zil.c revision 308595
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, 2015 by Delphix. All rights reserved. 24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25 * Copyright (c) 2014 Integros [integros.com] 26 */ 27 28/* Portions Copyright 2010 Robert Milkowski */ 29 30#include <sys/zfs_context.h> 31#include <sys/spa.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 44/* 45 * The zfs intent log (ZIL) saves transaction records of system calls 46 * that change the file system in memory with enough information 47 * to be able to replay them. These are stored in memory until 48 * either the DMU transaction group (txg) commits them to the stable pool 49 * and they can be discarded, or they are flushed to the stable log 50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 51 * requirement. In the event of a panic or power fail then those log 52 * records (transactions) are replayed. 53 * 54 * There is one ZIL per file system. Its on-disk (pool) format consists 55 * of 3 parts: 56 * 57 * - ZIL header 58 * - ZIL blocks 59 * - ZIL records 60 * 61 * A log record holds a system call transaction. Log blocks can 62 * hold many log records and the blocks are chained together. 63 * Each ZIL block contains a block pointer (blkptr_t) to the next 64 * ZIL block in the chain. The ZIL header points to the first 65 * block in the chain. Note there is not a fixed place in the pool 66 * to hold blocks. They are dynamically allocated and freed as 67 * needed from the blocks available. Figure X shows the ZIL structure: 68 */ 69 70/* 71 * Disable intent logging replay. This global ZIL switch affects all pools. 72 */ 73int zil_replay_disable = 0; 74SYSCTL_DECL(_vfs_zfs); 75SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN, 76 &zil_replay_disable, 0, "Disable intent logging replay"); 77 78/* 79 * Tunable parameter for debugging or performance analysis. Setting 80 * zfs_nocacheflush will cause corruption on power loss if a volatile 81 * out-of-order write cache is enabled. 82 */ 83boolean_t zfs_nocacheflush = B_FALSE; 84SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 85 &zfs_nocacheflush, 0, "Disable cache flush"); 86boolean_t zfs_trim_enabled = B_TRUE; 87SYSCTL_DECL(_vfs_zfs_trim); 88SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 89 "Enable ZFS TRIM"); 90 91static kmem_cache_t *zil_lwb_cache; 92 93#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 94 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 95 96 97/* 98 * ziltest is by and large an ugly hack, but very useful in 99 * checking replay without tedious work. 100 * When running ziltest we want to keep all itx's and so maintain 101 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 102 * We subtract TXG_CONCURRENT_STATES to allow for common code. 103 */ 104#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 105 106static int 107zil_bp_compare(const void *x1, const void *x2) 108{ 109 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 110 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 111 112 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 113 return (-1); 114 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 115 return (1); 116 117 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 118 return (-1); 119 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 120 return (1); 121 122 return (0); 123} 124 125static void 126zil_bp_tree_init(zilog_t *zilog) 127{ 128 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 129 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 130} 131 132static void 133zil_bp_tree_fini(zilog_t *zilog) 134{ 135 avl_tree_t *t = &zilog->zl_bp_tree; 136 zil_bp_node_t *zn; 137 void *cookie = NULL; 138 139 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 140 kmem_free(zn, sizeof (zil_bp_node_t)); 141 142 avl_destroy(t); 143} 144 145int 146zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 147{ 148 avl_tree_t *t = &zilog->zl_bp_tree; 149 const dva_t *dva; 150 zil_bp_node_t *zn; 151 avl_index_t where; 152 153 if (BP_IS_EMBEDDED(bp)) 154 return (0); 155 156 dva = BP_IDENTITY(bp); 157 158 if (avl_find(t, dva, &where) != NULL) 159 return (SET_ERROR(EEXIST)); 160 161 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 162 zn->zn_dva = *dva; 163 avl_insert(t, zn, where); 164 165 return (0); 166} 167 168static zil_header_t * 169zil_header_in_syncing_context(zilog_t *zilog) 170{ 171 return ((zil_header_t *)zilog->zl_header); 172} 173 174static void 175zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 176{ 177 zio_cksum_t *zc = &bp->blk_cksum; 178 179 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 180 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 181 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 182 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 183} 184 185/* 186 * Read a log block and make sure it's valid. 187 */ 188static int 189zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 190 char **end) 191{ 192 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 193 arc_flags_t aflags = ARC_FLAG_WAIT; 194 arc_buf_t *abuf = NULL; 195 zbookmark_phys_t zb; 196 int error; 197 198 if (zilog->zl_header->zh_claim_txg == 0) 199 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 200 201 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 202 zio_flags |= ZIO_FLAG_SPECULATIVE; 203 204 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 205 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 206 207 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 208 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 209 210 if (error == 0) { 211 zio_cksum_t cksum = bp->blk_cksum; 212 213 /* 214 * Validate the checksummed log block. 215 * 216 * Sequence numbers should be... sequential. The checksum 217 * verifier for the next block should be bp's checksum plus 1. 218 * 219 * Also check the log chain linkage and size used. 220 */ 221 cksum.zc_word[ZIL_ZC_SEQ]++; 222 223 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 224 zil_chain_t *zilc = abuf->b_data; 225 char *lr = (char *)(zilc + 1); 226 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 227 228 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 229 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 230 error = SET_ERROR(ECKSUM); 231 } else { 232 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 233 bcopy(lr, dst, len); 234 *end = (char *)dst + len; 235 *nbp = zilc->zc_next_blk; 236 } 237 } else { 238 char *lr = abuf->b_data; 239 uint64_t size = BP_GET_LSIZE(bp); 240 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 241 242 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 243 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 244 (zilc->zc_nused > (size - sizeof (*zilc)))) { 245 error = SET_ERROR(ECKSUM); 246 } else { 247 ASSERT3U(zilc->zc_nused, <=, 248 SPA_OLD_MAXBLOCKSIZE); 249 bcopy(lr, dst, zilc->zc_nused); 250 *end = (char *)dst + zilc->zc_nused; 251 *nbp = zilc->zc_next_blk; 252 } 253 } 254 255 arc_buf_destroy(abuf, &abuf); 256 } 257 258 return (error); 259} 260 261/* 262 * Read a TX_WRITE log data block. 263 */ 264static int 265zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 266{ 267 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 268 const blkptr_t *bp = &lr->lr_blkptr; 269 arc_flags_t aflags = ARC_FLAG_WAIT; 270 arc_buf_t *abuf = NULL; 271 zbookmark_phys_t zb; 272 int error; 273 274 if (BP_IS_HOLE(bp)) { 275 if (wbuf != NULL) 276 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 277 return (0); 278 } 279 280 if (zilog->zl_header->zh_claim_txg == 0) 281 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 282 283 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 284 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 285 286 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 287 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 288 289 if (error == 0) { 290 if (wbuf != NULL) 291 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 292 arc_buf_destroy(abuf, &abuf); 293 } 294 295 return (error); 296} 297 298/* 299 * Parse the intent log, and call parse_func for each valid record within. 300 */ 301int 302zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 303 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 304{ 305 const zil_header_t *zh = zilog->zl_header; 306 boolean_t claimed = !!zh->zh_claim_txg; 307 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 308 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 309 uint64_t max_blk_seq = 0; 310 uint64_t max_lr_seq = 0; 311 uint64_t blk_count = 0; 312 uint64_t lr_count = 0; 313 blkptr_t blk, next_blk; 314 char *lrbuf, *lrp; 315 int error = 0; 316 317 /* 318 * Old logs didn't record the maximum zh_claim_lr_seq. 319 */ 320 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 321 claim_lr_seq = UINT64_MAX; 322 323 /* 324 * Starting at the block pointed to by zh_log we read the log chain. 325 * For each block in the chain we strongly check that block to 326 * ensure its validity. We stop when an invalid block is found. 327 * For each block pointer in the chain we call parse_blk_func(). 328 * For each record in each valid block we call parse_lr_func(). 329 * If the log has been claimed, stop if we encounter a sequence 330 * number greater than the highest claimed sequence number. 331 */ 332 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 333 zil_bp_tree_init(zilog); 334 335 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 336 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 337 int reclen; 338 char *end; 339 340 if (blk_seq > claim_blk_seq) 341 break; 342 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 343 break; 344 ASSERT3U(max_blk_seq, <, blk_seq); 345 max_blk_seq = blk_seq; 346 blk_count++; 347 348 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 349 break; 350 351 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 352 if (error != 0) 353 break; 354 355 for (lrp = lrbuf; lrp < end; lrp += reclen) { 356 lr_t *lr = (lr_t *)lrp; 357 reclen = lr->lrc_reclen; 358 ASSERT3U(reclen, >=, sizeof (lr_t)); 359 if (lr->lrc_seq > claim_lr_seq) 360 goto done; 361 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 362 goto done; 363 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 364 max_lr_seq = lr->lrc_seq; 365 lr_count++; 366 } 367 } 368done: 369 zilog->zl_parse_error = error; 370 zilog->zl_parse_blk_seq = max_blk_seq; 371 zilog->zl_parse_lr_seq = max_lr_seq; 372 zilog->zl_parse_blk_count = blk_count; 373 zilog->zl_parse_lr_count = lr_count; 374 375 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 376 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 377 378 zil_bp_tree_fini(zilog); 379 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 380 381 return (error); 382} 383 384static int 385zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 386{ 387 /* 388 * Claim log block if not already committed and not already claimed. 389 * If tx == NULL, just verify that the block is claimable. 390 */ 391 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 392 zil_bp_tree_add(zilog, bp) != 0) 393 return (0); 394 395 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 396 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 397 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 398} 399 400static int 401zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 402{ 403 lr_write_t *lr = (lr_write_t *)lrc; 404 int error; 405 406 if (lrc->lrc_txtype != TX_WRITE) 407 return (0); 408 409 /* 410 * If the block is not readable, don't claim it. This can happen 411 * in normal operation when a log block is written to disk before 412 * some of the dmu_sync() blocks it points to. In this case, the 413 * transaction cannot have been committed to anyone (we would have 414 * waited for all writes to be stable first), so it is semantically 415 * correct to declare this the end of the log. 416 */ 417 if (lr->lr_blkptr.blk_birth >= first_txg && 418 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 419 return (error); 420 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 421} 422 423/* ARGSUSED */ 424static int 425zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 426{ 427 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 428 429 return (0); 430} 431 432static int 433zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 434{ 435 lr_write_t *lr = (lr_write_t *)lrc; 436 blkptr_t *bp = &lr->lr_blkptr; 437 438 /* 439 * If we previously claimed it, we need to free it. 440 */ 441 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 442 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 443 !BP_IS_HOLE(bp)) 444 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 445 446 return (0); 447} 448 449static lwb_t * 450zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 451{ 452 lwb_t *lwb; 453 454 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 455 lwb->lwb_zilog = zilog; 456 lwb->lwb_blk = *bp; 457 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 458 lwb->lwb_max_txg = txg; 459 lwb->lwb_zio = NULL; 460 lwb->lwb_tx = NULL; 461 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 462 lwb->lwb_nused = sizeof (zil_chain_t); 463 lwb->lwb_sz = BP_GET_LSIZE(bp); 464 } else { 465 lwb->lwb_nused = 0; 466 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 467 } 468 469 mutex_enter(&zilog->zl_lock); 470 list_insert_tail(&zilog->zl_lwb_list, lwb); 471 mutex_exit(&zilog->zl_lock); 472 473 return (lwb); 474} 475 476/* 477 * Called when we create in-memory log transactions so that we know 478 * to cleanup the itxs at the end of spa_sync(). 479 */ 480void 481zilog_dirty(zilog_t *zilog, uint64_t txg) 482{ 483 dsl_pool_t *dp = zilog->zl_dmu_pool; 484 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 485 486 if (ds->ds_is_snapshot) 487 panic("dirtying snapshot!"); 488 489 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 490 /* up the hold count until we can be written out */ 491 dmu_buf_add_ref(ds->ds_dbuf, zilog); 492 } 493} 494 495boolean_t 496zilog_is_dirty(zilog_t *zilog) 497{ 498 dsl_pool_t *dp = zilog->zl_dmu_pool; 499 500 for (int t = 0; t < TXG_SIZE; t++) { 501 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 502 return (B_TRUE); 503 } 504 return (B_FALSE); 505} 506 507/* 508 * Create an on-disk intent log. 509 */ 510static lwb_t * 511zil_create(zilog_t *zilog) 512{ 513 const zil_header_t *zh = zilog->zl_header; 514 lwb_t *lwb = NULL; 515 uint64_t txg = 0; 516 dmu_tx_t *tx = NULL; 517 blkptr_t blk; 518 int error = 0; 519 520 /* 521 * Wait for any previous destroy to complete. 522 */ 523 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 524 525 ASSERT(zh->zh_claim_txg == 0); 526 ASSERT(zh->zh_replay_seq == 0); 527 528 blk = zh->zh_log; 529 530 /* 531 * Allocate an initial log block if: 532 * - there isn't one already 533 * - the existing block is the wrong endianess 534 */ 535 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 536 tx = dmu_tx_create(zilog->zl_os); 537 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 538 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 539 txg = dmu_tx_get_txg(tx); 540 541 if (!BP_IS_HOLE(&blk)) { 542 zio_free_zil(zilog->zl_spa, txg, &blk); 543 BP_ZERO(&blk); 544 } 545 546 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 547 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 548 549 if (error == 0) 550 zil_init_log_chain(zilog, &blk); 551 } 552 553 /* 554 * Allocate a log write buffer (lwb) for the first log block. 555 */ 556 if (error == 0) 557 lwb = zil_alloc_lwb(zilog, &blk, txg); 558 559 /* 560 * If we just allocated the first log block, commit our transaction 561 * and wait for zil_sync() to stuff the block poiner into zh_log. 562 * (zh is part of the MOS, so we cannot modify it in open context.) 563 */ 564 if (tx != NULL) { 565 dmu_tx_commit(tx); 566 txg_wait_synced(zilog->zl_dmu_pool, txg); 567 } 568 569 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 570 571 return (lwb); 572} 573 574/* 575 * In one tx, free all log blocks and clear the log header. 576 * If keep_first is set, then we're replaying a log with no content. 577 * We want to keep the first block, however, so that the first 578 * synchronous transaction doesn't require a txg_wait_synced() 579 * in zil_create(). We don't need to txg_wait_synced() here either 580 * when keep_first is set, because both zil_create() and zil_destroy() 581 * will wait for any in-progress destroys to complete. 582 */ 583void 584zil_destroy(zilog_t *zilog, boolean_t keep_first) 585{ 586 const zil_header_t *zh = zilog->zl_header; 587 lwb_t *lwb; 588 dmu_tx_t *tx; 589 uint64_t txg; 590 591 /* 592 * Wait for any previous destroy to complete. 593 */ 594 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 595 596 zilog->zl_old_header = *zh; /* debugging aid */ 597 598 if (BP_IS_HOLE(&zh->zh_log)) 599 return; 600 601 tx = dmu_tx_create(zilog->zl_os); 602 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 603 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 604 txg = dmu_tx_get_txg(tx); 605 606 mutex_enter(&zilog->zl_lock); 607 608 ASSERT3U(zilog->zl_destroy_txg, <, txg); 609 zilog->zl_destroy_txg = txg; 610 zilog->zl_keep_first = keep_first; 611 612 if (!list_is_empty(&zilog->zl_lwb_list)) { 613 ASSERT(zh->zh_claim_txg == 0); 614 VERIFY(!keep_first); 615 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 616 list_remove(&zilog->zl_lwb_list, lwb); 617 if (lwb->lwb_buf != NULL) 618 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 619 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 620 kmem_cache_free(zil_lwb_cache, lwb); 621 } 622 } else if (!keep_first) { 623 zil_destroy_sync(zilog, tx); 624 } 625 mutex_exit(&zilog->zl_lock); 626 627 dmu_tx_commit(tx); 628} 629 630void 631zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 632{ 633 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 634 (void) zil_parse(zilog, zil_free_log_block, 635 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 636} 637 638int 639zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 640{ 641 dmu_tx_t *tx = txarg; 642 uint64_t first_txg = dmu_tx_get_txg(tx); 643 zilog_t *zilog; 644 zil_header_t *zh; 645 objset_t *os; 646 int error; 647 648 error = dmu_objset_own_obj(dp, ds->ds_object, 649 DMU_OST_ANY, B_FALSE, FTAG, &os); 650 if (error != 0) { 651 /* 652 * EBUSY indicates that the objset is inconsistent, in which 653 * case it can not have a ZIL. 654 */ 655 if (error != EBUSY) { 656 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 657 (unsigned long long)ds->ds_object, error); 658 } 659 return (0); 660 } 661 662 zilog = dmu_objset_zil(os); 663 zh = zil_header_in_syncing_context(zilog); 664 665 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 666 if (!BP_IS_HOLE(&zh->zh_log)) 667 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 668 BP_ZERO(&zh->zh_log); 669 dsl_dataset_dirty(dmu_objset_ds(os), tx); 670 dmu_objset_disown(os, FTAG); 671 return (0); 672 } 673 674 /* 675 * Claim all log blocks if we haven't already done so, and remember 676 * the highest claimed sequence number. This ensures that if we can 677 * read only part of the log now (e.g. due to a missing device), 678 * but we can read the entire log later, we will not try to replay 679 * or destroy beyond the last block we successfully claimed. 680 */ 681 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 682 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 683 (void) zil_parse(zilog, zil_claim_log_block, 684 zil_claim_log_record, tx, first_txg); 685 zh->zh_claim_txg = first_txg; 686 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 687 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 688 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 689 zh->zh_flags |= ZIL_REPLAY_NEEDED; 690 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 691 dsl_dataset_dirty(dmu_objset_ds(os), tx); 692 } 693 694 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 695 dmu_objset_disown(os, FTAG); 696 return (0); 697} 698 699/* 700 * Check the log by walking the log chain. 701 * Checksum errors are ok as they indicate the end of the chain. 702 * Any other error (no device or read failure) returns an error. 703 */ 704/* ARGSUSED */ 705int 706zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 707{ 708 zilog_t *zilog; 709 objset_t *os; 710 blkptr_t *bp; 711 int error; 712 713 ASSERT(tx == NULL); 714 715 error = dmu_objset_from_ds(ds, &os); 716 if (error != 0) { 717 cmn_err(CE_WARN, "can't open objset %llu, error %d", 718 (unsigned long long)ds->ds_object, error); 719 return (0); 720 } 721 722 zilog = dmu_objset_zil(os); 723 bp = (blkptr_t *)&zilog->zl_header->zh_log; 724 725 /* 726 * Check the first block and determine if it's on a log device 727 * which may have been removed or faulted prior to loading this 728 * pool. If so, there's no point in checking the rest of the log 729 * as its content should have already been synced to the pool. 730 */ 731 if (!BP_IS_HOLE(bp)) { 732 vdev_t *vd; 733 boolean_t valid = B_TRUE; 734 735 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 736 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 737 if (vd->vdev_islog && vdev_is_dead(vd)) 738 valid = vdev_log_state_valid(vd); 739 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 740 741 if (!valid) 742 return (0); 743 } 744 745 /* 746 * Because tx == NULL, zil_claim_log_block() will not actually claim 747 * any blocks, but just determine whether it is possible to do so. 748 * In addition to checking the log chain, zil_claim_log_block() 749 * will invoke zio_claim() with a done func of spa_claim_notify(), 750 * which will update spa_max_claim_txg. See spa_load() for details. 751 */ 752 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 753 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 754 755 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 756} 757 758static int 759zil_vdev_compare(const void *x1, const void *x2) 760{ 761 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 762 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 763 764 if (v1 < v2) 765 return (-1); 766 if (v1 > v2) 767 return (1); 768 769 return (0); 770} 771 772void 773zil_add_block(zilog_t *zilog, const blkptr_t *bp) 774{ 775 avl_tree_t *t = &zilog->zl_vdev_tree; 776 avl_index_t where; 777 zil_vdev_node_t *zv, zvsearch; 778 int ndvas = BP_GET_NDVAS(bp); 779 int i; 780 781 if (zfs_nocacheflush) 782 return; 783 784 ASSERT(zilog->zl_writer); 785 786 /* 787 * Even though we're zl_writer, we still need a lock because the 788 * zl_get_data() callbacks may have dmu_sync() done callbacks 789 * that will run concurrently. 790 */ 791 mutex_enter(&zilog->zl_vdev_lock); 792 for (i = 0; i < ndvas; i++) { 793 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 794 if (avl_find(t, &zvsearch, &where) == NULL) { 795 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 796 zv->zv_vdev = zvsearch.zv_vdev; 797 avl_insert(t, zv, where); 798 } 799 } 800 mutex_exit(&zilog->zl_vdev_lock); 801} 802 803static void 804zil_flush_vdevs(zilog_t *zilog) 805{ 806 spa_t *spa = zilog->zl_spa; 807 avl_tree_t *t = &zilog->zl_vdev_tree; 808 void *cookie = NULL; 809 zil_vdev_node_t *zv; 810 zio_t *zio; 811 812 ASSERT(zilog->zl_writer); 813 814 /* 815 * We don't need zl_vdev_lock here because we're the zl_writer, 816 * and all zl_get_data() callbacks are done. 817 */ 818 if (avl_numnodes(t) == 0) 819 return; 820 821 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 822 823 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 824 825 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 826 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 827 if (vd != NULL) 828 zio_flush(zio, vd); 829 kmem_free(zv, sizeof (*zv)); 830 } 831 832 /* 833 * Wait for all the flushes to complete. Not all devices actually 834 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 835 */ 836 (void) zio_wait(zio); 837 838 spa_config_exit(spa, SCL_STATE, FTAG); 839} 840 841/* 842 * Function called when a log block write completes 843 */ 844static void 845zil_lwb_write_done(zio_t *zio) 846{ 847 lwb_t *lwb = zio->io_private; 848 zilog_t *zilog = lwb->lwb_zilog; 849 dmu_tx_t *tx = lwb->lwb_tx; 850 851 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 852 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 853 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 854 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 855 ASSERT(!BP_IS_GANG(zio->io_bp)); 856 ASSERT(!BP_IS_HOLE(zio->io_bp)); 857 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 858 859 /* 860 * Ensure the lwb buffer pointer is cleared before releasing 861 * the txg. If we have had an allocation failure and 862 * the txg is waiting to sync then we want want zil_sync() 863 * to remove the lwb so that it's not picked up as the next new 864 * one in zil_commit_writer(). zil_sync() will only remove 865 * the lwb if lwb_buf is null. 866 */ 867 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 868 mutex_enter(&zilog->zl_lock); 869 lwb->lwb_buf = NULL; 870 lwb->lwb_tx = NULL; 871 mutex_exit(&zilog->zl_lock); 872 873 /* 874 * Now that we've written this log block, we have a stable pointer 875 * to the next block in the chain, so it's OK to let the txg in 876 * which we allocated the next block sync. 877 */ 878 dmu_tx_commit(tx); 879} 880 881/* 882 * Initialize the io for a log block. 883 */ 884static void 885zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 886{ 887 zbookmark_phys_t zb; 888 889 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 890 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 891 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 892 893 if (zilog->zl_root_zio == NULL) { 894 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 895 ZIO_FLAG_CANFAIL); 896 } 897 if (lwb->lwb_zio == NULL) { 898 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 899 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 900 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 901 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 902 } 903} 904 905/* 906 * Define a limited set of intent log block sizes. 907 * 908 * These must be a multiple of 4KB. Note only the amount used (again 909 * aligned to 4KB) actually gets written. However, we can't always just 910 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 911 */ 912uint64_t zil_block_buckets[] = { 913 4096, /* non TX_WRITE */ 914 8192+4096, /* data base */ 915 32*1024 + 4096, /* NFS writes */ 916 UINT64_MAX 917}; 918 919/* 920 * Use the slog as long as the logbias is 'latency' and the current commit size 921 * is less than the limit or the total list size is less than 2X the limit. 922 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 923 */ 924uint64_t zil_slog_limit = 1024 * 1024; 925#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 926 (((zilog)->zl_cur_used < zil_slog_limit) || \ 927 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 928 929/* 930 * Start a log block write and advance to the next log block. 931 * Calls are serialized. 932 */ 933static lwb_t * 934zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 935{ 936 lwb_t *nlwb = NULL; 937 zil_chain_t *zilc; 938 spa_t *spa = zilog->zl_spa; 939 blkptr_t *bp; 940 dmu_tx_t *tx; 941 uint64_t txg; 942 uint64_t zil_blksz, wsz; 943 int i, error; 944 945 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 946 zilc = (zil_chain_t *)lwb->lwb_buf; 947 bp = &zilc->zc_next_blk; 948 } else { 949 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 950 bp = &zilc->zc_next_blk; 951 } 952 953 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 954 955 /* 956 * Allocate the next block and save its address in this block 957 * before writing it in order to establish the log chain. 958 * Note that if the allocation of nlwb synced before we wrote 959 * the block that points at it (lwb), we'd leak it if we crashed. 960 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 961 * We dirty the dataset to ensure that zil_sync() will be called 962 * to clean up in the event of allocation failure or I/O failure. 963 */ 964 tx = dmu_tx_create(zilog->zl_os); 965 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 966 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 967 txg = dmu_tx_get_txg(tx); 968 969 lwb->lwb_tx = tx; 970 971 /* 972 * Log blocks are pre-allocated. Here we select the size of the next 973 * block, based on size used in the last block. 974 * - first find the smallest bucket that will fit the block from a 975 * limited set of block sizes. This is because it's faster to write 976 * blocks allocated from the same metaslab as they are adjacent or 977 * close. 978 * - next find the maximum from the new suggested size and an array of 979 * previous sizes. This lessens a picket fence effect of wrongly 980 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 981 * requests. 982 * 983 * Note we only write what is used, but we can't just allocate 984 * the maximum block size because we can exhaust the available 985 * pool log space. 986 */ 987 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 988 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 989 continue; 990 zil_blksz = zil_block_buckets[i]; 991 if (zil_blksz == UINT64_MAX) 992 zil_blksz = SPA_OLD_MAXBLOCKSIZE; 993 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 994 for (i = 0; i < ZIL_PREV_BLKS; i++) 995 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 996 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 997 998 BP_ZERO(bp); 999 /* pass the old blkptr in order to spread log blocks across devs */ 1000 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 1001 USE_SLOG(zilog)); 1002 if (error == 0) { 1003 ASSERT3U(bp->blk_birth, ==, txg); 1004 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1005 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1006 1007 /* 1008 * Allocate a new log write buffer (lwb). 1009 */ 1010 nlwb = zil_alloc_lwb(zilog, bp, txg); 1011 1012 /* Record the block for later vdev flushing */ 1013 zil_add_block(zilog, &lwb->lwb_blk); 1014 } 1015 1016 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1017 /* For Slim ZIL only write what is used. */ 1018 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1019 ASSERT3U(wsz, <=, lwb->lwb_sz); 1020 zio_shrink(lwb->lwb_zio, wsz); 1021 1022 } else { 1023 wsz = lwb->lwb_sz; 1024 } 1025 1026 zilc->zc_pad = 0; 1027 zilc->zc_nused = lwb->lwb_nused; 1028 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1029 1030 /* 1031 * clear unused data for security 1032 */ 1033 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1034 1035 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1036 1037 /* 1038 * If there was an allocation failure then nlwb will be null which 1039 * forces a txg_wait_synced(). 1040 */ 1041 return (nlwb); 1042} 1043 1044static lwb_t * 1045zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1046{ 1047 lr_t *lrc = &itx->itx_lr; /* common log record */ 1048 lr_write_t *lrw = (lr_write_t *)lrc; 1049 char *lr_buf; 1050 uint64_t txg = lrc->lrc_txg; 1051 uint64_t reclen = lrc->lrc_reclen; 1052 uint64_t dlen = 0; 1053 1054 if (lwb == NULL) 1055 return (NULL); 1056 1057 ASSERT(lwb->lwb_buf != NULL); 1058 ASSERT(zilog_is_dirty(zilog) || 1059 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1060 1061 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1062 dlen = P2ROUNDUP_TYPED( 1063 lrw->lr_length, sizeof (uint64_t), uint64_t); 1064 1065 zilog->zl_cur_used += (reclen + dlen); 1066 1067 zil_lwb_write_init(zilog, lwb); 1068 1069 /* 1070 * If this record won't fit in the current log block, start a new one. 1071 */ 1072 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1073 lwb = zil_lwb_write_start(zilog, lwb); 1074 if (lwb == NULL) 1075 return (NULL); 1076 zil_lwb_write_init(zilog, lwb); 1077 ASSERT(LWB_EMPTY(lwb)); 1078 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1079 txg_wait_synced(zilog->zl_dmu_pool, txg); 1080 return (lwb); 1081 } 1082 } 1083 1084 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1085 bcopy(lrc, lr_buf, reclen); 1086 lrc = (lr_t *)lr_buf; 1087 lrw = (lr_write_t *)lrc; 1088 1089 /* 1090 * If it's a write, fetch the data or get its blkptr as appropriate. 1091 */ 1092 if (lrc->lrc_txtype == TX_WRITE) { 1093 if (txg > spa_freeze_txg(zilog->zl_spa)) 1094 txg_wait_synced(zilog->zl_dmu_pool, txg); 1095 if (itx->itx_wr_state != WR_COPIED) { 1096 char *dbuf; 1097 int error; 1098 1099 if (dlen) { 1100 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1101 dbuf = lr_buf + reclen; 1102 lrw->lr_common.lrc_reclen += dlen; 1103 } else { 1104 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1105 dbuf = NULL; 1106 } 1107 error = zilog->zl_get_data( 1108 itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1109 if (error == EIO) { 1110 txg_wait_synced(zilog->zl_dmu_pool, txg); 1111 return (lwb); 1112 } 1113 if (error != 0) { 1114 ASSERT(error == ENOENT || error == EEXIST || 1115 error == EALREADY); 1116 return (lwb); 1117 } 1118 } 1119 } 1120 1121 /* 1122 * We're actually making an entry, so update lrc_seq to be the 1123 * log record sequence number. Note that this is generally not 1124 * equal to the itx sequence number because not all transactions 1125 * are synchronous, and sometimes spa_sync() gets there first. 1126 */ 1127 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1128 lwb->lwb_nused += reclen + dlen; 1129 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1130 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1131 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1132 1133 return (lwb); 1134} 1135 1136itx_t * 1137zil_itx_create(uint64_t txtype, size_t lrsize) 1138{ 1139 itx_t *itx; 1140 1141 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1142 1143 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1144 itx->itx_lr.lrc_txtype = txtype; 1145 itx->itx_lr.lrc_reclen = lrsize; 1146 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1147 itx->itx_lr.lrc_seq = 0; /* defensive */ 1148 itx->itx_sync = B_TRUE; /* default is synchronous */ 1149 1150 return (itx); 1151} 1152 1153void 1154zil_itx_destroy(itx_t *itx) 1155{ 1156 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1157} 1158 1159/* 1160 * Free up the sync and async itxs. The itxs_t has already been detached 1161 * so no locks are needed. 1162 */ 1163static void 1164zil_itxg_clean(itxs_t *itxs) 1165{ 1166 itx_t *itx; 1167 list_t *list; 1168 avl_tree_t *t; 1169 void *cookie; 1170 itx_async_node_t *ian; 1171 1172 list = &itxs->i_sync_list; 1173 while ((itx = list_head(list)) != NULL) { 1174 list_remove(list, itx); 1175 kmem_free(itx, offsetof(itx_t, itx_lr) + 1176 itx->itx_lr.lrc_reclen); 1177 } 1178 1179 cookie = NULL; 1180 t = &itxs->i_async_tree; 1181 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1182 list = &ian->ia_list; 1183 while ((itx = list_head(list)) != NULL) { 1184 list_remove(list, itx); 1185 kmem_free(itx, offsetof(itx_t, itx_lr) + 1186 itx->itx_lr.lrc_reclen); 1187 } 1188 list_destroy(list); 1189 kmem_free(ian, sizeof (itx_async_node_t)); 1190 } 1191 avl_destroy(t); 1192 1193 kmem_free(itxs, sizeof (itxs_t)); 1194} 1195 1196static int 1197zil_aitx_compare(const void *x1, const void *x2) 1198{ 1199 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1200 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1201 1202 if (o1 < o2) 1203 return (-1); 1204 if (o1 > o2) 1205 return (1); 1206 1207 return (0); 1208} 1209 1210/* 1211 * Remove all async itx with the given oid. 1212 */ 1213static void 1214zil_remove_async(zilog_t *zilog, uint64_t oid) 1215{ 1216 uint64_t otxg, txg; 1217 itx_async_node_t *ian; 1218 avl_tree_t *t; 1219 avl_index_t where; 1220 list_t clean_list; 1221 itx_t *itx; 1222 1223 ASSERT(oid != 0); 1224 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1225 1226 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1227 otxg = ZILTEST_TXG; 1228 else 1229 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1230 1231 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1232 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1233 1234 mutex_enter(&itxg->itxg_lock); 1235 if (itxg->itxg_txg != txg) { 1236 mutex_exit(&itxg->itxg_lock); 1237 continue; 1238 } 1239 1240 /* 1241 * Locate the object node and append its list. 1242 */ 1243 t = &itxg->itxg_itxs->i_async_tree; 1244 ian = avl_find(t, &oid, &where); 1245 if (ian != NULL) 1246 list_move_tail(&clean_list, &ian->ia_list); 1247 mutex_exit(&itxg->itxg_lock); 1248 } 1249 while ((itx = list_head(&clean_list)) != NULL) { 1250 list_remove(&clean_list, itx); 1251 kmem_free(itx, offsetof(itx_t, itx_lr) + 1252 itx->itx_lr.lrc_reclen); 1253 } 1254 list_destroy(&clean_list); 1255} 1256 1257void 1258zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1259{ 1260 uint64_t txg; 1261 itxg_t *itxg; 1262 itxs_t *itxs, *clean = NULL; 1263 1264 /* 1265 * Object ids can be re-instantiated in the next txg so 1266 * remove any async transactions to avoid future leaks. 1267 * This can happen if a fsync occurs on the re-instantiated 1268 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1269 * the new file data and flushes a write record for the old object. 1270 */ 1271 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1272 zil_remove_async(zilog, itx->itx_oid); 1273 1274 /* 1275 * Ensure the data of a renamed file is committed before the rename. 1276 */ 1277 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1278 zil_async_to_sync(zilog, itx->itx_oid); 1279 1280 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1281 txg = ZILTEST_TXG; 1282 else 1283 txg = dmu_tx_get_txg(tx); 1284 1285 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1286 mutex_enter(&itxg->itxg_lock); 1287 itxs = itxg->itxg_itxs; 1288 if (itxg->itxg_txg != txg) { 1289 if (itxs != NULL) { 1290 /* 1291 * The zil_clean callback hasn't got around to cleaning 1292 * this itxg. Save the itxs for release below. 1293 * This should be rare. 1294 */ 1295 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1296 itxg->itxg_sod = 0; 1297 clean = itxg->itxg_itxs; 1298 } 1299 ASSERT(itxg->itxg_sod == 0); 1300 itxg->itxg_txg = txg; 1301 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1302 1303 list_create(&itxs->i_sync_list, sizeof (itx_t), 1304 offsetof(itx_t, itx_node)); 1305 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1306 sizeof (itx_async_node_t), 1307 offsetof(itx_async_node_t, ia_node)); 1308 } 1309 if (itx->itx_sync) { 1310 list_insert_tail(&itxs->i_sync_list, itx); 1311 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1312 itxg->itxg_sod += itx->itx_sod; 1313 } else { 1314 avl_tree_t *t = &itxs->i_async_tree; 1315 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1316 itx_async_node_t *ian; 1317 avl_index_t where; 1318 1319 ian = avl_find(t, &foid, &where); 1320 if (ian == NULL) { 1321 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1322 list_create(&ian->ia_list, sizeof (itx_t), 1323 offsetof(itx_t, itx_node)); 1324 ian->ia_foid = foid; 1325 avl_insert(t, ian, where); 1326 } 1327 list_insert_tail(&ian->ia_list, itx); 1328 } 1329 1330 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1331 zilog_dirty(zilog, txg); 1332 mutex_exit(&itxg->itxg_lock); 1333 1334 /* Release the old itxs now we've dropped the lock */ 1335 if (clean != NULL) 1336 zil_itxg_clean(clean); 1337} 1338 1339/* 1340 * If there are any in-memory intent log transactions which have now been 1341 * synced then start up a taskq to free them. We should only do this after we 1342 * have written out the uberblocks (i.e. txg has been comitted) so that 1343 * don't inadvertently clean out in-memory log records that would be required 1344 * by zil_commit(). 1345 */ 1346void 1347zil_clean(zilog_t *zilog, uint64_t synced_txg) 1348{ 1349 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1350 itxs_t *clean_me; 1351 1352 mutex_enter(&itxg->itxg_lock); 1353 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1354 mutex_exit(&itxg->itxg_lock); 1355 return; 1356 } 1357 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1358 ASSERT(itxg->itxg_txg != 0); 1359 ASSERT(zilog->zl_clean_taskq != NULL); 1360 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1361 itxg->itxg_sod = 0; 1362 clean_me = itxg->itxg_itxs; 1363 itxg->itxg_itxs = NULL; 1364 itxg->itxg_txg = 0; 1365 mutex_exit(&itxg->itxg_lock); 1366 /* 1367 * Preferably start a task queue to free up the old itxs but 1368 * if taskq_dispatch can't allocate resources to do that then 1369 * free it in-line. This should be rare. Note, using TQ_SLEEP 1370 * created a bad performance problem. 1371 */ 1372 if (taskq_dispatch(zilog->zl_clean_taskq, 1373 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1374 zil_itxg_clean(clean_me); 1375} 1376 1377/* 1378 * Get the list of itxs to commit into zl_itx_commit_list. 1379 */ 1380static void 1381zil_get_commit_list(zilog_t *zilog) 1382{ 1383 uint64_t otxg, txg; 1384 list_t *commit_list = &zilog->zl_itx_commit_list; 1385 uint64_t push_sod = 0; 1386 1387 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1388 otxg = ZILTEST_TXG; 1389 else 1390 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1391 1392 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1393 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1394 1395 mutex_enter(&itxg->itxg_lock); 1396 if (itxg->itxg_txg != txg) { 1397 mutex_exit(&itxg->itxg_lock); 1398 continue; 1399 } 1400 1401 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1402 push_sod += itxg->itxg_sod; 1403 itxg->itxg_sod = 0; 1404 1405 mutex_exit(&itxg->itxg_lock); 1406 } 1407 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1408} 1409 1410/* 1411 * Move the async itxs for a specified object to commit into sync lists. 1412 */ 1413void 1414zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1415{ 1416 uint64_t otxg, txg; 1417 itx_async_node_t *ian; 1418 avl_tree_t *t; 1419 avl_index_t where; 1420 1421 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1422 otxg = ZILTEST_TXG; 1423 else 1424 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1425 1426 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1427 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1428 1429 mutex_enter(&itxg->itxg_lock); 1430 if (itxg->itxg_txg != txg) { 1431 mutex_exit(&itxg->itxg_lock); 1432 continue; 1433 } 1434 1435 /* 1436 * If a foid is specified then find that node and append its 1437 * list. Otherwise walk the tree appending all the lists 1438 * to the sync list. We add to the end rather than the 1439 * beginning to ensure the create has happened. 1440 */ 1441 t = &itxg->itxg_itxs->i_async_tree; 1442 if (foid != 0) { 1443 ian = avl_find(t, &foid, &where); 1444 if (ian != NULL) { 1445 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1446 &ian->ia_list); 1447 } 1448 } else { 1449 void *cookie = NULL; 1450 1451 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1452 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1453 &ian->ia_list); 1454 list_destroy(&ian->ia_list); 1455 kmem_free(ian, sizeof (itx_async_node_t)); 1456 } 1457 } 1458 mutex_exit(&itxg->itxg_lock); 1459 } 1460} 1461 1462static void 1463zil_commit_writer(zilog_t *zilog) 1464{ 1465 uint64_t txg; 1466 itx_t *itx; 1467 lwb_t *lwb; 1468 spa_t *spa = zilog->zl_spa; 1469 int error = 0; 1470 1471 ASSERT(zilog->zl_root_zio == NULL); 1472 1473 mutex_exit(&zilog->zl_lock); 1474 1475 zil_get_commit_list(zilog); 1476 1477 /* 1478 * Return if there's nothing to commit before we dirty the fs by 1479 * calling zil_create(). 1480 */ 1481 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1482 mutex_enter(&zilog->zl_lock); 1483 return; 1484 } 1485 1486 if (zilog->zl_suspend) { 1487 lwb = NULL; 1488 } else { 1489 lwb = list_tail(&zilog->zl_lwb_list); 1490 if (lwb == NULL) 1491 lwb = zil_create(zilog); 1492 } 1493 1494 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1495 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1496 txg = itx->itx_lr.lrc_txg; 1497 ASSERT(txg); 1498 1499 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1500 lwb = zil_lwb_commit(zilog, itx, lwb); 1501 list_remove(&zilog->zl_itx_commit_list, itx); 1502 kmem_free(itx, offsetof(itx_t, itx_lr) 1503 + itx->itx_lr.lrc_reclen); 1504 } 1505 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1506 1507 /* write the last block out */ 1508 if (lwb != NULL && lwb->lwb_zio != NULL) 1509 lwb = zil_lwb_write_start(zilog, lwb); 1510 1511 zilog->zl_cur_used = 0; 1512 1513 /* 1514 * Wait if necessary for the log blocks to be on stable storage. 1515 */ 1516 if (zilog->zl_root_zio) { 1517 error = zio_wait(zilog->zl_root_zio); 1518 zilog->zl_root_zio = NULL; 1519 zil_flush_vdevs(zilog); 1520 } 1521 1522 if (error || lwb == NULL) 1523 txg_wait_synced(zilog->zl_dmu_pool, 0); 1524 1525 mutex_enter(&zilog->zl_lock); 1526 1527 /* 1528 * Remember the highest committed log sequence number for ztest. 1529 * We only update this value when all the log writes succeeded, 1530 * because ztest wants to ASSERT that it got the whole log chain. 1531 */ 1532 if (error == 0 && lwb != NULL) 1533 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1534} 1535 1536/* 1537 * Commit zfs transactions to stable storage. 1538 * If foid is 0 push out all transactions, otherwise push only those 1539 * for that object or might reference that object. 1540 * 1541 * itxs are committed in batches. In a heavily stressed zil there will be 1542 * a commit writer thread who is writing out a bunch of itxs to the log 1543 * for a set of committing threads (cthreads) in the same batch as the writer. 1544 * Those cthreads are all waiting on the same cv for that batch. 1545 * 1546 * There will also be a different and growing batch of threads that are 1547 * waiting to commit (qthreads). When the committing batch completes 1548 * a transition occurs such that the cthreads exit and the qthreads become 1549 * cthreads. One of the new cthreads becomes the writer thread for the 1550 * batch. Any new threads arriving become new qthreads. 1551 * 1552 * Only 2 condition variables are needed and there's no transition 1553 * between the two cvs needed. They just flip-flop between qthreads 1554 * and cthreads. 1555 * 1556 * Using this scheme we can efficiently wakeup up only those threads 1557 * that have been committed. 1558 */ 1559void 1560zil_commit(zilog_t *zilog, uint64_t foid) 1561{ 1562 uint64_t mybatch; 1563 1564 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1565 return; 1566 1567 /* move the async itxs for the foid to the sync queues */ 1568 zil_async_to_sync(zilog, foid); 1569 1570 mutex_enter(&zilog->zl_lock); 1571 mybatch = zilog->zl_next_batch; 1572 while (zilog->zl_writer) { 1573 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1574 if (mybatch <= zilog->zl_com_batch) { 1575 mutex_exit(&zilog->zl_lock); 1576 return; 1577 } 1578 } 1579 1580 zilog->zl_next_batch++; 1581 zilog->zl_writer = B_TRUE; 1582 zil_commit_writer(zilog); 1583 zilog->zl_com_batch = mybatch; 1584 zilog->zl_writer = B_FALSE; 1585 mutex_exit(&zilog->zl_lock); 1586 1587 /* wake up one thread to become the next writer */ 1588 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1589 1590 /* wake up all threads waiting for this batch to be committed */ 1591 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1592} 1593 1594/* 1595 * Called in syncing context to free committed log blocks and update log header. 1596 */ 1597void 1598zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1599{ 1600 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1601 uint64_t txg = dmu_tx_get_txg(tx); 1602 spa_t *spa = zilog->zl_spa; 1603 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1604 lwb_t *lwb; 1605 1606 /* 1607 * We don't zero out zl_destroy_txg, so make sure we don't try 1608 * to destroy it twice. 1609 */ 1610 if (spa_sync_pass(spa) != 1) 1611 return; 1612 1613 mutex_enter(&zilog->zl_lock); 1614 1615 ASSERT(zilog->zl_stop_sync == 0); 1616 1617 if (*replayed_seq != 0) { 1618 ASSERT(zh->zh_replay_seq < *replayed_seq); 1619 zh->zh_replay_seq = *replayed_seq; 1620 *replayed_seq = 0; 1621 } 1622 1623 if (zilog->zl_destroy_txg == txg) { 1624 blkptr_t blk = zh->zh_log; 1625 1626 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1627 1628 bzero(zh, sizeof (zil_header_t)); 1629 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1630 1631 if (zilog->zl_keep_first) { 1632 /* 1633 * If this block was part of log chain that couldn't 1634 * be claimed because a device was missing during 1635 * zil_claim(), but that device later returns, 1636 * then this block could erroneously appear valid. 1637 * To guard against this, assign a new GUID to the new 1638 * log chain so it doesn't matter what blk points to. 1639 */ 1640 zil_init_log_chain(zilog, &blk); 1641 zh->zh_log = blk; 1642 } 1643 } 1644 1645 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1646 zh->zh_log = lwb->lwb_blk; 1647 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1648 break; 1649 list_remove(&zilog->zl_lwb_list, lwb); 1650 zio_free_zil(spa, txg, &lwb->lwb_blk); 1651 kmem_cache_free(zil_lwb_cache, lwb); 1652 1653 /* 1654 * If we don't have anything left in the lwb list then 1655 * we've had an allocation failure and we need to zero 1656 * out the zil_header blkptr so that we don't end 1657 * up freeing the same block twice. 1658 */ 1659 if (list_head(&zilog->zl_lwb_list) == NULL) 1660 BP_ZERO(&zh->zh_log); 1661 } 1662 mutex_exit(&zilog->zl_lock); 1663} 1664 1665void 1666zil_init(void) 1667{ 1668 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1669 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1670} 1671 1672void 1673zil_fini(void) 1674{ 1675 kmem_cache_destroy(zil_lwb_cache); 1676} 1677 1678void 1679zil_set_sync(zilog_t *zilog, uint64_t sync) 1680{ 1681 zilog->zl_sync = sync; 1682} 1683 1684void 1685zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1686{ 1687 zilog->zl_logbias = logbias; 1688} 1689 1690zilog_t * 1691zil_alloc(objset_t *os, zil_header_t *zh_phys) 1692{ 1693 zilog_t *zilog; 1694 1695 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1696 1697 zilog->zl_header = zh_phys; 1698 zilog->zl_os = os; 1699 zilog->zl_spa = dmu_objset_spa(os); 1700 zilog->zl_dmu_pool = dmu_objset_pool(os); 1701 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1702 zilog->zl_logbias = dmu_objset_logbias(os); 1703 zilog->zl_sync = dmu_objset_syncprop(os); 1704 zilog->zl_next_batch = 1; 1705 1706 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1707 1708 for (int i = 0; i < TXG_SIZE; i++) { 1709 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1710 MUTEX_DEFAULT, NULL); 1711 } 1712 1713 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1714 offsetof(lwb_t, lwb_node)); 1715 1716 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1717 offsetof(itx_t, itx_node)); 1718 1719 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1720 1721 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1722 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1723 1724 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1725 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1726 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1727 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1728 1729 return (zilog); 1730} 1731 1732void 1733zil_free(zilog_t *zilog) 1734{ 1735 zilog->zl_stop_sync = 1; 1736 1737 ASSERT0(zilog->zl_suspend); 1738 ASSERT0(zilog->zl_suspending); 1739 1740 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1741 list_destroy(&zilog->zl_lwb_list); 1742 1743 avl_destroy(&zilog->zl_vdev_tree); 1744 mutex_destroy(&zilog->zl_vdev_lock); 1745 1746 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1747 list_destroy(&zilog->zl_itx_commit_list); 1748 1749 for (int i = 0; i < TXG_SIZE; i++) { 1750 /* 1751 * It's possible for an itx to be generated that doesn't dirty 1752 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1753 * callback to remove the entry. We remove those here. 1754 * 1755 * Also free up the ziltest itxs. 1756 */ 1757 if (zilog->zl_itxg[i].itxg_itxs) 1758 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1759 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1760 } 1761 1762 mutex_destroy(&zilog->zl_lock); 1763 1764 cv_destroy(&zilog->zl_cv_writer); 1765 cv_destroy(&zilog->zl_cv_suspend); 1766 cv_destroy(&zilog->zl_cv_batch[0]); 1767 cv_destroy(&zilog->zl_cv_batch[1]); 1768 1769 kmem_free(zilog, sizeof (zilog_t)); 1770} 1771 1772/* 1773 * Open an intent log. 1774 */ 1775zilog_t * 1776zil_open(objset_t *os, zil_get_data_t *get_data) 1777{ 1778 zilog_t *zilog = dmu_objset_zil(os); 1779 1780 ASSERT(zilog->zl_clean_taskq == NULL); 1781 ASSERT(zilog->zl_get_data == NULL); 1782 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1783 1784 zilog->zl_get_data = get_data; 1785 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1786 2, 2, TASKQ_PREPOPULATE); 1787 1788 return (zilog); 1789} 1790 1791/* 1792 * Close an intent log. 1793 */ 1794void 1795zil_close(zilog_t *zilog) 1796{ 1797 lwb_t *lwb; 1798 uint64_t txg = 0; 1799 1800 zil_commit(zilog, 0); /* commit all itx */ 1801 1802 /* 1803 * The lwb_max_txg for the stubby lwb will reflect the last activity 1804 * for the zil. After a txg_wait_synced() on the txg we know all the 1805 * callbacks have occurred that may clean the zil. Only then can we 1806 * destroy the zl_clean_taskq. 1807 */ 1808 mutex_enter(&zilog->zl_lock); 1809 lwb = list_tail(&zilog->zl_lwb_list); 1810 if (lwb != NULL) 1811 txg = lwb->lwb_max_txg; 1812 mutex_exit(&zilog->zl_lock); 1813 if (txg) 1814 txg_wait_synced(zilog->zl_dmu_pool, txg); 1815 ASSERT(!zilog_is_dirty(zilog)); 1816 1817 taskq_destroy(zilog->zl_clean_taskq); 1818 zilog->zl_clean_taskq = NULL; 1819 zilog->zl_get_data = NULL; 1820 1821 /* 1822 * We should have only one LWB left on the list; remove it now. 1823 */ 1824 mutex_enter(&zilog->zl_lock); 1825 lwb = list_head(&zilog->zl_lwb_list); 1826 if (lwb != NULL) { 1827 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1828 list_remove(&zilog->zl_lwb_list, lwb); 1829 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1830 kmem_cache_free(zil_lwb_cache, lwb); 1831 } 1832 mutex_exit(&zilog->zl_lock); 1833} 1834 1835static char *suspend_tag = "zil suspending"; 1836 1837/* 1838 * Suspend an intent log. While in suspended mode, we still honor 1839 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1840 * On old version pools, we suspend the log briefly when taking a 1841 * snapshot so that it will have an empty intent log. 1842 * 1843 * Long holds are not really intended to be used the way we do here -- 1844 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1845 * could fail. Therefore we take pains to only put a long hold if it is 1846 * actually necessary. Fortunately, it will only be necessary if the 1847 * objset is currently mounted (or the ZVOL equivalent). In that case it 1848 * will already have a long hold, so we are not really making things any worse. 1849 * 1850 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1851 * zvol_state_t), and use their mechanism to prevent their hold from being 1852 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1853 * very little gain. 1854 * 1855 * if cookiep == NULL, this does both the suspend & resume. 1856 * Otherwise, it returns with the dataset "long held", and the cookie 1857 * should be passed into zil_resume(). 1858 */ 1859int 1860zil_suspend(const char *osname, void **cookiep) 1861{ 1862 objset_t *os; 1863 zilog_t *zilog; 1864 const zil_header_t *zh; 1865 int error; 1866 1867 error = dmu_objset_hold(osname, suspend_tag, &os); 1868 if (error != 0) 1869 return (error); 1870 zilog = dmu_objset_zil(os); 1871 1872 mutex_enter(&zilog->zl_lock); 1873 zh = zilog->zl_header; 1874 1875 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1876 mutex_exit(&zilog->zl_lock); 1877 dmu_objset_rele(os, suspend_tag); 1878 return (SET_ERROR(EBUSY)); 1879 } 1880 1881 /* 1882 * Don't put a long hold in the cases where we can avoid it. This 1883 * is when there is no cookie so we are doing a suspend & resume 1884 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1885 * for the suspend because it's already suspended, or there's no ZIL. 1886 */ 1887 if (cookiep == NULL && !zilog->zl_suspending && 1888 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1889 mutex_exit(&zilog->zl_lock); 1890 dmu_objset_rele(os, suspend_tag); 1891 return (0); 1892 } 1893 1894 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1895 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1896 1897 zilog->zl_suspend++; 1898 1899 if (zilog->zl_suspend > 1) { 1900 /* 1901 * Someone else is already suspending it. 1902 * Just wait for them to finish. 1903 */ 1904 1905 while (zilog->zl_suspending) 1906 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1907 mutex_exit(&zilog->zl_lock); 1908 1909 if (cookiep == NULL) 1910 zil_resume(os); 1911 else 1912 *cookiep = os; 1913 return (0); 1914 } 1915 1916 /* 1917 * If there is no pointer to an on-disk block, this ZIL must not 1918 * be active (e.g. filesystem not mounted), so there's nothing 1919 * to clean up. 1920 */ 1921 if (BP_IS_HOLE(&zh->zh_log)) { 1922 ASSERT(cookiep != NULL); /* fast path already handled */ 1923 1924 *cookiep = os; 1925 mutex_exit(&zilog->zl_lock); 1926 return (0); 1927 } 1928 1929 zilog->zl_suspending = B_TRUE; 1930 mutex_exit(&zilog->zl_lock); 1931 1932 zil_commit(zilog, 0); 1933 1934 zil_destroy(zilog, B_FALSE); 1935 1936 mutex_enter(&zilog->zl_lock); 1937 zilog->zl_suspending = B_FALSE; 1938 cv_broadcast(&zilog->zl_cv_suspend); 1939 mutex_exit(&zilog->zl_lock); 1940 1941 if (cookiep == NULL) 1942 zil_resume(os); 1943 else 1944 *cookiep = os; 1945 return (0); 1946} 1947 1948void 1949zil_resume(void *cookie) 1950{ 1951 objset_t *os = cookie; 1952 zilog_t *zilog = dmu_objset_zil(os); 1953 1954 mutex_enter(&zilog->zl_lock); 1955 ASSERT(zilog->zl_suspend != 0); 1956 zilog->zl_suspend--; 1957 mutex_exit(&zilog->zl_lock); 1958 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1959 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1960} 1961 1962typedef struct zil_replay_arg { 1963 zil_replay_func_t **zr_replay; 1964 void *zr_arg; 1965 boolean_t zr_byteswap; 1966 char *zr_lr; 1967} zil_replay_arg_t; 1968 1969static int 1970zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1971{ 1972 char name[ZFS_MAX_DATASET_NAME_LEN]; 1973 1974 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1975 1976 dmu_objset_name(zilog->zl_os, name); 1977 1978 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1979 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1980 (u_longlong_t)lr->lrc_seq, 1981 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1982 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1983 1984 return (error); 1985} 1986 1987static int 1988zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1989{ 1990 zil_replay_arg_t *zr = zra; 1991 const zil_header_t *zh = zilog->zl_header; 1992 uint64_t reclen = lr->lrc_reclen; 1993 uint64_t txtype = lr->lrc_txtype; 1994 int error = 0; 1995 1996 zilog->zl_replaying_seq = lr->lrc_seq; 1997 1998 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1999 return (0); 2000 2001 if (lr->lrc_txg < claim_txg) /* already committed */ 2002 return (0); 2003 2004 /* Strip case-insensitive bit, still present in log record */ 2005 txtype &= ~TX_CI; 2006 2007 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2008 return (zil_replay_error(zilog, lr, EINVAL)); 2009 2010 /* 2011 * If this record type can be logged out of order, the object 2012 * (lr_foid) may no longer exist. That's legitimate, not an error. 2013 */ 2014 if (TX_OOO(txtype)) { 2015 error = dmu_object_info(zilog->zl_os, 2016 ((lr_ooo_t *)lr)->lr_foid, NULL); 2017 if (error == ENOENT || error == EEXIST) 2018 return (0); 2019 } 2020 2021 /* 2022 * Make a copy of the data so we can revise and extend it. 2023 */ 2024 bcopy(lr, zr->zr_lr, reclen); 2025 2026 /* 2027 * If this is a TX_WRITE with a blkptr, suck in the data. 2028 */ 2029 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2030 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2031 zr->zr_lr + reclen); 2032 if (error != 0) 2033 return (zil_replay_error(zilog, lr, error)); 2034 } 2035 2036 /* 2037 * The log block containing this lr may have been byteswapped 2038 * so that we can easily examine common fields like lrc_txtype. 2039 * However, the log is a mix of different record types, and only the 2040 * replay vectors know how to byteswap their records. Therefore, if 2041 * the lr was byteswapped, undo it before invoking the replay vector. 2042 */ 2043 if (zr->zr_byteswap) 2044 byteswap_uint64_array(zr->zr_lr, reclen); 2045 2046 /* 2047 * We must now do two things atomically: replay this log record, 2048 * and update the log header sequence number to reflect the fact that 2049 * we did so. At the end of each replay function the sequence number 2050 * is updated if we are in replay mode. 2051 */ 2052 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2053 if (error != 0) { 2054 /* 2055 * The DMU's dnode layer doesn't see removes until the txg 2056 * commits, so a subsequent claim can spuriously fail with 2057 * EEXIST. So if we receive any error we try syncing out 2058 * any removes then retry the transaction. Note that we 2059 * specify B_FALSE for byteswap now, so we don't do it twice. 2060 */ 2061 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2062 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2063 if (error != 0) 2064 return (zil_replay_error(zilog, lr, error)); 2065 } 2066 return (0); 2067} 2068 2069/* ARGSUSED */ 2070static int 2071zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2072{ 2073 zilog->zl_replay_blks++; 2074 2075 return (0); 2076} 2077 2078/* 2079 * If this dataset has a non-empty intent log, replay it and destroy it. 2080 */ 2081void 2082zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2083{ 2084 zilog_t *zilog = dmu_objset_zil(os); 2085 const zil_header_t *zh = zilog->zl_header; 2086 zil_replay_arg_t zr; 2087 2088 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2089 zil_destroy(zilog, B_TRUE); 2090 return; 2091 } 2092 2093 zr.zr_replay = replay_func; 2094 zr.zr_arg = arg; 2095 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2096 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2097 2098 /* 2099 * Wait for in-progress removes to sync before starting replay. 2100 */ 2101 txg_wait_synced(zilog->zl_dmu_pool, 0); 2102 2103 zilog->zl_replay = B_TRUE; 2104 zilog->zl_replay_time = ddi_get_lbolt(); 2105 ASSERT(zilog->zl_replay_blks == 0); 2106 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2107 zh->zh_claim_txg); 2108 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2109 2110 zil_destroy(zilog, B_FALSE); 2111 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2112 zilog->zl_replay = B_FALSE; 2113} 2114 2115boolean_t 2116zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2117{ 2118 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2119 return (B_TRUE); 2120 2121 if (zilog->zl_replay) { 2122 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2123 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2124 zilog->zl_replaying_seq; 2125 return (B_TRUE); 2126 } 2127 2128 return (B_FALSE); 2129} 2130 2131/* ARGSUSED */ 2132int 2133zil_vdev_offline(const char *osname, void *arg) 2134{ 2135 int error; 2136 2137 error = zil_suspend(osname, NULL); 2138 if (error != 0) 2139 return (SET_ERROR(EEXIST)); 2140 return (0); 2141} 2142