zil.c revision 169028
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 2007 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26#pragma ident "%Z%%M% %I% %E% SMI" 27 28#include <sys/zfs_context.h> 29#include <sys/spa.h> 30#include <sys/dmu.h> 31#include <sys/zap.h> 32#include <sys/arc.h> 33#include <sys/stat.h> 34#include <sys/resource.h> 35#include <sys/zil.h> 36#include <sys/zil_impl.h> 37#include <sys/dsl_dataset.h> 38#include <sys/vdev.h> 39#include <sys/dmu_tx.h> 40 41/* 42 * The zfs intent log (ZIL) saves transaction records of system calls 43 * that change the file system in memory with enough information 44 * to be able to replay them. These are stored in memory until 45 * either the DMU transaction group (txg) commits them to the stable pool 46 * and they can be discarded, or they are flushed to the stable log 47 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 48 * requirement. In the event of a panic or power fail then those log 49 * records (transactions) are replayed. 50 * 51 * There is one ZIL per file system. Its on-disk (pool) format consists 52 * of 3 parts: 53 * 54 * - ZIL header 55 * - ZIL blocks 56 * - ZIL records 57 * 58 * A log record holds a system call transaction. Log blocks can 59 * hold many log records and the blocks are chained together. 60 * Each ZIL block contains a block pointer (blkptr_t) to the next 61 * ZIL block in the chain. The ZIL header points to the first 62 * block in the chain. Note there is not a fixed place in the pool 63 * to hold blocks. They are dynamically allocated and freed as 64 * needed from the blocks available. Figure X shows the ZIL structure: 65 */ 66 67/* 68 * This global ZIL switch affects all pools 69 */ 70int zil_disable = 0; /* disable intent logging */ 71SYSCTL_DECL(_vfs_zfs); 72TUNABLE_INT("vfs.zfs.zil_disable", &zil_disable); 73SYSCTL_INT(_vfs_zfs, OID_AUTO, zil_disable, CTLFLAG_RW, &zil_disable, 0, 74 "Disable ZFS Intent Log (ZIL)"); 75 76/* 77 * Tunable parameter for debugging or performance analysis. Setting 78 * zfs_nocacheflush will cause corruption on power loss if a volatile 79 * out-of-order write cache is enabled. 80 */ 81boolean_t zfs_nocacheflush = B_FALSE; 82TUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 83SYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 84 &zfs_nocacheflush, 0, "Disable cache flush"); 85 86static kmem_cache_t *zil_lwb_cache; 87 88static int 89zil_dva_compare(const void *x1, const void *x2) 90{ 91 const dva_t *dva1 = x1; 92 const dva_t *dva2 = x2; 93 94 if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 95 return (-1); 96 if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 97 return (1); 98 99 if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 100 return (-1); 101 if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 102 return (1); 103 104 return (0); 105} 106 107static void 108zil_dva_tree_init(avl_tree_t *t) 109{ 110 avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t), 111 offsetof(zil_dva_node_t, zn_node)); 112} 113 114static void 115zil_dva_tree_fini(avl_tree_t *t) 116{ 117 zil_dva_node_t *zn; 118 void *cookie = NULL; 119 120 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 121 kmem_free(zn, sizeof (zil_dva_node_t)); 122 123 avl_destroy(t); 124} 125 126static int 127zil_dva_tree_add(avl_tree_t *t, dva_t *dva) 128{ 129 zil_dva_node_t *zn; 130 avl_index_t where; 131 132 if (avl_find(t, dva, &where) != NULL) 133 return (EEXIST); 134 135 zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP); 136 zn->zn_dva = *dva; 137 avl_insert(t, zn, where); 138 139 return (0); 140} 141 142static zil_header_t * 143zil_header_in_syncing_context(zilog_t *zilog) 144{ 145 return ((zil_header_t *)zilog->zl_header); 146} 147 148static void 149zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 150{ 151 zio_cksum_t *zc = &bp->blk_cksum; 152 153 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 154 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 155 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 156 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 157} 158 159/* 160 * Read a log block, make sure it's valid, and byteswap it if necessary. 161 */ 162static int 163zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp) 164{ 165 blkptr_t blk = *bp; 166 zbookmark_t zb; 167 uint32_t aflags = ARC_WAIT; 168 int error; 169 170 zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET]; 171 zb.zb_object = 0; 172 zb.zb_level = -1; 173 zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ]; 174 175 *abufpp = NULL; 176 177 error = arc_read(NULL, zilog->zl_spa, &blk, byteswap_uint64_array, 178 arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL | 179 ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb); 180 181 if (error == 0) { 182 char *data = (*abufpp)->b_data; 183 uint64_t blksz = BP_GET_LSIZE(bp); 184 zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1; 185 zio_cksum_t cksum = bp->blk_cksum; 186 187 /* 188 * Sequence numbers should be... sequential. The checksum 189 * verifier for the next block should be bp's checksum plus 1. 190 */ 191 cksum.zc_word[ZIL_ZC_SEQ]++; 192 193 if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, sizeof (cksum))) 194 error = ESTALE; 195 else if (BP_IS_HOLE(&ztp->zit_next_blk)) 196 error = ENOENT; 197 else if (ztp->zit_nused > (blksz - sizeof (zil_trailer_t))) 198 error = EOVERFLOW; 199 200 if (error) { 201 VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1); 202 *abufpp = NULL; 203 } 204 } 205 206 dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid); 207 208 return (error); 209} 210 211/* 212 * Parse the intent log, and call parse_func for each valid record within. 213 * Return the highest sequence number. 214 */ 215uint64_t 216zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 217 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 218{ 219 const zil_header_t *zh = zilog->zl_header; 220 uint64_t claim_seq = zh->zh_claim_seq; 221 uint64_t seq = 0; 222 uint64_t max_seq = 0; 223 blkptr_t blk = zh->zh_log; 224 arc_buf_t *abuf; 225 char *lrbuf, *lrp; 226 zil_trailer_t *ztp; 227 int reclen, error; 228 229 if (BP_IS_HOLE(&blk)) 230 return (max_seq); 231 232 /* 233 * Starting at the block pointed to by zh_log we read the log chain. 234 * For each block in the chain we strongly check that block to 235 * ensure its validity. We stop when an invalid block is found. 236 * For each block pointer in the chain we call parse_blk_func(). 237 * For each record in each valid block we call parse_lr_func(). 238 * If the log has been claimed, stop if we encounter a sequence 239 * number greater than the highest claimed sequence number. 240 */ 241 zil_dva_tree_init(&zilog->zl_dva_tree); 242 for (;;) { 243 seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 244 245 if (claim_seq != 0 && seq > claim_seq) 246 break; 247 248 ASSERT(max_seq < seq); 249 max_seq = seq; 250 251 error = zil_read_log_block(zilog, &blk, &abuf); 252 253 if (parse_blk_func != NULL) 254 parse_blk_func(zilog, &blk, arg, txg); 255 256 if (error) 257 break; 258 259 lrbuf = abuf->b_data; 260 ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 261 blk = ztp->zit_next_blk; 262 263 if (parse_lr_func == NULL) { 264 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 265 continue; 266 } 267 268 for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { 269 lr_t *lr = (lr_t *)lrp; 270 reclen = lr->lrc_reclen; 271 ASSERT3U(reclen, >=, sizeof (lr_t)); 272 parse_lr_func(zilog, lr, arg, txg); 273 } 274 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 275 } 276 zil_dva_tree_fini(&zilog->zl_dva_tree); 277 278 return (max_seq); 279} 280 281/* ARGSUSED */ 282static void 283zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 284{ 285 spa_t *spa = zilog->zl_spa; 286 int err; 287 288 /* 289 * Claim log block if not already committed and not already claimed. 290 */ 291 if (bp->blk_birth >= first_txg && 292 zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) { 293 err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL)); 294 ASSERT(err == 0); 295 } 296} 297 298static void 299zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 300{ 301 if (lrc->lrc_txtype == TX_WRITE) { 302 lr_write_t *lr = (lr_write_t *)lrc; 303 zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg); 304 } 305} 306 307/* ARGSUSED */ 308static void 309zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 310{ 311 zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx)); 312} 313 314static void 315zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 316{ 317 /* 318 * If we previously claimed it, we need to free it. 319 */ 320 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) { 321 lr_write_t *lr = (lr_write_t *)lrc; 322 blkptr_t *bp = &lr->lr_blkptr; 323 if (bp->blk_birth >= claim_txg && 324 !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) { 325 (void) arc_free(NULL, zilog->zl_spa, 326 dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT); 327 } 328 } 329} 330 331/* 332 * Create an on-disk intent log. 333 */ 334static void 335zil_create(zilog_t *zilog) 336{ 337 const zil_header_t *zh = zilog->zl_header; 338 lwb_t *lwb; 339 uint64_t txg = 0; 340 dmu_tx_t *tx = NULL; 341 blkptr_t blk; 342 int error = 0; 343 344 /* 345 * Wait for any previous destroy to complete. 346 */ 347 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 348 349 ASSERT(zh->zh_claim_txg == 0); 350 ASSERT(zh->zh_replay_seq == 0); 351 352 blk = zh->zh_log; 353 354 /* 355 * If we don't already have an initial log block, allocate one now. 356 */ 357 if (BP_IS_HOLE(&blk)) { 358 tx = dmu_tx_create(zilog->zl_os); 359 (void) dmu_tx_assign(tx, TXG_WAIT); 360 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 361 txg = dmu_tx_get_txg(tx); 362 363 error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk, 364 NULL, txg); 365 366 if (error == 0) 367 zil_init_log_chain(zilog, &blk); 368 } 369 370 /* 371 * Allocate a log write buffer (lwb) for the first log block. 372 */ 373 if (error == 0) { 374 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 375 lwb->lwb_zilog = zilog; 376 lwb->lwb_blk = blk; 377 lwb->lwb_nused = 0; 378 lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); 379 lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); 380 lwb->lwb_max_txg = txg; 381 lwb->lwb_zio = NULL; 382 383 mutex_enter(&zilog->zl_lock); 384 list_insert_tail(&zilog->zl_lwb_list, lwb); 385 mutex_exit(&zilog->zl_lock); 386 } 387 388 /* 389 * If we just allocated the first log block, commit our transaction 390 * and wait for zil_sync() to stuff the block poiner into zh_log. 391 * (zh is part of the MOS, so we cannot modify it in open context.) 392 */ 393 if (tx != NULL) { 394 dmu_tx_commit(tx); 395 txg_wait_synced(zilog->zl_dmu_pool, txg); 396 } 397 398 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 399} 400 401/* 402 * In one tx, free all log blocks and clear the log header. 403 * If keep_first is set, then we're replaying a log with no content. 404 * We want to keep the first block, however, so that the first 405 * synchronous transaction doesn't require a txg_wait_synced() 406 * in zil_create(). We don't need to txg_wait_synced() here either 407 * when keep_first is set, because both zil_create() and zil_destroy() 408 * will wait for any in-progress destroys to complete. 409 */ 410void 411zil_destroy(zilog_t *zilog, boolean_t keep_first) 412{ 413 const zil_header_t *zh = zilog->zl_header; 414 lwb_t *lwb; 415 dmu_tx_t *tx; 416 uint64_t txg; 417 418 /* 419 * Wait for any previous destroy to complete. 420 */ 421 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 422 423 if (BP_IS_HOLE(&zh->zh_log)) 424 return; 425 426 tx = dmu_tx_create(zilog->zl_os); 427 (void) dmu_tx_assign(tx, TXG_WAIT); 428 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 429 txg = dmu_tx_get_txg(tx); 430 431 mutex_enter(&zilog->zl_lock); 432 433 ASSERT3U(zilog->zl_destroy_txg, <, txg); 434 zilog->zl_destroy_txg = txg; 435 zilog->zl_keep_first = keep_first; 436 437 if (!list_is_empty(&zilog->zl_lwb_list)) { 438 ASSERT(zh->zh_claim_txg == 0); 439 ASSERT(!keep_first); 440 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 441 list_remove(&zilog->zl_lwb_list, lwb); 442 if (lwb->lwb_buf != NULL) 443 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 444 zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg); 445 kmem_cache_free(zil_lwb_cache, lwb); 446 } 447 } else { 448 if (!keep_first) { 449 (void) zil_parse(zilog, zil_free_log_block, 450 zil_free_log_record, tx, zh->zh_claim_txg); 451 } 452 } 453 mutex_exit(&zilog->zl_lock); 454 455 dmu_tx_commit(tx); 456 457 if (keep_first) /* no need to wait in this case */ 458 return; 459 460 txg_wait_synced(zilog->zl_dmu_pool, txg); 461 ASSERT(BP_IS_HOLE(&zh->zh_log)); 462} 463 464int 465zil_claim(char *osname, void *txarg) 466{ 467 dmu_tx_t *tx = txarg; 468 uint64_t first_txg = dmu_tx_get_txg(tx); 469 zilog_t *zilog; 470 zil_header_t *zh; 471 objset_t *os; 472 int error; 473 474 error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_STANDARD, &os); 475 if (error) { 476 cmn_err(CE_WARN, "can't process intent log for %s", osname); 477 return (0); 478 } 479 480 zilog = dmu_objset_zil(os); 481 zh = zil_header_in_syncing_context(zilog); 482 483 /* 484 * Claim all log blocks if we haven't already done so, and remember 485 * the highest claimed sequence number. This ensures that if we can 486 * read only part of the log now (e.g. due to a missing device), 487 * but we can read the entire log later, we will not try to replay 488 * or destroy beyond the last block we successfully claimed. 489 */ 490 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 491 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 492 zh->zh_claim_txg = first_txg; 493 zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block, 494 zil_claim_log_record, tx, first_txg); 495 dsl_dataset_dirty(dmu_objset_ds(os), tx); 496 } 497 498 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 499 dmu_objset_close(os); 500 return (0); 501} 502 503void 504zil_add_vdev(zilog_t *zilog, uint64_t vdev) 505{ 506 zil_vdev_t *zv, *new; 507 uint64_t bmap_sz = sizeof (zilog->zl_vdev_bmap) << 3; 508 uchar_t *cp; 509 510 if (zfs_nocacheflush) 511 return; 512 513 if (vdev < bmap_sz) { 514 cp = zilog->zl_vdev_bmap + (vdev / 8); 515 atomic_or_8(cp, 1 << (vdev % 8)); 516 } else { 517 /* 518 * insert into ordered list 519 */ 520 mutex_enter(&zilog->zl_lock); 521 for (zv = list_head(&zilog->zl_vdev_list); zv != NULL; 522 zv = list_next(&zilog->zl_vdev_list, zv)) { 523 if (zv->vdev == vdev) { 524 /* duplicate found - just return */ 525 mutex_exit(&zilog->zl_lock); 526 return; 527 } 528 if (zv->vdev > vdev) { 529 /* insert before this entry */ 530 new = kmem_alloc(sizeof (zil_vdev_t), 531 KM_SLEEP); 532 new->vdev = vdev; 533 list_insert_before(&zilog->zl_vdev_list, 534 zv, new); 535 mutex_exit(&zilog->zl_lock); 536 return; 537 } 538 } 539 /* ran off end of list, insert at the end */ 540 ASSERT(zv == NULL); 541 new = kmem_alloc(sizeof (zil_vdev_t), KM_SLEEP); 542 new->vdev = vdev; 543 list_insert_tail(&zilog->zl_vdev_list, new); 544 mutex_exit(&zilog->zl_lock); 545 } 546} 547 548/* start an async flush of the write cache for this vdev */ 549void 550zil_flush_vdev(spa_t *spa, uint64_t vdev, zio_t **zio) 551{ 552 vdev_t *vd; 553 554 if (*zio == NULL) 555 *zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 556 557 vd = vdev_lookup_top(spa, vdev); 558 ASSERT(vd); 559 560 (void) zio_nowait(zio_ioctl(*zio, spa, vd, DKIOCFLUSHWRITECACHE, 561 NULL, NULL, ZIO_PRIORITY_NOW, 562 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY)); 563} 564 565void 566zil_flush_vdevs(zilog_t *zilog) 567{ 568 zil_vdev_t *zv; 569 zio_t *zio = NULL; 570 spa_t *spa = zilog->zl_spa; 571 uint64_t vdev; 572 uint8_t b; 573 int i, j; 574 575 ASSERT(zilog->zl_writer); 576 577 for (i = 0; i < sizeof (zilog->zl_vdev_bmap); i++) { 578 b = zilog->zl_vdev_bmap[i]; 579 if (b == 0) 580 continue; 581 for (j = 0; j < 8; j++) { 582 if (b & (1 << j)) { 583 vdev = (i << 3) + j; 584 zil_flush_vdev(spa, vdev, &zio); 585 } 586 } 587 zilog->zl_vdev_bmap[i] = 0; 588 } 589 590 while ((zv = list_head(&zilog->zl_vdev_list)) != NULL) { 591 zil_flush_vdev(spa, zv->vdev, &zio); 592 list_remove(&zilog->zl_vdev_list, zv); 593 kmem_free(zv, sizeof (zil_vdev_t)); 594 } 595 /* 596 * Wait for all the flushes to complete. Not all devices actually 597 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 598 */ 599 if (zio) 600 (void) zio_wait(zio); 601} 602 603/* 604 * Function called when a log block write completes 605 */ 606static void 607zil_lwb_write_done(zio_t *zio) 608{ 609 lwb_t *lwb = zio->io_private; 610 zilog_t *zilog = lwb->lwb_zilog; 611 612 /* 613 * Now that we've written this log block, we have a stable pointer 614 * to the next block in the chain, so it's OK to let the txg in 615 * which we allocated the next block sync. 616 */ 617 txg_rele_to_sync(&lwb->lwb_txgh); 618 619 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 620 mutex_enter(&zilog->zl_lock); 621 lwb->lwb_buf = NULL; 622 if (zio->io_error) { 623 zilog->zl_log_error = B_TRUE; 624 mutex_exit(&zilog->zl_lock); 625 return; 626 } 627 mutex_exit(&zilog->zl_lock); 628} 629 630/* 631 * Initialize the io for a log block. 632 * 633 * Note, we should not initialize the IO until we are about 634 * to use it, since zio_rewrite() does a spa_config_enter(). 635 */ 636static void 637zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 638{ 639 zbookmark_t zb; 640 641 zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET]; 642 zb.zb_object = 0; 643 zb.zb_level = -1; 644 zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 645 646 if (zilog->zl_root_zio == NULL) { 647 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 648 ZIO_FLAG_CANFAIL); 649 } 650 if (lwb->lwb_zio == NULL) { 651 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 652 ZIO_CHECKSUM_ZILOG, 0, &lwb->lwb_blk, lwb->lwb_buf, 653 lwb->lwb_sz, zil_lwb_write_done, lwb, 654 ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 655 } 656} 657 658/* 659 * Start a log block write and advance to the next log block. 660 * Calls are serialized. 661 */ 662static lwb_t * 663zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 664{ 665 lwb_t *nlwb; 666 zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; 667 spa_t *spa = zilog->zl_spa; 668 blkptr_t *bp = &ztp->zit_next_blk; 669 uint64_t txg; 670 uint64_t zil_blksz; 671 int error; 672 673 ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); 674 675 /* 676 * Allocate the next block and save its address in this block 677 * before writing it in order to establish the log chain. 678 * Note that if the allocation of nlwb synced before we wrote 679 * the block that points at it (lwb), we'd leak it if we crashed. 680 * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done(). 681 */ 682 txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh); 683 txg_rele_to_quiesce(&lwb->lwb_txgh); 684 685 /* 686 * Pick a ZIL blocksize. We request a size that is the 687 * maximum of the previous used size, the current used size and 688 * the amount waiting in the queue. 689 */ 690 zil_blksz = MAX(zilog->zl_prev_used, 691 zilog->zl_cur_used + sizeof (*ztp)); 692 zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); 693 zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); 694 if (zil_blksz > ZIL_MAX_BLKSZ) 695 zil_blksz = ZIL_MAX_BLKSZ; 696 697 BP_ZERO(bp); 698 /* pass the old blkptr in order to spread log blocks across devs */ 699 error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg); 700 if (error) { 701 dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg); 702 703 /* 704 * We dirty the dataset to ensure that zil_sync() will 705 * be called to remove this lwb from our zl_lwb_list. 706 * Failing to do so, may leave an lwb with a NULL lwb_buf 707 * hanging around on the zl_lwb_list. 708 */ 709 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 710 dmu_tx_commit(tx); 711 712 /* 713 * Since we've just experienced an allocation failure so we 714 * terminate the current lwb and send it on its way. 715 */ 716 ztp->zit_pad = 0; 717 ztp->zit_nused = lwb->lwb_nused; 718 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 719 zio_nowait(lwb->lwb_zio); 720 721 /* 722 * By returning NULL the caller will call tx_wait_synced() 723 */ 724 return (NULL); 725 } 726 727 ASSERT3U(bp->blk_birth, ==, txg); 728 ztp->zit_pad = 0; 729 ztp->zit_nused = lwb->lwb_nused; 730 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 731 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 732 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 733 734 /* 735 * Allocate a new log write buffer (lwb). 736 */ 737 nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 738 739 nlwb->lwb_zilog = zilog; 740 nlwb->lwb_blk = *bp; 741 nlwb->lwb_nused = 0; 742 nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); 743 nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); 744 nlwb->lwb_max_txg = txg; 745 nlwb->lwb_zio = NULL; 746 747 /* 748 * Put new lwb at the end of the log chain 749 */ 750 mutex_enter(&zilog->zl_lock); 751 list_insert_tail(&zilog->zl_lwb_list, nlwb); 752 mutex_exit(&zilog->zl_lock); 753 754 /* Record the vdev for later flushing */ 755 zil_add_vdev(zilog, DVA_GET_VDEV(BP_IDENTITY(&(lwb->lwb_blk)))); 756 757 /* 758 * kick off the write for the old log block 759 */ 760 dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg); 761 ASSERT(lwb->lwb_zio); 762 zio_nowait(lwb->lwb_zio); 763 764 return (nlwb); 765} 766 767static lwb_t * 768zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 769{ 770 lr_t *lrc = &itx->itx_lr; /* common log record */ 771 lr_write_t *lr = (lr_write_t *)lrc; 772 uint64_t txg = lrc->lrc_txg; 773 uint64_t reclen = lrc->lrc_reclen; 774 uint64_t dlen; 775 776 if (lwb == NULL) 777 return (NULL); 778 ASSERT(lwb->lwb_buf != NULL); 779 780 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 781 dlen = P2ROUNDUP_TYPED( 782 lr->lr_length, sizeof (uint64_t), uint64_t); 783 else 784 dlen = 0; 785 786 zilog->zl_cur_used += (reclen + dlen); 787 788 zil_lwb_write_init(zilog, lwb); 789 790 /* 791 * If this record won't fit in the current log block, start a new one. 792 */ 793 if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 794 lwb = zil_lwb_write_start(zilog, lwb); 795 if (lwb == NULL) 796 return (NULL); 797 zil_lwb_write_init(zilog, lwb); 798 ASSERT(lwb->lwb_nused == 0); 799 if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 800 txg_wait_synced(zilog->zl_dmu_pool, txg); 801 return (lwb); 802 } 803 } 804 805 /* 806 * Update the lrc_seq, to be log record sequence number. See zil.h 807 * Then copy the record to the log buffer. 808 */ 809 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 810 bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen); 811 812 /* 813 * If it's a write, fetch the data or get its blkptr as appropriate. 814 */ 815 if (lrc->lrc_txtype == TX_WRITE) { 816 if (txg > spa_freeze_txg(zilog->zl_spa)) 817 txg_wait_synced(zilog->zl_dmu_pool, txg); 818 if (itx->itx_wr_state != WR_COPIED) { 819 char *dbuf; 820 int error; 821 822 /* alignment is guaranteed */ 823 lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused); 824 if (dlen) { 825 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 826 dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen; 827 lr->lr_common.lrc_reclen += dlen; 828 } else { 829 ASSERT(itx->itx_wr_state == WR_INDIRECT); 830 dbuf = NULL; 831 } 832 error = zilog->zl_get_data( 833 itx->itx_private, lr, dbuf, lwb->lwb_zio); 834 if (error) { 835 ASSERT(error == ENOENT || error == EEXIST || 836 error == EALREADY); 837 return (lwb); 838 } 839 } 840 } 841 842 lwb->lwb_nused += reclen + dlen; 843 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 844 ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); 845 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 846 847 return (lwb); 848} 849 850itx_t * 851zil_itx_create(int txtype, size_t lrsize) 852{ 853 itx_t *itx; 854 855 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 856 857 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 858 itx->itx_lr.lrc_txtype = txtype; 859 itx->itx_lr.lrc_reclen = lrsize; 860 itx->itx_lr.lrc_seq = 0; /* defensive */ 861 862 return (itx); 863} 864 865uint64_t 866zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 867{ 868 uint64_t seq; 869 870 ASSERT(itx->itx_lr.lrc_seq == 0); 871 872 mutex_enter(&zilog->zl_lock); 873 list_insert_tail(&zilog->zl_itx_list, itx); 874 zilog->zl_itx_list_sz += itx->itx_lr.lrc_reclen; 875 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 876 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 877 mutex_exit(&zilog->zl_lock); 878 879 return (seq); 880} 881 882/* 883 * Free up all in-memory intent log transactions that have now been synced. 884 */ 885static void 886zil_itx_clean(zilog_t *zilog) 887{ 888 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 889 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 890 list_t clean_list; 891 itx_t *itx; 892 893 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 894 895 mutex_enter(&zilog->zl_lock); 896 /* wait for a log writer to finish walking list */ 897 while (zilog->zl_writer) { 898 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 899 } 900 901 /* 902 * Move the sync'd log transactions to a separate list so we can call 903 * kmem_free without holding the zl_lock. 904 * 905 * There is no need to set zl_writer as we don't drop zl_lock here 906 */ 907 while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 908 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 909 list_remove(&zilog->zl_itx_list, itx); 910 zilog->zl_itx_list_sz -= itx->itx_lr.lrc_reclen; 911 list_insert_tail(&clean_list, itx); 912 } 913 cv_broadcast(&zilog->zl_cv_writer); 914 mutex_exit(&zilog->zl_lock); 915 916 /* destroy sync'd log transactions */ 917 while ((itx = list_head(&clean_list)) != NULL) { 918 list_remove(&clean_list, itx); 919 kmem_free(itx, offsetof(itx_t, itx_lr) 920 + itx->itx_lr.lrc_reclen); 921 } 922 list_destroy(&clean_list); 923} 924 925/* 926 * If there are any in-memory intent log transactions which have now been 927 * synced then start up a taskq to free them. 928 */ 929void 930zil_clean(zilog_t *zilog) 931{ 932 itx_t *itx; 933 934 mutex_enter(&zilog->zl_lock); 935 itx = list_head(&zilog->zl_itx_list); 936 if ((itx != NULL) && 937 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 938 (void) taskq_dispatch(zilog->zl_clean_taskq, 939 (void (*)(void *))zil_itx_clean, zilog, TQ_NOSLEEP); 940 } 941 mutex_exit(&zilog->zl_lock); 942} 943 944void 945zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 946{ 947 uint64_t txg; 948 uint64_t reclen; 949 uint64_t commit_seq = 0; 950 itx_t *itx, *itx_next = (itx_t *)-1; 951 lwb_t *lwb; 952 spa_t *spa; 953 954 zilog->zl_writer = B_TRUE; 955 zilog->zl_root_zio = NULL; 956 spa = zilog->zl_spa; 957 958 if (zilog->zl_suspend) { 959 lwb = NULL; 960 } else { 961 lwb = list_tail(&zilog->zl_lwb_list); 962 if (lwb == NULL) { 963 /* 964 * Return if there's nothing to flush before we 965 * dirty the fs by calling zil_create() 966 */ 967 if (list_is_empty(&zilog->zl_itx_list)) { 968 zilog->zl_writer = B_FALSE; 969 return; 970 } 971 mutex_exit(&zilog->zl_lock); 972 zil_create(zilog); 973 mutex_enter(&zilog->zl_lock); 974 lwb = list_tail(&zilog->zl_lwb_list); 975 } 976 } 977 978 /* Loop through in-memory log transactions filling log blocks. */ 979 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 980 for (;;) { 981 /* 982 * Find the next itx to push: 983 * Push all transactions related to specified foid and all 984 * other transactions except TX_WRITE, TX_TRUNCATE, 985 * TX_SETATTR and TX_ACL for all other files. 986 */ 987 if (itx_next != (itx_t *)-1) 988 itx = itx_next; 989 else 990 itx = list_head(&zilog->zl_itx_list); 991 for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) { 992 if (foid == 0) /* push all foids? */ 993 break; 994 if (itx->itx_sync) /* push all O_[D]SYNC */ 995 break; 996 switch (itx->itx_lr.lrc_txtype) { 997 case TX_SETATTR: 998 case TX_WRITE: 999 case TX_TRUNCATE: 1000 case TX_ACL: 1001 /* lr_foid is same offset for these records */ 1002 if (((lr_write_t *)&itx->itx_lr)->lr_foid 1003 != foid) { 1004 continue; /* skip this record */ 1005 } 1006 } 1007 break; 1008 } 1009 if (itx == NULL) 1010 break; 1011 1012 reclen = itx->itx_lr.lrc_reclen; 1013 if ((itx->itx_lr.lrc_seq > seq) && 1014 ((lwb == NULL) || (lwb->lwb_nused == 0) || 1015 (lwb->lwb_nused + reclen > ZIL_BLK_DATA_SZ(lwb)))) { 1016 break; 1017 } 1018 1019 /* 1020 * Save the next pointer. Even though we soon drop 1021 * zl_lock all threads that may change the list 1022 * (another writer or zil_itx_clean) can't do so until 1023 * they have zl_writer. 1024 */ 1025 itx_next = list_next(&zilog->zl_itx_list, itx); 1026 list_remove(&zilog->zl_itx_list, itx); 1027 mutex_exit(&zilog->zl_lock); 1028 txg = itx->itx_lr.lrc_txg; 1029 ASSERT(txg); 1030 1031 if (txg > spa_last_synced_txg(spa) || 1032 txg > spa_freeze_txg(spa)) 1033 lwb = zil_lwb_commit(zilog, itx, lwb); 1034 kmem_free(itx, offsetof(itx_t, itx_lr) 1035 + itx->itx_lr.lrc_reclen); 1036 mutex_enter(&zilog->zl_lock); 1037 zilog->zl_itx_list_sz -= reclen; 1038 } 1039 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1040 /* determine commit sequence number */ 1041 itx = list_head(&zilog->zl_itx_list); 1042 if (itx) 1043 commit_seq = itx->itx_lr.lrc_seq; 1044 else 1045 commit_seq = zilog->zl_itx_seq; 1046 mutex_exit(&zilog->zl_lock); 1047 1048 /* write the last block out */ 1049 if (lwb != NULL && lwb->lwb_zio != NULL) 1050 lwb = zil_lwb_write_start(zilog, lwb); 1051 1052 zilog->zl_prev_used = zilog->zl_cur_used; 1053 zilog->zl_cur_used = 0; 1054 1055 /* 1056 * Wait if necessary for the log blocks to be on stable storage. 1057 */ 1058 if (zilog->zl_root_zio) { 1059 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1060 (void) zio_wait(zilog->zl_root_zio); 1061 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 1062 if (!zfs_nocacheflush) 1063 zil_flush_vdevs(zilog); 1064 } 1065 1066 if (zilog->zl_log_error || lwb == NULL) { 1067 zilog->zl_log_error = 0; 1068 txg_wait_synced(zilog->zl_dmu_pool, 0); 1069 } 1070 1071 mutex_enter(&zilog->zl_lock); 1072 zilog->zl_writer = B_FALSE; 1073 1074 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 1075 zilog->zl_commit_seq = commit_seq; 1076} 1077 1078/* 1079 * Push zfs transactions to stable storage up to the supplied sequence number. 1080 * If foid is 0 push out all transactions, otherwise push only those 1081 * for that file or might have been used to create that file. 1082 */ 1083void 1084zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1085{ 1086 if (zilog == NULL || seq == 0) 1087 return; 1088 1089 mutex_enter(&zilog->zl_lock); 1090 1091 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1092 1093 while (zilog->zl_writer) { 1094 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1095 if (seq < zilog->zl_commit_seq) { 1096 mutex_exit(&zilog->zl_lock); 1097 return; 1098 } 1099 } 1100 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 1101 /* wake up others waiting on the commit */ 1102 cv_broadcast(&zilog->zl_cv_writer); 1103 mutex_exit(&zilog->zl_lock); 1104} 1105 1106/* 1107 * Called in syncing context to free committed log blocks and update log header. 1108 */ 1109void 1110zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1111{ 1112 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1113 uint64_t txg = dmu_tx_get_txg(tx); 1114 spa_t *spa = zilog->zl_spa; 1115 lwb_t *lwb; 1116 1117 mutex_enter(&zilog->zl_lock); 1118 1119 ASSERT(zilog->zl_stop_sync == 0); 1120 1121 zh->zh_replay_seq = zilog->zl_replay_seq[txg & TXG_MASK]; 1122 1123 if (zilog->zl_destroy_txg == txg) { 1124 blkptr_t blk = zh->zh_log; 1125 1126 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1127 ASSERT(spa_sync_pass(spa) == 1); 1128 1129 bzero(zh, sizeof (zil_header_t)); 1130 bzero(zilog->zl_replay_seq, sizeof (zilog->zl_replay_seq)); 1131 1132 if (zilog->zl_keep_first) { 1133 /* 1134 * If this block was part of log chain that couldn't 1135 * be claimed because a device was missing during 1136 * zil_claim(), but that device later returns, 1137 * then this block could erroneously appear valid. 1138 * To guard against this, assign a new GUID to the new 1139 * log chain so it doesn't matter what blk points to. 1140 */ 1141 zil_init_log_chain(zilog, &blk); 1142 zh->zh_log = blk; 1143 } 1144 } 1145 1146 for (;;) { 1147 lwb = list_head(&zilog->zl_lwb_list); 1148 if (lwb == NULL) { 1149 mutex_exit(&zilog->zl_lock); 1150 return; 1151 } 1152 zh->zh_log = lwb->lwb_blk; 1153 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1154 break; 1155 list_remove(&zilog->zl_lwb_list, lwb); 1156 zio_free_blk(spa, &lwb->lwb_blk, txg); 1157 kmem_cache_free(zil_lwb_cache, lwb); 1158 1159 /* 1160 * If we don't have anything left in the lwb list then 1161 * we've had an allocation failure and we need to zero 1162 * out the zil_header blkptr so that we don't end 1163 * up freeing the same block twice. 1164 */ 1165 if (list_head(&zilog->zl_lwb_list) == NULL) 1166 BP_ZERO(&zh->zh_log); 1167 } 1168 mutex_exit(&zilog->zl_lock); 1169} 1170 1171void 1172zil_init(void) 1173{ 1174 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1175 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1176} 1177 1178void 1179zil_fini(void) 1180{ 1181 kmem_cache_destroy(zil_lwb_cache); 1182} 1183 1184zilog_t * 1185zil_alloc(objset_t *os, zil_header_t *zh_phys) 1186{ 1187 zilog_t *zilog; 1188 1189 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1190 1191 zilog->zl_header = zh_phys; 1192 zilog->zl_os = os; 1193 zilog->zl_spa = dmu_objset_spa(os); 1194 zilog->zl_dmu_pool = dmu_objset_pool(os); 1195 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1196 1197 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1198 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1199 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1200 1201 list_create(&zilog->zl_itx_list, sizeof (itx_t), 1202 offsetof(itx_t, itx_node)); 1203 1204 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1205 offsetof(lwb_t, lwb_node)); 1206 1207 list_create(&zilog->zl_vdev_list, sizeof (zil_vdev_t), 1208 offsetof(zil_vdev_t, vdev_seq_node)); 1209 1210 return (zilog); 1211} 1212 1213void 1214zil_free(zilog_t *zilog) 1215{ 1216 lwb_t *lwb; 1217 zil_vdev_t *zv; 1218 1219 zilog->zl_stop_sync = 1; 1220 1221 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1222 list_remove(&zilog->zl_lwb_list, lwb); 1223 if (lwb->lwb_buf != NULL) 1224 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1225 kmem_cache_free(zil_lwb_cache, lwb); 1226 } 1227 list_destroy(&zilog->zl_lwb_list); 1228 1229 while ((zv = list_head(&zilog->zl_vdev_list)) != NULL) { 1230 list_remove(&zilog->zl_vdev_list, zv); 1231 kmem_free(zv, sizeof (zil_vdev_t)); 1232 } 1233 list_destroy(&zilog->zl_vdev_list); 1234 1235 ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1236 list_destroy(&zilog->zl_itx_list); 1237 cv_destroy(&zilog->zl_cv_suspend); 1238 cv_destroy(&zilog->zl_cv_writer); 1239 mutex_destroy(&zilog->zl_lock); 1240 1241 kmem_free(zilog, sizeof (zilog_t)); 1242} 1243 1244/* 1245 * return true if the initial log block is not valid 1246 */ 1247static int 1248zil_empty(zilog_t *zilog) 1249{ 1250 const zil_header_t *zh = zilog->zl_header; 1251 arc_buf_t *abuf = NULL; 1252 1253 if (BP_IS_HOLE(&zh->zh_log)) 1254 return (1); 1255 1256 if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0) 1257 return (1); 1258 1259 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 1260 return (0); 1261} 1262 1263/* 1264 * Open an intent log. 1265 */ 1266zilog_t * 1267zil_open(objset_t *os, zil_get_data_t *get_data) 1268{ 1269 zilog_t *zilog = dmu_objset_zil(os); 1270 1271 zilog->zl_get_data = get_data; 1272 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1273 2, 2, TASKQ_PREPOPULATE); 1274 1275 return (zilog); 1276} 1277 1278/* 1279 * Close an intent log. 1280 */ 1281void 1282zil_close(zilog_t *zilog) 1283{ 1284 /* 1285 * If the log isn't already committed, mark the objset dirty 1286 * (so zil_sync() will be called) and wait for that txg to sync. 1287 */ 1288 if (!zil_is_committed(zilog)) { 1289 uint64_t txg; 1290 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1291 (void) dmu_tx_assign(tx, TXG_WAIT); 1292 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1293 txg = dmu_tx_get_txg(tx); 1294 dmu_tx_commit(tx); 1295 txg_wait_synced(zilog->zl_dmu_pool, txg); 1296 } 1297 1298 taskq_destroy(zilog->zl_clean_taskq); 1299 zilog->zl_clean_taskq = NULL; 1300 zilog->zl_get_data = NULL; 1301 1302 zil_itx_clean(zilog); 1303 ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1304} 1305 1306/* 1307 * Suspend an intent log. While in suspended mode, we still honor 1308 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1309 * We suspend the log briefly when taking a snapshot so that the snapshot 1310 * contains all the data it's supposed to, and has an empty intent log. 1311 */ 1312int 1313zil_suspend(zilog_t *zilog) 1314{ 1315 const zil_header_t *zh = zilog->zl_header; 1316 1317 mutex_enter(&zilog->zl_lock); 1318 if (zh->zh_claim_txg != 0) { /* unplayed log */ 1319 mutex_exit(&zilog->zl_lock); 1320 return (EBUSY); 1321 } 1322 if (zilog->zl_suspend++ != 0) { 1323 /* 1324 * Someone else already began a suspend. 1325 * Just wait for them to finish. 1326 */ 1327 while (zilog->zl_suspending) 1328 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1329 ASSERT(BP_IS_HOLE(&zh->zh_log)); 1330 mutex_exit(&zilog->zl_lock); 1331 return (0); 1332 } 1333 zilog->zl_suspending = B_TRUE; 1334 mutex_exit(&zilog->zl_lock); 1335 1336 zil_commit(zilog, UINT64_MAX, 0); 1337 1338 /* 1339 * Wait for any in-flight log writes to complete. 1340 */ 1341 mutex_enter(&zilog->zl_lock); 1342 while (zilog->zl_writer) 1343 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1344 mutex_exit(&zilog->zl_lock); 1345 1346 zil_destroy(zilog, B_FALSE); 1347 1348 mutex_enter(&zilog->zl_lock); 1349 ASSERT(BP_IS_HOLE(&zh->zh_log)); 1350 zilog->zl_suspending = B_FALSE; 1351 cv_broadcast(&zilog->zl_cv_suspend); 1352 mutex_exit(&zilog->zl_lock); 1353 1354 return (0); 1355} 1356 1357void 1358zil_resume(zilog_t *zilog) 1359{ 1360 mutex_enter(&zilog->zl_lock); 1361 ASSERT(zilog->zl_suspend != 0); 1362 zilog->zl_suspend--; 1363 mutex_exit(&zilog->zl_lock); 1364} 1365 1366typedef struct zil_replay_arg { 1367 objset_t *zr_os; 1368 zil_replay_func_t **zr_replay; 1369 void *zr_arg; 1370 uint64_t *zr_txgp; 1371 boolean_t zr_byteswap; 1372 char *zr_lrbuf; 1373} zil_replay_arg_t; 1374 1375static void 1376zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1377{ 1378 zil_replay_arg_t *zr = zra; 1379 const zil_header_t *zh = zilog->zl_header; 1380 uint64_t reclen = lr->lrc_reclen; 1381 uint64_t txtype = lr->lrc_txtype; 1382 char *name; 1383 int pass, error, sunk; 1384 1385 if (zilog->zl_stop_replay) 1386 return; 1387 1388 if (lr->lrc_txg < claim_txg) /* already committed */ 1389 return; 1390 1391 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1392 return; 1393 1394 /* 1395 * Make a copy of the data so we can revise and extend it. 1396 */ 1397 bcopy(lr, zr->zr_lrbuf, reclen); 1398 1399 /* 1400 * The log block containing this lr may have been byteswapped 1401 * so that we can easily examine common fields like lrc_txtype. 1402 * However, the log is a mix of different data types, and only the 1403 * replay vectors know how to byteswap their records. Therefore, if 1404 * the lr was byteswapped, undo it before invoking the replay vector. 1405 */ 1406 if (zr->zr_byteswap) 1407 byteswap_uint64_array(zr->zr_lrbuf, reclen); 1408 1409 /* 1410 * If this is a TX_WRITE with a blkptr, suck in the data. 1411 */ 1412 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1413 lr_write_t *lrw = (lr_write_t *)lr; 1414 blkptr_t *wbp = &lrw->lr_blkptr; 1415 uint64_t wlen = lrw->lr_length; 1416 char *wbuf = zr->zr_lrbuf + reclen; 1417 1418 if (BP_IS_HOLE(wbp)) { /* compressed to a hole */ 1419 bzero(wbuf, wlen); 1420 } else { 1421 /* 1422 * A subsequent write may have overwritten this block, 1423 * in which case wbp may have been been freed and 1424 * reallocated, and our read of wbp may fail with a 1425 * checksum error. We can safely ignore this because 1426 * the later write will provide the correct data. 1427 */ 1428 zbookmark_t zb; 1429 1430 zb.zb_objset = dmu_objset_id(zilog->zl_os); 1431 zb.zb_object = lrw->lr_foid; 1432 zb.zb_level = -1; 1433 zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp); 1434 1435 (void) zio_wait(zio_read(NULL, zilog->zl_spa, 1436 wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL, 1437 ZIO_PRIORITY_SYNC_READ, 1438 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb)); 1439 (void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen); 1440 } 1441 } 1442 1443 /* 1444 * We must now do two things atomically: replay this log record, 1445 * and update the log header to reflect the fact that we did so. 1446 * We use the DMU's ability to assign into a specific txg to do this. 1447 */ 1448 for (pass = 1, sunk = B_FALSE; /* CONSTANTCONDITION */; pass++) { 1449 uint64_t replay_txg; 1450 dmu_tx_t *replay_tx; 1451 1452 replay_tx = dmu_tx_create(zr->zr_os); 1453 error = dmu_tx_assign(replay_tx, TXG_WAIT); 1454 if (error) { 1455 dmu_tx_abort(replay_tx); 1456 break; 1457 } 1458 1459 replay_txg = dmu_tx_get_txg(replay_tx); 1460 1461 if (txtype == 0 || txtype >= TX_MAX_TYPE) { 1462 error = EINVAL; 1463 } else { 1464 /* 1465 * On the first pass, arrange for the replay vector 1466 * to fail its dmu_tx_assign(). That's the only way 1467 * to ensure that those code paths remain well tested. 1468 */ 1469 *zr->zr_txgp = replay_txg - (pass == 1); 1470 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf, 1471 zr->zr_byteswap); 1472 *zr->zr_txgp = TXG_NOWAIT; 1473 } 1474 1475 if (error == 0) { 1476 dsl_dataset_dirty(dmu_objset_ds(zr->zr_os), replay_tx); 1477 zilog->zl_replay_seq[replay_txg & TXG_MASK] = 1478 lr->lrc_seq; 1479 } 1480 1481 dmu_tx_commit(replay_tx); 1482 1483 if (!error) 1484 return; 1485 1486 /* 1487 * The DMU's dnode layer doesn't see removes until the txg 1488 * commits, so a subsequent claim can spuriously fail with 1489 * EEXIST. So if we receive any error other than ERESTART 1490 * we try syncing out any removes then retrying the 1491 * transaction. 1492 */ 1493 if (error != ERESTART && !sunk) { 1494 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1495 sunk = B_TRUE; 1496 continue; /* retry */ 1497 } 1498 1499 if (error != ERESTART) 1500 break; 1501 1502 if (pass != 1) 1503 txg_wait_open(spa_get_dsl(zilog->zl_spa), 1504 replay_txg + 1); 1505 1506 dprintf("pass %d, retrying\n", pass); 1507 } 1508 1509 ASSERT(error && error != ERESTART); 1510 name = kmem_alloc(MAXNAMELEN, KM_SLEEP); 1511 dmu_objset_name(zr->zr_os, name); 1512 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1513 "dataset %s, seq 0x%llx, txtype %llu\n", 1514 error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype); 1515 zilog->zl_stop_replay = 1; 1516 kmem_free(name, MAXNAMELEN); 1517} 1518 1519/* ARGSUSED */ 1520static void 1521zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 1522{ 1523 zilog->zl_replay_blks++; 1524} 1525 1526/* 1527 * If this dataset has a non-empty intent log, replay it and destroy it. 1528 */ 1529void 1530zil_replay(objset_t *os, void *arg, uint64_t *txgp, 1531 zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1532{ 1533 zilog_t *zilog = dmu_objset_zil(os); 1534 const zil_header_t *zh = zilog->zl_header; 1535 zil_replay_arg_t zr; 1536 1537 if (zil_empty(zilog)) { 1538 zil_destroy(zilog, B_TRUE); 1539 return; 1540 } 1541 //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 1542 1543 zr.zr_os = os; 1544 zr.zr_replay = replay_func; 1545 zr.zr_arg = arg; 1546 zr.zr_txgp = txgp; 1547 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1548 zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1549 1550 /* 1551 * Wait for in-progress removes to sync before starting replay. 1552 */ 1553 txg_wait_synced(zilog->zl_dmu_pool, 0); 1554 1555 zilog->zl_stop_replay = 0; 1556 zilog->zl_replay_time = lbolt; 1557 ASSERT(zilog->zl_replay_blks == 0); 1558 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1559 zh->zh_claim_txg); 1560 kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE); 1561 1562 zil_destroy(zilog, B_FALSE); 1563 //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 1564} 1565 1566/* 1567 * Report whether all transactions are committed 1568 */ 1569int 1570zil_is_committed(zilog_t *zilog) 1571{ 1572 lwb_t *lwb; 1573 int ret; 1574 1575 mutex_enter(&zilog->zl_lock); 1576 while (zilog->zl_writer) 1577 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1578 1579 /* recent unpushed intent log transactions? */ 1580 if (!list_is_empty(&zilog->zl_itx_list)) { 1581 ret = B_FALSE; 1582 goto out; 1583 } 1584 1585 /* intent log never used? */ 1586 lwb = list_head(&zilog->zl_lwb_list); 1587 if (lwb == NULL) { 1588 ret = B_TRUE; 1589 goto out; 1590 } 1591 1592 /* 1593 * more than 1 log buffer means zil_sync() hasn't yet freed 1594 * entries after a txg has committed 1595 */ 1596 if (list_next(&zilog->zl_lwb_list, lwb)) { 1597 ret = B_FALSE; 1598 goto out; 1599 } 1600 1601 ASSERT(zil_empty(zilog)); 1602 ret = B_TRUE; 1603out: 1604 cv_broadcast(&zilog->zl_cv_writer); 1605 mutex_exit(&zilog->zl_lock); 1606 return (ret); 1607} 1608