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 */ 26 27 /* Portions Copyright 2010 Robert Milkowski */ 28 29 #include <sys/zfs_context.h> 30 #include <sys/spa.h> 31 #include <sys/spa_impl.h> 32 #include <sys/dmu.h> 33 #include <sys/zap.h> 34 #include <sys/arc.h> 35 #include <sys/stat.h> 36 #include <sys/resource.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/abd.h> 44 45 /* 46 * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system 47 * calls that change the file system. Each itx has enough information to 48 * be able to replay them after a system crash, power loss, or 49 * equivalent failure mode. These are stored in memory until either: 50 * 51 * 1. they are committed to the pool by the DMU transaction group 52 * (txg), at which point they can be discarded; or 53 * 2. they are committed to the on-disk ZIL for the dataset being 54 * modified (e.g. due to an fsync, O_DSYNC, or other synchronous 55 * requirement). 56 * 57 * In the event of a crash or power loss, the itxs contained by each 58 * dataset's on-disk ZIL will be replayed when that dataset is first 59 * instantianted (e.g. if the dataset is a normal fileystem, when it is 60 * first mounted). 61 * 62 * As hinted at above, there is one ZIL per dataset (both the in-memory 63 * representation, and the on-disk representation). The on-disk format 64 * consists of 3 parts: 65 * 66 * - a single, per-dataset, ZIL header; which points to a chain of 67 * - zero or more ZIL blocks; each of which contains 68 * - zero or more ZIL records 69 * 70 * A ZIL record holds the information necessary to replay a single 71 * system call transaction. A ZIL block can hold many ZIL records, and 72 * the blocks are chained together, similarly to a singly linked list. 73 * 74 * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL 75 * block in the chain, and the ZIL header points to the first block in 76 * the chain. 77 * 78 * Note, there is not a fixed place in the pool to hold these ZIL 79 * blocks; they are dynamically allocated and freed as needed from the 80 * blocks available on the pool, though they can be preferentially 81 * allocated from a dedicated "log" vdev. 82 */ 83 84 /* 85 * This controls the amount of time that a ZIL block (lwb) will remain 86 * "open" when it isn't "full", and it has a thread waiting for it to be 87 * committed to stable storage. Please refer to the zil_commit_waiter() 88 * function (and the comments within it) for more details. 89 */ 90 int zfs_commit_timeout_pct = 5; 91 92 /* 93 * Disable intent logging replay. This global ZIL switch affects all pools. 94 */ 95 int zil_replay_disable = 0; 96 97 /* 98 * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to 99 * the disk(s) by the ZIL after an LWB write has completed. Setting this 100 * will cause ZIL corruption on power loss if a volatile out-of-order 101 * write cache is enabled. 102 */ 103 boolean_t zil_nocacheflush = B_FALSE; 104 105 /* 106 * Limit SLOG write size per commit executed with synchronous priority. 107 * Any writes above that will be executed with lower (asynchronous) priority 108 * to limit potential SLOG device abuse by single active ZIL writer. 109 */ 110 uint64_t zil_slog_bulk = 768 * 1024; 111 112 static kmem_cache_t *zil_lwb_cache; 113 static kmem_cache_t *zil_zcw_cache; 114 115 static void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 116 117 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 118 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 119 120 static int 121 zil_bp_compare(const void *x1, const void *x2) 122 { 123 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 124 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 125 126 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 127 return (-1); 128 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 129 return (1); 130 131 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 132 return (-1); 133 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 134 return (1); 135 136 return (0); 137 } 138 139 static void 140 zil_bp_tree_init(zilog_t *zilog) 141 { 142 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 143 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 144 } 145 146 static void 147 zil_bp_tree_fini(zilog_t *zilog) 148 { 149 avl_tree_t *t = &zilog->zl_bp_tree; 150 zil_bp_node_t *zn; 151 void *cookie = NULL; 152 153 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 154 kmem_free(zn, sizeof (zil_bp_node_t)); 155 156 avl_destroy(t); 157 } 158 159 int 160 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 161 { 162 avl_tree_t *t = &zilog->zl_bp_tree; 163 const dva_t *dva; 164 zil_bp_node_t *zn; 165 avl_index_t where; 166 167 if (BP_IS_EMBEDDED(bp)) 168 return (0); 169 170 dva = BP_IDENTITY(bp); 171 172 if (avl_find(t, dva, &where) != NULL) 173 return (SET_ERROR(EEXIST)); 174 175 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 176 zn->zn_dva = *dva; 177 avl_insert(t, zn, where); 178 179 return (0); 180 } 181 182 static zil_header_t * 183 zil_header_in_syncing_context(zilog_t *zilog) 184 { 185 return ((zil_header_t *)zilog->zl_header); 186 } 187 188 static void 189 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 190 { 191 zio_cksum_t *zc = &bp->blk_cksum; 192 193 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 194 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 195 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 196 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 197 } 198 199 /* 200 * Read a log block and make sure it's valid. 201 */ 202 static int 203 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 204 char **end) 205 { 206 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 207 arc_flags_t aflags = ARC_FLAG_WAIT; 208 arc_buf_t *abuf = NULL; 209 zbookmark_phys_t zb; 210 int error; 211 212 if (zilog->zl_header->zh_claim_txg == 0) 213 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 214 215 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 216 zio_flags |= ZIO_FLAG_SPECULATIVE; 217 218 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 219 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 220 221 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 222 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 223 224 if (error == 0) { 225 zio_cksum_t cksum = bp->blk_cksum; 226 227 /* 228 * Validate the checksummed log block. 229 * 230 * Sequence numbers should be... sequential. The checksum 231 * verifier for the next block should be bp's checksum plus 1. 232 * 233 * Also check the log chain linkage and size used. 234 */ 235 cksum.zc_word[ZIL_ZC_SEQ]++; 236 237 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 238 zil_chain_t *zilc = abuf->b_data; 239 char *lr = (char *)(zilc + 1); 240 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 241 242 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 243 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 244 error = SET_ERROR(ECKSUM); 245 } else { 246 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 247 bcopy(lr, dst, len); 248 *end = (char *)dst + len; 249 *nbp = zilc->zc_next_blk; 250 } 251 } else { 252 char *lr = abuf->b_data; 253 uint64_t size = BP_GET_LSIZE(bp); 254 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 255 256 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 257 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 258 (zilc->zc_nused > (size - sizeof (*zilc)))) { 259 error = SET_ERROR(ECKSUM); 260 } else { 261 ASSERT3U(zilc->zc_nused, <=, 262 SPA_OLD_MAXBLOCKSIZE); 263 bcopy(lr, dst, zilc->zc_nused); 264 *end = (char *)dst + zilc->zc_nused; 265 *nbp = zilc->zc_next_blk; 266 } 267 } 268 269 arc_buf_destroy(abuf, &abuf); 270 } 271 272 return (error); 273 } 274 275 /* 276 * Read a TX_WRITE log data block. 277 */ 278 static int 279 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 280 { 281 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 282 const blkptr_t *bp = &lr->lr_blkptr; 283 arc_flags_t aflags = ARC_FLAG_WAIT; 284 arc_buf_t *abuf = NULL; 285 zbookmark_phys_t zb; 286 int error; 287 288 if (BP_IS_HOLE(bp)) { 289 if (wbuf != NULL) 290 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 291 return (0); 292 } 293 294 if (zilog->zl_header->zh_claim_txg == 0) 295 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 296 297 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 298 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 299 300 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 301 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 302 303 if (error == 0) { 304 if (wbuf != NULL) 305 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 306 arc_buf_destroy(abuf, &abuf); 307 } 308 309 return (error); 310 } 311 312 /* 313 * Parse the intent log, and call parse_func for each valid record within. 314 */ 315 int 316 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 317 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 318 { 319 const zil_header_t *zh = zilog->zl_header; 320 boolean_t claimed = !!zh->zh_claim_txg; 321 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 322 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 323 uint64_t max_blk_seq = 0; 324 uint64_t max_lr_seq = 0; 325 uint64_t blk_count = 0; 326 uint64_t lr_count = 0; 327 blkptr_t blk, next_blk; 328 char *lrbuf, *lrp; 329 int error = 0; 330 331 /* 332 * Old logs didn't record the maximum zh_claim_lr_seq. 333 */ 334 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 335 claim_lr_seq = UINT64_MAX; 336 337 /* 338 * Starting at the block pointed to by zh_log we read the log chain. 339 * For each block in the chain we strongly check that block to 340 * ensure its validity. We stop when an invalid block is found. 341 * For each block pointer in the chain we call parse_blk_func(). 342 * For each record in each valid block we call parse_lr_func(). 343 * If the log has been claimed, stop if we encounter a sequence 344 * number greater than the highest claimed sequence number. 345 */ 346 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 347 zil_bp_tree_init(zilog); 348 349 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 350 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 351 int reclen; 352 char *end; 353 354 if (blk_seq > claim_blk_seq) 355 break; 356 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 357 break; 358 ASSERT3U(max_blk_seq, <, blk_seq); 359 max_blk_seq = blk_seq; 360 blk_count++; 361 362 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 363 break; 364 365 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 366 if (error != 0) 367 break; 368 369 for (lrp = lrbuf; lrp < end; lrp += reclen) { 370 lr_t *lr = (lr_t *)lrp; 371 reclen = lr->lrc_reclen; 372 ASSERT3U(reclen, >=, sizeof (lr_t)); 373 if (lr->lrc_seq > claim_lr_seq) 374 goto done; 375 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 376 goto done; 377 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 378 max_lr_seq = lr->lrc_seq; 379 lr_count++; 380 } 381 } 382 done: 383 zilog->zl_parse_error = error; 384 zilog->zl_parse_blk_seq = max_blk_seq; 385 zilog->zl_parse_lr_seq = max_lr_seq; 386 zilog->zl_parse_blk_count = blk_count; 387 zilog->zl_parse_lr_count = lr_count; 388 389 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 390 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 391 392 zil_bp_tree_fini(zilog); 393 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 394 395 return (error); 396 } 397 398 /* ARGSUSED */ 399 static int 400 zil_clear_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 401 { 402 ASSERT(!BP_IS_HOLE(bp)); 403 404 /* 405 * As we call this function from the context of a rewind to a 406 * checkpoint, each ZIL block whose txg is later than the txg 407 * that we rewind to is invalid. Thus, we return -1 so 408 * zil_parse() doesn't attempt to read it. 409 */ 410 if (bp->blk_birth >= first_txg) 411 return (-1); 412 413 if (zil_bp_tree_add(zilog, bp) != 0) 414 return (0); 415 416 zio_free(zilog->zl_spa, first_txg, bp); 417 return (0); 418 } 419 420 /* ARGSUSED */ 421 static int 422 zil_noop_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 423 { 424 return (0); 425 } 426 427 static int 428 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 429 { 430 /* 431 * Claim log block if not already committed and not already claimed. 432 * If tx == NULL, just verify that the block is claimable. 433 */ 434 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 435 zil_bp_tree_add(zilog, bp) != 0) 436 return (0); 437 438 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 439 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 440 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 441 } 442 443 static int 444 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 445 { 446 lr_write_t *lr = (lr_write_t *)lrc; 447 int error; 448 449 if (lrc->lrc_txtype != TX_WRITE) 450 return (0); 451 452 /* 453 * If the block is not readable, don't claim it. This can happen 454 * in normal operation when a log block is written to disk before 455 * some of the dmu_sync() blocks it points to. In this case, the 456 * transaction cannot have been committed to anyone (we would have 457 * waited for all writes to be stable first), so it is semantically 458 * correct to declare this the end of the log. 459 */ 460 if (lr->lr_blkptr.blk_birth >= first_txg && 461 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 462 return (error); 463 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 464 } 465 466 /* ARGSUSED */ 467 static int 468 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 469 { 470 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 471 472 return (0); 473 } 474 475 static int 476 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 477 { 478 lr_write_t *lr = (lr_write_t *)lrc; 479 blkptr_t *bp = &lr->lr_blkptr; 480 481 /* 482 * If we previously claimed it, we need to free it. 483 */ 484 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 485 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 486 !BP_IS_HOLE(bp)) 487 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 488 489 return (0); 490 } 491 492 static int 493 zil_lwb_vdev_compare(const void *x1, const void *x2) 494 { 495 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 496 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 497 498 if (v1 < v2) 499 return (-1); 500 if (v1 > v2) 501 return (1); 502 503 return (0); 504 } 505 506 static lwb_t * 507 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg) 508 { 509 lwb_t *lwb; 510 511 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 512 lwb->lwb_zilog = zilog; 513 lwb->lwb_blk = *bp; 514 lwb->lwb_slog = slog; 515 lwb->lwb_state = LWB_STATE_CLOSED; 516 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 517 lwb->lwb_max_txg = txg; 518 lwb->lwb_write_zio = NULL; 519 lwb->lwb_root_zio = NULL; 520 lwb->lwb_tx = NULL; 521 lwb->lwb_issued_timestamp = 0; 522 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 523 lwb->lwb_nused = sizeof (zil_chain_t); 524 lwb->lwb_sz = BP_GET_LSIZE(bp); 525 } else { 526 lwb->lwb_nused = 0; 527 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 528 } 529 530 mutex_enter(&zilog->zl_lock); 531 list_insert_tail(&zilog->zl_lwb_list, lwb); 532 mutex_exit(&zilog->zl_lock); 533 534 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); 535 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 536 VERIFY(list_is_empty(&lwb->lwb_waiters)); 537 538 return (lwb); 539 } 540 541 static void 542 zil_free_lwb(zilog_t *zilog, lwb_t *lwb) 543 { 544 ASSERT(MUTEX_HELD(&zilog->zl_lock)); 545 ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock)); 546 VERIFY(list_is_empty(&lwb->lwb_waiters)); 547 ASSERT(avl_is_empty(&lwb->lwb_vdev_tree)); 548 ASSERT3P(lwb->lwb_write_zio, ==, NULL); 549 ASSERT3P(lwb->lwb_root_zio, ==, NULL); 550 ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa)); 551 ASSERT(lwb->lwb_state == LWB_STATE_CLOSED || 552 lwb->lwb_state == LWB_STATE_DONE); 553 554 /* 555 * Clear the zilog's field to indicate this lwb is no longer 556 * valid, and prevent use-after-free errors. 557 */ 558 if (zilog->zl_last_lwb_opened == lwb) 559 zilog->zl_last_lwb_opened = NULL; 560 561 kmem_cache_free(zil_lwb_cache, lwb); 562 } 563 564 /* 565 * Called when we create in-memory log transactions so that we know 566 * to cleanup the itxs at the end of spa_sync(). 567 */ 568 void 569 zilog_dirty(zilog_t *zilog, uint64_t txg) 570 { 571 dsl_pool_t *dp = zilog->zl_dmu_pool; 572 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 573 574 ASSERT(spa_writeable(zilog->zl_spa)); 575 576 if (ds->ds_is_snapshot) 577 panic("dirtying snapshot!"); 578 579 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 580 /* up the hold count until we can be written out */ 581 dmu_buf_add_ref(ds->ds_dbuf, zilog); 582 583 zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg); 584 } 585 } 586 587 /* 588 * Determine if the zil is dirty in the specified txg. Callers wanting to 589 * ensure that the dirty state does not change must hold the itxg_lock for 590 * the specified txg. Holding the lock will ensure that the zil cannot be 591 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 592 * state. 593 */ 594 boolean_t 595 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 596 { 597 dsl_pool_t *dp = zilog->zl_dmu_pool; 598 599 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 600 return (B_TRUE); 601 return (B_FALSE); 602 } 603 604 /* 605 * Determine if the zil is dirty. The zil is considered dirty if it has 606 * any pending itx records that have not been cleaned by zil_clean(). 607 */ 608 boolean_t 609 zilog_is_dirty(zilog_t *zilog) 610 { 611 dsl_pool_t *dp = zilog->zl_dmu_pool; 612 613 for (int t = 0; t < TXG_SIZE; t++) { 614 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 615 return (B_TRUE); 616 } 617 return (B_FALSE); 618 } 619 620 /* 621 * Create an on-disk intent log. 622 */ 623 static lwb_t * 624 zil_create(zilog_t *zilog) 625 { 626 const zil_header_t *zh = zilog->zl_header; 627 lwb_t *lwb = NULL; 628 uint64_t txg = 0; 629 dmu_tx_t *tx = NULL; 630 blkptr_t blk; 631 int error = 0; 632 boolean_t slog = FALSE; 633 634 /* 635 * Wait for any previous destroy to complete. 636 */ 637 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 638 639 ASSERT(zh->zh_claim_txg == 0); 640 ASSERT(zh->zh_replay_seq == 0); 641 642 blk = zh->zh_log; 643 644 /* 645 * Allocate an initial log block if: 646 * - there isn't one already 647 * - the existing block is the wrong endianess 648 */ 649 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 650 tx = dmu_tx_create(zilog->zl_os); 651 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 652 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 653 txg = dmu_tx_get_txg(tx); 654 655 if (!BP_IS_HOLE(&blk)) { 656 zio_free(zilog->zl_spa, txg, &blk); 657 BP_ZERO(&blk); 658 } 659 660 error = zio_alloc_zil(zilog->zl_spa, 661 zilog->zl_os->os_dsl_dataset->ds_object, txg, &blk, NULL, 662 ZIL_MIN_BLKSZ, &slog); 663 664 if (error == 0) 665 zil_init_log_chain(zilog, &blk); 666 } 667 668 /* 669 * Allocate a log write block (lwb) for the first log block. 670 */ 671 if (error == 0) 672 lwb = zil_alloc_lwb(zilog, &blk, slog, txg); 673 674 /* 675 * If we just allocated the first log block, commit our transaction 676 * and wait for zil_sync() to stuff the block poiner into zh_log. 677 * (zh is part of the MOS, so we cannot modify it in open context.) 678 */ 679 if (tx != NULL) { 680 dmu_tx_commit(tx); 681 txg_wait_synced(zilog->zl_dmu_pool, txg); 682 } 683 684 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 685 686 return (lwb); 687 } 688 689 /* 690 * In one tx, free all log blocks and clear the log header. If keep_first 691 * is set, then we're replaying a log with no content. We want to keep the 692 * first block, however, so that the first synchronous transaction doesn't 693 * require a txg_wait_synced() in zil_create(). We don't need to 694 * txg_wait_synced() here either when keep_first is set, because both 695 * zil_create() and zil_destroy() will wait for any in-progress destroys 696 * to complete. 697 */ 698 void 699 zil_destroy(zilog_t *zilog, boolean_t keep_first) 700 { 701 const zil_header_t *zh = zilog->zl_header; 702 lwb_t *lwb; 703 dmu_tx_t *tx; 704 uint64_t txg; 705 706 /* 707 * Wait for any previous destroy to complete. 708 */ 709 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 710 711 zilog->zl_old_header = *zh; /* debugging aid */ 712 713 if (BP_IS_HOLE(&zh->zh_log)) 714 return; 715 716 tx = dmu_tx_create(zilog->zl_os); 717 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 718 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 719 txg = dmu_tx_get_txg(tx); 720 721 mutex_enter(&zilog->zl_lock); 722 723 ASSERT3U(zilog->zl_destroy_txg, <, txg); 724 zilog->zl_destroy_txg = txg; 725 zilog->zl_keep_first = keep_first; 726 727 if (!list_is_empty(&zilog->zl_lwb_list)) { 728 ASSERT(zh->zh_claim_txg == 0); 729 VERIFY(!keep_first); 730 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 731 list_remove(&zilog->zl_lwb_list, lwb); 732 if (lwb->lwb_buf != NULL) 733 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 734 zio_free(zilog->zl_spa, txg, &lwb->lwb_blk); 735 zil_free_lwb(zilog, lwb); 736 } 737 } else if (!keep_first) { 738 zil_destroy_sync(zilog, tx); 739 } 740 mutex_exit(&zilog->zl_lock); 741 742 dmu_tx_commit(tx); 743 } 744 745 void 746 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 747 { 748 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 749 (void) zil_parse(zilog, zil_free_log_block, 750 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 751 } 752 753 int 754 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 755 { 756 dmu_tx_t *tx = txarg; 757 zilog_t *zilog; 758 uint64_t first_txg; 759 zil_header_t *zh; 760 objset_t *os; 761 int error; 762 763 error = dmu_objset_own_obj(dp, ds->ds_object, 764 DMU_OST_ANY, B_FALSE, FTAG, &os); 765 if (error != 0) { 766 /* 767 * EBUSY indicates that the objset is inconsistent, in which 768 * case it can not have a ZIL. 769 */ 770 if (error != EBUSY) { 771 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 772 (unsigned long long)ds->ds_object, error); 773 } 774 return (0); 775 } 776 777 zilog = dmu_objset_zil(os); 778 zh = zil_header_in_syncing_context(zilog); 779 ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa)); 780 first_txg = spa_min_claim_txg(zilog->zl_spa); 781 782 /* 783 * If the spa_log_state is not set to be cleared, check whether 784 * the current uberblock is a checkpoint one and if the current 785 * header has been claimed before moving on. 786 * 787 * If the current uberblock is a checkpointed uberblock then 788 * one of the following scenarios took place: 789 * 790 * 1] We are currently rewinding to the checkpoint of the pool. 791 * 2] We crashed in the middle of a checkpoint rewind but we 792 * did manage to write the checkpointed uberblock to the 793 * vdev labels, so when we tried to import the pool again 794 * the checkpointed uberblock was selected from the import 795 * procedure. 796 * 797 * In both cases we want to zero out all the ZIL blocks, except 798 * the ones that have been claimed at the time of the checkpoint 799 * (their zh_claim_txg != 0). The reason is that these blocks 800 * may be corrupted since we may have reused their locations on 801 * disk after we took the checkpoint. 802 * 803 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier 804 * when we first figure out whether the current uberblock is 805 * checkpointed or not. Unfortunately, that would discard all 806 * the logs, including the ones that are claimed, and we would 807 * leak space. 808 */ 809 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR || 810 (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && 811 zh->zh_claim_txg == 0)) { 812 if (!BP_IS_HOLE(&zh->zh_log)) { 813 (void) zil_parse(zilog, zil_clear_log_block, 814 zil_noop_log_record, tx, first_txg); 815 } 816 BP_ZERO(&zh->zh_log); 817 dsl_dataset_dirty(dmu_objset_ds(os), tx); 818 dmu_objset_disown(os, FTAG); 819 return (0); 820 } 821 822 /* 823 * If we are not rewinding and opening the pool normally, then 824 * the min_claim_txg should be equal to the first txg of the pool. 825 */ 826 ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa)); 827 828 /* 829 * Claim all log blocks if we haven't already done so, and remember 830 * the highest claimed sequence number. This ensures that if we can 831 * read only part of the log now (e.g. due to a missing device), 832 * but we can read the entire log later, we will not try to replay 833 * or destroy beyond the last block we successfully claimed. 834 */ 835 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 836 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 837 (void) zil_parse(zilog, zil_claim_log_block, 838 zil_claim_log_record, tx, first_txg); 839 zh->zh_claim_txg = first_txg; 840 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 841 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 842 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 843 zh->zh_flags |= ZIL_REPLAY_NEEDED; 844 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 845 dsl_dataset_dirty(dmu_objset_ds(os), tx); 846 } 847 848 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 849 dmu_objset_disown(os, FTAG); 850 return (0); 851 } 852 853 /* 854 * Check the log by walking the log chain. 855 * Checksum errors are ok as they indicate the end of the chain. 856 * Any other error (no device or read failure) returns an error. 857 */ 858 /* ARGSUSED */ 859 int 860 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 861 { 862 zilog_t *zilog; 863 objset_t *os; 864 blkptr_t *bp; 865 int error; 866 867 ASSERT(tx == NULL); 868 869 error = dmu_objset_from_ds(ds, &os); 870 if (error != 0) { 871 cmn_err(CE_WARN, "can't open objset %llu, error %d", 872 (unsigned long long)ds->ds_object, error); 873 return (0); 874 } 875 876 zilog = dmu_objset_zil(os); 877 bp = (blkptr_t *)&zilog->zl_header->zh_log; 878 879 if (!BP_IS_HOLE(bp)) { 880 vdev_t *vd; 881 boolean_t valid = B_TRUE; 882 883 /* 884 * Check the first block and determine if it's on a log device 885 * which may have been removed or faulted prior to loading this 886 * pool. If so, there's no point in checking the rest of the 887 * log as its content should have already been synced to the 888 * pool. 889 */ 890 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 891 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 892 if (vd->vdev_islog && vdev_is_dead(vd)) 893 valid = vdev_log_state_valid(vd); 894 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 895 896 if (!valid) 897 return (0); 898 899 /* 900 * Check whether the current uberblock is checkpointed (e.g. 901 * we are rewinding) and whether the current header has been 902 * claimed or not. If it hasn't then skip verifying it. We 903 * do this because its ZIL blocks may be part of the pool's 904 * state before the rewind, which is no longer valid. 905 */ 906 zil_header_t *zh = zil_header_in_syncing_context(zilog); 907 if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 && 908 zh->zh_claim_txg == 0) 909 return (0); 910 } 911 912 /* 913 * Because tx == NULL, zil_claim_log_block() will not actually claim 914 * any blocks, but just determine whether it is possible to do so. 915 * In addition to checking the log chain, zil_claim_log_block() 916 * will invoke zio_claim() with a done func of spa_claim_notify(), 917 * which will update spa_max_claim_txg. See spa_load() for details. 918 */ 919 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 920 zilog->zl_header->zh_claim_txg ? -1ULL : 921 spa_min_claim_txg(os->os_spa)); 922 923 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 924 } 925 926 /* 927 * When an itx is "skipped", this function is used to properly mark the 928 * waiter as "done, and signal any thread(s) waiting on it. An itx can 929 * be skipped (and not committed to an lwb) for a variety of reasons, 930 * one of them being that the itx was committed via spa_sync(), prior to 931 * it being committed to an lwb; this can happen if a thread calling 932 * zil_commit() is racing with spa_sync(). 933 */ 934 static void 935 zil_commit_waiter_skip(zil_commit_waiter_t *zcw) 936 { 937 mutex_enter(&zcw->zcw_lock); 938 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 939 zcw->zcw_done = B_TRUE; 940 cv_broadcast(&zcw->zcw_cv); 941 mutex_exit(&zcw->zcw_lock); 942 } 943 944 /* 945 * This function is used when the given waiter is to be linked into an 946 * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb. 947 * At this point, the waiter will no longer be referenced by the itx, 948 * and instead, will be referenced by the lwb. 949 */ 950 static void 951 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb) 952 { 953 /* 954 * The lwb_waiters field of the lwb is protected by the zilog's 955 * zl_lock, thus it must be held when calling this function. 956 */ 957 ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock)); 958 959 mutex_enter(&zcw->zcw_lock); 960 ASSERT(!list_link_active(&zcw->zcw_node)); 961 ASSERT3P(zcw->zcw_lwb, ==, NULL); 962 ASSERT3P(lwb, !=, NULL); 963 ASSERT(lwb->lwb_state == LWB_STATE_OPENED || 964 lwb->lwb_state == LWB_STATE_ISSUED); 965 966 list_insert_tail(&lwb->lwb_waiters, zcw); 967 zcw->zcw_lwb = lwb; 968 mutex_exit(&zcw->zcw_lock); 969 } 970 971 /* 972 * This function is used when zio_alloc_zil() fails to allocate a ZIL 973 * block, and the given waiter must be linked to the "nolwb waiters" 974 * list inside of zil_process_commit_list(). 975 */ 976 static void 977 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb) 978 { 979 mutex_enter(&zcw->zcw_lock); 980 ASSERT(!list_link_active(&zcw->zcw_node)); 981 ASSERT3P(zcw->zcw_lwb, ==, NULL); 982 list_insert_tail(nolwb, zcw); 983 mutex_exit(&zcw->zcw_lock); 984 } 985 986 void 987 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp) 988 { 989 avl_tree_t *t = &lwb->lwb_vdev_tree; 990 avl_index_t where; 991 zil_vdev_node_t *zv, zvsearch; 992 int ndvas = BP_GET_NDVAS(bp); 993 int i; 994 995 if (zil_nocacheflush) 996 return; 997 998 mutex_enter(&lwb->lwb_vdev_lock); 999 for (i = 0; i < ndvas; i++) { 1000 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 1001 if (avl_find(t, &zvsearch, &where) == NULL) { 1002 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 1003 zv->zv_vdev = zvsearch.zv_vdev; 1004 avl_insert(t, zv, where); 1005 } 1006 } 1007 mutex_exit(&lwb->lwb_vdev_lock); 1008 } 1009 1010 void 1011 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg) 1012 { 1013 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1014 } 1015 1016 /* 1017 * This function is a called after all VDEVs associated with a given lwb 1018 * write have completed their DKIOCFLUSHWRITECACHE command; or as soon 1019 * as the lwb write completes, if "zil_nocacheflush" is set. 1020 * 1021 * The intention is for this function to be called as soon as the 1022 * contents of an lwb are considered "stable" on disk, and will survive 1023 * any sudden loss of power. At this point, any threads waiting for the 1024 * lwb to reach this state are signalled, and the "waiter" structures 1025 * are marked "done". 1026 */ 1027 static void 1028 zil_lwb_flush_vdevs_done(zio_t *zio) 1029 { 1030 lwb_t *lwb = zio->io_private; 1031 zilog_t *zilog = lwb->lwb_zilog; 1032 dmu_tx_t *tx = lwb->lwb_tx; 1033 zil_commit_waiter_t *zcw; 1034 1035 spa_config_exit(zilog->zl_spa, SCL_STATE, lwb); 1036 1037 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1038 1039 mutex_enter(&zilog->zl_lock); 1040 1041 /* 1042 * Ensure the lwb buffer pointer is cleared before releasing the 1043 * txg. If we have had an allocation failure and the txg is 1044 * waiting to sync then we want zil_sync() to remove the lwb so 1045 * that it's not picked up as the next new one in 1046 * zil_process_commit_list(). zil_sync() will only remove the 1047 * lwb if lwb_buf is null. 1048 */ 1049 lwb->lwb_buf = NULL; 1050 lwb->lwb_tx = NULL; 1051 1052 ASSERT3U(lwb->lwb_issued_timestamp, >, 0); 1053 zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp; 1054 1055 lwb->lwb_root_zio = NULL; 1056 lwb->lwb_state = LWB_STATE_DONE; 1057 1058 if (zilog->zl_last_lwb_opened == lwb) { 1059 /* 1060 * Remember the highest committed log sequence number 1061 * for ztest. We only update this value when all the log 1062 * writes succeeded, because ztest wants to ASSERT that 1063 * it got the whole log chain. 1064 */ 1065 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1066 } 1067 1068 while ((zcw = list_head(&lwb->lwb_waiters)) != NULL) { 1069 mutex_enter(&zcw->zcw_lock); 1070 1071 ASSERT(list_link_active(&zcw->zcw_node)); 1072 list_remove(&lwb->lwb_waiters, zcw); 1073 1074 ASSERT3P(zcw->zcw_lwb, ==, lwb); 1075 zcw->zcw_lwb = NULL; 1076 1077 zcw->zcw_zio_error = zio->io_error; 1078 1079 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 1080 zcw->zcw_done = B_TRUE; 1081 cv_broadcast(&zcw->zcw_cv); 1082 1083 mutex_exit(&zcw->zcw_lock); 1084 } 1085 1086 mutex_exit(&zilog->zl_lock); 1087 1088 /* 1089 * Now that we've written this log block, we have a stable pointer 1090 * to the next block in the chain, so it's OK to let the txg in 1091 * which we allocated the next block sync. 1092 */ 1093 dmu_tx_commit(tx); 1094 } 1095 1096 /* 1097 * This is called when an lwb write completes. This means, this specific 1098 * lwb was written to disk, and all dependent lwb have also been 1099 * written to disk. 1100 * 1101 * At this point, a DKIOCFLUSHWRITECACHE command hasn't been issued to 1102 * the VDEVs involved in writing out this specific lwb. The lwb will be 1103 * "done" once zil_lwb_flush_vdevs_done() is called, which occurs in the 1104 * zio completion callback for the lwb's root zio. 1105 */ 1106 static void 1107 zil_lwb_write_done(zio_t *zio) 1108 { 1109 lwb_t *lwb = zio->io_private; 1110 spa_t *spa = zio->io_spa; 1111 zilog_t *zilog = lwb->lwb_zilog; 1112 avl_tree_t *t = &lwb->lwb_vdev_tree; 1113 void *cookie = NULL; 1114 zil_vdev_node_t *zv; 1115 1116 ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0); 1117 1118 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 1119 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 1120 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 1121 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 1122 ASSERT(!BP_IS_GANG(zio->io_bp)); 1123 ASSERT(!BP_IS_HOLE(zio->io_bp)); 1124 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 1125 1126 abd_put(zio->io_abd); 1127 1128 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED); 1129 1130 mutex_enter(&zilog->zl_lock); 1131 lwb->lwb_write_zio = NULL; 1132 mutex_exit(&zilog->zl_lock); 1133 1134 if (avl_numnodes(t) == 0) 1135 return; 1136 1137 /* 1138 * If there was an IO error, we're not going to call zio_flush() 1139 * on these vdevs, so we simply empty the tree and free the 1140 * nodes. We avoid calling zio_flush() since there isn't any 1141 * good reason for doing so, after the lwb block failed to be 1142 * written out. 1143 */ 1144 if (zio->io_error != 0) { 1145 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) 1146 kmem_free(zv, sizeof (*zv)); 1147 return; 1148 } 1149 1150 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 1151 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 1152 if (vd != NULL) 1153 zio_flush(lwb->lwb_root_zio, vd); 1154 kmem_free(zv, sizeof (*zv)); 1155 } 1156 } 1157 1158 /* 1159 * This function's purpose is to "open" an lwb such that it is ready to 1160 * accept new itxs being committed to it. To do this, the lwb's zio 1161 * structures are created, and linked to the lwb. This function is 1162 * idempotent; if the passed in lwb has already been opened, this 1163 * function is essentially a no-op. 1164 */ 1165 static void 1166 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb) 1167 { 1168 zbookmark_phys_t zb; 1169 zio_priority_t prio; 1170 1171 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1172 ASSERT3P(lwb, !=, NULL); 1173 EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED); 1174 EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED); 1175 1176 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 1177 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 1178 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 1179 1180 if (lwb->lwb_root_zio == NULL) { 1181 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, 1182 BP_GET_LSIZE(&lwb->lwb_blk)); 1183 1184 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk) 1185 prio = ZIO_PRIORITY_SYNC_WRITE; 1186 else 1187 prio = ZIO_PRIORITY_ASYNC_WRITE; 1188 1189 lwb->lwb_root_zio = zio_root(zilog->zl_spa, 1190 zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL); 1191 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1192 1193 lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, 1194 zilog->zl_spa, 0, &lwb->lwb_blk, lwb_abd, 1195 BP_GET_LSIZE(&lwb->lwb_blk), zil_lwb_write_done, lwb, 1196 prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 1197 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1198 1199 lwb->lwb_state = LWB_STATE_OPENED; 1200 1201 mutex_enter(&zilog->zl_lock); 1202 1203 /* 1204 * The zilog's "zl_last_lwb_opened" field is used to 1205 * build the lwb/zio dependency chain, which is used to 1206 * preserve the ordering of lwb completions that is 1207 * required by the semantics of the ZIL. Each new lwb 1208 * zio becomes a parent of the "previous" lwb zio, such 1209 * that the new lwb's zio cannot complete until the 1210 * "previous" lwb's zio completes. 1211 * 1212 * This is required by the semantics of zil_commit(); 1213 * the commit waiters attached to the lwbs will be woken 1214 * in the lwb zio's completion callback, so this zio 1215 * dependency graph ensures the waiters are woken in the 1216 * correct order (the same order the lwbs were created). 1217 */ 1218 lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened; 1219 if (last_lwb_opened != NULL && 1220 last_lwb_opened->lwb_state != LWB_STATE_DONE) { 1221 ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED || 1222 last_lwb_opened->lwb_state == LWB_STATE_ISSUED); 1223 ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL); 1224 zio_add_child(lwb->lwb_root_zio, 1225 last_lwb_opened->lwb_root_zio); 1226 } 1227 zilog->zl_last_lwb_opened = lwb; 1228 1229 mutex_exit(&zilog->zl_lock); 1230 } 1231 1232 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1233 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1234 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1235 } 1236 1237 /* 1238 * Define a limited set of intent log block sizes. 1239 * 1240 * These must be a multiple of 4KB. Note only the amount used (again 1241 * aligned to 4KB) actually gets written. However, we can't always just 1242 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 1243 */ 1244 uint64_t zil_block_buckets[] = { 1245 4096, /* non TX_WRITE */ 1246 8192+4096, /* data base */ 1247 32*1024 + 4096, /* NFS writes */ 1248 UINT64_MAX 1249 }; 1250 1251 /* 1252 * Start a log block write and advance to the next log block. 1253 * Calls are serialized. 1254 */ 1255 static lwb_t * 1256 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb) 1257 { 1258 lwb_t *nlwb = NULL; 1259 zil_chain_t *zilc; 1260 spa_t *spa = zilog->zl_spa; 1261 blkptr_t *bp; 1262 dmu_tx_t *tx; 1263 uint64_t txg; 1264 uint64_t zil_blksz, wsz; 1265 int i, error; 1266 boolean_t slog; 1267 1268 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1269 ASSERT3P(lwb->lwb_root_zio, !=, NULL); 1270 ASSERT3P(lwb->lwb_write_zio, !=, NULL); 1271 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 1272 1273 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1274 zilc = (zil_chain_t *)lwb->lwb_buf; 1275 bp = &zilc->zc_next_blk; 1276 } else { 1277 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 1278 bp = &zilc->zc_next_blk; 1279 } 1280 1281 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 1282 1283 /* 1284 * Allocate the next block and save its address in this block 1285 * before writing it in order to establish the log chain. 1286 * Note that if the allocation of nlwb synced before we wrote 1287 * the block that points at it (lwb), we'd leak it if we crashed. 1288 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 1289 * We dirty the dataset to ensure that zil_sync() will be called 1290 * to clean up in the event of allocation failure or I/O failure. 1291 */ 1292 1293 tx = dmu_tx_create(zilog->zl_os); 1294 1295 /* 1296 * Since we are not going to create any new dirty data, and we 1297 * can even help with clearing the existing dirty data, we 1298 * should not be subject to the dirty data based delays. We 1299 * use TXG_NOTHROTTLE to bypass the delay mechanism. 1300 */ 1301 VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); 1302 1303 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1304 txg = dmu_tx_get_txg(tx); 1305 1306 lwb->lwb_tx = tx; 1307 1308 /* 1309 * Log blocks are pre-allocated. Here we select the size of the next 1310 * block, based on size used in the last block. 1311 * - first find the smallest bucket that will fit the block from a 1312 * limited set of block sizes. This is because it's faster to write 1313 * blocks allocated from the same metaslab as they are adjacent or 1314 * close. 1315 * - next find the maximum from the new suggested size and an array of 1316 * previous sizes. This lessens a picket fence effect of wrongly 1317 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 1318 * requests. 1319 * 1320 * Note we only write what is used, but we can't just allocate 1321 * the maximum block size because we can exhaust the available 1322 * pool log space. 1323 */ 1324 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1325 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 1326 continue; 1327 zil_blksz = zil_block_buckets[i]; 1328 if (zil_blksz == UINT64_MAX) 1329 zil_blksz = SPA_OLD_MAXBLOCKSIZE; 1330 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1331 for (i = 0; i < ZIL_PREV_BLKS; i++) 1332 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1333 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1334 1335 BP_ZERO(bp); 1336 1337 /* pass the old blkptr in order to spread log blocks across devs */ 1338 error = zio_alloc_zil(spa, zilog->zl_os->os_dsl_dataset->ds_object, 1339 txg, bp, &lwb->lwb_blk, zil_blksz, &slog); 1340 if (error == 0) { 1341 ASSERT3U(bp->blk_birth, ==, txg); 1342 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1343 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1344 1345 /* 1346 * Allocate a new log write block (lwb). 1347 */ 1348 nlwb = zil_alloc_lwb(zilog, bp, slog, txg); 1349 } 1350 1351 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1352 /* For Slim ZIL only write what is used. */ 1353 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1354 ASSERT3U(wsz, <=, lwb->lwb_sz); 1355 zio_shrink(lwb->lwb_write_zio, wsz); 1356 1357 } else { 1358 wsz = lwb->lwb_sz; 1359 } 1360 1361 zilc->zc_pad = 0; 1362 zilc->zc_nused = lwb->lwb_nused; 1363 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1364 1365 /* 1366 * clear unused data for security 1367 */ 1368 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1369 1370 spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER); 1371 1372 zil_lwb_add_block(lwb, &lwb->lwb_blk); 1373 lwb->lwb_issued_timestamp = gethrtime(); 1374 lwb->lwb_state = LWB_STATE_ISSUED; 1375 1376 zio_nowait(lwb->lwb_root_zio); 1377 zio_nowait(lwb->lwb_write_zio); 1378 1379 /* 1380 * If there was an allocation failure then nlwb will be null which 1381 * forces a txg_wait_synced(). 1382 */ 1383 return (nlwb); 1384 } 1385 1386 static lwb_t * 1387 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1388 { 1389 lr_t *lrcb, *lrc; 1390 lr_write_t *lrwb, *lrw; 1391 char *lr_buf; 1392 uint64_t dlen, dnow, lwb_sp, reclen, txg; 1393 1394 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1395 ASSERT3P(lwb, !=, NULL); 1396 ASSERT3P(lwb->lwb_buf, !=, NULL); 1397 1398 zil_lwb_write_open(zilog, lwb); 1399 1400 lrc = &itx->itx_lr; 1401 lrw = (lr_write_t *)lrc; 1402 1403 /* 1404 * A commit itx doesn't represent any on-disk state; instead 1405 * it's simply used as a place holder on the commit list, and 1406 * provides a mechanism for attaching a "commit waiter" onto the 1407 * correct lwb (such that the waiter can be signalled upon 1408 * completion of that lwb). Thus, we don't process this itx's 1409 * log record if it's a commit itx (these itx's don't have log 1410 * records), and instead link the itx's waiter onto the lwb's 1411 * list of waiters. 1412 * 1413 * For more details, see the comment above zil_commit(). 1414 */ 1415 if (lrc->lrc_txtype == TX_COMMIT) { 1416 mutex_enter(&zilog->zl_lock); 1417 zil_commit_waiter_link_lwb(itx->itx_private, lwb); 1418 itx->itx_private = NULL; 1419 mutex_exit(&zilog->zl_lock); 1420 return (lwb); 1421 } 1422 1423 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { 1424 dlen = P2ROUNDUP_TYPED( 1425 lrw->lr_length, sizeof (uint64_t), uint64_t); 1426 } else { 1427 dlen = 0; 1428 } 1429 reclen = lrc->lrc_reclen; 1430 zilog->zl_cur_used += (reclen + dlen); 1431 txg = lrc->lrc_txg; 1432 1433 ASSERT3U(zilog->zl_cur_used, <, UINT64_MAX - (reclen + dlen)); 1434 1435 cont: 1436 /* 1437 * If this record won't fit in the current log block, start a new one. 1438 * For WR_NEED_COPY optimize layout for minimal number of chunks. 1439 */ 1440 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1441 if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1442 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 || 1443 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) { 1444 lwb = zil_lwb_write_issue(zilog, lwb); 1445 if (lwb == NULL) 1446 return (NULL); 1447 zil_lwb_write_open(zilog, lwb); 1448 ASSERT(LWB_EMPTY(lwb)); 1449 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1450 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp); 1451 } 1452 1453 dnow = MIN(dlen, lwb_sp - reclen); 1454 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1455 bcopy(lrc, lr_buf, reclen); 1456 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */ 1457 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */ 1458 1459 /* 1460 * If it's a write, fetch the data or get its blkptr as appropriate. 1461 */ 1462 if (lrc->lrc_txtype == TX_WRITE) { 1463 if (txg > spa_freeze_txg(zilog->zl_spa)) 1464 txg_wait_synced(zilog->zl_dmu_pool, txg); 1465 if (itx->itx_wr_state != WR_COPIED) { 1466 char *dbuf; 1467 int error; 1468 1469 if (itx->itx_wr_state == WR_NEED_COPY) { 1470 dbuf = lr_buf + reclen; 1471 lrcb->lrc_reclen += dnow; 1472 if (lrwb->lr_length > dnow) 1473 lrwb->lr_length = dnow; 1474 lrw->lr_offset += dnow; 1475 lrw->lr_length -= dnow; 1476 } else { 1477 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1478 dbuf = NULL; 1479 } 1480 1481 /* 1482 * We pass in the "lwb_write_zio" rather than 1483 * "lwb_root_zio" so that the "lwb_write_zio" 1484 * becomes the parent of any zio's created by 1485 * the "zl_get_data" callback. The vdevs are 1486 * flushed after the "lwb_write_zio" completes, 1487 * so we want to make sure that completion 1488 * callback waits for these additional zio's, 1489 * such that the vdevs used by those zio's will 1490 * be included in the lwb's vdev tree, and those 1491 * vdevs will be properly flushed. If we passed 1492 * in "lwb_root_zio" here, then these additional 1493 * vdevs may not be flushed; e.g. if these zio's 1494 * completed after "lwb_write_zio" completed. 1495 */ 1496 error = zilog->zl_get_data(itx->itx_private, 1497 lrwb, dbuf, lwb, lwb->lwb_write_zio); 1498 1499 if (error == EIO) { 1500 txg_wait_synced(zilog->zl_dmu_pool, txg); 1501 return (lwb); 1502 } 1503 if (error != 0) { 1504 ASSERT(error == ENOENT || error == EEXIST || 1505 error == EALREADY); 1506 return (lwb); 1507 } 1508 } 1509 } 1510 1511 /* 1512 * We're actually making an entry, so update lrc_seq to be the 1513 * log record sequence number. Note that this is generally not 1514 * equal to the itx sequence number because not all transactions 1515 * are synchronous, and sometimes spa_sync() gets there first. 1516 */ 1517 lrcb->lrc_seq = ++zilog->zl_lr_seq; 1518 lwb->lwb_nused += reclen + dnow; 1519 1520 zil_lwb_add_txg(lwb, txg); 1521 1522 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1523 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1524 1525 dlen -= dnow; 1526 if (dlen > 0) { 1527 zilog->zl_cur_used += reclen; 1528 goto cont; 1529 } 1530 1531 return (lwb); 1532 } 1533 1534 itx_t * 1535 zil_itx_create(uint64_t txtype, size_t lrsize) 1536 { 1537 itx_t *itx; 1538 1539 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1540 1541 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1542 itx->itx_lr.lrc_txtype = txtype; 1543 itx->itx_lr.lrc_reclen = lrsize; 1544 itx->itx_lr.lrc_seq = 0; /* defensive */ 1545 itx->itx_sync = B_TRUE; /* default is synchronous */ 1546 1547 return (itx); 1548 } 1549 1550 void 1551 zil_itx_destroy(itx_t *itx) 1552 { 1553 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1554 } 1555 1556 /* 1557 * Free up the sync and async itxs. The itxs_t has already been detached 1558 * so no locks are needed. 1559 */ 1560 static void 1561 zil_itxg_clean(itxs_t *itxs) 1562 { 1563 itx_t *itx; 1564 list_t *list; 1565 avl_tree_t *t; 1566 void *cookie; 1567 itx_async_node_t *ian; 1568 1569 list = &itxs->i_sync_list; 1570 while ((itx = list_head(list)) != NULL) { 1571 /* 1572 * In the general case, commit itxs will not be found 1573 * here, as they'll be committed to an lwb via 1574 * zil_lwb_commit(), and free'd in that function. Having 1575 * said that, it is still possible for commit itxs to be 1576 * found here, due to the following race: 1577 * 1578 * - a thread calls zil_commit() which assigns the 1579 * commit itx to a per-txg i_sync_list 1580 * - zil_itxg_clean() is called (e.g. via spa_sync()) 1581 * while the waiter is still on the i_sync_list 1582 * 1583 * There's nothing to prevent syncing the txg while the 1584 * waiter is on the i_sync_list. This normally doesn't 1585 * happen because spa_sync() is slower than zil_commit(), 1586 * but if zil_commit() calls txg_wait_synced() (e.g. 1587 * because zil_create() or zil_commit_writer_stall() is 1588 * called) we will hit this case. 1589 */ 1590 if (itx->itx_lr.lrc_txtype == TX_COMMIT) 1591 zil_commit_waiter_skip(itx->itx_private); 1592 1593 list_remove(list, itx); 1594 zil_itx_destroy(itx); 1595 } 1596 1597 cookie = NULL; 1598 t = &itxs->i_async_tree; 1599 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1600 list = &ian->ia_list; 1601 while ((itx = list_head(list)) != NULL) { 1602 list_remove(list, itx); 1603 /* commit itxs should never be on the async lists. */ 1604 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 1605 zil_itx_destroy(itx); 1606 } 1607 list_destroy(list); 1608 kmem_free(ian, sizeof (itx_async_node_t)); 1609 } 1610 avl_destroy(t); 1611 1612 kmem_free(itxs, sizeof (itxs_t)); 1613 } 1614 1615 static int 1616 zil_aitx_compare(const void *x1, const void *x2) 1617 { 1618 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1619 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1620 1621 if (o1 < o2) 1622 return (-1); 1623 if (o1 > o2) 1624 return (1); 1625 1626 return (0); 1627 } 1628 1629 /* 1630 * Remove all async itx with the given oid. 1631 */ 1632 static void 1633 zil_remove_async(zilog_t *zilog, uint64_t oid) 1634 { 1635 uint64_t otxg, txg; 1636 itx_async_node_t *ian; 1637 avl_tree_t *t; 1638 avl_index_t where; 1639 list_t clean_list; 1640 itx_t *itx; 1641 1642 ASSERT(oid != 0); 1643 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1644 1645 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1646 otxg = ZILTEST_TXG; 1647 else 1648 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1649 1650 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1651 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1652 1653 mutex_enter(&itxg->itxg_lock); 1654 if (itxg->itxg_txg != txg) { 1655 mutex_exit(&itxg->itxg_lock); 1656 continue; 1657 } 1658 1659 /* 1660 * Locate the object node and append its list. 1661 */ 1662 t = &itxg->itxg_itxs->i_async_tree; 1663 ian = avl_find(t, &oid, &where); 1664 if (ian != NULL) 1665 list_move_tail(&clean_list, &ian->ia_list); 1666 mutex_exit(&itxg->itxg_lock); 1667 } 1668 while ((itx = list_head(&clean_list)) != NULL) { 1669 list_remove(&clean_list, itx); 1670 /* commit itxs should never be on the async lists. */ 1671 ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT); 1672 zil_itx_destroy(itx); 1673 } 1674 list_destroy(&clean_list); 1675 } 1676 1677 void 1678 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1679 { 1680 uint64_t txg; 1681 itxg_t *itxg; 1682 itxs_t *itxs, *clean = NULL; 1683 1684 /* 1685 * Object ids can be re-instantiated in the next txg so 1686 * remove any async transactions to avoid future leaks. 1687 * This can happen if a fsync occurs on the re-instantiated 1688 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1689 * the new file data and flushes a write record for the old object. 1690 */ 1691 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1692 zil_remove_async(zilog, itx->itx_oid); 1693 1694 /* 1695 * Ensure the data of a renamed file is committed before the rename. 1696 */ 1697 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1698 zil_async_to_sync(zilog, itx->itx_oid); 1699 1700 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1701 txg = ZILTEST_TXG; 1702 else 1703 txg = dmu_tx_get_txg(tx); 1704 1705 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1706 mutex_enter(&itxg->itxg_lock); 1707 itxs = itxg->itxg_itxs; 1708 if (itxg->itxg_txg != txg) { 1709 if (itxs != NULL) { 1710 /* 1711 * The zil_clean callback hasn't got around to cleaning 1712 * this itxg. Save the itxs for release below. 1713 * This should be rare. 1714 */ 1715 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for " 1716 "txg %llu", itxg->itxg_txg); 1717 clean = itxg->itxg_itxs; 1718 } 1719 itxg->itxg_txg = txg; 1720 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1721 1722 list_create(&itxs->i_sync_list, sizeof (itx_t), 1723 offsetof(itx_t, itx_node)); 1724 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1725 sizeof (itx_async_node_t), 1726 offsetof(itx_async_node_t, ia_node)); 1727 } 1728 if (itx->itx_sync) { 1729 list_insert_tail(&itxs->i_sync_list, itx); 1730 } else { 1731 avl_tree_t *t = &itxs->i_async_tree; 1732 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1733 itx_async_node_t *ian; 1734 avl_index_t where; 1735 1736 ian = avl_find(t, &foid, &where); 1737 if (ian == NULL) { 1738 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1739 list_create(&ian->ia_list, sizeof (itx_t), 1740 offsetof(itx_t, itx_node)); 1741 ian->ia_foid = foid; 1742 avl_insert(t, ian, where); 1743 } 1744 list_insert_tail(&ian->ia_list, itx); 1745 } 1746 1747 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1748 1749 /* 1750 * We don't want to dirty the ZIL using ZILTEST_TXG, because 1751 * zil_clean() will never be called using ZILTEST_TXG. Thus, we 1752 * need to be careful to always dirty the ZIL using the "real" 1753 * TXG (not itxg_txg) even when the SPA is frozen. 1754 */ 1755 zilog_dirty(zilog, dmu_tx_get_txg(tx)); 1756 mutex_exit(&itxg->itxg_lock); 1757 1758 /* Release the old itxs now we've dropped the lock */ 1759 if (clean != NULL) 1760 zil_itxg_clean(clean); 1761 } 1762 1763 /* 1764 * If there are any in-memory intent log transactions which have now been 1765 * synced then start up a taskq to free them. We should only do this after we 1766 * have written out the uberblocks (i.e. txg has been comitted) so that 1767 * don't inadvertently clean out in-memory log records that would be required 1768 * by zil_commit(). 1769 */ 1770 void 1771 zil_clean(zilog_t *zilog, uint64_t synced_txg) 1772 { 1773 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1774 itxs_t *clean_me; 1775 1776 ASSERT3U(synced_txg, <, ZILTEST_TXG); 1777 1778 mutex_enter(&itxg->itxg_lock); 1779 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1780 mutex_exit(&itxg->itxg_lock); 1781 return; 1782 } 1783 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1784 ASSERT3U(itxg->itxg_txg, !=, 0); 1785 clean_me = itxg->itxg_itxs; 1786 itxg->itxg_itxs = NULL; 1787 itxg->itxg_txg = 0; 1788 mutex_exit(&itxg->itxg_lock); 1789 /* 1790 * Preferably start a task queue to free up the old itxs but 1791 * if taskq_dispatch can't allocate resources to do that then 1792 * free it in-line. This should be rare. Note, using TQ_SLEEP 1793 * created a bad performance problem. 1794 */ 1795 ASSERT3P(zilog->zl_dmu_pool, !=, NULL); 1796 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL); 1797 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq, 1798 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL) 1799 zil_itxg_clean(clean_me); 1800 } 1801 1802 /* 1803 * This function will traverse the queue of itxs that need to be 1804 * committed, and move them onto the ZIL's zl_itx_commit_list. 1805 */ 1806 static void 1807 zil_get_commit_list(zilog_t *zilog) 1808 { 1809 uint64_t otxg, txg; 1810 list_t *commit_list = &zilog->zl_itx_commit_list; 1811 1812 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1813 1814 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1815 otxg = ZILTEST_TXG; 1816 else 1817 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1818 1819 /* 1820 * This is inherently racy, since there is nothing to prevent 1821 * the last synced txg from changing. That's okay since we'll 1822 * only commit things in the future. 1823 */ 1824 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1825 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1826 1827 mutex_enter(&itxg->itxg_lock); 1828 if (itxg->itxg_txg != txg) { 1829 mutex_exit(&itxg->itxg_lock); 1830 continue; 1831 } 1832 1833 /* 1834 * If we're adding itx records to the zl_itx_commit_list, 1835 * then the zil better be dirty in this "txg". We can assert 1836 * that here since we're holding the itxg_lock which will 1837 * prevent spa_sync from cleaning it. Once we add the itxs 1838 * to the zl_itx_commit_list we must commit it to disk even 1839 * if it's unnecessary (i.e. the txg was synced). 1840 */ 1841 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 1842 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1843 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1844 1845 mutex_exit(&itxg->itxg_lock); 1846 } 1847 } 1848 1849 /* 1850 * Move the async itxs for a specified object to commit into sync lists. 1851 */ 1852 static void 1853 zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1854 { 1855 uint64_t otxg, txg; 1856 itx_async_node_t *ian; 1857 avl_tree_t *t; 1858 avl_index_t where; 1859 1860 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1861 otxg = ZILTEST_TXG; 1862 else 1863 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1864 1865 /* 1866 * This is inherently racy, since there is nothing to prevent 1867 * the last synced txg from changing. 1868 */ 1869 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1870 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1871 1872 mutex_enter(&itxg->itxg_lock); 1873 if (itxg->itxg_txg != txg) { 1874 mutex_exit(&itxg->itxg_lock); 1875 continue; 1876 } 1877 1878 /* 1879 * If a foid is specified then find that node and append its 1880 * list. Otherwise walk the tree appending all the lists 1881 * to the sync list. We add to the end rather than the 1882 * beginning to ensure the create has happened. 1883 */ 1884 t = &itxg->itxg_itxs->i_async_tree; 1885 if (foid != 0) { 1886 ian = avl_find(t, &foid, &where); 1887 if (ian != NULL) { 1888 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1889 &ian->ia_list); 1890 } 1891 } else { 1892 void *cookie = NULL; 1893 1894 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1895 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1896 &ian->ia_list); 1897 list_destroy(&ian->ia_list); 1898 kmem_free(ian, sizeof (itx_async_node_t)); 1899 } 1900 } 1901 mutex_exit(&itxg->itxg_lock); 1902 } 1903 } 1904 1905 /* 1906 * This function will prune commit itxs that are at the head of the 1907 * commit list (it won't prune past the first non-commit itx), and 1908 * either: a) attach them to the last lwb that's still pending 1909 * completion, or b) skip them altogether. 1910 * 1911 * This is used as a performance optimization to prevent commit itxs 1912 * from generating new lwbs when it's unnecessary to do so. 1913 */ 1914 static void 1915 zil_prune_commit_list(zilog_t *zilog) 1916 { 1917 itx_t *itx; 1918 1919 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1920 1921 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1922 lr_t *lrc = &itx->itx_lr; 1923 if (lrc->lrc_txtype != TX_COMMIT) 1924 break; 1925 1926 mutex_enter(&zilog->zl_lock); 1927 1928 lwb_t *last_lwb = zilog->zl_last_lwb_opened; 1929 if (last_lwb == NULL || last_lwb->lwb_state == LWB_STATE_DONE) { 1930 /* 1931 * All of the itxs this waiter was waiting on 1932 * must have already completed (or there were 1933 * never any itx's for it to wait on), so it's 1934 * safe to skip this waiter and mark it done. 1935 */ 1936 zil_commit_waiter_skip(itx->itx_private); 1937 } else { 1938 zil_commit_waiter_link_lwb(itx->itx_private, last_lwb); 1939 itx->itx_private = NULL; 1940 } 1941 1942 mutex_exit(&zilog->zl_lock); 1943 1944 list_remove(&zilog->zl_itx_commit_list, itx); 1945 zil_itx_destroy(itx); 1946 } 1947 1948 IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT); 1949 } 1950 1951 static void 1952 zil_commit_writer_stall(zilog_t *zilog) 1953 { 1954 /* 1955 * When zio_alloc_zil() fails to allocate the next lwb block on 1956 * disk, we must call txg_wait_synced() to ensure all of the 1957 * lwbs in the zilog's zl_lwb_list are synced and then freed (in 1958 * zil_sync()), such that any subsequent ZIL writer (i.e. a call 1959 * to zil_process_commit_list()) will have to call zil_create(), 1960 * and start a new ZIL chain. 1961 * 1962 * Since zil_alloc_zil() failed, the lwb that was previously 1963 * issued does not have a pointer to the "next" lwb on disk. 1964 * Thus, if another ZIL writer thread was to allocate the "next" 1965 * on-disk lwb, that block could be leaked in the event of a 1966 * crash (because the previous lwb on-disk would not point to 1967 * it). 1968 * 1969 * We must hold the zilog's zl_issuer_lock while we do this, to 1970 * ensure no new threads enter zil_process_commit_list() until 1971 * all lwb's in the zl_lwb_list have been synced and freed 1972 * (which is achieved via the txg_wait_synced() call). 1973 */ 1974 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1975 txg_wait_synced(zilog->zl_dmu_pool, 0); 1976 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 1977 } 1978 1979 /* 1980 * This function will traverse the commit list, creating new lwbs as 1981 * needed, and committing the itxs from the commit list to these newly 1982 * created lwbs. Additionally, as a new lwb is created, the previous 1983 * lwb will be issued to the zio layer to be written to disk. 1984 */ 1985 static void 1986 zil_process_commit_list(zilog_t *zilog) 1987 { 1988 spa_t *spa = zilog->zl_spa; 1989 list_t nolwb_waiters; 1990 lwb_t *lwb; 1991 itx_t *itx; 1992 1993 ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); 1994 1995 /* 1996 * Return if there's nothing to commit before we dirty the fs by 1997 * calling zil_create(). 1998 */ 1999 if (list_head(&zilog->zl_itx_commit_list) == NULL) 2000 return; 2001 2002 list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t), 2003 offsetof(zil_commit_waiter_t, zcw_node)); 2004 2005 lwb = list_tail(&zilog->zl_lwb_list); 2006 if (lwb == NULL) { 2007 lwb = zil_create(zilog); 2008 } else { 2009 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2010 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE); 2011 } 2012 2013 while (itx = list_head(&zilog->zl_itx_commit_list)) { 2014 lr_t *lrc = &itx->itx_lr; 2015 uint64_t txg = lrc->lrc_txg; 2016 2017 ASSERT3U(txg, !=, 0); 2018 2019 if (lrc->lrc_txtype == TX_COMMIT) { 2020 DTRACE_PROBE2(zil__process__commit__itx, 2021 zilog_t *, zilog, itx_t *, itx); 2022 } else { 2023 DTRACE_PROBE2(zil__process__normal__itx, 2024 zilog_t *, zilog, itx_t *, itx); 2025 } 2026 2027 boolean_t synced = txg <= spa_last_synced_txg(spa); 2028 boolean_t frozen = txg > spa_freeze_txg(spa); 2029 2030 /* 2031 * If the txg of this itx has already been synced out, then 2032 * we don't need to commit this itx to an lwb. This is 2033 * because the data of this itx will have already been 2034 * written to the main pool. This is inherently racy, and 2035 * it's still ok to commit an itx whose txg has already 2036 * been synced; this will result in a write that's 2037 * unnecessary, but will do no harm. 2038 * 2039 * With that said, we always want to commit TX_COMMIT itxs 2040 * to an lwb, regardless of whether or not that itx's txg 2041 * has been synced out. We do this to ensure any OPENED lwb 2042 * will always have at least one zil_commit_waiter_t linked 2043 * to the lwb. 2044 * 2045 * As a counter-example, if we skipped TX_COMMIT itx's 2046 * whose txg had already been synced, the following 2047 * situation could occur if we happened to be racing with 2048 * spa_sync: 2049 * 2050 * 1. we commit a non-TX_COMMIT itx to an lwb, where the 2051 * itx's txg is 10 and the last synced txg is 9. 2052 * 2. spa_sync finishes syncing out txg 10. 2053 * 3. we move to the next itx in the list, it's a TX_COMMIT 2054 * whose txg is 10, so we skip it rather than committing 2055 * it to the lwb used in (1). 2056 * 2057 * If the itx that is skipped in (3) is the last TX_COMMIT 2058 * itx in the commit list, than it's possible for the lwb 2059 * used in (1) to remain in the OPENED state indefinitely. 2060 * 2061 * To prevent the above scenario from occuring, ensuring 2062 * that once an lwb is OPENED it will transition to ISSUED 2063 * and eventually DONE, we always commit TX_COMMIT itx's to 2064 * an lwb here, even if that itx's txg has already been 2065 * synced. 2066 * 2067 * Finally, if the pool is frozen, we _always_ commit the 2068 * itx. The point of freezing the pool is to prevent data 2069 * from being written to the main pool via spa_sync, and 2070 * instead rely solely on the ZIL to persistently store the 2071 * data; i.e. when the pool is frozen, the last synced txg 2072 * value can't be trusted. 2073 */ 2074 if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) { 2075 if (lwb != NULL) { 2076 lwb = zil_lwb_commit(zilog, itx, lwb); 2077 } else if (lrc->lrc_txtype == TX_COMMIT) { 2078 ASSERT3P(lwb, ==, NULL); 2079 zil_commit_waiter_link_nolwb( 2080 itx->itx_private, &nolwb_waiters); 2081 } 2082 } 2083 2084 list_remove(&zilog->zl_itx_commit_list, itx); 2085 zil_itx_destroy(itx); 2086 } 2087 2088 if (lwb == NULL) { 2089 /* 2090 * This indicates zio_alloc_zil() failed to allocate the 2091 * "next" lwb on-disk. When this happens, we must stall 2092 * the ZIL write pipeline; see the comment within 2093 * zil_commit_writer_stall() for more details. 2094 */ 2095 zil_commit_writer_stall(zilog); 2096 2097 /* 2098 * Additionally, we have to signal and mark the "nolwb" 2099 * waiters as "done" here, since without an lwb, we 2100 * can't do this via zil_lwb_flush_vdevs_done() like 2101 * normal. 2102 */ 2103 zil_commit_waiter_t *zcw; 2104 while (zcw = list_head(&nolwb_waiters)) { 2105 zil_commit_waiter_skip(zcw); 2106 list_remove(&nolwb_waiters, zcw); 2107 } 2108 } else { 2109 ASSERT(list_is_empty(&nolwb_waiters)); 2110 ASSERT3P(lwb, !=, NULL); 2111 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2112 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_DONE); 2113 2114 /* 2115 * At this point, the ZIL block pointed at by the "lwb" 2116 * variable is in one of the following states: "closed" 2117 * or "open". 2118 * 2119 * If its "closed", then no itxs have been committed to 2120 * it, so there's no point in issuing its zio (i.e. 2121 * it's "empty"). 2122 * 2123 * If its "open" state, then it contains one or more 2124 * itxs that eventually need to be committed to stable 2125 * storage. In this case we intentionally do not issue 2126 * the lwb's zio to disk yet, and instead rely on one of 2127 * the following two mechanisms for issuing the zio: 2128 * 2129 * 1. Ideally, there will be more ZIL activity occuring 2130 * on the system, such that this function will be 2131 * immediately called again (not necessarily by the same 2132 * thread) and this lwb's zio will be issued via 2133 * zil_lwb_commit(). This way, the lwb is guaranteed to 2134 * be "full" when it is issued to disk, and we'll make 2135 * use of the lwb's size the best we can. 2136 * 2137 * 2. If there isn't sufficient ZIL activity occuring on 2138 * the system, such that this lwb's zio isn't issued via 2139 * zil_lwb_commit(), zil_commit_waiter() will issue the 2140 * lwb's zio. If this occurs, the lwb is not guaranteed 2141 * to be "full" by the time its zio is issued, and means 2142 * the size of the lwb was "too large" given the amount 2143 * of ZIL activity occuring on the system at that time. 2144 * 2145 * We do this for a couple of reasons: 2146 * 2147 * 1. To try and reduce the number of IOPs needed to 2148 * write the same number of itxs. If an lwb has space 2149 * available in it's buffer for more itxs, and more itxs 2150 * will be committed relatively soon (relative to the 2151 * latency of performing a write), then it's beneficial 2152 * to wait for these "next" itxs. This way, more itxs 2153 * can be committed to stable storage with fewer writes. 2154 * 2155 * 2. To try and use the largest lwb block size that the 2156 * incoming rate of itxs can support. Again, this is to 2157 * try and pack as many itxs into as few lwbs as 2158 * possible, without significantly impacting the latency 2159 * of each individual itx. 2160 */ 2161 } 2162 } 2163 2164 /* 2165 * This function is responsible for ensuring the passed in commit waiter 2166 * (and associated commit itx) is committed to an lwb. If the waiter is 2167 * not already committed to an lwb, all itxs in the zilog's queue of 2168 * itxs will be processed. The assumption is the passed in waiter's 2169 * commit itx will found in the queue just like the other non-commit 2170 * itxs, such that when the entire queue is processed, the waiter will 2171 * have been commited to an lwb. 2172 * 2173 * The lwb associated with the passed in waiter is not guaranteed to 2174 * have been issued by the time this function completes. If the lwb is 2175 * not issued, we rely on future calls to zil_commit_writer() to issue 2176 * the lwb, or the timeout mechanism found in zil_commit_waiter(). 2177 */ 2178 static void 2179 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw) 2180 { 2181 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2182 ASSERT(spa_writeable(zilog->zl_spa)); 2183 2184 mutex_enter(&zilog->zl_issuer_lock); 2185 2186 if (zcw->zcw_lwb != NULL || zcw->zcw_done) { 2187 /* 2188 * It's possible that, while we were waiting to acquire 2189 * the "zl_issuer_lock", another thread committed this 2190 * waiter to an lwb. If that occurs, we bail out early, 2191 * without processing any of the zilog's queue of itxs. 2192 * 2193 * On certain workloads and system configurations, the 2194 * "zl_issuer_lock" can become highly contended. In an 2195 * attempt to reduce this contention, we immediately drop 2196 * the lock if the waiter has already been processed. 2197 * 2198 * We've measured this optimization to reduce CPU spent 2199 * contending on this lock by up to 5%, using a system 2200 * with 32 CPUs, low latency storage (~50 usec writes), 2201 * and 1024 threads performing sync writes. 2202 */ 2203 goto out; 2204 } 2205 2206 zil_get_commit_list(zilog); 2207 zil_prune_commit_list(zilog); 2208 zil_process_commit_list(zilog); 2209 2210 out: 2211 mutex_exit(&zilog->zl_issuer_lock); 2212 } 2213 2214 static void 2215 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw) 2216 { 2217 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2218 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2219 ASSERT3B(zcw->zcw_done, ==, B_FALSE); 2220 2221 lwb_t *lwb = zcw->zcw_lwb; 2222 ASSERT3P(lwb, !=, NULL); 2223 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED); 2224 2225 /* 2226 * If the lwb has already been issued by another thread, we can 2227 * immediately return since there's no work to be done (the 2228 * point of this function is to issue the lwb). Additionally, we 2229 * do this prior to acquiring the zl_issuer_lock, to avoid 2230 * acquiring it when it's not necessary to do so. 2231 */ 2232 if (lwb->lwb_state == LWB_STATE_ISSUED || 2233 lwb->lwb_state == LWB_STATE_DONE) 2234 return; 2235 2236 /* 2237 * In order to call zil_lwb_write_issue() we must hold the 2238 * zilog's "zl_issuer_lock". We can't simply acquire that lock, 2239 * since we're already holding the commit waiter's "zcw_lock", 2240 * and those two locks are aquired in the opposite order 2241 * elsewhere. 2242 */ 2243 mutex_exit(&zcw->zcw_lock); 2244 mutex_enter(&zilog->zl_issuer_lock); 2245 mutex_enter(&zcw->zcw_lock); 2246 2247 /* 2248 * Since we just dropped and re-acquired the commit waiter's 2249 * lock, we have to re-check to see if the waiter was marked 2250 * "done" during that process. If the waiter was marked "done", 2251 * the "lwb" pointer is no longer valid (it can be free'd after 2252 * the waiter is marked "done"), so without this check we could 2253 * wind up with a use-after-free error below. 2254 */ 2255 if (zcw->zcw_done) 2256 goto out; 2257 2258 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2259 2260 /* 2261 * We've already checked this above, but since we hadn't acquired 2262 * the zilog's zl_issuer_lock, we have to perform this check a 2263 * second time while holding the lock. 2264 * 2265 * We don't need to hold the zl_lock since the lwb cannot transition 2266 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb 2267 * _can_ transition from ISSUED to DONE, but it's OK to race with 2268 * that transition since we treat the lwb the same, whether it's in 2269 * the ISSUED or DONE states. 2270 * 2271 * The important thing, is we treat the lwb differently depending on 2272 * if it's ISSUED or OPENED, and block any other threads that might 2273 * attempt to issue this lwb. For that reason we hold the 2274 * zl_issuer_lock when checking the lwb_state; we must not call 2275 * zil_lwb_write_issue() if the lwb had already been issued. 2276 * 2277 * See the comment above the lwb_state_t structure definition for 2278 * more details on the lwb states, and locking requirements. 2279 */ 2280 if (lwb->lwb_state == LWB_STATE_ISSUED || 2281 lwb->lwb_state == LWB_STATE_DONE) 2282 goto out; 2283 2284 ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED); 2285 2286 /* 2287 * As described in the comments above zil_commit_waiter() and 2288 * zil_process_commit_list(), we need to issue this lwb's zio 2289 * since we've reached the commit waiter's timeout and it still 2290 * hasn't been issued. 2291 */ 2292 lwb_t *nlwb = zil_lwb_write_issue(zilog, lwb); 2293 2294 IMPLY(nlwb != NULL, lwb->lwb_state != LWB_STATE_OPENED); 2295 2296 /* 2297 * Since the lwb's zio hadn't been issued by the time this thread 2298 * reached its timeout, we reset the zilog's "zl_cur_used" field 2299 * to influence the zil block size selection algorithm. 2300 * 2301 * By having to issue the lwb's zio here, it means the size of the 2302 * lwb was too large, given the incoming throughput of itxs. By 2303 * setting "zl_cur_used" to zero, we communicate this fact to the 2304 * block size selection algorithm, so it can take this informaiton 2305 * into account, and potentially select a smaller size for the 2306 * next lwb block that is allocated. 2307 */ 2308 zilog->zl_cur_used = 0; 2309 2310 if (nlwb == NULL) { 2311 /* 2312 * When zil_lwb_write_issue() returns NULL, this 2313 * indicates zio_alloc_zil() failed to allocate the 2314 * "next" lwb on-disk. When this occurs, the ZIL write 2315 * pipeline must be stalled; see the comment within the 2316 * zil_commit_writer_stall() function for more details. 2317 * 2318 * We must drop the commit waiter's lock prior to 2319 * calling zil_commit_writer_stall() or else we can wind 2320 * up with the following deadlock: 2321 * 2322 * - This thread is waiting for the txg to sync while 2323 * holding the waiter's lock; txg_wait_synced() is 2324 * used within txg_commit_writer_stall(). 2325 * 2326 * - The txg can't sync because it is waiting for this 2327 * lwb's zio callback to call dmu_tx_commit(). 2328 * 2329 * - The lwb's zio callback can't call dmu_tx_commit() 2330 * because it's blocked trying to acquire the waiter's 2331 * lock, which occurs prior to calling dmu_tx_commit() 2332 */ 2333 mutex_exit(&zcw->zcw_lock); 2334 zil_commit_writer_stall(zilog); 2335 mutex_enter(&zcw->zcw_lock); 2336 } 2337 2338 out: 2339 mutex_exit(&zilog->zl_issuer_lock); 2340 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2341 } 2342 2343 /* 2344 * This function is responsible for performing the following two tasks: 2345 * 2346 * 1. its primary responsibility is to block until the given "commit 2347 * waiter" is considered "done". 2348 * 2349 * 2. its secondary responsibility is to issue the zio for the lwb that 2350 * the given "commit waiter" is waiting on, if this function has 2351 * waited "long enough" and the lwb is still in the "open" state. 2352 * 2353 * Given a sufficient amount of itxs being generated and written using 2354 * the ZIL, the lwb's zio will be issued via the zil_lwb_commit() 2355 * function. If this does not occur, this secondary responsibility will 2356 * ensure the lwb is issued even if there is not other synchronous 2357 * activity on the system. 2358 * 2359 * For more details, see zil_process_commit_list(); more specifically, 2360 * the comment at the bottom of that function. 2361 */ 2362 static void 2363 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw) 2364 { 2365 ASSERT(!MUTEX_HELD(&zilog->zl_lock)); 2366 ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock)); 2367 ASSERT(spa_writeable(zilog->zl_spa)); 2368 2369 mutex_enter(&zcw->zcw_lock); 2370 2371 /* 2372 * The timeout is scaled based on the lwb latency to avoid 2373 * significantly impacting the latency of each individual itx. 2374 * For more details, see the comment at the bottom of the 2375 * zil_process_commit_list() function. 2376 */ 2377 int pct = MAX(zfs_commit_timeout_pct, 1); 2378 hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100; 2379 hrtime_t wakeup = gethrtime() + sleep; 2380 boolean_t timedout = B_FALSE; 2381 2382 while (!zcw->zcw_done) { 2383 ASSERT(MUTEX_HELD(&zcw->zcw_lock)); 2384 2385 lwb_t *lwb = zcw->zcw_lwb; 2386 2387 /* 2388 * Usually, the waiter will have a non-NULL lwb field here, 2389 * but it's possible for it to be NULL as a result of 2390 * zil_commit() racing with spa_sync(). 2391 * 2392 * When zil_clean() is called, it's possible for the itxg 2393 * list (which may be cleaned via a taskq) to contain 2394 * commit itxs. When this occurs, the commit waiters linked 2395 * off of these commit itxs will not be committed to an 2396 * lwb. Additionally, these commit waiters will not be 2397 * marked done until zil_commit_waiter_skip() is called via 2398 * zil_itxg_clean(). 2399 * 2400 * Thus, it's possible for this commit waiter (i.e. the 2401 * "zcw" variable) to be found in this "in between" state; 2402 * where it's "zcw_lwb" field is NULL, and it hasn't yet 2403 * been skipped, so it's "zcw_done" field is still B_FALSE. 2404 */ 2405 IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED); 2406 2407 if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) { 2408 ASSERT3B(timedout, ==, B_FALSE); 2409 2410 /* 2411 * If the lwb hasn't been issued yet, then we 2412 * need to wait with a timeout, in case this 2413 * function needs to issue the lwb after the 2414 * timeout is reached; responsibility (2) from 2415 * the comment above this function. 2416 */ 2417 clock_t timeleft = cv_timedwait_hires(&zcw->zcw_cv, 2418 &zcw->zcw_lock, wakeup, USEC2NSEC(1), 2419 CALLOUT_FLAG_ABSOLUTE); 2420 2421 if (timeleft >= 0 || zcw->zcw_done) 2422 continue; 2423 2424 timedout = B_TRUE; 2425 zil_commit_waiter_timeout(zilog, zcw); 2426 2427 if (!zcw->zcw_done) { 2428 /* 2429 * If the commit waiter has already been 2430 * marked "done", it's possible for the 2431 * waiter's lwb structure to have already 2432 * been freed. Thus, we can only reliably 2433 * make these assertions if the waiter 2434 * isn't done. 2435 */ 2436 ASSERT3P(lwb, ==, zcw->zcw_lwb); 2437 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED); 2438 } 2439 } else { 2440 /* 2441 * If the lwb isn't open, then it must have already 2442 * been issued. In that case, there's no need to 2443 * use a timeout when waiting for the lwb to 2444 * complete. 2445 * 2446 * Additionally, if the lwb is NULL, the waiter 2447 * will soon be signalled and marked done via 2448 * zil_clean() and zil_itxg_clean(), so no timeout 2449 * is required. 2450 */ 2451 2452 IMPLY(lwb != NULL, 2453 lwb->lwb_state == LWB_STATE_ISSUED || 2454 lwb->lwb_state == LWB_STATE_DONE); 2455 cv_wait(&zcw->zcw_cv, &zcw->zcw_lock); 2456 } 2457 } 2458 2459 mutex_exit(&zcw->zcw_lock); 2460 } 2461 2462 static zil_commit_waiter_t * 2463 zil_alloc_commit_waiter() 2464 { 2465 zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP); 2466 2467 cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL); 2468 mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL); 2469 list_link_init(&zcw->zcw_node); 2470 zcw->zcw_lwb = NULL; 2471 zcw->zcw_done = B_FALSE; 2472 zcw->zcw_zio_error = 0; 2473 2474 return (zcw); 2475 } 2476 2477 static void 2478 zil_free_commit_waiter(zil_commit_waiter_t *zcw) 2479 { 2480 ASSERT(!list_link_active(&zcw->zcw_node)); 2481 ASSERT3P(zcw->zcw_lwb, ==, NULL); 2482 ASSERT3B(zcw->zcw_done, ==, B_TRUE); 2483 mutex_destroy(&zcw->zcw_lock); 2484 cv_destroy(&zcw->zcw_cv); 2485 kmem_cache_free(zil_zcw_cache, zcw); 2486 } 2487 2488 /* 2489 * This function is used to create a TX_COMMIT itx and assign it. This 2490 * way, it will be linked into the ZIL's list of synchronous itxs, and 2491 * then later committed to an lwb (or skipped) when 2492 * zil_process_commit_list() is called. 2493 */ 2494 static void 2495 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw) 2496 { 2497 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 2498 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 2499 2500 itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t)); 2501 itx->itx_sync = B_TRUE; 2502 itx->itx_private = zcw; 2503 2504 zil_itx_assign(zilog, itx, tx); 2505 2506 dmu_tx_commit(tx); 2507 } 2508 2509 /* 2510 * Commit ZFS Intent Log transactions (itxs) to stable storage. 2511 * 2512 * When writing ZIL transactions to the on-disk representation of the 2513 * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple 2514 * itxs can be committed to a single lwb. Once a lwb is written and 2515 * committed to stable storage (i.e. the lwb is written, and vdevs have 2516 * been flushed), each itx that was committed to that lwb is also 2517 * considered to be committed to stable storage. 2518 * 2519 * When an itx is committed to an lwb, the log record (lr_t) contained 2520 * by the itx is copied into the lwb's zio buffer, and once this buffer 2521 * is written to disk, it becomes an on-disk ZIL block. 2522 * 2523 * As itxs are generated, they're inserted into the ZIL's queue of 2524 * uncommitted itxs. The semantics of zil_commit() are such that it will 2525 * block until all itxs that were in the queue when it was called, are 2526 * committed to stable storage. 2527 * 2528 * If "foid" is zero, this means all "synchronous" and "asynchronous" 2529 * itxs, for all objects in the dataset, will be committed to stable 2530 * storage prior to zil_commit() returning. If "foid" is non-zero, all 2531 * "synchronous" itxs for all objects, but only "asynchronous" itxs 2532 * that correspond to the foid passed in, will be committed to stable 2533 * storage prior to zil_commit() returning. 2534 * 2535 * Generally speaking, when zil_commit() is called, the consumer doesn't 2536 * actually care about _all_ of the uncommitted itxs. Instead, they're 2537 * simply trying to waiting for a specific itx to be committed to disk, 2538 * but the interface(s) for interacting with the ZIL don't allow such 2539 * fine-grained communication. A better interface would allow a consumer 2540 * to create and assign an itx, and then pass a reference to this itx to 2541 * zil_commit(); such that zil_commit() would return as soon as that 2542 * specific itx was committed to disk (instead of waiting for _all_ 2543 * itxs to be committed). 2544 * 2545 * When a thread calls zil_commit() a special "commit itx" will be 2546 * generated, along with a corresponding "waiter" for this commit itx. 2547 * zil_commit() will wait on this waiter's CV, such that when the waiter 2548 * is marked done, and signalled, zil_commit() will return. 2549 * 2550 * This commit itx is inserted into the queue of uncommitted itxs. This 2551 * provides an easy mechanism for determining which itxs were in the 2552 * queue prior to zil_commit() having been called, and which itxs were 2553 * added after zil_commit() was called. 2554 * 2555 * The commit it is special; it doesn't have any on-disk representation. 2556 * When a commit itx is "committed" to an lwb, the waiter associated 2557 * with it is linked onto the lwb's list of waiters. Then, when that lwb 2558 * completes, each waiter on the lwb's list is marked done and signalled 2559 * -- allowing the thread waiting on the waiter to return from zil_commit(). 2560 * 2561 * It's important to point out a few critical factors that allow us 2562 * to make use of the commit itxs, commit waiters, per-lwb lists of 2563 * commit waiters, and zio completion callbacks like we're doing: 2564 * 2565 * 1. The list of waiters for each lwb is traversed, and each commit 2566 * waiter is marked "done" and signalled, in the zio completion 2567 * callback of the lwb's zio[*]. 2568 * 2569 * * Actually, the waiters are signalled in the zio completion 2570 * callback of the root zio for the DKIOCFLUSHWRITECACHE commands 2571 * that are sent to the vdevs upon completion of the lwb zio. 2572 * 2573 * 2. When the itxs are inserted into the ZIL's queue of uncommitted 2574 * itxs, the order in which they are inserted is preserved[*]; as 2575 * itxs are added to the queue, they are added to the tail of 2576 * in-memory linked lists. 2577 * 2578 * When committing the itxs to lwbs (to be written to disk), they 2579 * are committed in the same order in which the itxs were added to 2580 * the uncommitted queue's linked list(s); i.e. the linked list of 2581 * itxs to commit is traversed from head to tail, and each itx is 2582 * committed to an lwb in that order. 2583 * 2584 * * To clarify: 2585 * 2586 * - the order of "sync" itxs is preserved w.r.t. other 2587 * "sync" itxs, regardless of the corresponding objects. 2588 * - the order of "async" itxs is preserved w.r.t. other 2589 * "async" itxs corresponding to the same object. 2590 * - the order of "async" itxs is *not* preserved w.r.t. other 2591 * "async" itxs corresponding to different objects. 2592 * - the order of "sync" itxs w.r.t. "async" itxs (or vice 2593 * versa) is *not* preserved, even for itxs that correspond 2594 * to the same object. 2595 * 2596 * For more details, see: zil_itx_assign(), zil_async_to_sync(), 2597 * zil_get_commit_list(), and zil_process_commit_list(). 2598 * 2599 * 3. The lwbs represent a linked list of blocks on disk. Thus, any 2600 * lwb cannot be considered committed to stable storage, until its 2601 * "previous" lwb is also committed to stable storage. This fact, 2602 * coupled with the fact described above, means that itxs are 2603 * committed in (roughly) the order in which they were generated. 2604 * This is essential because itxs are dependent on prior itxs. 2605 * Thus, we *must not* deem an itx as being committed to stable 2606 * storage, until *all* prior itxs have also been committed to 2607 * stable storage. 2608 * 2609 * To enforce this ordering of lwb zio's, while still leveraging as 2610 * much of the underlying storage performance as possible, we rely 2611 * on two fundamental concepts: 2612 * 2613 * 1. The creation and issuance of lwb zio's is protected by 2614 * the zilog's "zl_issuer_lock", which ensures only a single 2615 * thread is creating and/or issuing lwb's at a time 2616 * 2. The "previous" lwb is a child of the "current" lwb 2617 * (leveraging the zio parent-child depenency graph) 2618 * 2619 * By relying on this parent-child zio relationship, we can have 2620 * many lwb zio's concurrently issued to the underlying storage, 2621 * but the order in which they complete will be the same order in 2622 * which they were created. 2623 */ 2624 void 2625 zil_commit(zilog_t *zilog, uint64_t foid) 2626 { 2627 /* 2628 * We should never attempt to call zil_commit on a snapshot for 2629 * a couple of reasons: 2630 * 2631 * 1. A snapshot may never be modified, thus it cannot have any 2632 * in-flight itxs that would have modified the dataset. 2633 * 2634 * 2. By design, when zil_commit() is called, a commit itx will 2635 * be assigned to this zilog; as a result, the zilog will be 2636 * dirtied. We must not dirty the zilog of a snapshot; there's 2637 * checks in the code that enforce this invariant, and will 2638 * cause a panic if it's not upheld. 2639 */ 2640 ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE); 2641 2642 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2643 return; 2644 2645 if (!spa_writeable(zilog->zl_spa)) { 2646 /* 2647 * If the SPA is not writable, there should never be any 2648 * pending itxs waiting to be committed to disk. If that 2649 * weren't true, we'd skip writing those itxs out, and 2650 * would break the sematics of zil_commit(); thus, we're 2651 * verifying that truth before we return to the caller. 2652 */ 2653 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 2654 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 2655 for (int i = 0; i < TXG_SIZE; i++) 2656 ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL); 2657 return; 2658 } 2659 2660 /* 2661 * If the ZIL is suspended, we don't want to dirty it by calling 2662 * zil_commit_itx_assign() below, nor can we write out 2663 * lwbs like would be done in zil_commit_write(). Thus, we 2664 * simply rely on txg_wait_synced() to maintain the necessary 2665 * semantics, and avoid calling those functions altogether. 2666 */ 2667 if (zilog->zl_suspend > 0) { 2668 txg_wait_synced(zilog->zl_dmu_pool, 0); 2669 return; 2670 } 2671 2672 zil_commit_impl(zilog, foid); 2673 } 2674 2675 void 2676 zil_commit_impl(zilog_t *zilog, uint64_t foid) 2677 { 2678 /* 2679 * Move the "async" itxs for the specified foid to the "sync" 2680 * queues, such that they will be later committed (or skipped) 2681 * to an lwb when zil_process_commit_list() is called. 2682 * 2683 * Since these "async" itxs must be committed prior to this 2684 * call to zil_commit returning, we must perform this operation 2685 * before we call zil_commit_itx_assign(). 2686 */ 2687 zil_async_to_sync(zilog, foid); 2688 2689 /* 2690 * We allocate a new "waiter" structure which will initially be 2691 * linked to the commit itx using the itx's "itx_private" field. 2692 * Since the commit itx doesn't represent any on-disk state, 2693 * when it's committed to an lwb, rather than copying the its 2694 * lr_t into the lwb's buffer, the commit itx's "waiter" will be 2695 * added to the lwb's list of waiters. Then, when the lwb is 2696 * committed to stable storage, each waiter in the lwb's list of 2697 * waiters will be marked "done", and signalled. 2698 * 2699 * We must create the waiter and assign the commit itx prior to 2700 * calling zil_commit_writer(), or else our specific commit itx 2701 * is not guaranteed to be committed to an lwb prior to calling 2702 * zil_commit_waiter(). 2703 */ 2704 zil_commit_waiter_t *zcw = zil_alloc_commit_waiter(); 2705 zil_commit_itx_assign(zilog, zcw); 2706 2707 zil_commit_writer(zilog, zcw); 2708 zil_commit_waiter(zilog, zcw); 2709 2710 if (zcw->zcw_zio_error != 0) { 2711 /* 2712 * If there was an error writing out the ZIL blocks that 2713 * this thread is waiting on, then we fallback to 2714 * relying on spa_sync() to write out the data this 2715 * thread is waiting on. Obviously this has performance 2716 * implications, but the expectation is for this to be 2717 * an exceptional case, and shouldn't occur often. 2718 */ 2719 DTRACE_PROBE2(zil__commit__io__error, 2720 zilog_t *, zilog, zil_commit_waiter_t *, zcw); 2721 txg_wait_synced(zilog->zl_dmu_pool, 0); 2722 } 2723 2724 zil_free_commit_waiter(zcw); 2725 } 2726 2727 /* 2728 * Called in syncing context to free committed log blocks and update log header. 2729 */ 2730 void 2731 zil_sync(zilog_t *zilog, dmu_tx_t *tx) 2732 { 2733 zil_header_t *zh = zil_header_in_syncing_context(zilog); 2734 uint64_t txg = dmu_tx_get_txg(tx); 2735 spa_t *spa = zilog->zl_spa; 2736 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 2737 lwb_t *lwb; 2738 2739 /* 2740 * We don't zero out zl_destroy_txg, so make sure we don't try 2741 * to destroy it twice. 2742 */ 2743 if (spa_sync_pass(spa) != 1) 2744 return; 2745 2746 mutex_enter(&zilog->zl_lock); 2747 2748 ASSERT(zilog->zl_stop_sync == 0); 2749 2750 if (*replayed_seq != 0) { 2751 ASSERT(zh->zh_replay_seq < *replayed_seq); 2752 zh->zh_replay_seq = *replayed_seq; 2753 *replayed_seq = 0; 2754 } 2755 2756 if (zilog->zl_destroy_txg == txg) { 2757 blkptr_t blk = zh->zh_log; 2758 2759 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 2760 2761 bzero(zh, sizeof (zil_header_t)); 2762 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 2763 2764 if (zilog->zl_keep_first) { 2765 /* 2766 * If this block was part of log chain that couldn't 2767 * be claimed because a device was missing during 2768 * zil_claim(), but that device later returns, 2769 * then this block could erroneously appear valid. 2770 * To guard against this, assign a new GUID to the new 2771 * log chain so it doesn't matter what blk points to. 2772 */ 2773 zil_init_log_chain(zilog, &blk); 2774 zh->zh_log = blk; 2775 } 2776 } 2777 2778 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 2779 zh->zh_log = lwb->lwb_blk; 2780 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 2781 break; 2782 list_remove(&zilog->zl_lwb_list, lwb); 2783 zio_free(spa, txg, &lwb->lwb_blk); 2784 zil_free_lwb(zilog, lwb); 2785 2786 /* 2787 * If we don't have anything left in the lwb list then 2788 * we've had an allocation failure and we need to zero 2789 * out the zil_header blkptr so that we don't end 2790 * up freeing the same block twice. 2791 */ 2792 if (list_head(&zilog->zl_lwb_list) == NULL) 2793 BP_ZERO(&zh->zh_log); 2794 } 2795 mutex_exit(&zilog->zl_lock); 2796 } 2797 2798 /* ARGSUSED */ 2799 static int 2800 zil_lwb_cons(void *vbuf, void *unused, int kmflag) 2801 { 2802 lwb_t *lwb = vbuf; 2803 list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t), 2804 offsetof(zil_commit_waiter_t, zcw_node)); 2805 avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare, 2806 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 2807 mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 2808 return (0); 2809 } 2810 2811 /* ARGSUSED */ 2812 static void 2813 zil_lwb_dest(void *vbuf, void *unused) 2814 { 2815 lwb_t *lwb = vbuf; 2816 mutex_destroy(&lwb->lwb_vdev_lock); 2817 avl_destroy(&lwb->lwb_vdev_tree); 2818 list_destroy(&lwb->lwb_waiters); 2819 } 2820 2821 void 2822 zil_init(void) 2823 { 2824 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 2825 sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0); 2826 2827 zil_zcw_cache = kmem_cache_create("zil_zcw_cache", 2828 sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 2829 } 2830 2831 void 2832 zil_fini(void) 2833 { 2834 kmem_cache_destroy(zil_zcw_cache); 2835 kmem_cache_destroy(zil_lwb_cache); 2836 } 2837 2838 void 2839 zil_set_sync(zilog_t *zilog, uint64_t sync) 2840 { 2841 zilog->zl_sync = sync; 2842 } 2843 2844 void 2845 zil_set_logbias(zilog_t *zilog, uint64_t logbias) 2846 { 2847 zilog->zl_logbias = logbias; 2848 } 2849 2850 zilog_t * 2851 zil_alloc(objset_t *os, zil_header_t *zh_phys) 2852 { 2853 zilog_t *zilog; 2854 2855 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 2856 2857 zilog->zl_header = zh_phys; 2858 zilog->zl_os = os; 2859 zilog->zl_spa = dmu_objset_spa(os); 2860 zilog->zl_dmu_pool = dmu_objset_pool(os); 2861 zilog->zl_destroy_txg = TXG_INITIAL - 1; 2862 zilog->zl_logbias = dmu_objset_logbias(os); 2863 zilog->zl_sync = dmu_objset_syncprop(os); 2864 zilog->zl_dirty_max_txg = 0; 2865 zilog->zl_last_lwb_opened = NULL; 2866 zilog->zl_last_lwb_latency = 0; 2867 2868 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 2869 mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL); 2870 2871 for (int i = 0; i < TXG_SIZE; i++) { 2872 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 2873 MUTEX_DEFAULT, NULL); 2874 } 2875 2876 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 2877 offsetof(lwb_t, lwb_node)); 2878 2879 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 2880 offsetof(itx_t, itx_node)); 2881 2882 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 2883 2884 return (zilog); 2885 } 2886 2887 void 2888 zil_free(zilog_t *zilog) 2889 { 2890 zilog->zl_stop_sync = 1; 2891 2892 ASSERT0(zilog->zl_suspend); 2893 ASSERT0(zilog->zl_suspending); 2894 2895 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 2896 list_destroy(&zilog->zl_lwb_list); 2897 2898 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 2899 list_destroy(&zilog->zl_itx_commit_list); 2900 2901 for (int i = 0; i < TXG_SIZE; i++) { 2902 /* 2903 * It's possible for an itx to be generated that doesn't dirty 2904 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 2905 * callback to remove the entry. We remove those here. 2906 * 2907 * Also free up the ziltest itxs. 2908 */ 2909 if (zilog->zl_itxg[i].itxg_itxs) 2910 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 2911 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 2912 } 2913 2914 mutex_destroy(&zilog->zl_issuer_lock); 2915 mutex_destroy(&zilog->zl_lock); 2916 2917 cv_destroy(&zilog->zl_cv_suspend); 2918 2919 kmem_free(zilog, sizeof (zilog_t)); 2920 } 2921 2922 /* 2923 * Open an intent log. 2924 */ 2925 zilog_t * 2926 zil_open(objset_t *os, zil_get_data_t *get_data) 2927 { 2928 zilog_t *zilog = dmu_objset_zil(os); 2929 2930 ASSERT3P(zilog->zl_get_data, ==, NULL); 2931 ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL); 2932 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 2933 2934 zilog->zl_get_data = get_data; 2935 2936 return (zilog); 2937 } 2938 2939 /* 2940 * Close an intent log. 2941 */ 2942 void 2943 zil_close(zilog_t *zilog) 2944 { 2945 lwb_t *lwb; 2946 uint64_t txg; 2947 2948 if (!dmu_objset_is_snapshot(zilog->zl_os)) { 2949 zil_commit(zilog, 0); 2950 } else { 2951 ASSERT3P(list_tail(&zilog->zl_lwb_list), ==, NULL); 2952 ASSERT0(zilog->zl_dirty_max_txg); 2953 ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE); 2954 } 2955 2956 mutex_enter(&zilog->zl_lock); 2957 lwb = list_tail(&zilog->zl_lwb_list); 2958 if (lwb == NULL) 2959 txg = zilog->zl_dirty_max_txg; 2960 else 2961 txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg); 2962 mutex_exit(&zilog->zl_lock); 2963 2964 /* 2965 * We need to use txg_wait_synced() to wait long enough for the 2966 * ZIL to be clean, and to wait for all pending lwbs to be 2967 * written out. 2968 */ 2969 if (txg != 0) 2970 txg_wait_synced(zilog->zl_dmu_pool, txg); 2971 2972 if (zilog_is_dirty(zilog)) 2973 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg); 2974 VERIFY(!zilog_is_dirty(zilog)); 2975 2976 zilog->zl_get_data = NULL; 2977 2978 /* 2979 * We should have only one lwb left on the list; remove it now. 2980 */ 2981 mutex_enter(&zilog->zl_lock); 2982 lwb = list_head(&zilog->zl_lwb_list); 2983 if (lwb != NULL) { 2984 ASSERT3P(lwb, ==, list_tail(&zilog->zl_lwb_list)); 2985 ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); 2986 list_remove(&zilog->zl_lwb_list, lwb); 2987 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 2988 zil_free_lwb(zilog, lwb); 2989 } 2990 mutex_exit(&zilog->zl_lock); 2991 } 2992 2993 static char *suspend_tag = "zil suspending"; 2994 2995 /* 2996 * Suspend an intent log. While in suspended mode, we still honor 2997 * synchronous semantics, but we rely on txg_wait_synced() to do it. 2998 * On old version pools, we suspend the log briefly when taking a 2999 * snapshot so that it will have an empty intent log. 3000 * 3001 * Long holds are not really intended to be used the way we do here -- 3002 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 3003 * could fail. Therefore we take pains to only put a long hold if it is 3004 * actually necessary. Fortunately, it will only be necessary if the 3005 * objset is currently mounted (or the ZVOL equivalent). In that case it 3006 * will already have a long hold, so we are not really making things any worse. 3007 * 3008 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 3009 * zvol_state_t), and use their mechanism to prevent their hold from being 3010 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 3011 * very little gain. 3012 * 3013 * if cookiep == NULL, this does both the suspend & resume. 3014 * Otherwise, it returns with the dataset "long held", and the cookie 3015 * should be passed into zil_resume(). 3016 */ 3017 int 3018 zil_suspend(const char *osname, void **cookiep) 3019 { 3020 objset_t *os; 3021 zilog_t *zilog; 3022 const zil_header_t *zh; 3023 int error; 3024 3025 error = dmu_objset_hold(osname, suspend_tag, &os); 3026 if (error != 0) 3027 return (error); 3028 zilog = dmu_objset_zil(os); 3029 3030 mutex_enter(&zilog->zl_lock); 3031 zh = zilog->zl_header; 3032 3033 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 3034 mutex_exit(&zilog->zl_lock); 3035 dmu_objset_rele(os, suspend_tag); 3036 return (SET_ERROR(EBUSY)); 3037 } 3038 3039 /* 3040 * Don't put a long hold in the cases where we can avoid it. This 3041 * is when there is no cookie so we are doing a suspend & resume 3042 * (i.e. called from zil_vdev_offline()), and there's nothing to do 3043 * for the suspend because it's already suspended, or there's no ZIL. 3044 */ 3045 if (cookiep == NULL && !zilog->zl_suspending && 3046 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 3047 mutex_exit(&zilog->zl_lock); 3048 dmu_objset_rele(os, suspend_tag); 3049 return (0); 3050 } 3051 3052 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 3053 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 3054 3055 zilog->zl_suspend++; 3056 3057 if (zilog->zl_suspend > 1) { 3058 /* 3059 * Someone else is already suspending it. 3060 * Just wait for them to finish. 3061 */ 3062 3063 while (zilog->zl_suspending) 3064 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 3065 mutex_exit(&zilog->zl_lock); 3066 3067 if (cookiep == NULL) 3068 zil_resume(os); 3069 else 3070 *cookiep = os; 3071 return (0); 3072 } 3073 3074 /* 3075 * If there is no pointer to an on-disk block, this ZIL must not 3076 * be active (e.g. filesystem not mounted), so there's nothing 3077 * to clean up. 3078 */ 3079 if (BP_IS_HOLE(&zh->zh_log)) { 3080 ASSERT(cookiep != NULL); /* fast path already handled */ 3081 3082 *cookiep = os; 3083 mutex_exit(&zilog->zl_lock); 3084 return (0); 3085 } 3086 3087 zilog->zl_suspending = B_TRUE; 3088 mutex_exit(&zilog->zl_lock); 3089 3090 /* 3091 * We need to use zil_commit_impl to ensure we wait for all 3092 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwb's to be committed 3093 * to disk before proceeding. If we used zil_commit instead, it 3094 * would just call txg_wait_synced(), because zl_suspend is set. 3095 * txg_wait_synced() doesn't wait for these lwb's to be 3096 * LWB_STATE_DONE before returning. 3097 */ 3098 zil_commit_impl(zilog, 0); 3099 3100 /* 3101 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use 3102 * txg_wait_synced() to ensure the data from the zilog has 3103 * migrated to the main pool before calling zil_destroy(). 3104 */ 3105 txg_wait_synced(zilog->zl_dmu_pool, 0); 3106 3107 zil_destroy(zilog, B_FALSE); 3108 3109 mutex_enter(&zilog->zl_lock); 3110 zilog->zl_suspending = B_FALSE; 3111 cv_broadcast(&zilog->zl_cv_suspend); 3112 mutex_exit(&zilog->zl_lock); 3113 3114 if (cookiep == NULL) 3115 zil_resume(os); 3116 else 3117 *cookiep = os; 3118 return (0); 3119 } 3120 3121 void 3122 zil_resume(void *cookie) 3123 { 3124 objset_t *os = cookie; 3125 zilog_t *zilog = dmu_objset_zil(os); 3126 3127 mutex_enter(&zilog->zl_lock); 3128 ASSERT(zilog->zl_suspend != 0); 3129 zilog->zl_suspend--; 3130 mutex_exit(&zilog->zl_lock); 3131 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 3132 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 3133 } 3134 3135 typedef struct zil_replay_arg { 3136 zil_replay_func_t **zr_replay; 3137 void *zr_arg; 3138 boolean_t zr_byteswap; 3139 char *zr_lr; 3140 } zil_replay_arg_t; 3141 3142 static int 3143 zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 3144 { 3145 char name[ZFS_MAX_DATASET_NAME_LEN]; 3146 3147 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 3148 3149 dmu_objset_name(zilog->zl_os, name); 3150 3151 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 3152 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 3153 (u_longlong_t)lr->lrc_seq, 3154 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 3155 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 3156 3157 return (error); 3158 } 3159 3160 static int 3161 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 3162 { 3163 zil_replay_arg_t *zr = zra; 3164 const zil_header_t *zh = zilog->zl_header; 3165 uint64_t reclen = lr->lrc_reclen; 3166 uint64_t txtype = lr->lrc_txtype; 3167 int error = 0; 3168 3169 zilog->zl_replaying_seq = lr->lrc_seq; 3170 3171 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 3172 return (0); 3173 3174 if (lr->lrc_txg < claim_txg) /* already committed */ 3175 return (0); 3176 3177 /* Strip case-insensitive bit, still present in log record */ 3178 txtype &= ~TX_CI; 3179 3180 if (txtype == 0 || txtype >= TX_MAX_TYPE) 3181 return (zil_replay_error(zilog, lr, EINVAL)); 3182 3183 /* 3184 * If this record type can be logged out of order, the object 3185 * (lr_foid) may no longer exist. That's legitimate, not an error. 3186 */ 3187 if (TX_OOO(txtype)) { 3188 error = dmu_object_info(zilog->zl_os, 3189 ((lr_ooo_t *)lr)->lr_foid, NULL); 3190 if (error == ENOENT || error == EEXIST) 3191 return (0); 3192 } 3193 3194 /* 3195 * Make a copy of the data so we can revise and extend it. 3196 */ 3197 bcopy(lr, zr->zr_lr, reclen); 3198 3199 /* 3200 * If this is a TX_WRITE with a blkptr, suck in the data. 3201 */ 3202 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 3203 error = zil_read_log_data(zilog, (lr_write_t *)lr, 3204 zr->zr_lr + reclen); 3205 if (error != 0) 3206 return (zil_replay_error(zilog, lr, error)); 3207 } 3208 3209 /* 3210 * The log block containing this lr may have been byteswapped 3211 * so that we can easily examine common fields like lrc_txtype. 3212 * However, the log is a mix of different record types, and only the 3213 * replay vectors know how to byteswap their records. Therefore, if 3214 * the lr was byteswapped, undo it before invoking the replay vector. 3215 */ 3216 if (zr->zr_byteswap) 3217 byteswap_uint64_array(zr->zr_lr, reclen); 3218 3219 /* 3220 * We must now do two things atomically: replay this log record, 3221 * and update the log header sequence number to reflect the fact that 3222 * we did so. At the end of each replay function the sequence number 3223 * is updated if we are in replay mode. 3224 */ 3225 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 3226 if (error != 0) { 3227 /* 3228 * The DMU's dnode layer doesn't see removes until the txg 3229 * commits, so a subsequent claim can spuriously fail with 3230 * EEXIST. So if we receive any error we try syncing out 3231 * any removes then retry the transaction. Note that we 3232 * specify B_FALSE for byteswap now, so we don't do it twice. 3233 */ 3234 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 3235 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 3236 if (error != 0) 3237 return (zil_replay_error(zilog, lr, error)); 3238 } 3239 return (0); 3240 } 3241 3242 /* ARGSUSED */ 3243 static int 3244 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 3245 { 3246 zilog->zl_replay_blks++; 3247 3248 return (0); 3249 } 3250 3251 /* 3252 * If this dataset has a non-empty intent log, replay it and destroy it. 3253 */ 3254 void 3255 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 3256 { 3257 zilog_t *zilog = dmu_objset_zil(os); 3258 const zil_header_t *zh = zilog->zl_header; 3259 zil_replay_arg_t zr; 3260 3261 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 3262 zil_destroy(zilog, B_TRUE); 3263 return; 3264 } 3265 3266 zr.zr_replay = replay_func; 3267 zr.zr_arg = arg; 3268 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 3269 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 3270 3271 /* 3272 * Wait for in-progress removes to sync before starting replay. 3273 */ 3274 txg_wait_synced(zilog->zl_dmu_pool, 0); 3275 3276 zilog->zl_replay = B_TRUE; 3277 zilog->zl_replay_time = ddi_get_lbolt(); 3278 ASSERT(zilog->zl_replay_blks == 0); 3279 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 3280 zh->zh_claim_txg); 3281 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 3282 3283 zil_destroy(zilog, B_FALSE); 3284 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 3285 zilog->zl_replay = B_FALSE; 3286 } 3287 3288 boolean_t 3289 zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 3290 { 3291 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 3292 return (B_TRUE); 3293 3294 if (zilog->zl_replay) { 3295 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 3296 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 3297 zilog->zl_replaying_seq; 3298 return (B_TRUE); 3299 } 3300 3301 return (B_FALSE); 3302 } 3303 3304 /* ARGSUSED */ 3305 int 3306 zil_reset(const char *osname, void *arg) 3307 { 3308 int error; 3309 3310 error = zil_suspend(osname, NULL); 3311 if (error != 0) 3312 return (SET_ERROR(EEXIST)); 3313 return (0); 3314 } 3315