zil.c revision 315384
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2016 by Delphix. All rights reserved. 24 * Copyright (c) 2014 Integros [integros.com] 25 */ 26 27/* Portions Copyright 2010 Robert Milkowski */ 28 29#include <sys/zfs_context.h> 30#include <sys/spa.h> 31#include <sys/dmu.h> 32#include <sys/zap.h> 33#include <sys/arc.h> 34#include <sys/stat.h> 35#include <sys/resource.h> 36#include <sys/zil.h> 37#include <sys/zil_impl.h> 38#include <sys/dsl_dataset.h> 39#include <sys/vdev_impl.h> 40#include <sys/dmu_tx.h> 41#include <sys/dsl_pool.h> 42 43/* 44 * The zfs intent log (ZIL) saves transaction records of system calls 45 * that change the file system in memory with enough information 46 * to be able to replay them. These are stored in memory until 47 * either the DMU transaction group (txg) commits them to the stable pool 48 * and they can be discarded, or they are flushed to the stable log 49 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 50 * requirement. In the event of a panic or power fail then those log 51 * records (transactions) are replayed. 52 * 53 * There is one ZIL per file system. Its on-disk (pool) format consists 54 * of 3 parts: 55 * 56 * - ZIL header 57 * - ZIL blocks 58 * - ZIL records 59 * 60 * A log record holds a system call transaction. Log blocks can 61 * hold many log records and the blocks are chained together. 62 * Each ZIL block contains a block pointer (blkptr_t) to the next 63 * ZIL block in the chain. The ZIL header points to the first 64 * block in the chain. Note there is not a fixed place in the pool 65 * to hold blocks. They are dynamically allocated and freed as 66 * needed from the blocks available. Figure X shows the ZIL structure: 67 */ 68 69/* 70 * Disable intent logging replay. This global ZIL switch affects all pools. 71 */ 72int zil_replay_disable = 0; 73SYSCTL_DECL(_vfs_zfs); 74SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RWTUN, 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; 83SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 84 &zfs_nocacheflush, 0, "Disable cache flush"); 85boolean_t zfs_trim_enabled = B_TRUE; 86SYSCTL_DECL(_vfs_zfs_trim); 87SYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 88 "Enable ZFS TRIM"); 89 90static kmem_cache_t *zil_lwb_cache; 91 92#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 93 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 94 95 96/* 97 * ziltest is by and large an ugly hack, but very useful in 98 * checking replay without tedious work. 99 * When running ziltest we want to keep all itx's and so maintain 100 * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 101 * We subtract TXG_CONCURRENT_STATES to allow for common code. 102 */ 103#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 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, 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_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 457 lwb->lwb_max_txg = txg; 458 lwb->lwb_zio = NULL; 459 lwb->lwb_tx = NULL; 460 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 461 lwb->lwb_nused = sizeof (zil_chain_t); 462 lwb->lwb_sz = BP_GET_LSIZE(bp); 463 } else { 464 lwb->lwb_nused = 0; 465 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 466 } 467 468 mutex_enter(&zilog->zl_lock); 469 list_insert_tail(&zilog->zl_lwb_list, lwb); 470 mutex_exit(&zilog->zl_lock); 471 472 return (lwb); 473} 474 475/* 476 * Called when we create in-memory log transactions so that we know 477 * to cleanup the itxs at the end of spa_sync(). 478 */ 479void 480zilog_dirty(zilog_t *zilog, uint64_t txg) 481{ 482 dsl_pool_t *dp = zilog->zl_dmu_pool; 483 dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 484 485 if (ds->ds_is_snapshot) 486 panic("dirtying snapshot!"); 487 488 if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 489 /* up the hold count until we can be written out */ 490 dmu_buf_add_ref(ds->ds_dbuf, zilog); 491 } 492} 493 494/* 495 * Determine if the zil is dirty in the specified txg. Callers wanting to 496 * ensure that the dirty state does not change must hold the itxg_lock for 497 * the specified txg. Holding the lock will ensure that the zil cannot be 498 * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 499 * state. 500 */ 501boolean_t 502zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 503{ 504 dsl_pool_t *dp = zilog->zl_dmu_pool; 505 506 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 507 return (B_TRUE); 508 return (B_FALSE); 509} 510 511/* 512 * Determine if the zil is dirty. The zil is considered dirty if it has 513 * any pending itx records that have not been cleaned by zil_clean(). 514 */ 515boolean_t 516zilog_is_dirty(zilog_t *zilog) 517{ 518 dsl_pool_t *dp = zilog->zl_dmu_pool; 519 520 for (int t = 0; t < TXG_SIZE; t++) { 521 if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 522 return (B_TRUE); 523 } 524 return (B_FALSE); 525} 526 527/* 528 * Create an on-disk intent log. 529 */ 530static lwb_t * 531zil_create(zilog_t *zilog) 532{ 533 const zil_header_t *zh = zilog->zl_header; 534 lwb_t *lwb = NULL; 535 uint64_t txg = 0; 536 dmu_tx_t *tx = NULL; 537 blkptr_t blk; 538 int error = 0; 539 540 /* 541 * Wait for any previous destroy to complete. 542 */ 543 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 544 545 ASSERT(zh->zh_claim_txg == 0); 546 ASSERT(zh->zh_replay_seq == 0); 547 548 blk = zh->zh_log; 549 550 /* 551 * Allocate an initial log block if: 552 * - there isn't one already 553 * - the existing block is the wrong endianess 554 */ 555 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 556 tx = dmu_tx_create(zilog->zl_os); 557 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 558 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 559 txg = dmu_tx_get_txg(tx); 560 561 if (!BP_IS_HOLE(&blk)) { 562 zio_free_zil(zilog->zl_spa, txg, &blk); 563 BP_ZERO(&blk); 564 } 565 566 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 567 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 568 569 if (error == 0) 570 zil_init_log_chain(zilog, &blk); 571 } 572 573 /* 574 * Allocate a log write buffer (lwb) for the first log block. 575 */ 576 if (error == 0) 577 lwb = zil_alloc_lwb(zilog, &blk, txg); 578 579 /* 580 * If we just allocated the first log block, commit our transaction 581 * and wait for zil_sync() to stuff the block poiner into zh_log. 582 * (zh is part of the MOS, so we cannot modify it in open context.) 583 */ 584 if (tx != NULL) { 585 dmu_tx_commit(tx); 586 txg_wait_synced(zilog->zl_dmu_pool, txg); 587 } 588 589 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 590 591 return (lwb); 592} 593 594/* 595 * In one tx, free all log blocks and clear the log header. 596 * If keep_first is set, then we're replaying a log with no content. 597 * We want to keep the first block, however, so that the first 598 * synchronous transaction doesn't require a txg_wait_synced() 599 * in zil_create(). We don't need to txg_wait_synced() here either 600 * when keep_first is set, because both zil_create() and zil_destroy() 601 * will wait for any in-progress destroys to complete. 602 */ 603void 604zil_destroy(zilog_t *zilog, boolean_t keep_first) 605{ 606 const zil_header_t *zh = zilog->zl_header; 607 lwb_t *lwb; 608 dmu_tx_t *tx; 609 uint64_t txg; 610 611 /* 612 * Wait for any previous destroy to complete. 613 */ 614 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 615 616 zilog->zl_old_header = *zh; /* debugging aid */ 617 618 if (BP_IS_HOLE(&zh->zh_log)) 619 return; 620 621 tx = dmu_tx_create(zilog->zl_os); 622 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 623 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 624 txg = dmu_tx_get_txg(tx); 625 626 mutex_enter(&zilog->zl_lock); 627 628 ASSERT3U(zilog->zl_destroy_txg, <, txg); 629 zilog->zl_destroy_txg = txg; 630 zilog->zl_keep_first = keep_first; 631 632 if (!list_is_empty(&zilog->zl_lwb_list)) { 633 ASSERT(zh->zh_claim_txg == 0); 634 VERIFY(!keep_first); 635 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 636 list_remove(&zilog->zl_lwb_list, lwb); 637 if (lwb->lwb_buf != NULL) 638 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 639 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 640 kmem_cache_free(zil_lwb_cache, lwb); 641 } 642 } else if (!keep_first) { 643 zil_destroy_sync(zilog, tx); 644 } 645 mutex_exit(&zilog->zl_lock); 646 647 dmu_tx_commit(tx); 648} 649 650void 651zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 652{ 653 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 654 (void) zil_parse(zilog, zil_free_log_block, 655 zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 656} 657 658int 659zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 660{ 661 dmu_tx_t *tx = txarg; 662 uint64_t first_txg = dmu_tx_get_txg(tx); 663 zilog_t *zilog; 664 zil_header_t *zh; 665 objset_t *os; 666 int error; 667 668 error = dmu_objset_own_obj(dp, ds->ds_object, 669 DMU_OST_ANY, B_FALSE, FTAG, &os); 670 if (error != 0) { 671 /* 672 * EBUSY indicates that the objset is inconsistent, in which 673 * case it can not have a ZIL. 674 */ 675 if (error != EBUSY) { 676 cmn_err(CE_WARN, "can't open objset for %llu, error %u", 677 (unsigned long long)ds->ds_object, error); 678 } 679 return (0); 680 } 681 682 zilog = dmu_objset_zil(os); 683 zh = zil_header_in_syncing_context(zilog); 684 685 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 686 if (!BP_IS_HOLE(&zh->zh_log)) 687 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 688 BP_ZERO(&zh->zh_log); 689 dsl_dataset_dirty(dmu_objset_ds(os), tx); 690 dmu_objset_disown(os, FTAG); 691 return (0); 692 } 693 694 /* 695 * Claim all log blocks if we haven't already done so, and remember 696 * the highest claimed sequence number. This ensures that if we can 697 * read only part of the log now (e.g. due to a missing device), 698 * but we can read the entire log later, we will not try to replay 699 * or destroy beyond the last block we successfully claimed. 700 */ 701 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 702 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 703 (void) zil_parse(zilog, zil_claim_log_block, 704 zil_claim_log_record, tx, first_txg); 705 zh->zh_claim_txg = first_txg; 706 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 707 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 708 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 709 zh->zh_flags |= ZIL_REPLAY_NEEDED; 710 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 711 dsl_dataset_dirty(dmu_objset_ds(os), tx); 712 } 713 714 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 715 dmu_objset_disown(os, FTAG); 716 return (0); 717} 718 719/* 720 * Check the log by walking the log chain. 721 * Checksum errors are ok as they indicate the end of the chain. 722 * Any other error (no device or read failure) returns an error. 723 */ 724/* ARGSUSED */ 725int 726zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 727{ 728 zilog_t *zilog; 729 objset_t *os; 730 blkptr_t *bp; 731 int error; 732 733 ASSERT(tx == NULL); 734 735 error = dmu_objset_from_ds(ds, &os); 736 if (error != 0) { 737 cmn_err(CE_WARN, "can't open objset %llu, error %d", 738 (unsigned long long)ds->ds_object, error); 739 return (0); 740 } 741 742 zilog = dmu_objset_zil(os); 743 bp = (blkptr_t *)&zilog->zl_header->zh_log; 744 745 /* 746 * Check the first block and determine if it's on a log device 747 * which may have been removed or faulted prior to loading this 748 * pool. If so, there's no point in checking the rest of the log 749 * as its content should have already been synced to the pool. 750 */ 751 if (!BP_IS_HOLE(bp)) { 752 vdev_t *vd; 753 boolean_t valid = B_TRUE; 754 755 spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 756 vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 757 if (vd->vdev_islog && vdev_is_dead(vd)) 758 valid = vdev_log_state_valid(vd); 759 spa_config_exit(os->os_spa, SCL_STATE, FTAG); 760 761 if (!valid) 762 return (0); 763 } 764 765 /* 766 * Because tx == NULL, zil_claim_log_block() will not actually claim 767 * any blocks, but just determine whether it is possible to do so. 768 * In addition to checking the log chain, zil_claim_log_block() 769 * will invoke zio_claim() with a done func of spa_claim_notify(), 770 * which will update spa_max_claim_txg. See spa_load() for details. 771 */ 772 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 773 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 774 775 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 776} 777 778static int 779zil_vdev_compare(const void *x1, const void *x2) 780{ 781 const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 782 const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 783 784 if (v1 < v2) 785 return (-1); 786 if (v1 > v2) 787 return (1); 788 789 return (0); 790} 791 792void 793zil_add_block(zilog_t *zilog, const blkptr_t *bp) 794{ 795 avl_tree_t *t = &zilog->zl_vdev_tree; 796 avl_index_t where; 797 zil_vdev_node_t *zv, zvsearch; 798 int ndvas = BP_GET_NDVAS(bp); 799 int i; 800 801 if (zfs_nocacheflush) 802 return; 803 804 ASSERT(zilog->zl_writer); 805 806 /* 807 * Even though we're zl_writer, we still need a lock because the 808 * zl_get_data() callbacks may have dmu_sync() done callbacks 809 * that will run concurrently. 810 */ 811 mutex_enter(&zilog->zl_vdev_lock); 812 for (i = 0; i < ndvas; i++) { 813 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 814 if (avl_find(t, &zvsearch, &where) == NULL) { 815 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 816 zv->zv_vdev = zvsearch.zv_vdev; 817 avl_insert(t, zv, where); 818 } 819 } 820 mutex_exit(&zilog->zl_vdev_lock); 821} 822 823static void 824zil_flush_vdevs(zilog_t *zilog) 825{ 826 spa_t *spa = zilog->zl_spa; 827 avl_tree_t *t = &zilog->zl_vdev_tree; 828 void *cookie = NULL; 829 zil_vdev_node_t *zv; 830 zio_t *zio = NULL; 831 832 ASSERT(zilog->zl_writer); 833 834 /* 835 * We don't need zl_vdev_lock here because we're the zl_writer, 836 * and all zl_get_data() callbacks are done. 837 */ 838 if (avl_numnodes(t) == 0) 839 return; 840 841 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 842 843 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 844 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 845 if (vd != NULL && !vd->vdev_nowritecache) { 846 if (zio == NULL) 847 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 848 zio_flush(zio, vd); 849 } 850 kmem_free(zv, sizeof (*zv)); 851 } 852 853 /* 854 * Wait for all the flushes to complete. Not all devices actually 855 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 856 */ 857 if (zio) 858 (void) zio_wait(zio); 859 860 spa_config_exit(spa, SCL_STATE, FTAG); 861} 862 863/* 864 * Function called when a log block write completes 865 */ 866static void 867zil_lwb_write_done(zio_t *zio) 868{ 869 lwb_t *lwb = zio->io_private; 870 zilog_t *zilog = lwb->lwb_zilog; 871 dmu_tx_t *tx = lwb->lwb_tx; 872 873 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 874 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 875 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 876 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 877 ASSERT(!BP_IS_GANG(zio->io_bp)); 878 ASSERT(!BP_IS_HOLE(zio->io_bp)); 879 ASSERT(BP_GET_FILL(zio->io_bp) == 0); 880 881 /* 882 * Ensure the lwb buffer pointer is cleared before releasing 883 * the txg. If we have had an allocation failure and 884 * the txg is waiting to sync then we want want zil_sync() 885 * to remove the lwb so that it's not picked up as the next new 886 * one in zil_commit_writer(). zil_sync() will only remove 887 * the lwb if lwb_buf is null. 888 */ 889 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 890 mutex_enter(&zilog->zl_lock); 891 lwb->lwb_buf = NULL; 892 lwb->lwb_tx = NULL; 893 mutex_exit(&zilog->zl_lock); 894 895 /* 896 * Now that we've written this log block, we have a stable pointer 897 * to the next block in the chain, so it's OK to let the txg in 898 * which we allocated the next block sync. 899 */ 900 dmu_tx_commit(tx); 901} 902 903/* 904 * Initialize the io for a log block. 905 */ 906static void 907zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 908{ 909 zbookmark_phys_t zb; 910 911 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 912 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 913 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 914 915 if (zilog->zl_root_zio == NULL) { 916 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 917 ZIO_FLAG_CANFAIL); 918 } 919 if (lwb->lwb_zio == NULL) { 920 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 921 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 922 zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 923 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 924 } 925} 926 927/* 928 * Define a limited set of intent log block sizes. 929 * 930 * These must be a multiple of 4KB. Note only the amount used (again 931 * aligned to 4KB) actually gets written. However, we can't always just 932 * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 933 */ 934uint64_t zil_block_buckets[] = { 935 4096, /* non TX_WRITE */ 936 8192+4096, /* data base */ 937 32*1024 + 4096, /* NFS writes */ 938 UINT64_MAX 939}; 940 941/* 942 * Use the slog as long as the logbias is 'latency' and the current commit size 943 * is less than the limit or the total list size is less than 2X the limit. 944 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 945 */ 946uint64_t zil_slog_limit = 1024 * 1024; 947#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 948 (((zilog)->zl_cur_used < zil_slog_limit) || \ 949 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 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) 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; 966 967 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 968 zilc = (zil_chain_t *)lwb->lwb_buf; 969 bp = &zilc->zc_next_blk; 970 } else { 971 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 972 bp = &zilc->zc_next_blk; 973 } 974 975 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 976 977 /* 978 * Allocate the next block and save its address in this block 979 * before writing it in order to establish the log chain. 980 * Note that if the allocation of nlwb synced before we wrote 981 * the block that points at it (lwb), we'd leak it if we crashed. 982 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 983 * We dirty the dataset to ensure that zil_sync() will be called 984 * to clean up in the event of allocation failure or I/O failure. 985 */ 986 tx = dmu_tx_create(zilog->zl_os); 987 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 988 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 989 txg = dmu_tx_get_txg(tx); 990 991 lwb->lwb_tx = tx; 992 993 /* 994 * Log blocks are pre-allocated. Here we select the size of the next 995 * block, based on size used in the last block. 996 * - first find the smallest bucket that will fit the block from a 997 * limited set of block sizes. This is because it's faster to write 998 * blocks allocated from the same metaslab as they are adjacent or 999 * close. 1000 * - next find the maximum from the new suggested size and an array of 1001 * previous sizes. This lessens a picket fence effect of wrongly 1002 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 1003 * requests. 1004 * 1005 * Note we only write what is used, but we can't just allocate 1006 * the maximum block size because we can exhaust the available 1007 * pool log space. 1008 */ 1009 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1010 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 1011 continue; 1012 zil_blksz = zil_block_buckets[i]; 1013 if (zil_blksz == UINT64_MAX) 1014 zil_blksz = SPA_OLD_MAXBLOCKSIZE; 1015 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1016 for (i = 0; i < ZIL_PREV_BLKS; i++) 1017 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1018 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1019 1020 BP_ZERO(bp); 1021 /* pass the old blkptr in order to spread log blocks across devs */ 1022 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 1023 USE_SLOG(zilog)); 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, 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 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1058 1059 /* 1060 * If there was an allocation failure then nlwb will be null which 1061 * forces a txg_wait_synced(). 1062 */ 1063 return (nlwb); 1064} 1065 1066static lwb_t * 1067zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1068{ 1069 lr_t *lrc = &itx->itx_lr; /* common log record */ 1070 lr_write_t *lrw = (lr_write_t *)lrc; 1071 char *lr_buf; 1072 uint64_t txg = lrc->lrc_txg; 1073 uint64_t reclen = lrc->lrc_reclen; 1074 uint64_t dlen = 0; 1075 1076 if (lwb == NULL) 1077 return (NULL); 1078 1079 ASSERT(lwb->lwb_buf != NULL); 1080 1081 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1082 dlen = P2ROUNDUP_TYPED( 1083 lrw->lr_length, sizeof (uint64_t), uint64_t); 1084 1085 zilog->zl_cur_used += (reclen + dlen); 1086 1087 zil_lwb_write_init(zilog, lwb); 1088 1089 /* 1090 * If this record won't fit in the current log block, start a new one. 1091 */ 1092 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1093 lwb = zil_lwb_write_start(zilog, lwb); 1094 if (lwb == NULL) 1095 return (NULL); 1096 zil_lwb_write_init(zilog, lwb); 1097 ASSERT(LWB_EMPTY(lwb)); 1098 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1099 txg_wait_synced(zilog->zl_dmu_pool, txg); 1100 return (lwb); 1101 } 1102 } 1103 1104 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1105 bcopy(lrc, lr_buf, reclen); 1106 lrc = (lr_t *)lr_buf; 1107 lrw = (lr_write_t *)lrc; 1108 1109 /* 1110 * If it's a write, fetch the data or get its blkptr as appropriate. 1111 */ 1112 if (lrc->lrc_txtype == TX_WRITE) { 1113 if (txg > spa_freeze_txg(zilog->zl_spa)) 1114 txg_wait_synced(zilog->zl_dmu_pool, txg); 1115 if (itx->itx_wr_state != WR_COPIED) { 1116 char *dbuf; 1117 int error; 1118 1119 if (dlen) { 1120 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1121 dbuf = lr_buf + reclen; 1122 lrw->lr_common.lrc_reclen += dlen; 1123 } else { 1124 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1125 dbuf = NULL; 1126 } 1127 error = zilog->zl_get_data( 1128 itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1129 if (error == EIO) { 1130 txg_wait_synced(zilog->zl_dmu_pool, txg); 1131 return (lwb); 1132 } 1133 if (error != 0) { 1134 ASSERT(error == ENOENT || error == EEXIST || 1135 error == EALREADY); 1136 return (lwb); 1137 } 1138 } 1139 } 1140 1141 /* 1142 * We're actually making an entry, so update lrc_seq to be the 1143 * log record sequence number. Note that this is generally not 1144 * equal to the itx sequence number because not all transactions 1145 * are synchronous, and sometimes spa_sync() gets there first. 1146 */ 1147 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1148 lwb->lwb_nused += reclen + dlen; 1149 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1150 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1151 ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1152 1153 return (lwb); 1154} 1155 1156itx_t * 1157zil_itx_create(uint64_t txtype, size_t lrsize) 1158{ 1159 itx_t *itx; 1160 1161 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1162 1163 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1164 itx->itx_lr.lrc_txtype = txtype; 1165 itx->itx_lr.lrc_reclen = lrsize; 1166 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1167 itx->itx_lr.lrc_seq = 0; /* defensive */ 1168 itx->itx_sync = B_TRUE; /* default is synchronous */ 1169 1170 return (itx); 1171} 1172 1173void 1174zil_itx_destroy(itx_t *itx) 1175{ 1176 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1177} 1178 1179/* 1180 * Free up the sync and async itxs. The itxs_t has already been detached 1181 * so no locks are needed. 1182 */ 1183static void 1184zil_itxg_clean(itxs_t *itxs) 1185{ 1186 itx_t *itx; 1187 list_t *list; 1188 avl_tree_t *t; 1189 void *cookie; 1190 itx_async_node_t *ian; 1191 1192 list = &itxs->i_sync_list; 1193 while ((itx = list_head(list)) != NULL) { 1194 list_remove(list, itx); 1195 kmem_free(itx, offsetof(itx_t, itx_lr) + 1196 itx->itx_lr.lrc_reclen); 1197 } 1198 1199 cookie = NULL; 1200 t = &itxs->i_async_tree; 1201 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1202 list = &ian->ia_list; 1203 while ((itx = list_head(list)) != NULL) { 1204 list_remove(list, itx); 1205 kmem_free(itx, offsetof(itx_t, itx_lr) + 1206 itx->itx_lr.lrc_reclen); 1207 } 1208 list_destroy(list); 1209 kmem_free(ian, sizeof (itx_async_node_t)); 1210 } 1211 avl_destroy(t); 1212 1213 kmem_free(itxs, sizeof (itxs_t)); 1214} 1215 1216static int 1217zil_aitx_compare(const void *x1, const void *x2) 1218{ 1219 const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1220 const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1221 1222 if (o1 < o2) 1223 return (-1); 1224 if (o1 > o2) 1225 return (1); 1226 1227 return (0); 1228} 1229 1230/* 1231 * Remove all async itx with the given oid. 1232 */ 1233static void 1234zil_remove_async(zilog_t *zilog, uint64_t oid) 1235{ 1236 uint64_t otxg, txg; 1237 itx_async_node_t *ian; 1238 avl_tree_t *t; 1239 avl_index_t where; 1240 list_t clean_list; 1241 itx_t *itx; 1242 1243 ASSERT(oid != 0); 1244 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1245 1246 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1247 otxg = ZILTEST_TXG; 1248 else 1249 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1250 1251 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1252 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1253 1254 mutex_enter(&itxg->itxg_lock); 1255 if (itxg->itxg_txg != txg) { 1256 mutex_exit(&itxg->itxg_lock); 1257 continue; 1258 } 1259 1260 /* 1261 * Locate the object node and append its list. 1262 */ 1263 t = &itxg->itxg_itxs->i_async_tree; 1264 ian = avl_find(t, &oid, &where); 1265 if (ian != NULL) 1266 list_move_tail(&clean_list, &ian->ia_list); 1267 mutex_exit(&itxg->itxg_lock); 1268 } 1269 while ((itx = list_head(&clean_list)) != NULL) { 1270 list_remove(&clean_list, itx); 1271 kmem_free(itx, offsetof(itx_t, itx_lr) + 1272 itx->itx_lr.lrc_reclen); 1273 } 1274 list_destroy(&clean_list); 1275} 1276 1277void 1278zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1279{ 1280 uint64_t txg; 1281 itxg_t *itxg; 1282 itxs_t *itxs, *clean = NULL; 1283 1284 /* 1285 * Object ids can be re-instantiated in the next txg so 1286 * remove any async transactions to avoid future leaks. 1287 * This can happen if a fsync occurs on the re-instantiated 1288 * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1289 * the new file data and flushes a write record for the old object. 1290 */ 1291 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1292 zil_remove_async(zilog, itx->itx_oid); 1293 1294 /* 1295 * Ensure the data of a renamed file is committed before the rename. 1296 */ 1297 if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1298 zil_async_to_sync(zilog, itx->itx_oid); 1299 1300 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1301 txg = ZILTEST_TXG; 1302 else 1303 txg = dmu_tx_get_txg(tx); 1304 1305 itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1306 mutex_enter(&itxg->itxg_lock); 1307 itxs = itxg->itxg_itxs; 1308 if (itxg->itxg_txg != txg) { 1309 if (itxs != NULL) { 1310 /* 1311 * The zil_clean callback hasn't got around to cleaning 1312 * this itxg. Save the itxs for release below. 1313 * This should be rare. 1314 */ 1315 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1316 itxg->itxg_sod = 0; 1317 clean = itxg->itxg_itxs; 1318 } 1319 ASSERT(itxg->itxg_sod == 0); 1320 itxg->itxg_txg = txg; 1321 itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1322 1323 list_create(&itxs->i_sync_list, sizeof (itx_t), 1324 offsetof(itx_t, itx_node)); 1325 avl_create(&itxs->i_async_tree, zil_aitx_compare, 1326 sizeof (itx_async_node_t), 1327 offsetof(itx_async_node_t, ia_node)); 1328 } 1329 if (itx->itx_sync) { 1330 list_insert_tail(&itxs->i_sync_list, itx); 1331 atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1332 itxg->itxg_sod += itx->itx_sod; 1333 } else { 1334 avl_tree_t *t = &itxs->i_async_tree; 1335 uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1336 itx_async_node_t *ian; 1337 avl_index_t where; 1338 1339 ian = avl_find(t, &foid, &where); 1340 if (ian == NULL) { 1341 ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1342 list_create(&ian->ia_list, sizeof (itx_t), 1343 offsetof(itx_t, itx_node)); 1344 ian->ia_foid = foid; 1345 avl_insert(t, ian, where); 1346 } 1347 list_insert_tail(&ian->ia_list, itx); 1348 } 1349 1350 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1351 zilog_dirty(zilog, txg); 1352 mutex_exit(&itxg->itxg_lock); 1353 1354 /* Release the old itxs now we've dropped the lock */ 1355 if (clean != NULL) 1356 zil_itxg_clean(clean); 1357} 1358 1359/* 1360 * If there are any in-memory intent log transactions which have now been 1361 * synced then start up a taskq to free them. We should only do this after we 1362 * have written out the uberblocks (i.e. txg has been comitted) so that 1363 * don't inadvertently clean out in-memory log records that would be required 1364 * by zil_commit(). 1365 */ 1366void 1367zil_clean(zilog_t *zilog, uint64_t synced_txg) 1368{ 1369 itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1370 itxs_t *clean_me; 1371 1372 mutex_enter(&itxg->itxg_lock); 1373 if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1374 mutex_exit(&itxg->itxg_lock); 1375 return; 1376 } 1377 ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1378 ASSERT(itxg->itxg_txg != 0); 1379 ASSERT(zilog->zl_clean_taskq != NULL); 1380 atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1381 itxg->itxg_sod = 0; 1382 clean_me = itxg->itxg_itxs; 1383 itxg->itxg_itxs = NULL; 1384 itxg->itxg_txg = 0; 1385 mutex_exit(&itxg->itxg_lock); 1386 /* 1387 * Preferably start a task queue to free up the old itxs but 1388 * if taskq_dispatch can't allocate resources to do that then 1389 * free it in-line. This should be rare. Note, using TQ_SLEEP 1390 * created a bad performance problem. 1391 */ 1392 if (taskq_dispatch(zilog->zl_clean_taskq, 1393 (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1394 zil_itxg_clean(clean_me); 1395} 1396 1397/* 1398 * Get the list of itxs to commit into zl_itx_commit_list. 1399 */ 1400static void 1401zil_get_commit_list(zilog_t *zilog) 1402{ 1403 uint64_t otxg, txg; 1404 list_t *commit_list = &zilog->zl_itx_commit_list; 1405 uint64_t push_sod = 0; 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 /* 1413 * This is inherently racy, since there is nothing to prevent 1414 * the last synced txg from changing. That's okay since we'll 1415 * only commit things in the future. 1416 */ 1417 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1418 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1419 1420 mutex_enter(&itxg->itxg_lock); 1421 if (itxg->itxg_txg != txg) { 1422 mutex_exit(&itxg->itxg_lock); 1423 continue; 1424 } 1425 1426 /* 1427 * If we're adding itx records to the zl_itx_commit_list, 1428 * then the zil better be dirty in this "txg". We can assert 1429 * that here since we're holding the itxg_lock which will 1430 * prevent spa_sync from cleaning it. Once we add the itxs 1431 * to the zl_itx_commit_list we must commit it to disk even 1432 * if it's unnecessary (i.e. the txg was synced). 1433 */ 1434 ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 1435 spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1436 list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1437 push_sod += itxg->itxg_sod; 1438 itxg->itxg_sod = 0; 1439 1440 mutex_exit(&itxg->itxg_lock); 1441 } 1442 atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1443} 1444 1445/* 1446 * Move the async itxs for a specified object to commit into sync lists. 1447 */ 1448void 1449zil_async_to_sync(zilog_t *zilog, uint64_t foid) 1450{ 1451 uint64_t otxg, txg; 1452 itx_async_node_t *ian; 1453 avl_tree_t *t; 1454 avl_index_t where; 1455 1456 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1457 otxg = ZILTEST_TXG; 1458 else 1459 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1460 1461 /* 1462 * This is inherently racy, since there is nothing to prevent 1463 * the last synced txg from changing. 1464 */ 1465 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1466 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1467 1468 mutex_enter(&itxg->itxg_lock); 1469 if (itxg->itxg_txg != txg) { 1470 mutex_exit(&itxg->itxg_lock); 1471 continue; 1472 } 1473 1474 /* 1475 * If a foid is specified then find that node and append its 1476 * list. Otherwise walk the tree appending all the lists 1477 * to the sync list. We add to the end rather than the 1478 * beginning to ensure the create has happened. 1479 */ 1480 t = &itxg->itxg_itxs->i_async_tree; 1481 if (foid != 0) { 1482 ian = avl_find(t, &foid, &where); 1483 if (ian != NULL) { 1484 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1485 &ian->ia_list); 1486 } 1487 } else { 1488 void *cookie = NULL; 1489 1490 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1491 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1492 &ian->ia_list); 1493 list_destroy(&ian->ia_list); 1494 kmem_free(ian, sizeof (itx_async_node_t)); 1495 } 1496 } 1497 mutex_exit(&itxg->itxg_lock); 1498 } 1499} 1500 1501static void 1502zil_commit_writer(zilog_t *zilog) 1503{ 1504 uint64_t txg; 1505 itx_t *itx; 1506 lwb_t *lwb; 1507 spa_t *spa = zilog->zl_spa; 1508 int error = 0; 1509 1510 ASSERT(zilog->zl_root_zio == NULL); 1511 1512 mutex_exit(&zilog->zl_lock); 1513 1514 zil_get_commit_list(zilog); 1515 1516 /* 1517 * Return if there's nothing to commit before we dirty the fs by 1518 * calling zil_create(). 1519 */ 1520 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1521 mutex_enter(&zilog->zl_lock); 1522 return; 1523 } 1524 1525 if (zilog->zl_suspend) { 1526 lwb = NULL; 1527 } else { 1528 lwb = list_tail(&zilog->zl_lwb_list); 1529 if (lwb == NULL) 1530 lwb = zil_create(zilog); 1531 } 1532 1533 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1534 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1535 txg = itx->itx_lr.lrc_txg; 1536 ASSERT3U(txg, !=, 0); 1537 1538 /* 1539 * This is inherently racy and may result in us writing 1540 * out a log block for a txg that was just synced. This is 1541 * ok since we'll end cleaning up that log block the next 1542 * time we call zil_sync(). 1543 */ 1544 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1545 lwb = zil_lwb_commit(zilog, itx, lwb); 1546 list_remove(&zilog->zl_itx_commit_list, itx); 1547 kmem_free(itx, offsetof(itx_t, itx_lr) 1548 + itx->itx_lr.lrc_reclen); 1549 } 1550 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1551 1552 /* write the last block out */ 1553 if (lwb != NULL && lwb->lwb_zio != NULL) 1554 lwb = zil_lwb_write_start(zilog, lwb); 1555 1556 zilog->zl_cur_used = 0; 1557 1558 /* 1559 * Wait if necessary for the log blocks to be on stable storage. 1560 */ 1561 if (zilog->zl_root_zio) { 1562 error = zio_wait(zilog->zl_root_zio); 1563 zilog->zl_root_zio = NULL; 1564 zil_flush_vdevs(zilog); 1565 } 1566 1567 if (error || lwb == NULL) 1568 txg_wait_synced(zilog->zl_dmu_pool, 0); 1569 1570 mutex_enter(&zilog->zl_lock); 1571 1572 /* 1573 * Remember the highest committed log sequence number for ztest. 1574 * We only update this value when all the log writes succeeded, 1575 * because ztest wants to ASSERT that it got the whole log chain. 1576 */ 1577 if (error == 0 && lwb != NULL) 1578 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1579} 1580 1581/* 1582 * Commit zfs transactions to stable storage. 1583 * If foid is 0 push out all transactions, otherwise push only those 1584 * for that object or might reference that object. 1585 * 1586 * itxs are committed in batches. In a heavily stressed zil there will be 1587 * a commit writer thread who is writing out a bunch of itxs to the log 1588 * for a set of committing threads (cthreads) in the same batch as the writer. 1589 * Those cthreads are all waiting on the same cv for that batch. 1590 * 1591 * There will also be a different and growing batch of threads that are 1592 * waiting to commit (qthreads). When the committing batch completes 1593 * a transition occurs such that the cthreads exit and the qthreads become 1594 * cthreads. One of the new cthreads becomes the writer thread for the 1595 * batch. Any new threads arriving become new qthreads. 1596 * 1597 * Only 2 condition variables are needed and there's no transition 1598 * between the two cvs needed. They just flip-flop between qthreads 1599 * and cthreads. 1600 * 1601 * Using this scheme we can efficiently wakeup up only those threads 1602 * that have been committed. 1603 */ 1604void 1605zil_commit(zilog_t *zilog, uint64_t foid) 1606{ 1607 uint64_t mybatch; 1608 1609 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1610 return; 1611 1612 /* move the async itxs for the foid to the sync queues */ 1613 zil_async_to_sync(zilog, foid); 1614 1615 mutex_enter(&zilog->zl_lock); 1616 mybatch = zilog->zl_next_batch; 1617 while (zilog->zl_writer) { 1618 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1619 if (mybatch <= zilog->zl_com_batch) { 1620 mutex_exit(&zilog->zl_lock); 1621 return; 1622 } 1623 } 1624 1625 zilog->zl_next_batch++; 1626 zilog->zl_writer = B_TRUE; 1627 zil_commit_writer(zilog); 1628 zilog->zl_com_batch = mybatch; 1629 zilog->zl_writer = B_FALSE; 1630 mutex_exit(&zilog->zl_lock); 1631 1632 /* wake up one thread to become the next writer */ 1633 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1634 1635 /* wake up all threads waiting for this batch to be committed */ 1636 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1637} 1638 1639/* 1640 * Called in syncing context to free committed log blocks and update log header. 1641 */ 1642void 1643zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1644{ 1645 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1646 uint64_t txg = dmu_tx_get_txg(tx); 1647 spa_t *spa = zilog->zl_spa; 1648 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1649 lwb_t *lwb; 1650 1651 /* 1652 * We don't zero out zl_destroy_txg, so make sure we don't try 1653 * to destroy it twice. 1654 */ 1655 if (spa_sync_pass(spa) != 1) 1656 return; 1657 1658 mutex_enter(&zilog->zl_lock); 1659 1660 ASSERT(zilog->zl_stop_sync == 0); 1661 1662 if (*replayed_seq != 0) { 1663 ASSERT(zh->zh_replay_seq < *replayed_seq); 1664 zh->zh_replay_seq = *replayed_seq; 1665 *replayed_seq = 0; 1666 } 1667 1668 if (zilog->zl_destroy_txg == txg) { 1669 blkptr_t blk = zh->zh_log; 1670 1671 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1672 1673 bzero(zh, sizeof (zil_header_t)); 1674 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1675 1676 if (zilog->zl_keep_first) { 1677 /* 1678 * If this block was part of log chain that couldn't 1679 * be claimed because a device was missing during 1680 * zil_claim(), but that device later returns, 1681 * then this block could erroneously appear valid. 1682 * To guard against this, assign a new GUID to the new 1683 * log chain so it doesn't matter what blk points to. 1684 */ 1685 zil_init_log_chain(zilog, &blk); 1686 zh->zh_log = blk; 1687 } 1688 } 1689 1690 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1691 zh->zh_log = lwb->lwb_blk; 1692 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1693 break; 1694 list_remove(&zilog->zl_lwb_list, lwb); 1695 zio_free_zil(spa, txg, &lwb->lwb_blk); 1696 kmem_cache_free(zil_lwb_cache, lwb); 1697 1698 /* 1699 * If we don't have anything left in the lwb list then 1700 * we've had an allocation failure and we need to zero 1701 * out the zil_header blkptr so that we don't end 1702 * up freeing the same block twice. 1703 */ 1704 if (list_head(&zilog->zl_lwb_list) == NULL) 1705 BP_ZERO(&zh->zh_log); 1706 } 1707 mutex_exit(&zilog->zl_lock); 1708} 1709 1710void 1711zil_init(void) 1712{ 1713 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1714 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1715} 1716 1717void 1718zil_fini(void) 1719{ 1720 kmem_cache_destroy(zil_lwb_cache); 1721} 1722 1723void 1724zil_set_sync(zilog_t *zilog, uint64_t sync) 1725{ 1726 zilog->zl_sync = sync; 1727} 1728 1729void 1730zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1731{ 1732 zilog->zl_logbias = logbias; 1733} 1734 1735zilog_t * 1736zil_alloc(objset_t *os, zil_header_t *zh_phys) 1737{ 1738 zilog_t *zilog; 1739 1740 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1741 1742 zilog->zl_header = zh_phys; 1743 zilog->zl_os = os; 1744 zilog->zl_spa = dmu_objset_spa(os); 1745 zilog->zl_dmu_pool = dmu_objset_pool(os); 1746 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1747 zilog->zl_logbias = dmu_objset_logbias(os); 1748 zilog->zl_sync = dmu_objset_syncprop(os); 1749 zilog->zl_next_batch = 1; 1750 1751 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1752 1753 for (int i = 0; i < TXG_SIZE; i++) { 1754 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1755 MUTEX_DEFAULT, NULL); 1756 } 1757 1758 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1759 offsetof(lwb_t, lwb_node)); 1760 1761 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1762 offsetof(itx_t, itx_node)); 1763 1764 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1765 1766 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1767 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1768 1769 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1770 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1771 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1772 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1773 1774 return (zilog); 1775} 1776 1777void 1778zil_free(zilog_t *zilog) 1779{ 1780 zilog->zl_stop_sync = 1; 1781 1782 ASSERT0(zilog->zl_suspend); 1783 ASSERT0(zilog->zl_suspending); 1784 1785 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1786 list_destroy(&zilog->zl_lwb_list); 1787 1788 avl_destroy(&zilog->zl_vdev_tree); 1789 mutex_destroy(&zilog->zl_vdev_lock); 1790 1791 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1792 list_destroy(&zilog->zl_itx_commit_list); 1793 1794 for (int i = 0; i < TXG_SIZE; i++) { 1795 /* 1796 * It's possible for an itx to be generated that doesn't dirty 1797 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1798 * callback to remove the entry. We remove those here. 1799 * 1800 * Also free up the ziltest itxs. 1801 */ 1802 if (zilog->zl_itxg[i].itxg_itxs) 1803 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1804 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1805 } 1806 1807 mutex_destroy(&zilog->zl_lock); 1808 1809 cv_destroy(&zilog->zl_cv_writer); 1810 cv_destroy(&zilog->zl_cv_suspend); 1811 cv_destroy(&zilog->zl_cv_batch[0]); 1812 cv_destroy(&zilog->zl_cv_batch[1]); 1813 1814 kmem_free(zilog, sizeof (zilog_t)); 1815} 1816 1817/* 1818 * Open an intent log. 1819 */ 1820zilog_t * 1821zil_open(objset_t *os, zil_get_data_t *get_data) 1822{ 1823 zilog_t *zilog = dmu_objset_zil(os); 1824 1825 ASSERT(zilog->zl_clean_taskq == NULL); 1826 ASSERT(zilog->zl_get_data == NULL); 1827 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1828 1829 zilog->zl_get_data = get_data; 1830 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1831 2, 2, TASKQ_PREPOPULATE); 1832 1833 return (zilog); 1834} 1835 1836/* 1837 * Close an intent log. 1838 */ 1839void 1840zil_close(zilog_t *zilog) 1841{ 1842 lwb_t *lwb; 1843 uint64_t txg = 0; 1844 1845 zil_commit(zilog, 0); /* commit all itx */ 1846 1847 /* 1848 * The lwb_max_txg for the stubby lwb will reflect the last activity 1849 * for the zil. After a txg_wait_synced() on the txg we know all the 1850 * callbacks have occurred that may clean the zil. Only then can we 1851 * destroy the zl_clean_taskq. 1852 */ 1853 mutex_enter(&zilog->zl_lock); 1854 lwb = list_tail(&zilog->zl_lwb_list); 1855 if (lwb != NULL) 1856 txg = lwb->lwb_max_txg; 1857 mutex_exit(&zilog->zl_lock); 1858 if (txg) 1859 txg_wait_synced(zilog->zl_dmu_pool, txg); 1860 1861 if (zilog_is_dirty(zilog)) 1862 zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg); 1863 VERIFY(!zilog_is_dirty(zilog)); 1864 1865 taskq_destroy(zilog->zl_clean_taskq); 1866 zilog->zl_clean_taskq = NULL; 1867 zilog->zl_get_data = NULL; 1868 1869 /* 1870 * We should have only one LWB left on the list; remove it now. 1871 */ 1872 mutex_enter(&zilog->zl_lock); 1873 lwb = list_head(&zilog->zl_lwb_list); 1874 if (lwb != NULL) { 1875 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1876 list_remove(&zilog->zl_lwb_list, lwb); 1877 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1878 kmem_cache_free(zil_lwb_cache, lwb); 1879 } 1880 mutex_exit(&zilog->zl_lock); 1881} 1882 1883static char *suspend_tag = "zil suspending"; 1884 1885/* 1886 * Suspend an intent log. While in suspended mode, we still honor 1887 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1888 * On old version pools, we suspend the log briefly when taking a 1889 * snapshot so that it will have an empty intent log. 1890 * 1891 * Long holds are not really intended to be used the way we do here -- 1892 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1893 * could fail. Therefore we take pains to only put a long hold if it is 1894 * actually necessary. Fortunately, it will only be necessary if the 1895 * objset is currently mounted (or the ZVOL equivalent). In that case it 1896 * will already have a long hold, so we are not really making things any worse. 1897 * 1898 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1899 * zvol_state_t), and use their mechanism to prevent their hold from being 1900 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1901 * very little gain. 1902 * 1903 * if cookiep == NULL, this does both the suspend & resume. 1904 * Otherwise, it returns with the dataset "long held", and the cookie 1905 * should be passed into zil_resume(). 1906 */ 1907int 1908zil_suspend(const char *osname, void **cookiep) 1909{ 1910 objset_t *os; 1911 zilog_t *zilog; 1912 const zil_header_t *zh; 1913 int error; 1914 1915 error = dmu_objset_hold(osname, suspend_tag, &os); 1916 if (error != 0) 1917 return (error); 1918 zilog = dmu_objset_zil(os); 1919 1920 mutex_enter(&zilog->zl_lock); 1921 zh = zilog->zl_header; 1922 1923 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1924 mutex_exit(&zilog->zl_lock); 1925 dmu_objset_rele(os, suspend_tag); 1926 return (SET_ERROR(EBUSY)); 1927 } 1928 1929 /* 1930 * Don't put a long hold in the cases where we can avoid it. This 1931 * is when there is no cookie so we are doing a suspend & resume 1932 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1933 * for the suspend because it's already suspended, or there's no ZIL. 1934 */ 1935 if (cookiep == NULL && !zilog->zl_suspending && 1936 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1937 mutex_exit(&zilog->zl_lock); 1938 dmu_objset_rele(os, suspend_tag); 1939 return (0); 1940 } 1941 1942 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1943 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1944 1945 zilog->zl_suspend++; 1946 1947 if (zilog->zl_suspend > 1) { 1948 /* 1949 * Someone else is already suspending it. 1950 * Just wait for them to finish. 1951 */ 1952 1953 while (zilog->zl_suspending) 1954 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1955 mutex_exit(&zilog->zl_lock); 1956 1957 if (cookiep == NULL) 1958 zil_resume(os); 1959 else 1960 *cookiep = os; 1961 return (0); 1962 } 1963 1964 /* 1965 * If there is no pointer to an on-disk block, this ZIL must not 1966 * be active (e.g. filesystem not mounted), so there's nothing 1967 * to clean up. 1968 */ 1969 if (BP_IS_HOLE(&zh->zh_log)) { 1970 ASSERT(cookiep != NULL); /* fast path already handled */ 1971 1972 *cookiep = os; 1973 mutex_exit(&zilog->zl_lock); 1974 return (0); 1975 } 1976 1977 zilog->zl_suspending = B_TRUE; 1978 mutex_exit(&zilog->zl_lock); 1979 1980 zil_commit(zilog, 0); 1981 1982 zil_destroy(zilog, B_FALSE); 1983 1984 mutex_enter(&zilog->zl_lock); 1985 zilog->zl_suspending = B_FALSE; 1986 cv_broadcast(&zilog->zl_cv_suspend); 1987 mutex_exit(&zilog->zl_lock); 1988 1989 if (cookiep == NULL) 1990 zil_resume(os); 1991 else 1992 *cookiep = os; 1993 return (0); 1994} 1995 1996void 1997zil_resume(void *cookie) 1998{ 1999 objset_t *os = cookie; 2000 zilog_t *zilog = dmu_objset_zil(os); 2001 2002 mutex_enter(&zilog->zl_lock); 2003 ASSERT(zilog->zl_suspend != 0); 2004 zilog->zl_suspend--; 2005 mutex_exit(&zilog->zl_lock); 2006 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 2007 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 2008} 2009 2010typedef struct zil_replay_arg { 2011 zil_replay_func_t **zr_replay; 2012 void *zr_arg; 2013 boolean_t zr_byteswap; 2014 char *zr_lr; 2015} zil_replay_arg_t; 2016 2017static int 2018zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 2019{ 2020 char name[ZFS_MAX_DATASET_NAME_LEN]; 2021 2022 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 2023 2024 dmu_objset_name(zilog->zl_os, name); 2025 2026 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 2027 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 2028 (u_longlong_t)lr->lrc_seq, 2029 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 2030 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 2031 2032 return (error); 2033} 2034 2035static int 2036zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 2037{ 2038 zil_replay_arg_t *zr = zra; 2039 const zil_header_t *zh = zilog->zl_header; 2040 uint64_t reclen = lr->lrc_reclen; 2041 uint64_t txtype = lr->lrc_txtype; 2042 int error = 0; 2043 2044 zilog->zl_replaying_seq = lr->lrc_seq; 2045 2046 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2047 return (0); 2048 2049 if (lr->lrc_txg < claim_txg) /* already committed */ 2050 return (0); 2051 2052 /* Strip case-insensitive bit, still present in log record */ 2053 txtype &= ~TX_CI; 2054 2055 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2056 return (zil_replay_error(zilog, lr, EINVAL)); 2057 2058 /* 2059 * If this record type can be logged out of order, the object 2060 * (lr_foid) may no longer exist. That's legitimate, not an error. 2061 */ 2062 if (TX_OOO(txtype)) { 2063 error = dmu_object_info(zilog->zl_os, 2064 ((lr_ooo_t *)lr)->lr_foid, NULL); 2065 if (error == ENOENT || error == EEXIST) 2066 return (0); 2067 } 2068 2069 /* 2070 * Make a copy of the data so we can revise and extend it. 2071 */ 2072 bcopy(lr, zr->zr_lr, reclen); 2073 2074 /* 2075 * If this is a TX_WRITE with a blkptr, suck in the data. 2076 */ 2077 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2078 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2079 zr->zr_lr + reclen); 2080 if (error != 0) 2081 return (zil_replay_error(zilog, lr, error)); 2082 } 2083 2084 /* 2085 * The log block containing this lr may have been byteswapped 2086 * so that we can easily examine common fields like lrc_txtype. 2087 * However, the log is a mix of different record types, and only the 2088 * replay vectors know how to byteswap their records. Therefore, if 2089 * the lr was byteswapped, undo it before invoking the replay vector. 2090 */ 2091 if (zr->zr_byteswap) 2092 byteswap_uint64_array(zr->zr_lr, reclen); 2093 2094 /* 2095 * We must now do two things atomically: replay this log record, 2096 * and update the log header sequence number to reflect the fact that 2097 * we did so. At the end of each replay function the sequence number 2098 * is updated if we are in replay mode. 2099 */ 2100 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2101 if (error != 0) { 2102 /* 2103 * The DMU's dnode layer doesn't see removes until the txg 2104 * commits, so a subsequent claim can spuriously fail with 2105 * EEXIST. So if we receive any error we try syncing out 2106 * any removes then retry the transaction. Note that we 2107 * specify B_FALSE for byteswap now, so we don't do it twice. 2108 */ 2109 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2110 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2111 if (error != 0) 2112 return (zil_replay_error(zilog, lr, error)); 2113 } 2114 return (0); 2115} 2116 2117/* ARGSUSED */ 2118static int 2119zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2120{ 2121 zilog->zl_replay_blks++; 2122 2123 return (0); 2124} 2125 2126/* 2127 * If this dataset has a non-empty intent log, replay it and destroy it. 2128 */ 2129void 2130zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2131{ 2132 zilog_t *zilog = dmu_objset_zil(os); 2133 const zil_header_t *zh = zilog->zl_header; 2134 zil_replay_arg_t zr; 2135 2136 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2137 zil_destroy(zilog, B_TRUE); 2138 return; 2139 } 2140 2141 zr.zr_replay = replay_func; 2142 zr.zr_arg = arg; 2143 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2144 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2145 2146 /* 2147 * Wait for in-progress removes to sync before starting replay. 2148 */ 2149 txg_wait_synced(zilog->zl_dmu_pool, 0); 2150 2151 zilog->zl_replay = B_TRUE; 2152 zilog->zl_replay_time = ddi_get_lbolt(); 2153 ASSERT(zilog->zl_replay_blks == 0); 2154 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2155 zh->zh_claim_txg); 2156 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2157 2158 zil_destroy(zilog, B_FALSE); 2159 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2160 zilog->zl_replay = B_FALSE; 2161} 2162 2163boolean_t 2164zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2165{ 2166 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2167 return (B_TRUE); 2168 2169 if (zilog->zl_replay) { 2170 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2171 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2172 zilog->zl_replaying_seq; 2173 return (B_TRUE); 2174 } 2175 2176 return (B_FALSE); 2177} 2178 2179/* ARGSUSED */ 2180int 2181zil_vdev_offline(const char *osname, void *arg) 2182{ 2183 int error; 2184 2185 error = zil_suspend(osname, NULL); 2186 if (error != 0) 2187 return (SET_ERROR(EEXIST)); 2188 return (0); 2189} 2190