zil.c revision 249921
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) 2013 by Delphix. All rights reserved. 24 */ 25 26/* Portions Copyright 2010 Robert Milkowski */ 27 28#include <sys/zfs_context.h> 29#include <sys/spa.h> 30#include <sys/dmu.h> 31#include <sys/zap.h> 32#include <sys/arc.h> 33#include <sys/stat.h> 34#include <sys/resource.h> 35#include <sys/zil.h> 36#include <sys/zil_impl.h> 37#include <sys/dsl_dataset.h> 38#include <sys/vdev_impl.h> 39#include <sys/dmu_tx.h> 40#include <sys/dsl_pool.h> 41 42/* 43 * The zfs intent log (ZIL) saves transaction records of system calls 44 * that change the file system in memory with enough information 45 * to be able to replay them. These are stored in memory until 46 * either the DMU transaction group (txg) commits them to the stable pool 47 * and they can be discarded, or they are flushed to the stable log 48 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 49 * requirement. In the event of a panic or power fail then those log 50 * records (transactions) are replayed. 51 * 52 * There is one ZIL per file system. Its on-disk (pool) format consists 53 * of 3 parts: 54 * 55 * - ZIL header 56 * - ZIL blocks 57 * - ZIL records 58 * 59 * A log record holds a system call transaction. Log blocks can 60 * hold many log records and the blocks are chained together. 61 * Each ZIL block contains a block pointer (blkptr_t) to the next 62 * ZIL block in the chain. The ZIL header points to the first 63 * block in the chain. Note there is not a fixed place in the pool 64 * to hold blocks. They are dynamically allocated and freed as 65 * needed from the blocks available. Figure X shows the ZIL structure: 66 */ 67 68/* 69 * This global ZIL switch affects all pools 70 */ 71int zil_replay_disable = 0; /* disable intent logging replay */ 72SYSCTL_DECL(_vfs_zfs); 73TUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 74SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 75 &zil_replay_disable, 0, "Disable intent logging replay"); 76 77/* 78 * Tunable parameter for debugging or performance analysis. Setting 79 * zfs_nocacheflush will cause corruption on power loss if a volatile 80 * out-of-order write cache is enabled. 81 */ 82boolean_t zfs_nocacheflush = B_FALSE; 83TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 84SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 85 &zfs_nocacheflush, 0, "Disable cache flush"); 86boolean_t zfs_trim_enabled = B_TRUE; 87SYSCTL_DECL(_vfs_zfs_trim); 88TUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 89SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 90 "Enable ZFS TRIM"); 91 92static kmem_cache_t *zil_lwb_cache; 93 94static void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 95 96#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 97 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 98 99 100/* 101 * ziltest is by and large an ugly hack, but very useful in 102 * checking replay without tedious work. 103 * When running ziltest we want to keep all itx's and so maintain 104 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 105 * We subtract TXG_CONCURRENT_STATES to allow for common code. 106 */ 107#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 108 109static int 110zil_bp_compare(const void *x1, const void *x2) 111{ 112 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 113 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 114 115 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 116 return (-1); 117 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 118 return (1); 119 120 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 121 return (-1); 122 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 123 return (1); 124 125 return (0); 126} 127 128static void 129zil_bp_tree_init(zilog_t *zilog) 130{ 131 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 132 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 133} 134 135static void 136zil_bp_tree_fini(zilog_t *zilog) 137{ 138 avl_tree_t *t = &zilog->zl_bp_tree; 139 zil_bp_node_t *zn; 140 void *cookie = NULL; 141 142 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 143 kmem_free(zn, sizeof (zil_bp_node_t)); 144 145 avl_destroy(t); 146} 147 148int 149zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 150{ 151 avl_tree_t *t = &zilog->zl_bp_tree; 152 const dva_t *dva = BP_IDENTITY(bp); 153 zil_bp_node_t *zn; 154 avl_index_t where; 155 156 if (avl_find(t, dva, &where) != NULL) 157 return (SET_ERROR(EEXIST)); 158 159 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 160 zn->zn_dva = *dva; 161 avl_insert(t, zn, where); 162 163 return (0); 164} 165 166static zil_header_t * 167zil_header_in_syncing_context(zilog_t *zilog) 168{ 169 return ((zil_header_t *)zilog->zl_header); 170} 171 172static void 173zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 174{ 175 zio_cksum_t *zc = &bp->blk_cksum; 176 177 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 178 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 179 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 180 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 181} 182 183/* 184 * Read a log block and make sure it's valid. 185 */ 186static int 187zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 188 char **end) 189{ 190 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 191 uint32_t aflags = ARC_WAIT; 192 arc_buf_t *abuf = NULL; 193 zbookmark_t zb; 194 int error; 195 196 if (zilog->zl_header->zh_claim_txg == 0) 197 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 198 199 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 200 zio_flags |= ZIO_FLAG_SPECULATIVE; 201 202 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 203 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 204 205 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 206 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 207 208 if (error == 0) { 209 zio_cksum_t cksum = bp->blk_cksum; 210 211 /* 212 * Validate the checksummed log block. 213 * 214 * Sequence numbers should be... sequential. The checksum 215 * verifier for the next block should be bp's checksum plus 1. 216 * 217 * Also check the log chain linkage and size used. 218 */ 219 cksum.zc_word[ZIL_ZC_SEQ]++; 220 221 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 222 zil_chain_t *zilc = abuf->b_data; 223 char *lr = (char *)(zilc + 1); 224 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 225 226 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 227 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 228 error = SET_ERROR(ECKSUM); 229 } else { 230 bcopy(lr, dst, len); 231 *end = (char *)dst + len; 232 *nbp = zilc->zc_next_blk; 233 } 234 } else { 235 char *lr = abuf->b_data; 236 uint64_t size = BP_GET_LSIZE(bp); 237 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 238 239 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 240 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 241 (zilc->zc_nused > (size - sizeof (*zilc)))) { 242 error = SET_ERROR(ECKSUM); 243 } else { 244 bcopy(lr, dst, zilc->zc_nused); 245 *end = (char *)dst + zilc->zc_nused; 246 *nbp = zilc->zc_next_blk; 247 } 248 } 249 250 VERIFY(arc_buf_remove_ref(abuf, &abuf)); 251 } 252 253 return (error); 254} 255 256/* 257 * Read a TX_WRITE log data block. 258 */ 259static int 260zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 261{ 262 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 263 const blkptr_t *bp = &lr->lr_blkptr; 264 uint32_t aflags = ARC_WAIT; 265 arc_buf_t *abuf = NULL; 266 zbookmark_t zb; 267 int error; 268 269 if (BP_IS_HOLE(bp)) { 270 if (wbuf != NULL) 271 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 272 return (0); 273 } 274 275 if (zilog->zl_header->zh_claim_txg == 0) 276 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 277 278 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 279 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 280 281 error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 282 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 283 284 if (error == 0) { 285 if (wbuf != NULL) 286 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 287 (void) arc_buf_remove_ref(abuf, &abuf); 288 } 289 290 return (error); 291} 292 293/* 294 * Parse the intent log, and call parse_func for each valid record within. 295 */ 296int 297zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 298 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 299{ 300 const zil_header_t *zh = zilog->zl_header; 301 boolean_t claimed = !!zh->zh_claim_txg; 302 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 303 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 304 uint64_t max_blk_seq = 0; 305 uint64_t max_lr_seq = 0; 306 uint64_t blk_count = 0; 307 uint64_t lr_count = 0; 308 blkptr_t blk, next_blk; 309 char *lrbuf, *lrp; 310 int error = 0; 311 312 /* 313 * Old logs didn't record the maximum zh_claim_lr_seq. 314 */ 315 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 316 claim_lr_seq = UINT64_MAX; 317 318 /* 319 * Starting at the block pointed to by zh_log we read the log chain. 320 * For each block in the chain we strongly check that block to 321 * ensure its validity. We stop when an invalid block is found. 322 * For each block pointer in the chain we call parse_blk_func(). 323 * For each record in each valid block we call parse_lr_func(). 324 * If the log has been claimed, stop if we encounter a sequence 325 * number greater than the highest claimed sequence number. 326 */ 327 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 328 zil_bp_tree_init(zilog); 329 330 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 331 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 332 int reclen; 333 char *end; 334 335 if (blk_seq > claim_blk_seq) 336 break; 337 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 338 break; 339 ASSERT3U(max_blk_seq, <, blk_seq); 340 max_blk_seq = blk_seq; 341 blk_count++; 342 343 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 344 break; 345 346 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 347 if (error != 0) 348 break; 349 350 for (lrp = lrbuf; lrp < end; lrp += reclen) { 351 lr_t *lr = (lr_t *)lrp; 352 reclen = lr->lrc_reclen; 353 ASSERT3U(reclen, >=, sizeof (lr_t)); 354 if (lr->lrc_seq > claim_lr_seq) 355 goto done; 356 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 357 goto done; 358 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 359 max_lr_seq = lr->lrc_seq; 360 lr_count++; 361 } 362 } 363done: 364 zilog->zl_parse_error = error; 365 zilog->zl_parse_blk_seq = max_blk_seq; 366 zilog->zl_parse_lr_seq = max_lr_seq; 367 zilog->zl_parse_blk_count = blk_count; 368 zilog->zl_parse_lr_count = lr_count; 369 370 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 371 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 372 373 zil_bp_tree_fini(zilog); 374 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 375 376 return (error); 377} 378 379static int 380zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 381{ 382 /* 383 * Claim log block if not already committed and not already claimed. 384 * If tx == NULL, just verify that the block is claimable. 385 */ 386 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 387 return (0); 388 389 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 390 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 391 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 392} 393 394static int 395zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 396{ 397 lr_write_t *lr = (lr_write_t *)lrc; 398 int error; 399 400 if (lrc->lrc_txtype != TX_WRITE) 401 return (0); 402 403 /* 404 * If the block is not readable, don't claim it. This can happen 405 * in normal operation when a log block is written to disk before 406 * some of the dmu_sync() blocks it points to. In this case, the 407 * transaction cannot have been committed to anyone (we would have 408 * waited for all writes to be stable first), so it is semantically 409 * correct to declare this the end of the log. 410 */ 411 if (lr->lr_blkptr.blk_birth >= first_txg && 412 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 413 return (error); 414 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 415} 416 417/* ARGSUSED */ 418static int 419zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 420{ 421 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 422 423 return (0); 424} 425 426static int 427zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 428{ 429 lr_write_t *lr = (lr_write_t *)lrc; 430 blkptr_t *bp = &lr->lr_blkptr; 431 432 /* 433 * If we previously claimed it, we need to free it. 434 */ 435 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 436 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 437 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 438 439 return (0); 440} 441 442static lwb_t * 443zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 444{ 445 lwb_t *lwb; 446 447 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 448 lwb->lwb_zilog = zilog; 449 lwb->lwb_blk = *bp; 450 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 451 lwb->lwb_max_txg = txg; 452 lwb->lwb_zio = NULL; 453 lwb->lwb_tx = NULL; 454 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 455 lwb->lwb_nused = sizeof (zil_chain_t); 456 lwb->lwb_sz = BP_GET_LSIZE(bp); 457 } else { 458 lwb->lwb_nused = 0; 459 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 460 } 461 462 mutex_enter(&zilog->zl_lock); 463 list_insert_tail(&zilog->zl_lwb_list, lwb); 464 mutex_exit(&zilog->zl_lock); 465 466 return (lwb); 467} 468 469/* 470 * Called when we create in-memory log transactions so that we know 471 * to cleanup the itxs at the end of spa_sync(). 472 */ 473void 474zilog_dirty(zilog_t *zilog, uint64_t txg) 475{ 476 dsl_pool_t *dp = zilog->zl_dmu_pool; 477 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 478 479 if (dsl_dataset_is_snapshot(ds)) 480 panic("dirtying snapshot!"); 481 482 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 483 /* up the hold count until we can be written out */ 484 dmu_buf_add_ref(ds->ds_dbuf, zilog); 485 } 486} 487 488boolean_t 489zilog_is_dirty(zilog_t *zilog) 490{ 491 dsl_pool_t *dp = zilog->zl_dmu_pool; 492 493 for (int t = 0; t < TXG_SIZE; t++) { 494 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 495 return (B_TRUE); 496 } 497 return (B_FALSE); 498} 499 500/* 501 * Create an on-disk intent log. 502 */ 503static lwb_t * 504zil_create(zilog_t *zilog) 505{ 506 const zil_header_t *zh = zilog->zl_header; 507 lwb_t *lwb = NULL; 508 uint64_t txg = 0; 509 dmu_tx_t *tx = NULL; 510 blkptr_t blk; 511 int error = 0; 512 513 /* 514 * Wait for any previous destroy to complete. 515 */ 516 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 517 518 ASSERT(zh->zh_claim_txg == 0); 519 ASSERT(zh->zh_replay_seq == 0); 520 521 blk = zh->zh_log; 522 523 /* 524 * Allocate an initial log block if: 525 * - there isn't one already 526 * - the existing block is the wrong endianess 527 */ 528 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 529 tx = dmu_tx_create(zilog->zl_os); 530 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 531 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 532 txg = dmu_tx_get_txg(tx); 533 534 if (!BP_IS_HOLE(&blk)) { 535 zio_free_zil(zilog->zl_spa, txg, &blk); 536 BP_ZERO(&blk); 537 } 538 539 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 540 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 541 542 if (error == 0) 543 zil_init_log_chain(zilog, &blk); 544 } 545 546 /* 547 * Allocate a log write buffer (lwb) for the first log block. 548 */ 549 if (error == 0) 550 lwb = zil_alloc_lwb(zilog, &blk, txg); 551 552 /* 553 * If we just allocated the first log block, commit our transaction 554 * and wait for zil_sync() to stuff the block poiner into zh_log. 555 * (zh is part of the MOS, so we cannot modify it in open context.) 556 */ 557 if (tx != NULL) { 558 dmu_tx_commit(tx); 559 txg_wait_synced(zilog->zl_dmu_pool, txg); 560 } 561 562 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 563 564 return (lwb); 565} 566 567/* 568 * In one tx, free all log blocks and clear the log header. 569 * If keep_first is set, then we're replaying a log with no content. 570 * We want to keep the first block, however, so that the first 571 * synchronous transaction doesn't require a txg_wait_synced() 572 * in zil_create(). We don't need to txg_wait_synced() here either 573 * when keep_first is set, because both zil_create() and zil_destroy() 574 * will wait for any in-progress destroys to complete. 575 */ 576void 577zil_destroy(zilog_t *zilog, boolean_t keep_first) 578{ 579 const zil_header_t *zh = zilog->zl_header; 580 lwb_t *lwb; 581 dmu_tx_t *tx; 582 uint64_t txg; 583 584 /* 585 * Wait for any previous destroy to complete. 586 */ 587 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 588 589 zilog->zl_old_header = *zh; /* debugging aid */ 590 591 if (BP_IS_HOLE(&zh->zh_log)) 592 return; 593 594 tx = dmu_tx_create(zilog->zl_os); 595 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 596 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 597 txg = dmu_tx_get_txg(tx); 598 599 mutex_enter(&zilog->zl_lock); 600 601 ASSERT3U(zilog->zl_destroy_txg, <, txg); 602 zilog->zl_destroy_txg = txg; 603 zilog->zl_keep_first = keep_first; 604 605 if (!list_is_empty(&zilog->zl_lwb_list)) { 606 ASSERT(zh->zh_claim_txg == 0); 607 VERIFY(!keep_first); 608 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 609 list_remove(&zilog->zl_lwb_list, lwb); 610 if (lwb->lwb_buf != NULL) 611 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 612 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 613 kmem_cache_free(zil_lwb_cache, lwb); 614 } 615 } else if (!keep_first) { 616 zil_destroy_sync(zilog, tx); 617 } 618 mutex_exit(&zilog->zl_lock); 619 620 dmu_tx_commit(tx); 621} 622 623void 624zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 625{ 626 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 627 (void) zil_parse(zilog, zil_free_log_block, 628 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 629} 630 631int 632zil_claim(const char *osname, void *txarg) 633{ 634 dmu_tx_t *tx = txarg; 635 uint64_t first_txg = dmu_tx_get_txg(tx); 636 zilog_t *zilog; 637 zil_header_t *zh; 638 objset_t *os; 639 int error; 640 641 error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); 642 if (error != 0) { 643 cmn_err(CE_WARN, "can't open objset for %s", osname); 644 return (0); 645 } 646 647 zilog = dmu_objset_zil(os); 648 zh = zil_header_in_syncing_context(zilog); 649 650 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 651 if (!BP_IS_HOLE(&zh->zh_log)) 652 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 653 BP_ZERO(&zh->zh_log); 654 dsl_dataset_dirty(dmu_objset_ds(os), tx); 655 dmu_objset_disown(os, FTAG); 656 return (0); 657 } 658 659 /* 660 * Claim all log blocks if we haven't already done so, and remember 661 * the highest claimed sequence number. This ensures that if we can 662 * read only part of the log now (e.g. due to a missing device), 663 * but we can read the entire log later, we will not try to replay 664 * or destroy beyond the last block we successfully claimed. 665 */ 666 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 667 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 668 (void) zil_parse(zilog, zil_claim_log_block, 669 zil_claim_log_record, tx, first_txg); 670 zh->zh_claim_txg = first_txg; 671 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 672 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 673 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 674 zh->zh_flags |= ZIL_REPLAY_NEEDED; 675 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 676 dsl_dataset_dirty(dmu_objset_ds(os), tx); 677 } 678 679 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 680 dmu_objset_disown(os, FTAG); 681 return (0); 682} 683 684/* 685 * Check the log by walking the log chain. 686 * Checksum errors are ok as they indicate the end of the chain. 687 * Any other error (no device or read failure) returns an error. 688 */ 689int 690zil_check_log_chain(const char *osname, void *tx) 691{ 692 zilog_t *zilog; 693 objset_t *os; 694 blkptr_t *bp; 695 int error; 696 697 ASSERT(tx == NULL); 698 699 error = dmu_objset_hold(osname, FTAG, &os); 700 if (error != 0) { 701 cmn_err(CE_WARN, "can't open objset for %s", osname); 702 return (0); 703 } 704 705 zilog = dmu_objset_zil(os); 706 bp = (blkptr_t *)&zilog->zl_header->zh_log; 707 708 /* 709 * Check the first block and determine if it's on a log device 710 * which may have been removed or faulted prior to loading this 711 * pool. If so, there's no point in checking the rest of the log 712 * as its content should have already been synced to the pool. 713 */ 714 if (!BP_IS_HOLE(bp)) { 715 vdev_t *vd; 716 boolean_t valid = B_TRUE; 717 718 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 719 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 720 if (vd->vdev_islog && vdev_is_dead(vd)) 721 valid = vdev_log_state_valid(vd); 722 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 723 724 if (!valid) { 725 dmu_objset_rele(os, FTAG); 726 return (0); 727 } 728 } 729 730 /* 731 * Because tx == NULL, zil_claim_log_block() will not actually claim 732 * any blocks, but just determine whether it is possible to do so. 733 * In addition to checking the log chain, zil_claim_log_block() 734 * will invoke zio_claim() with a done func of spa_claim_notify(), 735 * which will update spa_max_claim_txg. See spa_load() for details. 736 */ 737 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 738 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 739 740 dmu_objset_rele(os, FTAG); 741 742 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 743} 744 745static int 746zil_vdev_compare(const void *x1, const void *x2) 747{ 748 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 749 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 750 751 if (v1 < v2) 752 return (-1); 753 if (v1 > v2) 754 return (1); 755 756 return (0); 757} 758 759void 760zil_add_block(zilog_t *zilog, const blkptr_t *bp) 761{ 762 avl_tree_t *t = &zilog->zl_vdev_tree; 763 avl_index_t where; 764 zil_vdev_node_t *zv, zvsearch; 765 int ndvas = BP_GET_NDVAS(bp); 766 int i; 767 768 if (zfs_nocacheflush) 769 return; 770 771 ASSERT(zilog->zl_writer); 772 773 /* 774 * Even though we're zl_writer, we still need a lock because the 775 * zl_get_data() callbacks may have dmu_sync() done callbacks 776 * that will run concurrently. 777 */ 778 mutex_enter(&zilog->zl_vdev_lock); 779 for (i = 0; i < ndvas; i++) { 780 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 781 if (avl_find(t, &zvsearch, &where) == NULL) { 782 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 783 zv->zv_vdev = zvsearch.zv_vdev; 784 avl_insert(t, zv, where); 785 } 786 } 787 mutex_exit(&zilog->zl_vdev_lock); 788} 789 790static void 791zil_flush_vdevs(zilog_t *zilog) 792{ 793 spa_t *spa = zilog->zl_spa; 794 avl_tree_t *t = &zilog->zl_vdev_tree; 795 void *cookie = NULL; 796 zil_vdev_node_t *zv; 797 zio_t *zio; 798 799 ASSERT(zilog->zl_writer); 800 801 /* 802 * We don't need zl_vdev_lock here because we're the zl_writer, 803 * and all zl_get_data() callbacks are done. 804 */ 805 if (avl_numnodes(t) == 0) 806 return; 807 808 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 809 810 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 811 812 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 813 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 814 if (vd != NULL) 815 zio_flush(zio, vd); 816 kmem_free(zv, sizeof (*zv)); 817 } 818 819 /* 820 * Wait for all the flushes to complete. Not all devices actually 821 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 822 */ 823 (void) zio_wait(zio); 824 825 spa_config_exit(spa, SCL_STATE, FTAG); 826} 827 828/* 829 * Function called when a log block write completes 830 */ 831static void 832zil_lwb_write_done(zio_t *zio) 833{ 834 lwb_t *lwb = zio->io_private; 835 zilog_t *zilog = lwb->lwb_zilog; 836 dmu_tx_t *tx = lwb->lwb_tx; 837 838 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 839 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 840 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 841 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 842 ASSERT(!BP_IS_GANG(zio->io_bp)); 843 ASSERT(!BP_IS_HOLE(zio->io_bp)); 844 ASSERT(zio->io_bp->blk_fill == 0); 845 846 /* 847 * Ensure the lwb buffer pointer is cleared before releasing 848 * the txg. If we have had an allocation failure and 849 * the txg is waiting to sync then we want want zil_sync() 850 * to remove the lwb so that it's not picked up as the next new 851 * one in zil_commit_writer(). zil_sync() will only remove 852 * the lwb if lwb_buf is null. 853 */ 854 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 855 mutex_enter(&zilog->zl_lock); 856 lwb->lwb_buf = NULL; 857 lwb->lwb_tx = NULL; 858 mutex_exit(&zilog->zl_lock); 859 860 /* 861 * Now that we've written this log block, we have a stable pointer 862 * to the next block in the chain, so it's OK to let the txg in 863 * which we allocated the next block sync. 864 */ 865 dmu_tx_commit(tx); 866} 867 868/* 869 * Initialize the io for a log block. 870 */ 871static void 872zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 873{ 874 zbookmark_t zb; 875 876 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 877 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 878 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 879 880 if (zilog->zl_root_zio == NULL) { 881 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 882 ZIO_FLAG_CANFAIL); 883 } 884 if (lwb->lwb_zio == NULL) { 885 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 886 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 887 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 888 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 889 } 890} 891 892/* 893 * Define a limited set of intent log block sizes. 894 * These must be a multiple of 4KB. Note only the amount used (again 895 * aligned to 4KB) actually gets written. However, we can't always just 896 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 897 */ 898uint64_t zil_block_buckets[] = { 899 4096, /* non TX_WRITE */ 900 8192+4096, /* data base */ 901 32*1024 + 4096, /* NFS writes */ 902 UINT64_MAX 903}; 904 905/* 906 * Use the slog as long as the logbias is 'latency' and the current commit size 907 * is less than the limit or the total list size is less than 2X the limit. 908 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 909 */ 910uint64_t zil_slog_limit = 1024 * 1024; 911#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 912 (((zilog)->zl_cur_used < zil_slog_limit) || \ 913 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 914 915/* 916 * Start a log block write and advance to the next log block. 917 * Calls are serialized. 918 */ 919static lwb_t * 920zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 921{ 922 lwb_t *nlwb = NULL; 923 zil_chain_t *zilc; 924 spa_t *spa = zilog->zl_spa; 925 blkptr_t *bp; 926 dmu_tx_t *tx; 927 uint64_t txg; 928 uint64_t zil_blksz, wsz; 929 int i, error; 930 931 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 932 zilc = (zil_chain_t *)lwb->lwb_buf; 933 bp = &zilc->zc_next_blk; 934 } else { 935 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 936 bp = &zilc->zc_next_blk; 937 } 938 939 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 940 941 /* 942 * Allocate the next block and save its address in this block 943 * before writing it in order to establish the log chain. 944 * Note that if the allocation of nlwb synced before we wrote 945 * the block that points at it (lwb), we'd leak it if we crashed. 946 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 947 * We dirty the dataset to ensure that zil_sync() will be called 948 * to clean up in the event of allocation failure or I/O failure. 949 */ 950 tx = dmu_tx_create(zilog->zl_os); 951 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 952 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 953 txg = dmu_tx_get_txg(tx); 954 955 lwb->lwb_tx = tx; 956 957 /* 958 * Log blocks are pre-allocated. Here we select the size of the next 959 * block, based on size used in the last block. 960 * - first find the smallest bucket that will fit the block from a 961 * limited set of block sizes. This is because it's faster to write 962 * blocks allocated from the same metaslab as they are adjacent or 963 * close. 964 * - next find the maximum from the new suggested size and an array of 965 * previous sizes. This lessens a picket fence effect of wrongly 966 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 967 * requests. 968 * 969 * Note we only write what is used, but we can't just allocate 970 * the maximum block size because we can exhaust the available 971 * pool log space. 972 */ 973 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 974 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 975 continue; 976 zil_blksz = zil_block_buckets[i]; 977 if (zil_blksz == UINT64_MAX) 978 zil_blksz = SPA_MAXBLOCKSIZE; 979 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 980 for (i = 0; i < ZIL_PREV_BLKS; i++) 981 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 982 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 983 984 BP_ZERO(bp); 985 /* pass the old blkptr in order to spread log blocks across devs */ 986 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 987 USE_SLOG(zilog)); 988 if (error == 0) { 989 ASSERT3U(bp->blk_birth, ==, txg); 990 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 991 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 992 993 /* 994 * Allocate a new log write buffer (lwb). 995 */ 996 nlwb = zil_alloc_lwb(zilog, bp, txg); 997 998 /* Record the block for later vdev flushing */ 999 zil_add_block(zilog, &lwb->lwb_blk); 1000 } 1001 1002 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1003 /* For Slim ZIL only write what is used. */ 1004 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1005 ASSERT3U(wsz, <=, lwb->lwb_sz); 1006 zio_shrink(lwb->lwb_zio, wsz); 1007 1008 } else { 1009 wsz = lwb->lwb_sz; 1010 } 1011 1012 zilc->zc_pad = 0; 1013 zilc->zc_nused = lwb->lwb_nused; 1014 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1015 1016 /* 1017 * clear unused data for security 1018 */ 1019 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1020 1021 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1022 1023 /* 1024 * If there was an allocation failure then nlwb will be null which 1025 * forces a txg_wait_synced(). 1026 */ 1027 return (nlwb); 1028} 1029 1030static lwb_t * 1031zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1032{ 1033 lr_t *lrc = &itx->itx_lr; /* common log record */ 1034 lr_write_t *lrw = (lr_write_t *)lrc; 1035 char *lr_buf; 1036 uint64_t txg = lrc->lrc_txg; 1037 uint64_t reclen = lrc->lrc_reclen; 1038 uint64_t dlen = 0; 1039 1040 if (lwb == NULL) 1041 return (NULL); 1042 1043 ASSERT(lwb->lwb_buf != NULL); 1044 ASSERT(zilog_is_dirty(zilog) || 1045 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1046 1047 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1048 dlen = P2ROUNDUP_TYPED( 1049 lrw->lr_length, sizeof (uint64_t), uint64_t); 1050 1051 zilog->zl_cur_used += (reclen + dlen); 1052 1053 zil_lwb_write_init(zilog, lwb); 1054 1055 /* 1056 * If this record won't fit in the current log block, start a new one. 1057 */ 1058 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1059 lwb = zil_lwb_write_start(zilog, lwb); 1060 if (lwb == NULL) 1061 return (NULL); 1062 zil_lwb_write_init(zilog, lwb); 1063 ASSERT(LWB_EMPTY(lwb)); 1064 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1065 txg_wait_synced(zilog->zl_dmu_pool, txg); 1066 return (lwb); 1067 } 1068 } 1069 1070 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1071 bcopy(lrc, lr_buf, reclen); 1072 lrc = (lr_t *)lr_buf; 1073 lrw = (lr_write_t *)lrc; 1074 1075 /* 1076 * If it's a write, fetch the data or get its blkptr as appropriate. 1077 */ 1078 if (lrc->lrc_txtype == TX_WRITE) { 1079 if (txg > spa_freeze_txg(zilog->zl_spa)) 1080 txg_wait_synced(zilog->zl_dmu_pool, txg); 1081 if (itx->itx_wr_state != WR_COPIED) { 1082 char *dbuf; 1083 int error; 1084 1085 if (dlen) { 1086 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1087 dbuf = lr_buf + reclen; 1088 lrw->lr_common.lrc_reclen += dlen; 1089 } else { 1090 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1091 dbuf = NULL; 1092 } 1093 error = zilog->zl_get_data( 1094 itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1095 if (error == EIO) { 1096 txg_wait_synced(zilog->zl_dmu_pool, txg); 1097 return (lwb); 1098 } 1099 if (error != 0) { 1100 ASSERT(error == ENOENT || error == EEXIST || 1101 error == EALREADY); 1102 return (lwb); 1103 } 1104 } 1105 } 1106 1107 /* 1108 * We're actually making an entry, so update lrc_seq to be the 1109 * log record sequence number. Note that this is generally not 1110 * equal to the itx sequence number because not all transactions 1111 * are synchronous, and sometimes spa_sync() gets there first. 1112 */ 1113 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1114 lwb->lwb_nused += reclen + dlen; 1115 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1116 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1117 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1118 1119 return (lwb); 1120} 1121 1122itx_t * 1123zil_itx_create(uint64_t txtype, size_t lrsize) 1124{ 1125 itx_t *itx; 1126 1127 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1128 1129 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1130 itx->itx_lr.lrc_txtype = txtype; 1131 itx->itx_lr.lrc_reclen = lrsize; 1132 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1133 itx->itx_lr.lrc_seq = 0; /* defensive */ 1134 itx->itx_sync = B_TRUE; /* default is synchronous */ 1135 1136 return (itx); 1137} 1138 1139void 1140zil_itx_destroy(itx_t *itx) 1141{ 1142 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1143} 1144 1145/* 1146 * Free up the sync and async itxs. The itxs_t has already been detached 1147 * so no locks are needed. 1148 */ 1149static void 1150zil_itxg_clean(itxs_t *itxs) 1151{ 1152 itx_t *itx; 1153 list_t *list; 1154 avl_tree_t *t; 1155 void *cookie; 1156 itx_async_node_t *ian; 1157 1158 list = &itxs->i_sync_list; 1159 while ((itx = list_head(list)) != NULL) { 1160 list_remove(list, itx); 1161 kmem_free(itx, offsetof(itx_t, itx_lr) + 1162 itx->itx_lr.lrc_reclen); 1163 } 1164 1165 cookie = NULL; 1166 t = &itxs->i_async_tree; 1167 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1168 list = &ian->ia_list; 1169 while ((itx = list_head(list)) != NULL) { 1170 list_remove(list, itx); 1171 kmem_free(itx, offsetof(itx_t, itx_lr) + 1172 itx->itx_lr.lrc_reclen); 1173 } 1174 list_destroy(list); 1175 kmem_free(ian, sizeof (itx_async_node_t)); 1176 } 1177 avl_destroy(t); 1178 1179 kmem_free(itxs, sizeof (itxs_t)); 1180} 1181 1182static int 1183zil_aitx_compare(const void *x1, const void *x2) 1184{ 1185 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1186 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1187 1188 if (o1 < o2) 1189 return (-1); 1190 if (o1 > o2) 1191 return (1); 1192 1193 return (0); 1194} 1195 1196/* 1197 * Remove all async itx with the given oid. 1198 */ 1199static void 1200zil_remove_async(zilog_t *zilog, uint64_t oid) 1201{ 1202 uint64_t otxg, txg; 1203 itx_async_node_t *ian; 1204 avl_tree_t *t; 1205 avl_index_t where; 1206 list_t clean_list; 1207 itx_t *itx; 1208 1209 ASSERT(oid != 0); 1210 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1211 1212 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1213 otxg = ZILTEST_TXG; 1214 else 1215 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1216 1217 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1218 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1219 1220 mutex_enter(&itxg->itxg_lock); 1221 if (itxg->itxg_txg != txg) { 1222 mutex_exit(&itxg->itxg_lock); 1223 continue; 1224 } 1225 1226 /* 1227 * Locate the object node and append its list. 1228 */ 1229 t = &itxg->itxg_itxs->i_async_tree; 1230 ian = avl_find(t, &oid, &where); 1231 if (ian != NULL) 1232 list_move_tail(&clean_list, &ian->ia_list); 1233 mutex_exit(&itxg->itxg_lock); 1234 } 1235 while ((itx = list_head(&clean_list)) != NULL) { 1236 list_remove(&clean_list, itx); 1237 kmem_free(itx, offsetof(itx_t, itx_lr) + 1238 itx->itx_lr.lrc_reclen); 1239 } 1240 list_destroy(&clean_list); 1241} 1242 1243void 1244zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1245{ 1246 uint64_t txg; 1247 itxg_t *itxg; 1248 itxs_t *itxs, *clean = NULL; 1249 1250 /* 1251 * Object ids can be re-instantiated in the next txg so 1252 * remove any async transactions to avoid future leaks. 1253 * This can happen if a fsync occurs on the re-instantiated 1254 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1255 * the new file data and flushes a write record for the old object. 1256 */ 1257 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1258 zil_remove_async(zilog, itx->itx_oid); 1259 1260 /* 1261 * Ensure the data of a renamed file is committed before the rename. 1262 */ 1263 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1264 zil_async_to_sync(zilog, itx->itx_oid); 1265 1266 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1267 txg = ZILTEST_TXG; 1268 else 1269 txg = dmu_tx_get_txg(tx); 1270 1271 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1272 mutex_enter(&itxg->itxg_lock); 1273 itxs = itxg->itxg_itxs; 1274 if (itxg->itxg_txg != txg) { 1275 if (itxs != NULL) { 1276 /* 1277 * The zil_clean callback hasn't got around to cleaning 1278 * this itxg. Save the itxs for release below. 1279 * This should be rare. 1280 */ 1281 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1282 itxg->itxg_sod = 0; 1283 clean = itxg->itxg_itxs; 1284 } 1285 ASSERT(itxg->itxg_sod == 0); 1286 itxg->itxg_txg = txg; 1287 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1288 1289 list_create(&itxs->i_sync_list, sizeof (itx_t), 1290 offsetof(itx_t, itx_node)); 1291 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1292 sizeof (itx_async_node_t), 1293 offsetof(itx_async_node_t, ia_node)); 1294 } 1295 if (itx->itx_sync) { 1296 list_insert_tail(&itxs->i_sync_list, itx); 1297 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1298 itxg->itxg_sod += itx->itx_sod; 1299 } else { 1300 avl_tree_t *t = &itxs->i_async_tree; 1301 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1302 itx_async_node_t *ian; 1303 avl_index_t where; 1304 1305 ian = avl_find(t, &foid, &where); 1306 if (ian == NULL) { 1307 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1308 list_create(&ian->ia_list, sizeof (itx_t), 1309 offsetof(itx_t, itx_node)); 1310 ian->ia_foid = foid; 1311 avl_insert(t, ian, where); 1312 } 1313 list_insert_tail(&ian->ia_list, itx); 1314 } 1315 1316 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1317 zilog_dirty(zilog, txg); 1318 mutex_exit(&itxg->itxg_lock); 1319 1320 /* Release the old itxs now we've dropped the lock */ 1321 if (clean != NULL) 1322 zil_itxg_clean(clean); 1323} 1324 1325/* 1326 * If there are any in-memory intent log transactions which have now been 1327 * synced then start up a taskq to free them. We should only do this after we 1328 * have written out the uberblocks (i.e. txg has been comitted) so that 1329 * don't inadvertently clean out in-memory log records that would be required 1330 * by zil_commit(). 1331 */ 1332void 1333zil_clean(zilog_t *zilog, uint64_t synced_txg) 1334{ 1335 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1336 itxs_t *clean_me; 1337 1338 mutex_enter(&itxg->itxg_lock); 1339 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1340 mutex_exit(&itxg->itxg_lock); 1341 return; 1342 } 1343 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1344 ASSERT(itxg->itxg_txg != 0); 1345 ASSERT(zilog->zl_clean_taskq != NULL); 1346 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1347 itxg->itxg_sod = 0; 1348 clean_me = itxg->itxg_itxs; 1349 itxg->itxg_itxs = NULL; 1350 itxg->itxg_txg = 0; 1351 mutex_exit(&itxg->itxg_lock); 1352 /* 1353 * Preferably start a task queue to free up the old itxs but 1354 * if taskq_dispatch can't allocate resources to do that then 1355 * free it in-line. This should be rare. Note, using TQ_SLEEP 1356 * created a bad performance problem. 1357 */ 1358 if (taskq_dispatch(zilog->zl_clean_taskq, 1359 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1360 zil_itxg_clean(clean_me); 1361} 1362 1363/* 1364 * Get the list of itxs to commit into zl_itx_commit_list. 1365 */ 1366static void 1367zil_get_commit_list(zilog_t *zilog) 1368{ 1369 uint64_t otxg, txg; 1370 list_t *commit_list = &zilog->zl_itx_commit_list; 1371 uint64_t push_sod = 0; 1372 1373 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1374 otxg = ZILTEST_TXG; 1375 else 1376 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1377 1378 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1379 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1380 1381 mutex_enter(&itxg->itxg_lock); 1382 if (itxg->itxg_txg != txg) { 1383 mutex_exit(&itxg->itxg_lock); 1384 continue; 1385 } 1386 1387 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1388 push_sod += itxg->itxg_sod; 1389 itxg->itxg_sod = 0; 1390 1391 mutex_exit(&itxg->itxg_lock); 1392 } 1393 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1394} 1395 1396/* 1397 * Move the async itxs for a specified object to commit into sync lists. 1398 */ 1399static void 1400zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1401{ 1402 uint64_t otxg, txg; 1403 itx_async_node_t *ian; 1404 avl_tree_t *t; 1405 avl_index_t where; 1406 1407 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1408 otxg = ZILTEST_TXG; 1409 else 1410 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1411 1412 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1413 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1414 1415 mutex_enter(&itxg->itxg_lock); 1416 if (itxg->itxg_txg != txg) { 1417 mutex_exit(&itxg->itxg_lock); 1418 continue; 1419 } 1420 1421 /* 1422 * If a foid is specified then find that node and append its 1423 * list. Otherwise walk the tree appending all the lists 1424 * to the sync list. We add to the end rather than the 1425 * beginning to ensure the create has happened. 1426 */ 1427 t = &itxg->itxg_itxs->i_async_tree; 1428 if (foid != 0) { 1429 ian = avl_find(t, &foid, &where); 1430 if (ian != NULL) { 1431 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1432 &ian->ia_list); 1433 } 1434 } else { 1435 void *cookie = NULL; 1436 1437 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1438 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1439 &ian->ia_list); 1440 list_destroy(&ian->ia_list); 1441 kmem_free(ian, sizeof (itx_async_node_t)); 1442 } 1443 } 1444 mutex_exit(&itxg->itxg_lock); 1445 } 1446} 1447 1448static void 1449zil_commit_writer(zilog_t *zilog) 1450{ 1451 uint64_t txg; 1452 itx_t *itx; 1453 lwb_t *lwb; 1454 spa_t *spa = zilog->zl_spa; 1455 int error = 0; 1456 1457 ASSERT(zilog->zl_root_zio == NULL); 1458 1459 mutex_exit(&zilog->zl_lock); 1460 1461 zil_get_commit_list(zilog); 1462 1463 /* 1464 * Return if there's nothing to commit before we dirty the fs by 1465 * calling zil_create(). 1466 */ 1467 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1468 mutex_enter(&zilog->zl_lock); 1469 return; 1470 } 1471 1472 if (zilog->zl_suspend) { 1473 lwb = NULL; 1474 } else { 1475 lwb = list_tail(&zilog->zl_lwb_list); 1476 if (lwb == NULL) 1477 lwb = zil_create(zilog); 1478 } 1479 1480 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1481 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1482 txg = itx->itx_lr.lrc_txg; 1483 ASSERT(txg); 1484 1485 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1486 lwb = zil_lwb_commit(zilog, itx, lwb); 1487 list_remove(&zilog->zl_itx_commit_list, itx); 1488 kmem_free(itx, offsetof(itx_t, itx_lr) 1489 + itx->itx_lr.lrc_reclen); 1490 } 1491 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1492 1493 /* write the last block out */ 1494 if (lwb != NULL && lwb->lwb_zio != NULL) 1495 lwb = zil_lwb_write_start(zilog, lwb); 1496 1497 zilog->zl_cur_used = 0; 1498 1499 /* 1500 * Wait if necessary for the log blocks to be on stable storage. 1501 */ 1502 if (zilog->zl_root_zio) { 1503 error = zio_wait(zilog->zl_root_zio); 1504 zilog->zl_root_zio = NULL; 1505 zil_flush_vdevs(zilog); 1506 } 1507 1508 if (error || lwb == NULL) 1509 txg_wait_synced(zilog->zl_dmu_pool, 0); 1510 1511 mutex_enter(&zilog->zl_lock); 1512 1513 /* 1514 * Remember the highest committed log sequence number for ztest. 1515 * We only update this value when all the log writes succeeded, 1516 * because ztest wants to ASSERT that it got the whole log chain. 1517 */ 1518 if (error == 0 && lwb != NULL) 1519 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1520} 1521 1522/* 1523 * Commit zfs transactions to stable storage. 1524 * If foid is 0 push out all transactions, otherwise push only those 1525 * for that object or might reference that object. 1526 * 1527 * itxs are committed in batches. In a heavily stressed zil there will be 1528 * a commit writer thread who is writing out a bunch of itxs to the log 1529 * for a set of committing threads (cthreads) in the same batch as the writer. 1530 * Those cthreads are all waiting on the same cv for that batch. 1531 * 1532 * There will also be a different and growing batch of threads that are 1533 * waiting to commit (qthreads). When the committing batch completes 1534 * a transition occurs such that the cthreads exit and the qthreads become 1535 * cthreads. One of the new cthreads becomes the writer thread for the 1536 * batch. Any new threads arriving become new qthreads. 1537 * 1538 * Only 2 condition variables are needed and there's no transition 1539 * between the two cvs needed. They just flip-flop between qthreads 1540 * and cthreads. 1541 * 1542 * Using this scheme we can efficiently wakeup up only those threads 1543 * that have been committed. 1544 */ 1545void 1546zil_commit(zilog_t *zilog, uint64_t foid) 1547{ 1548 uint64_t mybatch; 1549 1550 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1551 return; 1552 1553 /* move the async itxs for the foid to the sync queues */ 1554 zil_async_to_sync(zilog, foid); 1555 1556 mutex_enter(&zilog->zl_lock); 1557 mybatch = zilog->zl_next_batch; 1558 while (zilog->zl_writer) { 1559 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1560 if (mybatch <= zilog->zl_com_batch) { 1561 mutex_exit(&zilog->zl_lock); 1562 return; 1563 } 1564 } 1565 1566 zilog->zl_next_batch++; 1567 zilog->zl_writer = B_TRUE; 1568 zil_commit_writer(zilog); 1569 zilog->zl_com_batch = mybatch; 1570 zilog->zl_writer = B_FALSE; 1571 mutex_exit(&zilog->zl_lock); 1572 1573 /* wake up one thread to become the next writer */ 1574 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1575 1576 /* wake up all threads waiting for this batch to be committed */ 1577 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1578} 1579 1580/* 1581 * Called in syncing context to free committed log blocks and update log header. 1582 */ 1583void 1584zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1585{ 1586 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1587 uint64_t txg = dmu_tx_get_txg(tx); 1588 spa_t *spa = zilog->zl_spa; 1589 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1590 lwb_t *lwb; 1591 1592 /* 1593 * We don't zero out zl_destroy_txg, so make sure we don't try 1594 * to destroy it twice. 1595 */ 1596 if (spa_sync_pass(spa) != 1) 1597 return; 1598 1599 mutex_enter(&zilog->zl_lock); 1600 1601 ASSERT(zilog->zl_stop_sync == 0); 1602 1603 if (*replayed_seq != 0) { 1604 ASSERT(zh->zh_replay_seq < *replayed_seq); 1605 zh->zh_replay_seq = *replayed_seq; 1606 *replayed_seq = 0; 1607 } 1608 1609 if (zilog->zl_destroy_txg == txg) { 1610 blkptr_t blk = zh->zh_log; 1611 1612 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1613 1614 bzero(zh, sizeof (zil_header_t)); 1615 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1616 1617 if (zilog->zl_keep_first) { 1618 /* 1619 * If this block was part of log chain that couldn't 1620 * be claimed because a device was missing during 1621 * zil_claim(), but that device later returns, 1622 * then this block could erroneously appear valid. 1623 * To guard against this, assign a new GUID to the new 1624 * log chain so it doesn't matter what blk points to. 1625 */ 1626 zil_init_log_chain(zilog, &blk); 1627 zh->zh_log = blk; 1628 } 1629 } 1630 1631 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1632 zh->zh_log = lwb->lwb_blk; 1633 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1634 break; 1635 list_remove(&zilog->zl_lwb_list, lwb); 1636 zio_free_zil(spa, txg, &lwb->lwb_blk); 1637 kmem_cache_free(zil_lwb_cache, lwb); 1638 1639 /* 1640 * If we don't have anything left in the lwb list then 1641 * we've had an allocation failure and we need to zero 1642 * out the zil_header blkptr so that we don't end 1643 * up freeing the same block twice. 1644 */ 1645 if (list_head(&zilog->zl_lwb_list) == NULL) 1646 BP_ZERO(&zh->zh_log); 1647 } 1648 mutex_exit(&zilog->zl_lock); 1649} 1650 1651void 1652zil_init(void) 1653{ 1654 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1655 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1656} 1657 1658void 1659zil_fini(void) 1660{ 1661 kmem_cache_destroy(zil_lwb_cache); 1662} 1663 1664void 1665zil_set_sync(zilog_t *zilog, uint64_t sync) 1666{ 1667 zilog->zl_sync = sync; 1668} 1669 1670void 1671zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1672{ 1673 zilog->zl_logbias = logbias; 1674} 1675 1676zilog_t * 1677zil_alloc(objset_t *os, zil_header_t *zh_phys) 1678{ 1679 zilog_t *zilog; 1680 1681 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1682 1683 zilog->zl_header = zh_phys; 1684 zilog->zl_os = os; 1685 zilog->zl_spa = dmu_objset_spa(os); 1686 zilog->zl_dmu_pool = dmu_objset_pool(os); 1687 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1688 zilog->zl_logbias = dmu_objset_logbias(os); 1689 zilog->zl_sync = dmu_objset_syncprop(os); 1690 zilog->zl_next_batch = 1; 1691 1692 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1693 1694 for (int i = 0; i < TXG_SIZE; i++) { 1695 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1696 MUTEX_DEFAULT, NULL); 1697 } 1698 1699 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1700 offsetof(lwb_t, lwb_node)); 1701 1702 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1703 offsetof(itx_t, itx_node)); 1704 1705 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1706 1707 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1708 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1709 1710 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1711 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1712 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1713 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1714 1715 return (zilog); 1716} 1717 1718void 1719zil_free(zilog_t *zilog) 1720{ 1721 zilog->zl_stop_sync = 1; 1722 1723 ASSERT0(zilog->zl_suspend); 1724 ASSERT0(zilog->zl_suspending); 1725 1726 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1727 list_destroy(&zilog->zl_lwb_list); 1728 1729 avl_destroy(&zilog->zl_vdev_tree); 1730 mutex_destroy(&zilog->zl_vdev_lock); 1731 1732 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1733 list_destroy(&zilog->zl_itx_commit_list); 1734 1735 for (int i = 0; i < TXG_SIZE; i++) { 1736 /* 1737 * It's possible for an itx to be generated that doesn't dirty 1738 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1739 * callback to remove the entry. We remove those here. 1740 * 1741 * Also free up the ziltest itxs. 1742 */ 1743 if (zilog->zl_itxg[i].itxg_itxs) 1744 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1745 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1746 } 1747 1748 mutex_destroy(&zilog->zl_lock); 1749 1750 cv_destroy(&zilog->zl_cv_writer); 1751 cv_destroy(&zilog->zl_cv_suspend); 1752 cv_destroy(&zilog->zl_cv_batch[0]); 1753 cv_destroy(&zilog->zl_cv_batch[1]); 1754 1755 kmem_free(zilog, sizeof (zilog_t)); 1756} 1757 1758/* 1759 * Open an intent log. 1760 */ 1761zilog_t * 1762zil_open(objset_t *os, zil_get_data_t *get_data) 1763{ 1764 zilog_t *zilog = dmu_objset_zil(os); 1765 1766 ASSERT(zilog->zl_clean_taskq == NULL); 1767 ASSERT(zilog->zl_get_data == NULL); 1768 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1769 1770 zilog->zl_get_data = get_data; 1771 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1772 2, 2, TASKQ_PREPOPULATE); 1773 1774 return (zilog); 1775} 1776 1777/* 1778 * Close an intent log. 1779 */ 1780void 1781zil_close(zilog_t *zilog) 1782{ 1783 lwb_t *lwb; 1784 uint64_t txg = 0; 1785 1786 zil_commit(zilog, 0); /* commit all itx */ 1787 1788 /* 1789 * The lwb_max_txg for the stubby lwb will reflect the last activity 1790 * for the zil. After a txg_wait_synced() on the txg we know all the 1791 * callbacks have occurred that may clean the zil. Only then can we 1792 * destroy the zl_clean_taskq. 1793 */ 1794 mutex_enter(&zilog->zl_lock); 1795 lwb = list_tail(&zilog->zl_lwb_list); 1796 if (lwb != NULL) 1797 txg = lwb->lwb_max_txg; 1798 mutex_exit(&zilog->zl_lock); 1799 if (txg) 1800 txg_wait_synced(zilog->zl_dmu_pool, txg); 1801 ASSERT(!zilog_is_dirty(zilog)); 1802 1803 taskq_destroy(zilog->zl_clean_taskq); 1804 zilog->zl_clean_taskq = NULL; 1805 zilog->zl_get_data = NULL; 1806 1807 /* 1808 * We should have only one LWB left on the list; remove it now. 1809 */ 1810 mutex_enter(&zilog->zl_lock); 1811 lwb = list_head(&zilog->zl_lwb_list); 1812 if (lwb != NULL) { 1813 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1814 list_remove(&zilog->zl_lwb_list, lwb); 1815 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1816 kmem_cache_free(zil_lwb_cache, lwb); 1817 } 1818 mutex_exit(&zilog->zl_lock); 1819} 1820 1821static char *suspend_tag = "zil suspending"; 1822 1823/* 1824 * Suspend an intent log. While in suspended mode, we still honor 1825 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1826 * On old version pools, we suspend the log briefly when taking a 1827 * snapshot so that it will have an empty intent log. 1828 * 1829 * Long holds are not really intended to be used the way we do here -- 1830 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1831 * could fail. Therefore we take pains to only put a long hold if it is 1832 * actually necessary. Fortunately, it will only be necessary if the 1833 * objset is currently mounted (or the ZVOL equivalent). In that case it 1834 * will already have a long hold, so we are not really making things any worse. 1835 * 1836 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1837 * zvol_state_t), and use their mechanism to prevent their hold from being 1838 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1839 * very little gain. 1840 * 1841 * if cookiep == NULL, this does both the suspend & resume. 1842 * Otherwise, it returns with the dataset "long held", and the cookie 1843 * should be passed into zil_resume(). 1844 */ 1845int 1846zil_suspend(const char *osname, void **cookiep) 1847{ 1848 objset_t *os; 1849 zilog_t *zilog; 1850 const zil_header_t *zh; 1851 int error; 1852 1853 error = dmu_objset_hold(osname, suspend_tag, &os); 1854 if (error != 0) 1855 return (error); 1856 zilog = dmu_objset_zil(os); 1857 1858 mutex_enter(&zilog->zl_lock); 1859 zh = zilog->zl_header; 1860 1861 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1862 mutex_exit(&zilog->zl_lock); 1863 dmu_objset_rele(os, suspend_tag); 1864 return (SET_ERROR(EBUSY)); 1865 } 1866 1867 /* 1868 * Don't put a long hold in the cases where we can avoid it. This 1869 * is when there is no cookie so we are doing a suspend & resume 1870 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1871 * for the suspend because it's already suspended, or there's no ZIL. 1872 */ 1873 if (cookiep == NULL && !zilog->zl_suspending && 1874 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1875 mutex_exit(&zilog->zl_lock); 1876 dmu_objset_rele(os, suspend_tag); 1877 return (0); 1878 } 1879 1880 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1881 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1882 1883 zilog->zl_suspend++; 1884 1885 if (zilog->zl_suspend > 1) { 1886 /* 1887 * Someone else is already suspending it. 1888 * Just wait for them to finish. 1889 */ 1890 1891 while (zilog->zl_suspending) 1892 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1893 mutex_exit(&zilog->zl_lock); 1894 1895 if (cookiep == NULL) 1896 zil_resume(os); 1897 else 1898 *cookiep = os; 1899 return (0); 1900 } 1901 1902 /* 1903 * If there is no pointer to an on-disk block, this ZIL must not 1904 * be active (e.g. filesystem not mounted), so there's nothing 1905 * to clean up. 1906 */ 1907 if (BP_IS_HOLE(&zh->zh_log)) { 1908 ASSERT(cookiep != NULL); /* fast path already handled */ 1909 1910 *cookiep = os; 1911 mutex_exit(&zilog->zl_lock); 1912 return (0); 1913 } 1914 1915 zilog->zl_suspending = B_TRUE; 1916 mutex_exit(&zilog->zl_lock); 1917 1918 zil_commit(zilog, 0); 1919 1920 zil_destroy(zilog, B_FALSE); 1921 1922 mutex_enter(&zilog->zl_lock); 1923 zilog->zl_suspending = B_FALSE; 1924 cv_broadcast(&zilog->zl_cv_suspend); 1925 mutex_exit(&zilog->zl_lock); 1926 1927 if (cookiep == NULL) 1928 zil_resume(os); 1929 else 1930 *cookiep = os; 1931 return (0); 1932} 1933 1934void 1935zil_resume(void *cookie) 1936{ 1937 objset_t *os = cookie; 1938 zilog_t *zilog = dmu_objset_zil(os); 1939 1940 mutex_enter(&zilog->zl_lock); 1941 ASSERT(zilog->zl_suspend != 0); 1942 zilog->zl_suspend--; 1943 mutex_exit(&zilog->zl_lock); 1944 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1945 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1946} 1947 1948typedef struct zil_replay_arg { 1949 zil_replay_func_t **zr_replay; 1950 void *zr_arg; 1951 boolean_t zr_byteswap; 1952 char *zr_lr; 1953} zil_replay_arg_t; 1954 1955static int 1956zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1957{ 1958 char name[MAXNAMELEN]; 1959 1960 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1961 1962 dmu_objset_name(zilog->zl_os, name); 1963 1964 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1965 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1966 (u_longlong_t)lr->lrc_seq, 1967 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1968 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1969 1970 return (error); 1971} 1972 1973static int 1974zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1975{ 1976 zil_replay_arg_t *zr = zra; 1977 const zil_header_t *zh = zilog->zl_header; 1978 uint64_t reclen = lr->lrc_reclen; 1979 uint64_t txtype = lr->lrc_txtype; 1980 int error = 0; 1981 1982 zilog->zl_replaying_seq = lr->lrc_seq; 1983 1984 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1985 return (0); 1986 1987 if (lr->lrc_txg < claim_txg) /* already committed */ 1988 return (0); 1989 1990 /* Strip case-insensitive bit, still present in log record */ 1991 txtype &= ~TX_CI; 1992 1993 if (txtype == 0 || txtype >= TX_MAX_TYPE) 1994 return (zil_replay_error(zilog, lr, EINVAL)); 1995 1996 /* 1997 * If this record type can be logged out of order, the object 1998 * (lr_foid) may no longer exist. That's legitimate, not an error. 1999 */ 2000 if (TX_OOO(txtype)) { 2001 error = dmu_object_info(zilog->zl_os, 2002 ((lr_ooo_t *)lr)->lr_foid, NULL); 2003 if (error == ENOENT || error == EEXIST) 2004 return (0); 2005 } 2006 2007 /* 2008 * Make a copy of the data so we can revise and extend it. 2009 */ 2010 bcopy(lr, zr->zr_lr, reclen); 2011 2012 /* 2013 * If this is a TX_WRITE with a blkptr, suck in the data. 2014 */ 2015 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2016 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2017 zr->zr_lr + reclen); 2018 if (error != 0) 2019 return (zil_replay_error(zilog, lr, error)); 2020 } 2021 2022 /* 2023 * The log block containing this lr may have been byteswapped 2024 * so that we can easily examine common fields like lrc_txtype. 2025 * However, the log is a mix of different record types, and only the 2026 * replay vectors know how to byteswap their records. Therefore, if 2027 * the lr was byteswapped, undo it before invoking the replay vector. 2028 */ 2029 if (zr->zr_byteswap) 2030 byteswap_uint64_array(zr->zr_lr, reclen); 2031 2032 /* 2033 * We must now do two things atomically: replay this log record, 2034 * and update the log header sequence number to reflect the fact that 2035 * we did so. At the end of each replay function the sequence number 2036 * is updated if we are in replay mode. 2037 */ 2038 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2039 if (error != 0) { 2040 /* 2041 * The DMU's dnode layer doesn't see removes until the txg 2042 * commits, so a subsequent claim can spuriously fail with 2043 * EEXIST. So if we receive any error we try syncing out 2044 * any removes then retry the transaction. Note that we 2045 * specify B_FALSE for byteswap now, so we don't do it twice. 2046 */ 2047 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2048 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2049 if (error != 0) 2050 return (zil_replay_error(zilog, lr, error)); 2051 } 2052 return (0); 2053} 2054 2055/* ARGSUSED */ 2056static int 2057zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2058{ 2059 zilog->zl_replay_blks++; 2060 2061 return (0); 2062} 2063 2064/* 2065 * If this dataset has a non-empty intent log, replay it and destroy it. 2066 */ 2067void 2068zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2069{ 2070 zilog_t *zilog = dmu_objset_zil(os); 2071 const zil_header_t *zh = zilog->zl_header; 2072 zil_replay_arg_t zr; 2073 2074 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2075 zil_destroy(zilog, B_TRUE); 2076 return; 2077 } 2078 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 2079 2080 zr.zr_replay = replay_func; 2081 zr.zr_arg = arg; 2082 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2083 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2084 2085 /* 2086 * Wait for in-progress removes to sync before starting replay. 2087 */ 2088 txg_wait_synced(zilog->zl_dmu_pool, 0); 2089 2090 zilog->zl_replay = B_TRUE; 2091 zilog->zl_replay_time = ddi_get_lbolt(); 2092 ASSERT(zilog->zl_replay_blks == 0); 2093 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2094 zh->zh_claim_txg); 2095 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2096 2097 zil_destroy(zilog, B_FALSE); 2098 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2099 zilog->zl_replay = B_FALSE; 2100 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2101} 2102 2103boolean_t 2104zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2105{ 2106 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2107 return (B_TRUE); 2108 2109 if (zilog->zl_replay) { 2110 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2111 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2112 zilog->zl_replaying_seq; 2113 return (B_TRUE); 2114 } 2115 2116 return (B_FALSE); 2117} 2118 2119/* ARGSUSED */ 2120int 2121zil_vdev_offline(const char *osname, void *arg) 2122{ 2123 int error; 2124 2125 error = zil_suspend(osname, NULL); 2126 if (error != 0) 2127 return (SET_ERROR(EEXIST)); 2128 return (0); 2129} 2130