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_tx = NULL; 563 lwb->lwb_issued_timestamp = 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 dmu_tx_t *tx = lwb->lwb_tx; 1187 zil_commit_waiter_t *zcw; 1188 itx_t *itx; 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 * Ensure the lwb buffer pointer is cleared before releasing the 1198 * txg. If we have had an allocation failure and the txg is 1199 * waiting to sync then we want zil_sync() to remove the lwb so 1200 * that it's not picked up as the next new one in 1201 * zil_process_commit_list(). zil_sync() will only remove the 1202 * lwb if lwb_buf is null. 1203 */ 1204 lwb->lwb_buf = NULL; 1205 lwb->lwb_tx = NULL; 1206 1207 ASSERT3U(lwb->lwb_issued_timestamp, >, 0); 1208 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp; 1209 1210 lwb->lwb_root_zio = NULL; 1211 1212 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE); 1213 lwb->lwb_state = LWB_STATE_FLUSH_DONE; 1214 1215 if (zilog->zl_last_lwb_opened == lwb) { 1216 /* 1217 * Remember the highest committed log sequence number 1218 * for ztest. We only update this value when all the log 1219 * writes succeeded, because ztest wants to ASSERT that 1220 * it got the whole log chain. 1221 */ 1222 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1223 } 1224 1225 while ((itx = list_head(&lwb->lwb_itxs)) != NULL) { 1226 list_remove(&lwb->lwb_itxs, itx); 1227 zil_itx_destroy(itx); 1228 } 1229 1230 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) { 1231 mutex_enter(&zcw->zcw_lock); 1232 1233 ASSERT(list_link_active(&zcw->zcw_node)); 1234 list_remove(&lwb->lwb_waiters, zcw); 1235 1236 ASSERT3P(zcw->zcw_lwb, ==, lwb); 1237 zcw->zcw_lwb = NULL; 1238 /* 1239 * We expect any ZIO errors from child ZIOs to have been 1240 * propagated "up" to this specific LWB's root ZIO, in 1241 * order for this error handling to work correctly. This 1242 * includes ZIO errors from either this LWB's write or 1243 * flush, as well as any errors from other dependent LWBs 1244 * (e.g. a root LWB ZIO that might be a child of this LWB). 1245 * 1246 * With that said, it's important to note that LWB flush 1247 * errors are not propagated up to the LWB root ZIO. 1248 * This is incorrect behavior, and results in VDEV flush 1249 * errors not being handled correctly here. See the 1250 * comment above the call to "zio_flush" for details. 1251 */ 1252 1253 zcw->zcw_zio_error = zio->io_error; 1254 1255 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 1256 zcw->zcw_done = B_TRUE; 1257 cv_broadcast(&zcw->zcw_cv); 1258 1259 mutex_exit(&zcw->zcw_lock); 1260 } 1261 1262 mutex_exit(&zilog->zl_lock); 1263 1264 /* 1265 * Now that we've written this log block, we have a stable pointer 1266 * to the next block in the chain, so it's OK to let the txg in 1267 * which we allocated the next block sync. 1268 */ 1269 dmu_tx_commit(tx); 1270 } 1271 1272 /* 1273 * This is called when an lwb's write zio completes. The callback's 1274 * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs 1275 * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved 1276 * in writing out this specific lwb's data, and in the case that cache 1277 * flushes have been deferred, vdevs involved in writing the data for 1278 * previous lwbs. The writes corresponding to all the vdevs in the 1279 * lwb_vdev_tree will have completed by the time this is called, due to 1280 * the zio dependencies configured in zil_lwb_set_zio_dependency(), 1281 * which takes deferred flushes into account. The lwb will be "done" 1282 * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio 1283 * completion callback for the lwb's root zio. 1284 */ 1285 static void 1286 zil_lwb_write_done(zio_t *zio) 1287 { 1288 lwb_t *lwb = zio->io_private; 1289 spa_t *spa = zio->io_spa; 1290 zilog_t *zilog = lwb->lwb_zilog; 1291 avl_tree_t *t = &lwb->lwb_vdev_tree; 1292 void *cookie = NULL; 1293 zil_vdev_node_t *zv; 1294 lwb_t *nlwb; 1295 1296 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0); 1297 1298 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1299 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 1300 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 1301 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 1302 ASSERT(!BP_IS_GANG(zio->io_bp)); 1303 ASSERT(!BP_IS_HOLE(zio->io_bp)); 1304 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 1305 1306 abd_free(zio->io_abd); 1307 1308 mutex_enter(&zilog->zl_lock); 1309 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED); 1310 lwb->lwb_state = LWB_STATE_WRITE_DONE; 1311 lwb->lwb_write_zio = NULL; 1312 lwb->lwb_fastwrite = FALSE; 1313 nlwb = list_next(&zilog->zl_lwb_list, lwb); 1314 mutex_exit(&zilog->zl_lock); 1315 1316 if (avl_numnodes(t) == 0) 1317 return; 1318 1319 /* 1320 * If there was an IO error, we're not going to call zio_flush() 1321 * on these vdevs, so we simply empty the tree and free the 1322 * nodes. We avoid calling zio_flush() since there isn't any 1323 * good reason for doing so, after the lwb block failed to be 1324 * written out. 1325 * 1326 * Additionally, we don't perform any further error handling at 1327 * this point (e.g. setting "zcw_zio_error" appropriately), as 1328 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus, 1329 * we expect any error seen here, to have been propagated to 1330 * that function). 1331 */ 1332 if (zio->io_error != 0) { 1333 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) 1334 kmem_free(zv, sizeof (*zv)); 1335 return; 1336 } 1337 1338 /* 1339 * If this lwb does not have any threads waiting for it to 1340 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE 1341 * command to the vdevs written to by "this" lwb, and instead 1342 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE 1343 * command for those vdevs. Thus, we merge the vdev tree of 1344 * "this" lwb with the vdev tree of the "next" lwb in the list, 1345 * and assume the "next" lwb will handle flushing the vdevs (or 1346 * deferring the flush(s) again). 1347 * 1348 * This is a useful performance optimization, especially for 1349 * workloads with lots of async write activity and few sync 1350 * write and/or fsync activity, as it has the potential to 1351 * coalesce multiple flush commands to a vdev into one. 1352 */ 1353 if (list_head(&lwb->lwb_waiters) == NULL && nlwb != NULL) { 1354 zil_lwb_flush_defer(lwb, nlwb); 1355 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 1356 return; 1357 } 1358 1359 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 1360 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 1361 if (vd != NULL) { 1362 /* 1363 * The "ZIO_FLAG_DONT_PROPAGATE" is currently 1364 * always used within "zio_flush". This means, 1365 * any errors when flushing the vdev(s), will 1366 * (unfortunately) not be handled correctly, 1367 * since these "zio_flush" errors will not be 1368 * propagated up to "zil_lwb_flush_vdevs_done". 1369 */ 1370 zio_flush(lwb->lwb_root_zio, vd); 1371 } 1372 kmem_free(zv, sizeof (*zv)); 1373 } 1374 } 1375 1376 static void 1377 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb) 1378 { 1379 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened; 1380 1381 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1382 ASSERT(MUTEX_HELD(&zilog->zl_lock)); 1383 1384 /* 1385 * The zilog's "zl_last_lwb_opened" field is used to build the 1386 * lwb/zio dependency chain, which is used to preserve the 1387 * ordering of lwb completions that is required by the semantics 1388 * of the ZIL. Each new lwb zio becomes a parent of the 1389 * "previous" lwb zio, such that the new lwb's zio cannot 1390 * complete until the "previous" lwb's zio completes. 1391 * 1392 * This is required by the semantics of zil_commit(); the commit 1393 * waiters attached to the lwbs will be woken in the lwb zio's 1394 * completion callback, so this zio dependency graph ensures the 1395 * waiters are woken in the correct order (the same order the 1396 * lwbs were created). 1397 */ 1398 if (last_lwb_opened != NULL && 1399 last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) { 1400 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1401 last_lwb_opened->lwb_state == LWB_STATE_ISSUED || 1402 last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE); 1403 1404 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL); 1405 zio_add_child(lwb->lwb_root_zio, 1406 last_lwb_opened->lwb_root_zio); 1407 1408 /* 1409 * If the previous lwb's write hasn't already completed, 1410 * we also want to order the completion of the lwb write 1411 * zios (above, we only order the completion of the lwb 1412 * root zios). This is required because of how we can 1413 * defer the DKIOCFLUSHWRITECACHE commands for each lwb. 1414 * 1415 * When the DKIOCFLUSHWRITECACHE commands are deferred, 1416 * the previous lwb will rely on this lwb to flush the 1417 * vdevs written to by that previous lwb. Thus, we need 1418 * to ensure this lwb doesn't issue the flush until 1419 * after the previous lwb's write completes. We ensure 1420 * this ordering by setting the zio parent/child 1421 * relationship here. 1422 * 1423 * Without this relationship on the lwb's write zio, 1424 * it's possible for this lwb's write to complete prior 1425 * to the previous lwb's write completing; and thus, the 1426 * vdevs for the previous lwb would be flushed prior to 1427 * that lwb's data being written to those vdevs (the 1428 * vdevs are flushed in the lwb write zio's completion 1429 * handler, zil_lwb_write_done()). 1430 */ 1431 if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) { 1432 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1433 last_lwb_opened->lwb_state == LWB_STATE_ISSUED); 1434 1435 ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL); 1436 zio_add_child(lwb->lwb_write_zio, 1437 last_lwb_opened->lwb_write_zio); 1438 } 1439 } 1440 } 1441 1442 1443 /* 1444 * This function's purpose is to "open" an lwb such that it is ready to 1445 * accept new itxs being committed to it. To do this, the lwb's zio 1446 * structures are created, and linked to the lwb. This function is 1447 * idempotent; if the passed in lwb has already been opened, this 1448 * function is essentially a no-op. 1449 */ 1450 static void 1451 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb) 1452 { 1453 zbookmark_phys_t zb; 1454 zio_priority_t prio; 1455 1456 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1457 ASSERT3P(lwb, !=, NULL); 1458 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED); 1459 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED); 1460 1461 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 1462 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 1463 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 1464 1465 /* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */ 1466 mutex_enter(&zilog->zl_lock); 1467 if (lwb->lwb_root_zio == NULL) { 1468 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, 1469 BP_GET_LSIZE(&lwb->lwb_blk)); 1470 1471 if (!lwb->lwb_fastwrite) { 1472 metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk); 1473 lwb->lwb_fastwrite = 1; 1474 } 1475 1476 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk) 1477 prio = ZIO_PRIORITY_SYNC_WRITE; 1478 else 1479 prio = ZIO_PRIORITY_ASYNC_WRITE; 1480 1481 lwb->lwb_root_zio = zio_root(zilog->zl_spa, 1482 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL); 1483 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1484 1485 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, 1486 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd, 1487 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb, 1488 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb); 1489 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1490 1491 lwb->lwb_state = LWB_STATE_OPENED; 1492 1493 zil_lwb_set_zio_dependency(zilog, lwb); 1494 zilog->zl_last_lwb_opened = lwb; 1495 } 1496 mutex_exit(&zilog->zl_lock); 1497 1498 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1499 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1500 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1501 } 1502 1503 /* 1504 * Define a limited set of intent log block sizes. 1505 * 1506 * These must be a multiple of 4KB. Note only the amount used (again 1507 * aligned to 4KB) actually gets written. However, we can't always just 1508 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 1509 */ 1510 static const struct { 1511 uint64_t limit; 1512 uint64_t blksz; 1513 } zil_block_buckets[] = { 1514 { 4096, 4096 }, /* non TX_WRITE */ 1515 { 8192 + 4096, 8192 + 4096 }, /* database */ 1516 { 32768 + 4096, 32768 + 4096 }, /* NFS writes */ 1517 { 65536 + 4096, 65536 + 4096 }, /* 64KB writes */ 1518 { 131072, 131072 }, /* < 128KB writes */ 1519 { 131072 +4096, 65536 + 4096 }, /* 128KB writes */ 1520 { UINT64_MAX, SPA_OLD_MAXBLOCKSIZE}, /* > 128KB writes */ 1521 }; 1522 1523 /* 1524 * Maximum block size used by the ZIL. This is picked up when the ZIL is 1525 * initialized. Otherwise this should not be used directly; see 1526 * zl_max_block_size instead. 1527 */ 1528 static int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE; 1529 1530 /* 1531 * Start a log block write and advance to the next log block. 1532 * Calls are serialized. 1533 */ 1534 static lwb_t * 1535 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb) 1536 { 1537 lwb_t *nlwb = NULL; 1538 zil_chain_t *zilc; 1539 spa_t *spa = zilog->zl_spa; 1540 blkptr_t *bp; 1541 dmu_tx_t *tx; 1542 uint64_t txg; 1543 uint64_t zil_blksz, wsz; 1544 int i, error; 1545 boolean_t slog; 1546 1547 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1548 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1549 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1550 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1551 1552 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1553 zilc = (zil_chain_t *)lwb->lwb_buf; 1554 bp = &zilc->zc_next_blk; 1555 } else { 1556 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 1557 bp = &zilc->zc_next_blk; 1558 } 1559 1560 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 1561 1562 /* 1563 * Allocate the next block and save its address in this block 1564 * before writing it in order to establish the log chain. 1565 * Note that if the allocation of nlwb synced before we wrote 1566 * the block that points at it (lwb), we'd leak it if we crashed. 1567 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 1568 * We dirty the dataset to ensure that zil_sync() will be called 1569 * to clean up in the event of allocation failure or I/O failure. 1570 */ 1571 1572 tx = dmu_tx_create(zilog->zl_os); 1573 1574 /* 1575 * Since we are not going to create any new dirty data, and we 1576 * can even help with clearing the existing dirty data, we 1577 * should not be subject to the dirty data based delays. We 1578 * use TXG_NOTHROTTLE to bypass the delay mechanism. 1579 */ 1580 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); 1581 1582 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1583 txg = dmu_tx_get_txg(tx); 1584 1585 lwb->lwb_tx = tx; 1586 1587 /* 1588 * Log blocks are pre-allocated. Here we select the size of the next 1589 * block, based on size used in the last block. 1590 * - first find the smallest bucket that will fit the block from a 1591 * limited set of block sizes. This is because it's faster to write 1592 * blocks allocated from the same metaslab as they are adjacent or 1593 * close. 1594 * - next find the maximum from the new suggested size and an array of 1595 * previous sizes. This lessens a picket fence effect of wrongly 1596 * guessing the size if we have a stream of say 2k, 64k, 2k, 64k 1597 * requests. 1598 * 1599 * Note we only write what is used, but we can't just allocate 1600 * the maximum block size because we can exhaust the available 1601 * pool log space. 1602 */ 1603 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1604 for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++) 1605 continue; 1606 zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size); 1607 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1608 for (i = 0; i < ZIL_PREV_BLKS; i++) 1609 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1610 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1611 1612 BP_ZERO(bp); 1613 error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog); 1614 if (slog) { 1615 ZIL_STAT_BUMP(zil_itx_metaslab_slog_count); 1616 ZIL_STAT_INCR(zil_itx_metaslab_slog_bytes, lwb->lwb_nused); 1617 } else { 1618 ZIL_STAT_BUMP(zil_itx_metaslab_normal_count); 1619 ZIL_STAT_INCR(zil_itx_metaslab_normal_bytes, lwb->lwb_nused); 1620 } 1621 if (error == 0) { 1622 ASSERT3U(bp->blk_birth, ==, txg); 1623 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1624 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1625 1626 /* 1627 * Allocate a new log write block (lwb). 1628 */ 1629 nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE); 1630 } 1631 1632 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1633 /* For Slim ZIL only write what is used. */ 1634 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1635 ASSERT3U(wsz, <=, lwb->lwb_sz); 1636 zio_shrink(lwb->lwb_write_zio, wsz); 1637 1638 } else { 1639 wsz = lwb->lwb_sz; 1640 } 1641 1642 zilc->zc_pad = 0; 1643 zilc->zc_nused = lwb->lwb_nused; 1644 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1645 1646 /* 1647 * clear unused data for security 1648 */ 1649 memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused); 1650 1651 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER); 1652 1653 zil_lwb_add_block(lwb, &lwb->lwb_blk); 1654 lwb->lwb_issued_timestamp = gethrtime(); 1655 lwb->lwb_state = LWB_STATE_ISSUED; 1656 1657 zio_nowait(lwb->lwb_root_zio); 1658 zio_nowait(lwb->lwb_write_zio); 1659 1660 /* 1661 * If there was an allocation failure then nlwb will be null which 1662 * forces a txg_wait_synced(). 1663 */ 1664 return (nlwb); 1665 } 1666 1667 /* 1668 * Maximum amount of write data that can be put into single log block. 1669 */ 1670 uint64_t 1671 zil_max_log_data(zilog_t *zilog) 1672 { 1673 return (zilog->zl_max_block_size - 1674 sizeof (zil_chain_t) - sizeof (lr_write_t)); 1675 } 1676 1677 /* 1678 * Maximum amount of log space we agree to waste to reduce number of 1679 * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%). 1680 */ 1681 static inline uint64_t 1682 zil_max_waste_space(zilog_t *zilog) 1683 { 1684 return (zil_max_log_data(zilog) / 8); 1685 } 1686 1687 /* 1688 * Maximum amount of write data for WR_COPIED. For correctness, consumers 1689 * must fall back to WR_NEED_COPY if we can't fit the entire record into one 1690 * maximum sized log block, because each WR_COPIED record must fit in a 1691 * single log block. For space efficiency, we want to fit two records into a 1692 * max-sized log block. 1693 */ 1694 uint64_t 1695 zil_max_copied_data(zilog_t *zilog) 1696 { 1697 return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 - 1698 sizeof (lr_write_t)); 1699 } 1700 1701 static lwb_t * 1702 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1703 { 1704 lr_t *lrcb, *lrc; 1705 lr_write_t *lrwb, *lrw; 1706 char *lr_buf; 1707 uint64_t dlen, dnow, dpad, lwb_sp, reclen, txg, max_log_data; 1708 1709 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1710 ASSERT3P(lwb, !=, NULL); 1711 ASSERT3P(lwb->lwb_buf, !=, NULL); 1712 1713 zil_lwb_write_open(zilog, lwb); 1714 1715 lrc = &itx->itx_lr; 1716 lrw = (lr_write_t *)lrc; 1717 1718 /* 1719 * A commit itx doesn't represent any on-disk state; instead 1720 * it's simply used as a place holder on the commit list, and 1721 * provides a mechanism for attaching a "commit waiter" onto the 1722 * correct lwb (such that the waiter can be signalled upon 1723 * completion of that lwb). Thus, we don't process this itx's 1724 * log record if it's a commit itx (these itx's don't have log 1725 * records), and instead link the itx's waiter onto the lwb's 1726 * list of waiters. 1727 * 1728 * For more details, see the comment above zil_commit(). 1729 */ 1730 if (lrc->lrc_txtype == TX_COMMIT) { 1731 mutex_enter(&zilog->zl_lock); 1732 zil_commit_waiter_link_lwb(itx->itx_private, lwb); 1733 itx->itx_private = NULL; 1734 mutex_exit(&zilog->zl_lock); 1735 return (lwb); 1736 } 1737 1738 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { 1739 dlen = P2ROUNDUP_TYPED( 1740 lrw->lr_length, sizeof (uint64_t), uint64_t); 1741 dpad = dlen - lrw->lr_length; 1742 } else { 1743 dlen = dpad = 0; 1744 } 1745 reclen = lrc->lrc_reclen; 1746 zilog->zl_cur_used += (reclen + dlen); 1747 txg = lrc->lrc_txg; 1748 1749 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen)); 1750 1751 cont: 1752 /* 1753 * If this record won't fit in the current log block, start a new one. 1754 * For WR_NEED_COPY optimize layout for minimal number of chunks. 1755 */ 1756 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1757 max_log_data = zil_max_log_data(zilog); 1758 if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1759 lwb_sp < zil_max_waste_space(zilog) && 1760 (dlen % max_log_data == 0 || 1761 lwb_sp < reclen + dlen % max_log_data))) { 1762 lwb = zil_lwb_write_issue(zilog, lwb); 1763 if (lwb == NULL) 1764 return (NULL); 1765 zil_lwb_write_open(zilog, lwb); 1766 ASSERT(LWB_EMPTY(lwb)); 1767 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1768 1769 /* 1770 * There must be enough space in the new, empty log block to 1771 * hold reclen. For WR_COPIED, we need to fit the whole 1772 * record in one block, and reclen is the header size + the 1773 * data size. For WR_NEED_COPY, we can create multiple 1774 * records, splitting the data into multiple blocks, so we 1775 * only need to fit one word of data per block; in this case 1776 * reclen is just the header size (no data). 1777 */ 1778 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp); 1779 } 1780 1781 dnow = MIN(dlen, lwb_sp - reclen); 1782 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1783 memcpy(lr_buf, lrc, reclen); 1784 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */ 1785 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */ 1786 1787 ZIL_STAT_BUMP(zil_itx_count); 1788 1789 /* 1790 * If it's a write, fetch the data or get its blkptr as appropriate. 1791 */ 1792 if (lrc->lrc_txtype == TX_WRITE) { 1793 if (txg > spa_freeze_txg(zilog->zl_spa)) 1794 txg_wait_synced(zilog->zl_dmu_pool, txg); 1795 if (itx->itx_wr_state == WR_COPIED) { 1796 ZIL_STAT_BUMP(zil_itx_copied_count); 1797 ZIL_STAT_INCR(zil_itx_copied_bytes, lrw->lr_length); 1798 } else { 1799 char *dbuf; 1800 int error; 1801 1802 if (itx->itx_wr_state == WR_NEED_COPY) { 1803 dbuf = lr_buf + reclen; 1804 lrcb->lrc_reclen += dnow; 1805 if (lrwb->lr_length > dnow) 1806 lrwb->lr_length = dnow; 1807 lrw->lr_offset += dnow; 1808 lrw->lr_length -= dnow; 1809 ZIL_STAT_BUMP(zil_itx_needcopy_count); 1810 ZIL_STAT_INCR(zil_itx_needcopy_bytes, dnow); 1811 } else { 1812 ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT); 1813 dbuf = NULL; 1814 ZIL_STAT_BUMP(zil_itx_indirect_count); 1815 ZIL_STAT_INCR(zil_itx_indirect_bytes, 1816 lrw->lr_length); 1817 } 1818 1819 /* 1820 * We pass in the "lwb_write_zio" rather than 1821 * "lwb_root_zio" so that the "lwb_write_zio" 1822 * becomes the parent of any zio's created by 1823 * the "zl_get_data" callback. The vdevs are 1824 * flushed after the "lwb_write_zio" completes, 1825 * so we want to make sure that completion 1826 * callback waits for these additional zio's, 1827 * such that the vdevs used by those zio's will 1828 * be included in the lwb's vdev tree, and those 1829 * vdevs will be properly flushed. If we passed 1830 * in "lwb_root_zio" here, then these additional 1831 * vdevs may not be flushed; e.g. if these zio's 1832 * completed after "lwb_write_zio" completed. 1833 */ 1834 error = zilog->zl_get_data(itx->itx_private, 1835 itx->itx_gen, lrwb, dbuf, lwb, 1836 lwb->lwb_write_zio); 1837 if (dbuf != NULL && error == 0 && dnow == dlen) 1838 /* Zero any padding bytes in the last block. */ 1839 memset((char *)dbuf + lrwb->lr_length, 0, dpad); 1840 1841 if (error == EIO) { 1842 txg_wait_synced(zilog->zl_dmu_pool, txg); 1843 return (lwb); 1844 } 1845 if (error != 0) { 1846 ASSERT(error == ENOENT || error == EEXIST || 1847 error == EALREADY); 1848 return (lwb); 1849 } 1850 } 1851 } 1852 1853 /* 1854 * We're actually making an entry, so update lrc_seq to be the 1855 * log record sequence number. Note that this is generally not 1856 * equal to the itx sequence number because not all transactions 1857 * are synchronous, and sometimes spa_sync() gets there first. 1858 */ 1859 lrcb->lrc_seq = ++zilog->zl_lr_seq; 1860 lwb->lwb_nused += reclen + dnow; 1861 1862 zil_lwb_add_txg(lwb, txg); 1863 1864 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1865 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1866 1867 dlen -= dnow; 1868 if (dlen > 0) { 1869 zilog->zl_cur_used += reclen; 1870 goto cont; 1871 } 1872 1873 return (lwb); 1874 } 1875 1876 itx_t * 1877 zil_itx_create(uint64_t txtype, size_t olrsize) 1878 { 1879 size_t itxsize, lrsize; 1880 itx_t *itx; 1881 1882 lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t); 1883 itxsize = offsetof(itx_t, itx_lr) + lrsize; 1884 1885 itx = zio_data_buf_alloc(itxsize); 1886 itx->itx_lr.lrc_txtype = txtype; 1887 itx->itx_lr.lrc_reclen = lrsize; 1888 itx->itx_lr.lrc_seq = 0; /* defensive */ 1889 memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize); 1890 itx->itx_sync = B_TRUE; /* default is synchronous */ 1891 itx->itx_callback = NULL; 1892 itx->itx_callback_data = NULL; 1893 itx->itx_size = itxsize; 1894 1895 return (itx); 1896 } 1897 1898 void 1899 zil_itx_destroy(itx_t *itx) 1900 { 1901 IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL); 1902 IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 1903 1904 if (itx->itx_callback != NULL) 1905 itx->itx_callback(itx->itx_callback_data); 1906 1907 zio_data_buf_free(itx, itx->itx_size); 1908 } 1909 1910 /* 1911 * Free up the sync and async itxs. The itxs_t has already been detached 1912 * so no locks are needed. 1913 */ 1914 static void 1915 zil_itxg_clean(void *arg) 1916 { 1917 itx_t *itx; 1918 list_t *list; 1919 avl_tree_t *t; 1920 void *cookie; 1921 itxs_t *itxs = arg; 1922 itx_async_node_t *ian; 1923 1924 list = &itxs->i_sync_list; 1925 while ((itx = list_head(list)) != NULL) { 1926 /* 1927 * In the general case, commit itxs will not be found 1928 * here, as they'll be committed to an lwb via 1929 * zil_lwb_commit(), and free'd in that function. Having 1930 * said that, it is still possible for commit itxs to be 1931 * found here, due to the following race: 1932 * 1933 * - a thread calls zil_commit() which assigns the 1934 * commit itx to a per-txg i_sync_list 1935 * - zil_itxg_clean() is called (e.g. via spa_sync()) 1936 * while the waiter is still on the i_sync_list 1937 * 1938 * There's nothing to prevent syncing the txg while the 1939 * waiter is on the i_sync_list. This normally doesn't 1940 * happen because spa_sync() is slower than zil_commit(), 1941 * but if zil_commit() calls txg_wait_synced() (e.g. 1942 * because zil_create() or zil_commit_writer_stall() is 1943 * called) we will hit this case. 1944 */ 1945 if (itx->itx_lr.lrc_txtype == TX_COMMIT) 1946 zil_commit_waiter_skip(itx->itx_private); 1947 1948 list_remove(list, itx); 1949 zil_itx_destroy(itx); 1950 } 1951 1952 cookie = NULL; 1953 t = &itxs->i_async_tree; 1954 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1955 list = &ian->ia_list; 1956 while ((itx = list_head(list)) != NULL) { 1957 list_remove(list, itx); 1958 /* commit itxs should never be on the async lists. */ 1959 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 1960 zil_itx_destroy(itx); 1961 } 1962 list_destroy(list); 1963 kmem_free(ian, sizeof (itx_async_node_t)); 1964 } 1965 avl_destroy(t); 1966 1967 kmem_free(itxs, sizeof (itxs_t)); 1968 } 1969 1970 static int 1971 zil_aitx_compare(const void *x1, const void *x2) 1972 { 1973 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1974 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1975 1976 return (TREE_CMP(o1, o2)); 1977 } 1978 1979 /* 1980 * Remove all async itx with the given oid. 1981 */ 1982 void 1983 zil_remove_async(zilog_t *zilog, uint64_t oid) 1984 { 1985 uint64_t otxg, txg; 1986 itx_async_node_t *ian; 1987 avl_tree_t *t; 1988 avl_index_t where; 1989 list_t clean_list; 1990 itx_t *itx; 1991 1992 ASSERT(oid != 0); 1993 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1994 1995 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1996 otxg = ZILTEST_TXG; 1997 else 1998 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1999 2000 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2001 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2002 2003 mutex_enter(&itxg->itxg_lock); 2004 if (itxg->itxg_txg != txg) { 2005 mutex_exit(&itxg->itxg_lock); 2006 continue; 2007 } 2008 2009 /* 2010 * Locate the object node and append its list. 2011 */ 2012 t = &itxg->itxg_itxs->i_async_tree; 2013 ian = avl_find(t, &oid, &where); 2014 if (ian != NULL) 2015 list_move_tail(&clean_list, &ian->ia_list); 2016 mutex_exit(&itxg->itxg_lock); 2017 } 2018 while ((itx = list_head(&clean_list)) != NULL) { 2019 list_remove(&clean_list, itx); 2020 /* commit itxs should never be on the async lists. */ 2021 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 2022 zil_itx_destroy(itx); 2023 } 2024 list_destroy(&clean_list); 2025 } 2026 2027 void 2028 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 2029 { 2030 uint64_t txg; 2031 itxg_t *itxg; 2032 itxs_t *itxs, *clean = NULL; 2033 2034 /* 2035 * Ensure the data of a renamed file is committed before the rename. 2036 */ 2037 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 2038 zil_async_to_sync(zilog, itx->itx_oid); 2039 2040 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 2041 txg = ZILTEST_TXG; 2042 else 2043 txg = dmu_tx_get_txg(tx); 2044 2045 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2046 mutex_enter(&itxg->itxg_lock); 2047 itxs = itxg->itxg_itxs; 2048 if (itxg->itxg_txg != txg) { 2049 if (itxs != NULL) { 2050 /* 2051 * The zil_clean callback hasn't got around to cleaning 2052 * this itxg. Save the itxs for release below. 2053 * This should be rare. 2054 */ 2055 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for " 2056 "txg %llu", (u_longlong_t)itxg->itxg_txg); 2057 clean = itxg->itxg_itxs; 2058 } 2059 itxg->itxg_txg = txg; 2060 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), 2061 KM_SLEEP); 2062 2063 list_create(&itxs->i_sync_list, sizeof (itx_t), 2064 offsetof(itx_t, itx_node)); 2065 avl_create(&itxs->i_async_tree, zil_aitx_compare, 2066 sizeof (itx_async_node_t), 2067 offsetof(itx_async_node_t, ia_node)); 2068 } 2069 if (itx->itx_sync) { 2070 list_insert_tail(&itxs->i_sync_list, itx); 2071 } else { 2072 avl_tree_t *t = &itxs->i_async_tree; 2073 uint64_t foid = 2074 LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid); 2075 itx_async_node_t *ian; 2076 avl_index_t where; 2077 2078 ian = avl_find(t, &foid, &where); 2079 if (ian == NULL) { 2080 ian = kmem_alloc(sizeof (itx_async_node_t), 2081 KM_SLEEP); 2082 list_create(&ian->ia_list, sizeof (itx_t), 2083 offsetof(itx_t, itx_node)); 2084 ian->ia_foid = foid; 2085 avl_insert(t, ian, where); 2086 } 2087 list_insert_tail(&ian->ia_list, itx); 2088 } 2089 2090 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 2091 2092 /* 2093 * We don't want to dirty the ZIL using ZILTEST_TXG, because 2094 * zil_clean() will never be called using ZILTEST_TXG. Thus, we 2095 * need to be careful to always dirty the ZIL using the "real" 2096 * TXG (not itxg_txg) even when the SPA is frozen. 2097 */ 2098 zilog_dirty(zilog, dmu_tx_get_txg(tx)); 2099 mutex_exit(&itxg->itxg_lock); 2100 2101 /* Release the old itxs now we've dropped the lock */ 2102 if (clean != NULL) 2103 zil_itxg_clean(clean); 2104 } 2105 2106 /* 2107 * If there are any in-memory intent log transactions which have now been 2108 * synced then start up a taskq to free them. We should only do this after we 2109 * have written out the uberblocks (i.e. txg has been committed) so that 2110 * don't inadvertently clean out in-memory log records that would be required 2111 * by zil_commit(). 2112 */ 2113 void 2114 zil_clean(zilog_t *zilog, uint64_t synced_txg) 2115 { 2116 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 2117 itxs_t *clean_me; 2118 2119 ASSERT3U(synced_txg, <, ZILTEST_TXG); 2120 2121 mutex_enter(&itxg->itxg_lock); 2122 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 2123 mutex_exit(&itxg->itxg_lock); 2124 return; 2125 } 2126 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 2127 ASSERT3U(itxg->itxg_txg, !=, 0); 2128 clean_me = itxg->itxg_itxs; 2129 itxg->itxg_itxs = NULL; 2130 itxg->itxg_txg = 0; 2131 mutex_exit(&itxg->itxg_lock); 2132 /* 2133 * Preferably start a task queue to free up the old itxs but 2134 * if taskq_dispatch can't allocate resources to do that then 2135 * free it in-line. This should be rare. Note, using TQ_SLEEP 2136 * created a bad performance problem. 2137 */ 2138 ASSERT3P(zilog->zl_dmu_pool, !=, NULL); 2139 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL); 2140 taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq, 2141 zil_itxg_clean, clean_me, TQ_NOSLEEP); 2142 if (id == TASKQID_INVALID) 2143 zil_itxg_clean(clean_me); 2144 } 2145 2146 /* 2147 * This function will traverse the queue of itxs that need to be 2148 * committed, and move them onto the ZIL's zl_itx_commit_list. 2149 */ 2150 static void 2151 zil_get_commit_list(zilog_t *zilog) 2152 { 2153 uint64_t otxg, txg; 2154 list_t *commit_list = &zilog->zl_itx_commit_list; 2155 2156 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2157 2158 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2159 otxg = ZILTEST_TXG; 2160 else 2161 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2162 2163 /* 2164 * This is inherently racy, since there is nothing to prevent 2165 * the last synced txg from changing. That's okay since we'll 2166 * only commit things in the future. 2167 */ 2168 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2169 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2170 2171 mutex_enter(&itxg->itxg_lock); 2172 if (itxg->itxg_txg != txg) { 2173 mutex_exit(&itxg->itxg_lock); 2174 continue; 2175 } 2176 2177 /* 2178 * If we're adding itx records to the zl_itx_commit_list, 2179 * then the zil better be dirty in this "txg". We can assert 2180 * that here since we're holding the itxg_lock which will 2181 * prevent spa_sync from cleaning it. Once we add the itxs 2182 * to the zl_itx_commit_list we must commit it to disk even 2183 * if it's unnecessary (i.e. the txg was synced). 2184 */ 2185 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 2186 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 2187 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 2188 2189 mutex_exit(&itxg->itxg_lock); 2190 } 2191 } 2192 2193 /* 2194 * Move the async itxs for a specified object to commit into sync lists. 2195 */ 2196 void 2197 zil_async_to_sync(zilog_t *zilog, uint64_t foid) 2198 { 2199 uint64_t otxg, txg; 2200 itx_async_node_t *ian; 2201 avl_tree_t *t; 2202 avl_index_t where; 2203 2204 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 2205 otxg = ZILTEST_TXG; 2206 else 2207 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 2208 2209 /* 2210 * This is inherently racy, since there is nothing to prevent 2211 * the last synced txg from changing. 2212 */ 2213 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 2214 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 2215 2216 mutex_enter(&itxg->itxg_lock); 2217 if (itxg->itxg_txg != txg) { 2218 mutex_exit(&itxg->itxg_lock); 2219 continue; 2220 } 2221 2222 /* 2223 * If a foid is specified then find that node and append its 2224 * list. Otherwise walk the tree appending all the lists 2225 * to the sync list. We add to the end rather than the 2226 * beginning to ensure the create has happened. 2227 */ 2228 t = &itxg->itxg_itxs->i_async_tree; 2229 if (foid != 0) { 2230 ian = avl_find(t, &foid, &where); 2231 if (ian != NULL) { 2232 list_move_tail(&itxg->itxg_itxs->i_sync_list, 2233 &ian->ia_list); 2234 } 2235 } else { 2236 void *cookie = NULL; 2237 2238 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 2239 list_move_tail(&itxg->itxg_itxs->i_sync_list, 2240 &ian->ia_list); 2241 list_destroy(&ian->ia_list); 2242 kmem_free(ian, sizeof (itx_async_node_t)); 2243 } 2244 } 2245 mutex_exit(&itxg->itxg_lock); 2246 } 2247 } 2248 2249 /* 2250 * This function will prune commit itxs that are at the head of the 2251 * commit list (it won't prune past the first non-commit itx), and 2252 * either: a) attach them to the last lwb that's still pending 2253 * completion, or b) skip them altogether. 2254 * 2255 * This is used as a performance optimization to prevent commit itxs 2256 * from generating new lwbs when it's unnecessary to do so. 2257 */ 2258 static void 2259 zil_prune_commit_list(zilog_t *zilog) 2260 { 2261 itx_t *itx; 2262 2263 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2264 2265 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { 2266 lr_t *lrc = &itx->itx_lr; 2267 if (lrc->lrc_txtype != TX_COMMIT) 2268 break; 2269 2270 mutex_enter(&zilog->zl_lock); 2271 2272 lwb_t *last_lwb = zilog->zl_last_lwb_opened; 2273 if (last_lwb == NULL || 2274 last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) { 2275 /* 2276 * All of the itxs this waiter was waiting on 2277 * must have already completed (or there were 2278 * never any itx's for it to wait on), so it's 2279 * safe to skip this waiter and mark it done. 2280 */ 2281 zil_commit_waiter_skip(itx->itx_private); 2282 } else { 2283 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb); 2284 itx->itx_private = NULL; 2285 } 2286 2287 mutex_exit(&zilog->zl_lock); 2288 2289 list_remove(&zilog->zl_itx_commit_list, itx); 2290 zil_itx_destroy(itx); 2291 } 2292 2293 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 2294 } 2295 2296 static void 2297 zil_commit_writer_stall(zilog_t *zilog) 2298 { 2299 /* 2300 * When zio_alloc_zil() fails to allocate the next lwb block on 2301 * disk, we must call txg_wait_synced() to ensure all of the 2302 * lwbs in the zilog's zl_lwb_list are synced and then freed (in 2303 * zil_sync()), such that any subsequent ZIL writer (i.e. a call 2304 * to zil_process_commit_list()) will have to call zil_create(), 2305 * and start a new ZIL chain. 2306 * 2307 * Since zil_alloc_zil() failed, the lwb that was previously 2308 * issued does not have a pointer to the "next" lwb on disk. 2309 * Thus, if another ZIL writer thread was to allocate the "next" 2310 * on-disk lwb, that block could be leaked in the event of a 2311 * crash (because the previous lwb on-disk would not point to 2312 * it). 2313 * 2314 * We must hold the zilog's zl_issuer_lock while we do this, to 2315 * ensure no new threads enter zil_process_commit_list() until 2316 * all lwb's in the zl_lwb_list have been synced and freed 2317 * (which is achieved via the txg_wait_synced() call). 2318 */ 2319 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2320 txg_wait_synced(zilog->zl_dmu_pool, 0); 2321 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 2322 } 2323 2324 /* 2325 * This function will traverse the commit list, creating new lwbs as 2326 * needed, and committing the itxs from the commit list to these newly 2327 * created lwbs. Additionally, as a new lwb is created, the previous 2328 * lwb will be issued to the zio layer to be written to disk. 2329 */ 2330 static void 2331 zil_process_commit_list(zilog_t *zilog) 2332 { 2333 spa_t *spa = zilog->zl_spa; 2334 list_t nolwb_itxs; 2335 list_t nolwb_waiters; 2336 lwb_t *lwb; 2337 itx_t *itx; 2338 2339 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 2340 2341 /* 2342 * Return if there's nothing to commit before we dirty the fs by 2343 * calling zil_create(). 2344 */ 2345 if (list_head(&zilog->zl_itx_commit_list) == NULL) 2346 return; 2347 2348 list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); 2349 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t), 2350 offsetof(zil_commit_waiter_t, zcw_node)); 2351 2352 lwb = list_tail(&zilog->zl_lwb_list); 2353 if (lwb == NULL) { 2354 lwb = zil_create(zilog); 2355 } else { 2356 /* 2357 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will 2358 * have already been created (zl_lwb_list not empty). 2359 */ 2360 zil_commit_activate_saxattr_feature(zilog); 2361 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2362 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 2363 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 2364 } 2365 2366 while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { 2367 lr_t *lrc = &itx->itx_lr; 2368 uint64_t txg = lrc->lrc_txg; 2369 2370 ASSERT3U(txg, !=, 0); 2371 2372 if (lrc->lrc_txtype == TX_COMMIT) { 2373 DTRACE_PROBE2(zil__process__commit__itx, 2374 zilog_t *, zilog, itx_t *, itx); 2375 } else { 2376 DTRACE_PROBE2(zil__process__normal__itx, 2377 zilog_t *, zilog, itx_t *, itx); 2378 } 2379 2380 list_remove(&zilog->zl_itx_commit_list, itx); 2381 2382 boolean_t synced = txg <= spa_last_synced_txg(spa); 2383 boolean_t frozen = txg > spa_freeze_txg(spa); 2384 2385 /* 2386 * If the txg of this itx has already been synced out, then 2387 * we don't need to commit this itx to an lwb. This is 2388 * because the data of this itx will have already been 2389 * written to the main pool. This is inherently racy, and 2390 * it's still ok to commit an itx whose txg has already 2391 * been synced; this will result in a write that's 2392 * unnecessary, but will do no harm. 2393 * 2394 * With that said, we always want to commit TX_COMMIT itxs 2395 * to an lwb, regardless of whether or not that itx's txg 2396 * has been synced out. We do this to ensure any OPENED lwb 2397 * will always have at least one zil_commit_waiter_t linked 2398 * to the lwb. 2399 * 2400 * As a counter-example, if we skipped TX_COMMIT itx's 2401 * whose txg had already been synced, the following 2402 * situation could occur if we happened to be racing with 2403 * spa_sync: 2404 * 2405 * 1. We commit a non-TX_COMMIT itx to an lwb, where the 2406 * itx's txg is 10 and the last synced txg is 9. 2407 * 2. spa_sync finishes syncing out txg 10. 2408 * 3. We move to the next itx in the list, it's a TX_COMMIT 2409 * whose txg is 10, so we skip it rather than committing 2410 * it to the lwb used in (1). 2411 * 2412 * If the itx that is skipped in (3) is the last TX_COMMIT 2413 * itx in the commit list, than it's possible for the lwb 2414 * used in (1) to remain in the OPENED state indefinitely. 2415 * 2416 * To prevent the above scenario from occurring, ensuring 2417 * that once an lwb is OPENED it will transition to ISSUED 2418 * and eventually DONE, we always commit TX_COMMIT itx's to 2419 * an lwb here, even if that itx's txg has already been 2420 * synced. 2421 * 2422 * Finally, if the pool is frozen, we _always_ commit the 2423 * itx. The point of freezing the pool is to prevent data 2424 * from being written to the main pool via spa_sync, and 2425 * instead rely solely on the ZIL to persistently store the 2426 * data; i.e. when the pool is frozen, the last synced txg 2427 * value can't be trusted. 2428 */ 2429 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) { 2430 if (lwb != NULL) { 2431 lwb = zil_lwb_commit(zilog, itx, lwb); 2432 2433 if (lwb == NULL) 2434 list_insert_tail(&nolwb_itxs, itx); 2435 else 2436 list_insert_tail(&lwb->lwb_itxs, itx); 2437 } else { 2438 if (lrc->lrc_txtype == TX_COMMIT) { 2439 zil_commit_waiter_link_nolwb( 2440 itx->itx_private, &nolwb_waiters); 2441 } 2442 2443 list_insert_tail(&nolwb_itxs, itx); 2444 } 2445 } else { 2446 ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT); 2447 zil_itx_destroy(itx); 2448 } 2449 } 2450 2451 if (lwb == NULL) { 2452 /* 2453 * This indicates zio_alloc_zil() failed to allocate the 2454 * "next" lwb on-disk. When this happens, we must stall 2455 * the ZIL write pipeline; see the comment within 2456 * zil_commit_writer_stall() for more details. 2457 */ 2458 zil_commit_writer_stall(zilog); 2459 2460 /* 2461 * Additionally, we have to signal and mark the "nolwb" 2462 * waiters as "done" here, since without an lwb, we 2463 * can't do this via zil_lwb_flush_vdevs_done() like 2464 * normal. 2465 */ 2466 zil_commit_waiter_t *zcw; 2467 while ((zcw = list_head(&nolwb_waiters)) != NULL) { 2468 zil_commit_waiter_skip(zcw); 2469 list_remove(&nolwb_waiters, zcw); 2470 } 2471 2472 /* 2473 * And finally, we have to destroy the itx's that 2474 * couldn't be committed to an lwb; this will also call 2475 * the itx's callback if one exists for the itx. 2476 */ 2477 while ((itx = list_head(&nolwb_itxs)) != NULL) { 2478 list_remove(&nolwb_itxs, itx); 2479 zil_itx_destroy(itx); 2480 } 2481 } else { 2482 ASSERT(list_is_empty(&nolwb_waiters)); 2483 ASSERT3P(lwb, !=, NULL); 2484 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2485 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); 2486 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); 2487 2488 /* 2489 * At this point, the ZIL block pointed at by the "lwb" 2490 * variable is in one of the following states: "closed" 2491 * or "open". 2492 * 2493 * If it's "closed", then no itxs have been committed to 2494 * it, so there's no point in issuing its zio (i.e. it's 2495 * "empty"). 2496 * 2497 * If it's "open", then it contains one or more itxs that 2498 * eventually need to be committed to stable storage. In 2499 * this case we intentionally do not issue the lwb's zio 2500 * to disk yet, and instead rely on one of the following 2501 * two mechanisms for issuing the zio: 2502 * 2503 * 1. Ideally, there will be more ZIL activity occurring 2504 * on the system, such that this function will be 2505 * immediately called again (not necessarily by the same 2506 * thread) and this lwb's zio will be issued via 2507 * zil_lwb_commit(). This way, the lwb is guaranteed to 2508 * be "full" when it is issued to disk, and we'll make 2509 * use of the lwb's size the best we can. 2510 * 2511 * 2. If there isn't sufficient ZIL activity occurring on 2512 * the system, such that this lwb's zio isn't issued via 2513 * zil_lwb_commit(), zil_commit_waiter() will issue the 2514 * lwb's zio. If this occurs, the lwb is not guaranteed 2515 * to be "full" by the time its zio is issued, and means 2516 * the size of the lwb was "too large" given the amount 2517 * of ZIL activity occurring on the system at that time. 2518 * 2519 * We do this for a couple of reasons: 2520 * 2521 * 1. To try and reduce the number of IOPs needed to 2522 * write the same number of itxs. If an lwb has space 2523 * available in its buffer for more itxs, and more itxs 2524 * will be committed relatively soon (relative to the 2525 * latency of performing a write), then it's beneficial 2526 * to wait for these "next" itxs. This way, more itxs 2527 * can be committed to stable storage with fewer writes. 2528 * 2529 * 2. To try and use the largest lwb block size that the 2530 * incoming rate of itxs can support. Again, this is to 2531 * try and pack as many itxs into as few lwbs as 2532 * possible, without significantly impacting the latency 2533 * of each individual itx. 2534 */ 2535 } 2536 } 2537 2538 /* 2539 * This function is responsible for ensuring the passed in commit waiter 2540 * (and associated commit itx) is committed to an lwb. If the waiter is 2541 * not already committed to an lwb, all itxs in the zilog's queue of 2542 * itxs will be processed. The assumption is the passed in waiter's 2543 * commit itx will found in the queue just like the other non-commit 2544 * itxs, such that when the entire queue is processed, the waiter will 2545 * have been committed to an lwb. 2546 * 2547 * The lwb associated with the passed in waiter is not guaranteed to 2548 * have been issued by the time this function completes. If the lwb is 2549 * not issued, we rely on future calls to zil_commit_writer() to issue 2550 * the lwb, or the timeout mechanism found in zil_commit_waiter(). 2551 */ 2552 static void 2553 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw) 2554 { 2555 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2556 ASSERT(spa_writeable(zilog->zl_spa)); 2557 2558 mutex_enter(&zilog->zl_issuer_lock); 2559 2560 if (zcw->zcw_lwb != NULL || zcw->zcw_done) { 2561 /* 2562 * It's possible that, while we were waiting to acquire 2563 * the "zl_issuer_lock", another thread committed this 2564 * waiter to an lwb. If that occurs, we bail out early, 2565 * without processing any of the zilog's queue of itxs. 2566 * 2567 * On certain workloads and system configurations, the 2568 * "zl_issuer_lock" can become highly contended. In an 2569 * attempt to reduce this contention, we immediately drop 2570 * the lock if the waiter has already been processed. 2571 * 2572 * We've measured this optimization to reduce CPU spent 2573 * contending on this lock by up to 5%, using a system 2574 * with 32 CPUs, low latency storage (~50 usec writes), 2575 * and 1024 threads performing sync writes. 2576 */ 2577 goto out; 2578 } 2579 2580 ZIL_STAT_BUMP(zil_commit_writer_count); 2581 2582 zil_get_commit_list(zilog); 2583 zil_prune_commit_list(zilog); 2584 zil_process_commit_list(zilog); 2585 2586 out: 2587 mutex_exit(&zilog->zl_issuer_lock); 2588 } 2589 2590 static void 2591 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw) 2592 { 2593 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2594 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2595 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 2596 2597 lwb_t *lwb = zcw->zcw_lwb; 2598 ASSERT3P(lwb, !=, NULL); 2599 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED); 2600 2601 /* 2602 * If the lwb has already been issued by another thread, we can 2603 * immediately return since there's no work to be done (the 2604 * point of this function is to issue the lwb). Additionally, we 2605 * do this prior to acquiring the zl_issuer_lock, to avoid 2606 * acquiring it when it's not necessary to do so. 2607 */ 2608 if (lwb->lwb_state == LWB_STATE_ISSUED || 2609 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2610 lwb->lwb_state == LWB_STATE_FLUSH_DONE) 2611 return; 2612 2613 /* 2614 * In order to call zil_lwb_write_issue() we must hold the 2615 * zilog's "zl_issuer_lock". We can't simply acquire that lock, 2616 * since we're already holding the commit waiter's "zcw_lock", 2617 * and those two locks are acquired in the opposite order 2618 * elsewhere. 2619 */ 2620 mutex_exit(&zcw->zcw_lock); 2621 mutex_enter(&zilog->zl_issuer_lock); 2622 mutex_enter(&zcw->zcw_lock); 2623 2624 /* 2625 * Since we just dropped and re-acquired the commit waiter's 2626 * lock, we have to re-check to see if the waiter was marked 2627 * "done" during that process. If the waiter was marked "done", 2628 * the "lwb" pointer is no longer valid (it can be free'd after 2629 * the waiter is marked "done"), so without this check we could 2630 * wind up with a use-after-free error below. 2631 */ 2632 if (zcw->zcw_done) 2633 goto out; 2634 2635 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2636 2637 /* 2638 * We've already checked this above, but since we hadn't acquired 2639 * the zilog's zl_issuer_lock, we have to perform this check a 2640 * second time while holding the lock. 2641 * 2642 * We don't need to hold the zl_lock since the lwb cannot transition 2643 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb 2644 * _can_ transition from ISSUED to DONE, but it's OK to race with 2645 * that transition since we treat the lwb the same, whether it's in 2646 * the ISSUED or DONE states. 2647 * 2648 * The important thing, is we treat the lwb differently depending on 2649 * if it's ISSUED or OPENED, and block any other threads that might 2650 * attempt to issue this lwb. For that reason we hold the 2651 * zl_issuer_lock when checking the lwb_state; we must not call 2652 * zil_lwb_write_issue() if the lwb had already been issued. 2653 * 2654 * See the comment above the lwb_state_t structure definition for 2655 * more details on the lwb states, and locking requirements. 2656 */ 2657 if (lwb->lwb_state == LWB_STATE_ISSUED || 2658 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2659 lwb->lwb_state == LWB_STATE_FLUSH_DONE) 2660 goto out; 2661 2662 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 2663 2664 /* 2665 * As described in the comments above zil_commit_waiter() and 2666 * zil_process_commit_list(), we need to issue this lwb's zio 2667 * since we've reached the commit waiter's timeout and it still 2668 * hasn't been issued. 2669 */ 2670 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb); 2671 2672 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED); 2673 2674 /* 2675 * Since the lwb's zio hadn't been issued by the time this thread 2676 * reached its timeout, we reset the zilog's "zl_cur_used" field 2677 * to influence the zil block size selection algorithm. 2678 * 2679 * By having to issue the lwb's zio here, it means the size of the 2680 * lwb was too large, given the incoming throughput of itxs. By 2681 * setting "zl_cur_used" to zero, we communicate this fact to the 2682 * block size selection algorithm, so it can take this information 2683 * into account, and potentially select a smaller size for the 2684 * next lwb block that is allocated. 2685 */ 2686 zilog->zl_cur_used = 0; 2687 2688 if (nlwb == NULL) { 2689 /* 2690 * When zil_lwb_write_issue() returns NULL, this 2691 * indicates zio_alloc_zil() failed to allocate the 2692 * "next" lwb on-disk. When this occurs, the ZIL write 2693 * pipeline must be stalled; see the comment within the 2694 * zil_commit_writer_stall() function for more details. 2695 * 2696 * We must drop the commit waiter's lock prior to 2697 * calling zil_commit_writer_stall() or else we can wind 2698 * up with the following deadlock: 2699 * 2700 * - This thread is waiting for the txg to sync while 2701 * holding the waiter's lock; txg_wait_synced() is 2702 * used within txg_commit_writer_stall(). 2703 * 2704 * - The txg can't sync because it is waiting for this 2705 * lwb's zio callback to call dmu_tx_commit(). 2706 * 2707 * - The lwb's zio callback can't call dmu_tx_commit() 2708 * because it's blocked trying to acquire the waiter's 2709 * lock, which occurs prior to calling dmu_tx_commit() 2710 */ 2711 mutex_exit(&zcw->zcw_lock); 2712 zil_commit_writer_stall(zilog); 2713 mutex_enter(&zcw->zcw_lock); 2714 } 2715 2716 out: 2717 mutex_exit(&zilog->zl_issuer_lock); 2718 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2719 } 2720 2721 /* 2722 * This function is responsible for performing the following two tasks: 2723 * 2724 * 1. its primary responsibility is to block until the given "commit 2725 * waiter" is considered "done". 2726 * 2727 * 2. its secondary responsibility is to issue the zio for the lwb that 2728 * the given "commit waiter" is waiting on, if this function has 2729 * waited "long enough" and the lwb is still in the "open" state. 2730 * 2731 * Given a sufficient amount of itxs being generated and written using 2732 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit() 2733 * function. If this does not occur, this secondary responsibility will 2734 * ensure the lwb is issued even if there is not other synchronous 2735 * activity on the system. 2736 * 2737 * For more details, see zil_process_commit_list(); more specifically, 2738 * the comment at the bottom of that function. 2739 */ 2740 static void 2741 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw) 2742 { 2743 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2744 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2745 ASSERT(spa_writeable(zilog->zl_spa)); 2746 2747 mutex_enter(&zcw->zcw_lock); 2748 2749 /* 2750 * The timeout is scaled based on the lwb latency to avoid 2751 * significantly impacting the latency of each individual itx. 2752 * For more details, see the comment at the bottom of the 2753 * zil_process_commit_list() function. 2754 */ 2755 int pct = MAX(zfs_commit_timeout_pct, 1); 2756 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100; 2757 hrtime_t wakeup = gethrtime() + sleep; 2758 boolean_t timedout = B_FALSE; 2759 2760 while (!zcw->zcw_done) { 2761 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2762 2763 lwb_t *lwb = zcw->zcw_lwb; 2764 2765 /* 2766 * Usually, the waiter will have a non-NULL lwb field here, 2767 * but it's possible for it to be NULL as a result of 2768 * zil_commit() racing with spa_sync(). 2769 * 2770 * When zil_clean() is called, it's possible for the itxg 2771 * list (which may be cleaned via a taskq) to contain 2772 * commit itxs. When this occurs, the commit waiters linked 2773 * off of these commit itxs will not be committed to an 2774 * lwb. Additionally, these commit waiters will not be 2775 * marked done until zil_commit_waiter_skip() is called via 2776 * zil_itxg_clean(). 2777 * 2778 * Thus, it's possible for this commit waiter (i.e. the 2779 * "zcw" variable) to be found in this "in between" state; 2780 * where it's "zcw_lwb" field is NULL, and it hasn't yet 2781 * been skipped, so it's "zcw_done" field is still B_FALSE. 2782 */ 2783 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED); 2784 2785 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) { 2786 ASSERT3B(timedout, ==, B_FALSE); 2787 2788 /* 2789 * If the lwb hasn't been issued yet, then we 2790 * need to wait with a timeout, in case this 2791 * function needs to issue the lwb after the 2792 * timeout is reached; responsibility (2) from 2793 * the comment above this function. 2794 */ 2795 int rc = cv_timedwait_hires(&zcw->zcw_cv, 2796 &zcw->zcw_lock, wakeup, USEC2NSEC(1), 2797 CALLOUT_FLAG_ABSOLUTE); 2798 2799 if (rc != -1 || zcw->zcw_done) 2800 continue; 2801 2802 timedout = B_TRUE; 2803 zil_commit_waiter_timeout(zilog, zcw); 2804 2805 if (!zcw->zcw_done) { 2806 /* 2807 * If the commit waiter has already been 2808 * marked "done", it's possible for the 2809 * waiter's lwb structure to have already 2810 * been freed. Thus, we can only reliably 2811 * make these assertions if the waiter 2812 * isn't done. 2813 */ 2814 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2815 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED); 2816 } 2817 } else { 2818 /* 2819 * If the lwb isn't open, then it must have already 2820 * been issued. In that case, there's no need to 2821 * use a timeout when waiting for the lwb to 2822 * complete. 2823 * 2824 * Additionally, if the lwb is NULL, the waiter 2825 * will soon be signaled and marked done via 2826 * zil_clean() and zil_itxg_clean(), so no timeout 2827 * is required. 2828 */ 2829 2830 IMPLY(lwb != NULL, 2831 lwb->lwb_state == LWB_STATE_ISSUED || 2832 lwb->lwb_state == LWB_STATE_WRITE_DONE || 2833 lwb->lwb_state == LWB_STATE_FLUSH_DONE); 2834 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock); 2835 } 2836 } 2837 2838 mutex_exit(&zcw->zcw_lock); 2839 } 2840 2841 static zil_commit_waiter_t * 2842 zil_alloc_commit_waiter(void) 2843 { 2844 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP); 2845 2846 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL); 2847 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL); 2848 list_link_init(&zcw->zcw_node); 2849 zcw->zcw_lwb = NULL; 2850 zcw->zcw_done = B_FALSE; 2851 zcw->zcw_zio_error = 0; 2852 2853 return (zcw); 2854 } 2855 2856 static void 2857 zil_free_commit_waiter(zil_commit_waiter_t *zcw) 2858 { 2859 ASSERT(!list_link_active(&zcw->zcw_node)); 2860 ASSERT3P(zcw->zcw_lwb, ==, NULL); 2861 ASSERT3B(zcw->zcw_done, ==, B_TRUE); 2862 mutex_destroy(&zcw->zcw_lock); 2863 cv_destroy(&zcw->zcw_cv); 2864 kmem_cache_free(zil_zcw_cache, zcw); 2865 } 2866 2867 /* 2868 * This function is used to create a TX_COMMIT itx and assign it. This 2869 * way, it will be linked into the ZIL's list of synchronous itxs, and 2870 * then later committed to an lwb (or skipped) when 2871 * zil_process_commit_list() is called. 2872 */ 2873 static void 2874 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw) 2875 { 2876 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 2877 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 2878 2879 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t)); 2880 itx->itx_sync = B_TRUE; 2881 itx->itx_private = zcw; 2882 2883 zil_itx_assign(zilog, itx, tx); 2884 2885 dmu_tx_commit(tx); 2886 } 2887 2888 /* 2889 * Commit ZFS Intent Log transactions (itxs) to stable storage. 2890 * 2891 * When writing ZIL transactions to the on-disk representation of the 2892 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple 2893 * itxs can be committed to a single lwb. Once a lwb is written and 2894 * committed to stable storage (i.e. the lwb is written, and vdevs have 2895 * been flushed), each itx that was committed to that lwb is also 2896 * considered to be committed to stable storage. 2897 * 2898 * When an itx is committed to an lwb, the log record (lr_t) contained 2899 * by the itx is copied into the lwb's zio buffer, and once this buffer 2900 * is written to disk, it becomes an on-disk ZIL block. 2901 * 2902 * As itxs are generated, they're inserted into the ZIL's queue of 2903 * uncommitted itxs. The semantics of zil_commit() are such that it will 2904 * block until all itxs that were in the queue when it was called, are 2905 * committed to stable storage. 2906 * 2907 * If "foid" is zero, this means all "synchronous" and "asynchronous" 2908 * itxs, for all objects in the dataset, will be committed to stable 2909 * storage prior to zil_commit() returning. If "foid" is non-zero, all 2910 * "synchronous" itxs for all objects, but only "asynchronous" itxs 2911 * that correspond to the foid passed in, will be committed to stable 2912 * storage prior to zil_commit() returning. 2913 * 2914 * Generally speaking, when zil_commit() is called, the consumer doesn't 2915 * actually care about _all_ of the uncommitted itxs. Instead, they're 2916 * simply trying to waiting for a specific itx to be committed to disk, 2917 * but the interface(s) for interacting with the ZIL don't allow such 2918 * fine-grained communication. A better interface would allow a consumer 2919 * to create and assign an itx, and then pass a reference to this itx to 2920 * zil_commit(); such that zil_commit() would return as soon as that 2921 * specific itx was committed to disk (instead of waiting for _all_ 2922 * itxs to be committed). 2923 * 2924 * When a thread calls zil_commit() a special "commit itx" will be 2925 * generated, along with a corresponding "waiter" for this commit itx. 2926 * zil_commit() will wait on this waiter's CV, such that when the waiter 2927 * is marked done, and signaled, zil_commit() will return. 2928 * 2929 * This commit itx is inserted into the queue of uncommitted itxs. This 2930 * provides an easy mechanism for determining which itxs were in the 2931 * queue prior to zil_commit() having been called, and which itxs were 2932 * added after zil_commit() was called. 2933 * 2934 * The commit it is special; it doesn't have any on-disk representation. 2935 * When a commit itx is "committed" to an lwb, the waiter associated 2936 * with it is linked onto the lwb's list of waiters. Then, when that lwb 2937 * completes, each waiter on the lwb's list is marked done and signaled 2938 * -- allowing the thread waiting on the waiter to return from zil_commit(). 2939 * 2940 * It's important to point out a few critical factors that allow us 2941 * to make use of the commit itxs, commit waiters, per-lwb lists of 2942 * commit waiters, and zio completion callbacks like we're doing: 2943 * 2944 * 1. The list of waiters for each lwb is traversed, and each commit 2945 * waiter is marked "done" and signaled, in the zio completion 2946 * callback of the lwb's zio[*]. 2947 * 2948 * * Actually, the waiters are signaled in the zio completion 2949 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands 2950 * that are sent to the vdevs upon completion of the lwb zio. 2951 * 2952 * 2. When the itxs are inserted into the ZIL's queue of uncommitted 2953 * itxs, the order in which they are inserted is preserved[*]; as 2954 * itxs are added to the queue, they are added to the tail of 2955 * in-memory linked lists. 2956 * 2957 * When committing the itxs to lwbs (to be written to disk), they 2958 * are committed in the same order in which the itxs were added to 2959 * the uncommitted queue's linked list(s); i.e. the linked list of 2960 * itxs to commit is traversed from head to tail, and each itx is 2961 * committed to an lwb in that order. 2962 * 2963 * * To clarify: 2964 * 2965 * - the order of "sync" itxs is preserved w.r.t. other 2966 * "sync" itxs, regardless of the corresponding objects. 2967 * - the order of "async" itxs is preserved w.r.t. other 2968 * "async" itxs corresponding to the same object. 2969 * - the order of "async" itxs is *not* preserved w.r.t. other 2970 * "async" itxs corresponding to different objects. 2971 * - the order of "sync" itxs w.r.t. "async" itxs (or vice 2972 * versa) is *not* preserved, even for itxs that correspond 2973 * to the same object. 2974 * 2975 * For more details, see: zil_itx_assign(), zil_async_to_sync(), 2976 * zil_get_commit_list(), and zil_process_commit_list(). 2977 * 2978 * 3. The lwbs represent a linked list of blocks on disk. Thus, any 2979 * lwb cannot be considered committed to stable storage, until its 2980 * "previous" lwb is also committed to stable storage. This fact, 2981 * coupled with the fact described above, means that itxs are 2982 * committed in (roughly) the order in which they were generated. 2983 * This is essential because itxs are dependent on prior itxs. 2984 * Thus, we *must not* deem an itx as being committed to stable 2985 * storage, until *all* prior itxs have also been committed to 2986 * stable storage. 2987 * 2988 * To enforce this ordering of lwb zio's, while still leveraging as 2989 * much of the underlying storage performance as possible, we rely 2990 * on two fundamental concepts: 2991 * 2992 * 1. The creation and issuance of lwb zio's is protected by 2993 * the zilog's "zl_issuer_lock", which ensures only a single 2994 * thread is creating and/or issuing lwb's at a time 2995 * 2. The "previous" lwb is a child of the "current" lwb 2996 * (leveraging the zio parent-child dependency graph) 2997 * 2998 * By relying on this parent-child zio relationship, we can have 2999 * many lwb zio's concurrently issued to the underlying storage, 3000 * but the order in which they complete will be the same order in 3001 * which they were created. 3002 */ 3003 void 3004 zil_commit(zilog_t *zilog, uint64_t foid) 3005 { 3006 /* 3007 * We should never attempt to call zil_commit on a snapshot for 3008 * a couple of reasons: 3009 * 3010 * 1. A snapshot may never be modified, thus it cannot have any 3011 * in-flight itxs that would have modified the dataset. 3012 * 3013 * 2. By design, when zil_commit() is called, a commit itx will 3014 * be assigned to this zilog; as a result, the zilog will be 3015 * dirtied. We must not dirty the zilog of a snapshot; there's 3016 * checks in the code that enforce this invariant, and will 3017 * cause a panic if it's not upheld. 3018 */ 3019 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE); 3020 3021 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 3022 return; 3023 3024 if (!spa_writeable(zilog->zl_spa)) { 3025 /* 3026 * If the SPA is not writable, there should never be any 3027 * pending itxs waiting to be committed to disk. If that 3028 * weren't true, we'd skip writing those itxs out, and 3029 * would break the semantics of zil_commit(); thus, we're 3030 * verifying that truth before we return to the caller. 3031 */ 3032 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3033 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 3034 for (int i = 0; i < TXG_SIZE; i++) 3035 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL); 3036 return; 3037 } 3038 3039 /* 3040 * If the ZIL is suspended, we don't want to dirty it by calling 3041 * zil_commit_itx_assign() below, nor can we write out 3042 * lwbs like would be done in zil_commit_write(). Thus, we 3043 * simply rely on txg_wait_synced() to maintain the necessary 3044 * semantics, and avoid calling those functions altogether. 3045 */ 3046 if (zilog->zl_suspend > 0) { 3047 txg_wait_synced(zilog->zl_dmu_pool, 0); 3048 return; 3049 } 3050 3051 zil_commit_impl(zilog, foid); 3052 } 3053 3054 void 3055 zil_commit_impl(zilog_t *zilog, uint64_t foid) 3056 { 3057 ZIL_STAT_BUMP(zil_commit_count); 3058 3059 /* 3060 * Move the "async" itxs for the specified foid to the "sync" 3061 * queues, such that they will be later committed (or skipped) 3062 * to an lwb when zil_process_commit_list() is called. 3063 * 3064 * Since these "async" itxs must be committed prior to this 3065 * call to zil_commit returning, we must perform this operation 3066 * before we call zil_commit_itx_assign(). 3067 */ 3068 zil_async_to_sync(zilog, foid); 3069 3070 /* 3071 * We allocate a new "waiter" structure which will initially be 3072 * linked to the commit itx using the itx's "itx_private" field. 3073 * Since the commit itx doesn't represent any on-disk state, 3074 * when it's committed to an lwb, rather than copying the its 3075 * lr_t into the lwb's buffer, the commit itx's "waiter" will be 3076 * added to the lwb's list of waiters. Then, when the lwb is 3077 * committed to stable storage, each waiter in the lwb's list of 3078 * waiters will be marked "done", and signalled. 3079 * 3080 * We must create the waiter and assign the commit itx prior to 3081 * calling zil_commit_writer(), or else our specific commit itx 3082 * is not guaranteed to be committed to an lwb prior to calling 3083 * zil_commit_waiter(). 3084 */ 3085 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter(); 3086 zil_commit_itx_assign(zilog, zcw); 3087 3088 zil_commit_writer(zilog, zcw); 3089 zil_commit_waiter(zilog, zcw); 3090 3091 if (zcw->zcw_zio_error != 0) { 3092 /* 3093 * If there was an error writing out the ZIL blocks that 3094 * this thread is waiting on, then we fallback to 3095 * relying on spa_sync() to write out the data this 3096 * thread is waiting on. Obviously this has performance 3097 * implications, but the expectation is for this to be 3098 * an exceptional case, and shouldn't occur often. 3099 */ 3100 DTRACE_PROBE2(zil__commit__io__error, 3101 zilog_t *, zilog, zil_commit_waiter_t *, zcw); 3102 txg_wait_synced(zilog->zl_dmu_pool, 0); 3103 } 3104 3105 zil_free_commit_waiter(zcw); 3106 } 3107 3108 /* 3109 * Called in syncing context to free committed log blocks and update log header. 3110 */ 3111 void 3112 zil_sync(zilog_t *zilog, dmu_tx_t *tx) 3113 { 3114 zil_header_t *zh = zil_header_in_syncing_context(zilog); 3115 uint64_t txg = dmu_tx_get_txg(tx); 3116 spa_t *spa = zilog->zl_spa; 3117 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 3118 lwb_t *lwb; 3119 3120 /* 3121 * We don't zero out zl_destroy_txg, so make sure we don't try 3122 * to destroy it twice. 3123 */ 3124 if (spa_sync_pass(spa) != 1) 3125 return; 3126 3127 mutex_enter(&zilog->zl_lock); 3128 3129 ASSERT(zilog->zl_stop_sync == 0); 3130 3131 if (*replayed_seq != 0) { 3132 ASSERT(zh->zh_replay_seq < *replayed_seq); 3133 zh->zh_replay_seq = *replayed_seq; 3134 *replayed_seq = 0; 3135 } 3136 3137 if (zilog->zl_destroy_txg == txg) { 3138 blkptr_t blk = zh->zh_log; 3139 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 3140 3141 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 3142 3143 memset(zh, 0, sizeof (zil_header_t)); 3144 memset(zilog->zl_replayed_seq, 0, 3145 sizeof (zilog->zl_replayed_seq)); 3146 3147 if (zilog->zl_keep_first) { 3148 /* 3149 * If this block was part of log chain that couldn't 3150 * be claimed because a device was missing during 3151 * zil_claim(), but that device later returns, 3152 * then this block could erroneously appear valid. 3153 * To guard against this, assign a new GUID to the new 3154 * log chain so it doesn't matter what blk points to. 3155 */ 3156 zil_init_log_chain(zilog, &blk); 3157 zh->zh_log = blk; 3158 } else { 3159 /* 3160 * A destroyed ZIL chain can't contain any TX_SETSAXATTR 3161 * records. So, deactivate the feature for this dataset. 3162 * We activate it again when we start a new ZIL chain. 3163 */ 3164 if (dsl_dataset_feature_is_active(ds, 3165 SPA_FEATURE_ZILSAXATTR)) 3166 dsl_dataset_deactivate_feature(ds, 3167 SPA_FEATURE_ZILSAXATTR, tx); 3168 } 3169 } 3170 3171 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 3172 zh->zh_log = lwb->lwb_blk; 3173 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 3174 break; 3175 list_remove(&zilog->zl_lwb_list, lwb); 3176 zio_free(spa, txg, &lwb->lwb_blk); 3177 zil_free_lwb(zilog, lwb); 3178 3179 /* 3180 * If we don't have anything left in the lwb list then 3181 * we've had an allocation failure and we need to zero 3182 * out the zil_header blkptr so that we don't end 3183 * up freeing the same block twice. 3184 */ 3185 if (list_head(&zilog->zl_lwb_list) == NULL) 3186 BP_ZERO(&zh->zh_log); 3187 } 3188 3189 /* 3190 * Remove fastwrite on any blocks that have been pre-allocated for 3191 * the next commit. This prevents fastwrite counter pollution by 3192 * unused, long-lived LWBs. 3193 */ 3194 for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) { 3195 if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) { 3196 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); 3197 lwb->lwb_fastwrite = 0; 3198 } 3199 } 3200 3201 mutex_exit(&zilog->zl_lock); 3202 } 3203 3204 static int 3205 zil_lwb_cons(void *vbuf, void *unused, int kmflag) 3206 { 3207 (void) unused, (void) kmflag; 3208 lwb_t *lwb = vbuf; 3209 list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node)); 3210 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t), 3211 offsetof(zil_commit_waiter_t, zcw_node)); 3212 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare, 3213 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 3214 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 3215 return (0); 3216 } 3217 3218 static void 3219 zil_lwb_dest(void *vbuf, void *unused) 3220 { 3221 (void) unused; 3222 lwb_t *lwb = vbuf; 3223 mutex_destroy(&lwb->lwb_vdev_lock); 3224 avl_destroy(&lwb->lwb_vdev_tree); 3225 list_destroy(&lwb->lwb_waiters); 3226 list_destroy(&lwb->lwb_itxs); 3227 } 3228 3229 void 3230 zil_init(void) 3231 { 3232 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 3233 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0); 3234 3235 zil_zcw_cache = kmem_cache_create("zil_zcw_cache", 3236 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 3237 3238 zil_ksp = kstat_create("zfs", 0, "zil", "misc", 3239 KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t), 3240 KSTAT_FLAG_VIRTUAL); 3241 3242 if (zil_ksp != NULL) { 3243 zil_ksp->ks_data = &zil_stats; 3244 kstat_install(zil_ksp); 3245 } 3246 } 3247 3248 void 3249 zil_fini(void) 3250 { 3251 kmem_cache_destroy(zil_zcw_cache); 3252 kmem_cache_destroy(zil_lwb_cache); 3253 3254 if (zil_ksp != NULL) { 3255 kstat_delete(zil_ksp); 3256 zil_ksp = NULL; 3257 } 3258 } 3259 3260 void 3261 zil_set_sync(zilog_t *zilog, uint64_t sync) 3262 { 3263 zilog->zl_sync = sync; 3264 } 3265 3266 void 3267 zil_set_logbias(zilog_t *zilog, uint64_t logbias) 3268 { 3269 zilog->zl_logbias = logbias; 3270 } 3271 3272 zilog_t * 3273 zil_alloc(objset_t *os, zil_header_t *zh_phys) 3274 { 3275 zilog_t *zilog; 3276 3277 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 3278 3279 zilog->zl_header = zh_phys; 3280 zilog->zl_os = os; 3281 zilog->zl_spa = dmu_objset_spa(os); 3282 zilog->zl_dmu_pool = dmu_objset_pool(os); 3283 zilog->zl_destroy_txg = TXG_INITIAL - 1; 3284 zilog->zl_logbias = dmu_objset_logbias(os); 3285 zilog->zl_sync = dmu_objset_syncprop(os); 3286 zilog->zl_dirty_max_txg = 0; 3287 zilog->zl_last_lwb_opened = NULL; 3288 zilog->zl_last_lwb_latency = 0; 3289 zilog->zl_max_block_size = zil_maxblocksize; 3290 3291 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 3292 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL); 3293 3294 for (int i = 0; i < TXG_SIZE; i++) { 3295 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 3296 MUTEX_DEFAULT, NULL); 3297 } 3298 3299 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 3300 offsetof(lwb_t, lwb_node)); 3301 3302 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 3303 offsetof(itx_t, itx_node)); 3304 3305 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 3306 3307 return (zilog); 3308 } 3309 3310 void 3311 zil_free(zilog_t *zilog) 3312 { 3313 int i; 3314 3315 zilog->zl_stop_sync = 1; 3316 3317 ASSERT0(zilog->zl_suspend); 3318 ASSERT0(zilog->zl_suspending); 3319 3320 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3321 list_destroy(&zilog->zl_lwb_list); 3322 3323 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 3324 list_destroy(&zilog->zl_itx_commit_list); 3325 3326 for (i = 0; i < TXG_SIZE; i++) { 3327 /* 3328 * It's possible for an itx to be generated that doesn't dirty 3329 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 3330 * callback to remove the entry. We remove those here. 3331 * 3332 * Also free up the ziltest itxs. 3333 */ 3334 if (zilog->zl_itxg[i].itxg_itxs) 3335 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 3336 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 3337 } 3338 3339 mutex_destroy(&zilog->zl_issuer_lock); 3340 mutex_destroy(&zilog->zl_lock); 3341 3342 cv_destroy(&zilog->zl_cv_suspend); 3343 3344 kmem_free(zilog, sizeof (zilog_t)); 3345 } 3346 3347 /* 3348 * Open an intent log. 3349 */ 3350 zilog_t * 3351 zil_open(objset_t *os, zil_get_data_t *get_data) 3352 { 3353 zilog_t *zilog = dmu_objset_zil(os); 3354 3355 ASSERT3P(zilog->zl_get_data, ==, NULL); 3356 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 3357 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 3358 3359 zilog->zl_get_data = get_data; 3360 3361 return (zilog); 3362 } 3363 3364 /* 3365 * Close an intent log. 3366 */ 3367 void 3368 zil_close(zilog_t *zilog) 3369 { 3370 lwb_t *lwb; 3371 uint64_t txg; 3372 3373 if (!dmu_objset_is_snapshot(zilog->zl_os)) { 3374 zil_commit(zilog, 0); 3375 } else { 3376 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 3377 ASSERT0(zilog->zl_dirty_max_txg); 3378 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE); 3379 } 3380 3381 mutex_enter(&zilog->zl_lock); 3382 lwb = list_tail(&zilog->zl_lwb_list); 3383 if (lwb == NULL) 3384 txg = zilog->zl_dirty_max_txg; 3385 else 3386 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg); 3387 mutex_exit(&zilog->zl_lock); 3388 3389 /* 3390 * We need to use txg_wait_synced() to wait long enough for the 3391 * ZIL to be clean, and to wait for all pending lwbs to be 3392 * written out. 3393 */ 3394 if (txg != 0) 3395 txg_wait_synced(zilog->zl_dmu_pool, txg); 3396 3397 if (zilog_is_dirty(zilog)) 3398 zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog, 3399 (u_longlong_t)txg); 3400 if (txg < spa_freeze_txg(zilog->zl_spa)) 3401 VERIFY(!zilog_is_dirty(zilog)); 3402 3403 zilog->zl_get_data = NULL; 3404 3405 /* 3406 * We should have only one lwb left on the list; remove it now. 3407 */ 3408 mutex_enter(&zilog->zl_lock); 3409 lwb = list_head(&zilog->zl_lwb_list); 3410 if (lwb != NULL) { 3411 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list)); 3412 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 3413 3414 if (lwb->lwb_fastwrite) 3415 metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk); 3416 3417 list_remove(&zilog->zl_lwb_list, lwb); 3418 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 3419 zil_free_lwb(zilog, lwb); 3420 } 3421 mutex_exit(&zilog->zl_lock); 3422 } 3423 3424 static char *suspend_tag = "zil suspending"; 3425 3426 /* 3427 * Suspend an intent log. While in suspended mode, we still honor 3428 * synchronous semantics, but we rely on txg_wait_synced() to do it. 3429 * On old version pools, we suspend the log briefly when taking a 3430 * snapshot so that it will have an empty intent log. 3431 * 3432 * Long holds are not really intended to be used the way we do here -- 3433 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 3434 * could fail. Therefore we take pains to only put a long hold if it is 3435 * actually necessary. Fortunately, it will only be necessary if the 3436 * objset is currently mounted (or the ZVOL equivalent). In that case it 3437 * will already have a long hold, so we are not really making things any worse. 3438 * 3439 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 3440 * zvol_state_t), and use their mechanism to prevent their hold from being 3441 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 3442 * very little gain. 3443 * 3444 * if cookiep == NULL, this does both the suspend & resume. 3445 * Otherwise, it returns with the dataset "long held", and the cookie 3446 * should be passed into zil_resume(). 3447 */ 3448 int 3449 zil_suspend(const char *osname, void **cookiep) 3450 { 3451 objset_t *os; 3452 zilog_t *zilog; 3453 const zil_header_t *zh; 3454 int error; 3455 3456 error = dmu_objset_hold(osname, suspend_tag, &os); 3457 if (error != 0) 3458 return (error); 3459 zilog = dmu_objset_zil(os); 3460 3461 mutex_enter(&zilog->zl_lock); 3462 zh = zilog->zl_header; 3463 3464 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 3465 mutex_exit(&zilog->zl_lock); 3466 dmu_objset_rele(os, suspend_tag); 3467 return (SET_ERROR(EBUSY)); 3468 } 3469 3470 /* 3471 * Don't put a long hold in the cases where we can avoid it. This 3472 * is when there is no cookie so we are doing a suspend & resume 3473 * (i.e. called from zil_vdev_offline()), and there's nothing to do 3474 * for the suspend because it's already suspended, or there's no ZIL. 3475 */ 3476 if (cookiep == NULL && !zilog->zl_suspending && 3477 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 3478 mutex_exit(&zilog->zl_lock); 3479 dmu_objset_rele(os, suspend_tag); 3480 return (0); 3481 } 3482 3483 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 3484 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 3485 3486 zilog->zl_suspend++; 3487 3488 if (zilog->zl_suspend > 1) { 3489 /* 3490 * Someone else is already suspending it. 3491 * Just wait for them to finish. 3492 */ 3493 3494 while (zilog->zl_suspending) 3495 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 3496 mutex_exit(&zilog->zl_lock); 3497 3498 if (cookiep == NULL) 3499 zil_resume(os); 3500 else 3501 *cookiep = os; 3502 return (0); 3503 } 3504 3505 /* 3506 * If there is no pointer to an on-disk block, this ZIL must not 3507 * be active (e.g. filesystem not mounted), so there's nothing 3508 * to clean up. 3509 */ 3510 if (BP_IS_HOLE(&zh->zh_log)) { 3511 ASSERT(cookiep != NULL); /* fast path already handled */ 3512 3513 *cookiep = os; 3514 mutex_exit(&zilog->zl_lock); 3515 return (0); 3516 } 3517 3518 /* 3519 * The ZIL has work to do. Ensure that the associated encryption 3520 * key will remain mapped while we are committing the log by 3521 * grabbing a reference to it. If the key isn't loaded we have no 3522 * choice but to return an error until the wrapping key is loaded. 3523 */ 3524 if (os->os_encrypted && 3525 dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) { 3526 zilog->zl_suspend--; 3527 mutex_exit(&zilog->zl_lock); 3528 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3529 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3530 return (SET_ERROR(EACCES)); 3531 } 3532 3533 zilog->zl_suspending = B_TRUE; 3534 mutex_exit(&zilog->zl_lock); 3535 3536 /* 3537 * We need to use zil_commit_impl to ensure we wait for all 3538 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed 3539 * to disk before proceeding. If we used zil_commit instead, it 3540 * would just call txg_wait_synced(), because zl_suspend is set. 3541 * txg_wait_synced() doesn't wait for these lwb's to be 3542 * LWB_STATE_FLUSH_DONE before returning. 3543 */ 3544 zil_commit_impl(zilog, 0); 3545 3546 /* 3547 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we 3548 * use txg_wait_synced() to ensure the data from the zilog has 3549 * migrated to the main pool before calling zil_destroy(). 3550 */ 3551 txg_wait_synced(zilog->zl_dmu_pool, 0); 3552 3553 zil_destroy(zilog, B_FALSE); 3554 3555 mutex_enter(&zilog->zl_lock); 3556 zilog->zl_suspending = B_FALSE; 3557 cv_broadcast(&zilog->zl_cv_suspend); 3558 mutex_exit(&zilog->zl_lock); 3559 3560 if (os->os_encrypted) 3561 dsl_dataset_remove_key_mapping(dmu_objset_ds(os)); 3562 3563 if (cookiep == NULL) 3564 zil_resume(os); 3565 else 3566 *cookiep = os; 3567 return (0); 3568 } 3569 3570 void 3571 zil_resume(void *cookie) 3572 { 3573 objset_t *os = cookie; 3574 zilog_t *zilog = dmu_objset_zil(os); 3575 3576 mutex_enter(&zilog->zl_lock); 3577 ASSERT(zilog->zl_suspend != 0); 3578 zilog->zl_suspend--; 3579 mutex_exit(&zilog->zl_lock); 3580 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3581 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3582 } 3583 3584 typedef struct zil_replay_arg { 3585 zil_replay_func_t *const *zr_replay; 3586 void *zr_arg; 3587 boolean_t zr_byteswap; 3588 char *zr_lr; 3589 } zil_replay_arg_t; 3590 3591 static int 3592 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error) 3593 { 3594 char name[ZFS_MAX_DATASET_NAME_LEN]; 3595 3596 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 3597 3598 dmu_objset_name(zilog->zl_os, name); 3599 3600 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 3601 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 3602 (u_longlong_t)lr->lrc_seq, 3603 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 3604 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 3605 3606 return (error); 3607 } 3608 3609 static int 3610 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra, 3611 uint64_t claim_txg) 3612 { 3613 zil_replay_arg_t *zr = zra; 3614 const zil_header_t *zh = zilog->zl_header; 3615 uint64_t reclen = lr->lrc_reclen; 3616 uint64_t txtype = lr->lrc_txtype; 3617 int error = 0; 3618 3619 zilog->zl_replaying_seq = lr->lrc_seq; 3620 3621 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 3622 return (0); 3623 3624 if (lr->lrc_txg < claim_txg) /* already committed */ 3625 return (0); 3626 3627 /* Strip case-insensitive bit, still present in log record */ 3628 txtype &= ~TX_CI; 3629 3630 if (txtype == 0 || txtype >= TX_MAX_TYPE) 3631 return (zil_replay_error(zilog, lr, EINVAL)); 3632 3633 /* 3634 * If this record type can be logged out of order, the object 3635 * (lr_foid) may no longer exist. That's legitimate, not an error. 3636 */ 3637 if (TX_OOO(txtype)) { 3638 error = dmu_object_info(zilog->zl_os, 3639 LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL); 3640 if (error == ENOENT || error == EEXIST) 3641 return (0); 3642 } 3643 3644 /* 3645 * Make a copy of the data so we can revise and extend it. 3646 */ 3647 memcpy(zr->zr_lr, lr, reclen); 3648 3649 /* 3650 * If this is a TX_WRITE with a blkptr, suck in the data. 3651 */ 3652 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 3653 error = zil_read_log_data(zilog, (lr_write_t *)lr, 3654 zr->zr_lr + reclen); 3655 if (error != 0) 3656 return (zil_replay_error(zilog, lr, error)); 3657 } 3658 3659 /* 3660 * The log block containing this lr may have been byteswapped 3661 * so that we can easily examine common fields like lrc_txtype. 3662 * However, the log is a mix of different record types, and only the 3663 * replay vectors know how to byteswap their records. Therefore, if 3664 * the lr was byteswapped, undo it before invoking the replay vector. 3665 */ 3666 if (zr->zr_byteswap) 3667 byteswap_uint64_array(zr->zr_lr, reclen); 3668 3669 /* 3670 * We must now do two things atomically: replay this log record, 3671 * and update the log header sequence number to reflect the fact that 3672 * we did so. At the end of each replay function the sequence number 3673 * is updated if we are in replay mode. 3674 */ 3675 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 3676 if (error != 0) { 3677 /* 3678 * The DMU's dnode layer doesn't see removes until the txg 3679 * commits, so a subsequent claim can spuriously fail with 3680 * EEXIST. So if we receive any error we try syncing out 3681 * any removes then retry the transaction. Note that we 3682 * specify B_FALSE for byteswap now, so we don't do it twice. 3683 */ 3684 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 3685 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 3686 if (error != 0) 3687 return (zil_replay_error(zilog, lr, error)); 3688 } 3689 return (0); 3690 } 3691 3692 static int 3693 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg) 3694 { 3695 (void) bp, (void) arg, (void) claim_txg; 3696 3697 zilog->zl_replay_blks++; 3698 3699 return (0); 3700 } 3701 3702 /* 3703 * If this dataset has a non-empty intent log, replay it and destroy it. 3704 */ 3705 void 3706 zil_replay(objset_t *os, void *arg, 3707 zil_replay_func_t *const replay_func[TX_MAX_TYPE]) 3708 { 3709 zilog_t *zilog = dmu_objset_zil(os); 3710 const zil_header_t *zh = zilog->zl_header; 3711 zil_replay_arg_t zr; 3712 3713 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 3714 zil_destroy(zilog, B_TRUE); 3715 return; 3716 } 3717 3718 zr.zr_replay = replay_func; 3719 zr.zr_arg = arg; 3720 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 3721 zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 3722 3723 /* 3724 * Wait for in-progress removes to sync before starting replay. 3725 */ 3726 txg_wait_synced(zilog->zl_dmu_pool, 0); 3727 3728 zilog->zl_replay = B_TRUE; 3729 zilog->zl_replay_time = ddi_get_lbolt(); 3730 ASSERT(zilog->zl_replay_blks == 0); 3731 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 3732 zh->zh_claim_txg, B_TRUE); 3733 vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 3734 3735 zil_destroy(zilog, B_FALSE); 3736 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 3737 zilog->zl_replay = B_FALSE; 3738 } 3739 3740 boolean_t 3741 zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 3742 { 3743 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 3744 return (B_TRUE); 3745 3746 if (zilog->zl_replay) { 3747 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 3748 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 3749 zilog->zl_replaying_seq; 3750 return (B_TRUE); 3751 } 3752 3753 return (B_FALSE); 3754 } 3755 3756 int 3757 zil_reset(const char *osname, void *arg) 3758 { 3759 (void) arg; 3760 3761 int error = zil_suspend(osname, NULL); 3762 /* EACCES means crypto key not loaded */ 3763 if ((error == EACCES) || (error == EBUSY)) 3764 return (SET_ERROR(error)); 3765 if (error != 0) 3766 return (SET_ERROR(EEXIST)); 3767 return (0); 3768 } 3769 3770 EXPORT_SYMBOL(zil_alloc); 3771 EXPORT_SYMBOL(zil_free); 3772 EXPORT_SYMBOL(zil_open); 3773 EXPORT_SYMBOL(zil_close); 3774 EXPORT_SYMBOL(zil_replay); 3775 EXPORT_SYMBOL(zil_replaying); 3776 EXPORT_SYMBOL(zil_destroy); 3777 EXPORT_SYMBOL(zil_destroy_sync); 3778 EXPORT_SYMBOL(zil_itx_create); 3779 EXPORT_SYMBOL(zil_itx_destroy); 3780 EXPORT_SYMBOL(zil_itx_assign); 3781 EXPORT_SYMBOL(zil_commit); 3782 EXPORT_SYMBOL(zil_claim); 3783 EXPORT_SYMBOL(zil_check_log_chain); 3784 EXPORT_SYMBOL(zil_sync); 3785 EXPORT_SYMBOL(zil_clean); 3786 EXPORT_SYMBOL(zil_suspend); 3787 EXPORT_SYMBOL(zil_resume); 3788 EXPORT_SYMBOL(zil_lwb_add_block); 3789 EXPORT_SYMBOL(zil_bp_tree_add); 3790 EXPORT_SYMBOL(zil_set_sync); 3791 EXPORT_SYMBOL(zil_set_logbias); 3792 3793 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, INT, ZMOD_RW, 3794 "ZIL block open timeout percentage"); 3795 3796 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW, 3797 "Disable intent logging replay"); 3798 3799 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW, 3800 "Disable ZIL cache flushes"); 3801 3802 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, ULONG, ZMOD_RW, 3803 "Limit in bytes slog sync writes per commit"); 3804 3805 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, INT, ZMOD_RW, 3806 "Limit in bytes of ZIL log block size"); 3807