1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 */ 24 25 /* Portions Copyright 2010 Robert Milkowski */ 26 27 #include <sys/zfs_context.h> 28 #include <sys/spa.h> 29 #include <sys/dmu.h> 30 #include <sys/zap.h> 31 #include <sys/arc.h> 32 #include <sys/stat.h> 33 #include <sys/resource.h> 34 #include <sys/zil.h> 35 #include <sys/zil_impl.h> 36 #include <sys/dsl_dataset.h> 37 #include <sys/vdev.h> 38 #include <sys/dmu_tx.h> 39 40 /* 41 * The zfs intent log (ZIL) saves transaction records of system calls 42 * that change the file system in memory with enough information 43 * to be able to replay them. These are stored in memory until 44 * either the DMU transaction group (txg) commits them to the stable pool 45 * and they can be discarded, or they are flushed to the stable log 46 * (also in the pool) due to a fsync, O_DSYNC or other synchronous 47 * requirement. In the event of a panic or power fail then those log 48 * records (transactions) are replayed. 49 * 50 * There is one ZIL per file system. Its on-disk (pool) format consists 51 * of 3 parts: 52 * 53 * - ZIL header 54 * - ZIL blocks 55 * - ZIL records 56 * 57 * A log record holds a system call transaction. Log blocks can 58 * hold many log records and the blocks are chained together. 59 * Each ZIL block contains a block pointer (blkptr_t) to the next 60 * ZIL block in the chain. The ZIL header points to the first 61 * block in the chain. Note there is not a fixed place in the pool 62 * to hold blocks. They are dynamically allocated and freed as 63 * needed from the blocks available. Figure X shows the ZIL structure: 64 */ 65 66 /* 67 * This global ZIL switch affects all pools 68 */ 69 int zil_replay_disable = 0; /* disable intent logging replay */ 70 71 /* 72 * Tunable parameter for debugging or performance analysis. Setting 73 * zfs_nocacheflush will cause corruption on power loss if a volatile 74 * out-of-order write cache is enabled. 75 */ 76 boolean_t zfs_nocacheflush = B_FALSE; 77 78 static kmem_cache_t *zil_lwb_cache; 79 80 static boolean_t zil_empty(zilog_t *zilog); 81 82 #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 83 sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 84 85 86 static int 87 zil_bp_compare(const void *x1, const void *x2) 88 { 89 const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 90 const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 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 } 104 105 static void 106 zil_bp_tree_init(zilog_t *zilog) 107 { 108 avl_create(&zilog->zl_bp_tree, zil_bp_compare, 109 sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 110 } 111 112 static void 113 zil_bp_tree_fini(zilog_t *zilog) 114 { 115 avl_tree_t *t = &zilog->zl_bp_tree; 116 zil_bp_node_t *zn; 117 void *cookie = NULL; 118 119 while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 120 kmem_free(zn, sizeof (zil_bp_node_t)); 121 122 avl_destroy(t); 123 } 124 125 int 126 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 127 { 128 avl_tree_t *t = &zilog->zl_bp_tree; 129 const dva_t *dva = BP_IDENTITY(bp); 130 zil_bp_node_t *zn; 131 avl_index_t where; 132 133 if (avl_find(t, dva, &where) != NULL) 134 return (EEXIST); 135 136 zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 137 zn->zn_dva = *dva; 138 avl_insert(t, zn, where); 139 140 return (0); 141 } 142 143 static zil_header_t * 144 zil_header_in_syncing_context(zilog_t *zilog) 145 { 146 return ((zil_header_t *)zilog->zl_header); 147 } 148 149 static void 150 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 151 { 152 zio_cksum_t *zc = &bp->blk_cksum; 153 154 zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 155 zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 156 zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 157 zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 158 } 159 160 /* 161 * Read a log block and make sure it's valid. 162 */ 163 static int 164 zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 165 char **end) 166 { 167 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 168 uint32_t aflags = ARC_WAIT; 169 arc_buf_t *abuf = NULL; 170 zbookmark_t zb; 171 int error; 172 173 if (zilog->zl_header->zh_claim_txg == 0) 174 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 175 176 if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 177 zio_flags |= ZIO_FLAG_SPECULATIVE; 178 179 SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 180 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 181 182 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 183 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 184 185 if (error == 0) { 186 zio_cksum_t cksum = bp->blk_cksum; 187 188 /* 189 * Validate the checksummed log block. 190 * 191 * Sequence numbers should be... sequential. The checksum 192 * verifier for the next block should be bp's checksum plus 1. 193 * 194 * Also check the log chain linkage and size used. 195 */ 196 cksum.zc_word[ZIL_ZC_SEQ]++; 197 198 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 199 zil_chain_t *zilc = abuf->b_data; 200 char *lr = (char *)(zilc + 1); 201 uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 202 203 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 204 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 205 error = ECKSUM; 206 } else { 207 bcopy(lr, dst, len); 208 *end = (char *)dst + len; 209 *nbp = zilc->zc_next_blk; 210 } 211 } else { 212 char *lr = abuf->b_data; 213 uint64_t size = BP_GET_LSIZE(bp); 214 zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 215 216 if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 217 sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 218 (zilc->zc_nused > (size - sizeof (*zilc)))) { 219 error = ECKSUM; 220 } else { 221 bcopy(lr, dst, zilc->zc_nused); 222 *end = (char *)dst + zilc->zc_nused; 223 *nbp = zilc->zc_next_blk; 224 } 225 } 226 227 VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 228 } 229 230 return (error); 231 } 232 233 /* 234 * Read a TX_WRITE log data block. 235 */ 236 static int 237 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 238 { 239 enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 240 const blkptr_t *bp = &lr->lr_blkptr; 241 uint32_t aflags = ARC_WAIT; 242 arc_buf_t *abuf = NULL; 243 zbookmark_t zb; 244 int error; 245 246 if (BP_IS_HOLE(bp)) { 247 if (wbuf != NULL) 248 bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 249 return (0); 250 } 251 252 if (zilog->zl_header->zh_claim_txg == 0) 253 zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 254 255 SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 256 ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 257 258 error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 259 ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 260 261 if (error == 0) { 262 if (wbuf != NULL) 263 bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 264 (void) arc_buf_remove_ref(abuf, &abuf); 265 } 266 267 return (error); 268 } 269 270 /* 271 * Parse the intent log, and call parse_func for each valid record within. 272 */ 273 int 274 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 275 zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 276 { 277 const zil_header_t *zh = zilog->zl_header; 278 boolean_t claimed = !!zh->zh_claim_txg; 279 uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 280 uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 281 uint64_t max_blk_seq = 0; 282 uint64_t max_lr_seq = 0; 283 uint64_t blk_count = 0; 284 uint64_t lr_count = 0; 285 blkptr_t blk, next_blk; 286 char *lrbuf, *lrp; 287 int error = 0; 288 289 /* 290 * Old logs didn't record the maximum zh_claim_lr_seq. 291 */ 292 if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 293 claim_lr_seq = UINT64_MAX; 294 295 /* 296 * Starting at the block pointed to by zh_log we read the log chain. 297 * For each block in the chain we strongly check that block to 298 * ensure its validity. We stop when an invalid block is found. 299 * For each block pointer in the chain we call parse_blk_func(). 300 * For each record in each valid block we call parse_lr_func(). 301 * If the log has been claimed, stop if we encounter a sequence 302 * number greater than the highest claimed sequence number. 303 */ 304 lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 305 zil_bp_tree_init(zilog); 306 307 for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 308 uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 309 int reclen; 310 char *end; 311 312 if (blk_seq > claim_blk_seq) 313 break; 314 if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 315 break; 316 ASSERT3U(max_blk_seq, <, blk_seq); 317 max_blk_seq = blk_seq; 318 blk_count++; 319 320 if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 321 break; 322 323 error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 324 if (error) 325 break; 326 327 for (lrp = lrbuf; lrp < end; lrp += reclen) { 328 lr_t *lr = (lr_t *)lrp; 329 reclen = lr->lrc_reclen; 330 ASSERT3U(reclen, >=, sizeof (lr_t)); 331 if (lr->lrc_seq > claim_lr_seq) 332 goto done; 333 if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 334 goto done; 335 ASSERT3U(max_lr_seq, <, lr->lrc_seq); 336 max_lr_seq = lr->lrc_seq; 337 lr_count++; 338 } 339 } 340 done: 341 zilog->zl_parse_error = error; 342 zilog->zl_parse_blk_seq = max_blk_seq; 343 zilog->zl_parse_lr_seq = max_lr_seq; 344 zilog->zl_parse_blk_count = blk_count; 345 zilog->zl_parse_lr_count = lr_count; 346 347 ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 348 (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 349 350 zil_bp_tree_fini(zilog); 351 zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 352 353 return (error); 354 } 355 356 static int 357 zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 358 { 359 /* 360 * Claim log block if not already committed and not already claimed. 361 * If tx == NULL, just verify that the block is claimable. 362 */ 363 if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 364 return (0); 365 366 return (zio_wait(zio_claim(NULL, zilog->zl_spa, 367 tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 368 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 369 } 370 371 static int 372 zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 373 { 374 lr_write_t *lr = (lr_write_t *)lrc; 375 int error; 376 377 if (lrc->lrc_txtype != TX_WRITE) 378 return (0); 379 380 /* 381 * If the block is not readable, don't claim it. This can happen 382 * in normal operation when a log block is written to disk before 383 * some of the dmu_sync() blocks it points to. In this case, the 384 * transaction cannot have been committed to anyone (we would have 385 * waited for all writes to be stable first), so it is semantically 386 * correct to declare this the end of the log. 387 */ 388 if (lr->lr_blkptr.blk_birth >= first_txg && 389 (error = zil_read_log_data(zilog, lr, NULL)) != 0) 390 return (error); 391 return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 392 } 393 394 /* ARGSUSED */ 395 static int 396 zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 397 { 398 zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 399 400 return (0); 401 } 402 403 static int 404 zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 405 { 406 lr_write_t *lr = (lr_write_t *)lrc; 407 blkptr_t *bp = &lr->lr_blkptr; 408 409 /* 410 * If we previously claimed it, we need to free it. 411 */ 412 if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 413 bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 414 zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 415 416 return (0); 417 } 418 419 static lwb_t * 420 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 421 { 422 lwb_t *lwb; 423 424 lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 425 lwb->lwb_zilog = zilog; 426 lwb->lwb_blk = *bp; 427 lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 428 lwb->lwb_max_txg = txg; 429 lwb->lwb_zio = NULL; 430 lwb->lwb_tx = NULL; 431 if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 432 lwb->lwb_nused = sizeof (zil_chain_t); 433 lwb->lwb_sz = BP_GET_LSIZE(bp); 434 } else { 435 lwb->lwb_nused = 0; 436 lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 437 } 438 439 mutex_enter(&zilog->zl_lock); 440 list_insert_tail(&zilog->zl_lwb_list, lwb); 441 mutex_exit(&zilog->zl_lock); 442 443 return (lwb); 444 } 445 446 /* 447 * Create an on-disk intent log. 448 */ 449 static lwb_t * 450 zil_create(zilog_t *zilog) 451 { 452 const zil_header_t *zh = zilog->zl_header; 453 lwb_t *lwb = NULL; 454 uint64_t txg = 0; 455 dmu_tx_t *tx = NULL; 456 blkptr_t blk; 457 int error = 0; 458 459 /* 460 * Wait for any previous destroy to complete. 461 */ 462 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 463 464 ASSERT(zh->zh_claim_txg == 0); 465 ASSERT(zh->zh_replay_seq == 0); 466 467 blk = zh->zh_log; 468 469 /* 470 * Allocate an initial log block if: 471 * - there isn't one already 472 * - the existing block is the wrong endianess 473 */ 474 if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 475 tx = dmu_tx_create(zilog->zl_os); 476 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 477 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 478 txg = dmu_tx_get_txg(tx); 479 480 if (!BP_IS_HOLE(&blk)) { 481 zio_free_zil(zilog->zl_spa, txg, &blk); 482 BP_ZERO(&blk); 483 } 484 485 error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 486 ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 487 488 if (error == 0) 489 zil_init_log_chain(zilog, &blk); 490 } 491 492 /* 493 * Allocate a log write buffer (lwb) for the first log block. 494 */ 495 if (error == 0) 496 lwb = zil_alloc_lwb(zilog, &blk, txg); 497 498 /* 499 * If we just allocated the first log block, commit our transaction 500 * and wait for zil_sync() to stuff the block poiner into zh_log. 501 * (zh is part of the MOS, so we cannot modify it in open context.) 502 */ 503 if (tx != NULL) { 504 dmu_tx_commit(tx); 505 txg_wait_synced(zilog->zl_dmu_pool, txg); 506 } 507 508 ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 509 510 return (lwb); 511 } 512 513 /* 514 * In one tx, free all log blocks and clear the log header. 515 * If keep_first is set, then we're replaying a log with no content. 516 * We want to keep the first block, however, so that the first 517 * synchronous transaction doesn't require a txg_wait_synced() 518 * in zil_create(). We don't need to txg_wait_synced() here either 519 * when keep_first is set, because both zil_create() and zil_destroy() 520 * will wait for any in-progress destroys to complete. 521 */ 522 void 523 zil_destroy(zilog_t *zilog, boolean_t keep_first) 524 { 525 const zil_header_t *zh = zilog->zl_header; 526 lwb_t *lwb; 527 dmu_tx_t *tx; 528 uint64_t txg; 529 530 /* 531 * Wait for any previous destroy to complete. 532 */ 533 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 534 535 zilog->zl_old_header = *zh; /* debugging aid */ 536 537 if (BP_IS_HOLE(&zh->zh_log)) 538 return; 539 540 tx = dmu_tx_create(zilog->zl_os); 541 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 542 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 543 txg = dmu_tx_get_txg(tx); 544 545 mutex_enter(&zilog->zl_lock); 546 547 ASSERT3U(zilog->zl_destroy_txg, <, txg); 548 zilog->zl_destroy_txg = txg; 549 zilog->zl_keep_first = keep_first; 550 551 if (!list_is_empty(&zilog->zl_lwb_list)) { 552 ASSERT(zh->zh_claim_txg == 0); 553 ASSERT(!keep_first); 554 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 555 list_remove(&zilog->zl_lwb_list, lwb); 556 if (lwb->lwb_buf != NULL) 557 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 558 zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 559 kmem_cache_free(zil_lwb_cache, lwb); 560 } 561 } else if (!keep_first) { 562 (void) zil_parse(zilog, zil_free_log_block, 563 zil_free_log_record, tx, zh->zh_claim_txg); 564 } 565 mutex_exit(&zilog->zl_lock); 566 567 dmu_tx_commit(tx); 568 } 569 570 int 571 zil_claim(const char *osname, void *txarg) 572 { 573 dmu_tx_t *tx = txarg; 574 uint64_t first_txg = dmu_tx_get_txg(tx); 575 zilog_t *zilog; 576 zil_header_t *zh; 577 objset_t *os; 578 int error; 579 580 error = dmu_objset_hold(osname, FTAG, &os); 581 if (error) { 582 cmn_err(CE_WARN, "can't open objset for %s", osname); 583 return (0); 584 } 585 586 zilog = dmu_objset_zil(os); 587 zh = zil_header_in_syncing_context(zilog); 588 589 if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 590 if (!BP_IS_HOLE(&zh->zh_log)) 591 zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 592 BP_ZERO(&zh->zh_log); 593 dsl_dataset_dirty(dmu_objset_ds(os), tx); 594 dmu_objset_rele(os, FTAG); 595 return (0); 596 } 597 598 /* 599 * Claim all log blocks if we haven't already done so, and remember 600 * the highest claimed sequence number. This ensures that if we can 601 * read only part of the log now (e.g. due to a missing device), 602 * but we can read the entire log later, we will not try to replay 603 * or destroy beyond the last block we successfully claimed. 604 */ 605 ASSERT3U(zh->zh_claim_txg, <=, first_txg); 606 if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 607 (void) zil_parse(zilog, zil_claim_log_block, 608 zil_claim_log_record, tx, first_txg); 609 zh->zh_claim_txg = first_txg; 610 zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 611 zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 612 if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 613 zh->zh_flags |= ZIL_REPLAY_NEEDED; 614 zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 615 dsl_dataset_dirty(dmu_objset_ds(os), tx); 616 } 617 618 ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 619 dmu_objset_rele(os, FTAG); 620 return (0); 621 } 622 623 /* 624 * Check the log by walking the log chain. 625 * Checksum errors are ok as they indicate the end of the chain. 626 * Any other error (no device or read failure) returns an error. 627 */ 628 int 629 zil_check_log_chain(const char *osname, void *tx) 630 { 631 zilog_t *zilog; 632 objset_t *os; 633 int error; 634 635 ASSERT(tx == NULL); 636 637 error = dmu_objset_hold(osname, FTAG, &os); 638 if (error) { 639 cmn_err(CE_WARN, "can't open objset for %s", osname); 640 return (0); 641 } 642 643 zilog = dmu_objset_zil(os); 644 645 /* 646 * Because tx == NULL, zil_claim_log_block() will not actually claim 647 * any blocks, but just determine whether it is possible to do so. 648 * In addition to checking the log chain, zil_claim_log_block() 649 * will invoke zio_claim() with a done func of spa_claim_notify(), 650 * which will update spa_max_claim_txg. See spa_load() for details. 651 */ 652 error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 653 zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 654 655 dmu_objset_rele(os, FTAG); 656 657 return ((error == ECKSUM || error == ENOENT) ? 0 : error); 658 } 659 660 static int 661 zil_vdev_compare(const void *x1, const void *x2) 662 { 663 uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 664 uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 665 666 if (v1 < v2) 667 return (-1); 668 if (v1 > v2) 669 return (1); 670 671 return (0); 672 } 673 674 void 675 zil_add_block(zilog_t *zilog, const blkptr_t *bp) 676 { 677 avl_tree_t *t = &zilog->zl_vdev_tree; 678 avl_index_t where; 679 zil_vdev_node_t *zv, zvsearch; 680 int ndvas = BP_GET_NDVAS(bp); 681 int i; 682 683 if (zfs_nocacheflush) 684 return; 685 686 ASSERT(zilog->zl_writer); 687 688 /* 689 * Even though we're zl_writer, we still need a lock because the 690 * zl_get_data() callbacks may have dmu_sync() done callbacks 691 * that will run concurrently. 692 */ 693 mutex_enter(&zilog->zl_vdev_lock); 694 for (i = 0; i < ndvas; i++) { 695 zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 696 if (avl_find(t, &zvsearch, &where) == NULL) { 697 zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 698 zv->zv_vdev = zvsearch.zv_vdev; 699 avl_insert(t, zv, where); 700 } 701 } 702 mutex_exit(&zilog->zl_vdev_lock); 703 } 704 705 void 706 zil_flush_vdevs(zilog_t *zilog) 707 { 708 spa_t *spa = zilog->zl_spa; 709 avl_tree_t *t = &zilog->zl_vdev_tree; 710 void *cookie = NULL; 711 zil_vdev_node_t *zv; 712 zio_t *zio; 713 714 ASSERT(zilog->zl_writer); 715 716 /* 717 * We don't need zl_vdev_lock here because we're the zl_writer, 718 * and all zl_get_data() callbacks are done. 719 */ 720 if (avl_numnodes(t) == 0) 721 return; 722 723 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 724 725 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 726 727 while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 728 vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 729 if (vd != NULL) 730 zio_flush(zio, vd); 731 kmem_free(zv, sizeof (*zv)); 732 } 733 734 /* 735 * Wait for all the flushes to complete. Not all devices actually 736 * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 737 */ 738 (void) zio_wait(zio); 739 740 spa_config_exit(spa, SCL_STATE, FTAG); 741 } 742 743 /* 744 * Function called when a log block write completes 745 */ 746 static void 747 zil_lwb_write_done(zio_t *zio) 748 { 749 lwb_t *lwb = zio->io_private; 750 zilog_t *zilog = lwb->lwb_zilog; 751 dmu_tx_t *tx = lwb->lwb_tx; 752 753 ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 754 ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 755 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 756 ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 757 ASSERT(!BP_IS_GANG(zio->io_bp)); 758 ASSERT(!BP_IS_HOLE(zio->io_bp)); 759 ASSERT(zio->io_bp->blk_fill == 0); 760 761 /* 762 * Ensure the lwb buffer pointer is cleared before releasing 763 * the txg. If we have had an allocation failure and 764 * the txg is waiting to sync then we want want zil_sync() 765 * to remove the lwb so that it's not picked up as the next new 766 * one in zil_commit_writer(). zil_sync() will only remove 767 * the lwb if lwb_buf is null. 768 */ 769 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 770 mutex_enter(&zilog->zl_lock); 771 lwb->lwb_buf = NULL; 772 lwb->lwb_tx = NULL; 773 mutex_exit(&zilog->zl_lock); 774 775 /* 776 * Now that we've written this log block, we have a stable pointer 777 * to the next block in the chain, so it's OK to let the txg in 778 * which we allocated the next block sync. 779 */ 780 dmu_tx_commit(tx); 781 } 782 783 /* 784 * Initialize the io for a log block. 785 */ 786 static void 787 zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 788 { 789 zbookmark_t zb; 790 791 SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 792 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 793 lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 794 795 if (zilog->zl_root_zio == NULL) { 796 zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 797 ZIO_FLAG_CANFAIL); 798 } 799 if (lwb->lwb_zio == NULL) { 800 lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 801 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 802 zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 803 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 804 } 805 } 806 807 /* 808 * Define a limited set of intent log block sizes. 809 * These must be a multiple of 4KB. Note only the amount used (again 810 * aligned to 4KB) actually gets written. However, we can't always just 811 * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 812 */ 813 uint64_t zil_block_buckets[] = { 814 4096, /* non TX_WRITE */ 815 8192+4096, /* data base */ 816 32*1024 + 4096, /* NFS writes */ 817 UINT64_MAX 818 }; 819 820 /* 821 * Use the slog as long as the logbias is 'latency' and the current commit size 822 * is less than the limit or the total list size is less than 2X the limit. 823 * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 824 */ 825 uint64_t zil_slog_limit = 1024 * 1024; 826 #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 827 (((zilog)->zl_cur_used < zil_slog_limit) || \ 828 ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 829 830 /* 831 * Start a log block write and advance to the next log block. 832 * Calls are serialized. 833 */ 834 static lwb_t * 835 zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 836 { 837 lwb_t *nlwb = NULL; 838 zil_chain_t *zilc; 839 spa_t *spa = zilog->zl_spa; 840 blkptr_t *bp; 841 dmu_tx_t *tx; 842 uint64_t txg; 843 uint64_t zil_blksz, wsz; 844 int i, error; 845 846 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 847 zilc = (zil_chain_t *)lwb->lwb_buf; 848 bp = &zilc->zc_next_blk; 849 } else { 850 zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 851 bp = &zilc->zc_next_blk; 852 } 853 854 ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 855 856 /* 857 * Allocate the next block and save its address in this block 858 * before writing it in order to establish the log chain. 859 * Note that if the allocation of nlwb synced before we wrote 860 * the block that points at it (lwb), we'd leak it if we crashed. 861 * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 862 * We dirty the dataset to ensure that zil_sync() will be called 863 * to clean up in the event of allocation failure or I/O failure. 864 */ 865 tx = dmu_tx_create(zilog->zl_os); 866 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 867 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 868 txg = dmu_tx_get_txg(tx); 869 870 lwb->lwb_tx = tx; 871 872 /* 873 * Log blocks are pre-allocated. Here we select the size of the next 874 * block, based on size used in the last block. 875 * - first find the smallest bucket that will fit the block from a 876 * limited set of block sizes. This is because it's faster to write 877 * blocks allocated from the same metaslab as they are adjacent or 878 * close. 879 * - next find the maximum from the new suggested size and an array of 880 * previous sizes. This lessens a picket fence effect of wrongly 881 * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 882 * requests. 883 * 884 * Note we only write what is used, but we can't just allocate 885 * the maximum block size because we can exhaust the available 886 * pool log space. 887 */ 888 zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 889 for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 890 continue; 891 zil_blksz = zil_block_buckets[i]; 892 if (zil_blksz == UINT64_MAX) 893 zil_blksz = SPA_MAXBLOCKSIZE; 894 zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 895 for (i = 0; i < ZIL_PREV_BLKS; i++) 896 zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 897 zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 898 899 BP_ZERO(bp); 900 /* pass the old blkptr in order to spread log blocks across devs */ 901 error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 902 USE_SLOG(zilog)); 903 if (!error) { 904 ASSERT3U(bp->blk_birth, ==, txg); 905 bp->blk_cksum = lwb->lwb_blk.blk_cksum; 906 bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 907 908 /* 909 * Allocate a new log write buffer (lwb). 910 */ 911 nlwb = zil_alloc_lwb(zilog, bp, txg); 912 913 /* Record the block for later vdev flushing */ 914 zil_add_block(zilog, &lwb->lwb_blk); 915 } 916 917 if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 918 /* For Slim ZIL only write what is used. */ 919 wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 920 ASSERT3U(wsz, <=, lwb->lwb_sz); 921 zio_shrink(lwb->lwb_zio, wsz); 922 923 } else { 924 wsz = lwb->lwb_sz; 925 } 926 927 zilc->zc_pad = 0; 928 zilc->zc_nused = lwb->lwb_nused; 929 zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 930 931 /* 932 * clear unused data for security 933 */ 934 bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 935 936 zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 937 938 /* 939 * If there was an allocation failure then nlwb will be null which 940 * forces a txg_wait_synced(). 941 */ 942 return (nlwb); 943 } 944 945 static lwb_t * 946 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 947 { 948 lr_t *lrc = &itx->itx_lr; /* common log record */ 949 lr_write_t *lrw = (lr_write_t *)lrc; 950 char *lr_buf; 951 uint64_t txg = lrc->lrc_txg; 952 uint64_t reclen = lrc->lrc_reclen; 953 uint64_t dlen = 0; 954 955 if (lwb == NULL) 956 return (NULL); 957 958 ASSERT(lwb->lwb_buf != NULL); 959 960 if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 961 dlen = P2ROUNDUP_TYPED( 962 lrw->lr_length, sizeof (uint64_t), uint64_t); 963 964 zilog->zl_cur_used += (reclen + dlen); 965 966 zil_lwb_write_init(zilog, lwb); 967 968 /* 969 * If this record won't fit in the current log block, start a new one. 970 */ 971 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 972 lwb = zil_lwb_write_start(zilog, lwb); 973 if (lwb == NULL) 974 return (NULL); 975 zil_lwb_write_init(zilog, lwb); 976 ASSERT(LWB_EMPTY(lwb)); 977 if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 978 txg_wait_synced(zilog->zl_dmu_pool, txg); 979 return (lwb); 980 } 981 } 982 983 lr_buf = lwb->lwb_buf + lwb->lwb_nused; 984 bcopy(lrc, lr_buf, reclen); 985 lrc = (lr_t *)lr_buf; 986 lrw = (lr_write_t *)lrc; 987 988 /* 989 * If it's a write, fetch the data or get its blkptr as appropriate. 990 */ 991 if (lrc->lrc_txtype == TX_WRITE) { 992 if (txg > spa_freeze_txg(zilog->zl_spa)) 993 txg_wait_synced(zilog->zl_dmu_pool, txg); 994 if (itx->itx_wr_state != WR_COPIED) { 995 char *dbuf; 996 int error; 997 998 if (dlen) { 999 ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1000 dbuf = lr_buf + reclen; 1001 lrw->lr_common.lrc_reclen += dlen; 1002 } else { 1003 ASSERT(itx->itx_wr_state == WR_INDIRECT); 1004 dbuf = NULL; 1005 } 1006 error = zilog->zl_get_data( 1007 itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1008 if (error == EIO) { 1009 txg_wait_synced(zilog->zl_dmu_pool, txg); 1010 return (lwb); 1011 } 1012 if (error) { 1013 ASSERT(error == ENOENT || error == EEXIST || 1014 error == EALREADY); 1015 return (lwb); 1016 } 1017 } 1018 } 1019 1020 /* 1021 * We're actually making an entry, so update lrc_seq to be the 1022 * log record sequence number. Note that this is generally not 1023 * equal to the itx sequence number because not all transactions 1024 * are synchronous, and sometimes spa_sync() gets there first. 1025 */ 1026 lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1027 lwb->lwb_nused += reclen + dlen; 1028 lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1029 ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1030 ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 1031 1032 return (lwb); 1033 } 1034 1035 itx_t * 1036 zil_itx_create(uint64_t txtype, size_t lrsize) 1037 { 1038 itx_t *itx; 1039 1040 lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1041 1042 itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1043 itx->itx_lr.lrc_txtype = txtype; 1044 itx->itx_lr.lrc_reclen = lrsize; 1045 itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1046 itx->itx_lr.lrc_seq = 0; /* defensive */ 1047 1048 return (itx); 1049 } 1050 1051 void 1052 zil_itx_destroy(itx_t *itx) 1053 { 1054 kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1055 } 1056 1057 uint64_t 1058 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1059 { 1060 uint64_t seq; 1061 1062 ASSERT(itx->itx_lr.lrc_seq == 0); 1063 ASSERT(!zilog->zl_replay); 1064 1065 mutex_enter(&zilog->zl_lock); 1066 list_insert_tail(&zilog->zl_itx_list, itx); 1067 zilog->zl_itx_list_sz += itx->itx_sod; 1068 itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1069 itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 1070 mutex_exit(&zilog->zl_lock); 1071 1072 return (seq); 1073 } 1074 1075 /* 1076 * Free up all in-memory intent log transactions that have now been synced. 1077 */ 1078 static void 1079 zil_itx_clean(zilog_t *zilog) 1080 { 1081 uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 1082 uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 1083 list_t clean_list; 1084 itx_t *itx; 1085 1086 list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1087 1088 mutex_enter(&zilog->zl_lock); 1089 /* wait for a log writer to finish walking list */ 1090 while (zilog->zl_writer) { 1091 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1092 } 1093 1094 /* 1095 * Move the sync'd log transactions to a separate list so we can call 1096 * kmem_free without holding the zl_lock. 1097 * 1098 * There is no need to set zl_writer as we don't drop zl_lock here 1099 */ 1100 while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 1101 itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 1102 list_remove(&zilog->zl_itx_list, itx); 1103 zilog->zl_itx_list_sz -= itx->itx_sod; 1104 list_insert_tail(&clean_list, itx); 1105 } 1106 cv_broadcast(&zilog->zl_cv_writer); 1107 mutex_exit(&zilog->zl_lock); 1108 1109 /* destroy sync'd log transactions */ 1110 while ((itx = list_head(&clean_list)) != NULL) { 1111 list_remove(&clean_list, itx); 1112 zil_itx_destroy(itx); 1113 } 1114 list_destroy(&clean_list); 1115 } 1116 1117 /* 1118 * If there are any in-memory intent log transactions which have now been 1119 * synced then start up a taskq to free them. 1120 */ 1121 void 1122 zil_clean(zilog_t *zilog) 1123 { 1124 itx_t *itx; 1125 1126 mutex_enter(&zilog->zl_lock); 1127 itx = list_head(&zilog->zl_itx_list); 1128 if ((itx != NULL) && 1129 (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 1130 (void) taskq_dispatch(zilog->zl_clean_taskq, 1131 (task_func_t *)zil_itx_clean, zilog, TQ_NOSLEEP); 1132 } 1133 mutex_exit(&zilog->zl_lock); 1134 } 1135 1136 static void 1137 zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 1138 { 1139 uint64_t txg; 1140 uint64_t commit_seq = 0; 1141 itx_t *itx, *itx_next; 1142 lwb_t *lwb; 1143 spa_t *spa; 1144 int error = 0; 1145 1146 zilog->zl_writer = B_TRUE; 1147 ASSERT(zilog->zl_root_zio == NULL); 1148 spa = zilog->zl_spa; 1149 1150 if (zilog->zl_suspend) { 1151 lwb = NULL; 1152 } else { 1153 lwb = list_tail(&zilog->zl_lwb_list); 1154 if (lwb == NULL) { 1155 /* 1156 * Return if there's nothing to flush before we 1157 * dirty the fs by calling zil_create() 1158 */ 1159 if (list_is_empty(&zilog->zl_itx_list)) { 1160 zilog->zl_writer = B_FALSE; 1161 return; 1162 } 1163 mutex_exit(&zilog->zl_lock); 1164 lwb = zil_create(zilog); 1165 mutex_enter(&zilog->zl_lock); 1166 } 1167 } 1168 ASSERT(lwb == NULL || lwb->lwb_zio == NULL); 1169 1170 /* Loop through in-memory log transactions filling log blocks. */ 1171 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1172 1173 for (itx = list_head(&zilog->zl_itx_list); itx; itx = itx_next) { 1174 /* 1175 * Save the next pointer. Even though we drop zl_lock below, 1176 * all threads that can remove itx list entries (other writers 1177 * and zil_itx_clean()) can't do so until they have zl_writer. 1178 */ 1179 itx_next = list_next(&zilog->zl_itx_list, itx); 1180 1181 /* 1182 * Determine whether to push this itx. 1183 * Push all transactions related to specified foid and 1184 * all other transactions except those that can be logged 1185 * out of order (TX_WRITE, TX_TRUNCATE, TX_SETATTR, TX_ACL) 1186 * for all other files. 1187 * 1188 * If foid == 0 (meaning "push all foids") or 1189 * itx->itx_sync is set (meaning O_[D]SYNC), push regardless. 1190 */ 1191 if (foid != 0 && !itx->itx_sync && 1192 TX_OOO(itx->itx_lr.lrc_txtype) && 1193 ((lr_ooo_t *)&itx->itx_lr)->lr_foid != foid) 1194 continue; /* skip this record */ 1195 1196 if ((itx->itx_lr.lrc_seq > seq) && 1197 ((lwb == NULL) || (LWB_EMPTY(lwb)) || 1198 (lwb->lwb_nused + itx->itx_sod > lwb->lwb_sz))) 1199 break; 1200 1201 list_remove(&zilog->zl_itx_list, itx); 1202 zilog->zl_itx_list_sz -= itx->itx_sod; 1203 1204 mutex_exit(&zilog->zl_lock); 1205 1206 txg = itx->itx_lr.lrc_txg; 1207 ASSERT(txg); 1208 1209 if (txg > spa_last_synced_txg(spa) || 1210 txg > spa_freeze_txg(spa)) 1211 lwb = zil_lwb_commit(zilog, itx, lwb); 1212 1213 zil_itx_destroy(itx); 1214 1215 mutex_enter(&zilog->zl_lock); 1216 } 1217 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1218 /* determine commit sequence number */ 1219 itx = list_head(&zilog->zl_itx_list); 1220 if (itx) 1221 commit_seq = itx->itx_lr.lrc_seq - 1; 1222 else 1223 commit_seq = zilog->zl_itx_seq; 1224 mutex_exit(&zilog->zl_lock); 1225 1226 /* write the last block out */ 1227 if (lwb != NULL && lwb->lwb_zio != NULL) 1228 lwb = zil_lwb_write_start(zilog, lwb); 1229 1230 zilog->zl_prev_used = zilog->zl_cur_used; 1231 zilog->zl_cur_used = 0; 1232 1233 /* 1234 * Wait if necessary for the log blocks to be on stable storage. 1235 */ 1236 if (zilog->zl_root_zio) { 1237 DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1238 error = zio_wait(zilog->zl_root_zio); 1239 zilog->zl_root_zio = NULL; 1240 DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 1241 zil_flush_vdevs(zilog); 1242 } 1243 1244 if (error || lwb == NULL) 1245 txg_wait_synced(zilog->zl_dmu_pool, 0); 1246 1247 mutex_enter(&zilog->zl_lock); 1248 zilog->zl_writer = B_FALSE; 1249 1250 ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 1251 zilog->zl_commit_seq = commit_seq; 1252 1253 /* 1254 * Remember the highest committed log sequence number for ztest. 1255 * We only update this value when all the log writes succeeded, 1256 * because ztest wants to ASSERT that it got the whole log chain. 1257 */ 1258 if (error == 0 && lwb != NULL) 1259 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1260 } 1261 1262 /* 1263 * Push zfs transactions to stable storage up to the supplied sequence number. 1264 * If foid is 0 push out all transactions, otherwise push only those 1265 * for that file or might have been used to create that file. 1266 */ 1267 void 1268 zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1269 { 1270 if (zilog->zl_sync == ZFS_SYNC_DISABLED || seq == 0) 1271 return; 1272 1273 mutex_enter(&zilog->zl_lock); 1274 1275 seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1276 1277 while (zilog->zl_writer) { 1278 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1279 if (seq <= zilog->zl_commit_seq) { 1280 mutex_exit(&zilog->zl_lock); 1281 return; 1282 } 1283 } 1284 zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 1285 /* wake up others waiting on the commit */ 1286 cv_broadcast(&zilog->zl_cv_writer); 1287 mutex_exit(&zilog->zl_lock); 1288 } 1289 1290 /* 1291 * Report whether all transactions are committed. 1292 */ 1293 static boolean_t 1294 zil_is_committed(zilog_t *zilog) 1295 { 1296 lwb_t *lwb; 1297 boolean_t committed; 1298 1299 mutex_enter(&zilog->zl_lock); 1300 1301 while (zilog->zl_writer) 1302 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1303 1304 if (!list_is_empty(&zilog->zl_itx_list)) 1305 committed = B_FALSE; /* unpushed transactions */ 1306 else if ((lwb = list_head(&zilog->zl_lwb_list)) == NULL) 1307 committed = B_TRUE; /* intent log never used */ 1308 else if (list_next(&zilog->zl_lwb_list, lwb) != NULL) 1309 committed = B_FALSE; /* zil_sync() not done yet */ 1310 else 1311 committed = B_TRUE; /* everything synced */ 1312 1313 mutex_exit(&zilog->zl_lock); 1314 return (committed); 1315 } 1316 1317 /* 1318 * Called in syncing context to free committed log blocks and update log header. 1319 */ 1320 void 1321 zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1322 { 1323 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1324 uint64_t txg = dmu_tx_get_txg(tx); 1325 spa_t *spa = zilog->zl_spa; 1326 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1327 lwb_t *lwb; 1328 1329 /* 1330 * We don't zero out zl_destroy_txg, so make sure we don't try 1331 * to destroy it twice. 1332 */ 1333 if (spa_sync_pass(spa) != 1) 1334 return; 1335 1336 mutex_enter(&zilog->zl_lock); 1337 1338 ASSERT(zilog->zl_stop_sync == 0); 1339 1340 if (*replayed_seq != 0) { 1341 ASSERT(zh->zh_replay_seq < *replayed_seq); 1342 zh->zh_replay_seq = *replayed_seq; 1343 *replayed_seq = 0; 1344 } 1345 1346 if (zilog->zl_destroy_txg == txg) { 1347 blkptr_t blk = zh->zh_log; 1348 1349 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1350 1351 bzero(zh, sizeof (zil_header_t)); 1352 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1353 1354 if (zilog->zl_keep_first) { 1355 /* 1356 * If this block was part of log chain that couldn't 1357 * be claimed because a device was missing during 1358 * zil_claim(), but that device later returns, 1359 * then this block could erroneously appear valid. 1360 * To guard against this, assign a new GUID to the new 1361 * log chain so it doesn't matter what blk points to. 1362 */ 1363 zil_init_log_chain(zilog, &blk); 1364 zh->zh_log = blk; 1365 } 1366 } 1367 1368 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1369 zh->zh_log = lwb->lwb_blk; 1370 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1371 break; 1372 list_remove(&zilog->zl_lwb_list, lwb); 1373 zio_free_zil(spa, txg, &lwb->lwb_blk); 1374 kmem_cache_free(zil_lwb_cache, lwb); 1375 1376 /* 1377 * If we don't have anything left in the lwb list then 1378 * we've had an allocation failure and we need to zero 1379 * out the zil_header blkptr so that we don't end 1380 * up freeing the same block twice. 1381 */ 1382 if (list_head(&zilog->zl_lwb_list) == NULL) 1383 BP_ZERO(&zh->zh_log); 1384 } 1385 mutex_exit(&zilog->zl_lock); 1386 } 1387 1388 void 1389 zil_init(void) 1390 { 1391 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1392 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1393 } 1394 1395 void 1396 zil_fini(void) 1397 { 1398 kmem_cache_destroy(zil_lwb_cache); 1399 } 1400 1401 void 1402 zil_set_sync(zilog_t *zilog, uint64_t sync) 1403 { 1404 zilog->zl_sync = sync; 1405 } 1406 1407 void 1408 zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1409 { 1410 zilog->zl_logbias = logbias; 1411 } 1412 1413 zilog_t * 1414 zil_alloc(objset_t *os, zil_header_t *zh_phys) 1415 { 1416 zilog_t *zilog; 1417 1418 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1419 1420 zilog->zl_header = zh_phys; 1421 zilog->zl_os = os; 1422 zilog->zl_spa = dmu_objset_spa(os); 1423 zilog->zl_dmu_pool = dmu_objset_pool(os); 1424 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1425 zilog->zl_logbias = dmu_objset_logbias(os); 1426 zilog->zl_sync = dmu_objset_syncprop(os); 1427 1428 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1429 1430 list_create(&zilog->zl_itx_list, sizeof (itx_t), 1431 offsetof(itx_t, itx_node)); 1432 1433 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1434 offsetof(lwb_t, lwb_node)); 1435 1436 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1437 1438 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1439 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1440 1441 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1442 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1443 1444 return (zilog); 1445 } 1446 1447 void 1448 zil_free(zilog_t *zilog) 1449 { 1450 lwb_t *lwb; 1451 1452 zilog->zl_stop_sync = 1; 1453 1454 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1455 list_remove(&zilog->zl_lwb_list, lwb); 1456 if (lwb->lwb_buf != NULL) 1457 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1458 kmem_cache_free(zil_lwb_cache, lwb); 1459 } 1460 list_destroy(&zilog->zl_lwb_list); 1461 1462 avl_destroy(&zilog->zl_vdev_tree); 1463 mutex_destroy(&zilog->zl_vdev_lock); 1464 1465 ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1466 list_destroy(&zilog->zl_itx_list); 1467 mutex_destroy(&zilog->zl_lock); 1468 1469 cv_destroy(&zilog->zl_cv_writer); 1470 cv_destroy(&zilog->zl_cv_suspend); 1471 1472 kmem_free(zilog, sizeof (zilog_t)); 1473 } 1474 1475 /* 1476 * Open an intent log. 1477 */ 1478 zilog_t * 1479 zil_open(objset_t *os, zil_get_data_t *get_data) 1480 { 1481 zilog_t *zilog = dmu_objset_zil(os); 1482 1483 zilog->zl_get_data = get_data; 1484 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1485 2, 2, TASKQ_PREPOPULATE); 1486 1487 return (zilog); 1488 } 1489 1490 /* 1491 * Close an intent log. 1492 */ 1493 void 1494 zil_close(zilog_t *zilog) 1495 { 1496 /* 1497 * If the log isn't already committed, mark the objset dirty 1498 * (so zil_sync() will be called) and wait for that txg to sync. 1499 */ 1500 if (!zil_is_committed(zilog)) { 1501 uint64_t txg; 1502 dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1503 VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 1504 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1505 txg = dmu_tx_get_txg(tx); 1506 dmu_tx_commit(tx); 1507 txg_wait_synced(zilog->zl_dmu_pool, txg); 1508 } 1509 1510 taskq_destroy(zilog->zl_clean_taskq); 1511 zilog->zl_clean_taskq = NULL; 1512 zilog->zl_get_data = NULL; 1513 1514 zil_itx_clean(zilog); 1515 ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1516 } 1517 1518 /* 1519 * Suspend an intent log. While in suspended mode, we still honor 1520 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1521 * We suspend the log briefly when taking a snapshot so that the snapshot 1522 * contains all the data it's supposed to, and has an empty intent log. 1523 */ 1524 int 1525 zil_suspend(zilog_t *zilog) 1526 { 1527 const zil_header_t *zh = zilog->zl_header; 1528 1529 mutex_enter(&zilog->zl_lock); 1530 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1531 mutex_exit(&zilog->zl_lock); 1532 return (EBUSY); 1533 } 1534 if (zilog->zl_suspend++ != 0) { 1535 /* 1536 * Someone else already began a suspend. 1537 * Just wait for them to finish. 1538 */ 1539 while (zilog->zl_suspending) 1540 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1541 mutex_exit(&zilog->zl_lock); 1542 return (0); 1543 } 1544 zilog->zl_suspending = B_TRUE; 1545 mutex_exit(&zilog->zl_lock); 1546 1547 zil_commit(zilog, UINT64_MAX, 0); 1548 1549 /* 1550 * Wait for any in-flight log writes to complete. 1551 */ 1552 mutex_enter(&zilog->zl_lock); 1553 while (zilog->zl_writer) 1554 cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1555 mutex_exit(&zilog->zl_lock); 1556 1557 zil_destroy(zilog, B_FALSE); 1558 1559 mutex_enter(&zilog->zl_lock); 1560 zilog->zl_suspending = B_FALSE; 1561 cv_broadcast(&zilog->zl_cv_suspend); 1562 mutex_exit(&zilog->zl_lock); 1563 1564 return (0); 1565 } 1566 1567 void 1568 zil_resume(zilog_t *zilog) 1569 { 1570 mutex_enter(&zilog->zl_lock); 1571 ASSERT(zilog->zl_suspend != 0); 1572 zilog->zl_suspend--; 1573 mutex_exit(&zilog->zl_lock); 1574 } 1575 1576 typedef struct zil_replay_arg { 1577 zil_replay_func_t **zr_replay; 1578 void *zr_arg; 1579 boolean_t zr_byteswap; 1580 char *zr_lr; 1581 } zil_replay_arg_t; 1582 1583 static int 1584 zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1585 { 1586 char name[MAXNAMELEN]; 1587 1588 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1589 1590 dmu_objset_name(zilog->zl_os, name); 1591 1592 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1593 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1594 (u_longlong_t)lr->lrc_seq, 1595 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1596 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1597 1598 return (error); 1599 } 1600 1601 static int 1602 zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1603 { 1604 zil_replay_arg_t *zr = zra; 1605 const zil_header_t *zh = zilog->zl_header; 1606 uint64_t reclen = lr->lrc_reclen; 1607 uint64_t txtype = lr->lrc_txtype; 1608 int error = 0; 1609 1610 zilog->zl_replaying_seq = lr->lrc_seq; 1611 1612 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1613 return (0); 1614 1615 if (lr->lrc_txg < claim_txg) /* already committed */ 1616 return (0); 1617 1618 /* Strip case-insensitive bit, still present in log record */ 1619 txtype &= ~TX_CI; 1620 1621 if (txtype == 0 || txtype >= TX_MAX_TYPE) 1622 return (zil_replay_error(zilog, lr, EINVAL)); 1623 1624 /* 1625 * If this record type can be logged out of order, the object 1626 * (lr_foid) may no longer exist. That's legitimate, not an error. 1627 */ 1628 if (TX_OOO(txtype)) { 1629 error = dmu_object_info(zilog->zl_os, 1630 ((lr_ooo_t *)lr)->lr_foid, NULL); 1631 if (error == ENOENT || error == EEXIST) 1632 return (0); 1633 } 1634 1635 /* 1636 * Make a copy of the data so we can revise and extend it. 1637 */ 1638 bcopy(lr, zr->zr_lr, reclen); 1639 1640 /* 1641 * If this is a TX_WRITE with a blkptr, suck in the data. 1642 */ 1643 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1644 error = zil_read_log_data(zilog, (lr_write_t *)lr, 1645 zr->zr_lr + reclen); 1646 if (error) 1647 return (zil_replay_error(zilog, lr, error)); 1648 } 1649 1650 /* 1651 * The log block containing this lr may have been byteswapped 1652 * so that we can easily examine common fields like lrc_txtype. 1653 * However, the log is a mix of different record types, and only the 1654 * replay vectors know how to byteswap their records. Therefore, if 1655 * the lr was byteswapped, undo it before invoking the replay vector. 1656 */ 1657 if (zr->zr_byteswap) 1658 byteswap_uint64_array(zr->zr_lr, reclen); 1659 1660 /* 1661 * We must now do two things atomically: replay this log record, 1662 * and update the log header sequence number to reflect the fact that 1663 * we did so. At the end of each replay function the sequence number 1664 * is updated if we are in replay mode. 1665 */ 1666 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 1667 if (error) { 1668 /* 1669 * The DMU's dnode layer doesn't see removes until the txg 1670 * commits, so a subsequent claim can spuriously fail with 1671 * EEXIST. So if we receive any error we try syncing out 1672 * any removes then retry the transaction. Note that we 1673 * specify B_FALSE for byteswap now, so we don't do it twice. 1674 */ 1675 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1676 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 1677 if (error) 1678 return (zil_replay_error(zilog, lr, error)); 1679 } 1680 return (0); 1681 } 1682 1683 /* ARGSUSED */ 1684 static int 1685 zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 1686 { 1687 zilog->zl_replay_blks++; 1688 1689 return (0); 1690 } 1691 1692 /* 1693 * If this dataset has a non-empty intent log, replay it and destroy it. 1694 */ 1695 void 1696 zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1697 { 1698 zilog_t *zilog = dmu_objset_zil(os); 1699 const zil_header_t *zh = zilog->zl_header; 1700 zil_replay_arg_t zr; 1701 1702 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1703 zil_destroy(zilog, B_TRUE); 1704 return; 1705 } 1706 1707 zr.zr_replay = replay_func; 1708 zr.zr_arg = arg; 1709 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1710 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1711 1712 /* 1713 * Wait for in-progress removes to sync before starting replay. 1714 */ 1715 txg_wait_synced(zilog->zl_dmu_pool, 0); 1716 1717 zilog->zl_replay = B_TRUE; 1718 zilog->zl_replay_time = ddi_get_lbolt(); 1719 ASSERT(zilog->zl_replay_blks == 0); 1720 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1721 zh->zh_claim_txg); 1722 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 1723 1724 zil_destroy(zilog, B_FALSE); 1725 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 1726 zilog->zl_replay = B_FALSE; 1727 } 1728 1729 boolean_t 1730 zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 1731 { 1732 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1733 return (B_TRUE); 1734 1735 if (zilog->zl_replay) { 1736 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1737 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 1738 zilog->zl_replaying_seq; 1739 return (B_TRUE); 1740 } 1741 1742 return (B_FALSE); 1743 } 1744 1745 /* ARGSUSED */ 1746 int 1747 zil_vdev_offline(const char *osname, void *arg) 1748 { 1749 objset_t *os; 1750 zilog_t *zilog; 1751 int error; 1752 1753 error = dmu_objset_hold(osname, FTAG, &os); 1754 if (error) 1755 return (error); 1756 1757 zilog = dmu_objset_zil(os); 1758 if (zil_suspend(zilog) != 0) 1759 error = EEXIST; 1760 else 1761 zil_resume(zilog); 1762 dmu_objset_rele(os, FTAG); 1763 return (error); 1764 } 1765