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