zil.c revision 268657
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, 2014 by Delphix. All rights reserved. 24 */ 25 26/* Portions Copyright 2010 Robert Milkowski */ 27 28#include <sys/zfs_context.h> 29#include <sys/spa.h> 30#include <sys/dmu.h> 31#include <sys/zap.h> 32#include <sys/arc.h> 33#include <sys/stat.h> 34#include <sys/resource.h> 35#include <sys/zil.h> 36#include <sys/zil_impl.h> 37#include <sys/dsl_dataset.h> 38#include <sys/vdev_impl.h> 39#include <sys/dmu_tx.h> 40#include <sys/dsl_pool.h> 41 42/* 43 * The zfs intent log (ZIL) saves transaction records of system calls 44 * that change the file system in memory with enough information 45 * to be able to replay them. These are stored in memory until 46 * either the DMU transaction group (txg) commits them to the stable pool 47 * and they can be discarded, or they are flushed to the stable log 48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 49 * requirement. In the event of a panic or power fail then those log 50 * records (transactions) are replayed. 51 * 52 * There is one ZIL per file system. Its on-disk (pool) format consists 53 * of 3 parts: 54 * 55 * - ZIL header 56 * - ZIL blocks 57 * - ZIL records 58 * 59 * A log record holds a system call transaction. Log blocks can 60 * hold many log records and the blocks are chained together. 61 * Each ZIL block contains a block pointer (blkptr_t) to the next 62 * ZIL block in the chain. The ZIL header points to the first 63 * block in the chain. Note there is not a fixed place in the pool 64 * to hold blocks. They are dynamically allocated and freed as 65 * needed from the blocks available. Figure X shows the ZIL structure: 66 */ 67 68/* 69 * Disable intent logging replay. This global ZIL switch affects all pools. 70 */ 71int zil_replay_disable = 0; 72SYSCTL_DECL(_vfs_zfs); 73TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 74SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 75 &zil_replay_disable, 0, "Disable intent logging replay"); 76 77/* 78 * Tunable parameter for debugging or performance analysis. Setting 79 * zfs_nocacheflush will cause corruption on power loss if a volatile 80 * out-of-order write cache is enabled. 81 */ 82boolean_t zfs_nocacheflush = B_FALSE; 83TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 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); 88TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 89SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 90 "Enable ZFS TRIM"); 91 92static kmem_cache_t *zil_lwb_cache; 93 94static void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 95 96#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 97 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 98 99 100/* 101 * ziltest is by and large an ugly hack, but very useful in 102 * checking replay without tedious work. 103 * When running ziltest we want to keep all itx's and so maintain 104 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 105 * We subtract TXG_CONCURRENT_STATES to allow for common code. 106 */ 107#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 108 109static int 110zil_bp_compare(const void *x1, const void *x2) 111{ 112 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 113 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 114 115 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 116 return (-1); 117 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 118 return (1); 119 120 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 121 return (-1); 122 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 123 return (1); 124 125 return (0); 126} 127 128static void 129zil_bp_tree_init(zilog_t *zilog) 130{ 131 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 132 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 133} 134 135static void 136zil_bp_tree_fini(zilog_t *zilog) 137{ 138 avl_tree_t *t = &zilog->zl_bp_tree; 139 zil_bp_node_t *zn; 140 void *cookie = NULL; 141 142 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 143 kmem_free(zn, sizeof (zil_bp_node_t)); 144 145 avl_destroy(t); 146} 147 148int 149zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 150{ 151 avl_tree_t *t = &zilog->zl_bp_tree; 152 const dva_t *dva; 153 zil_bp_node_t *zn; 154 avl_index_t where; 155 156 if (BP_IS_EMBEDDED(bp)) 157 return (0); 158 159 dva = BP_IDENTITY(bp); 160 161 if (avl_find(t, dva, &where) != NULL) 162 return (SET_ERROR(EEXIST)); 163 164 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 165 zn->zn_dva = *dva; 166 avl_insert(t, zn, where); 167 168 return (0); 169} 170 171static zil_header_t * 172zil_header_in_syncing_context(zilog_t *zilog) 173{ 174 return ((zil_header_t *)zilog->zl_header); 175} 176 177static void 178zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 179{ 180 zio_cksum_t *zc = &bp->blk_cksum; 181 182 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 183 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 184 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 185 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 186} 187 188/* 189 * Read a log block and make sure it's valid. 190 */ 191static int 192zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 193 char **end) 194{ 195 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 196 uint32_t aflags = ARC_WAIT; 197 arc_buf_t *abuf = NULL; 198 zbookmark_phys_t zb; 199 int error; 200 201 if (zilog->zl_header->zh_claim_txg == 0) 202 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 203 204 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 205 zio_flags |= ZIO_FLAG_SPECULATIVE; 206 207 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 208 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 209 210 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 211 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 212 213 if (error == 0) { 214 zio_cksum_t cksum = bp->blk_cksum; 215 216 /* 217 * Validate the checksummed log block. 218 * 219 * Sequence numbers should be... sequential. The checksum 220 * verifier for the next block should be bp's checksum plus 1. 221 * 222 * Also check the log chain linkage and size used. 223 */ 224 cksum.zc_word[ZIL_ZC_SEQ]++; 225 226 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 227 zil_chain_t *zilc = abuf->b_data; 228 char *lr = (char *)(zilc + 1); 229 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 230 231 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 232 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 233 error = SET_ERROR(ECKSUM); 234 } else { 235 bcopy(lr, dst, len); 236 *end = (char *)dst + len; 237 *nbp = zilc->zc_next_blk; 238 } 239 } else { 240 char *lr = abuf->b_data; 241 uint64_t size = BP_GET_LSIZE(bp); 242 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 243 244 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 245 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 246 (zilc->zc_nused > (size - sizeof (*zilc)))) { 247 error = SET_ERROR(ECKSUM); 248 } else { 249 bcopy(lr, dst, zilc->zc_nused); 250 *end = (char *)dst + zilc->zc_nused; 251 *nbp = zilc->zc_next_blk; 252 } 253 } 254 255 VERIFY(arc_buf_remove_ref(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 uint32_t aflags = ARC_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 (void) arc_buf_remove_ref(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_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_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 (dsl_dataset_is_snapshot(ds)) 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(const char *osname, 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(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); 649 if (error != 0) { 650 cmn_err(CE_WARN, "can't open objset for %s", osname); 651 return (0); 652 } 653 654 zilog = dmu_objset_zil(os); 655 zh = zil_header_in_syncing_context(zilog); 656 657 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 658 if (!BP_IS_HOLE(&zh->zh_log)) 659 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 660 BP_ZERO(&zh->zh_log); 661 dsl_dataset_dirty(dmu_objset_ds(os), tx); 662 dmu_objset_disown(os, FTAG); 663 return (0); 664 } 665 666 /* 667 * Claim all log blocks if we haven't already done so, and remember 668 * the highest claimed sequence number. This ensures that if we can 669 * read only part of the log now (e.g. due to a missing device), 670 * but we can read the entire log later, we will not try to replay 671 * or destroy beyond the last block we successfully claimed. 672 */ 673 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 674 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 675 (void) zil_parse(zilog, zil_claim_log_block, 676 zil_claim_log_record, tx, first_txg); 677 zh->zh_claim_txg = first_txg; 678 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 679 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 680 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 681 zh->zh_flags |= ZIL_REPLAY_NEEDED; 682 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 683 dsl_dataset_dirty(dmu_objset_ds(os), tx); 684 } 685 686 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 687 dmu_objset_disown(os, FTAG); 688 return (0); 689} 690 691/* 692 * Check the log by walking the log chain. 693 * Checksum errors are ok as they indicate the end of the chain. 694 * Any other error (no device or read failure) returns an error. 695 */ 696int 697zil_check_log_chain(const char *osname, void *tx) 698{ 699 zilog_t *zilog; 700 objset_t *os; 701 blkptr_t *bp; 702 int error; 703 704 ASSERT(tx == NULL); 705 706 error = dmu_objset_hold(osname, FTAG, &os); 707 if (error != 0) { 708 cmn_err(CE_WARN, "can't open objset for %s", osname); 709 return (0); 710 } 711 712 zilog = dmu_objset_zil(os); 713 bp = (blkptr_t *)&zilog->zl_header->zh_log; 714 715 /* 716 * Check the first block and determine if it's on a log device 717 * which may have been removed or faulted prior to loading this 718 * pool. If so, there's no point in checking the rest of the log 719 * as its content should have already been synced to the pool. 720 */ 721 if (!BP_IS_HOLE(bp)) { 722 vdev_t *vd; 723 boolean_t valid = B_TRUE; 724 725 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 726 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 727 if (vd->vdev_islog && vdev_is_dead(vd)) 728 valid = vdev_log_state_valid(vd); 729 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 730 731 if (!valid) { 732 dmu_objset_rele(os, FTAG); 733 return (0); 734 } 735 } 736 737 /* 738 * Because tx == NULL, zil_claim_log_block() will not actually claim 739 * any blocks, but just determine whether it is possible to do so. 740 * In addition to checking the log chain, zil_claim_log_block() 741 * will invoke zio_claim() with a done func of spa_claim_notify(), 742 * which will update spa_max_claim_txg. See spa_load() for details. 743 */ 744 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 745 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 746 747 dmu_objset_rele(os, FTAG); 748 749 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 750} 751 752static int 753zil_vdev_compare(const void *x1, const void *x2) 754{ 755 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 756 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 757 758 if (v1 < v2) 759 return (-1); 760 if (v1 > v2) 761 return (1); 762 763 return (0); 764} 765 766void 767zil_add_block(zilog_t *zilog, const blkptr_t *bp) 768{ 769 avl_tree_t *t = &zilog->zl_vdev_tree; 770 avl_index_t where; 771 zil_vdev_node_t *zv, zvsearch; 772 int ndvas = BP_GET_NDVAS(bp); 773 int i; 774 775 if (zfs_nocacheflush) 776 return; 777 778 ASSERT(zilog->zl_writer); 779 780 /* 781 * Even though we're zl_writer, we still need a lock because the 782 * zl_get_data() callbacks may have dmu_sync() done callbacks 783 * that will run concurrently. 784 */ 785 mutex_enter(&zilog->zl_vdev_lock); 786 for (i = 0; i < ndvas; i++) { 787 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 788 if (avl_find(t, &zvsearch, &where) == NULL) { 789 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 790 zv->zv_vdev = zvsearch.zv_vdev; 791 avl_insert(t, zv, where); 792 } 793 } 794 mutex_exit(&zilog->zl_vdev_lock); 795} 796 797static void 798zil_flush_vdevs(zilog_t *zilog) 799{ 800 spa_t *spa = zilog->zl_spa; 801 avl_tree_t *t = &zilog->zl_vdev_tree; 802 void *cookie = NULL; 803 zil_vdev_node_t *zv; 804 zio_t *zio; 805 806 ASSERT(zilog->zl_writer); 807 808 /* 809 * We don't need zl_vdev_lock here because we're the zl_writer, 810 * and all zl_get_data() callbacks are done. 811 */ 812 if (avl_numnodes(t) == 0) 813 return; 814 815 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 816 817 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 818 819 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 820 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 821 if (vd != NULL) 822 zio_flush(zio, vd); 823 kmem_free(zv, sizeof (*zv)); 824 } 825 826 /* 827 * Wait for all the flushes to complete. Not all devices actually 828 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 829 */ 830 (void) zio_wait(zio); 831 832 spa_config_exit(spa, SCL_STATE, FTAG); 833} 834 835/* 836 * Function called when a log block write completes 837 */ 838static void 839zil_lwb_write_done(zio_t *zio) 840{ 841 lwb_t *lwb = zio->io_private; 842 zilog_t *zilog = lwb->lwb_zilog; 843 dmu_tx_t *tx = lwb->lwb_tx; 844 845 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 846 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 847 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 848 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 849 ASSERT(!BP_IS_GANG(zio->io_bp)); 850 ASSERT(!BP_IS_HOLE(zio->io_bp)); 851 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 852 853 /* 854 * Ensure the lwb buffer pointer is cleared before releasing 855 * the txg. If we have had an allocation failure and 856 * the txg is waiting to sync then we want want zil_sync() 857 * to remove the lwb so that it's not picked up as the next new 858 * one in zil_commit_writer(). zil_sync() will only remove 859 * the lwb if lwb_buf is null. 860 */ 861 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 862 mutex_enter(&zilog->zl_lock); 863 lwb->lwb_buf = NULL; 864 lwb->lwb_tx = NULL; 865 mutex_exit(&zilog->zl_lock); 866 867 /* 868 * Now that we've written this log block, we have a stable pointer 869 * to the next block in the chain, so it's OK to let the txg in 870 * which we allocated the next block sync. 871 */ 872 dmu_tx_commit(tx); 873} 874 875/* 876 * Initialize the io for a log block. 877 */ 878static void 879zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 880{ 881 zbookmark_phys_t zb; 882 883 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 884 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 885 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 886 887 if (zilog->zl_root_zio == NULL) { 888 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 889 ZIO_FLAG_CANFAIL); 890 } 891 if (lwb->lwb_zio == NULL) { 892 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 893 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 894 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 895 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 896 } 897} 898 899/* 900 * Define a limited set of intent log block sizes. 901 * 902 * These must be a multiple of 4KB. Note only the amount used (again 903 * aligned to 4KB) actually gets written. However, we can't always just 904 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 905 */ 906uint64_t zil_block_buckets[] = { 907 4096, /* non TX_WRITE */ 908 8192+4096, /* data base */ 909 32*1024 + 4096, /* NFS writes */ 910 UINT64_MAX 911}; 912 913/* 914 * Use the slog as long as the logbias is 'latency' and the current commit size 915 * is less than the limit or the total list size is less than 2X the limit. 916 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 917 */ 918uint64_t zil_slog_limit = 1024 * 1024; 919#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 920 (((zilog)->zl_cur_used < zil_slog_limit) || \ 921 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 922 923/* 924 * Start a log block write and advance to the next log block. 925 * Calls are serialized. 926 */ 927static lwb_t * 928zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 929{ 930 lwb_t *nlwb = NULL; 931 zil_chain_t *zilc; 932 spa_t *spa = zilog->zl_spa; 933 blkptr_t *bp; 934 dmu_tx_t *tx; 935 uint64_t txg; 936 uint64_t zil_blksz, wsz; 937 int i, error; 938 939 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 940 zilc = (zil_chain_t *)lwb->lwb_buf; 941 bp = &zilc->zc_next_blk; 942 } else { 943 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 944 bp = &zilc->zc_next_blk; 945 } 946 947 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 948 949 /* 950 * Allocate the next block and save its address in this block 951 * before writing it in order to establish the log chain. 952 * Note that if the allocation of nlwb synced before we wrote 953 * the block that points at it (lwb), we'd leak it if we crashed. 954 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 955 * We dirty the dataset to ensure that zil_sync() will be called 956 * to clean up in the event of allocation failure or I/O failure. 957 */ 958 tx = dmu_tx_create(zilog->zl_os); 959 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 960 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 961 txg = dmu_tx_get_txg(tx); 962 963 lwb->lwb_tx = tx; 964 965 /* 966 * Log blocks are pre-allocated. Here we select the size of the next 967 * block, based on size used in the last block. 968 * - first find the smallest bucket that will fit the block from a 969 * limited set of block sizes. This is because it's faster to write 970 * blocks allocated from the same metaslab as they are adjacent or 971 * close. 972 * - next find the maximum from the new suggested size and an array of 973 * previous sizes. This lessens a picket fence effect of wrongly 974 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 975 * requests. 976 * 977 * Note we only write what is used, but we can't just allocate 978 * the maximum block size because we can exhaust the available 979 * pool log space. 980 */ 981 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 982 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 983 continue; 984 zil_blksz = zil_block_buckets[i]; 985 if (zil_blksz == UINT64_MAX) 986 zil_blksz = SPA_MAXBLOCKSIZE; 987 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 988 for (i = 0; i < ZIL_PREV_BLKS; i++) 989 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 990 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 991 992 BP_ZERO(bp); 993 /* pass the old blkptr in order to spread log blocks across devs */ 994 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 995 USE_SLOG(zilog)); 996 if (error == 0) { 997 ASSERT3U(bp->blk_birth, ==, txg); 998 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 999 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1000 1001 /* 1002 * Allocate a new log write buffer (lwb). 1003 */ 1004 nlwb = zil_alloc_lwb(zilog, bp, txg); 1005 1006 /* Record the block for later vdev flushing */ 1007 zil_add_block(zilog, &lwb->lwb_blk); 1008 } 1009 1010 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1011 /* For Slim ZIL only write what is used. */ 1012 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1013 ASSERT3U(wsz, <=, lwb->lwb_sz); 1014 zio_shrink(lwb->lwb_zio, wsz); 1015 1016 } else { 1017 wsz = lwb->lwb_sz; 1018 } 1019 1020 zilc->zc_pad = 0; 1021 zilc->zc_nused = lwb->lwb_nused; 1022 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1023 1024 /* 1025 * clear unused data for security 1026 */ 1027 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1028 1029 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1030 1031 /* 1032 * If there was an allocation failure then nlwb will be null which 1033 * forces a txg_wait_synced(). 1034 */ 1035 return (nlwb); 1036} 1037 1038static lwb_t * 1039zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1040{ 1041 lr_t *lrc = &itx->itx_lr; /* common log record */ 1042 lr_write_t *lrw = (lr_write_t *)lrc; 1043 char *lr_buf; 1044 uint64_t txg = lrc->lrc_txg; 1045 uint64_t reclen = lrc->lrc_reclen; 1046 uint64_t dlen = 0; 1047 1048 if (lwb == NULL) 1049 return (NULL); 1050 1051 ASSERT(lwb->lwb_buf != NULL); 1052 ASSERT(zilog_is_dirty(zilog) || 1053 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1054 1055 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1056 dlen = P2ROUNDUP_TYPED( 1057 lrw->lr_length, sizeof (uint64_t), uint64_t); 1058 1059 zilog->zl_cur_used += (reclen + dlen); 1060 1061 zil_lwb_write_init(zilog, lwb); 1062 1063 /* 1064 * If this record won't fit in the current log block, start a new one. 1065 */ 1066 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1067 lwb = zil_lwb_write_start(zilog, lwb); 1068 if (lwb == NULL) 1069 return (NULL); 1070 zil_lwb_write_init(zilog, lwb); 1071 ASSERT(LWB_EMPTY(lwb)); 1072 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1073 txg_wait_synced(zilog->zl_dmu_pool, txg); 1074 return (lwb); 1075 } 1076 } 1077 1078 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1079 bcopy(lrc, lr_buf, reclen); 1080 lrc = (lr_t *)lr_buf; 1081 lrw = (lr_write_t *)lrc; 1082 1083 /* 1084 * If it's a write, fetch the data or get its blkptr as appropriate. 1085 */ 1086 if (lrc->lrc_txtype == TX_WRITE) { 1087 if (txg > spa_freeze_txg(zilog->zl_spa)) 1088 txg_wait_synced(zilog->zl_dmu_pool, txg); 1089 if (itx->itx_wr_state != WR_COPIED) { 1090 char *dbuf; 1091 int error; 1092 1093 if (dlen) { 1094 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1095 dbuf = lr_buf + reclen; 1096 lrw->lr_common.lrc_reclen += dlen; 1097 } else { 1098 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1099 dbuf = NULL; 1100 } 1101 error = zilog->zl_get_data( 1102 itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1103 if (error == EIO) { 1104 txg_wait_synced(zilog->zl_dmu_pool, txg); 1105 return (lwb); 1106 } 1107 if (error != 0) { 1108 ASSERT(error == ENOENT || error == EEXIST || 1109 error == EALREADY); 1110 return (lwb); 1111 } 1112 } 1113 } 1114 1115 /* 1116 * We're actually making an entry, so update lrc_seq to be the 1117 * log record sequence number. Note that this is generally not 1118 * equal to the itx sequence number because not all transactions 1119 * are synchronous, and sometimes spa_sync() gets there first. 1120 */ 1121 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1122 lwb->lwb_nused += reclen + dlen; 1123 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1124 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1125 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1126 1127 return (lwb); 1128} 1129 1130itx_t * 1131zil_itx_create(uint64_t txtype, size_t lrsize) 1132{ 1133 itx_t *itx; 1134 1135 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1136 1137 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1138 itx->itx_lr.lrc_txtype = txtype; 1139 itx->itx_lr.lrc_reclen = lrsize; 1140 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1141 itx->itx_lr.lrc_seq = 0; /* defensive */ 1142 itx->itx_sync = B_TRUE; /* default is synchronous */ 1143 1144 return (itx); 1145} 1146 1147void 1148zil_itx_destroy(itx_t *itx) 1149{ 1150 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1151} 1152 1153/* 1154 * Free up the sync and async itxs. The itxs_t has already been detached 1155 * so no locks are needed. 1156 */ 1157static void 1158zil_itxg_clean(itxs_t *itxs) 1159{ 1160 itx_t *itx; 1161 list_t *list; 1162 avl_tree_t *t; 1163 void *cookie; 1164 itx_async_node_t *ian; 1165 1166 list = &itxs->i_sync_list; 1167 while ((itx = list_head(list)) != NULL) { 1168 list_remove(list, itx); 1169 kmem_free(itx, offsetof(itx_t, itx_lr) + 1170 itx->itx_lr.lrc_reclen); 1171 } 1172 1173 cookie = NULL; 1174 t = &itxs->i_async_tree; 1175 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1176 list = &ian->ia_list; 1177 while ((itx = list_head(list)) != NULL) { 1178 list_remove(list, itx); 1179 kmem_free(itx, offsetof(itx_t, itx_lr) + 1180 itx->itx_lr.lrc_reclen); 1181 } 1182 list_destroy(list); 1183 kmem_free(ian, sizeof (itx_async_node_t)); 1184 } 1185 avl_destroy(t); 1186 1187 kmem_free(itxs, sizeof (itxs_t)); 1188} 1189 1190static int 1191zil_aitx_compare(const void *x1, const void *x2) 1192{ 1193 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1194 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1195 1196 if (o1 < o2) 1197 return (-1); 1198 if (o1 > o2) 1199 return (1); 1200 1201 return (0); 1202} 1203 1204/* 1205 * Remove all async itx with the given oid. 1206 */ 1207static void 1208zil_remove_async(zilog_t *zilog, uint64_t oid) 1209{ 1210 uint64_t otxg, txg; 1211 itx_async_node_t *ian; 1212 avl_tree_t *t; 1213 avl_index_t where; 1214 list_t clean_list; 1215 itx_t *itx; 1216 1217 ASSERT(oid != 0); 1218 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1219 1220 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1221 otxg = ZILTEST_TXG; 1222 else 1223 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1224 1225 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1226 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1227 1228 mutex_enter(&itxg->itxg_lock); 1229 if (itxg->itxg_txg != txg) { 1230 mutex_exit(&itxg->itxg_lock); 1231 continue; 1232 } 1233 1234 /* 1235 * Locate the object node and append its list. 1236 */ 1237 t = &itxg->itxg_itxs->i_async_tree; 1238 ian = avl_find(t, &oid, &where); 1239 if (ian != NULL) 1240 list_move_tail(&clean_list, &ian->ia_list); 1241 mutex_exit(&itxg->itxg_lock); 1242 } 1243 while ((itx = list_head(&clean_list)) != NULL) { 1244 list_remove(&clean_list, itx); 1245 kmem_free(itx, offsetof(itx_t, itx_lr) + 1246 itx->itx_lr.lrc_reclen); 1247 } 1248 list_destroy(&clean_list); 1249} 1250 1251void 1252zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1253{ 1254 uint64_t txg; 1255 itxg_t *itxg; 1256 itxs_t *itxs, *clean = NULL; 1257 1258 /* 1259 * Object ids can be re-instantiated in the next txg so 1260 * remove any async transactions to avoid future leaks. 1261 * This can happen if a fsync occurs on the re-instantiated 1262 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1263 * the new file data and flushes a write record for the old object. 1264 */ 1265 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1266 zil_remove_async(zilog, itx->itx_oid); 1267 1268 /* 1269 * Ensure the data of a renamed file is committed before the rename. 1270 */ 1271 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1272 zil_async_to_sync(zilog, itx->itx_oid); 1273 1274 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1275 txg = ZILTEST_TXG; 1276 else 1277 txg = dmu_tx_get_txg(tx); 1278 1279 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1280 mutex_enter(&itxg->itxg_lock); 1281 itxs = itxg->itxg_itxs; 1282 if (itxg->itxg_txg != txg) { 1283 if (itxs != NULL) { 1284 /* 1285 * The zil_clean callback hasn't got around to cleaning 1286 * this itxg. Save the itxs for release below. 1287 * This should be rare. 1288 */ 1289 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1290 itxg->itxg_sod = 0; 1291 clean = itxg->itxg_itxs; 1292 } 1293 ASSERT(itxg->itxg_sod == 0); 1294 itxg->itxg_txg = txg; 1295 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1296 1297 list_create(&itxs->i_sync_list, sizeof (itx_t), 1298 offsetof(itx_t, itx_node)); 1299 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1300 sizeof (itx_async_node_t), 1301 offsetof(itx_async_node_t, ia_node)); 1302 } 1303 if (itx->itx_sync) { 1304 list_insert_tail(&itxs->i_sync_list, itx); 1305 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1306 itxg->itxg_sod += itx->itx_sod; 1307 } else { 1308 avl_tree_t *t = &itxs->i_async_tree; 1309 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1310 itx_async_node_t *ian; 1311 avl_index_t where; 1312 1313 ian = avl_find(t, &foid, &where); 1314 if (ian == NULL) { 1315 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1316 list_create(&ian->ia_list, sizeof (itx_t), 1317 offsetof(itx_t, itx_node)); 1318 ian->ia_foid = foid; 1319 avl_insert(t, ian, where); 1320 } 1321 list_insert_tail(&ian->ia_list, itx); 1322 } 1323 1324 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1325 zilog_dirty(zilog, txg); 1326 mutex_exit(&itxg->itxg_lock); 1327 1328 /* Release the old itxs now we've dropped the lock */ 1329 if (clean != NULL) 1330 zil_itxg_clean(clean); 1331} 1332 1333/* 1334 * If there are any in-memory intent log transactions which have now been 1335 * synced then start up a taskq to free them. We should only do this after we 1336 * have written out the uberblocks (i.e. txg has been comitted) so that 1337 * don't inadvertently clean out in-memory log records that would be required 1338 * by zil_commit(). 1339 */ 1340void 1341zil_clean(zilog_t *zilog, uint64_t synced_txg) 1342{ 1343 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1344 itxs_t *clean_me; 1345 1346 mutex_enter(&itxg->itxg_lock); 1347 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1348 mutex_exit(&itxg->itxg_lock); 1349 return; 1350 } 1351 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1352 ASSERT(itxg->itxg_txg != 0); 1353 ASSERT(zilog->zl_clean_taskq != NULL); 1354 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1355 itxg->itxg_sod = 0; 1356 clean_me = itxg->itxg_itxs; 1357 itxg->itxg_itxs = NULL; 1358 itxg->itxg_txg = 0; 1359 mutex_exit(&itxg->itxg_lock); 1360 /* 1361 * Preferably start a task queue to free up the old itxs but 1362 * if taskq_dispatch can't allocate resources to do that then 1363 * free it in-line. This should be rare. Note, using TQ_SLEEP 1364 * created a bad performance problem. 1365 */ 1366 if (taskq_dispatch(zilog->zl_clean_taskq, 1367 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1368 zil_itxg_clean(clean_me); 1369} 1370 1371/* 1372 * Get the list of itxs to commit into zl_itx_commit_list. 1373 */ 1374static void 1375zil_get_commit_list(zilog_t *zilog) 1376{ 1377 uint64_t otxg, txg; 1378 list_t *commit_list = &zilog->zl_itx_commit_list; 1379 uint64_t push_sod = 0; 1380 1381 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1382 otxg = ZILTEST_TXG; 1383 else 1384 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1385 1386 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1387 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1388 1389 mutex_enter(&itxg->itxg_lock); 1390 if (itxg->itxg_txg != txg) { 1391 mutex_exit(&itxg->itxg_lock); 1392 continue; 1393 } 1394 1395 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1396 push_sod += itxg->itxg_sod; 1397 itxg->itxg_sod = 0; 1398 1399 mutex_exit(&itxg->itxg_lock); 1400 } 1401 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1402} 1403 1404/* 1405 * Move the async itxs for a specified object to commit into sync lists. 1406 */ 1407static void 1408zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1409{ 1410 uint64_t otxg, txg; 1411 itx_async_node_t *ian; 1412 avl_tree_t *t; 1413 avl_index_t where; 1414 1415 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1416 otxg = ZILTEST_TXG; 1417 else 1418 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1419 1420 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1421 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1422 1423 mutex_enter(&itxg->itxg_lock); 1424 if (itxg->itxg_txg != txg) { 1425 mutex_exit(&itxg->itxg_lock); 1426 continue; 1427 } 1428 1429 /* 1430 * If a foid is specified then find that node and append its 1431 * list. Otherwise walk the tree appending all the lists 1432 * to the sync list. We add to the end rather than the 1433 * beginning to ensure the create has happened. 1434 */ 1435 t = &itxg->itxg_itxs->i_async_tree; 1436 if (foid != 0) { 1437 ian = avl_find(t, &foid, &where); 1438 if (ian != NULL) { 1439 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1440 &ian->ia_list); 1441 } 1442 } else { 1443 void *cookie = NULL; 1444 1445 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1446 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1447 &ian->ia_list); 1448 list_destroy(&ian->ia_list); 1449 kmem_free(ian, sizeof (itx_async_node_t)); 1450 } 1451 } 1452 mutex_exit(&itxg->itxg_lock); 1453 } 1454} 1455 1456static void 1457zil_commit_writer(zilog_t *zilog) 1458{ 1459 uint64_t txg; 1460 itx_t *itx; 1461 lwb_t *lwb; 1462 spa_t *spa = zilog->zl_spa; 1463 int error = 0; 1464 1465 ASSERT(zilog->zl_root_zio == NULL); 1466 1467 mutex_exit(&zilog->zl_lock); 1468 1469 zil_get_commit_list(zilog); 1470 1471 /* 1472 * Return if there's nothing to commit before we dirty the fs by 1473 * calling zil_create(). 1474 */ 1475 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1476 mutex_enter(&zilog->zl_lock); 1477 return; 1478 } 1479 1480 if (zilog->zl_suspend) { 1481 lwb = NULL; 1482 } else { 1483 lwb = list_tail(&zilog->zl_lwb_list); 1484 if (lwb == NULL) 1485 lwb = zil_create(zilog); 1486 } 1487 1488 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1489 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1490 txg = itx->itx_lr.lrc_txg; 1491 ASSERT(txg); 1492 1493 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1494 lwb = zil_lwb_commit(zilog, itx, lwb); 1495 list_remove(&zilog->zl_itx_commit_list, itx); 1496 kmem_free(itx, offsetof(itx_t, itx_lr) 1497 + itx->itx_lr.lrc_reclen); 1498 } 1499 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1500 1501 /* write the last block out */ 1502 if (lwb != NULL && lwb->lwb_zio != NULL) 1503 lwb = zil_lwb_write_start(zilog, lwb); 1504 1505 zilog->zl_cur_used = 0; 1506 1507 /* 1508 * Wait if necessary for the log blocks to be on stable storage. 1509 */ 1510 if (zilog->zl_root_zio) { 1511 error = zio_wait(zilog->zl_root_zio); 1512 zilog->zl_root_zio = NULL; 1513 zil_flush_vdevs(zilog); 1514 } 1515 1516 if (error || lwb == NULL) 1517 txg_wait_synced(zilog->zl_dmu_pool, 0); 1518 1519 mutex_enter(&zilog->zl_lock); 1520 1521 /* 1522 * Remember the highest committed log sequence number for ztest. 1523 * We only update this value when all the log writes succeeded, 1524 * because ztest wants to ASSERT that it got the whole log chain. 1525 */ 1526 if (error == 0 && lwb != NULL) 1527 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1528} 1529 1530/* 1531 * Commit zfs transactions to stable storage. 1532 * If foid is 0 push out all transactions, otherwise push only those 1533 * for that object or might reference that object. 1534 * 1535 * itxs are committed in batches. In a heavily stressed zil there will be 1536 * a commit writer thread who is writing out a bunch of itxs to the log 1537 * for a set of committing threads (cthreads) in the same batch as the writer. 1538 * Those cthreads are all waiting on the same cv for that batch. 1539 * 1540 * There will also be a different and growing batch of threads that are 1541 * waiting to commit (qthreads). When the committing batch completes 1542 * a transition occurs such that the cthreads exit and the qthreads become 1543 * cthreads. One of the new cthreads becomes the writer thread for the 1544 * batch. Any new threads arriving become new qthreads. 1545 * 1546 * Only 2 condition variables are needed and there's no transition 1547 * between the two cvs needed. They just flip-flop between qthreads 1548 * and cthreads. 1549 * 1550 * Using this scheme we can efficiently wakeup up only those threads 1551 * that have been committed. 1552 */ 1553void 1554zil_commit(zilog_t *zilog, uint64_t foid) 1555{ 1556 uint64_t mybatch; 1557 1558 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1559 return; 1560 1561 /* move the async itxs for the foid to the sync queues */ 1562 zil_async_to_sync(zilog, foid); 1563 1564 mutex_enter(&zilog->zl_lock); 1565 mybatch = zilog->zl_next_batch; 1566 while (zilog->zl_writer) { 1567 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1568 if (mybatch <= zilog->zl_com_batch) { 1569 mutex_exit(&zilog->zl_lock); 1570 return; 1571 } 1572 } 1573 1574 zilog->zl_next_batch++; 1575 zilog->zl_writer = B_TRUE; 1576 zil_commit_writer(zilog); 1577 zilog->zl_com_batch = mybatch; 1578 zilog->zl_writer = B_FALSE; 1579 mutex_exit(&zilog->zl_lock); 1580 1581 /* wake up one thread to become the next writer */ 1582 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1583 1584 /* wake up all threads waiting for this batch to be committed */ 1585 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1586} 1587 1588/* 1589 * Called in syncing context to free committed log blocks and update log header. 1590 */ 1591void 1592zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1593{ 1594 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1595 uint64_t txg = dmu_tx_get_txg(tx); 1596 spa_t *spa = zilog->zl_spa; 1597 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1598 lwb_t *lwb; 1599 1600 /* 1601 * We don't zero out zl_destroy_txg, so make sure we don't try 1602 * to destroy it twice. 1603 */ 1604 if (spa_sync_pass(spa) != 1) 1605 return; 1606 1607 mutex_enter(&zilog->zl_lock); 1608 1609 ASSERT(zilog->zl_stop_sync == 0); 1610 1611 if (*replayed_seq != 0) { 1612 ASSERT(zh->zh_replay_seq < *replayed_seq); 1613 zh->zh_replay_seq = *replayed_seq; 1614 *replayed_seq = 0; 1615 } 1616 1617 if (zilog->zl_destroy_txg == txg) { 1618 blkptr_t blk = zh->zh_log; 1619 1620 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1621 1622 bzero(zh, sizeof (zil_header_t)); 1623 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1624 1625 if (zilog->zl_keep_first) { 1626 /* 1627 * If this block was part of log chain that couldn't 1628 * be claimed because a device was missing during 1629 * zil_claim(), but that device later returns, 1630 * then this block could erroneously appear valid. 1631 * To guard against this, assign a new GUID to the new 1632 * log chain so it doesn't matter what blk points to. 1633 */ 1634 zil_init_log_chain(zilog, &blk); 1635 zh->zh_log = blk; 1636 } 1637 } 1638 1639 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1640 zh->zh_log = lwb->lwb_blk; 1641 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1642 break; 1643 list_remove(&zilog->zl_lwb_list, lwb); 1644 zio_free_zil(spa, txg, &lwb->lwb_blk); 1645 kmem_cache_free(zil_lwb_cache, lwb); 1646 1647 /* 1648 * If we don't have anything left in the lwb list then 1649 * we've had an allocation failure and we need to zero 1650 * out the zil_header blkptr so that we don't end 1651 * up freeing the same block twice. 1652 */ 1653 if (list_head(&zilog->zl_lwb_list) == NULL) 1654 BP_ZERO(&zh->zh_log); 1655 } 1656 mutex_exit(&zilog->zl_lock); 1657} 1658 1659void 1660zil_init(void) 1661{ 1662 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1663 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1664} 1665 1666void 1667zil_fini(void) 1668{ 1669 kmem_cache_destroy(zil_lwb_cache); 1670} 1671 1672void 1673zil_set_sync(zilog_t *zilog, uint64_t sync) 1674{ 1675 zilog->zl_sync = sync; 1676} 1677 1678void 1679zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1680{ 1681 zilog->zl_logbias = logbias; 1682} 1683 1684zilog_t * 1685zil_alloc(objset_t *os, zil_header_t *zh_phys) 1686{ 1687 zilog_t *zilog; 1688 1689 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1690 1691 zilog->zl_header = zh_phys; 1692 zilog->zl_os = os; 1693 zilog->zl_spa = dmu_objset_spa(os); 1694 zilog->zl_dmu_pool = dmu_objset_pool(os); 1695 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1696 zilog->zl_logbias = dmu_objset_logbias(os); 1697 zilog->zl_sync = dmu_objset_syncprop(os); 1698 zilog->zl_next_batch = 1; 1699 1700 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1701 1702 for (int i = 0; i < TXG_SIZE; i++) { 1703 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1704 MUTEX_DEFAULT, NULL); 1705 } 1706 1707 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1708 offsetof(lwb_t, lwb_node)); 1709 1710 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1711 offsetof(itx_t, itx_node)); 1712 1713 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1714 1715 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1716 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1717 1718 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1719 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1720 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1721 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1722 1723 return (zilog); 1724} 1725 1726void 1727zil_free(zilog_t *zilog) 1728{ 1729 zilog->zl_stop_sync = 1; 1730 1731 ASSERT0(zilog->zl_suspend); 1732 ASSERT0(zilog->zl_suspending); 1733 1734 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1735 list_destroy(&zilog->zl_lwb_list); 1736 1737 avl_destroy(&zilog->zl_vdev_tree); 1738 mutex_destroy(&zilog->zl_vdev_lock); 1739 1740 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1741 list_destroy(&zilog->zl_itx_commit_list); 1742 1743 for (int i = 0; i < TXG_SIZE; i++) { 1744 /* 1745 * It's possible for an itx to be generated that doesn't dirty 1746 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1747 * callback to remove the entry. We remove those here. 1748 * 1749 * Also free up the ziltest itxs. 1750 */ 1751 if (zilog->zl_itxg[i].itxg_itxs) 1752 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1753 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1754 } 1755 1756 mutex_destroy(&zilog->zl_lock); 1757 1758 cv_destroy(&zilog->zl_cv_writer); 1759 cv_destroy(&zilog->zl_cv_suspend); 1760 cv_destroy(&zilog->zl_cv_batch[0]); 1761 cv_destroy(&zilog->zl_cv_batch[1]); 1762 1763 kmem_free(zilog, sizeof (zilog_t)); 1764} 1765 1766/* 1767 * Open an intent log. 1768 */ 1769zilog_t * 1770zil_open(objset_t *os, zil_get_data_t *get_data) 1771{ 1772 zilog_t *zilog = dmu_objset_zil(os); 1773 1774 ASSERT(zilog->zl_clean_taskq == NULL); 1775 ASSERT(zilog->zl_get_data == NULL); 1776 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1777 1778 zilog->zl_get_data = get_data; 1779 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1780 2, 2, TASKQ_PREPOPULATE); 1781 1782 return (zilog); 1783} 1784 1785/* 1786 * Close an intent log. 1787 */ 1788void 1789zil_close(zilog_t *zilog) 1790{ 1791 lwb_t *lwb; 1792 uint64_t txg = 0; 1793 1794 zil_commit(zilog, 0); /* commit all itx */ 1795 1796 /* 1797 * The lwb_max_txg for the stubby lwb will reflect the last activity 1798 * for the zil. After a txg_wait_synced() on the txg we know all the 1799 * callbacks have occurred that may clean the zil. Only then can we 1800 * destroy the zl_clean_taskq. 1801 */ 1802 mutex_enter(&zilog->zl_lock); 1803 lwb = list_tail(&zilog->zl_lwb_list); 1804 if (lwb != NULL) 1805 txg = lwb->lwb_max_txg; 1806 mutex_exit(&zilog->zl_lock); 1807 if (txg) 1808 txg_wait_synced(zilog->zl_dmu_pool, txg); 1809 ASSERT(!zilog_is_dirty(zilog)); 1810 1811 taskq_destroy(zilog->zl_clean_taskq); 1812 zilog->zl_clean_taskq = NULL; 1813 zilog->zl_get_data = NULL; 1814 1815 /* 1816 * We should have only one LWB left on the list; remove it now. 1817 */ 1818 mutex_enter(&zilog->zl_lock); 1819 lwb = list_head(&zilog->zl_lwb_list); 1820 if (lwb != NULL) { 1821 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1822 list_remove(&zilog->zl_lwb_list, lwb); 1823 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1824 kmem_cache_free(zil_lwb_cache, lwb); 1825 } 1826 mutex_exit(&zilog->zl_lock); 1827} 1828 1829static char *suspend_tag = "zil suspending"; 1830 1831/* 1832 * Suspend an intent log. While in suspended mode, we still honor 1833 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1834 * On old version pools, we suspend the log briefly when taking a 1835 * snapshot so that it will have an empty intent log. 1836 * 1837 * Long holds are not really intended to be used the way we do here -- 1838 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1839 * could fail. Therefore we take pains to only put a long hold if it is 1840 * actually necessary. Fortunately, it will only be necessary if the 1841 * objset is currently mounted (or the ZVOL equivalent). In that case it 1842 * will already have a long hold, so we are not really making things any worse. 1843 * 1844 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1845 * zvol_state_t), and use their mechanism to prevent their hold from being 1846 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1847 * very little gain. 1848 * 1849 * if cookiep == NULL, this does both the suspend & resume. 1850 * Otherwise, it returns with the dataset "long held", and the cookie 1851 * should be passed into zil_resume(). 1852 */ 1853int 1854zil_suspend(const char *osname, void **cookiep) 1855{ 1856 objset_t *os; 1857 zilog_t *zilog; 1858 const zil_header_t *zh; 1859 int error; 1860 1861 error = dmu_objset_hold(osname, suspend_tag, &os); 1862 if (error != 0) 1863 return (error); 1864 zilog = dmu_objset_zil(os); 1865 1866 mutex_enter(&zilog->zl_lock); 1867 zh = zilog->zl_header; 1868 1869 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1870 mutex_exit(&zilog->zl_lock); 1871 dmu_objset_rele(os, suspend_tag); 1872 return (SET_ERROR(EBUSY)); 1873 } 1874 1875 /* 1876 * Don't put a long hold in the cases where we can avoid it. This 1877 * is when there is no cookie so we are doing a suspend & resume 1878 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1879 * for the suspend because it's already suspended, or there's no ZIL. 1880 */ 1881 if (cookiep == NULL && !zilog->zl_suspending && 1882 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1883 mutex_exit(&zilog->zl_lock); 1884 dmu_objset_rele(os, suspend_tag); 1885 return (0); 1886 } 1887 1888 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1889 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1890 1891 zilog->zl_suspend++; 1892 1893 if (zilog->zl_suspend > 1) { 1894 /* 1895 * Someone else is already suspending it. 1896 * Just wait for them to finish. 1897 */ 1898 1899 while (zilog->zl_suspending) 1900 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1901 mutex_exit(&zilog->zl_lock); 1902 1903 if (cookiep == NULL) 1904 zil_resume(os); 1905 else 1906 *cookiep = os; 1907 return (0); 1908 } 1909 1910 /* 1911 * If there is no pointer to an on-disk block, this ZIL must not 1912 * be active (e.g. filesystem not mounted), so there's nothing 1913 * to clean up. 1914 */ 1915 if (BP_IS_HOLE(&zh->zh_log)) { 1916 ASSERT(cookiep != NULL); /* fast path already handled */ 1917 1918 *cookiep = os; 1919 mutex_exit(&zilog->zl_lock); 1920 return (0); 1921 } 1922 1923 zilog->zl_suspending = B_TRUE; 1924 mutex_exit(&zilog->zl_lock); 1925 1926 zil_commit(zilog, 0); 1927 1928 zil_destroy(zilog, B_FALSE); 1929 1930 mutex_enter(&zilog->zl_lock); 1931 zilog->zl_suspending = B_FALSE; 1932 cv_broadcast(&zilog->zl_cv_suspend); 1933 mutex_exit(&zilog->zl_lock); 1934 1935 if (cookiep == NULL) 1936 zil_resume(os); 1937 else 1938 *cookiep = os; 1939 return (0); 1940} 1941 1942void 1943zil_resume(void *cookie) 1944{ 1945 objset_t *os = cookie; 1946 zilog_t *zilog = dmu_objset_zil(os); 1947 1948 mutex_enter(&zilog->zl_lock); 1949 ASSERT(zilog->zl_suspend != 0); 1950 zilog->zl_suspend--; 1951 mutex_exit(&zilog->zl_lock); 1952 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1953 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1954} 1955 1956typedef struct zil_replay_arg { 1957 zil_replay_func_t **zr_replay; 1958 void *zr_arg; 1959 boolean_t zr_byteswap; 1960 char *zr_lr; 1961} zil_replay_arg_t; 1962 1963static int 1964zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1965{ 1966 char name[MAXNAMELEN]; 1967 1968 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1969 1970 dmu_objset_name(zilog->zl_os, name); 1971 1972 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1973 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1974 (u_longlong_t)lr->lrc_seq, 1975 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1976 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1977 1978 return (error); 1979} 1980 1981static int 1982zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1983{ 1984 zil_replay_arg_t *zr = zra; 1985 const zil_header_t *zh = zilog->zl_header; 1986 uint64_t reclen = lr->lrc_reclen; 1987 uint64_t txtype = lr->lrc_txtype; 1988 int error = 0; 1989 1990 zilog->zl_replaying_seq = lr->lrc_seq; 1991 1992 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1993 return (0); 1994 1995 if (lr->lrc_txg < claim_txg) /* already committed */ 1996 return (0); 1997 1998 /* Strip case-insensitive bit, still present in log record */ 1999 txtype &= ~TX_CI; 2000 2001 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2002 return (zil_replay_error(zilog, lr, EINVAL)); 2003 2004 /* 2005 * If this record type can be logged out of order, the object 2006 * (lr_foid) may no longer exist. That's legitimate, not an error. 2007 */ 2008 if (TX_OOO(txtype)) { 2009 error = dmu_object_info(zilog->zl_os, 2010 ((lr_ooo_t *)lr)->lr_foid, NULL); 2011 if (error == ENOENT || error == EEXIST) 2012 return (0); 2013 } 2014 2015 /* 2016 * Make a copy of the data so we can revise and extend it. 2017 */ 2018 bcopy(lr, zr->zr_lr, reclen); 2019 2020 /* 2021 * If this is a TX_WRITE with a blkptr, suck in the data. 2022 */ 2023 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2024 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2025 zr->zr_lr + reclen); 2026 if (error != 0) 2027 return (zil_replay_error(zilog, lr, error)); 2028 } 2029 2030 /* 2031 * The log block containing this lr may have been byteswapped 2032 * so that we can easily examine common fields like lrc_txtype. 2033 * However, the log is a mix of different record types, and only the 2034 * replay vectors know how to byteswap their records. Therefore, if 2035 * the lr was byteswapped, undo it before invoking the replay vector. 2036 */ 2037 if (zr->zr_byteswap) 2038 byteswap_uint64_array(zr->zr_lr, reclen); 2039 2040 /* 2041 * We must now do two things atomically: replay this log record, 2042 * and update the log header sequence number to reflect the fact that 2043 * we did so. At the end of each replay function the sequence number 2044 * is updated if we are in replay mode. 2045 */ 2046 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2047 if (error != 0) { 2048 /* 2049 * The DMU's dnode layer doesn't see removes until the txg 2050 * commits, so a subsequent claim can spuriously fail with 2051 * EEXIST. So if we receive any error we try syncing out 2052 * any removes then retry the transaction. Note that we 2053 * specify B_FALSE for byteswap now, so we don't do it twice. 2054 */ 2055 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2056 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2057 if (error != 0) 2058 return (zil_replay_error(zilog, lr, error)); 2059 } 2060 return (0); 2061} 2062 2063/* ARGSUSED */ 2064static int 2065zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2066{ 2067 zilog->zl_replay_blks++; 2068 2069 return (0); 2070} 2071 2072/* 2073 * If this dataset has a non-empty intent log, replay it and destroy it. 2074 */ 2075void 2076zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2077{ 2078 zilog_t *zilog = dmu_objset_zil(os); 2079 const zil_header_t *zh = zilog->zl_header; 2080 zil_replay_arg_t zr; 2081 2082 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2083 zil_destroy(zilog, B_TRUE); 2084 return; 2085 } 2086 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 2087 2088 zr.zr_replay = replay_func; 2089 zr.zr_arg = arg; 2090 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2091 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2092 2093 /* 2094 * Wait for in-progress removes to sync before starting replay. 2095 */ 2096 txg_wait_synced(zilog->zl_dmu_pool, 0); 2097 2098 zilog->zl_replay = B_TRUE; 2099 zilog->zl_replay_time = ddi_get_lbolt(); 2100 ASSERT(zilog->zl_replay_blks == 0); 2101 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2102 zh->zh_claim_txg); 2103 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2104 2105 zil_destroy(zilog, B_FALSE); 2106 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2107 zilog->zl_replay = B_FALSE; 2108 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2109} 2110 2111boolean_t 2112zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2113{ 2114 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2115 return (B_TRUE); 2116 2117 if (zilog->zl_replay) { 2118 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2119 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2120 zilog->zl_replaying_seq; 2121 return (B_TRUE); 2122 } 2123 2124 return (B_FALSE); 2125} 2126 2127/* ARGSUSED */ 2128int 2129zil_vdev_offline(const char *osname, void *arg) 2130{ 2131 int error; 2132 2133 error = zil_suspend(osname, NULL); 2134 if (error != 0) 2135 return (SET_ERROR(EEXIST)); 2136 return (0); 2137} 2138