zil.c revision 209962
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 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26#include <sys/zfs_context.h> 27#include <sys/spa.h> 28#include <sys/dmu.h> 29#include <sys/zap.h> 30#include <sys/arc.h> 31#include <sys/stat.h> 32#include <sys/resource.h> 33#include <sys/zil.h> 34#include <sys/zil_impl.h> 35#include <sys/dsl_dataset.h> 36#include <sys/vdev.h> 37#include <sys/dmu_tx.h> 38 39/* 40 * The zfs intent log (ZIL) saves transaction records of system calls 41 * that change the file system in memory with enough information 42 * to be able to replay them. These are stored in memory until 43 * either the DMU transaction group (txg) commits them to the stable pool 44 * and they can be discarded, or they are flushed to the stable log 45 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 46 * requirement. In the event of a panic or power fail then those log 47 * records (transactions) are replayed. 48 * 49 * There is one ZIL per file system. Its on-disk (pool) format consists 50 * of 3 parts: 51 * 52 * - ZIL header 53 * - ZIL blocks 54 * - ZIL records 55 * 56 * A log record holds a system call transaction. Log blocks can 57 * hold many log records and the blocks are chained together. 58 * Each ZIL block contains a block pointer (blkptr_t) to the next 59 * ZIL block in the chain. The ZIL header points to the first 60 * block in the chain. Note there is not a fixed place in the pool 61 * to hold blocks. They are dynamically allocated and freed as 62 * needed from the blocks available. Figure X shows the ZIL structure: 63 */ 64 65/* 66 * This global ZIL switch affects all pools 67 */ 68int zil_disable = 0; /* disable intent logging */ 69SYSCTL_DECL(_vfs_zfs); 70TUNABLE_INT("vfs.zfs.zil_disable", &zil_disable); 71SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_disable, CTLFLAG_RW, &zil_disable, 0, 72 "Disable ZFS Intent Log (ZIL)"); 73 74/* 75 * Tunable parameter for debugging or performance analysis. Setting 76 * zfs_nocacheflush will cause corruption on power loss if a volatile 77 * out-of-order write cache is enabled. 78 */ 79boolean_t zfs_nocacheflush = B_FALSE; 80TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 81SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 82 &zfs_nocacheflush, 0, "Disable cache flush"); 83 84static kmem_cache_t *zil_lwb_cache; 85 86static int 87zil_dva_compare(const void *x1, const void *x2) 88{ 89 const dva_t *dva1 = x1; 90 const dva_t *dva2 = x2; 91 92 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 93 return (-1); 94 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 95 return (1); 96 97 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 98 return (-1); 99 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 100 return (1); 101 102 return (0); 103} 104 105static void 106zil_dva_tree_init(avl_tree_t *t) 107{ 108 avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t), 109 offsetof(zil_dva_node_t, zn_node)); 110} 111 112static void 113zil_dva_tree_fini(avl_tree_t *t) 114{ 115 zil_dva_node_t *zn; 116 void *cookie = NULL; 117 118 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 119 kmem_free(zn, sizeof (zil_dva_node_t)); 120 121 avl_destroy(t); 122} 123 124static int 125zil_dva_tree_add(avl_tree_t *t, dva_t *dva) 126{ 127 zil_dva_node_t *zn; 128 avl_index_t where; 129 130 if (avl_find(t, dva, &where) != NULL) 131 return (EEXIST); 132 133 zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP); 134 zn->zn_dva = *dva; 135 avl_insert(t, zn, where); 136 137 return (0); 138} 139 140static zil_header_t * 141zil_header_in_syncing_context(zilog_t *zilog) 142{ 143 return ((zil_header_t *)zilog->zl_header); 144} 145 146static void 147zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 148{ 149 zio_cksum_t *zc = &bp->blk_cksum; 150 151 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 152 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 153 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 154 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 155} 156 157/* 158 * Read a log block, make sure it's valid, and byteswap it if necessary. 159 */ 160static int 161zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp) 162{ 163 blkptr_t blk = *bp; 164 zbookmark_t zb; 165 uint32_t aflags = ARC_WAIT; 166 int error; 167 168 zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET]; 169 zb.zb_object = 0; 170 zb.zb_level = -1; 171 zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ]; 172 173 *abufpp = NULL; 174 175 /* 176 * We shouldn't be doing any scrubbing while we're doing log 177 * replay, it's OK to not lock. 178 */ 179 error = arc_read_nolock(NULL, zilog->zl_spa, &blk, 180 arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL | 181 ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb); 182 183 if (error == 0) { 184 char *data = (*abufpp)->b_data; 185 uint64_t blksz = BP_GET_LSIZE(bp); 186 zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1; 187 zio_cksum_t cksum = bp->blk_cksum; 188 189 /* 190 * Validate the checksummed log block. 191 * 192 * Sequence numbers should be... sequential. The checksum 193 * verifier for the next block should be bp's checksum plus 1. 194 * 195 * Also check the log chain linkage and size used. 196 */ 197 cksum.zc_word[ZIL_ZC_SEQ]++; 198 199 if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, 200 sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) || 201 (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))) { 202 error = ECKSUM; 203 } 204 205 if (error) { 206 VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1); 207 *abufpp = NULL; 208 } 209 } 210 211 dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid); 212 213 return (error); 214} 215 216/* 217 * Parse the intent log, and call parse_func for each valid record within. 218 * Return the highest sequence number. 219 */ 220uint64_t 221zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 222 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 223{ 224 const zil_header_t *zh = zilog->zl_header; 225 uint64_t claim_seq = zh->zh_claim_seq; 226 uint64_t seq = 0; 227 uint64_t max_seq = 0; 228 blkptr_t blk = zh->zh_log; 229 arc_buf_t *abuf; 230 char *lrbuf, *lrp; 231 zil_trailer_t *ztp; 232 int reclen, error; 233 234 if (BP_IS_HOLE(&blk)) 235 return (max_seq); 236 237 /* 238 * Starting at the block pointed to by zh_log we read the log chain. 239 * For each block in the chain we strongly check that block to 240 * ensure its validity. We stop when an invalid block is found. 241 * For each block pointer in the chain we call parse_blk_func(). 242 * For each record in each valid block we call parse_lr_func(). 243 * If the log has been claimed, stop if we encounter a sequence 244 * number greater than the highest claimed sequence number. 245 */ 246 zil_dva_tree_init(&zilog->zl_dva_tree); 247 for (;;) { 248 seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 249 250 if (claim_seq != 0 && seq > claim_seq) 251 break; 252 253 ASSERT(max_seq < seq); 254 max_seq = seq; 255 256 error = zil_read_log_block(zilog, &blk, &abuf); 257 258 if (parse_blk_func != NULL) 259 parse_blk_func(zilog, &blk, arg, txg); 260 261 if (error) 262 break; 263 264 lrbuf = abuf->b_data; 265 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 266 blk = ztp->zit_next_blk; 267 268 if (parse_lr_func == NULL) { 269 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 270 continue; 271 } 272 273 for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { 274 lr_t *lr = (lr_t *)lrp; 275 reclen = lr->lrc_reclen; 276 ASSERT3U(reclen, >=, sizeof (lr_t)); 277 parse_lr_func(zilog, lr, arg, txg); 278 } 279 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 280 } 281 zil_dva_tree_fini(&zilog->zl_dva_tree); 282 283 return (max_seq); 284} 285 286/* ARGSUSED */ 287static void 288zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 289{ 290 spa_t *spa = zilog->zl_spa; 291 int err; 292 293 /* 294 * Claim log block if not already committed and not already claimed. 295 */ 296 if (bp->blk_birth >= first_txg && 297 zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) { 298 err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL, 299 ZIO_FLAG_MUSTSUCCEED)); 300 ASSERT(err == 0); 301 } 302} 303 304static void 305zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 306{ 307 if (lrc->lrc_txtype == TX_WRITE) { 308 lr_write_t *lr = (lr_write_t *)lrc; 309 zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg); 310 } 311} 312 313/* ARGSUSED */ 314static void 315zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 316{ 317 zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx)); 318} 319 320static void 321zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 322{ 323 /* 324 * If we previously claimed it, we need to free it. 325 */ 326 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) { 327 lr_write_t *lr = (lr_write_t *)lrc; 328 blkptr_t *bp = &lr->lr_blkptr; 329 if (bp->blk_birth >= claim_txg && 330 !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) { 331 (void) arc_free(NULL, zilog->zl_spa, 332 dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT); 333 } 334 } 335} 336 337/* 338 * Create an on-disk intent log. 339 */ 340static void 341zil_create(zilog_t *zilog) 342{ 343 const zil_header_t *zh = zilog->zl_header; 344 lwb_t *lwb; 345 uint64_t txg = 0; 346 dmu_tx_t *tx = NULL; 347 blkptr_t blk; 348 int error = 0; 349 350 /* 351 * Wait for any previous destroy to complete. 352 */ 353 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 354 355 ASSERT(zh->zh_claim_txg == 0); 356 ASSERT(zh->zh_replay_seq == 0); 357 358 blk = zh->zh_log; 359 360 /* 361 * If we don't already have an initial log block or we have one 362 * but it's the wrong endianness then allocate one. 363 */ 364 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 365 tx = dmu_tx_create(zilog->zl_os); 366 (void) dmu_tx_assign(tx, TXG_WAIT); 367 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 368 txg = dmu_tx_get_txg(tx); 369 370 if (!BP_IS_HOLE(&blk)) { 371 zio_free_blk(zilog->zl_spa, &blk, txg); 372 BP_ZERO(&blk); 373 } 374 375 error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk, 376 NULL, txg); 377 378 if (error == 0) 379 zil_init_log_chain(zilog, &blk); 380 } 381 382 /* 383 * Allocate a log write buffer (lwb) for the first log block. 384 */ 385 if (error == 0) { 386 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 387 lwb->lwb_zilog = zilog; 388 lwb->lwb_blk = blk; 389 lwb->lwb_nused = 0; 390 lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); 391 lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); 392 lwb->lwb_max_txg = txg; 393 lwb->lwb_zio = NULL; 394 395 mutex_enter(&zilog->zl_lock); 396 list_insert_tail(&zilog->zl_lwb_list, lwb); 397 mutex_exit(&zilog->zl_lock); 398 } 399 400 /* 401 * If we just allocated the first log block, commit our transaction 402 * and wait for zil_sync() to stuff the block poiner into zh_log. 403 * (zh is part of the MOS, so we cannot modify it in open context.) 404 */ 405 if (tx != NULL) { 406 dmu_tx_commit(tx); 407 txg_wait_synced(zilog->zl_dmu_pool, txg); 408 } 409 410 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 411} 412 413/* 414 * In one tx, free all log blocks and clear the log header. 415 * If keep_first is set, then we're replaying a log with no content. 416 * We want to keep the first block, however, so that the first 417 * synchronous transaction doesn't require a txg_wait_synced() 418 * in zil_create(). We don't need to txg_wait_synced() here either 419 * when keep_first is set, because both zil_create() and zil_destroy() 420 * will wait for any in-progress destroys to complete. 421 */ 422void 423zil_destroy(zilog_t *zilog, boolean_t keep_first) 424{ 425 const zil_header_t *zh = zilog->zl_header; 426 lwb_t *lwb; 427 dmu_tx_t *tx; 428 uint64_t txg; 429 430 /* 431 * Wait for any previous destroy to complete. 432 */ 433 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 434 435 if (BP_IS_HOLE(&zh->zh_log)) 436 return; 437 438 tx = dmu_tx_create(zilog->zl_os); 439 (void) dmu_tx_assign(tx, TXG_WAIT); 440 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 441 txg = dmu_tx_get_txg(tx); 442 443 mutex_enter(&zilog->zl_lock); 444 445 /* 446 * It is possible for the ZIL to get the previously mounted zilog 447 * structure of the same dataset if quickly remounted and the dbuf 448 * eviction has not completed. In this case we can see a non 449 * empty lwb list and keep_first will be set. We fix this by 450 * clearing the keep_first. This will be slower but it's very rare. 451 */ 452 if (!list_is_empty(&zilog->zl_lwb_list) && keep_first) 453 keep_first = B_FALSE; 454 455 ASSERT3U(zilog->zl_destroy_txg, <, txg); 456 zilog->zl_destroy_txg = txg; 457 zilog->zl_keep_first = keep_first; 458 459 if (!list_is_empty(&zilog->zl_lwb_list)) { 460 ASSERT(zh->zh_claim_txg == 0); 461 ASSERT(!keep_first); 462 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 463 list_remove(&zilog->zl_lwb_list, lwb); 464 if (lwb->lwb_buf != NULL) 465 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 466 zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg); 467 kmem_cache_free(zil_lwb_cache, lwb); 468 } 469 } else { 470 if (!keep_first) { 471 (void) zil_parse(zilog, zil_free_log_block, 472 zil_free_log_record, tx, zh->zh_claim_txg); 473 } 474 } 475 mutex_exit(&zilog->zl_lock); 476 477 dmu_tx_commit(tx); 478} 479 480/* 481 * return true if the initial log block is not valid 482 */ 483static boolean_t 484zil_empty(zilog_t *zilog) 485{ 486 const zil_header_t *zh = zilog->zl_header; 487 arc_buf_t *abuf = NULL; 488 489 if (BP_IS_HOLE(&zh->zh_log)) 490 return (B_TRUE); 491 492 if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0) 493 return (B_TRUE); 494 495 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 496 return (B_FALSE); 497} 498 499int 500zil_claim(char *osname, void *txarg) 501{ 502 dmu_tx_t *tx = txarg; 503 uint64_t first_txg = dmu_tx_get_txg(tx); 504 zilog_t *zilog; 505 zil_header_t *zh; 506 objset_t *os; 507 int error; 508 509 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 510 if (error) { 511 cmn_err(CE_WARN, "can't open objset for %s", osname); 512 return (0); 513 } 514 515 zilog = dmu_objset_zil(os); 516 zh = zil_header_in_syncing_context(zilog); 517 518 /* 519 * Record here whether the zil has any records to replay. 520 * If the header block pointer is null or the block points 521 * to the stubby then we know there are no valid log records. 522 * We use the header to store this state as the the zilog gets 523 * freed later in dmu_objset_close(). 524 * The flags (and the rest of the header fields) are cleared in 525 * zil_sync() as a result of a zil_destroy(), after replaying the log. 526 * 527 * Note, the intent log can be empty but still need the 528 * stubby to be claimed. 529 */ 530 if (!zil_empty(zilog)) 531 zh->zh_flags |= ZIL_REPLAY_NEEDED; 532 533 /* 534 * Claim all log blocks if we haven't already done so, and remember 535 * the highest claimed sequence number. This ensures that if we can 536 * read only part of the log now (e.g. due to a missing device), 537 * but we can read the entire log later, we will not try to replay 538 * or destroy beyond the last block we successfully claimed. 539 */ 540 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 541 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 542 zh->zh_claim_txg = first_txg; 543 zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block, 544 zil_claim_log_record, tx, first_txg); 545 dsl_dataset_dirty(dmu_objset_ds(os), tx); 546 } 547 548 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 549 dmu_objset_close(os); 550 return (0); 551} 552 553/* 554 * Check the log by walking the log chain. 555 * Checksum errors are ok as they indicate the end of the chain. 556 * Any other error (no device or read failure) returns an error. 557 */ 558/* ARGSUSED */ 559int 560zil_check_log_chain(char *osname, void *txarg) 561{ 562 zilog_t *zilog; 563 zil_header_t *zh; 564 blkptr_t blk; 565 arc_buf_t *abuf; 566 objset_t *os; 567 char *lrbuf; 568 zil_trailer_t *ztp; 569 int error; 570 571 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 572 if (error) { 573 cmn_err(CE_WARN, "can't open objset for %s", osname); 574 return (0); 575 } 576 577 zilog = dmu_objset_zil(os); 578 zh = zil_header_in_syncing_context(zilog); 579 blk = zh->zh_log; 580 if (BP_IS_HOLE(&blk)) { 581 dmu_objset_close(os); 582 return (0); /* no chain */ 583 } 584 585 for (;;) { 586 error = zil_read_log_block(zilog, &blk, &abuf); 587 if (error) 588 break; 589 lrbuf = abuf->b_data; 590 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 591 blk = ztp->zit_next_blk; 592 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 593 } 594 dmu_objset_close(os); 595 if (error == ECKSUM) 596 return (0); /* normal end of chain */ 597 return (error); 598} 599 600/* 601 * Clear a log chain 602 */ 603/* ARGSUSED */ 604int 605zil_clear_log_chain(char *osname, void *txarg) 606{ 607 zilog_t *zilog; 608 zil_header_t *zh; 609 objset_t *os; 610 dmu_tx_t *tx; 611 int error; 612 613 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 614 if (error) { 615 cmn_err(CE_WARN, "can't open objset for %s", osname); 616 return (0); 617 } 618 619 zilog = dmu_objset_zil(os); 620 tx = dmu_tx_create(zilog->zl_os); 621 (void) dmu_tx_assign(tx, TXG_WAIT); 622 zh = zil_header_in_syncing_context(zilog); 623 BP_ZERO(&zh->zh_log); 624 dsl_dataset_dirty(dmu_objset_ds(os), tx); 625 dmu_tx_commit(tx); 626 dmu_objset_close(os); 627 return (0); 628} 629 630static int 631zil_vdev_compare(const void *x1, const void *x2) 632{ 633 uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 634 uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 635 636 if (v1 < v2) 637 return (-1); 638 if (v1 > v2) 639 return (1); 640 641 return (0); 642} 643 644void 645zil_add_block(zilog_t *zilog, blkptr_t *bp) 646{ 647 avl_tree_t *t = &zilog->zl_vdev_tree; 648 avl_index_t where; 649 zil_vdev_node_t *zv, zvsearch; 650 int ndvas = BP_GET_NDVAS(bp); 651 int i; 652 653 if (zfs_nocacheflush) 654 return; 655 656 ASSERT(zilog->zl_writer); 657 658 /* 659 * Even though we're zl_writer, we still need a lock because the 660 * zl_get_data() callbacks may have dmu_sync() done callbacks 661 * that will run concurrently. 662 */ 663 mutex_enter(&zilog->zl_vdev_lock); 664 for (i = 0; i < ndvas; i++) { 665 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 666 if (avl_find(t, &zvsearch, &where) == NULL) { 667 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 668 zv->zv_vdev = zvsearch.zv_vdev; 669 avl_insert(t, zv, where); 670 } 671 } 672 mutex_exit(&zilog->zl_vdev_lock); 673} 674 675void 676zil_flush_vdevs(zilog_t *zilog) 677{ 678 spa_t *spa = zilog->zl_spa; 679 avl_tree_t *t = &zilog->zl_vdev_tree; 680 void *cookie = NULL; 681 zil_vdev_node_t *zv; 682 zio_t *zio; 683 684 ASSERT(zilog->zl_writer); 685 686 /* 687 * We don't need zl_vdev_lock here because we're the zl_writer, 688 * and all zl_get_data() callbacks are done. 689 */ 690 if (avl_numnodes(t) == 0) 691 return; 692 693 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 694 695 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 696 697 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 698 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 699 if (vd != NULL) 700 zio_flush(zio, vd); 701 kmem_free(zv, sizeof (*zv)); 702 } 703 704 /* 705 * Wait for all the flushes to complete. Not all devices actually 706 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 707 */ 708 (void) zio_wait(zio); 709 710 spa_config_exit(spa, SCL_STATE, FTAG); 711} 712 713/* 714 * Function called when a log block write completes 715 */ 716static void 717zil_lwb_write_done(zio_t *zio) 718{ 719 lwb_t *lwb = zio->io_private; 720 zilog_t *zilog = lwb->lwb_zilog; 721 722 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 723 ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG); 724 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 725 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 726 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 727 ASSERT(!BP_IS_GANG(zio->io_bp)); 728 ASSERT(!BP_IS_HOLE(zio->io_bp)); 729 ASSERT(zio->io_bp->blk_fill == 0); 730 731 /* 732 * Ensure the lwb buffer pointer is cleared before releasing 733 * the txg. If we have had an allocation failure and 734 * the txg is waiting to sync then we want want zil_sync() 735 * to remove the lwb so that it's not picked up as the next new 736 * one in zil_commit_writer(). zil_sync() will only remove 737 * the lwb if lwb_buf is null. 738 */ 739 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 740 mutex_enter(&zilog->zl_lock); 741 lwb->lwb_buf = NULL; 742 if (zio->io_error) 743 zilog->zl_log_error = B_TRUE; 744 745 /* 746 * Now that we've written this log block, we have a stable pointer 747 * to the next block in the chain, so it's OK to let the txg in 748 * which we allocated the next block sync. We still have the 749 * zl_lock to ensure zil_sync doesn't kmem free the lwb. 750 */ 751 txg_rele_to_sync(&lwb->lwb_txgh); 752 mutex_exit(&zilog->zl_lock); 753} 754 755/* 756 * Initialize the io for a log block. 757 */ 758static void 759zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 760{ 761 zbookmark_t zb; 762 763 zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET]; 764 zb.zb_object = 0; 765 zb.zb_level = -1; 766 zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 767 768 if (zilog->zl_root_zio == NULL) { 769 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 770 ZIO_FLAG_CANFAIL); 771 } 772 if (lwb->lwb_zio == NULL) { 773 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 774 0, &lwb->lwb_blk, lwb->lwb_buf, 775 lwb->lwb_sz, zil_lwb_write_done, lwb, 776 ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_CANFAIL, &zb); 777 } 778} 779 780/* 781 * Start a log block write and advance to the next log block. 782 * Calls are serialized. 783 */ 784static lwb_t * 785zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 786{ 787 lwb_t *nlwb; 788 zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; 789 spa_t *spa = zilog->zl_spa; 790 blkptr_t *bp = &ztp->zit_next_blk; 791 uint64_t txg; 792 uint64_t zil_blksz; 793 int error; 794 795 ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); 796 797 /* 798 * Allocate the next block and save its address in this block 799 * before writing it in order to establish the log chain. 800 * Note that if the allocation of nlwb synced before we wrote 801 * the block that points at it (lwb), we'd leak it if we crashed. 802 * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done(). 803 */ 804 txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh); 805 txg_rele_to_quiesce(&lwb->lwb_txgh); 806 807 /* 808 * Pick a ZIL blocksize. We request a size that is the 809 * maximum of the previous used size, the current used size and 810 * the amount waiting in the queue. 811 */ 812 zil_blksz = MAX(zilog->zl_prev_used, 813 zilog->zl_cur_used + sizeof (*ztp)); 814 zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); 815 zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); 816 if (zil_blksz > ZIL_MAX_BLKSZ) 817 zil_blksz = ZIL_MAX_BLKSZ; 818 819 BP_ZERO(bp); 820 /* pass the old blkptr in order to spread log blocks across devs */ 821 error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg); 822 if (error) { 823 dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg); 824 825 /* 826 * We dirty the dataset to ensure that zil_sync() will 827 * be called to remove this lwb from our zl_lwb_list. 828 * Failing to do so, may leave an lwb with a NULL lwb_buf 829 * hanging around on the zl_lwb_list. 830 */ 831 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 832 dmu_tx_commit(tx); 833 834 /* 835 * Since we've just experienced an allocation failure so we 836 * terminate the current lwb and send it on its way. 837 */ 838 ztp->zit_pad = 0; 839 ztp->zit_nused = lwb->lwb_nused; 840 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 841 zio_nowait(lwb->lwb_zio); 842 843 /* 844 * By returning NULL the caller will call tx_wait_synced() 845 */ 846 return (NULL); 847 } 848 849 ASSERT3U(bp->blk_birth, ==, txg); 850 ztp->zit_pad = 0; 851 ztp->zit_nused = lwb->lwb_nused; 852 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 853 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 854 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 855 856 /* 857 * Allocate a new log write buffer (lwb). 858 */ 859 nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 860 861 nlwb->lwb_zilog = zilog; 862 nlwb->lwb_blk = *bp; 863 nlwb->lwb_nused = 0; 864 nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); 865 nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); 866 nlwb->lwb_max_txg = txg; 867 nlwb->lwb_zio = NULL; 868 869 /* 870 * Put new lwb at the end of the log chain 871 */ 872 mutex_enter(&zilog->zl_lock); 873 list_insert_tail(&zilog->zl_lwb_list, nlwb); 874 mutex_exit(&zilog->zl_lock); 875 876 /* Record the block for later vdev flushing */ 877 zil_add_block(zilog, &lwb->lwb_blk); 878 879 /* 880 * kick off the write for the old log block 881 */ 882 dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg); 883 ASSERT(lwb->lwb_zio); 884 zio_nowait(lwb->lwb_zio); 885 886 return (nlwb); 887} 888 889static lwb_t * 890zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 891{ 892 lr_t *lrc = &itx->itx_lr; /* common log record */ 893 lr_write_t *lr = (lr_write_t *)lrc; 894 uint64_t txg = lrc->lrc_txg; 895 uint64_t reclen = lrc->lrc_reclen; 896 uint64_t dlen; 897 898 if (lwb == NULL) 899 return (NULL); 900 ASSERT(lwb->lwb_buf != NULL); 901 902 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 903 dlen = P2ROUNDUP_TYPED( 904 lr->lr_length, sizeof (uint64_t), uint64_t); 905 else 906 dlen = 0; 907 908 zilog->zl_cur_used += (reclen + dlen); 909 910 zil_lwb_write_init(zilog, lwb); 911 912 /* 913 * If this record won't fit in the current log block, start a new one. 914 */ 915 if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 916 lwb = zil_lwb_write_start(zilog, lwb); 917 if (lwb == NULL) 918 return (NULL); 919 zil_lwb_write_init(zilog, lwb); 920 ASSERT(lwb->lwb_nused == 0); 921 if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 922 txg_wait_synced(zilog->zl_dmu_pool, txg); 923 return (lwb); 924 } 925 } 926 927 /* 928 * Update the lrc_seq, to be log record sequence number. See zil.h 929 * Then copy the record to the log buffer. 930 */ 931 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 932 bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen); 933 934 /* 935 * If it's a write, fetch the data or get its blkptr as appropriate. 936 */ 937 if (lrc->lrc_txtype == TX_WRITE) { 938 if (txg > spa_freeze_txg(zilog->zl_spa)) 939 txg_wait_synced(zilog->zl_dmu_pool, txg); 940 if (itx->itx_wr_state != WR_COPIED) { 941 char *dbuf; 942 int error; 943 944 /* alignment is guaranteed */ 945 lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused); 946 if (dlen) { 947 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 948 dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen; 949 lr->lr_common.lrc_reclen += dlen; 950 } else { 951 ASSERT(itx->itx_wr_state == WR_INDIRECT); 952 dbuf = NULL; 953 } 954 error = zilog->zl_get_data( 955 itx->itx_private, lr, dbuf, lwb->lwb_zio); 956 if (error) { 957 ASSERT(error == ENOENT || error == EEXIST || 958 error == EALREADY); 959 return (lwb); 960 } 961 } 962 } 963 964 lwb->lwb_nused += reclen + dlen; 965 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 966 ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); 967 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 968 969 return (lwb); 970} 971 972itx_t * 973zil_itx_create(uint64_t txtype, size_t lrsize) 974{ 975 itx_t *itx; 976 977 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 978 979 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 980 itx->itx_lr.lrc_txtype = txtype; 981 itx->itx_lr.lrc_reclen = lrsize; 982 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 983 itx->itx_lr.lrc_seq = 0; /* defensive */ 984 985 return (itx); 986} 987 988uint64_t 989zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 990{ 991 uint64_t seq; 992 993 ASSERT(itx->itx_lr.lrc_seq == 0); 994 995 mutex_enter(&zilog->zl_lock); 996 list_insert_tail(&zilog->zl_itx_list, itx); 997 zilog->zl_itx_list_sz += itx->itx_sod; 998 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 999 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 1000 mutex_exit(&zilog->zl_lock); 1001 1002 return (seq); 1003} 1004 1005/* 1006 * Free up all in-memory intent log transactions that have now been synced. 1007 */ 1008static void 1009zil_itx_clean(zilog_t *zilog) 1010{ 1011 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 1012 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 1013 list_t clean_list; 1014 itx_t *itx; 1015 1016 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1017 1018 mutex_enter(&zilog->zl_lock); 1019 /* wait for a log writer to finish walking list */ 1020 while (zilog->zl_writer) { 1021 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1022 } 1023 1024 /* 1025 * Move the sync'd log transactions to a separate list so we can call 1026 * kmem_free without holding the zl_lock. 1027 * 1028 * There is no need to set zl_writer as we don't drop zl_lock here 1029 */ 1030 while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 1031 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 1032 list_remove(&zilog->zl_itx_list, itx); 1033 zilog->zl_itx_list_sz -= itx->itx_sod; 1034 list_insert_tail(&clean_list, itx); 1035 } 1036 cv_broadcast(&zilog->zl_cv_writer); 1037 mutex_exit(&zilog->zl_lock); 1038 1039 /* destroy sync'd log transactions */ 1040 while ((itx = list_head(&clean_list)) != NULL) { 1041 list_remove(&clean_list, itx); 1042 kmem_free(itx, offsetof(itx_t, itx_lr) 1043 + itx->itx_lr.lrc_reclen); 1044 } 1045 list_destroy(&clean_list); 1046} 1047 1048/* 1049 * If there are any in-memory intent log transactions which have now been 1050 * synced then start up a taskq to free them. 1051 */ 1052void 1053zil_clean(zilog_t *zilog) 1054{ 1055 itx_t *itx; 1056 1057 mutex_enter(&zilog->zl_lock); 1058 itx = list_head(&zilog->zl_itx_list); 1059 if ((itx != NULL) && 1060 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 1061 (void) taskq_dispatch(zilog->zl_clean_taskq, 1062 (task_func_t *)zil_itx_clean, zilog, TQ_SLEEP); 1063 } 1064 mutex_exit(&zilog->zl_lock); 1065} 1066 1067static void 1068zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 1069{ 1070 uint64_t txg; 1071 uint64_t commit_seq = 0; 1072 itx_t *itx, *itx_next = (itx_t *)-1; 1073 lwb_t *lwb; 1074 spa_t *spa; 1075 1076 zilog->zl_writer = B_TRUE; 1077 ASSERT(zilog->zl_root_zio == NULL); 1078 spa = zilog->zl_spa; 1079 1080 if (zilog->zl_suspend) { 1081 lwb = NULL; 1082 } else { 1083 lwb = list_tail(&zilog->zl_lwb_list); 1084 if (lwb == NULL) { 1085 /* 1086 * Return if there's nothing to flush before we 1087 * dirty the fs by calling zil_create() 1088 */ 1089 if (list_is_empty(&zilog->zl_itx_list)) { 1090 zilog->zl_writer = B_FALSE; 1091 return; 1092 } 1093 mutex_exit(&zilog->zl_lock); 1094 zil_create(zilog); 1095 mutex_enter(&zilog->zl_lock); 1096 lwb = list_tail(&zilog->zl_lwb_list); 1097 } 1098 } 1099 1100 /* Loop through in-memory log transactions filling log blocks. */ 1101 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1102 for (;;) { 1103 /* 1104 * Find the next itx to push: 1105 * Push all transactions related to specified foid and all 1106 * other transactions except TX_WRITE, TX_TRUNCATE, 1107 * TX_SETATTR and TX_ACL for all other files. 1108 */ 1109 if (itx_next != (itx_t *)-1) 1110 itx = itx_next; 1111 else 1112 itx = list_head(&zilog->zl_itx_list); 1113 for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) { 1114 if (foid == 0) /* push all foids? */ 1115 break; 1116 if (itx->itx_sync) /* push all O_[D]SYNC */ 1117 break; 1118 switch (itx->itx_lr.lrc_txtype) { 1119 case TX_SETATTR: 1120 case TX_WRITE: 1121 case TX_TRUNCATE: 1122 case TX_ACL: 1123 /* lr_foid is same offset for these records */ 1124 if (((lr_write_t *)&itx->itx_lr)->lr_foid 1125 != foid) { 1126 continue; /* skip this record */ 1127 } 1128 } 1129 break; 1130 } 1131 if (itx == NULL) 1132 break; 1133 1134 if ((itx->itx_lr.lrc_seq > seq) && 1135 ((lwb == NULL) || (lwb->lwb_nused == 0) || 1136 (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) { 1137 break; 1138 } 1139 1140 /* 1141 * Save the next pointer. Even though we soon drop 1142 * zl_lock all threads that may change the list 1143 * (another writer or zil_itx_clean) can't do so until 1144 * they have zl_writer. 1145 */ 1146 itx_next = list_next(&zilog->zl_itx_list, itx); 1147 list_remove(&zilog->zl_itx_list, itx); 1148 zilog->zl_itx_list_sz -= itx->itx_sod; 1149 mutex_exit(&zilog->zl_lock); 1150 txg = itx->itx_lr.lrc_txg; 1151 ASSERT(txg); 1152 1153 if (txg > spa_last_synced_txg(spa) || 1154 txg > spa_freeze_txg(spa)) 1155 lwb = zil_lwb_commit(zilog, itx, lwb); 1156 kmem_free(itx, offsetof(itx_t, itx_lr) 1157 + itx->itx_lr.lrc_reclen); 1158 mutex_enter(&zilog->zl_lock); 1159 } 1160 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1161 /* determine commit sequence number */ 1162 itx = list_head(&zilog->zl_itx_list); 1163 if (itx) 1164 commit_seq = itx->itx_lr.lrc_seq; 1165 else 1166 commit_seq = zilog->zl_itx_seq; 1167 mutex_exit(&zilog->zl_lock); 1168 1169 /* write the last block out */ 1170 if (lwb != NULL && lwb->lwb_zio != NULL) 1171 lwb = zil_lwb_write_start(zilog, lwb); 1172 1173 zilog->zl_prev_used = zilog->zl_cur_used; 1174 zilog->zl_cur_used = 0; 1175 1176 /* 1177 * Wait if necessary for the log blocks to be on stable storage. 1178 */ 1179 if (zilog->zl_root_zio) { 1180 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1181 (void) zio_wait(zilog->zl_root_zio); 1182 zilog->zl_root_zio = NULL; 1183 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 1184 zil_flush_vdevs(zilog); 1185 } 1186 1187 if (zilog->zl_log_error || lwb == NULL) { 1188 zilog->zl_log_error = 0; 1189 txg_wait_synced(zilog->zl_dmu_pool, 0); 1190 } 1191 1192 mutex_enter(&zilog->zl_lock); 1193 zilog->zl_writer = B_FALSE; 1194 1195 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 1196 zilog->zl_commit_seq = commit_seq; 1197} 1198 1199/* 1200 * Push zfs transactions to stable storage up to the supplied sequence number. 1201 * If foid is 0 push out all transactions, otherwise push only those 1202 * for that file or might have been used to create that file. 1203 */ 1204void 1205zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1206{ 1207 if (zilog == NULL || seq == 0) 1208 return; 1209 1210 mutex_enter(&zilog->zl_lock); 1211 1212 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1213 1214 while (zilog->zl_writer) { 1215 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1216 if (seq < zilog->zl_commit_seq) { 1217 mutex_exit(&zilog->zl_lock); 1218 return; 1219 } 1220 } 1221 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 1222 /* wake up others waiting on the commit */ 1223 cv_broadcast(&zilog->zl_cv_writer); 1224 mutex_exit(&zilog->zl_lock); 1225} 1226 1227/* 1228 * Called in syncing context to free committed log blocks and update log header. 1229 */ 1230void 1231zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1232{ 1233 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1234 uint64_t txg = dmu_tx_get_txg(tx); 1235 spa_t *spa = zilog->zl_spa; 1236 lwb_t *lwb; 1237 1238 /* 1239 * We don't zero out zl_destroy_txg, so make sure we don't try 1240 * to destroy it twice. 1241 */ 1242 if (spa_sync_pass(spa) != 1) 1243 return; 1244 1245 mutex_enter(&zilog->zl_lock); 1246 1247 ASSERT(zilog->zl_stop_sync == 0); 1248 1249 zh->zh_replay_seq = zilog->zl_replayed_seq[txg & TXG_MASK]; 1250 1251 if (zilog->zl_destroy_txg == txg) { 1252 blkptr_t blk = zh->zh_log; 1253 1254 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1255 1256 bzero(zh, sizeof (zil_header_t)); 1257 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1258 1259 if (zilog->zl_keep_first) { 1260 /* 1261 * If this block was part of log chain that couldn't 1262 * be claimed because a device was missing during 1263 * zil_claim(), but that device later returns, 1264 * then this block could erroneously appear valid. 1265 * To guard against this, assign a new GUID to the new 1266 * log chain so it doesn't matter what blk points to. 1267 */ 1268 zil_init_log_chain(zilog, &blk); 1269 zh->zh_log = blk; 1270 } 1271 } 1272 1273 for (;;) { 1274 lwb = list_head(&zilog->zl_lwb_list); 1275 if (lwb == NULL) { 1276 mutex_exit(&zilog->zl_lock); 1277 return; 1278 } 1279 zh->zh_log = lwb->lwb_blk; 1280 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1281 break; 1282 list_remove(&zilog->zl_lwb_list, lwb); 1283 zio_free_blk(spa, &lwb->lwb_blk, txg); 1284 kmem_cache_free(zil_lwb_cache, lwb); 1285 1286 /* 1287 * If we don't have anything left in the lwb list then 1288 * we've had an allocation failure and we need to zero 1289 * out the zil_header blkptr so that we don't end 1290 * up freeing the same block twice. 1291 */ 1292 if (list_head(&zilog->zl_lwb_list) == NULL) 1293 BP_ZERO(&zh->zh_log); 1294 } 1295 mutex_exit(&zilog->zl_lock); 1296} 1297 1298void 1299zil_init(void) 1300{ 1301 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1302 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1303} 1304 1305void 1306zil_fini(void) 1307{ 1308 kmem_cache_destroy(zil_lwb_cache); 1309} 1310 1311zilog_t * 1312zil_alloc(objset_t *os, zil_header_t *zh_phys) 1313{ 1314 zilog_t *zilog; 1315 1316 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1317 1318 zilog->zl_header = zh_phys; 1319 zilog->zl_os = os; 1320 zilog->zl_spa = dmu_objset_spa(os); 1321 zilog->zl_dmu_pool = dmu_objset_pool(os); 1322 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1323 1324 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1325 1326 list_create(&zilog->zl_itx_list, sizeof (itx_t), 1327 offsetof(itx_t, itx_node)); 1328 1329 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1330 offsetof(lwb_t, lwb_node)); 1331 1332 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1333 1334 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1335 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1336 1337 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1338 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1339 1340 return (zilog); 1341} 1342 1343void 1344zil_free(zilog_t *zilog) 1345{ 1346 lwb_t *lwb; 1347 1348 zilog->zl_stop_sync = 1; 1349 1350 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1351 list_remove(&zilog->zl_lwb_list, lwb); 1352 if (lwb->lwb_buf != NULL) 1353 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1354 kmem_cache_free(zil_lwb_cache, lwb); 1355 } 1356 list_destroy(&zilog->zl_lwb_list); 1357 1358 avl_destroy(&zilog->zl_vdev_tree); 1359 mutex_destroy(&zilog->zl_vdev_lock); 1360 1361 ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1362 list_destroy(&zilog->zl_itx_list); 1363 mutex_destroy(&zilog->zl_lock); 1364 1365 cv_destroy(&zilog->zl_cv_writer); 1366 cv_destroy(&zilog->zl_cv_suspend); 1367 1368 kmem_free(zilog, sizeof (zilog_t)); 1369} 1370 1371/* 1372 * Open an intent log. 1373 */ 1374zilog_t * 1375zil_open(objset_t *os, zil_get_data_t *get_data) 1376{ 1377 zilog_t *zilog = dmu_objset_zil(os); 1378 1379 zilog->zl_get_data = get_data; 1380 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1381 2, 2, TASKQ_PREPOPULATE); 1382 1383 return (zilog); 1384} 1385 1386/* 1387 * Close an intent log. 1388 */ 1389void 1390zil_close(zilog_t *zilog) 1391{ 1392 /* 1393 * If the log isn't already committed, mark the objset dirty 1394 * (so zil_sync() will be called) and wait for that txg to sync. 1395 */ 1396 if (!zil_is_committed(zilog)) { 1397 uint64_t txg; 1398 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1399 (void) dmu_tx_assign(tx, TXG_WAIT); 1400 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1401 txg = dmu_tx_get_txg(tx); 1402 dmu_tx_commit(tx); 1403 txg_wait_synced(zilog->zl_dmu_pool, txg); 1404 } 1405 1406 taskq_destroy(zilog->zl_clean_taskq); 1407 zilog->zl_clean_taskq = NULL; 1408 zilog->zl_get_data = NULL; 1409 1410 zil_itx_clean(zilog); 1411 ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1412} 1413 1414/* 1415 * Suspend an intent log. While in suspended mode, we still honor 1416 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1417 * We suspend the log briefly when taking a snapshot so that the snapshot 1418 * contains all the data it's supposed to, and has an empty intent log. 1419 */ 1420int 1421zil_suspend(zilog_t *zilog) 1422{ 1423 const zil_header_t *zh = zilog->zl_header; 1424 1425 mutex_enter(&zilog->zl_lock); 1426 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1427 mutex_exit(&zilog->zl_lock); 1428 return (EBUSY); 1429 } 1430 if (zilog->zl_suspend++ != 0) { 1431 /* 1432 * Someone else already began a suspend. 1433 * Just wait for them to finish. 1434 */ 1435 while (zilog->zl_suspending) 1436 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1437 mutex_exit(&zilog->zl_lock); 1438 return (0); 1439 } 1440 zilog->zl_suspending = B_TRUE; 1441 mutex_exit(&zilog->zl_lock); 1442 1443 zil_commit(zilog, UINT64_MAX, 0); 1444 1445 /* 1446 * Wait for any in-flight log writes to complete. 1447 */ 1448 mutex_enter(&zilog->zl_lock); 1449 while (zilog->zl_writer) 1450 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1451 mutex_exit(&zilog->zl_lock); 1452 1453 zil_destroy(zilog, B_FALSE); 1454 1455 mutex_enter(&zilog->zl_lock); 1456 zilog->zl_suspending = B_FALSE; 1457 cv_broadcast(&zilog->zl_cv_suspend); 1458 mutex_exit(&zilog->zl_lock); 1459 1460 return (0); 1461} 1462 1463void 1464zil_resume(zilog_t *zilog) 1465{ 1466 mutex_enter(&zilog->zl_lock); 1467 ASSERT(zilog->zl_suspend != 0); 1468 zilog->zl_suspend--; 1469 mutex_exit(&zilog->zl_lock); 1470} 1471 1472/* 1473 * Read in the data for the dmu_sync()ed block, and change the log 1474 * record to write this whole block. 1475 */ 1476void 1477zil_get_replay_data(zilog_t *zilog, lr_write_t *lr) 1478{ 1479 blkptr_t *wbp = &lr->lr_blkptr; 1480 char *wbuf = (char *)(lr + 1); /* data follows lr_write_t */ 1481 uint64_t blksz; 1482 1483 if (BP_IS_HOLE(wbp)) { /* compressed to a hole */ 1484 blksz = BP_GET_LSIZE(&lr->lr_blkptr); 1485 /* 1486 * If the blksz is zero then we must be replaying a log 1487 * from an version prior to setting the blksize of null blocks. 1488 * So we just zero the actual write size reqeusted. 1489 */ 1490 if (blksz == 0) { 1491 bzero(wbuf, lr->lr_length); 1492 return; 1493 } 1494 bzero(wbuf, blksz); 1495 } else { 1496 /* 1497 * A subsequent write may have overwritten this block, in which 1498 * case wbp may have been been freed and reallocated, and our 1499 * read of wbp may fail with a checksum error. We can safely 1500 * ignore this because the later write will provide the 1501 * correct data. 1502 */ 1503 zbookmark_t zb; 1504 1505 zb.zb_objset = dmu_objset_id(zilog->zl_os); 1506 zb.zb_object = lr->lr_foid; 1507 zb.zb_level = 0; 1508 zb.zb_blkid = -1; /* unknown */ 1509 1510 blksz = BP_GET_LSIZE(&lr->lr_blkptr); 1511 (void) zio_wait(zio_read(NULL, zilog->zl_spa, wbp, wbuf, blksz, 1512 NULL, NULL, ZIO_PRIORITY_SYNC_READ, 1513 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb)); 1514 } 1515 lr->lr_offset -= lr->lr_offset % blksz; 1516 lr->lr_length = blksz; 1517} 1518 1519typedef struct zil_replay_arg { 1520 objset_t *zr_os; 1521 zil_replay_func_t **zr_replay; 1522 void *zr_arg; 1523 boolean_t zr_byteswap; 1524 char *zr_lrbuf; 1525} zil_replay_arg_t; 1526 1527static void 1528zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1529{ 1530 zil_replay_arg_t *zr = zra; 1531 const zil_header_t *zh = zilog->zl_header; 1532 uint64_t reclen = lr->lrc_reclen; 1533 uint64_t txtype = lr->lrc_txtype; 1534 char *name; 1535 int pass, error; 1536 1537 if (!zilog->zl_replay) /* giving up */ 1538 return; 1539 1540 if (lr->lrc_txg < claim_txg) /* already committed */ 1541 return; 1542 1543 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1544 return; 1545 1546 /* Strip case-insensitive bit, still present in log record */ 1547 txtype &= ~TX_CI; 1548 1549 if (txtype == 0 || txtype >= TX_MAX_TYPE) { 1550 error = EINVAL; 1551 goto bad; 1552 } 1553 1554 /* 1555 * Make a copy of the data so we can revise and extend it. 1556 */ 1557 bcopy(lr, zr->zr_lrbuf, reclen); 1558 1559 /* 1560 * The log block containing this lr may have been byteswapped 1561 * so that we can easily examine common fields like lrc_txtype. 1562 * However, the log is a mix of different data types, and only the 1563 * replay vectors know how to byteswap their records. Therefore, if 1564 * the lr was byteswapped, undo it before invoking the replay vector. 1565 */ 1566 if (zr->zr_byteswap) 1567 byteswap_uint64_array(zr->zr_lrbuf, reclen); 1568 1569 /* 1570 * We must now do two things atomically: replay this log record, 1571 * and update the log header sequence number to reflect the fact that 1572 * we did so. At the end of each replay function the sequence number 1573 * is updated if we are in replay mode. 1574 */ 1575 for (pass = 1; pass <= 2; pass++) { 1576 zilog->zl_replaying_seq = lr->lrc_seq; 1577 /* Only byteswap (if needed) on the 1st pass. */ 1578 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf, 1579 zr->zr_byteswap && pass == 1); 1580 1581 if (!error) 1582 return; 1583 1584 /* 1585 * The DMU's dnode layer doesn't see removes until the txg 1586 * commits, so a subsequent claim can spuriously fail with 1587 * EEXIST. So if we receive any error we try syncing out 1588 * any removes then retry the transaction. 1589 */ 1590 if (pass == 1) 1591 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1592 } 1593 1594bad: 1595 ASSERT(error); 1596 name = kmem_alloc(MAXNAMELEN, KM_SLEEP); 1597 dmu_objset_name(zr->zr_os, name); 1598 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1599 "dataset %s, seq 0x%llx, txtype %llu %s\n", 1600 error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype, 1601 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1602 zilog->zl_replay = B_FALSE; 1603 kmem_free(name, MAXNAMELEN); 1604} 1605 1606/* ARGSUSED */ 1607static void 1608zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 1609{ 1610 zilog->zl_replay_blks++; 1611} 1612 1613/* 1614 * If this dataset has a non-empty intent log, replay it and destroy it. 1615 */ 1616void 1617zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1618{ 1619 zilog_t *zilog = dmu_objset_zil(os); 1620 const zil_header_t *zh = zilog->zl_header; 1621 zil_replay_arg_t zr; 1622 1623 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1624 zil_destroy(zilog, B_TRUE); 1625 return; 1626 } 1627 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 1628 1629 zr.zr_os = os; 1630 zr.zr_replay = replay_func; 1631 zr.zr_arg = arg; 1632 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1633 zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1634 1635 /* 1636 * Wait for in-progress removes to sync before starting replay. 1637 */ 1638 txg_wait_synced(zilog->zl_dmu_pool, 0); 1639 1640 zilog->zl_replay = B_TRUE; 1641 zilog->zl_replay_time = LBOLT; 1642 ASSERT(zilog->zl_replay_blks == 0); 1643 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1644 zh->zh_claim_txg); 1645 kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE); 1646 1647 zil_destroy(zilog, B_FALSE); 1648 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 1649 zilog->zl_replay = B_FALSE; 1650 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 1651} 1652 1653/* 1654 * Report whether all transactions are committed 1655 */ 1656int 1657zil_is_committed(zilog_t *zilog) 1658{ 1659 lwb_t *lwb; 1660 int ret; 1661 1662 mutex_enter(&zilog->zl_lock); 1663 while (zilog->zl_writer) 1664 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1665 1666 /* recent unpushed intent log transactions? */ 1667 if (!list_is_empty(&zilog->zl_itx_list)) { 1668 ret = B_FALSE; 1669 goto out; 1670 } 1671 1672 /* intent log never used? */ 1673 lwb = list_head(&zilog->zl_lwb_list); 1674 if (lwb == NULL) { 1675 ret = B_TRUE; 1676 goto out; 1677 } 1678 1679 /* 1680 * more than 1 log buffer means zil_sync() hasn't yet freed 1681 * entries after a txg has committed 1682 */ 1683 if (list_next(&zilog->zl_lwb_list, lwb)) { 1684 ret = B_FALSE; 1685 goto out; 1686 } 1687 1688 ASSERT(zil_empty(zilog)); 1689 ret = B_TRUE; 1690out: 1691 cv_broadcast(&zilog->zl_cv_writer); 1692 mutex_exit(&zilog->zl_lock); 1693 return (ret); 1694} 1695