1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5fe9cf88cSperrin * Common Development and Distribution License (the "License"). 6fe9cf88cSperrin * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 22ab69d62fSMatthew Ahrens * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 26fa9e4066Sahrens #include <sys/zfs_context.h> 27fa9e4066Sahrens #include <sys/spa.h> 28*e6ca193dSGeorge Wilson #include <sys/spa_impl.h> 29fa9e4066Sahrens #include <sys/dmu.h> 30fa9e4066Sahrens #include <sys/zap.h> 31fa9e4066Sahrens #include <sys/arc.h> 32fa9e4066Sahrens #include <sys/stat.h> 33fa9e4066Sahrens #include <sys/resource.h> 34fa9e4066Sahrens #include <sys/zil.h> 35fa9e4066Sahrens #include <sys/zil_impl.h> 36fa9e4066Sahrens #include <sys/dsl_dataset.h> 37fa9e4066Sahrens #include <sys/vdev.h> 38d63d470bSgw #include <sys/dmu_tx.h> 39fa9e4066Sahrens 40fa9e4066Sahrens /* 41fa9e4066Sahrens * The zfs intent log (ZIL) saves transaction records of system calls 42fa9e4066Sahrens * that change the file system in memory with enough information 43fa9e4066Sahrens * to be able to replay them. These are stored in memory until 44fa9e4066Sahrens * either the DMU transaction group (txg) commits them to the stable pool 45fa9e4066Sahrens * and they can be discarded, or they are flushed to the stable log 46fa9e4066Sahrens * (also in the pool) due to a fsync, O_DSYNC or other synchronous 47fa9e4066Sahrens * requirement. In the event of a panic or power fail then those log 48fa9e4066Sahrens * records (transactions) are replayed. 49fa9e4066Sahrens * 50fa9e4066Sahrens * There is one ZIL per file system. Its on-disk (pool) format consists 51fa9e4066Sahrens * of 3 parts: 52fa9e4066Sahrens * 53fa9e4066Sahrens * - ZIL header 54fa9e4066Sahrens * - ZIL blocks 55fa9e4066Sahrens * - ZIL records 56fa9e4066Sahrens * 57fa9e4066Sahrens * A log record holds a system call transaction. Log blocks can 58fa9e4066Sahrens * hold many log records and the blocks are chained together. 59fa9e4066Sahrens * Each ZIL block contains a block pointer (blkptr_t) to the next 60fa9e4066Sahrens * ZIL block in the chain. The ZIL header points to the first 61fa9e4066Sahrens * block in the chain. Note there is not a fixed place in the pool 62fa9e4066Sahrens * to hold blocks. They are dynamically allocated and freed as 63fa9e4066Sahrens * needed from the blocks available. Figure X shows the ZIL structure: 64fa9e4066Sahrens */ 65fa9e4066Sahrens 66fa9e4066Sahrens /* 67416e0cd8Sek * This global ZIL switch affects all pools 68fa9e4066Sahrens */ 69fa9e4066Sahrens int zil_disable = 0; /* disable intent logging */ 70416e0cd8Sek 71416e0cd8Sek /* 72416e0cd8Sek * Tunable parameter for debugging or performance analysis. Setting 73416e0cd8Sek * zfs_nocacheflush will cause corruption on power loss if a volatile 74416e0cd8Sek * out-of-order write cache is enabled. 75416e0cd8Sek */ 76416e0cd8Sek boolean_t zfs_nocacheflush = B_FALSE; 77fa9e4066Sahrens 78fa9e4066Sahrens static kmem_cache_t *zil_lwb_cache; 79fa9e4066Sahrens 80fa9e4066Sahrens static int 81fa9e4066Sahrens zil_dva_compare(const void *x1, const void *x2) 82fa9e4066Sahrens { 83fa9e4066Sahrens const dva_t *dva1 = x1; 84fa9e4066Sahrens const dva_t *dva2 = x2; 85fa9e4066Sahrens 86fa9e4066Sahrens if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 87fa9e4066Sahrens return (-1); 88fa9e4066Sahrens if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 89fa9e4066Sahrens return (1); 90fa9e4066Sahrens 91fa9e4066Sahrens if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 92fa9e4066Sahrens return (-1); 93fa9e4066Sahrens if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 94fa9e4066Sahrens return (1); 95fa9e4066Sahrens 96fa9e4066Sahrens return (0); 97fa9e4066Sahrens } 98fa9e4066Sahrens 99fa9e4066Sahrens static void 100fa9e4066Sahrens zil_dva_tree_init(avl_tree_t *t) 101fa9e4066Sahrens { 102fa9e4066Sahrens avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t), 103fa9e4066Sahrens offsetof(zil_dva_node_t, zn_node)); 104fa9e4066Sahrens } 105fa9e4066Sahrens 106fa9e4066Sahrens static void 107fa9e4066Sahrens zil_dva_tree_fini(avl_tree_t *t) 108fa9e4066Sahrens { 109fa9e4066Sahrens zil_dva_node_t *zn; 110fa9e4066Sahrens void *cookie = NULL; 111fa9e4066Sahrens 112fa9e4066Sahrens while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 113fa9e4066Sahrens kmem_free(zn, sizeof (zil_dva_node_t)); 114fa9e4066Sahrens 115fa9e4066Sahrens avl_destroy(t); 116fa9e4066Sahrens } 117fa9e4066Sahrens 118fa9e4066Sahrens static int 119fa9e4066Sahrens zil_dva_tree_add(avl_tree_t *t, dva_t *dva) 120fa9e4066Sahrens { 121fa9e4066Sahrens zil_dva_node_t *zn; 122fa9e4066Sahrens avl_index_t where; 123fa9e4066Sahrens 124fa9e4066Sahrens if (avl_find(t, dva, &where) != NULL) 125fa9e4066Sahrens return (EEXIST); 126fa9e4066Sahrens 127fa9e4066Sahrens zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP); 128fa9e4066Sahrens zn->zn_dva = *dva; 129fa9e4066Sahrens avl_insert(t, zn, where); 130fa9e4066Sahrens 131fa9e4066Sahrens return (0); 132fa9e4066Sahrens } 133fa9e4066Sahrens 134d80c45e0Sbonwick static zil_header_t * 135d80c45e0Sbonwick zil_header_in_syncing_context(zilog_t *zilog) 136d80c45e0Sbonwick { 137d80c45e0Sbonwick return ((zil_header_t *)zilog->zl_header); 138d80c45e0Sbonwick } 139d80c45e0Sbonwick 140d80c45e0Sbonwick static void 141d80c45e0Sbonwick zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 142d80c45e0Sbonwick { 143d80c45e0Sbonwick zio_cksum_t *zc = &bp->blk_cksum; 144d80c45e0Sbonwick 145d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 146d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 147d80c45e0Sbonwick zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 148d80c45e0Sbonwick zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 149d80c45e0Sbonwick } 150d80c45e0Sbonwick 151fa9e4066Sahrens /* 152fa9e4066Sahrens * Read a log block, make sure it's valid, and byteswap it if necessary. 153fa9e4066Sahrens */ 154fa9e4066Sahrens static int 155d80c45e0Sbonwick zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp) 156fa9e4066Sahrens { 157d80c45e0Sbonwick blkptr_t blk = *bp; 158ea8dc4b6Seschrock zbookmark_t zb; 15913506d1eSmaybee uint32_t aflags = ARC_WAIT; 160fa9e4066Sahrens int error; 161fa9e4066Sahrens 162d80c45e0Sbonwick zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET]; 163ea8dc4b6Seschrock zb.zb_object = 0; 164ea8dc4b6Seschrock zb.zb_level = -1; 165d80c45e0Sbonwick zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ]; 166ea8dc4b6Seschrock 167d80c45e0Sbonwick *abufpp = NULL; 168fa9e4066Sahrens 169088f3894Sahrens /* 170088f3894Sahrens * We shouldn't be doing any scrubbing while we're doing log 171088f3894Sahrens * replay, it's OK to not lock. 172088f3894Sahrens */ 173088f3894Sahrens error = arc_read_nolock(NULL, zilog->zl_spa, &blk, 174d80c45e0Sbonwick arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL | 17513506d1eSmaybee ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb); 176fa9e4066Sahrens 177d80c45e0Sbonwick if (error == 0) { 178d80c45e0Sbonwick char *data = (*abufpp)->b_data; 179d80c45e0Sbonwick uint64_t blksz = BP_GET_LSIZE(bp); 180d80c45e0Sbonwick zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1; 181d80c45e0Sbonwick zio_cksum_t cksum = bp->blk_cksum; 182fa9e4066Sahrens 183d80c45e0Sbonwick /* 184f5e6e722SNeil Perrin * Validate the checksummed log block. 185f5e6e722SNeil Perrin * 186d80c45e0Sbonwick * Sequence numbers should be... sequential. The checksum 187d80c45e0Sbonwick * verifier for the next block should be bp's checksum plus 1. 188f5e6e722SNeil Perrin * 189f5e6e722SNeil Perrin * Also check the log chain linkage and size used. 190d80c45e0Sbonwick */ 191d80c45e0Sbonwick cksum.zc_word[ZIL_ZC_SEQ]++; 192d80c45e0Sbonwick 193f5e6e722SNeil Perrin if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, 194f5e6e722SNeil Perrin sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) || 195f5e6e722SNeil Perrin (ztp->zit_nused > (blksz - sizeof (zil_trailer_t)))) { 196f5e6e722SNeil Perrin error = ECKSUM; 197f5e6e722SNeil Perrin } 198fa9e4066Sahrens 199d80c45e0Sbonwick if (error) { 200d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1); 201d80c45e0Sbonwick *abufpp = NULL; 202d80c45e0Sbonwick } 203fa9e4066Sahrens } 204fa9e4066Sahrens 205d80c45e0Sbonwick dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid); 206fa9e4066Sahrens 207d80c45e0Sbonwick return (error); 208fa9e4066Sahrens } 209fa9e4066Sahrens 210fa9e4066Sahrens /* 211fa9e4066Sahrens * Parse the intent log, and call parse_func for each valid record within. 212d80c45e0Sbonwick * Return the highest sequence number. 213fa9e4066Sahrens */ 214d80c45e0Sbonwick uint64_t 215fa9e4066Sahrens zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 216fa9e4066Sahrens zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 217fa9e4066Sahrens { 218d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 219d80c45e0Sbonwick uint64_t claim_seq = zh->zh_claim_seq; 220d80c45e0Sbonwick uint64_t seq = 0; 221d80c45e0Sbonwick uint64_t max_seq = 0; 222d80c45e0Sbonwick blkptr_t blk = zh->zh_log; 223d80c45e0Sbonwick arc_buf_t *abuf; 224fa9e4066Sahrens char *lrbuf, *lrp; 225fa9e4066Sahrens zil_trailer_t *ztp; 226fa9e4066Sahrens int reclen, error; 227fa9e4066Sahrens 228fa9e4066Sahrens if (BP_IS_HOLE(&blk)) 229d80c45e0Sbonwick return (max_seq); 230fa9e4066Sahrens 231fa9e4066Sahrens /* 232fa9e4066Sahrens * Starting at the block pointed to by zh_log we read the log chain. 233fa9e4066Sahrens * For each block in the chain we strongly check that block to 234fa9e4066Sahrens * ensure its validity. We stop when an invalid block is found. 235fa9e4066Sahrens * For each block pointer in the chain we call parse_blk_func(). 236fa9e4066Sahrens * For each record in each valid block we call parse_lr_func(). 237d80c45e0Sbonwick * If the log has been claimed, stop if we encounter a sequence 238d80c45e0Sbonwick * number greater than the highest claimed sequence number. 239fa9e4066Sahrens */ 240fa9e4066Sahrens zil_dva_tree_init(&zilog->zl_dva_tree); 241fa9e4066Sahrens for (;;) { 242d80c45e0Sbonwick seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 243d80c45e0Sbonwick 244d80c45e0Sbonwick if (claim_seq != 0 && seq > claim_seq) 245d80c45e0Sbonwick break; 246d80c45e0Sbonwick 247d80c45e0Sbonwick ASSERT(max_seq < seq); 248d80c45e0Sbonwick max_seq = seq; 249d80c45e0Sbonwick 250d80c45e0Sbonwick error = zil_read_log_block(zilog, &blk, &abuf); 251fa9e4066Sahrens 252fa9e4066Sahrens if (parse_blk_func != NULL) 253fa9e4066Sahrens parse_blk_func(zilog, &blk, arg, txg); 254fa9e4066Sahrens 255fa9e4066Sahrens if (error) 256fa9e4066Sahrens break; 257fa9e4066Sahrens 258d80c45e0Sbonwick lrbuf = abuf->b_data; 259fa9e4066Sahrens ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 260fa9e4066Sahrens blk = ztp->zit_next_blk; 261fa9e4066Sahrens 262d80c45e0Sbonwick if (parse_lr_func == NULL) { 263d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 264fa9e4066Sahrens continue; 265d80c45e0Sbonwick } 266fa9e4066Sahrens 267fa9e4066Sahrens for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { 268fa9e4066Sahrens lr_t *lr = (lr_t *)lrp; 269fa9e4066Sahrens reclen = lr->lrc_reclen; 270fa9e4066Sahrens ASSERT3U(reclen, >=, sizeof (lr_t)); 271fa9e4066Sahrens parse_lr_func(zilog, lr, arg, txg); 272fa9e4066Sahrens } 273d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 274fa9e4066Sahrens } 275fa9e4066Sahrens zil_dva_tree_fini(&zilog->zl_dva_tree); 276d80c45e0Sbonwick 277d80c45e0Sbonwick return (max_seq); 278fa9e4066Sahrens } 279fa9e4066Sahrens 280fa9e4066Sahrens /* ARGSUSED */ 281fa9e4066Sahrens static void 282fa9e4066Sahrens zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 283fa9e4066Sahrens { 284fa9e4066Sahrens spa_t *spa = zilog->zl_spa; 285fa9e4066Sahrens int err; 286fa9e4066Sahrens 287fa9e4066Sahrens /* 288fa9e4066Sahrens * Claim log block if not already committed and not already claimed. 289fa9e4066Sahrens */ 290fa9e4066Sahrens if (bp->blk_birth >= first_txg && 291fa9e4066Sahrens zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) { 292e14bb325SJeff Bonwick err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL, 293e14bb325SJeff Bonwick ZIO_FLAG_MUSTSUCCEED)); 294fa9e4066Sahrens ASSERT(err == 0); 295fa9e4066Sahrens } 296fa9e4066Sahrens } 297fa9e4066Sahrens 298fa9e4066Sahrens static void 299fa9e4066Sahrens zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 300fa9e4066Sahrens { 301fa9e4066Sahrens if (lrc->lrc_txtype == TX_WRITE) { 302fa9e4066Sahrens lr_write_t *lr = (lr_write_t *)lrc; 303fa9e4066Sahrens zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg); 304fa9e4066Sahrens } 305fa9e4066Sahrens } 306fa9e4066Sahrens 307fa9e4066Sahrens /* ARGSUSED */ 308fa9e4066Sahrens static void 309fa9e4066Sahrens zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 310fa9e4066Sahrens { 311fa9e4066Sahrens zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx)); 312fa9e4066Sahrens } 313fa9e4066Sahrens 314fa9e4066Sahrens static void 315fa9e4066Sahrens zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 316fa9e4066Sahrens { 317fa9e4066Sahrens /* 318fa9e4066Sahrens * If we previously claimed it, we need to free it. 319fa9e4066Sahrens */ 320fa9e4066Sahrens if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) { 321fa9e4066Sahrens lr_write_t *lr = (lr_write_t *)lrc; 322fa9e4066Sahrens blkptr_t *bp = &lr->lr_blkptr; 323fa9e4066Sahrens if (bp->blk_birth >= claim_txg && 324fa9e4066Sahrens !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) { 325fa9e4066Sahrens (void) arc_free(NULL, zilog->zl_spa, 326fa9e4066Sahrens dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT); 327fa9e4066Sahrens } 328fa9e4066Sahrens } 329fa9e4066Sahrens } 330fa9e4066Sahrens 331fa9e4066Sahrens /* 332fa9e4066Sahrens * Create an on-disk intent log. 333fa9e4066Sahrens */ 334fa9e4066Sahrens static void 335fa9e4066Sahrens zil_create(zilog_t *zilog) 336fa9e4066Sahrens { 337d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 338fa9e4066Sahrens lwb_t *lwb; 339d80c45e0Sbonwick uint64_t txg = 0; 340d80c45e0Sbonwick dmu_tx_t *tx = NULL; 341fa9e4066Sahrens blkptr_t blk; 342d80c45e0Sbonwick int error = 0; 343fa9e4066Sahrens 344fa9e4066Sahrens /* 345d80c45e0Sbonwick * Wait for any previous destroy to complete. 346fa9e4066Sahrens */ 347d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 348d80c45e0Sbonwick 349d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 350d80c45e0Sbonwick ASSERT(zh->zh_replay_seq == 0); 351d80c45e0Sbonwick 352d80c45e0Sbonwick blk = zh->zh_log; 353fa9e4066Sahrens 354fa9e4066Sahrens /* 355899217ddSNeil Perrin * If we don't already have an initial log block or we have one 356899217ddSNeil Perrin * but it's the wrong endianness then allocate one. 357fa9e4066Sahrens */ 358899217ddSNeil Perrin if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 359d80c45e0Sbonwick tx = dmu_tx_create(zilog->zl_os); 360d80c45e0Sbonwick (void) dmu_tx_assign(tx, TXG_WAIT); 361d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 362d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 363d80c45e0Sbonwick 364899217ddSNeil Perrin if (!BP_IS_HOLE(&blk)) { 365899217ddSNeil Perrin zio_free_blk(zilog->zl_spa, &blk, txg); 366899217ddSNeil Perrin BP_ZERO(&blk); 367899217ddSNeil Perrin } 368899217ddSNeil Perrin 36967bd71c6Sperrin error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk, 37067bd71c6Sperrin NULL, txg); 371d80c45e0Sbonwick 372d80c45e0Sbonwick if (error == 0) 373d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 37413f5297eSperrin } 375fa9e4066Sahrens 376d80c45e0Sbonwick /* 377d80c45e0Sbonwick * Allocate a log write buffer (lwb) for the first log block. 378d80c45e0Sbonwick */ 379d80c45e0Sbonwick if (error == 0) { 380fa9e4066Sahrens lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 381fa9e4066Sahrens lwb->lwb_zilog = zilog; 382fa9e4066Sahrens lwb->lwb_blk = blk; 383fa9e4066Sahrens lwb->lwb_nused = 0; 384fa9e4066Sahrens lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); 385fa9e4066Sahrens lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); 386fa9e4066Sahrens lwb->lwb_max_txg = txg; 387c5c6ffa0Smaybee lwb->lwb_zio = NULL; 388c5c6ffa0Smaybee 389fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 390fa9e4066Sahrens list_insert_tail(&zilog->zl_lwb_list, lwb); 391fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 392fa9e4066Sahrens } 393fa9e4066Sahrens 394d80c45e0Sbonwick /* 395d80c45e0Sbonwick * If we just allocated the first log block, commit our transaction 396d80c45e0Sbonwick * and wait for zil_sync() to stuff the block poiner into zh_log. 397d80c45e0Sbonwick * (zh is part of the MOS, so we cannot modify it in open context.) 398d80c45e0Sbonwick */ 399d80c45e0Sbonwick if (tx != NULL) { 400d80c45e0Sbonwick dmu_tx_commit(tx); 40113f5297eSperrin txg_wait_synced(zilog->zl_dmu_pool, txg); 402d80c45e0Sbonwick } 403d80c45e0Sbonwick 404d80c45e0Sbonwick ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 405fa9e4066Sahrens } 406fa9e4066Sahrens 407fa9e4066Sahrens /* 408fa9e4066Sahrens * In one tx, free all log blocks and clear the log header. 409d80c45e0Sbonwick * If keep_first is set, then we're replaying a log with no content. 410d80c45e0Sbonwick * We want to keep the first block, however, so that the first 411d80c45e0Sbonwick * synchronous transaction doesn't require a txg_wait_synced() 412d80c45e0Sbonwick * in zil_create(). We don't need to txg_wait_synced() here either 413d80c45e0Sbonwick * when keep_first is set, because both zil_create() and zil_destroy() 414d80c45e0Sbonwick * will wait for any in-progress destroys to complete. 415fa9e4066Sahrens */ 416fa9e4066Sahrens void 417d80c45e0Sbonwick zil_destroy(zilog_t *zilog, boolean_t keep_first) 418fa9e4066Sahrens { 419d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 420d80c45e0Sbonwick lwb_t *lwb; 421fa9e4066Sahrens dmu_tx_t *tx; 422fa9e4066Sahrens uint64_t txg; 423fa9e4066Sahrens 424d80c45e0Sbonwick /* 425d80c45e0Sbonwick * Wait for any previous destroy to complete. 426d80c45e0Sbonwick */ 427d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 428fa9e4066Sahrens 429d80c45e0Sbonwick if (BP_IS_HOLE(&zh->zh_log)) 430fa9e4066Sahrens return; 431fa9e4066Sahrens 432fa9e4066Sahrens tx = dmu_tx_create(zilog->zl_os); 433fa9e4066Sahrens (void) dmu_tx_assign(tx, TXG_WAIT); 434fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 435fa9e4066Sahrens txg = dmu_tx_get_txg(tx); 436fa9e4066Sahrens 437d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 438d80c45e0Sbonwick 439587c2a35Sperrin /* 440587c2a35Sperrin * It is possible for the ZIL to get the previously mounted zilog 441587c2a35Sperrin * structure of the same dataset if quickly remounted and the dbuf 442587c2a35Sperrin * eviction has not completed. In this case we can see a non 443587c2a35Sperrin * empty lwb list and keep_first will be set. We fix this by 444587c2a35Sperrin * clearing the keep_first. This will be slower but it's very rare. 445587c2a35Sperrin */ 446587c2a35Sperrin if (!list_is_empty(&zilog->zl_lwb_list) && keep_first) 447587c2a35Sperrin keep_first = B_FALSE; 448587c2a35Sperrin 449d80c45e0Sbonwick ASSERT3U(zilog->zl_destroy_txg, <, txg); 450fa9e4066Sahrens zilog->zl_destroy_txg = txg; 451d80c45e0Sbonwick zilog->zl_keep_first = keep_first; 452d80c45e0Sbonwick 453d80c45e0Sbonwick if (!list_is_empty(&zilog->zl_lwb_list)) { 454d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 455d80c45e0Sbonwick ASSERT(!keep_first); 456d80c45e0Sbonwick while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 457d80c45e0Sbonwick list_remove(&zilog->zl_lwb_list, lwb); 458d80c45e0Sbonwick if (lwb->lwb_buf != NULL) 459d80c45e0Sbonwick zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 460d80c45e0Sbonwick zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg); 461d80c45e0Sbonwick kmem_cache_free(zil_lwb_cache, lwb); 462d80c45e0Sbonwick } 463d80c45e0Sbonwick } else { 464d80c45e0Sbonwick if (!keep_first) { 465d80c45e0Sbonwick (void) zil_parse(zilog, zil_free_log_block, 466d80c45e0Sbonwick zil_free_log_record, tx, zh->zh_claim_txg); 467d80c45e0Sbonwick } 468d80c45e0Sbonwick } 469b19a79ecSperrin mutex_exit(&zilog->zl_lock); 470fa9e4066Sahrens 471fa9e4066Sahrens dmu_tx_commit(tx); 472fa9e4066Sahrens } 473fa9e4066Sahrens 4743589c4f0SNeil Perrin /* 4753589c4f0SNeil Perrin * return true if the initial log block is not valid 4763589c4f0SNeil Perrin */ 4773589c4f0SNeil Perrin static boolean_t 4783589c4f0SNeil Perrin zil_empty(zilog_t *zilog) 4793589c4f0SNeil Perrin { 4803589c4f0SNeil Perrin const zil_header_t *zh = zilog->zl_header; 4813589c4f0SNeil Perrin arc_buf_t *abuf = NULL; 4823589c4f0SNeil Perrin 4833589c4f0SNeil Perrin if (BP_IS_HOLE(&zh->zh_log)) 4843589c4f0SNeil Perrin return (B_TRUE); 4853589c4f0SNeil Perrin 4863589c4f0SNeil Perrin if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0) 4873589c4f0SNeil Perrin return (B_TRUE); 4883589c4f0SNeil Perrin 4893589c4f0SNeil Perrin VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 4903589c4f0SNeil Perrin return (B_FALSE); 4913589c4f0SNeil Perrin } 4923589c4f0SNeil Perrin 4931d452cf5Sahrens int 494fa9e4066Sahrens zil_claim(char *osname, void *txarg) 495fa9e4066Sahrens { 496fa9e4066Sahrens dmu_tx_t *tx = txarg; 497fa9e4066Sahrens uint64_t first_txg = dmu_tx_get_txg(tx); 498fa9e4066Sahrens zilog_t *zilog; 499fa9e4066Sahrens zil_header_t *zh; 500fa9e4066Sahrens objset_t *os; 501fa9e4066Sahrens int error; 502fa9e4066Sahrens 503745cd3c5Smaybee error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 504fa9e4066Sahrens if (error) { 505b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 5061d452cf5Sahrens return (0); 507fa9e4066Sahrens } 508fa9e4066Sahrens 509fa9e4066Sahrens zilog = dmu_objset_zil(os); 510d80c45e0Sbonwick zh = zil_header_in_syncing_context(zilog); 511fa9e4066Sahrens 512*e6ca193dSGeorge Wilson if (zilog->zl_spa->spa_log_state == SPA_LOG_CLEAR) { 513*e6ca193dSGeorge Wilson if (!BP_IS_HOLE(&zh->zh_log)) 514*e6ca193dSGeorge Wilson zio_free_blk(zilog->zl_spa, &zh->zh_log, first_txg); 515*e6ca193dSGeorge Wilson BP_ZERO(&zh->zh_log); 516*e6ca193dSGeorge Wilson dsl_dataset_dirty(dmu_objset_ds(os), tx); 517*e6ca193dSGeorge Wilson } 518*e6ca193dSGeorge Wilson 5193589c4f0SNeil Perrin /* 5203589c4f0SNeil Perrin * Record here whether the zil has any records to replay. 5213589c4f0SNeil Perrin * If the header block pointer is null or the block points 5223589c4f0SNeil Perrin * to the stubby then we know there are no valid log records. 5233589c4f0SNeil Perrin * We use the header to store this state as the the zilog gets 5243589c4f0SNeil Perrin * freed later in dmu_objset_close(). 5253589c4f0SNeil Perrin * The flags (and the rest of the header fields) are cleared in 5263589c4f0SNeil Perrin * zil_sync() as a result of a zil_destroy(), after replaying the log. 5273589c4f0SNeil Perrin * 5283589c4f0SNeil Perrin * Note, the intent log can be empty but still need the 5293589c4f0SNeil Perrin * stubby to be claimed. 5303589c4f0SNeil Perrin */ 531*e6ca193dSGeorge Wilson if (!zil_empty(zilog)) { 5323589c4f0SNeil Perrin zh->zh_flags |= ZIL_REPLAY_NEEDED; 533*e6ca193dSGeorge Wilson dsl_dataset_dirty(dmu_objset_ds(os), tx); 534*e6ca193dSGeorge Wilson } 5353589c4f0SNeil Perrin 536fa9e4066Sahrens /* 537d80c45e0Sbonwick * Claim all log blocks if we haven't already done so, and remember 538d80c45e0Sbonwick * the highest claimed sequence number. This ensures that if we can 539d80c45e0Sbonwick * read only part of the log now (e.g. due to a missing device), 540d80c45e0Sbonwick * but we can read the entire log later, we will not try to replay 541d80c45e0Sbonwick * or destroy beyond the last block we successfully claimed. 542fa9e4066Sahrens */ 543fa9e4066Sahrens ASSERT3U(zh->zh_claim_txg, <=, first_txg); 544fa9e4066Sahrens if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 545fa9e4066Sahrens zh->zh_claim_txg = first_txg; 546d80c45e0Sbonwick zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block, 547d80c45e0Sbonwick zil_claim_log_record, tx, first_txg); 548fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(os), tx); 549fa9e4066Sahrens } 550d80c45e0Sbonwick 551fa9e4066Sahrens ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 552fa9e4066Sahrens dmu_objset_close(os); 5531d452cf5Sahrens return (0); 554b87f3af3Sperrin } 555b87f3af3Sperrin 556b87f3af3Sperrin /* 557b87f3af3Sperrin * Check the log by walking the log chain. 558b87f3af3Sperrin * Checksum errors are ok as they indicate the end of the chain. 559b87f3af3Sperrin * Any other error (no device or read failure) returns an error. 560b87f3af3Sperrin */ 561b87f3af3Sperrin /* ARGSUSED */ 562b87f3af3Sperrin int 563b87f3af3Sperrin zil_check_log_chain(char *osname, void *txarg) 564b87f3af3Sperrin { 565b87f3af3Sperrin zilog_t *zilog; 566b87f3af3Sperrin zil_header_t *zh; 567b87f3af3Sperrin blkptr_t blk; 568b87f3af3Sperrin arc_buf_t *abuf; 569b87f3af3Sperrin objset_t *os; 570b87f3af3Sperrin char *lrbuf; 571b87f3af3Sperrin zil_trailer_t *ztp; 572b87f3af3Sperrin int error; 573b87f3af3Sperrin 574b87f3af3Sperrin error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 575b87f3af3Sperrin if (error) { 576b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 577b87f3af3Sperrin return (0); 578b87f3af3Sperrin } 579b87f3af3Sperrin 580b87f3af3Sperrin zilog = dmu_objset_zil(os); 581b87f3af3Sperrin zh = zil_header_in_syncing_context(zilog); 582b87f3af3Sperrin blk = zh->zh_log; 583b87f3af3Sperrin if (BP_IS_HOLE(&blk)) { 584b87f3af3Sperrin dmu_objset_close(os); 585b87f3af3Sperrin return (0); /* no chain */ 586b87f3af3Sperrin } 587b87f3af3Sperrin 588b87f3af3Sperrin for (;;) { 589b87f3af3Sperrin error = zil_read_log_block(zilog, &blk, &abuf); 590b87f3af3Sperrin if (error) 591b87f3af3Sperrin break; 592b87f3af3Sperrin lrbuf = abuf->b_data; 593b87f3af3Sperrin ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 594b87f3af3Sperrin blk = ztp->zit_next_blk; 595b87f3af3Sperrin VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 596b87f3af3Sperrin } 597b87f3af3Sperrin dmu_objset_close(os); 598b87f3af3Sperrin if (error == ECKSUM) 599b87f3af3Sperrin return (0); /* normal end of chain */ 600b87f3af3Sperrin return (error); 601b87f3af3Sperrin } 602b87f3af3Sperrin 60317f17c2dSbonwick static int 60417f17c2dSbonwick zil_vdev_compare(const void *x1, const void *x2) 60517f17c2dSbonwick { 606c0d19b33Sperrin uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 607c0d19b33Sperrin uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 60817f17c2dSbonwick 60917f17c2dSbonwick if (v1 < v2) 61017f17c2dSbonwick return (-1); 61117f17c2dSbonwick if (v1 > v2) 61217f17c2dSbonwick return (1); 61317f17c2dSbonwick 61417f17c2dSbonwick return (0); 61517f17c2dSbonwick } 61617f17c2dSbonwick 617fa9e4066Sahrens void 61817f17c2dSbonwick zil_add_block(zilog_t *zilog, blkptr_t *bp) 619fa9e4066Sahrens { 62017f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 62117f17c2dSbonwick avl_index_t where; 62217f17c2dSbonwick zil_vdev_node_t *zv, zvsearch; 62317f17c2dSbonwick int ndvas = BP_GET_NDVAS(bp); 62417f17c2dSbonwick int i; 625fa9e4066Sahrens 626416e0cd8Sek if (zfs_nocacheflush) 627fa9e4066Sahrens return; 628fa9e4066Sahrens 62917f17c2dSbonwick ASSERT(zilog->zl_writer); 63017f17c2dSbonwick 63117f17c2dSbonwick /* 63217f17c2dSbonwick * Even though we're zl_writer, we still need a lock because the 63317f17c2dSbonwick * zl_get_data() callbacks may have dmu_sync() done callbacks 63417f17c2dSbonwick * that will run concurrently. 63517f17c2dSbonwick */ 63617f17c2dSbonwick mutex_enter(&zilog->zl_vdev_lock); 63717f17c2dSbonwick for (i = 0; i < ndvas; i++) { 63817f17c2dSbonwick zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 63917f17c2dSbonwick if (avl_find(t, &zvsearch, &where) == NULL) { 64017f17c2dSbonwick zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 64117f17c2dSbonwick zv->zv_vdev = zvsearch.zv_vdev; 64217f17c2dSbonwick avl_insert(t, zv, where); 64367bd71c6Sperrin } 64467bd71c6Sperrin } 64517f17c2dSbonwick mutex_exit(&zilog->zl_vdev_lock); 646fa9e4066Sahrens } 647fa9e4066Sahrens 64867bd71c6Sperrin void 64967bd71c6Sperrin zil_flush_vdevs(zilog_t *zilog) 65067bd71c6Sperrin { 65167bd71c6Sperrin spa_t *spa = zilog->zl_spa; 65217f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 65317f17c2dSbonwick void *cookie = NULL; 65417f17c2dSbonwick zil_vdev_node_t *zv; 65517f17c2dSbonwick zio_t *zio; 656fa9e4066Sahrens 65767bd71c6Sperrin ASSERT(zilog->zl_writer); 65867bd71c6Sperrin 65917f17c2dSbonwick /* 66017f17c2dSbonwick * We don't need zl_vdev_lock here because we're the zl_writer, 66117f17c2dSbonwick * and all zl_get_data() callbacks are done. 66217f17c2dSbonwick */ 66317f17c2dSbonwick if (avl_numnodes(t) == 0) 66417f17c2dSbonwick return; 66517f17c2dSbonwick 666e14bb325SJeff Bonwick spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 66717f17c2dSbonwick 668e14bb325SJeff Bonwick zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 669fa9e4066Sahrens 67017f17c2dSbonwick while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 67117f17c2dSbonwick vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 67217f17c2dSbonwick if (vd != NULL) 67317f17c2dSbonwick zio_flush(zio, vd); 67417f17c2dSbonwick kmem_free(zv, sizeof (*zv)); 67567bd71c6Sperrin } 67617f17c2dSbonwick 677fa9e4066Sahrens /* 678fa9e4066Sahrens * Wait for all the flushes to complete. Not all devices actually 679fa9e4066Sahrens * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 680fa9e4066Sahrens */ 68117f17c2dSbonwick (void) zio_wait(zio); 68217f17c2dSbonwick 683e14bb325SJeff Bonwick spa_config_exit(spa, SCL_STATE, FTAG); 684fa9e4066Sahrens } 685fa9e4066Sahrens 686fa9e4066Sahrens /* 687fa9e4066Sahrens * Function called when a log block write completes 688fa9e4066Sahrens */ 689fa9e4066Sahrens static void 690fa9e4066Sahrens zil_lwb_write_done(zio_t *zio) 691fa9e4066Sahrens { 692fa9e4066Sahrens lwb_t *lwb = zio->io_private; 693fa9e4066Sahrens zilog_t *zilog = lwb->lwb_zilog; 694fa9e4066Sahrens 695e14bb325SJeff Bonwick ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 696e14bb325SJeff Bonwick ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG); 697e14bb325SJeff Bonwick ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 698e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 699e14bb325SJeff Bonwick ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 700e14bb325SJeff Bonwick ASSERT(!BP_IS_GANG(zio->io_bp)); 701e14bb325SJeff Bonwick ASSERT(!BP_IS_HOLE(zio->io_bp)); 702e14bb325SJeff Bonwick ASSERT(zio->io_bp->blk_fill == 0); 703e14bb325SJeff Bonwick 704fa9e4066Sahrens /* 705ef0d8e11SNeil Perrin * Ensure the lwb buffer pointer is cleared before releasing 706ef0d8e11SNeil Perrin * the txg. If we have had an allocation failure and 707ef0d8e11SNeil Perrin * the txg is waiting to sync then we want want zil_sync() 708ef0d8e11SNeil Perrin * to remove the lwb so that it's not picked up as the next new 709ef0d8e11SNeil Perrin * one in zil_commit_writer(). zil_sync() will only remove 710ef0d8e11SNeil Perrin * the lwb if lwb_buf is null. 711fa9e4066Sahrens */ 712fa9e4066Sahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 713fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 714fa9e4066Sahrens lwb->lwb_buf = NULL; 7158654d025Sperrin if (zio->io_error) 716fa9e4066Sahrens zilog->zl_log_error = B_TRUE; 717fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 718ef0d8e11SNeil Perrin 719ef0d8e11SNeil Perrin /* 720ef0d8e11SNeil Perrin * Now that we've written this log block, we have a stable pointer 721ef0d8e11SNeil Perrin * to the next block in the chain, so it's OK to let the txg in 722ef0d8e11SNeil Perrin * which we allocated the next block sync. 723ef0d8e11SNeil Perrin */ 724ef0d8e11SNeil Perrin txg_rele_to_sync(&lwb->lwb_txgh); 725fa9e4066Sahrens } 726fa9e4066Sahrens 727c5c6ffa0Smaybee /* 728c5c6ffa0Smaybee * Initialize the io for a log block. 729c5c6ffa0Smaybee */ 730c5c6ffa0Smaybee static void 731c5c6ffa0Smaybee zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 732c5c6ffa0Smaybee { 733c5c6ffa0Smaybee zbookmark_t zb; 734c5c6ffa0Smaybee 735c5c6ffa0Smaybee zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET]; 736c5c6ffa0Smaybee zb.zb_object = 0; 737c5c6ffa0Smaybee zb.zb_level = -1; 738c5c6ffa0Smaybee zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 739c5c6ffa0Smaybee 740b19a79ecSperrin if (zilog->zl_root_zio == NULL) { 741b19a79ecSperrin zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 742b19a79ecSperrin ZIO_FLAG_CANFAIL); 743b19a79ecSperrin } 74467bd71c6Sperrin if (lwb->lwb_zio == NULL) { 74567bd71c6Sperrin lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 746*e6ca193dSGeorge Wilson 0, &lwb->lwb_blk, lwb->lwb_buf, lwb->lwb_sz, 747*e6ca193dSGeorge Wilson zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 748*e6ca193dSGeorge Wilson ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb); 74967bd71c6Sperrin } 750c5c6ffa0Smaybee } 751c5c6ffa0Smaybee 752fa9e4066Sahrens /* 753fa9e4066Sahrens * Start a log block write and advance to the next log block. 754fa9e4066Sahrens * Calls are serialized. 755fa9e4066Sahrens */ 756fa9e4066Sahrens static lwb_t * 757fa9e4066Sahrens zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 758fa9e4066Sahrens { 759fa9e4066Sahrens lwb_t *nlwb; 760fa9e4066Sahrens zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; 761d80c45e0Sbonwick spa_t *spa = zilog->zl_spa; 762d80c45e0Sbonwick blkptr_t *bp = &ztp->zit_next_blk; 763fa9e4066Sahrens uint64_t txg; 764fa9e4066Sahrens uint64_t zil_blksz; 765fa9e4066Sahrens int error; 766fa9e4066Sahrens 767fa9e4066Sahrens ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); 768fa9e4066Sahrens 769fa9e4066Sahrens /* 770fa9e4066Sahrens * Allocate the next block and save its address in this block 771fa9e4066Sahrens * before writing it in order to establish the log chain. 772fa9e4066Sahrens * Note that if the allocation of nlwb synced before we wrote 773fa9e4066Sahrens * the block that points at it (lwb), we'd leak it if we crashed. 774fa9e4066Sahrens * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done(). 775fa9e4066Sahrens */ 776fa9e4066Sahrens txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh); 777fa9e4066Sahrens txg_rele_to_quiesce(&lwb->lwb_txgh); 778fa9e4066Sahrens 779fa9e4066Sahrens /* 78022ac5be4Sperrin * Pick a ZIL blocksize. We request a size that is the 78122ac5be4Sperrin * maximum of the previous used size, the current used size and 78222ac5be4Sperrin * the amount waiting in the queue. 783fa9e4066Sahrens */ 784c5c6ffa0Smaybee zil_blksz = MAX(zilog->zl_prev_used, 785c5c6ffa0Smaybee zilog->zl_cur_used + sizeof (*ztp)); 78622ac5be4Sperrin zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); 787b4d654b0Sperrin zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); 78822ac5be4Sperrin if (zil_blksz > ZIL_MAX_BLKSZ) 78922ac5be4Sperrin zil_blksz = ZIL_MAX_BLKSZ; 790fa9e4066Sahrens 79167bd71c6Sperrin BP_ZERO(bp); 79267bd71c6Sperrin /* pass the old blkptr in order to spread log blocks across devs */ 79367bd71c6Sperrin error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg); 794fa9e4066Sahrens if (error) { 795d63d470bSgw dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg); 796d63d470bSgw 797d63d470bSgw /* 798d63d470bSgw * We dirty the dataset to ensure that zil_sync() will 799d63d470bSgw * be called to remove this lwb from our zl_lwb_list. 800d63d470bSgw * Failing to do so, may leave an lwb with a NULL lwb_buf 801d63d470bSgw * hanging around on the zl_lwb_list. 802d63d470bSgw */ 803d63d470bSgw dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 804aeb1c1b6Sgw dmu_tx_commit(tx); 805d63d470bSgw 806d63d470bSgw /* 807d63d470bSgw * Since we've just experienced an allocation failure so we 808d63d470bSgw * terminate the current lwb and send it on its way. 809d63d470bSgw */ 810d63d470bSgw ztp->zit_pad = 0; 811d63d470bSgw ztp->zit_nused = lwb->lwb_nused; 812d63d470bSgw ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 813d63d470bSgw zio_nowait(lwb->lwb_zio); 814d63d470bSgw 815ea8dc4b6Seschrock /* 816ea8dc4b6Seschrock * By returning NULL the caller will call tx_wait_synced() 817ea8dc4b6Seschrock */ 818fa9e4066Sahrens return (NULL); 819fa9e4066Sahrens } 820fa9e4066Sahrens 821d80c45e0Sbonwick ASSERT3U(bp->blk_birth, ==, txg); 822ea8dc4b6Seschrock ztp->zit_pad = 0; 823fa9e4066Sahrens ztp->zit_nused = lwb->lwb_nused; 824fa9e4066Sahrens ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 825d80c45e0Sbonwick bp->blk_cksum = lwb->lwb_blk.blk_cksum; 826d80c45e0Sbonwick bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 827fa9e4066Sahrens 828fa9e4066Sahrens /* 829fa9e4066Sahrens * Allocate a new log write buffer (lwb). 830fa9e4066Sahrens */ 831fa9e4066Sahrens nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 832fa9e4066Sahrens 833fa9e4066Sahrens nlwb->lwb_zilog = zilog; 834d80c45e0Sbonwick nlwb->lwb_blk = *bp; 835fa9e4066Sahrens nlwb->lwb_nused = 0; 836fa9e4066Sahrens nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); 837fa9e4066Sahrens nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); 838fa9e4066Sahrens nlwb->lwb_max_txg = txg; 839c5c6ffa0Smaybee nlwb->lwb_zio = NULL; 840fa9e4066Sahrens 841fa9e4066Sahrens /* 84267bd71c6Sperrin * Put new lwb at the end of the log chain 843fa9e4066Sahrens */ 844fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 845fa9e4066Sahrens list_insert_tail(&zilog->zl_lwb_list, nlwb); 846fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 847fa9e4066Sahrens 84817f17c2dSbonwick /* Record the block for later vdev flushing */ 84917f17c2dSbonwick zil_add_block(zilog, &lwb->lwb_blk); 85067bd71c6Sperrin 851fa9e4066Sahrens /* 852c5c6ffa0Smaybee * kick off the write for the old log block 853fa9e4066Sahrens */ 854c5c6ffa0Smaybee dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg); 85567bd71c6Sperrin ASSERT(lwb->lwb_zio); 856c5c6ffa0Smaybee zio_nowait(lwb->lwb_zio); 857fa9e4066Sahrens 858fa9e4066Sahrens return (nlwb); 859fa9e4066Sahrens } 860fa9e4066Sahrens 861fa9e4066Sahrens static lwb_t * 862fa9e4066Sahrens zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 863fa9e4066Sahrens { 864fa9e4066Sahrens lr_t *lrc = &itx->itx_lr; /* common log record */ 865c5c6ffa0Smaybee lr_write_t *lr = (lr_write_t *)lrc; 866fa9e4066Sahrens uint64_t txg = lrc->lrc_txg; 867fa9e4066Sahrens uint64_t reclen = lrc->lrc_reclen; 868c5c6ffa0Smaybee uint64_t dlen; 869fa9e4066Sahrens 870fa9e4066Sahrens if (lwb == NULL) 871fa9e4066Sahrens return (NULL); 872fa9e4066Sahrens ASSERT(lwb->lwb_buf != NULL); 873fa9e4066Sahrens 874c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 875c5c6ffa0Smaybee dlen = P2ROUNDUP_TYPED( 876c5c6ffa0Smaybee lr->lr_length, sizeof (uint64_t), uint64_t); 877c5c6ffa0Smaybee else 878c5c6ffa0Smaybee dlen = 0; 879fa9e4066Sahrens 880104e2ed7Sperrin zilog->zl_cur_used += (reclen + dlen); 88122ac5be4Sperrin 88267bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 88367bd71c6Sperrin 884fa9e4066Sahrens /* 885fa9e4066Sahrens * If this record won't fit in the current log block, start a new one. 886fa9e4066Sahrens */ 887104e2ed7Sperrin if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 888fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 889c5c6ffa0Smaybee if (lwb == NULL) 890fa9e4066Sahrens return (NULL); 89167bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 892ea8dc4b6Seschrock ASSERT(lwb->lwb_nused == 0); 893104e2ed7Sperrin if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 894fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, txg); 895fa9e4066Sahrens return (lwb); 896fa9e4066Sahrens } 897fa9e4066Sahrens } 898fa9e4066Sahrens 899b19a79ecSperrin /* 900b19a79ecSperrin * Update the lrc_seq, to be log record sequence number. See zil.h 901b19a79ecSperrin * Then copy the record to the log buffer. 902b19a79ecSperrin */ 903b19a79ecSperrin lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 904fa9e4066Sahrens bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen); 905c5c6ffa0Smaybee 906c5c6ffa0Smaybee /* 907c5c6ffa0Smaybee * If it's a write, fetch the data or get its blkptr as appropriate. 908c5c6ffa0Smaybee */ 909c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE) { 910c5c6ffa0Smaybee if (txg > spa_freeze_txg(zilog->zl_spa)) 911c5c6ffa0Smaybee txg_wait_synced(zilog->zl_dmu_pool, txg); 912c5c6ffa0Smaybee if (itx->itx_wr_state != WR_COPIED) { 913c5c6ffa0Smaybee char *dbuf; 914c5c6ffa0Smaybee int error; 915c5c6ffa0Smaybee 916c5c6ffa0Smaybee /* alignment is guaranteed */ 917c5c6ffa0Smaybee lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused); 918c5c6ffa0Smaybee if (dlen) { 919c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_NEED_COPY); 920c5c6ffa0Smaybee dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen; 921c5c6ffa0Smaybee lr->lr_common.lrc_reclen += dlen; 922c5c6ffa0Smaybee } else { 923c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_INDIRECT); 924c5c6ffa0Smaybee dbuf = NULL; 925c5c6ffa0Smaybee } 926c5c6ffa0Smaybee error = zilog->zl_get_data( 927c5c6ffa0Smaybee itx->itx_private, lr, dbuf, lwb->lwb_zio); 928c5c6ffa0Smaybee if (error) { 929c5c6ffa0Smaybee ASSERT(error == ENOENT || error == EEXIST || 930c5c6ffa0Smaybee error == EALREADY); 931c5c6ffa0Smaybee return (lwb); 932c5c6ffa0Smaybee } 933c5c6ffa0Smaybee } 934104e2ed7Sperrin } 935c5c6ffa0Smaybee 936c5c6ffa0Smaybee lwb->lwb_nused += reclen + dlen; 937fa9e4066Sahrens lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 938fa9e4066Sahrens ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); 939fa9e4066Sahrens ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 940fa9e4066Sahrens 941fa9e4066Sahrens return (lwb); 942fa9e4066Sahrens } 943fa9e4066Sahrens 944fa9e4066Sahrens itx_t * 945da6c28aaSamw zil_itx_create(uint64_t txtype, size_t lrsize) 946fa9e4066Sahrens { 947fa9e4066Sahrens itx_t *itx; 948fa9e4066Sahrens 949b4d654b0Sperrin lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 950fa9e4066Sahrens 951fa9e4066Sahrens itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 952fa9e4066Sahrens itx->itx_lr.lrc_txtype = txtype; 953fa9e4066Sahrens itx->itx_lr.lrc_reclen = lrsize; 954abf76b6eSperrin itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 955fa9e4066Sahrens itx->itx_lr.lrc_seq = 0; /* defensive */ 956fa9e4066Sahrens 957fa9e4066Sahrens return (itx); 958fa9e4066Sahrens } 959fa9e4066Sahrens 960fa9e4066Sahrens uint64_t 961fa9e4066Sahrens zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 962fa9e4066Sahrens { 963fa9e4066Sahrens uint64_t seq; 964fa9e4066Sahrens 965fa9e4066Sahrens ASSERT(itx->itx_lr.lrc_seq == 0); 966fa9e4066Sahrens 967fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 968fa9e4066Sahrens list_insert_tail(&zilog->zl_itx_list, itx); 969abf76b6eSperrin zilog->zl_itx_list_sz += itx->itx_sod; 970fa9e4066Sahrens itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 971fa9e4066Sahrens itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 972fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 973fa9e4066Sahrens 974fa9e4066Sahrens return (seq); 975fa9e4066Sahrens } 976fa9e4066Sahrens 977fa9e4066Sahrens /* 978fa9e4066Sahrens * Free up all in-memory intent log transactions that have now been synced. 979fa9e4066Sahrens */ 980fa9e4066Sahrens static void 981fa9e4066Sahrens zil_itx_clean(zilog_t *zilog) 982fa9e4066Sahrens { 983fa9e4066Sahrens uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 984fa9e4066Sahrens uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 985a584ef65Sjohansen list_t clean_list; 986fa9e4066Sahrens itx_t *itx; 987fa9e4066Sahrens 988a584ef65Sjohansen list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 989a584ef65Sjohansen 990fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 991b19a79ecSperrin /* wait for a log writer to finish walking list */ 992b19a79ecSperrin while (zilog->zl_writer) { 993b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 994b19a79ecSperrin } 995a584ef65Sjohansen 996a584ef65Sjohansen /* 997a584ef65Sjohansen * Move the sync'd log transactions to a separate list so we can call 998a584ef65Sjohansen * kmem_free without holding the zl_lock. 999a584ef65Sjohansen * 1000a584ef65Sjohansen * There is no need to set zl_writer as we don't drop zl_lock here 1001a584ef65Sjohansen */ 1002fa9e4066Sahrens while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 1003fa9e4066Sahrens itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 1004fa9e4066Sahrens list_remove(&zilog->zl_itx_list, itx); 1005abf76b6eSperrin zilog->zl_itx_list_sz -= itx->itx_sod; 1006a584ef65Sjohansen list_insert_tail(&clean_list, itx); 1007fa9e4066Sahrens } 100867bd71c6Sperrin cv_broadcast(&zilog->zl_cv_writer); 1009fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1010a584ef65Sjohansen 1011a584ef65Sjohansen /* destroy sync'd log transactions */ 1012a584ef65Sjohansen while ((itx = list_head(&clean_list)) != NULL) { 1013a584ef65Sjohansen list_remove(&clean_list, itx); 1014a584ef65Sjohansen kmem_free(itx, offsetof(itx_t, itx_lr) 1015a584ef65Sjohansen + itx->itx_lr.lrc_reclen); 1016a584ef65Sjohansen } 1017a584ef65Sjohansen list_destroy(&clean_list); 1018fa9e4066Sahrens } 1019fa9e4066Sahrens 1020b19a79ecSperrin /* 102167bd71c6Sperrin * If there are any in-memory intent log transactions which have now been 102267bd71c6Sperrin * synced then start up a taskq to free them. 1023b19a79ecSperrin */ 1024fa9e4066Sahrens void 1025fa9e4066Sahrens zil_clean(zilog_t *zilog) 1026fa9e4066Sahrens { 102767bd71c6Sperrin itx_t *itx; 102867bd71c6Sperrin 1029fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 103067bd71c6Sperrin itx = list_head(&zilog->zl_itx_list); 103167bd71c6Sperrin if ((itx != NULL) && 103267bd71c6Sperrin (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 1033fa9e4066Sahrens (void) taskq_dispatch(zilog->zl_clean_taskq, 10349d3574bfSNeil Perrin (task_func_t *)zil_itx_clean, zilog, TQ_SLEEP); 103567bd71c6Sperrin } 1036fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1037fa9e4066Sahrens } 1038fa9e4066Sahrens 1039e14bb325SJeff Bonwick static void 1040b19a79ecSperrin zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 1041fa9e4066Sahrens { 1042fa9e4066Sahrens uint64_t txg; 104367bd71c6Sperrin uint64_t commit_seq = 0; 1044b19a79ecSperrin itx_t *itx, *itx_next = (itx_t *)-1; 1045fa9e4066Sahrens lwb_t *lwb; 1046fa9e4066Sahrens spa_t *spa; 1047fa9e4066Sahrens 1048fa9e4066Sahrens zilog->zl_writer = B_TRUE; 1049e14bb325SJeff Bonwick ASSERT(zilog->zl_root_zio == NULL); 1050b19a79ecSperrin spa = zilog->zl_spa; 1051fa9e4066Sahrens 1052fa9e4066Sahrens if (zilog->zl_suspend) { 1053fa9e4066Sahrens lwb = NULL; 1054fa9e4066Sahrens } else { 1055fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 1056fa9e4066Sahrens if (lwb == NULL) { 1057b19a79ecSperrin /* 1058b19a79ecSperrin * Return if there's nothing to flush before we 1059b19a79ecSperrin * dirty the fs by calling zil_create() 1060b19a79ecSperrin */ 1061b19a79ecSperrin if (list_is_empty(&zilog->zl_itx_list)) { 1062b19a79ecSperrin zilog->zl_writer = B_FALSE; 1063b19a79ecSperrin return; 1064b19a79ecSperrin } 1065fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1066fa9e4066Sahrens zil_create(zilog); 1067fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1068fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 1069fa9e4066Sahrens } 1070fa9e4066Sahrens } 1071fa9e4066Sahrens 107267bd71c6Sperrin /* Loop through in-memory log transactions filling log blocks. */ 1073b19a79ecSperrin DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1074fa9e4066Sahrens for (;;) { 1075b19a79ecSperrin /* 1076b19a79ecSperrin * Find the next itx to push: 1077b19a79ecSperrin * Push all transactions related to specified foid and all 1078b19a79ecSperrin * other transactions except TX_WRITE, TX_TRUNCATE, 1079b19a79ecSperrin * TX_SETATTR and TX_ACL for all other files. 1080b19a79ecSperrin */ 1081b19a79ecSperrin if (itx_next != (itx_t *)-1) 1082b19a79ecSperrin itx = itx_next; 1083b19a79ecSperrin else 1084b19a79ecSperrin itx = list_head(&zilog->zl_itx_list); 1085b19a79ecSperrin for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) { 1086b19a79ecSperrin if (foid == 0) /* push all foids? */ 1087b19a79ecSperrin break; 108867bd71c6Sperrin if (itx->itx_sync) /* push all O_[D]SYNC */ 108967bd71c6Sperrin break; 1090b19a79ecSperrin switch (itx->itx_lr.lrc_txtype) { 1091b19a79ecSperrin case TX_SETATTR: 1092b19a79ecSperrin case TX_WRITE: 1093b19a79ecSperrin case TX_TRUNCATE: 1094b19a79ecSperrin case TX_ACL: 1095b19a79ecSperrin /* lr_foid is same offset for these records */ 1096b19a79ecSperrin if (((lr_write_t *)&itx->itx_lr)->lr_foid 1097b19a79ecSperrin != foid) { 1098b19a79ecSperrin continue; /* skip this record */ 1099b19a79ecSperrin } 1100b19a79ecSperrin } 1101b19a79ecSperrin break; 1102b19a79ecSperrin } 1103fa9e4066Sahrens if (itx == NULL) 1104fa9e4066Sahrens break; 1105fa9e4066Sahrens 1106fa9e4066Sahrens if ((itx->itx_lr.lrc_seq > seq) && 1107b19a79ecSperrin ((lwb == NULL) || (lwb->lwb_nused == 0) || 1108abf76b6eSperrin (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) { 1109fa9e4066Sahrens break; 111067bd71c6Sperrin } 1111fa9e4066Sahrens 1112b19a79ecSperrin /* 1113b19a79ecSperrin * Save the next pointer. Even though we soon drop 1114b19a79ecSperrin * zl_lock all threads that may change the list 1115b19a79ecSperrin * (another writer or zil_itx_clean) can't do so until 1116b19a79ecSperrin * they have zl_writer. 1117b19a79ecSperrin */ 1118b19a79ecSperrin itx_next = list_next(&zilog->zl_itx_list, itx); 1119fa9e4066Sahrens list_remove(&zilog->zl_itx_list, itx); 1120abf76b6eSperrin zilog->zl_itx_list_sz -= itx->itx_sod; 112167bd71c6Sperrin mutex_exit(&zilog->zl_lock); 1122fa9e4066Sahrens txg = itx->itx_lr.lrc_txg; 1123fa9e4066Sahrens ASSERT(txg); 1124fa9e4066Sahrens 1125fa9e4066Sahrens if (txg > spa_last_synced_txg(spa) || 1126fa9e4066Sahrens txg > spa_freeze_txg(spa)) 1127fa9e4066Sahrens lwb = zil_lwb_commit(zilog, itx, lwb); 1128fa9e4066Sahrens kmem_free(itx, offsetof(itx_t, itx_lr) 1129fa9e4066Sahrens + itx->itx_lr.lrc_reclen); 1130fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1131fa9e4066Sahrens } 1132b19a79ecSperrin DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 113367bd71c6Sperrin /* determine commit sequence number */ 113467bd71c6Sperrin itx = list_head(&zilog->zl_itx_list); 113567bd71c6Sperrin if (itx) 113667bd71c6Sperrin commit_seq = itx->itx_lr.lrc_seq; 113767bd71c6Sperrin else 113867bd71c6Sperrin commit_seq = zilog->zl_itx_seq; 1139fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1140fa9e4066Sahrens 1141fa9e4066Sahrens /* write the last block out */ 114267bd71c6Sperrin if (lwb != NULL && lwb->lwb_zio != NULL) 1143fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 1144fa9e4066Sahrens 114522ac5be4Sperrin zilog->zl_prev_used = zilog->zl_cur_used; 114622ac5be4Sperrin zilog->zl_cur_used = 0; 1147fa9e4066Sahrens 1148fa9e4066Sahrens /* 1149b19a79ecSperrin * Wait if necessary for the log blocks to be on stable storage. 1150fa9e4066Sahrens */ 1151b19a79ecSperrin if (zilog->zl_root_zio) { 1152b19a79ecSperrin DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1153b19a79ecSperrin (void) zio_wait(zilog->zl_root_zio); 1154e14bb325SJeff Bonwick zilog->zl_root_zio = NULL; 1155b19a79ecSperrin DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 115617f17c2dSbonwick zil_flush_vdevs(zilog); 1157fa9e4066Sahrens } 115822ac5be4Sperrin 1159fa9e4066Sahrens if (zilog->zl_log_error || lwb == NULL) { 1160fa9e4066Sahrens zilog->zl_log_error = 0; 1161fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1162fa9e4066Sahrens } 116367bd71c6Sperrin 116467bd71c6Sperrin mutex_enter(&zilog->zl_lock); 116522ac5be4Sperrin zilog->zl_writer = B_FALSE; 116667bd71c6Sperrin 116767bd71c6Sperrin ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 116867bd71c6Sperrin zilog->zl_commit_seq = commit_seq; 1169b19a79ecSperrin } 1170b19a79ecSperrin 1171b19a79ecSperrin /* 1172b19a79ecSperrin * Push zfs transactions to stable storage up to the supplied sequence number. 1173b19a79ecSperrin * If foid is 0 push out all transactions, otherwise push only those 1174b19a79ecSperrin * for that file or might have been used to create that file. 1175b19a79ecSperrin */ 1176b19a79ecSperrin void 1177b19a79ecSperrin zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1178b19a79ecSperrin { 1179b19a79ecSperrin if (zilog == NULL || seq == 0) 1180b19a79ecSperrin return; 1181b19a79ecSperrin 1182b19a79ecSperrin mutex_enter(&zilog->zl_lock); 1183b19a79ecSperrin 1184b19a79ecSperrin seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1185b19a79ecSperrin 118667bd71c6Sperrin while (zilog->zl_writer) { 1187b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 118867bd71c6Sperrin if (seq < zilog->zl_commit_seq) { 118967bd71c6Sperrin mutex_exit(&zilog->zl_lock); 119067bd71c6Sperrin return; 119167bd71c6Sperrin } 119267bd71c6Sperrin } 1193b19a79ecSperrin zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 119467bd71c6Sperrin /* wake up others waiting on the commit */ 119567bd71c6Sperrin cv_broadcast(&zilog->zl_cv_writer); 119667bd71c6Sperrin mutex_exit(&zilog->zl_lock); 1197fa9e4066Sahrens } 1198fa9e4066Sahrens 1199fa9e4066Sahrens /* 1200fa9e4066Sahrens * Called in syncing context to free committed log blocks and update log header. 1201fa9e4066Sahrens */ 1202fa9e4066Sahrens void 1203fa9e4066Sahrens zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1204fa9e4066Sahrens { 1205d80c45e0Sbonwick zil_header_t *zh = zil_header_in_syncing_context(zilog); 1206fa9e4066Sahrens uint64_t txg = dmu_tx_get_txg(tx); 1207fa9e4066Sahrens spa_t *spa = zilog->zl_spa; 1208fa9e4066Sahrens lwb_t *lwb; 1209fa9e4066Sahrens 121014843421SMatthew Ahrens /* 121114843421SMatthew Ahrens * We don't zero out zl_destroy_txg, so make sure we don't try 121214843421SMatthew Ahrens * to destroy it twice. 121314843421SMatthew Ahrens */ 121414843421SMatthew Ahrens if (spa_sync_pass(spa) != 1) 121514843421SMatthew Ahrens return; 121614843421SMatthew Ahrens 1217d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1218d80c45e0Sbonwick 1219fa9e4066Sahrens ASSERT(zilog->zl_stop_sync == 0); 1220fa9e4066Sahrens 12211209a471SNeil Perrin zh->zh_replay_seq = zilog->zl_replayed_seq[txg & TXG_MASK]; 1222fa9e4066Sahrens 1223fa9e4066Sahrens if (zilog->zl_destroy_txg == txg) { 1224d80c45e0Sbonwick blkptr_t blk = zh->zh_log; 1225d80c45e0Sbonwick 1226d80c45e0Sbonwick ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1227d80c45e0Sbonwick 1228d80c45e0Sbonwick bzero(zh, sizeof (zil_header_t)); 12291209a471SNeil Perrin bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1230d80c45e0Sbonwick 1231d80c45e0Sbonwick if (zilog->zl_keep_first) { 1232d80c45e0Sbonwick /* 1233d80c45e0Sbonwick * If this block was part of log chain that couldn't 1234d80c45e0Sbonwick * be claimed because a device was missing during 1235d80c45e0Sbonwick * zil_claim(), but that device later returns, 1236d80c45e0Sbonwick * then this block could erroneously appear valid. 1237d80c45e0Sbonwick * To guard against this, assign a new GUID to the new 1238d80c45e0Sbonwick * log chain so it doesn't matter what blk points to. 1239d80c45e0Sbonwick */ 1240d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 1241d80c45e0Sbonwick zh->zh_log = blk; 1242d80c45e0Sbonwick } 1243fa9e4066Sahrens } 1244fa9e4066Sahrens 1245*e6ca193dSGeorge Wilson while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1246b19a79ecSperrin zh->zh_log = lwb->lwb_blk; 1247fa9e4066Sahrens if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1248fa9e4066Sahrens break; 1249fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1250fa9e4066Sahrens zio_free_blk(spa, &lwb->lwb_blk, txg); 1251fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1252d63d470bSgw 1253d63d470bSgw /* 1254d63d470bSgw * If we don't have anything left in the lwb list then 1255d63d470bSgw * we've had an allocation failure and we need to zero 1256d63d470bSgw * out the zil_header blkptr so that we don't end 1257d63d470bSgw * up freeing the same block twice. 1258d63d470bSgw */ 1259d63d470bSgw if (list_head(&zilog->zl_lwb_list) == NULL) 1260d63d470bSgw BP_ZERO(&zh->zh_log); 1261fa9e4066Sahrens } 1262fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1263fa9e4066Sahrens } 1264fa9e4066Sahrens 1265fa9e4066Sahrens void 1266fa9e4066Sahrens zil_init(void) 1267fa9e4066Sahrens { 1268fa9e4066Sahrens zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 12695ad82045Snd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1270fa9e4066Sahrens } 1271fa9e4066Sahrens 1272fa9e4066Sahrens void 1273fa9e4066Sahrens zil_fini(void) 1274fa9e4066Sahrens { 1275fa9e4066Sahrens kmem_cache_destroy(zil_lwb_cache); 1276fa9e4066Sahrens } 1277fa9e4066Sahrens 1278fa9e4066Sahrens zilog_t * 1279fa9e4066Sahrens zil_alloc(objset_t *os, zil_header_t *zh_phys) 1280fa9e4066Sahrens { 1281fa9e4066Sahrens zilog_t *zilog; 1282fa9e4066Sahrens 1283fa9e4066Sahrens zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1284fa9e4066Sahrens 1285fa9e4066Sahrens zilog->zl_header = zh_phys; 1286fa9e4066Sahrens zilog->zl_os = os; 1287fa9e4066Sahrens zilog->zl_spa = dmu_objset_spa(os); 1288fa9e4066Sahrens zilog->zl_dmu_pool = dmu_objset_pool(os); 1289d80c45e0Sbonwick zilog->zl_destroy_txg = TXG_INITIAL - 1; 1290fa9e4066Sahrens 12915ad82045Snd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 12925ad82045Snd 1293fa9e4066Sahrens list_create(&zilog->zl_itx_list, sizeof (itx_t), 1294fa9e4066Sahrens offsetof(itx_t, itx_node)); 1295fa9e4066Sahrens 1296fa9e4066Sahrens list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1297fa9e4066Sahrens offsetof(lwb_t, lwb_node)); 1298fa9e4066Sahrens 129917f17c2dSbonwick mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 130017f17c2dSbonwick 130117f17c2dSbonwick avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 130217f17c2dSbonwick sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1303fa9e4066Sahrens 1304b7b97454Sperrin cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1305b7b97454Sperrin cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1306b7b97454Sperrin 1307fa9e4066Sahrens return (zilog); 1308fa9e4066Sahrens } 1309fa9e4066Sahrens 1310fa9e4066Sahrens void 1311fa9e4066Sahrens zil_free(zilog_t *zilog) 1312fa9e4066Sahrens { 1313fa9e4066Sahrens lwb_t *lwb; 1314fa9e4066Sahrens 1315fa9e4066Sahrens zilog->zl_stop_sync = 1; 1316fa9e4066Sahrens 1317fa9e4066Sahrens while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1318fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1319fa9e4066Sahrens if (lwb->lwb_buf != NULL) 1320fa9e4066Sahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1321fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1322fa9e4066Sahrens } 1323fa9e4066Sahrens list_destroy(&zilog->zl_lwb_list); 1324fa9e4066Sahrens 132517f17c2dSbonwick avl_destroy(&zilog->zl_vdev_tree); 132617f17c2dSbonwick mutex_destroy(&zilog->zl_vdev_lock); 1327fa9e4066Sahrens 1328fa9e4066Sahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1329fa9e4066Sahrens list_destroy(&zilog->zl_itx_list); 13305ad82045Snd mutex_destroy(&zilog->zl_lock); 1331fa9e4066Sahrens 1332b7b97454Sperrin cv_destroy(&zilog->zl_cv_writer); 1333b7b97454Sperrin cv_destroy(&zilog->zl_cv_suspend); 1334b7b97454Sperrin 1335fa9e4066Sahrens kmem_free(zilog, sizeof (zilog_t)); 1336fa9e4066Sahrens } 1337fa9e4066Sahrens 1338fa9e4066Sahrens /* 1339fa9e4066Sahrens * Open an intent log. 1340fa9e4066Sahrens */ 1341fa9e4066Sahrens zilog_t * 1342fa9e4066Sahrens zil_open(objset_t *os, zil_get_data_t *get_data) 1343fa9e4066Sahrens { 1344fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1345fa9e4066Sahrens 1346fa9e4066Sahrens zilog->zl_get_data = get_data; 1347fa9e4066Sahrens zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1348fa9e4066Sahrens 2, 2, TASKQ_PREPOPULATE); 1349fa9e4066Sahrens 1350fa9e4066Sahrens return (zilog); 1351fa9e4066Sahrens } 1352fa9e4066Sahrens 1353fa9e4066Sahrens /* 1354fa9e4066Sahrens * Close an intent log. 1355fa9e4066Sahrens */ 1356fa9e4066Sahrens void 1357fa9e4066Sahrens zil_close(zilog_t *zilog) 1358fa9e4066Sahrens { 1359d80c45e0Sbonwick /* 1360d80c45e0Sbonwick * If the log isn't already committed, mark the objset dirty 1361d80c45e0Sbonwick * (so zil_sync() will be called) and wait for that txg to sync. 1362d80c45e0Sbonwick */ 1363d80c45e0Sbonwick if (!zil_is_committed(zilog)) { 1364d80c45e0Sbonwick uint64_t txg; 1365d80c45e0Sbonwick dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1366d80c45e0Sbonwick (void) dmu_tx_assign(tx, TXG_WAIT); 1367d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1368d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 1369d80c45e0Sbonwick dmu_tx_commit(tx); 1370d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, txg); 1371d80c45e0Sbonwick } 1372d80c45e0Sbonwick 1373fa9e4066Sahrens taskq_destroy(zilog->zl_clean_taskq); 1374fa9e4066Sahrens zilog->zl_clean_taskq = NULL; 1375fa9e4066Sahrens zilog->zl_get_data = NULL; 1376fa9e4066Sahrens 1377fa9e4066Sahrens zil_itx_clean(zilog); 1378fa9e4066Sahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1379fa9e4066Sahrens } 1380fa9e4066Sahrens 1381fa9e4066Sahrens /* 1382fa9e4066Sahrens * Suspend an intent log. While in suspended mode, we still honor 1383fa9e4066Sahrens * synchronous semantics, but we rely on txg_wait_synced() to do it. 1384fa9e4066Sahrens * We suspend the log briefly when taking a snapshot so that the snapshot 1385fa9e4066Sahrens * contains all the data it's supposed to, and has an empty intent log. 1386fa9e4066Sahrens */ 1387fa9e4066Sahrens int 1388fa9e4066Sahrens zil_suspend(zilog_t *zilog) 1389fa9e4066Sahrens { 1390d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1391fa9e4066Sahrens 1392fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 13933589c4f0SNeil Perrin if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1394fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1395fa9e4066Sahrens return (EBUSY); 1396fa9e4066Sahrens } 1397d80c45e0Sbonwick if (zilog->zl_suspend++ != 0) { 1398d80c45e0Sbonwick /* 1399d80c45e0Sbonwick * Someone else already began a suspend. 1400d80c45e0Sbonwick * Just wait for them to finish. 1401d80c45e0Sbonwick */ 1402d80c45e0Sbonwick while (zilog->zl_suspending) 1403d80c45e0Sbonwick cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1404d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1405d80c45e0Sbonwick return (0); 1406d80c45e0Sbonwick } 1407d80c45e0Sbonwick zilog->zl_suspending = B_TRUE; 1408fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1409fa9e4066Sahrens 1410b19a79ecSperrin zil_commit(zilog, UINT64_MAX, 0); 1411fa9e4066Sahrens 1412b19a79ecSperrin /* 1413b19a79ecSperrin * Wait for any in-flight log writes to complete. 1414b19a79ecSperrin */ 1415fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1416b19a79ecSperrin while (zilog->zl_writer) 1417b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1418fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1419fa9e4066Sahrens 1420d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1421d80c45e0Sbonwick 1422d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1423d80c45e0Sbonwick zilog->zl_suspending = B_FALSE; 1424d80c45e0Sbonwick cv_broadcast(&zilog->zl_cv_suspend); 1425d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1426fa9e4066Sahrens 1427fa9e4066Sahrens return (0); 1428fa9e4066Sahrens } 1429fa9e4066Sahrens 1430fa9e4066Sahrens void 1431fa9e4066Sahrens zil_resume(zilog_t *zilog) 1432fa9e4066Sahrens { 1433fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1434fa9e4066Sahrens ASSERT(zilog->zl_suspend != 0); 1435fa9e4066Sahrens zilog->zl_suspend--; 1436fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1437fa9e4066Sahrens } 1438fa9e4066Sahrens 1439fa9e4066Sahrens typedef struct zil_replay_arg { 1440fa9e4066Sahrens objset_t *zr_os; 1441fa9e4066Sahrens zil_replay_func_t **zr_replay; 1442fa9e4066Sahrens void *zr_arg; 1443fa9e4066Sahrens boolean_t zr_byteswap; 1444fa9e4066Sahrens char *zr_lrbuf; 1445fa9e4066Sahrens } zil_replay_arg_t; 1446fa9e4066Sahrens 1447fa9e4066Sahrens static void 1448fa9e4066Sahrens zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1449fa9e4066Sahrens { 1450fa9e4066Sahrens zil_replay_arg_t *zr = zra; 1451d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1452fa9e4066Sahrens uint64_t reclen = lr->lrc_reclen; 1453fa9e4066Sahrens uint64_t txtype = lr->lrc_txtype; 145467bd71c6Sperrin char *name; 14551209a471SNeil Perrin int pass, error; 1456fa9e4066Sahrens 14571209a471SNeil Perrin if (!zilog->zl_replay) /* giving up */ 1458fa9e4066Sahrens return; 1459fa9e4066Sahrens 1460fa9e4066Sahrens if (lr->lrc_txg < claim_txg) /* already committed */ 1461fa9e4066Sahrens return; 1462fa9e4066Sahrens 1463fa9e4066Sahrens if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1464fa9e4066Sahrens return; 1465fa9e4066Sahrens 1466da6c28aaSamw /* Strip case-insensitive bit, still present in log record */ 1467da6c28aaSamw txtype &= ~TX_CI; 1468da6c28aaSamw 14691209a471SNeil Perrin if (txtype == 0 || txtype >= TX_MAX_TYPE) { 14701209a471SNeil Perrin error = EINVAL; 14711209a471SNeil Perrin goto bad; 14721209a471SNeil Perrin } 14731209a471SNeil Perrin 1474fa9e4066Sahrens /* 1475fa9e4066Sahrens * Make a copy of the data so we can revise and extend it. 1476fa9e4066Sahrens */ 1477fa9e4066Sahrens bcopy(lr, zr->zr_lrbuf, reclen); 1478fa9e4066Sahrens 1479fa9e4066Sahrens /* 1480fa9e4066Sahrens * The log block containing this lr may have been byteswapped 1481fa9e4066Sahrens * so that we can easily examine common fields like lrc_txtype. 1482fa9e4066Sahrens * However, the log is a mix of different data types, and only the 1483fa9e4066Sahrens * replay vectors know how to byteswap their records. Therefore, if 1484fa9e4066Sahrens * the lr was byteswapped, undo it before invoking the replay vector. 1485fa9e4066Sahrens */ 1486fa9e4066Sahrens if (zr->zr_byteswap) 1487fa9e4066Sahrens byteswap_uint64_array(zr->zr_lrbuf, reclen); 1488fa9e4066Sahrens 1489fa9e4066Sahrens /* 1490fa9e4066Sahrens * If this is a TX_WRITE with a blkptr, suck in the data. 1491fa9e4066Sahrens */ 1492fa9e4066Sahrens if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1493fa9e4066Sahrens lr_write_t *lrw = (lr_write_t *)lr; 1494fa9e4066Sahrens blkptr_t *wbp = &lrw->lr_blkptr; 1495fa9e4066Sahrens uint64_t wlen = lrw->lr_length; 1496fa9e4066Sahrens char *wbuf = zr->zr_lrbuf + reclen; 1497fa9e4066Sahrens 1498fa9e4066Sahrens if (BP_IS_HOLE(wbp)) { /* compressed to a hole */ 1499fa9e4066Sahrens bzero(wbuf, wlen); 1500fa9e4066Sahrens } else { 1501fa9e4066Sahrens /* 1502fa9e4066Sahrens * A subsequent write may have overwritten this block, 1503fa9e4066Sahrens * in which case wbp may have been been freed and 1504fa9e4066Sahrens * reallocated, and our read of wbp may fail with a 1505fa9e4066Sahrens * checksum error. We can safely ignore this because 1506fa9e4066Sahrens * the later write will provide the correct data. 1507fa9e4066Sahrens */ 1508ea8dc4b6Seschrock zbookmark_t zb; 1509ea8dc4b6Seschrock 1510ea8dc4b6Seschrock zb.zb_objset = dmu_objset_id(zilog->zl_os); 1511ea8dc4b6Seschrock zb.zb_object = lrw->lr_foid; 1512ea8dc4b6Seschrock zb.zb_level = -1; 1513ea8dc4b6Seschrock zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp); 1514ea8dc4b6Seschrock 1515fa9e4066Sahrens (void) zio_wait(zio_read(NULL, zilog->zl_spa, 1516fa9e4066Sahrens wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL, 1517fa9e4066Sahrens ZIO_PRIORITY_SYNC_READ, 1518ea8dc4b6Seschrock ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb)); 1519fa9e4066Sahrens (void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen); 1520fa9e4066Sahrens } 1521fa9e4066Sahrens } 1522fa9e4066Sahrens 1523fa9e4066Sahrens /* 1524fa9e4066Sahrens * We must now do two things atomically: replay this log record, 15251209a471SNeil Perrin * and update the log header sequence number to reflect the fact that 15261209a471SNeil Perrin * we did so. At the end of each replay function the sequence number 15271209a471SNeil Perrin * is updated if we are in replay mode. 1528fa9e4066Sahrens */ 15291209a471SNeil Perrin for (pass = 1; pass <= 2; pass++) { 15301209a471SNeil Perrin zilog->zl_replaying_seq = lr->lrc_seq; 15311209a471SNeil Perrin /* Only byteswap (if needed) on the 1st pass. */ 15321209a471SNeil Perrin error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf, 15331209a471SNeil Perrin zr->zr_byteswap && pass == 1); 1534fa9e4066Sahrens 153567bd71c6Sperrin if (!error) 153667bd71c6Sperrin return; 153767bd71c6Sperrin 153867bd71c6Sperrin /* 153967bd71c6Sperrin * The DMU's dnode layer doesn't see removes until the txg 154067bd71c6Sperrin * commits, so a subsequent claim can spuriously fail with 15411209a471SNeil Perrin * EEXIST. So if we receive any error we try syncing out 15421209a471SNeil Perrin * any removes then retry the transaction. 154367bd71c6Sperrin */ 15441209a471SNeil Perrin if (pass == 1) 154567bd71c6Sperrin txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1546fa9e4066Sahrens } 1547fa9e4066Sahrens 1548d30e856aSNeil Perrin bad: 15491209a471SNeil Perrin ASSERT(error); 155067bd71c6Sperrin name = kmem_alloc(MAXNAMELEN, KM_SLEEP); 155167bd71c6Sperrin dmu_objset_name(zr->zr_os, name); 155267bd71c6Sperrin cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1553da6c28aaSamw "dataset %s, seq 0x%llx, txtype %llu %s\n", 1554da6c28aaSamw error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype, 1555da6c28aaSamw (lr->lrc_txtype & TX_CI) ? "CI" : ""); 15561209a471SNeil Perrin zilog->zl_replay = B_FALSE; 155767bd71c6Sperrin kmem_free(name, MAXNAMELEN); 155867bd71c6Sperrin } 1559fa9e4066Sahrens 156067bd71c6Sperrin /* ARGSUSED */ 156167bd71c6Sperrin static void 156267bd71c6Sperrin zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 156367bd71c6Sperrin { 156467bd71c6Sperrin zilog->zl_replay_blks++; 1565fa9e4066Sahrens } 1566fa9e4066Sahrens 1567fa9e4066Sahrens /* 156813f5297eSperrin * If this dataset has a non-empty intent log, replay it and destroy it. 1569fa9e4066Sahrens */ 1570fa9e4066Sahrens void 15711209a471SNeil Perrin zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1572fa9e4066Sahrens { 1573fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1574d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1575d80c45e0Sbonwick zil_replay_arg_t zr; 157613f5297eSperrin 15773589c4f0SNeil Perrin if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1578d80c45e0Sbonwick zil_destroy(zilog, B_TRUE); 157913f5297eSperrin return; 158013f5297eSperrin } 1581fa9e4066Sahrens 1582fa9e4066Sahrens zr.zr_os = os; 1583fa9e4066Sahrens zr.zr_replay = replay_func; 1584fa9e4066Sahrens zr.zr_arg = arg; 1585d80c45e0Sbonwick zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1586fa9e4066Sahrens zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1587fa9e4066Sahrens 1588fa9e4066Sahrens /* 1589fa9e4066Sahrens * Wait for in-progress removes to sync before starting replay. 1590fa9e4066Sahrens */ 1591fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1592fa9e4066Sahrens 15931209a471SNeil Perrin zilog->zl_replay = B_TRUE; 159467bd71c6Sperrin zilog->zl_replay_time = lbolt; 159567bd71c6Sperrin ASSERT(zilog->zl_replay_blks == 0); 159667bd71c6Sperrin (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1597d80c45e0Sbonwick zh->zh_claim_txg); 1598fa9e4066Sahrens kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE); 1599fa9e4066Sahrens 1600d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1601a4611edeSahrens txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 16021209a471SNeil Perrin zilog->zl_replay = B_FALSE; 1603fa9e4066Sahrens } 1604436b2950Sperrin 1605436b2950Sperrin /* 1606436b2950Sperrin * Report whether all transactions are committed 1607436b2950Sperrin */ 1608436b2950Sperrin int 1609436b2950Sperrin zil_is_committed(zilog_t *zilog) 1610436b2950Sperrin { 1611436b2950Sperrin lwb_t *lwb; 1612b19a79ecSperrin int ret; 1613b19a79ecSperrin 1614b19a79ecSperrin mutex_enter(&zilog->zl_lock); 1615b19a79ecSperrin while (zilog->zl_writer) 1616b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1617b19a79ecSperrin 1618b19a79ecSperrin /* recent unpushed intent log transactions? */ 1619b19a79ecSperrin if (!list_is_empty(&zilog->zl_itx_list)) { 1620b19a79ecSperrin ret = B_FALSE; 1621b19a79ecSperrin goto out; 1622b19a79ecSperrin } 1623436b2950Sperrin 1624b19a79ecSperrin /* intent log never used? */ 1625b19a79ecSperrin lwb = list_head(&zilog->zl_lwb_list); 1626b19a79ecSperrin if (lwb == NULL) { 1627b19a79ecSperrin ret = B_TRUE; 1628b19a79ecSperrin goto out; 1629b19a79ecSperrin } 1630436b2950Sperrin 1631436b2950Sperrin /* 1632b19a79ecSperrin * more than 1 log buffer means zil_sync() hasn't yet freed 1633b19a79ecSperrin * entries after a txg has committed 1634436b2950Sperrin */ 1635b19a79ecSperrin if (list_next(&zilog->zl_lwb_list, lwb)) { 1636b19a79ecSperrin ret = B_FALSE; 1637b19a79ecSperrin goto out; 1638b19a79ecSperrin } 1639b19a79ecSperrin 1640436b2950Sperrin ASSERT(zil_empty(zilog)); 1641b19a79ecSperrin ret = B_TRUE; 1642b19a79ecSperrin out: 1643b19a79ecSperrin cv_broadcast(&zilog->zl_cv_writer); 1644b19a79ecSperrin mutex_exit(&zilog->zl_lock); 1645b19a79ecSperrin return (ret); 1646436b2950Sperrin } 1647*e6ca193dSGeorge Wilson 1648*e6ca193dSGeorge Wilson /* ARGSUSED */ 1649*e6ca193dSGeorge Wilson int 1650*e6ca193dSGeorge Wilson zil_vdev_offline(char *osname, void *arg) 1651*e6ca193dSGeorge Wilson { 1652*e6ca193dSGeorge Wilson objset_t *os; 1653*e6ca193dSGeorge Wilson zilog_t *zilog; 1654*e6ca193dSGeorge Wilson int error; 1655*e6ca193dSGeorge Wilson 1656*e6ca193dSGeorge Wilson error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 1657*e6ca193dSGeorge Wilson if (error) 1658*e6ca193dSGeorge Wilson return (error); 1659*e6ca193dSGeorge Wilson 1660*e6ca193dSGeorge Wilson zilog = dmu_objset_zil(os); 1661*e6ca193dSGeorge Wilson if (zil_suspend(zilog) != 0) 1662*e6ca193dSGeorge Wilson error = EEXIST; 1663*e6ca193dSGeorge Wilson else 1664*e6ca193dSGeorge Wilson zil_resume(zilog); 1665*e6ca193dSGeorge Wilson dmu_objset_close(os); 1666*e6ca193dSGeorge Wilson return (error); 1667*e6ca193dSGeorge Wilson } 1668