zil.c revision 320496
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, 2016 by Delphix. All rights reserved. 24 * Copyright (c) 2014 Integros [integros.com] 25 */ 26 27/* Portions Copyright 2010 Robert Milkowski */ 28 29#include <sys/zfs_context.h> 30#include <sys/spa.h> 31#include <sys/dmu.h> 32#include <sys/zap.h> 33#include <sys/arc.h> 34#include <sys/stat.h> 35#include <sys/resource.h> 36#include <sys/zil.h> 37#include <sys/zil_impl.h> 38#include <sys/dsl_dataset.h> 39#include <sys/vdev_impl.h> 40#include <sys/dmu_tx.h> 41#include <sys/dsl_pool.h> 42 43/* 44 * The zfs intent log (ZIL) saves transaction records of system calls 45 * that change the file system in memory with enough information 46 * to be able to replay them. These are stored in memory until 47 * either the DMU transaction group (txg) commits them to the stable pool 48 * and they can be discarded, or they are flushed to the stable log 49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 50 * requirement. In the event of a panic or power fail then those log 51 * records (transactions) are replayed. 52 * 53 * There is one ZIL per file system. Its on-disk (pool) format consists 54 * of 3 parts: 55 * 56 * - ZIL header 57 * - ZIL blocks 58 * - ZIL records 59 * 60 * A log record holds a system call transaction. Log blocks can 61 * hold many log records and the blocks are chained together. 62 * Each ZIL block contains a block pointer (blkptr_t) to the next 63 * ZIL block in the chain. The ZIL header points to the first 64 * block in the chain. Note there is not a fixed place in the pool 65 * to hold blocks. They are dynamically allocated and freed as 66 * needed from the blocks available. Figure X shows the ZIL structure: 67 */ 68 69/* 70 * Disable intent logging replay. This global ZIL switch affects all pools. 71 */ 72int zil_replay_disable = 0; 73SYSCTL_DECL(_vfs_zfs); 74TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 75SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 76 &zil_replay_disable, 0, "Disable intent logging replay"); 77 78/* 79 * Tunable parameter for debugging or performance analysis. Setting 80 * zfs_nocacheflush will cause corruption on power loss if a volatile 81 * out-of-order write cache is enabled. 82 */ 83boolean_t zfs_nocacheflush = B_FALSE; 84TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 85SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 86 &zfs_nocacheflush, 0, "Disable cache flush"); 87boolean_t zfs_trim_enabled = B_TRUE; 88SYSCTL_DECL(_vfs_zfs_trim); 89TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 90SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 91 "Enable ZFS TRIM"); 92 93/* 94 * Limit SLOG write size per commit executed with synchronous priority. 95 * Any writes above that executed with lower (asynchronous) priority to 96 * limit potential SLOG device abuse by single active ZIL writer. 97 */ 98uint64_t zil_slog_limit = 768 * 1024; 99SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_limit, CTLFLAG_RWTUN, 100 &zil_slog_limit, 0, "Maximal SLOG commit size with sync priority"); 101 102static kmem_cache_t *zil_lwb_cache; 103 104#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 105 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 106 107 108/* 109 * ziltest is by and large an ugly hack, but very useful in 110 * checking replay without tedious work. 111 * When running ziltest we want to keep all itx's and so maintain 112 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 113 * We subtract TXG_CONCURRENT_STATES to allow for common code. 114 */ 115#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 116 117static int 118zil_bp_compare(const void *x1, const void *x2) 119{ 120 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 121 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 122 123 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 124 return (-1); 125 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 126 return (1); 127 128 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 129 return (-1); 130 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 131 return (1); 132 133 return (0); 134} 135 136static void 137zil_bp_tree_init(zilog_t *zilog) 138{ 139 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 140 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 141} 142 143static void 144zil_bp_tree_fini(zilog_t *zilog) 145{ 146 avl_tree_t *t = &zilog->zl_bp_tree; 147 zil_bp_node_t *zn; 148 void *cookie = NULL; 149 150 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 151 kmem_free(zn, sizeof (zil_bp_node_t)); 152 153 avl_destroy(t); 154} 155 156int 157zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 158{ 159 avl_tree_t *t = &zilog->zl_bp_tree; 160 const dva_t *dva; 161 zil_bp_node_t *zn; 162 avl_index_t where; 163 164 if (BP_IS_EMBEDDED(bp)) 165 return (0); 166 167 dva = BP_IDENTITY(bp); 168 169 if (avl_find(t, dva, &where) != NULL) 170 return (SET_ERROR(EEXIST)); 171 172 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 173 zn->zn_dva = *dva; 174 avl_insert(t, zn, where); 175 176 return (0); 177} 178 179static zil_header_t * 180zil_header_in_syncing_context(zilog_t *zilog) 181{ 182 return ((zil_header_t *)zilog->zl_header); 183} 184 185static void 186zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 187{ 188 zio_cksum_t *zc = &bp->blk_cksum; 189 190 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 191 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 192 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 193 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 194} 195 196/* 197 * Read a log block and make sure it's valid. 198 */ 199static int 200zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 201 char **end) 202{ 203 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 204 arc_flags_t aflags = ARC_FLAG_WAIT; 205 arc_buf_t *abuf = NULL; 206 zbookmark_phys_t zb; 207 int error; 208 209 if (zilog->zl_header->zh_claim_txg == 0) 210 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 211 212 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 213 zio_flags |= ZIO_FLAG_SPECULATIVE; 214 215 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 216 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 217 218 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 219 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 220 221 if (error == 0) { 222 zio_cksum_t cksum = bp->blk_cksum; 223 224 /* 225 * Validate the checksummed log block. 226 * 227 * Sequence numbers should be... sequential. The checksum 228 * verifier for the next block should be bp's checksum plus 1. 229 * 230 * Also check the log chain linkage and size used. 231 */ 232 cksum.zc_word[ZIL_ZC_SEQ]++; 233 234 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 235 zil_chain_t *zilc = abuf->b_data; 236 char *lr = (char *)(zilc + 1); 237 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 238 239 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 240 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 241 error = SET_ERROR(ECKSUM); 242 } else { 243 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 244 bcopy(lr, dst, len); 245 *end = (char *)dst + len; 246 *nbp = zilc->zc_next_blk; 247 } 248 } else { 249 char *lr = abuf->b_data; 250 uint64_t size = BP_GET_LSIZE(bp); 251 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 252 253 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 254 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 255 (zilc->zc_nused > (size - sizeof (*zilc)))) { 256 error = SET_ERROR(ECKSUM); 257 } else { 258 ASSERT3U(zilc->zc_nused, <=, 259 SPA_OLD_MAXBLOCKSIZE); 260 bcopy(lr, dst, zilc->zc_nused); 261 *end = (char *)dst + zilc->zc_nused; 262 *nbp = zilc->zc_next_blk; 263 } 264 } 265 266 arc_buf_destroy(abuf, &abuf); 267 } 268 269 return (error); 270} 271 272/* 273 * Read a TX_WRITE log data block. 274 */ 275static int 276zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 277{ 278 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 279 const blkptr_t *bp = &lr->lr_blkptr; 280 arc_flags_t aflags = ARC_FLAG_WAIT; 281 arc_buf_t *abuf = NULL; 282 zbookmark_phys_t zb; 283 int error; 284 285 if (BP_IS_HOLE(bp)) { 286 if (wbuf != NULL) 287 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 288 return (0); 289 } 290 291 if (zilog->zl_header->zh_claim_txg == 0) 292 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 293 294 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 295 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 296 297 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 298 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 299 300 if (error == 0) { 301 if (wbuf != NULL) 302 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 303 arc_buf_destroy(abuf, &abuf); 304 } 305 306 return (error); 307} 308 309/* 310 * Parse the intent log, and call parse_func for each valid record within. 311 */ 312int 313zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 314 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 315{ 316 const zil_header_t *zh = zilog->zl_header; 317 boolean_t claimed = !!zh->zh_claim_txg; 318 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 319 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 320 uint64_t max_blk_seq = 0; 321 uint64_t max_lr_seq = 0; 322 uint64_t blk_count = 0; 323 uint64_t lr_count = 0; 324 blkptr_t blk, next_blk; 325 char *lrbuf, *lrp; 326 int error = 0; 327 328 /* 329 * Old logs didn't record the maximum zh_claim_lr_seq. 330 */ 331 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 332 claim_lr_seq = UINT64_MAX; 333 334 /* 335 * Starting at the block pointed to by zh_log we read the log chain. 336 * For each block in the chain we strongly check that block to 337 * ensure its validity. We stop when an invalid block is found. 338 * For each block pointer in the chain we call parse_blk_func(). 339 * For each record in each valid block we call parse_lr_func(). 340 * If the log has been claimed, stop if we encounter a sequence 341 * number greater than the highest claimed sequence number. 342 */ 343 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 344 zil_bp_tree_init(zilog); 345 346 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 347 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 348 int reclen; 349 char *end; 350 351 if (blk_seq > claim_blk_seq) 352 break; 353 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 354 break; 355 ASSERT3U(max_blk_seq, <, blk_seq); 356 max_blk_seq = blk_seq; 357 blk_count++; 358 359 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 360 break; 361 362 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 363 if (error != 0) 364 break; 365 366 for (lrp = lrbuf; lrp < end; lrp += reclen) { 367 lr_t *lr = (lr_t *)lrp; 368 reclen = lr->lrc_reclen; 369 ASSERT3U(reclen, >=, sizeof (lr_t)); 370 if (lr->lrc_seq > claim_lr_seq) 371 goto done; 372 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 373 goto done; 374 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 375 max_lr_seq = lr->lrc_seq; 376 lr_count++; 377 } 378 } 379done: 380 zilog->zl_parse_error = error; 381 zilog->zl_parse_blk_seq = max_blk_seq; 382 zilog->zl_parse_lr_seq = max_lr_seq; 383 zilog->zl_parse_blk_count = blk_count; 384 zilog->zl_parse_lr_count = lr_count; 385 386 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 387 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 388 389 zil_bp_tree_fini(zilog); 390 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 391 392 return (error); 393} 394 395static int 396zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 397{ 398 /* 399 * Claim log block if not already committed and not already claimed. 400 * If tx == NULL, just verify that the block is claimable. 401 */ 402 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 403 zil_bp_tree_add(zilog, bp) != 0) 404 return (0); 405 406 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 407 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 408 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 409} 410 411static int 412zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 413{ 414 lr_write_t *lr = (lr_write_t *)lrc; 415 int error; 416 417 if (lrc->lrc_txtype != TX_WRITE) 418 return (0); 419 420 /* 421 * If the block is not readable, don't claim it. This can happen 422 * in normal operation when a log block is written to disk before 423 * some of the dmu_sync() blocks it points to. In this case, the 424 * transaction cannot have been committed to anyone (we would have 425 * waited for all writes to be stable first), so it is semantically 426 * correct to declare this the end of the log. 427 */ 428 if (lr->lr_blkptr.blk_birth >= first_txg && 429 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 430 return (error); 431 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 432} 433 434/* ARGSUSED */ 435static int 436zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 437{ 438 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 439 440 return (0); 441} 442 443static int 444zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 445{ 446 lr_write_t *lr = (lr_write_t *)lrc; 447 blkptr_t *bp = &lr->lr_blkptr; 448 449 /* 450 * If we previously claimed it, we need to free it. 451 */ 452 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 453 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 454 !BP_IS_HOLE(bp)) 455 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 456 457 return (0); 458} 459 460static lwb_t * 461zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg) 462{ 463 lwb_t *lwb; 464 465 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 466 lwb->lwb_zilog = zilog; 467 lwb->lwb_blk = *bp; 468 lwb->lwb_slog = slog; 469 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 470 lwb->lwb_max_txg = txg; 471 lwb->lwb_zio = NULL; 472 lwb->lwb_tx = NULL; 473 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 474 lwb->lwb_nused = sizeof (zil_chain_t); 475 lwb->lwb_sz = BP_GET_LSIZE(bp); 476 } else { 477 lwb->lwb_nused = 0; 478 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 479 } 480 481 mutex_enter(&zilog->zl_lock); 482 list_insert_tail(&zilog->zl_lwb_list, lwb); 483 mutex_exit(&zilog->zl_lock); 484 485 return (lwb); 486} 487 488/* 489 * Called when we create in-memory log transactions so that we know 490 * to cleanup the itxs at the end of spa_sync(). 491 */ 492void 493zilog_dirty(zilog_t *zilog, uint64_t txg) 494{ 495 dsl_pool_t *dp = zilog->zl_dmu_pool; 496 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 497 498 if (ds->ds_is_snapshot) 499 panic("dirtying snapshot!"); 500 501 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 502 /* up the hold count until we can be written out */ 503 dmu_buf_add_ref(ds->ds_dbuf, zilog); 504 } 505} 506 507/* 508 * Determine if the zil is dirty in the specified txg. Callers wanting to 509 * ensure that the dirty state does not change must hold the itxg_lock for 510 * the specified txg. Holding the lock will ensure that the zil cannot be 511 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 512 * state. 513 */ 514boolean_t 515zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 516{ 517 dsl_pool_t *dp = zilog->zl_dmu_pool; 518 519 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 520 return (B_TRUE); 521 return (B_FALSE); 522} 523 524/* 525 * Determine if the zil is dirty. The zil is considered dirty if it has 526 * any pending itx records that have not been cleaned by zil_clean(). 527 */ 528boolean_t 529zilog_is_dirty(zilog_t *zilog) 530{ 531 dsl_pool_t *dp = zilog->zl_dmu_pool; 532 533 for (int t = 0; t < TXG_SIZE; t++) { 534 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 535 return (B_TRUE); 536 } 537 return (B_FALSE); 538} 539 540/* 541 * Create an on-disk intent log. 542 */ 543static lwb_t * 544zil_create(zilog_t *zilog) 545{ 546 const zil_header_t *zh = zilog->zl_header; 547 lwb_t *lwb = NULL; 548 uint64_t txg = 0; 549 dmu_tx_t *tx = NULL; 550 blkptr_t blk; 551 int error = 0; 552 boolean_t slog = FALSE; 553 554 /* 555 * Wait for any previous destroy to complete. 556 */ 557 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 558 559 ASSERT(zh->zh_claim_txg == 0); 560 ASSERT(zh->zh_replay_seq == 0); 561 562 blk = zh->zh_log; 563 564 /* 565 * Allocate an initial log block if: 566 * - there isn't one already 567 * - the existing block is the wrong endianess 568 */ 569 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 570 tx = dmu_tx_create(zilog->zl_os); 571 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 572 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 573 txg = dmu_tx_get_txg(tx); 574 575 if (!BP_IS_HOLE(&blk)) { 576 zio_free_zil(zilog->zl_spa, txg, &blk); 577 BP_ZERO(&blk); 578 } 579 580 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 581 ZIL_MIN_BLKSZ, &slog); 582 583 if (error == 0) 584 zil_init_log_chain(zilog, &blk); 585 } 586 587 /* 588 * Allocate a log write buffer (lwb) for the first log block. 589 */ 590 if (error == 0) 591 lwb = zil_alloc_lwb(zilog, &blk, slog, txg); 592 593 /* 594 * If we just allocated the first log block, commit our transaction 595 * and wait for zil_sync() to stuff the block poiner into zh_log. 596 * (zh is part of the MOS, so we cannot modify it in open context.) 597 */ 598 if (tx != NULL) { 599 dmu_tx_commit(tx); 600 txg_wait_synced(zilog->zl_dmu_pool, txg); 601 } 602 603 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 604 605 return (lwb); 606} 607 608/* 609 * In one tx, free all log blocks and clear the log header. 610 * If keep_first is set, then we're replaying a log with no content. 611 * We want to keep the first block, however, so that the first 612 * synchronous transaction doesn't require a txg_wait_synced() 613 * in zil_create(). We don't need to txg_wait_synced() here either 614 * when keep_first is set, because both zil_create() and zil_destroy() 615 * will wait for any in-progress destroys to complete. 616 */ 617void 618zil_destroy(zilog_t *zilog, boolean_t keep_first) 619{ 620 const zil_header_t *zh = zilog->zl_header; 621 lwb_t *lwb; 622 dmu_tx_t *tx; 623 uint64_t txg; 624 625 /* 626 * Wait for any previous destroy to complete. 627 */ 628 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 629 630 zilog->zl_old_header = *zh; /* debugging aid */ 631 632 if (BP_IS_HOLE(&zh->zh_log)) 633 return; 634 635 tx = dmu_tx_create(zilog->zl_os); 636 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 637 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 638 txg = dmu_tx_get_txg(tx); 639 640 mutex_enter(&zilog->zl_lock); 641 642 ASSERT3U(zilog->zl_destroy_txg, <, txg); 643 zilog->zl_destroy_txg = txg; 644 zilog->zl_keep_first = keep_first; 645 646 if (!list_is_empty(&zilog->zl_lwb_list)) { 647 ASSERT(zh->zh_claim_txg == 0); 648 VERIFY(!keep_first); 649 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 650 list_remove(&zilog->zl_lwb_list, lwb); 651 if (lwb->lwb_buf != NULL) 652 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 653 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 654 kmem_cache_free(zil_lwb_cache, lwb); 655 } 656 } else if (!keep_first) { 657 zil_destroy_sync(zilog, tx); 658 } 659 mutex_exit(&zilog->zl_lock); 660 661 dmu_tx_commit(tx); 662} 663 664void 665zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 666{ 667 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 668 (void) zil_parse(zilog, zil_free_log_block, 669 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 670} 671 672int 673zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 674{ 675 dmu_tx_t *tx = txarg; 676 uint64_t first_txg = dmu_tx_get_txg(tx); 677 zilog_t *zilog; 678 zil_header_t *zh; 679 objset_t *os; 680 int error; 681 682 error = dmu_objset_own_obj(dp, ds->ds_object, 683 DMU_OST_ANY, B_FALSE, FTAG, &os); 684 if (error != 0) { 685 /* 686 * EBUSY indicates that the objset is inconsistent, in which 687 * case it can not have a ZIL. 688 */ 689 if (error != EBUSY) { 690 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 691 (unsigned long long)ds->ds_object, error); 692 } 693 return (0); 694 } 695 696 zilog = dmu_objset_zil(os); 697 zh = zil_header_in_syncing_context(zilog); 698 699 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 700 if (!BP_IS_HOLE(&zh->zh_log)) 701 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 702 BP_ZERO(&zh->zh_log); 703 dsl_dataset_dirty(dmu_objset_ds(os), tx); 704 dmu_objset_disown(os, FTAG); 705 return (0); 706 } 707 708 /* 709 * Claim all log blocks if we haven't already done so, and remember 710 * the highest claimed sequence number. This ensures that if we can 711 * read only part of the log now (e.g. due to a missing device), 712 * but we can read the entire log later, we will not try to replay 713 * or destroy beyond the last block we successfully claimed. 714 */ 715 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 716 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 717 (void) zil_parse(zilog, zil_claim_log_block, 718 zil_claim_log_record, tx, first_txg); 719 zh->zh_claim_txg = first_txg; 720 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 721 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 722 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 723 zh->zh_flags |= ZIL_REPLAY_NEEDED; 724 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 725 dsl_dataset_dirty(dmu_objset_ds(os), tx); 726 } 727 728 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 729 dmu_objset_disown(os, FTAG); 730 return (0); 731} 732 733/* 734 * Check the log by walking the log chain. 735 * Checksum errors are ok as they indicate the end of the chain. 736 * Any other error (no device or read failure) returns an error. 737 */ 738/* ARGSUSED */ 739int 740zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 741{ 742 zilog_t *zilog; 743 objset_t *os; 744 blkptr_t *bp; 745 int error; 746 747 ASSERT(tx == NULL); 748 749 error = dmu_objset_from_ds(ds, &os); 750 if (error != 0) { 751 cmn_err(CE_WARN, "can't open objset %llu, error %d", 752 (unsigned long long)ds->ds_object, error); 753 return (0); 754 } 755 756 zilog = dmu_objset_zil(os); 757 bp = (blkptr_t *)&zilog->zl_header->zh_log; 758 759 /* 760 * Check the first block and determine if it's on a log device 761 * which may have been removed or faulted prior to loading this 762 * pool. If so, there's no point in checking the rest of the log 763 * as its content should have already been synced to the pool. 764 */ 765 if (!BP_IS_HOLE(bp)) { 766 vdev_t *vd; 767 boolean_t valid = B_TRUE; 768 769 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 770 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 771 if (vd->vdev_islog && vdev_is_dead(vd)) 772 valid = vdev_log_state_valid(vd); 773 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 774 775 if (!valid) 776 return (0); 777 } 778 779 /* 780 * Because tx == NULL, zil_claim_log_block() will not actually claim 781 * any blocks, but just determine whether it is possible to do so. 782 * In addition to checking the log chain, zil_claim_log_block() 783 * will invoke zio_claim() with a done func of spa_claim_notify(), 784 * which will update spa_max_claim_txg. See spa_load() for details. 785 */ 786 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 787 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 788 789 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 790} 791 792static int 793zil_vdev_compare(const void *x1, const void *x2) 794{ 795 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 796 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 797 798 if (v1 < v2) 799 return (-1); 800 if (v1 > v2) 801 return (1); 802 803 return (0); 804} 805 806void 807zil_add_block(zilog_t *zilog, const blkptr_t *bp) 808{ 809 avl_tree_t *t = &zilog->zl_vdev_tree; 810 avl_index_t where; 811 zil_vdev_node_t *zv, zvsearch; 812 int ndvas = BP_GET_NDVAS(bp); 813 int i; 814 815 if (zfs_nocacheflush) 816 return; 817 818 ASSERT(zilog->zl_writer); 819 820 /* 821 * Even though we're zl_writer, we still need a lock because the 822 * zl_get_data() callbacks may have dmu_sync() done callbacks 823 * that will run concurrently. 824 */ 825 mutex_enter(&zilog->zl_vdev_lock); 826 for (i = 0; i < ndvas; i++) { 827 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 828 if (avl_find(t, &zvsearch, &where) == NULL) { 829 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 830 zv->zv_vdev = zvsearch.zv_vdev; 831 avl_insert(t, zv, where); 832 } 833 } 834 mutex_exit(&zilog->zl_vdev_lock); 835} 836 837static void 838zil_flush_vdevs(zilog_t *zilog) 839{ 840 spa_t *spa = zilog->zl_spa; 841 avl_tree_t *t = &zilog->zl_vdev_tree; 842 void *cookie = NULL; 843 zil_vdev_node_t *zv; 844 zio_t *zio = NULL; 845 846 ASSERT(zilog->zl_writer); 847 848 /* 849 * We don't need zl_vdev_lock here because we're the zl_writer, 850 * and all zl_get_data() callbacks are done. 851 */ 852 if (avl_numnodes(t) == 0) 853 return; 854 855 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 856 857 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 858 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 859 if (vd != NULL && !vd->vdev_nowritecache) { 860 if (zio == NULL) 861 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 862 zio_flush(zio, vd); 863 } 864 kmem_free(zv, sizeof (*zv)); 865 } 866 867 /* 868 * Wait for all the flushes to complete. Not all devices actually 869 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 870 */ 871 if (zio) 872 (void) zio_wait(zio); 873 874 spa_config_exit(spa, SCL_STATE, FTAG); 875} 876 877/* 878 * Function called when a log block write completes 879 */ 880static void 881zil_lwb_write_done(zio_t *zio) 882{ 883 lwb_t *lwb = zio->io_private; 884 zilog_t *zilog = lwb->lwb_zilog; 885 dmu_tx_t *tx = lwb->lwb_tx; 886 887 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 888 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 889 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 890 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 891 ASSERT(!BP_IS_GANG(zio->io_bp)); 892 ASSERT(!BP_IS_HOLE(zio->io_bp)); 893 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 894 895 /* 896 * Ensure the lwb buffer pointer is cleared before releasing 897 * the txg. If we have had an allocation failure and 898 * the txg is waiting to sync then we want want zil_sync() 899 * to remove the lwb so that it's not picked up as the next new 900 * one in zil_commit_writer(). zil_sync() will only remove 901 * the lwb if lwb_buf is null. 902 */ 903 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 904 mutex_enter(&zilog->zl_lock); 905 lwb->lwb_buf = NULL; 906 lwb->lwb_tx = NULL; 907 mutex_exit(&zilog->zl_lock); 908 909 /* 910 * Now that we've written this log block, we have a stable pointer 911 * to the next block in the chain, so it's OK to let the txg in 912 * which we allocated the next block sync. 913 */ 914 dmu_tx_commit(tx); 915} 916 917/* 918 * Initialize the io for a log block. 919 */ 920static void 921zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 922{ 923 zbookmark_phys_t zb; 924 zio_priority_t prio; 925 926 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 927 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 928 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 929 930 if (zilog->zl_root_zio == NULL) { 931 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 932 ZIO_FLAG_CANFAIL); 933 } 934 if (lwb->lwb_zio == NULL) { 935 if (zilog->zl_cur_used <= zil_slog_limit || !lwb->lwb_slog) 936 prio = ZIO_PRIORITY_SYNC_WRITE; 937 else 938 prio = ZIO_PRIORITY_ASYNC_WRITE; 939 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 940 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 941 zil_lwb_write_done, lwb, prio, 942 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 943 } 944} 945 946/* 947 * Define a limited set of intent log block sizes. 948 * 949 * These must be a multiple of 4KB. Note only the amount used (again 950 * aligned to 4KB) actually gets written. However, we can't always just 951 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 952 */ 953uint64_t zil_block_buckets[] = { 954 4096, /* non TX_WRITE */ 955 8192+4096, /* data base */ 956 32*1024 + 4096, /* NFS writes */ 957 UINT64_MAX 958}; 959 960/* 961 * Start a log block write and advance to the next log block. 962 * Calls are serialized. 963 */ 964static lwb_t * 965zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb, boolean_t last) 966{ 967 lwb_t *nlwb = NULL; 968 zil_chain_t *zilc; 969 spa_t *spa = zilog->zl_spa; 970 blkptr_t *bp; 971 dmu_tx_t *tx; 972 uint64_t txg; 973 uint64_t zil_blksz, wsz; 974 int i, error; 975 boolean_t slog; 976 977 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 978 zilc = (zil_chain_t *)lwb->lwb_buf; 979 bp = &zilc->zc_next_blk; 980 } else { 981 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 982 bp = &zilc->zc_next_blk; 983 } 984 985 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 986 987 /* 988 * Allocate the next block and save its address in this block 989 * before writing it in order to establish the log chain. 990 * Note that if the allocation of nlwb synced before we wrote 991 * the block that points at it (lwb), we'd leak it if we crashed. 992 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 993 * We dirty the dataset to ensure that zil_sync() will be called 994 * to clean up in the event of allocation failure or I/O failure. 995 */ 996 tx = dmu_tx_create(zilog->zl_os); 997 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 998 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 999 txg = dmu_tx_get_txg(tx); 1000 1001 lwb->lwb_tx = tx; 1002 1003 /* 1004 * Log blocks are pre-allocated. Here we select the size of the next 1005 * block, based on size used in the last block. 1006 * - first find the smallest bucket that will fit the block from a 1007 * limited set of block sizes. This is because it's faster to write 1008 * blocks allocated from the same metaslab as they are adjacent or 1009 * close. 1010 * - next find the maximum from the new suggested size and an array of 1011 * previous sizes. This lessens a picket fence effect of wrongly 1012 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 1013 * requests. 1014 * 1015 * Note we only write what is used, but we can't just allocate 1016 * the maximum block size because we can exhaust the available 1017 * pool log space. 1018 */ 1019 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1020 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 1021 continue; 1022 zil_blksz = zil_block_buckets[i]; 1023 if (zil_blksz == UINT64_MAX) 1024 zil_blksz = SPA_OLD_MAXBLOCKSIZE; 1025 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1026 for (i = 0; i < ZIL_PREV_BLKS; i++) 1027 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1028 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1029 1030 BP_ZERO(bp); 1031 /* pass the old blkptr in order to spread log blocks across devs */ 1032 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog); 1033 if (error == 0) { 1034 ASSERT3U(bp->blk_birth, ==, txg); 1035 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1036 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1037 1038 /* 1039 * Allocate a new log write buffer (lwb). 1040 */ 1041 nlwb = zil_alloc_lwb(zilog, bp, slog, txg); 1042 1043 /* Record the block for later vdev flushing */ 1044 zil_add_block(zilog, &lwb->lwb_blk); 1045 } 1046 1047 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1048 /* For Slim ZIL only write what is used. */ 1049 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1050 ASSERT3U(wsz, <=, lwb->lwb_sz); 1051 zio_shrink(lwb->lwb_zio, wsz); 1052 1053 } else { 1054 wsz = lwb->lwb_sz; 1055 } 1056 1057 zilc->zc_pad = 0; 1058 zilc->zc_nused = lwb->lwb_nused; 1059 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1060 1061 /* 1062 * clear unused data for security 1063 */ 1064 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1065 1066 if (last) 1067 lwb->lwb_zio->io_pipeline &= ~ZIO_STAGE_ISSUE_ASYNC; 1068 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1069 1070 /* 1071 * If there was an allocation failure then nlwb will be null which 1072 * forces a txg_wait_synced(). 1073 */ 1074 return (nlwb); 1075} 1076 1077static lwb_t * 1078zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1079{ 1080 lr_t *lrcb, *lrc = &itx->itx_lr; /* common log record */ 1081 lr_write_t *lrwb, *lrw = (lr_write_t *)lrc; 1082 char *lr_buf; 1083 uint64_t txg = lrc->lrc_txg; 1084 uint64_t reclen = lrc->lrc_reclen; 1085 uint64_t dlen = 0; 1086 uint64_t dnow, lwb_sp; 1087 1088 if (lwb == NULL) 1089 return (NULL); 1090 1091 ASSERT(lwb->lwb_buf != NULL); 1092 1093 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1094 dlen = P2ROUNDUP_TYPED( 1095 lrw->lr_length, sizeof (uint64_t), uint64_t); 1096 1097 zilog->zl_cur_used += (reclen + dlen); 1098 1099 zil_lwb_write_init(zilog, lwb); 1100 1101cont: 1102 /* 1103 * If this record won't fit in the current log block, start a new one. 1104 * For WR_NEED_COPY optimize layout for minimal number of chunks, but 1105 * try to keep wasted space withing reasonable range (12%). 1106 */ 1107 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1108 if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1109 lwb_sp < ZIL_MAX_LOG_DATA / 8 && (dlen % ZIL_MAX_LOG_DATA == 0 || 1110 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) { 1111 lwb = zil_lwb_write_start(zilog, lwb, B_FALSE); 1112 if (lwb == NULL) 1113 return (NULL); 1114 zil_lwb_write_init(zilog, lwb); 1115 ASSERT(LWB_EMPTY(lwb)); 1116 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1117 ASSERT3U(reclen + MIN(dlen, sizeof(uint64_t)), <=, lwb_sp); 1118 } 1119 1120 dnow = MIN(dlen, lwb_sp - reclen); 1121 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1122 bcopy(lrc, lr_buf, reclen); 1123 lrcb = (lr_t *)lr_buf; 1124 lrwb = (lr_write_t *)lrcb; 1125 1126 /* 1127 * If it's a write, fetch the data or get its blkptr as appropriate. 1128 */ 1129 if (lrc->lrc_txtype == TX_WRITE) { 1130 if (txg > spa_freeze_txg(zilog->zl_spa)) 1131 txg_wait_synced(zilog->zl_dmu_pool, txg); 1132 if (itx->itx_wr_state != WR_COPIED) { 1133 char *dbuf; 1134 int error; 1135 1136 if (itx->itx_wr_state == WR_NEED_COPY) { 1137 dbuf = lr_buf + reclen; 1138 lrcb->lrc_reclen += dnow; 1139 if (lrwb->lr_length > dnow) 1140 lrwb->lr_length = dnow; 1141 lrw->lr_offset += dnow; 1142 lrw->lr_length -= dnow; 1143 } else { 1144 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1145 dbuf = NULL; 1146 } 1147 error = zilog->zl_get_data( 1148 itx->itx_private, lrwb, dbuf, lwb->lwb_zio); 1149 if (error == EIO) { 1150 txg_wait_synced(zilog->zl_dmu_pool, txg); 1151 return (lwb); 1152 } 1153 if (error != 0) { 1154 ASSERT(error == ENOENT || error == EEXIST || 1155 error == EALREADY); 1156 return (lwb); 1157 } 1158 } 1159 } 1160 1161 /* 1162 * We're actually making an entry, so update lrc_seq to be the 1163 * log record sequence number. Note that this is generally not 1164 * equal to the itx sequence number because not all transactions 1165 * are synchronous, and sometimes spa_sync() gets there first. 1166 */ 1167 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1168 lwb->lwb_nused += reclen + dnow; 1169 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1170 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1171 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1172 1173 dlen -= dnow; 1174 if (dlen > 0) { 1175 zilog->zl_cur_used += reclen; 1176 goto cont; 1177 } 1178 1179 return (lwb); 1180} 1181 1182itx_t * 1183zil_itx_create(uint64_t txtype, size_t lrsize) 1184{ 1185 itx_t *itx; 1186 1187 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1188 1189 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1190 itx->itx_lr.lrc_txtype = txtype; 1191 itx->itx_lr.lrc_reclen = lrsize; 1192 itx->itx_lr.lrc_seq = 0; /* defensive */ 1193 itx->itx_sync = B_TRUE; /* default is synchronous */ 1194 1195 return (itx); 1196} 1197 1198void 1199zil_itx_destroy(itx_t *itx) 1200{ 1201 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1202} 1203 1204/* 1205 * Free up the sync and async itxs. The itxs_t has already been detached 1206 * so no locks are needed. 1207 */ 1208static void 1209zil_itxg_clean(itxs_t *itxs) 1210{ 1211 itx_t *itx; 1212 list_t *list; 1213 avl_tree_t *t; 1214 void *cookie; 1215 itx_async_node_t *ian; 1216 1217 list = &itxs->i_sync_list; 1218 while ((itx = list_head(list)) != NULL) { 1219 list_remove(list, itx); 1220 kmem_free(itx, offsetof(itx_t, itx_lr) + 1221 itx->itx_lr.lrc_reclen); 1222 } 1223 1224 cookie = NULL; 1225 t = &itxs->i_async_tree; 1226 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1227 list = &ian->ia_list; 1228 while ((itx = list_head(list)) != NULL) { 1229 list_remove(list, itx); 1230 kmem_free(itx, offsetof(itx_t, itx_lr) + 1231 itx->itx_lr.lrc_reclen); 1232 } 1233 list_destroy(list); 1234 kmem_free(ian, sizeof (itx_async_node_t)); 1235 } 1236 avl_destroy(t); 1237 1238 kmem_free(itxs, sizeof (itxs_t)); 1239} 1240 1241static int 1242zil_aitx_compare(const void *x1, const void *x2) 1243{ 1244 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1245 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1246 1247 if (o1 < o2) 1248 return (-1); 1249 if (o1 > o2) 1250 return (1); 1251 1252 return (0); 1253} 1254 1255/* 1256 * Remove all async itx with the given oid. 1257 */ 1258static void 1259zil_remove_async(zilog_t *zilog, uint64_t oid) 1260{ 1261 uint64_t otxg, txg; 1262 itx_async_node_t *ian; 1263 avl_tree_t *t; 1264 avl_index_t where; 1265 list_t clean_list; 1266 itx_t *itx; 1267 1268 ASSERT(oid != 0); 1269 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1270 1271 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1272 otxg = ZILTEST_TXG; 1273 else 1274 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1275 1276 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1277 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1278 1279 mutex_enter(&itxg->itxg_lock); 1280 if (itxg->itxg_txg != txg) { 1281 mutex_exit(&itxg->itxg_lock); 1282 continue; 1283 } 1284 1285 /* 1286 * Locate the object node and append its list. 1287 */ 1288 t = &itxg->itxg_itxs->i_async_tree; 1289 ian = avl_find(t, &oid, &where); 1290 if (ian != NULL) 1291 list_move_tail(&clean_list, &ian->ia_list); 1292 mutex_exit(&itxg->itxg_lock); 1293 } 1294 while ((itx = list_head(&clean_list)) != NULL) { 1295 list_remove(&clean_list, itx); 1296 kmem_free(itx, offsetof(itx_t, itx_lr) + 1297 itx->itx_lr.lrc_reclen); 1298 } 1299 list_destroy(&clean_list); 1300} 1301 1302void 1303zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1304{ 1305 uint64_t txg; 1306 itxg_t *itxg; 1307 itxs_t *itxs, *clean = NULL; 1308 1309 /* 1310 * Object ids can be re-instantiated in the next txg so 1311 * remove any async transactions to avoid future leaks. 1312 * This can happen if a fsync occurs on the re-instantiated 1313 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1314 * the new file data and flushes a write record for the old object. 1315 */ 1316 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1317 zil_remove_async(zilog, itx->itx_oid); 1318 1319 /* 1320 * Ensure the data of a renamed file is committed before the rename. 1321 */ 1322 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1323 zil_async_to_sync(zilog, itx->itx_oid); 1324 1325 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1326 txg = ZILTEST_TXG; 1327 else 1328 txg = dmu_tx_get_txg(tx); 1329 1330 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1331 mutex_enter(&itxg->itxg_lock); 1332 itxs = itxg->itxg_itxs; 1333 if (itxg->itxg_txg != txg) { 1334 if (itxs != NULL) { 1335 /* 1336 * The zil_clean callback hasn't got around to cleaning 1337 * this itxg. Save the itxs for release below. 1338 * This should be rare. 1339 */ 1340 clean = itxg->itxg_itxs; 1341 } 1342 itxg->itxg_txg = txg; 1343 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1344 1345 list_create(&itxs->i_sync_list, sizeof (itx_t), 1346 offsetof(itx_t, itx_node)); 1347 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1348 sizeof (itx_async_node_t), 1349 offsetof(itx_async_node_t, ia_node)); 1350 } 1351 if (itx->itx_sync) { 1352 list_insert_tail(&itxs->i_sync_list, itx); 1353 } else { 1354 avl_tree_t *t = &itxs->i_async_tree; 1355 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1356 itx_async_node_t *ian; 1357 avl_index_t where; 1358 1359 ian = avl_find(t, &foid, &where); 1360 if (ian == NULL) { 1361 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1362 list_create(&ian->ia_list, sizeof (itx_t), 1363 offsetof(itx_t, itx_node)); 1364 ian->ia_foid = foid; 1365 avl_insert(t, ian, where); 1366 } 1367 list_insert_tail(&ian->ia_list, itx); 1368 } 1369 1370 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1371 zilog_dirty(zilog, txg); 1372 mutex_exit(&itxg->itxg_lock); 1373 1374 /* Release the old itxs now we've dropped the lock */ 1375 if (clean != NULL) 1376 zil_itxg_clean(clean); 1377} 1378 1379/* 1380 * If there are any in-memory intent log transactions which have now been 1381 * synced then start up a taskq to free them. We should only do this after we 1382 * have written out the uberblocks (i.e. txg has been comitted) so that 1383 * don't inadvertently clean out in-memory log records that would be required 1384 * by zil_commit(). 1385 */ 1386void 1387zil_clean(zilog_t *zilog, uint64_t synced_txg) 1388{ 1389 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1390 itxs_t *clean_me; 1391 1392 mutex_enter(&itxg->itxg_lock); 1393 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1394 mutex_exit(&itxg->itxg_lock); 1395 return; 1396 } 1397 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1398 ASSERT(itxg->itxg_txg != 0); 1399 ASSERT(zilog->zl_clean_taskq != NULL); 1400 clean_me = itxg->itxg_itxs; 1401 itxg->itxg_itxs = NULL; 1402 itxg->itxg_txg = 0; 1403 mutex_exit(&itxg->itxg_lock); 1404 /* 1405 * Preferably start a task queue to free up the old itxs but 1406 * if taskq_dispatch can't allocate resources to do that then 1407 * free it in-line. This should be rare. Note, using TQ_SLEEP 1408 * created a bad performance problem. 1409 */ 1410 if (taskq_dispatch(zilog->zl_clean_taskq, 1411 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1412 zil_itxg_clean(clean_me); 1413} 1414 1415/* 1416 * Get the list of itxs to commit into zl_itx_commit_list. 1417 */ 1418static void 1419zil_get_commit_list(zilog_t *zilog) 1420{ 1421 uint64_t otxg, txg; 1422 list_t *commit_list = &zilog->zl_itx_commit_list; 1423 1424 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1425 otxg = ZILTEST_TXG; 1426 else 1427 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1428 1429 /* 1430 * This is inherently racy, since there is nothing to prevent 1431 * the last synced txg from changing. That's okay since we'll 1432 * only commit things in the future. 1433 */ 1434 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1435 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1436 1437 mutex_enter(&itxg->itxg_lock); 1438 if (itxg->itxg_txg != txg) { 1439 mutex_exit(&itxg->itxg_lock); 1440 continue; 1441 } 1442 1443 /* 1444 * If we're adding itx records to the zl_itx_commit_list, 1445 * then the zil better be dirty in this "txg". We can assert 1446 * that here since we're holding the itxg_lock which will 1447 * prevent spa_sync from cleaning it. Once we add the itxs 1448 * to the zl_itx_commit_list we must commit it to disk even 1449 * if it's unnecessary (i.e. the txg was synced). 1450 */ 1451 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 1452 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1453 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1454 1455 mutex_exit(&itxg->itxg_lock); 1456 } 1457} 1458 1459/* 1460 * Move the async itxs for a specified object to commit into sync lists. 1461 */ 1462void 1463zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1464{ 1465 uint64_t otxg, txg; 1466 itx_async_node_t *ian; 1467 avl_tree_t *t; 1468 avl_index_t where; 1469 1470 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1471 otxg = ZILTEST_TXG; 1472 else 1473 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1474 1475 /* 1476 * This is inherently racy, since there is nothing to prevent 1477 * the last synced txg from changing. 1478 */ 1479 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1480 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1481 1482 mutex_enter(&itxg->itxg_lock); 1483 if (itxg->itxg_txg != txg) { 1484 mutex_exit(&itxg->itxg_lock); 1485 continue; 1486 } 1487 1488 /* 1489 * If a foid is specified then find that node and append its 1490 * list. Otherwise walk the tree appending all the lists 1491 * to the sync list. We add to the end rather than the 1492 * beginning to ensure the create has happened. 1493 */ 1494 t = &itxg->itxg_itxs->i_async_tree; 1495 if (foid != 0) { 1496 ian = avl_find(t, &foid, &where); 1497 if (ian != NULL) { 1498 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1499 &ian->ia_list); 1500 } 1501 } else { 1502 void *cookie = NULL; 1503 1504 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1505 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1506 &ian->ia_list); 1507 list_destroy(&ian->ia_list); 1508 kmem_free(ian, sizeof (itx_async_node_t)); 1509 } 1510 } 1511 mutex_exit(&itxg->itxg_lock); 1512 } 1513} 1514 1515static void 1516zil_commit_writer(zilog_t *zilog) 1517{ 1518 uint64_t txg; 1519 itx_t *itx; 1520 lwb_t *lwb; 1521 spa_t *spa = zilog->zl_spa; 1522 int error = 0; 1523 1524 ASSERT(zilog->zl_root_zio == NULL); 1525 1526 mutex_exit(&zilog->zl_lock); 1527 1528 zil_get_commit_list(zilog); 1529 1530 /* 1531 * Return if there's nothing to commit before we dirty the fs by 1532 * calling zil_create(). 1533 */ 1534 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1535 mutex_enter(&zilog->zl_lock); 1536 return; 1537 } 1538 1539 if (zilog->zl_suspend) { 1540 lwb = NULL; 1541 } else { 1542 lwb = list_tail(&zilog->zl_lwb_list); 1543 if (lwb == NULL) 1544 lwb = zil_create(zilog); 1545 } 1546 1547 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1548 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1549 txg = itx->itx_lr.lrc_txg; 1550 ASSERT3U(txg, !=, 0); 1551 1552 /* 1553 * This is inherently racy and may result in us writing 1554 * out a log block for a txg that was just synced. This is 1555 * ok since we'll end cleaning up that log block the next 1556 * time we call zil_sync(). 1557 */ 1558 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1559 lwb = zil_lwb_commit(zilog, itx, lwb); 1560 list_remove(&zilog->zl_itx_commit_list, itx); 1561 kmem_free(itx, offsetof(itx_t, itx_lr) 1562 + itx->itx_lr.lrc_reclen); 1563 } 1564 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1565 1566 /* write the last block out */ 1567 if (lwb != NULL && lwb->lwb_zio != NULL) 1568 lwb = zil_lwb_write_start(zilog, lwb, B_TRUE); 1569 1570 zilog->zl_cur_used = 0; 1571 1572 /* 1573 * Wait if necessary for the log blocks to be on stable storage. 1574 */ 1575 if (zilog->zl_root_zio) { 1576 error = zio_wait(zilog->zl_root_zio); 1577 zilog->zl_root_zio = NULL; 1578 zil_flush_vdevs(zilog); 1579 } 1580 1581 if (error || lwb == NULL) 1582 txg_wait_synced(zilog->zl_dmu_pool, 0); 1583 1584 mutex_enter(&zilog->zl_lock); 1585 1586 /* 1587 * Remember the highest committed log sequence number for ztest. 1588 * We only update this value when all the log writes succeeded, 1589 * because ztest wants to ASSERT that it got the whole log chain. 1590 */ 1591 if (error == 0 && lwb != NULL) 1592 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1593} 1594 1595/* 1596 * Commit zfs transactions to stable storage. 1597 * If foid is 0 push out all transactions, otherwise push only those 1598 * for that object or might reference that object. 1599 * 1600 * itxs are committed in batches. In a heavily stressed zil there will be 1601 * a commit writer thread who is writing out a bunch of itxs to the log 1602 * for a set of committing threads (cthreads) in the same batch as the writer. 1603 * Those cthreads are all waiting on the same cv for that batch. 1604 * 1605 * There will also be a different and growing batch of threads that are 1606 * waiting to commit (qthreads). When the committing batch completes 1607 * a transition occurs such that the cthreads exit and the qthreads become 1608 * cthreads. One of the new cthreads becomes the writer thread for the 1609 * batch. Any new threads arriving become new qthreads. 1610 * 1611 * Only 2 condition variables are needed and there's no transition 1612 * between the two cvs needed. They just flip-flop between qthreads 1613 * and cthreads. 1614 * 1615 * Using this scheme we can efficiently wakeup up only those threads 1616 * that have been committed. 1617 */ 1618void 1619zil_commit(zilog_t *zilog, uint64_t foid) 1620{ 1621 uint64_t mybatch; 1622 1623 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1624 return; 1625 1626 /* move the async itxs for the foid to the sync queues */ 1627 zil_async_to_sync(zilog, foid); 1628 1629 mutex_enter(&zilog->zl_lock); 1630 mybatch = zilog->zl_next_batch; 1631 while (zilog->zl_writer) { 1632 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1633 if (mybatch <= zilog->zl_com_batch) { 1634 mutex_exit(&zilog->zl_lock); 1635 return; 1636 } 1637 } 1638 1639 zilog->zl_next_batch++; 1640 zilog->zl_writer = B_TRUE; 1641 zil_commit_writer(zilog); 1642 zilog->zl_com_batch = mybatch; 1643 zilog->zl_writer = B_FALSE; 1644 mutex_exit(&zilog->zl_lock); 1645 1646 /* wake up one thread to become the next writer */ 1647 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1648 1649 /* wake up all threads waiting for this batch to be committed */ 1650 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1651} 1652 1653/* 1654 * Called in syncing context to free committed log blocks and update log header. 1655 */ 1656void 1657zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1658{ 1659 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1660 uint64_t txg = dmu_tx_get_txg(tx); 1661 spa_t *spa = zilog->zl_spa; 1662 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1663 lwb_t *lwb; 1664 1665 /* 1666 * We don't zero out zl_destroy_txg, so make sure we don't try 1667 * to destroy it twice. 1668 */ 1669 if (spa_sync_pass(spa) != 1) 1670 return; 1671 1672 mutex_enter(&zilog->zl_lock); 1673 1674 ASSERT(zilog->zl_stop_sync == 0); 1675 1676 if (*replayed_seq != 0) { 1677 ASSERT(zh->zh_replay_seq < *replayed_seq); 1678 zh->zh_replay_seq = *replayed_seq; 1679 *replayed_seq = 0; 1680 } 1681 1682 if (zilog->zl_destroy_txg == txg) { 1683 blkptr_t blk = zh->zh_log; 1684 1685 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1686 1687 bzero(zh, sizeof (zil_header_t)); 1688 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1689 1690 if (zilog->zl_keep_first) { 1691 /* 1692 * If this block was part of log chain that couldn't 1693 * be claimed because a device was missing during 1694 * zil_claim(), but that device later returns, 1695 * then this block could erroneously appear valid. 1696 * To guard against this, assign a new GUID to the new 1697 * log chain so it doesn't matter what blk points to. 1698 */ 1699 zil_init_log_chain(zilog, &blk); 1700 zh->zh_log = blk; 1701 } 1702 } 1703 1704 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1705 zh->zh_log = lwb->lwb_blk; 1706 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1707 break; 1708 list_remove(&zilog->zl_lwb_list, lwb); 1709 zio_free_zil(spa, txg, &lwb->lwb_blk); 1710 kmem_cache_free(zil_lwb_cache, lwb); 1711 1712 /* 1713 * If we don't have anything left in the lwb list then 1714 * we've had an allocation failure and we need to zero 1715 * out the zil_header blkptr so that we don't end 1716 * up freeing the same block twice. 1717 */ 1718 if (list_head(&zilog->zl_lwb_list) == NULL) 1719 BP_ZERO(&zh->zh_log); 1720 } 1721 mutex_exit(&zilog->zl_lock); 1722} 1723 1724void 1725zil_init(void) 1726{ 1727 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1728 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1729} 1730 1731void 1732zil_fini(void) 1733{ 1734 kmem_cache_destroy(zil_lwb_cache); 1735} 1736 1737void 1738zil_set_sync(zilog_t *zilog, uint64_t sync) 1739{ 1740 zilog->zl_sync = sync; 1741} 1742 1743void 1744zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1745{ 1746 zilog->zl_logbias = logbias; 1747} 1748 1749zilog_t * 1750zil_alloc(objset_t *os, zil_header_t *zh_phys) 1751{ 1752 zilog_t *zilog; 1753 1754 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1755 1756 zilog->zl_header = zh_phys; 1757 zilog->zl_os = os; 1758 zilog->zl_spa = dmu_objset_spa(os); 1759 zilog->zl_dmu_pool = dmu_objset_pool(os); 1760 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1761 zilog->zl_logbias = dmu_objset_logbias(os); 1762 zilog->zl_sync = dmu_objset_syncprop(os); 1763 zilog->zl_next_batch = 1; 1764 1765 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1766 1767 for (int i = 0; i < TXG_SIZE; i++) { 1768 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1769 MUTEX_DEFAULT, NULL); 1770 } 1771 1772 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1773 offsetof(lwb_t, lwb_node)); 1774 1775 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1776 offsetof(itx_t, itx_node)); 1777 1778 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1779 1780 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1781 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1782 1783 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1784 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1785 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1786 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1787 1788 return (zilog); 1789} 1790 1791void 1792zil_free(zilog_t *zilog) 1793{ 1794 zilog->zl_stop_sync = 1; 1795 1796 ASSERT0(zilog->zl_suspend); 1797 ASSERT0(zilog->zl_suspending); 1798 1799 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1800 list_destroy(&zilog->zl_lwb_list); 1801 1802 avl_destroy(&zilog->zl_vdev_tree); 1803 mutex_destroy(&zilog->zl_vdev_lock); 1804 1805 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1806 list_destroy(&zilog->zl_itx_commit_list); 1807 1808 for (int i = 0; i < TXG_SIZE; i++) { 1809 /* 1810 * It's possible for an itx to be generated that doesn't dirty 1811 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1812 * callback to remove the entry. We remove those here. 1813 * 1814 * Also free up the ziltest itxs. 1815 */ 1816 if (zilog->zl_itxg[i].itxg_itxs) 1817 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1818 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1819 } 1820 1821 mutex_destroy(&zilog->zl_lock); 1822 1823 cv_destroy(&zilog->zl_cv_writer); 1824 cv_destroy(&zilog->zl_cv_suspend); 1825 cv_destroy(&zilog->zl_cv_batch[0]); 1826 cv_destroy(&zilog->zl_cv_batch[1]); 1827 1828 kmem_free(zilog, sizeof (zilog_t)); 1829} 1830 1831/* 1832 * Open an intent log. 1833 */ 1834zilog_t * 1835zil_open(objset_t *os, zil_get_data_t *get_data) 1836{ 1837 zilog_t *zilog = dmu_objset_zil(os); 1838 1839 ASSERT(zilog->zl_clean_taskq == NULL); 1840 ASSERT(zilog->zl_get_data == NULL); 1841 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1842 1843 zilog->zl_get_data = get_data; 1844 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1845 2, 2, TASKQ_PREPOPULATE); 1846 1847 return (zilog); 1848} 1849 1850/* 1851 * Close an intent log. 1852 */ 1853void 1854zil_close(zilog_t *zilog) 1855{ 1856 lwb_t *lwb; 1857 uint64_t txg = 0; 1858 1859 zil_commit(zilog, 0); /* commit all itx */ 1860 1861 /* 1862 * The lwb_max_txg for the stubby lwb will reflect the last activity 1863 * for the zil. After a txg_wait_synced() on the txg we know all the 1864 * callbacks have occurred that may clean the zil. Only then can we 1865 * destroy the zl_clean_taskq. 1866 */ 1867 mutex_enter(&zilog->zl_lock); 1868 lwb = list_tail(&zilog->zl_lwb_list); 1869 if (lwb != NULL) 1870 txg = lwb->lwb_max_txg; 1871 mutex_exit(&zilog->zl_lock); 1872 if (txg) 1873 txg_wait_synced(zilog->zl_dmu_pool, txg); 1874 1875 if (zilog_is_dirty(zilog)) 1876 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg); 1877 VERIFY(!zilog_is_dirty(zilog)); 1878 1879 taskq_destroy(zilog->zl_clean_taskq); 1880 zilog->zl_clean_taskq = NULL; 1881 zilog->zl_get_data = NULL; 1882 1883 /* 1884 * We should have only one LWB left on the list; remove it now. 1885 */ 1886 mutex_enter(&zilog->zl_lock); 1887 lwb = list_head(&zilog->zl_lwb_list); 1888 if (lwb != NULL) { 1889 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1890 list_remove(&zilog->zl_lwb_list, lwb); 1891 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1892 kmem_cache_free(zil_lwb_cache, lwb); 1893 } 1894 mutex_exit(&zilog->zl_lock); 1895} 1896 1897static char *suspend_tag = "zil suspending"; 1898 1899/* 1900 * Suspend an intent log. While in suspended mode, we still honor 1901 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1902 * On old version pools, we suspend the log briefly when taking a 1903 * snapshot so that it will have an empty intent log. 1904 * 1905 * Long holds are not really intended to be used the way we do here -- 1906 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1907 * could fail. Therefore we take pains to only put a long hold if it is 1908 * actually necessary. Fortunately, it will only be necessary if the 1909 * objset is currently mounted (or the ZVOL equivalent). In that case it 1910 * will already have a long hold, so we are not really making things any worse. 1911 * 1912 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1913 * zvol_state_t), and use their mechanism to prevent their hold from being 1914 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1915 * very little gain. 1916 * 1917 * if cookiep == NULL, this does both the suspend & resume. 1918 * Otherwise, it returns with the dataset "long held", and the cookie 1919 * should be passed into zil_resume(). 1920 */ 1921int 1922zil_suspend(const char *osname, void **cookiep) 1923{ 1924 objset_t *os; 1925 zilog_t *zilog; 1926 const zil_header_t *zh; 1927 int error; 1928 1929 error = dmu_objset_hold(osname, suspend_tag, &os); 1930 if (error != 0) 1931 return (error); 1932 zilog = dmu_objset_zil(os); 1933 1934 mutex_enter(&zilog->zl_lock); 1935 zh = zilog->zl_header; 1936 1937 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1938 mutex_exit(&zilog->zl_lock); 1939 dmu_objset_rele(os, suspend_tag); 1940 return (SET_ERROR(EBUSY)); 1941 } 1942 1943 /* 1944 * Don't put a long hold in the cases where we can avoid it. This 1945 * is when there is no cookie so we are doing a suspend & resume 1946 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1947 * for the suspend because it's already suspended, or there's no ZIL. 1948 */ 1949 if (cookiep == NULL && !zilog->zl_suspending && 1950 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1951 mutex_exit(&zilog->zl_lock); 1952 dmu_objset_rele(os, suspend_tag); 1953 return (0); 1954 } 1955 1956 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1957 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1958 1959 zilog->zl_suspend++; 1960 1961 if (zilog->zl_suspend > 1) { 1962 /* 1963 * Someone else is already suspending it. 1964 * Just wait for them to finish. 1965 */ 1966 1967 while (zilog->zl_suspending) 1968 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1969 mutex_exit(&zilog->zl_lock); 1970 1971 if (cookiep == NULL) 1972 zil_resume(os); 1973 else 1974 *cookiep = os; 1975 return (0); 1976 } 1977 1978 /* 1979 * If there is no pointer to an on-disk block, this ZIL must not 1980 * be active (e.g. filesystem not mounted), so there's nothing 1981 * to clean up. 1982 */ 1983 if (BP_IS_HOLE(&zh->zh_log)) { 1984 ASSERT(cookiep != NULL); /* fast path already handled */ 1985 1986 *cookiep = os; 1987 mutex_exit(&zilog->zl_lock); 1988 return (0); 1989 } 1990 1991 zilog->zl_suspending = B_TRUE; 1992 mutex_exit(&zilog->zl_lock); 1993 1994 zil_commit(zilog, 0); 1995 1996 zil_destroy(zilog, B_FALSE); 1997 1998 mutex_enter(&zilog->zl_lock); 1999 zilog->zl_suspending = B_FALSE; 2000 cv_broadcast(&zilog->zl_cv_suspend); 2001 mutex_exit(&zilog->zl_lock); 2002 2003 if (cookiep == NULL) 2004 zil_resume(os); 2005 else 2006 *cookiep = os; 2007 return (0); 2008} 2009 2010void 2011zil_resume(void *cookie) 2012{ 2013 objset_t *os = cookie; 2014 zilog_t *zilog = dmu_objset_zil(os); 2015 2016 mutex_enter(&zilog->zl_lock); 2017 ASSERT(zilog->zl_suspend != 0); 2018 zilog->zl_suspend--; 2019 mutex_exit(&zilog->zl_lock); 2020 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 2021 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 2022} 2023 2024typedef struct zil_replay_arg { 2025 zil_replay_func_t **zr_replay; 2026 void *zr_arg; 2027 boolean_t zr_byteswap; 2028 char *zr_lr; 2029} zil_replay_arg_t; 2030 2031static int 2032zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 2033{ 2034 char name[ZFS_MAX_DATASET_NAME_LEN]; 2035 2036 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 2037 2038 dmu_objset_name(zilog->zl_os, name); 2039 2040 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 2041 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 2042 (u_longlong_t)lr->lrc_seq, 2043 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 2044 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 2045 2046 return (error); 2047} 2048 2049static int 2050zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 2051{ 2052 zil_replay_arg_t *zr = zra; 2053 const zil_header_t *zh = zilog->zl_header; 2054 uint64_t reclen = lr->lrc_reclen; 2055 uint64_t txtype = lr->lrc_txtype; 2056 int error = 0; 2057 2058 zilog->zl_replaying_seq = lr->lrc_seq; 2059 2060 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2061 return (0); 2062 2063 if (lr->lrc_txg < claim_txg) /* already committed */ 2064 return (0); 2065 2066 /* Strip case-insensitive bit, still present in log record */ 2067 txtype &= ~TX_CI; 2068 2069 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2070 return (zil_replay_error(zilog, lr, EINVAL)); 2071 2072 /* 2073 * If this record type can be logged out of order, the object 2074 * (lr_foid) may no longer exist. That's legitimate, not an error. 2075 */ 2076 if (TX_OOO(txtype)) { 2077 error = dmu_object_info(zilog->zl_os, 2078 ((lr_ooo_t *)lr)->lr_foid, NULL); 2079 if (error == ENOENT || error == EEXIST) 2080 return (0); 2081 } 2082 2083 /* 2084 * Make a copy of the data so we can revise and extend it. 2085 */ 2086 bcopy(lr, zr->zr_lr, reclen); 2087 2088 /* 2089 * If this is a TX_WRITE with a blkptr, suck in the data. 2090 */ 2091 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2092 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2093 zr->zr_lr + reclen); 2094 if (error != 0) 2095 return (zil_replay_error(zilog, lr, error)); 2096 } 2097 2098 /* 2099 * The log block containing this lr may have been byteswapped 2100 * so that we can easily examine common fields like lrc_txtype. 2101 * However, the log is a mix of different record types, and only the 2102 * replay vectors know how to byteswap their records. Therefore, if 2103 * the lr was byteswapped, undo it before invoking the replay vector. 2104 */ 2105 if (zr->zr_byteswap) 2106 byteswap_uint64_array(zr->zr_lr, reclen); 2107 2108 /* 2109 * We must now do two things atomically: replay this log record, 2110 * and update the log header sequence number to reflect the fact that 2111 * we did so. At the end of each replay function the sequence number 2112 * is updated if we are in replay mode. 2113 */ 2114 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2115 if (error != 0) { 2116 /* 2117 * The DMU's dnode layer doesn't see removes until the txg 2118 * commits, so a subsequent claim can spuriously fail with 2119 * EEXIST. So if we receive any error we try syncing out 2120 * any removes then retry the transaction. Note that we 2121 * specify B_FALSE for byteswap now, so we don't do it twice. 2122 */ 2123 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2124 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2125 if (error != 0) 2126 return (zil_replay_error(zilog, lr, error)); 2127 } 2128 return (0); 2129} 2130 2131/* ARGSUSED */ 2132static int 2133zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2134{ 2135 zilog->zl_replay_blks++; 2136 2137 return (0); 2138} 2139 2140/* 2141 * If this dataset has a non-empty intent log, replay it and destroy it. 2142 */ 2143void 2144zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2145{ 2146 zilog_t *zilog = dmu_objset_zil(os); 2147 const zil_header_t *zh = zilog->zl_header; 2148 zil_replay_arg_t zr; 2149 2150 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2151 zil_destroy(zilog, B_TRUE); 2152 return; 2153 } 2154 2155 zr.zr_replay = replay_func; 2156 zr.zr_arg = arg; 2157 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2158 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2159 2160 /* 2161 * Wait for in-progress removes to sync before starting replay. 2162 */ 2163 txg_wait_synced(zilog->zl_dmu_pool, 0); 2164 2165 zilog->zl_replay = B_TRUE; 2166 zilog->zl_replay_time = ddi_get_lbolt(); 2167 ASSERT(zilog->zl_replay_blks == 0); 2168 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2169 zh->zh_claim_txg); 2170 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2171 2172 zil_destroy(zilog, B_FALSE); 2173 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2174 zilog->zl_replay = B_FALSE; 2175} 2176 2177boolean_t 2178zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2179{ 2180 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2181 return (B_TRUE); 2182 2183 if (zilog->zl_replay) { 2184 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2185 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2186 zilog->zl_replaying_seq; 2187 return (B_TRUE); 2188 } 2189 2190 return (B_FALSE); 2191} 2192 2193/* ARGSUSED */ 2194int 2195zil_vdev_offline(const char *osname, void *arg) 2196{ 2197 int error; 2198 2199 error = zil_suspend(osname, NULL); 2200 if (error != 0) 2201 return (SET_ERROR(EEXIST)); 2202 return (0); 2203} 2204