1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 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 47 /* 48 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system 49 * calls that change the file system. Each itx has enough information to 50 * be able to replay them after a system crash, power loss, or 51 * equivalent failure mode. These are stored in memory until either: 52 * 53 * 1. they are committed to the pool by the DMU transaction group 54 * (txg), at which point they can be discarded; or 55 * 2. they are committed to the on-disk ZIL for the dataset being 56 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous 57 * requirement). 58 * 59 * In the event of a crash or power loss, the itxs contained by each 60 * dataset's on-disk ZIL will be replayed when that dataset is first 61 * instantiated (e.g. if the dataset is a normal filesystem, when it is 62 * first mounted). 63 * 64 * As hinted at above, there is one ZIL per dataset (both the in-memory 65 * representation, and the on-disk representation). The on-disk format 66 * consists of 3 parts: 67 * 68 * - a single, per-dataset, ZIL header; which points to a chain of 69 * - zero or more ZIL blocks; each of which contains 70 * - zero or more ZIL records 71 * 72 * A ZIL record holds the information necessary to replay a single 73 * system call transaction. A ZIL block can hold many ZIL records, and 74 * the blocks are chained together, similarly to a singly linked list. 75 * 76 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL 77 * block in the chain, and the ZIL header points to the first block in 78 * the chain. 79 * 80 * Note, there is not a fixed place in the pool to hold these ZIL 81 * blocks; they are dynamically allocated and freed as needed from the 82 * blocks available on the pool, though they can be preferentially 83 * allocated from a dedicated "log" vdev. 84 */ 85 86 /* 87 * This controls the amount of time that a ZIL block (lwb) will remain 88 * "open" when it isn't "full", and it has a thread waiting for it to be 89 * committed to stable storage. Please refer to the zil_commit_waiter() 90 * function (and the comments within it) for more details. 91 */ 92 int zfs_commit_timeout_pct = 5; 93 94 /* 95 * See zil.h for more information about these fields. 96 */ 97 zil_stats_t zil_stats = { 98 { "zil_commit_count", KSTAT_DATA_UINT64 }, 99 { "zil_commit_writer_count", KSTAT_DATA_UINT64 }, 100 { "zil_itx_count", KSTAT_DATA_UINT64 }, 101 { "zil_itx_indirect_count", KSTAT_DATA_UINT64 }, 102 { "zil_itx_indirect_bytes", KSTAT_DATA_UINT64 }, 103 { "zil_itx_copied_count", KSTAT_DATA_UINT64 }, 104 { "zil_itx_copied_bytes", KSTAT_DATA_UINT64 }, 105 { "zil_itx_needcopy_count", KSTAT_DATA_UINT64 }, 106 { "zil_itx_needcopy_bytes", KSTAT_DATA_UINT64 }, 107 { "zil_itx_metaslab_normal_count", KSTAT_DATA_UINT64 }, 108 { "zil_itx_metaslab_normal_bytes", KSTAT_DATA_UINT64 }, 109 { "zil_itx_metaslab_slog_count", KSTAT_DATA_UINT64 }, 110 { "zil_itx_metaslab_slog_bytes", KSTAT_DATA_UINT64 }, 111 }; 112 113 static kstat_t *zil_ksp; 114 115 /* 116 * Disable intent logging replay. This global ZIL switch affects all pools. 117 */ 118 int zil_replay_disable = 0; 119 120 /* 121 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to 122 * the disk(s) by the ZIL after an LWB write has completed. Setting this 123 * will cause ZIL corruption on power loss if a volatile out-of-order 124 * write cache is enabled. 125 */ 126 int zil_nocacheflush = 0; 127 128 /* 129 * Limit SLOG write size per commit executed with synchronous priority. 130 * Any writes above that will be executed with lower (asynchronous) priority 131 * to limit potential SLOG device abuse by single active ZIL writer. 132 */ 133 unsigned long zil_slog_bulk = 768 * 1024; 134 135 static kmem_cache_t *zil_lwb_cache; 136 static kmem_cache_t *zil_zcw_cache; 137 138 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 139 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 140 141 static int 142 zil_bp_compare(const void *x1, const void *x2) 143 { 144 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 145 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 146 147 int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2)); 148 if (likely(cmp)) 149 return (cmp); 150 151 return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2))); 152 } 153 154 static void 155 zil_bp_tree_init(zilog_t *zilog) 156 { 157 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 158 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 159 } 160 161 static void 162 zil_bp_tree_fini(zilog_t *zilog) 163 { 164 avl_tree_t *t = &zilog->zl_bp_tree; 165 zil_bp_node_t *zn; 166 void *cookie = NULL; 167 168 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 169 kmem_free(zn, sizeof (zil_bp_node_t)); 170 171 avl_destroy(t); 172 } 173 174 int 175 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 176 { 177 avl_tree_t *t = &zilog->zl_bp_tree; 178 const dva_t *dva; 179 zil_bp_node_t *zn; 180 avl_index_t where; 181 182 if (BP_IS_EMBEDDED(bp)) 183 return (0); 184 185 dva = BP_IDENTITY(bp); 186 187 if (avl_find(t, dva, &where) != NULL) 188 return (SET_ERROR(EEXIST)); 189 190 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 191 zn->zn_dva = *dva; 192 avl_insert(t, zn, where); 193 194 return (0); 195 } 196 197 static zil_header_t * 198 zil_header_in_syncing_context(zilog_t *zilog) 199 { 200 return ((zil_header_t *)zilog->zl_header); 201 } 202 203 static void 204 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 205 { 206 zio_cksum_t *zc = &bp->blk_cksum; 207 208 (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0], 209 sizeof (zc->zc_word[ZIL_ZC_GUID_0])); 210 (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1], 211 sizeof (zc->zc_word[ZIL_ZC_GUID_1])); 212 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 213 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 214 } 215 216 /* 217 * Read a log block and make sure it's valid. 218 */ 219 static int 220 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp, 221 blkptr_t *nbp, void *dst, char **end) 222 { 223 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 224 arc_flags_t aflags = ARC_FLAG_WAIT; 225 arc_buf_t *abuf = NULL; 226 zbookmark_phys_t zb; 227 int error; 228 229 if (zilog->zl_header->zh_claim_txg == 0) 230 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 231 232 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 233 zio_flags |= ZIO_FLAG_SPECULATIVE; 234 235 if (!decrypt) 236 zio_flags |= ZIO_FLAG_RAW; 237 238 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 239 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 240 241 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, 242 &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 243 244 if (error == 0) { 245 zio_cksum_t cksum = bp->blk_cksum; 246 247 /* 248 * Validate the checksummed log block. 249 * 250 * Sequence numbers should be... sequential. The checksum 251 * verifier for the next block should be bp's checksum plus 1. 252 * 253 * Also check the log chain linkage and size used. 254 */ 255 cksum.zc_word[ZIL_ZC_SEQ]++; 256 257 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 258 zil_chain_t *zilc = abuf->b_data; 259 char *lr = (char *)(zilc + 1); 260 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 261 262 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 263 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 264 error = SET_ERROR(ECKSUM); 265 } else { 266 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 267 bcopy(lr, dst, len); 268 *end = (char *)dst + len; 269 *nbp = zilc->zc_next_blk; 270 } 271 } else { 272 char *lr = abuf->b_data; 273 uint64_t size = BP_GET_LSIZE(bp); 274 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 275 276 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 277 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 278 (zilc->zc_nused > (size - sizeof (*zilc)))) { 279 error = SET_ERROR(ECKSUM); 280 } else { 281 ASSERT3U(zilc->zc_nused, <=, 282 SPA_OLD_MAXBLOCKSIZE); 283 bcopy(lr, dst, zilc->zc_nused); 284 *end = (char *)dst + zilc->zc_nused; 285 *nbp = zilc->zc_next_blk; 286 } 287 } 288 289 arc_buf_destroy(abuf, &abuf); 290 } 291 292 return (error); 293 } 294 295 /* 296 * Read a TX_WRITE log data block. 297 */ 298 static int 299 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 300 { 301 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 302 const blkptr_t *bp = &lr->lr_blkptr; 303 arc_flags_t aflags = ARC_FLAG_WAIT; 304 arc_buf_t *abuf = NULL; 305 zbookmark_phys_t zb; 306 int error; 307 308 if (BP_IS_HOLE(bp)) { 309 if (wbuf != NULL) 310 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 311 return (0); 312 } 313 314 if (zilog->zl_header->zh_claim_txg == 0) 315 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 316 317 /* 318 * If we are not using the resulting data, we are just checking that 319 * it hasn't been corrupted so we don't need to waste CPU time 320 * decompressing and decrypting it. 321 */ 322 if (wbuf == NULL) 323 zio_flags |= ZIO_FLAG_RAW; 324 325 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 326 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 327 328 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 329 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 330 331 if (error == 0) { 332 if (wbuf != NULL) 333 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 334 arc_buf_destroy(abuf, &abuf); 335 } 336 337 return (error); 338 } 339 340 /* 341 * Parse the intent log, and call parse_func for each valid record within. 342 */ 343 int 344 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 345 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg, 346 boolean_t decrypt) 347 { 348 const zil_header_t *zh = zilog->zl_header; 349 boolean_t claimed = !!zh->zh_claim_txg; 350 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 351 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 352 uint64_t max_blk_seq = 0; 353 uint64_t max_lr_seq = 0; 354 uint64_t blk_count = 0; 355 uint64_t lr_count = 0; 356 blkptr_t blk, next_blk; 357 char *lrbuf, *lrp; 358 int error = 0; 359 360 bzero(&next_blk, sizeof (blkptr_t)); 361 362 /* 363 * Old logs didn't record the maximum zh_claim_lr_seq. 364 */ 365 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 366 claim_lr_seq = UINT64_MAX; 367 368 /* 369 * Starting at the block pointed to by zh_log we read the log chain. 370 * For each block in the chain we strongly check that block to 371 * ensure its validity. We stop when an invalid block is found. 372 * For each block pointer in the chain we call parse_blk_func(). 373 * For each record in each valid block we call parse_lr_func(). 374 * If the log has been claimed, stop if we encounter a sequence 375 * number greater than the highest claimed sequence number. 376 */ 377 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 378 zil_bp_tree_init(zilog); 379 380 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 381 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 382 int reclen; 383 char *end = NULL; 384 385 if (blk_seq > claim_blk_seq) 386 break; 387 388 error = parse_blk_func(zilog, &blk, arg, txg); 389 if (error != 0) 390 break; 391 ASSERT3U(max_blk_seq, <, blk_seq); 392 max_blk_seq = blk_seq; 393 blk_count++; 394 395 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 396 break; 397 398 error = zil_read_log_block(zilog, decrypt, &blk, &next_blk, 399 lrbuf, &end); 400 if (error != 0) 401 break; 402 403 for (lrp = lrbuf; lrp < end; lrp += reclen) { 404 lr_t *lr = (lr_t *)lrp; 405 reclen = lr->lrc_reclen; 406 ASSERT3U(reclen, >=, sizeof (lr_t)); 407 if (lr->lrc_seq > claim_lr_seq) 408 goto done; 409 410 error = parse_lr_func(zilog, lr, arg, txg); 411 if (error != 0) 412 goto done; 413 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 414 max_lr_seq = lr->lrc_seq; 415 lr_count++; 416 } 417 } 418 done: 419 zilog->zl_parse_error = error; 420 zilog->zl_parse_blk_seq = max_blk_seq; 421 zilog->zl_parse_lr_seq = max_lr_seq; 422 zilog->zl_parse_blk_count = blk_count; 423 zilog->zl_parse_lr_count = lr_count; 424 425 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 426 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq) || 427 (decrypt && error == EIO)); 428 429 zil_bp_tree_fini(zilog); 430 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 431 432 return (error); 433 } 434 435 /* ARGSUSED */ 436 static int 437 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, 438 uint64_t first_txg) 439 { 440 ASSERT(!BP_IS_HOLE(bp)); 441 442 /* 443 * As we call this function from the context of a rewind to a 444 * checkpoint, each ZIL block whose txg is later than the txg 445 * that we rewind to is invalid. Thus, we return -1 so 446 * zil_parse() doesn't attempt to read it. 447 */ 448 if (bp->blk_birth >= first_txg) 449 return (-1); 450 451 if (zil_bp_tree_add(zilog, bp) != 0) 452 return (0); 453 454 zio_free(zilog->zl_spa, first_txg, bp); 455 return (0); 456 } 457 458 /* ARGSUSED */ 459 static int 460 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, 461 uint64_t first_txg) 462 { 463 return (0); 464 } 465 466 static int 467 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, 468 uint64_t first_txg) 469 { 470 /* 471 * Claim log block if not already committed and not already claimed. 472 * If tx == NULL, just verify that the block is claimable. 473 */ 474 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 475 zil_bp_tree_add(zilog, bp) != 0) 476 return (0); 477 478 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 479 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 480 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 481 } 482 483 static int 484 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, 485 uint64_t first_txg) 486 { 487 lr_write_t *lr = (lr_write_t *)lrc; 488 int error; 489 490 if (lrc->lrc_txtype != TX_WRITE) 491 return (0); 492 493 /* 494 * If the block is not readable, don't claim it. This can happen 495 * in normal operation when a log block is written to disk before 496 * some of the dmu_sync() blocks it points to. In this case, the 497 * transaction cannot have been committed to anyone (we would have 498 * waited for all writes to be stable first), so it is semantically 499 * correct to declare this the end of the log. 500 */ 501 if (lr->lr_blkptr.blk_birth >= first_txg) { 502 error = zil_read_log_data(zilog, lr, NULL); 503 if (error != 0) 504 return (error); 505 } 506 507 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 508 } 509 510 /* ARGSUSED */ 511 static int 512 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, 513 uint64_t claim_txg) 514 { 515 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 516 517 return (0); 518 } 519 520 static int 521 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, 522 uint64_t claim_txg) 523 { 524 lr_write_t *lr = (lr_write_t *)lrc; 525 blkptr_t *bp = &lr->lr_blkptr; 526 527 /* 528 * If we previously claimed it, we need to free it. 529 */ 530 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 531 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 532 !BP_IS_HOLE(bp)) 533 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 534 535 return (0); 536 } 537 538 static int 539 zil_lwb_vdev_compare(const void *x1, const void *x2) 540 { 541 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 542 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 543 544 return (TREE_CMP(v1, v2)); 545 } 546 547 static lwb_t * 548 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg, 549 boolean_t fastwrite) 550 { 551 lwb_t *lwb; 552 553 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 554 lwb->lwb_zilog = zilog; 555 lwb->lwb_blk = *bp; 556 lwb->lwb_fastwrite = fastwrite; 557 lwb->lwb_slog = slog; 558 lwb->lwb_state = LWB_STATE_CLOSED; 559 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 560 lwb->lwb_max_txg = txg; 561 lwb->lwb_write_zio = NULL; 562 lwb->lwb_root_zio = NULL; 563 lwb->lwb_tx = NULL; 564 lwb->lwb_issued_timestamp = 0; 565 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 566 lwb->lwb_nused = sizeof (zil_chain_t); 567 lwb->lwb_sz = BP_GET_LSIZE(bp); 568 } else { 569 lwb->lwb_nused = 0; 570 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 571 } 572 573 mutex_enter(&zilog->zl_lock); 574 list_insert_tail(&zilog->zl_lwb_list, lwb); 575 mutex_exit(&zilog->zl_lock); 576 577 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); 578 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 579 VERIFY(list_is_empty(&lwb->lwb_waiters)); 580 VERIFY(list_is_empty(&lwb->lwb_itxs)); 581 582 return (lwb); 583 } 584 585 static void 586 zil_free_lwb(zilog_t *zilog, lwb_t *lwb) 587 { 588 ASSERT(MUTEX_HELD(&zilog->zl_lock)); 589 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); 590 VERIFY(list_is_empty(&lwb->lwb_waiters)); 591 VERIFY(list_is_empty(&lwb->lwb_itxs)); 592 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 593 ASSERT3P(lwb->lwb_write_zio, ==, NULL); 594 ASSERT3P(lwb->lwb_root_zio, ==, NULL); 595 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa)); 596 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED || 597 lwb->lwb_state == LWB_STATE_FLUSH_DONE); 598 599 /* 600 * Clear the zilog's field to indicate this lwb is no longer 601 * valid, and prevent use-after-free errors. 602 */ 603 if (zilog->zl_last_lwb_opened == lwb) 604 zilog->zl_last_lwb_opened = NULL; 605 606 kmem_cache_free(zil_lwb_cache, lwb); 607 } 608 609 /* 610 * Called when we create in-memory log transactions so that we know 611 * to cleanup the itxs at the end of spa_sync(). 612 */ 613 static void 614 zilog_dirty(zilog_t *zilog, uint64_t txg) 615 { 616 dsl_pool_t *dp = zilog->zl_dmu_pool; 617 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 618 619 ASSERT(spa_writeable(zilog->zl_spa)); 620 621 if (ds->ds_is_snapshot) 622 panic("dirtying snapshot!"); 623 624 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 625 /* up the hold count until we can be written out */ 626 dmu_buf_add_ref(ds->ds_dbuf, zilog); 627 628 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg); 629 } 630 } 631 632 /* 633 * Determine if the zil is dirty in the specified txg. Callers wanting to 634 * ensure that the dirty state does not change must hold the itxg_lock for 635 * the specified txg. Holding the lock will ensure that the zil cannot be 636 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 637 * state. 638 */ 639 static boolean_t __maybe_unused 640 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 641 { 642 dsl_pool_t *dp = zilog->zl_dmu_pool; 643 644 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 645 return (B_TRUE); 646 return (B_FALSE); 647 } 648 649 /* 650 * Determine if the zil is dirty. The zil is considered dirty if it has 651 * any pending itx records that have not been cleaned by zil_clean(). 652 */ 653 static boolean_t 654 zilog_is_dirty(zilog_t *zilog) 655 { 656 dsl_pool_t *dp = zilog->zl_dmu_pool; 657 658 for (int t = 0; t < TXG_SIZE; t++) { 659 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 660 return (B_TRUE); 661 } 662 return (B_FALSE); 663 } 664 665 /* 666 * Create an on-disk intent log. 667 */ 668 static lwb_t * 669 zil_create(zilog_t *zilog) 670 { 671 const zil_header_t *zh = zilog->zl_header; 672 lwb_t *lwb = NULL; 673 uint64_t txg = 0; 674 dmu_tx_t *tx = NULL; 675 blkptr_t blk; 676 int error = 0; 677 boolean_t fastwrite = FALSE; 678 boolean_t slog = FALSE; 679 680 /* 681 * Wait for any previous destroy to complete. 682 */ 683 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 684 685 ASSERT(zh->zh_claim_txg == 0); 686 ASSERT(zh->zh_replay_seq == 0); 687 688 blk = zh->zh_log; 689 690 /* 691 * Allocate an initial log block if: 692 * - there isn't one already 693 * - the existing block is the wrong endianness 694 */ 695 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 696 tx = dmu_tx_create(zilog->zl_os); 697 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 698 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 699 txg = dmu_tx_get_txg(tx); 700 701 if (!BP_IS_HOLE(&blk)) { 702 zio_free(zilog->zl_spa, txg, &blk); 703 BP_ZERO(&blk); 704 } 705 706 error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk, 707 ZIL_MIN_BLKSZ, &slog); 708 fastwrite = TRUE; 709 710 if (error == 0) 711 zil_init_log_chain(zilog, &blk); 712 } 713 714 /* 715 * Allocate a log write block (lwb) for the first log block. 716 */ 717 if (error == 0) 718 lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite); 719 720 /* 721 * If we just allocated the first log block, commit our transaction 722 * and wait for zil_sync() to stuff the block pointer into zh_log. 723 * (zh is part of the MOS, so we cannot modify it in open context.) 724 */ 725 if (tx != NULL) { 726 dmu_tx_commit(tx); 727 txg_wait_synced(zilog->zl_dmu_pool, txg); 728 } 729 730 ASSERT(error != 0 || bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 731 IMPLY(error == 0, lwb != NULL); 732 733 return (lwb); 734 } 735 736 /* 737 * In one tx, free all log blocks and clear the log header. If keep_first 738 * is set, then we're replaying a log with no content. We want to keep the 739 * first block, however, so that the first synchronous transaction doesn't 740 * require a txg_wait_synced() in zil_create(). We don't need to 741 * txg_wait_synced() here either when keep_first is set, because both 742 * zil_create() and zil_destroy() will wait for any in-progress destroys 743 * to complete. 744 */ 745 void 746 zil_destroy(zilog_t *zilog, boolean_t keep_first) 747 { 748 const zil_header_t *zh = zilog->zl_header; 749 lwb_t *lwb; 750 dmu_tx_t *tx; 751 uint64_t txg; 752 753 /* 754 * Wait for any previous destroy to complete. 755 */ 756 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 757 758 zilog->zl_old_header = *zh; /* debugging aid */ 759 760 if (BP_IS_HOLE(&zh->zh_log)) 761 return; 762 763 tx = dmu_tx_create(zilog->zl_os); 764 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 765 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 766 txg = dmu_tx_get_txg(tx); 767 768 mutex_enter(&zilog->zl_lock); 769 770 ASSERT3U(zilog->zl_destroy_txg, <, txg); 771 zilog->zl_destroy_txg = txg; 772 zilog->zl_keep_first = keep_first; 773 774 if (!list_is_empty(&zilog->zl_lwb_list)) { 775 ASSERT(zh->zh_claim_txg == 0); 776 VERIFY(!keep_first); 777 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 778 if (lwb->lwb_fastwrite) 779 metaslab_fastwrite_unmark(zilog->zl_spa, 780 &lwb->lwb_blk); 781 782 list_remove(&zilog->zl_lwb_list, lwb); 783 if (lwb->lwb_buf != NULL) 784 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 785 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk); 786 zil_free_lwb(zilog, lwb); 787 } 788 } else if (!keep_first) { 789 zil_destroy_sync(zilog, tx); 790 } 791 mutex_exit(&zilog->zl_lock); 792 793 dmu_tx_commit(tx); 794 } 795 796 void 797 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 798 { 799 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 800 (void) zil_parse(zilog, zil_free_log_block, 801 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE); 802 } 803 804 int 805 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 806 { 807 dmu_tx_t *tx = txarg; 808 zilog_t *zilog; 809 uint64_t first_txg; 810 zil_header_t *zh; 811 objset_t *os; 812 int error; 813 814 error = dmu_objset_own_obj(dp, ds->ds_object, 815 DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os); 816 if (error != 0) { 817 /* 818 * EBUSY indicates that the objset is inconsistent, in which 819 * case it can not have a ZIL. 820 */ 821 if (error != EBUSY) { 822 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 823 (unsigned long long)ds->ds_object, error); 824 } 825 826 return (0); 827 } 828 829 zilog = dmu_objset_zil(os); 830 zh = zil_header_in_syncing_context(zilog); 831 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa)); 832 first_txg = spa_min_claim_txg(zilog->zl_spa); 833 834 /* 835 * If the spa_log_state is not set to be cleared, check whether 836 * the current uberblock is a checkpoint one and if the current 837 * header has been claimed before moving on. 838 * 839 * If the current uberblock is a checkpointed uberblock then 840 * one of the following scenarios took place: 841 * 842 * 1] We are currently rewinding to the checkpoint of the pool. 843 * 2] We crashed in the middle of a checkpoint rewind but we 844 * did manage to write the checkpointed uberblock to the 845 * vdev labels, so when we tried to import the pool again 846 * the checkpointed uberblock was selected from the import 847 * procedure. 848 * 849 * In both cases we want to zero out all the ZIL blocks, except 850 * the ones that have been claimed at the time of the checkpoint 851 * (their zh_claim_txg != 0). The reason is that these blocks 852 * may be corrupted since we may have reused their locations on 853 * disk after we took the checkpoint. 854 * 855 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier 856 * when we first figure out whether the current uberblock is 857 * checkpointed or not. Unfortunately, that would discard all 858 * the logs, including the ones that are claimed, and we would 859 * leak space. 860 */ 861 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR || 862 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && 863 zh->zh_claim_txg == 0)) { 864 if (!BP_IS_HOLE(&zh->zh_log)) { 865 (void) zil_parse(zilog, zil_clear_log_block, 866 zil_noop_log_record, tx, first_txg, B_FALSE); 867 } 868 BP_ZERO(&zh->zh_log); 869 if (os->os_encrypted) 870 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE; 871 dsl_dataset_dirty(dmu_objset_ds(os), tx); 872 dmu_objset_disown(os, B_FALSE, FTAG); 873 return (0); 874 } 875 876 /* 877 * If we are not rewinding and opening the pool normally, then 878 * the min_claim_txg should be equal to the first txg of the pool. 879 */ 880 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa)); 881 882 /* 883 * Claim all log blocks if we haven't already done so, and remember 884 * the highest claimed sequence number. This ensures that if we can 885 * read only part of the log now (e.g. due to a missing device), 886 * but we can read the entire log later, we will not try to replay 887 * or destroy beyond the last block we successfully claimed. 888 */ 889 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 890 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 891 (void) zil_parse(zilog, zil_claim_log_block, 892 zil_claim_log_record, tx, first_txg, B_FALSE); 893 zh->zh_claim_txg = first_txg; 894 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 895 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 896 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 897 zh->zh_flags |= ZIL_REPLAY_NEEDED; 898 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 899 if (os->os_encrypted) 900 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE; 901 dsl_dataset_dirty(dmu_objset_ds(os), tx); 902 } 903 904 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 905 dmu_objset_disown(os, B_FALSE, FTAG); 906 return (0); 907 } 908 909 /* 910 * Check the log by walking the log chain. 911 * Checksum errors are ok as they indicate the end of the chain. 912 * Any other error (no device or read failure) returns an error. 913 */ 914 /* ARGSUSED */ 915 int 916 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 917 { 918 zilog_t *zilog; 919 objset_t *os; 920 blkptr_t *bp; 921 int error; 922 923 ASSERT(tx == NULL); 924 925 error = dmu_objset_from_ds(ds, &os); 926 if (error != 0) { 927 cmn_err(CE_WARN, "can't open objset %llu, error %d", 928 (unsigned long long)ds->ds_object, error); 929 return (0); 930 } 931 932 zilog = dmu_objset_zil(os); 933 bp = (blkptr_t *)&zilog->zl_header->zh_log; 934 935 if (!BP_IS_HOLE(bp)) { 936 vdev_t *vd; 937 boolean_t valid = B_TRUE; 938 939 /* 940 * Check the first block and determine if it's on a log device 941 * which may have been removed or faulted prior to loading this 942 * pool. If so, there's no point in checking the rest of the 943 * log as its content should have already been synced to the 944 * pool. 945 */ 946 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 947 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 948 if (vd->vdev_islog && vdev_is_dead(vd)) 949 valid = vdev_log_state_valid(vd); 950 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 951 952 if (!valid) 953 return (0); 954 955 /* 956 * Check whether the current uberblock is checkpointed (e.g. 957 * we are rewinding) and whether the current header has been 958 * claimed or not. If it hasn't then skip verifying it. We 959 * do this because its ZIL blocks may be part of the pool's 960 * state before the rewind, which is no longer valid. 961 */ 962 zil_header_t *zh = zil_header_in_syncing_context(zilog); 963 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && 964 zh->zh_claim_txg == 0) 965 return (0); 966 } 967 968 /* 969 * Because tx == NULL, zil_claim_log_block() will not actually claim 970 * any blocks, but just determine whether it is possible to do so. 971 * In addition to checking the log chain, zil_claim_log_block() 972 * will invoke zio_claim() with a done func of spa_claim_notify(), 973 * which will update spa_max_claim_txg. See spa_load() for details. 974 */ 975 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 976 zilog->zl_header->zh_claim_txg ? -1ULL : 977 spa_min_claim_txg(os->os_spa), B_FALSE); 978 979 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 980 } 981 982 /* 983 * When an itx is "skipped", this function is used to properly mark the 984 * waiter as "done, and signal any thread(s) waiting on it. An itx can 985 * be skipped (and not committed to an lwb) for a variety of reasons, 986 * one of them being that the itx was committed via spa_sync(), prior to 987 * it being committed to an lwb; this can happen if a thread calling 988 * zil_commit() is racing with spa_sync(). 989 */ 990 static void 991 zil_commit_waiter_skip(zil_commit_waiter_t *zcw) 992 { 993 mutex_enter(&zcw->zcw_lock); 994 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 995 zcw->zcw_done = B_TRUE; 996 cv_broadcast(&zcw->zcw_cv); 997 mutex_exit(&zcw->zcw_lock); 998 } 999 1000 /* 1001 * This function is used when the given waiter is to be linked into an 1002 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb. 1003 * At this point, the waiter will no longer be referenced by the itx, 1004 * and instead, will be referenced by the lwb. 1005 */ 1006 static void 1007 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb) 1008 { 1009 /* 1010 * The lwb_waiters field of the lwb is protected by the zilog's 1011 * zl_lock, thus it must be held when calling this function. 1012 */ 1013 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock)); 1014 1015 mutex_enter(&zcw->zcw_lock); 1016 ASSERT(!list_link_active(&zcw->zcw_node)); 1017 ASSERT3P(zcw->zcw_lwb, ==, NULL); 1018 ASSERT3P(lwb, !=, NULL); 1019 ASSERT(lwb->lwb_state == LWB_STATE_OPENED || 1020 lwb->lwb_state == LWB_STATE_ISSUED || 1021 lwb->lwb_state == LWB_STATE_WRITE_DONE); 1022 1023 list_insert_tail(&lwb->lwb_waiters, zcw); 1024 zcw->zcw_lwb = lwb; 1025 mutex_exit(&zcw->zcw_lock); 1026 } 1027 1028 /* 1029 * This function is used when zio_alloc_zil() fails to allocate a ZIL 1030 * block, and the given waiter must be linked to the "nolwb waiters" 1031 * list inside of zil_process_commit_list(). 1032 */ 1033 static void 1034 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb) 1035 { 1036 mutex_enter(&zcw->zcw_lock); 1037 ASSERT(!list_link_active(&zcw->zcw_node)); 1038 ASSERT3P(zcw->zcw_lwb, ==, NULL); 1039 list_insert_tail(nolwb, zcw); 1040 mutex_exit(&zcw->zcw_lock); 1041 } 1042 1043 void 1044 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp) 1045 { 1046 avl_tree_t *t = &lwb->lwb_vdev_tree; 1047 avl_index_t where; 1048 zil_vdev_node_t *zv, zvsearch; 1049 int ndvas = BP_GET_NDVAS(bp); 1050 int i; 1051 1052 if (zil_nocacheflush) 1053 return; 1054 1055 mutex_enter(&lwb->lwb_vdev_lock); 1056 for (i = 0; i < ndvas; i++) { 1057 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 1058 if (avl_find(t, &zvsearch, &where) == NULL) { 1059 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 1060 zv->zv_vdev = zvsearch.zv_vdev; 1061 avl_insert(t, zv, where); 1062 } 1063 } 1064 mutex_exit(&lwb->lwb_vdev_lock); 1065 } 1066 1067 static void 1068 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb) 1069 { 1070 avl_tree_t *src = &lwb->lwb_vdev_tree; 1071 avl_tree_t *dst = &nlwb->lwb_vdev_tree; 1072 void *cookie = NULL; 1073 zil_vdev_node_t *zv; 1074 1075 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); 1076 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 1077 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 1078 1079 /* 1080 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does 1081 * not need the protection of lwb_vdev_lock (it will only be modified 1082 * while holding zilog->zl_lock) as its writes and those of its 1083 * children have all completed. The younger 'nlwb' may be waiting on 1084 * future writes to additional vdevs. 1085 */ 1086 mutex_enter(&nlwb->lwb_vdev_lock); 1087 /* 1088 * Tear down the 'lwb' vdev tree, ensuring that entries which do not 1089 * exist in 'nlwb' are moved to it, freeing any would-be duplicates. 1090 */ 1091 while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) { 1092 avl_index_t where; 1093 1094 if (avl_find(dst, zv, &where) == NULL) { 1095 avl_insert(dst, zv, where); 1096 } else { 1097 kmem_free(zv, sizeof (*zv)); 1098 } 1099 } 1100 mutex_exit(&nlwb->lwb_vdev_lock); 1101 } 1102 1103 void 1104 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg) 1105 { 1106 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1107 } 1108 1109 /* 1110 * This function is a called after all vdevs associated with a given lwb 1111 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon 1112 * as the lwb write completes, if "zil_nocacheflush" is set. Further, 1113 * all "previous" lwb's will have completed before this function is 1114 * called; i.e. this function is called for all previous lwbs before 1115 * it's called for "this" lwb (enforced via zio the dependencies 1116 * configured in zil_lwb_set_zio_dependency()). 1117 * 1118 * The intention is for this function to be called as soon as the 1119 * contents of an lwb are considered "stable" on disk, and will survive 1120 * any sudden loss of power. At this point, any threads waiting for the 1121 * lwb to reach this state are signalled, and the "waiter" structures 1122 * are marked "done". 1123 */ 1124 static void 1125 zil_lwb_flush_vdevs_done(zio_t *zio) 1126 { 1127 lwb_t *lwb = zio->io_private; 1128 zilog_t *zilog = lwb->lwb_zilog; 1129 dmu_tx_t *tx = lwb->lwb_tx; 1130 zil_commit_waiter_t *zcw; 1131 itx_t *itx; 1132 1133 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb); 1134 1135 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1136 1137 mutex_enter(&zilog->zl_lock); 1138 1139 /* 1140 * Ensure the lwb buffer pointer is cleared before releasing the 1141 * txg. If we have had an allocation failure and the txg is 1142 * waiting to sync then we want zil_sync() to remove the lwb so 1143 * that it's not picked up as the next new one in 1144 * zil_process_commit_list(). zil_sync() will only remove the 1145 * lwb if lwb_buf is null. 1146 */ 1147 lwb->lwb_buf = NULL; 1148 lwb->lwb_tx = NULL; 1149 1150 ASSERT3U(lwb->lwb_issued_timestamp, >, 0); 1151 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp; 1152 1153 lwb->lwb_root_zio = NULL; 1154 1155 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); 1156 lwb->lwb_state = LWB_STATE_FLUSH_DONE; 1157 1158 if (zilog->zl_last_lwb_opened == lwb) { 1159 /* 1160 * Remember the highest committed log sequence number 1161 * for ztest. We only update this value when all the log 1162 * writes succeeded, because ztest wants to ASSERT that 1163 * it got the whole log chain. 1164 */ 1165 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1166 } 1167 1168 while ((itx = list_head(&lwb->lwb_itxs)) != NULL) { 1169 list_remove(&lwb->lwb_itxs, itx); 1170 zil_itx_destroy(itx); 1171 } 1172 1173 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) { 1174 mutex_enter(&zcw->zcw_lock); 1175 1176 ASSERT(list_link_active(&zcw->zcw_node)); 1177 list_remove(&lwb->lwb_waiters, zcw); 1178 1179 ASSERT3P(zcw->zcw_lwb, ==, lwb); 1180 zcw->zcw_lwb = NULL; 1181 /* 1182 * We expect any ZIO errors from child ZIOs to have been 1183 * propagated "up" to this specific LWB's root ZIO, in 1184 * order for this error handling to work correctly. This 1185 * includes ZIO errors from either this LWB's write or 1186 * flush, as well as any errors from other dependent LWBs 1187 * (e.g. a root LWB ZIO that might be a child of this LWB). 1188 * 1189 * With that said, it's important to note that LWB flush 1190 * errors are not propagated up to the LWB root ZIO. 1191 * This is incorrect behavior, and results in VDEV flush 1192 * errors not being handled correctly here. See the 1193 * comment above the call to "zio_flush" for details. 1194 */ 1195 1196 zcw->zcw_zio_error = zio->io_error; 1197 1198 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 1199 zcw->zcw_done = B_TRUE; 1200 cv_broadcast(&zcw->zcw_cv); 1201 1202 mutex_exit(&zcw->zcw_lock); 1203 } 1204 1205 mutex_exit(&zilog->zl_lock); 1206 1207 /* 1208 * Now that we've written this log block, we have a stable pointer 1209 * to the next block in the chain, so it's OK to let the txg in 1210 * which we allocated the next block sync. 1211 */ 1212 dmu_tx_commit(tx); 1213 } 1214 1215 /* 1216 * This is called when an lwb's write zio completes. The callback's 1217 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs 1218 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved 1219 * in writing out this specific lwb's data, and in the case that cache 1220 * flushes have been deferred, vdevs involved in writing the data for 1221 * previous lwbs. The writes corresponding to all the vdevs in the 1222 * lwb_vdev_tree will have completed by the time this is called, due to 1223 * the zio dependencies configured in zil_lwb_set_zio_dependency(), 1224 * which takes deferred flushes into account. The lwb will be "done" 1225 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio 1226 * completion callback for the lwb's root zio. 1227 */ 1228 static void 1229 zil_lwb_write_done(zio_t *zio) 1230 { 1231 lwb_t *lwb = zio->io_private; 1232 spa_t *spa = zio->io_spa; 1233 zilog_t *zilog = lwb->lwb_zilog; 1234 avl_tree_t *t = &lwb->lwb_vdev_tree; 1235 void *cookie = NULL; 1236 zil_vdev_node_t *zv; 1237 lwb_t *nlwb; 1238 1239 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0); 1240 1241 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1242 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 1243 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 1244 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 1245 ASSERT(!BP_IS_GANG(zio->io_bp)); 1246 ASSERT(!BP_IS_HOLE(zio->io_bp)); 1247 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 1248 1249 abd_free(zio->io_abd); 1250 1251 mutex_enter(&zilog->zl_lock); 1252 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED); 1253 lwb->lwb_state = LWB_STATE_WRITE_DONE; 1254 lwb->lwb_write_zio = NULL; 1255 lwb->lwb_fastwrite = FALSE; 1256 nlwb = list_next(&zilog->zl_lwb_list, lwb); 1257 mutex_exit(&zilog->zl_lock); 1258 1259 if (avl_numnodes(t) == 0) 1260 return; 1261 1262 /* 1263 * If there was an IO error, we're not going to call zio_flush() 1264 * on these vdevs, so we simply empty the tree and free the 1265 * nodes. We avoid calling zio_flush() since there isn't any 1266 * good reason for doing so, after the lwb block failed to be 1267 * written out. 1268 * 1269 * Additionally, we don't perform any further error handling at 1270 * this point (e.g. setting "zcw_zio_error" appropriately), as 1271 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus, 1272 * we expect any error seen here, to have been propagated to 1273 * that function). 1274 */ 1275 if (zio->io_error != 0) { 1276 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) 1277 kmem_free(zv, sizeof (*zv)); 1278 return; 1279 } 1280 1281 /* 1282 * If this lwb does not have any threads waiting for it to 1283 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE 1284 * command to the vdevs written to by "this" lwb, and instead 1285 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE 1286 * command for those vdevs. Thus, we merge the vdev tree of 1287 * "this" lwb with the vdev tree of the "next" lwb in the list, 1288 * and assume the "next" lwb will handle flushing the vdevs (or 1289 * deferring the flush(s) again). 1290 * 1291 * This is a useful performance optimization, especially for 1292 * workloads with lots of async write activity and few sync 1293 * write and/or fsync activity, as it has the potential to 1294 * coalesce multiple flush commands to a vdev into one. 1295 */ 1296 if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) { 1297 zil_lwb_flush_defer(lwb, nlwb); 1298 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 1299 return; 1300 } 1301 1302 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 1303 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 1304 if (vd != NULL) { 1305 /* 1306 * The "ZIO_FLAG_DONT_PROPAGATE" is currently 1307 * always used within "zio_flush". This means, 1308 * any errors when flushing the vdev(s), will 1309 * (unfortunately) not be handled correctly, 1310 * since these "zio_flush" errors will not be 1311 * propagated up to "zil_lwb_flush_vdevs_done". 1312 */ 1313 zio_flush(lwb->lwb_root_zio, vd); 1314 } 1315 kmem_free(zv, sizeof (*zv)); 1316 } 1317 } 1318 1319 static void 1320 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb) 1321 { 1322 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened; 1323 1324 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1325 ASSERT(MUTEX_HELD(&zilog->zl_lock)); 1326 1327 /* 1328 * The zilog's "zl_last_lwb_opened" field is used to build the 1329 * lwb/zio dependency chain, which is used to preserve the 1330 * ordering of lwb completions that is required by the semantics 1331 * of the ZIL. Each new lwb zio becomes a parent of the 1332 * "previous" lwb zio, such that the new lwb's zio cannot 1333 * complete until the "previous" lwb's zio completes. 1334 * 1335 * This is required by the semantics of zil_commit(); the commit 1336 * waiters attached to the lwbs will be woken in the lwb zio's 1337 * completion callback, so this zio dependency graph ensures the 1338 * waiters are woken in the correct order (the same order the 1339 * lwbs were created). 1340 */ 1341 if (last_lwb_opened != NULL && 1342 last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) { 1343 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1344 last_lwb_opened->lwb_state == LWB_STATE_ISSUED || 1345 last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE); 1346 1347 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL); 1348 zio_add_child(lwb->lwb_root_zio, 1349 last_lwb_opened->lwb_root_zio); 1350 1351 /* 1352 * If the previous lwb's write hasn't already completed, 1353 * we also want to order the completion of the lwb write 1354 * zios (above, we only order the completion of the lwb 1355 * root zios). This is required because of how we can 1356 * defer the DKIOCFLUSHWRITECACHE commands for each lwb. 1357 * 1358 * When the DKIOCFLUSHWRITECACHE commands are deferred, 1359 * the previous lwb will rely on this lwb to flush the 1360 * vdevs written to by that previous lwb. Thus, we need 1361 * to ensure this lwb doesn't issue the flush until 1362 * after the previous lwb's write completes. We ensure 1363 * this ordering by setting the zio parent/child 1364 * relationship here. 1365 * 1366 * Without this relationship on the lwb's write zio, 1367 * it's possible for this lwb's write to complete prior 1368 * to the previous lwb's write completing; and thus, the 1369 * vdevs for the previous lwb would be flushed prior to 1370 * that lwb's data being written to those vdevs (the 1371 * vdevs are flushed in the lwb write zio's completion 1372 * handler, zil_lwb_write_done()). 1373 */ 1374 if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) { 1375 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1376 last_lwb_opened->lwb_state == LWB_STATE_ISSUED); 1377 1378 ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL); 1379 zio_add_child(lwb->lwb_write_zio, 1380 last_lwb_opened->lwb_write_zio); 1381 } 1382 } 1383 } 1384 1385 1386 /* 1387 * This function's purpose is to "open" an lwb such that it is ready to 1388 * accept new itxs being committed to it. To do this, the lwb's zio 1389 * structures are created, and linked to the lwb. This function is 1390 * idempotent; if the passed in lwb has already been opened, this 1391 * function is essentially a no-op. 1392 */ 1393 static void 1394 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb) 1395 { 1396 zbookmark_phys_t zb; 1397 zio_priority_t prio; 1398 1399 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1400 ASSERT3P(lwb, !=, NULL); 1401 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED); 1402 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED); 1403 1404 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 1405 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 1406 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 1407 1408 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */ 1409 mutex_enter(&zilog->zl_lock); 1410 if (lwb->lwb_root_zio == NULL) { 1411 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, 1412 BP_GET_LSIZE(&lwb->lwb_blk)); 1413 1414 if (!lwb->lwb_fastwrite) { 1415 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk); 1416 lwb->lwb_fastwrite = 1; 1417 } 1418 1419 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk) 1420 prio = ZIO_PRIORITY_SYNC_WRITE; 1421 else 1422 prio = ZIO_PRIORITY_ASYNC_WRITE; 1423 1424 lwb->lwb_root_zio = zio_root(zilog->zl_spa, 1425 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL); 1426 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1427 1428 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, 1429 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd, 1430 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb, 1431 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb); 1432 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1433 1434 lwb->lwb_state = LWB_STATE_OPENED; 1435 1436 zil_lwb_set_zio_dependency(zilog, lwb); 1437 zilog->zl_last_lwb_opened = lwb; 1438 } 1439 mutex_exit(&zilog->zl_lock); 1440 1441 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1442 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1443 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1444 } 1445 1446 /* 1447 * Define a limited set of intent log block sizes. 1448 * 1449 * These must be a multiple of 4KB. Note only the amount used (again 1450 * aligned to 4KB) actually gets written. However, we can't always just 1451 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 1452 */ 1453 struct { 1454 uint64_t limit; 1455 uint64_t blksz; 1456 } zil_block_buckets[] = { 1457 { 4096, 4096 }, /* non TX_WRITE */ 1458 { 8192 + 4096, 8192 + 4096 }, /* database */ 1459 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */ 1460 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */ 1461 { 131072, 131072 }, /* < 128KB writes */ 1462 { 131072 +4096, 65536 + 4096 }, /* 128KB writes */ 1463 { UINT64_MAX, SPA_OLD_MAXBLOCKSIZE}, /* > 128KB writes */ 1464 }; 1465 1466 /* 1467 * Maximum block size used by the ZIL. This is picked up when the ZIL is 1468 * initialized. Otherwise this should not be used directly; see 1469 * zl_max_block_size instead. 1470 */ 1471 int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE; 1472 1473 /* 1474 * Start a log block write and advance to the next log block. 1475 * Calls are serialized. 1476 */ 1477 static lwb_t * 1478 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb) 1479 { 1480 lwb_t *nlwb = NULL; 1481 zil_chain_t *zilc; 1482 spa_t *spa = zilog->zl_spa; 1483 blkptr_t *bp; 1484 dmu_tx_t *tx; 1485 uint64_t txg; 1486 uint64_t zil_blksz, wsz; 1487 int i, error; 1488 boolean_t slog; 1489 1490 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1491 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1492 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1493 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1494 1495 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1496 zilc = (zil_chain_t *)lwb->lwb_buf; 1497 bp = &zilc->zc_next_blk; 1498 } else { 1499 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 1500 bp = &zilc->zc_next_blk; 1501 } 1502 1503 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 1504 1505 /* 1506 * Allocate the next block and save its address in this block 1507 * before writing it in order to establish the log chain. 1508 * Note that if the allocation of nlwb synced before we wrote 1509 * the block that points at it (lwb), we'd leak it if we crashed. 1510 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 1511 * We dirty the dataset to ensure that zil_sync() will be called 1512 * to clean up in the event of allocation failure or I/O failure. 1513 */ 1514 1515 tx = dmu_tx_create(zilog->zl_os); 1516 1517 /* 1518 * Since we are not going to create any new dirty data, and we 1519 * can even help with clearing the existing dirty data, we 1520 * should not be subject to the dirty data based delays. We 1521 * use TXG_NOTHROTTLE to bypass the delay mechanism. 1522 */ 1523 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); 1524 1525 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1526 txg = dmu_tx_get_txg(tx); 1527 1528 lwb->lwb_tx = tx; 1529 1530 /* 1531 * Log blocks are pre-allocated. Here we select the size of the next 1532 * block, based on size used in the last block. 1533 * - first find the smallest bucket that will fit the block from a 1534 * limited set of block sizes. This is because it's faster to write 1535 * blocks allocated from the same metaslab as they are adjacent or 1536 * close. 1537 * - next find the maximum from the new suggested size and an array of 1538 * previous sizes. This lessens a picket fence effect of wrongly 1539 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k 1540 * requests. 1541 * 1542 * Note we only write what is used, but we can't just allocate 1543 * the maximum block size because we can exhaust the available 1544 * pool log space. 1545 */ 1546 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1547 for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++) 1548 continue; 1549 zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size); 1550 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1551 for (i = 0; i < ZIL_PREV_BLKS; i++) 1552 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1553 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1554 1555 BP_ZERO(bp); 1556 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog); 1557 if (slog) { 1558 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count); 1559 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused); 1560 } else { 1561 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count); 1562 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused); 1563 } 1564 if (error == 0) { 1565 ASSERT3U(bp->blk_birth, ==, txg); 1566 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1567 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1568 1569 /* 1570 * Allocate a new log write block (lwb). 1571 */ 1572 nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE); 1573 } 1574 1575 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1576 /* For Slim ZIL only write what is used. */ 1577 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1578 ASSERT3U(wsz, <=, lwb->lwb_sz); 1579 zio_shrink(lwb->lwb_write_zio, wsz); 1580 1581 } else { 1582 wsz = lwb->lwb_sz; 1583 } 1584 1585 zilc->zc_pad = 0; 1586 zilc->zc_nused = lwb->lwb_nused; 1587 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1588 1589 /* 1590 * clear unused data for security 1591 */ 1592 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1593 1594 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER); 1595 1596 zil_lwb_add_block(lwb, &lwb->lwb_blk); 1597 lwb->lwb_issued_timestamp = gethrtime(); 1598 lwb->lwb_state = LWB_STATE_ISSUED; 1599 1600 zio_nowait(lwb->lwb_root_zio); 1601 zio_nowait(lwb->lwb_write_zio); 1602 1603 /* 1604 * If there was an allocation failure then nlwb will be null which 1605 * forces a txg_wait_synced(). 1606 */ 1607 return (nlwb); 1608 } 1609 1610 /* 1611 * Maximum amount of write data that can be put into single log block. 1612 */ 1613 uint64_t 1614 zil_max_log_data(zilog_t *zilog) 1615 { 1616 return (zilog->zl_max_block_size - 1617 sizeof (zil_chain_t) - sizeof (lr_write_t)); 1618 } 1619 1620 /* 1621 * Maximum amount of log space we agree to waste to reduce number of 1622 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%). 1623 */ 1624 static inline uint64_t 1625 zil_max_waste_space(zilog_t *zilog) 1626 { 1627 return (zil_max_log_data(zilog) / 8); 1628 } 1629 1630 /* 1631 * Maximum amount of write data for WR_COPIED. For correctness, consumers 1632 * must fall back to WR_NEED_COPY if we can't fit the entire record into one 1633 * maximum sized log block, because each WR_COPIED record must fit in a 1634 * single log block. For space efficiency, we want to fit two records into a 1635 * max-sized log block. 1636 */ 1637 uint64_t 1638 zil_max_copied_data(zilog_t *zilog) 1639 { 1640 return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 - 1641 sizeof (lr_write_t)); 1642 } 1643 1644 static lwb_t * 1645 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1646 { 1647 lr_t *lrcb, *lrc; 1648 lr_write_t *lrwb, *lrw; 1649 char *lr_buf; 1650 uint64_t dlen, dnow, dpad, lwb_sp, reclen, txg, max_log_data; 1651 1652 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1653 ASSERT3P(lwb, !=, NULL); 1654 ASSERT3P(lwb->lwb_buf, !=, NULL); 1655 1656 zil_lwb_write_open(zilog, lwb); 1657 1658 lrc = &itx->itx_lr; 1659 lrw = (lr_write_t *)lrc; 1660 1661 /* 1662 * A commit itx doesn't represent any on-disk state; instead 1663 * it's simply used as a place holder on the commit list, and 1664 * provides a mechanism for attaching a "commit waiter" onto the 1665 * correct lwb (such that the waiter can be signalled upon 1666 * completion of that lwb). Thus, we don't process this itx's 1667 * log record if it's a commit itx (these itx's don't have log 1668 * records), and instead link the itx's waiter onto the lwb's 1669 * list of waiters. 1670 * 1671 * For more details, see the comment above zil_commit(). 1672 */ 1673 if (lrc->lrc_txtype == TX_COMMIT) { 1674 mutex_enter(&zilog->zl_lock); 1675 zil_commit_waiter_link_lwb(itx->itx_private, lwb); 1676 itx->itx_private = NULL; 1677 mutex_exit(&zilog->zl_lock); 1678 return (lwb); 1679 } 1680 1681 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { 1682 dlen = P2ROUNDUP_TYPED( 1683 lrw->lr_length, sizeof (uint64_t), uint64_t); 1684 dpad = dlen - lrw->lr_length; 1685 } else { 1686 dlen = dpad = 0; 1687 } 1688 reclen = lrc->lrc_reclen; 1689 zilog->zl_cur_used += (reclen + dlen); 1690 txg = lrc->lrc_txg; 1691 1692 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen)); 1693 1694 cont: 1695 /* 1696 * If this record won't fit in the current log block, start a new one. 1697 * For WR_NEED_COPY optimize layout for minimal number of chunks. 1698 */ 1699 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1700 max_log_data = zil_max_log_data(zilog); 1701 if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1702 lwb_sp < zil_max_waste_space(zilog) && 1703 (dlen % max_log_data == 0 || 1704 lwb_sp < reclen + dlen % max_log_data))) { 1705 lwb = zil_lwb_write_issue(zilog, lwb); 1706 if (lwb == NULL) 1707 return (NULL); 1708 zil_lwb_write_open(zilog, lwb); 1709 ASSERT(LWB_EMPTY(lwb)); 1710 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1711 1712 /* 1713 * There must be enough space in the new, empty log block to 1714 * hold reclen. For WR_COPIED, we need to fit the whole 1715 * record in one block, and reclen is the header size + the 1716 * data size. For WR_NEED_COPY, we can create multiple 1717 * records, splitting the data into multiple blocks, so we 1718 * only need to fit one word of data per block; in this case 1719 * reclen is just the header size (no data). 1720 */ 1721 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp); 1722 } 1723 1724 dnow = MIN(dlen, lwb_sp - reclen); 1725 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1726 bcopy(lrc, lr_buf, reclen); 1727 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */ 1728 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */ 1729 1730 ZIL_STAT_BUMP(zil_itx_count); 1731 1732 /* 1733 * If it's a write, fetch the data or get its blkptr as appropriate. 1734 */ 1735 if (lrc->lrc_txtype == TX_WRITE) { 1736 if (txg > spa_freeze_txg(zilog->zl_spa)) 1737 txg_wait_synced(zilog->zl_dmu_pool, txg); 1738 if (itx->itx_wr_state == WR_COPIED) { 1739 ZIL_STAT_BUMP(zil_itx_copied_count); 1740 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length); 1741 } else { 1742 char *dbuf; 1743 int error; 1744 1745 if (itx->itx_wr_state == WR_NEED_COPY) { 1746 dbuf = lr_buf + reclen; 1747 lrcb->lrc_reclen += dnow; 1748 if (lrwb->lr_length > dnow) 1749 lrwb->lr_length = dnow; 1750 lrw->lr_offset += dnow; 1751 lrw->lr_length -= dnow; 1752 ZIL_STAT_BUMP(zil_itx_needcopy_count); 1753 ZIL_STAT_INCR(zil_itx_needcopy_bytes, dnow); 1754 } else { 1755 ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT); 1756 dbuf = NULL; 1757 ZIL_STAT_BUMP(zil_itx_indirect_count); 1758 ZIL_STAT_INCR(zil_itx_indirect_bytes, 1759 lrw->lr_length); 1760 } 1761 1762 /* 1763 * We pass in the "lwb_write_zio" rather than 1764 * "lwb_root_zio" so that the "lwb_write_zio" 1765 * becomes the parent of any zio's created by 1766 * the "zl_get_data" callback. The vdevs are 1767 * flushed after the "lwb_write_zio" completes, 1768 * so we want to make sure that completion 1769 * callback waits for these additional zio's, 1770 * such that the vdevs used by those zio's will 1771 * be included in the lwb's vdev tree, and those 1772 * vdevs will be properly flushed. If we passed 1773 * in "lwb_root_zio" here, then these additional 1774 * vdevs may not be flushed; e.g. if these zio's 1775 * completed after "lwb_write_zio" completed. 1776 */ 1777 error = zilog->zl_get_data(itx->itx_private, 1778 itx->itx_gen, lrwb, dbuf, lwb, 1779 lwb->lwb_write_zio); 1780 if (dbuf != NULL && error == 0 && dnow == dlen) 1781 /* Zero any padding bytes in the last block. */ 1782 bzero((char *)dbuf + lrwb->lr_length, dpad); 1783 1784 if (error == EIO) { 1785 txg_wait_synced(zilog->zl_dmu_pool, txg); 1786 return (lwb); 1787 } 1788 if (error != 0) { 1789 ASSERT(error == ENOENT || error == EEXIST || 1790 error == EALREADY); 1791 return (lwb); 1792 } 1793 } 1794 } 1795 1796 /* 1797 * We're actually making an entry, so update lrc_seq to be the 1798 * log record sequence number. Note that this is generally not 1799 * equal to the itx sequence number because not all transactions 1800 * are synchronous, and sometimes spa_sync() gets there first. 1801 */ 1802 lrcb->lrc_seq = ++zilog->zl_lr_seq; 1803 lwb->lwb_nused += reclen + dnow; 1804 1805 zil_lwb_add_txg(lwb, txg); 1806 1807 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1808 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1809 1810 dlen -= dnow; 1811 if (dlen > 0) { 1812 zilog->zl_cur_used += reclen; 1813 goto cont; 1814 } 1815 1816 return (lwb); 1817 } 1818 1819 itx_t * 1820 zil_itx_create(uint64_t txtype, size_t olrsize) 1821 { 1822 size_t itxsize, lrsize; 1823 itx_t *itx; 1824 1825 lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t); 1826 itxsize = offsetof(itx_t, itx_lr) + lrsize; 1827 1828 itx = zio_data_buf_alloc(itxsize); 1829 itx->itx_lr.lrc_txtype = txtype; 1830 itx->itx_lr.lrc_reclen = lrsize; 1831 itx->itx_lr.lrc_seq = 0; /* defensive */ 1832 bzero((char *)&itx->itx_lr + olrsize, lrsize - olrsize); 1833 itx->itx_sync = B_TRUE; /* default is synchronous */ 1834 itx->itx_callback = NULL; 1835 itx->itx_callback_data = NULL; 1836 itx->itx_size = itxsize; 1837 1838 return (itx); 1839 } 1840 1841 void 1842 zil_itx_destroy(itx_t *itx) 1843 { 1844 IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL); 1845 IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 1846 1847 if (itx->itx_callback != NULL) 1848 itx->itx_callback(itx->itx_callback_data); 1849 1850 zio_data_buf_free(itx, itx->itx_size); 1851 } 1852 1853 /* 1854 * Free up the sync and async itxs. The itxs_t has already been detached 1855 * so no locks are needed. 1856 */ 1857 static void 1858 zil_itxg_clean(void *arg) 1859 { 1860 itx_t *itx; 1861 list_t *list; 1862 avl_tree_t *t; 1863 void *cookie; 1864 itxs_t *itxs = arg; 1865 itx_async_node_t *ian; 1866 1867 list = &itxs->i_sync_list; 1868 while ((itx = list_head(list)) != NULL) { 1869 /* 1870 * In the general case, commit itxs will not be found 1871 * here, as they'll be committed to an lwb via 1872 * zil_lwb_commit(), and free'd in that function. Having 1873 * said that, it is still possible for commit itxs to be 1874 * found here, due to the following race: 1875 * 1876 * - a thread calls zil_commit() which assigns the 1877 * commit itx to a per-txg i_sync_list 1878 * - zil_itxg_clean() is called (e.g. via spa_sync()) 1879 * while the waiter is still on the i_sync_list 1880 * 1881 * There's nothing to prevent syncing the txg while the 1882 * waiter is on the i_sync_list. This normally doesn't 1883 * happen because spa_sync() is slower than zil_commit(), 1884 * but if zil_commit() calls txg_wait_synced() (e.g. 1885 * because zil_create() or zil_commit_writer_stall() is 1886 * called) we will hit this case. 1887 */ 1888 if (itx->itx_lr.lrc_txtype == TX_COMMIT) 1889 zil_commit_waiter_skip(itx->itx_private); 1890 1891 list_remove(list, itx); 1892 zil_itx_destroy(itx); 1893 } 1894 1895 cookie = NULL; 1896 t = &itxs->i_async_tree; 1897 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1898 list = &ian->ia_list; 1899 while ((itx = list_head(list)) != NULL) { 1900 list_remove(list, itx); 1901 /* commit itxs should never be on the async lists. */ 1902 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 1903 zil_itx_destroy(itx); 1904 } 1905 list_destroy(list); 1906 kmem_free(ian, sizeof (itx_async_node_t)); 1907 } 1908 avl_destroy(t); 1909 1910 kmem_free(itxs, sizeof (itxs_t)); 1911 } 1912 1913 static int 1914 zil_aitx_compare(const void *x1, const void *x2) 1915 { 1916 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1917 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1918 1919 return (TREE_CMP(o1, o2)); 1920 } 1921 1922 /* 1923 * Remove all async itx with the given oid. 1924 */ 1925 void 1926 zil_remove_async(zilog_t *zilog, uint64_t oid) 1927 { 1928 uint64_t otxg, txg; 1929 itx_async_node_t *ian; 1930 avl_tree_t *t; 1931 avl_index_t where; 1932 list_t clean_list; 1933 itx_t *itx; 1934 1935 ASSERT(oid != 0); 1936 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1937 1938 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1939 otxg = ZILTEST_TXG; 1940 else 1941 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1942 1943 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1944 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1945 1946 mutex_enter(&itxg->itxg_lock); 1947 if (itxg->itxg_txg != txg) { 1948 mutex_exit(&itxg->itxg_lock); 1949 continue; 1950 } 1951 1952 /* 1953 * Locate the object node and append its list. 1954 */ 1955 t = &itxg->itxg_itxs->i_async_tree; 1956 ian = avl_find(t, &oid, &where); 1957 if (ian != NULL) 1958 list_move_tail(&clean_list, &ian->ia_list); 1959 mutex_exit(&itxg->itxg_lock); 1960 } 1961 while ((itx = list_head(&clean_list)) != NULL) { 1962 list_remove(&clean_list, itx); 1963 /* commit itxs should never be on the async lists. */ 1964 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 1965 zil_itx_destroy(itx); 1966 } 1967 list_destroy(&clean_list); 1968 } 1969 1970 void 1971 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1972 { 1973 uint64_t txg; 1974 itxg_t *itxg; 1975 itxs_t *itxs, *clean = NULL; 1976 1977 /* 1978 * Ensure the data of a renamed file is committed before the rename. 1979 */ 1980 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1981 zil_async_to_sync(zilog, itx->itx_oid); 1982 1983 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1984 txg = ZILTEST_TXG; 1985 else 1986 txg = dmu_tx_get_txg(tx); 1987 1988 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1989 mutex_enter(&itxg->itxg_lock); 1990 itxs = itxg->itxg_itxs; 1991 if (itxg->itxg_txg != txg) { 1992 if (itxs != NULL) { 1993 /* 1994 * The zil_clean callback hasn't got around to cleaning 1995 * this itxg. Save the itxs for release below. 1996 * This should be rare. 1997 */ 1998 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for " 1999 "txg %llu", (u_longlong_t)itxg->itxg_txg); 2000 clean = itxg->itxg_itxs; 2001 } 2002 itxg->itxg_txg = txg; 2003 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), 2004 KM_SLEEP); 2005 2006 list_create(&itxs->i_sync_list, sizeof (itx_t), 2007 offsetof(itx_t, itx_node)); 2008 avl_create(&itxs->i_async_tree, zil_aitx_compare, 2009 sizeof (itx_async_node_t), 2010 offsetof(itx_async_node_t, ia_node)); 2011 } 2012 if (itx->itx_sync) { 2013 list_insert_tail(&itxs->i_sync_list, itx); 2014 } else { 2015 avl_tree_t *t = &itxs->i_async_tree; 2016 uint64_t foid = 2017 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid); 2018 itx_async_node_t *ian; 2019 avl_index_t where; 2020 2021 ian = avl_find(t, &foid, &where); 2022 if (ian == NULL) { 2023 ian = kmem_alloc(sizeof (itx_async_node_t), 2024 KM_SLEEP); 2025 list_create(&ian->ia_list, sizeof (itx_t), 2026 offsetof(itx_t, itx_node)); 2027 ian->ia_foid = foid; 2028 avl_insert(t, ian, where); 2029 } 2030 list_insert_tail(&ian->ia_list, itx); 2031 } 2032 2033 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 2034 2035 /* 2036 * We don't want to dirty the ZIL using ZILTEST_TXG, because 2037 * zil_clean() will never be called using ZILTEST_TXG. Thus, we 2038 * need to be careful to always dirty the ZIL using the "real" 2039 * TXG (not itxg_txg) even when the SPA is frozen. 2040 */ 2041 zilog_dirty(zilog, dmu_tx_get_txg(tx)); 2042 mutex_exit(&itxg->itxg_lock); 2043 2044 /* Release the old itxs now we've dropped the lock */ 2045 if (clean != NULL) 2046 zil_itxg_clean(clean); 2047 } 2048 2049 /* 2050 * If there are any in-memory intent log transactions which have now been 2051 * synced then start up a taskq to free them. We should only do this after we 2052 * have written out the uberblocks (i.e. txg has been committed) so that 2053 * don't inadvertently clean out in-memory log records that would be required 2054 * by zil_commit(). 2055 */ 2056 void 2057 zil_clean(zilog_t *zilog, uint64_t synced_txg) 2058 { 2059 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 2060 itxs_t *clean_me; 2061 2062 ASSERT3U(synced_txg, <, ZILTEST_TXG); 2063 2064 mutex_enter(&itxg->itxg_lock); 2065 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 2066 mutex_exit(&itxg->itxg_lock); 2067 return; 2068 } 2069 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 2070 ASSERT3U(itxg->itxg_txg, !=, 0); 2071 clean_me = itxg->itxg_itxs; 2072 itxg->itxg_itxs = NULL; 2073 itxg->itxg_txg = 0; 2074 mutex_exit(&itxg->itxg_lock); 2075 /* 2076 * Preferably start a task queue to free up the old itxs but 2077 * if taskq_dispatch can't allocate resources to do that then 2078 * free it in-line. This should be rare. Note, using TQ_SLEEP 2079 * created a bad performance problem. 2080 */ 2081 ASSERT3P(zilog->zl_dmu_pool, !=, NULL); 2082 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL); 2083 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq, 2084 zil_itxg_clean, clean_me, TQ_NOSLEEP); 2085 if (id == TASKQID_INVALID) 2086 zil_itxg_clean(clean_me); 2087 } 2088 2089 /* 2090 * This function will traverse the queue of itxs that need to be 2091 * committed, and move them onto the ZIL's zl_itx_commit_list. 2092 */ 2093 static void 2094 zil_get_commit_list(zilog_t *zilog) 2095 { 2096 uint64_t otxg, txg; 2097 list_t *commit_list = &zilog->zl_itx_commit_list; 2098 2099 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2100 2101 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2102 otxg = ZILTEST_TXG; 2103 else 2104 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2105 2106 /* 2107 * This is inherently racy, since there is nothing to prevent 2108 * the last synced txg from changing. That's okay since we'll 2109 * only commit things in the future. 2110 */ 2111 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2112 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2113 2114 mutex_enter(&itxg->itxg_lock); 2115 if (itxg->itxg_txg != txg) { 2116 mutex_exit(&itxg->itxg_lock); 2117 continue; 2118 } 2119 2120 /* 2121 * If we're adding itx records to the zl_itx_commit_list, 2122 * then the zil better be dirty in this "txg". We can assert 2123 * that here since we're holding the itxg_lock which will 2124 * prevent spa_sync from cleaning it. Once we add the itxs 2125 * to the zl_itx_commit_list we must commit it to disk even 2126 * if it's unnecessary (i.e. the txg was synced). 2127 */ 2128 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 2129 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 2130 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 2131 2132 mutex_exit(&itxg->itxg_lock); 2133 } 2134 } 2135 2136 /* 2137 * Move the async itxs for a specified object to commit into sync lists. 2138 */ 2139 void 2140 zil_async_to_sync(zilog_t *zilog, uint64_t foid) 2141 { 2142 uint64_t otxg, txg; 2143 itx_async_node_t *ian; 2144 avl_tree_t *t; 2145 avl_index_t where; 2146 2147 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2148 otxg = ZILTEST_TXG; 2149 else 2150 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2151 2152 /* 2153 * This is inherently racy, since there is nothing to prevent 2154 * the last synced txg from changing. 2155 */ 2156 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2157 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2158 2159 mutex_enter(&itxg->itxg_lock); 2160 if (itxg->itxg_txg != txg) { 2161 mutex_exit(&itxg->itxg_lock); 2162 continue; 2163 } 2164 2165 /* 2166 * If a foid is specified then find that node and append its 2167 * list. Otherwise walk the tree appending all the lists 2168 * to the sync list. We add to the end rather than the 2169 * beginning to ensure the create has happened. 2170 */ 2171 t = &itxg->itxg_itxs->i_async_tree; 2172 if (foid != 0) { 2173 ian = avl_find(t, &foid, &where); 2174 if (ian != NULL) { 2175 list_move_tail(&itxg->itxg_itxs->i_sync_list, 2176 &ian->ia_list); 2177 } 2178 } else { 2179 void *cookie = NULL; 2180 2181 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 2182 list_move_tail(&itxg->itxg_itxs->i_sync_list, 2183 &ian->ia_list); 2184 list_destroy(&ian->ia_list); 2185 kmem_free(ian, sizeof (itx_async_node_t)); 2186 } 2187 } 2188 mutex_exit(&itxg->itxg_lock); 2189 } 2190 } 2191 2192 /* 2193 * This function will prune commit itxs that are at the head of the 2194 * commit list (it won't prune past the first non-commit itx), and 2195 * either: a) attach them to the last lwb that's still pending 2196 * completion, or b) skip them altogether. 2197 * 2198 * This is used as a performance optimization to prevent commit itxs 2199 * from generating new lwbs when it's unnecessary to do so. 2200 */ 2201 static void 2202 zil_prune_commit_list(zilog_t *zilog) 2203 { 2204 itx_t *itx; 2205 2206 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2207 2208 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { 2209 lr_t *lrc = &itx->itx_lr; 2210 if (lrc->lrc_txtype != TX_COMMIT) 2211 break; 2212 2213 mutex_enter(&zilog->zl_lock); 2214 2215 lwb_t *last_lwb = zilog->zl_last_lwb_opened; 2216 if (last_lwb == NULL || 2217 last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) { 2218 /* 2219 * All of the itxs this waiter was waiting on 2220 * must have already completed (or there were 2221 * never any itx's for it to wait on), so it's 2222 * safe to skip this waiter and mark it done. 2223 */ 2224 zil_commit_waiter_skip(itx->itx_private); 2225 } else { 2226 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb); 2227 itx->itx_private = NULL; 2228 } 2229 2230 mutex_exit(&zilog->zl_lock); 2231 2232 list_remove(&zilog->zl_itx_commit_list, itx); 2233 zil_itx_destroy(itx); 2234 } 2235 2236 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 2237 } 2238 2239 static void 2240 zil_commit_writer_stall(zilog_t *zilog) 2241 { 2242 /* 2243 * When zio_alloc_zil() fails to allocate the next lwb block on 2244 * disk, we must call txg_wait_synced() to ensure all of the 2245 * lwbs in the zilog's zl_lwb_list are synced and then freed (in 2246 * zil_sync()), such that any subsequent ZIL writer (i.e. a call 2247 * to zil_process_commit_list()) will have to call zil_create(), 2248 * and start a new ZIL chain. 2249 * 2250 * Since zil_alloc_zil() failed, the lwb that was previously 2251 * issued does not have a pointer to the "next" lwb on disk. 2252 * Thus, if another ZIL writer thread was to allocate the "next" 2253 * on-disk lwb, that block could be leaked in the event of a 2254 * crash (because the previous lwb on-disk would not point to 2255 * it). 2256 * 2257 * We must hold the zilog's zl_issuer_lock while we do this, to 2258 * ensure no new threads enter zil_process_commit_list() until 2259 * all lwb's in the zl_lwb_list have been synced and freed 2260 * (which is achieved via the txg_wait_synced() call). 2261 */ 2262 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2263 txg_wait_synced(zilog->zl_dmu_pool, 0); 2264 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 2265 } 2266 2267 /* 2268 * This function will traverse the commit list, creating new lwbs as 2269 * needed, and committing the itxs from the commit list to these newly 2270 * created lwbs. Additionally, as a new lwb is created, the previous 2271 * lwb will be issued to the zio layer to be written to disk. 2272 */ 2273 static void 2274 zil_process_commit_list(zilog_t *zilog) 2275 { 2276 spa_t *spa = zilog->zl_spa; 2277 list_t nolwb_itxs; 2278 list_t nolwb_waiters; 2279 lwb_t *lwb; 2280 itx_t *itx; 2281 2282 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2283 2284 /* 2285 * Return if there's nothing to commit before we dirty the fs by 2286 * calling zil_create(). 2287 */ 2288 if (list_head(&zilog->zl_itx_commit_list) == NULL) 2289 return; 2290 2291 list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); 2292 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t), 2293 offsetof(zil_commit_waiter_t, zcw_node)); 2294 2295 lwb = list_tail(&zilog->zl_lwb_list); 2296 if (lwb == NULL) { 2297 lwb = zil_create(zilog); 2298 } else { 2299 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2300 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 2301 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 2302 } 2303 2304 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { 2305 lr_t *lrc = &itx->itx_lr; 2306 uint64_t txg = lrc->lrc_txg; 2307 2308 ASSERT3U(txg, !=, 0); 2309 2310 if (lrc->lrc_txtype == TX_COMMIT) { 2311 DTRACE_PROBE2(zil__process__commit__itx, 2312 zilog_t *, zilog, itx_t *, itx); 2313 } else { 2314 DTRACE_PROBE2(zil__process__normal__itx, 2315 zilog_t *, zilog, itx_t *, itx); 2316 } 2317 2318 list_remove(&zilog->zl_itx_commit_list, itx); 2319 2320 boolean_t synced = txg <= spa_last_synced_txg(spa); 2321 boolean_t frozen = txg > spa_freeze_txg(spa); 2322 2323 /* 2324 * If the txg of this itx has already been synced out, then 2325 * we don't need to commit this itx to an lwb. This is 2326 * because the data of this itx will have already been 2327 * written to the main pool. This is inherently racy, and 2328 * it's still ok to commit an itx whose txg has already 2329 * been synced; this will result in a write that's 2330 * unnecessary, but will do no harm. 2331 * 2332 * With that said, we always want to commit TX_COMMIT itxs 2333 * to an lwb, regardless of whether or not that itx's txg 2334 * has been synced out. We do this to ensure any OPENED lwb 2335 * will always have at least one zil_commit_waiter_t linked 2336 * to the lwb. 2337 * 2338 * As a counter-example, if we skipped TX_COMMIT itx's 2339 * whose txg had already been synced, the following 2340 * situation could occur if we happened to be racing with 2341 * spa_sync: 2342 * 2343 * 1. We commit a non-TX_COMMIT itx to an lwb, where the 2344 * itx's txg is 10 and the last synced txg is 9. 2345 * 2. spa_sync finishes syncing out txg 10. 2346 * 3. We move to the next itx in the list, it's a TX_COMMIT 2347 * whose txg is 10, so we skip it rather than committing 2348 * it to the lwb used in (1). 2349 * 2350 * If the itx that is skipped in (3) is the last TX_COMMIT 2351 * itx in the commit list, than it's possible for the lwb 2352 * used in (1) to remain in the OPENED state indefinitely. 2353 * 2354 * To prevent the above scenario from occurring, ensuring 2355 * that once an lwb is OPENED it will transition to ISSUED 2356 * and eventually DONE, we always commit TX_COMMIT itx's to 2357 * an lwb here, even if that itx's txg has already been 2358 * synced. 2359 * 2360 * Finally, if the pool is frozen, we _always_ commit the 2361 * itx. The point of freezing the pool is to prevent data 2362 * from being written to the main pool via spa_sync, and 2363 * instead rely solely on the ZIL to persistently store the 2364 * data; i.e. when the pool is frozen, the last synced txg 2365 * value can't be trusted. 2366 */ 2367 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) { 2368 if (lwb != NULL) { 2369 lwb = zil_lwb_commit(zilog, itx, lwb); 2370 2371 if (lwb == NULL) 2372 list_insert_tail(&nolwb_itxs, itx); 2373 else 2374 list_insert_tail(&lwb->lwb_itxs, itx); 2375 } else { 2376 if (lrc->lrc_txtype == TX_COMMIT) { 2377 zil_commit_waiter_link_nolwb( 2378 itx->itx_private, &nolwb_waiters); 2379 } 2380 2381 list_insert_tail(&nolwb_itxs, itx); 2382 } 2383 } else { 2384 ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT); 2385 zil_itx_destroy(itx); 2386 } 2387 } 2388 2389 if (lwb == NULL) { 2390 /* 2391 * This indicates zio_alloc_zil() failed to allocate the 2392 * "next" lwb on-disk. When this happens, we must stall 2393 * the ZIL write pipeline; see the comment within 2394 * zil_commit_writer_stall() for more details. 2395 */ 2396 zil_commit_writer_stall(zilog); 2397 2398 /* 2399 * Additionally, we have to signal and mark the "nolwb" 2400 * waiters as "done" here, since without an lwb, we 2401 * can't do this via zil_lwb_flush_vdevs_done() like 2402 * normal. 2403 */ 2404 zil_commit_waiter_t *zcw; 2405 while ((zcw = list_head(&nolwb_waiters)) != NULL) { 2406 zil_commit_waiter_skip(zcw); 2407 list_remove(&nolwb_waiters, zcw); 2408 } 2409 2410 /* 2411 * And finally, we have to destroy the itx's that 2412 * couldn't be committed to an lwb; this will also call 2413 * the itx's callback if one exists for the itx. 2414 */ 2415 while ((itx = list_head(&nolwb_itxs)) != NULL) { 2416 list_remove(&nolwb_itxs, itx); 2417 zil_itx_destroy(itx); 2418 } 2419 } else { 2420 ASSERT(list_is_empty(&nolwb_waiters)); 2421 ASSERT3P(lwb, !=, NULL); 2422 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2423 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 2424 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 2425 2426 /* 2427 * At this point, the ZIL block pointed at by the "lwb" 2428 * variable is in one of the following states: "closed" 2429 * or "open". 2430 * 2431 * If it's "closed", then no itxs have been committed to 2432 * it, so there's no point in issuing its zio (i.e. it's 2433 * "empty"). 2434 * 2435 * If it's "open", then it contains one or more itxs that 2436 * eventually need to be committed to stable storage. In 2437 * this case we intentionally do not issue the lwb's zio 2438 * to disk yet, and instead rely on one of the following 2439 * two mechanisms for issuing the zio: 2440 * 2441 * 1. Ideally, there will be more ZIL activity occurring 2442 * on the system, such that this function will be 2443 * immediately called again (not necessarily by the same 2444 * thread) and this lwb's zio will be issued via 2445 * zil_lwb_commit(). This way, the lwb is guaranteed to 2446 * be "full" when it is issued to disk, and we'll make 2447 * use of the lwb's size the best we can. 2448 * 2449 * 2. If there isn't sufficient ZIL activity occurring on 2450 * the system, such that this lwb's zio isn't issued via 2451 * zil_lwb_commit(), zil_commit_waiter() will issue the 2452 * lwb's zio. If this occurs, the lwb is not guaranteed 2453 * to be "full" by the time its zio is issued, and means 2454 * the size of the lwb was "too large" given the amount 2455 * of ZIL activity occurring on the system at that time. 2456 * 2457 * We do this for a couple of reasons: 2458 * 2459 * 1. To try and reduce the number of IOPs needed to 2460 * write the same number of itxs. If an lwb has space 2461 * available in its buffer for more itxs, and more itxs 2462 * will be committed relatively soon (relative to the 2463 * latency of performing a write), then it's beneficial 2464 * to wait for these "next" itxs. This way, more itxs 2465 * can be committed to stable storage with fewer writes. 2466 * 2467 * 2. To try and use the largest lwb block size that the 2468 * incoming rate of itxs can support. Again, this is to 2469 * try and pack as many itxs into as few lwbs as 2470 * possible, without significantly impacting the latency 2471 * of each individual itx. 2472 */ 2473 } 2474 } 2475 2476 /* 2477 * This function is responsible for ensuring the passed in commit waiter 2478 * (and associated commit itx) is committed to an lwb. If the waiter is 2479 * not already committed to an lwb, all itxs in the zilog's queue of 2480 * itxs will be processed. The assumption is the passed in waiter's 2481 * commit itx will found in the queue just like the other non-commit 2482 * itxs, such that when the entire queue is processed, the waiter will 2483 * have been committed to an lwb. 2484 * 2485 * The lwb associated with the passed in waiter is not guaranteed to 2486 * have been issued by the time this function completes. If the lwb is 2487 * not issued, we rely on future calls to zil_commit_writer() to issue 2488 * the lwb, or the timeout mechanism found in zil_commit_waiter(). 2489 */ 2490 static void 2491 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw) 2492 { 2493 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2494 ASSERT(spa_writeable(zilog->zl_spa)); 2495 2496 mutex_enter(&zilog->zl_issuer_lock); 2497 2498 if (zcw->zcw_lwb != NULL || zcw->zcw_done) { 2499 /* 2500 * It's possible that, while we were waiting to acquire 2501 * the "zl_issuer_lock", another thread committed this 2502 * waiter to an lwb. If that occurs, we bail out early, 2503 * without processing any of the zilog's queue of itxs. 2504 * 2505 * On certain workloads and system configurations, the 2506 * "zl_issuer_lock" can become highly contended. In an 2507 * attempt to reduce this contention, we immediately drop 2508 * the lock if the waiter has already been processed. 2509 * 2510 * We've measured this optimization to reduce CPU spent 2511 * contending on this lock by up to 5%, using a system 2512 * with 32 CPUs, low latency storage (~50 usec writes), 2513 * and 1024 threads performing sync writes. 2514 */ 2515 goto out; 2516 } 2517 2518 ZIL_STAT_BUMP(zil_commit_writer_count); 2519 2520 zil_get_commit_list(zilog); 2521 zil_prune_commit_list(zilog); 2522 zil_process_commit_list(zilog); 2523 2524 out: 2525 mutex_exit(&zilog->zl_issuer_lock); 2526 } 2527 2528 static void 2529 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw) 2530 { 2531 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2532 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2533 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 2534 2535 lwb_t *lwb = zcw->zcw_lwb; 2536 ASSERT3P(lwb, !=, NULL); 2537 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED); 2538 2539 /* 2540 * If the lwb has already been issued by another thread, we can 2541 * immediately return since there's no work to be done (the 2542 * point of this function is to issue the lwb). Additionally, we 2543 * do this prior to acquiring the zl_issuer_lock, to avoid 2544 * acquiring it when it's not necessary to do so. 2545 */ 2546 if (lwb->lwb_state == LWB_STATE_ISSUED || 2547 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2548 lwb->lwb_state == LWB_STATE_FLUSH_DONE) 2549 return; 2550 2551 /* 2552 * In order to call zil_lwb_write_issue() we must hold the 2553 * zilog's "zl_issuer_lock". We can't simply acquire that lock, 2554 * since we're already holding the commit waiter's "zcw_lock", 2555 * and those two locks are acquired in the opposite order 2556 * elsewhere. 2557 */ 2558 mutex_exit(&zcw->zcw_lock); 2559 mutex_enter(&zilog->zl_issuer_lock); 2560 mutex_enter(&zcw->zcw_lock); 2561 2562 /* 2563 * Since we just dropped and re-acquired the commit waiter's 2564 * lock, we have to re-check to see if the waiter was marked 2565 * "done" during that process. If the waiter was marked "done", 2566 * the "lwb" pointer is no longer valid (it can be free'd after 2567 * the waiter is marked "done"), so without this check we could 2568 * wind up with a use-after-free error below. 2569 */ 2570 if (zcw->zcw_done) 2571 goto out; 2572 2573 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2574 2575 /* 2576 * We've already checked this above, but since we hadn't acquired 2577 * the zilog's zl_issuer_lock, we have to perform this check a 2578 * second time while holding the lock. 2579 * 2580 * We don't need to hold the zl_lock since the lwb cannot transition 2581 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb 2582 * _can_ transition from ISSUED to DONE, but it's OK to race with 2583 * that transition since we treat the lwb the same, whether it's in 2584 * the ISSUED or DONE states. 2585 * 2586 * The important thing, is we treat the lwb differently depending on 2587 * if it's ISSUED or OPENED, and block any other threads that might 2588 * attempt to issue this lwb. For that reason we hold the 2589 * zl_issuer_lock when checking the lwb_state; we must not call 2590 * zil_lwb_write_issue() if the lwb had already been issued. 2591 * 2592 * See the comment above the lwb_state_t structure definition for 2593 * more details on the lwb states, and locking requirements. 2594 */ 2595 if (lwb->lwb_state == LWB_STATE_ISSUED || 2596 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2597 lwb->lwb_state == LWB_STATE_FLUSH_DONE) 2598 goto out; 2599 2600 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 2601 2602 /* 2603 * As described in the comments above zil_commit_waiter() and 2604 * zil_process_commit_list(), we need to issue this lwb's zio 2605 * since we've reached the commit waiter's timeout and it still 2606 * hasn't been issued. 2607 */ 2608 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb); 2609 2610 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED); 2611 2612 /* 2613 * Since the lwb's zio hadn't been issued by the time this thread 2614 * reached its timeout, we reset the zilog's "zl_cur_used" field 2615 * to influence the zil block size selection algorithm. 2616 * 2617 * By having to issue the lwb's zio here, it means the size of the 2618 * lwb was too large, given the incoming throughput of itxs. By 2619 * setting "zl_cur_used" to zero, we communicate this fact to the 2620 * block size selection algorithm, so it can take this information 2621 * into account, and potentially select a smaller size for the 2622 * next lwb block that is allocated. 2623 */ 2624 zilog->zl_cur_used = 0; 2625 2626 if (nlwb == NULL) { 2627 /* 2628 * When zil_lwb_write_issue() returns NULL, this 2629 * indicates zio_alloc_zil() failed to allocate the 2630 * "next" lwb on-disk. When this occurs, the ZIL write 2631 * pipeline must be stalled; see the comment within the 2632 * zil_commit_writer_stall() function for more details. 2633 * 2634 * We must drop the commit waiter's lock prior to 2635 * calling zil_commit_writer_stall() or else we can wind 2636 * up with the following deadlock: 2637 * 2638 * - This thread is waiting for the txg to sync while 2639 * holding the waiter's lock; txg_wait_synced() is 2640 * used within txg_commit_writer_stall(). 2641 * 2642 * - The txg can't sync because it is waiting for this 2643 * lwb's zio callback to call dmu_tx_commit(). 2644 * 2645 * - The lwb's zio callback can't call dmu_tx_commit() 2646 * because it's blocked trying to acquire the waiter's 2647 * lock, which occurs prior to calling dmu_tx_commit() 2648 */ 2649 mutex_exit(&zcw->zcw_lock); 2650 zil_commit_writer_stall(zilog); 2651 mutex_enter(&zcw->zcw_lock); 2652 } 2653 2654 out: 2655 mutex_exit(&zilog->zl_issuer_lock); 2656 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2657 } 2658 2659 /* 2660 * This function is responsible for performing the following two tasks: 2661 * 2662 * 1. its primary responsibility is to block until the given "commit 2663 * waiter" is considered "done". 2664 * 2665 * 2. its secondary responsibility is to issue the zio for the lwb that 2666 * the given "commit waiter" is waiting on, if this function has 2667 * waited "long enough" and the lwb is still in the "open" state. 2668 * 2669 * Given a sufficient amount of itxs being generated and written using 2670 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit() 2671 * function. If this does not occur, this secondary responsibility will 2672 * ensure the lwb is issued even if there is not other synchronous 2673 * activity on the system. 2674 * 2675 * For more details, see zil_process_commit_list(); more specifically, 2676 * the comment at the bottom of that function. 2677 */ 2678 static void 2679 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw) 2680 { 2681 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2682 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2683 ASSERT(spa_writeable(zilog->zl_spa)); 2684 2685 mutex_enter(&zcw->zcw_lock); 2686 2687 /* 2688 * The timeout is scaled based on the lwb latency to avoid 2689 * significantly impacting the latency of each individual itx. 2690 * For more details, see the comment at the bottom of the 2691 * zil_process_commit_list() function. 2692 */ 2693 int pct = MAX(zfs_commit_timeout_pct, 1); 2694 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100; 2695 hrtime_t wakeup = gethrtime() + sleep; 2696 boolean_t timedout = B_FALSE; 2697 2698 while (!zcw->zcw_done) { 2699 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2700 2701 lwb_t *lwb = zcw->zcw_lwb; 2702 2703 /* 2704 * Usually, the waiter will have a non-NULL lwb field here, 2705 * but it's possible for it to be NULL as a result of 2706 * zil_commit() racing with spa_sync(). 2707 * 2708 * When zil_clean() is called, it's possible for the itxg 2709 * list (which may be cleaned via a taskq) to contain 2710 * commit itxs. When this occurs, the commit waiters linked 2711 * off of these commit itxs will not be committed to an 2712 * lwb. Additionally, these commit waiters will not be 2713 * marked done until zil_commit_waiter_skip() is called via 2714 * zil_itxg_clean(). 2715 * 2716 * Thus, it's possible for this commit waiter (i.e. the 2717 * "zcw" variable) to be found in this "in between" state; 2718 * where it's "zcw_lwb" field is NULL, and it hasn't yet 2719 * been skipped, so it's "zcw_done" field is still B_FALSE. 2720 */ 2721 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED); 2722 2723 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) { 2724 ASSERT3B(timedout, ==, B_FALSE); 2725 2726 /* 2727 * If the lwb hasn't been issued yet, then we 2728 * need to wait with a timeout, in case this 2729 * function needs to issue the lwb after the 2730 * timeout is reached; responsibility (2) from 2731 * the comment above this function. 2732 */ 2733 int rc = cv_timedwait_hires(&zcw->zcw_cv, 2734 &zcw->zcw_lock, wakeup, USEC2NSEC(1), 2735 CALLOUT_FLAG_ABSOLUTE); 2736 2737 if (rc != -1 || zcw->zcw_done) 2738 continue; 2739 2740 timedout = B_TRUE; 2741 zil_commit_waiter_timeout(zilog, zcw); 2742 2743 if (!zcw->zcw_done) { 2744 /* 2745 * If the commit waiter has already been 2746 * marked "done", it's possible for the 2747 * waiter's lwb structure to have already 2748 * been freed. Thus, we can only reliably 2749 * make these assertions if the waiter 2750 * isn't done. 2751 */ 2752 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2753 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED); 2754 } 2755 } else { 2756 /* 2757 * If the lwb isn't open, then it must have already 2758 * been issued. In that case, there's no need to 2759 * use a timeout when waiting for the lwb to 2760 * complete. 2761 * 2762 * Additionally, if the lwb is NULL, the waiter 2763 * will soon be signaled and marked done via 2764 * zil_clean() and zil_itxg_clean(), so no timeout 2765 * is required. 2766 */ 2767 2768 IMPLY(lwb != NULL, 2769 lwb->lwb_state == LWB_STATE_ISSUED || 2770 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2771 lwb->lwb_state == LWB_STATE_FLUSH_DONE); 2772 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock); 2773 } 2774 } 2775 2776 mutex_exit(&zcw->zcw_lock); 2777 } 2778 2779 static zil_commit_waiter_t * 2780 zil_alloc_commit_waiter(void) 2781 { 2782 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP); 2783 2784 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL); 2785 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL); 2786 list_link_init(&zcw->zcw_node); 2787 zcw->zcw_lwb = NULL; 2788 zcw->zcw_done = B_FALSE; 2789 zcw->zcw_zio_error = 0; 2790 2791 return (zcw); 2792 } 2793 2794 static void 2795 zil_free_commit_waiter(zil_commit_waiter_t *zcw) 2796 { 2797 ASSERT(!list_link_active(&zcw->zcw_node)); 2798 ASSERT3P(zcw->zcw_lwb, ==, NULL); 2799 ASSERT3B(zcw->zcw_done, ==, B_TRUE); 2800 mutex_destroy(&zcw->zcw_lock); 2801 cv_destroy(&zcw->zcw_cv); 2802 kmem_cache_free(zil_zcw_cache, zcw); 2803 } 2804 2805 /* 2806 * This function is used to create a TX_COMMIT itx and assign it. This 2807 * way, it will be linked into the ZIL's list of synchronous itxs, and 2808 * then later committed to an lwb (or skipped) when 2809 * zil_process_commit_list() is called. 2810 */ 2811 static void 2812 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw) 2813 { 2814 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 2815 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 2816 2817 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t)); 2818 itx->itx_sync = B_TRUE; 2819 itx->itx_private = zcw; 2820 2821 zil_itx_assign(zilog, itx, tx); 2822 2823 dmu_tx_commit(tx); 2824 } 2825 2826 /* 2827 * Commit ZFS Intent Log transactions (itxs) to stable storage. 2828 * 2829 * When writing ZIL transactions to the on-disk representation of the 2830 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple 2831 * itxs can be committed to a single lwb. Once a lwb is written and 2832 * committed to stable storage (i.e. the lwb is written, and vdevs have 2833 * been flushed), each itx that was committed to that lwb is also 2834 * considered to be committed to stable storage. 2835 * 2836 * When an itx is committed to an lwb, the log record (lr_t) contained 2837 * by the itx is copied into the lwb's zio buffer, and once this buffer 2838 * is written to disk, it becomes an on-disk ZIL block. 2839 * 2840 * As itxs are generated, they're inserted into the ZIL's queue of 2841 * uncommitted itxs. The semantics of zil_commit() are such that it will 2842 * block until all itxs that were in the queue when it was called, are 2843 * committed to stable storage. 2844 * 2845 * If "foid" is zero, this means all "synchronous" and "asynchronous" 2846 * itxs, for all objects in the dataset, will be committed to stable 2847 * storage prior to zil_commit() returning. If "foid" is non-zero, all 2848 * "synchronous" itxs for all objects, but only "asynchronous" itxs 2849 * that correspond to the foid passed in, will be committed to stable 2850 * storage prior to zil_commit() returning. 2851 * 2852 * Generally speaking, when zil_commit() is called, the consumer doesn't 2853 * actually care about _all_ of the uncommitted itxs. Instead, they're 2854 * simply trying to waiting for a specific itx to be committed to disk, 2855 * but the interface(s) for interacting with the ZIL don't allow such 2856 * fine-grained communication. A better interface would allow a consumer 2857 * to create and assign an itx, and then pass a reference to this itx to 2858 * zil_commit(); such that zil_commit() would return as soon as that 2859 * specific itx was committed to disk (instead of waiting for _all_ 2860 * itxs to be committed). 2861 * 2862 * When a thread calls zil_commit() a special "commit itx" will be 2863 * generated, along with a corresponding "waiter" for this commit itx. 2864 * zil_commit() will wait on this waiter's CV, such that when the waiter 2865 * is marked done, and signaled, zil_commit() will return. 2866 * 2867 * This commit itx is inserted into the queue of uncommitted itxs. This 2868 * provides an easy mechanism for determining which itxs were in the 2869 * queue prior to zil_commit() having been called, and which itxs were 2870 * added after zil_commit() was called. 2871 * 2872 * The commit it is special; it doesn't have any on-disk representation. 2873 * When a commit itx is "committed" to an lwb, the waiter associated 2874 * with it is linked onto the lwb's list of waiters. Then, when that lwb 2875 * completes, each waiter on the lwb's list is marked done and signaled 2876 * -- allowing the thread waiting on the waiter to return from zil_commit(). 2877 * 2878 * It's important to point out a few critical factors that allow us 2879 * to make use of the commit itxs, commit waiters, per-lwb lists of 2880 * commit waiters, and zio completion callbacks like we're doing: 2881 * 2882 * 1. The list of waiters for each lwb is traversed, and each commit 2883 * waiter is marked "done" and signaled, in the zio completion 2884 * callback of the lwb's zio[*]. 2885 * 2886 * * Actually, the waiters are signaled in the zio completion 2887 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands 2888 * that are sent to the vdevs upon completion of the lwb zio. 2889 * 2890 * 2. When the itxs are inserted into the ZIL's queue of uncommitted 2891 * itxs, the order in which they are inserted is preserved[*]; as 2892 * itxs are added to the queue, they are added to the tail of 2893 * in-memory linked lists. 2894 * 2895 * When committing the itxs to lwbs (to be written to disk), they 2896 * are committed in the same order in which the itxs were added to 2897 * the uncommitted queue's linked list(s); i.e. the linked list of 2898 * itxs to commit is traversed from head to tail, and each itx is 2899 * committed to an lwb in that order. 2900 * 2901 * * To clarify: 2902 * 2903 * - the order of "sync" itxs is preserved w.r.t. other 2904 * "sync" itxs, regardless of the corresponding objects. 2905 * - the order of "async" itxs is preserved w.r.t. other 2906 * "async" itxs corresponding to the same object. 2907 * - the order of "async" itxs is *not* preserved w.r.t. other 2908 * "async" itxs corresponding to different objects. 2909 * - the order of "sync" itxs w.r.t. "async" itxs (or vice 2910 * versa) is *not* preserved, even for itxs that correspond 2911 * to the same object. 2912 * 2913 * For more details, see: zil_itx_assign(), zil_async_to_sync(), 2914 * zil_get_commit_list(), and zil_process_commit_list(). 2915 * 2916 * 3. The lwbs represent a linked list of blocks on disk. Thus, any 2917 * lwb cannot be considered committed to stable storage, until its 2918 * "previous" lwb is also committed to stable storage. This fact, 2919 * coupled with the fact described above, means that itxs are 2920 * committed in (roughly) the order in which they were generated. 2921 * This is essential because itxs are dependent on prior itxs. 2922 * Thus, we *must not* deem an itx as being committed to stable 2923 * storage, until *all* prior itxs have also been committed to 2924 * stable storage. 2925 * 2926 * To enforce this ordering of lwb zio's, while still leveraging as 2927 * much of the underlying storage performance as possible, we rely 2928 * on two fundamental concepts: 2929 * 2930 * 1. The creation and issuance of lwb zio's is protected by 2931 * the zilog's "zl_issuer_lock", which ensures only a single 2932 * thread is creating and/or issuing lwb's at a time 2933 * 2. The "previous" lwb is a child of the "current" lwb 2934 * (leveraging the zio parent-child dependency graph) 2935 * 2936 * By relying on this parent-child zio relationship, we can have 2937 * many lwb zio's concurrently issued to the underlying storage, 2938 * but the order in which they complete will be the same order in 2939 * which they were created. 2940 */ 2941 void 2942 zil_commit(zilog_t *zilog, uint64_t foid) 2943 { 2944 /* 2945 * We should never attempt to call zil_commit on a snapshot for 2946 * a couple of reasons: 2947 * 2948 * 1. A snapshot may never be modified, thus it cannot have any 2949 * in-flight itxs that would have modified the dataset. 2950 * 2951 * 2. By design, when zil_commit() is called, a commit itx will 2952 * be assigned to this zilog; as a result, the zilog will be 2953 * dirtied. We must not dirty the zilog of a snapshot; there's 2954 * checks in the code that enforce this invariant, and will 2955 * cause a panic if it's not upheld. 2956 */ 2957 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE); 2958 2959 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2960 return; 2961 2962 if (!spa_writeable(zilog->zl_spa)) { 2963 /* 2964 * If the SPA is not writable, there should never be any 2965 * pending itxs waiting to be committed to disk. If that 2966 * weren't true, we'd skip writing those itxs out, and 2967 * would break the semantics of zil_commit(); thus, we're 2968 * verifying that truth before we return to the caller. 2969 */ 2970 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 2971 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 2972 for (int i = 0; i < TXG_SIZE; i++) 2973 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL); 2974 return; 2975 } 2976 2977 /* 2978 * If the ZIL is suspended, we don't want to dirty it by calling 2979 * zil_commit_itx_assign() below, nor can we write out 2980 * lwbs like would be done in zil_commit_write(). Thus, we 2981 * simply rely on txg_wait_synced() to maintain the necessary 2982 * semantics, and avoid calling those functions altogether. 2983 */ 2984 if (zilog->zl_suspend > 0) { 2985 txg_wait_synced(zilog->zl_dmu_pool, 0); 2986 return; 2987 } 2988 2989 zil_commit_impl(zilog, foid); 2990 } 2991 2992 void 2993 zil_commit_impl(zilog_t *zilog, uint64_t foid) 2994 { 2995 ZIL_STAT_BUMP(zil_commit_count); 2996 2997 /* 2998 * Move the "async" itxs for the specified foid to the "sync" 2999 * queues, such that they will be later committed (or skipped) 3000 * to an lwb when zil_process_commit_list() is called. 3001 * 3002 * Since these "async" itxs must be committed prior to this 3003 * call to zil_commit returning, we must perform this operation 3004 * before we call zil_commit_itx_assign(). 3005 */ 3006 zil_async_to_sync(zilog, foid); 3007 3008 /* 3009 * We allocate a new "waiter" structure which will initially be 3010 * linked to the commit itx using the itx's "itx_private" field. 3011 * Since the commit itx doesn't represent any on-disk state, 3012 * when it's committed to an lwb, rather than copying the its 3013 * lr_t into the lwb's buffer, the commit itx's "waiter" will be 3014 * added to the lwb's list of waiters. Then, when the lwb is 3015 * committed to stable storage, each waiter in the lwb's list of 3016 * waiters will be marked "done", and signalled. 3017 * 3018 * We must create the waiter and assign the commit itx prior to 3019 * calling zil_commit_writer(), or else our specific commit itx 3020 * is not guaranteed to be committed to an lwb prior to calling 3021 * zil_commit_waiter(). 3022 */ 3023 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter(); 3024 zil_commit_itx_assign(zilog, zcw); 3025 3026 zil_commit_writer(zilog, zcw); 3027 zil_commit_waiter(zilog, zcw); 3028 3029 if (zcw->zcw_zio_error != 0) { 3030 /* 3031 * If there was an error writing out the ZIL blocks that 3032 * this thread is waiting on, then we fallback to 3033 * relying on spa_sync() to write out the data this 3034 * thread is waiting on. Obviously this has performance 3035 * implications, but the expectation is for this to be 3036 * an exceptional case, and shouldn't occur often. 3037 */ 3038 DTRACE_PROBE2(zil__commit__io__error, 3039 zilog_t *, zilog, zil_commit_waiter_t *, zcw); 3040 txg_wait_synced(zilog->zl_dmu_pool, 0); 3041 } 3042 3043 zil_free_commit_waiter(zcw); 3044 } 3045 3046 /* 3047 * Called in syncing context to free committed log blocks and update log header. 3048 */ 3049 void 3050 zil_sync(zilog_t *zilog, dmu_tx_t *tx) 3051 { 3052 zil_header_t *zh = zil_header_in_syncing_context(zilog); 3053 uint64_t txg = dmu_tx_get_txg(tx); 3054 spa_t *spa = zilog->zl_spa; 3055 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 3056 lwb_t *lwb; 3057 3058 /* 3059 * We don't zero out zl_destroy_txg, so make sure we don't try 3060 * to destroy it twice. 3061 */ 3062 if (spa_sync_pass(spa) != 1) 3063 return; 3064 3065 mutex_enter(&zilog->zl_lock); 3066 3067 ASSERT(zilog->zl_stop_sync == 0); 3068 3069 if (*replayed_seq != 0) { 3070 ASSERT(zh->zh_replay_seq < *replayed_seq); 3071 zh->zh_replay_seq = *replayed_seq; 3072 *replayed_seq = 0; 3073 } 3074 3075 if (zilog->zl_destroy_txg == txg) { 3076 blkptr_t blk = zh->zh_log; 3077 3078 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 3079 3080 bzero(zh, sizeof (zil_header_t)); 3081 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 3082 3083 if (zilog->zl_keep_first) { 3084 /* 3085 * If this block was part of log chain that couldn't 3086 * be claimed because a device was missing during 3087 * zil_claim(), but that device later returns, 3088 * then this block could erroneously appear valid. 3089 * To guard against this, assign a new GUID to the new 3090 * log chain so it doesn't matter what blk points to. 3091 */ 3092 zil_init_log_chain(zilog, &blk); 3093 zh->zh_log = blk; 3094 } 3095 } 3096 3097 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 3098 zh->zh_log = lwb->lwb_blk; 3099 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 3100 break; 3101 list_remove(&zilog->zl_lwb_list, lwb); 3102 zio_free(spa, txg, &lwb->lwb_blk); 3103 zil_free_lwb(zilog, lwb); 3104 3105 /* 3106 * If we don't have anything left in the lwb list then 3107 * we've had an allocation failure and we need to zero 3108 * out the zil_header blkptr so that we don't end 3109 * up freeing the same block twice. 3110 */ 3111 if (list_head(&zilog->zl_lwb_list) == NULL) 3112 BP_ZERO(&zh->zh_log); 3113 } 3114 3115 /* 3116 * Remove fastwrite on any blocks that have been pre-allocated for 3117 * the next commit. This prevents fastwrite counter pollution by 3118 * unused, long-lived LWBs. 3119 */ 3120 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) { 3121 if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) { 3122 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); 3123 lwb->lwb_fastwrite = 0; 3124 } 3125 } 3126 3127 mutex_exit(&zilog->zl_lock); 3128 } 3129 3130 /* ARGSUSED */ 3131 static int 3132 zil_lwb_cons(void *vbuf, void *unused, int kmflag) 3133 { 3134 lwb_t *lwb = vbuf; 3135 list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); 3136 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t), 3137 offsetof(zil_commit_waiter_t, zcw_node)); 3138 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare, 3139 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 3140 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 3141 return (0); 3142 } 3143 3144 /* ARGSUSED */ 3145 static void 3146 zil_lwb_dest(void *vbuf, void *unused) 3147 { 3148 lwb_t *lwb = vbuf; 3149 mutex_destroy(&lwb->lwb_vdev_lock); 3150 avl_destroy(&lwb->lwb_vdev_tree); 3151 list_destroy(&lwb->lwb_waiters); 3152 list_destroy(&lwb->lwb_itxs); 3153 } 3154 3155 void 3156 zil_init(void) 3157 { 3158 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 3159 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0); 3160 3161 zil_zcw_cache = kmem_cache_create("zil_zcw_cache", 3162 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 3163 3164 zil_ksp = kstat_create("zfs", 0, "zil", "misc", 3165 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t), 3166 KSTAT_FLAG_VIRTUAL); 3167 3168 if (zil_ksp != NULL) { 3169 zil_ksp->ks_data = &zil_stats; 3170 kstat_install(zil_ksp); 3171 } 3172 } 3173 3174 void 3175 zil_fini(void) 3176 { 3177 kmem_cache_destroy(zil_zcw_cache); 3178 kmem_cache_destroy(zil_lwb_cache); 3179 3180 if (zil_ksp != NULL) { 3181 kstat_delete(zil_ksp); 3182 zil_ksp = NULL; 3183 } 3184 } 3185 3186 void 3187 zil_set_sync(zilog_t *zilog, uint64_t sync) 3188 { 3189 zilog->zl_sync = sync; 3190 } 3191 3192 void 3193 zil_set_logbias(zilog_t *zilog, uint64_t logbias) 3194 { 3195 zilog->zl_logbias = logbias; 3196 } 3197 3198 zilog_t * 3199 zil_alloc(objset_t *os, zil_header_t *zh_phys) 3200 { 3201 zilog_t *zilog; 3202 3203 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 3204 3205 zilog->zl_header = zh_phys; 3206 zilog->zl_os = os; 3207 zilog->zl_spa = dmu_objset_spa(os); 3208 zilog->zl_dmu_pool = dmu_objset_pool(os); 3209 zilog->zl_destroy_txg = TXG_INITIAL - 1; 3210 zilog->zl_logbias = dmu_objset_logbias(os); 3211 zilog->zl_sync = dmu_objset_syncprop(os); 3212 zilog->zl_dirty_max_txg = 0; 3213 zilog->zl_last_lwb_opened = NULL; 3214 zilog->zl_last_lwb_latency = 0; 3215 zilog->zl_max_block_size = zil_maxblocksize; 3216 3217 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 3218 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL); 3219 3220 for (int i = 0; i < TXG_SIZE; i++) { 3221 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 3222 MUTEX_DEFAULT, NULL); 3223 } 3224 3225 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 3226 offsetof(lwb_t, lwb_node)); 3227 3228 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 3229 offsetof(itx_t, itx_node)); 3230 3231 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 3232 3233 return (zilog); 3234 } 3235 3236 void 3237 zil_free(zilog_t *zilog) 3238 { 3239 int i; 3240 3241 zilog->zl_stop_sync = 1; 3242 3243 ASSERT0(zilog->zl_suspend); 3244 ASSERT0(zilog->zl_suspending); 3245 3246 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3247 list_destroy(&zilog->zl_lwb_list); 3248 3249 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 3250 list_destroy(&zilog->zl_itx_commit_list); 3251 3252 for (i = 0; i < TXG_SIZE; i++) { 3253 /* 3254 * It's possible for an itx to be generated that doesn't dirty 3255 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 3256 * callback to remove the entry. We remove those here. 3257 * 3258 * Also free up the ziltest itxs. 3259 */ 3260 if (zilog->zl_itxg[i].itxg_itxs) 3261 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 3262 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 3263 } 3264 3265 mutex_destroy(&zilog->zl_issuer_lock); 3266 mutex_destroy(&zilog->zl_lock); 3267 3268 cv_destroy(&zilog->zl_cv_suspend); 3269 3270 kmem_free(zilog, sizeof (zilog_t)); 3271 } 3272 3273 /* 3274 * Open an intent log. 3275 */ 3276 zilog_t * 3277 zil_open(objset_t *os, zil_get_data_t *get_data) 3278 { 3279 zilog_t *zilog = dmu_objset_zil(os); 3280 3281 ASSERT3P(zilog->zl_get_data, ==, NULL); 3282 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 3283 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3284 3285 zilog->zl_get_data = get_data; 3286 3287 return (zilog); 3288 } 3289 3290 /* 3291 * Close an intent log. 3292 */ 3293 void 3294 zil_close(zilog_t *zilog) 3295 { 3296 lwb_t *lwb; 3297 uint64_t txg; 3298 3299 if (!dmu_objset_is_snapshot(zilog->zl_os)) { 3300 zil_commit(zilog, 0); 3301 } else { 3302 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 3303 ASSERT0(zilog->zl_dirty_max_txg); 3304 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE); 3305 } 3306 3307 mutex_enter(&zilog->zl_lock); 3308 lwb = list_tail(&zilog->zl_lwb_list); 3309 if (lwb == NULL) 3310 txg = zilog->zl_dirty_max_txg; 3311 else 3312 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg); 3313 mutex_exit(&zilog->zl_lock); 3314 3315 /* 3316 * We need to use txg_wait_synced() to wait long enough for the 3317 * ZIL to be clean, and to wait for all pending lwbs to be 3318 * written out. 3319 */ 3320 if (txg != 0) 3321 txg_wait_synced(zilog->zl_dmu_pool, txg); 3322 3323 if (zilog_is_dirty(zilog)) 3324 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog, 3325 (u_longlong_t)txg); 3326 if (txg < spa_freeze_txg(zilog->zl_spa)) 3327 VERIFY(!zilog_is_dirty(zilog)); 3328 3329 zilog->zl_get_data = NULL; 3330 3331 /* 3332 * We should have only one lwb left on the list; remove it now. 3333 */ 3334 mutex_enter(&zilog->zl_lock); 3335 lwb = list_head(&zilog->zl_lwb_list); 3336 if (lwb != NULL) { 3337 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list)); 3338 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 3339 3340 if (lwb->lwb_fastwrite) 3341 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); 3342 3343 list_remove(&zilog->zl_lwb_list, lwb); 3344 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 3345 zil_free_lwb(zilog, lwb); 3346 } 3347 mutex_exit(&zilog->zl_lock); 3348 } 3349 3350 static char *suspend_tag = "zil suspending"; 3351 3352 /* 3353 * Suspend an intent log. While in suspended mode, we still honor 3354 * synchronous semantics, but we rely on txg_wait_synced() to do it. 3355 * On old version pools, we suspend the log briefly when taking a 3356 * snapshot so that it will have an empty intent log. 3357 * 3358 * Long holds are not really intended to be used the way we do here -- 3359 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 3360 * could fail. Therefore we take pains to only put a long hold if it is 3361 * actually necessary. Fortunately, it will only be necessary if the 3362 * objset is currently mounted (or the ZVOL equivalent). In that case it 3363 * will already have a long hold, so we are not really making things any worse. 3364 * 3365 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 3366 * zvol_state_t), and use their mechanism to prevent their hold from being 3367 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 3368 * very little gain. 3369 * 3370 * if cookiep == NULL, this does both the suspend & resume. 3371 * Otherwise, it returns with the dataset "long held", and the cookie 3372 * should be passed into zil_resume(). 3373 */ 3374 int 3375 zil_suspend(const char *osname, void **cookiep) 3376 { 3377 objset_t *os; 3378 zilog_t *zilog; 3379 const zil_header_t *zh; 3380 int error; 3381 3382 error = dmu_objset_hold(osname, suspend_tag, &os); 3383 if (error != 0) 3384 return (error); 3385 zilog = dmu_objset_zil(os); 3386 3387 mutex_enter(&zilog->zl_lock); 3388 zh = zilog->zl_header; 3389 3390 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 3391 mutex_exit(&zilog->zl_lock); 3392 dmu_objset_rele(os, suspend_tag); 3393 return (SET_ERROR(EBUSY)); 3394 } 3395 3396 /* 3397 * Don't put a long hold in the cases where we can avoid it. This 3398 * is when there is no cookie so we are doing a suspend & resume 3399 * (i.e. called from zil_vdev_offline()), and there's nothing to do 3400 * for the suspend because it's already suspended, or there's no ZIL. 3401 */ 3402 if (cookiep == NULL && !zilog->zl_suspending && 3403 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 3404 mutex_exit(&zilog->zl_lock); 3405 dmu_objset_rele(os, suspend_tag); 3406 return (0); 3407 } 3408 3409 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 3410 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 3411 3412 zilog->zl_suspend++; 3413 3414 if (zilog->zl_suspend > 1) { 3415 /* 3416 * Someone else is already suspending it. 3417 * Just wait for them to finish. 3418 */ 3419 3420 while (zilog->zl_suspending) 3421 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 3422 mutex_exit(&zilog->zl_lock); 3423 3424 if (cookiep == NULL) 3425 zil_resume(os); 3426 else 3427 *cookiep = os; 3428 return (0); 3429 } 3430 3431 /* 3432 * If there is no pointer to an on-disk block, this ZIL must not 3433 * be active (e.g. filesystem not mounted), so there's nothing 3434 * to clean up. 3435 */ 3436 if (BP_IS_HOLE(&zh->zh_log)) { 3437 ASSERT(cookiep != NULL); /* fast path already handled */ 3438 3439 *cookiep = os; 3440 mutex_exit(&zilog->zl_lock); 3441 return (0); 3442 } 3443 3444 /* 3445 * The ZIL has work to do. Ensure that the associated encryption 3446 * key will remain mapped while we are committing the log by 3447 * grabbing a reference to it. If the key isn't loaded we have no 3448 * choice but to return an error until the wrapping key is loaded. 3449 */ 3450 if (os->os_encrypted && 3451 dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) { 3452 zilog->zl_suspend--; 3453 mutex_exit(&zilog->zl_lock); 3454 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3455 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3456 return (SET_ERROR(EACCES)); 3457 } 3458 3459 zilog->zl_suspending = B_TRUE; 3460 mutex_exit(&zilog->zl_lock); 3461 3462 /* 3463 * We need to use zil_commit_impl to ensure we wait for all 3464 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed 3465 * to disk before proceeding. If we used zil_commit instead, it 3466 * would just call txg_wait_synced(), because zl_suspend is set. 3467 * txg_wait_synced() doesn't wait for these lwb's to be 3468 * LWB_STATE_FLUSH_DONE before returning. 3469 */ 3470 zil_commit_impl(zilog, 0); 3471 3472 /* 3473 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we 3474 * use txg_wait_synced() to ensure the data from the zilog has 3475 * migrated to the main pool before calling zil_destroy(). 3476 */ 3477 txg_wait_synced(zilog->zl_dmu_pool, 0); 3478 3479 zil_destroy(zilog, B_FALSE); 3480 3481 mutex_enter(&zilog->zl_lock); 3482 zilog->zl_suspending = B_FALSE; 3483 cv_broadcast(&zilog->zl_cv_suspend); 3484 mutex_exit(&zilog->zl_lock); 3485 3486 if (os->os_encrypted) 3487 dsl_dataset_remove_key_mapping(dmu_objset_ds(os)); 3488 3489 if (cookiep == NULL) 3490 zil_resume(os); 3491 else 3492 *cookiep = os; 3493 return (0); 3494 } 3495 3496 void 3497 zil_resume(void *cookie) 3498 { 3499 objset_t *os = cookie; 3500 zilog_t *zilog = dmu_objset_zil(os); 3501 3502 mutex_enter(&zilog->zl_lock); 3503 ASSERT(zilog->zl_suspend != 0); 3504 zilog->zl_suspend--; 3505 mutex_exit(&zilog->zl_lock); 3506 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3507 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3508 } 3509 3510 typedef struct zil_replay_arg { 3511 zil_replay_func_t **zr_replay; 3512 void *zr_arg; 3513 boolean_t zr_byteswap; 3514 char *zr_lr; 3515 } zil_replay_arg_t; 3516 3517 static int 3518 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error) 3519 { 3520 char name[ZFS_MAX_DATASET_NAME_LEN]; 3521 3522 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 3523 3524 dmu_objset_name(zilog->zl_os, name); 3525 3526 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 3527 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 3528 (u_longlong_t)lr->lrc_seq, 3529 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 3530 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 3531 3532 return (error); 3533 } 3534 3535 static int 3536 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra, 3537 uint64_t claim_txg) 3538 { 3539 zil_replay_arg_t *zr = zra; 3540 const zil_header_t *zh = zilog->zl_header; 3541 uint64_t reclen = lr->lrc_reclen; 3542 uint64_t txtype = lr->lrc_txtype; 3543 int error = 0; 3544 3545 zilog->zl_replaying_seq = lr->lrc_seq; 3546 3547 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 3548 return (0); 3549 3550 if (lr->lrc_txg < claim_txg) /* already committed */ 3551 return (0); 3552 3553 /* Strip case-insensitive bit, still present in log record */ 3554 txtype &= ~TX_CI; 3555 3556 if (txtype == 0 || txtype >= TX_MAX_TYPE) 3557 return (zil_replay_error(zilog, lr, EINVAL)); 3558 3559 /* 3560 * If this record type can be logged out of order, the object 3561 * (lr_foid) may no longer exist. That's legitimate, not an error. 3562 */ 3563 if (TX_OOO(txtype)) { 3564 error = dmu_object_info(zilog->zl_os, 3565 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL); 3566 if (error == ENOENT || error == EEXIST) 3567 return (0); 3568 } 3569 3570 /* 3571 * Make a copy of the data so we can revise and extend it. 3572 */ 3573 bcopy(lr, zr->zr_lr, reclen); 3574 3575 /* 3576 * If this is a TX_WRITE with a blkptr, suck in the data. 3577 */ 3578 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 3579 error = zil_read_log_data(zilog, (lr_write_t *)lr, 3580 zr->zr_lr + reclen); 3581 if (error != 0) 3582 return (zil_replay_error(zilog, lr, error)); 3583 } 3584 3585 /* 3586 * The log block containing this lr may have been byteswapped 3587 * so that we can easily examine common fields like lrc_txtype. 3588 * However, the log is a mix of different record types, and only the 3589 * replay vectors know how to byteswap their records. Therefore, if 3590 * the lr was byteswapped, undo it before invoking the replay vector. 3591 */ 3592 if (zr->zr_byteswap) 3593 byteswap_uint64_array(zr->zr_lr, reclen); 3594 3595 /* 3596 * We must now do two things atomically: replay this log record, 3597 * and update the log header sequence number to reflect the fact that 3598 * we did so. At the end of each replay function the sequence number 3599 * is updated if we are in replay mode. 3600 */ 3601 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 3602 if (error != 0) { 3603 /* 3604 * The DMU's dnode layer doesn't see removes until the txg 3605 * commits, so a subsequent claim can spuriously fail with 3606 * EEXIST. So if we receive any error we try syncing out 3607 * any removes then retry the transaction. Note that we 3608 * specify B_FALSE for byteswap now, so we don't do it twice. 3609 */ 3610 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 3611 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 3612 if (error != 0) 3613 return (zil_replay_error(zilog, lr, error)); 3614 } 3615 return (0); 3616 } 3617 3618 /* ARGSUSED */ 3619 static int 3620 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg) 3621 { 3622 zilog->zl_replay_blks++; 3623 3624 return (0); 3625 } 3626 3627 /* 3628 * If this dataset has a non-empty intent log, replay it and destroy it. 3629 */ 3630 void 3631 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 3632 { 3633 zilog_t *zilog = dmu_objset_zil(os); 3634 const zil_header_t *zh = zilog->zl_header; 3635 zil_replay_arg_t zr; 3636 3637 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 3638 zil_destroy(zilog, B_TRUE); 3639 return; 3640 } 3641 3642 zr.zr_replay = replay_func; 3643 zr.zr_arg = arg; 3644 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 3645 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 3646 3647 /* 3648 * Wait for in-progress removes to sync before starting replay. 3649 */ 3650 txg_wait_synced(zilog->zl_dmu_pool, 0); 3651 3652 zilog->zl_replay = B_TRUE; 3653 zilog->zl_replay_time = ddi_get_lbolt(); 3654 ASSERT(zilog->zl_replay_blks == 0); 3655 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 3656 zh->zh_claim_txg, B_TRUE); 3657 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 3658 3659 zil_destroy(zilog, B_FALSE); 3660 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 3661 zilog->zl_replay = B_FALSE; 3662 } 3663 3664 boolean_t 3665 zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 3666 { 3667 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 3668 return (B_TRUE); 3669 3670 if (zilog->zl_replay) { 3671 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 3672 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 3673 zilog->zl_replaying_seq; 3674 return (B_TRUE); 3675 } 3676 3677 return (B_FALSE); 3678 } 3679 3680 /* ARGSUSED */ 3681 int 3682 zil_reset(const char *osname, void *arg) 3683 { 3684 int error; 3685 3686 error = zil_suspend(osname, NULL); 3687 /* EACCES means crypto key not loaded */ 3688 if ((error == EACCES) || (error == EBUSY)) 3689 return (SET_ERROR(error)); 3690 if (error != 0) 3691 return (SET_ERROR(EEXIST)); 3692 return (0); 3693 } 3694 3695 EXPORT_SYMBOL(zil_alloc); 3696 EXPORT_SYMBOL(zil_free); 3697 EXPORT_SYMBOL(zil_open); 3698 EXPORT_SYMBOL(zil_close); 3699 EXPORT_SYMBOL(zil_replay); 3700 EXPORT_SYMBOL(zil_replaying); 3701 EXPORT_SYMBOL(zil_destroy); 3702 EXPORT_SYMBOL(zil_destroy_sync); 3703 EXPORT_SYMBOL(zil_itx_create); 3704 EXPORT_SYMBOL(zil_itx_destroy); 3705 EXPORT_SYMBOL(zil_itx_assign); 3706 EXPORT_SYMBOL(zil_commit); 3707 EXPORT_SYMBOL(zil_claim); 3708 EXPORT_SYMBOL(zil_check_log_chain); 3709 EXPORT_SYMBOL(zil_sync); 3710 EXPORT_SYMBOL(zil_clean); 3711 EXPORT_SYMBOL(zil_suspend); 3712 EXPORT_SYMBOL(zil_resume); 3713 EXPORT_SYMBOL(zil_lwb_add_block); 3714 EXPORT_SYMBOL(zil_bp_tree_add); 3715 EXPORT_SYMBOL(zil_set_sync); 3716 EXPORT_SYMBOL(zil_set_logbias); 3717 3718 /* BEGIN CSTYLED */ 3719 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, INT, ZMOD_RW, 3720 "ZIL block open timeout percentage"); 3721 3722 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW, 3723 "Disable intent logging replay"); 3724 3725 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW, 3726 "Disable ZIL cache flushes"); 3727 3728 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, ULONG, ZMOD_RW, 3729 "Limit in bytes slog sync writes per commit"); 3730 3731 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, INT, ZMOD_RW, 3732 "Limit in bytes of ZIL log block size"); 3733 /* END CSTYLED */ 3734