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 static int zfs_commit_timeout_pct = 5; 93 94 /* 95 * See zil.h for more information about these fields. 96 */ 97 static 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 static 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 static unsigned long zil_slog_bulk = 768 * 1024; 134 135 static kmem_cache_t *zil_lwb_cache; 136 static kmem_cache_t *zil_zcw_cache; 137 138 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 139 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 140 141 static int 142 zil_bp_compare(const void *x1, const void *x2) 143 { 144 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 145 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 146 147 int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2)); 148 if (likely(cmp)) 149 return (cmp); 150 151 return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2))); 152 } 153 154 static void 155 zil_bp_tree_init(zilog_t *zilog) 156 { 157 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 158 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 159 } 160 161 static void 162 zil_bp_tree_fini(zilog_t *zilog) 163 { 164 avl_tree_t *t = &zilog->zl_bp_tree; 165 zil_bp_node_t *zn; 166 void *cookie = NULL; 167 168 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 169 kmem_free(zn, sizeof (zil_bp_node_t)); 170 171 avl_destroy(t); 172 } 173 174 int 175 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 176 { 177 avl_tree_t *t = &zilog->zl_bp_tree; 178 const dva_t *dva; 179 zil_bp_node_t *zn; 180 avl_index_t where; 181 182 if (BP_IS_EMBEDDED(bp)) 183 return (0); 184 185 dva = BP_IDENTITY(bp); 186 187 if (avl_find(t, dva, &where) != NULL) 188 return (SET_ERROR(EEXIST)); 189 190 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 191 zn->zn_dva = *dva; 192 avl_insert(t, zn, where); 193 194 return (0); 195 } 196 197 static zil_header_t * 198 zil_header_in_syncing_context(zilog_t *zilog) 199 { 200 return ((zil_header_t *)zilog->zl_header); 201 } 202 203 static void 204 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 205 { 206 zio_cksum_t *zc = &bp->blk_cksum; 207 208 (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0], 209 sizeof (zc->zc_word[ZIL_ZC_GUID_0])); 210 (void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1], 211 sizeof (zc->zc_word[ZIL_ZC_GUID_1])); 212 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 213 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 214 } 215 216 /* 217 * Read a log block and make sure it's valid. 218 */ 219 static int 220 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp, 221 blkptr_t *nbp, void *dst, char **end) 222 { 223 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 224 arc_flags_t aflags = ARC_FLAG_WAIT; 225 arc_buf_t *abuf = NULL; 226 zbookmark_phys_t zb; 227 int error; 228 229 if (zilog->zl_header->zh_claim_txg == 0) 230 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 231 232 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 233 zio_flags |= ZIO_FLAG_SPECULATIVE; 234 235 if (!decrypt) 236 zio_flags |= ZIO_FLAG_RAW; 237 238 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 239 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 240 241 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, 242 &abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 243 244 if (error == 0) { 245 zio_cksum_t cksum = bp->blk_cksum; 246 247 /* 248 * Validate the checksummed log block. 249 * 250 * Sequence numbers should be... sequential. The checksum 251 * verifier for the next block should be bp's checksum plus 1. 252 * 253 * Also check the log chain linkage and size used. 254 */ 255 cksum.zc_word[ZIL_ZC_SEQ]++; 256 257 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 258 zil_chain_t *zilc = abuf->b_data; 259 char *lr = (char *)(zilc + 1); 260 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 261 262 if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 263 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 264 error = SET_ERROR(ECKSUM); 265 } else { 266 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 267 memcpy(dst, lr, len); 268 *end = (char *)dst + len; 269 *nbp = zilc->zc_next_blk; 270 } 271 } else { 272 char *lr = abuf->b_data; 273 uint64_t size = BP_GET_LSIZE(bp); 274 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 275 276 if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 277 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 278 (zilc->zc_nused > (size - sizeof (*zilc)))) { 279 error = SET_ERROR(ECKSUM); 280 } else { 281 ASSERT3U(zilc->zc_nused, <=, 282 SPA_OLD_MAXBLOCKSIZE); 283 memcpy(dst, lr, zilc->zc_nused); 284 *end = (char *)dst + zilc->zc_nused; 285 *nbp = zilc->zc_next_blk; 286 } 287 } 288 289 arc_buf_destroy(abuf, &abuf); 290 } 291 292 return (error); 293 } 294 295 /* 296 * Read a TX_WRITE log data block. 297 */ 298 static int 299 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 300 { 301 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 302 const blkptr_t *bp = &lr->lr_blkptr; 303 arc_flags_t aflags = ARC_FLAG_WAIT; 304 arc_buf_t *abuf = NULL; 305 zbookmark_phys_t zb; 306 int error; 307 308 if (BP_IS_HOLE(bp)) { 309 if (wbuf != NULL) 310 memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 311 return (0); 312 } 313 314 if (zilog->zl_header->zh_claim_txg == 0) 315 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 316 317 /* 318 * If we are not using the resulting data, we are just checking that 319 * it hasn't been corrupted so we don't need to waste CPU time 320 * decompressing and decrypting it. 321 */ 322 if (wbuf == NULL) 323 zio_flags |= ZIO_FLAG_RAW; 324 325 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 326 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 327 328 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 329 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 330 331 if (error == 0) { 332 if (wbuf != NULL) 333 memcpy(wbuf, abuf->b_data, arc_buf_size(abuf)); 334 arc_buf_destroy(abuf, &abuf); 335 } 336 337 return (error); 338 } 339 340 /* 341 * Parse the intent log, and call parse_func for each valid record within. 342 */ 343 int 344 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 345 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg, 346 boolean_t decrypt) 347 { 348 const zil_header_t *zh = zilog->zl_header; 349 boolean_t claimed = !!zh->zh_claim_txg; 350 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 351 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 352 uint64_t max_blk_seq = 0; 353 uint64_t max_lr_seq = 0; 354 uint64_t blk_count = 0; 355 uint64_t lr_count = 0; 356 blkptr_t blk, next_blk = {{{{0}}}}; 357 char *lrbuf, *lrp; 358 int error = 0; 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 static int 434 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, 435 uint64_t first_txg) 436 { 437 (void) tx; 438 ASSERT(!BP_IS_HOLE(bp)); 439 440 /* 441 * As we call this function from the context of a rewind to a 442 * checkpoint, each ZIL block whose txg is later than the txg 443 * that we rewind to is invalid. Thus, we return -1 so 444 * zil_parse() doesn't attempt to read it. 445 */ 446 if (bp->blk_birth >= first_txg) 447 return (-1); 448 449 if (zil_bp_tree_add(zilog, bp) != 0) 450 return (0); 451 452 zio_free(zilog->zl_spa, first_txg, bp); 453 return (0); 454 } 455 456 static int 457 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, 458 uint64_t first_txg) 459 { 460 (void) zilog, (void) lrc, (void) tx, (void) first_txg; 461 return (0); 462 } 463 464 static int 465 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, 466 uint64_t first_txg) 467 { 468 /* 469 * Claim log block if not already committed and not already claimed. 470 * If tx == NULL, just verify that the block is claimable. 471 */ 472 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 473 zil_bp_tree_add(zilog, bp) != 0) 474 return (0); 475 476 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 477 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 478 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 479 } 480 481 static int 482 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, 483 uint64_t first_txg) 484 { 485 lr_write_t *lr = (lr_write_t *)lrc; 486 int error; 487 488 if (lrc->lrc_txtype != TX_WRITE) 489 return (0); 490 491 /* 492 * If the block is not readable, don't claim it. This can happen 493 * in normal operation when a log block is written to disk before 494 * some of the dmu_sync() blocks it points to. In this case, the 495 * transaction cannot have been committed to anyone (we would have 496 * waited for all writes to be stable first), so it is semantically 497 * correct to declare this the end of the log. 498 */ 499 if (lr->lr_blkptr.blk_birth >= first_txg) { 500 error = zil_read_log_data(zilog, lr, NULL); 501 if (error != 0) 502 return (error); 503 } 504 505 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 506 } 507 508 static int 509 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx, 510 uint64_t claim_txg) 511 { 512 (void) claim_txg; 513 514 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 515 516 return (0); 517 } 518 519 static int 520 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx, 521 uint64_t claim_txg) 522 { 523 lr_write_t *lr = (lr_write_t *)lrc; 524 blkptr_t *bp = &lr->lr_blkptr; 525 526 /* 527 * If we previously claimed it, we need to free it. 528 */ 529 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 530 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 531 !BP_IS_HOLE(bp)) 532 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 533 534 return (0); 535 } 536 537 static int 538 zil_lwb_vdev_compare(const void *x1, const void *x2) 539 { 540 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 541 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 542 543 return (TREE_CMP(v1, v2)); 544 } 545 546 static lwb_t * 547 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg, 548 boolean_t fastwrite) 549 { 550 lwb_t *lwb; 551 552 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 553 lwb->lwb_zilog = zilog; 554 lwb->lwb_blk = *bp; 555 lwb->lwb_fastwrite = fastwrite; 556 lwb->lwb_slog = slog; 557 lwb->lwb_state = LWB_STATE_CLOSED; 558 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 559 lwb->lwb_max_txg = txg; 560 lwb->lwb_write_zio = NULL; 561 lwb->lwb_root_zio = NULL; 562 lwb->lwb_issued_timestamp = 0; 563 lwb->lwb_issued_txg = 0; 564 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 565 lwb->lwb_nused = sizeof (zil_chain_t); 566 lwb->lwb_sz = BP_GET_LSIZE(bp); 567 } else { 568 lwb->lwb_nused = 0; 569 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 570 } 571 572 mutex_enter(&zilog->zl_lock); 573 list_insert_tail(&zilog->zl_lwb_list, lwb); 574 mutex_exit(&zilog->zl_lock); 575 576 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); 577 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 578 VERIFY(list_is_empty(&lwb->lwb_waiters)); 579 VERIFY(list_is_empty(&lwb->lwb_itxs)); 580 581 return (lwb); 582 } 583 584 static void 585 zil_free_lwb(zilog_t *zilog, lwb_t *lwb) 586 { 587 ASSERT(MUTEX_HELD(&zilog->zl_lock)); 588 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); 589 VERIFY(list_is_empty(&lwb->lwb_waiters)); 590 VERIFY(list_is_empty(&lwb->lwb_itxs)); 591 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 592 ASSERT3P(lwb->lwb_write_zio, ==, NULL); 593 ASSERT3P(lwb->lwb_root_zio, ==, NULL); 594 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa)); 595 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED || 596 lwb->lwb_state == LWB_STATE_FLUSH_DONE); 597 598 /* 599 * Clear the zilog's field to indicate this lwb is no longer 600 * valid, and prevent use-after-free errors. 601 */ 602 if (zilog->zl_last_lwb_opened == lwb) 603 zilog->zl_last_lwb_opened = NULL; 604 605 kmem_cache_free(zil_lwb_cache, lwb); 606 } 607 608 /* 609 * Called when we create in-memory log transactions so that we know 610 * to cleanup the itxs at the end of spa_sync(). 611 */ 612 static void 613 zilog_dirty(zilog_t *zilog, uint64_t txg) 614 { 615 dsl_pool_t *dp = zilog->zl_dmu_pool; 616 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 617 618 ASSERT(spa_writeable(zilog->zl_spa)); 619 620 if (ds->ds_is_snapshot) 621 panic("dirtying snapshot!"); 622 623 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 624 /* up the hold count until we can be written out */ 625 dmu_buf_add_ref(ds->ds_dbuf, zilog); 626 627 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg); 628 } 629 } 630 631 /* 632 * Determine if the zil is dirty in the specified txg. Callers wanting to 633 * ensure that the dirty state does not change must hold the itxg_lock for 634 * the specified txg. Holding the lock will ensure that the zil cannot be 635 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 636 * state. 637 */ 638 static boolean_t __maybe_unused 639 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 640 { 641 dsl_pool_t *dp = zilog->zl_dmu_pool; 642 643 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 644 return (B_TRUE); 645 return (B_FALSE); 646 } 647 648 /* 649 * Determine if the zil is dirty. The zil is considered dirty if it has 650 * any pending itx records that have not been cleaned by zil_clean(). 651 */ 652 static boolean_t 653 zilog_is_dirty(zilog_t *zilog) 654 { 655 dsl_pool_t *dp = zilog->zl_dmu_pool; 656 657 for (int t = 0; t < TXG_SIZE; t++) { 658 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 659 return (B_TRUE); 660 } 661 return (B_FALSE); 662 } 663 664 /* 665 * Its called in zil_commit context (zil_process_commit_list()/zil_create()). 666 * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled. 667 * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every 668 * zil_commit. 669 */ 670 static void 671 zil_commit_activate_saxattr_feature(zilog_t *zilog) 672 { 673 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 674 uint64_t txg = 0; 675 dmu_tx_t *tx = NULL; 676 677 if (spa_feature_is_enabled(zilog->zl_spa, 678 SPA_FEATURE_ZILSAXATTR) && 679 dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL && 680 !dsl_dataset_feature_is_active(ds, 681 SPA_FEATURE_ZILSAXATTR)) { 682 tx = dmu_tx_create(zilog->zl_os); 683 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 684 dsl_dataset_dirty(ds, tx); 685 txg = dmu_tx_get_txg(tx); 686 687 mutex_enter(&ds->ds_lock); 688 ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] = 689 (void *)B_TRUE; 690 mutex_exit(&ds->ds_lock); 691 dmu_tx_commit(tx); 692 txg_wait_synced(zilog->zl_dmu_pool, txg); 693 } 694 } 695 696 /* 697 * Create an on-disk intent log. 698 */ 699 static lwb_t * 700 zil_create(zilog_t *zilog) 701 { 702 const zil_header_t *zh = zilog->zl_header; 703 lwb_t *lwb = NULL; 704 uint64_t txg = 0; 705 dmu_tx_t *tx = NULL; 706 blkptr_t blk; 707 int error = 0; 708 boolean_t fastwrite = FALSE; 709 boolean_t slog = FALSE; 710 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 711 712 713 /* 714 * Wait for any previous destroy to complete. 715 */ 716 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 717 718 ASSERT(zh->zh_claim_txg == 0); 719 ASSERT(zh->zh_replay_seq == 0); 720 721 blk = zh->zh_log; 722 723 /* 724 * Allocate an initial log block if: 725 * - there isn't one already 726 * - the existing block is the wrong endianness 727 */ 728 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 729 tx = dmu_tx_create(zilog->zl_os); 730 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 731 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 732 txg = dmu_tx_get_txg(tx); 733 734 if (!BP_IS_HOLE(&blk)) { 735 zio_free(zilog->zl_spa, txg, &blk); 736 BP_ZERO(&blk); 737 } 738 739 error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk, 740 ZIL_MIN_BLKSZ, &slog); 741 fastwrite = TRUE; 742 743 if (error == 0) 744 zil_init_log_chain(zilog, &blk); 745 } 746 747 /* 748 * Allocate a log write block (lwb) for the first log block. 749 */ 750 if (error == 0) 751 lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite); 752 753 /* 754 * If we just allocated the first log block, commit our transaction 755 * and wait for zil_sync() to stuff the block pointer into zh_log. 756 * (zh is part of the MOS, so we cannot modify it in open context.) 757 */ 758 if (tx != NULL) { 759 /* 760 * If "zilsaxattr" feature is enabled on zpool, then activate 761 * it now when we're creating the ZIL chain. We can't wait with 762 * this until we write the first xattr log record because we 763 * need to wait for the feature activation to sync out. 764 */ 765 if (spa_feature_is_enabled(zilog->zl_spa, 766 SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) != 767 DMU_OST_ZVOL) { 768 mutex_enter(&ds->ds_lock); 769 ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] = 770 (void *)B_TRUE; 771 mutex_exit(&ds->ds_lock); 772 } 773 774 dmu_tx_commit(tx); 775 txg_wait_synced(zilog->zl_dmu_pool, txg); 776 } else { 777 /* 778 * This branch covers the case where we enable the feature on a 779 * zpool that has existing ZIL headers. 780 */ 781 zil_commit_activate_saxattr_feature(zilog); 782 } 783 IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) && 784 dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL, 785 dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)); 786 787 ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 788 IMPLY(error == 0, lwb != NULL); 789 790 return (lwb); 791 } 792 793 /* 794 * In one tx, free all log blocks and clear the log header. If keep_first 795 * is set, then we're replaying a log with no content. We want to keep the 796 * first block, however, so that the first synchronous transaction doesn't 797 * require a txg_wait_synced() in zil_create(). We don't need to 798 * txg_wait_synced() here either when keep_first is set, because both 799 * zil_create() and zil_destroy() will wait for any in-progress destroys 800 * to complete. 801 */ 802 void 803 zil_destroy(zilog_t *zilog, boolean_t keep_first) 804 { 805 const zil_header_t *zh = zilog->zl_header; 806 lwb_t *lwb; 807 dmu_tx_t *tx; 808 uint64_t txg; 809 810 /* 811 * Wait for any previous destroy to complete. 812 */ 813 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 814 815 zilog->zl_old_header = *zh; /* debugging aid */ 816 817 if (BP_IS_HOLE(&zh->zh_log)) 818 return; 819 820 tx = dmu_tx_create(zilog->zl_os); 821 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 822 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 823 txg = dmu_tx_get_txg(tx); 824 825 mutex_enter(&zilog->zl_lock); 826 827 ASSERT3U(zilog->zl_destroy_txg, <, txg); 828 zilog->zl_destroy_txg = txg; 829 zilog->zl_keep_first = keep_first; 830 831 if (!list_is_empty(&zilog->zl_lwb_list)) { 832 ASSERT(zh->zh_claim_txg == 0); 833 VERIFY(!keep_first); 834 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 835 if (lwb->lwb_fastwrite) 836 metaslab_fastwrite_unmark(zilog->zl_spa, 837 &lwb->lwb_blk); 838 839 list_remove(&zilog->zl_lwb_list, lwb); 840 if (lwb->lwb_buf != NULL) 841 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 842 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk); 843 zil_free_lwb(zilog, lwb); 844 } 845 } else if (!keep_first) { 846 zil_destroy_sync(zilog, tx); 847 } 848 mutex_exit(&zilog->zl_lock); 849 850 dmu_tx_commit(tx); 851 } 852 853 void 854 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 855 { 856 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 857 (void) zil_parse(zilog, zil_free_log_block, 858 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE); 859 } 860 861 int 862 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 863 { 864 dmu_tx_t *tx = txarg; 865 zilog_t *zilog; 866 uint64_t first_txg; 867 zil_header_t *zh; 868 objset_t *os; 869 int error; 870 871 error = dmu_objset_own_obj(dp, ds->ds_object, 872 DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os); 873 if (error != 0) { 874 /* 875 * EBUSY indicates that the objset is inconsistent, in which 876 * case it can not have a ZIL. 877 */ 878 if (error != EBUSY) { 879 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 880 (unsigned long long)ds->ds_object, error); 881 } 882 883 return (0); 884 } 885 886 zilog = dmu_objset_zil(os); 887 zh = zil_header_in_syncing_context(zilog); 888 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa)); 889 first_txg = spa_min_claim_txg(zilog->zl_spa); 890 891 /* 892 * If the spa_log_state is not set to be cleared, check whether 893 * the current uberblock is a checkpoint one and if the current 894 * header has been claimed before moving on. 895 * 896 * If the current uberblock is a checkpointed uberblock then 897 * one of the following scenarios took place: 898 * 899 * 1] We are currently rewinding to the checkpoint of the pool. 900 * 2] We crashed in the middle of a checkpoint rewind but we 901 * did manage to write the checkpointed uberblock to the 902 * vdev labels, so when we tried to import the pool again 903 * the checkpointed uberblock was selected from the import 904 * procedure. 905 * 906 * In both cases we want to zero out all the ZIL blocks, except 907 * the ones that have been claimed at the time of the checkpoint 908 * (their zh_claim_txg != 0). The reason is that these blocks 909 * may be corrupted since we may have reused their locations on 910 * disk after we took the checkpoint. 911 * 912 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier 913 * when we first figure out whether the current uberblock is 914 * checkpointed or not. Unfortunately, that would discard all 915 * the logs, including the ones that are claimed, and we would 916 * leak space. 917 */ 918 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR || 919 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && 920 zh->zh_claim_txg == 0)) { 921 if (!BP_IS_HOLE(&zh->zh_log)) { 922 (void) zil_parse(zilog, zil_clear_log_block, 923 zil_noop_log_record, tx, first_txg, B_FALSE); 924 } 925 BP_ZERO(&zh->zh_log); 926 if (os->os_encrypted) 927 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE; 928 dsl_dataset_dirty(dmu_objset_ds(os), tx); 929 dmu_objset_disown(os, B_FALSE, FTAG); 930 return (0); 931 } 932 933 /* 934 * If we are not rewinding and opening the pool normally, then 935 * the min_claim_txg should be equal to the first txg of the pool. 936 */ 937 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa)); 938 939 /* 940 * Claim all log blocks if we haven't already done so, and remember 941 * the highest claimed sequence number. This ensures that if we can 942 * read only part of the log now (e.g. due to a missing device), 943 * but we can read the entire log later, we will not try to replay 944 * or destroy beyond the last block we successfully claimed. 945 */ 946 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 947 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 948 (void) zil_parse(zilog, zil_claim_log_block, 949 zil_claim_log_record, tx, first_txg, B_FALSE); 950 zh->zh_claim_txg = first_txg; 951 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 952 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 953 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 954 zh->zh_flags |= ZIL_REPLAY_NEEDED; 955 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 956 if (os->os_encrypted) 957 os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE; 958 dsl_dataset_dirty(dmu_objset_ds(os), tx); 959 } 960 961 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 962 dmu_objset_disown(os, B_FALSE, FTAG); 963 return (0); 964 } 965 966 /* 967 * Check the log by walking the log chain. 968 * Checksum errors are ok as they indicate the end of the chain. 969 * Any other error (no device or read failure) returns an error. 970 */ 971 int 972 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 973 { 974 (void) dp; 975 zilog_t *zilog; 976 objset_t *os; 977 blkptr_t *bp; 978 int error; 979 980 ASSERT(tx == NULL); 981 982 error = dmu_objset_from_ds(ds, &os); 983 if (error != 0) { 984 cmn_err(CE_WARN, "can't open objset %llu, error %d", 985 (unsigned long long)ds->ds_object, error); 986 return (0); 987 } 988 989 zilog = dmu_objset_zil(os); 990 bp = (blkptr_t *)&zilog->zl_header->zh_log; 991 992 if (!BP_IS_HOLE(bp)) { 993 vdev_t *vd; 994 boolean_t valid = B_TRUE; 995 996 /* 997 * Check the first block and determine if it's on a log device 998 * which may have been removed or faulted prior to loading this 999 * pool. If so, there's no point in checking the rest of the 1000 * log as its content should have already been synced to the 1001 * pool. 1002 */ 1003 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 1004 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 1005 if (vd->vdev_islog && vdev_is_dead(vd)) 1006 valid = vdev_log_state_valid(vd); 1007 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 1008 1009 if (!valid) 1010 return (0); 1011 1012 /* 1013 * Check whether the current uberblock is checkpointed (e.g. 1014 * we are rewinding) and whether the current header has been 1015 * claimed or not. If it hasn't then skip verifying it. We 1016 * do this because its ZIL blocks may be part of the pool's 1017 * state before the rewind, which is no longer valid. 1018 */ 1019 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1020 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && 1021 zh->zh_claim_txg == 0) 1022 return (0); 1023 } 1024 1025 /* 1026 * Because tx == NULL, zil_claim_log_block() will not actually claim 1027 * any blocks, but just determine whether it is possible to do so. 1028 * In addition to checking the log chain, zil_claim_log_block() 1029 * will invoke zio_claim() with a done func of spa_claim_notify(), 1030 * which will update spa_max_claim_txg. See spa_load() for details. 1031 */ 1032 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 1033 zilog->zl_header->zh_claim_txg ? -1ULL : 1034 spa_min_claim_txg(os->os_spa), B_FALSE); 1035 1036 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 1037 } 1038 1039 /* 1040 * When an itx is "skipped", this function is used to properly mark the 1041 * waiter as "done, and signal any thread(s) waiting on it. An itx can 1042 * be skipped (and not committed to an lwb) for a variety of reasons, 1043 * one of them being that the itx was committed via spa_sync(), prior to 1044 * it being committed to an lwb; this can happen if a thread calling 1045 * zil_commit() is racing with spa_sync(). 1046 */ 1047 static void 1048 zil_commit_waiter_skip(zil_commit_waiter_t *zcw) 1049 { 1050 mutex_enter(&zcw->zcw_lock); 1051 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 1052 zcw->zcw_done = B_TRUE; 1053 cv_broadcast(&zcw->zcw_cv); 1054 mutex_exit(&zcw->zcw_lock); 1055 } 1056 1057 /* 1058 * This function is used when the given waiter is to be linked into an 1059 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb. 1060 * At this point, the waiter will no longer be referenced by the itx, 1061 * and instead, will be referenced by the lwb. 1062 */ 1063 static void 1064 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb) 1065 { 1066 /* 1067 * The lwb_waiters field of the lwb is protected by the zilog's 1068 * zl_lock, thus it must be held when calling this function. 1069 */ 1070 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock)); 1071 1072 mutex_enter(&zcw->zcw_lock); 1073 ASSERT(!list_link_active(&zcw->zcw_node)); 1074 ASSERT3P(zcw->zcw_lwb, ==, NULL); 1075 ASSERT3P(lwb, !=, NULL); 1076 ASSERT(lwb->lwb_state == LWB_STATE_OPENED || 1077 lwb->lwb_state == LWB_STATE_ISSUED || 1078 lwb->lwb_state == LWB_STATE_WRITE_DONE); 1079 1080 list_insert_tail(&lwb->lwb_waiters, zcw); 1081 zcw->zcw_lwb = lwb; 1082 mutex_exit(&zcw->zcw_lock); 1083 } 1084 1085 /* 1086 * This function is used when zio_alloc_zil() fails to allocate a ZIL 1087 * block, and the given waiter must be linked to the "nolwb waiters" 1088 * list inside of zil_process_commit_list(). 1089 */ 1090 static void 1091 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb) 1092 { 1093 mutex_enter(&zcw->zcw_lock); 1094 ASSERT(!list_link_active(&zcw->zcw_node)); 1095 ASSERT3P(zcw->zcw_lwb, ==, NULL); 1096 list_insert_tail(nolwb, zcw); 1097 mutex_exit(&zcw->zcw_lock); 1098 } 1099 1100 void 1101 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp) 1102 { 1103 avl_tree_t *t = &lwb->lwb_vdev_tree; 1104 avl_index_t where; 1105 zil_vdev_node_t *zv, zvsearch; 1106 int ndvas = BP_GET_NDVAS(bp); 1107 int i; 1108 1109 if (zil_nocacheflush) 1110 return; 1111 1112 mutex_enter(&lwb->lwb_vdev_lock); 1113 for (i = 0; i < ndvas; i++) { 1114 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 1115 if (avl_find(t, &zvsearch, &where) == NULL) { 1116 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 1117 zv->zv_vdev = zvsearch.zv_vdev; 1118 avl_insert(t, zv, where); 1119 } 1120 } 1121 mutex_exit(&lwb->lwb_vdev_lock); 1122 } 1123 1124 static void 1125 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb) 1126 { 1127 avl_tree_t *src = &lwb->lwb_vdev_tree; 1128 avl_tree_t *dst = &nlwb->lwb_vdev_tree; 1129 void *cookie = NULL; 1130 zil_vdev_node_t *zv; 1131 1132 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); 1133 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 1134 ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 1135 1136 /* 1137 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does 1138 * not need the protection of lwb_vdev_lock (it will only be modified 1139 * while holding zilog->zl_lock) as its writes and those of its 1140 * children have all completed. The younger 'nlwb' may be waiting on 1141 * future writes to additional vdevs. 1142 */ 1143 mutex_enter(&nlwb->lwb_vdev_lock); 1144 /* 1145 * Tear down the 'lwb' vdev tree, ensuring that entries which do not 1146 * exist in 'nlwb' are moved to it, freeing any would-be duplicates. 1147 */ 1148 while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) { 1149 avl_index_t where; 1150 1151 if (avl_find(dst, zv, &where) == NULL) { 1152 avl_insert(dst, zv, where); 1153 } else { 1154 kmem_free(zv, sizeof (*zv)); 1155 } 1156 } 1157 mutex_exit(&nlwb->lwb_vdev_lock); 1158 } 1159 1160 void 1161 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg) 1162 { 1163 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1164 } 1165 1166 /* 1167 * This function is a called after all vdevs associated with a given lwb 1168 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon 1169 * as the lwb write completes, if "zil_nocacheflush" is set. Further, 1170 * all "previous" lwb's will have completed before this function is 1171 * called; i.e. this function is called for all previous lwbs before 1172 * it's called for "this" lwb (enforced via zio the dependencies 1173 * configured in zil_lwb_set_zio_dependency()). 1174 * 1175 * The intention is for this function to be called as soon as the 1176 * contents of an lwb are considered "stable" on disk, and will survive 1177 * any sudden loss of power. At this point, any threads waiting for the 1178 * lwb to reach this state are signalled, and the "waiter" structures 1179 * are marked "done". 1180 */ 1181 static void 1182 zil_lwb_flush_vdevs_done(zio_t *zio) 1183 { 1184 lwb_t *lwb = zio->io_private; 1185 zilog_t *zilog = lwb->lwb_zilog; 1186 zil_commit_waiter_t *zcw; 1187 itx_t *itx; 1188 uint64_t txg; 1189 1190 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb); 1191 1192 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1193 1194 mutex_enter(&zilog->zl_lock); 1195 1196 /* 1197 * If we have had an allocation failure and the txg is 1198 * waiting to sync then we want zil_sync() to remove the lwb so 1199 * that it's not picked up as the next new one in 1200 * zil_process_commit_list(). zil_sync() will only remove the 1201 * lwb if lwb_buf is null. 1202 */ 1203 lwb->lwb_buf = NULL; 1204 1205 ASSERT3U(lwb->lwb_issued_timestamp, >, 0); 1206 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp; 1207 1208 lwb->lwb_root_zio = NULL; 1209 1210 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); 1211 lwb->lwb_state = LWB_STATE_FLUSH_DONE; 1212 1213 if (zilog->zl_last_lwb_opened == lwb) { 1214 /* 1215 * Remember the highest committed log sequence number 1216 * for ztest. We only update this value when all the log 1217 * writes succeeded, because ztest wants to ASSERT that 1218 * it got the whole log chain. 1219 */ 1220 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1221 } 1222 1223 while ((itx = list_head(&lwb->lwb_itxs)) != NULL) { 1224 list_remove(&lwb->lwb_itxs, itx); 1225 zil_itx_destroy(itx); 1226 } 1227 1228 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) { 1229 mutex_enter(&zcw->zcw_lock); 1230 1231 ASSERT(list_link_active(&zcw->zcw_node)); 1232 list_remove(&lwb->lwb_waiters, zcw); 1233 1234 ASSERT3P(zcw->zcw_lwb, ==, lwb); 1235 zcw->zcw_lwb = NULL; 1236 /* 1237 * We expect any ZIO errors from child ZIOs to have been 1238 * propagated "up" to this specific LWB's root ZIO, in 1239 * order for this error handling to work correctly. This 1240 * includes ZIO errors from either this LWB's write or 1241 * flush, as well as any errors from other dependent LWBs 1242 * (e.g. a root LWB ZIO that might be a child of this LWB). 1243 * 1244 * With that said, it's important to note that LWB flush 1245 * errors are not propagated up to the LWB root ZIO. 1246 * This is incorrect behavior, and results in VDEV flush 1247 * errors not being handled correctly here. See the 1248 * comment above the call to "zio_flush" for details. 1249 */ 1250 1251 zcw->zcw_zio_error = zio->io_error; 1252 1253 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 1254 zcw->zcw_done = B_TRUE; 1255 cv_broadcast(&zcw->zcw_cv); 1256 1257 mutex_exit(&zcw->zcw_lock); 1258 } 1259 1260 mutex_exit(&zilog->zl_lock); 1261 1262 mutex_enter(&zilog->zl_lwb_io_lock); 1263 txg = lwb->lwb_issued_txg; 1264 ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0); 1265 zilog->zl_lwb_inflight[txg & TXG_MASK]--; 1266 if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0) 1267 cv_broadcast(&zilog->zl_lwb_io_cv); 1268 mutex_exit(&zilog->zl_lwb_io_lock); 1269 } 1270 1271 /* 1272 * Wait for the completion of all issued write/flush of that txg provided. 1273 * It guarantees zil_lwb_flush_vdevs_done() is called and returned. 1274 */ 1275 static void 1276 zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg) 1277 { 1278 ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa)); 1279 1280 mutex_enter(&zilog->zl_lwb_io_lock); 1281 while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0) 1282 cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock); 1283 mutex_exit(&zilog->zl_lwb_io_lock); 1284 1285 #ifdef ZFS_DEBUG 1286 mutex_enter(&zilog->zl_lock); 1287 mutex_enter(&zilog->zl_lwb_io_lock); 1288 lwb_t *lwb = list_head(&zilog->zl_lwb_list); 1289 while (lwb != NULL && lwb->lwb_max_txg <= txg) { 1290 if (lwb->lwb_issued_txg <= txg) { 1291 ASSERT(lwb->lwb_state != LWB_STATE_ISSUED); 1292 ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE); 1293 IMPLY(lwb->lwb_issued_txg > 0, 1294 lwb->lwb_state == LWB_STATE_FLUSH_DONE); 1295 } 1296 IMPLY(lwb->lwb_state == LWB_STATE_FLUSH_DONE, 1297 lwb->lwb_buf == NULL); 1298 lwb = list_next(&zilog->zl_lwb_list, lwb); 1299 } 1300 mutex_exit(&zilog->zl_lwb_io_lock); 1301 mutex_exit(&zilog->zl_lock); 1302 #endif 1303 } 1304 1305 /* 1306 * This is called when an lwb's write zio completes. The callback's 1307 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs 1308 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved 1309 * in writing out this specific lwb's data, and in the case that cache 1310 * flushes have been deferred, vdevs involved in writing the data for 1311 * previous lwbs. The writes corresponding to all the vdevs in the 1312 * lwb_vdev_tree will have completed by the time this is called, due to 1313 * the zio dependencies configured in zil_lwb_set_zio_dependency(), 1314 * which takes deferred flushes into account. The lwb will be "done" 1315 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio 1316 * completion callback for the lwb's root zio. 1317 */ 1318 static void 1319 zil_lwb_write_done(zio_t *zio) 1320 { 1321 lwb_t *lwb = zio->io_private; 1322 spa_t *spa = zio->io_spa; 1323 zilog_t *zilog = lwb->lwb_zilog; 1324 avl_tree_t *t = &lwb->lwb_vdev_tree; 1325 void *cookie = NULL; 1326 zil_vdev_node_t *zv; 1327 lwb_t *nlwb; 1328 1329 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0); 1330 1331 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1332 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 1333 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 1334 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 1335 ASSERT(!BP_IS_GANG(zio->io_bp)); 1336 ASSERT(!BP_IS_HOLE(zio->io_bp)); 1337 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 1338 1339 abd_free(zio->io_abd); 1340 1341 mutex_enter(&zilog->zl_lock); 1342 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED); 1343 lwb->lwb_state = LWB_STATE_WRITE_DONE; 1344 lwb->lwb_write_zio = NULL; 1345 lwb->lwb_fastwrite = FALSE; 1346 nlwb = list_next(&zilog->zl_lwb_list, lwb); 1347 mutex_exit(&zilog->zl_lock); 1348 1349 if (avl_numnodes(t) == 0) 1350 return; 1351 1352 /* 1353 * If there was an IO error, we're not going to call zio_flush() 1354 * on these vdevs, so we simply empty the tree and free the 1355 * nodes. We avoid calling zio_flush() since there isn't any 1356 * good reason for doing so, after the lwb block failed to be 1357 * written out. 1358 * 1359 * Additionally, we don't perform any further error handling at 1360 * this point (e.g. setting "zcw_zio_error" appropriately), as 1361 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus, 1362 * we expect any error seen here, to have been propagated to 1363 * that function). 1364 */ 1365 if (zio->io_error != 0) { 1366 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) 1367 kmem_free(zv, sizeof (*zv)); 1368 return; 1369 } 1370 1371 /* 1372 * If this lwb does not have any threads waiting for it to 1373 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE 1374 * command to the vdevs written to by "this" lwb, and instead 1375 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE 1376 * command for those vdevs. Thus, we merge the vdev tree of 1377 * "this" lwb with the vdev tree of the "next" lwb in the list, 1378 * and assume the "next" lwb will handle flushing the vdevs (or 1379 * deferring the flush(s) again). 1380 * 1381 * This is a useful performance optimization, especially for 1382 * workloads with lots of async write activity and few sync 1383 * write and/or fsync activity, as it has the potential to 1384 * coalesce multiple flush commands to a vdev into one. 1385 */ 1386 if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) { 1387 zil_lwb_flush_defer(lwb, nlwb); 1388 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 1389 return; 1390 } 1391 1392 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 1393 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 1394 if (vd != NULL) { 1395 /* 1396 * The "ZIO_FLAG_DONT_PROPAGATE" is currently 1397 * always used within "zio_flush". This means, 1398 * any errors when flushing the vdev(s), will 1399 * (unfortunately) not be handled correctly, 1400 * since these "zio_flush" errors will not be 1401 * propagated up to "zil_lwb_flush_vdevs_done". 1402 */ 1403 zio_flush(lwb->lwb_root_zio, vd); 1404 } 1405 kmem_free(zv, sizeof (*zv)); 1406 } 1407 } 1408 1409 static void 1410 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb) 1411 { 1412 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened; 1413 1414 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1415 ASSERT(MUTEX_HELD(&zilog->zl_lock)); 1416 1417 /* 1418 * The zilog's "zl_last_lwb_opened" field is used to build the 1419 * lwb/zio dependency chain, which is used to preserve the 1420 * ordering of lwb completions that is required by the semantics 1421 * of the ZIL. Each new lwb zio becomes a parent of the 1422 * "previous" lwb zio, such that the new lwb's zio cannot 1423 * complete until the "previous" lwb's zio completes. 1424 * 1425 * This is required by the semantics of zil_commit(); the commit 1426 * waiters attached to the lwbs will be woken in the lwb zio's 1427 * completion callback, so this zio dependency graph ensures the 1428 * waiters are woken in the correct order (the same order the 1429 * lwbs were created). 1430 */ 1431 if (last_lwb_opened != NULL && 1432 last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) { 1433 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1434 last_lwb_opened->lwb_state == LWB_STATE_ISSUED || 1435 last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE); 1436 1437 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL); 1438 zio_add_child(lwb->lwb_root_zio, 1439 last_lwb_opened->lwb_root_zio); 1440 1441 /* 1442 * If the previous lwb's write hasn't already completed, 1443 * we also want to order the completion of the lwb write 1444 * zios (above, we only order the completion of the lwb 1445 * root zios). This is required because of how we can 1446 * defer the DKIOCFLUSHWRITECACHE commands for each lwb. 1447 * 1448 * When the DKIOCFLUSHWRITECACHE commands are deferred, 1449 * the previous lwb will rely on this lwb to flush the 1450 * vdevs written to by that previous lwb. Thus, we need 1451 * to ensure this lwb doesn't issue the flush until 1452 * after the previous lwb's write completes. We ensure 1453 * this ordering by setting the zio parent/child 1454 * relationship here. 1455 * 1456 * Without this relationship on the lwb's write zio, 1457 * it's possible for this lwb's write to complete prior 1458 * to the previous lwb's write completing; and thus, the 1459 * vdevs for the previous lwb would be flushed prior to 1460 * that lwb's data being written to those vdevs (the 1461 * vdevs are flushed in the lwb write zio's completion 1462 * handler, zil_lwb_write_done()). 1463 */ 1464 if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) { 1465 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1466 last_lwb_opened->lwb_state == LWB_STATE_ISSUED); 1467 1468 ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL); 1469 zio_add_child(lwb->lwb_write_zio, 1470 last_lwb_opened->lwb_write_zio); 1471 } 1472 } 1473 } 1474 1475 1476 /* 1477 * This function's purpose is to "open" an lwb such that it is ready to 1478 * accept new itxs being committed to it. To do this, the lwb's zio 1479 * structures are created, and linked to the lwb. This function is 1480 * idempotent; if the passed in lwb has already been opened, this 1481 * function is essentially a no-op. 1482 */ 1483 static void 1484 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb) 1485 { 1486 zbookmark_phys_t zb; 1487 zio_priority_t prio; 1488 1489 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1490 ASSERT3P(lwb, !=, NULL); 1491 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED); 1492 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED); 1493 1494 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 1495 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 1496 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 1497 1498 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */ 1499 mutex_enter(&zilog->zl_lock); 1500 if (lwb->lwb_root_zio == NULL) { 1501 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, 1502 BP_GET_LSIZE(&lwb->lwb_blk)); 1503 1504 if (!lwb->lwb_fastwrite) { 1505 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk); 1506 lwb->lwb_fastwrite = 1; 1507 } 1508 1509 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk) 1510 prio = ZIO_PRIORITY_SYNC_WRITE; 1511 else 1512 prio = ZIO_PRIORITY_ASYNC_WRITE; 1513 1514 lwb->lwb_root_zio = zio_root(zilog->zl_spa, 1515 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL); 1516 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1517 1518 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, 1519 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd, 1520 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb, 1521 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb); 1522 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1523 1524 lwb->lwb_state = LWB_STATE_OPENED; 1525 1526 zil_lwb_set_zio_dependency(zilog, lwb); 1527 zilog->zl_last_lwb_opened = lwb; 1528 } 1529 mutex_exit(&zilog->zl_lock); 1530 1531 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1532 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1533 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1534 } 1535 1536 /* 1537 * Define a limited set of intent log block sizes. 1538 * 1539 * These must be a multiple of 4KB. Note only the amount used (again 1540 * aligned to 4KB) actually gets written. However, we can't always just 1541 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 1542 */ 1543 static const struct { 1544 uint64_t limit; 1545 uint64_t blksz; 1546 } zil_block_buckets[] = { 1547 { 4096, 4096 }, /* non TX_WRITE */ 1548 { 8192 + 4096, 8192 + 4096 }, /* database */ 1549 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */ 1550 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */ 1551 { 131072, 131072 }, /* < 128KB writes */ 1552 { 131072 +4096, 65536 + 4096 }, /* 128KB writes */ 1553 { UINT64_MAX, SPA_OLD_MAXBLOCKSIZE}, /* > 128KB writes */ 1554 }; 1555 1556 /* 1557 * Maximum block size used by the ZIL. This is picked up when the ZIL is 1558 * initialized. Otherwise this should not be used directly; see 1559 * zl_max_block_size instead. 1560 */ 1561 static int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE; 1562 1563 /* 1564 * Start a log block write and advance to the next log block. 1565 * Calls are serialized. 1566 */ 1567 static lwb_t * 1568 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb) 1569 { 1570 lwb_t *nlwb = NULL; 1571 zil_chain_t *zilc; 1572 spa_t *spa = zilog->zl_spa; 1573 blkptr_t *bp; 1574 dmu_tx_t *tx; 1575 uint64_t txg; 1576 uint64_t zil_blksz, wsz; 1577 int i, error; 1578 boolean_t slog; 1579 1580 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1581 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1582 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1583 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1584 1585 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1586 zilc = (zil_chain_t *)lwb->lwb_buf; 1587 bp = &zilc->zc_next_blk; 1588 } else { 1589 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 1590 bp = &zilc->zc_next_blk; 1591 } 1592 1593 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 1594 1595 /* 1596 * Allocate the next block and save its address in this block 1597 * before writing it in order to establish the log chain. 1598 */ 1599 1600 tx = dmu_tx_create(zilog->zl_os); 1601 1602 /* 1603 * Since we are not going to create any new dirty data, and we 1604 * can even help with clearing the existing dirty data, we 1605 * should not be subject to the dirty data based delays. We 1606 * use TXG_NOTHROTTLE to bypass the delay mechanism. 1607 */ 1608 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); 1609 1610 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1611 txg = dmu_tx_get_txg(tx); 1612 1613 mutex_enter(&zilog->zl_lwb_io_lock); 1614 lwb->lwb_issued_txg = txg; 1615 zilog->zl_lwb_inflight[txg & TXG_MASK]++; 1616 zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg); 1617 mutex_exit(&zilog->zl_lwb_io_lock); 1618 1619 /* 1620 * Log blocks are pre-allocated. Here we select the size of the next 1621 * block, based on size used in the last block. 1622 * - first find the smallest bucket that will fit the block from a 1623 * limited set of block sizes. This is because it's faster to write 1624 * blocks allocated from the same metaslab as they are adjacent or 1625 * close. 1626 * - next find the maximum from the new suggested size and an array of 1627 * previous sizes. This lessens a picket fence effect of wrongly 1628 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k 1629 * requests. 1630 * 1631 * Note we only write what is used, but we can't just allocate 1632 * the maximum block size because we can exhaust the available 1633 * pool log space. 1634 */ 1635 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1636 for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++) 1637 continue; 1638 zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size); 1639 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1640 for (i = 0; i < ZIL_PREV_BLKS; i++) 1641 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1642 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1643 1644 BP_ZERO(bp); 1645 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog); 1646 if (slog) { 1647 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count); 1648 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused); 1649 } else { 1650 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count); 1651 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused); 1652 } 1653 if (error == 0) { 1654 ASSERT3U(bp->blk_birth, ==, txg); 1655 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1656 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1657 1658 /* 1659 * Allocate a new log write block (lwb). 1660 */ 1661 nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE); 1662 } 1663 1664 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1665 /* For Slim ZIL only write what is used. */ 1666 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1667 ASSERT3U(wsz, <=, lwb->lwb_sz); 1668 zio_shrink(lwb->lwb_write_zio, wsz); 1669 1670 } else { 1671 wsz = lwb->lwb_sz; 1672 } 1673 1674 zilc->zc_pad = 0; 1675 zilc->zc_nused = lwb->lwb_nused; 1676 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1677 1678 /* 1679 * clear unused data for security 1680 */ 1681 memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused); 1682 1683 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER); 1684 1685 zil_lwb_add_block(lwb, &lwb->lwb_blk); 1686 lwb->lwb_issued_timestamp = gethrtime(); 1687 lwb->lwb_state = LWB_STATE_ISSUED; 1688 1689 zio_nowait(lwb->lwb_root_zio); 1690 zio_nowait(lwb->lwb_write_zio); 1691 1692 dmu_tx_commit(tx); 1693 1694 /* 1695 * If there was an allocation failure then nlwb will be null which 1696 * forces a txg_wait_synced(). 1697 */ 1698 return (nlwb); 1699 } 1700 1701 /* 1702 * Maximum amount of write data that can be put into single log block. 1703 */ 1704 uint64_t 1705 zil_max_log_data(zilog_t *zilog) 1706 { 1707 return (zilog->zl_max_block_size - 1708 sizeof (zil_chain_t) - sizeof (lr_write_t)); 1709 } 1710 1711 /* 1712 * Maximum amount of log space we agree to waste to reduce number of 1713 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%). 1714 */ 1715 static inline uint64_t 1716 zil_max_waste_space(zilog_t *zilog) 1717 { 1718 return (zil_max_log_data(zilog) / 8); 1719 } 1720 1721 /* 1722 * Maximum amount of write data for WR_COPIED. For correctness, consumers 1723 * must fall back to WR_NEED_COPY if we can't fit the entire record into one 1724 * maximum sized log block, because each WR_COPIED record must fit in a 1725 * single log block. For space efficiency, we want to fit two records into a 1726 * max-sized log block. 1727 */ 1728 uint64_t 1729 zil_max_copied_data(zilog_t *zilog) 1730 { 1731 return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 - 1732 sizeof (lr_write_t)); 1733 } 1734 1735 static lwb_t * 1736 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1737 { 1738 lr_t *lrcb, *lrc; 1739 lr_write_t *lrwb, *lrw; 1740 char *lr_buf; 1741 uint64_t dlen, dnow, dpad, lwb_sp, reclen, txg, max_log_data; 1742 1743 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1744 ASSERT3P(lwb, !=, NULL); 1745 ASSERT3P(lwb->lwb_buf, !=, NULL); 1746 1747 zil_lwb_write_open(zilog, lwb); 1748 1749 lrc = &itx->itx_lr; 1750 lrw = (lr_write_t *)lrc; 1751 1752 /* 1753 * A commit itx doesn't represent any on-disk state; instead 1754 * it's simply used as a place holder on the commit list, and 1755 * provides a mechanism for attaching a "commit waiter" onto the 1756 * correct lwb (such that the waiter can be signalled upon 1757 * completion of that lwb). Thus, we don't process this itx's 1758 * log record if it's a commit itx (these itx's don't have log 1759 * records), and instead link the itx's waiter onto the lwb's 1760 * list of waiters. 1761 * 1762 * For more details, see the comment above zil_commit(). 1763 */ 1764 if (lrc->lrc_txtype == TX_COMMIT) { 1765 mutex_enter(&zilog->zl_lock); 1766 zil_commit_waiter_link_lwb(itx->itx_private, lwb); 1767 itx->itx_private = NULL; 1768 mutex_exit(&zilog->zl_lock); 1769 return (lwb); 1770 } 1771 1772 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { 1773 dlen = P2ROUNDUP_TYPED( 1774 lrw->lr_length, sizeof (uint64_t), uint64_t); 1775 dpad = dlen - lrw->lr_length; 1776 } else { 1777 dlen = dpad = 0; 1778 } 1779 reclen = lrc->lrc_reclen; 1780 zilog->zl_cur_used += (reclen + dlen); 1781 txg = lrc->lrc_txg; 1782 1783 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen)); 1784 1785 cont: 1786 /* 1787 * If this record won't fit in the current log block, start a new one. 1788 * For WR_NEED_COPY optimize layout for minimal number of chunks. 1789 */ 1790 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1791 max_log_data = zil_max_log_data(zilog); 1792 if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1793 lwb_sp < zil_max_waste_space(zilog) && 1794 (dlen % max_log_data == 0 || 1795 lwb_sp < reclen + dlen % max_log_data))) { 1796 lwb = zil_lwb_write_issue(zilog, lwb); 1797 if (lwb == NULL) 1798 return (NULL); 1799 zil_lwb_write_open(zilog, lwb); 1800 ASSERT(LWB_EMPTY(lwb)); 1801 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1802 1803 /* 1804 * There must be enough space in the new, empty log block to 1805 * hold reclen. For WR_COPIED, we need to fit the whole 1806 * record in one block, and reclen is the header size + the 1807 * data size. For WR_NEED_COPY, we can create multiple 1808 * records, splitting the data into multiple blocks, so we 1809 * only need to fit one word of data per block; in this case 1810 * reclen is just the header size (no data). 1811 */ 1812 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp); 1813 } 1814 1815 dnow = MIN(dlen, lwb_sp - reclen); 1816 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1817 memcpy(lr_buf, lrc, reclen); 1818 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */ 1819 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */ 1820 1821 ZIL_STAT_BUMP(zil_itx_count); 1822 1823 /* 1824 * If it's a write, fetch the data or get its blkptr as appropriate. 1825 */ 1826 if (lrc->lrc_txtype == TX_WRITE) { 1827 if (txg > spa_freeze_txg(zilog->zl_spa)) 1828 txg_wait_synced(zilog->zl_dmu_pool, txg); 1829 if (itx->itx_wr_state == WR_COPIED) { 1830 ZIL_STAT_BUMP(zil_itx_copied_count); 1831 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length); 1832 } else { 1833 char *dbuf; 1834 int error; 1835 1836 if (itx->itx_wr_state == WR_NEED_COPY) { 1837 dbuf = lr_buf + reclen; 1838 lrcb->lrc_reclen += dnow; 1839 if (lrwb->lr_length > dnow) 1840 lrwb->lr_length = dnow; 1841 lrw->lr_offset += dnow; 1842 lrw->lr_length -= dnow; 1843 ZIL_STAT_BUMP(zil_itx_needcopy_count); 1844 ZIL_STAT_INCR(zil_itx_needcopy_bytes, dnow); 1845 } else { 1846 ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT); 1847 dbuf = NULL; 1848 ZIL_STAT_BUMP(zil_itx_indirect_count); 1849 ZIL_STAT_INCR(zil_itx_indirect_bytes, 1850 lrw->lr_length); 1851 } 1852 1853 /* 1854 * We pass in the "lwb_write_zio" rather than 1855 * "lwb_root_zio" so that the "lwb_write_zio" 1856 * becomes the parent of any zio's created by 1857 * the "zl_get_data" callback. The vdevs are 1858 * flushed after the "lwb_write_zio" completes, 1859 * so we want to make sure that completion 1860 * callback waits for these additional zio's, 1861 * such that the vdevs used by those zio's will 1862 * be included in the lwb's vdev tree, and those 1863 * vdevs will be properly flushed. If we passed 1864 * in "lwb_root_zio" here, then these additional 1865 * vdevs may not be flushed; e.g. if these zio's 1866 * completed after "lwb_write_zio" completed. 1867 */ 1868 error = zilog->zl_get_data(itx->itx_private, 1869 itx->itx_gen, lrwb, dbuf, lwb, 1870 lwb->lwb_write_zio); 1871 if (dbuf != NULL && error == 0 && dnow == dlen) 1872 /* Zero any padding bytes in the last block. */ 1873 memset((char *)dbuf + lrwb->lr_length, 0, dpad); 1874 1875 if (error == EIO) { 1876 txg_wait_synced(zilog->zl_dmu_pool, txg); 1877 return (lwb); 1878 } 1879 if (error != 0) { 1880 ASSERT(error == ENOENT || error == EEXIST || 1881 error == EALREADY); 1882 return (lwb); 1883 } 1884 } 1885 } 1886 1887 /* 1888 * We're actually making an entry, so update lrc_seq to be the 1889 * log record sequence number. Note that this is generally not 1890 * equal to the itx sequence number because not all transactions 1891 * are synchronous, and sometimes spa_sync() gets there first. 1892 */ 1893 lrcb->lrc_seq = ++zilog->zl_lr_seq; 1894 lwb->lwb_nused += reclen + dnow; 1895 1896 zil_lwb_add_txg(lwb, txg); 1897 1898 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1899 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1900 1901 dlen -= dnow; 1902 if (dlen > 0) { 1903 zilog->zl_cur_used += reclen; 1904 goto cont; 1905 } 1906 1907 return (lwb); 1908 } 1909 1910 itx_t * 1911 zil_itx_create(uint64_t txtype, size_t olrsize) 1912 { 1913 size_t itxsize, lrsize; 1914 itx_t *itx; 1915 1916 lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t); 1917 itxsize = offsetof(itx_t, itx_lr) + lrsize; 1918 1919 itx = zio_data_buf_alloc(itxsize); 1920 itx->itx_lr.lrc_txtype = txtype; 1921 itx->itx_lr.lrc_reclen = lrsize; 1922 itx->itx_lr.lrc_seq = 0; /* defensive */ 1923 memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize); 1924 itx->itx_sync = B_TRUE; /* default is synchronous */ 1925 itx->itx_callback = NULL; 1926 itx->itx_callback_data = NULL; 1927 itx->itx_size = itxsize; 1928 1929 return (itx); 1930 } 1931 1932 void 1933 zil_itx_destroy(itx_t *itx) 1934 { 1935 IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL); 1936 IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 1937 1938 if (itx->itx_callback != NULL) 1939 itx->itx_callback(itx->itx_callback_data); 1940 1941 zio_data_buf_free(itx, itx->itx_size); 1942 } 1943 1944 /* 1945 * Free up the sync and async itxs. The itxs_t has already been detached 1946 * so no locks are needed. 1947 */ 1948 static void 1949 zil_itxg_clean(void *arg) 1950 { 1951 itx_t *itx; 1952 list_t *list; 1953 avl_tree_t *t; 1954 void *cookie; 1955 itxs_t *itxs = arg; 1956 itx_async_node_t *ian; 1957 1958 list = &itxs->i_sync_list; 1959 while ((itx = list_head(list)) != NULL) { 1960 /* 1961 * In the general case, commit itxs will not be found 1962 * here, as they'll be committed to an lwb via 1963 * zil_lwb_commit(), and free'd in that function. Having 1964 * said that, it is still possible for commit itxs to be 1965 * found here, due to the following race: 1966 * 1967 * - a thread calls zil_commit() which assigns the 1968 * commit itx to a per-txg i_sync_list 1969 * - zil_itxg_clean() is called (e.g. via spa_sync()) 1970 * while the waiter is still on the i_sync_list 1971 * 1972 * There's nothing to prevent syncing the txg while the 1973 * waiter is on the i_sync_list. This normally doesn't 1974 * happen because spa_sync() is slower than zil_commit(), 1975 * but if zil_commit() calls txg_wait_synced() (e.g. 1976 * because zil_create() or zil_commit_writer_stall() is 1977 * called) we will hit this case. 1978 */ 1979 if (itx->itx_lr.lrc_txtype == TX_COMMIT) 1980 zil_commit_waiter_skip(itx->itx_private); 1981 1982 list_remove(list, itx); 1983 zil_itx_destroy(itx); 1984 } 1985 1986 cookie = NULL; 1987 t = &itxs->i_async_tree; 1988 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1989 list = &ian->ia_list; 1990 while ((itx = list_head(list)) != NULL) { 1991 list_remove(list, itx); 1992 /* commit itxs should never be on the async lists. */ 1993 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 1994 zil_itx_destroy(itx); 1995 } 1996 list_destroy(list); 1997 kmem_free(ian, sizeof (itx_async_node_t)); 1998 } 1999 avl_destroy(t); 2000 2001 kmem_free(itxs, sizeof (itxs_t)); 2002 } 2003 2004 static int 2005 zil_aitx_compare(const void *x1, const void *x2) 2006 { 2007 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 2008 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 2009 2010 return (TREE_CMP(o1, o2)); 2011 } 2012 2013 /* 2014 * Remove all async itx with the given oid. 2015 */ 2016 void 2017 zil_remove_async(zilog_t *zilog, uint64_t oid) 2018 { 2019 uint64_t otxg, txg; 2020 itx_async_node_t *ian; 2021 avl_tree_t *t; 2022 avl_index_t where; 2023 list_t clean_list; 2024 itx_t *itx; 2025 2026 ASSERT(oid != 0); 2027 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 2028 2029 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2030 otxg = ZILTEST_TXG; 2031 else 2032 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2033 2034 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2035 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2036 2037 mutex_enter(&itxg->itxg_lock); 2038 if (itxg->itxg_txg != txg) { 2039 mutex_exit(&itxg->itxg_lock); 2040 continue; 2041 } 2042 2043 /* 2044 * Locate the object node and append its list. 2045 */ 2046 t = &itxg->itxg_itxs->i_async_tree; 2047 ian = avl_find(t, &oid, &where); 2048 if (ian != NULL) 2049 list_move_tail(&clean_list, &ian->ia_list); 2050 mutex_exit(&itxg->itxg_lock); 2051 } 2052 while ((itx = list_head(&clean_list)) != NULL) { 2053 list_remove(&clean_list, itx); 2054 /* commit itxs should never be on the async lists. */ 2055 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 2056 zil_itx_destroy(itx); 2057 } 2058 list_destroy(&clean_list); 2059 } 2060 2061 void 2062 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 2063 { 2064 uint64_t txg; 2065 itxg_t *itxg; 2066 itxs_t *itxs, *clean = NULL; 2067 2068 /* 2069 * Ensure the data of a renamed file is committed before the rename. 2070 */ 2071 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 2072 zil_async_to_sync(zilog, itx->itx_oid); 2073 2074 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 2075 txg = ZILTEST_TXG; 2076 else 2077 txg = dmu_tx_get_txg(tx); 2078 2079 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2080 mutex_enter(&itxg->itxg_lock); 2081 itxs = itxg->itxg_itxs; 2082 if (itxg->itxg_txg != txg) { 2083 if (itxs != NULL) { 2084 /* 2085 * The zil_clean callback hasn't got around to cleaning 2086 * this itxg. Save the itxs for release below. 2087 * This should be rare. 2088 */ 2089 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for " 2090 "txg %llu", (u_longlong_t)itxg->itxg_txg); 2091 clean = itxg->itxg_itxs; 2092 } 2093 itxg->itxg_txg = txg; 2094 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), 2095 KM_SLEEP); 2096 2097 list_create(&itxs->i_sync_list, sizeof (itx_t), 2098 offsetof(itx_t, itx_node)); 2099 avl_create(&itxs->i_async_tree, zil_aitx_compare, 2100 sizeof (itx_async_node_t), 2101 offsetof(itx_async_node_t, ia_node)); 2102 } 2103 if (itx->itx_sync) { 2104 list_insert_tail(&itxs->i_sync_list, itx); 2105 } else { 2106 avl_tree_t *t = &itxs->i_async_tree; 2107 uint64_t foid = 2108 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid); 2109 itx_async_node_t *ian; 2110 avl_index_t where; 2111 2112 ian = avl_find(t, &foid, &where); 2113 if (ian == NULL) { 2114 ian = kmem_alloc(sizeof (itx_async_node_t), 2115 KM_SLEEP); 2116 list_create(&ian->ia_list, sizeof (itx_t), 2117 offsetof(itx_t, itx_node)); 2118 ian->ia_foid = foid; 2119 avl_insert(t, ian, where); 2120 } 2121 list_insert_tail(&ian->ia_list, itx); 2122 } 2123 2124 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 2125 2126 /* 2127 * We don't want to dirty the ZIL using ZILTEST_TXG, because 2128 * zil_clean() will never be called using ZILTEST_TXG. Thus, we 2129 * need to be careful to always dirty the ZIL using the "real" 2130 * TXG (not itxg_txg) even when the SPA is frozen. 2131 */ 2132 zilog_dirty(zilog, dmu_tx_get_txg(tx)); 2133 mutex_exit(&itxg->itxg_lock); 2134 2135 /* Release the old itxs now we've dropped the lock */ 2136 if (clean != NULL) 2137 zil_itxg_clean(clean); 2138 } 2139 2140 /* 2141 * If there are any in-memory intent log transactions which have now been 2142 * synced then start up a taskq to free them. We should only do this after we 2143 * have written out the uberblocks (i.e. txg has been committed) so that 2144 * don't inadvertently clean out in-memory log records that would be required 2145 * by zil_commit(). 2146 */ 2147 void 2148 zil_clean(zilog_t *zilog, uint64_t synced_txg) 2149 { 2150 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 2151 itxs_t *clean_me; 2152 2153 ASSERT3U(synced_txg, <, ZILTEST_TXG); 2154 2155 mutex_enter(&itxg->itxg_lock); 2156 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 2157 mutex_exit(&itxg->itxg_lock); 2158 return; 2159 } 2160 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 2161 ASSERT3U(itxg->itxg_txg, !=, 0); 2162 clean_me = itxg->itxg_itxs; 2163 itxg->itxg_itxs = NULL; 2164 itxg->itxg_txg = 0; 2165 mutex_exit(&itxg->itxg_lock); 2166 /* 2167 * Preferably start a task queue to free up the old itxs but 2168 * if taskq_dispatch can't allocate resources to do that then 2169 * free it in-line. This should be rare. Note, using TQ_SLEEP 2170 * created a bad performance problem. 2171 */ 2172 ASSERT3P(zilog->zl_dmu_pool, !=, NULL); 2173 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL); 2174 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq, 2175 zil_itxg_clean, clean_me, TQ_NOSLEEP); 2176 if (id == TASKQID_INVALID) 2177 zil_itxg_clean(clean_me); 2178 } 2179 2180 /* 2181 * This function will traverse the queue of itxs that need to be 2182 * committed, and move them onto the ZIL's zl_itx_commit_list. 2183 */ 2184 static void 2185 zil_get_commit_list(zilog_t *zilog) 2186 { 2187 uint64_t otxg, txg; 2188 list_t *commit_list = &zilog->zl_itx_commit_list; 2189 2190 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2191 2192 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2193 otxg = ZILTEST_TXG; 2194 else 2195 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2196 2197 /* 2198 * This is inherently racy, since there is nothing to prevent 2199 * the last synced txg from changing. That's okay since we'll 2200 * only commit things in the future. 2201 */ 2202 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2203 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2204 2205 mutex_enter(&itxg->itxg_lock); 2206 if (itxg->itxg_txg != txg) { 2207 mutex_exit(&itxg->itxg_lock); 2208 continue; 2209 } 2210 2211 /* 2212 * If we're adding itx records to the zl_itx_commit_list, 2213 * then the zil better be dirty in this "txg". We can assert 2214 * that here since we're holding the itxg_lock which will 2215 * prevent spa_sync from cleaning it. Once we add the itxs 2216 * to the zl_itx_commit_list we must commit it to disk even 2217 * if it's unnecessary (i.e. the txg was synced). 2218 */ 2219 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 2220 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 2221 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 2222 2223 mutex_exit(&itxg->itxg_lock); 2224 } 2225 } 2226 2227 /* 2228 * Move the async itxs for a specified object to commit into sync lists. 2229 */ 2230 void 2231 zil_async_to_sync(zilog_t *zilog, uint64_t foid) 2232 { 2233 uint64_t otxg, txg; 2234 itx_async_node_t *ian; 2235 avl_tree_t *t; 2236 avl_index_t where; 2237 2238 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2239 otxg = ZILTEST_TXG; 2240 else 2241 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2242 2243 /* 2244 * This is inherently racy, since there is nothing to prevent 2245 * the last synced txg from changing. 2246 */ 2247 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2248 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2249 2250 mutex_enter(&itxg->itxg_lock); 2251 if (itxg->itxg_txg != txg) { 2252 mutex_exit(&itxg->itxg_lock); 2253 continue; 2254 } 2255 2256 /* 2257 * If a foid is specified then find that node and append its 2258 * list. Otherwise walk the tree appending all the lists 2259 * to the sync list. We add to the end rather than the 2260 * beginning to ensure the create has happened. 2261 */ 2262 t = &itxg->itxg_itxs->i_async_tree; 2263 if (foid != 0) { 2264 ian = avl_find(t, &foid, &where); 2265 if (ian != NULL) { 2266 list_move_tail(&itxg->itxg_itxs->i_sync_list, 2267 &ian->ia_list); 2268 } 2269 } else { 2270 void *cookie = NULL; 2271 2272 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 2273 list_move_tail(&itxg->itxg_itxs->i_sync_list, 2274 &ian->ia_list); 2275 list_destroy(&ian->ia_list); 2276 kmem_free(ian, sizeof (itx_async_node_t)); 2277 } 2278 } 2279 mutex_exit(&itxg->itxg_lock); 2280 } 2281 } 2282 2283 /* 2284 * This function will prune commit itxs that are at the head of the 2285 * commit list (it won't prune past the first non-commit itx), and 2286 * either: a) attach them to the last lwb that's still pending 2287 * completion, or b) skip them altogether. 2288 * 2289 * This is used as a performance optimization to prevent commit itxs 2290 * from generating new lwbs when it's unnecessary to do so. 2291 */ 2292 static void 2293 zil_prune_commit_list(zilog_t *zilog) 2294 { 2295 itx_t *itx; 2296 2297 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2298 2299 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { 2300 lr_t *lrc = &itx->itx_lr; 2301 if (lrc->lrc_txtype != TX_COMMIT) 2302 break; 2303 2304 mutex_enter(&zilog->zl_lock); 2305 2306 lwb_t *last_lwb = zilog->zl_last_lwb_opened; 2307 if (last_lwb == NULL || 2308 last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) { 2309 /* 2310 * All of the itxs this waiter was waiting on 2311 * must have already completed (or there were 2312 * never any itx's for it to wait on), so it's 2313 * safe to skip this waiter and mark it done. 2314 */ 2315 zil_commit_waiter_skip(itx->itx_private); 2316 } else { 2317 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb); 2318 itx->itx_private = NULL; 2319 } 2320 2321 mutex_exit(&zilog->zl_lock); 2322 2323 list_remove(&zilog->zl_itx_commit_list, itx); 2324 zil_itx_destroy(itx); 2325 } 2326 2327 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 2328 } 2329 2330 static void 2331 zil_commit_writer_stall(zilog_t *zilog) 2332 { 2333 /* 2334 * When zio_alloc_zil() fails to allocate the next lwb block on 2335 * disk, we must call txg_wait_synced() to ensure all of the 2336 * lwbs in the zilog's zl_lwb_list are synced and then freed (in 2337 * zil_sync()), such that any subsequent ZIL writer (i.e. a call 2338 * to zil_process_commit_list()) will have to call zil_create(), 2339 * and start a new ZIL chain. 2340 * 2341 * Since zil_alloc_zil() failed, the lwb that was previously 2342 * issued does not have a pointer to the "next" lwb on disk. 2343 * Thus, if another ZIL writer thread was to allocate the "next" 2344 * on-disk lwb, that block could be leaked in the event of a 2345 * crash (because the previous lwb on-disk would not point to 2346 * it). 2347 * 2348 * We must hold the zilog's zl_issuer_lock while we do this, to 2349 * ensure no new threads enter zil_process_commit_list() until 2350 * all lwb's in the zl_lwb_list have been synced and freed 2351 * (which is achieved via the txg_wait_synced() call). 2352 */ 2353 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2354 txg_wait_synced(zilog->zl_dmu_pool, 0); 2355 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 2356 } 2357 2358 /* 2359 * This function will traverse the commit list, creating new lwbs as 2360 * needed, and committing the itxs from the commit list to these newly 2361 * created lwbs. Additionally, as a new lwb is created, the previous 2362 * lwb will be issued to the zio layer to be written to disk. 2363 */ 2364 static void 2365 zil_process_commit_list(zilog_t *zilog) 2366 { 2367 spa_t *spa = zilog->zl_spa; 2368 list_t nolwb_itxs; 2369 list_t nolwb_waiters; 2370 lwb_t *lwb; 2371 itx_t *itx; 2372 2373 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2374 2375 /* 2376 * Return if there's nothing to commit before we dirty the fs by 2377 * calling zil_create(). 2378 */ 2379 if (list_head(&zilog->zl_itx_commit_list) == NULL) 2380 return; 2381 2382 list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); 2383 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t), 2384 offsetof(zil_commit_waiter_t, zcw_node)); 2385 2386 lwb = list_tail(&zilog->zl_lwb_list); 2387 if (lwb == NULL) { 2388 lwb = zil_create(zilog); 2389 } else { 2390 /* 2391 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will 2392 * have already been created (zl_lwb_list not empty). 2393 */ 2394 zil_commit_activate_saxattr_feature(zilog); 2395 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2396 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 2397 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 2398 } 2399 2400 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { 2401 lr_t *lrc = &itx->itx_lr; 2402 uint64_t txg = lrc->lrc_txg; 2403 2404 ASSERT3U(txg, !=, 0); 2405 2406 if (lrc->lrc_txtype == TX_COMMIT) { 2407 DTRACE_PROBE2(zil__process__commit__itx, 2408 zilog_t *, zilog, itx_t *, itx); 2409 } else { 2410 DTRACE_PROBE2(zil__process__normal__itx, 2411 zilog_t *, zilog, itx_t *, itx); 2412 } 2413 2414 list_remove(&zilog->zl_itx_commit_list, itx); 2415 2416 boolean_t synced = txg <= spa_last_synced_txg(spa); 2417 boolean_t frozen = txg > spa_freeze_txg(spa); 2418 2419 /* 2420 * If the txg of this itx has already been synced out, then 2421 * we don't need to commit this itx to an lwb. This is 2422 * because the data of this itx will have already been 2423 * written to the main pool. This is inherently racy, and 2424 * it's still ok to commit an itx whose txg has already 2425 * been synced; this will result in a write that's 2426 * unnecessary, but will do no harm. 2427 * 2428 * With that said, we always want to commit TX_COMMIT itxs 2429 * to an lwb, regardless of whether or not that itx's txg 2430 * has been synced out. We do this to ensure any OPENED lwb 2431 * will always have at least one zil_commit_waiter_t linked 2432 * to the lwb. 2433 * 2434 * As a counter-example, if we skipped TX_COMMIT itx's 2435 * whose txg had already been synced, the following 2436 * situation could occur if we happened to be racing with 2437 * spa_sync: 2438 * 2439 * 1. We commit a non-TX_COMMIT itx to an lwb, where the 2440 * itx's txg is 10 and the last synced txg is 9. 2441 * 2. spa_sync finishes syncing out txg 10. 2442 * 3. We move to the next itx in the list, it's a TX_COMMIT 2443 * whose txg is 10, so we skip it rather than committing 2444 * it to the lwb used in (1). 2445 * 2446 * If the itx that is skipped in (3) is the last TX_COMMIT 2447 * itx in the commit list, than it's possible for the lwb 2448 * used in (1) to remain in the OPENED state indefinitely. 2449 * 2450 * To prevent the above scenario from occurring, ensuring 2451 * that once an lwb is OPENED it will transition to ISSUED 2452 * and eventually DONE, we always commit TX_COMMIT itx's to 2453 * an lwb here, even if that itx's txg has already been 2454 * synced. 2455 * 2456 * Finally, if the pool is frozen, we _always_ commit the 2457 * itx. The point of freezing the pool is to prevent data 2458 * from being written to the main pool via spa_sync, and 2459 * instead rely solely on the ZIL to persistently store the 2460 * data; i.e. when the pool is frozen, the last synced txg 2461 * value can't be trusted. 2462 */ 2463 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) { 2464 if (lwb != NULL) { 2465 lwb = zil_lwb_commit(zilog, itx, lwb); 2466 2467 if (lwb == NULL) 2468 list_insert_tail(&nolwb_itxs, itx); 2469 else 2470 list_insert_tail(&lwb->lwb_itxs, itx); 2471 } else { 2472 if (lrc->lrc_txtype == TX_COMMIT) { 2473 zil_commit_waiter_link_nolwb( 2474 itx->itx_private, &nolwb_waiters); 2475 } 2476 2477 list_insert_tail(&nolwb_itxs, itx); 2478 } 2479 } else { 2480 ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT); 2481 zil_itx_destroy(itx); 2482 } 2483 } 2484 2485 if (lwb == NULL) { 2486 /* 2487 * This indicates zio_alloc_zil() failed to allocate the 2488 * "next" lwb on-disk. When this happens, we must stall 2489 * the ZIL write pipeline; see the comment within 2490 * zil_commit_writer_stall() for more details. 2491 */ 2492 zil_commit_writer_stall(zilog); 2493 2494 /* 2495 * Additionally, we have to signal and mark the "nolwb" 2496 * waiters as "done" here, since without an lwb, we 2497 * can't do this via zil_lwb_flush_vdevs_done() like 2498 * normal. 2499 */ 2500 zil_commit_waiter_t *zcw; 2501 while ((zcw = list_head(&nolwb_waiters)) != NULL) { 2502 zil_commit_waiter_skip(zcw); 2503 list_remove(&nolwb_waiters, zcw); 2504 } 2505 2506 /* 2507 * And finally, we have to destroy the itx's that 2508 * couldn't be committed to an lwb; this will also call 2509 * the itx's callback if one exists for the itx. 2510 */ 2511 while ((itx = list_head(&nolwb_itxs)) != NULL) { 2512 list_remove(&nolwb_itxs, itx); 2513 zil_itx_destroy(itx); 2514 } 2515 } else { 2516 ASSERT(list_is_empty(&nolwb_waiters)); 2517 ASSERT3P(lwb, !=, NULL); 2518 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2519 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 2520 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 2521 2522 /* 2523 * At this point, the ZIL block pointed at by the "lwb" 2524 * variable is in one of the following states: "closed" 2525 * or "open". 2526 * 2527 * If it's "closed", then no itxs have been committed to 2528 * it, so there's no point in issuing its zio (i.e. it's 2529 * "empty"). 2530 * 2531 * If it's "open", then it contains one or more itxs that 2532 * eventually need to be committed to stable storage. In 2533 * this case we intentionally do not issue the lwb's zio 2534 * to disk yet, and instead rely on one of the following 2535 * two mechanisms for issuing the zio: 2536 * 2537 * 1. Ideally, there will be more ZIL activity occurring 2538 * on the system, such that this function will be 2539 * immediately called again (not necessarily by the same 2540 * thread) and this lwb's zio will be issued via 2541 * zil_lwb_commit(). This way, the lwb is guaranteed to 2542 * be "full" when it is issued to disk, and we'll make 2543 * use of the lwb's size the best we can. 2544 * 2545 * 2. If there isn't sufficient ZIL activity occurring on 2546 * the system, such that this lwb's zio isn't issued via 2547 * zil_lwb_commit(), zil_commit_waiter() will issue the 2548 * lwb's zio. If this occurs, the lwb is not guaranteed 2549 * to be "full" by the time its zio is issued, and means 2550 * the size of the lwb was "too large" given the amount 2551 * of ZIL activity occurring on the system at that time. 2552 * 2553 * We do this for a couple of reasons: 2554 * 2555 * 1. To try and reduce the number of IOPs needed to 2556 * write the same number of itxs. If an lwb has space 2557 * available in its buffer for more itxs, and more itxs 2558 * will be committed relatively soon (relative to the 2559 * latency of performing a write), then it's beneficial 2560 * to wait for these "next" itxs. This way, more itxs 2561 * can be committed to stable storage with fewer writes. 2562 * 2563 * 2. To try and use the largest lwb block size that the 2564 * incoming rate of itxs can support. Again, this is to 2565 * try and pack as many itxs into as few lwbs as 2566 * possible, without significantly impacting the latency 2567 * of each individual itx. 2568 */ 2569 } 2570 } 2571 2572 /* 2573 * This function is responsible for ensuring the passed in commit waiter 2574 * (and associated commit itx) is committed to an lwb. If the waiter is 2575 * not already committed to an lwb, all itxs in the zilog's queue of 2576 * itxs will be processed. The assumption is the passed in waiter's 2577 * commit itx will found in the queue just like the other non-commit 2578 * itxs, such that when the entire queue is processed, the waiter will 2579 * have been committed to an lwb. 2580 * 2581 * The lwb associated with the passed in waiter is not guaranteed to 2582 * have been issued by the time this function completes. If the lwb is 2583 * not issued, we rely on future calls to zil_commit_writer() to issue 2584 * the lwb, or the timeout mechanism found in zil_commit_waiter(). 2585 */ 2586 static void 2587 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw) 2588 { 2589 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2590 ASSERT(spa_writeable(zilog->zl_spa)); 2591 2592 mutex_enter(&zilog->zl_issuer_lock); 2593 2594 if (zcw->zcw_lwb != NULL || zcw->zcw_done) { 2595 /* 2596 * It's possible that, while we were waiting to acquire 2597 * the "zl_issuer_lock", another thread committed this 2598 * waiter to an lwb. If that occurs, we bail out early, 2599 * without processing any of the zilog's queue of itxs. 2600 * 2601 * On certain workloads and system configurations, the 2602 * "zl_issuer_lock" can become highly contended. In an 2603 * attempt to reduce this contention, we immediately drop 2604 * the lock if the waiter has already been processed. 2605 * 2606 * We've measured this optimization to reduce CPU spent 2607 * contending on this lock by up to 5%, using a system 2608 * with 32 CPUs, low latency storage (~50 usec writes), 2609 * and 1024 threads performing sync writes. 2610 */ 2611 goto out; 2612 } 2613 2614 ZIL_STAT_BUMP(zil_commit_writer_count); 2615 2616 zil_get_commit_list(zilog); 2617 zil_prune_commit_list(zilog); 2618 zil_process_commit_list(zilog); 2619 2620 out: 2621 mutex_exit(&zilog->zl_issuer_lock); 2622 } 2623 2624 static void 2625 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw) 2626 { 2627 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2628 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2629 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 2630 2631 lwb_t *lwb = zcw->zcw_lwb; 2632 ASSERT3P(lwb, !=, NULL); 2633 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED); 2634 2635 /* 2636 * If the lwb has already been issued by another thread, we can 2637 * immediately return since there's no work to be done (the 2638 * point of this function is to issue the lwb). Additionally, we 2639 * do this prior to acquiring the zl_issuer_lock, to avoid 2640 * acquiring it when it's not necessary to do so. 2641 */ 2642 if (lwb->lwb_state == LWB_STATE_ISSUED || 2643 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2644 lwb->lwb_state == LWB_STATE_FLUSH_DONE) 2645 return; 2646 2647 /* 2648 * In order to call zil_lwb_write_issue() we must hold the 2649 * zilog's "zl_issuer_lock". We can't simply acquire that lock, 2650 * since we're already holding the commit waiter's "zcw_lock", 2651 * and those two locks are acquired in the opposite order 2652 * elsewhere. 2653 */ 2654 mutex_exit(&zcw->zcw_lock); 2655 mutex_enter(&zilog->zl_issuer_lock); 2656 mutex_enter(&zcw->zcw_lock); 2657 2658 /* 2659 * Since we just dropped and re-acquired the commit waiter's 2660 * lock, we have to re-check to see if the waiter was marked 2661 * "done" during that process. If the waiter was marked "done", 2662 * the "lwb" pointer is no longer valid (it can be free'd after 2663 * the waiter is marked "done"), so without this check we could 2664 * wind up with a use-after-free error below. 2665 */ 2666 if (zcw->zcw_done) 2667 goto out; 2668 2669 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2670 2671 /* 2672 * We've already checked this above, but since we hadn't acquired 2673 * the zilog's zl_issuer_lock, we have to perform this check a 2674 * second time while holding the lock. 2675 * 2676 * We don't need to hold the zl_lock since the lwb cannot transition 2677 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb 2678 * _can_ transition from ISSUED to DONE, but it's OK to race with 2679 * that transition since we treat the lwb the same, whether it's in 2680 * the ISSUED or DONE states. 2681 * 2682 * The important thing, is we treat the lwb differently depending on 2683 * if it's ISSUED or OPENED, and block any other threads that might 2684 * attempt to issue this lwb. For that reason we hold the 2685 * zl_issuer_lock when checking the lwb_state; we must not call 2686 * zil_lwb_write_issue() if the lwb had already been issued. 2687 * 2688 * See the comment above the lwb_state_t structure definition for 2689 * more details on the lwb states, and locking requirements. 2690 */ 2691 if (lwb->lwb_state == LWB_STATE_ISSUED || 2692 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2693 lwb->lwb_state == LWB_STATE_FLUSH_DONE) 2694 goto out; 2695 2696 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 2697 2698 /* 2699 * As described in the comments above zil_commit_waiter() and 2700 * zil_process_commit_list(), we need to issue this lwb's zio 2701 * since we've reached the commit waiter's timeout and it still 2702 * hasn't been issued. 2703 */ 2704 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb); 2705 2706 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED); 2707 2708 /* 2709 * Since the lwb's zio hadn't been issued by the time this thread 2710 * reached its timeout, we reset the zilog's "zl_cur_used" field 2711 * to influence the zil block size selection algorithm. 2712 * 2713 * By having to issue the lwb's zio here, it means the size of the 2714 * lwb was too large, given the incoming throughput of itxs. By 2715 * setting "zl_cur_used" to zero, we communicate this fact to the 2716 * block size selection algorithm, so it can take this information 2717 * into account, and potentially select a smaller size for the 2718 * next lwb block that is allocated. 2719 */ 2720 zilog->zl_cur_used = 0; 2721 2722 if (nlwb == NULL) { 2723 /* 2724 * When zil_lwb_write_issue() returns NULL, this 2725 * indicates zio_alloc_zil() failed to allocate the 2726 * "next" lwb on-disk. When this occurs, the ZIL write 2727 * pipeline must be stalled; see the comment within the 2728 * zil_commit_writer_stall() function for more details. 2729 * 2730 * We must drop the commit waiter's lock prior to 2731 * calling zil_commit_writer_stall() or else we can wind 2732 * up with the following deadlock: 2733 * 2734 * - This thread is waiting for the txg to sync while 2735 * holding the waiter's lock; txg_wait_synced() is 2736 * used within txg_commit_writer_stall(). 2737 * 2738 * - The txg can't sync because it is waiting for this 2739 * lwb's zio callback to call dmu_tx_commit(). 2740 * 2741 * - The lwb's zio callback can't call dmu_tx_commit() 2742 * because it's blocked trying to acquire the waiter's 2743 * lock, which occurs prior to calling dmu_tx_commit() 2744 */ 2745 mutex_exit(&zcw->zcw_lock); 2746 zil_commit_writer_stall(zilog); 2747 mutex_enter(&zcw->zcw_lock); 2748 } 2749 2750 out: 2751 mutex_exit(&zilog->zl_issuer_lock); 2752 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2753 } 2754 2755 /* 2756 * This function is responsible for performing the following two tasks: 2757 * 2758 * 1. its primary responsibility is to block until the given "commit 2759 * waiter" is considered "done". 2760 * 2761 * 2. its secondary responsibility is to issue the zio for the lwb that 2762 * the given "commit waiter" is waiting on, if this function has 2763 * waited "long enough" and the lwb is still in the "open" state. 2764 * 2765 * Given a sufficient amount of itxs being generated and written using 2766 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit() 2767 * function. If this does not occur, this secondary responsibility will 2768 * ensure the lwb is issued even if there is not other synchronous 2769 * activity on the system. 2770 * 2771 * For more details, see zil_process_commit_list(); more specifically, 2772 * the comment at the bottom of that function. 2773 */ 2774 static void 2775 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw) 2776 { 2777 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2778 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2779 ASSERT(spa_writeable(zilog->zl_spa)); 2780 2781 mutex_enter(&zcw->zcw_lock); 2782 2783 /* 2784 * The timeout is scaled based on the lwb latency to avoid 2785 * significantly impacting the latency of each individual itx. 2786 * For more details, see the comment at the bottom of the 2787 * zil_process_commit_list() function. 2788 */ 2789 int pct = MAX(zfs_commit_timeout_pct, 1); 2790 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100; 2791 hrtime_t wakeup = gethrtime() + sleep; 2792 boolean_t timedout = B_FALSE; 2793 2794 while (!zcw->zcw_done) { 2795 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2796 2797 lwb_t *lwb = zcw->zcw_lwb; 2798 2799 /* 2800 * Usually, the waiter will have a non-NULL lwb field here, 2801 * but it's possible for it to be NULL as a result of 2802 * zil_commit() racing with spa_sync(). 2803 * 2804 * When zil_clean() is called, it's possible for the itxg 2805 * list (which may be cleaned via a taskq) to contain 2806 * commit itxs. When this occurs, the commit waiters linked 2807 * off of these commit itxs will not be committed to an 2808 * lwb. Additionally, these commit waiters will not be 2809 * marked done until zil_commit_waiter_skip() is called via 2810 * zil_itxg_clean(). 2811 * 2812 * Thus, it's possible for this commit waiter (i.e. the 2813 * "zcw" variable) to be found in this "in between" state; 2814 * where it's "zcw_lwb" field is NULL, and it hasn't yet 2815 * been skipped, so it's "zcw_done" field is still B_FALSE. 2816 */ 2817 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED); 2818 2819 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) { 2820 ASSERT3B(timedout, ==, B_FALSE); 2821 2822 /* 2823 * If the lwb hasn't been issued yet, then we 2824 * need to wait with a timeout, in case this 2825 * function needs to issue the lwb after the 2826 * timeout is reached; responsibility (2) from 2827 * the comment above this function. 2828 */ 2829 int rc = cv_timedwait_hires(&zcw->zcw_cv, 2830 &zcw->zcw_lock, wakeup, USEC2NSEC(1), 2831 CALLOUT_FLAG_ABSOLUTE); 2832 2833 if (rc != -1 || zcw->zcw_done) 2834 continue; 2835 2836 timedout = B_TRUE; 2837 zil_commit_waiter_timeout(zilog, zcw); 2838 2839 if (!zcw->zcw_done) { 2840 /* 2841 * If the commit waiter has already been 2842 * marked "done", it's possible for the 2843 * waiter's lwb structure to have already 2844 * been freed. Thus, we can only reliably 2845 * make these assertions if the waiter 2846 * isn't done. 2847 */ 2848 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2849 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED); 2850 } 2851 } else { 2852 /* 2853 * If the lwb isn't open, then it must have already 2854 * been issued. In that case, there's no need to 2855 * use a timeout when waiting for the lwb to 2856 * complete. 2857 * 2858 * Additionally, if the lwb is NULL, the waiter 2859 * will soon be signaled and marked done via 2860 * zil_clean() and zil_itxg_clean(), so no timeout 2861 * is required. 2862 */ 2863 2864 IMPLY(lwb != NULL, 2865 lwb->lwb_state == LWB_STATE_ISSUED || 2866 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2867 lwb->lwb_state == LWB_STATE_FLUSH_DONE); 2868 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock); 2869 } 2870 } 2871 2872 mutex_exit(&zcw->zcw_lock); 2873 } 2874 2875 static zil_commit_waiter_t * 2876 zil_alloc_commit_waiter(void) 2877 { 2878 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP); 2879 2880 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL); 2881 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL); 2882 list_link_init(&zcw->zcw_node); 2883 zcw->zcw_lwb = NULL; 2884 zcw->zcw_done = B_FALSE; 2885 zcw->zcw_zio_error = 0; 2886 2887 return (zcw); 2888 } 2889 2890 static void 2891 zil_free_commit_waiter(zil_commit_waiter_t *zcw) 2892 { 2893 ASSERT(!list_link_active(&zcw->zcw_node)); 2894 ASSERT3P(zcw->zcw_lwb, ==, NULL); 2895 ASSERT3B(zcw->zcw_done, ==, B_TRUE); 2896 mutex_destroy(&zcw->zcw_lock); 2897 cv_destroy(&zcw->zcw_cv); 2898 kmem_cache_free(zil_zcw_cache, zcw); 2899 } 2900 2901 /* 2902 * This function is used to create a TX_COMMIT itx and assign it. This 2903 * way, it will be linked into the ZIL's list of synchronous itxs, and 2904 * then later committed to an lwb (or skipped) when 2905 * zil_process_commit_list() is called. 2906 */ 2907 static void 2908 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw) 2909 { 2910 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 2911 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 2912 2913 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t)); 2914 itx->itx_sync = B_TRUE; 2915 itx->itx_private = zcw; 2916 2917 zil_itx_assign(zilog, itx, tx); 2918 2919 dmu_tx_commit(tx); 2920 } 2921 2922 /* 2923 * Commit ZFS Intent Log transactions (itxs) to stable storage. 2924 * 2925 * When writing ZIL transactions to the on-disk representation of the 2926 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple 2927 * itxs can be committed to a single lwb. Once a lwb is written and 2928 * committed to stable storage (i.e. the lwb is written, and vdevs have 2929 * been flushed), each itx that was committed to that lwb is also 2930 * considered to be committed to stable storage. 2931 * 2932 * When an itx is committed to an lwb, the log record (lr_t) contained 2933 * by the itx is copied into the lwb's zio buffer, and once this buffer 2934 * is written to disk, it becomes an on-disk ZIL block. 2935 * 2936 * As itxs are generated, they're inserted into the ZIL's queue of 2937 * uncommitted itxs. The semantics of zil_commit() are such that it will 2938 * block until all itxs that were in the queue when it was called, are 2939 * committed to stable storage. 2940 * 2941 * If "foid" is zero, this means all "synchronous" and "asynchronous" 2942 * itxs, for all objects in the dataset, will be committed to stable 2943 * storage prior to zil_commit() returning. If "foid" is non-zero, all 2944 * "synchronous" itxs for all objects, but only "asynchronous" itxs 2945 * that correspond to the foid passed in, will be committed to stable 2946 * storage prior to zil_commit() returning. 2947 * 2948 * Generally speaking, when zil_commit() is called, the consumer doesn't 2949 * actually care about _all_ of the uncommitted itxs. Instead, they're 2950 * simply trying to waiting for a specific itx to be committed to disk, 2951 * but the interface(s) for interacting with the ZIL don't allow such 2952 * fine-grained communication. A better interface would allow a consumer 2953 * to create and assign an itx, and then pass a reference to this itx to 2954 * zil_commit(); such that zil_commit() would return as soon as that 2955 * specific itx was committed to disk (instead of waiting for _all_ 2956 * itxs to be committed). 2957 * 2958 * When a thread calls zil_commit() a special "commit itx" will be 2959 * generated, along with a corresponding "waiter" for this commit itx. 2960 * zil_commit() will wait on this waiter's CV, such that when the waiter 2961 * is marked done, and signaled, zil_commit() will return. 2962 * 2963 * This commit itx is inserted into the queue of uncommitted itxs. This 2964 * provides an easy mechanism for determining which itxs were in the 2965 * queue prior to zil_commit() having been called, and which itxs were 2966 * added after zil_commit() was called. 2967 * 2968 * The commit itx is special; it doesn't have any on-disk representation. 2969 * When a commit itx is "committed" to an lwb, the waiter associated 2970 * with it is linked onto the lwb's list of waiters. Then, when that lwb 2971 * completes, each waiter on the lwb's list is marked done and signaled 2972 * -- allowing the thread waiting on the waiter to return from zil_commit(). 2973 * 2974 * It's important to point out a few critical factors that allow us 2975 * to make use of the commit itxs, commit waiters, per-lwb lists of 2976 * commit waiters, and zio completion callbacks like we're doing: 2977 * 2978 * 1. The list of waiters for each lwb is traversed, and each commit 2979 * waiter is marked "done" and signaled, in the zio completion 2980 * callback of the lwb's zio[*]. 2981 * 2982 * * Actually, the waiters are signaled in the zio completion 2983 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands 2984 * that are sent to the vdevs upon completion of the lwb zio. 2985 * 2986 * 2. When the itxs are inserted into the ZIL's queue of uncommitted 2987 * itxs, the order in which they are inserted is preserved[*]; as 2988 * itxs are added to the queue, they are added to the tail of 2989 * in-memory linked lists. 2990 * 2991 * When committing the itxs to lwbs (to be written to disk), they 2992 * are committed in the same order in which the itxs were added to 2993 * the uncommitted queue's linked list(s); i.e. the linked list of 2994 * itxs to commit is traversed from head to tail, and each itx is 2995 * committed to an lwb in that order. 2996 * 2997 * * To clarify: 2998 * 2999 * - the order of "sync" itxs is preserved w.r.t. other 3000 * "sync" itxs, regardless of the corresponding objects. 3001 * - the order of "async" itxs is preserved w.r.t. other 3002 * "async" itxs corresponding to the same object. 3003 * - the order of "async" itxs is *not* preserved w.r.t. other 3004 * "async" itxs corresponding to different objects. 3005 * - the order of "sync" itxs w.r.t. "async" itxs (or vice 3006 * versa) is *not* preserved, even for itxs that correspond 3007 * to the same object. 3008 * 3009 * For more details, see: zil_itx_assign(), zil_async_to_sync(), 3010 * zil_get_commit_list(), and zil_process_commit_list(). 3011 * 3012 * 3. The lwbs represent a linked list of blocks on disk. Thus, any 3013 * lwb cannot be considered committed to stable storage, until its 3014 * "previous" lwb is also committed to stable storage. This fact, 3015 * coupled with the fact described above, means that itxs are 3016 * committed in (roughly) the order in which they were generated. 3017 * This is essential because itxs are dependent on prior itxs. 3018 * Thus, we *must not* deem an itx as being committed to stable 3019 * storage, until *all* prior itxs have also been committed to 3020 * stable storage. 3021 * 3022 * To enforce this ordering of lwb zio's, while still leveraging as 3023 * much of the underlying storage performance as possible, we rely 3024 * on two fundamental concepts: 3025 * 3026 * 1. The creation and issuance of lwb zio's is protected by 3027 * the zilog's "zl_issuer_lock", which ensures only a single 3028 * thread is creating and/or issuing lwb's at a time 3029 * 2. The "previous" lwb is a child of the "current" lwb 3030 * (leveraging the zio parent-child dependency graph) 3031 * 3032 * By relying on this parent-child zio relationship, we can have 3033 * many lwb zio's concurrently issued to the underlying storage, 3034 * but the order in which they complete will be the same order in 3035 * which they were created. 3036 */ 3037 void 3038 zil_commit(zilog_t *zilog, uint64_t foid) 3039 { 3040 /* 3041 * We should never attempt to call zil_commit on a snapshot for 3042 * a couple of reasons: 3043 * 3044 * 1. A snapshot may never be modified, thus it cannot have any 3045 * in-flight itxs that would have modified the dataset. 3046 * 3047 * 2. By design, when zil_commit() is called, a commit itx will 3048 * be assigned to this zilog; as a result, the zilog will be 3049 * dirtied. We must not dirty the zilog of a snapshot; there's 3050 * checks in the code that enforce this invariant, and will 3051 * cause a panic if it's not upheld. 3052 */ 3053 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE); 3054 3055 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 3056 return; 3057 3058 if (!spa_writeable(zilog->zl_spa)) { 3059 /* 3060 * If the SPA is not writable, there should never be any 3061 * pending itxs waiting to be committed to disk. If that 3062 * weren't true, we'd skip writing those itxs out, and 3063 * would break the semantics of zil_commit(); thus, we're 3064 * verifying that truth before we return to the caller. 3065 */ 3066 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3067 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 3068 for (int i = 0; i < TXG_SIZE; i++) 3069 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL); 3070 return; 3071 } 3072 3073 /* 3074 * If the ZIL is suspended, we don't want to dirty it by calling 3075 * zil_commit_itx_assign() below, nor can we write out 3076 * lwbs like would be done in zil_commit_write(). Thus, we 3077 * simply rely on txg_wait_synced() to maintain the necessary 3078 * semantics, and avoid calling those functions altogether. 3079 */ 3080 if (zilog->zl_suspend > 0) { 3081 txg_wait_synced(zilog->zl_dmu_pool, 0); 3082 return; 3083 } 3084 3085 zil_commit_impl(zilog, foid); 3086 } 3087 3088 void 3089 zil_commit_impl(zilog_t *zilog, uint64_t foid) 3090 { 3091 ZIL_STAT_BUMP(zil_commit_count); 3092 3093 /* 3094 * Move the "async" itxs for the specified foid to the "sync" 3095 * queues, such that they will be later committed (or skipped) 3096 * to an lwb when zil_process_commit_list() is called. 3097 * 3098 * Since these "async" itxs must be committed prior to this 3099 * call to zil_commit returning, we must perform this operation 3100 * before we call zil_commit_itx_assign(). 3101 */ 3102 zil_async_to_sync(zilog, foid); 3103 3104 /* 3105 * We allocate a new "waiter" structure which will initially be 3106 * linked to the commit itx using the itx's "itx_private" field. 3107 * Since the commit itx doesn't represent any on-disk state, 3108 * when it's committed to an lwb, rather than copying the its 3109 * lr_t into the lwb's buffer, the commit itx's "waiter" will be 3110 * added to the lwb's list of waiters. Then, when the lwb is 3111 * committed to stable storage, each waiter in the lwb's list of 3112 * waiters will be marked "done", and signalled. 3113 * 3114 * We must create the waiter and assign the commit itx prior to 3115 * calling zil_commit_writer(), or else our specific commit itx 3116 * is not guaranteed to be committed to an lwb prior to calling 3117 * zil_commit_waiter(). 3118 */ 3119 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter(); 3120 zil_commit_itx_assign(zilog, zcw); 3121 3122 zil_commit_writer(zilog, zcw); 3123 zil_commit_waiter(zilog, zcw); 3124 3125 if (zcw->zcw_zio_error != 0) { 3126 /* 3127 * If there was an error writing out the ZIL blocks that 3128 * this thread is waiting on, then we fallback to 3129 * relying on spa_sync() to write out the data this 3130 * thread is waiting on. Obviously this has performance 3131 * implications, but the expectation is for this to be 3132 * an exceptional case, and shouldn't occur often. 3133 */ 3134 DTRACE_PROBE2(zil__commit__io__error, 3135 zilog_t *, zilog, zil_commit_waiter_t *, zcw); 3136 txg_wait_synced(zilog->zl_dmu_pool, 0); 3137 } 3138 3139 zil_free_commit_waiter(zcw); 3140 } 3141 3142 /* 3143 * Called in syncing context to free committed log blocks and update log header. 3144 */ 3145 void 3146 zil_sync(zilog_t *zilog, dmu_tx_t *tx) 3147 { 3148 zil_header_t *zh = zil_header_in_syncing_context(zilog); 3149 uint64_t txg = dmu_tx_get_txg(tx); 3150 spa_t *spa = zilog->zl_spa; 3151 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 3152 lwb_t *lwb; 3153 3154 /* 3155 * We don't zero out zl_destroy_txg, so make sure we don't try 3156 * to destroy it twice. 3157 */ 3158 if (spa_sync_pass(spa) != 1) 3159 return; 3160 3161 zil_lwb_flush_wait_all(zilog, txg); 3162 3163 mutex_enter(&zilog->zl_lock); 3164 3165 ASSERT(zilog->zl_stop_sync == 0); 3166 3167 if (*replayed_seq != 0) { 3168 ASSERT(zh->zh_replay_seq < *replayed_seq); 3169 zh->zh_replay_seq = *replayed_seq; 3170 *replayed_seq = 0; 3171 } 3172 3173 if (zilog->zl_destroy_txg == txg) { 3174 blkptr_t blk = zh->zh_log; 3175 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 3176 3177 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 3178 3179 memset(zh, 0, sizeof (zil_header_t)); 3180 memset(zilog->zl_replayed_seq, 0, 3181 sizeof (zilog->zl_replayed_seq)); 3182 3183 if (zilog->zl_keep_first) { 3184 /* 3185 * If this block was part of log chain that couldn't 3186 * be claimed because a device was missing during 3187 * zil_claim(), but that device later returns, 3188 * then this block could erroneously appear valid. 3189 * To guard against this, assign a new GUID to the new 3190 * log chain so it doesn't matter what blk points to. 3191 */ 3192 zil_init_log_chain(zilog, &blk); 3193 zh->zh_log = blk; 3194 } else { 3195 /* 3196 * A destroyed ZIL chain can't contain any TX_SETSAXATTR 3197 * records. So, deactivate the feature for this dataset. 3198 * We activate it again when we start a new ZIL chain. 3199 */ 3200 if (dsl_dataset_feature_is_active(ds, 3201 SPA_FEATURE_ZILSAXATTR)) 3202 dsl_dataset_deactivate_feature(ds, 3203 SPA_FEATURE_ZILSAXATTR, tx); 3204 } 3205 } 3206 3207 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 3208 zh->zh_log = lwb->lwb_blk; 3209 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 3210 break; 3211 list_remove(&zilog->zl_lwb_list, lwb); 3212 zio_free(spa, txg, &lwb->lwb_blk); 3213 zil_free_lwb(zilog, lwb); 3214 3215 /* 3216 * If we don't have anything left in the lwb list then 3217 * we've had an allocation failure and we need to zero 3218 * out the zil_header blkptr so that we don't end 3219 * up freeing the same block twice. 3220 */ 3221 if (list_head(&zilog->zl_lwb_list) == NULL) 3222 BP_ZERO(&zh->zh_log); 3223 } 3224 3225 /* 3226 * Remove fastwrite on any blocks that have been pre-allocated for 3227 * the next commit. This prevents fastwrite counter pollution by 3228 * unused, long-lived LWBs. 3229 */ 3230 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) { 3231 if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) { 3232 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); 3233 lwb->lwb_fastwrite = 0; 3234 } 3235 } 3236 3237 mutex_exit(&zilog->zl_lock); 3238 } 3239 3240 static int 3241 zil_lwb_cons(void *vbuf, void *unused, int kmflag) 3242 { 3243 (void) unused, (void) kmflag; 3244 lwb_t *lwb = vbuf; 3245 list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); 3246 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t), 3247 offsetof(zil_commit_waiter_t, zcw_node)); 3248 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare, 3249 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 3250 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 3251 return (0); 3252 } 3253 3254 static void 3255 zil_lwb_dest(void *vbuf, void *unused) 3256 { 3257 (void) unused; 3258 lwb_t *lwb = vbuf; 3259 mutex_destroy(&lwb->lwb_vdev_lock); 3260 avl_destroy(&lwb->lwb_vdev_tree); 3261 list_destroy(&lwb->lwb_waiters); 3262 list_destroy(&lwb->lwb_itxs); 3263 } 3264 3265 void 3266 zil_init(void) 3267 { 3268 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 3269 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0); 3270 3271 zil_zcw_cache = kmem_cache_create("zil_zcw_cache", 3272 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 3273 3274 zil_ksp = kstat_create("zfs", 0, "zil", "misc", 3275 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t), 3276 KSTAT_FLAG_VIRTUAL); 3277 3278 if (zil_ksp != NULL) { 3279 zil_ksp->ks_data = &zil_stats; 3280 kstat_install(zil_ksp); 3281 } 3282 } 3283 3284 void 3285 zil_fini(void) 3286 { 3287 kmem_cache_destroy(zil_zcw_cache); 3288 kmem_cache_destroy(zil_lwb_cache); 3289 3290 if (zil_ksp != NULL) { 3291 kstat_delete(zil_ksp); 3292 zil_ksp = NULL; 3293 } 3294 } 3295 3296 void 3297 zil_set_sync(zilog_t *zilog, uint64_t sync) 3298 { 3299 zilog->zl_sync = sync; 3300 } 3301 3302 void 3303 zil_set_logbias(zilog_t *zilog, uint64_t logbias) 3304 { 3305 zilog->zl_logbias = logbias; 3306 } 3307 3308 zilog_t * 3309 zil_alloc(objset_t *os, zil_header_t *zh_phys) 3310 { 3311 zilog_t *zilog; 3312 3313 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 3314 3315 zilog->zl_header = zh_phys; 3316 zilog->zl_os = os; 3317 zilog->zl_spa = dmu_objset_spa(os); 3318 zilog->zl_dmu_pool = dmu_objset_pool(os); 3319 zilog->zl_destroy_txg = TXG_INITIAL - 1; 3320 zilog->zl_logbias = dmu_objset_logbias(os); 3321 zilog->zl_sync = dmu_objset_syncprop(os); 3322 zilog->zl_dirty_max_txg = 0; 3323 zilog->zl_last_lwb_opened = NULL; 3324 zilog->zl_last_lwb_latency = 0; 3325 zilog->zl_max_block_size = zil_maxblocksize; 3326 3327 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 3328 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL); 3329 mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL); 3330 3331 for (int i = 0; i < TXG_SIZE; i++) { 3332 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 3333 MUTEX_DEFAULT, NULL); 3334 } 3335 3336 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 3337 offsetof(lwb_t, lwb_node)); 3338 3339 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 3340 offsetof(itx_t, itx_node)); 3341 3342 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 3343 cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL); 3344 3345 return (zilog); 3346 } 3347 3348 void 3349 zil_free(zilog_t *zilog) 3350 { 3351 int i; 3352 3353 zilog->zl_stop_sync = 1; 3354 3355 ASSERT0(zilog->zl_suspend); 3356 ASSERT0(zilog->zl_suspending); 3357 3358 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3359 list_destroy(&zilog->zl_lwb_list); 3360 3361 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 3362 list_destroy(&zilog->zl_itx_commit_list); 3363 3364 for (i = 0; i < TXG_SIZE; i++) { 3365 /* 3366 * It's possible for an itx to be generated that doesn't dirty 3367 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 3368 * callback to remove the entry. We remove those here. 3369 * 3370 * Also free up the ziltest itxs. 3371 */ 3372 if (zilog->zl_itxg[i].itxg_itxs) 3373 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 3374 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 3375 } 3376 3377 mutex_destroy(&zilog->zl_issuer_lock); 3378 mutex_destroy(&zilog->zl_lock); 3379 mutex_destroy(&zilog->zl_lwb_io_lock); 3380 3381 cv_destroy(&zilog->zl_cv_suspend); 3382 cv_destroy(&zilog->zl_lwb_io_cv); 3383 3384 kmem_free(zilog, sizeof (zilog_t)); 3385 } 3386 3387 /* 3388 * Open an intent log. 3389 */ 3390 zilog_t * 3391 zil_open(objset_t *os, zil_get_data_t *get_data) 3392 { 3393 zilog_t *zilog = dmu_objset_zil(os); 3394 3395 ASSERT3P(zilog->zl_get_data, ==, NULL); 3396 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 3397 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3398 3399 zilog->zl_get_data = get_data; 3400 3401 return (zilog); 3402 } 3403 3404 /* 3405 * Close an intent log. 3406 */ 3407 void 3408 zil_close(zilog_t *zilog) 3409 { 3410 lwb_t *lwb; 3411 uint64_t txg; 3412 3413 if (!dmu_objset_is_snapshot(zilog->zl_os)) { 3414 zil_commit(zilog, 0); 3415 } else { 3416 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 3417 ASSERT0(zilog->zl_dirty_max_txg); 3418 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE); 3419 } 3420 3421 mutex_enter(&zilog->zl_lock); 3422 lwb = list_tail(&zilog->zl_lwb_list); 3423 if (lwb == NULL) 3424 txg = zilog->zl_dirty_max_txg; 3425 else 3426 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg); 3427 mutex_exit(&zilog->zl_lock); 3428 3429 /* 3430 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends 3431 * on the time when the dmu_tx transaction is assigned in 3432 * zil_lwb_write_issue(). 3433 */ 3434 mutex_enter(&zilog->zl_lwb_io_lock); 3435 txg = MAX(zilog->zl_lwb_max_issued_txg, txg); 3436 mutex_exit(&zilog->zl_lwb_io_lock); 3437 3438 /* 3439 * We need to use txg_wait_synced() to wait until that txg is synced. 3440 * zil_sync() will guarantee all lwbs up to that txg have been 3441 * written out, flushed, and cleaned. 3442 */ 3443 if (txg != 0) 3444 txg_wait_synced(zilog->zl_dmu_pool, txg); 3445 3446 if (zilog_is_dirty(zilog)) 3447 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog, 3448 (u_longlong_t)txg); 3449 if (txg < spa_freeze_txg(zilog->zl_spa)) 3450 VERIFY(!zilog_is_dirty(zilog)); 3451 3452 zilog->zl_get_data = NULL; 3453 3454 /* 3455 * We should have only one lwb left on the list; remove it now. 3456 */ 3457 mutex_enter(&zilog->zl_lock); 3458 lwb = list_head(&zilog->zl_lwb_list); 3459 if (lwb != NULL) { 3460 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list)); 3461 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 3462 3463 if (lwb->lwb_fastwrite) 3464 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); 3465 3466 list_remove(&zilog->zl_lwb_list, lwb); 3467 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 3468 zil_free_lwb(zilog, lwb); 3469 } 3470 mutex_exit(&zilog->zl_lock); 3471 } 3472 3473 static char *suspend_tag = "zil suspending"; 3474 3475 /* 3476 * Suspend an intent log. While in suspended mode, we still honor 3477 * synchronous semantics, but we rely on txg_wait_synced() to do it. 3478 * On old version pools, we suspend the log briefly when taking a 3479 * snapshot so that it will have an empty intent log. 3480 * 3481 * Long holds are not really intended to be used the way we do here -- 3482 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 3483 * could fail. Therefore we take pains to only put a long hold if it is 3484 * actually necessary. Fortunately, it will only be necessary if the 3485 * objset is currently mounted (or the ZVOL equivalent). In that case it 3486 * will already have a long hold, so we are not really making things any worse. 3487 * 3488 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 3489 * zvol_state_t), and use their mechanism to prevent their hold from being 3490 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 3491 * very little gain. 3492 * 3493 * if cookiep == NULL, this does both the suspend & resume. 3494 * Otherwise, it returns with the dataset "long held", and the cookie 3495 * should be passed into zil_resume(). 3496 */ 3497 int 3498 zil_suspend(const char *osname, void **cookiep) 3499 { 3500 objset_t *os; 3501 zilog_t *zilog; 3502 const zil_header_t *zh; 3503 int error; 3504 3505 error = dmu_objset_hold(osname, suspend_tag, &os); 3506 if (error != 0) 3507 return (error); 3508 zilog = dmu_objset_zil(os); 3509 3510 mutex_enter(&zilog->zl_lock); 3511 zh = zilog->zl_header; 3512 3513 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 3514 mutex_exit(&zilog->zl_lock); 3515 dmu_objset_rele(os, suspend_tag); 3516 return (SET_ERROR(EBUSY)); 3517 } 3518 3519 /* 3520 * Don't put a long hold in the cases where we can avoid it. This 3521 * is when there is no cookie so we are doing a suspend & resume 3522 * (i.e. called from zil_vdev_offline()), and there's nothing to do 3523 * for the suspend because it's already suspended, or there's no ZIL. 3524 */ 3525 if (cookiep == NULL && !zilog->zl_suspending && 3526 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 3527 mutex_exit(&zilog->zl_lock); 3528 dmu_objset_rele(os, suspend_tag); 3529 return (0); 3530 } 3531 3532 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 3533 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 3534 3535 zilog->zl_suspend++; 3536 3537 if (zilog->zl_suspend > 1) { 3538 /* 3539 * Someone else is already suspending it. 3540 * Just wait for them to finish. 3541 */ 3542 3543 while (zilog->zl_suspending) 3544 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 3545 mutex_exit(&zilog->zl_lock); 3546 3547 if (cookiep == NULL) 3548 zil_resume(os); 3549 else 3550 *cookiep = os; 3551 return (0); 3552 } 3553 3554 /* 3555 * If there is no pointer to an on-disk block, this ZIL must not 3556 * be active (e.g. filesystem not mounted), so there's nothing 3557 * to clean up. 3558 */ 3559 if (BP_IS_HOLE(&zh->zh_log)) { 3560 ASSERT(cookiep != NULL); /* fast path already handled */ 3561 3562 *cookiep = os; 3563 mutex_exit(&zilog->zl_lock); 3564 return (0); 3565 } 3566 3567 /* 3568 * The ZIL has work to do. Ensure that the associated encryption 3569 * key will remain mapped while we are committing the log by 3570 * grabbing a reference to it. If the key isn't loaded we have no 3571 * choice but to return an error until the wrapping key is loaded. 3572 */ 3573 if (os->os_encrypted && 3574 dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) { 3575 zilog->zl_suspend--; 3576 mutex_exit(&zilog->zl_lock); 3577 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3578 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3579 return (SET_ERROR(EACCES)); 3580 } 3581 3582 zilog->zl_suspending = B_TRUE; 3583 mutex_exit(&zilog->zl_lock); 3584 3585 /* 3586 * We need to use zil_commit_impl to ensure we wait for all 3587 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed 3588 * to disk before proceeding. If we used zil_commit instead, it 3589 * would just call txg_wait_synced(), because zl_suspend is set. 3590 * txg_wait_synced() doesn't wait for these lwb's to be 3591 * LWB_STATE_FLUSH_DONE before returning. 3592 */ 3593 zil_commit_impl(zilog, 0); 3594 3595 /* 3596 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we 3597 * use txg_wait_synced() to ensure the data from the zilog has 3598 * migrated to the main pool before calling zil_destroy(). 3599 */ 3600 txg_wait_synced(zilog->zl_dmu_pool, 0); 3601 3602 zil_destroy(zilog, B_FALSE); 3603 3604 mutex_enter(&zilog->zl_lock); 3605 zilog->zl_suspending = B_FALSE; 3606 cv_broadcast(&zilog->zl_cv_suspend); 3607 mutex_exit(&zilog->zl_lock); 3608 3609 if (os->os_encrypted) 3610 dsl_dataset_remove_key_mapping(dmu_objset_ds(os)); 3611 3612 if (cookiep == NULL) 3613 zil_resume(os); 3614 else 3615 *cookiep = os; 3616 return (0); 3617 } 3618 3619 void 3620 zil_resume(void *cookie) 3621 { 3622 objset_t *os = cookie; 3623 zilog_t *zilog = dmu_objset_zil(os); 3624 3625 mutex_enter(&zilog->zl_lock); 3626 ASSERT(zilog->zl_suspend != 0); 3627 zilog->zl_suspend--; 3628 mutex_exit(&zilog->zl_lock); 3629 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3630 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3631 } 3632 3633 typedef struct zil_replay_arg { 3634 zil_replay_func_t *const *zr_replay; 3635 void *zr_arg; 3636 boolean_t zr_byteswap; 3637 char *zr_lr; 3638 } zil_replay_arg_t; 3639 3640 static int 3641 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error) 3642 { 3643 char name[ZFS_MAX_DATASET_NAME_LEN]; 3644 3645 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 3646 3647 dmu_objset_name(zilog->zl_os, name); 3648 3649 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 3650 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 3651 (u_longlong_t)lr->lrc_seq, 3652 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 3653 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 3654 3655 return (error); 3656 } 3657 3658 static int 3659 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra, 3660 uint64_t claim_txg) 3661 { 3662 zil_replay_arg_t *zr = zra; 3663 const zil_header_t *zh = zilog->zl_header; 3664 uint64_t reclen = lr->lrc_reclen; 3665 uint64_t txtype = lr->lrc_txtype; 3666 int error = 0; 3667 3668 zilog->zl_replaying_seq = lr->lrc_seq; 3669 3670 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 3671 return (0); 3672 3673 if (lr->lrc_txg < claim_txg) /* already committed */ 3674 return (0); 3675 3676 /* Strip case-insensitive bit, still present in log record */ 3677 txtype &= ~TX_CI; 3678 3679 if (txtype == 0 || txtype >= TX_MAX_TYPE) 3680 return (zil_replay_error(zilog, lr, EINVAL)); 3681 3682 /* 3683 * If this record type can be logged out of order, the object 3684 * (lr_foid) may no longer exist. That's legitimate, not an error. 3685 */ 3686 if (TX_OOO(txtype)) { 3687 error = dmu_object_info(zilog->zl_os, 3688 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL); 3689 if (error == ENOENT || error == EEXIST) 3690 return (0); 3691 } 3692 3693 /* 3694 * Make a copy of the data so we can revise and extend it. 3695 */ 3696 memcpy(zr->zr_lr, lr, reclen); 3697 3698 /* 3699 * If this is a TX_WRITE with a blkptr, suck in the data. 3700 */ 3701 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 3702 error = zil_read_log_data(zilog, (lr_write_t *)lr, 3703 zr->zr_lr + reclen); 3704 if (error != 0) 3705 return (zil_replay_error(zilog, lr, error)); 3706 } 3707 3708 /* 3709 * The log block containing this lr may have been byteswapped 3710 * so that we can easily examine common fields like lrc_txtype. 3711 * However, the log is a mix of different record types, and only the 3712 * replay vectors know how to byteswap their records. Therefore, if 3713 * the lr was byteswapped, undo it before invoking the replay vector. 3714 */ 3715 if (zr->zr_byteswap) 3716 byteswap_uint64_array(zr->zr_lr, reclen); 3717 3718 /* 3719 * We must now do two things atomically: replay this log record, 3720 * and update the log header sequence number to reflect the fact that 3721 * we did so. At the end of each replay function the sequence number 3722 * is updated if we are in replay mode. 3723 */ 3724 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 3725 if (error != 0) { 3726 /* 3727 * The DMU's dnode layer doesn't see removes until the txg 3728 * commits, so a subsequent claim can spuriously fail with 3729 * EEXIST. So if we receive any error we try syncing out 3730 * any removes then retry the transaction. Note that we 3731 * specify B_FALSE for byteswap now, so we don't do it twice. 3732 */ 3733 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 3734 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 3735 if (error != 0) 3736 return (zil_replay_error(zilog, lr, error)); 3737 } 3738 return (0); 3739 } 3740 3741 static int 3742 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg) 3743 { 3744 (void) bp, (void) arg, (void) claim_txg; 3745 3746 zilog->zl_replay_blks++; 3747 3748 return (0); 3749 } 3750 3751 /* 3752 * If this dataset has a non-empty intent log, replay it and destroy it. 3753 */ 3754 void 3755 zil_replay(objset_t *os, void *arg, 3756 zil_replay_func_t *const replay_func[TX_MAX_TYPE]) 3757 { 3758 zilog_t *zilog = dmu_objset_zil(os); 3759 const zil_header_t *zh = zilog->zl_header; 3760 zil_replay_arg_t zr; 3761 3762 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 3763 zil_destroy(zilog, B_TRUE); 3764 return; 3765 } 3766 3767 zr.zr_replay = replay_func; 3768 zr.zr_arg = arg; 3769 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 3770 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 3771 3772 /* 3773 * Wait for in-progress removes to sync before starting replay. 3774 */ 3775 txg_wait_synced(zilog->zl_dmu_pool, 0); 3776 3777 zilog->zl_replay = B_TRUE; 3778 zilog->zl_replay_time = ddi_get_lbolt(); 3779 ASSERT(zilog->zl_replay_blks == 0); 3780 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 3781 zh->zh_claim_txg, B_TRUE); 3782 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 3783 3784 zil_destroy(zilog, B_FALSE); 3785 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 3786 zilog->zl_replay = B_FALSE; 3787 } 3788 3789 boolean_t 3790 zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 3791 { 3792 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 3793 return (B_TRUE); 3794 3795 if (zilog->zl_replay) { 3796 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 3797 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 3798 zilog->zl_replaying_seq; 3799 return (B_TRUE); 3800 } 3801 3802 return (B_FALSE); 3803 } 3804 3805 int 3806 zil_reset(const char *osname, void *arg) 3807 { 3808 (void) arg; 3809 3810 int error = zil_suspend(osname, NULL); 3811 /* EACCES means crypto key not loaded */ 3812 if ((error == EACCES) || (error == EBUSY)) 3813 return (SET_ERROR(error)); 3814 if (error != 0) 3815 return (SET_ERROR(EEXIST)); 3816 return (0); 3817 } 3818 3819 EXPORT_SYMBOL(zil_alloc); 3820 EXPORT_SYMBOL(zil_free); 3821 EXPORT_SYMBOL(zil_open); 3822 EXPORT_SYMBOL(zil_close); 3823 EXPORT_SYMBOL(zil_replay); 3824 EXPORT_SYMBOL(zil_replaying); 3825 EXPORT_SYMBOL(zil_destroy); 3826 EXPORT_SYMBOL(zil_destroy_sync); 3827 EXPORT_SYMBOL(zil_itx_create); 3828 EXPORT_SYMBOL(zil_itx_destroy); 3829 EXPORT_SYMBOL(zil_itx_assign); 3830 EXPORT_SYMBOL(zil_commit); 3831 EXPORT_SYMBOL(zil_claim); 3832 EXPORT_SYMBOL(zil_check_log_chain); 3833 EXPORT_SYMBOL(zil_sync); 3834 EXPORT_SYMBOL(zil_clean); 3835 EXPORT_SYMBOL(zil_suspend); 3836 EXPORT_SYMBOL(zil_resume); 3837 EXPORT_SYMBOL(zil_lwb_add_block); 3838 EXPORT_SYMBOL(zil_bp_tree_add); 3839 EXPORT_SYMBOL(zil_set_sync); 3840 EXPORT_SYMBOL(zil_set_logbias); 3841 3842 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, INT, ZMOD_RW, 3843 "ZIL block open timeout percentage"); 3844 3845 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW, 3846 "Disable intent logging replay"); 3847 3848 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW, 3849 "Disable ZIL cache flushes"); 3850 3851 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, ULONG, ZMOD_RW, 3852 "Limit in bytes slog sync writes per commit"); 3853 3854 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, INT, ZMOD_RW, 3855 "Limit in bytes of ZIL log block size"); 3856