zil.c revision 276081
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 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 236 bcopy(lr, dst, len); 237 *end = (char *)dst + len; 238 *nbp = zilc->zc_next_blk; 239 } 240 } else { 241 char *lr = abuf->b_data; 242 uint64_t size = BP_GET_LSIZE(bp); 243 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 244 245 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 246 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 247 (zilc->zc_nused > (size - sizeof (*zilc)))) { 248 error = SET_ERROR(ECKSUM); 249 } else { 250 ASSERT3U(zilc->zc_nused, <=, 251 SPA_OLD_MAXBLOCKSIZE); 252 bcopy(lr, dst, zilc->zc_nused); 253 *end = (char *)dst + zilc->zc_nused; 254 *nbp = zilc->zc_next_blk; 255 } 256 } 257 258 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 259 } 260 261 return (error); 262} 263 264/* 265 * Read a TX_WRITE log data block. 266 */ 267static int 268zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 269{ 270 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 271 const blkptr_t *bp = &lr->lr_blkptr; 272 uint32_t aflags = ARC_WAIT; 273 arc_buf_t *abuf = NULL; 274 zbookmark_phys_t zb; 275 int error; 276 277 if (BP_IS_HOLE(bp)) { 278 if (wbuf != NULL) 279 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 280 return (0); 281 } 282 283 if (zilog->zl_header->zh_claim_txg == 0) 284 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 285 286 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 287 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 288 289 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 290 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 291 292 if (error == 0) { 293 if (wbuf != NULL) 294 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 295 (void) arc_buf_remove_ref(abuf, &abuf); 296 } 297 298 return (error); 299} 300 301/* 302 * Parse the intent log, and call parse_func for each valid record within. 303 */ 304int 305zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 306 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 307{ 308 const zil_header_t *zh = zilog->zl_header; 309 boolean_t claimed = !!zh->zh_claim_txg; 310 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 311 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 312 uint64_t max_blk_seq = 0; 313 uint64_t max_lr_seq = 0; 314 uint64_t blk_count = 0; 315 uint64_t lr_count = 0; 316 blkptr_t blk, next_blk; 317 char *lrbuf, *lrp; 318 int error = 0; 319 320 /* 321 * Old logs didn't record the maximum zh_claim_lr_seq. 322 */ 323 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 324 claim_lr_seq = UINT64_MAX; 325 326 /* 327 * Starting at the block pointed to by zh_log we read the log chain. 328 * For each block in the chain we strongly check that block to 329 * ensure its validity. We stop when an invalid block is found. 330 * For each block pointer in the chain we call parse_blk_func(). 331 * For each record in each valid block we call parse_lr_func(). 332 * If the log has been claimed, stop if we encounter a sequence 333 * number greater than the highest claimed sequence number. 334 */ 335 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 336 zil_bp_tree_init(zilog); 337 338 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 339 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 340 int reclen; 341 char *end; 342 343 if (blk_seq > claim_blk_seq) 344 break; 345 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 346 break; 347 ASSERT3U(max_blk_seq, <, blk_seq); 348 max_blk_seq = blk_seq; 349 blk_count++; 350 351 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 352 break; 353 354 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 355 if (error != 0) 356 break; 357 358 for (lrp = lrbuf; lrp < end; lrp += reclen) { 359 lr_t *lr = (lr_t *)lrp; 360 reclen = lr->lrc_reclen; 361 ASSERT3U(reclen, >=, sizeof (lr_t)); 362 if (lr->lrc_seq > claim_lr_seq) 363 goto done; 364 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 365 goto done; 366 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 367 max_lr_seq = lr->lrc_seq; 368 lr_count++; 369 } 370 } 371done: 372 zilog->zl_parse_error = error; 373 zilog->zl_parse_blk_seq = max_blk_seq; 374 zilog->zl_parse_lr_seq = max_lr_seq; 375 zilog->zl_parse_blk_count = blk_count; 376 zilog->zl_parse_lr_count = lr_count; 377 378 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 379 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 380 381 zil_bp_tree_fini(zilog); 382 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 383 384 return (error); 385} 386 387static int 388zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 389{ 390 /* 391 * Claim log block if not already committed and not already claimed. 392 * If tx == NULL, just verify that the block is claimable. 393 */ 394 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 395 zil_bp_tree_add(zilog, bp) != 0) 396 return (0); 397 398 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 399 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 400 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 401} 402 403static int 404zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 405{ 406 lr_write_t *lr = (lr_write_t *)lrc; 407 int error; 408 409 if (lrc->lrc_txtype != TX_WRITE) 410 return (0); 411 412 /* 413 * If the block is not readable, don't claim it. This can happen 414 * in normal operation when a log block is written to disk before 415 * some of the dmu_sync() blocks it points to. In this case, the 416 * transaction cannot have been committed to anyone (we would have 417 * waited for all writes to be stable first), so it is semantically 418 * correct to declare this the end of the log. 419 */ 420 if (lr->lr_blkptr.blk_birth >= first_txg && 421 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 422 return (error); 423 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 424} 425 426/* ARGSUSED */ 427static int 428zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 429{ 430 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 431 432 return (0); 433} 434 435static int 436zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 437{ 438 lr_write_t *lr = (lr_write_t *)lrc; 439 blkptr_t *bp = &lr->lr_blkptr; 440 441 /* 442 * If we previously claimed it, we need to free it. 443 */ 444 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 445 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 446 !BP_IS_HOLE(bp)) 447 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 448 449 return (0); 450} 451 452static lwb_t * 453zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 454{ 455 lwb_t *lwb; 456 457 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 458 lwb->lwb_zilog = zilog; 459 lwb->lwb_blk = *bp; 460 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 461 lwb->lwb_max_txg = txg; 462 lwb->lwb_zio = NULL; 463 lwb->lwb_tx = NULL; 464 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 465 lwb->lwb_nused = sizeof (zil_chain_t); 466 lwb->lwb_sz = BP_GET_LSIZE(bp); 467 } else { 468 lwb->lwb_nused = 0; 469 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 470 } 471 472 mutex_enter(&zilog->zl_lock); 473 list_insert_tail(&zilog->zl_lwb_list, lwb); 474 mutex_exit(&zilog->zl_lock); 475 476 return (lwb); 477} 478 479/* 480 * Called when we create in-memory log transactions so that we know 481 * to cleanup the itxs at the end of spa_sync(). 482 */ 483void 484zilog_dirty(zilog_t *zilog, uint64_t txg) 485{ 486 dsl_pool_t *dp = zilog->zl_dmu_pool; 487 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 488 489 if (dsl_dataset_is_snapshot(ds)) 490 panic("dirtying snapshot!"); 491 492 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 493 /* up the hold count until we can be written out */ 494 dmu_buf_add_ref(ds->ds_dbuf, zilog); 495 } 496} 497 498boolean_t 499zilog_is_dirty(zilog_t *zilog) 500{ 501 dsl_pool_t *dp = zilog->zl_dmu_pool; 502 503 for (int t = 0; t < TXG_SIZE; t++) { 504 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 505 return (B_TRUE); 506 } 507 return (B_FALSE); 508} 509 510/* 511 * Create an on-disk intent log. 512 */ 513static lwb_t * 514zil_create(zilog_t *zilog) 515{ 516 const zil_header_t *zh = zilog->zl_header; 517 lwb_t *lwb = NULL; 518 uint64_t txg = 0; 519 dmu_tx_t *tx = NULL; 520 blkptr_t blk; 521 int error = 0; 522 523 /* 524 * Wait for any previous destroy to complete. 525 */ 526 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 527 528 ASSERT(zh->zh_claim_txg == 0); 529 ASSERT(zh->zh_replay_seq == 0); 530 531 blk = zh->zh_log; 532 533 /* 534 * Allocate an initial log block if: 535 * - there isn't one already 536 * - the existing block is the wrong endianess 537 */ 538 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 539 tx = dmu_tx_create(zilog->zl_os); 540 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 541 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 542 txg = dmu_tx_get_txg(tx); 543 544 if (!BP_IS_HOLE(&blk)) { 545 zio_free_zil(zilog->zl_spa, txg, &blk); 546 BP_ZERO(&blk); 547 } 548 549 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 550 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 551 552 if (error == 0) 553 zil_init_log_chain(zilog, &blk); 554 } 555 556 /* 557 * Allocate a log write buffer (lwb) for the first log block. 558 */ 559 if (error == 0) 560 lwb = zil_alloc_lwb(zilog, &blk, txg); 561 562 /* 563 * If we just allocated the first log block, commit our transaction 564 * and wait for zil_sync() to stuff the block poiner into zh_log. 565 * (zh is part of the MOS, so we cannot modify it in open context.) 566 */ 567 if (tx != NULL) { 568 dmu_tx_commit(tx); 569 txg_wait_synced(zilog->zl_dmu_pool, txg); 570 } 571 572 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 573 574 return (lwb); 575} 576 577/* 578 * In one tx, free all log blocks and clear the log header. 579 * If keep_first is set, then we're replaying a log with no content. 580 * We want to keep the first block, however, so that the first 581 * synchronous transaction doesn't require a txg_wait_synced() 582 * in zil_create(). We don't need to txg_wait_synced() here either 583 * when keep_first is set, because both zil_create() and zil_destroy() 584 * will wait for any in-progress destroys to complete. 585 */ 586void 587zil_destroy(zilog_t *zilog, boolean_t keep_first) 588{ 589 const zil_header_t *zh = zilog->zl_header; 590 lwb_t *lwb; 591 dmu_tx_t *tx; 592 uint64_t txg; 593 594 /* 595 * Wait for any previous destroy to complete. 596 */ 597 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 598 599 zilog->zl_old_header = *zh; /* debugging aid */ 600 601 if (BP_IS_HOLE(&zh->zh_log)) 602 return; 603 604 tx = dmu_tx_create(zilog->zl_os); 605 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 606 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 607 txg = dmu_tx_get_txg(tx); 608 609 mutex_enter(&zilog->zl_lock); 610 611 ASSERT3U(zilog->zl_destroy_txg, <, txg); 612 zilog->zl_destroy_txg = txg; 613 zilog->zl_keep_first = keep_first; 614 615 if (!list_is_empty(&zilog->zl_lwb_list)) { 616 ASSERT(zh->zh_claim_txg == 0); 617 VERIFY(!keep_first); 618 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 619 list_remove(&zilog->zl_lwb_list, lwb); 620 if (lwb->lwb_buf != NULL) 621 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 622 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 623 kmem_cache_free(zil_lwb_cache, lwb); 624 } 625 } else if (!keep_first) { 626 zil_destroy_sync(zilog, tx); 627 } 628 mutex_exit(&zilog->zl_lock); 629 630 dmu_tx_commit(tx); 631} 632 633void 634zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 635{ 636 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 637 (void) zil_parse(zilog, zil_free_log_block, 638 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 639} 640 641int 642zil_claim(const char *osname, void *txarg) 643{ 644 dmu_tx_t *tx = txarg; 645 uint64_t first_txg = dmu_tx_get_txg(tx); 646 zilog_t *zilog; 647 zil_header_t *zh; 648 objset_t *os; 649 int error; 650 651 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); 652 if (error != 0) { 653 /* 654 * EBUSY indicates that the objset is inconsistent, in which 655 * case it can not have a ZIL. 656 */ 657 if (error != EBUSY) { 658 cmn_err(CE_WARN, "can't open objset for %s, error %u", 659 osname, error); 660 } 661 return (0); 662 } 663 664 zilog = dmu_objset_zil(os); 665 zh = zil_header_in_syncing_context(zilog); 666 667 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 668 if (!BP_IS_HOLE(&zh->zh_log)) 669 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 670 BP_ZERO(&zh->zh_log); 671 dsl_dataset_dirty(dmu_objset_ds(os), tx); 672 dmu_objset_disown(os, FTAG); 673 return (0); 674 } 675 676 /* 677 * Claim all log blocks if we haven't already done so, and remember 678 * the highest claimed sequence number. This ensures that if we can 679 * read only part of the log now (e.g. due to a missing device), 680 * but we can read the entire log later, we will not try to replay 681 * or destroy beyond the last block we successfully claimed. 682 */ 683 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 684 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 685 (void) zil_parse(zilog, zil_claim_log_block, 686 zil_claim_log_record, tx, first_txg); 687 zh->zh_claim_txg = first_txg; 688 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 689 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 690 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 691 zh->zh_flags |= ZIL_REPLAY_NEEDED; 692 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 693 dsl_dataset_dirty(dmu_objset_ds(os), tx); 694 } 695 696 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 697 dmu_objset_disown(os, FTAG); 698 return (0); 699} 700 701/* 702 * Check the log by walking the log chain. 703 * Checksum errors are ok as they indicate the end of the chain. 704 * Any other error (no device or read failure) returns an error. 705 */ 706int 707zil_check_log_chain(const char *osname, void *tx) 708{ 709 zilog_t *zilog; 710 objset_t *os; 711 blkptr_t *bp; 712 int error; 713 714 ASSERT(tx == NULL); 715 716 error = dmu_objset_hold(osname, FTAG, &os); 717 if (error != 0) { 718 cmn_err(CE_WARN, "can't open objset for %s", osname); 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 dmu_objset_rele(os, FTAG); 743 return (0); 744 } 745 } 746 747 /* 748 * Because tx == NULL, zil_claim_log_block() will not actually claim 749 * any blocks, but just determine whether it is possible to do so. 750 * In addition to checking the log chain, zil_claim_log_block() 751 * will invoke zio_claim() with a done func of spa_claim_notify(), 752 * which will update spa_max_claim_txg. See spa_load() for details. 753 */ 754 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 755 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 756 757 dmu_objset_rele(os, FTAG); 758 759 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 760} 761 762static int 763zil_vdev_compare(const void *x1, const void *x2) 764{ 765 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 766 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 767 768 if (v1 < v2) 769 return (-1); 770 if (v1 > v2) 771 return (1); 772 773 return (0); 774} 775 776void 777zil_add_block(zilog_t *zilog, const blkptr_t *bp) 778{ 779 avl_tree_t *t = &zilog->zl_vdev_tree; 780 avl_index_t where; 781 zil_vdev_node_t *zv, zvsearch; 782 int ndvas = BP_GET_NDVAS(bp); 783 int i; 784 785 if (zfs_nocacheflush) 786 return; 787 788 ASSERT(zilog->zl_writer); 789 790 /* 791 * Even though we're zl_writer, we still need a lock because the 792 * zl_get_data() callbacks may have dmu_sync() done callbacks 793 * that will run concurrently. 794 */ 795 mutex_enter(&zilog->zl_vdev_lock); 796 for (i = 0; i < ndvas; i++) { 797 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 798 if (avl_find(t, &zvsearch, &where) == NULL) { 799 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 800 zv->zv_vdev = zvsearch.zv_vdev; 801 avl_insert(t, zv, where); 802 } 803 } 804 mutex_exit(&zilog->zl_vdev_lock); 805} 806 807static void 808zil_flush_vdevs(zilog_t *zilog) 809{ 810 spa_t *spa = zilog->zl_spa; 811 avl_tree_t *t = &zilog->zl_vdev_tree; 812 void *cookie = NULL; 813 zil_vdev_node_t *zv; 814 zio_t *zio; 815 816 ASSERT(zilog->zl_writer); 817 818 /* 819 * We don't need zl_vdev_lock here because we're the zl_writer, 820 * and all zl_get_data() callbacks are done. 821 */ 822 if (avl_numnodes(t) == 0) 823 return; 824 825 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 826 827 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 828 829 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 830 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 831 if (vd != NULL) 832 zio_flush(zio, vd); 833 kmem_free(zv, sizeof (*zv)); 834 } 835 836 /* 837 * Wait for all the flushes to complete. Not all devices actually 838 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 839 */ 840 (void) zio_wait(zio); 841 842 spa_config_exit(spa, SCL_STATE, FTAG); 843} 844 845/* 846 * Function called when a log block write completes 847 */ 848static void 849zil_lwb_write_done(zio_t *zio) 850{ 851 lwb_t *lwb = zio->io_private; 852 zilog_t *zilog = lwb->lwb_zilog; 853 dmu_tx_t *tx = lwb->lwb_tx; 854 855 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 856 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 857 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 858 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 859 ASSERT(!BP_IS_GANG(zio->io_bp)); 860 ASSERT(!BP_IS_HOLE(zio->io_bp)); 861 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 862 863 /* 864 * Ensure the lwb buffer pointer is cleared before releasing 865 * the txg. If we have had an allocation failure and 866 * the txg is waiting to sync then we want want zil_sync() 867 * to remove the lwb so that it's not picked up as the next new 868 * one in zil_commit_writer(). zil_sync() will only remove 869 * the lwb if lwb_buf is null. 870 */ 871 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 872 mutex_enter(&zilog->zl_lock); 873 lwb->lwb_buf = NULL; 874 lwb->lwb_tx = NULL; 875 mutex_exit(&zilog->zl_lock); 876 877 /* 878 * Now that we've written this log block, we have a stable pointer 879 * to the next block in the chain, so it's OK to let the txg in 880 * which we allocated the next block sync. 881 */ 882 dmu_tx_commit(tx); 883} 884 885/* 886 * Initialize the io for a log block. 887 */ 888static void 889zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 890{ 891 zbookmark_phys_t zb; 892 893 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 894 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 895 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 896 897 if (zilog->zl_root_zio == NULL) { 898 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 899 ZIO_FLAG_CANFAIL); 900 } 901 if (lwb->lwb_zio == NULL) { 902 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 903 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 904 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 905 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 906 } 907} 908 909/* 910 * Define a limited set of intent log block sizes. 911 * 912 * These must be a multiple of 4KB. Note only the amount used (again 913 * aligned to 4KB) actually gets written. However, we can't always just 914 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 915 */ 916uint64_t zil_block_buckets[] = { 917 4096, /* non TX_WRITE */ 918 8192+4096, /* data base */ 919 32*1024 + 4096, /* NFS writes */ 920 UINT64_MAX 921}; 922 923/* 924 * Use the slog as long as the logbias is 'latency' and the current commit size 925 * is less than the limit or the total list size is less than 2X the limit. 926 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 927 */ 928uint64_t zil_slog_limit = 1024 * 1024; 929#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 930 (((zilog)->zl_cur_used < zil_slog_limit) || \ 931 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 932 933/* 934 * Start a log block write and advance to the next log block. 935 * Calls are serialized. 936 */ 937static lwb_t * 938zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 939{ 940 lwb_t *nlwb = NULL; 941 zil_chain_t *zilc; 942 spa_t *spa = zilog->zl_spa; 943 blkptr_t *bp; 944 dmu_tx_t *tx; 945 uint64_t txg; 946 uint64_t zil_blksz, wsz; 947 int i, error; 948 949 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 950 zilc = (zil_chain_t *)lwb->lwb_buf; 951 bp = &zilc->zc_next_blk; 952 } else { 953 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 954 bp = &zilc->zc_next_blk; 955 } 956 957 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 958 959 /* 960 * Allocate the next block and save its address in this block 961 * before writing it in order to establish the log chain. 962 * Note that if the allocation of nlwb synced before we wrote 963 * the block that points at it (lwb), we'd leak it if we crashed. 964 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 965 * We dirty the dataset to ensure that zil_sync() will be called 966 * to clean up in the event of allocation failure or I/O failure. 967 */ 968 tx = dmu_tx_create(zilog->zl_os); 969 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 970 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 971 txg = dmu_tx_get_txg(tx); 972 973 lwb->lwb_tx = tx; 974 975 /* 976 * Log blocks are pre-allocated. Here we select the size of the next 977 * block, based on size used in the last block. 978 * - first find the smallest bucket that will fit the block from a 979 * limited set of block sizes. This is because it's faster to write 980 * blocks allocated from the same metaslab as they are adjacent or 981 * close. 982 * - next find the maximum from the new suggested size and an array of 983 * previous sizes. This lessens a picket fence effect of wrongly 984 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 985 * requests. 986 * 987 * Note we only write what is used, but we can't just allocate 988 * the maximum block size because we can exhaust the available 989 * pool log space. 990 */ 991 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 992 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 993 continue; 994 zil_blksz = zil_block_buckets[i]; 995 if (zil_blksz == UINT64_MAX) 996 zil_blksz = SPA_OLD_MAXBLOCKSIZE; 997 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 998 for (i = 0; i < ZIL_PREV_BLKS; i++) 999 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1000 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1001 1002 BP_ZERO(bp); 1003 /* pass the old blkptr in order to spread log blocks across devs */ 1004 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 1005 USE_SLOG(zilog)); 1006 if (error == 0) { 1007 ASSERT3U(bp->blk_birth, ==, txg); 1008 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1009 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1010 1011 /* 1012 * Allocate a new log write buffer (lwb). 1013 */ 1014 nlwb = zil_alloc_lwb(zilog, bp, txg); 1015 1016 /* Record the block for later vdev flushing */ 1017 zil_add_block(zilog, &lwb->lwb_blk); 1018 } 1019 1020 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1021 /* For Slim ZIL only write what is used. */ 1022 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1023 ASSERT3U(wsz, <=, lwb->lwb_sz); 1024 zio_shrink(lwb->lwb_zio, wsz); 1025 1026 } else { 1027 wsz = lwb->lwb_sz; 1028 } 1029 1030 zilc->zc_pad = 0; 1031 zilc->zc_nused = lwb->lwb_nused; 1032 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1033 1034 /* 1035 * clear unused data for security 1036 */ 1037 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1038 1039 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1040 1041 /* 1042 * If there was an allocation failure then nlwb will be null which 1043 * forces a txg_wait_synced(). 1044 */ 1045 return (nlwb); 1046} 1047 1048static lwb_t * 1049zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1050{ 1051 lr_t *lrc = &itx->itx_lr; /* common log record */ 1052 lr_write_t *lrw = (lr_write_t *)lrc; 1053 char *lr_buf; 1054 uint64_t txg = lrc->lrc_txg; 1055 uint64_t reclen = lrc->lrc_reclen; 1056 uint64_t dlen = 0; 1057 1058 if (lwb == NULL) 1059 return (NULL); 1060 1061 ASSERT(lwb->lwb_buf != NULL); 1062 ASSERT(zilog_is_dirty(zilog) || 1063 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1064 1065 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1066 dlen = P2ROUNDUP_TYPED( 1067 lrw->lr_length, sizeof (uint64_t), uint64_t); 1068 1069 zilog->zl_cur_used += (reclen + dlen); 1070 1071 zil_lwb_write_init(zilog, lwb); 1072 1073 /* 1074 * If this record won't fit in the current log block, start a new one. 1075 */ 1076 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1077 lwb = zil_lwb_write_start(zilog, lwb); 1078 if (lwb == NULL) 1079 return (NULL); 1080 zil_lwb_write_init(zilog, lwb); 1081 ASSERT(LWB_EMPTY(lwb)); 1082 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1083 txg_wait_synced(zilog->zl_dmu_pool, txg); 1084 return (lwb); 1085 } 1086 } 1087 1088 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1089 bcopy(lrc, lr_buf, reclen); 1090 lrc = (lr_t *)lr_buf; 1091 lrw = (lr_write_t *)lrc; 1092 1093 /* 1094 * If it's a write, fetch the data or get its blkptr as appropriate. 1095 */ 1096 if (lrc->lrc_txtype == TX_WRITE) { 1097 if (txg > spa_freeze_txg(zilog->zl_spa)) 1098 txg_wait_synced(zilog->zl_dmu_pool, txg); 1099 if (itx->itx_wr_state != WR_COPIED) { 1100 char *dbuf; 1101 int error; 1102 1103 if (dlen) { 1104 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1105 dbuf = lr_buf + reclen; 1106 lrw->lr_common.lrc_reclen += dlen; 1107 } else { 1108 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1109 dbuf = NULL; 1110 } 1111 error = zilog->zl_get_data( 1112 itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1113 if (error == EIO) { 1114 txg_wait_synced(zilog->zl_dmu_pool, txg); 1115 return (lwb); 1116 } 1117 if (error != 0) { 1118 ASSERT(error == ENOENT || error == EEXIST || 1119 error == EALREADY); 1120 return (lwb); 1121 } 1122 } 1123 } 1124 1125 /* 1126 * We're actually making an entry, so update lrc_seq to be the 1127 * log record sequence number. Note that this is generally not 1128 * equal to the itx sequence number because not all transactions 1129 * are synchronous, and sometimes spa_sync() gets there first. 1130 */ 1131 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1132 lwb->lwb_nused += reclen + dlen; 1133 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1134 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1135 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1136 1137 return (lwb); 1138} 1139 1140itx_t * 1141zil_itx_create(uint64_t txtype, size_t lrsize) 1142{ 1143 itx_t *itx; 1144 1145 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1146 1147 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1148 itx->itx_lr.lrc_txtype = txtype; 1149 itx->itx_lr.lrc_reclen = lrsize; 1150 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1151 itx->itx_lr.lrc_seq = 0; /* defensive */ 1152 itx->itx_sync = B_TRUE; /* default is synchronous */ 1153 1154 return (itx); 1155} 1156 1157void 1158zil_itx_destroy(itx_t *itx) 1159{ 1160 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1161} 1162 1163/* 1164 * Free up the sync and async itxs. The itxs_t has already been detached 1165 * so no locks are needed. 1166 */ 1167static void 1168zil_itxg_clean(itxs_t *itxs) 1169{ 1170 itx_t *itx; 1171 list_t *list; 1172 avl_tree_t *t; 1173 void *cookie; 1174 itx_async_node_t *ian; 1175 1176 list = &itxs->i_sync_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 1183 cookie = NULL; 1184 t = &itxs->i_async_tree; 1185 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1186 list = &ian->ia_list; 1187 while ((itx = list_head(list)) != NULL) { 1188 list_remove(list, itx); 1189 kmem_free(itx, offsetof(itx_t, itx_lr) + 1190 itx->itx_lr.lrc_reclen); 1191 } 1192 list_destroy(list); 1193 kmem_free(ian, sizeof (itx_async_node_t)); 1194 } 1195 avl_destroy(t); 1196 1197 kmem_free(itxs, sizeof (itxs_t)); 1198} 1199 1200static int 1201zil_aitx_compare(const void *x1, const void *x2) 1202{ 1203 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1204 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1205 1206 if (o1 < o2) 1207 return (-1); 1208 if (o1 > o2) 1209 return (1); 1210 1211 return (0); 1212} 1213 1214/* 1215 * Remove all async itx with the given oid. 1216 */ 1217static void 1218zil_remove_async(zilog_t *zilog, uint64_t oid) 1219{ 1220 uint64_t otxg, txg; 1221 itx_async_node_t *ian; 1222 avl_tree_t *t; 1223 avl_index_t where; 1224 list_t clean_list; 1225 itx_t *itx; 1226 1227 ASSERT(oid != 0); 1228 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1229 1230 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1231 otxg = ZILTEST_TXG; 1232 else 1233 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1234 1235 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1236 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1237 1238 mutex_enter(&itxg->itxg_lock); 1239 if (itxg->itxg_txg != txg) { 1240 mutex_exit(&itxg->itxg_lock); 1241 continue; 1242 } 1243 1244 /* 1245 * Locate the object node and append its list. 1246 */ 1247 t = &itxg->itxg_itxs->i_async_tree; 1248 ian = avl_find(t, &oid, &where); 1249 if (ian != NULL) 1250 list_move_tail(&clean_list, &ian->ia_list); 1251 mutex_exit(&itxg->itxg_lock); 1252 } 1253 while ((itx = list_head(&clean_list)) != NULL) { 1254 list_remove(&clean_list, itx); 1255 kmem_free(itx, offsetof(itx_t, itx_lr) + 1256 itx->itx_lr.lrc_reclen); 1257 } 1258 list_destroy(&clean_list); 1259} 1260 1261void 1262zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1263{ 1264 uint64_t txg; 1265 itxg_t *itxg; 1266 itxs_t *itxs, *clean = NULL; 1267 1268 /* 1269 * Object ids can be re-instantiated in the next txg so 1270 * remove any async transactions to avoid future leaks. 1271 * This can happen if a fsync occurs on the re-instantiated 1272 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1273 * the new file data and flushes a write record for the old object. 1274 */ 1275 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1276 zil_remove_async(zilog, itx->itx_oid); 1277 1278 /* 1279 * Ensure the data of a renamed file is committed before the rename. 1280 */ 1281 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1282 zil_async_to_sync(zilog, itx->itx_oid); 1283 1284 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1285 txg = ZILTEST_TXG; 1286 else 1287 txg = dmu_tx_get_txg(tx); 1288 1289 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1290 mutex_enter(&itxg->itxg_lock); 1291 itxs = itxg->itxg_itxs; 1292 if (itxg->itxg_txg != txg) { 1293 if (itxs != NULL) { 1294 /* 1295 * The zil_clean callback hasn't got around to cleaning 1296 * this itxg. Save the itxs for release below. 1297 * This should be rare. 1298 */ 1299 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1300 itxg->itxg_sod = 0; 1301 clean = itxg->itxg_itxs; 1302 } 1303 ASSERT(itxg->itxg_sod == 0); 1304 itxg->itxg_txg = txg; 1305 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1306 1307 list_create(&itxs->i_sync_list, sizeof (itx_t), 1308 offsetof(itx_t, itx_node)); 1309 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1310 sizeof (itx_async_node_t), 1311 offsetof(itx_async_node_t, ia_node)); 1312 } 1313 if (itx->itx_sync) { 1314 list_insert_tail(&itxs->i_sync_list, itx); 1315 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1316 itxg->itxg_sod += itx->itx_sod; 1317 } else { 1318 avl_tree_t *t = &itxs->i_async_tree; 1319 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1320 itx_async_node_t *ian; 1321 avl_index_t where; 1322 1323 ian = avl_find(t, &foid, &where); 1324 if (ian == NULL) { 1325 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1326 list_create(&ian->ia_list, sizeof (itx_t), 1327 offsetof(itx_t, itx_node)); 1328 ian->ia_foid = foid; 1329 avl_insert(t, ian, where); 1330 } 1331 list_insert_tail(&ian->ia_list, itx); 1332 } 1333 1334 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1335 zilog_dirty(zilog, txg); 1336 mutex_exit(&itxg->itxg_lock); 1337 1338 /* Release the old itxs now we've dropped the lock */ 1339 if (clean != NULL) 1340 zil_itxg_clean(clean); 1341} 1342 1343/* 1344 * If there are any in-memory intent log transactions which have now been 1345 * synced then start up a taskq to free them. We should only do this after we 1346 * have written out the uberblocks (i.e. txg has been comitted) so that 1347 * don't inadvertently clean out in-memory log records that would be required 1348 * by zil_commit(). 1349 */ 1350void 1351zil_clean(zilog_t *zilog, uint64_t synced_txg) 1352{ 1353 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1354 itxs_t *clean_me; 1355 1356 mutex_enter(&itxg->itxg_lock); 1357 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1358 mutex_exit(&itxg->itxg_lock); 1359 return; 1360 } 1361 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1362 ASSERT(itxg->itxg_txg != 0); 1363 ASSERT(zilog->zl_clean_taskq != NULL); 1364 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1365 itxg->itxg_sod = 0; 1366 clean_me = itxg->itxg_itxs; 1367 itxg->itxg_itxs = NULL; 1368 itxg->itxg_txg = 0; 1369 mutex_exit(&itxg->itxg_lock); 1370 /* 1371 * Preferably start a task queue to free up the old itxs but 1372 * if taskq_dispatch can't allocate resources to do that then 1373 * free it in-line. This should be rare. Note, using TQ_SLEEP 1374 * created a bad performance problem. 1375 */ 1376 if (taskq_dispatch(zilog->zl_clean_taskq, 1377 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1378 zil_itxg_clean(clean_me); 1379} 1380 1381/* 1382 * Get the list of itxs to commit into zl_itx_commit_list. 1383 */ 1384static void 1385zil_get_commit_list(zilog_t *zilog) 1386{ 1387 uint64_t otxg, txg; 1388 list_t *commit_list = &zilog->zl_itx_commit_list; 1389 uint64_t push_sod = 0; 1390 1391 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1392 otxg = ZILTEST_TXG; 1393 else 1394 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1395 1396 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1397 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1398 1399 mutex_enter(&itxg->itxg_lock); 1400 if (itxg->itxg_txg != txg) { 1401 mutex_exit(&itxg->itxg_lock); 1402 continue; 1403 } 1404 1405 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1406 push_sod += itxg->itxg_sod; 1407 itxg->itxg_sod = 0; 1408 1409 mutex_exit(&itxg->itxg_lock); 1410 } 1411 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1412} 1413 1414/* 1415 * Move the async itxs for a specified object to commit into sync lists. 1416 */ 1417static void 1418zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1419{ 1420 uint64_t otxg, txg; 1421 itx_async_node_t *ian; 1422 avl_tree_t *t; 1423 avl_index_t where; 1424 1425 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1426 otxg = ZILTEST_TXG; 1427 else 1428 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1429 1430 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1431 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1432 1433 mutex_enter(&itxg->itxg_lock); 1434 if (itxg->itxg_txg != txg) { 1435 mutex_exit(&itxg->itxg_lock); 1436 continue; 1437 } 1438 1439 /* 1440 * If a foid is specified then find that node and append its 1441 * list. Otherwise walk the tree appending all the lists 1442 * to the sync list. We add to the end rather than the 1443 * beginning to ensure the create has happened. 1444 */ 1445 t = &itxg->itxg_itxs->i_async_tree; 1446 if (foid != 0) { 1447 ian = avl_find(t, &foid, &where); 1448 if (ian != NULL) { 1449 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1450 &ian->ia_list); 1451 } 1452 } else { 1453 void *cookie = NULL; 1454 1455 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1456 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1457 &ian->ia_list); 1458 list_destroy(&ian->ia_list); 1459 kmem_free(ian, sizeof (itx_async_node_t)); 1460 } 1461 } 1462 mutex_exit(&itxg->itxg_lock); 1463 } 1464} 1465 1466static void 1467zil_commit_writer(zilog_t *zilog) 1468{ 1469 uint64_t txg; 1470 itx_t *itx; 1471 lwb_t *lwb; 1472 spa_t *spa = zilog->zl_spa; 1473 int error = 0; 1474 1475 ASSERT(zilog->zl_root_zio == NULL); 1476 1477 mutex_exit(&zilog->zl_lock); 1478 1479 zil_get_commit_list(zilog); 1480 1481 /* 1482 * Return if there's nothing to commit before we dirty the fs by 1483 * calling zil_create(). 1484 */ 1485 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1486 mutex_enter(&zilog->zl_lock); 1487 return; 1488 } 1489 1490 if (zilog->zl_suspend) { 1491 lwb = NULL; 1492 } else { 1493 lwb = list_tail(&zilog->zl_lwb_list); 1494 if (lwb == NULL) 1495 lwb = zil_create(zilog); 1496 } 1497 1498 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1499 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1500 txg = itx->itx_lr.lrc_txg; 1501 ASSERT(txg); 1502 1503 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1504 lwb = zil_lwb_commit(zilog, itx, lwb); 1505 list_remove(&zilog->zl_itx_commit_list, itx); 1506 kmem_free(itx, offsetof(itx_t, itx_lr) 1507 + itx->itx_lr.lrc_reclen); 1508 } 1509 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1510 1511 /* write the last block out */ 1512 if (lwb != NULL && lwb->lwb_zio != NULL) 1513 lwb = zil_lwb_write_start(zilog, lwb); 1514 1515 zilog->zl_cur_used = 0; 1516 1517 /* 1518 * Wait if necessary for the log blocks to be on stable storage. 1519 */ 1520 if (zilog->zl_root_zio) { 1521 error = zio_wait(zilog->zl_root_zio); 1522 zilog->zl_root_zio = NULL; 1523 zil_flush_vdevs(zilog); 1524 } 1525 1526 if (error || lwb == NULL) 1527 txg_wait_synced(zilog->zl_dmu_pool, 0); 1528 1529 mutex_enter(&zilog->zl_lock); 1530 1531 /* 1532 * Remember the highest committed log sequence number for ztest. 1533 * We only update this value when all the log writes succeeded, 1534 * because ztest wants to ASSERT that it got the whole log chain. 1535 */ 1536 if (error == 0 && lwb != NULL) 1537 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1538} 1539 1540/* 1541 * Commit zfs transactions to stable storage. 1542 * If foid is 0 push out all transactions, otherwise push only those 1543 * for that object or might reference that object. 1544 * 1545 * itxs are committed in batches. In a heavily stressed zil there will be 1546 * a commit writer thread who is writing out a bunch of itxs to the log 1547 * for a set of committing threads (cthreads) in the same batch as the writer. 1548 * Those cthreads are all waiting on the same cv for that batch. 1549 * 1550 * There will also be a different and growing batch of threads that are 1551 * waiting to commit (qthreads). When the committing batch completes 1552 * a transition occurs such that the cthreads exit and the qthreads become 1553 * cthreads. One of the new cthreads becomes the writer thread for the 1554 * batch. Any new threads arriving become new qthreads. 1555 * 1556 * Only 2 condition variables are needed and there's no transition 1557 * between the two cvs needed. They just flip-flop between qthreads 1558 * and cthreads. 1559 * 1560 * Using this scheme we can efficiently wakeup up only those threads 1561 * that have been committed. 1562 */ 1563void 1564zil_commit(zilog_t *zilog, uint64_t foid) 1565{ 1566 uint64_t mybatch; 1567 1568 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1569 return; 1570 1571 /* move the async itxs for the foid to the sync queues */ 1572 zil_async_to_sync(zilog, foid); 1573 1574 mutex_enter(&zilog->zl_lock); 1575 mybatch = zilog->zl_next_batch; 1576 while (zilog->zl_writer) { 1577 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1578 if (mybatch <= zilog->zl_com_batch) { 1579 mutex_exit(&zilog->zl_lock); 1580 return; 1581 } 1582 } 1583 1584 zilog->zl_next_batch++; 1585 zilog->zl_writer = B_TRUE; 1586 zil_commit_writer(zilog); 1587 zilog->zl_com_batch = mybatch; 1588 zilog->zl_writer = B_FALSE; 1589 mutex_exit(&zilog->zl_lock); 1590 1591 /* wake up one thread to become the next writer */ 1592 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1593 1594 /* wake up all threads waiting for this batch to be committed */ 1595 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1596} 1597 1598/* 1599 * Called in syncing context to free committed log blocks and update log header. 1600 */ 1601void 1602zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1603{ 1604 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1605 uint64_t txg = dmu_tx_get_txg(tx); 1606 spa_t *spa = zilog->zl_spa; 1607 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1608 lwb_t *lwb; 1609 1610 /* 1611 * We don't zero out zl_destroy_txg, so make sure we don't try 1612 * to destroy it twice. 1613 */ 1614 if (spa_sync_pass(spa) != 1) 1615 return; 1616 1617 mutex_enter(&zilog->zl_lock); 1618 1619 ASSERT(zilog->zl_stop_sync == 0); 1620 1621 if (*replayed_seq != 0) { 1622 ASSERT(zh->zh_replay_seq < *replayed_seq); 1623 zh->zh_replay_seq = *replayed_seq; 1624 *replayed_seq = 0; 1625 } 1626 1627 if (zilog->zl_destroy_txg == txg) { 1628 blkptr_t blk = zh->zh_log; 1629 1630 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1631 1632 bzero(zh, sizeof (zil_header_t)); 1633 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1634 1635 if (zilog->zl_keep_first) { 1636 /* 1637 * If this block was part of log chain that couldn't 1638 * be claimed because a device was missing during 1639 * zil_claim(), but that device later returns, 1640 * then this block could erroneously appear valid. 1641 * To guard against this, assign a new GUID to the new 1642 * log chain so it doesn't matter what blk points to. 1643 */ 1644 zil_init_log_chain(zilog, &blk); 1645 zh->zh_log = blk; 1646 } 1647 } 1648 1649 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1650 zh->zh_log = lwb->lwb_blk; 1651 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1652 break; 1653 list_remove(&zilog->zl_lwb_list, lwb); 1654 zio_free_zil(spa, txg, &lwb->lwb_blk); 1655 kmem_cache_free(zil_lwb_cache, lwb); 1656 1657 /* 1658 * If we don't have anything left in the lwb list then 1659 * we've had an allocation failure and we need to zero 1660 * out the zil_header blkptr so that we don't end 1661 * up freeing the same block twice. 1662 */ 1663 if (list_head(&zilog->zl_lwb_list) == NULL) 1664 BP_ZERO(&zh->zh_log); 1665 } 1666 mutex_exit(&zilog->zl_lock); 1667} 1668 1669void 1670zil_init(void) 1671{ 1672 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1673 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1674} 1675 1676void 1677zil_fini(void) 1678{ 1679 kmem_cache_destroy(zil_lwb_cache); 1680} 1681 1682void 1683zil_set_sync(zilog_t *zilog, uint64_t sync) 1684{ 1685 zilog->zl_sync = sync; 1686} 1687 1688void 1689zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1690{ 1691 zilog->zl_logbias = logbias; 1692} 1693 1694zilog_t * 1695zil_alloc(objset_t *os, zil_header_t *zh_phys) 1696{ 1697 zilog_t *zilog; 1698 1699 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1700 1701 zilog->zl_header = zh_phys; 1702 zilog->zl_os = os; 1703 zilog->zl_spa = dmu_objset_spa(os); 1704 zilog->zl_dmu_pool = dmu_objset_pool(os); 1705 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1706 zilog->zl_logbias = dmu_objset_logbias(os); 1707 zilog->zl_sync = dmu_objset_syncprop(os); 1708 zilog->zl_next_batch = 1; 1709 1710 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1711 1712 for (int i = 0; i < TXG_SIZE; i++) { 1713 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1714 MUTEX_DEFAULT, NULL); 1715 } 1716 1717 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1718 offsetof(lwb_t, lwb_node)); 1719 1720 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1721 offsetof(itx_t, itx_node)); 1722 1723 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1724 1725 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1726 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1727 1728 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1729 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1730 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1731 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1732 1733 return (zilog); 1734} 1735 1736void 1737zil_free(zilog_t *zilog) 1738{ 1739 zilog->zl_stop_sync = 1; 1740 1741 ASSERT0(zilog->zl_suspend); 1742 ASSERT0(zilog->zl_suspending); 1743 1744 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1745 list_destroy(&zilog->zl_lwb_list); 1746 1747 avl_destroy(&zilog->zl_vdev_tree); 1748 mutex_destroy(&zilog->zl_vdev_lock); 1749 1750 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1751 list_destroy(&zilog->zl_itx_commit_list); 1752 1753 for (int i = 0; i < TXG_SIZE; i++) { 1754 /* 1755 * It's possible for an itx to be generated that doesn't dirty 1756 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1757 * callback to remove the entry. We remove those here. 1758 * 1759 * Also free up the ziltest itxs. 1760 */ 1761 if (zilog->zl_itxg[i].itxg_itxs) 1762 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1763 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1764 } 1765 1766 mutex_destroy(&zilog->zl_lock); 1767 1768 cv_destroy(&zilog->zl_cv_writer); 1769 cv_destroy(&zilog->zl_cv_suspend); 1770 cv_destroy(&zilog->zl_cv_batch[0]); 1771 cv_destroy(&zilog->zl_cv_batch[1]); 1772 1773 kmem_free(zilog, sizeof (zilog_t)); 1774} 1775 1776/* 1777 * Open an intent log. 1778 */ 1779zilog_t * 1780zil_open(objset_t *os, zil_get_data_t *get_data) 1781{ 1782 zilog_t *zilog = dmu_objset_zil(os); 1783 1784 ASSERT(zilog->zl_clean_taskq == NULL); 1785 ASSERT(zilog->zl_get_data == NULL); 1786 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1787 1788 zilog->zl_get_data = get_data; 1789 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1790 2, 2, TASKQ_PREPOPULATE); 1791 1792 return (zilog); 1793} 1794 1795/* 1796 * Close an intent log. 1797 */ 1798void 1799zil_close(zilog_t *zilog) 1800{ 1801 lwb_t *lwb; 1802 uint64_t txg = 0; 1803 1804 zil_commit(zilog, 0); /* commit all itx */ 1805 1806 /* 1807 * The lwb_max_txg for the stubby lwb will reflect the last activity 1808 * for the zil. After a txg_wait_synced() on the txg we know all the 1809 * callbacks have occurred that may clean the zil. Only then can we 1810 * destroy the zl_clean_taskq. 1811 */ 1812 mutex_enter(&zilog->zl_lock); 1813 lwb = list_tail(&zilog->zl_lwb_list); 1814 if (lwb != NULL) 1815 txg = lwb->lwb_max_txg; 1816 mutex_exit(&zilog->zl_lock); 1817 if (txg) 1818 txg_wait_synced(zilog->zl_dmu_pool, txg); 1819 ASSERT(!zilog_is_dirty(zilog)); 1820 1821 taskq_destroy(zilog->zl_clean_taskq); 1822 zilog->zl_clean_taskq = NULL; 1823 zilog->zl_get_data = NULL; 1824 1825 /* 1826 * We should have only one LWB left on the list; remove it now. 1827 */ 1828 mutex_enter(&zilog->zl_lock); 1829 lwb = list_head(&zilog->zl_lwb_list); 1830 if (lwb != NULL) { 1831 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1832 list_remove(&zilog->zl_lwb_list, lwb); 1833 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1834 kmem_cache_free(zil_lwb_cache, lwb); 1835 } 1836 mutex_exit(&zilog->zl_lock); 1837} 1838 1839static char *suspend_tag = "zil suspending"; 1840 1841/* 1842 * Suspend an intent log. While in suspended mode, we still honor 1843 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1844 * On old version pools, we suspend the log briefly when taking a 1845 * snapshot so that it will have an empty intent log. 1846 * 1847 * Long holds are not really intended to be used the way we do here -- 1848 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1849 * could fail. Therefore we take pains to only put a long hold if it is 1850 * actually necessary. Fortunately, it will only be necessary if the 1851 * objset is currently mounted (or the ZVOL equivalent). In that case it 1852 * will already have a long hold, so we are not really making things any worse. 1853 * 1854 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1855 * zvol_state_t), and use their mechanism to prevent their hold from being 1856 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1857 * very little gain. 1858 * 1859 * if cookiep == NULL, this does both the suspend & resume. 1860 * Otherwise, it returns with the dataset "long held", and the cookie 1861 * should be passed into zil_resume(). 1862 */ 1863int 1864zil_suspend(const char *osname, void **cookiep) 1865{ 1866 objset_t *os; 1867 zilog_t *zilog; 1868 const zil_header_t *zh; 1869 int error; 1870 1871 error = dmu_objset_hold(osname, suspend_tag, &os); 1872 if (error != 0) 1873 return (error); 1874 zilog = dmu_objset_zil(os); 1875 1876 mutex_enter(&zilog->zl_lock); 1877 zh = zilog->zl_header; 1878 1879 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1880 mutex_exit(&zilog->zl_lock); 1881 dmu_objset_rele(os, suspend_tag); 1882 return (SET_ERROR(EBUSY)); 1883 } 1884 1885 /* 1886 * Don't put a long hold in the cases where we can avoid it. This 1887 * is when there is no cookie so we are doing a suspend & resume 1888 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1889 * for the suspend because it's already suspended, or there's no ZIL. 1890 */ 1891 if (cookiep == NULL && !zilog->zl_suspending && 1892 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1893 mutex_exit(&zilog->zl_lock); 1894 dmu_objset_rele(os, suspend_tag); 1895 return (0); 1896 } 1897 1898 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1899 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1900 1901 zilog->zl_suspend++; 1902 1903 if (zilog->zl_suspend > 1) { 1904 /* 1905 * Someone else is already suspending it. 1906 * Just wait for them to finish. 1907 */ 1908 1909 while (zilog->zl_suspending) 1910 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1911 mutex_exit(&zilog->zl_lock); 1912 1913 if (cookiep == NULL) 1914 zil_resume(os); 1915 else 1916 *cookiep = os; 1917 return (0); 1918 } 1919 1920 /* 1921 * If there is no pointer to an on-disk block, this ZIL must not 1922 * be active (e.g. filesystem not mounted), so there's nothing 1923 * to clean up. 1924 */ 1925 if (BP_IS_HOLE(&zh->zh_log)) { 1926 ASSERT(cookiep != NULL); /* fast path already handled */ 1927 1928 *cookiep = os; 1929 mutex_exit(&zilog->zl_lock); 1930 return (0); 1931 } 1932 1933 zilog->zl_suspending = B_TRUE; 1934 mutex_exit(&zilog->zl_lock); 1935 1936 zil_commit(zilog, 0); 1937 1938 zil_destroy(zilog, B_FALSE); 1939 1940 mutex_enter(&zilog->zl_lock); 1941 zilog->zl_suspending = B_FALSE; 1942 cv_broadcast(&zilog->zl_cv_suspend); 1943 mutex_exit(&zilog->zl_lock); 1944 1945 if (cookiep == NULL) 1946 zil_resume(os); 1947 else 1948 *cookiep = os; 1949 return (0); 1950} 1951 1952void 1953zil_resume(void *cookie) 1954{ 1955 objset_t *os = cookie; 1956 zilog_t *zilog = dmu_objset_zil(os); 1957 1958 mutex_enter(&zilog->zl_lock); 1959 ASSERT(zilog->zl_suspend != 0); 1960 zilog->zl_suspend--; 1961 mutex_exit(&zilog->zl_lock); 1962 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1963 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1964} 1965 1966typedef struct zil_replay_arg { 1967 zil_replay_func_t **zr_replay; 1968 void *zr_arg; 1969 boolean_t zr_byteswap; 1970 char *zr_lr; 1971} zil_replay_arg_t; 1972 1973static int 1974zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1975{ 1976 char name[MAXNAMELEN]; 1977 1978 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1979 1980 dmu_objset_name(zilog->zl_os, name); 1981 1982 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1983 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1984 (u_longlong_t)lr->lrc_seq, 1985 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1986 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1987 1988 return (error); 1989} 1990 1991static int 1992zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1993{ 1994 zil_replay_arg_t *zr = zra; 1995 const zil_header_t *zh = zilog->zl_header; 1996 uint64_t reclen = lr->lrc_reclen; 1997 uint64_t txtype = lr->lrc_txtype; 1998 int error = 0; 1999 2000 zilog->zl_replaying_seq = lr->lrc_seq; 2001 2002 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2003 return (0); 2004 2005 if (lr->lrc_txg < claim_txg) /* already committed */ 2006 return (0); 2007 2008 /* Strip case-insensitive bit, still present in log record */ 2009 txtype &= ~TX_CI; 2010 2011 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2012 return (zil_replay_error(zilog, lr, EINVAL)); 2013 2014 /* 2015 * If this record type can be logged out of order, the object 2016 * (lr_foid) may no longer exist. That's legitimate, not an error. 2017 */ 2018 if (TX_OOO(txtype)) { 2019 error = dmu_object_info(zilog->zl_os, 2020 ((lr_ooo_t *)lr)->lr_foid, NULL); 2021 if (error == ENOENT || error == EEXIST) 2022 return (0); 2023 } 2024 2025 /* 2026 * Make a copy of the data so we can revise and extend it. 2027 */ 2028 bcopy(lr, zr->zr_lr, reclen); 2029 2030 /* 2031 * If this is a TX_WRITE with a blkptr, suck in the data. 2032 */ 2033 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2034 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2035 zr->zr_lr + reclen); 2036 if (error != 0) 2037 return (zil_replay_error(zilog, lr, error)); 2038 } 2039 2040 /* 2041 * The log block containing this lr may have been byteswapped 2042 * so that we can easily examine common fields like lrc_txtype. 2043 * However, the log is a mix of different record types, and only the 2044 * replay vectors know how to byteswap their records. Therefore, if 2045 * the lr was byteswapped, undo it before invoking the replay vector. 2046 */ 2047 if (zr->zr_byteswap) 2048 byteswap_uint64_array(zr->zr_lr, reclen); 2049 2050 /* 2051 * We must now do two things atomically: replay this log record, 2052 * and update the log header sequence number to reflect the fact that 2053 * we did so. At the end of each replay function the sequence number 2054 * is updated if we are in replay mode. 2055 */ 2056 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2057 if (error != 0) { 2058 /* 2059 * The DMU's dnode layer doesn't see removes until the txg 2060 * commits, so a subsequent claim can spuriously fail with 2061 * EEXIST. So if we receive any error we try syncing out 2062 * any removes then retry the transaction. Note that we 2063 * specify B_FALSE for byteswap now, so we don't do it twice. 2064 */ 2065 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2066 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2067 if (error != 0) 2068 return (zil_replay_error(zilog, lr, error)); 2069 } 2070 return (0); 2071} 2072 2073/* ARGSUSED */ 2074static int 2075zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2076{ 2077 zilog->zl_replay_blks++; 2078 2079 return (0); 2080} 2081 2082/* 2083 * If this dataset has a non-empty intent log, replay it and destroy it. 2084 */ 2085void 2086zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2087{ 2088 zilog_t *zilog = dmu_objset_zil(os); 2089 const zil_header_t *zh = zilog->zl_header; 2090 zil_replay_arg_t zr; 2091 2092 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2093 zil_destroy(zilog, B_TRUE); 2094 return; 2095 } 2096 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 2097 2098 zr.zr_replay = replay_func; 2099 zr.zr_arg = arg; 2100 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2101 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2102 2103 /* 2104 * Wait for in-progress removes to sync before starting replay. 2105 */ 2106 txg_wait_synced(zilog->zl_dmu_pool, 0); 2107 2108 zilog->zl_replay = B_TRUE; 2109 zilog->zl_replay_time = ddi_get_lbolt(); 2110 ASSERT(zilog->zl_replay_blks == 0); 2111 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2112 zh->zh_claim_txg); 2113 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2114 2115 zil_destroy(zilog, B_FALSE); 2116 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2117 zilog->zl_replay = B_FALSE; 2118 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2119} 2120 2121boolean_t 2122zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2123{ 2124 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2125 return (B_TRUE); 2126 2127 if (zilog->zl_replay) { 2128 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2129 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2130 zilog->zl_replaying_seq; 2131 return (B_TRUE); 2132 } 2133 2134 return (B_FALSE); 2135} 2136 2137/* ARGSUSED */ 2138int 2139zil_vdev_offline(const char *osname, void *arg) 2140{ 2141 int error; 2142 2143 error = zil_suspend(osname, NULL); 2144 if (error != 0) 2145 return (SET_ERROR(EEXIST)); 2146 return (0); 2147} 2148