zil.c revision 324205
1330449Seadler/* 2124758Semax * CDDL HEADER START 3124758Semax * 4330449Seadler * The contents of this file are subject to the terms of the 5330449Seadler * Common Development and Distribution License (the "License"). 6124758Semax * You may not use this file except in compliance with the License. 7124758Semax * 8124758Semax * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9124758Semax * or http://www.opensolaris.org/os/licensing. 10124758Semax * See the License for the specific language governing permissions 11124758Semax * and limitations under the License. 12124758Semax * 13124758Semax * When distributing Covered Code, include this CDDL HEADER in each 14124758Semax * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15124758Semax * If applicable, add the following below this CDDL HEADER, with the 16124758Semax * fields enclosed by brackets "[]" replaced with your own identifying 17124758Semax * information: Portions Copyright [yyyy] [name of copyright owner] 18124758Semax * 19124758Semax * CDDL HEADER END 20124758Semax */ 21124758Semax/* 22124758Semax * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23124758Semax * Copyright (c) 2011, 2017 by Delphix. All rights reserved. 24124758Semax * Copyright (c) 2014 Integros [integros.com] 25124758Semax */ 26124758Semax 27124758Semax/* Portions Copyright 2010 Robert Milkowski */ 28124758Semax 29124758Semax#include <sys/zfs_context.h> 30124758Semax#include <sys/spa.h> 31124758Semax#include <sys/dmu.h> 32124758Semax#include <sys/zap.h> 33124758Semax#include <sys/arc.h> 34124758Semax#include <sys/stat.h> 35124758Semax#include <sys/resource.h> 36124758Semax#include <sys/zil.h> 37124758Semax#include <sys/zil_impl.h> 38124758Semax#include <sys/dsl_dataset.h> 39124758Semax#include <sys/vdev_impl.h> 40124758Semax#include <sys/dmu_tx.h> 41124758Semax#include <sys/dsl_pool.h> 42124758Semax#include <sys/abd.h> 43124758Semax 44124758Semax/* 45124758Semax * The zfs intent log (ZIL) saves transaction records of system calls 46124758Semax * that change the file system in memory with enough information 47124758Semax * to be able to replay them. These are stored in memory until 48124758Semax * either the DMU transaction group (txg) commits them to the stable pool 49124758Semax * and they can be discarded, or they are flushed to the stable log 50 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 51 * requirement. In the event of a panic or power fail then those log 52 * records (transactions) are replayed. 53 * 54 * There is one ZIL per file system. Its on-disk (pool) format consists 55 * of 3 parts: 56 * 57 * - ZIL header 58 * - ZIL blocks 59 * - ZIL records 60 * 61 * A log record holds a system call transaction. Log blocks can 62 * hold many log records and the blocks are chained together. 63 * Each ZIL block contains a block pointer (blkptr_t) to the next 64 * ZIL block in the chain. The ZIL header points to the first 65 * block in the chain. Note there is not a fixed place in the pool 66 * to hold blocks. They are dynamically allocated and freed as 67 * needed from the blocks available. Figure X shows the ZIL structure: 68 */ 69 70/* 71 * Disable intent logging replay. This global ZIL switch affects all pools. 72 */ 73int zil_replay_disable = 0; 74SYSCTL_DECL(_vfs_zfs); 75SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN, 76 &zil_replay_disable, 0, "Disable intent logging replay"); 77 78/* 79 * Tunable parameter for debugging or performance analysis. Setting 80 * zfs_nocacheflush will cause corruption on power loss if a volatile 81 * out-of-order write cache is enabled. 82 */ 83boolean_t zfs_nocacheflush = B_FALSE; 84SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 85 &zfs_nocacheflush, 0, "Disable cache flush"); 86boolean_t zfs_trim_enabled = B_TRUE; 87SYSCTL_DECL(_vfs_zfs_trim); 88SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 89 "Enable ZFS TRIM"); 90 91/* 92 * Limit SLOG write size per commit executed with synchronous priority. 93 * Any writes above that will be executed with lower (asynchronous) priority 94 * to limit potential SLOG device abuse by single active ZIL writer. 95 */ 96uint64_t zil_slog_bulk = 768 * 1024; 97SYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_bulk, CTLFLAG_RWTUN, 98 &zil_slog_bulk, 0, "Maximal SLOG commit size with sync priority"); 99 100static kmem_cache_t *zil_lwb_cache; 101 102#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 103 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 104 105static int 106zil_bp_compare(const void *x1, const void *x2) 107{ 108 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 109 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 110 111 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 112 return (-1); 113 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 114 return (1); 115 116 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 117 return (-1); 118 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 119 return (1); 120 121 return (0); 122} 123 124static void 125zil_bp_tree_init(zilog_t *zilog) 126{ 127 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 128 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 129} 130 131static void 132zil_bp_tree_fini(zilog_t *zilog) 133{ 134 avl_tree_t *t = &zilog->zl_bp_tree; 135 zil_bp_node_t *zn; 136 void *cookie = NULL; 137 138 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 139 kmem_free(zn, sizeof (zil_bp_node_t)); 140 141 avl_destroy(t); 142} 143 144int 145zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 146{ 147 avl_tree_t *t = &zilog->zl_bp_tree; 148 const dva_t *dva; 149 zil_bp_node_t *zn; 150 avl_index_t where; 151 152 if (BP_IS_EMBEDDED(bp)) 153 return (0); 154 155 dva = BP_IDENTITY(bp); 156 157 if (avl_find(t, dva, &where) != NULL) 158 return (SET_ERROR(EEXIST)); 159 160 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 161 zn->zn_dva = *dva; 162 avl_insert(t, zn, where); 163 164 return (0); 165} 166 167static zil_header_t * 168zil_header_in_syncing_context(zilog_t *zilog) 169{ 170 return ((zil_header_t *)zilog->zl_header); 171} 172 173static void 174zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 175{ 176 zio_cksum_t *zc = &bp->blk_cksum; 177 178 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 179 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 180 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 181 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 182} 183 184/* 185 * Read a log block and make sure it's valid. 186 */ 187static int 188zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 189 char **end) 190{ 191 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 192 arc_flags_t aflags = ARC_FLAG_WAIT; 193 arc_buf_t *abuf = NULL; 194 zbookmark_phys_t zb; 195 int error; 196 197 if (zilog->zl_header->zh_claim_txg == 0) 198 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 199 200 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 201 zio_flags |= ZIO_FLAG_SPECULATIVE; 202 203 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 204 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 205 206 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 207 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 208 209 if (error == 0) { 210 zio_cksum_t cksum = bp->blk_cksum; 211 212 /* 213 * Validate the checksummed log block. 214 * 215 * Sequence numbers should be... sequential. The checksum 216 * verifier for the next block should be bp's checksum plus 1. 217 * 218 * Also check the log chain linkage and size used. 219 */ 220 cksum.zc_word[ZIL_ZC_SEQ]++; 221 222 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 223 zil_chain_t *zilc = abuf->b_data; 224 char *lr = (char *)(zilc + 1); 225 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 226 227 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 228 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 229 error = SET_ERROR(ECKSUM); 230 } else { 231 ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 232 bcopy(lr, dst, len); 233 *end = (char *)dst + len; 234 *nbp = zilc->zc_next_blk; 235 } 236 } else { 237 char *lr = abuf->b_data; 238 uint64_t size = BP_GET_LSIZE(bp); 239 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 240 241 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 242 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 243 (zilc->zc_nused > (size - sizeof (*zilc)))) { 244 error = SET_ERROR(ECKSUM); 245 } else { 246 ASSERT3U(zilc->zc_nused, <=, 247 SPA_OLD_MAXBLOCKSIZE); 248 bcopy(lr, dst, zilc->zc_nused); 249 *end = (char *)dst + zilc->zc_nused; 250 *nbp = zilc->zc_next_blk; 251 } 252 } 253 254 arc_buf_destroy(abuf, &abuf); 255 } 256 257 return (error); 258} 259 260/* 261 * Read a TX_WRITE log data block. 262 */ 263static int 264zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 265{ 266 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 267 const blkptr_t *bp = &lr->lr_blkptr; 268 arc_flags_t aflags = ARC_FLAG_WAIT; 269 arc_buf_t *abuf = NULL; 270 zbookmark_phys_t zb; 271 int error; 272 273 if (BP_IS_HOLE(bp)) { 274 if (wbuf != NULL) 275 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 276 return (0); 277 } 278 279 if (zilog->zl_header->zh_claim_txg == 0) 280 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 281 282 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 283 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 284 285 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 286 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 287 288 if (error == 0) { 289 if (wbuf != NULL) 290 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 291 arc_buf_destroy(abuf, &abuf); 292 } 293 294 return (error); 295} 296 297/* 298 * Parse the intent log, and call parse_func for each valid record within. 299 */ 300int 301zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 302 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 303{ 304 const zil_header_t *zh = zilog->zl_header; 305 boolean_t claimed = !!zh->zh_claim_txg; 306 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 307 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 308 uint64_t max_blk_seq = 0; 309 uint64_t max_lr_seq = 0; 310 uint64_t blk_count = 0; 311 uint64_t lr_count = 0; 312 blkptr_t blk, next_blk; 313 char *lrbuf, *lrp; 314 int error = 0; 315 316 /* 317 * Old logs didn't record the maximum zh_claim_lr_seq. 318 */ 319 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 320 claim_lr_seq = UINT64_MAX; 321 322 /* 323 * Starting at the block pointed to by zh_log we read the log chain. 324 * For each block in the chain we strongly check that block to 325 * ensure its validity. We stop when an invalid block is found. 326 * For each block pointer in the chain we call parse_blk_func(). 327 * For each record in each valid block we call parse_lr_func(). 328 * If the log has been claimed, stop if we encounter a sequence 329 * number greater than the highest claimed sequence number. 330 */ 331 lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 332 zil_bp_tree_init(zilog); 333 334 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 335 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 336 int reclen; 337 char *end; 338 339 if (blk_seq > claim_blk_seq) 340 break; 341 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 342 break; 343 ASSERT3U(max_blk_seq, <, blk_seq); 344 max_blk_seq = blk_seq; 345 blk_count++; 346 347 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 348 break; 349 350 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 351 if (error != 0) 352 break; 353 354 for (lrp = lrbuf; lrp < end; lrp += reclen) { 355 lr_t *lr = (lr_t *)lrp; 356 reclen = lr->lrc_reclen; 357 ASSERT3U(reclen, >=, sizeof (lr_t)); 358 if (lr->lrc_seq > claim_lr_seq) 359 goto done; 360 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 361 goto done; 362 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 363 max_lr_seq = lr->lrc_seq; 364 lr_count++; 365 } 366 } 367done: 368 zilog->zl_parse_error = error; 369 zilog->zl_parse_blk_seq = max_blk_seq; 370 zilog->zl_parse_lr_seq = max_lr_seq; 371 zilog->zl_parse_blk_count = blk_count; 372 zilog->zl_parse_lr_count = lr_count; 373 374 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 375 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 376 377 zil_bp_tree_fini(zilog); 378 zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 379 380 return (error); 381} 382 383static int 384zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 385{ 386 /* 387 * Claim log block if not already committed and not already claimed. 388 * If tx == NULL, just verify that the block is claimable. 389 */ 390 if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 391 zil_bp_tree_add(zilog, bp) != 0) 392 return (0); 393 394 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 395 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 396 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 397} 398 399static int 400zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 401{ 402 lr_write_t *lr = (lr_write_t *)lrc; 403 int error; 404 405 if (lrc->lrc_txtype != TX_WRITE) 406 return (0); 407 408 /* 409 * If the block is not readable, don't claim it. This can happen 410 * in normal operation when a log block is written to disk before 411 * some of the dmu_sync() blocks it points to. In this case, the 412 * transaction cannot have been committed to anyone (we would have 413 * waited for all writes to be stable first), so it is semantically 414 * correct to declare this the end of the log. 415 */ 416 if (lr->lr_blkptr.blk_birth >= first_txg && 417 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 418 return (error); 419 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 420} 421 422/* ARGSUSED */ 423static int 424zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 425{ 426 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 427 428 return (0); 429} 430 431static int 432zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 433{ 434 lr_write_t *lr = (lr_write_t *)lrc; 435 blkptr_t *bp = &lr->lr_blkptr; 436 437 /* 438 * If we previously claimed it, we need to free it. 439 */ 440 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 441 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 442 !BP_IS_HOLE(bp)) 443 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 444 445 return (0); 446} 447 448static lwb_t * 449zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg) 450{ 451 lwb_t *lwb; 452 453 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 454 lwb->lwb_zilog = zilog; 455 lwb->lwb_blk = *bp; 456 lwb->lwb_slog = slog; 457 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 458 lwb->lwb_max_txg = txg; 459 lwb->lwb_zio = NULL; 460 lwb->lwb_tx = NULL; 461 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 462 lwb->lwb_nused = sizeof (zil_chain_t); 463 lwb->lwb_sz = BP_GET_LSIZE(bp); 464 } else { 465 lwb->lwb_nused = 0; 466 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 467 } 468 469 mutex_enter(&zilog->zl_lock); 470 list_insert_tail(&zilog->zl_lwb_list, lwb); 471 mutex_exit(&zilog->zl_lock); 472 473 return (lwb); 474} 475 476/* 477 * Called when we create in-memory log transactions so that we know 478 * to cleanup the itxs at the end of spa_sync(). 479 */ 480void 481zilog_dirty(zilog_t *zilog, uint64_t txg) 482{ 483 dsl_pool_t *dp = zilog->zl_dmu_pool; 484 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 485 486 if (ds->ds_is_snapshot) 487 panic("dirtying snapshot!"); 488 489 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 490 /* up the hold count until we can be written out */ 491 dmu_buf_add_ref(ds->ds_dbuf, zilog); 492 } 493} 494 495/* 496 * Determine if the zil is dirty in the specified txg. Callers wanting to 497 * ensure that the dirty state does not change must hold the itxg_lock for 498 * the specified txg. Holding the lock will ensure that the zil cannot be 499 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 500 * state. 501 */ 502boolean_t 503zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 504{ 505 dsl_pool_t *dp = zilog->zl_dmu_pool; 506 507 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 508 return (B_TRUE); 509 return (B_FALSE); 510} 511 512/* 513 * Determine if the zil is dirty. The zil is considered dirty if it has 514 * any pending itx records that have not been cleaned by zil_clean(). 515 */ 516boolean_t 517zilog_is_dirty(zilog_t *zilog) 518{ 519 dsl_pool_t *dp = zilog->zl_dmu_pool; 520 521 for (int t = 0; t < TXG_SIZE; t++) { 522 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 523 return (B_TRUE); 524 } 525 return (B_FALSE); 526} 527 528/* 529 * Create an on-disk intent log. 530 */ 531static lwb_t * 532zil_create(zilog_t *zilog) 533{ 534 const zil_header_t *zh = zilog->zl_header; 535 lwb_t *lwb = NULL; 536 uint64_t txg = 0; 537 dmu_tx_t *tx = NULL; 538 blkptr_t blk; 539 int error = 0; 540 boolean_t slog = FALSE; 541 542 /* 543 * Wait for any previous destroy to complete. 544 */ 545 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 546 547 ASSERT(zh->zh_claim_txg == 0); 548 ASSERT(zh->zh_replay_seq == 0); 549 550 blk = zh->zh_log; 551 552 /* 553 * Allocate an initial log block if: 554 * - there isn't one already 555 * - the existing block is the wrong endianess 556 */ 557 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 558 tx = dmu_tx_create(zilog->zl_os); 559 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 560 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 561 txg = dmu_tx_get_txg(tx); 562 563 if (!BP_IS_HOLE(&blk)) { 564 zio_free_zil(zilog->zl_spa, txg, &blk); 565 BP_ZERO(&blk); 566 } 567 568 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 569 ZIL_MIN_BLKSZ, &slog); 570 571 if (error == 0) 572 zil_init_log_chain(zilog, &blk); 573 } 574 575 /* 576 * Allocate a log write buffer (lwb) for the first log block. 577 */ 578 if (error == 0) 579 lwb = zil_alloc_lwb(zilog, &blk, slog, txg); 580 581 /* 582 * If we just allocated the first log block, commit our transaction 583 * and wait for zil_sync() to stuff the block poiner into zh_log. 584 * (zh is part of the MOS, so we cannot modify it in open context.) 585 */ 586 if (tx != NULL) { 587 dmu_tx_commit(tx); 588 txg_wait_synced(zilog->zl_dmu_pool, txg); 589 } 590 591 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 592 593 return (lwb); 594} 595 596/* 597 * In one tx, free all log blocks and clear the log header. 598 * If keep_first is set, then we're replaying a log with no content. 599 * We want to keep the first block, however, so that the first 600 * synchronous transaction doesn't require a txg_wait_synced() 601 * in zil_create(). We don't need to txg_wait_synced() here either 602 * when keep_first is set, because both zil_create() and zil_destroy() 603 * will wait for any in-progress destroys to complete. 604 */ 605void 606zil_destroy(zilog_t *zilog, boolean_t keep_first) 607{ 608 const zil_header_t *zh = zilog->zl_header; 609 lwb_t *lwb; 610 dmu_tx_t *tx; 611 uint64_t txg; 612 613 /* 614 * Wait for any previous destroy to complete. 615 */ 616 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 617 618 zilog->zl_old_header = *zh; /* debugging aid */ 619 620 if (BP_IS_HOLE(&zh->zh_log)) 621 return; 622 623 tx = dmu_tx_create(zilog->zl_os); 624 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 625 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 626 txg = dmu_tx_get_txg(tx); 627 628 mutex_enter(&zilog->zl_lock); 629 630 ASSERT3U(zilog->zl_destroy_txg, <, txg); 631 zilog->zl_destroy_txg = txg; 632 zilog->zl_keep_first = keep_first; 633 634 if (!list_is_empty(&zilog->zl_lwb_list)) { 635 ASSERT(zh->zh_claim_txg == 0); 636 VERIFY(!keep_first); 637 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 638 list_remove(&zilog->zl_lwb_list, lwb); 639 if (lwb->lwb_buf != NULL) 640 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 641 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 642 kmem_cache_free(zil_lwb_cache, lwb); 643 } 644 } else if (!keep_first) { 645 zil_destroy_sync(zilog, tx); 646 } 647 mutex_exit(&zilog->zl_lock); 648 649 dmu_tx_commit(tx); 650} 651 652void 653zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 654{ 655 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 656 (void) zil_parse(zilog, zil_free_log_block, 657 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 658} 659 660int 661zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 662{ 663 dmu_tx_t *tx = txarg; 664 uint64_t first_txg = dmu_tx_get_txg(tx); 665 zilog_t *zilog; 666 zil_header_t *zh; 667 objset_t *os; 668 int error; 669 670 error = dmu_objset_own_obj(dp, ds->ds_object, 671 DMU_OST_ANY, B_FALSE, FTAG, &os); 672 if (error != 0) { 673 /* 674 * EBUSY indicates that the objset is inconsistent, in which 675 * case it can not have a ZIL. 676 */ 677 if (error != EBUSY) { 678 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 679 (unsigned long long)ds->ds_object, error); 680 } 681 return (0); 682 } 683 684 zilog = dmu_objset_zil(os); 685 zh = zil_header_in_syncing_context(zilog); 686 687 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 688 if (!BP_IS_HOLE(&zh->zh_log)) 689 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 690 BP_ZERO(&zh->zh_log); 691 dsl_dataset_dirty(dmu_objset_ds(os), tx); 692 dmu_objset_disown(os, FTAG); 693 return (0); 694 } 695 696 /* 697 * Claim all log blocks if we haven't already done so, and remember 698 * the highest claimed sequence number. This ensures that if we can 699 * read only part of the log now (e.g. due to a missing device), 700 * but we can read the entire log later, we will not try to replay 701 * or destroy beyond the last block we successfully claimed. 702 */ 703 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 704 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 705 (void) zil_parse(zilog, zil_claim_log_block, 706 zil_claim_log_record, tx, first_txg); 707 zh->zh_claim_txg = first_txg; 708 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 709 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 710 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 711 zh->zh_flags |= ZIL_REPLAY_NEEDED; 712 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 713 dsl_dataset_dirty(dmu_objset_ds(os), tx); 714 } 715 716 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 717 dmu_objset_disown(os, FTAG); 718 return (0); 719} 720 721/* 722 * Check the log by walking the log chain. 723 * Checksum errors are ok as they indicate the end of the chain. 724 * Any other error (no device or read failure) returns an error. 725 */ 726/* ARGSUSED */ 727int 728zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 729{ 730 zilog_t *zilog; 731 objset_t *os; 732 blkptr_t *bp; 733 int error; 734 735 ASSERT(tx == NULL); 736 737 error = dmu_objset_from_ds(ds, &os); 738 if (error != 0) { 739 cmn_err(CE_WARN, "can't open objset %llu, error %d", 740 (unsigned long long)ds->ds_object, error); 741 return (0); 742 } 743 744 zilog = dmu_objset_zil(os); 745 bp = (blkptr_t *)&zilog->zl_header->zh_log; 746 747 /* 748 * Check the first block and determine if it's on a log device 749 * which may have been removed or faulted prior to loading this 750 * pool. If so, there's no point in checking the rest of the log 751 * as its content should have already been synced to the pool. 752 */ 753 if (!BP_IS_HOLE(bp)) { 754 vdev_t *vd; 755 boolean_t valid = B_TRUE; 756 757 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 758 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 759 if (vd->vdev_islog && vdev_is_dead(vd)) 760 valid = vdev_log_state_valid(vd); 761 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 762 763 if (!valid) 764 return (0); 765 } 766 767 /* 768 * Because tx == NULL, zil_claim_log_block() will not actually claim 769 * any blocks, but just determine whether it is possible to do so. 770 * In addition to checking the log chain, zil_claim_log_block() 771 * will invoke zio_claim() with a done func of spa_claim_notify(), 772 * which will update spa_max_claim_txg. See spa_load() for details. 773 */ 774 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 775 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 776 777 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 778} 779 780static int 781zil_vdev_compare(const void *x1, const void *x2) 782{ 783 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 784 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 785 786 if (v1 < v2) 787 return (-1); 788 if (v1 > v2) 789 return (1); 790 791 return (0); 792} 793 794void 795zil_add_block(zilog_t *zilog, const blkptr_t *bp) 796{ 797 avl_tree_t *t = &zilog->zl_vdev_tree; 798 avl_index_t where; 799 zil_vdev_node_t *zv, zvsearch; 800 int ndvas = BP_GET_NDVAS(bp); 801 int i; 802 803 if (zfs_nocacheflush) 804 return; 805 806 ASSERT(zilog->zl_writer); 807 808 /* 809 * Even though we're zl_writer, we still need a lock because the 810 * zl_get_data() callbacks may have dmu_sync() done callbacks 811 * that will run concurrently. 812 */ 813 mutex_enter(&zilog->zl_vdev_lock); 814 for (i = 0; i < ndvas; i++) { 815 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 816 if (avl_find(t, &zvsearch, &where) == NULL) { 817 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 818 zv->zv_vdev = zvsearch.zv_vdev; 819 avl_insert(t, zv, where); 820 } 821 } 822 mutex_exit(&zilog->zl_vdev_lock); 823} 824 825static void 826zil_flush_vdevs(zilog_t *zilog) 827{ 828 spa_t *spa = zilog->zl_spa; 829 avl_tree_t *t = &zilog->zl_vdev_tree; 830 void *cookie = NULL; 831 zil_vdev_node_t *zv; 832 zio_t *zio = NULL; 833 834 ASSERT(zilog->zl_writer); 835 836 /* 837 * We don't need zl_vdev_lock here because we're the zl_writer, 838 * and all zl_get_data() callbacks are done. 839 */ 840 if (avl_numnodes(t) == 0) 841 return; 842 843 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 844 845 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 846 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 847 if (vd != NULL && !vd->vdev_nowritecache) { 848 if (zio == NULL) 849 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 850 zio_flush(zio, vd); 851 } 852 kmem_free(zv, sizeof (*zv)); 853 } 854 855 /* 856 * Wait for all the flushes to complete. Not all devices actually 857 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 858 */ 859 if (zio) 860 (void) zio_wait(zio); 861 862 spa_config_exit(spa, SCL_STATE, FTAG); 863} 864 865/* 866 * Function called when a log block write completes 867 */ 868static void 869zil_lwb_write_done(zio_t *zio) 870{ 871 lwb_t *lwb = zio->io_private; 872 zilog_t *zilog = lwb->lwb_zilog; 873 dmu_tx_t *tx = lwb->lwb_tx; 874 875 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 876 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 877 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 878 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 879 ASSERT(!BP_IS_GANG(zio->io_bp)); 880 ASSERT(!BP_IS_HOLE(zio->io_bp)); 881 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 882 883 /* 884 * Ensure the lwb buffer pointer is cleared before releasing 885 * the txg. If we have had an allocation failure and 886 * the txg is waiting to sync then we want want zil_sync() 887 * to remove the lwb so that it's not picked up as the next new 888 * one in zil_commit_writer(). zil_sync() will only remove 889 * the lwb if lwb_buf is null. 890 */ 891 abd_put(zio->io_abd); 892 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 893 mutex_enter(&zilog->zl_lock); 894 lwb->lwb_buf = NULL; 895 lwb->lwb_tx = NULL; 896 mutex_exit(&zilog->zl_lock); 897 898 /* 899 * Now that we've written this log block, we have a stable pointer 900 * to the next block in the chain, so it's OK to let the txg in 901 * which we allocated the next block sync. 902 */ 903 dmu_tx_commit(tx); 904} 905 906/* 907 * Initialize the io for a log block. 908 */ 909static void 910zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 911{ 912 zbookmark_phys_t zb; 913 zio_priority_t prio; 914 915 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 916 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 917 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 918 919 if (zilog->zl_root_zio == NULL) { 920 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 921 ZIO_FLAG_CANFAIL); 922 } 923 if (lwb->lwb_zio == NULL) { 924 abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf, 925 BP_GET_LSIZE(&lwb->lwb_blk)); 926 if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk) 927 prio = ZIO_PRIORITY_SYNC_WRITE; 928 else 929 prio = ZIO_PRIORITY_ASYNC_WRITE; 930 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 931 0, &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk), 932 zil_lwb_write_done, lwb, prio, 933 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 934 } 935} 936 937/* 938 * Define a limited set of intent log block sizes. 939 * 940 * These must be a multiple of 4KB. Note only the amount used (again 941 * aligned to 4KB) actually gets written. However, we can't always just 942 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 943 */ 944uint64_t zil_block_buckets[] = { 945 4096, /* non TX_WRITE */ 946 8192+4096, /* data base */ 947 32*1024 + 4096, /* NFS writes */ 948 UINT64_MAX 949}; 950 951/* 952 * Start a log block write and advance to the next log block. 953 * Calls are serialized. 954 */ 955static lwb_t * 956zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb, boolean_t last) 957{ 958 lwb_t *nlwb = NULL; 959 zil_chain_t *zilc; 960 spa_t *spa = zilog->zl_spa; 961 blkptr_t *bp; 962 dmu_tx_t *tx; 963 uint64_t txg; 964 uint64_t zil_blksz, wsz; 965 int i, error; 966 boolean_t slog; 967 968 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 969 zilc = (zil_chain_t *)lwb->lwb_buf; 970 bp = &zilc->zc_next_blk; 971 } else { 972 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 973 bp = &zilc->zc_next_blk; 974 } 975 976 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 977 978 /* 979 * Allocate the next block and save its address in this block 980 * before writing it in order to establish the log chain. 981 * Note that if the allocation of nlwb synced before we wrote 982 * the block that points at it (lwb), we'd leak it if we crashed. 983 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 984 * We dirty the dataset to ensure that zil_sync() will be called 985 * to clean up in the event of allocation failure or I/O failure. 986 */ 987 tx = dmu_tx_create(zilog->zl_os); 988 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 989 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 990 txg = dmu_tx_get_txg(tx); 991 992 lwb->lwb_tx = tx; 993 994 /* 995 * Log blocks are pre-allocated. Here we select the size of the next 996 * block, based on size used in the last block. 997 * - first find the smallest bucket that will fit the block from a 998 * limited set of block sizes. This is because it's faster to write 999 * blocks allocated from the same metaslab as they are adjacent or 1000 * close. 1001 * - next find the maximum from the new suggested size and an array of 1002 * previous sizes. This lessens a picket fence effect of wrongly 1003 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 1004 * requests. 1005 * 1006 * Note we only write what is used, but we can't just allocate 1007 * the maximum block size because we can exhaust the available 1008 * pool log space. 1009 */ 1010 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1011 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 1012 continue; 1013 zil_blksz = zil_block_buckets[i]; 1014 if (zil_blksz == UINT64_MAX) 1015 zil_blksz = SPA_OLD_MAXBLOCKSIZE; 1016 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1017 for (i = 0; i < ZIL_PREV_BLKS; i++) 1018 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1019 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1020 1021 BP_ZERO(bp); 1022 /* pass the old blkptr in order to spread log blocks across devs */ 1023 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog); 1024 if (error == 0) { 1025 ASSERT3U(bp->blk_birth, ==, txg); 1026 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1027 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1028 1029 /* 1030 * Allocate a new log write buffer (lwb). 1031 */ 1032 nlwb = zil_alloc_lwb(zilog, bp, slog, txg); 1033 1034 /* Record the block for later vdev flushing */ 1035 zil_add_block(zilog, &lwb->lwb_blk); 1036 } 1037 1038 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1039 /* For Slim ZIL only write what is used. */ 1040 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1041 ASSERT3U(wsz, <=, lwb->lwb_sz); 1042 zio_shrink(lwb->lwb_zio, wsz); 1043 1044 } else { 1045 wsz = lwb->lwb_sz; 1046 } 1047 1048 zilc->zc_pad = 0; 1049 zilc->zc_nused = lwb->lwb_nused; 1050 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1051 1052 /* 1053 * clear unused data for security 1054 */ 1055 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1056 1057 if (last) 1058 lwb->lwb_zio->io_pipeline &= ~ZIO_STAGE_ISSUE_ASYNC; 1059 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1060 1061 /* 1062 * If there was an allocation failure then nlwb will be null which 1063 * forces a txg_wait_synced(). 1064 */ 1065 return (nlwb); 1066} 1067 1068static lwb_t * 1069zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1070{ 1071 lr_t *lrcb, *lrc; 1072 lr_write_t *lrwb, *lrw; 1073 char *lr_buf; 1074 uint64_t dlen, dnow, lwb_sp, reclen, txg; 1075 1076 if (lwb == NULL) 1077 return (NULL); 1078 1079 ASSERT(lwb->lwb_buf != NULL); 1080 1081 lrc = &itx->itx_lr; /* Common log record inside itx. */ 1082 lrw = (lr_write_t *)lrc; /* Write log record inside itx. */ 1083 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) { 1084 dlen = P2ROUNDUP_TYPED( 1085 lrw->lr_length, sizeof (uint64_t), uint64_t); 1086 } else { 1087 dlen = 0; 1088 } 1089 reclen = lrc->lrc_reclen; 1090 zilog->zl_cur_used += (reclen + dlen); 1091 txg = lrc->lrc_txg; 1092 1093 zil_lwb_write_init(zilog, lwb); 1094 1095cont: 1096 /* 1097 * If this record won't fit in the current log block, start a new one. 1098 * For WR_NEED_COPY optimize layout for minimal number of chunks. 1099 */ 1100 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1101 if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1102 lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 || 1103 lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) { 1104 lwb = zil_lwb_write_start(zilog, lwb, B_FALSE); 1105 if (lwb == NULL) 1106 return (NULL); 1107 zil_lwb_write_init(zilog, lwb); 1108 ASSERT(LWB_EMPTY(lwb)); 1109 lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1110 ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp); 1111 } 1112 1113 dnow = MIN(dlen, lwb_sp - reclen); 1114 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1115 bcopy(lrc, lr_buf, reclen); 1116 lrcb = (lr_t *)lr_buf; /* Like lrc, but inside lwb. */ 1117 lrwb = (lr_write_t *)lrcb; /* Like lrw, but inside lwb. */ 1118 1119 /* 1120 * If it's a write, fetch the data or get its blkptr as appropriate. 1121 */ 1122 if (lrc->lrc_txtype == TX_WRITE) { 1123 if (txg > spa_freeze_txg(zilog->zl_spa)) 1124 txg_wait_synced(zilog->zl_dmu_pool, txg); 1125 if (itx->itx_wr_state != WR_COPIED) { 1126 char *dbuf; 1127 int error; 1128 1129 if (itx->itx_wr_state == WR_NEED_COPY) { 1130 dbuf = lr_buf + reclen; 1131 lrcb->lrc_reclen += dnow; 1132 if (lrwb->lr_length > dnow) 1133 lrwb->lr_length = dnow; 1134 lrw->lr_offset += dnow; 1135 lrw->lr_length -= dnow; 1136 } else { 1137 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1138 dbuf = NULL; 1139 } 1140 error = zilog->zl_get_data( 1141 itx->itx_private, lrwb, dbuf, lwb->lwb_zio); 1142 if (error == EIO) { 1143 txg_wait_synced(zilog->zl_dmu_pool, txg); 1144 return (lwb); 1145 } 1146 if (error != 0) { 1147 ASSERT(error == ENOENT || error == EEXIST || 1148 error == EALREADY); 1149 return (lwb); 1150 } 1151 } 1152 } 1153 1154 /* 1155 * We're actually making an entry, so update lrc_seq to be the 1156 * log record sequence number. Note that this is generally not 1157 * equal to the itx sequence number because not all transactions 1158 * are synchronous, and sometimes spa_sync() gets there first. 1159 */ 1160 lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1161 lwb->lwb_nused += reclen + dnow; 1162 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1163 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1164 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1165 1166 dlen -= dnow; 1167 if (dlen > 0) { 1168 zilog->zl_cur_used += reclen; 1169 goto cont; 1170 } 1171 1172 return (lwb); 1173} 1174 1175itx_t * 1176zil_itx_create(uint64_t txtype, size_t lrsize) 1177{ 1178 itx_t *itx; 1179 1180 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1181 1182 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1183 itx->itx_lr.lrc_txtype = txtype; 1184 itx->itx_lr.lrc_reclen = lrsize; 1185 itx->itx_lr.lrc_seq = 0; /* defensive */ 1186 itx->itx_sync = B_TRUE; /* default is synchronous */ 1187 1188 return (itx); 1189} 1190 1191void 1192zil_itx_destroy(itx_t *itx) 1193{ 1194 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1195} 1196 1197/* 1198 * Free up the sync and async itxs. The itxs_t has already been detached 1199 * so no locks are needed. 1200 */ 1201static void 1202zil_itxg_clean(itxs_t *itxs) 1203{ 1204 itx_t *itx; 1205 list_t *list; 1206 avl_tree_t *t; 1207 void *cookie; 1208 itx_async_node_t *ian; 1209 1210 list = &itxs->i_sync_list; 1211 while ((itx = list_head(list)) != NULL) { 1212 list_remove(list, itx); 1213 kmem_free(itx, offsetof(itx_t, itx_lr) + 1214 itx->itx_lr.lrc_reclen); 1215 } 1216 1217 cookie = NULL; 1218 t = &itxs->i_async_tree; 1219 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1220 list = &ian->ia_list; 1221 while ((itx = list_head(list)) != NULL) { 1222 list_remove(list, itx); 1223 kmem_free(itx, offsetof(itx_t, itx_lr) + 1224 itx->itx_lr.lrc_reclen); 1225 } 1226 list_destroy(list); 1227 kmem_free(ian, sizeof (itx_async_node_t)); 1228 } 1229 avl_destroy(t); 1230 1231 kmem_free(itxs, sizeof (itxs_t)); 1232} 1233 1234static int 1235zil_aitx_compare(const void *x1, const void *x2) 1236{ 1237 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1238 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1239 1240 if (o1 < o2) 1241 return (-1); 1242 if (o1 > o2) 1243 return (1); 1244 1245 return (0); 1246} 1247 1248/* 1249 * Remove all async itx with the given oid. 1250 */ 1251static void 1252zil_remove_async(zilog_t *zilog, uint64_t oid) 1253{ 1254 uint64_t otxg, txg; 1255 itx_async_node_t *ian; 1256 avl_tree_t *t; 1257 avl_index_t where; 1258 list_t clean_list; 1259 itx_t *itx; 1260 1261 ASSERT(oid != 0); 1262 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1263 1264 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1265 otxg = ZILTEST_TXG; 1266 else 1267 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1268 1269 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1270 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1271 1272 mutex_enter(&itxg->itxg_lock); 1273 if (itxg->itxg_txg != txg) { 1274 mutex_exit(&itxg->itxg_lock); 1275 continue; 1276 } 1277 1278 /* 1279 * Locate the object node and append its list. 1280 */ 1281 t = &itxg->itxg_itxs->i_async_tree; 1282 ian = avl_find(t, &oid, &where); 1283 if (ian != NULL) 1284 list_move_tail(&clean_list, &ian->ia_list); 1285 mutex_exit(&itxg->itxg_lock); 1286 } 1287 while ((itx = list_head(&clean_list)) != NULL) { 1288 list_remove(&clean_list, itx); 1289 kmem_free(itx, offsetof(itx_t, itx_lr) + 1290 itx->itx_lr.lrc_reclen); 1291 } 1292 list_destroy(&clean_list); 1293} 1294 1295void 1296zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1297{ 1298 uint64_t txg; 1299 itxg_t *itxg; 1300 itxs_t *itxs, *clean = NULL; 1301 1302 /* 1303 * Object ids can be re-instantiated in the next txg so 1304 * remove any async transactions to avoid future leaks. 1305 * This can happen if a fsync occurs on the re-instantiated 1306 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1307 * the new file data and flushes a write record for the old object. 1308 */ 1309 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1310 zil_remove_async(zilog, itx->itx_oid); 1311 1312 /* 1313 * Ensure the data of a renamed file is committed before the rename. 1314 */ 1315 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1316 zil_async_to_sync(zilog, itx->itx_oid); 1317 1318 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1319 txg = ZILTEST_TXG; 1320 else 1321 txg = dmu_tx_get_txg(tx); 1322 1323 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1324 mutex_enter(&itxg->itxg_lock); 1325 itxs = itxg->itxg_itxs; 1326 if (itxg->itxg_txg != txg) { 1327 if (itxs != NULL) { 1328 /* 1329 * The zil_clean callback hasn't got around to cleaning 1330 * this itxg. Save the itxs for release below. 1331 * This should be rare. 1332 */ 1333 zfs_dbgmsg("zil_itx_assign: missed itx cleanup for " 1334 "txg %llu", itxg->itxg_txg); 1335 clean = itxg->itxg_itxs; 1336 } 1337 itxg->itxg_txg = txg; 1338 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1339 1340 list_create(&itxs->i_sync_list, sizeof (itx_t), 1341 offsetof(itx_t, itx_node)); 1342 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1343 sizeof (itx_async_node_t), 1344 offsetof(itx_async_node_t, ia_node)); 1345 } 1346 if (itx->itx_sync) { 1347 list_insert_tail(&itxs->i_sync_list, itx); 1348 } else { 1349 avl_tree_t *t = &itxs->i_async_tree; 1350 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1351 itx_async_node_t *ian; 1352 avl_index_t where; 1353 1354 ian = avl_find(t, &foid, &where); 1355 if (ian == NULL) { 1356 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1357 list_create(&ian->ia_list, sizeof (itx_t), 1358 offsetof(itx_t, itx_node)); 1359 ian->ia_foid = foid; 1360 avl_insert(t, ian, where); 1361 } 1362 list_insert_tail(&ian->ia_list, itx); 1363 } 1364 1365 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1366 zilog_dirty(zilog, txg); 1367 mutex_exit(&itxg->itxg_lock); 1368 1369 /* Release the old itxs now we've dropped the lock */ 1370 if (clean != NULL) 1371 zil_itxg_clean(clean); 1372} 1373 1374/* 1375 * If there are any in-memory intent log transactions which have now been 1376 * synced then start up a taskq to free them. We should only do this after we 1377 * have written out the uberblocks (i.e. txg has been comitted) so that 1378 * don't inadvertently clean out in-memory log records that would be required 1379 * by zil_commit(). 1380 */ 1381void 1382zil_clean(zilog_t *zilog, uint64_t synced_txg) 1383{ 1384 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1385 itxs_t *clean_me; 1386 1387 mutex_enter(&itxg->itxg_lock); 1388 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1389 mutex_exit(&itxg->itxg_lock); 1390 return; 1391 } 1392 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1393 ASSERT3U(itxg->itxg_txg, !=, 0); 1394 clean_me = itxg->itxg_itxs; 1395 itxg->itxg_itxs = NULL; 1396 itxg->itxg_txg = 0; 1397 mutex_exit(&itxg->itxg_lock); 1398 /* 1399 * Preferably start a task queue to free up the old itxs but 1400 * if taskq_dispatch can't allocate resources to do that then 1401 * free it in-line. This should be rare. Note, using TQ_SLEEP 1402 * created a bad performance problem. 1403 */ 1404 ASSERT3P(zilog->zl_dmu_pool, !=, NULL); 1405 ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL); 1406 if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq, 1407 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1408 zil_itxg_clean(clean_me); 1409} 1410 1411/* 1412 * Get the list of itxs to commit into zl_itx_commit_list. 1413 */ 1414static void 1415zil_get_commit_list(zilog_t *zilog) 1416{ 1417 uint64_t otxg, txg; 1418 list_t *commit_list = &zilog->zl_itx_commit_list; 1419 1420 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1421 otxg = ZILTEST_TXG; 1422 else 1423 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1424 1425 /* 1426 * This is inherently racy, since there is nothing to prevent 1427 * the last synced txg from changing. That's okay since we'll 1428 * only commit things in the future. 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 we're adding itx records to the zl_itx_commit_list, 1441 * then the zil better be dirty in this "txg". We can assert 1442 * that here since we're holding the itxg_lock which will 1443 * prevent spa_sync from cleaning it. Once we add the itxs 1444 * to the zl_itx_commit_list we must commit it to disk even 1445 * if it's unnecessary (i.e. the txg was synced). 1446 */ 1447 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 1448 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1449 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1450 1451 mutex_exit(&itxg->itxg_lock); 1452 } 1453} 1454 1455/* 1456 * Move the async itxs for a specified object to commit into sync lists. 1457 */ 1458void 1459zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1460{ 1461 uint64_t otxg, txg; 1462 itx_async_node_t *ian; 1463 avl_tree_t *t; 1464 avl_index_t where; 1465 1466 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1467 otxg = ZILTEST_TXG; 1468 else 1469 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1470 1471 /* 1472 * This is inherently racy, since there is nothing to prevent 1473 * the last synced txg from changing. 1474 */ 1475 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1476 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1477 1478 mutex_enter(&itxg->itxg_lock); 1479 if (itxg->itxg_txg != txg) { 1480 mutex_exit(&itxg->itxg_lock); 1481 continue; 1482 } 1483 1484 /* 1485 * If a foid is specified then find that node and append its 1486 * list. Otherwise walk the tree appending all the lists 1487 * to the sync list. We add to the end rather than the 1488 * beginning to ensure the create has happened. 1489 */ 1490 t = &itxg->itxg_itxs->i_async_tree; 1491 if (foid != 0) { 1492 ian = avl_find(t, &foid, &where); 1493 if (ian != NULL) { 1494 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1495 &ian->ia_list); 1496 } 1497 } else { 1498 void *cookie = NULL; 1499 1500 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1501 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1502 &ian->ia_list); 1503 list_destroy(&ian->ia_list); 1504 kmem_free(ian, sizeof (itx_async_node_t)); 1505 } 1506 } 1507 mutex_exit(&itxg->itxg_lock); 1508 } 1509} 1510 1511static void 1512zil_commit_writer(zilog_t *zilog) 1513{ 1514 uint64_t txg; 1515 itx_t *itx; 1516 lwb_t *lwb; 1517 spa_t *spa = zilog->zl_spa; 1518 int error = 0; 1519 1520 ASSERT(zilog->zl_root_zio == NULL); 1521 1522 mutex_exit(&zilog->zl_lock); 1523 1524 zil_get_commit_list(zilog); 1525 1526 /* 1527 * Return if there's nothing to commit before we dirty the fs by 1528 * calling zil_create(). 1529 */ 1530 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1531 mutex_enter(&zilog->zl_lock); 1532 return; 1533 } 1534 1535 if (zilog->zl_suspend) { 1536 lwb = NULL; 1537 } else { 1538 lwb = list_tail(&zilog->zl_lwb_list); 1539 if (lwb == NULL) 1540 lwb = zil_create(zilog); 1541 } 1542 1543 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1544 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1545 txg = itx->itx_lr.lrc_txg; 1546 ASSERT3U(txg, !=, 0); 1547 1548 /* 1549 * This is inherently racy and may result in us writing 1550 * out a log block for a txg that was just synced. This is 1551 * ok since we'll end cleaning up that log block the next 1552 * time we call zil_sync(). 1553 */ 1554 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1555 lwb = zil_lwb_commit(zilog, itx, lwb); 1556 list_remove(&zilog->zl_itx_commit_list, itx); 1557 kmem_free(itx, offsetof(itx_t, itx_lr) 1558 + itx->itx_lr.lrc_reclen); 1559 } 1560 1561 /* write the last block out */ 1562 if (lwb != NULL && lwb->lwb_zio != NULL) 1563 lwb = zil_lwb_write_start(zilog, lwb, B_TRUE); 1564 1565 zilog->zl_cur_used = 0; 1566 1567 /* 1568 * Wait if necessary for the log blocks to be on stable storage. 1569 */ 1570 if (zilog->zl_root_zio) { 1571 error = zio_wait(zilog->zl_root_zio); 1572 zilog->zl_root_zio = NULL; 1573 zil_flush_vdevs(zilog); 1574 } 1575 1576 if (error || lwb == NULL) 1577 txg_wait_synced(zilog->zl_dmu_pool, 0); 1578 1579 mutex_enter(&zilog->zl_lock); 1580 1581 /* 1582 * Remember the highest committed log sequence number for ztest. 1583 * We only update this value when all the log writes succeeded, 1584 * because ztest wants to ASSERT that it got the whole log chain. 1585 */ 1586 if (error == 0 && lwb != NULL) 1587 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1588} 1589 1590/* 1591 * Commit zfs transactions to stable storage. 1592 * If foid is 0 push out all transactions, otherwise push only those 1593 * for that object or might reference that object. 1594 * 1595 * itxs are committed in batches. In a heavily stressed zil there will be 1596 * a commit writer thread who is writing out a bunch of itxs to the log 1597 * for a set of committing threads (cthreads) in the same batch as the writer. 1598 * Those cthreads are all waiting on the same cv for that batch. 1599 * 1600 * There will also be a different and growing batch of threads that are 1601 * waiting to commit (qthreads). When the committing batch completes 1602 * a transition occurs such that the cthreads exit and the qthreads become 1603 * cthreads. One of the new cthreads becomes the writer thread for the 1604 * batch. Any new threads arriving become new qthreads. 1605 * 1606 * Only 2 condition variables are needed and there's no transition 1607 * between the two cvs needed. They just flip-flop between qthreads 1608 * and cthreads. 1609 * 1610 * Using this scheme we can efficiently wakeup up only those threads 1611 * that have been committed. 1612 */ 1613void 1614zil_commit(zilog_t *zilog, uint64_t foid) 1615{ 1616 uint64_t mybatch; 1617 1618 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1619 return; 1620 1621 /* move the async itxs for the foid to the sync queues */ 1622 zil_async_to_sync(zilog, foid); 1623 1624 mutex_enter(&zilog->zl_lock); 1625 mybatch = zilog->zl_next_batch; 1626 while (zilog->zl_writer) { 1627 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1628 if (mybatch <= zilog->zl_com_batch) { 1629 mutex_exit(&zilog->zl_lock); 1630 return; 1631 } 1632 } 1633 1634 zilog->zl_next_batch++; 1635 zilog->zl_writer = B_TRUE; 1636 zil_commit_writer(zilog); 1637 zilog->zl_com_batch = mybatch; 1638 zilog->zl_writer = B_FALSE; 1639 mutex_exit(&zilog->zl_lock); 1640 1641 /* wake up one thread to become the next writer */ 1642 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1643 1644 /* wake up all threads waiting for this batch to be committed */ 1645 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1646} 1647 1648/* 1649 * Called in syncing context to free committed log blocks and update log header. 1650 */ 1651void 1652zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1653{ 1654 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1655 uint64_t txg = dmu_tx_get_txg(tx); 1656 spa_t *spa = zilog->zl_spa; 1657 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1658 lwb_t *lwb; 1659 1660 /* 1661 * We don't zero out zl_destroy_txg, so make sure we don't try 1662 * to destroy it twice. 1663 */ 1664 if (spa_sync_pass(spa) != 1) 1665 return; 1666 1667 mutex_enter(&zilog->zl_lock); 1668 1669 ASSERT(zilog->zl_stop_sync == 0); 1670 1671 if (*replayed_seq != 0) { 1672 ASSERT(zh->zh_replay_seq < *replayed_seq); 1673 zh->zh_replay_seq = *replayed_seq; 1674 *replayed_seq = 0; 1675 } 1676 1677 if (zilog->zl_destroy_txg == txg) { 1678 blkptr_t blk = zh->zh_log; 1679 1680 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1681 1682 bzero(zh, sizeof (zil_header_t)); 1683 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1684 1685 if (zilog->zl_keep_first) { 1686 /* 1687 * If this block was part of log chain that couldn't 1688 * be claimed because a device was missing during 1689 * zil_claim(), but that device later returns, 1690 * then this block could erroneously appear valid. 1691 * To guard against this, assign a new GUID to the new 1692 * log chain so it doesn't matter what blk points to. 1693 */ 1694 zil_init_log_chain(zilog, &blk); 1695 zh->zh_log = blk; 1696 } 1697 } 1698 1699 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1700 zh->zh_log = lwb->lwb_blk; 1701 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1702 break; 1703 list_remove(&zilog->zl_lwb_list, lwb); 1704 zio_free_zil(spa, txg, &lwb->lwb_blk); 1705 kmem_cache_free(zil_lwb_cache, lwb); 1706 1707 /* 1708 * If we don't have anything left in the lwb list then 1709 * we've had an allocation failure and we need to zero 1710 * out the zil_header blkptr so that we don't end 1711 * up freeing the same block twice. 1712 */ 1713 if (list_head(&zilog->zl_lwb_list) == NULL) 1714 BP_ZERO(&zh->zh_log); 1715 } 1716 mutex_exit(&zilog->zl_lock); 1717} 1718 1719void 1720zil_init(void) 1721{ 1722 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1723 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1724} 1725 1726void 1727zil_fini(void) 1728{ 1729 kmem_cache_destroy(zil_lwb_cache); 1730} 1731 1732void 1733zil_set_sync(zilog_t *zilog, uint64_t sync) 1734{ 1735 zilog->zl_sync = sync; 1736} 1737 1738void 1739zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1740{ 1741 zilog->zl_logbias = logbias; 1742} 1743 1744zilog_t * 1745zil_alloc(objset_t *os, zil_header_t *zh_phys) 1746{ 1747 zilog_t *zilog; 1748 1749 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1750 1751 zilog->zl_header = zh_phys; 1752 zilog->zl_os = os; 1753 zilog->zl_spa = dmu_objset_spa(os); 1754 zilog->zl_dmu_pool = dmu_objset_pool(os); 1755 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1756 zilog->zl_logbias = dmu_objset_logbias(os); 1757 zilog->zl_sync = dmu_objset_syncprop(os); 1758 zilog->zl_next_batch = 1; 1759 1760 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1761 1762 for (int i = 0; i < TXG_SIZE; i++) { 1763 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1764 MUTEX_DEFAULT, NULL); 1765 } 1766 1767 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1768 offsetof(lwb_t, lwb_node)); 1769 1770 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1771 offsetof(itx_t, itx_node)); 1772 1773 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1774 1775 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1776 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1777 1778 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1779 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1780 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1781 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1782 1783 return (zilog); 1784} 1785 1786void 1787zil_free(zilog_t *zilog) 1788{ 1789 zilog->zl_stop_sync = 1; 1790 1791 ASSERT0(zilog->zl_suspend); 1792 ASSERT0(zilog->zl_suspending); 1793 1794 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1795 list_destroy(&zilog->zl_lwb_list); 1796 1797 avl_destroy(&zilog->zl_vdev_tree); 1798 mutex_destroy(&zilog->zl_vdev_lock); 1799 1800 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1801 list_destroy(&zilog->zl_itx_commit_list); 1802 1803 for (int i = 0; i < TXG_SIZE; i++) { 1804 /* 1805 * It's possible for an itx to be generated that doesn't dirty 1806 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1807 * callback to remove the entry. We remove those here. 1808 * 1809 * Also free up the ziltest itxs. 1810 */ 1811 if (zilog->zl_itxg[i].itxg_itxs) 1812 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1813 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1814 } 1815 1816 mutex_destroy(&zilog->zl_lock); 1817 1818 cv_destroy(&zilog->zl_cv_writer); 1819 cv_destroy(&zilog->zl_cv_suspend); 1820 cv_destroy(&zilog->zl_cv_batch[0]); 1821 cv_destroy(&zilog->zl_cv_batch[1]); 1822 1823 kmem_free(zilog, sizeof (zilog_t)); 1824} 1825 1826/* 1827 * Open an intent log. 1828 */ 1829zilog_t * 1830zil_open(objset_t *os, zil_get_data_t *get_data) 1831{ 1832 zilog_t *zilog = dmu_objset_zil(os); 1833 1834 ASSERT(zilog->zl_get_data == NULL); 1835 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1836 1837 zilog->zl_get_data = get_data; 1838 1839 return (zilog); 1840} 1841 1842/* 1843 * Close an intent log. 1844 */ 1845void 1846zil_close(zilog_t *zilog) 1847{ 1848 lwb_t *lwb; 1849 uint64_t txg = 0; 1850 1851 zil_commit(zilog, 0); /* commit all itx */ 1852 1853 /* 1854 * The lwb_max_txg for the stubby lwb will reflect the last activity 1855 * for the zil. After a txg_wait_synced() on the txg we know all the 1856 * callbacks have occurred that may clean the zil. Only then can we 1857 * destroy the zl_clean_taskq. 1858 */ 1859 mutex_enter(&zilog->zl_lock); 1860 lwb = list_tail(&zilog->zl_lwb_list); 1861 if (lwb != NULL) 1862 txg = lwb->lwb_max_txg; 1863 mutex_exit(&zilog->zl_lock); 1864 if (txg) 1865 txg_wait_synced(zilog->zl_dmu_pool, txg); 1866 1867 if (zilog_is_dirty(zilog)) 1868 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg); 1869 VERIFY(!zilog_is_dirty(zilog)); 1870 1871 zilog->zl_get_data = NULL; 1872 1873 /* 1874 * We should have only one LWB left on the list; remove it now. 1875 */ 1876 mutex_enter(&zilog->zl_lock); 1877 lwb = list_head(&zilog->zl_lwb_list); 1878 if (lwb != NULL) { 1879 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1880 list_remove(&zilog->zl_lwb_list, lwb); 1881 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1882 kmem_cache_free(zil_lwb_cache, lwb); 1883 } 1884 mutex_exit(&zilog->zl_lock); 1885} 1886 1887static char *suspend_tag = "zil suspending"; 1888 1889/* 1890 * Suspend an intent log. While in suspended mode, we still honor 1891 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1892 * On old version pools, we suspend the log briefly when taking a 1893 * snapshot so that it will have an empty intent log. 1894 * 1895 * Long holds are not really intended to be used the way we do here -- 1896 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1897 * could fail. Therefore we take pains to only put a long hold if it is 1898 * actually necessary. Fortunately, it will only be necessary if the 1899 * objset is currently mounted (or the ZVOL equivalent). In that case it 1900 * will already have a long hold, so we are not really making things any worse. 1901 * 1902 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1903 * zvol_state_t), and use their mechanism to prevent their hold from being 1904 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1905 * very little gain. 1906 * 1907 * if cookiep == NULL, this does both the suspend & resume. 1908 * Otherwise, it returns with the dataset "long held", and the cookie 1909 * should be passed into zil_resume(). 1910 */ 1911int 1912zil_suspend(const char *osname, void **cookiep) 1913{ 1914 objset_t *os; 1915 zilog_t *zilog; 1916 const zil_header_t *zh; 1917 int error; 1918 1919 error = dmu_objset_hold(osname, suspend_tag, &os); 1920 if (error != 0) 1921 return (error); 1922 zilog = dmu_objset_zil(os); 1923 1924 mutex_enter(&zilog->zl_lock); 1925 zh = zilog->zl_header; 1926 1927 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1928 mutex_exit(&zilog->zl_lock); 1929 dmu_objset_rele(os, suspend_tag); 1930 return (SET_ERROR(EBUSY)); 1931 } 1932 1933 /* 1934 * Don't put a long hold in the cases where we can avoid it. This 1935 * is when there is no cookie so we are doing a suspend & resume 1936 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1937 * for the suspend because it's already suspended, or there's no ZIL. 1938 */ 1939 if (cookiep == NULL && !zilog->zl_suspending && 1940 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1941 mutex_exit(&zilog->zl_lock); 1942 dmu_objset_rele(os, suspend_tag); 1943 return (0); 1944 } 1945 1946 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1947 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1948 1949 zilog->zl_suspend++; 1950 1951 if (zilog->zl_suspend > 1) { 1952 /* 1953 * Someone else is already suspending it. 1954 * Just wait for them to finish. 1955 */ 1956 1957 while (zilog->zl_suspending) 1958 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1959 mutex_exit(&zilog->zl_lock); 1960 1961 if (cookiep == NULL) 1962 zil_resume(os); 1963 else 1964 *cookiep = os; 1965 return (0); 1966 } 1967 1968 /* 1969 * If there is no pointer to an on-disk block, this ZIL must not 1970 * be active (e.g. filesystem not mounted), so there's nothing 1971 * to clean up. 1972 */ 1973 if (BP_IS_HOLE(&zh->zh_log)) { 1974 ASSERT(cookiep != NULL); /* fast path already handled */ 1975 1976 *cookiep = os; 1977 mutex_exit(&zilog->zl_lock); 1978 return (0); 1979 } 1980 1981 zilog->zl_suspending = B_TRUE; 1982 mutex_exit(&zilog->zl_lock); 1983 1984 zil_commit(zilog, 0); 1985 1986 zil_destroy(zilog, B_FALSE); 1987 1988 mutex_enter(&zilog->zl_lock); 1989 zilog->zl_suspending = B_FALSE; 1990 cv_broadcast(&zilog->zl_cv_suspend); 1991 mutex_exit(&zilog->zl_lock); 1992 1993 if (cookiep == NULL) 1994 zil_resume(os); 1995 else 1996 *cookiep = os; 1997 return (0); 1998} 1999 2000void 2001zil_resume(void *cookie) 2002{ 2003 objset_t *os = cookie; 2004 zilog_t *zilog = dmu_objset_zil(os); 2005 2006 mutex_enter(&zilog->zl_lock); 2007 ASSERT(zilog->zl_suspend != 0); 2008 zilog->zl_suspend--; 2009 mutex_exit(&zilog->zl_lock); 2010 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 2011 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 2012} 2013 2014typedef struct zil_replay_arg { 2015 zil_replay_func_t **zr_replay; 2016 void *zr_arg; 2017 boolean_t zr_byteswap; 2018 char *zr_lr; 2019} zil_replay_arg_t; 2020 2021static int 2022zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 2023{ 2024 char name[ZFS_MAX_DATASET_NAME_LEN]; 2025 2026 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 2027 2028 dmu_objset_name(zilog->zl_os, name); 2029 2030 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 2031 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 2032 (u_longlong_t)lr->lrc_seq, 2033 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 2034 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 2035 2036 return (error); 2037} 2038 2039static int 2040zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 2041{ 2042 zil_replay_arg_t *zr = zra; 2043 const zil_header_t *zh = zilog->zl_header; 2044 uint64_t reclen = lr->lrc_reclen; 2045 uint64_t txtype = lr->lrc_txtype; 2046 int error = 0; 2047 2048 zilog->zl_replaying_seq = lr->lrc_seq; 2049 2050 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2051 return (0); 2052 2053 if (lr->lrc_txg < claim_txg) /* already committed */ 2054 return (0); 2055 2056 /* Strip case-insensitive bit, still present in log record */ 2057 txtype &= ~TX_CI; 2058 2059 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2060 return (zil_replay_error(zilog, lr, EINVAL)); 2061 2062 /* 2063 * If this record type can be logged out of order, the object 2064 * (lr_foid) may no longer exist. That's legitimate, not an error. 2065 */ 2066 if (TX_OOO(txtype)) { 2067 error = dmu_object_info(zilog->zl_os, 2068 ((lr_ooo_t *)lr)->lr_foid, NULL); 2069 if (error == ENOENT || error == EEXIST) 2070 return (0); 2071 } 2072 2073 /* 2074 * Make a copy of the data so we can revise and extend it. 2075 */ 2076 bcopy(lr, zr->zr_lr, reclen); 2077 2078 /* 2079 * If this is a TX_WRITE with a blkptr, suck in the data. 2080 */ 2081 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2082 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2083 zr->zr_lr + reclen); 2084 if (error != 0) 2085 return (zil_replay_error(zilog, lr, error)); 2086 } 2087 2088 /* 2089 * The log block containing this lr may have been byteswapped 2090 * so that we can easily examine common fields like lrc_txtype. 2091 * However, the log is a mix of different record types, and only the 2092 * replay vectors know how to byteswap their records. Therefore, if 2093 * the lr was byteswapped, undo it before invoking the replay vector. 2094 */ 2095 if (zr->zr_byteswap) 2096 byteswap_uint64_array(zr->zr_lr, reclen); 2097 2098 /* 2099 * We must now do two things atomically: replay this log record, 2100 * and update the log header sequence number to reflect the fact that 2101 * we did so. At the end of each replay function the sequence number 2102 * is updated if we are in replay mode. 2103 */ 2104 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2105 if (error != 0) { 2106 /* 2107 * The DMU's dnode layer doesn't see removes until the txg 2108 * commits, so a subsequent claim can spuriously fail with 2109 * EEXIST. So if we receive any error we try syncing out 2110 * any removes then retry the transaction. Note that we 2111 * specify B_FALSE for byteswap now, so we don't do it twice. 2112 */ 2113 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2114 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2115 if (error != 0) 2116 return (zil_replay_error(zilog, lr, error)); 2117 } 2118 return (0); 2119} 2120 2121/* ARGSUSED */ 2122static int 2123zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2124{ 2125 zilog->zl_replay_blks++; 2126 2127 return (0); 2128} 2129 2130/* 2131 * If this dataset has a non-empty intent log, replay it and destroy it. 2132 */ 2133void 2134zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2135{ 2136 zilog_t *zilog = dmu_objset_zil(os); 2137 const zil_header_t *zh = zilog->zl_header; 2138 zil_replay_arg_t zr; 2139 2140 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2141 zil_destroy(zilog, B_TRUE); 2142 return; 2143 } 2144 2145 zr.zr_replay = replay_func; 2146 zr.zr_arg = arg; 2147 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2148 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2149 2150 /* 2151 * Wait for in-progress removes to sync before starting replay. 2152 */ 2153 txg_wait_synced(zilog->zl_dmu_pool, 0); 2154 2155 zilog->zl_replay = B_TRUE; 2156 zilog->zl_replay_time = ddi_get_lbolt(); 2157 ASSERT(zilog->zl_replay_blks == 0); 2158 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2159 zh->zh_claim_txg); 2160 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2161 2162 zil_destroy(zilog, B_FALSE); 2163 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2164 zilog->zl_replay = B_FALSE; 2165} 2166 2167boolean_t 2168zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2169{ 2170 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2171 return (B_TRUE); 2172 2173 if (zilog->zl_replay) { 2174 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2175 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2176 zilog->zl_replaying_seq; 2177 return (B_TRUE); 2178 } 2179 2180 return (B_FALSE); 2181} 2182 2183/* ARGSUSED */ 2184int 2185zil_vdev_offline(const char *osname, void *arg) 2186{ 2187 int error; 2188 2189 error = zil_suspend(osname, NULL); 2190 if (error != 0) 2191 return (SET_ERROR(EEXIST)); 2192 return (0); 2193} 2194