zil.c revision 307266
1238104Sdes/* 2238104Sdes * CDDL HEADER START 3238104Sdes * 4238104Sdes * The contents of this file are subject to the terms of the 5238104Sdes * Common Development and Distribution License (the "License"). 6238104Sdes * You may not use this file except in compliance with the License. 7238104Sdes * 8238104Sdes * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9238104Sdes * or http://www.opensolaris.org/os/licensing. 10238104Sdes * See the License for the specific language governing permissions 11238104Sdes * and limitations under the License. 12238104Sdes * 13238104Sdes * When distributing Covered Code, include this CDDL HEADER in each 14238104Sdes * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15238104Sdes * If applicable, add the following below this CDDL HEADER, with the 16238104Sdes * fields enclosed by brackets "[]" replaced with your own identifying 17238104Sdes * information: Portions Copyright [yyyy] [name of copyright owner] 18238104Sdes * 19238104Sdes * CDDL HEADER END 20238104Sdes */ 21238104Sdes/* 22238104Sdes * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23238104Sdes * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 24238104Sdes * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25238104Sdes * Copyright (c) 2014 Integros [integros.com] 26238104Sdes */ 27238104Sdes 28238104Sdes/* Portions Copyright 2010 Robert Milkowski */ 29238104Sdes 30238104Sdes#include <sys/zfs_context.h> 31238104Sdes#include <sys/spa.h> 32238104Sdes#include <sys/dmu.h> 33238104Sdes#include <sys/zap.h> 34238104Sdes#include <sys/arc.h> 35238104Sdes#include <sys/stat.h> 36238104Sdes#include <sys/resource.h> 37238104Sdes#include <sys/zil.h> 38238104Sdes#include <sys/zil_impl.h> 39238104Sdes#include <sys/dsl_dataset.h> 40238104Sdes#include <sys/vdev_impl.h> 41238104Sdes#include <sys/dmu_tx.h> 42238104Sdes#include <sys/dsl_pool.h> 43238104Sdes 44238104Sdes/* 45238104Sdes * The zfs intent log (ZIL) saves transaction records of system calls 46238104Sdes * that change the file system in memory with enough information 47238104Sdes * to be able to replay them. These are stored in memory until 48238104Sdes * either the DMU transaction group (txg) commits them to the stable pool 49238104Sdes * and they can be discarded, or they are flushed to the stable log 50238104Sdes * (also in the pool) due to a fsync, O_DSYNC or other synchronous 51238104Sdes * requirement. In the event of a panic or power fail then those log 52238104Sdes * records (transactions) are replayed. 53238104Sdes * 54238104Sdes * There is one ZIL per file system. Its on-disk (pool) format consists 55238104Sdes * of 3 parts: 56238104Sdes * 57238104Sdes * - ZIL header 58238104Sdes * - ZIL blocks 59238104Sdes * - ZIL records 60238104Sdes * 61238104Sdes * A log record holds a system call transaction. Log blocks can 62238104Sdes * hold many log records and the blocks are chained together. 63238104Sdes * Each ZIL block contains a block pointer (blkptr_t) to the next 64238104Sdes * ZIL block in the chain. The ZIL header points to the first 65238104Sdes * block in the chain. Note there is not a fixed place in the pool 66238104Sdes * to hold blocks. They are dynamically allocated and freed as 67238104Sdes * needed from the blocks available. Figure X shows the ZIL structure: 68238104Sdes */ 69238104Sdes 70238104Sdes/* 71238104Sdes * Disable intent logging replay. This global ZIL switch affects all pools. 72238104Sdes */ 73238104Sdesint zil_replay_disable = 0; 74238104SdesSYSCTL_DECL(_vfs_zfs); 75238104SdesTUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 76238104SdesSYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 77238104Sdes &zil_replay_disable, 0, "Disable intent logging replay"); 78238104Sdes 79238104Sdes/* 80238104Sdes * Tunable parameter for debugging or performance analysis. Setting 81238104Sdes * zfs_nocacheflush will cause corruption on power loss if a volatile 82238104Sdes * out-of-order write cache is enabled. 83238104Sdes */ 84238104Sdesboolean_t zfs_nocacheflush = B_FALSE; 85238104SdesTUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 86238104SdesSYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 87238104Sdes &zfs_nocacheflush, 0, "Disable cache flush"); 88238104Sdesboolean_t zfs_trim_enabled = B_TRUE; 89238104SdesSYSCTL_DECL(_vfs_zfs_trim); 90238104SdesTUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 91238104SdesSYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 92238104Sdes "Enable ZFS TRIM"); 93238104Sdes 94238104Sdesstatic kmem_cache_t *zil_lwb_cache; 95238104Sdes 96238104Sdesstatic void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 97238104Sdes 98238104Sdes#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 99238104Sdes sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 100238104Sdes 101238104Sdes 102238104Sdes/* 103238104Sdes * ziltest is by and large an ugly hack, but very useful in 104238104Sdes * checking replay without tedious work. 105238104Sdes * When running ziltest we want to keep all itx's and so maintain 106238104Sdes * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 107238104Sdes * We subtract TXG_CONCURRENT_STATES to allow for common code. 108238104Sdes */ 109238104Sdes#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 110238104Sdes 111238104Sdesstatic int 112238104Sdeszil_bp_compare(const void *x1, const void *x2) 113238104Sdes{ 114238104Sdes const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 115238104Sdes const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 116238104Sdes 117238104Sdes if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 118238104Sdes return (-1); 119238104Sdes if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 120238104Sdes return (1); 121238104Sdes 122238104Sdes if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 123238104Sdes return (-1); 124238104Sdes if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 125238104Sdes return (1); 126238104Sdes 127238104Sdes return (0); 128238104Sdes} 129238104Sdes 130238104Sdesstatic void 131238104Sdeszil_bp_tree_init(zilog_t *zilog) 132238104Sdes{ 133238104Sdes avl_create(&zilog->zl_bp_tree, zil_bp_compare, 134238104Sdes sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 135238104Sdes} 136238104Sdes 137238104Sdesstatic void 138238104Sdeszil_bp_tree_fini(zilog_t *zilog) 139238104Sdes{ 140238104Sdes avl_tree_t *t = &zilog->zl_bp_tree; 141238104Sdes zil_bp_node_t *zn; 142238104Sdes void *cookie = NULL; 143238104Sdes 144238104Sdes while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 145238104Sdes kmem_free(zn, sizeof (zil_bp_node_t)); 146238104Sdes 147238104Sdes avl_destroy(t); 148238104Sdes} 149238104Sdes 150238104Sdesint 151238104Sdeszil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 152238104Sdes{ 153238104Sdes avl_tree_t *t = &zilog->zl_bp_tree; 154238104Sdes const dva_t *dva; 155238104Sdes zil_bp_node_t *zn; 156238104Sdes avl_index_t where; 157238104Sdes 158238104Sdes if (BP_IS_EMBEDDED(bp)) 159238104Sdes return (0); 160238104Sdes 161238104Sdes dva = BP_IDENTITY(bp); 162238104Sdes 163238104Sdes if (avl_find(t, dva, &where) != NULL) 164238104Sdes return (SET_ERROR(EEXIST)); 165238104Sdes 166238104Sdes zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 167238104Sdes zn->zn_dva = *dva; 168238104Sdes avl_insert(t, zn, where); 169238104Sdes 170238104Sdes return (0); 171238104Sdes} 172238104Sdes 173238104Sdesstatic zil_header_t * 174238104Sdeszil_header_in_syncing_context(zilog_t *zilog) 175238104Sdes{ 176238104Sdes return ((zil_header_t *)zilog->zl_header); 177238104Sdes} 178238104Sdes 179238104Sdesstatic void 180238104Sdeszil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 181238104Sdes{ 182238104Sdes zio_cksum_t *zc = &bp->blk_cksum; 183238104Sdes 184238104Sdes zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 185238104Sdes zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 186238104Sdes zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 187238104Sdes zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 188238104Sdes} 189238104Sdes 190238104Sdes/* 191238104Sdes * Read a log block and make sure it's valid. 192238104Sdes */ 193238104Sdesstatic int 194238104Sdeszil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 195238104Sdes char **end) 196238104Sdes{ 197238104Sdes enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 198238104Sdes arc_flags_t aflags = ARC_FLAG_WAIT; 199238104Sdes arc_buf_t *abuf = NULL; 200238104Sdes zbookmark_phys_t zb; 201238104Sdes int error; 202238104Sdes 203238104Sdes if (zilog->zl_header->zh_claim_txg == 0) 204238104Sdes zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 205238104Sdes 206238104Sdes if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 207238104Sdes zio_flags |= ZIO_FLAG_SPECULATIVE; 208238104Sdes 209238104Sdes SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 210238104Sdes ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 211238104Sdes 212238104Sdes error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 213238104Sdes ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 214238104Sdes 215238104Sdes if (error == 0) { 216238104Sdes zio_cksum_t cksum = bp->blk_cksum; 217238104Sdes 218238104Sdes /* 219238104Sdes * Validate the checksummed log block. 220238104Sdes * 221238104Sdes * Sequence numbers should be... sequential. The checksum 222238104Sdes * verifier for the next block should be bp's checksum plus 1. 223238104Sdes * 224238104Sdes * Also check the log chain linkage and size used. 225238104Sdes */ 226238104Sdes cksum.zc_word[ZIL_ZC_SEQ]++; 227238104Sdes 228238104Sdes if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 229238104Sdes zil_chain_t *zilc = abuf->b_data; 230238104Sdes char *lr = (char *)(zilc + 1); 231238104Sdes uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 232238104Sdes 233238104Sdes if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 234238104Sdes sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 235238104Sdes error = SET_ERROR(ECKSUM); 236238104Sdes } else { 237238104Sdes ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 238238104Sdes bcopy(lr, dst, len); 239238104Sdes *end = (char *)dst + len; 240238104Sdes *nbp = zilc->zc_next_blk; 241238104Sdes } 242238104Sdes } else { 243238104Sdes char *lr = abuf->b_data; 244238104Sdes uint64_t size = BP_GET_LSIZE(bp); 245238104Sdes zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 246238104Sdes 247238104Sdes if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 248238104Sdes sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 249238104Sdes (zilc->zc_nused > (size - sizeof (*zilc)))) { 250238104Sdes error = SET_ERROR(ECKSUM); 251238104Sdes } else { 252238104Sdes ASSERT3U(zilc->zc_nused, <=, 253238104Sdes SPA_OLD_MAXBLOCKSIZE); 254238104Sdes bcopy(lr, dst, zilc->zc_nused); 255238104Sdes *end = (char *)dst + zilc->zc_nused; 256238104Sdes *nbp = zilc->zc_next_blk; 257238104Sdes } 258238104Sdes } 259238104Sdes 260238104Sdes arc_buf_destroy(abuf, &abuf); 261238104Sdes } 262238104Sdes 263238104Sdes return (error); 264238104Sdes} 265238104Sdes 266238104Sdes/* 267238104Sdes * Read a TX_WRITE log data block. 268238104Sdes */ 269238104Sdesstatic int 270238104Sdeszil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 271238104Sdes{ 272238104Sdes enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 273238104Sdes const blkptr_t *bp = &lr->lr_blkptr; 274238104Sdes arc_flags_t aflags = ARC_FLAG_WAIT; 275238104Sdes arc_buf_t *abuf = NULL; 276238104Sdes zbookmark_phys_t zb; 277238104Sdes int error; 278238104Sdes 279238104Sdes if (BP_IS_HOLE(bp)) { 280238104Sdes if (wbuf != NULL) 281238104Sdes bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 282238104Sdes return (0); 283238104Sdes } 284238104Sdes 285238104Sdes if (zilog->zl_header->zh_claim_txg == 0) 286238104Sdes zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 287238104Sdes 288238104Sdes SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 289238104Sdes ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 290238104Sdes 291238104Sdes error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 292238104Sdes ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 293238104Sdes 294238104Sdes if (error == 0) { 295238104Sdes if (wbuf != NULL) 296238104Sdes bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 297238104Sdes arc_buf_destroy(abuf, &abuf); 298238104Sdes } 299238104Sdes 300238104Sdes return (error); 301238104Sdes} 302238104Sdes 303238104Sdes/* 304238104Sdes * Parse the intent log, and call parse_func for each valid record within. 305238104Sdes */ 306238104Sdesint 307238104Sdeszil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 308238104Sdes zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 309238104Sdes{ 310238104Sdes const zil_header_t *zh = zilog->zl_header; 311238104Sdes boolean_t claimed = !!zh->zh_claim_txg; 312238104Sdes uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 313238104Sdes uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 314238104Sdes uint64_t max_blk_seq = 0; 315238104Sdes uint64_t max_lr_seq = 0; 316238104Sdes uint64_t blk_count = 0; 317238104Sdes uint64_t lr_count = 0; 318238104Sdes blkptr_t blk, next_blk; 319238104Sdes char *lrbuf, *lrp; 320238104Sdes int error = 0; 321238104Sdes 322238104Sdes /* 323238104Sdes * Old logs didn't record the maximum zh_claim_lr_seq. 324238104Sdes */ 325238104Sdes if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 326238104Sdes claim_lr_seq = UINT64_MAX; 327238104Sdes 328238104Sdes /* 329238104Sdes * Starting at the block pointed to by zh_log we read the log chain. 330238104Sdes * For each block in the chain we strongly check that block to 331238104Sdes * ensure its validity. We stop when an invalid block is found. 332238104Sdes * For each block pointer in the chain we call parse_blk_func(). 333238104Sdes * For each record in each valid block we call parse_lr_func(). 334238104Sdes * If the log has been claimed, stop if we encounter a sequence 335238104Sdes * number greater than the highest claimed sequence number. 336238104Sdes */ 337238104Sdes lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 338238104Sdes zil_bp_tree_init(zilog); 339238104Sdes 340238104Sdes for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 341238104Sdes uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 342238104Sdes int reclen; 343238104Sdes char *end; 344238104Sdes 345238104Sdes if (blk_seq > claim_blk_seq) 346238104Sdes break; 347238104Sdes if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 348238104Sdes break; 349238104Sdes ASSERT3U(max_blk_seq, <, blk_seq); 350238104Sdes max_blk_seq = blk_seq; 351238104Sdes blk_count++; 352238104Sdes 353238104Sdes if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 354238104Sdes break; 355238104Sdes 356238104Sdes error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 357238104Sdes if (error != 0) 358238104Sdes break; 359238104Sdes 360238104Sdes for (lrp = lrbuf; lrp < end; lrp += reclen) { 361238104Sdes lr_t *lr = (lr_t *)lrp; 362238104Sdes reclen = lr->lrc_reclen; 363238104Sdes ASSERT3U(reclen, >=, sizeof (lr_t)); 364238104Sdes if (lr->lrc_seq > claim_lr_seq) 365238104Sdes goto done; 366238104Sdes if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 367238104Sdes goto done; 368238104Sdes ASSERT3U(max_lr_seq, <, lr->lrc_seq); 369238104Sdes max_lr_seq = lr->lrc_seq; 370238104Sdes lr_count++; 371238104Sdes } 372238104Sdes } 373238104Sdesdone: 374238104Sdes zilog->zl_parse_error = error; 375238104Sdes zilog->zl_parse_blk_seq = max_blk_seq; 376238104Sdes zilog->zl_parse_lr_seq = max_lr_seq; 377238104Sdes zilog->zl_parse_blk_count = blk_count; 378238104Sdes zilog->zl_parse_lr_count = lr_count; 379238104Sdes 380238104Sdes ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 381238104Sdes (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 382238104Sdes 383238104Sdes zil_bp_tree_fini(zilog); 384238104Sdes zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 385238104Sdes 386238104Sdes return (error); 387238104Sdes} 388238104Sdes 389238104Sdesstatic int 390238104Sdeszil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 391238104Sdes{ 392238104Sdes /* 393238104Sdes * Claim log block if not already committed and not already claimed. 394238104Sdes * If tx == NULL, just verify that the block is claimable. 395238104Sdes */ 396238104Sdes if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 397238104Sdes zil_bp_tree_add(zilog, bp) != 0) 398238104Sdes return (0); 399238104Sdes 400238104Sdes return (zio_wait(zio_claim(NULL, zilog->zl_spa, 401238104Sdes tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 402238104Sdes ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 403238104Sdes} 404238104Sdes 405238104Sdesstatic int 406238104Sdeszil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 407238104Sdes{ 408238104Sdes lr_write_t *lr = (lr_write_t *)lrc; 409238104Sdes int error; 410238104Sdes 411238104Sdes if (lrc->lrc_txtype != TX_WRITE) 412238104Sdes return (0); 413238104Sdes 414238104Sdes /* 415238104Sdes * If the block is not readable, don't claim it. This can happen 416238104Sdes * in normal operation when a log block is written to disk before 417238104Sdes * some of the dmu_sync() blocks it points to. In this case, the 418238104Sdes * transaction cannot have been committed to anyone (we would have 419238104Sdes * waited for all writes to be stable first), so it is semantically 420238104Sdes * correct to declare this the end of the log. 421238104Sdes */ 422238104Sdes if (lr->lr_blkptr.blk_birth >= first_txg && 423238104Sdes (error = zil_read_log_data(zilog, lr, NULL)) != 0) 424238104Sdes return (error); 425238104Sdes return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 426238104Sdes} 427238104Sdes 428238104Sdes/* ARGSUSED */ 429238104Sdesstatic int 430238104Sdeszil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 431238104Sdes{ 432238104Sdes zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 433238104Sdes 434238104Sdes return (0); 435238104Sdes} 436238104Sdes 437238104Sdesstatic int 438238104Sdeszil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 439238104Sdes{ 440238104Sdes lr_write_t *lr = (lr_write_t *)lrc; 441238104Sdes blkptr_t *bp = &lr->lr_blkptr; 442238104Sdes 443238104Sdes /* 444238104Sdes * If we previously claimed it, we need to free it. 445238104Sdes */ 446238104Sdes if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 447238104Sdes bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 448238104Sdes !BP_IS_HOLE(bp)) 449238104Sdes zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 450238104Sdes 451238104Sdes return (0); 452238104Sdes} 453238104Sdes 454238104Sdesstatic lwb_t * 455238104Sdeszil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 456238104Sdes{ 457238104Sdes lwb_t *lwb; 458238104Sdes 459238104Sdes lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 460238104Sdes lwb->lwb_zilog = zilog; 461238104Sdes lwb->lwb_blk = *bp; 462238104Sdes lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 463238104Sdes lwb->lwb_max_txg = txg; 464238104Sdes lwb->lwb_zio = NULL; 465238104Sdes lwb->lwb_tx = NULL; 466238104Sdes if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 467238104Sdes lwb->lwb_nused = sizeof (zil_chain_t); 468238104Sdes lwb->lwb_sz = BP_GET_LSIZE(bp); 469238104Sdes } else { 470238104Sdes lwb->lwb_nused = 0; 471238104Sdes lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 472238104Sdes } 473238104Sdes 474238104Sdes mutex_enter(&zilog->zl_lock); 475238104Sdes list_insert_tail(&zilog->zl_lwb_list, lwb); 476238104Sdes mutex_exit(&zilog->zl_lock); 477238104Sdes 478238104Sdes return (lwb); 479238104Sdes} 480238104Sdes 481238104Sdes/* 482238104Sdes * Called when we create in-memory log transactions so that we know 483238104Sdes * to cleanup the itxs at the end of spa_sync(). 484238104Sdes */ 485238104Sdesvoid 486238104Sdeszilog_dirty(zilog_t *zilog, uint64_t txg) 487238104Sdes{ 488238104Sdes dsl_pool_t *dp = zilog->zl_dmu_pool; 489238104Sdes dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 490238104Sdes 491238104Sdes if (ds->ds_is_snapshot) 492238104Sdes panic("dirtying snapshot!"); 493238104Sdes 494238104Sdes if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 495238104Sdes /* up the hold count until we can be written out */ 496238104Sdes dmu_buf_add_ref(ds->ds_dbuf, zilog); 497238104Sdes } 498238104Sdes} 499238104Sdes 500238104Sdesboolean_t 501238104Sdeszilog_is_dirty(zilog_t *zilog) 502238104Sdes{ 503238104Sdes dsl_pool_t *dp = zilog->zl_dmu_pool; 504238104Sdes 505238104Sdes for (int t = 0; t < TXG_SIZE; t++) { 506238104Sdes if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 507238104Sdes return (B_TRUE); 508238104Sdes } 509238104Sdes return (B_FALSE); 510238104Sdes} 511238104Sdes 512238104Sdes/* 513238104Sdes * Create an on-disk intent log. 514238104Sdes */ 515238104Sdesstatic lwb_t * 516238104Sdeszil_create(zilog_t *zilog) 517238104Sdes{ 518238104Sdes const zil_header_t *zh = zilog->zl_header; 519238104Sdes lwb_t *lwb = NULL; 520238104Sdes uint64_t txg = 0; 521238104Sdes dmu_tx_t *tx = NULL; 522238104Sdes blkptr_t blk; 523238104Sdes int error = 0; 524238104Sdes 525238104Sdes /* 526238104Sdes * Wait for any previous destroy to complete. 527238104Sdes */ 528238104Sdes txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 529238104Sdes 530238104Sdes ASSERT(zh->zh_claim_txg == 0); 531238104Sdes ASSERT(zh->zh_replay_seq == 0); 532238104Sdes 533238104Sdes blk = zh->zh_log; 534238104Sdes 535238104Sdes /* 536238104Sdes * Allocate an initial log block if: 537238104Sdes * - there isn't one already 538238104Sdes * - the existing block is the wrong endianess 539238104Sdes */ 540238104Sdes if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 541238104Sdes tx = dmu_tx_create(zilog->zl_os); 542238104Sdes VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 543238104Sdes dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 544238104Sdes txg = dmu_tx_get_txg(tx); 545238104Sdes 546238104Sdes if (!BP_IS_HOLE(&blk)) { 547238104Sdes zio_free_zil(zilog->zl_spa, txg, &blk); 548238104Sdes BP_ZERO(&blk); 549238104Sdes } 550238104Sdes 551238104Sdes error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 552238104Sdes ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 553238104Sdes 554238104Sdes if (error == 0) 555238104Sdes zil_init_log_chain(zilog, &blk); 556238104Sdes } 557238104Sdes 558238104Sdes /* 559238104Sdes * Allocate a log write buffer (lwb) for the first log block. 560238104Sdes */ 561238104Sdes if (error == 0) 562238104Sdes lwb = zil_alloc_lwb(zilog, &blk, txg); 563238104Sdes 564238104Sdes /* 565238104Sdes * If we just allocated the first log block, commit our transaction 566238104Sdes * and wait for zil_sync() to stuff the block poiner into zh_log. 567238104Sdes * (zh is part of the MOS, so we cannot modify it in open context.) 568238104Sdes */ 569238104Sdes if (tx != NULL) { 570238104Sdes dmu_tx_commit(tx); 571238104Sdes txg_wait_synced(zilog->zl_dmu_pool, txg); 572238104Sdes } 573238104Sdes 574238104Sdes ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 575238104Sdes 576238104Sdes return (lwb); 577238104Sdes} 578238104Sdes 579238104Sdes/* 580238104Sdes * In one tx, free all log blocks and clear the log header. 581238104Sdes * If keep_first is set, then we're replaying a log with no content. 582238104Sdes * We want to keep the first block, however, so that the first 583238104Sdes * synchronous transaction doesn't require a txg_wait_synced() 584238104Sdes * in zil_create(). We don't need to txg_wait_synced() here either 585238104Sdes * when keep_first is set, because both zil_create() and zil_destroy() 586238104Sdes * will wait for any in-progress destroys to complete. 587238104Sdes */ 588238104Sdesvoid 589238104Sdeszil_destroy(zilog_t *zilog, boolean_t keep_first) 590238104Sdes{ 591238104Sdes const zil_header_t *zh = zilog->zl_header; 592238104Sdes lwb_t *lwb; 593238104Sdes dmu_tx_t *tx; 594238104Sdes uint64_t txg; 595238104Sdes 596238104Sdes /* 597238104Sdes * Wait for any previous destroy to complete. 598238104Sdes */ 599238104Sdes txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 600238104Sdes 601238104Sdes zilog->zl_old_header = *zh; /* debugging aid */ 602238104Sdes 603238104Sdes if (BP_IS_HOLE(&zh->zh_log)) 604238104Sdes return; 605238104Sdes 606238104Sdes tx = dmu_tx_create(zilog->zl_os); 607238104Sdes VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 608238104Sdes dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 609238104Sdes txg = dmu_tx_get_txg(tx); 610238104Sdes 611238104Sdes mutex_enter(&zilog->zl_lock); 612238104Sdes 613238104Sdes ASSERT3U(zilog->zl_destroy_txg, <, txg); 614238104Sdes zilog->zl_destroy_txg = txg; 615238104Sdes zilog->zl_keep_first = keep_first; 616238104Sdes 617238104Sdes if (!list_is_empty(&zilog->zl_lwb_list)) { 618238104Sdes ASSERT(zh->zh_claim_txg == 0); 619238104Sdes VERIFY(!keep_first); 620238104Sdes while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 621238104Sdes list_remove(&zilog->zl_lwb_list, lwb); 622238104Sdes if (lwb->lwb_buf != NULL) 623238104Sdes zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 624238104Sdes zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 625238104Sdes kmem_cache_free(zil_lwb_cache, lwb); 626238104Sdes } 627238104Sdes } else if (!keep_first) { 628238104Sdes zil_destroy_sync(zilog, tx); 629238104Sdes } 630238104Sdes mutex_exit(&zilog->zl_lock); 631238104Sdes 632238104Sdes dmu_tx_commit(tx); 633238104Sdes} 634238104Sdes 635238104Sdesvoid 636238104Sdeszil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 637238104Sdes{ 638238104Sdes ASSERT(list_is_empty(&zilog->zl_lwb_list)); 639238104Sdes (void) zil_parse(zilog, zil_free_log_block, 640238104Sdes zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 641238104Sdes} 642238104Sdes 643238104Sdesint 644238104Sdeszil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 645238104Sdes{ 646238104Sdes dmu_tx_t *tx = txarg; 647238104Sdes uint64_t first_txg = dmu_tx_get_txg(tx); 648238104Sdes zilog_t *zilog; 649238104Sdes zil_header_t *zh; 650238104Sdes objset_t *os; 651238104Sdes int error; 652238104Sdes 653238104Sdes error = dmu_objset_own_obj(dp, ds->ds_object, 654238104Sdes DMU_OST_ANY, B_FALSE, FTAG, &os); 655238104Sdes if (error != 0) { 656238104Sdes /* 657238104Sdes * EBUSY indicates that the objset is inconsistent, in which 658238104Sdes * case it can not have a ZIL. 659238104Sdes */ 660238104Sdes if (error != EBUSY) { 661238104Sdes cmn_err(CE_WARN, "can't open objset for %llu, error %u", 662238104Sdes (unsigned long long)ds->ds_object, error); 663238104Sdes } 664238104Sdes return (0); 665238104Sdes } 666238104Sdes 667238104Sdes zilog = dmu_objset_zil(os); 668238104Sdes zh = zil_header_in_syncing_context(zilog); 669238104Sdes 670238104Sdes if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 671238104Sdes if (!BP_IS_HOLE(&zh->zh_log)) 672238104Sdes zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 673238104Sdes BP_ZERO(&zh->zh_log); 674238104Sdes dsl_dataset_dirty(dmu_objset_ds(os), tx); 675238104Sdes dmu_objset_disown(os, FTAG); 676238104Sdes return (0); 677238104Sdes } 678238104Sdes 679238104Sdes /* 680238104Sdes * Claim all log blocks if we haven't already done so, and remember 681238104Sdes * the highest claimed sequence number. This ensures that if we can 682238104Sdes * read only part of the log now (e.g. due to a missing device), 683238104Sdes * but we can read the entire log later, we will not try to replay 684238104Sdes * or destroy beyond the last block we successfully claimed. 685238104Sdes */ 686238104Sdes ASSERT3U(zh->zh_claim_txg, <=, first_txg); 687238104Sdes if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 688238104Sdes (void) zil_parse(zilog, zil_claim_log_block, 689238104Sdes zil_claim_log_record, tx, first_txg); 690238104Sdes zh->zh_claim_txg = first_txg; 691238104Sdes zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 692238104Sdes zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 693238104Sdes if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 694238104Sdes zh->zh_flags |= ZIL_REPLAY_NEEDED; 695238104Sdes zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 696238104Sdes dsl_dataset_dirty(dmu_objset_ds(os), tx); 697238104Sdes } 698238104Sdes 699238104Sdes ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 700238104Sdes dmu_objset_disown(os, FTAG); 701238104Sdes return (0); 702238104Sdes} 703238104Sdes 704238104Sdes/* 705238104Sdes * Check the log by walking the log chain. 706238104Sdes * Checksum errors are ok as they indicate the end of the chain. 707238104Sdes * Any other error (no device or read failure) returns an error. 708238104Sdes */ 709238104Sdes/* ARGSUSED */ 710238104Sdesint 711238104Sdeszil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 712238104Sdes{ 713238104Sdes zilog_t *zilog; 714238104Sdes objset_t *os; 715238104Sdes blkptr_t *bp; 716238104Sdes int error; 717238104Sdes 718238104Sdes ASSERT(tx == NULL); 719238104Sdes 720238104Sdes error = dmu_objset_from_ds(ds, &os); 721238104Sdes if (error != 0) { 722238104Sdes cmn_err(CE_WARN, "can't open objset %llu, error %d", 723238104Sdes (unsigned long long)ds->ds_object, error); 724238104Sdes return (0); 725238104Sdes } 726238104Sdes 727238104Sdes zilog = dmu_objset_zil(os); 728238104Sdes bp = (blkptr_t *)&zilog->zl_header->zh_log; 729238104Sdes 730238104Sdes /* 731238104Sdes * Check the first block and determine if it's on a log device 732238104Sdes * which may have been removed or faulted prior to loading this 733238104Sdes * pool. If so, there's no point in checking the rest of the log 734238104Sdes * as its content should have already been synced to the pool. 735238104Sdes */ 736238104Sdes if (!BP_IS_HOLE(bp)) { 737238104Sdes vdev_t *vd; 738238104Sdes boolean_t valid = B_TRUE; 739238104Sdes 740238104Sdes spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 741238104Sdes vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 742238104Sdes if (vd->vdev_islog && vdev_is_dead(vd)) 743238104Sdes valid = vdev_log_state_valid(vd); 744238104Sdes spa_config_exit(os->os_spa, SCL_STATE, FTAG); 745238104Sdes 746238104Sdes if (!valid) 747238104Sdes return (0); 748238104Sdes } 749238104Sdes 750238104Sdes /* 751238104Sdes * Because tx == NULL, zil_claim_log_block() will not actually claim 752238104Sdes * any blocks, but just determine whether it is possible to do so. 753238104Sdes * In addition to checking the log chain, zil_claim_log_block() 754238104Sdes * will invoke zio_claim() with a done func of spa_claim_notify(), 755238104Sdes * which will update spa_max_claim_txg. See spa_load() for details. 756238104Sdes */ 757238104Sdes error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 758238104Sdes zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 759238104Sdes 760238104Sdes return ((error == ECKSUM || error == ENOENT) ? 0 : error); 761238104Sdes} 762238104Sdes 763238104Sdesstatic int 764238104Sdeszil_vdev_compare(const void *x1, const void *x2) 765238104Sdes{ 766238104Sdes const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 767238104Sdes const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 768238104Sdes 769238104Sdes if (v1 < v2) 770238104Sdes return (-1); 771238104Sdes if (v1 > v2) 772238104Sdes return (1); 773238104Sdes 774246827Sdes return (0); 775238104Sdes} 776238104Sdes 777238104Sdesvoid 778238104Sdeszil_add_block(zilog_t *zilog, const blkptr_t *bp) 779238104Sdes{ 780238104Sdes avl_tree_t *t = &zilog->zl_vdev_tree; 781238104Sdes avl_index_t where; 782238104Sdes zil_vdev_node_t *zv, zvsearch; 783238104Sdes int ndvas = BP_GET_NDVAS(bp); 784238104Sdes int i; 785238104Sdes 786238104Sdes if (zfs_nocacheflush) 787238104Sdes return; 788238104Sdes 789238104Sdes ASSERT(zilog->zl_writer); 790238104Sdes 791238104Sdes /* 792238104Sdes * Even though we're zl_writer, we still need a lock because the 793238104Sdes * zl_get_data() callbacks may have dmu_sync() done callbacks 794238104Sdes * that will run concurrently. 795238104Sdes */ 796238104Sdes mutex_enter(&zilog->zl_vdev_lock); 797238104Sdes for (i = 0; i < ndvas; i++) { 798238104Sdes zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 799238104Sdes if (avl_find(t, &zvsearch, &where) == NULL) { 800238104Sdes zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 801238104Sdes zv->zv_vdev = zvsearch.zv_vdev; 802238104Sdes avl_insert(t, zv, where); 803238104Sdes } 804238104Sdes } 805238104Sdes mutex_exit(&zilog->zl_vdev_lock); 806238104Sdes} 807238104Sdes 808238104Sdesstatic void 809238104Sdeszil_flush_vdevs(zilog_t *zilog) 810238104Sdes{ 811238104Sdes spa_t *spa = zilog->zl_spa; 812238104Sdes avl_tree_t *t = &zilog->zl_vdev_tree; 813238104Sdes void *cookie = NULL; 814238104Sdes zil_vdev_node_t *zv; 815238104Sdes zio_t *zio; 816238104Sdes 817238104Sdes ASSERT(zilog->zl_writer); 818238104Sdes 819238104Sdes /* 820238104Sdes * We don't need zl_vdev_lock here because we're the zl_writer, 821238104Sdes * and all zl_get_data() callbacks are done. 822238104Sdes */ 823238104Sdes if (avl_numnodes(t) == 0) 824238104Sdes return; 825238104Sdes 826238104Sdes spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 827238104Sdes 828238104Sdes zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 829238104Sdes 830238104Sdes while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 831238104Sdes vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 832238104Sdes if (vd != NULL) 833238104Sdes zio_flush(zio, vd); 834238104Sdes kmem_free(zv, sizeof (*zv)); 835238104Sdes } 836238104Sdes 837238104Sdes /* 838238104Sdes * Wait for all the flushes to complete. Not all devices actually 839238104Sdes * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 840238104Sdes */ 841238104Sdes (void) zio_wait(zio); 842238104Sdes 843238104Sdes spa_config_exit(spa, SCL_STATE, FTAG); 844238104Sdes} 845238104Sdes 846238104Sdes/* 847238104Sdes * Function called when a log block write completes 848238104Sdes */ 849238104Sdesstatic void 850238104Sdeszil_lwb_write_done(zio_t *zio) 851238104Sdes{ 852238104Sdes lwb_t *lwb = zio->io_private; 853238104Sdes zilog_t *zilog = lwb->lwb_zilog; 854238104Sdes dmu_tx_t *tx = lwb->lwb_tx; 855238104Sdes 856238104Sdes ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 857238104Sdes ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 858238104Sdes ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 859238104Sdes ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 860238104Sdes ASSERT(!BP_IS_GANG(zio->io_bp)); 861238104Sdes ASSERT(!BP_IS_HOLE(zio->io_bp)); 862238104Sdes ASSERT(BP_GET_FILL(zio->io_bp) == 0); 863238104Sdes 864238104Sdes /* 865238104Sdes * Ensure the lwb buffer pointer is cleared before releasing 866238104Sdes * the txg. If we have had an allocation failure and 867238104Sdes * the txg is waiting to sync then we want want zil_sync() 868238104Sdes * to remove the lwb so that it's not picked up as the next new 869238104Sdes * one in zil_commit_writer(). zil_sync() will only remove 870238104Sdes * the lwb if lwb_buf is null. 871238104Sdes */ 872238104Sdes zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 873238104Sdes mutex_enter(&zilog->zl_lock); 874238104Sdes lwb->lwb_buf = NULL; 875238104Sdes lwb->lwb_tx = NULL; 876238104Sdes mutex_exit(&zilog->zl_lock); 877238104Sdes 878238104Sdes /* 879238104Sdes * Now that we've written this log block, we have a stable pointer 880238104Sdes * to the next block in the chain, so it's OK to let the txg in 881238104Sdes * which we allocated the next block sync. 882238104Sdes */ 883238104Sdes dmu_tx_commit(tx); 884238104Sdes} 885238104Sdes 886238104Sdes/* 887238104Sdes * Initialize the io for a log block. 888238104Sdes */ 889238104Sdesstatic void 890238104Sdeszil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 891238104Sdes{ 892238104Sdes zbookmark_phys_t zb; 893238104Sdes 894238104Sdes SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 895238104Sdes ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 896238104Sdes lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 897238104Sdes 898238104Sdes if (zilog->zl_root_zio == NULL) { 899238104Sdes zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 900238104Sdes ZIO_FLAG_CANFAIL); 901238104Sdes } 902238104Sdes if (lwb->lwb_zio == NULL) { 903238104Sdes lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 904238104Sdes 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 905238104Sdes zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 906238104Sdes ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 907238104Sdes } 908238104Sdes} 909238104Sdes 910238104Sdes/* 911238104Sdes * Define a limited set of intent log block sizes. 912238104Sdes * 913238104Sdes * These must be a multiple of 4KB. Note only the amount used (again 914238104Sdes * aligned to 4KB) actually gets written. However, we can't always just 915238104Sdes * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 916238104Sdes */ 917238104Sdesuint64_t zil_block_buckets[] = { 918238104Sdes 4096, /* non TX_WRITE */ 919238104Sdes 8192+4096, /* data base */ 920238104Sdes 32*1024 + 4096, /* NFS writes */ 921238104Sdes UINT64_MAX 922238104Sdes}; 923238104Sdes 924238104Sdes/* 925238104Sdes * Use the slog as long as the logbias is 'latency' and the current commit size 926238104Sdes * is less than the limit or the total list size is less than 2X the limit. 927238104Sdes * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 928238104Sdes */ 929238104Sdesuint64_t zil_slog_limit = 1024 * 1024; 930238104Sdes#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 931238104Sdes (((zilog)->zl_cur_used < zil_slog_limit) || \ 932238104Sdes ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 933238104Sdes 934238104Sdes/* 935238104Sdes * Start a log block write and advance to the next log block. 936238104Sdes * Calls are serialized. 937238104Sdes */ 938238104Sdesstatic lwb_t * 939238104Sdeszil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 940238104Sdes{ 941238104Sdes lwb_t *nlwb = NULL; 942238104Sdes zil_chain_t *zilc; 943238104Sdes spa_t *spa = zilog->zl_spa; 944238104Sdes blkptr_t *bp; 945238104Sdes dmu_tx_t *tx; 946238104Sdes uint64_t txg; 947238104Sdes uint64_t zil_blksz, wsz; 948238104Sdes int i, error; 949238104Sdes 950238104Sdes if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 951238104Sdes zilc = (zil_chain_t *)lwb->lwb_buf; 952238104Sdes bp = &zilc->zc_next_blk; 953238104Sdes } else { 954238104Sdes zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 955238104Sdes bp = &zilc->zc_next_blk; 956238104Sdes } 957238104Sdes 958238104Sdes ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 959238104Sdes 960238104Sdes /* 961238104Sdes * Allocate the next block and save its address in this block 962238104Sdes * before writing it in order to establish the log chain. 963238104Sdes * Note that if the allocation of nlwb synced before we wrote 964238104Sdes * the block that points at it (lwb), we'd leak it if we crashed. 965238104Sdes * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 966238104Sdes * We dirty the dataset to ensure that zil_sync() will be called 967238104Sdes * to clean up in the event of allocation failure or I/O failure. 968238104Sdes */ 969238104Sdes tx = dmu_tx_create(zilog->zl_os); 970238104Sdes VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 971238104Sdes dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 972238104Sdes txg = dmu_tx_get_txg(tx); 973238104Sdes 974238104Sdes lwb->lwb_tx = tx; 975238104Sdes 976238104Sdes /* 977238104Sdes * Log blocks are pre-allocated. Here we select the size of the next 978238104Sdes * block, based on size used in the last block. 979238104Sdes * - first find the smallest bucket that will fit the block from a 980238104Sdes * limited set of block sizes. This is because it's faster to write 981238104Sdes * blocks allocated from the same metaslab as they are adjacent or 982238104Sdes * close. 983238104Sdes * - next find the maximum from the new suggested size and an array of 984238104Sdes * previous sizes. This lessens a picket fence effect of wrongly 985238104Sdes * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 986238104Sdes * requests. 987238104Sdes * 988238104Sdes * Note we only write what is used, but we can't just allocate 989238104Sdes * the maximum block size because we can exhaust the available 990238104Sdes * pool log space. 991238104Sdes */ 992238104Sdes zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 993238104Sdes for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 994238104Sdes continue; 995238104Sdes zil_blksz = zil_block_buckets[i]; 996238104Sdes if (zil_blksz == UINT64_MAX) 997238104Sdes zil_blksz = SPA_OLD_MAXBLOCKSIZE; 998238104Sdes zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 999238104Sdes for (i = 0; i < ZIL_PREV_BLKS; i++) 1000238104Sdes zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1001238104Sdes zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1002238104Sdes 1003238104Sdes BP_ZERO(bp); 1004238104Sdes /* pass the old blkptr in order to spread log blocks across devs */ 1005238104Sdes error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 1006238104Sdes USE_SLOG(zilog)); 1007238104Sdes if (error == 0) { 1008238104Sdes ASSERT3U(bp->blk_birth, ==, txg); 1009238104Sdes bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1010238104Sdes bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1011238104Sdes 1012238104Sdes /* 1013238104Sdes * Allocate a new log write buffer (lwb). 1014238104Sdes */ 1015238104Sdes nlwb = zil_alloc_lwb(zilog, bp, txg); 1016238104Sdes 1017238104Sdes /* Record the block for later vdev flushing */ 1018238104Sdes zil_add_block(zilog, &lwb->lwb_blk); 1019238104Sdes } 1020238104Sdes 1021238104Sdes if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1022238104Sdes /* For Slim ZIL only write what is used. */ 1023238104Sdes wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1024238104Sdes ASSERT3U(wsz, <=, lwb->lwb_sz); 1025238104Sdes zio_shrink(lwb->lwb_zio, wsz); 1026238104Sdes 1027238104Sdes } else { 1028238104Sdes wsz = lwb->lwb_sz; 1029238104Sdes } 1030238104Sdes 1031238104Sdes zilc->zc_pad = 0; 1032238104Sdes zilc->zc_nused = lwb->lwb_nused; 1033238104Sdes zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1034238104Sdes 1035238104Sdes /* 1036238104Sdes * clear unused data for security 1037238104Sdes */ 1038238104Sdes bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1039238104Sdes 1040238104Sdes zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1041238104Sdes 1042238104Sdes /* 1043238104Sdes * If there was an allocation failure then nlwb will be null which 1044238104Sdes * forces a txg_wait_synced(). 1045238104Sdes */ 1046238104Sdes return (nlwb); 1047238104Sdes} 1048238104Sdes 1049238104Sdesstatic lwb_t * 1050238104Sdeszil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1051238104Sdes{ 1052238104Sdes lr_t *lrc = &itx->itx_lr; /* common log record */ 1053238104Sdes lr_write_t *lrw = (lr_write_t *)lrc; 1054238104Sdes char *lr_buf; 1055238104Sdes uint64_t txg = lrc->lrc_txg; 1056238104Sdes uint64_t reclen = lrc->lrc_reclen; 1057238104Sdes uint64_t dlen = 0; 1058238104Sdes 1059238104Sdes if (lwb == NULL) 1060238104Sdes return (NULL); 1061238104Sdes 1062238104Sdes ASSERT(lwb->lwb_buf != NULL); 1063238104Sdes ASSERT(zilog_is_dirty(zilog) || 1064238104Sdes spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1065238104Sdes 1066238104Sdes if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1067238104Sdes dlen = P2ROUNDUP_TYPED( 1068238104Sdes lrw->lr_length, sizeof (uint64_t), uint64_t); 1069238104Sdes 1070238104Sdes zilog->zl_cur_used += (reclen + dlen); 1071238104Sdes 1072238104Sdes zil_lwb_write_init(zilog, lwb); 1073238104Sdes 1074238104Sdes /* 1075238104Sdes * If this record won't fit in the current log block, start a new one. 1076238104Sdes */ 1077238104Sdes if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1078238104Sdes lwb = zil_lwb_write_start(zilog, lwb); 1079238104Sdes if (lwb == NULL) 1080238104Sdes return (NULL); 1081238104Sdes zil_lwb_write_init(zilog, lwb); 1082238104Sdes ASSERT(LWB_EMPTY(lwb)); 1083238104Sdes if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1084238104Sdes txg_wait_synced(zilog->zl_dmu_pool, txg); 1085238104Sdes return (lwb); 1086238104Sdes } 1087238104Sdes } 1088238104Sdes 1089238104Sdes lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1090238104Sdes bcopy(lrc, lr_buf, reclen); 1091238104Sdes lrc = (lr_t *)lr_buf; 1092238104Sdes lrw = (lr_write_t *)lrc; 1093238104Sdes 1094238104Sdes /* 1095238104Sdes * If it's a write, fetch the data or get its blkptr as appropriate. 1096238104Sdes */ 1097238104Sdes if (lrc->lrc_txtype == TX_WRITE) { 1098238104Sdes if (txg > spa_freeze_txg(zilog->zl_spa)) 1099238104Sdes txg_wait_synced(zilog->zl_dmu_pool, txg); 1100238104Sdes if (itx->itx_wr_state != WR_COPIED) { 1101238104Sdes char *dbuf; 1102238104Sdes int error; 1103238104Sdes 1104238104Sdes if (dlen) { 1105238104Sdes ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1106238104Sdes dbuf = lr_buf + reclen; 1107238104Sdes lrw->lr_common.lrc_reclen += dlen; 1108238104Sdes } else { 1109238104Sdes ASSERT(itx->itx_wr_state == WR_INDIRECT); 1110238104Sdes dbuf = NULL; 1111238104Sdes } 1112238104Sdes error = zilog->zl_get_data( 1113238104Sdes itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1114238104Sdes if (error == EIO) { 1115238104Sdes txg_wait_synced(zilog->zl_dmu_pool, txg); 1116238104Sdes return (lwb); 1117238104Sdes } 1118238104Sdes if (error != 0) { 1119238104Sdes ASSERT(error == ENOENT || error == EEXIST || 1120238104Sdes error == EALREADY); 1121238104Sdes return (lwb); 1122238104Sdes } 1123238104Sdes } 1124238104Sdes } 1125238104Sdes 1126238104Sdes /* 1127238104Sdes * We're actually making an entry, so update lrc_seq to be the 1128238104Sdes * log record sequence number. Note that this is generally not 1129238104Sdes * equal to the itx sequence number because not all transactions 1130238104Sdes * are synchronous, and sometimes spa_sync() gets there first. 1131238104Sdes */ 1132238104Sdes lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1133238104Sdes lwb->lwb_nused += reclen + dlen; 1134238104Sdes lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1135238104Sdes ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1136238104Sdes ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1137238104Sdes 1138238104Sdes return (lwb); 1139238104Sdes} 1140238104Sdes 1141238104Sdesitx_t * 1142238104Sdeszil_itx_create(uint64_t txtype, size_t lrsize) 1143238104Sdes{ 1144238104Sdes itx_t *itx; 1145238104Sdes 1146238104Sdes lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1147238104Sdes 1148238104Sdes itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1149238104Sdes itx->itx_lr.lrc_txtype = txtype; 1150238104Sdes itx->itx_lr.lrc_reclen = lrsize; 1151238104Sdes itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1152238104Sdes itx->itx_lr.lrc_seq = 0; /* defensive */ 1153238104Sdes itx->itx_sync = B_TRUE; /* default is synchronous */ 1154238104Sdes 1155238104Sdes return (itx); 1156238104Sdes} 1157238104Sdes 1158238104Sdesvoid 1159238104Sdeszil_itx_destroy(itx_t *itx) 1160238104Sdes{ 1161238104Sdes kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1162238104Sdes} 1163238104Sdes 1164238104Sdes/* 1165238104Sdes * Free up the sync and async itxs. The itxs_t has already been detached 1166238104Sdes * so no locks are needed. 1167238104Sdes */ 1168238104Sdesstatic void 1169238104Sdeszil_itxg_clean(itxs_t *itxs) 1170238104Sdes{ 1171238104Sdes itx_t *itx; 1172238104Sdes list_t *list; 1173238104Sdes avl_tree_t *t; 1174238104Sdes void *cookie; 1175238104Sdes itx_async_node_t *ian; 1176238104Sdes 1177238104Sdes list = &itxs->i_sync_list; 1178238104Sdes while ((itx = list_head(list)) != NULL) { 1179238104Sdes list_remove(list, itx); 1180238104Sdes kmem_free(itx, offsetof(itx_t, itx_lr) + 1181238104Sdes itx->itx_lr.lrc_reclen); 1182238104Sdes } 1183238104Sdes 1184238104Sdes cookie = NULL; 1185238104Sdes t = &itxs->i_async_tree; 1186238104Sdes while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1187238104Sdes list = &ian->ia_list; 1188238104Sdes while ((itx = list_head(list)) != NULL) { 1189238104Sdes list_remove(list, itx); 1190238104Sdes kmem_free(itx, offsetof(itx_t, itx_lr) + 1191238104Sdes itx->itx_lr.lrc_reclen); 1192238104Sdes } 1193238104Sdes list_destroy(list); 1194238104Sdes kmem_free(ian, sizeof (itx_async_node_t)); 1195238104Sdes } 1196238104Sdes avl_destroy(t); 1197238104Sdes 1198238104Sdes kmem_free(itxs, sizeof (itxs_t)); 1199238104Sdes} 1200238104Sdes 1201238104Sdesstatic int 1202238104Sdeszil_aitx_compare(const void *x1, const void *x2) 1203238104Sdes{ 1204238104Sdes const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1205238104Sdes const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1206238104Sdes 1207238104Sdes if (o1 < o2) 1208238104Sdes return (-1); 1209238104Sdes if (o1 > o2) 1210238104Sdes return (1); 1211238104Sdes 1212238104Sdes return (0); 1213238104Sdes} 1214238104Sdes 1215238104Sdes/* 1216238104Sdes * Remove all async itx with the given oid. 1217238104Sdes */ 1218238104Sdesstatic void 1219238104Sdeszil_remove_async(zilog_t *zilog, uint64_t oid) 1220238104Sdes{ 1221238104Sdes uint64_t otxg, txg; 1222238104Sdes itx_async_node_t *ian; 1223238104Sdes avl_tree_t *t; 1224238104Sdes avl_index_t where; 1225238104Sdes list_t clean_list; 1226238104Sdes itx_t *itx; 1227238104Sdes 1228238104Sdes ASSERT(oid != 0); 1229238104Sdes list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1230238104Sdes 1231238104Sdes if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1232238104Sdes otxg = ZILTEST_TXG; 1233238104Sdes else 1234238104Sdes otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1235238104Sdes 1236238104Sdes for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1237238104Sdes itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1238238104Sdes 1239238104Sdes mutex_enter(&itxg->itxg_lock); 1240238104Sdes if (itxg->itxg_txg != txg) { 1241238104Sdes mutex_exit(&itxg->itxg_lock); 1242238104Sdes continue; 1243238104Sdes } 1244238104Sdes 1245238104Sdes /* 1246238104Sdes * Locate the object node and append its list. 1247238104Sdes */ 1248238104Sdes t = &itxg->itxg_itxs->i_async_tree; 1249238104Sdes ian = avl_find(t, &oid, &where); 1250238104Sdes if (ian != NULL) 1251238104Sdes list_move_tail(&clean_list, &ian->ia_list); 1252238104Sdes mutex_exit(&itxg->itxg_lock); 1253238104Sdes } 1254238104Sdes while ((itx = list_head(&clean_list)) != NULL) { 1255238104Sdes list_remove(&clean_list, itx); 1256238104Sdes kmem_free(itx, offsetof(itx_t, itx_lr) + 1257238104Sdes itx->itx_lr.lrc_reclen); 1258238104Sdes } 1259238104Sdes list_destroy(&clean_list); 1260238104Sdes} 1261238104Sdes 1262238104Sdesvoid 1263238104Sdeszil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1264238104Sdes{ 1265238104Sdes uint64_t txg; 1266238104Sdes itxg_t *itxg; 1267238104Sdes itxs_t *itxs, *clean = NULL; 1268238104Sdes 1269238104Sdes /* 1270238104Sdes * Object ids can be re-instantiated in the next txg so 1271238104Sdes * remove any async transactions to avoid future leaks. 1272238104Sdes * This can happen if a fsync occurs on the re-instantiated 1273238104Sdes * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1274238104Sdes * the new file data and flushes a write record for the old object. 1275238104Sdes */ 1276238104Sdes if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1277238104Sdes zil_remove_async(zilog, itx->itx_oid); 1278238104Sdes 1279238104Sdes /* 1280238104Sdes * Ensure the data of a renamed file is committed before the rename. 1281238104Sdes */ 1282238104Sdes if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1283238104Sdes zil_async_to_sync(zilog, itx->itx_oid); 1284238104Sdes 1285238104Sdes if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1286238104Sdes txg = ZILTEST_TXG; 1287238104Sdes else 1288238104Sdes txg = dmu_tx_get_txg(tx); 1289238104Sdes 1290238104Sdes itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1291238104Sdes mutex_enter(&itxg->itxg_lock); 1292238104Sdes itxs = itxg->itxg_itxs; 1293238104Sdes if (itxg->itxg_txg != txg) { 1294238104Sdes if (itxs != NULL) { 1295238104Sdes /* 1296238104Sdes * The zil_clean callback hasn't got around to cleaning 1297238104Sdes * this itxg. Save the itxs for release below. 1298238104Sdes * This should be rare. 1299238104Sdes */ 1300238104Sdes atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1301238104Sdes itxg->itxg_sod = 0; 1302238104Sdes clean = itxg->itxg_itxs; 1303238104Sdes } 1304238104Sdes ASSERT(itxg->itxg_sod == 0); 1305238104Sdes itxg->itxg_txg = txg; 1306238104Sdes itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1307238104Sdes 1308238104Sdes list_create(&itxs->i_sync_list, sizeof (itx_t), 1309238104Sdes offsetof(itx_t, itx_node)); 1310238104Sdes avl_create(&itxs->i_async_tree, zil_aitx_compare, 1311238104Sdes sizeof (itx_async_node_t), 1312238104Sdes offsetof(itx_async_node_t, ia_node)); 1313238104Sdes } 1314238104Sdes if (itx->itx_sync) { 1315238104Sdes list_insert_tail(&itxs->i_sync_list, itx); 1316238104Sdes atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1317238104Sdes itxg->itxg_sod += itx->itx_sod; 1318238104Sdes } else { 1319238104Sdes avl_tree_t *t = &itxs->i_async_tree; 1320238104Sdes uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1321238104Sdes itx_async_node_t *ian; 1322238104Sdes avl_index_t where; 1323238104Sdes 1324238104Sdes ian = avl_find(t, &foid, &where); 1325238104Sdes if (ian == NULL) { 1326238104Sdes ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1327238104Sdes list_create(&ian->ia_list, sizeof (itx_t), 1328238104Sdes offsetof(itx_t, itx_node)); 1329238104Sdes ian->ia_foid = foid; 1330238104Sdes avl_insert(t, ian, where); 1331238104Sdes } 1332238104Sdes list_insert_tail(&ian->ia_list, itx); 1333238104Sdes } 1334238104Sdes 1335238104Sdes itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1336238104Sdes zilog_dirty(zilog, txg); 1337238104Sdes mutex_exit(&itxg->itxg_lock); 1338238104Sdes 1339238104Sdes /* Release the old itxs now we've dropped the lock */ 1340238104Sdes if (clean != NULL) 1341238104Sdes zil_itxg_clean(clean); 1342238104Sdes} 1343238104Sdes 1344238104Sdes/* 1345238104Sdes * If there are any in-memory intent log transactions which have now been 1346238104Sdes * synced then start up a taskq to free them. We should only do this after we 1347238104Sdes * have written out the uberblocks (i.e. txg has been comitted) so that 1348238104Sdes * don't inadvertently clean out in-memory log records that would be required 1349238104Sdes * by zil_commit(). 1350238104Sdes */ 1351238104Sdesvoid 1352238104Sdeszil_clean(zilog_t *zilog, uint64_t synced_txg) 1353238104Sdes{ 1354238104Sdes itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1355238104Sdes itxs_t *clean_me; 1356238104Sdes 1357238104Sdes mutex_enter(&itxg->itxg_lock); 1358238104Sdes if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1359238104Sdes mutex_exit(&itxg->itxg_lock); 1360238104Sdes return; 1361238104Sdes } 1362238104Sdes ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1363238104Sdes ASSERT(itxg->itxg_txg != 0); 1364238104Sdes ASSERT(zilog->zl_clean_taskq != NULL); 1365238104Sdes atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1366238104Sdes itxg->itxg_sod = 0; 1367238104Sdes clean_me = itxg->itxg_itxs; 1368238104Sdes itxg->itxg_itxs = NULL; 1369238104Sdes itxg->itxg_txg = 0; 1370238104Sdes mutex_exit(&itxg->itxg_lock); 1371238104Sdes /* 1372238104Sdes * Preferably start a task queue to free up the old itxs but 1373238104Sdes * if taskq_dispatch can't allocate resources to do that then 1374238104Sdes * free it in-line. This should be rare. Note, using TQ_SLEEP 1375238104Sdes * created a bad performance problem. 1376238104Sdes */ 1377238104Sdes if (taskq_dispatch(zilog->zl_clean_taskq, 1378238104Sdes (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1379238104Sdes zil_itxg_clean(clean_me); 1380238104Sdes} 1381238104Sdes 1382238104Sdes/* 1383238104Sdes * Get the list of itxs to commit into zl_itx_commit_list. 1384238104Sdes */ 1385238104Sdesstatic void 1386238104Sdeszil_get_commit_list(zilog_t *zilog) 1387238104Sdes{ 1388238104Sdes uint64_t otxg, txg; 1389238104Sdes list_t *commit_list = &zilog->zl_itx_commit_list; 1390238104Sdes uint64_t push_sod = 0; 1391238104Sdes 1392238104Sdes if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1393238104Sdes otxg = ZILTEST_TXG; 1394238104Sdes else 1395238104Sdes otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1396238104Sdes 1397238104Sdes for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1398238104Sdes itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1399238104Sdes 1400238104Sdes mutex_enter(&itxg->itxg_lock); 1401238104Sdes if (itxg->itxg_txg != txg) { 1402238104Sdes mutex_exit(&itxg->itxg_lock); 1403238104Sdes continue; 1404238104Sdes } 1405238104Sdes 1406238104Sdes list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1407238104Sdes push_sod += itxg->itxg_sod; 1408238104Sdes itxg->itxg_sod = 0; 1409238104Sdes 1410238104Sdes mutex_exit(&itxg->itxg_lock); 1411238104Sdes } 1412238104Sdes atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1413238104Sdes} 1414238104Sdes 1415238104Sdes/* 1416238104Sdes * Move the async itxs for a specified object to commit into sync lists. 1417238104Sdes */ 1418238104Sdesstatic void 1419238104Sdeszil_async_to_sync(zilog_t *zilog, uint64_t foid) 1420238104Sdes{ 1421 uint64_t otxg, txg; 1422 itx_async_node_t *ian; 1423 avl_tree_t *t; 1424 avl_index_t where; 1425 1426 if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1427 otxg = ZILTEST_TXG; 1428 else 1429 otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1430 1431 for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1432 itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1433 1434 mutex_enter(&itxg->itxg_lock); 1435 if (itxg->itxg_txg != txg) { 1436 mutex_exit(&itxg->itxg_lock); 1437 continue; 1438 } 1439 1440 /* 1441 * If a foid is specified then find that node and append its 1442 * list. Otherwise walk the tree appending all the lists 1443 * to the sync list. We add to the end rather than the 1444 * beginning to ensure the create has happened. 1445 */ 1446 t = &itxg->itxg_itxs->i_async_tree; 1447 if (foid != 0) { 1448 ian = avl_find(t, &foid, &where); 1449 if (ian != NULL) { 1450 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1451 &ian->ia_list); 1452 } 1453 } else { 1454 void *cookie = NULL; 1455 1456 while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1457 list_move_tail(&itxg->itxg_itxs->i_sync_list, 1458 &ian->ia_list); 1459 list_destroy(&ian->ia_list); 1460 kmem_free(ian, sizeof (itx_async_node_t)); 1461 } 1462 } 1463 mutex_exit(&itxg->itxg_lock); 1464 } 1465} 1466 1467static void 1468zil_commit_writer(zilog_t *zilog) 1469{ 1470 uint64_t txg; 1471 itx_t *itx; 1472 lwb_t *lwb; 1473 spa_t *spa = zilog->zl_spa; 1474 int error = 0; 1475 1476 ASSERT(zilog->zl_root_zio == NULL); 1477 1478 mutex_exit(&zilog->zl_lock); 1479 1480 zil_get_commit_list(zilog); 1481 1482 /* 1483 * Return if there's nothing to commit before we dirty the fs by 1484 * calling zil_create(). 1485 */ 1486 if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1487 mutex_enter(&zilog->zl_lock); 1488 return; 1489 } 1490 1491 if (zilog->zl_suspend) { 1492 lwb = NULL; 1493 } else { 1494 lwb = list_tail(&zilog->zl_lwb_list); 1495 if (lwb == NULL) 1496 lwb = zil_create(zilog); 1497 } 1498 1499 DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1500 while (itx = list_head(&zilog->zl_itx_commit_list)) { 1501 txg = itx->itx_lr.lrc_txg; 1502 ASSERT(txg); 1503 1504 if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1505 lwb = zil_lwb_commit(zilog, itx, lwb); 1506 list_remove(&zilog->zl_itx_commit_list, itx); 1507 kmem_free(itx, offsetof(itx_t, itx_lr) 1508 + itx->itx_lr.lrc_reclen); 1509 } 1510 DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1511 1512 /* write the last block out */ 1513 if (lwb != NULL && lwb->lwb_zio != NULL) 1514 lwb = zil_lwb_write_start(zilog, lwb); 1515 1516 zilog->zl_cur_used = 0; 1517 1518 /* 1519 * Wait if necessary for the log blocks to be on stable storage. 1520 */ 1521 if (zilog->zl_root_zio) { 1522 error = zio_wait(zilog->zl_root_zio); 1523 zilog->zl_root_zio = NULL; 1524 zil_flush_vdevs(zilog); 1525 } 1526 1527 if (error || lwb == NULL) 1528 txg_wait_synced(zilog->zl_dmu_pool, 0); 1529 1530 mutex_enter(&zilog->zl_lock); 1531 1532 /* 1533 * Remember the highest committed log sequence number for ztest. 1534 * We only update this value when all the log writes succeeded, 1535 * because ztest wants to ASSERT that it got the whole log chain. 1536 */ 1537 if (error == 0 && lwb != NULL) 1538 zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1539} 1540 1541/* 1542 * Commit zfs transactions to stable storage. 1543 * If foid is 0 push out all transactions, otherwise push only those 1544 * for that object or might reference that object. 1545 * 1546 * itxs are committed in batches. In a heavily stressed zil there will be 1547 * a commit writer thread who is writing out a bunch of itxs to the log 1548 * for a set of committing threads (cthreads) in the same batch as the writer. 1549 * Those cthreads are all waiting on the same cv for that batch. 1550 * 1551 * There will also be a different and growing batch of threads that are 1552 * waiting to commit (qthreads). When the committing batch completes 1553 * a transition occurs such that the cthreads exit and the qthreads become 1554 * cthreads. One of the new cthreads becomes the writer thread for the 1555 * batch. Any new threads arriving become new qthreads. 1556 * 1557 * Only 2 condition variables are needed and there's no transition 1558 * between the two cvs needed. They just flip-flop between qthreads 1559 * and cthreads. 1560 * 1561 * Using this scheme we can efficiently wakeup up only those threads 1562 * that have been committed. 1563 */ 1564void 1565zil_commit(zilog_t *zilog, uint64_t foid) 1566{ 1567 uint64_t mybatch; 1568 1569 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1570 return; 1571 1572 /* move the async itxs for the foid to the sync queues */ 1573 zil_async_to_sync(zilog, foid); 1574 1575 mutex_enter(&zilog->zl_lock); 1576 mybatch = zilog->zl_next_batch; 1577 while (zilog->zl_writer) { 1578 cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1579 if (mybatch <= zilog->zl_com_batch) { 1580 mutex_exit(&zilog->zl_lock); 1581 return; 1582 } 1583 } 1584 1585 zilog->zl_next_batch++; 1586 zilog->zl_writer = B_TRUE; 1587 zil_commit_writer(zilog); 1588 zilog->zl_com_batch = mybatch; 1589 zilog->zl_writer = B_FALSE; 1590 mutex_exit(&zilog->zl_lock); 1591 1592 /* wake up one thread to become the next writer */ 1593 cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1594 1595 /* wake up all threads waiting for this batch to be committed */ 1596 cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1597} 1598 1599/* 1600 * Called in syncing context to free committed log blocks and update log header. 1601 */ 1602void 1603zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1604{ 1605 zil_header_t *zh = zil_header_in_syncing_context(zilog); 1606 uint64_t txg = dmu_tx_get_txg(tx); 1607 spa_t *spa = zilog->zl_spa; 1608 uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1609 lwb_t *lwb; 1610 1611 /* 1612 * We don't zero out zl_destroy_txg, so make sure we don't try 1613 * to destroy it twice. 1614 */ 1615 if (spa_sync_pass(spa) != 1) 1616 return; 1617 1618 mutex_enter(&zilog->zl_lock); 1619 1620 ASSERT(zilog->zl_stop_sync == 0); 1621 1622 if (*replayed_seq != 0) { 1623 ASSERT(zh->zh_replay_seq < *replayed_seq); 1624 zh->zh_replay_seq = *replayed_seq; 1625 *replayed_seq = 0; 1626 } 1627 1628 if (zilog->zl_destroy_txg == txg) { 1629 blkptr_t blk = zh->zh_log; 1630 1631 ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1632 1633 bzero(zh, sizeof (zil_header_t)); 1634 bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1635 1636 if (zilog->zl_keep_first) { 1637 /* 1638 * If this block was part of log chain that couldn't 1639 * be claimed because a device was missing during 1640 * zil_claim(), but that device later returns, 1641 * then this block could erroneously appear valid. 1642 * To guard against this, assign a new GUID to the new 1643 * log chain so it doesn't matter what blk points to. 1644 */ 1645 zil_init_log_chain(zilog, &blk); 1646 zh->zh_log = blk; 1647 } 1648 } 1649 1650 while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1651 zh->zh_log = lwb->lwb_blk; 1652 if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1653 break; 1654 list_remove(&zilog->zl_lwb_list, lwb); 1655 zio_free_zil(spa, txg, &lwb->lwb_blk); 1656 kmem_cache_free(zil_lwb_cache, lwb); 1657 1658 /* 1659 * If we don't have anything left in the lwb list then 1660 * we've had an allocation failure and we need to zero 1661 * out the zil_header blkptr so that we don't end 1662 * up freeing the same block twice. 1663 */ 1664 if (list_head(&zilog->zl_lwb_list) == NULL) 1665 BP_ZERO(&zh->zh_log); 1666 } 1667 mutex_exit(&zilog->zl_lock); 1668} 1669 1670void 1671zil_init(void) 1672{ 1673 zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1674 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1675} 1676 1677void 1678zil_fini(void) 1679{ 1680 kmem_cache_destroy(zil_lwb_cache); 1681} 1682 1683void 1684zil_set_sync(zilog_t *zilog, uint64_t sync) 1685{ 1686 zilog->zl_sync = sync; 1687} 1688 1689void 1690zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1691{ 1692 zilog->zl_logbias = logbias; 1693} 1694 1695zilog_t * 1696zil_alloc(objset_t *os, zil_header_t *zh_phys) 1697{ 1698 zilog_t *zilog; 1699 1700 zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1701 1702 zilog->zl_header = zh_phys; 1703 zilog->zl_os = os; 1704 zilog->zl_spa = dmu_objset_spa(os); 1705 zilog->zl_dmu_pool = dmu_objset_pool(os); 1706 zilog->zl_destroy_txg = TXG_INITIAL - 1; 1707 zilog->zl_logbias = dmu_objset_logbias(os); 1708 zilog->zl_sync = dmu_objset_syncprop(os); 1709 zilog->zl_next_batch = 1; 1710 1711 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1712 1713 for (int i = 0; i < TXG_SIZE; i++) { 1714 mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1715 MUTEX_DEFAULT, NULL); 1716 } 1717 1718 list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1719 offsetof(lwb_t, lwb_node)); 1720 1721 list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1722 offsetof(itx_t, itx_node)); 1723 1724 mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1725 1726 avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1727 sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1728 1729 cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1730 cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1731 cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1732 cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1733 1734 return (zilog); 1735} 1736 1737void 1738zil_free(zilog_t *zilog) 1739{ 1740 zilog->zl_stop_sync = 1; 1741 1742 ASSERT0(zilog->zl_suspend); 1743 ASSERT0(zilog->zl_suspending); 1744 1745 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1746 list_destroy(&zilog->zl_lwb_list); 1747 1748 avl_destroy(&zilog->zl_vdev_tree); 1749 mutex_destroy(&zilog->zl_vdev_lock); 1750 1751 ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1752 list_destroy(&zilog->zl_itx_commit_list); 1753 1754 for (int i = 0; i < TXG_SIZE; i++) { 1755 /* 1756 * It's possible for an itx to be generated that doesn't dirty 1757 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1758 * callback to remove the entry. We remove those here. 1759 * 1760 * Also free up the ziltest itxs. 1761 */ 1762 if (zilog->zl_itxg[i].itxg_itxs) 1763 zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1764 mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1765 } 1766 1767 mutex_destroy(&zilog->zl_lock); 1768 1769 cv_destroy(&zilog->zl_cv_writer); 1770 cv_destroy(&zilog->zl_cv_suspend); 1771 cv_destroy(&zilog->zl_cv_batch[0]); 1772 cv_destroy(&zilog->zl_cv_batch[1]); 1773 1774 kmem_free(zilog, sizeof (zilog_t)); 1775} 1776 1777/* 1778 * Open an intent log. 1779 */ 1780zilog_t * 1781zil_open(objset_t *os, zil_get_data_t *get_data) 1782{ 1783 zilog_t *zilog = dmu_objset_zil(os); 1784 1785 ASSERT(zilog->zl_clean_taskq == NULL); 1786 ASSERT(zilog->zl_get_data == NULL); 1787 ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1788 1789 zilog->zl_get_data = get_data; 1790 zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1791 2, 2, TASKQ_PREPOPULATE); 1792 1793 return (zilog); 1794} 1795 1796/* 1797 * Close an intent log. 1798 */ 1799void 1800zil_close(zilog_t *zilog) 1801{ 1802 lwb_t *lwb; 1803 uint64_t txg = 0; 1804 1805 zil_commit(zilog, 0); /* commit all itx */ 1806 1807 /* 1808 * The lwb_max_txg for the stubby lwb will reflect the last activity 1809 * for the zil. After a txg_wait_synced() on the txg we know all the 1810 * callbacks have occurred that may clean the zil. Only then can we 1811 * destroy the zl_clean_taskq. 1812 */ 1813 mutex_enter(&zilog->zl_lock); 1814 lwb = list_tail(&zilog->zl_lwb_list); 1815 if (lwb != NULL) 1816 txg = lwb->lwb_max_txg; 1817 mutex_exit(&zilog->zl_lock); 1818 if (txg) 1819 txg_wait_synced(zilog->zl_dmu_pool, txg); 1820 ASSERT(!zilog_is_dirty(zilog)); 1821 1822 taskq_destroy(zilog->zl_clean_taskq); 1823 zilog->zl_clean_taskq = NULL; 1824 zilog->zl_get_data = NULL; 1825 1826 /* 1827 * We should have only one LWB left on the list; remove it now. 1828 */ 1829 mutex_enter(&zilog->zl_lock); 1830 lwb = list_head(&zilog->zl_lwb_list); 1831 if (lwb != NULL) { 1832 ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1833 list_remove(&zilog->zl_lwb_list, lwb); 1834 zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1835 kmem_cache_free(zil_lwb_cache, lwb); 1836 } 1837 mutex_exit(&zilog->zl_lock); 1838} 1839 1840static char *suspend_tag = "zil suspending"; 1841 1842/* 1843 * Suspend an intent log. While in suspended mode, we still honor 1844 * synchronous semantics, but we rely on txg_wait_synced() to do it. 1845 * On old version pools, we suspend the log briefly when taking a 1846 * snapshot so that it will have an empty intent log. 1847 * 1848 * Long holds are not really intended to be used the way we do here -- 1849 * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1850 * could fail. Therefore we take pains to only put a long hold if it is 1851 * actually necessary. Fortunately, it will only be necessary if the 1852 * objset is currently mounted (or the ZVOL equivalent). In that case it 1853 * will already have a long hold, so we are not really making things any worse. 1854 * 1855 * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1856 * zvol_state_t), and use their mechanism to prevent their hold from being 1857 * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1858 * very little gain. 1859 * 1860 * if cookiep == NULL, this does both the suspend & resume. 1861 * Otherwise, it returns with the dataset "long held", and the cookie 1862 * should be passed into zil_resume(). 1863 */ 1864int 1865zil_suspend(const char *osname, void **cookiep) 1866{ 1867 objset_t *os; 1868 zilog_t *zilog; 1869 const zil_header_t *zh; 1870 int error; 1871 1872 error = dmu_objset_hold(osname, suspend_tag, &os); 1873 if (error != 0) 1874 return (error); 1875 zilog = dmu_objset_zil(os); 1876 1877 mutex_enter(&zilog->zl_lock); 1878 zh = zilog->zl_header; 1879 1880 if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1881 mutex_exit(&zilog->zl_lock); 1882 dmu_objset_rele(os, suspend_tag); 1883 return (SET_ERROR(EBUSY)); 1884 } 1885 1886 /* 1887 * Don't put a long hold in the cases where we can avoid it. This 1888 * is when there is no cookie so we are doing a suspend & resume 1889 * (i.e. called from zil_vdev_offline()), and there's nothing to do 1890 * for the suspend because it's already suspended, or there's no ZIL. 1891 */ 1892 if (cookiep == NULL && !zilog->zl_suspending && 1893 (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1894 mutex_exit(&zilog->zl_lock); 1895 dmu_objset_rele(os, suspend_tag); 1896 return (0); 1897 } 1898 1899 dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1900 dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1901 1902 zilog->zl_suspend++; 1903 1904 if (zilog->zl_suspend > 1) { 1905 /* 1906 * Someone else is already suspending it. 1907 * Just wait for them to finish. 1908 */ 1909 1910 while (zilog->zl_suspending) 1911 cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1912 mutex_exit(&zilog->zl_lock); 1913 1914 if (cookiep == NULL) 1915 zil_resume(os); 1916 else 1917 *cookiep = os; 1918 return (0); 1919 } 1920 1921 /* 1922 * If there is no pointer to an on-disk block, this ZIL must not 1923 * be active (e.g. filesystem not mounted), so there's nothing 1924 * to clean up. 1925 */ 1926 if (BP_IS_HOLE(&zh->zh_log)) { 1927 ASSERT(cookiep != NULL); /* fast path already handled */ 1928 1929 *cookiep = os; 1930 mutex_exit(&zilog->zl_lock); 1931 return (0); 1932 } 1933 1934 zilog->zl_suspending = B_TRUE; 1935 mutex_exit(&zilog->zl_lock); 1936 1937 zil_commit(zilog, 0); 1938 1939 zil_destroy(zilog, B_FALSE); 1940 1941 mutex_enter(&zilog->zl_lock); 1942 zilog->zl_suspending = B_FALSE; 1943 cv_broadcast(&zilog->zl_cv_suspend); 1944 mutex_exit(&zilog->zl_lock); 1945 1946 if (cookiep == NULL) 1947 zil_resume(os); 1948 else 1949 *cookiep = os; 1950 return (0); 1951} 1952 1953void 1954zil_resume(void *cookie) 1955{ 1956 objset_t *os = cookie; 1957 zilog_t *zilog = dmu_objset_zil(os); 1958 1959 mutex_enter(&zilog->zl_lock); 1960 ASSERT(zilog->zl_suspend != 0); 1961 zilog->zl_suspend--; 1962 mutex_exit(&zilog->zl_lock); 1963 dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1964 dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1965} 1966 1967typedef struct zil_replay_arg { 1968 zil_replay_func_t **zr_replay; 1969 void *zr_arg; 1970 boolean_t zr_byteswap; 1971 char *zr_lr; 1972} zil_replay_arg_t; 1973 1974static int 1975zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1976{ 1977 char name[ZFS_MAX_DATASET_NAME_LEN]; 1978 1979 zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1980 1981 dmu_objset_name(zilog->zl_os, name); 1982 1983 cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1984 "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1985 (u_longlong_t)lr->lrc_seq, 1986 (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1987 (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1988 1989 return (error); 1990} 1991 1992static int 1993zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1994{ 1995 zil_replay_arg_t *zr = zra; 1996 const zil_header_t *zh = zilog->zl_header; 1997 uint64_t reclen = lr->lrc_reclen; 1998 uint64_t txtype = lr->lrc_txtype; 1999 int error = 0; 2000 2001 zilog->zl_replaying_seq = lr->lrc_seq; 2002 2003 if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2004 return (0); 2005 2006 if (lr->lrc_txg < claim_txg) /* already committed */ 2007 return (0); 2008 2009 /* Strip case-insensitive bit, still present in log record */ 2010 txtype &= ~TX_CI; 2011 2012 if (txtype == 0 || txtype >= TX_MAX_TYPE) 2013 return (zil_replay_error(zilog, lr, EINVAL)); 2014 2015 /* 2016 * If this record type can be logged out of order, the object 2017 * (lr_foid) may no longer exist. That's legitimate, not an error. 2018 */ 2019 if (TX_OOO(txtype)) { 2020 error = dmu_object_info(zilog->zl_os, 2021 ((lr_ooo_t *)lr)->lr_foid, NULL); 2022 if (error == ENOENT || error == EEXIST) 2023 return (0); 2024 } 2025 2026 /* 2027 * Make a copy of the data so we can revise and extend it. 2028 */ 2029 bcopy(lr, zr->zr_lr, reclen); 2030 2031 /* 2032 * If this is a TX_WRITE with a blkptr, suck in the data. 2033 */ 2034 if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2035 error = zil_read_log_data(zilog, (lr_write_t *)lr, 2036 zr->zr_lr + reclen); 2037 if (error != 0) 2038 return (zil_replay_error(zilog, lr, error)); 2039 } 2040 2041 /* 2042 * The log block containing this lr may have been byteswapped 2043 * so that we can easily examine common fields like lrc_txtype. 2044 * However, the log is a mix of different record types, and only the 2045 * replay vectors know how to byteswap their records. Therefore, if 2046 * the lr was byteswapped, undo it before invoking the replay vector. 2047 */ 2048 if (zr->zr_byteswap) 2049 byteswap_uint64_array(zr->zr_lr, reclen); 2050 2051 /* 2052 * We must now do two things atomically: replay this log record, 2053 * and update the log header sequence number to reflect the fact that 2054 * we did so. At the end of each replay function the sequence number 2055 * is updated if we are in replay mode. 2056 */ 2057 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2058 if (error != 0) { 2059 /* 2060 * The DMU's dnode layer doesn't see removes until the txg 2061 * commits, so a subsequent claim can spuriously fail with 2062 * EEXIST. So if we receive any error we try syncing out 2063 * any removes then retry the transaction. Note that we 2064 * specify B_FALSE for byteswap now, so we don't do it twice. 2065 */ 2066 txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2067 error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2068 if (error != 0) 2069 return (zil_replay_error(zilog, lr, error)); 2070 } 2071 return (0); 2072} 2073 2074/* ARGSUSED */ 2075static int 2076zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2077{ 2078 zilog->zl_replay_blks++; 2079 2080 return (0); 2081} 2082 2083/* 2084 * If this dataset has a non-empty intent log, replay it and destroy it. 2085 */ 2086void 2087zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2088{ 2089 zilog_t *zilog = dmu_objset_zil(os); 2090 const zil_header_t *zh = zilog->zl_header; 2091 zil_replay_arg_t zr; 2092 2093 if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2094 zil_destroy(zilog, B_TRUE); 2095 return; 2096 } 2097 2098 zr.zr_replay = replay_func; 2099 zr.zr_arg = arg; 2100 zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2101 zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2102 2103 /* 2104 * Wait for in-progress removes to sync before starting replay. 2105 */ 2106 txg_wait_synced(zilog->zl_dmu_pool, 0); 2107 2108 zilog->zl_replay = B_TRUE; 2109 zilog->zl_replay_time = ddi_get_lbolt(); 2110 ASSERT(zilog->zl_replay_blks == 0); 2111 (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2112 zh->zh_claim_txg); 2113 kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2114 2115 zil_destroy(zilog, B_FALSE); 2116 txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2117 zilog->zl_replay = B_FALSE; 2118} 2119 2120boolean_t 2121zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2122{ 2123 if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2124 return (B_TRUE); 2125 2126 if (zilog->zl_replay) { 2127 dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2128 zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2129 zilog->zl_replaying_seq; 2130 return (B_TRUE); 2131 } 2132 2133 return (B_FALSE); 2134} 2135 2136/* ARGSUSED */ 2137int 2138zil_vdev_offline(const char *osname, void *arg) 2139{ 2140 int error; 2141 2142 error = zil_suspend(osname, NULL); 2143 if (error != 0) 2144 return (SET_ERROR(EEXIST)); 2145 return (0); 2146} 2147