
1037 * first block, however, so that the first synchronous transaction doesn't
1038 * require a txg_wait_synced() in zil_create(). We don't need to
1039 * txg_wait_synced() here either when keep_first is set, because both
1040 * zil_create() and zil_destroy() will wait for any in-progress destroys
1041 * to complete.
1042 * Return B_TRUE if there were any entries to replay.
1043 */
1044 boolean_t
zil_destroy(zilog_t * zilog,boolean_t keep_first)1045 zil_destroy(zilog_t *zilog, boolean_t keep_first)
1046 {
1047 const zil_header_t *zh = zilog->zl_header;
1048 lwb_t *lwb;
1049 dmu_tx_t *tx;
1050 uint64_t txg;
1051
1052 /*
1053 * Wait for any previous destroy to complete.
1054 */
1055 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1056
1057 zilog->zl_old_header = *zh; /* debugging aid */
1058
1059 if (BP_IS_HOLE(&zh->zh_log))
1060 return (B_FALSE);
1061
1062 tx = dmu_tx_create(zilog->zl_os);
1063 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1064 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1065 txg = dmu_tx_get_txg(tx);
1066
1067 mutex_enter(&zilog->zl_lock);
1068
1069 ASSERT3U(zilog->zl_destroy_txg, <, txg);
1070 zilog->zl_destroy_txg = txg;
1071 zilog->zl_keep_first = keep_first;
1072
1073 if (!list_is_empty(&zilog->zl_lwb_list)) {
1074 ASSERT(zh->zh_claim_txg == 0);
1075 VERIFY(!keep_first);
1076 while ((lwb = list_remove_head(&zilog->zl_lwb_list)) != NULL) {
1077 if (lwb->lwb_buf != NULL)
1078 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1079 if (!BP_IS_HOLE(&lwb->lwb_blk))
1080 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
1081 zil_free_lwb(zilog, lwb);
1082 }
1083 } else if (!keep_first) {
1084 zil_destroy_sync(zilog, tx);
1085 }
1086 mutex_exit(&zilog->zl_lock);
1087
1088 dmu_tx_commit(tx);
1089
1090 return (B_TRUE);
1091 }
1092
1093 void
zil_destroy_sync(zilog_t * zilog,dmu_tx_t * tx)1094 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
1095 {
1096 ASSERT(list_is_empty(&zilog->zl_lwb_list));
1097 (void) zil_parse(zilog, zil_free_log_block,
1098 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
1099 }
1100
1101 int
zil_claim(dsl_pool_t * dp,dsl_dataset_t * ds,void * txarg)1102 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
1103 {
1104 dmu_tx_t *tx = txarg;
1105 zilog_t *zilog;
1106 uint64_t first_txg;
1107 zil_header_t *zh;
1108 objset_t *os;
1109 int error;
1110
1111 error = dmu_objset_own_obj(dp, ds->ds_object,
1112 DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
1113 if (error != 0) {
1114 /*
1115 * EBUSY indicates that the objset is inconsistent, in which
1116 * case it can not have a ZIL.
1117 */
1118 if (error != EBUSY) {
1119 cmn_err(CE_WARN, "can't open objset for %llu, error %u",
1120 (unsigned long long)ds->ds_object, error);
1121 }
1122
1123 return (0);
1124 }
1125
1126 zilog = dmu_objset_zil(os);
1127 zh = zil_header_in_syncing_context(zilog);
1128 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
1129 first_txg = spa_min_claim_txg(zilog->zl_spa);
1130
1131 /*
1132 * If the spa_log_state is not set to be cleared, check whether
1133 * the current uberblock is a checkpoint one and if the current
1134 * header has been claimed before moving on.
1135 *
1136 * If the current uberblock is a checkpointed uberblock then
1137 * one of the following scenarios took place:
1138 *
1139 * 1] We are currently rewinding to the checkpoint of the pool.
1140 * 2] We crashed in the middle of a checkpoint rewind but we
1141 * did manage to write the checkpointed uberblock to the
1142 * vdev labels, so when we tried to import the pool again
1143 * the checkpointed uberblock was selected from the import
1144 * procedure.
1145 *
1146 * In both cases we want to zero out all the ZIL blocks, except
1147 * the ones that have been claimed at the time of the checkpoint
1148 * (their zh_claim_txg != 0). The reason is that these blocks
1149 * may be corrupted since we may have reused their locations on
1150 * disk after we took the checkpoint.
1151 *
1152 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1153 * when we first figure out whether the current uberblock is
1154 * checkpointed or not. Unfortunately, that would discard all
1155 * the logs, including the ones that are claimed, and we would
1156 * leak space.
1157 */
1158 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
1159 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1160 zh->zh_claim_txg == 0)) {
1161 if (!BP_IS_HOLE(&zh->zh_log)) {
1162 (void) zil_parse(zilog, zil_clear_log_block,
1163 zil_noop_log_record, tx, first_txg, B_FALSE);
1164 }
1165 BP_ZERO(&zh->zh_log);
1166 if (os->os_encrypted)
1167 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1168 dsl_dataset_dirty(dmu_objset_ds(os), tx);
1169 dmu_objset_disown(os, B_FALSE, FTAG);
1170 return (0);
1171 }
1172
1173 /*
1174 * If we are not rewinding and opening the pool normally, then
1175 * the min_claim_txg should be equal to the first txg of the pool.
1176 */
1177 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
1178
1179 /*
1180 * Claim all log blocks if we haven't already done so, and remember
1181 * the highest claimed sequence number. This ensures that if we can
1182 * read only part of the log now (e.g. due to a missing device),
1183 * but we can read the entire log later, we will not try to replay
1184 * or destroy beyond the last block we successfully claimed.
1185 */
1186 ASSERT3U(zh->zh_claim_txg, <=, first_txg);
1187 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
1188 (void) zil_parse(zilog, zil_claim_log_block,
1189 zil_claim_log_record, tx, first_txg, B_FALSE);
1190 zh->zh_claim_txg = first_txg;
1191 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
1192 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
1193 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
1194 zh->zh_flags |= ZIL_REPLAY_NEEDED;
1195 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
1196 if (os->os_encrypted)
1197 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1198 dsl_dataset_dirty(dmu_objset_ds(os), tx);
1199 }
1200
1201 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
1202 dmu_objset_disown(os, B_FALSE, FTAG);
1203 return (0);
1204 }
1205
1206 /*
1207 * Check the log by walking the log chain.
1208 * Checksum errors are ok as they indicate the end of the chain.
1209 * Any other error (no device or read failure) returns an error.
1210 */
1211 int
zil_check_log_chain(dsl_pool_t * dp,dsl_dataset_t * ds,void * tx)1212 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
1213 {
1214 (void) dp;
1215 zilog_t *zilog;
1216 objset_t *os;
1217 blkptr_t *bp;
1218 int error;
1219
1220 ASSERT(tx == NULL);
1221
1222 error = dmu_objset_from_ds(ds, &os);
1223 if (error != 0) {
1224 cmn_err(CE_WARN, "can't open objset %llu, error %d",
1225 (unsigned long long)ds->ds_object, error);
1226 return (0);
1227 }
1228
1229 zilog = dmu_objset_zil(os);
1230 bp = (blkptr_t *)&zilog->zl_header->zh_log;
1231
1232 if (!BP_IS_HOLE(bp)) {
1233 vdev_t *vd;
1234 boolean_t valid = B_TRUE;
1235
1236 /*
1237 * Check the first block and determine if it's on a log device
1238 * which may have been removed or faulted prior to loading this
1239 * pool. If so, there's no point in checking the rest of the
1240 * log as its content should have already been synced to the
1241 * pool.
1242 */
1243 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
1244 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
1245 if (vd->vdev_islog && vdev_is_dead(vd))
1246 valid = vdev_log_state_valid(vd);
1247 spa_config_exit(os->os_spa, SCL_STATE, FTAG);
1248
1249 if (!valid)
1250 return (0);
1251
1252 /*
1253 * Check whether the current uberblock is checkpointed (e.g.
1254 * we are rewinding) and whether the current header has been
1255 * claimed or not. If it hasn't then skip verifying it. We
1256 * do this because its ZIL blocks may be part of the pool's
1257 * state before the rewind, which is no longer valid.
1258 */
1259 zil_header_t *zh = zil_header_in_syncing_context(zilog);
1260 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1261 zh->zh_claim_txg == 0)
1262 return (0);
1263 }
1264
1265 /*
1266 * Because tx == NULL, zil_claim_log_block() will not actually claim
1267 * any blocks, but just determine whether it is possible to do so.
1268 * In addition to checking the log chain, zil_claim_log_block()
1269 * will invoke zio_claim() with a done func of spa_claim_notify(),
1270 * which will update spa_max_claim_txg. See spa_load() for details.
1271 */
1272 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
1273 zilog->zl_header->zh_claim_txg ? -1ULL :
1274 spa_min_claim_txg(os->os_spa), B_FALSE);
1275
1276 return ((error == ECKSUM || error == ENOENT) ? 0 : error);
1277 }
1278
1279 /*
1280 * When an itx is "skipped", this function is used to properly mark the
1281 * waiter as "done, and signal any thread(s) waiting on it. An itx can
1282 * be skipped (and not committed to an lwb) for a variety of reasons,
1283 * one of them being that the itx was committed via spa_sync(), prior to
1284 * it being committed to an lwb; this can happen if a thread calling
1285 * zil_commit() is racing with spa_sync().
1286 */
1287 static void
zil_commit_waiter_skip(zil_commit_waiter_t * zcw)1288 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
1289 {
1290 mutex_enter(&zcw->zcw_lock);
1291 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1292 zcw->zcw_done = B_TRUE;
1293 cv_broadcast(&zcw->zcw_cv);
1294 mutex_exit(&zcw->zcw_lock);
1295 }
1296
1297 /*
1298 * This function is used when the given waiter is to be linked into an
1299 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1300 * At this point, the waiter will no longer be referenced by the itx,
1301 * and instead, will be referenced by the lwb.
1302 */
1303 static void
zil_commit_waiter_link_lwb(zil_commit_waiter_t * zcw,lwb_t * lwb)1304 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1305 {
1306 /*
1307 * The lwb_waiters field of the lwb is protected by the zilog's
1308 * zl_issuer_lock while the lwb is open and zl_lock otherwise.
1309 * zl_issuer_lock also protects leaving the open state.
1310 * zcw_lwb setting is protected by zl_issuer_lock and state !=
1311 * flush_done, which transition is protected by zl_lock.
1312 */
1313 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_issuer_lock));
1314 IMPLY(lwb->lwb_state != LWB_STATE_OPENED,
1315 MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1316 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW);
1317 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1318
1319 ASSERT(!list_link_active(&zcw->zcw_node));
1320 list_insert_tail(&lwb->lwb_waiters, zcw);
1321 ASSERT3P(zcw->zcw_lwb, ==, NULL);
1322 zcw->zcw_lwb = lwb;
1323 }
1324
1325 /*
1326 * This function is used when zio_alloc_zil() fails to allocate a ZIL
1327 * block, and the given waiter must be linked to the "nolwb waiters"
1328 * list inside of zil_process_commit_list().
1329 */
1330 static void
zil_commit_waiter_link_nolwb(zil_commit_waiter_t * zcw,list_t * nolwb)1331 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1332 {
1333 ASSERT(!list_link_active(&zcw->zcw_node));
1334 list_insert_tail(nolwb, zcw);
1335 ASSERT3P(zcw->zcw_lwb, ==, NULL);
1336 }
1337
1338 void
zil_lwb_add_block(lwb_t * lwb,const blkptr_t * bp)1339 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1340 {
1341 avl_tree_t *t = &lwb->lwb_vdev_tree;
1342 avl_index_t where;
1343 zil_vdev_node_t *zv, zvsearch;
1344 int ndvas = BP_GET_NDVAS(bp);
1345 int i;
1346
1347 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1348 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1349
1350 if (zil_nocacheflush)
1351 return;
1352
1353 mutex_enter(&lwb->lwb_vdev_lock);
1354 for (i = 0; i < ndvas; i++) {
1355 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1356 if (avl_find(t, &zvsearch, &where) == NULL) {
1357 zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1358 zv->zv_vdev = zvsearch.zv_vdev;
1359 avl_insert(t, zv, where);
1360 }
1361 }
1362 mutex_exit(&lwb->lwb_vdev_lock);
1363 }
1364
1365 static void
zil_lwb_flush_defer(lwb_t * lwb,lwb_t * nlwb)1366 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1367 {
1368 avl_tree_t *src = &lwb->lwb_vdev_tree;
1369 avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1370 void *cookie = NULL;
1371 zil_vdev_node_t *zv;
1372
1373 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1374 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1375 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1376
1377 /*
1378 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1379 * not need the protection of lwb_vdev_lock (it will only be modified
1380 * while holding zilog->zl_lock) as its writes and those of its
1381 * children have all completed. The younger 'nlwb' may be waiting on
1382 * future writes to additional vdevs.
1383 */
1384 mutex_enter(&nlwb->lwb_vdev_lock);
1385 /*
1386 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1387 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1388 */
1389 while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1390 avl_index_t where;
1391
1392 if (avl_find(dst, zv, &where) == NULL) {
1393 avl_insert(dst, zv, where);
1394 } else {
1395 kmem_free(zv, sizeof (*zv));
1396 }
1397 }
1398 mutex_exit(&nlwb->lwb_vdev_lock);
1399 }
1400
1401 void
zil_lwb_add_txg(lwb_t * lwb,uint64_t txg)1402 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1403 {
1404 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1405 }
1406
1407 /*
1408 * This function is a called after all vdevs associated with a given lwb write
1409 * have completed their flush command; or as soon as the lwb write completes,
1410 * if "zil_nocacheflush" is set. Further, all "previous" lwb's will have
1411 * completed before this function is called; i.e. this function is called for
1412 * all previous lwbs before it's called for "this" lwb (enforced via zio the
1413 * dependencies configured in zil_lwb_set_zio_dependency()).
1414 *
1415 * The intention is for this function to be called as soon as the contents of
1416 * an lwb are considered "stable" on disk, and will survive any sudden loss of
1417 * power. At this point, any threads waiting for the lwb to reach this state
1418 * are signalled, and the "waiter" structures are marked "done".
1419 */
1420 static void
zil_lwb_flush_vdevs_done(zio_t * zio)1421 zil_lwb_flush_vdevs_done(zio_t *zio)
1422 {
1423 lwb_t *lwb = zio->io_private;
1424 zilog_t *zilog = lwb->lwb_zilog;
1425 zil_commit_waiter_t *zcw;
1426 itx_t *itx;
1427
1428 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1429
1430 hrtime_t t = gethrtime() - lwb->lwb_issued_timestamp;
1431
1432 mutex_enter(&zilog->zl_lock);
1433
1434 zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 7 + t) / 8;
1435
1436 lwb->lwb_root_zio = NULL;
1437
1438 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1439 lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1440
1441 if (zilog->zl_last_lwb_opened == lwb) {
1442 /*
1443 * Remember the highest committed log sequence number
1444 * for ztest. We only update this value when all the log
1445 * writes succeeded, because ztest wants to ASSERT that
1446 * it got the whole log chain.
1447 */
1448 zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1449 }
1450
1451 while ((itx = list_remove_head(&lwb->lwb_itxs)) != NULL)
1452 zil_itx_destroy(itx);
1453
1454 while ((zcw = list_remove_head(&lwb->lwb_waiters)) != NULL) {
1455 mutex_enter(&zcw->zcw_lock);
1456
1457 ASSERT3P(zcw->zcw_lwb, ==, lwb);
1458 zcw->zcw_lwb = NULL;
1459 /*
1460 * We expect any ZIO errors from child ZIOs to have been
1461 * propagated "up" to this specific LWB's root ZIO, in
1462 * order for this error handling to work correctly. This
1463 * includes ZIO errors from either this LWB's write or
1464 * flush, as well as any errors from other dependent LWBs
1465 * (e.g. a root LWB ZIO that might be a child of this LWB).
1466 *
1467 * With that said, it's important to note that LWB flush
1468 * errors are not propagated up to the LWB root ZIO.
1469 * This is incorrect behavior, and results in VDEV flush
1470 * errors not being handled correctly here. See the
1471 * comment above the call to "zio_flush" for details.
1472 */
1473
1474 zcw->zcw_zio_error = zio->io_error;
1475
1476 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1477 zcw->zcw_done = B_TRUE;
1478 cv_broadcast(&zcw->zcw_cv);
1479
1480 mutex_exit(&zcw->zcw_lock);
1481 }
1482
1483 uint64_t txg = lwb->lwb_issued_txg;
1484
1485 /* Once we drop the lock, lwb may be freed by zil_sync(). */
1486 mutex_exit(&zilog->zl_lock);
1487
1488 mutex_enter(&zilog->zl_lwb_io_lock);
1489 ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0);
1490 zilog->zl_lwb_inflight[txg & TXG_MASK]--;
1491 if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0)
1492 cv_broadcast(&zilog->zl_lwb_io_cv);
1493 mutex_exit(&zilog->zl_lwb_io_lock);
1494 }
1495
1496 /*
1497 * Wait for the completion of all issued write/flush of that txg provided.
1498 * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1499 */
1500 static void
zil_lwb_flush_wait_all(zilog_t * zilog,uint64_t txg)1501 zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg)
1502 {
1503 ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa));
1504
1505 mutex_enter(&zilog->zl_lwb_io_lock);
1506 while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0)
1507 cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock);
1508 mutex_exit(&zilog->zl_lwb_io_lock);
1509
1510 #ifdef ZFS_DEBUG
1511 mutex_enter(&zilog->zl_lock);
1512 mutex_enter(&zilog->zl_lwb_io_lock);
1513 lwb_t *lwb = list_head(&zilog->zl_lwb_list);
1514 while (lwb != NULL) {
1515 if (lwb->lwb_issued_txg <= txg) {
1516 ASSERT(lwb->lwb_state != LWB_STATE_ISSUED);
1517 ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE);
1518 IMPLY(lwb->lwb_issued_txg > 0,
1519 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
1520 }
1521 IMPLY(lwb->lwb_state == LWB_STATE_WRITE_DONE ||
1522 lwb->lwb_state == LWB_STATE_FLUSH_DONE,
1523 lwb->lwb_buf == NULL);
1524 lwb = list_next(&zilog->zl_lwb_list, lwb);
1525 }
1526 mutex_exit(&zilog->zl_lwb_io_lock);
1527 mutex_exit(&zilog->zl_lock);
1528 #endif
1529 }
1530
1531 /*
1532 * This is called when an lwb's write zio completes. The callback's purpose is
1533 * to issue the flush commands for the vdevs in the lwb's lwb_vdev_tree. The
1534 * tree will contain the vdevs involved in writing out this specific lwb's
1535 * data, and in the case that cache flushes have been deferred, vdevs involved
1536 * in writing the data for previous lwbs. The writes corresponding to all the
1537 * vdevs in the lwb_vdev_tree will have completed by the time this is called,
1538 * due to the zio dependencies configured in zil_lwb_set_zio_dependency(),
1539 * which takes deferred flushes into account. The lwb will be "done" once
1540 * zil_lwb_flush_vdevs_done() is called, which occurs in the zio completion
1541 * callback for the lwb's root zio.
1542 */
1543 static void
zil_lwb_write_done(zio_t * zio)1544 zil_lwb_write_done(zio_t *zio)
1545 {
1546 lwb_t *lwb = zio->io_private;
1547 spa_t *spa = zio->io_spa;
1548 zilog_t *zilog = lwb->lwb_zilog;
1549 avl_tree_t *t = &lwb->lwb_vdev_tree;
1550 void *cookie = NULL;
1551 zil_vdev_node_t *zv;
1552 lwb_t *nlwb;
1553
1554 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1555
1556 abd_free(zio->io_abd);
1557 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1558 lwb->lwb_buf = NULL;
1559
1560 mutex_enter(&zilog->zl_lock);
1561 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1562 lwb->lwb_state = LWB_STATE_WRITE_DONE;
1563 lwb->lwb_child_zio = NULL;
1564 lwb->lwb_write_zio = NULL;
1565
1566 /*
1567 * If nlwb is not yet issued, zil_lwb_set_zio_dependency() is not
1568 * called for it yet, and when it will be, it won't be able to make
1569 * its write ZIO a parent this ZIO. In such case we can not defer
1570 * our flushes or below may be a race between the done callbacks.
1571 */
1572 nlwb = list_next(&zilog->zl_lwb_list, lwb);
1573 if (nlwb && nlwb->lwb_state != LWB_STATE_ISSUED)
1574 nlwb = NULL;
1575 mutex_exit(&zilog->zl_lock);
1576
1577 if (avl_numnodes(t) == 0)
1578 return;
1579
1580 /*
1581 * If there was an IO error, we're not going to call zio_flush()
1582 * on these vdevs, so we simply empty the tree and free the
1583 * nodes. We avoid calling zio_flush() since there isn't any
1584 * good reason for doing so, after the lwb block failed to be
1585 * written out.
1586 *
1587 * Additionally, we don't perform any further error handling at
1588 * this point (e.g. setting "zcw_zio_error" appropriately), as
1589 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1590 * we expect any error seen here, to have been propagated to
1591 * that function).
1592 */
1593 if (zio->io_error != 0) {
1594 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1595 kmem_free(zv, sizeof (*zv));
1596 return;
1597 }
1598
1599 /*
1600 * If this lwb does not have any threads waiting for it to complete, we
1601 * want to defer issuing the flush command to the vdevs written to by
1602 * "this" lwb, and instead rely on the "next" lwb to handle the flush
1603 * command for those vdevs. Thus, we merge the vdev tree of "this" lwb
1604 * with the vdev tree of the "next" lwb in the list, and assume the
1605 * "next" lwb will handle flushing the vdevs (or deferring the flush(s)
1606 * again).
1607 *
1608 * This is a useful performance optimization, especially for workloads
1609 * with lots of async write activity and few sync write and/or fsync
1610 * activity, as it has the potential to coalesce multiple flush
1611 * commands to a vdev into one.
1612 */
1613 if (list_is_empty(&lwb->lwb_waiters) && nlwb != NULL) {
1614 zil_lwb_flush_defer(lwb, nlwb);
1615 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1616 return;
1617 }
1618
1619 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1620 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1621 if (vd != NULL) {
1622 /*
1623 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1624 * always used within "zio_flush". This means,
1625 * any errors when flushing the vdev(s), will
1626 * (unfortunately) not be handled correctly,
1627 * since these "zio_flush" errors will not be
1628 * propagated up to "zil_lwb_flush_vdevs_done".
1629 */
1630 zio_flush(lwb->lwb_root_zio, vd);
1631 }
1632 kmem_free(zv, sizeof (*zv));
1633 }
1634 }
1635
1636 /*
1637 * Build the zio dependency chain, which is used to preserve the ordering of
1638 * lwb completions that is required by the semantics of the ZIL. Each new lwb
1639 * zio becomes a parent of the previous lwb zio, such that the new lwb's zio
1640 * cannot complete until the previous lwb's zio completes.
1641 *
1642 * This is required by the semantics of zil_commit(): the commit waiters
1643 * attached to the lwbs will be woken in the lwb zio's completion callback,
1644 * so this zio dependency graph ensures the waiters are woken in the correct
1645 * order (the same order the lwbs were created).
1646 */
1647 static void
zil_lwb_set_zio_dependency(zilog_t * zilog,lwb_t * lwb)1648 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1649 {
1650 ASSERT(MUTEX_HELD(&zilog->zl_lock));
1651
1652 lwb_t *prev_lwb = list_prev(&zilog->zl_lwb_list, lwb);
1653 if (prev_lwb == NULL ||
1654 prev_lwb->lwb_state == LWB_STATE_FLUSH_DONE)
1655 return;
1656
1657 /*
1658 * If the previous lwb's write hasn't already completed, we also want
1659 * to order the completion of the lwb write zios (above, we only order
1660 * the completion of the lwb root zios). This is required because of
1661 * how we can defer the flush commands for each lwb.
1662 *
1663 * When the flush commands are deferred, the previous lwb will rely on
1664 * this lwb to flush the vdevs written to by that previous lwb. Thus,
1665 * we need to ensure this lwb doesn't issue the flush until after the
1666 * previous lwb's write completes. We ensure this ordering by setting
1667 * the zio parent/child relationship here.
1668 *
1669 * Without this relationship on the lwb's write zio, it's possible for
1670 * this lwb's write to complete prior to the previous lwb's write
1671 * completing; and thus, the vdevs for the previous lwb would be
1672 * flushed prior to that lwb's data being written to those vdevs (the
1673 * vdevs are flushed in the lwb write zio's completion handler,
1674 * zil_lwb_write_done()).
1675 */
1676 if (prev_lwb->lwb_state == LWB_STATE_ISSUED) {
1677 ASSERT3P(prev_lwb->lwb_write_zio, !=, NULL);
1678 zio_add_child(lwb->lwb_write_zio, prev_lwb->lwb_write_zio);
1679 } else {
1680 ASSERT3S(prev_lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1681 }
1682
1683 ASSERT3P(prev_lwb->lwb_root_zio, !=, NULL);
1684 zio_add_child(lwb->lwb_root_zio, prev_lwb->lwb_root_zio);
1685 }
1686
1687
1688 /*
1689 * This function's purpose is to "open" an lwb such that it is ready to
1690 * accept new itxs being committed to it. This function is idempotent; if
1691 * the passed in lwb has already been opened, it is essentially a no-op.
1692 */
1693 static void
zil_lwb_write_open(zilog_t * zilog,lwb_t * lwb)1694 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1695 {
1696 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1697
1698 if (lwb->lwb_state != LWB_STATE_NEW) {
1699 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1700 return;
1701 }
1702
1703 mutex_enter(&zilog->zl_lock);
1704 lwb->lwb_state = LWB_STATE_OPENED;
1705 zilog->zl_last_lwb_opened = lwb;
1706 mutex_exit(&zilog->zl_lock);
1707 }
1708
1709 /*
1710 * Maximum block size used by the ZIL. This is picked up when the ZIL is
1711 * initialized. Otherwise this should not be used directly; see
1712 * zl_max_block_size instead.
1713 */
1714 static uint_t zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1715
1716 /*
1717 * Plan splitting of the provided burst size between several blocks.
1718 */
1719 static uint_t
zil_lwb_plan(zilog_t * zilog,uint64_t size,uint_t * minsize)1720 zil_lwb_plan(zilog_t *zilog, uint64_t size, uint_t *minsize)
1721 {
1722 uint_t md = zilog->zl_max_block_size - sizeof (zil_chain_t);
1723
1724 if (size <= md) {
1725 /*
1726 * Small bursts are written as-is in one block.
1727 */
1728 *minsize = size;
1729 return (size);
1730 } else if (size > 8 * md) {
1731 /*
1732 * Big bursts use maximum blocks. The first block size
1733 * is hard to predict, but it does not really matter.
1734 */
1735 *minsize = 0;
1736 return (md);
1737 }
1738
1739 /*
1740 * Medium bursts try to divide evenly to better utilize several SLOG
1741 * VDEVs. The first block size we predict assuming the worst case of
1742 * maxing out others. Fall back to using maximum blocks if due to
1743 * large records or wasted space we can not predict anything better.
1744 */
1745 uint_t s = size;
1746 uint_t n = DIV_ROUND_UP(s, md - sizeof (lr_write_t));
1747 uint_t chunk = DIV_ROUND_UP(s, n);
1748 uint_t waste = zil_max_waste_space(zilog);
1749 waste = MAX(waste, zilog->zl_cur_max);
1750 if (chunk <= md - waste) {
1751 *minsize = MAX(s - (md - waste) * (n - 1), waste);
1752 return (chunk);
1753 } else {
1754 *minsize = 0;
1755 return (md);
1756 }
1757 }
1758
1759 /*
1760 * Try to predict next block size based on previous history. Make prediction
1761 * sufficient for 7 of 8 previous bursts. Don't try to save if the saving is
1762 * less then 50%, extra writes may cost more, but we don't want single spike
1763 * to badly affect our predictions.
1764 */
1765 static uint_t
zil_lwb_predict(zilog_t * zilog)1766 zil_lwb_predict(zilog_t *zilog)
1767 {
1768 uint_t m, o;
1769
1770 /* If we are in the middle of a burst, take it into account also. */
1771 if (zilog->zl_cur_size > 0) {
1772 o = zil_lwb_plan(zilog, zilog->zl_cur_size, &m);
1773 } else {
1774 o = UINT_MAX;
1775 m = 0;
1776 }
1777
1778 /* Find minimum optimal size. We don't need to go below that. */
1779 for (int i = 0; i < ZIL_BURSTS; i++)
1780 o = MIN(o, zilog->zl_prev_opt[i]);
1781
1782 /* Find two biggest minimal first block sizes above the optimal. */
1783 uint_t m1 = MAX(m, o), m2 = o;
1784 for (int i = 0; i < ZIL_BURSTS; i++) {
1785 m = zilog->zl_prev_min[i];
1786 if (m >= m1) {
1787 m2 = m1;
1788 m1 = m;
1789 } else if (m > m2) {
1790 m2 = m;
1791 }
1792 }
1793
1794 /*
1795 * If second minimum size gives 50% saving -- use it. It may cost us
1796 * one additional write later, but the space saving is just too big.
1797 */
1798 return ((m1 < m2 * 2) ? m1 : m2);
1799 }
1800
1801 /*
1802 * Close the log block for being issued and allocate the next one.
1803 * Has to be called under zl_issuer_lock to chain more lwbs.
1804 */
1805 static lwb_t *
zil_lwb_write_close(zilog_t * zilog,lwb_t * lwb,lwb_state_t state)1806 zil_lwb_write_close(zilog_t *zilog, lwb_t *lwb, lwb_state_t state)
1807 {
1808 uint64_t blksz, plan, plan2;
1809
1810 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1811 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1812 lwb->lwb_state = LWB_STATE_CLOSED;
1813
1814 /*
1815 * If there was an allocation failure then returned NULL will trigger
1816 * zil_commit_writer_stall() at the caller. This is inherently racy,
1817 * since allocation may not have happened yet.
1818 */
1819 if (lwb->lwb_error != 0)
1820 return (NULL);
1821
1822 /*
1823 * Log blocks are pre-allocated. Here we select the size of the next
1824 * block, based on what's left of this burst and the previous history.
1825 * While we try to only write used part of the block, we can't just
1826 * always allocate the maximum block size because we can exhaust all
1827 * available pool log space, so we try to be reasonable.
1828 */
1829 if (zilog->zl_cur_left > 0) {
1830 /*
1831 * We are in the middle of a burst and know how much is left.
1832 * But if workload is multi-threaded there may be more soon.
1833 * Try to predict what can it be and plan for the worst case.
1834 */
1835 uint_t m;
1836 plan = zil_lwb_plan(zilog, zilog->zl_cur_left, &m);
1837 if (zilog->zl_parallel) {
1838 plan2 = zil_lwb_plan(zilog, zilog->zl_cur_left +
1839 zil_lwb_predict(zilog), &m);
1840 if (plan < plan2)
1841 plan = plan2;
1842 }
1843 } else {
1844 /*
1845 * The previous burst is done and we can only predict what
1846 * will come next.
1847 */
1848 plan = zil_lwb_predict(zilog);
1849 }
1850 blksz = plan + sizeof (zil_chain_t);
1851 blksz = P2ROUNDUP_TYPED(blksz, ZIL_MIN_BLKSZ, uint64_t);
1852 blksz = MIN(blksz, zilog->zl_max_block_size);
1853 DTRACE_PROBE3(zil__block__size, zilog_t *, zilog, uint64_t, blksz,
1854 uint64_t, plan);
1855
1856 return (zil_alloc_lwb(zilog, blksz, NULL, 0, 0, state));
1857 }
1858
1859 /*
1860 * Finalize previously closed block and issue the write zio.
1861 */
1862 static void
zil_lwb_write_issue(zilog_t * zilog,lwb_t * lwb)1863 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1864 {
1865 spa_t *spa = zilog->zl_spa;
1866 zil_chain_t *zilc;
1867 boolean_t slog;
1868 zbookmark_phys_t zb;
1869 zio_priority_t prio;
1870 int error;
1871
1872 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED);
1873
1874 /* Actually fill the lwb with the data. */
1875 for (itx_t *itx = list_head(&lwb->lwb_itxs); itx;
1876 itx = list_next(&lwb->lwb_itxs, itx))
1877 zil_lwb_commit(zilog, lwb, itx);
1878 lwb->lwb_nused = lwb->lwb_nfilled;
1879 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax);
1880
1881 lwb->lwb_root_zio = zio_root(spa, zil_lwb_flush_vdevs_done, lwb,
1882 ZIO_FLAG_CANFAIL);
1883
1884 /*
1885 * The lwb is now ready to be issued, but it can be only if it already
1886 * got its block pointer allocated or the allocation has failed.
1887 * Otherwise leave it as-is, relying on some other thread to issue it
1888 * after allocating its block pointer via calling zil_lwb_write_issue()
1889 * for the previous lwb(s) in the chain.
1890 */
1891 mutex_enter(&zilog->zl_lock);
1892 lwb->lwb_state = LWB_STATE_READY;
1893 if (BP_IS_HOLE(&lwb->lwb_blk) && lwb->lwb_error == 0) {
1894 mutex_exit(&zilog->zl_lock);
1895 return;
1896 }
1897 mutex_exit(&zilog->zl_lock);
1898
1899 next_lwb:
1900 if (lwb->lwb_slim)
1901 zilc = (zil_chain_t *)lwb->lwb_buf;
1902 else
1903 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_nmax);
1904 int wsz = lwb->lwb_sz;
1905 if (lwb->lwb_error == 0) {
1906 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, lwb->lwb_sz);
1907 if (!lwb->lwb_slog || zilog->zl_cur_size <= zil_slog_bulk)
1908 prio = ZIO_PRIORITY_SYNC_WRITE;
1909 else
1910 prio = ZIO_PRIORITY_ASYNC_WRITE;
1911 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1912 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1913 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1914 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, spa, 0,
1915 &lwb->lwb_blk, lwb_abd, lwb->lwb_sz, zil_lwb_write_done,
1916 lwb, prio, ZIO_FLAG_CANFAIL, &zb);
1917 zil_lwb_add_block(lwb, &lwb->lwb_blk);
1918
1919 if (lwb->lwb_slim) {
1920 /* For Slim ZIL only write what is used. */
1921 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ,
1922 int);
1923 ASSERT3S(wsz, <=, lwb->lwb_sz);
1924 zio_shrink(lwb->lwb_write_zio, wsz);
1925 wsz = lwb->lwb_write_zio->io_size;
1926 }
1927 memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused);
1928 zilc->zc_pad = 0;
1929 zilc->zc_nused = lwb->lwb_nused;
1930 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1931 } else {
1932 /*
1933 * We can't write the lwb if there was an allocation failure,
1934 * so create a null zio instead just to maintain dependencies.
1935 */
1936 lwb->lwb_write_zio = zio_null(lwb->lwb_root_zio, spa, NULL,
1937 zil_lwb_write_done, lwb, ZIO_FLAG_CANFAIL);
1938 lwb->lwb_write_zio->io_error = lwb->lwb_error;
1939 }
1940 if (lwb->lwb_child_zio)
1941 zio_add_child(lwb->lwb_write_zio, lwb->lwb_child_zio);
1942
1943 /*
1944 * Open transaction to allocate the next block pointer.
1945 */
1946 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
1947 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1948 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1949 uint64_t txg = dmu_tx_get_txg(tx);
1950
1951 /*
1952 * Allocate next the block pointer unless we are already in error.
1953 */
1954 lwb_t *nlwb = list_next(&zilog->zl_lwb_list, lwb);
1955 blkptr_t *bp = &zilc->zc_next_blk;
1956 BP_ZERO(bp);
1957 error = lwb->lwb_error;
1958 if (error == 0) {
1959 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, nlwb->lwb_sz,
1960 &slog);
1961 }
1962 if (error == 0) {
1963 ASSERT3U(BP_GET_LOGICAL_BIRTH(bp), ==, txg);
1964 BP_SET_CHECKSUM(bp, nlwb->lwb_slim ? ZIO_CHECKSUM_ZILOG2 :
1965 ZIO_CHECKSUM_ZILOG);
1966 bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1967 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1968 }
1969
1970 /*
1971 * Reduce TXG open time by incrementing inflight counter and committing
1972 * the transaciton. zil_sync() will wait for it to return to zero.
1973 */
1974 mutex_enter(&zilog->zl_lwb_io_lock);
1975 lwb->lwb_issued_txg = txg;
1976 zilog->zl_lwb_inflight[txg & TXG_MASK]++;
1977 zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg);
1978 mutex_exit(&zilog->zl_lwb_io_lock);
1979 dmu_tx_commit(tx);
1980
1981 spa_config_enter(spa, SCL_STATE, lwb, RW_READER);
1982
1983 /*
1984 * We've completed all potentially blocking operations. Update the
1985 * nlwb and allow it proceed without possible lock order reversals.
1986 */
1987 mutex_enter(&zilog->zl_lock);
1988 zil_lwb_set_zio_dependency(zilog, lwb);
1989 lwb->lwb_state = LWB_STATE_ISSUED;
1990
1991 if (nlwb) {
1992 nlwb->lwb_blk = *bp;
1993 nlwb->lwb_error = error;
1994 nlwb->lwb_slog = slog;
1995 nlwb->lwb_alloc_txg = txg;
1996 if (nlwb->lwb_state != LWB_STATE_READY)
1997 nlwb = NULL;
1998 }
1999 mutex_exit(&zilog->zl_lock);
2000
2001 if (lwb->lwb_slog) {
2002 ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count);
2003 ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes,
2004 lwb->lwb_nused);
2005 ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_write,
2006 wsz);
2007 ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_alloc,
2008 BP_GET_LSIZE(&lwb->lwb_blk));
2009 } else {
2010 ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count);
2011 ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes,
2012 lwb->lwb_nused);
2013 ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_write,
2014 wsz);
2015 ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_alloc,
2016 BP_GET_LSIZE(&lwb->lwb_blk));
2017 }
2018 lwb->lwb_issued_timestamp = gethrtime();
2019 if (lwb->lwb_child_zio)
2020 zio_nowait(lwb->lwb_child_zio);
2021 zio_nowait(lwb->lwb_write_zio);
2022 zio_nowait(lwb->lwb_root_zio);
2023
2024 /*
2025 * If nlwb was ready when we gave it the block pointer,
2026 * it is on us to issue it and possibly following ones.
2027 */
2028 lwb = nlwb;
2029 if (lwb)
2030 goto next_lwb;
2031 }
2032
2033 /*
2034 * Maximum amount of data that can be put into single log block.
2035 */
2036 uint64_t
zil_max_log_data(zilog_t * zilog,size_t hdrsize)2037 zil_max_log_data(zilog_t *zilog, size_t hdrsize)
2038 {
2039 return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize);
2040 }
2041
2042 /*
2043 * Maximum amount of log space we agree to waste to reduce number of
2044 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~6%).
2045 */
2046 static inline uint64_t
zil_max_waste_space(zilog_t * zilog)2047 zil_max_waste_space(zilog_t *zilog)
2048 {
2049 return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 16);
2050 }
2051
2052 /*
2053 * Maximum amount of write data for WR_COPIED. For correctness, consumers
2054 * must fall back to WR_NEED_COPY if we can't fit the entire record into one
2055 * maximum sized log block, because each WR_COPIED record must fit in a
2056 * single log block. Below that it is a tradeoff of additional memory copy
2057 * and possibly worse log space efficiency vs additional range lock/unlock.
2058 */
2059 static uint_t zil_maxcopied = 7680;
2060
2061 uint64_t
zil_max_copied_data(zilog_t * zilog)2062 zil_max_copied_data(zilog_t *zilog)
2063 {
2064 uint64_t max_data = zil_max_log_data(zilog, sizeof (lr_write_t));
2065 return (MIN(max_data, zil_maxcopied));
2066 }
2067
2068 static uint64_t
zil_itx_record_size(itx_t * itx)2069 zil_itx_record_size(itx_t *itx)
2070 {
2071 lr_t *lr = &itx->itx_lr;
2072
2073 if (lr->lrc_txtype == TX_COMMIT)
2074 return (0);
2075 ASSERT3U(lr->lrc_reclen, >=, sizeof (lr_t));
2076 return (lr->lrc_reclen);
2077 }
2078
2079 static uint64_t
zil_itx_data_size(itx_t * itx)2080 zil_itx_data_size(itx_t *itx)
2081 {
2082 lr_t *lr = &itx->itx_lr;
2083 lr_write_t *lrw = (lr_write_t *)lr;
2084
2085 if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
2086 ASSERT3U(lr->lrc_reclen, ==, sizeof (lr_write_t));
2087 return (P2ROUNDUP_TYPED(lrw->lr_length, sizeof (uint64_t),
2088 uint64_t));
2089 }
2090 return (0);
2091 }
2092
2093 static uint64_t
zil_itx_full_size(itx_t * itx)2094 zil_itx_full_size(itx_t *itx)
2095 {
2096 lr_t *lr = &itx->itx_lr;
2097
2098 if (lr->lrc_txtype == TX_COMMIT)
2099 return (0);
2100 ASSERT3U(lr->lrc_reclen, >=, sizeof (lr_t));
2101 return (lr->lrc_reclen + zil_itx_data_size(itx));
2102 }
2103
2104 /*
2105 * Estimate space needed in the lwb for the itx. Allocate more lwbs or
2106 * split the itx as needed, but don't touch the actual transaction data.
2107 * Has to be called under zl_issuer_lock to call zil_lwb_write_close()
2108 * to chain more lwbs.
2109 */
2110 static lwb_t *
zil_lwb_assign(zilog_t * zilog,lwb_t * lwb,itx_t * itx,list_t * ilwbs)2111 zil_lwb_assign(zilog_t *zilog, lwb_t *lwb, itx_t *itx, list_t *ilwbs)
2112 {
2113 itx_t *citx;
2114 lr_t *lr, *clr;
2115 lr_write_t *lrw;
2116 uint64_t dlen, dnow, lwb_sp, reclen, max_log_data;
2117
2118 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2119 ASSERT3P(lwb, !=, NULL);
2120 ASSERT3P(lwb->lwb_buf, !=, NULL);
2121
2122 zil_lwb_write_open(zilog, lwb);
2123
2124 lr = &itx->itx_lr;
2125 lrw = (lr_write_t *)lr;
2126
2127 /*
2128 * A commit itx doesn't represent any on-disk state; instead
2129 * it's simply used as a place holder on the commit list, and
2130 * provides a mechanism for attaching a "commit waiter" onto the
2131 * correct lwb (such that the waiter can be signalled upon
2132 * completion of that lwb). Thus, we don't process this itx's
2133 * log record if it's a commit itx (these itx's don't have log
2134 * records), and instead link the itx's waiter onto the lwb's
2135 * list of waiters.
2136 *
2137 * For more details, see the comment above zil_commit().
2138 */
2139 if (lr->lrc_txtype == TX_COMMIT) {
2140 zil_commit_waiter_link_lwb(itx->itx_private, lwb);
2141 list_insert_tail(&lwb->lwb_itxs, itx);
2142 return (lwb);
2143 }
2144
2145 reclen = lr->lrc_reclen;
2146 ASSERT3U(reclen, >=, sizeof (lr_t));
2147 ASSERT3U(reclen, <=, zil_max_log_data(zilog, 0));
2148 dlen = zil_itx_data_size(itx);
2149
2150 cont:
2151 /*
2152 * If this record won't fit in the current log block, start a new one.
2153 * For WR_NEED_COPY optimize layout for minimal number of chunks.
2154 */
2155 lwb_sp = lwb->lwb_nmax - lwb->lwb_nused;
2156 max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t));
2157 if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
2158 lwb_sp < zil_max_waste_space(zilog) &&
2159 (dlen % max_log_data == 0 ||
2160 lwb_sp < reclen + dlen % max_log_data))) {
2161 list_insert_tail(ilwbs, lwb);
2162 lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_OPENED);
2163 if (lwb == NULL)
2164 return (NULL);
2165 lwb_sp = lwb->lwb_nmax - lwb->lwb_nused;
2166 }
2167
2168 /*
2169 * There must be enough space in the log block to hold reclen.
2170 * For WR_COPIED, we need to fit the whole record in one block,
2171 * and reclen is the write record header size + the data size.
2172 * For WR_NEED_COPY, we can create multiple records, splitting
2173 * the data into multiple blocks, so we only need to fit one
2174 * word of data per block; in this case reclen is just the header
2175 * size (no data).
2176 */
2177 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
2178
2179 dnow = MIN(dlen, lwb_sp - reclen);
2180 if (dlen > dnow) {
2181 ASSERT3U(lr->lrc_txtype, ==, TX_WRITE);
2182 ASSERT3U(itx->itx_wr_state, ==, WR_NEED_COPY);
2183 citx = zil_itx_clone(itx);
2184 clr = &citx->itx_lr;
2185 lr_write_t *clrw = (lr_write_t *)clr;
2186 clrw->lr_length = dnow;
2187 lrw->lr_offset += dnow;
2188 lrw->lr_length -= dnow;
2189 zilog->zl_cur_left -= dnow;
2190 } else {
2191 citx = itx;
2192 clr = lr;
2193 }
2194
2195 /*
2196 * We're actually making an entry, so update lrc_seq to be the
2197 * log record sequence number. Note that this is generally not
2198 * equal to the itx sequence number because not all transactions
2199 * are synchronous, and sometimes spa_sync() gets there first.
2200 */
2201 clr->lrc_seq = ++zilog->zl_lr_seq;
2202
2203 lwb->lwb_nused += reclen + dnow;
2204 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_nmax);
2205 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
2206
2207 zil_lwb_add_txg(lwb, lr->lrc_txg);
2208 list_insert_tail(&lwb->lwb_itxs, citx);
2209
2210 dlen -= dnow;
2211 if (dlen > 0)
2212 goto cont;
2213
2214 if (lr->lrc_txtype == TX_WRITE &&
2215 lr->lrc_txg > spa_freeze_txg(zilog->zl_spa))
2216 txg_wait_synced(zilog->zl_dmu_pool, lr->lrc_txg);
2217
2218 return (lwb);
2219 }
2220
2221 /*
2222 * Fill the actual transaction data into the lwb, following zil_lwb_assign().
2223 * Does not require locking.
2224 */
2225 static void
zil_lwb_commit(zilog_t * zilog,lwb_t * lwb,itx_t * itx)2226 zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx)
2227 {
2228 lr_t *lr, *lrb;
2229 lr_write_t *lrw, *lrwb;
2230 char *lr_buf;
2231 uint64_t dlen, reclen;
2232
2233 lr = &itx->itx_lr;
2234 lrw = (lr_write_t *)lr;
2235
2236 if (lr->lrc_txtype == TX_COMMIT)
2237 return;
2238
2239 reclen = lr->lrc_reclen;
2240 dlen = zil_itx_data_size(itx);
2241 ASSERT3U(reclen + dlen, <=, lwb->lwb_nused - lwb->lwb_nfilled);
2242
2243 lr_buf = lwb->lwb_buf + lwb->lwb_nfilled;
2244 memcpy(lr_buf, lr, reclen);
2245 lrb = (lr_t *)lr_buf; /* Like lr, but inside lwb. */
2246 lrwb = (lr_write_t *)lrb; /* Like lrw, but inside lwb. */
2247
2248 ZIL_STAT_BUMP(zilog, zil_itx_count);
2249
2250 /*
2251 * If it's a write, fetch the data or get its blkptr as appropriate.
2252 */
2253 if (lr->lrc_txtype == TX_WRITE) {
2254 if (itx->itx_wr_state == WR_COPIED) {
2255 ZIL_STAT_BUMP(zilog, zil_itx_copied_count);
2256 ZIL_STAT_INCR(zilog, zil_itx_copied_bytes,
2257 lrw->lr_length);
2258 } else {
2259 char *dbuf;
2260 int error;
2261
2262 if (itx->itx_wr_state == WR_NEED_COPY) {
2263 dbuf = lr_buf + reclen;
2264 lrb->lrc_reclen += dlen;
2265 ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count);
2266 ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes,
2267 dlen);
2268 } else {
2269 ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
2270 dbuf = NULL;
2271 ZIL_STAT_BUMP(zilog, zil_itx_indirect_count);
2272 ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes,
2273 lrw->lr_length);
2274 if (lwb->lwb_child_zio == NULL) {
2275 lwb->lwb_child_zio = zio_null(NULL,
2276 zilog->zl_spa, NULL, NULL, NULL,
2277 ZIO_FLAG_CANFAIL);
2278 }
2279 }
2280
2281 /*
2282 * The "lwb_child_zio" we pass in will become a child of
2283 * "lwb_write_zio", when one is created, so one will be
2284 * a parent of any zio's created by the "zl_get_data".
2285 * This way "lwb_write_zio" will first wait for children
2286 * block pointers before own writing, and then for their
2287 * writing completion before the vdev cache flushing.
2288 */
2289 error = zilog->zl_get_data(itx->itx_private,
2290 itx->itx_gen, lrwb, dbuf, lwb,
2291 lwb->lwb_child_zio);
2292 if (dbuf != NULL && error == 0) {
2293 /* Zero any padding bytes in the last block. */
2294 memset((char *)dbuf + lrwb->lr_length, 0,
2295 dlen - lrwb->lr_length);
2296 }
2297
2298 /*
2299 * Typically, the only return values we should see from
2300 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
2301 * EALREADY. However, it is also possible to see other
2302 * error values such as ENOSPC or EINVAL from
2303 * dmu_read() -> dnode_hold() -> dnode_hold_impl() or
2304 * ENXIO as well as a multitude of others from the
2305 * block layer through dmu_buf_hold() -> dbuf_read()
2306 * -> zio_wait(), as well as through dmu_read() ->
2307 * dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
2308 * zio_wait(). When these errors happen, we can assume
2309 * that neither an immediate write nor an indirect
2310 * write occurred, so we need to fall back to
2311 * txg_wait_synced(). This is unusual, so we print to
2312 * dmesg whenever one of these errors occurs.
2313 */
2314 switch (error) {
2315 case 0:
2316 break;
2317 default:
2318 cmn_err(CE_WARN, "zil_lwb_commit() received "
2319 "unexpected error %d from ->zl_get_data()"
2320 ". Falling back to txg_wait_synced().",
2321 error);
2322 zfs_fallthrough;
2323 case EIO:
2324 txg_wait_synced(zilog->zl_dmu_pool,
2325 lr->lrc_txg);
2326 zfs_fallthrough;
2327 case ENOENT:
2328 zfs_fallthrough;
2329 case EEXIST:
2330 zfs_fallthrough;
2331 case EALREADY:
2332 return;
2333 }
2334 }
2335 }
2336
2337 lwb->lwb_nfilled += reclen + dlen;
2338 ASSERT3S(lwb->lwb_nfilled, <=, lwb->lwb_nused);
2339 ASSERT0(P2PHASE(lwb->lwb_nfilled, sizeof (uint64_t)));
2340 }
2341
2342 itx_t *
zil_itx_create(uint64_t txtype,size_t olrsize)2343 zil_itx_create(uint64_t txtype, size_t olrsize)
2344 {
2345 size_t itxsize, lrsize;
2346 itx_t *itx;
2347
2348 ASSERT3U(olrsize, >=, sizeof (lr_t));
2349 lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
2350 ASSERT3U(lrsize, >=, olrsize);
2351 itxsize = offsetof(itx_t, itx_lr) + lrsize;
2352
2353 itx = zio_data_buf_alloc(itxsize);
2354 itx->itx_lr.lrc_txtype = txtype;
2355 itx->itx_lr.lrc_reclen = lrsize;
2356 itx->itx_lr.lrc_seq = 0; /* defensive */
2357 memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize);
2358 itx->itx_sync = B_TRUE; /* default is synchronous */
2359 itx->itx_callback = NULL;
2360 itx->itx_callback_data = NULL;
2361 itx->itx_size = itxsize;
2362
2363 return (itx);
2364 }
2365
2366 static itx_t *
zil_itx_clone(itx_t * oitx)2367 zil_itx_clone(itx_t *oitx)
2368 {
2369 ASSERT3U(oitx->itx_size, >=, sizeof (itx_t));
2370 ASSERT3U(oitx->itx_size, ==,
2371 offsetof(itx_t, itx_lr) + oitx->itx_lr.lrc_reclen);
2372
2373 itx_t *itx = zio_data_buf_alloc(oitx->itx_size);
2374 memcpy(itx, oitx, oitx->itx_size);
2375 itx->itx_callback = NULL;
2376 itx->itx_callback_data = NULL;
2377 return (itx);
2378 }
2379
2380 void
zil_itx_destroy(itx_t * itx)2381 zil_itx_destroy(itx_t *itx)
2382 {
2383 ASSERT3U(itx->itx_size, >=, sizeof (itx_t));
2384 ASSERT3U(itx->itx_lr.lrc_reclen, ==,
2385 itx->itx_size - offsetof(itx_t, itx_lr));
2386 IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
2387 IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2388
2389 if (itx->itx_callback != NULL)
2390 itx->itx_callback(itx->itx_callback_data);
2391
2392 zio_data_buf_free(itx, itx->itx_size);
2393 }
2394
2395 /*
2396 * Free up the sync and async itxs. The itxs_t has already been detached
2397 * so no locks are needed.
2398 */
2399 static void
zil_itxg_clean(void * arg)2400 zil_itxg_clean(void *arg)
2401 {
2402 itx_t *itx;
2403 list_t *list;
2404 avl_tree_t *t;
2405 void *cookie;
2406 itxs_t *itxs = arg;
2407 itx_async_node_t *ian;
2408
2409 list = &itxs->i_sync_list;
2410 while ((itx = list_remove_head(list)) != NULL) {
2411 /*
2412 * In the general case, commit itxs will not be found
2413 * here, as they'll be committed to an lwb via
2414 * zil_lwb_assign(), and free'd in that function. Having
2415 * said that, it is still possible for commit itxs to be
2416 * found here, due to the following race:
2417 *
2418 * - a thread calls zil_commit() which assigns the
2419 * commit itx to a per-txg i_sync_list
2420 * - zil_itxg_clean() is called (e.g. via spa_sync())
2421 * while the waiter is still on the i_sync_list
2422 *
2423 * There's nothing to prevent syncing the txg while the
2424 * waiter is on the i_sync_list. This normally doesn't
2425 * happen because spa_sync() is slower than zil_commit(),
2426 * but if zil_commit() calls txg_wait_synced() (e.g.
2427 * because zil_create() or zil_commit_writer_stall() is
2428 * called) we will hit this case.
2429 */
2430 if (itx->itx_lr.lrc_txtype == TX_COMMIT)
2431 zil_commit_waiter_skip(itx->itx_private);
2432
2433 zil_itx_destroy(itx);
2434 }
2435
2436 cookie = NULL;
2437 t = &itxs->i_async_tree;
2438 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2439 list = &ian->ia_list;
2440 while ((itx = list_remove_head(list)) != NULL) {
2441 /* commit itxs should never be on the async lists. */
2442 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2443 zil_itx_destroy(itx);
2444 }
2445 list_destroy(list);
2446 kmem_free(ian, sizeof (itx_async_node_t));
2447 }
2448 avl_destroy(t);
2449
2450 kmem_free(itxs, sizeof (itxs_t));
2451 }
2452
2453 static int
zil_aitx_compare(const void * x1,const void * x2)2454 zil_aitx_compare(const void *x1, const void *x2)
2455 {
2456 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
2457 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
2458
2459 return (TREE_CMP(o1, o2));
2460 }
2461
2462 /*
2463 * Remove all async itx with the given oid.
2464 */
2465 void
zil_remove_async(zilog_t * zilog,uint64_t oid)2466 zil_remove_async(zilog_t *zilog, uint64_t oid)
2467 {
2468 uint64_t otxg, txg;
2469 itx_async_node_t *ian, ian_search;
2470 avl_tree_t *t;
2471 avl_index_t where;
2472 list_t clean_list;
2473 itx_t *itx;
2474
2475 ASSERT(oid != 0);
2476 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
2477
2478 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2479 otxg = ZILTEST_TXG;
2480 else
2481 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2482
2483 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2484 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2485
2486 mutex_enter(&itxg->itxg_lock);
2487 if (itxg->itxg_txg != txg) {
2488 mutex_exit(&itxg->itxg_lock);
2489 continue;
2490 }
2491
2492 /*
2493 * Locate the object node and append its list.
2494 */
2495 t = &itxg->itxg_itxs->i_async_tree;
2496 ian_search.ia_foid = oid;
2497 ian = avl_find(t, &ian_search, &where);
2498 if (ian != NULL)
2499 list_move_tail(&clean_list, &ian->ia_list);
2500 mutex_exit(&itxg->itxg_lock);
2501 }
2502 while ((itx = list_remove_head(&clean_list)) != NULL) {
2503 /* commit itxs should never be on the async lists. */
2504 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2505 zil_itx_destroy(itx);
2506 }
2507 list_destroy(&clean_list);
2508 }
2509
2510 void
zil_itx_assign(zilog_t * zilog,itx_t * itx,dmu_tx_t * tx)2511 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
2512 {
2513 uint64_t txg;
2514 itxg_t *itxg;
2515 itxs_t *itxs, *clean = NULL;
2516
2517 /*
2518 * Ensure the data of a renamed file is committed before the rename.
2519 */
2520 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
2521 zil_async_to_sync(zilog, itx->itx_oid);
2522
2523 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
2524 txg = ZILTEST_TXG;
2525 else
2526 txg = dmu_tx_get_txg(tx);
2527
2528 itxg = &zilog->zl_itxg[txg & TXG_MASK];
2529 mutex_enter(&itxg->itxg_lock);
2530 itxs = itxg->itxg_itxs;
2531 if (itxg->itxg_txg != txg) {
2532 if (itxs != NULL) {
2533 /*
2534 * The zil_clean callback hasn't got around to cleaning
2535 * this itxg. Save the itxs for release below.
2536 * This should be rare.
2537 */
2538 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2539 "txg %llu", (u_longlong_t)itxg->itxg_txg);
2540 clean = itxg->itxg_itxs;
2541 }
2542 itxg->itxg_txg = txg;
2543 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2544 KM_SLEEP);
2545
2546 list_create(&itxs->i_sync_list, sizeof (itx_t),
2547 offsetof(itx_t, itx_node));
2548 avl_create(&itxs->i_async_tree, zil_aitx_compare,
2549 sizeof (itx_async_node_t),
2550 offsetof(itx_async_node_t, ia_node));
2551 }
2552 if (itx->itx_sync) {
2553 list_insert_tail(&itxs->i_sync_list, itx);
2554 } else {
2555 avl_tree_t *t = &itxs->i_async_tree;
2556 uint64_t foid =
2557 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2558 itx_async_node_t *ian;
2559 avl_index_t where;
2560
2561 ian = avl_find(t, &foid, &where);
2562 if (ian == NULL) {
2563 ian = kmem_alloc(sizeof (itx_async_node_t),
2564 KM_SLEEP);
2565 list_create(&ian->ia_list, sizeof (itx_t),
2566 offsetof(itx_t, itx_node));
2567 ian->ia_foid = foid;
2568 avl_insert(t, ian, where);
2569 }
2570 list_insert_tail(&ian->ia_list, itx);
2571 }
2572
2573 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2574
2575 /*
2576 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2577 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2578 * need to be careful to always dirty the ZIL using the "real"
2579 * TXG (not itxg_txg) even when the SPA is frozen.
2580 */
2581 zilog_dirty(zilog, dmu_tx_get_txg(tx));
2582 mutex_exit(&itxg->itxg_lock);
2583
2584 /* Release the old itxs now we've dropped the lock */
2585 if (clean != NULL)
2586 zil_itxg_clean(clean);
2587 }
2588
2589 /*
2590 * If there are any in-memory intent log transactions which have now been
2591 * synced then start up a taskq to free them. We should only do this after we
2592 * have written out the uberblocks (i.e. txg has been committed) so that
2593 * don't inadvertently clean out in-memory log records that would be required
2594 * by zil_commit().
2595 */
2596 void
zil_clean(zilog_t * zilog,uint64_t synced_txg)2597 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2598 {
2599 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2600 itxs_t *clean_me;
2601
2602 ASSERT3U(synced_txg, <, ZILTEST_TXG);
2603
2604 mutex_enter(&itxg->itxg_lock);
2605 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2606 mutex_exit(&itxg->itxg_lock);
2607 return;
2608 }
2609 ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2610 ASSERT3U(itxg->itxg_txg, !=, 0);
2611 clean_me = itxg->itxg_itxs;
2612 itxg->itxg_itxs = NULL;
2613 itxg->itxg_txg = 0;
2614 mutex_exit(&itxg->itxg_lock);
2615 /*
2616 * Preferably start a task queue to free up the old itxs but
2617 * if taskq_dispatch can't allocate resources to do that then
2618 * free it in-line. This should be rare. Note, using TQ_SLEEP
2619 * created a bad performance problem.
2620 */
2621 ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2622 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2623 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2624 zil_itxg_clean, clean_me, TQ_NOSLEEP);
2625 if (id == TASKQID_INVALID)
2626 zil_itxg_clean(clean_me);
2627 }
2628
2629 /*
2630 * This function will traverse the queue of itxs that need to be
2631 * committed, and move them onto the ZIL's zl_itx_commit_list.
2632 */
2633 static uint64_t
zil_get_commit_list(zilog_t * zilog)2634 zil_get_commit_list(zilog_t *zilog)
2635 {
2636 uint64_t otxg, txg, wtxg = 0;
2637 list_t *commit_list = &zilog->zl_itx_commit_list;
2638
2639 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2640
2641 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2642 otxg = ZILTEST_TXG;
2643 else
2644 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2645
2646 /*
2647 * This is inherently racy, since there is nothing to prevent
2648 * the last synced txg from changing. That's okay since we'll
2649 * only commit things in the future.
2650 */
2651 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2652 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2653
2654 mutex_enter(&itxg->itxg_lock);
2655 if (itxg->itxg_txg != txg) {
2656 mutex_exit(&itxg->itxg_lock);
2657 continue;
2658 }
2659
2660 /*
2661 * If we're adding itx records to the zl_itx_commit_list,
2662 * then the zil better be dirty in this "txg". We can assert
2663 * that here since we're holding the itxg_lock which will
2664 * prevent spa_sync from cleaning it. Once we add the itxs
2665 * to the zl_itx_commit_list we must commit it to disk even
2666 * if it's unnecessary (i.e. the txg was synced).
2667 */
2668 ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2669 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2670 list_t *sync_list = &itxg->itxg_itxs->i_sync_list;
2671 itx_t *itx = NULL;
2672 if (unlikely(zilog->zl_suspend > 0)) {
2673 /*
2674 * ZIL was just suspended, but we lost the race.
2675 * Allow all earlier itxs to be committed, but ask
2676 * caller to do txg_wait_synced(txg) for any new.
2677 */
2678 if (!list_is_empty(sync_list))
2679 wtxg = MAX(wtxg, txg);
2680 } else {
2681 itx = list_head(sync_list);
2682 list_move_tail(commit_list, sync_list);
2683 }
2684
2685 mutex_exit(&itxg->itxg_lock);
2686
2687 while (itx != NULL) {
2688 uint64_t s = zil_itx_full_size(itx);
2689 zilog->zl_cur_size += s;
2690 zilog->zl_cur_left += s;
2691 s = zil_itx_record_size(itx);
2692 zilog->zl_cur_max = MAX(zilog->zl_cur_max, s);
2693 itx = list_next(commit_list, itx);
2694 }
2695 }
2696 return (wtxg);
2697 }
2698
2699 /*
2700 * Move the async itxs for a specified object to commit into sync lists.
2701 */
2702 void
zil_async_to_sync(zilog_t * zilog,uint64_t foid)2703 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2704 {
2705 uint64_t otxg, txg;
2706 itx_async_node_t *ian, ian_search;
2707 avl_tree_t *t;
2708 avl_index_t where;
2709
2710 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2711 otxg = ZILTEST_TXG;
2712 else
2713 otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2714
2715 /*
2716 * This is inherently racy, since there is nothing to prevent
2717 * the last synced txg from changing.
2718 */
2719 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2720 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2721
2722 mutex_enter(&itxg->itxg_lock);
2723 if (itxg->itxg_txg != txg) {
2724 mutex_exit(&itxg->itxg_lock);
2725 continue;
2726 }
2727
2728 /*
2729 * If a foid is specified then find that node and append its
2730 * list. Otherwise walk the tree appending all the lists
2731 * to the sync list. We add to the end rather than the
2732 * beginning to ensure the create has happened.
2733 */
2734 t = &itxg->itxg_itxs->i_async_tree;
2735 if (foid != 0) {
2736 ian_search.ia_foid = foid;
2737 ian = avl_find(t, &ian_search, &where);
2738 if (ian != NULL) {
2739 list_move_tail(&itxg->itxg_itxs->i_sync_list,
2740 &ian->ia_list);
2741 }
2742 } else {
2743 void *cookie = NULL;
2744
2745 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2746 list_move_tail(&itxg->itxg_itxs->i_sync_list,
2747 &ian->ia_list);
2748 list_destroy(&ian->ia_list);
2749 kmem_free(ian, sizeof (itx_async_node_t));
2750 }
2751 }
2752 mutex_exit(&itxg->itxg_lock);
2753 }
2754 }
2755
2756 /*
2757 * This function will prune commit itxs that are at the head of the
2758 * commit list (it won't prune past the first non-commit itx), and
2759 * either: a) attach them to the last lwb that's still pending
2760 * completion, or b) skip them altogether.
2761 *
2762 * This is used as a performance optimization to prevent commit itxs
2763 * from generating new lwbs when it's unnecessary to do so.
2764 */
2765 static void
zil_prune_commit_list(zilog_t * zilog)2766 zil_prune_commit_list(zilog_t *zilog)
2767 {
2768 itx_t *itx;
2769
2770 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2771
2772 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2773 lr_t *lrc = &itx->itx_lr;
2774 if (lrc->lrc_txtype != TX_COMMIT)
2775 break;
2776
2777 mutex_enter(&zilog->zl_lock);
2778
2779 lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2780 if (last_lwb == NULL ||
2781 last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2782 /*
2783 * All of the itxs this waiter was waiting on
2784 * must have already completed (or there were
2785 * never any itx's for it to wait on), so it's
2786 * safe to skip this waiter and mark it done.
2787 */
2788 zil_commit_waiter_skip(itx->itx_private);
2789 } else {
2790 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2791 }
2792
2793 mutex_exit(&zilog->zl_lock);
2794
2795 list_remove(&zilog->zl_itx_commit_list, itx);
2796 zil_itx_destroy(itx);
2797 }
2798
2799 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2800 }
2801
2802 static void
zil_commit_writer_stall(zilog_t * zilog)2803 zil_commit_writer_stall(zilog_t *zilog)
2804 {
2805 /*
2806 * When zio_alloc_zil() fails to allocate the next lwb block on
2807 * disk, we must call txg_wait_synced() to ensure all of the
2808 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2809 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2810 * to zil_process_commit_list()) will have to call zil_create(),
2811 * and start a new ZIL chain.
2812 *
2813 * Since zil_alloc_zil() failed, the lwb that was previously
2814 * issued does not have a pointer to the "next" lwb on disk.
2815 * Thus, if another ZIL writer thread was to allocate the "next"
2816 * on-disk lwb, that block could be leaked in the event of a
2817 * crash (because the previous lwb on-disk would not point to
2818 * it).
2819 *
2820 * We must hold the zilog's zl_issuer_lock while we do this, to
2821 * ensure no new threads enter zil_process_commit_list() until
2822 * all lwb's in the zl_lwb_list have been synced and freed
2823 * (which is achieved via the txg_wait_synced() call).
2824 */
2825 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2826 txg_wait_synced(zilog->zl_dmu_pool, 0);
2827 ASSERT(list_is_empty(&zilog->zl_lwb_list));
2828 }
2829
2830 static void
zil_burst_done(zilog_t * zilog)2831 zil_burst_done(zilog_t *zilog)
2832 {
2833 if (!list_is_empty(&zilog->zl_itx_commit_list) ||
2834 zilog->zl_cur_size == 0)
2835 return;
2836
2837 if (zilog->zl_parallel)
2838 zilog->zl_parallel--;
2839
2840 uint_t r = (zilog->zl_prev_rotor + 1) & (ZIL_BURSTS - 1);
2841 zilog->zl_prev_rotor = r;
2842 zilog->zl_prev_opt[r] = zil_lwb_plan(zilog, zilog->zl_cur_size,
2843 &zilog->zl_prev_min[r]);
2844
2845 zilog->zl_cur_size = 0;
2846 zilog->zl_cur_max = 0;
2847 zilog->zl_cur_left = 0;
2848 }
2849
2850 /*
2851 * This function will traverse the commit list, creating new lwbs as
2852 * needed, and committing the itxs from the commit list to these newly
2853 * created lwbs. Additionally, as a new lwb is created, the previous
2854 * lwb will be issued to the zio layer to be written to disk.
2855 */
2856 static void
zil_process_commit_list(zilog_t * zilog,zil_commit_waiter_t * zcw,list_t * ilwbs)2857 zil_process_commit_list(zilog_t *zilog, zil_commit_waiter_t *zcw, list_t *ilwbs)
2858 {
2859 spa_t *spa = zilog->zl_spa;
2860 list_t nolwb_itxs;
2861 list_t nolwb_waiters;
2862 lwb_t *lwb, *plwb;
2863 itx_t *itx;
2864
2865 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2866
2867 /*
2868 * Return if there's nothing to commit before we dirty the fs by
2869 * calling zil_create().
2870 */
2871 if (list_is_empty(&zilog->zl_itx_commit_list))
2872 return;
2873
2874 list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2875 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2876 offsetof(zil_commit_waiter_t, zcw_node));
2877
2878 lwb = list_tail(&zilog->zl_lwb_list);
2879 if (lwb == NULL) {
2880 lwb = zil_create(zilog);
2881 } else {
2882 /*
2883 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2884 * have already been created (zl_lwb_list not empty).
2885 */
2886 zil_commit_activate_saxattr_feature(zilog);
2887 ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
2888 lwb->lwb_state == LWB_STATE_OPENED);
2889
2890 /*
2891 * If the lwb is still opened, it means the workload is really
2892 * multi-threaded and we won the chance of write aggregation.
2893 * If it is not opened yet, but previous lwb is still not
2894 * flushed, it still means the workload is multi-threaded, but
2895 * there was too much time between the commits to aggregate, so
2896 * we try aggregation next times, but without too much hopes.
2897 */
2898 if (lwb->lwb_state == LWB_STATE_OPENED) {
2899 zilog->zl_parallel = ZIL_BURSTS;
2900 } else if ((plwb = list_prev(&zilog->zl_lwb_list, lwb))
2901 != NULL && plwb->lwb_state != LWB_STATE_FLUSH_DONE) {
2902 zilog->zl_parallel = MAX(zilog->zl_parallel,
2903 ZIL_BURSTS / 2);
2904 }
2905 }
2906
2907 while ((itx = list_remove_head(&zilog->zl_itx_commit_list)) != NULL) {
2908 lr_t *lrc = &itx->itx_lr;
2909 uint64_t txg = lrc->lrc_txg;
2910
2911 ASSERT3U(txg, !=, 0);
2912
2913 if (lrc->lrc_txtype == TX_COMMIT) {
2914 DTRACE_PROBE2(zil__process__commit__itx,
2915 zilog_t *, zilog, itx_t *, itx);
2916 } else {
2917 DTRACE_PROBE2(zil__process__normal__itx,
2918 zilog_t *, zilog, itx_t *, itx);
2919 }
2920
2921 boolean_t synced = txg <= spa_last_synced_txg(spa);
2922 boolean_t frozen = txg > spa_freeze_txg(spa);
2923
2924 /*
2925 * If the txg of this itx has already been synced out, then
2926 * we don't need to commit this itx to an lwb. This is
2927 * because the data of this itx will have already been
2928 * written to the main pool. This is inherently racy, and
2929 * it's still ok to commit an itx whose txg has already
2930 * been synced; this will result in a write that's
2931 * unnecessary, but will do no harm.
2932 *
2933 * With that said, we always want to commit TX_COMMIT itxs
2934 * to an lwb, regardless of whether or not that itx's txg
2935 * has been synced out. We do this to ensure any OPENED lwb
2936 * will always have at least one zil_commit_waiter_t linked
2937 * to the lwb.
2938 *
2939 * As a counter-example, if we skipped TX_COMMIT itx's
2940 * whose txg had already been synced, the following
2941 * situation could occur if we happened to be racing with
2942 * spa_sync:
2943 *
2944 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2945 * itx's txg is 10 and the last synced txg is 9.
2946 * 2. spa_sync finishes syncing out txg 10.
2947 * 3. We move to the next itx in the list, it's a TX_COMMIT
2948 * whose txg is 10, so we skip it rather than committing
2949 * it to the lwb used in (1).
2950 *
2951 * If the itx that is skipped in (3) is the last TX_COMMIT
2952 * itx in the commit list, than it's possible for the lwb
2953 * used in (1) to remain in the OPENED state indefinitely.
2954 *
2955 * To prevent the above scenario from occurring, ensuring
2956 * that once an lwb is OPENED it will transition to ISSUED
2957 * and eventually DONE, we always commit TX_COMMIT itx's to
2958 * an lwb here, even if that itx's txg has already been
2959 * synced.
2960 *
2961 * Finally, if the pool is frozen, we _always_ commit the
2962 * itx. The point of freezing the pool is to prevent data
2963 * from being written to the main pool via spa_sync, and
2964 * instead rely solely on the ZIL to persistently store the
2965 * data; i.e. when the pool is frozen, the last synced txg
2966 * value can't be trusted.
2967 */
2968 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2969 if (lwb != NULL) {
2970 lwb = zil_lwb_assign(zilog, lwb, itx, ilwbs);
2971 if (lwb == NULL) {
2972 list_insert_tail(&nolwb_itxs, itx);
2973 } else if ((zcw->zcw_lwb != NULL &&
2974 zcw->zcw_lwb != lwb) || zcw->zcw_done) {
2975 /*
2976 * Our lwb is done, leave the rest of
2977 * itx list to somebody else who care.
2978 */
2979 zilog->zl_parallel = ZIL_BURSTS;
2980 zilog->zl_cur_left -=
2981 zil_itx_full_size(itx);
2982 break;
2983 }
2984 } else {
2985 if (lrc->lrc_txtype == TX_COMMIT) {
2986 zil_commit_waiter_link_nolwb(
2987 itx->itx_private, &nolwb_waiters);
2988 }
2989 list_insert_tail(&nolwb_itxs, itx);
2990 }
2991 zilog->zl_cur_left -= zil_itx_full_size(itx);
2992 } else {
2993 ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2994 zilog->zl_cur_left -= zil_itx_full_size(itx);
2995 zil_itx_destroy(itx);
2996 }
2997 }
2998
2999 if (lwb == NULL) {
3000 /*
3001 * This indicates zio_alloc_zil() failed to allocate the
3002 * "next" lwb on-disk. When this happens, we must stall
3003 * the ZIL write pipeline; see the comment within
3004 * zil_commit_writer_stall() for more details.
3005 */
3006 while ((lwb = list_remove_head(ilwbs)) != NULL)
3007 zil_lwb_write_issue(zilog, lwb);
3008 zil_commit_writer_stall(zilog);
3009
3010 /*
3011 * Additionally, we have to signal and mark the "nolwb"
3012 * waiters as "done" here, since without an lwb, we
3013 * can't do this via zil_lwb_flush_vdevs_done() like
3014 * normal.
3015 */
3016 zil_commit_waiter_t *zcw;
3017 while ((zcw = list_remove_head(&nolwb_waiters)) != NULL)
3018 zil_commit_waiter_skip(zcw);
3019
3020 /*
3021 * And finally, we have to destroy the itx's that
3022 * couldn't be committed to an lwb; this will also call
3023 * the itx's callback if one exists for the itx.
3024 */
3025 while ((itx = list_remove_head(&nolwb_itxs)) != NULL)
3026 zil_itx_destroy(itx);
3027 } else {
3028 ASSERT(list_is_empty(&nolwb_waiters));
3029 ASSERT3P(lwb, !=, NULL);
3030 ASSERT(lwb->lwb_state == LWB_STATE_NEW ||
3031 lwb->lwb_state == LWB_STATE_OPENED);
3032
3033 /*
3034 * At this point, the ZIL block pointed at by the "lwb"
3035 * variable is in "new" or "opened" state.
3036 *
3037 * If it's "new", then no itxs have been committed to it, so
3038 * there's no point in issuing its zio (i.e. it's "empty").
3039 *
3040 * If it's "opened", then it contains one or more itxs that
3041 * eventually need to be committed to stable storage. In
3042 * this case we intentionally do not issue the lwb's zio
3043 * to disk yet, and instead rely on one of the following
3044 * two mechanisms for issuing the zio:
3045 *
3046 * 1. Ideally, there will be more ZIL activity occurring on
3047 * the system, such that this function will be immediately
3048 * called again by different thread and this lwb will be
3049 * closed by zil_lwb_assign(). This way, the lwb will be
3050 * "full" when it is issued to disk, and we'll make use of
3051 * the lwb's size the best we can.
3052 *
3053 * 2. If there isn't sufficient ZIL activity occurring on
3054 * the system, zil_commit_waiter() will close it and issue
3055 * the zio. If this occurs, the lwb is not guaranteed
3056 * to be "full" by the time its zio is issued, and means
3057 * the size of the lwb was "too large" given the amount
3058 * of ZIL activity occurring on the system at that time.
3059 *
3060 * We do this for a couple of reasons:
3061 *
3062 * 1. To try and reduce the number of IOPs needed to
3063 * write the same number of itxs. If an lwb has space
3064 * available in its buffer for more itxs, and more itxs
3065 * will be committed relatively soon (relative to the
3066 * latency of performing a write), then it's beneficial
3067 * to wait for these "next" itxs. This way, more itxs
3068 * can be committed to stable storage with fewer writes.
3069 *
3070 * 2. To try and use the largest lwb block size that the
3071 * incoming rate of itxs can support. Again, this is to
3072 * try and pack as many itxs into as few lwbs as
3073 * possible, without significantly impacting the latency
3074 * of each individual itx.
3075 */
3076 if (lwb->lwb_state == LWB_STATE_OPENED && !zilog->zl_parallel) {
3077 zil_burst_done(zilog);
3078 list_insert_tail(ilwbs, lwb);
3079 lwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW);
3080 if (lwb == NULL) {
3081 while ((lwb = list_remove_head(ilwbs)) != NULL)
3082 zil_lwb_write_issue(zilog, lwb);
3083 zil_commit_writer_stall(zilog);
3084 }
3085 }
3086 }
3087 }
3088
3089 /*
3090 * This function is responsible for ensuring the passed in commit waiter
3091 * (and associated commit itx) is committed to an lwb. If the waiter is
3092 * not already committed to an lwb, all itxs in the zilog's queue of
3093 * itxs will be processed. The assumption is the passed in waiter's
3094 * commit itx will found in the queue just like the other non-commit
3095 * itxs, such that when the entire queue is processed, the waiter will
3096 * have been committed to an lwb.
3097 *
3098 * The lwb associated with the passed in waiter is not guaranteed to
3099 * have been issued by the time this function completes. If the lwb is
3100 * not issued, we rely on future calls to zil_commit_writer() to issue
3101 * the lwb, or the timeout mechanism found in zil_commit_waiter().
3102 */
3103 static uint64_t
zil_commit_writer(zilog_t * zilog,zil_commit_waiter_t * zcw)3104 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
3105 {
3106 list_t ilwbs;
3107 lwb_t *lwb;
3108 uint64_t wtxg = 0;
3109
3110 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3111 ASSERT(spa_writeable(zilog->zl_spa));
3112
3113 list_create(&ilwbs, sizeof (lwb_t), offsetof(lwb_t, lwb_issue_node));
3114 mutex_enter(&zilog->zl_issuer_lock);
3115
3116 if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
3117 /*
3118 * It's possible that, while we were waiting to acquire
3119 * the "zl_issuer_lock", another thread committed this
3120 * waiter to an lwb. If that occurs, we bail out early,
3121 * without processing any of the zilog's queue of itxs.
3122 *
3123 * On certain workloads and system configurations, the
3124 * "zl_issuer_lock" can become highly contended. In an
3125 * attempt to reduce this contention, we immediately drop
3126 * the lock if the waiter has already been processed.
3127 *
3128 * We've measured this optimization to reduce CPU spent
3129 * contending on this lock by up to 5%, using a system
3130 * with 32 CPUs, low latency storage (~50 usec writes),
3131 * and 1024 threads performing sync writes.
3132 */
3133 goto out;
3134 }
3135
3136 ZIL_STAT_BUMP(zilog, zil_commit_writer_count);
3137
3138 wtxg = zil_get_commit_list(zilog);
3139 zil_prune_commit_list(zilog);
3140 zil_process_commit_list(zilog, zcw, &ilwbs);
3141
3142 out:
3143 mutex_exit(&zilog->zl_issuer_lock);
3144 while ((lwb = list_remove_head(&ilwbs)) != NULL)
3145 zil_lwb_write_issue(zilog, lwb);
3146 list_destroy(&ilwbs);
3147 return (wtxg);
3148 }
3149
3150 static void
zil_commit_waiter_timeout(zilog_t * zilog,zil_commit_waiter_t * zcw)3151 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
3152 {
3153 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3154 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3155 ASSERT3B(zcw->zcw_done, ==, B_FALSE);
3156
3157 lwb_t *lwb = zcw->zcw_lwb;
3158 ASSERT3P(lwb, !=, NULL);
3159 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_NEW);
3160
3161 /*
3162 * If the lwb has already been issued by another thread, we can
3163 * immediately return since there's no work to be done (the
3164 * point of this function is to issue the lwb). Additionally, we
3165 * do this prior to acquiring the zl_issuer_lock, to avoid
3166 * acquiring it when it's not necessary to do so.
3167 */
3168 if (lwb->lwb_state != LWB_STATE_OPENED)
3169 return;
3170
3171 /*
3172 * In order to call zil_lwb_write_close() we must hold the
3173 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
3174 * since we're already holding the commit waiter's "zcw_lock",
3175 * and those two locks are acquired in the opposite order
3176 * elsewhere.
3177 */
3178 mutex_exit(&zcw->zcw_lock);
3179 mutex_enter(&zilog->zl_issuer_lock);
3180 mutex_enter(&zcw->zcw_lock);
3181
3182 /*
3183 * Since we just dropped and re-acquired the commit waiter's
3184 * lock, we have to re-check to see if the waiter was marked
3185 * "done" during that process. If the waiter was marked "done",
3186 * the "lwb" pointer is no longer valid (it can be free'd after
3187 * the waiter is marked "done"), so without this check we could
3188 * wind up with a use-after-free error below.
3189 */
3190 if (zcw->zcw_done) {
3191 mutex_exit(&zilog->zl_issuer_lock);
3192 return;
3193 }
3194
3195 ASSERT3P(lwb, ==, zcw->zcw_lwb);
3196
3197 /*
3198 * We've already checked this above, but since we hadn't acquired
3199 * the zilog's zl_issuer_lock, we have to perform this check a
3200 * second time while holding the lock.
3201 *
3202 * We don't need to hold the zl_lock since the lwb cannot transition
3203 * from OPENED to CLOSED while we hold the zl_issuer_lock. The lwb
3204 * _can_ transition from CLOSED to DONE, but it's OK to race with
3205 * that transition since we treat the lwb the same, whether it's in
3206 * the CLOSED, ISSUED or DONE states.
3207 *
3208 * The important thing, is we treat the lwb differently depending on
3209 * if it's OPENED or CLOSED, and block any other threads that might
3210 * attempt to close/issue this lwb. For that reason we hold the
3211 * zl_issuer_lock when checking the lwb_state; we must not call
3212 * zil_lwb_write_close() if the lwb had already been closed/issued.
3213 *
3214 * See the comment above the lwb_state_t structure definition for
3215 * more details on the lwb states, and locking requirements.
3216 */
3217 if (lwb->lwb_state != LWB_STATE_OPENED) {
3218 mutex_exit(&zilog->zl_issuer_lock);
3219 return;
3220 }
3221
3222 /*
3223 * We do not need zcw_lock once we hold zl_issuer_lock and know lwb
3224 * is still open. But we have to drop it to avoid a deadlock in case
3225 * callback of zio issued by zil_lwb_write_issue() try to get it,
3226 * while zil_lwb_write_issue() is blocked on attempt to issue next
3227 * lwb it found in LWB_STATE_READY state.
3228 */
3229 mutex_exit(&zcw->zcw_lock);
3230
3231 /*
3232 * As described in the comments above zil_commit_waiter() and
3233 * zil_process_commit_list(), we need to issue this lwb's zio
3234 * since we've reached the commit waiter's timeout and it still
3235 * hasn't been issued.
3236 */
3237 zil_burst_done(zilog);
3238 lwb_t *nlwb = zil_lwb_write_close(zilog, lwb, LWB_STATE_NEW);
3239
3240 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_CLOSED);
3241
3242 if (nlwb == NULL) {
3243 /*
3244 * When zil_lwb_write_close() returns NULL, this
3245 * indicates zio_alloc_zil() failed to allocate the
3246 * "next" lwb on-disk. When this occurs, the ZIL write
3247 * pipeline must be stalled; see the comment within the
3248 * zil_commit_writer_stall() function for more details.
3249 */
3250 zil_lwb_write_issue(zilog, lwb);
3251 zil_commit_writer_stall(zilog);
3252 mutex_exit(&zilog->zl_issuer_lock);
3253 } else {
3254 mutex_exit(&zilog->zl_issuer_lock);
3255 zil_lwb_write_issue(zilog, lwb);
3256 }
3257 mutex_enter(&zcw->zcw_lock);
3258 }
3259
3260 /*
3261 * This function is responsible for performing the following two tasks:
3262 *
3263 * 1. its primary responsibility is to block until the given "commit
3264 * waiter" is considered "done".
3265 *
3266 * 2. its secondary responsibility is to issue the zio for the lwb that
3267 * the given "commit waiter" is waiting on, if this function has
3268 * waited "long enough" and the lwb is still in the "open" state.
3269 *
3270 * Given a sufficient amount of itxs being generated and written using
3271 * the ZIL, the lwb's zio will be issued via the zil_lwb_assign()
3272 * function. If this does not occur, this secondary responsibility will
3273 * ensure the lwb is issued even if there is not other synchronous
3274 * activity on the system.
3275 *
3276 * For more details, see zil_process_commit_list(); more specifically,
3277 * the comment at the bottom of that function.
3278 */
3279 static void
zil_commit_waiter(zilog_t * zilog,zil_commit_waiter_t * zcw)3280 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
3281 {
3282 ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3283 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3284 ASSERT(spa_writeable(zilog->zl_spa));
3285
3286 mutex_enter(&zcw->zcw_lock);
3287
3288 /*
3289 * The timeout is scaled based on the lwb latency to avoid
3290 * significantly impacting the latency of each individual itx.
3291 * For more details, see the comment at the bottom of the
3292 * zil_process_commit_list() function.
3293 */
3294 int pct = MAX(zfs_commit_timeout_pct, 1);
3295 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
3296 hrtime_t wakeup = gethrtime() + sleep;
3297 boolean_t timedout = B_FALSE;
3298
3299 while (!zcw->zcw_done) {
3300 ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3301
3302 lwb_t *lwb = zcw->zcw_lwb;
3303
3304 /*
3305 * Usually, the waiter will have a non-NULL lwb field here,
3306 * but it's possible for it to be NULL as a result of
3307 * zil_commit() racing with spa_sync().
3308 *
3309 * When zil_clean() is called, it's possible for the itxg
3310 * list (which may be cleaned via a taskq) to contain
3311 * commit itxs. When this occurs, the commit waiters linked
3312 * off of these commit itxs will not be committed to an
3313 * lwb. Additionally, these commit waiters will not be
3314 * marked done until zil_commit_waiter_skip() is called via
3315 * zil_itxg_clean().
3316 *
3317 * Thus, it's possible for this commit waiter (i.e. the
3318 * "zcw" variable) to be found in this "in between" state;
3319 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3320 * been skipped, so it's "zcw_done" field is still B_FALSE.
3321 */
3322 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_NEW);
3323
3324 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
3325 ASSERT3B(timedout, ==, B_FALSE);
3326
3327 /*
3328 * If the lwb hasn't been issued yet, then we
3329 * need to wait with a timeout, in case this
3330 * function needs to issue the lwb after the
3331 * timeout is reached; responsibility (2) from
3332 * the comment above this function.
3333 */
3334 int rc = cv_timedwait_hires(&zcw->zcw_cv,
3335 &zcw->zcw_lock, wakeup, USEC2NSEC(1),
3336 CALLOUT_FLAG_ABSOLUTE);
3337
3338 if (rc != -1 || zcw->zcw_done)
3339 continue;
3340
3341 timedout = B_TRUE;
3342 zil_commit_waiter_timeout(zilog, zcw);
3343
3344 if (!zcw->zcw_done) {
3345 /*
3346 * If the commit waiter has already been
3347 * marked "done", it's possible for the
3348 * waiter's lwb structure to have already
3349 * been freed. Thus, we can only reliably
3350 * make these assertions if the waiter
3351 * isn't done.
3352 */
3353 ASSERT3P(lwb, ==, zcw->zcw_lwb);
3354 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
3355 }
3356 } else {
3357 /*
3358 * If the lwb isn't open, then it must have already
3359 * been issued. In that case, there's no need to
3360 * use a timeout when waiting for the lwb to
3361 * complete.
3362 *
3363 * Additionally, if the lwb is NULL, the waiter
3364 * will soon be signaled and marked done via
3365 * zil_clean() and zil_itxg_clean(), so no timeout
3366 * is required.
3367 */
3368
3369 IMPLY(lwb != NULL,
3370 lwb->lwb_state == LWB_STATE_CLOSED ||
3371 lwb->lwb_state == LWB_STATE_READY ||
3372 lwb->lwb_state == LWB_STATE_ISSUED ||
3373 lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3374 lwb->lwb_state == LWB_STATE_FLUSH_DONE);
3375 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
3376 }
3377 }
3378
3379 mutex_exit(&zcw->zcw_lock);
3380 }
3381
3382 static zil_commit_waiter_t *
zil_alloc_commit_waiter(void)3383 zil_alloc_commit_waiter(void)
3384 {
3385 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
3386
3387 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
3388 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
3389 list_link_init(&zcw->zcw_node);
3390 zcw->zcw_lwb = NULL;
3391 zcw->zcw_done = B_FALSE;
3392 zcw->zcw_zio_error = 0;
3393
3394 return (zcw);
3395 }
3396
3397 static void
zil_free_commit_waiter(zil_commit_waiter_t * zcw)3398 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
3399 {
3400 ASSERT(!list_link_active(&zcw->zcw_node));
3401 ASSERT3P(zcw->zcw_lwb, ==, NULL);
3402 ASSERT3B(zcw->zcw_done, ==, B_TRUE);
3403 mutex_destroy(&zcw->zcw_lock);
3404 cv_destroy(&zcw->zcw_cv);
3405 kmem_cache_free(zil_zcw_cache, zcw);
3406 }
3407
3408 /*
3409 * This function is used to create a TX_COMMIT itx and assign it. This
3410 * way, it will be linked into the ZIL's list of synchronous itxs, and
3411 * then later committed to an lwb (or skipped) when
3412 * zil_process_commit_list() is called.
3413 */
3414 static void
zil_commit_itx_assign(zilog_t * zilog,zil_commit_waiter_t * zcw)3415 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
3416 {
3417 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
3418
3419 /*
3420 * Since we are not going to create any new dirty data, and we
3421 * can even help with clearing the existing dirty data, we
3422 * should not be subject to the dirty data based delays. We
3423 * use TXG_NOTHROTTLE to bypass the delay mechanism.
3424 */
3425 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
3426
3427 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
3428 itx->itx_sync = B_TRUE;
3429 itx->itx_private = zcw;
3430
3431 zil_itx_assign(zilog, itx, tx);
3432
3433 dmu_tx_commit(tx);
3434 }
3435
3436 /*
3437 * Commit ZFS Intent Log transactions (itxs) to stable storage.
3438 *
3439 * When writing ZIL transactions to the on-disk representation of the
3440 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3441 * itxs can be committed to a single lwb. Once a lwb is written and
3442 * committed to stable storage (i.e. the lwb is written, and vdevs have
3443 * been flushed), each itx that was committed to that lwb is also
3444 * considered to be committed to stable storage.
3445 *
3446 * When an itx is committed to an lwb, the log record (lr_t) contained
3447 * by the itx is copied into the lwb's zio buffer, and once this buffer
3448 * is written to disk, it becomes an on-disk ZIL block.
3449 *
3450 * As itxs are generated, they're inserted into the ZIL's queue of
3451 * uncommitted itxs. The semantics of zil_commit() are such that it will
3452 * block until all itxs that were in the queue when it was called, are
3453 * committed to stable storage.
3454 *
3455 * If "foid" is zero, this means all "synchronous" and "asynchronous"
3456 * itxs, for all objects in the dataset, will be committed to stable
3457 * storage prior to zil_commit() returning. If "foid" is non-zero, all
3458 * "synchronous" itxs for all objects, but only "asynchronous" itxs
3459 * that correspond to the foid passed in, will be committed to stable
3460 * storage prior to zil_commit() returning.
3461 *
3462 * Generally speaking, when zil_commit() is called, the consumer doesn't
3463 * actually care about _all_ of the uncommitted itxs. Instead, they're
3464 * simply trying to waiting for a specific itx to be committed to disk,
3465 * but the interface(s) for interacting with the ZIL don't allow such
3466 * fine-grained communication. A better interface would allow a consumer
3467 * to create and assign an itx, and then pass a reference to this itx to
3468 * zil_commit(); such that zil_commit() would return as soon as that
3469 * specific itx was committed to disk (instead of waiting for _all_
3470 * itxs to be committed).
3471 *
3472 * When a thread calls zil_commit() a special "commit itx" will be
3473 * generated, along with a corresponding "waiter" for this commit itx.
3474 * zil_commit() will wait on this waiter's CV, such that when the waiter
3475 * is marked done, and signaled, zil_commit() will return.
3476 *
3477 * This commit itx is inserted into the queue of uncommitted itxs. This
3478 * provides an easy mechanism for determining which itxs were in the
3479 * queue prior to zil_commit() having been called, and which itxs were
3480 * added after zil_commit() was called.
3481 *
3482 * The commit itx is special; it doesn't have any on-disk representation.
3483 * When a commit itx is "committed" to an lwb, the waiter associated
3484 * with it is linked onto the lwb's list of waiters. Then, when that lwb
3485 * completes, each waiter on the lwb's list is marked done and signaled
3486 * -- allowing the thread waiting on the waiter to return from zil_commit().
3487 *
3488 * It's important to point out a few critical factors that allow us
3489 * to make use of the commit itxs, commit waiters, per-lwb lists of
3490 * commit waiters, and zio completion callbacks like we're doing:
3491 *
3492 * 1. The list of waiters for each lwb is traversed, and each commit
3493 * waiter is marked "done" and signaled, in the zio completion
3494 * callback of the lwb's zio[*].
3495 *
3496 * * Actually, the waiters are signaled in the zio completion
3497 * callback of the root zio for the flush commands that are sent to
3498 * the vdevs upon completion of the lwb zio.
3499 *
3500 * 2. When the itxs are inserted into the ZIL's queue of uncommitted
3501 * itxs, the order in which they are inserted is preserved[*]; as
3502 * itxs are added to the queue, they are added to the tail of
3503 * in-memory linked lists.
3504 *
3505 * When committing the itxs to lwbs (to be written to disk), they
3506 * are committed in the same order in which the itxs were added to
3507 * the uncommitted queue's linked list(s); i.e. the linked list of
3508 * itxs to commit is traversed from head to tail, and each itx is
3509 * committed to an lwb in that order.
3510 *
3511 * * To clarify:
3512 *
3513 * - the order of "sync" itxs is preserved w.r.t. other
3514 * "sync" itxs, regardless of the corresponding objects.
3515 * - the order of "async" itxs is preserved w.r.t. other
3516 * "async" itxs corresponding to the same object.
3517 * - the order of "async" itxs is *not* preserved w.r.t. other
3518 * "async" itxs corresponding to different objects.
3519 * - the order of "sync" itxs w.r.t. "async" itxs (or vice
3520 * versa) is *not* preserved, even for itxs that correspond
3521 * to the same object.
3522 *
3523 * For more details, see: zil_itx_assign(), zil_async_to_sync(),
3524 * zil_get_commit_list(), and zil_process_commit_list().
3525 *
3526 * 3. The lwbs represent a linked list of blocks on disk. Thus, any
3527 * lwb cannot be considered committed to stable storage, until its
3528 * "previous" lwb is also committed to stable storage. This fact,
3529 * coupled with the fact described above, means that itxs are
3530 * committed in (roughly) the order in which they were generated.
3531 * This is essential because itxs are dependent on prior itxs.
3532 * Thus, we *must not* deem an itx as being committed to stable
3533 * storage, until *all* prior itxs have also been committed to
3534 * stable storage.
3535 *
3536 * To enforce this ordering of lwb zio's, while still leveraging as
3537 * much of the underlying storage performance as possible, we rely
3538 * on two fundamental concepts:
3539 *
3540 * 1. The creation and issuance of lwb zio's is protected by
3541 * the zilog's "zl_issuer_lock", which ensures only a single
3542 * thread is creating and/or issuing lwb's at a time
3543 * 2. The "previous" lwb is a child of the "current" lwb
3544 * (leveraging the zio parent-child dependency graph)
3545 *
3546 * By relying on this parent-child zio relationship, we can have
3547 * many lwb zio's concurrently issued to the underlying storage,
3548 * but the order in which they complete will be the same order in
3549 * which they were created.
3550 */
3551 void
zil_commit(zilog_t * zilog,uint64_t foid)3552 zil_commit(zilog_t *zilog, uint64_t foid)
3553 {
3554 /*
3555 * We should never attempt to call zil_commit on a snapshot for
3556 * a couple of reasons:
3557 *
3558 * 1. A snapshot may never be modified, thus it cannot have any
3559 * in-flight itxs that would have modified the dataset.
3560 *
3561 * 2. By design, when zil_commit() is called, a commit itx will
3562 * be assigned to this zilog; as a result, the zilog will be
3563 * dirtied. We must not dirty the zilog of a snapshot; there's
3564 * checks in the code that enforce this invariant, and will
3565 * cause a panic if it's not upheld.
3566 */
3567 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
3568
3569 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3570 return;
3571
3572 if (!spa_writeable(zilog->zl_spa)) {
3573 /*
3574 * If the SPA is not writable, there should never be any
3575 * pending itxs waiting to be committed to disk. If that
3576 * weren't true, we'd skip writing those itxs out, and
3577 * would break the semantics of zil_commit(); thus, we're
3578 * verifying that truth before we return to the caller.
3579 */
3580 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3581 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3582 for (int i = 0; i < TXG_SIZE; i++)
3583 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
3584 return;
3585 }
3586
3587 /*
3588 * If the ZIL is suspended, we don't want to dirty it by calling
3589 * zil_commit_itx_assign() below, nor can we write out
3590 * lwbs like would be done in zil_commit_write(). Thus, we
3591 * simply rely on txg_wait_synced() to maintain the necessary
3592 * semantics, and avoid calling those functions altogether.
3593 */
3594 if (zilog->zl_suspend > 0) {
3595 txg_wait_synced(zilog->zl_dmu_pool, 0);
3596 return;
3597 }
3598
3599 zil_commit_impl(zilog, foid);
3600 }
3601
3602 void
zil_commit_impl(zilog_t * zilog,uint64_t foid)3603 zil_commit_impl(zilog_t *zilog, uint64_t foid)
3604 {
3605 ZIL_STAT_BUMP(zilog, zil_commit_count);
3606
3607 /*
3608 * Move the "async" itxs for the specified foid to the "sync"
3609 * queues, such that they will be later committed (or skipped)
3610 * to an lwb when zil_process_commit_list() is called.
3611 *
3612 * Since these "async" itxs must be committed prior to this
3613 * call to zil_commit returning, we must perform this operation
3614 * before we call zil_commit_itx_assign().
3615 */
3616 zil_async_to_sync(zilog, foid);
3617
3618 /*
3619 * We allocate a new "waiter" structure which will initially be
3620 * linked to the commit itx using the itx's "itx_private" field.
3621 * Since the commit itx doesn't represent any on-disk state,
3622 * when it's committed to an lwb, rather than copying the its
3623 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3624 * added to the lwb's list of waiters. Then, when the lwb is
3625 * committed to stable storage, each waiter in the lwb's list of
3626 * waiters will be marked "done", and signalled.
3627 *
3628 * We must create the waiter and assign the commit itx prior to
3629 * calling zil_commit_writer(), or else our specific commit itx
3630 * is not guaranteed to be committed to an lwb prior to calling
3631 * zil_commit_waiter().
3632 */
3633 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
3634 zil_commit_itx_assign(zilog, zcw);
3635
3636 uint64_t wtxg = zil_commit_writer(zilog, zcw);
3637 zil_commit_waiter(zilog, zcw);
3638
3639 if (zcw->zcw_zio_error != 0) {
3640 /*
3641 * If there was an error writing out the ZIL blocks that
3642 * this thread is waiting on, then we fallback to
3643 * relying on spa_sync() to write out the data this
3644 * thread is waiting on. Obviously this has performance
3645 * implications, but the expectation is for this to be
3646 * an exceptional case, and shouldn't occur often.
3647 */
3648 DTRACE_PROBE2(zil__commit__io__error,
3649 zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3650 txg_wait_synced(zilog->zl_dmu_pool, 0);
3651 } else if (wtxg != 0) {
3652 txg_wait_synced(zilog->zl_dmu_pool, wtxg);
3653 }
3654
3655 zil_free_commit_waiter(zcw);
3656 }
3657
3658 /*
3659 * Called in syncing context to free committed log blocks and update log header.
3660 */
3661 void
zil_sync(zilog_t * zilog,dmu_tx_t * tx)3662 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3663 {
3664 zil_header_t *zh = zil_header_in_syncing_context(zilog);
3665 uint64_t txg = dmu_tx_get_txg(tx);
3666 spa_t *spa = zilog->zl_spa;
3667 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3668 lwb_t *lwb;
3669
3670 /*
3671 * We don't zero out zl_destroy_txg, so make sure we don't try
3672 * to destroy it twice.
3673 */
3674 if (spa_sync_pass(spa) != 1)
3675 return;
3676
3677 zil_lwb_flush_wait_all(zilog, txg);
3678
3679 mutex_enter(&zilog->zl_lock);
3680
3681 ASSERT(zilog->zl_stop_sync == 0);
3682
3683 if (*replayed_seq != 0) {
3684 ASSERT(zh->zh_replay_seq < *replayed_seq);
3685 zh->zh_replay_seq = *replayed_seq;
3686 *replayed_seq = 0;
3687 }
3688
3689 if (zilog->zl_destroy_txg == txg) {
3690 blkptr_t blk = zh->zh_log;
3691 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
3692
3693 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3694
3695 memset(zh, 0, sizeof (zil_header_t));
3696 memset(zilog->zl_replayed_seq, 0,
3697 sizeof (zilog->zl_replayed_seq));
3698
3699 if (zilog->zl_keep_first) {
3700 /*
3701 * If this block was part of log chain that couldn't
3702 * be claimed because a device was missing during
3703 * zil_claim(), but that device later returns,
3704 * then this block could erroneously appear valid.
3705 * To guard against this, assign a new GUID to the new
3706 * log chain so it doesn't matter what blk points to.
3707 */
3708 zil_init_log_chain(zilog, &blk);
3709 zh->zh_log = blk;
3710 } else {
3711 /*
3712 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3713 * records. So, deactivate the feature for this dataset.
3714 * We activate it again when we start a new ZIL chain.
3715 */
3716 if (dsl_dataset_feature_is_active(ds,
3717 SPA_FEATURE_ZILSAXATTR))
3718 dsl_dataset_deactivate_feature(ds,
3719 SPA_FEATURE_ZILSAXATTR, tx);
3720 }
3721 }
3722
3723 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3724 zh->zh_log = lwb->lwb_blk;
3725 if (lwb->lwb_state != LWB_STATE_FLUSH_DONE ||
3726 lwb->lwb_alloc_txg > txg || lwb->lwb_max_txg > txg)
3727 break;
3728 list_remove(&zilog->zl_lwb_list, lwb);
3729 if (!BP_IS_HOLE(&lwb->lwb_blk))
3730 zio_free(spa, txg, &lwb->lwb_blk);
3731 zil_free_lwb(zilog, lwb);
3732
3733 /*
3734 * If we don't have anything left in the lwb list then
3735 * we've had an allocation failure and we need to zero
3736 * out the zil_header blkptr so that we don't end
3737 * up freeing the same block twice.
3738 */
3739 if (list_is_empty(&zilog->zl_lwb_list))
3740 BP_ZERO(&zh->zh_log);
3741 }
3742
3743 mutex_exit(&zilog->zl_lock);
3744 }
3745
3746 static int
zil_lwb_cons(void * vbuf,void * unused,int kmflag)3747 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3748 {
3749 (void) unused, (void) kmflag;
3750 lwb_t *lwb = vbuf;
3751 list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3752 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3753 offsetof(zil_commit_waiter_t, zcw_node));
3754 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3755 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3756 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3757 return (0);
3758 }
3759
3760 static void
zil_lwb_dest(void * vbuf,void * unused)3761 zil_lwb_dest(void *vbuf, void *unused)
3762 {
3763 (void) unused;
3764 lwb_t *lwb = vbuf;
3765 mutex_destroy(&lwb->lwb_vdev_lock);
3766 avl_destroy(&lwb->lwb_vdev_tree);
3767 list_destroy(&lwb->lwb_waiters);
3768 list_destroy(&lwb->lwb_itxs);
3769 }
3770
3771 void
zil_init(void)3772 zil_init(void)
3773 {
3774 zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3775 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3776
3777 zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3778 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3779
3780 zil_sums_init(&zil_sums_global);
3781 zil_kstats_global = kstat_create("zfs", 0, "zil", "misc",
3782 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3783 KSTAT_FLAG_VIRTUAL);
3784
3785 if (zil_kstats_global != NULL) {
3786 zil_kstats_global->ks_data = &zil_stats;
3787 zil_kstats_global->ks_update = zil_kstats_global_update;
3788 zil_kstats_global->ks_private = NULL;
3789 kstat_install(zil_kstats_global);
3790 }
3791 }
3792
3793 void
zil_fini(void)3794 zil_fini(void)
3795 {
3796 kmem_cache_destroy(zil_zcw_cache);
3797 kmem_cache_destroy(zil_lwb_cache);
3798
3799 if (zil_kstats_global != NULL) {
3800 kstat_delete(zil_kstats_global);
3801 zil_kstats_global = NULL;
3802 }
3803
3804 zil_sums_fini(&zil_sums_global);
3805 }
3806
3807 void
zil_set_sync(zilog_t * zilog,uint64_t sync)3808 zil_set_sync(zilog_t *zilog, uint64_t sync)
3809 {
3810 zilog->zl_sync = sync;
3811 }
3812
3813 void
zil_set_logbias(zilog_t * zilog,uint64_t logbias)3814 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3815 {
3816 zilog->zl_logbias = logbias;
3817 }
3818
3819 zilog_t *
zil_alloc(objset_t * os,zil_header_t * zh_phys)3820 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3821 {
3822 zilog_t *zilog;
3823
3824 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3825
3826 zilog->zl_header = zh_phys;
3827 zilog->zl_os = os;
3828 zilog->zl_spa = dmu_objset_spa(os);
3829 zilog->zl_dmu_pool = dmu_objset_pool(os);
3830 zilog->zl_destroy_txg = TXG_INITIAL - 1;
3831 zilog->zl_logbias = dmu_objset_logbias(os);
3832 zilog->zl_sync = dmu_objset_syncprop(os);
3833 zilog->zl_dirty_max_txg = 0;
3834 zilog->zl_last_lwb_opened = NULL;
3835 zilog->zl_last_lwb_latency = 0;
3836 zilog->zl_max_block_size = MIN(MAX(P2ALIGN_TYPED(zil_maxblocksize,
3837 ZIL_MIN_BLKSZ, uint64_t), ZIL_MIN_BLKSZ),
3838 spa_maxblocksize(dmu_objset_spa(os)));
3839
3840 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3841 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3842 mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL);
3843
3844 for (int i = 0; i < TXG_SIZE; i++) {
3845 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3846 MUTEX_DEFAULT, NULL);
3847 }
3848
3849 list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3850 offsetof(lwb_t, lwb_node));
3851
3852 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3853 offsetof(itx_t, itx_node));
3854
3855 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3856 cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL);
3857
3858 for (int i = 0; i < ZIL_BURSTS; i++) {
3859 zilog->zl_prev_opt[i] = zilog->zl_max_block_size -
3860 sizeof (zil_chain_t);
3861 }
3862
3863 return (zilog);
3864 }
3865
3866 void
zil_free(zilog_t * zilog)3867 zil_free(zilog_t *zilog)
3868 {
3869 int i;
3870
3871 zilog->zl_stop_sync = 1;
3872
3873 ASSERT0(zilog->zl_suspend);
3874 ASSERT0(zilog->zl_suspending);
3875
3876 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3877 list_destroy(&zilog->zl_lwb_list);
3878
3879 ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3880 list_destroy(&zilog->zl_itx_commit_list);
3881
3882 for (i = 0; i < TXG_SIZE; i++) {
3883 /*
3884 * It's possible for an itx to be generated that doesn't dirty
3885 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3886 * callback to remove the entry. We remove those here.
3887 *
3888 * Also free up the ziltest itxs.
3889 */
3890 if (zilog->zl_itxg[i].itxg_itxs)
3891 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3892 mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3893 }
3894
3895 mutex_destroy(&zilog->zl_issuer_lock);
3896 mutex_destroy(&zilog->zl_lock);
3897 mutex_destroy(&zilog->zl_lwb_io_lock);
3898
3899 cv_destroy(&zilog->zl_cv_suspend);
3900 cv_destroy(&zilog->zl_lwb_io_cv);
3901
3902 kmem_free(zilog, sizeof (zilog_t));
3903 }
3904
3905 /*
3906 * Open an intent log.
3907 */
3908 zilog_t *
zil_open(objset_t * os,zil_get_data_t * get_data,zil_sums_t * zil_sums)3909 zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums)
3910 {
3911 zilog_t *zilog = dmu_objset_zil(os);
3912
3913 ASSERT3P(zilog->zl_get_data, ==, NULL);
3914 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3915 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3916
3917 zilog->zl_get_data = get_data;
3918 zilog->zl_sums = zil_sums;
3919
3920 return (zilog);
3921 }
3922
3923 /*
3924 * Close an intent log.
3925 */
3926 void
zil_close(zilog_t * zilog)3927 zil_close(zilog_t *zilog)
3928 {
3929 lwb_t *lwb;
3930 uint64_t txg;
3931
3932 if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3933 zil_commit(zilog, 0);
3934 } else {
3935 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3936 ASSERT0(zilog->zl_dirty_max_txg);
3937 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3938 }
3939
3940 mutex_enter(&zilog->zl_lock);
3941 txg = zilog->zl_dirty_max_txg;
3942 lwb = list_tail(&zilog->zl_lwb_list);
3943 if (lwb != NULL) {
3944 txg = MAX(txg, lwb->lwb_alloc_txg);
3945 txg = MAX(txg, lwb->lwb_max_txg);
3946 }
3947 mutex_exit(&zilog->zl_lock);
3948
3949 /*
3950 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3951 * on the time when the dmu_tx transaction is assigned in
3952 * zil_lwb_write_issue().
3953 */
3954 mutex_enter(&zilog->zl_lwb_io_lock);
3955 txg = MAX(zilog->zl_lwb_max_issued_txg, txg);
3956 mutex_exit(&zilog->zl_lwb_io_lock);
3957
3958 /*
3959 * We need to use txg_wait_synced() to wait until that txg is synced.
3960 * zil_sync() will guarantee all lwbs up to that txg have been
3961 * written out, flushed, and cleaned.
3962 */
3963 if (txg != 0)
3964 txg_wait_synced(zilog->zl_dmu_pool, txg);
3965
3966 if (zilog_is_dirty(zilog))
3967 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
3968 (u_longlong_t)txg);
3969 if (txg < spa_freeze_txg(zilog->zl_spa))
3970 VERIFY(!zilog_is_dirty(zilog));
3971
3972 zilog->zl_get_data = NULL;
3973
3974 /*
3975 * We should have only one lwb left on the list; remove it now.
3976 */
3977 mutex_enter(&zilog->zl_lock);
3978 lwb = list_remove_head(&zilog->zl_lwb_list);
3979 if (lwb != NULL) {
3980 ASSERT(list_is_empty(&zilog->zl_lwb_list));
3981 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_NEW);
3982 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3983 zil_free_lwb(zilog, lwb);
3984 }
3985 mutex_exit(&zilog->zl_lock);
3986 }
3987
3988 static const char *suspend_tag = "zil suspending";
3989
3990 /*
3991 * Suspend an intent log. While in suspended mode, we still honor
3992 * synchronous semantics, but we rely on txg_wait_synced() to do it.
3993 * On old version pools, we suspend the log briefly when taking a
3994 * snapshot so that it will have an empty intent log.
3995 *
3996 * Long holds are not really intended to be used the way we do here --
3997 * held for such a short time. A concurrent caller of dsl_dataset_long_held()
3998 * could fail. Therefore we take pains to only put a long hold if it is
3999 * actually necessary. Fortunately, it will only be necessary if the
4000 * objset is currently mounted (or the ZVOL equivalent). In that case it
4001 * will already have a long hold, so we are not really making things any worse.
4002 *
4003 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
4004 * zvol_state_t), and use their mechanism to prevent their hold from being
4005 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for
4006 * very little gain.
4007 *
4008 * if cookiep == NULL, this does both the suspend & resume.
4009 * Otherwise, it returns with the dataset "long held", and the cookie
4010 * should be passed into zil_resume().
4011 */
4012 int
zil_suspend(const char * osname,void ** cookiep)4013 zil_suspend(const char *osname, void **cookiep)
4014 {
4015 objset_t *os;
4016 zilog_t *zilog;
4017 const zil_header_t *zh;
4018 int error;
4019
4020 error = dmu_objset_hold(osname, suspend_tag, &os);
4021 if (error != 0)
4022 return (error);
4023 zilog = dmu_objset_zil(os);
4024
4025 mutex_enter(&zilog->zl_lock);
4026 zh = zilog->zl_header;
4027
4028 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */
4029 mutex_exit(&zilog->zl_lock);
4030 dmu_objset_rele(os, suspend_tag);
4031 return (SET_ERROR(EBUSY));
4032 }
4033
4034 /*
4035 * Don't put a long hold in the cases where we can avoid it. This
4036 * is when there is no cookie so we are doing a suspend & resume
4037 * (i.e. called from zil_vdev_offline()), and there's nothing to do
4038 * for the suspend because it's already suspended, or there's no ZIL.
4039 */
4040 if (cookiep == NULL && !zilog->zl_suspending &&
4041 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
4042 mutex_exit(&zilog->zl_lock);
4043 dmu_objset_rele(os, suspend_tag);
4044 return (0);
4045 }
4046
4047 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
4048 dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
4049
4050 zilog->zl_suspend++;
4051
4052 if (zilog->zl_suspend > 1) {
4053 /*
4054 * Someone else is already suspending it.
4055 * Just wait for them to finish.
4056 */
4057
4058 while (zilog->zl_suspending)
4059 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
4060 mutex_exit(&zilog->zl_lock);
4061
4062 if (cookiep == NULL)
4063 zil_resume(os);
4064 else
4065 *cookiep = os;
4066 return (0);
4067 }
4068
4069 /*
4070 * If there is no pointer to an on-disk block, this ZIL must not
4071 * be active (e.g. filesystem not mounted), so there's nothing
4072 * to clean up.
4073 */
4074 if (BP_IS_HOLE(&zh->zh_log)) {
4075 ASSERT(cookiep != NULL); /* fast path already handled */
4076
4077 *cookiep = os;
4078 mutex_exit(&zilog->zl_lock);
4079 return (0);
4080 }
4081
4082 /*
4083 * The ZIL has work to do. Ensure that the associated encryption
4084 * key will remain mapped while we are committing the log by
4085 * grabbing a reference to it. If the key isn't loaded we have no
4086 * choice but to return an error until the wrapping key is loaded.
4087 */
4088 if (os->os_encrypted &&
4089 dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
4090 zilog->zl_suspend--;
4091 mutex_exit(&zilog->zl_lock);
4092 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
4093 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
4094 return (SET_ERROR(EACCES));
4095 }
4096
4097 zilog->zl_suspending = B_TRUE;
4098 mutex_exit(&zilog->zl_lock);
4099
4100 /*
4101 * We need to use zil_commit_impl to ensure we wait for all
4102 * LWB_STATE_OPENED, _CLOSED and _READY lwbs to be committed
4103 * to disk before proceeding. If we used zil_commit instead, it
4104 * would just call txg_wait_synced(), because zl_suspend is set.
4105 * txg_wait_synced() doesn't wait for these lwb's to be
4106 * LWB_STATE_FLUSH_DONE before returning.
4107 */
4108 zil_commit_impl(zilog, 0);
4109
4110 /*
4111 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
4112 * use txg_wait_synced() to ensure the data from the zilog has
4113 * migrated to the main pool before calling zil_destroy().
4114 */
4115 txg_wait_synced(zilog->zl_dmu_pool, 0);
4116
4117 zil_destroy(zilog, B_FALSE);
4118
4119 mutex_enter(&zilog->zl_lock);
4120 zilog->zl_suspending = B_FALSE;
4121 cv_broadcast(&zilog->zl_cv_suspend);
4122 mutex_exit(&zilog->zl_lock);
4123
4124 if (os->os_encrypted)
4125 dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
4126
4127 if (cookiep == NULL)
4128 zil_resume(os);
4129 else
4130 *cookiep = os;
4131 return (0);
4132 }
4133
4134 void
zil_resume(void * cookie)4135 zil_resume(void *cookie)
4136 {
4137 objset_t *os = cookie;
4138 zilog_t *zilog = dmu_objset_zil(os);
4139
4140 mutex_enter(&zilog->zl_lock);
4141 ASSERT(zilog->zl_suspend != 0);
4142 zilog->zl_suspend--;
4143 mutex_exit(&zilog->zl_lock);
4144 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
4145 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
4146 }
4147
4148 typedef struct zil_replay_arg {
4149 zil_replay_func_t *const *zr_replay;
4150 void *zr_arg;
4151 boolean_t zr_byteswap;
4152 char *zr_lr;
4153 } zil_replay_arg_t;
4154
4155 static int
zil_replay_error(zilog_t * zilog,const lr_t * lr,int error)4156 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
4157 {
4158 char name[ZFS_MAX_DATASET_NAME_LEN];
4159
4160 zilog->zl_replaying_seq--; /* didn't actually replay this one */
4161
4162 dmu_objset_name(zilog->zl_os, name);
4163
4164 cmn_err(CE_WARN, "ZFS replay transaction error %d, "
4165 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
4166 (u_longlong_t)lr->lrc_seq,
4167 (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
4168 (lr->lrc_txtype & TX_CI) ? "CI" : "");
4169
4170 return (error);
4171 }
4172
4173 static int
zil_replay_log_record(zilog_t * zilog,const lr_t * lr,void * zra,uint64_t claim_txg)4174 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
4175 uint64_t claim_txg)
4176 {
4177 zil_replay_arg_t *zr = zra;
4178 const zil_header_t *zh = zilog->zl_header;
4179 uint64_t reclen = lr->lrc_reclen;
4180 uint64_t txtype = lr->lrc_txtype;
4181 int error = 0;
4182
4183 zilog->zl_replaying_seq = lr->lrc_seq;
4184
4185 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */
4186 return (0);
4187
4188 if (lr->lrc_txg < claim_txg) /* already committed */
4189 return (0);
4190
4191 /* Strip case-insensitive bit, still present in log record */
4192 txtype &= ~TX_CI;
4193
4194 if (txtype == 0 || txtype >= TX_MAX_TYPE)
4195 return (zil_replay_error(zilog, lr, EINVAL));
4196
4197 /*
4198 * If this record type can be logged out of order, the object
4199 * (lr_foid) may no longer exist. That's legitimate, not an error.
4200 */
4201 if (TX_OOO(txtype)) {
4202 error = dmu_object_info(zilog->zl_os,
4203 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
4204 if (error == ENOENT || error == EEXIST)
4205 return (0);
4206 }
4207
4208 /*
4209 * Make a copy of the data so we can revise and extend it.
4210 */
4211 memcpy(zr->zr_lr, lr, reclen);
4212
4213 /*
4214 * If this is a TX_WRITE with a blkptr, suck in the data.
4215 */
4216 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
4217 error = zil_read_log_data(zilog, (lr_write_t *)lr,
4218 zr->zr_lr + reclen);
4219 if (error != 0)
4220 return (zil_replay_error(zilog, lr, error));
4221 }
4222
4223 /*
4224 * The log block containing this lr may have been byteswapped
4225 * so that we can easily examine common fields like lrc_txtype.
4226 * However, the log is a mix of different record types, and only the
4227 * replay vectors know how to byteswap their records. Therefore, if
4228 * the lr was byteswapped, undo it before invoking the replay vector.
4229 */
4230 if (zr->zr_byteswap)
4231 byteswap_uint64_array(zr->zr_lr, reclen);
4232
4233 /*
4234 * We must now do two things atomically: replay this log record,
4235 * and update the log header sequence number to reflect the fact that
4236 * we did so. At the end of each replay function the sequence number
4237 * is updated if we are in replay mode.
4238 */
4239 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
4240 if (error != 0) {
4241 /*
4242 * The DMU's dnode layer doesn't see removes until the txg
4243 * commits, so a subsequent claim can spuriously fail with
4244 * EEXIST. So if we receive any error we try syncing out
4245 * any removes then retry the transaction. Note that we
4246 * specify B_FALSE for byteswap now, so we don't do it twice.
4247 */
4248 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
4249 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
4250 if (error != 0)
4251 return (zil_replay_error(zilog, lr, error));
4252 }
4253 return (0);
4254 }
4255
4256 static int
zil_incr_blks(zilog_t * zilog,const blkptr_t * bp,void * arg,uint64_t claim_txg)4257 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
4258 {
4259 (void) bp, (void) arg, (void) claim_txg;
4260
4261 zilog->zl_replay_blks++;
4262
4263 return (0);
4264 }
4265
4266 /*
4267 * If this dataset has a non-empty intent log, replay it and destroy it.
4268 * Return B_TRUE if there were any entries to replay.
4269 */
4270 boolean_t
zil_replay(objset_t * os,void * arg,zil_replay_func_t * const replay_func[TX_MAX_TYPE])4271 zil_replay(objset_t *os, void *arg,
4272 zil_replay_func_t *const replay_func[TX_MAX_TYPE])
4273 {
4274 zilog_t *zilog = dmu_objset_zil(os);
4275 const zil_header_t *zh = zilog->zl_header;
4276 zil_replay_arg_t zr;
4277
4278 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
4279 return (zil_destroy(zilog, B_TRUE));
4280 }
4281
4282 zr.zr_replay = replay_func;
4283 zr.zr_arg = arg;
4284 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
4285 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
4286
4287 /*
4288 * Wait for in-progress removes to sync before starting replay.
4289 */
4290 txg_wait_synced(zilog->zl_dmu_pool, 0);
4291
4292 zilog->zl_replay = B_TRUE;
4293 zilog->zl_replay_time = ddi_get_lbolt();
4294 ASSERT(zilog->zl_replay_blks == 0);
4295 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
4296 zh->zh_claim_txg, B_TRUE);
4297 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
4298
4299 zil_destroy(zilog, B_FALSE);
4300 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
4301 zilog->zl_replay = B_FALSE;
4302
4303 return (B_TRUE);
4304 }
4305
4306 boolean_t
zil_replaying(zilog_t * zilog,dmu_tx_t * tx)4307 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
4308 {
4309 if (zilog->zl_sync == ZFS_SYNC_DISABLED)
4310 return (B_TRUE);
4311
4312 if (zilog->zl_replay) {
4313 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
4314 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
4315 zilog->zl_replaying_seq;
4316 return (B_TRUE);
4317 }
4318
4319 return (B_FALSE);
4320 }
4321
4322 int
zil_reset(const char * osname,void * arg)4323 zil_reset(const char *osname, void *arg)
4324 {
4325 (void) arg;
4326
4327 int error = zil_suspend(osname, NULL);
4328 /* EACCES means crypto key not loaded */
4329 if ((error == EACCES) || (error == EBUSY))
4330 return (SET_ERROR(error));
4331 if (error != 0)
4332 return (SET_ERROR(EEXIST));
4333 return (0);
4334 }
4335
4336 EXPORT_SYMBOL(zil_alloc);
4337 EXPORT_SYMBOL(zil_free);
4338 EXPORT_SYMBOL(zil_open);
4339 EXPORT_SYMBOL(zil_close);
4340 EXPORT_SYMBOL(zil_replay);
4341 EXPORT_SYMBOL(zil_replaying);
4342 EXPORT_SYMBOL(zil_destroy);
4343 EXPORT_SYMBOL(zil_destroy_sync);
4344 EXPORT_SYMBOL(zil_itx_create);
4345 EXPORT_SYMBOL(zil_itx_destroy);
4346 EXPORT_SYMBOL(zil_itx_assign);
4347 EXPORT_SYMBOL(zil_commit);
4348 EXPORT_SYMBOL(zil_claim);
4349 EXPORT_SYMBOL(zil_check_log_chain);
4350 EXPORT_SYMBOL(zil_sync);
4351 EXPORT_SYMBOL(zil_clean);
4352 EXPORT_SYMBOL(zil_suspend);
4353 EXPORT_SYMBOL(zil_resume);
4354 EXPORT_SYMBOL(zil_lwb_add_block);
4355 EXPORT_SYMBOL(zil_bp_tree_add);
4356 EXPORT_SYMBOL(zil_set_sync);
4357 EXPORT_SYMBOL(zil_set_logbias);
4358 EXPORT_SYMBOL(zil_sums_init);
4359 EXPORT_SYMBOL(zil_sums_fini);
4360 EXPORT_SYMBOL(zil_kstat_values_update);
4361
4362 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW,
4363 "ZIL block open timeout percentage");
4364
4365 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
4366 "Disable intent logging replay");
4367
4368 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
4369 "Disable ZIL cache flushes");
4370
4371 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, U64, ZMOD_RW,
4372 "Limit in bytes slog sync writes per commit");
4373
4374 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW,
4375 "Limit in bytes of ZIL log block size");
4376
4377 ZFS_MODULE_PARAM(zfs_zil, zil_, maxcopied, UINT, ZMOD_RW,
4378 "Limit in bytes WR_COPIED size");
4379