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 /* 2298f34ea2Sperrin * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 26fa9e4066Sahrens #pragma ident "%Z%%M% %I% %E% SMI" 27fa9e4066Sahrens 28fa9e4066Sahrens #include <sys/zfs_context.h> 29fa9e4066Sahrens #include <sys/spa.h> 30fa9e4066Sahrens #include <sys/dmu.h> 31fa9e4066Sahrens #include <sys/zap.h> 32fa9e4066Sahrens #include <sys/arc.h> 33fa9e4066Sahrens #include <sys/stat.h> 34fa9e4066Sahrens #include <sys/resource.h> 35fa9e4066Sahrens #include <sys/zil.h> 36fa9e4066Sahrens #include <sys/zil_impl.h> 37fa9e4066Sahrens #include <sys/dsl_dataset.h> 38fa9e4066Sahrens #include <sys/vdev.h> 39d63d470bSgw #include <sys/dmu_tx.h> 40fa9e4066Sahrens 41fa9e4066Sahrens /* 42fa9e4066Sahrens * The zfs intent log (ZIL) saves transaction records of system calls 43fa9e4066Sahrens * that change the file system in memory with enough information 44fa9e4066Sahrens * to be able to replay them. These are stored in memory until 45fa9e4066Sahrens * either the DMU transaction group (txg) commits them to the stable pool 46fa9e4066Sahrens * and they can be discarded, or they are flushed to the stable log 47fa9e4066Sahrens * (also in the pool) due to a fsync, O_DSYNC or other synchronous 48fa9e4066Sahrens * requirement. In the event of a panic or power fail then those log 49fa9e4066Sahrens * records (transactions) are replayed. 50fa9e4066Sahrens * 51fa9e4066Sahrens * There is one ZIL per file system. Its on-disk (pool) format consists 52fa9e4066Sahrens * of 3 parts: 53fa9e4066Sahrens * 54fa9e4066Sahrens * - ZIL header 55fa9e4066Sahrens * - ZIL blocks 56fa9e4066Sahrens * - ZIL records 57fa9e4066Sahrens * 58fa9e4066Sahrens * A log record holds a system call transaction. Log blocks can 59fa9e4066Sahrens * hold many log records and the blocks are chained together. 60fa9e4066Sahrens * Each ZIL block contains a block pointer (blkptr_t) to the next 61fa9e4066Sahrens * ZIL block in the chain. The ZIL header points to the first 62fa9e4066Sahrens * block in the chain. Note there is not a fixed place in the pool 63fa9e4066Sahrens * to hold blocks. They are dynamically allocated and freed as 64fa9e4066Sahrens * needed from the blocks available. Figure X shows the ZIL structure: 65fa9e4066Sahrens */ 66fa9e4066Sahrens 67fa9e4066Sahrens /* 68416e0cd8Sek * This global ZIL switch affects all pools 69fa9e4066Sahrens */ 70fa9e4066Sahrens int zil_disable = 0; /* disable intent logging */ 71416e0cd8Sek 72416e0cd8Sek /* 73416e0cd8Sek * Tunable parameter for debugging or performance analysis. Setting 74416e0cd8Sek * zfs_nocacheflush will cause corruption on power loss if a volatile 75416e0cd8Sek * out-of-order write cache is enabled. 76416e0cd8Sek */ 77416e0cd8Sek boolean_t zfs_nocacheflush = B_FALSE; 78fa9e4066Sahrens 79fa9e4066Sahrens static kmem_cache_t *zil_lwb_cache; 80fa9e4066Sahrens 81fa9e4066Sahrens static int 82fa9e4066Sahrens zil_dva_compare(const void *x1, const void *x2) 83fa9e4066Sahrens { 84fa9e4066Sahrens const dva_t *dva1 = x1; 85fa9e4066Sahrens const dva_t *dva2 = x2; 86fa9e4066Sahrens 87fa9e4066Sahrens if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 88fa9e4066Sahrens return (-1); 89fa9e4066Sahrens if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 90fa9e4066Sahrens return (1); 91fa9e4066Sahrens 92fa9e4066Sahrens if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 93fa9e4066Sahrens return (-1); 94fa9e4066Sahrens if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 95fa9e4066Sahrens return (1); 96fa9e4066Sahrens 97fa9e4066Sahrens return (0); 98fa9e4066Sahrens } 99fa9e4066Sahrens 100fa9e4066Sahrens static void 101fa9e4066Sahrens zil_dva_tree_init(avl_tree_t *t) 102fa9e4066Sahrens { 103fa9e4066Sahrens avl_create(t, zil_dva_compare, sizeof (zil_dva_node_t), 104fa9e4066Sahrens offsetof(zil_dva_node_t, zn_node)); 105fa9e4066Sahrens } 106fa9e4066Sahrens 107fa9e4066Sahrens static void 108fa9e4066Sahrens zil_dva_tree_fini(avl_tree_t *t) 109fa9e4066Sahrens { 110fa9e4066Sahrens zil_dva_node_t *zn; 111fa9e4066Sahrens void *cookie = NULL; 112fa9e4066Sahrens 113fa9e4066Sahrens while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 114fa9e4066Sahrens kmem_free(zn, sizeof (zil_dva_node_t)); 115fa9e4066Sahrens 116fa9e4066Sahrens avl_destroy(t); 117fa9e4066Sahrens } 118fa9e4066Sahrens 119fa9e4066Sahrens static int 120fa9e4066Sahrens zil_dva_tree_add(avl_tree_t *t, dva_t *dva) 121fa9e4066Sahrens { 122fa9e4066Sahrens zil_dva_node_t *zn; 123fa9e4066Sahrens avl_index_t where; 124fa9e4066Sahrens 125fa9e4066Sahrens if (avl_find(t, dva, &where) != NULL) 126fa9e4066Sahrens return (EEXIST); 127fa9e4066Sahrens 128fa9e4066Sahrens zn = kmem_alloc(sizeof (zil_dva_node_t), KM_SLEEP); 129fa9e4066Sahrens zn->zn_dva = *dva; 130fa9e4066Sahrens avl_insert(t, zn, where); 131fa9e4066Sahrens 132fa9e4066Sahrens return (0); 133fa9e4066Sahrens } 134fa9e4066Sahrens 135d80c45e0Sbonwick static zil_header_t * 136d80c45e0Sbonwick zil_header_in_syncing_context(zilog_t *zilog) 137d80c45e0Sbonwick { 138d80c45e0Sbonwick return ((zil_header_t *)zilog->zl_header); 139d80c45e0Sbonwick } 140d80c45e0Sbonwick 141d80c45e0Sbonwick static void 142d80c45e0Sbonwick zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 143d80c45e0Sbonwick { 144d80c45e0Sbonwick zio_cksum_t *zc = &bp->blk_cksum; 145d80c45e0Sbonwick 146d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 147d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 148d80c45e0Sbonwick zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 149d80c45e0Sbonwick zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 150d80c45e0Sbonwick } 151d80c45e0Sbonwick 152fa9e4066Sahrens /* 153fa9e4066Sahrens * Read a log block, make sure it's valid, and byteswap it if necessary. 154fa9e4066Sahrens */ 155fa9e4066Sahrens static int 156d80c45e0Sbonwick zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, arc_buf_t **abufpp) 157fa9e4066Sahrens { 158d80c45e0Sbonwick blkptr_t blk = *bp; 159ea8dc4b6Seschrock zbookmark_t zb; 16013506d1eSmaybee uint32_t aflags = ARC_WAIT; 161fa9e4066Sahrens int error; 162fa9e4066Sahrens 163d80c45e0Sbonwick zb.zb_objset = bp->blk_cksum.zc_word[ZIL_ZC_OBJSET]; 164ea8dc4b6Seschrock zb.zb_object = 0; 165ea8dc4b6Seschrock zb.zb_level = -1; 166d80c45e0Sbonwick zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ]; 167ea8dc4b6Seschrock 168d80c45e0Sbonwick *abufpp = NULL; 169fa9e4066Sahrens 170088f3894Sahrens /* 171088f3894Sahrens * We shouldn't be doing any scrubbing while we're doing log 172088f3894Sahrens * replay, it's OK to not lock. 173088f3894Sahrens */ 174088f3894Sahrens error = arc_read_nolock(NULL, zilog->zl_spa, &blk, 175d80c45e0Sbonwick arc_getbuf_func, abufpp, ZIO_PRIORITY_SYNC_READ, ZIO_FLAG_CANFAIL | 17613506d1eSmaybee ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB, &aflags, &zb); 177fa9e4066Sahrens 178d80c45e0Sbonwick if (error == 0) { 179d80c45e0Sbonwick char *data = (*abufpp)->b_data; 180d80c45e0Sbonwick uint64_t blksz = BP_GET_LSIZE(bp); 181d80c45e0Sbonwick zil_trailer_t *ztp = (zil_trailer_t *)(data + blksz) - 1; 182d80c45e0Sbonwick zio_cksum_t cksum = bp->blk_cksum; 183fa9e4066Sahrens 184d80c45e0Sbonwick /* 185d80c45e0Sbonwick * Sequence numbers should be... sequential. The checksum 186d80c45e0Sbonwick * verifier for the next block should be bp's checksum plus 1. 187d80c45e0Sbonwick */ 188d80c45e0Sbonwick cksum.zc_word[ZIL_ZC_SEQ]++; 189d80c45e0Sbonwick 190d80c45e0Sbonwick if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, sizeof (cksum))) 191d80c45e0Sbonwick error = ESTALE; 192d80c45e0Sbonwick else if (BP_IS_HOLE(&ztp->zit_next_blk)) 193d80c45e0Sbonwick error = ENOENT; 194d80c45e0Sbonwick else if (ztp->zit_nused > (blksz - sizeof (zil_trailer_t))) 195d80c45e0Sbonwick error = EOVERFLOW; 196fa9e4066Sahrens 197d80c45e0Sbonwick if (error) { 198d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(*abufpp, abufpp) == 1); 199d80c45e0Sbonwick *abufpp = NULL; 200d80c45e0Sbonwick } 201fa9e4066Sahrens } 202fa9e4066Sahrens 203d80c45e0Sbonwick dprintf("error %d on %llu:%llu\n", error, zb.zb_objset, zb.zb_blkid); 204fa9e4066Sahrens 205d80c45e0Sbonwick return (error); 206fa9e4066Sahrens } 207fa9e4066Sahrens 208fa9e4066Sahrens /* 209fa9e4066Sahrens * Parse the intent log, and call parse_func for each valid record within. 210d80c45e0Sbonwick * Return the highest sequence number. 211fa9e4066Sahrens */ 212d80c45e0Sbonwick uint64_t 213fa9e4066Sahrens zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 214fa9e4066Sahrens zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 215fa9e4066Sahrens { 216d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 217d80c45e0Sbonwick uint64_t claim_seq = zh->zh_claim_seq; 218d80c45e0Sbonwick uint64_t seq = 0; 219d80c45e0Sbonwick uint64_t max_seq = 0; 220d80c45e0Sbonwick blkptr_t blk = zh->zh_log; 221d80c45e0Sbonwick arc_buf_t *abuf; 222fa9e4066Sahrens char *lrbuf, *lrp; 223fa9e4066Sahrens zil_trailer_t *ztp; 224fa9e4066Sahrens int reclen, error; 225fa9e4066Sahrens 226fa9e4066Sahrens if (BP_IS_HOLE(&blk)) 227d80c45e0Sbonwick return (max_seq); 228fa9e4066Sahrens 229fa9e4066Sahrens /* 230fa9e4066Sahrens * Starting at the block pointed to by zh_log we read the log chain. 231fa9e4066Sahrens * For each block in the chain we strongly check that block to 232fa9e4066Sahrens * ensure its validity. We stop when an invalid block is found. 233fa9e4066Sahrens * For each block pointer in the chain we call parse_blk_func(). 234fa9e4066Sahrens * For each record in each valid block we call parse_lr_func(). 235d80c45e0Sbonwick * If the log has been claimed, stop if we encounter a sequence 236d80c45e0Sbonwick * number greater than the highest claimed sequence number. 237fa9e4066Sahrens */ 238fa9e4066Sahrens zil_dva_tree_init(&zilog->zl_dva_tree); 239fa9e4066Sahrens for (;;) { 240d80c45e0Sbonwick seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 241d80c45e0Sbonwick 242d80c45e0Sbonwick if (claim_seq != 0 && seq > claim_seq) 243d80c45e0Sbonwick break; 244d80c45e0Sbonwick 245d80c45e0Sbonwick ASSERT(max_seq < seq); 246d80c45e0Sbonwick max_seq = seq; 247d80c45e0Sbonwick 248d80c45e0Sbonwick error = zil_read_log_block(zilog, &blk, &abuf); 249fa9e4066Sahrens 250fa9e4066Sahrens if (parse_blk_func != NULL) 251fa9e4066Sahrens parse_blk_func(zilog, &blk, arg, txg); 252fa9e4066Sahrens 253fa9e4066Sahrens if (error) 254fa9e4066Sahrens break; 255fa9e4066Sahrens 256d80c45e0Sbonwick lrbuf = abuf->b_data; 257fa9e4066Sahrens ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 258fa9e4066Sahrens blk = ztp->zit_next_blk; 259fa9e4066Sahrens 260d80c45e0Sbonwick if (parse_lr_func == NULL) { 261d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 262fa9e4066Sahrens continue; 263d80c45e0Sbonwick } 264fa9e4066Sahrens 265fa9e4066Sahrens for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { 266fa9e4066Sahrens lr_t *lr = (lr_t *)lrp; 267fa9e4066Sahrens reclen = lr->lrc_reclen; 268fa9e4066Sahrens ASSERT3U(reclen, >=, sizeof (lr_t)); 269fa9e4066Sahrens parse_lr_func(zilog, lr, arg, txg); 270fa9e4066Sahrens } 271d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 272fa9e4066Sahrens } 273fa9e4066Sahrens zil_dva_tree_fini(&zilog->zl_dva_tree); 274d80c45e0Sbonwick 275d80c45e0Sbonwick return (max_seq); 276fa9e4066Sahrens } 277fa9e4066Sahrens 278fa9e4066Sahrens /* ARGSUSED */ 279fa9e4066Sahrens static void 280fa9e4066Sahrens zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 281fa9e4066Sahrens { 282fa9e4066Sahrens spa_t *spa = zilog->zl_spa; 283fa9e4066Sahrens int err; 284fa9e4066Sahrens 285fa9e4066Sahrens /* 286fa9e4066Sahrens * Claim log block if not already committed and not already claimed. 287fa9e4066Sahrens */ 288fa9e4066Sahrens if (bp->blk_birth >= first_txg && 289fa9e4066Sahrens zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp)) == 0) { 290fa9e4066Sahrens err = zio_wait(zio_claim(NULL, spa, first_txg, bp, NULL, NULL)); 291fa9e4066Sahrens ASSERT(err == 0); 292fa9e4066Sahrens } 293fa9e4066Sahrens } 294fa9e4066Sahrens 295fa9e4066Sahrens static void 296fa9e4066Sahrens zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 297fa9e4066Sahrens { 298fa9e4066Sahrens if (lrc->lrc_txtype == TX_WRITE) { 299fa9e4066Sahrens lr_write_t *lr = (lr_write_t *)lrc; 300fa9e4066Sahrens zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg); 301fa9e4066Sahrens } 302fa9e4066Sahrens } 303fa9e4066Sahrens 304fa9e4066Sahrens /* ARGSUSED */ 305fa9e4066Sahrens static void 306fa9e4066Sahrens zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 307fa9e4066Sahrens { 308fa9e4066Sahrens zio_free_blk(zilog->zl_spa, bp, dmu_tx_get_txg(tx)); 309fa9e4066Sahrens } 310fa9e4066Sahrens 311fa9e4066Sahrens static void 312fa9e4066Sahrens zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 313fa9e4066Sahrens { 314fa9e4066Sahrens /* 315fa9e4066Sahrens * If we previously claimed it, we need to free it. 316fa9e4066Sahrens */ 317fa9e4066Sahrens if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE) { 318fa9e4066Sahrens lr_write_t *lr = (lr_write_t *)lrc; 319fa9e4066Sahrens blkptr_t *bp = &lr->lr_blkptr; 320fa9e4066Sahrens if (bp->blk_birth >= claim_txg && 321fa9e4066Sahrens !zil_dva_tree_add(&zilog->zl_dva_tree, BP_IDENTITY(bp))) { 322fa9e4066Sahrens (void) arc_free(NULL, zilog->zl_spa, 323fa9e4066Sahrens dmu_tx_get_txg(tx), bp, NULL, NULL, ARC_WAIT); 324fa9e4066Sahrens } 325fa9e4066Sahrens } 326fa9e4066Sahrens } 327fa9e4066Sahrens 328fa9e4066Sahrens /* 329fa9e4066Sahrens * Create an on-disk intent log. 330fa9e4066Sahrens */ 331fa9e4066Sahrens static void 332fa9e4066Sahrens zil_create(zilog_t *zilog) 333fa9e4066Sahrens { 334d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 335fa9e4066Sahrens lwb_t *lwb; 336d80c45e0Sbonwick uint64_t txg = 0; 337d80c45e0Sbonwick dmu_tx_t *tx = NULL; 338fa9e4066Sahrens blkptr_t blk; 339d80c45e0Sbonwick int error = 0; 340fa9e4066Sahrens 341fa9e4066Sahrens /* 342d80c45e0Sbonwick * Wait for any previous destroy to complete. 343fa9e4066Sahrens */ 344d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 345d80c45e0Sbonwick 346d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 347d80c45e0Sbonwick ASSERT(zh->zh_replay_seq == 0); 348d80c45e0Sbonwick 349d80c45e0Sbonwick blk = zh->zh_log; 350fa9e4066Sahrens 351fa9e4066Sahrens /* 352d80c45e0Sbonwick * If we don't already have an initial log block, allocate one now. 353fa9e4066Sahrens */ 354d80c45e0Sbonwick if (BP_IS_HOLE(&blk)) { 355d80c45e0Sbonwick tx = dmu_tx_create(zilog->zl_os); 356d80c45e0Sbonwick (void) dmu_tx_assign(tx, TXG_WAIT); 357d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 358d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 359d80c45e0Sbonwick 36067bd71c6Sperrin error = zio_alloc_blk(zilog->zl_spa, ZIL_MIN_BLKSZ, &blk, 36167bd71c6Sperrin NULL, txg); 362d80c45e0Sbonwick 363d80c45e0Sbonwick if (error == 0) 364d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 36513f5297eSperrin } 366fa9e4066Sahrens 367d80c45e0Sbonwick /* 368d80c45e0Sbonwick * Allocate a log write buffer (lwb) for the first log block. 369d80c45e0Sbonwick */ 370d80c45e0Sbonwick if (error == 0) { 371fa9e4066Sahrens lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 372fa9e4066Sahrens lwb->lwb_zilog = zilog; 373fa9e4066Sahrens lwb->lwb_blk = blk; 374fa9e4066Sahrens lwb->lwb_nused = 0; 375fa9e4066Sahrens lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); 376fa9e4066Sahrens lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); 377fa9e4066Sahrens lwb->lwb_max_txg = txg; 378c5c6ffa0Smaybee lwb->lwb_zio = NULL; 379c5c6ffa0Smaybee 380fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 381fa9e4066Sahrens list_insert_tail(&zilog->zl_lwb_list, lwb); 382fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 383fa9e4066Sahrens } 384fa9e4066Sahrens 385d80c45e0Sbonwick /* 386d80c45e0Sbonwick * If we just allocated the first log block, commit our transaction 387d80c45e0Sbonwick * and wait for zil_sync() to stuff the block poiner into zh_log. 388d80c45e0Sbonwick * (zh is part of the MOS, so we cannot modify it in open context.) 389d80c45e0Sbonwick */ 390d80c45e0Sbonwick if (tx != NULL) { 391d80c45e0Sbonwick dmu_tx_commit(tx); 39213f5297eSperrin txg_wait_synced(zilog->zl_dmu_pool, txg); 393d80c45e0Sbonwick } 394d80c45e0Sbonwick 395d80c45e0Sbonwick ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 396fa9e4066Sahrens } 397fa9e4066Sahrens 398fa9e4066Sahrens /* 399fa9e4066Sahrens * In one tx, free all log blocks and clear the log header. 400d80c45e0Sbonwick * If keep_first is set, then we're replaying a log with no content. 401d80c45e0Sbonwick * We want to keep the first block, however, so that the first 402d80c45e0Sbonwick * synchronous transaction doesn't require a txg_wait_synced() 403d80c45e0Sbonwick * in zil_create(). We don't need to txg_wait_synced() here either 404d80c45e0Sbonwick * when keep_first is set, because both zil_create() and zil_destroy() 405d80c45e0Sbonwick * will wait for any in-progress destroys to complete. 406fa9e4066Sahrens */ 407fa9e4066Sahrens void 408d80c45e0Sbonwick zil_destroy(zilog_t *zilog, boolean_t keep_first) 409fa9e4066Sahrens { 410d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 411d80c45e0Sbonwick lwb_t *lwb; 412fa9e4066Sahrens dmu_tx_t *tx; 413fa9e4066Sahrens uint64_t txg; 414fa9e4066Sahrens 415d80c45e0Sbonwick /* 416d80c45e0Sbonwick * Wait for any previous destroy to complete. 417d80c45e0Sbonwick */ 418d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 419fa9e4066Sahrens 420d80c45e0Sbonwick if (BP_IS_HOLE(&zh->zh_log)) 421fa9e4066Sahrens return; 422fa9e4066Sahrens 423fa9e4066Sahrens tx = dmu_tx_create(zilog->zl_os); 424fa9e4066Sahrens (void) dmu_tx_assign(tx, TXG_WAIT); 425fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 426fa9e4066Sahrens txg = dmu_tx_get_txg(tx); 427fa9e4066Sahrens 428d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 429d80c45e0Sbonwick 430587c2a35Sperrin /* 431587c2a35Sperrin * It is possible for the ZIL to get the previously mounted zilog 432587c2a35Sperrin * structure of the same dataset if quickly remounted and the dbuf 433587c2a35Sperrin * eviction has not completed. In this case we can see a non 434587c2a35Sperrin * empty lwb list and keep_first will be set. We fix this by 435587c2a35Sperrin * clearing the keep_first. This will be slower but it's very rare. 436587c2a35Sperrin */ 437587c2a35Sperrin if (!list_is_empty(&zilog->zl_lwb_list) && keep_first) 438587c2a35Sperrin keep_first = B_FALSE; 439587c2a35Sperrin 440d80c45e0Sbonwick ASSERT3U(zilog->zl_destroy_txg, <, txg); 441fa9e4066Sahrens zilog->zl_destroy_txg = txg; 442d80c45e0Sbonwick zilog->zl_keep_first = keep_first; 443d80c45e0Sbonwick 444d80c45e0Sbonwick if (!list_is_empty(&zilog->zl_lwb_list)) { 445d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 446d80c45e0Sbonwick ASSERT(!keep_first); 447d80c45e0Sbonwick while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 448d80c45e0Sbonwick list_remove(&zilog->zl_lwb_list, lwb); 449d80c45e0Sbonwick if (lwb->lwb_buf != NULL) 450d80c45e0Sbonwick zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 451d80c45e0Sbonwick zio_free_blk(zilog->zl_spa, &lwb->lwb_blk, txg); 452d80c45e0Sbonwick kmem_cache_free(zil_lwb_cache, lwb); 453d80c45e0Sbonwick } 454d80c45e0Sbonwick } else { 455d80c45e0Sbonwick if (!keep_first) { 456d80c45e0Sbonwick (void) zil_parse(zilog, zil_free_log_block, 457d80c45e0Sbonwick zil_free_log_record, tx, zh->zh_claim_txg); 458d80c45e0Sbonwick } 459d80c45e0Sbonwick } 460b19a79ecSperrin mutex_exit(&zilog->zl_lock); 461fa9e4066Sahrens 462fa9e4066Sahrens dmu_tx_commit(tx); 463fa9e4066Sahrens } 464fa9e4066Sahrens 4653a8a1de4Sperrin /* 4663a8a1de4Sperrin * zil_rollback_destroy() is only called by the rollback code. 4673a8a1de4Sperrin * We already have a syncing tx. Rollback has exclusive access to the 4683a8a1de4Sperrin * dataset, so we don't have to worry about concurrent zil access. 4693a8a1de4Sperrin * The actual freeing of any log blocks occurs in zil_sync() later in 4703a8a1de4Sperrin * this txg syncing phase. 4713a8a1de4Sperrin */ 4723a8a1de4Sperrin void 4733a8a1de4Sperrin zil_rollback_destroy(zilog_t *zilog, dmu_tx_t *tx) 4743a8a1de4Sperrin { 4753a8a1de4Sperrin const zil_header_t *zh = zilog->zl_header; 4763a8a1de4Sperrin uint64_t txg; 4773a8a1de4Sperrin 4783a8a1de4Sperrin if (BP_IS_HOLE(&zh->zh_log)) 4793a8a1de4Sperrin return; 4803a8a1de4Sperrin 4813a8a1de4Sperrin txg = dmu_tx_get_txg(tx); 4823a8a1de4Sperrin ASSERT3U(zilog->zl_destroy_txg, <, txg); 4833a8a1de4Sperrin zilog->zl_destroy_txg = txg; 4843a8a1de4Sperrin zilog->zl_keep_first = B_FALSE; 4853a8a1de4Sperrin 48698f34ea2Sperrin /* 48798f34ea2Sperrin * Ensure there's no outstanding ZIL IO. No lwbs or just the 48898f34ea2Sperrin * unused one that allocated in advance is ok. 48998f34ea2Sperrin */ 49098f34ea2Sperrin ASSERT(zilog->zl_lwb_list.list_head.list_next == 49198f34ea2Sperrin zilog->zl_lwb_list.list_head.list_prev); 4923a8a1de4Sperrin (void) zil_parse(zilog, zil_free_log_block, zil_free_log_record, 4933a8a1de4Sperrin tx, zh->zh_claim_txg); 4943a8a1de4Sperrin } 4953a8a1de4Sperrin 4961d452cf5Sahrens int 497fa9e4066Sahrens zil_claim(char *osname, void *txarg) 498fa9e4066Sahrens { 499fa9e4066Sahrens dmu_tx_t *tx = txarg; 500fa9e4066Sahrens uint64_t first_txg = dmu_tx_get_txg(tx); 501fa9e4066Sahrens zilog_t *zilog; 502fa9e4066Sahrens zil_header_t *zh; 503fa9e4066Sahrens objset_t *os; 504fa9e4066Sahrens int error; 505fa9e4066Sahrens 506745cd3c5Smaybee error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 507fa9e4066Sahrens if (error) { 508*b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 5091d452cf5Sahrens return (0); 510fa9e4066Sahrens } 511fa9e4066Sahrens 512fa9e4066Sahrens zilog = dmu_objset_zil(os); 513d80c45e0Sbonwick zh = zil_header_in_syncing_context(zilog); 514fa9e4066Sahrens 515fa9e4066Sahrens /* 516d80c45e0Sbonwick * Claim all log blocks if we haven't already done so, and remember 517d80c45e0Sbonwick * the highest claimed sequence number. This ensures that if we can 518d80c45e0Sbonwick * read only part of the log now (e.g. due to a missing device), 519d80c45e0Sbonwick * but we can read the entire log later, we will not try to replay 520d80c45e0Sbonwick * or destroy beyond the last block we successfully claimed. 521fa9e4066Sahrens */ 522fa9e4066Sahrens ASSERT3U(zh->zh_claim_txg, <=, first_txg); 523fa9e4066Sahrens if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 524fa9e4066Sahrens zh->zh_claim_txg = first_txg; 525d80c45e0Sbonwick zh->zh_claim_seq = zil_parse(zilog, zil_claim_log_block, 526d80c45e0Sbonwick zil_claim_log_record, tx, first_txg); 527fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(os), tx); 528fa9e4066Sahrens } 529d80c45e0Sbonwick 530fa9e4066Sahrens ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 531fa9e4066Sahrens dmu_objset_close(os); 5321d452cf5Sahrens return (0); 533*b87f3af3Sperrin } 534*b87f3af3Sperrin 535*b87f3af3Sperrin /* 536*b87f3af3Sperrin * Check the log by walking the log chain. 537*b87f3af3Sperrin * Checksum errors are ok as they indicate the end of the chain. 538*b87f3af3Sperrin * Any other error (no device or read failure) returns an error. 539*b87f3af3Sperrin */ 540*b87f3af3Sperrin /* ARGSUSED */ 541*b87f3af3Sperrin int 542*b87f3af3Sperrin zil_check_log_chain(char *osname, void *txarg) 543*b87f3af3Sperrin { 544*b87f3af3Sperrin zilog_t *zilog; 545*b87f3af3Sperrin zil_header_t *zh; 546*b87f3af3Sperrin blkptr_t blk; 547*b87f3af3Sperrin arc_buf_t *abuf; 548*b87f3af3Sperrin objset_t *os; 549*b87f3af3Sperrin char *lrbuf; 550*b87f3af3Sperrin zil_trailer_t *ztp; 551*b87f3af3Sperrin int error; 552*b87f3af3Sperrin 553*b87f3af3Sperrin error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 554*b87f3af3Sperrin if (error) { 555*b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 556*b87f3af3Sperrin return (0); 557*b87f3af3Sperrin } 558*b87f3af3Sperrin 559*b87f3af3Sperrin zilog = dmu_objset_zil(os); 560*b87f3af3Sperrin zh = zil_header_in_syncing_context(zilog); 561*b87f3af3Sperrin blk = zh->zh_log; 562*b87f3af3Sperrin if (BP_IS_HOLE(&blk)) { 563*b87f3af3Sperrin dmu_objset_close(os); 564*b87f3af3Sperrin return (0); /* no chain */ 565*b87f3af3Sperrin } 566*b87f3af3Sperrin 567*b87f3af3Sperrin for (;;) { 568*b87f3af3Sperrin error = zil_read_log_block(zilog, &blk, &abuf); 569*b87f3af3Sperrin if (error) 570*b87f3af3Sperrin break; 571*b87f3af3Sperrin lrbuf = abuf->b_data; 572*b87f3af3Sperrin ztp = (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 573*b87f3af3Sperrin blk = ztp->zit_next_blk; 574*b87f3af3Sperrin VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 575*b87f3af3Sperrin } 576*b87f3af3Sperrin dmu_objset_close(os); 577*b87f3af3Sperrin if (error == ECKSUM) 578*b87f3af3Sperrin return (0); /* normal end of chain */ 579*b87f3af3Sperrin return (error); 580*b87f3af3Sperrin } 581*b87f3af3Sperrin 582*b87f3af3Sperrin /* 583*b87f3af3Sperrin * Clear a log chain 584*b87f3af3Sperrin */ 585*b87f3af3Sperrin /* ARGSUSED */ 586*b87f3af3Sperrin int 587*b87f3af3Sperrin zil_clear_log_chain(char *osname, void *txarg) 588*b87f3af3Sperrin { 589*b87f3af3Sperrin zilog_t *zilog; 590*b87f3af3Sperrin zil_header_t *zh; 591*b87f3af3Sperrin objset_t *os; 592*b87f3af3Sperrin dmu_tx_t *tx; 593*b87f3af3Sperrin int error; 594*b87f3af3Sperrin 595*b87f3af3Sperrin error = dmu_objset_open(osname, DMU_OST_ANY, DS_MODE_USER, &os); 596*b87f3af3Sperrin if (error) { 597*b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 598*b87f3af3Sperrin return (0); 599*b87f3af3Sperrin } 600*b87f3af3Sperrin 601*b87f3af3Sperrin zilog = dmu_objset_zil(os); 602*b87f3af3Sperrin tx = dmu_tx_create(zilog->zl_os); 603*b87f3af3Sperrin (void) dmu_tx_assign(tx, TXG_WAIT); 604*b87f3af3Sperrin zh = zil_header_in_syncing_context(zilog); 605*b87f3af3Sperrin BP_ZERO(&zh->zh_log); 606*b87f3af3Sperrin dsl_dataset_dirty(dmu_objset_ds(os), tx); 607*b87f3af3Sperrin dmu_tx_commit(tx); 608*b87f3af3Sperrin dmu_objset_close(os); 609*b87f3af3Sperrin return (0); 610fa9e4066Sahrens } 611fa9e4066Sahrens 61217f17c2dSbonwick static int 61317f17c2dSbonwick zil_vdev_compare(const void *x1, const void *x2) 61417f17c2dSbonwick { 615c0d19b33Sperrin uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 616c0d19b33Sperrin uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 61717f17c2dSbonwick 61817f17c2dSbonwick if (v1 < v2) 61917f17c2dSbonwick return (-1); 62017f17c2dSbonwick if (v1 > v2) 62117f17c2dSbonwick return (1); 62217f17c2dSbonwick 62317f17c2dSbonwick return (0); 62417f17c2dSbonwick } 62517f17c2dSbonwick 626fa9e4066Sahrens void 62717f17c2dSbonwick zil_add_block(zilog_t *zilog, blkptr_t *bp) 628fa9e4066Sahrens { 62917f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 63017f17c2dSbonwick avl_index_t where; 63117f17c2dSbonwick zil_vdev_node_t *zv, zvsearch; 63217f17c2dSbonwick int ndvas = BP_GET_NDVAS(bp); 63317f17c2dSbonwick int i; 634fa9e4066Sahrens 635416e0cd8Sek if (zfs_nocacheflush) 636fa9e4066Sahrens return; 637fa9e4066Sahrens 63817f17c2dSbonwick ASSERT(zilog->zl_writer); 63917f17c2dSbonwick 64017f17c2dSbonwick /* 64117f17c2dSbonwick * Even though we're zl_writer, we still need a lock because the 64217f17c2dSbonwick * zl_get_data() callbacks may have dmu_sync() done callbacks 64317f17c2dSbonwick * that will run concurrently. 64417f17c2dSbonwick */ 64517f17c2dSbonwick mutex_enter(&zilog->zl_vdev_lock); 64617f17c2dSbonwick for (i = 0; i < ndvas; i++) { 64717f17c2dSbonwick zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 64817f17c2dSbonwick if (avl_find(t, &zvsearch, &where) == NULL) { 64917f17c2dSbonwick zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 65017f17c2dSbonwick zv->zv_vdev = zvsearch.zv_vdev; 65117f17c2dSbonwick avl_insert(t, zv, where); 65267bd71c6Sperrin } 65367bd71c6Sperrin } 65417f17c2dSbonwick mutex_exit(&zilog->zl_vdev_lock); 655fa9e4066Sahrens } 656fa9e4066Sahrens 65767bd71c6Sperrin void 65867bd71c6Sperrin zil_flush_vdevs(zilog_t *zilog) 65967bd71c6Sperrin { 66067bd71c6Sperrin spa_t *spa = zilog->zl_spa; 66117f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 66217f17c2dSbonwick void *cookie = NULL; 66317f17c2dSbonwick zil_vdev_node_t *zv; 66417f17c2dSbonwick zio_t *zio; 665fa9e4066Sahrens 66667bd71c6Sperrin ASSERT(zilog->zl_writer); 66767bd71c6Sperrin 66817f17c2dSbonwick /* 66917f17c2dSbonwick * We don't need zl_vdev_lock here because we're the zl_writer, 67017f17c2dSbonwick * and all zl_get_data() callbacks are done. 67117f17c2dSbonwick */ 67217f17c2dSbonwick if (avl_numnodes(t) == 0) 67317f17c2dSbonwick return; 67417f17c2dSbonwick 67517f17c2dSbonwick spa_config_enter(spa, RW_READER, FTAG); 67617f17c2dSbonwick 67717f17c2dSbonwick zio = zio_root(spa, NULL, NULL, 67817f17c2dSbonwick ZIO_FLAG_CONFIG_HELD | ZIO_FLAG_CANFAIL); 679fa9e4066Sahrens 68017f17c2dSbonwick while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 68117f17c2dSbonwick vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 68217f17c2dSbonwick if (vd != NULL) 68317f17c2dSbonwick zio_flush(zio, vd); 68417f17c2dSbonwick kmem_free(zv, sizeof (*zv)); 68567bd71c6Sperrin } 68617f17c2dSbonwick 687fa9e4066Sahrens /* 688fa9e4066Sahrens * Wait for all the flushes to complete. Not all devices actually 689fa9e4066Sahrens * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 690fa9e4066Sahrens */ 69117f17c2dSbonwick (void) zio_wait(zio); 69217f17c2dSbonwick 69317f17c2dSbonwick spa_config_exit(spa, FTAG); 694fa9e4066Sahrens } 695fa9e4066Sahrens 696fa9e4066Sahrens /* 697fa9e4066Sahrens * Function called when a log block write completes 698fa9e4066Sahrens */ 699fa9e4066Sahrens static void 700fa9e4066Sahrens zil_lwb_write_done(zio_t *zio) 701fa9e4066Sahrens { 702fa9e4066Sahrens lwb_t *lwb = zio->io_private; 703fa9e4066Sahrens zilog_t *zilog = lwb->lwb_zilog; 704fa9e4066Sahrens 705fa9e4066Sahrens /* 706fa9e4066Sahrens * Now that we've written this log block, we have a stable pointer 707fa9e4066Sahrens * to the next block in the chain, so it's OK to let the txg in 708fa9e4066Sahrens * which we allocated the next block sync. 709fa9e4066Sahrens */ 710fa9e4066Sahrens txg_rele_to_sync(&lwb->lwb_txgh); 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); 718fa9e4066Sahrens } 719fa9e4066Sahrens 720c5c6ffa0Smaybee /* 721c5c6ffa0Smaybee * Initialize the io for a log block. 722c5c6ffa0Smaybee * 723c5c6ffa0Smaybee * Note, we should not initialize the IO until we are about 724c5c6ffa0Smaybee * to use it, since zio_rewrite() does a spa_config_enter(). 725c5c6ffa0Smaybee */ 726c5c6ffa0Smaybee static void 727c5c6ffa0Smaybee zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 728c5c6ffa0Smaybee { 729c5c6ffa0Smaybee zbookmark_t zb; 730c5c6ffa0Smaybee 731c5c6ffa0Smaybee zb.zb_objset = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET]; 732c5c6ffa0Smaybee zb.zb_object = 0; 733c5c6ffa0Smaybee zb.zb_level = -1; 734c5c6ffa0Smaybee zb.zb_blkid = lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 735c5c6ffa0Smaybee 736b19a79ecSperrin if (zilog->zl_root_zio == NULL) { 737b19a79ecSperrin zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 738b19a79ecSperrin ZIO_FLAG_CANFAIL); 739b19a79ecSperrin } 74067bd71c6Sperrin if (lwb->lwb_zio == NULL) { 74167bd71c6Sperrin lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 7426203a313Sperrin ZIO_CHECKSUM_ZILOG, 0, &lwb->lwb_blk, lwb->lwb_buf, 74367bd71c6Sperrin lwb->lwb_sz, zil_lwb_write_done, lwb, 7448654d025Sperrin ZIO_PRIORITY_LOG_WRITE, ZIO_FLAG_CANFAIL, &zb); 74567bd71c6Sperrin } 746c5c6ffa0Smaybee } 747c5c6ffa0Smaybee 748fa9e4066Sahrens /* 749fa9e4066Sahrens * Start a log block write and advance to the next log block. 750fa9e4066Sahrens * Calls are serialized. 751fa9e4066Sahrens */ 752fa9e4066Sahrens static lwb_t * 753fa9e4066Sahrens zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 754fa9e4066Sahrens { 755fa9e4066Sahrens lwb_t *nlwb; 756fa9e4066Sahrens zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; 757d80c45e0Sbonwick spa_t *spa = zilog->zl_spa; 758d80c45e0Sbonwick blkptr_t *bp = &ztp->zit_next_blk; 759fa9e4066Sahrens uint64_t txg; 760fa9e4066Sahrens uint64_t zil_blksz; 761fa9e4066Sahrens int error; 762fa9e4066Sahrens 763fa9e4066Sahrens ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); 764fa9e4066Sahrens 765fa9e4066Sahrens /* 766fa9e4066Sahrens * Allocate the next block and save its address in this block 767fa9e4066Sahrens * before writing it in order to establish the log chain. 768fa9e4066Sahrens * Note that if the allocation of nlwb synced before we wrote 769fa9e4066Sahrens * the block that points at it (lwb), we'd leak it if we crashed. 770fa9e4066Sahrens * Therefore, we don't do txg_rele_to_sync() until zil_lwb_write_done(). 771fa9e4066Sahrens */ 772fa9e4066Sahrens txg = txg_hold_open(zilog->zl_dmu_pool, &lwb->lwb_txgh); 773fa9e4066Sahrens txg_rele_to_quiesce(&lwb->lwb_txgh); 774fa9e4066Sahrens 775fa9e4066Sahrens /* 77622ac5be4Sperrin * Pick a ZIL blocksize. We request a size that is the 77722ac5be4Sperrin * maximum of the previous used size, the current used size and 77822ac5be4Sperrin * the amount waiting in the queue. 779fa9e4066Sahrens */ 780c5c6ffa0Smaybee zil_blksz = MAX(zilog->zl_prev_used, 781c5c6ffa0Smaybee zilog->zl_cur_used + sizeof (*ztp)); 78222ac5be4Sperrin zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); 783b4d654b0Sperrin zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); 78422ac5be4Sperrin if (zil_blksz > ZIL_MAX_BLKSZ) 78522ac5be4Sperrin zil_blksz = ZIL_MAX_BLKSZ; 786fa9e4066Sahrens 78767bd71c6Sperrin BP_ZERO(bp); 78867bd71c6Sperrin /* pass the old blkptr in order to spread log blocks across devs */ 78967bd71c6Sperrin error = zio_alloc_blk(spa, zil_blksz, bp, &lwb->lwb_blk, txg); 790fa9e4066Sahrens if (error) { 791d63d470bSgw dmu_tx_t *tx = dmu_tx_create_assigned(zilog->zl_dmu_pool, txg); 792d63d470bSgw 793d63d470bSgw /* 794d63d470bSgw * We dirty the dataset to ensure that zil_sync() will 795d63d470bSgw * be called to remove this lwb from our zl_lwb_list. 796d63d470bSgw * Failing to do so, may leave an lwb with a NULL lwb_buf 797d63d470bSgw * hanging around on the zl_lwb_list. 798d63d470bSgw */ 799d63d470bSgw dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 800aeb1c1b6Sgw dmu_tx_commit(tx); 801d63d470bSgw 802d63d470bSgw /* 803d63d470bSgw * Since we've just experienced an allocation failure so we 804d63d470bSgw * terminate the current lwb and send it on its way. 805d63d470bSgw */ 806d63d470bSgw ztp->zit_pad = 0; 807d63d470bSgw ztp->zit_nused = lwb->lwb_nused; 808d63d470bSgw ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 809d63d470bSgw zio_nowait(lwb->lwb_zio); 810d63d470bSgw 811ea8dc4b6Seschrock /* 812ea8dc4b6Seschrock * By returning NULL the caller will call tx_wait_synced() 813ea8dc4b6Seschrock */ 814fa9e4066Sahrens return (NULL); 815fa9e4066Sahrens } 816fa9e4066Sahrens 817d80c45e0Sbonwick ASSERT3U(bp->blk_birth, ==, txg); 818ea8dc4b6Seschrock ztp->zit_pad = 0; 819fa9e4066Sahrens ztp->zit_nused = lwb->lwb_nused; 820fa9e4066Sahrens ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 821d80c45e0Sbonwick bp->blk_cksum = lwb->lwb_blk.blk_cksum; 822d80c45e0Sbonwick bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 823fa9e4066Sahrens 824fa9e4066Sahrens /* 825fa9e4066Sahrens * Allocate a new log write buffer (lwb). 826fa9e4066Sahrens */ 827fa9e4066Sahrens nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 828fa9e4066Sahrens 829fa9e4066Sahrens nlwb->lwb_zilog = zilog; 830d80c45e0Sbonwick nlwb->lwb_blk = *bp; 831fa9e4066Sahrens nlwb->lwb_nused = 0; 832fa9e4066Sahrens nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); 833fa9e4066Sahrens nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); 834fa9e4066Sahrens nlwb->lwb_max_txg = txg; 835c5c6ffa0Smaybee nlwb->lwb_zio = NULL; 836fa9e4066Sahrens 837fa9e4066Sahrens /* 83867bd71c6Sperrin * Put new lwb at the end of the log chain 839fa9e4066Sahrens */ 840fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 841fa9e4066Sahrens list_insert_tail(&zilog->zl_lwb_list, nlwb); 842fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 843fa9e4066Sahrens 84417f17c2dSbonwick /* Record the block for later vdev flushing */ 84517f17c2dSbonwick zil_add_block(zilog, &lwb->lwb_blk); 84667bd71c6Sperrin 847fa9e4066Sahrens /* 848c5c6ffa0Smaybee * kick off the write for the old log block 849fa9e4066Sahrens */ 850c5c6ffa0Smaybee dprintf_bp(&lwb->lwb_blk, "lwb %p txg %llu: ", lwb, txg); 85167bd71c6Sperrin ASSERT(lwb->lwb_zio); 852c5c6ffa0Smaybee zio_nowait(lwb->lwb_zio); 853fa9e4066Sahrens 854fa9e4066Sahrens return (nlwb); 855fa9e4066Sahrens } 856fa9e4066Sahrens 857fa9e4066Sahrens static lwb_t * 858fa9e4066Sahrens zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 859fa9e4066Sahrens { 860fa9e4066Sahrens lr_t *lrc = &itx->itx_lr; /* common log record */ 861c5c6ffa0Smaybee lr_write_t *lr = (lr_write_t *)lrc; 862fa9e4066Sahrens uint64_t txg = lrc->lrc_txg; 863fa9e4066Sahrens uint64_t reclen = lrc->lrc_reclen; 864c5c6ffa0Smaybee uint64_t dlen; 865fa9e4066Sahrens 866fa9e4066Sahrens if (lwb == NULL) 867fa9e4066Sahrens return (NULL); 868fa9e4066Sahrens ASSERT(lwb->lwb_buf != NULL); 869fa9e4066Sahrens 870c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 871c5c6ffa0Smaybee dlen = P2ROUNDUP_TYPED( 872c5c6ffa0Smaybee lr->lr_length, sizeof (uint64_t), uint64_t); 873c5c6ffa0Smaybee else 874c5c6ffa0Smaybee dlen = 0; 875fa9e4066Sahrens 876104e2ed7Sperrin zilog->zl_cur_used += (reclen + dlen); 87722ac5be4Sperrin 87867bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 87967bd71c6Sperrin 880fa9e4066Sahrens /* 881fa9e4066Sahrens * If this record won't fit in the current log block, start a new one. 882fa9e4066Sahrens */ 883104e2ed7Sperrin if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 884fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 885c5c6ffa0Smaybee if (lwb == NULL) 886fa9e4066Sahrens return (NULL); 88767bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 888ea8dc4b6Seschrock ASSERT(lwb->lwb_nused == 0); 889104e2ed7Sperrin if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 890fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, txg); 891fa9e4066Sahrens return (lwb); 892fa9e4066Sahrens } 893fa9e4066Sahrens } 894fa9e4066Sahrens 895b19a79ecSperrin /* 896b19a79ecSperrin * Update the lrc_seq, to be log record sequence number. See zil.h 897b19a79ecSperrin * Then copy the record to the log buffer. 898b19a79ecSperrin */ 899b19a79ecSperrin lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 900fa9e4066Sahrens bcopy(lrc, lwb->lwb_buf + lwb->lwb_nused, reclen); 901c5c6ffa0Smaybee 902c5c6ffa0Smaybee /* 903c5c6ffa0Smaybee * If it's a write, fetch the data or get its blkptr as appropriate. 904c5c6ffa0Smaybee */ 905c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE) { 906c5c6ffa0Smaybee if (txg > spa_freeze_txg(zilog->zl_spa)) 907c5c6ffa0Smaybee txg_wait_synced(zilog->zl_dmu_pool, txg); 908c5c6ffa0Smaybee if (itx->itx_wr_state != WR_COPIED) { 909c5c6ffa0Smaybee char *dbuf; 910c5c6ffa0Smaybee int error; 911c5c6ffa0Smaybee 912c5c6ffa0Smaybee /* alignment is guaranteed */ 913c5c6ffa0Smaybee lr = (lr_write_t *)(lwb->lwb_buf + lwb->lwb_nused); 914c5c6ffa0Smaybee if (dlen) { 915c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_NEED_COPY); 916c5c6ffa0Smaybee dbuf = lwb->lwb_buf + lwb->lwb_nused + reclen; 917c5c6ffa0Smaybee lr->lr_common.lrc_reclen += dlen; 918c5c6ffa0Smaybee } else { 919c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_INDIRECT); 920c5c6ffa0Smaybee dbuf = NULL; 921c5c6ffa0Smaybee } 922c5c6ffa0Smaybee error = zilog->zl_get_data( 923c5c6ffa0Smaybee itx->itx_private, lr, dbuf, lwb->lwb_zio); 924c5c6ffa0Smaybee if (error) { 925c5c6ffa0Smaybee ASSERT(error == ENOENT || error == EEXIST || 926c5c6ffa0Smaybee error == EALREADY); 927c5c6ffa0Smaybee return (lwb); 928c5c6ffa0Smaybee } 929c5c6ffa0Smaybee } 930104e2ed7Sperrin } 931c5c6ffa0Smaybee 932c5c6ffa0Smaybee lwb->lwb_nused += reclen + dlen; 933fa9e4066Sahrens lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 934fa9e4066Sahrens ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); 935fa9e4066Sahrens ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 936fa9e4066Sahrens 937fa9e4066Sahrens return (lwb); 938fa9e4066Sahrens } 939fa9e4066Sahrens 940fa9e4066Sahrens itx_t * 941da6c28aaSamw zil_itx_create(uint64_t txtype, size_t lrsize) 942fa9e4066Sahrens { 943fa9e4066Sahrens itx_t *itx; 944fa9e4066Sahrens 945b4d654b0Sperrin lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 946fa9e4066Sahrens 947fa9e4066Sahrens itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 948fa9e4066Sahrens itx->itx_lr.lrc_txtype = txtype; 949fa9e4066Sahrens itx->itx_lr.lrc_reclen = lrsize; 950abf76b6eSperrin itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 951fa9e4066Sahrens itx->itx_lr.lrc_seq = 0; /* defensive */ 952fa9e4066Sahrens 953fa9e4066Sahrens return (itx); 954fa9e4066Sahrens } 955fa9e4066Sahrens 956fa9e4066Sahrens uint64_t 957fa9e4066Sahrens zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 958fa9e4066Sahrens { 959fa9e4066Sahrens uint64_t seq; 960fa9e4066Sahrens 961fa9e4066Sahrens ASSERT(itx->itx_lr.lrc_seq == 0); 962fa9e4066Sahrens 963fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 964fa9e4066Sahrens list_insert_tail(&zilog->zl_itx_list, itx); 965abf76b6eSperrin zilog->zl_itx_list_sz += itx->itx_sod; 966fa9e4066Sahrens itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 967fa9e4066Sahrens itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 968fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 969fa9e4066Sahrens 970fa9e4066Sahrens return (seq); 971fa9e4066Sahrens } 972fa9e4066Sahrens 973fa9e4066Sahrens /* 974fa9e4066Sahrens * Free up all in-memory intent log transactions that have now been synced. 975fa9e4066Sahrens */ 976fa9e4066Sahrens static void 977fa9e4066Sahrens zil_itx_clean(zilog_t *zilog) 978fa9e4066Sahrens { 979fa9e4066Sahrens uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 980fa9e4066Sahrens uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 981a584ef65Sjohansen list_t clean_list; 982fa9e4066Sahrens itx_t *itx; 983fa9e4066Sahrens 984a584ef65Sjohansen list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 985a584ef65Sjohansen 986fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 987b19a79ecSperrin /* wait for a log writer to finish walking list */ 988b19a79ecSperrin while (zilog->zl_writer) { 989b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 990b19a79ecSperrin } 991a584ef65Sjohansen 992a584ef65Sjohansen /* 993a584ef65Sjohansen * Move the sync'd log transactions to a separate list so we can call 994a584ef65Sjohansen * kmem_free without holding the zl_lock. 995a584ef65Sjohansen * 996a584ef65Sjohansen * There is no need to set zl_writer as we don't drop zl_lock here 997a584ef65Sjohansen */ 998fa9e4066Sahrens while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 999fa9e4066Sahrens itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 1000fa9e4066Sahrens list_remove(&zilog->zl_itx_list, itx); 1001abf76b6eSperrin zilog->zl_itx_list_sz -= itx->itx_sod; 1002a584ef65Sjohansen list_insert_tail(&clean_list, itx); 1003fa9e4066Sahrens } 100467bd71c6Sperrin cv_broadcast(&zilog->zl_cv_writer); 1005fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1006a584ef65Sjohansen 1007a584ef65Sjohansen /* destroy sync'd log transactions */ 1008a584ef65Sjohansen while ((itx = list_head(&clean_list)) != NULL) { 1009a584ef65Sjohansen list_remove(&clean_list, itx); 1010a584ef65Sjohansen kmem_free(itx, offsetof(itx_t, itx_lr) 1011a584ef65Sjohansen + itx->itx_lr.lrc_reclen); 1012a584ef65Sjohansen } 1013a584ef65Sjohansen list_destroy(&clean_list); 1014fa9e4066Sahrens } 1015fa9e4066Sahrens 1016b19a79ecSperrin /* 101767bd71c6Sperrin * If there are any in-memory intent log transactions which have now been 101867bd71c6Sperrin * synced then start up a taskq to free them. 1019b19a79ecSperrin */ 1020fa9e4066Sahrens void 1021fa9e4066Sahrens zil_clean(zilog_t *zilog) 1022fa9e4066Sahrens { 102367bd71c6Sperrin itx_t *itx; 102467bd71c6Sperrin 1025fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 102667bd71c6Sperrin itx = list_head(&zilog->zl_itx_list); 102767bd71c6Sperrin if ((itx != NULL) && 102867bd71c6Sperrin (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 1029fa9e4066Sahrens (void) taskq_dispatch(zilog->zl_clean_taskq, 1030fa9e4066Sahrens (void (*)(void *))zil_itx_clean, zilog, TQ_NOSLEEP); 103167bd71c6Sperrin } 1032fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1033fa9e4066Sahrens } 1034fa9e4066Sahrens 1035fa9e4066Sahrens void 1036b19a79ecSperrin zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 1037fa9e4066Sahrens { 1038fa9e4066Sahrens uint64_t txg; 103967bd71c6Sperrin uint64_t commit_seq = 0; 1040b19a79ecSperrin itx_t *itx, *itx_next = (itx_t *)-1; 1041fa9e4066Sahrens lwb_t *lwb; 1042fa9e4066Sahrens spa_t *spa; 1043fa9e4066Sahrens 1044fa9e4066Sahrens zilog->zl_writer = B_TRUE; 1045b19a79ecSperrin zilog->zl_root_zio = NULL; 1046b19a79ecSperrin spa = zilog->zl_spa; 1047fa9e4066Sahrens 1048fa9e4066Sahrens if (zilog->zl_suspend) { 1049fa9e4066Sahrens lwb = NULL; 1050fa9e4066Sahrens } else { 1051fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 1052fa9e4066Sahrens if (lwb == NULL) { 1053b19a79ecSperrin /* 1054b19a79ecSperrin * Return if there's nothing to flush before we 1055b19a79ecSperrin * dirty the fs by calling zil_create() 1056b19a79ecSperrin */ 1057b19a79ecSperrin if (list_is_empty(&zilog->zl_itx_list)) { 1058b19a79ecSperrin zilog->zl_writer = B_FALSE; 1059b19a79ecSperrin return; 1060b19a79ecSperrin } 1061fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1062fa9e4066Sahrens zil_create(zilog); 1063fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1064fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 1065fa9e4066Sahrens } 1066fa9e4066Sahrens } 1067fa9e4066Sahrens 106867bd71c6Sperrin /* Loop through in-memory log transactions filling log blocks. */ 1069b19a79ecSperrin DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1070fa9e4066Sahrens for (;;) { 1071b19a79ecSperrin /* 1072b19a79ecSperrin * Find the next itx to push: 1073b19a79ecSperrin * Push all transactions related to specified foid and all 1074b19a79ecSperrin * other transactions except TX_WRITE, TX_TRUNCATE, 1075b19a79ecSperrin * TX_SETATTR and TX_ACL for all other files. 1076b19a79ecSperrin */ 1077b19a79ecSperrin if (itx_next != (itx_t *)-1) 1078b19a79ecSperrin itx = itx_next; 1079b19a79ecSperrin else 1080b19a79ecSperrin itx = list_head(&zilog->zl_itx_list); 1081b19a79ecSperrin for (; itx != NULL; itx = list_next(&zilog->zl_itx_list, itx)) { 1082b19a79ecSperrin if (foid == 0) /* push all foids? */ 1083b19a79ecSperrin break; 108467bd71c6Sperrin if (itx->itx_sync) /* push all O_[D]SYNC */ 108567bd71c6Sperrin break; 1086b19a79ecSperrin switch (itx->itx_lr.lrc_txtype) { 1087b19a79ecSperrin case TX_SETATTR: 1088b19a79ecSperrin case TX_WRITE: 1089b19a79ecSperrin case TX_TRUNCATE: 1090b19a79ecSperrin case TX_ACL: 1091b19a79ecSperrin /* lr_foid is same offset for these records */ 1092b19a79ecSperrin if (((lr_write_t *)&itx->itx_lr)->lr_foid 1093b19a79ecSperrin != foid) { 1094b19a79ecSperrin continue; /* skip this record */ 1095b19a79ecSperrin } 1096b19a79ecSperrin } 1097b19a79ecSperrin break; 1098b19a79ecSperrin } 1099fa9e4066Sahrens if (itx == NULL) 1100fa9e4066Sahrens break; 1101fa9e4066Sahrens 1102fa9e4066Sahrens if ((itx->itx_lr.lrc_seq > seq) && 1103b19a79ecSperrin ((lwb == NULL) || (lwb->lwb_nused == 0) || 1104abf76b6eSperrin (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) { 1105fa9e4066Sahrens break; 110667bd71c6Sperrin } 1107fa9e4066Sahrens 1108b19a79ecSperrin /* 1109b19a79ecSperrin * Save the next pointer. Even though we soon drop 1110b19a79ecSperrin * zl_lock all threads that may change the list 1111b19a79ecSperrin * (another writer or zil_itx_clean) can't do so until 1112b19a79ecSperrin * they have zl_writer. 1113b19a79ecSperrin */ 1114b19a79ecSperrin itx_next = list_next(&zilog->zl_itx_list, itx); 1115fa9e4066Sahrens list_remove(&zilog->zl_itx_list, itx); 1116abf76b6eSperrin zilog->zl_itx_list_sz -= itx->itx_sod; 111767bd71c6Sperrin mutex_exit(&zilog->zl_lock); 1118fa9e4066Sahrens txg = itx->itx_lr.lrc_txg; 1119fa9e4066Sahrens ASSERT(txg); 1120fa9e4066Sahrens 1121fa9e4066Sahrens if (txg > spa_last_synced_txg(spa) || 1122fa9e4066Sahrens txg > spa_freeze_txg(spa)) 1123fa9e4066Sahrens lwb = zil_lwb_commit(zilog, itx, lwb); 1124fa9e4066Sahrens kmem_free(itx, offsetof(itx_t, itx_lr) 1125fa9e4066Sahrens + itx->itx_lr.lrc_reclen); 1126fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1127fa9e4066Sahrens } 1128b19a79ecSperrin DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 112967bd71c6Sperrin /* determine commit sequence number */ 113067bd71c6Sperrin itx = list_head(&zilog->zl_itx_list); 113167bd71c6Sperrin if (itx) 113267bd71c6Sperrin commit_seq = itx->itx_lr.lrc_seq; 113367bd71c6Sperrin else 113467bd71c6Sperrin commit_seq = zilog->zl_itx_seq; 1135fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1136fa9e4066Sahrens 1137fa9e4066Sahrens /* write the last block out */ 113867bd71c6Sperrin if (lwb != NULL && lwb->lwb_zio != NULL) 1139fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 1140fa9e4066Sahrens 114122ac5be4Sperrin zilog->zl_prev_used = zilog->zl_cur_used; 114222ac5be4Sperrin zilog->zl_cur_used = 0; 1143fa9e4066Sahrens 1144fa9e4066Sahrens /* 1145b19a79ecSperrin * Wait if necessary for the log blocks to be on stable storage. 1146fa9e4066Sahrens */ 1147b19a79ecSperrin if (zilog->zl_root_zio) { 1148b19a79ecSperrin DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1149b19a79ecSperrin (void) zio_wait(zilog->zl_root_zio); 1150b19a79ecSperrin DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 115117f17c2dSbonwick zil_flush_vdevs(zilog); 1152fa9e4066Sahrens } 115322ac5be4Sperrin 1154fa9e4066Sahrens if (zilog->zl_log_error || lwb == NULL) { 1155fa9e4066Sahrens zilog->zl_log_error = 0; 1156fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1157fa9e4066Sahrens } 115867bd71c6Sperrin 115967bd71c6Sperrin mutex_enter(&zilog->zl_lock); 116022ac5be4Sperrin zilog->zl_writer = B_FALSE; 116167bd71c6Sperrin 116267bd71c6Sperrin ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 116367bd71c6Sperrin zilog->zl_commit_seq = commit_seq; 1164b19a79ecSperrin } 1165b19a79ecSperrin 1166b19a79ecSperrin /* 1167b19a79ecSperrin * Push zfs transactions to stable storage up to the supplied sequence number. 1168b19a79ecSperrin * If foid is 0 push out all transactions, otherwise push only those 1169b19a79ecSperrin * for that file or might have been used to create that file. 1170b19a79ecSperrin */ 1171b19a79ecSperrin void 1172b19a79ecSperrin zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1173b19a79ecSperrin { 1174b19a79ecSperrin if (zilog == NULL || seq == 0) 1175b19a79ecSperrin return; 1176b19a79ecSperrin 1177b19a79ecSperrin mutex_enter(&zilog->zl_lock); 1178b19a79ecSperrin 1179b19a79ecSperrin seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1180b19a79ecSperrin 118167bd71c6Sperrin while (zilog->zl_writer) { 1182b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 118367bd71c6Sperrin if (seq < zilog->zl_commit_seq) { 118467bd71c6Sperrin mutex_exit(&zilog->zl_lock); 118567bd71c6Sperrin return; 118667bd71c6Sperrin } 118767bd71c6Sperrin } 1188b19a79ecSperrin zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 118967bd71c6Sperrin /* wake up others waiting on the commit */ 119067bd71c6Sperrin cv_broadcast(&zilog->zl_cv_writer); 119167bd71c6Sperrin mutex_exit(&zilog->zl_lock); 1192fa9e4066Sahrens } 1193fa9e4066Sahrens 1194fa9e4066Sahrens /* 1195fa9e4066Sahrens * Called in syncing context to free committed log blocks and update log header. 1196fa9e4066Sahrens */ 1197fa9e4066Sahrens void 1198fa9e4066Sahrens zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1199fa9e4066Sahrens { 1200d80c45e0Sbonwick zil_header_t *zh = zil_header_in_syncing_context(zilog); 1201fa9e4066Sahrens uint64_t txg = dmu_tx_get_txg(tx); 1202fa9e4066Sahrens spa_t *spa = zilog->zl_spa; 1203fa9e4066Sahrens lwb_t *lwb; 1204fa9e4066Sahrens 1205d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1206d80c45e0Sbonwick 1207fa9e4066Sahrens ASSERT(zilog->zl_stop_sync == 0); 1208fa9e4066Sahrens 1209d80c45e0Sbonwick zh->zh_replay_seq = zilog->zl_replay_seq[txg & TXG_MASK]; 1210fa9e4066Sahrens 1211fa9e4066Sahrens if (zilog->zl_destroy_txg == txg) { 1212d80c45e0Sbonwick blkptr_t blk = zh->zh_log; 1213d80c45e0Sbonwick 1214d80c45e0Sbonwick ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1215d80c45e0Sbonwick ASSERT(spa_sync_pass(spa) == 1); 1216d80c45e0Sbonwick 1217d80c45e0Sbonwick bzero(zh, sizeof (zil_header_t)); 1218fa9e4066Sahrens bzero(zilog->zl_replay_seq, sizeof (zilog->zl_replay_seq)); 1219d80c45e0Sbonwick 1220d80c45e0Sbonwick if (zilog->zl_keep_first) { 1221d80c45e0Sbonwick /* 1222d80c45e0Sbonwick * If this block was part of log chain that couldn't 1223d80c45e0Sbonwick * be claimed because a device was missing during 1224d80c45e0Sbonwick * zil_claim(), but that device later returns, 1225d80c45e0Sbonwick * then this block could erroneously appear valid. 1226d80c45e0Sbonwick * To guard against this, assign a new GUID to the new 1227d80c45e0Sbonwick * log chain so it doesn't matter what blk points to. 1228d80c45e0Sbonwick */ 1229d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 1230d80c45e0Sbonwick zh->zh_log = blk; 1231d80c45e0Sbonwick } 1232fa9e4066Sahrens } 1233fa9e4066Sahrens 1234fa9e4066Sahrens for (;;) { 1235fa9e4066Sahrens lwb = list_head(&zilog->zl_lwb_list); 1236fa9e4066Sahrens if (lwb == NULL) { 1237fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1238fa9e4066Sahrens return; 1239fa9e4066Sahrens } 1240b19a79ecSperrin zh->zh_log = lwb->lwb_blk; 1241fa9e4066Sahrens if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1242fa9e4066Sahrens break; 1243fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1244fa9e4066Sahrens zio_free_blk(spa, &lwb->lwb_blk, txg); 1245fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1246d63d470bSgw 1247d63d470bSgw /* 1248d63d470bSgw * If we don't have anything left in the lwb list then 1249d63d470bSgw * we've had an allocation failure and we need to zero 1250d63d470bSgw * out the zil_header blkptr so that we don't end 1251d63d470bSgw * up freeing the same block twice. 1252d63d470bSgw */ 1253d63d470bSgw if (list_head(&zilog->zl_lwb_list) == NULL) 1254d63d470bSgw BP_ZERO(&zh->zh_log); 1255fa9e4066Sahrens } 1256fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1257fa9e4066Sahrens } 1258fa9e4066Sahrens 1259fa9e4066Sahrens void 1260fa9e4066Sahrens zil_init(void) 1261fa9e4066Sahrens { 1262fa9e4066Sahrens zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 12635ad82045Snd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1264fa9e4066Sahrens } 1265fa9e4066Sahrens 1266fa9e4066Sahrens void 1267fa9e4066Sahrens zil_fini(void) 1268fa9e4066Sahrens { 1269fa9e4066Sahrens kmem_cache_destroy(zil_lwb_cache); 1270fa9e4066Sahrens } 1271fa9e4066Sahrens 1272fa9e4066Sahrens zilog_t * 1273fa9e4066Sahrens zil_alloc(objset_t *os, zil_header_t *zh_phys) 1274fa9e4066Sahrens { 1275fa9e4066Sahrens zilog_t *zilog; 1276fa9e4066Sahrens 1277fa9e4066Sahrens zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1278fa9e4066Sahrens 1279fa9e4066Sahrens zilog->zl_header = zh_phys; 1280fa9e4066Sahrens zilog->zl_os = os; 1281fa9e4066Sahrens zilog->zl_spa = dmu_objset_spa(os); 1282fa9e4066Sahrens zilog->zl_dmu_pool = dmu_objset_pool(os); 1283d80c45e0Sbonwick zilog->zl_destroy_txg = TXG_INITIAL - 1; 1284fa9e4066Sahrens 12855ad82045Snd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 12865ad82045Snd 1287fa9e4066Sahrens list_create(&zilog->zl_itx_list, sizeof (itx_t), 1288fa9e4066Sahrens offsetof(itx_t, itx_node)); 1289fa9e4066Sahrens 1290fa9e4066Sahrens list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1291fa9e4066Sahrens offsetof(lwb_t, lwb_node)); 1292fa9e4066Sahrens 129317f17c2dSbonwick mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 129417f17c2dSbonwick 129517f17c2dSbonwick avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 129617f17c2dSbonwick sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1297fa9e4066Sahrens 1298b7b97454Sperrin cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1299b7b97454Sperrin cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1300b7b97454Sperrin 1301fa9e4066Sahrens return (zilog); 1302fa9e4066Sahrens } 1303fa9e4066Sahrens 1304fa9e4066Sahrens void 1305fa9e4066Sahrens zil_free(zilog_t *zilog) 1306fa9e4066Sahrens { 1307fa9e4066Sahrens lwb_t *lwb; 1308fa9e4066Sahrens 1309fa9e4066Sahrens zilog->zl_stop_sync = 1; 1310fa9e4066Sahrens 1311fa9e4066Sahrens while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1312fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1313fa9e4066Sahrens if (lwb->lwb_buf != NULL) 1314fa9e4066Sahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1315fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1316fa9e4066Sahrens } 1317fa9e4066Sahrens list_destroy(&zilog->zl_lwb_list); 1318fa9e4066Sahrens 131917f17c2dSbonwick avl_destroy(&zilog->zl_vdev_tree); 132017f17c2dSbonwick mutex_destroy(&zilog->zl_vdev_lock); 1321fa9e4066Sahrens 1322fa9e4066Sahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1323fa9e4066Sahrens list_destroy(&zilog->zl_itx_list); 13245ad82045Snd mutex_destroy(&zilog->zl_lock); 1325fa9e4066Sahrens 1326b7b97454Sperrin cv_destroy(&zilog->zl_cv_writer); 1327b7b97454Sperrin cv_destroy(&zilog->zl_cv_suspend); 1328b7b97454Sperrin 1329fa9e4066Sahrens kmem_free(zilog, sizeof (zilog_t)); 1330fa9e4066Sahrens } 1331fa9e4066Sahrens 133213f5297eSperrin /* 1333436b2950Sperrin * return true if the initial log block is not valid 133413f5297eSperrin */ 133513f5297eSperrin static int 133613f5297eSperrin zil_empty(zilog_t *zilog) 133713f5297eSperrin { 1338d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1339d80c45e0Sbonwick arc_buf_t *abuf = NULL; 134013f5297eSperrin 1341d80c45e0Sbonwick if (BP_IS_HOLE(&zh->zh_log)) 134213f5297eSperrin return (1); 134313f5297eSperrin 1344d80c45e0Sbonwick if (zil_read_log_block(zilog, &zh->zh_log, &abuf) != 0) 1345d80c45e0Sbonwick return (1); 1346d80c45e0Sbonwick 1347d80c45e0Sbonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 1348d80c45e0Sbonwick return (0); 134913f5297eSperrin } 135013f5297eSperrin 1351fa9e4066Sahrens /* 1352fa9e4066Sahrens * Open an intent log. 1353fa9e4066Sahrens */ 1354fa9e4066Sahrens zilog_t * 1355fa9e4066Sahrens zil_open(objset_t *os, zil_get_data_t *get_data) 1356fa9e4066Sahrens { 1357fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1358fa9e4066Sahrens 1359fa9e4066Sahrens zilog->zl_get_data = get_data; 1360fa9e4066Sahrens zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1361fa9e4066Sahrens 2, 2, TASKQ_PREPOPULATE); 1362fa9e4066Sahrens 1363fa9e4066Sahrens return (zilog); 1364fa9e4066Sahrens } 1365fa9e4066Sahrens 1366fa9e4066Sahrens /* 1367fa9e4066Sahrens * Close an intent log. 1368fa9e4066Sahrens */ 1369fa9e4066Sahrens void 1370fa9e4066Sahrens zil_close(zilog_t *zilog) 1371fa9e4066Sahrens { 1372d80c45e0Sbonwick /* 1373d80c45e0Sbonwick * If the log isn't already committed, mark the objset dirty 1374d80c45e0Sbonwick * (so zil_sync() will be called) and wait for that txg to sync. 1375d80c45e0Sbonwick */ 1376d80c45e0Sbonwick if (!zil_is_committed(zilog)) { 1377d80c45e0Sbonwick uint64_t txg; 1378d80c45e0Sbonwick dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1379d80c45e0Sbonwick (void) dmu_tx_assign(tx, TXG_WAIT); 1380d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1381d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 1382d80c45e0Sbonwick dmu_tx_commit(tx); 1383d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, txg); 1384d80c45e0Sbonwick } 1385d80c45e0Sbonwick 1386fa9e4066Sahrens taskq_destroy(zilog->zl_clean_taskq); 1387fa9e4066Sahrens zilog->zl_clean_taskq = NULL; 1388fa9e4066Sahrens zilog->zl_get_data = NULL; 1389fa9e4066Sahrens 1390fa9e4066Sahrens zil_itx_clean(zilog); 1391fa9e4066Sahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1392fa9e4066Sahrens } 1393fa9e4066Sahrens 1394fa9e4066Sahrens /* 1395fa9e4066Sahrens * Suspend an intent log. While in suspended mode, we still honor 1396fa9e4066Sahrens * synchronous semantics, but we rely on txg_wait_synced() to do it. 1397fa9e4066Sahrens * We suspend the log briefly when taking a snapshot so that the snapshot 1398fa9e4066Sahrens * contains all the data it's supposed to, and has an empty intent log. 1399fa9e4066Sahrens */ 1400fa9e4066Sahrens int 1401fa9e4066Sahrens zil_suspend(zilog_t *zilog) 1402fa9e4066Sahrens { 1403d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1404fa9e4066Sahrens 1405fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1406d80c45e0Sbonwick if (zh->zh_claim_txg != 0) { /* unplayed log */ 1407fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1408fa9e4066Sahrens return (EBUSY); 1409fa9e4066Sahrens } 1410d80c45e0Sbonwick if (zilog->zl_suspend++ != 0) { 1411d80c45e0Sbonwick /* 1412d80c45e0Sbonwick * Someone else already began a suspend. 1413d80c45e0Sbonwick * Just wait for them to finish. 1414d80c45e0Sbonwick */ 1415d80c45e0Sbonwick while (zilog->zl_suspending) 1416d80c45e0Sbonwick cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1417d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1418d80c45e0Sbonwick return (0); 1419d80c45e0Sbonwick } 1420d80c45e0Sbonwick zilog->zl_suspending = B_TRUE; 1421fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1422fa9e4066Sahrens 1423b19a79ecSperrin zil_commit(zilog, UINT64_MAX, 0); 1424fa9e4066Sahrens 1425b19a79ecSperrin /* 1426b19a79ecSperrin * Wait for any in-flight log writes to complete. 1427b19a79ecSperrin */ 1428fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1429b19a79ecSperrin while (zilog->zl_writer) 1430b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1431fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1432fa9e4066Sahrens 1433d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1434d80c45e0Sbonwick 1435d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1436d80c45e0Sbonwick zilog->zl_suspending = B_FALSE; 1437d80c45e0Sbonwick cv_broadcast(&zilog->zl_cv_suspend); 1438d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1439fa9e4066Sahrens 1440fa9e4066Sahrens return (0); 1441fa9e4066Sahrens } 1442fa9e4066Sahrens 1443fa9e4066Sahrens void 1444fa9e4066Sahrens zil_resume(zilog_t *zilog) 1445fa9e4066Sahrens { 1446fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1447fa9e4066Sahrens ASSERT(zilog->zl_suspend != 0); 1448fa9e4066Sahrens zilog->zl_suspend--; 1449fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1450fa9e4066Sahrens } 1451fa9e4066Sahrens 1452fa9e4066Sahrens typedef struct zil_replay_arg { 1453fa9e4066Sahrens objset_t *zr_os; 1454fa9e4066Sahrens zil_replay_func_t **zr_replay; 1455fa9e4066Sahrens void *zr_arg; 1456fa9e4066Sahrens uint64_t *zr_txgp; 1457fa9e4066Sahrens boolean_t zr_byteswap; 1458fa9e4066Sahrens char *zr_lrbuf; 1459fa9e4066Sahrens } zil_replay_arg_t; 1460fa9e4066Sahrens 1461fa9e4066Sahrens static void 1462fa9e4066Sahrens zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1463fa9e4066Sahrens { 1464fa9e4066Sahrens zil_replay_arg_t *zr = zra; 1465d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1466fa9e4066Sahrens uint64_t reclen = lr->lrc_reclen; 1467fa9e4066Sahrens uint64_t txtype = lr->lrc_txtype; 146867bd71c6Sperrin char *name; 146967bd71c6Sperrin int pass, error, sunk; 1470fa9e4066Sahrens 1471fa9e4066Sahrens if (zilog->zl_stop_replay) 1472fa9e4066Sahrens return; 1473fa9e4066Sahrens 1474fa9e4066Sahrens if (lr->lrc_txg < claim_txg) /* already committed */ 1475fa9e4066Sahrens return; 1476fa9e4066Sahrens 1477fa9e4066Sahrens if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1478fa9e4066Sahrens return; 1479fa9e4066Sahrens 1480da6c28aaSamw /* Strip case-insensitive bit, still present in log record */ 1481da6c28aaSamw txtype &= ~TX_CI; 1482da6c28aaSamw 1483fa9e4066Sahrens /* 1484fa9e4066Sahrens * Make a copy of the data so we can revise and extend it. 1485fa9e4066Sahrens */ 1486fa9e4066Sahrens bcopy(lr, zr->zr_lrbuf, reclen); 1487fa9e4066Sahrens 1488fa9e4066Sahrens /* 1489fa9e4066Sahrens * The log block containing this lr may have been byteswapped 1490fa9e4066Sahrens * so that we can easily examine common fields like lrc_txtype. 1491fa9e4066Sahrens * However, the log is a mix of different data types, and only the 1492fa9e4066Sahrens * replay vectors know how to byteswap their records. Therefore, if 1493fa9e4066Sahrens * the lr was byteswapped, undo it before invoking the replay vector. 1494fa9e4066Sahrens */ 1495fa9e4066Sahrens if (zr->zr_byteswap) 1496fa9e4066Sahrens byteswap_uint64_array(zr->zr_lrbuf, reclen); 1497fa9e4066Sahrens 1498fa9e4066Sahrens /* 1499fa9e4066Sahrens * If this is a TX_WRITE with a blkptr, suck in the data. 1500fa9e4066Sahrens */ 1501fa9e4066Sahrens if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1502fa9e4066Sahrens lr_write_t *lrw = (lr_write_t *)lr; 1503fa9e4066Sahrens blkptr_t *wbp = &lrw->lr_blkptr; 1504fa9e4066Sahrens uint64_t wlen = lrw->lr_length; 1505fa9e4066Sahrens char *wbuf = zr->zr_lrbuf + reclen; 1506fa9e4066Sahrens 1507fa9e4066Sahrens if (BP_IS_HOLE(wbp)) { /* compressed to a hole */ 1508fa9e4066Sahrens bzero(wbuf, wlen); 1509fa9e4066Sahrens } else { 1510fa9e4066Sahrens /* 1511fa9e4066Sahrens * A subsequent write may have overwritten this block, 1512fa9e4066Sahrens * in which case wbp may have been been freed and 1513fa9e4066Sahrens * reallocated, and our read of wbp may fail with a 1514fa9e4066Sahrens * checksum error. We can safely ignore this because 1515fa9e4066Sahrens * the later write will provide the correct data. 1516fa9e4066Sahrens */ 1517ea8dc4b6Seschrock zbookmark_t zb; 1518ea8dc4b6Seschrock 1519ea8dc4b6Seschrock zb.zb_objset = dmu_objset_id(zilog->zl_os); 1520ea8dc4b6Seschrock zb.zb_object = lrw->lr_foid; 1521ea8dc4b6Seschrock zb.zb_level = -1; 1522ea8dc4b6Seschrock zb.zb_blkid = lrw->lr_offset / BP_GET_LSIZE(wbp); 1523ea8dc4b6Seschrock 1524fa9e4066Sahrens (void) zio_wait(zio_read(NULL, zilog->zl_spa, 1525fa9e4066Sahrens wbp, wbuf, BP_GET_LSIZE(wbp), NULL, NULL, 1526fa9e4066Sahrens ZIO_PRIORITY_SYNC_READ, 1527ea8dc4b6Seschrock ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, &zb)); 1528fa9e4066Sahrens (void) memmove(wbuf, wbuf + lrw->lr_blkoff, wlen); 1529fa9e4066Sahrens } 1530fa9e4066Sahrens } 1531fa9e4066Sahrens 1532fa9e4066Sahrens /* 1533fa9e4066Sahrens * We must now do two things atomically: replay this log record, 1534fa9e4066Sahrens * and update the log header to reflect the fact that we did so. 1535fa9e4066Sahrens * We use the DMU's ability to assign into a specific txg to do this. 1536fa9e4066Sahrens */ 153767bd71c6Sperrin for (pass = 1, sunk = B_FALSE; /* CONSTANTCONDITION */; pass++) { 1538fa9e4066Sahrens uint64_t replay_txg; 1539fa9e4066Sahrens dmu_tx_t *replay_tx; 1540fa9e4066Sahrens 1541fa9e4066Sahrens replay_tx = dmu_tx_create(zr->zr_os); 1542fa9e4066Sahrens error = dmu_tx_assign(replay_tx, TXG_WAIT); 1543fa9e4066Sahrens if (error) { 1544fa9e4066Sahrens dmu_tx_abort(replay_tx); 1545fa9e4066Sahrens break; 1546fa9e4066Sahrens } 1547fa9e4066Sahrens 1548fa9e4066Sahrens replay_txg = dmu_tx_get_txg(replay_tx); 1549fa9e4066Sahrens 1550fa9e4066Sahrens if (txtype == 0 || txtype >= TX_MAX_TYPE) { 1551fa9e4066Sahrens error = EINVAL; 1552fa9e4066Sahrens } else { 1553fa9e4066Sahrens /* 1554fa9e4066Sahrens * On the first pass, arrange for the replay vector 1555fa9e4066Sahrens * to fail its dmu_tx_assign(). That's the only way 1556fa9e4066Sahrens * to ensure that those code paths remain well tested. 1557cf43fb6bSperrin * 1558cf43fb6bSperrin * Only byteswap (if needed) on the 1st pass. 1559fa9e4066Sahrens */ 1560fa9e4066Sahrens *zr->zr_txgp = replay_txg - (pass == 1); 1561fa9e4066Sahrens error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf, 1562cf43fb6bSperrin zr->zr_byteswap && pass == 1); 1563fa9e4066Sahrens *zr->zr_txgp = TXG_NOWAIT; 1564fa9e4066Sahrens } 1565fa9e4066Sahrens 1566fa9e4066Sahrens if (error == 0) { 1567fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(zr->zr_os), replay_tx); 1568fa9e4066Sahrens zilog->zl_replay_seq[replay_txg & TXG_MASK] = 1569fa9e4066Sahrens lr->lrc_seq; 1570fa9e4066Sahrens } 1571fa9e4066Sahrens 1572fa9e4066Sahrens dmu_tx_commit(replay_tx); 1573fa9e4066Sahrens 157467bd71c6Sperrin if (!error) 157567bd71c6Sperrin return; 157667bd71c6Sperrin 157767bd71c6Sperrin /* 157867bd71c6Sperrin * The DMU's dnode layer doesn't see removes until the txg 157967bd71c6Sperrin * commits, so a subsequent claim can spuriously fail with 158067bd71c6Sperrin * EEXIST. So if we receive any error other than ERESTART 158167bd71c6Sperrin * we try syncing out any removes then retrying the 158267bd71c6Sperrin * transaction. 158367bd71c6Sperrin */ 158467bd71c6Sperrin if (error != ERESTART && !sunk) { 158567bd71c6Sperrin txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 158667bd71c6Sperrin sunk = B_TRUE; 158767bd71c6Sperrin continue; /* retry */ 158867bd71c6Sperrin } 158967bd71c6Sperrin 1590fa9e4066Sahrens if (error != ERESTART) 1591fa9e4066Sahrens break; 1592fa9e4066Sahrens 1593fa9e4066Sahrens if (pass != 1) 1594fa9e4066Sahrens txg_wait_open(spa_get_dsl(zilog->zl_spa), 1595fa9e4066Sahrens replay_txg + 1); 1596fa9e4066Sahrens 1597fa9e4066Sahrens dprintf("pass %d, retrying\n", pass); 1598fa9e4066Sahrens } 1599fa9e4066Sahrens 160067bd71c6Sperrin ASSERT(error && error != ERESTART); 160167bd71c6Sperrin name = kmem_alloc(MAXNAMELEN, KM_SLEEP); 160267bd71c6Sperrin dmu_objset_name(zr->zr_os, name); 160367bd71c6Sperrin cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1604da6c28aaSamw "dataset %s, seq 0x%llx, txtype %llu %s\n", 1605da6c28aaSamw error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype, 1606da6c28aaSamw (lr->lrc_txtype & TX_CI) ? "CI" : ""); 160767bd71c6Sperrin zilog->zl_stop_replay = 1; 160867bd71c6Sperrin kmem_free(name, MAXNAMELEN); 160967bd71c6Sperrin } 1610fa9e4066Sahrens 161167bd71c6Sperrin /* ARGSUSED */ 161267bd71c6Sperrin static void 161367bd71c6Sperrin zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 161467bd71c6Sperrin { 161567bd71c6Sperrin zilog->zl_replay_blks++; 1616fa9e4066Sahrens } 1617fa9e4066Sahrens 1618fa9e4066Sahrens /* 161913f5297eSperrin * If this dataset has a non-empty intent log, replay it and destroy it. 1620fa9e4066Sahrens */ 1621fa9e4066Sahrens void 1622fa9e4066Sahrens zil_replay(objset_t *os, void *arg, uint64_t *txgp, 1623893a6d32Sahrens zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1624fa9e4066Sahrens { 1625fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1626d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1627d80c45e0Sbonwick zil_replay_arg_t zr; 162813f5297eSperrin 162913f5297eSperrin if (zil_empty(zilog)) { 1630d80c45e0Sbonwick zil_destroy(zilog, B_TRUE); 163113f5297eSperrin return; 163213f5297eSperrin } 1633fa9e4066Sahrens 1634fa9e4066Sahrens zr.zr_os = os; 1635fa9e4066Sahrens zr.zr_replay = replay_func; 1636fa9e4066Sahrens zr.zr_arg = arg; 1637fa9e4066Sahrens zr.zr_txgp = txgp; 1638d80c45e0Sbonwick zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1639fa9e4066Sahrens zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1640fa9e4066Sahrens 1641fa9e4066Sahrens /* 1642fa9e4066Sahrens * Wait for in-progress removes to sync before starting replay. 1643fa9e4066Sahrens */ 1644fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1645fa9e4066Sahrens 1646fa9e4066Sahrens zilog->zl_stop_replay = 0; 164767bd71c6Sperrin zilog->zl_replay_time = lbolt; 164867bd71c6Sperrin ASSERT(zilog->zl_replay_blks == 0); 164967bd71c6Sperrin (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1650d80c45e0Sbonwick zh->zh_claim_txg); 1651fa9e4066Sahrens kmem_free(zr.zr_lrbuf, 2 * SPA_MAXBLOCKSIZE); 1652fa9e4066Sahrens 1653d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1654a4611edeSahrens txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 1655fa9e4066Sahrens } 1656436b2950Sperrin 1657436b2950Sperrin /* 1658436b2950Sperrin * Report whether all transactions are committed 1659436b2950Sperrin */ 1660436b2950Sperrin int 1661436b2950Sperrin zil_is_committed(zilog_t *zilog) 1662436b2950Sperrin { 1663436b2950Sperrin lwb_t *lwb; 1664b19a79ecSperrin int ret; 1665b19a79ecSperrin 1666b19a79ecSperrin mutex_enter(&zilog->zl_lock); 1667b19a79ecSperrin while (zilog->zl_writer) 1668b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1669b19a79ecSperrin 1670b19a79ecSperrin /* recent unpushed intent log transactions? */ 1671b19a79ecSperrin if (!list_is_empty(&zilog->zl_itx_list)) { 1672b19a79ecSperrin ret = B_FALSE; 1673b19a79ecSperrin goto out; 1674b19a79ecSperrin } 1675436b2950Sperrin 1676b19a79ecSperrin /* intent log never used? */ 1677b19a79ecSperrin lwb = list_head(&zilog->zl_lwb_list); 1678b19a79ecSperrin if (lwb == NULL) { 1679b19a79ecSperrin ret = B_TRUE; 1680b19a79ecSperrin goto out; 1681b19a79ecSperrin } 1682436b2950Sperrin 1683436b2950Sperrin /* 1684b19a79ecSperrin * more than 1 log buffer means zil_sync() hasn't yet freed 1685b19a79ecSperrin * entries after a txg has committed 1686436b2950Sperrin */ 1687b19a79ecSperrin if (list_next(&zilog->zl_lwb_list, lwb)) { 1688b19a79ecSperrin ret = B_FALSE; 1689b19a79ecSperrin goto out; 1690b19a79ecSperrin } 1691b19a79ecSperrin 1692436b2950Sperrin ASSERT(zil_empty(zilog)); 1693b19a79ecSperrin ret = B_TRUE; 1694b19a79ecSperrin out: 1695b19a79ecSperrin cv_broadcast(&zilog->zl_cv_writer); 1696b19a79ecSperrin mutex_exit(&zilog->zl_lock); 1697b19a79ecSperrin return (ret); 1698436b2950Sperrin } 1699