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> 28fa9e4066Sahrens #include <sys/dmu.h> 29fa9e4066Sahrens #include <sys/zap.h> 30fa9e4066Sahrens #include <sys/arc.h> 31fa9e4066Sahrens #include <sys/stat.h> 32fa9e4066Sahrens #include <sys/resource.h> 33fa9e4066Sahrens #include <sys/zil.h> 34fa9e4066Sahrens #include <sys/zil_impl.h> 35fa9e4066Sahrens #include <sys/dsl_dataset.h> 36fa9e4066Sahrens #include <sys/vdev.h> 37d63d470bSgw #include <sys/dmu_tx.h> 38fa9e4066Sahrens 39fa9e4066Sahrens /* 40fa9e4066Sahrens * The zfs intent log (ZIL) saves transaction records of system calls 41fa9e4066Sahrens * that change the file system in memory with enough information 42fa9e4066Sahrens * to be able to replay them. These are stored in memory until 43fa9e4066Sahrens * either the DMU transaction group (txg) commits them to the stable pool 44fa9e4066Sahrens * and they can be discarded, or they are flushed to the stable log 45fa9e4066Sahrens * (also in the pool) due to a fsync, O_DSYNC or other synchronous 46fa9e4066Sahrens * requirement. In the event of a panic or power fail then those log 47fa9e4066Sahrens * records (transactions) are replayed. 48fa9e4066Sahrens * 49fa9e4066Sahrens * There is one ZIL per file system. Its on-disk (pool) format consists 50fa9e4066Sahrens * of 3 parts: 51fa9e4066Sahrens * 52fa9e4066Sahrens * - ZIL header 53fa9e4066Sahrens * - ZIL blocks 54fa9e4066Sahrens * - ZIL records 55fa9e4066Sahrens * 56fa9e4066Sahrens * A log record holds a system call transaction. Log blocks can 57fa9e4066Sahrens * hold many log records and the blocks are chained together. 58fa9e4066Sahrens * Each ZIL block contains a block pointer (blkptr_t) to the next 59fa9e4066Sahrens * ZIL block in the chain. The ZIL header points to the first 60fa9e4066Sahrens * block in the chain. Note there is not a fixed place in the pool 61fa9e4066Sahrens * to hold blocks. They are dynamically allocated and freed as 62fa9e4066Sahrens * needed from the blocks available. Figure X shows the ZIL structure: 63fa9e4066Sahrens */ 64fa9e4066Sahrens 65fa9e4066Sahrens /* 66416e0cd8Sek * This global ZIL switch affects all pools 67fa9e4066Sahrens */ 68fa9e4066Sahrens int zil_disable = 0; /* disable intent logging */ 69416e0cd8Sek 70416e0cd8Sek /* 71416e0cd8Sek * Tunable parameter for debugging or performance analysis. Setting 72416e0cd8Sek * zfs_nocacheflush will cause corruption on power loss if a volatile 73416e0cd8Sek * out-of-order write cache is enabled. 74416e0cd8Sek */ 75416e0cd8Sek boolean_t zfs_nocacheflush = B_FALSE; 76fa9e4066Sahrens 77fa9e4066Sahrens static kmem_cache_t *zil_lwb_cache; 78fa9e4066Sahrens 798f18d1faSGeorge Wilson static boolean_t zil_empty(zilog_t *zilog); 808f18d1faSGeorge Wilson 81fa9e4066Sahrens static int 82b24ab676SJeff Bonwick zil_bp_compare(const void *x1, const void *x2) 83fa9e4066Sahrens { 84b24ab676SJeff Bonwick const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 85b24ab676SJeff Bonwick const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 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 101b24ab676SJeff Bonwick zil_bp_tree_init(zilog_t *zilog) 102fa9e4066Sahrens { 103b24ab676SJeff Bonwick avl_create(&zilog->zl_bp_tree, zil_bp_compare, 104b24ab676SJeff Bonwick sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 105fa9e4066Sahrens } 106fa9e4066Sahrens 107fa9e4066Sahrens static void 108b24ab676SJeff Bonwick zil_bp_tree_fini(zilog_t *zilog) 109fa9e4066Sahrens { 110b24ab676SJeff Bonwick avl_tree_t *t = &zilog->zl_bp_tree; 111b24ab676SJeff Bonwick zil_bp_node_t *zn; 112fa9e4066Sahrens void *cookie = NULL; 113fa9e4066Sahrens 114fa9e4066Sahrens while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 115b24ab676SJeff Bonwick kmem_free(zn, sizeof (zil_bp_node_t)); 116fa9e4066Sahrens 117fa9e4066Sahrens avl_destroy(t); 118fa9e4066Sahrens } 119fa9e4066Sahrens 120b24ab676SJeff Bonwick int 121b24ab676SJeff Bonwick zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 122fa9e4066Sahrens { 123b24ab676SJeff Bonwick avl_tree_t *t = &zilog->zl_bp_tree; 124b24ab676SJeff Bonwick const dva_t *dva = BP_IDENTITY(bp); 125b24ab676SJeff Bonwick zil_bp_node_t *zn; 126fa9e4066Sahrens avl_index_t where; 127fa9e4066Sahrens 128fa9e4066Sahrens if (avl_find(t, dva, &where) != NULL) 129fa9e4066Sahrens return (EEXIST); 130fa9e4066Sahrens 131b24ab676SJeff Bonwick zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 132fa9e4066Sahrens zn->zn_dva = *dva; 133fa9e4066Sahrens avl_insert(t, zn, where); 134fa9e4066Sahrens 135fa9e4066Sahrens return (0); 136fa9e4066Sahrens } 137fa9e4066Sahrens 138d80c45e0Sbonwick static zil_header_t * 139d80c45e0Sbonwick zil_header_in_syncing_context(zilog_t *zilog) 140d80c45e0Sbonwick { 141d80c45e0Sbonwick return ((zil_header_t *)zilog->zl_header); 142d80c45e0Sbonwick } 143d80c45e0Sbonwick 144d80c45e0Sbonwick static void 145d80c45e0Sbonwick zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 146d80c45e0Sbonwick { 147d80c45e0Sbonwick zio_cksum_t *zc = &bp->blk_cksum; 148d80c45e0Sbonwick 149d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 150d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 151d80c45e0Sbonwick zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 152d80c45e0Sbonwick zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 153d80c45e0Sbonwick } 154d80c45e0Sbonwick 155fa9e4066Sahrens /* 156b24ab676SJeff Bonwick * Read a log block and make sure it's valid. 157fa9e4066Sahrens */ 158fa9e4066Sahrens static int 159b24ab676SJeff Bonwick zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst) 160fa9e4066Sahrens { 161b24ab676SJeff Bonwick enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 16213506d1eSmaybee uint32_t aflags = ARC_WAIT; 163b24ab676SJeff Bonwick arc_buf_t *abuf = NULL; 164b24ab676SJeff Bonwick zbookmark_t zb; 165fa9e4066Sahrens int error; 166fa9e4066Sahrens 167b24ab676SJeff Bonwick if (zilog->zl_header->zh_claim_txg == 0) 168b24ab676SJeff Bonwick zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 169ea8dc4b6Seschrock 170b24ab676SJeff Bonwick if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 171b24ab676SJeff Bonwick zio_flags |= ZIO_FLAG_SPECULATIVE; 172fa9e4066Sahrens 173b24ab676SJeff Bonwick SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 174b24ab676SJeff Bonwick ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 175b24ab676SJeff Bonwick 176b24ab676SJeff Bonwick error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 177b24ab676SJeff Bonwick ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 178fa9e4066Sahrens 179d80c45e0Sbonwick if (error == 0) { 180b24ab676SJeff Bonwick char *data = abuf->b_data; 181b24ab676SJeff Bonwick uint64_t size = BP_GET_LSIZE(bp); 182b24ab676SJeff Bonwick zil_trailer_t *ztp = (zil_trailer_t *)(data + size) - 1; 183d80c45e0Sbonwick zio_cksum_t cksum = bp->blk_cksum; 184fa9e4066Sahrens 185b24ab676SJeff Bonwick bcopy(data, dst, size); 186b24ab676SJeff Bonwick *nbp = ztp->zit_next_blk; 187b24ab676SJeff Bonwick 188d80c45e0Sbonwick /* 189f5e6e722SNeil Perrin * Validate the checksummed log block. 190f5e6e722SNeil Perrin * 191d80c45e0Sbonwick * Sequence numbers should be... sequential. The checksum 192d80c45e0Sbonwick * verifier for the next block should be bp's checksum plus 1. 193f5e6e722SNeil Perrin * 194f5e6e722SNeil Perrin * Also check the log chain linkage and size used. 195d80c45e0Sbonwick */ 196d80c45e0Sbonwick cksum.zc_word[ZIL_ZC_SEQ]++; 197d80c45e0Sbonwick 198f5e6e722SNeil Perrin if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, 199f5e6e722SNeil Perrin sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) || 200b24ab676SJeff Bonwick (ztp->zit_nused > (size - sizeof (zil_trailer_t)))) 201f5e6e722SNeil Perrin error = ECKSUM; 202fa9e4066Sahrens 203b24ab676SJeff Bonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 204fa9e4066Sahrens } 205fa9e4066Sahrens 206b24ab676SJeff Bonwick return (error); 207b24ab676SJeff Bonwick } 208b24ab676SJeff Bonwick 209b24ab676SJeff Bonwick /* 210b24ab676SJeff Bonwick * Read a TX_WRITE log data block. 211b24ab676SJeff Bonwick */ 212b24ab676SJeff Bonwick static int 213b24ab676SJeff Bonwick zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 214b24ab676SJeff Bonwick { 215b24ab676SJeff Bonwick enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 216b24ab676SJeff Bonwick const blkptr_t *bp = &lr->lr_blkptr; 217b24ab676SJeff Bonwick uint32_t aflags = ARC_WAIT; 218b24ab676SJeff Bonwick arc_buf_t *abuf = NULL; 219b24ab676SJeff Bonwick zbookmark_t zb; 220b24ab676SJeff Bonwick int error; 221b24ab676SJeff Bonwick 222b24ab676SJeff Bonwick if (BP_IS_HOLE(bp)) { 223b24ab676SJeff Bonwick if (wbuf != NULL) 224b24ab676SJeff Bonwick bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 225b24ab676SJeff Bonwick return (0); 226b24ab676SJeff Bonwick } 227b24ab676SJeff Bonwick 228b24ab676SJeff Bonwick if (zilog->zl_header->zh_claim_txg == 0) 229b24ab676SJeff Bonwick zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 230b24ab676SJeff Bonwick 231b24ab676SJeff Bonwick SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 232b24ab676SJeff Bonwick ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 233b24ab676SJeff Bonwick 234b24ab676SJeff Bonwick error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 235b24ab676SJeff Bonwick ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 236b24ab676SJeff Bonwick 237b24ab676SJeff Bonwick if (error == 0) { 238b24ab676SJeff Bonwick if (wbuf != NULL) 239b24ab676SJeff Bonwick bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 240b24ab676SJeff Bonwick (void) arc_buf_remove_ref(abuf, &abuf); 241b24ab676SJeff Bonwick } 242fa9e4066Sahrens 243d80c45e0Sbonwick return (error); 244fa9e4066Sahrens } 245fa9e4066Sahrens 246fa9e4066Sahrens /* 247fa9e4066Sahrens * Parse the intent log, and call parse_func for each valid record within. 248fa9e4066Sahrens */ 249b24ab676SJeff Bonwick int 250fa9e4066Sahrens zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 251fa9e4066Sahrens zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 252fa9e4066Sahrens { 253d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 254b24ab676SJeff Bonwick boolean_t claimed = !!zh->zh_claim_txg; 255b24ab676SJeff Bonwick uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 256b24ab676SJeff Bonwick uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 257b24ab676SJeff Bonwick uint64_t max_blk_seq = 0; 258b24ab676SJeff Bonwick uint64_t max_lr_seq = 0; 259b24ab676SJeff Bonwick uint64_t blk_count = 0; 260b24ab676SJeff Bonwick uint64_t lr_count = 0; 261b24ab676SJeff Bonwick blkptr_t blk, next_blk; 262fa9e4066Sahrens char *lrbuf, *lrp; 263b24ab676SJeff Bonwick int error = 0; 264fa9e4066Sahrens 265b24ab676SJeff Bonwick /* 266b24ab676SJeff Bonwick * Old logs didn't record the maximum zh_claim_lr_seq. 267b24ab676SJeff Bonwick */ 268b24ab676SJeff Bonwick if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 269b24ab676SJeff Bonwick claim_lr_seq = UINT64_MAX; 270fa9e4066Sahrens 271fa9e4066Sahrens /* 272fa9e4066Sahrens * Starting at the block pointed to by zh_log we read the log chain. 273fa9e4066Sahrens * For each block in the chain we strongly check that block to 274fa9e4066Sahrens * ensure its validity. We stop when an invalid block is found. 275fa9e4066Sahrens * For each block pointer in the chain we call parse_blk_func(). 276fa9e4066Sahrens * For each record in each valid block we call parse_lr_func(). 277d80c45e0Sbonwick * If the log has been claimed, stop if we encounter a sequence 278d80c45e0Sbonwick * number greater than the highest claimed sequence number. 279fa9e4066Sahrens */ 280b24ab676SJeff Bonwick lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 281b24ab676SJeff Bonwick zil_bp_tree_init(zilog); 282d80c45e0Sbonwick 283b24ab676SJeff Bonwick for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 284b24ab676SJeff Bonwick zil_trailer_t *ztp = 285b24ab676SJeff Bonwick (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 286b24ab676SJeff Bonwick uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 287b24ab676SJeff Bonwick int reclen; 288d80c45e0Sbonwick 289b24ab676SJeff Bonwick if (blk_seq > claim_blk_seq) 290b24ab676SJeff Bonwick break; 291b24ab676SJeff Bonwick if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 292b24ab676SJeff Bonwick break; 293b24ab676SJeff Bonwick ASSERT(max_blk_seq < blk_seq); 294b24ab676SJeff Bonwick max_blk_seq = blk_seq; 295b24ab676SJeff Bonwick blk_count++; 296fa9e4066Sahrens 297b24ab676SJeff Bonwick if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 298b24ab676SJeff Bonwick break; 299fa9e4066Sahrens 300b24ab676SJeff Bonwick error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf); 301fa9e4066Sahrens if (error) 302fa9e4066Sahrens break; 303fa9e4066Sahrens 304fa9e4066Sahrens for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { 305fa9e4066Sahrens lr_t *lr = (lr_t *)lrp; 306fa9e4066Sahrens reclen = lr->lrc_reclen; 307fa9e4066Sahrens ASSERT3U(reclen, >=, sizeof (lr_t)); 308b24ab676SJeff Bonwick if (lr->lrc_seq > claim_lr_seq) 309b24ab676SJeff Bonwick goto done; 310b24ab676SJeff Bonwick if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 311b24ab676SJeff Bonwick goto done; 312b24ab676SJeff Bonwick ASSERT(max_lr_seq < lr->lrc_seq); 313b24ab676SJeff Bonwick max_lr_seq = lr->lrc_seq; 314b24ab676SJeff Bonwick lr_count++; 315fa9e4066Sahrens } 316fa9e4066Sahrens } 317b24ab676SJeff Bonwick done: 318b24ab676SJeff Bonwick zilog->zl_parse_error = error; 319b24ab676SJeff Bonwick zilog->zl_parse_blk_seq = max_blk_seq; 320b24ab676SJeff Bonwick zilog->zl_parse_lr_seq = max_lr_seq; 321b24ab676SJeff Bonwick zilog->zl_parse_blk_count = blk_count; 322b24ab676SJeff Bonwick zilog->zl_parse_lr_count = lr_count; 323b24ab676SJeff Bonwick 324b24ab676SJeff Bonwick ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 325b24ab676SJeff Bonwick (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 326d80c45e0Sbonwick 327b24ab676SJeff Bonwick zil_bp_tree_fini(zilog); 328b24ab676SJeff Bonwick zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 329b24ab676SJeff Bonwick 330b24ab676SJeff Bonwick return (error); 331fa9e4066Sahrens } 332fa9e4066Sahrens 333b24ab676SJeff Bonwick static int 334fa9e4066Sahrens zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 335fa9e4066Sahrens { 336fa9e4066Sahrens /* 337fa9e4066Sahrens * Claim log block if not already committed and not already claimed. 338b24ab676SJeff Bonwick * If tx == NULL, just verify that the block is claimable. 339fa9e4066Sahrens */ 340b24ab676SJeff Bonwick if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 341b24ab676SJeff Bonwick return (0); 342b24ab676SJeff Bonwick 343b24ab676SJeff Bonwick return (zio_wait(zio_claim(NULL, zilog->zl_spa, 344b24ab676SJeff Bonwick tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 345b24ab676SJeff Bonwick ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 346fa9e4066Sahrens } 347fa9e4066Sahrens 348b24ab676SJeff Bonwick static int 349fa9e4066Sahrens zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 350fa9e4066Sahrens { 351b24ab676SJeff Bonwick lr_write_t *lr = (lr_write_t *)lrc; 352b24ab676SJeff Bonwick int error; 353b24ab676SJeff Bonwick 354b24ab676SJeff Bonwick if (lrc->lrc_txtype != TX_WRITE) 355b24ab676SJeff Bonwick return (0); 356b24ab676SJeff Bonwick 357b24ab676SJeff Bonwick /* 358b24ab676SJeff Bonwick * If the block is not readable, don't claim it. This can happen 359b24ab676SJeff Bonwick * in normal operation when a log block is written to disk before 360b24ab676SJeff Bonwick * some of the dmu_sync() blocks it points to. In this case, the 361b24ab676SJeff Bonwick * transaction cannot have been committed to anyone (we would have 362b24ab676SJeff Bonwick * waited for all writes to be stable first), so it is semantically 363b24ab676SJeff Bonwick * correct to declare this the end of the log. 364b24ab676SJeff Bonwick */ 365b24ab676SJeff Bonwick if (lr->lr_blkptr.blk_birth >= first_txg && 366b24ab676SJeff Bonwick (error = zil_read_log_data(zilog, lr, NULL)) != 0) 367b24ab676SJeff Bonwick return (error); 368b24ab676SJeff Bonwick 369b24ab676SJeff Bonwick return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 370fa9e4066Sahrens } 371fa9e4066Sahrens 372fa9e4066Sahrens /* ARGSUSED */ 373b24ab676SJeff Bonwick static int 374fa9e4066Sahrens zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 375fa9e4066Sahrens { 376b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 377b24ab676SJeff Bonwick 378b24ab676SJeff Bonwick return (0); 379fa9e4066Sahrens } 380fa9e4066Sahrens 381b24ab676SJeff Bonwick static int 382fa9e4066Sahrens zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 383fa9e4066Sahrens { 384b24ab676SJeff Bonwick lr_write_t *lr = (lr_write_t *)lrc; 385b24ab676SJeff Bonwick blkptr_t *bp = &lr->lr_blkptr; 386b24ab676SJeff Bonwick 387fa9e4066Sahrens /* 388fa9e4066Sahrens * If we previously claimed it, we need to free it. 389fa9e4066Sahrens */ 390b24ab676SJeff Bonwick if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 391b24ab676SJeff Bonwick bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 392b24ab676SJeff Bonwick zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 393b24ab676SJeff Bonwick 394b24ab676SJeff Bonwick return (0); 395fa9e4066Sahrens } 396fa9e4066Sahrens 397fa9e4066Sahrens /* 398fa9e4066Sahrens * Create an on-disk intent log. 399fa9e4066Sahrens */ 400fa9e4066Sahrens static void 401fa9e4066Sahrens zil_create(zilog_t *zilog) 402fa9e4066Sahrens { 403d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 404fa9e4066Sahrens lwb_t *lwb; 405d80c45e0Sbonwick uint64_t txg = 0; 406d80c45e0Sbonwick dmu_tx_t *tx = NULL; 407fa9e4066Sahrens blkptr_t blk; 408d80c45e0Sbonwick int error = 0; 409fa9e4066Sahrens 410fa9e4066Sahrens /* 411d80c45e0Sbonwick * Wait for any previous destroy to complete. 412fa9e4066Sahrens */ 413d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 414d80c45e0Sbonwick 415d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 416d80c45e0Sbonwick ASSERT(zh->zh_replay_seq == 0); 417d80c45e0Sbonwick 418d80c45e0Sbonwick blk = zh->zh_log; 419fa9e4066Sahrens 420fa9e4066Sahrens /* 421899217ddSNeil Perrin * If we don't already have an initial log block or we have one 422899217ddSNeil Perrin * but it's the wrong endianness then allocate one. 423fa9e4066Sahrens */ 424899217ddSNeil Perrin if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 425d80c45e0Sbonwick tx = dmu_tx_create(zilog->zl_os); 426b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 427d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 428d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 429d80c45e0Sbonwick 430899217ddSNeil Perrin if (!BP_IS_HOLE(&blk)) { 431b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, txg, &blk); 432899217ddSNeil Perrin BP_ZERO(&blk); 433899217ddSNeil Perrin } 434899217ddSNeil Perrin 435b24ab676SJeff Bonwick error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 436b24ab676SJeff Bonwick ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 437d80c45e0Sbonwick 438d80c45e0Sbonwick if (error == 0) 439d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 44013f5297eSperrin } 441fa9e4066Sahrens 442d80c45e0Sbonwick /* 443d80c45e0Sbonwick * Allocate a log write buffer (lwb) for the first log block. 444d80c45e0Sbonwick */ 445d80c45e0Sbonwick if (error == 0) { 446fa9e4066Sahrens lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 447fa9e4066Sahrens lwb->lwb_zilog = zilog; 448fa9e4066Sahrens lwb->lwb_blk = blk; 449fa9e4066Sahrens lwb->lwb_nused = 0; 450fa9e4066Sahrens lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); 451fa9e4066Sahrens lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); 452fa9e4066Sahrens lwb->lwb_max_txg = txg; 453c5c6ffa0Smaybee lwb->lwb_zio = NULL; 454b24ab676SJeff Bonwick lwb->lwb_tx = NULL; 455c5c6ffa0Smaybee 456fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 457fa9e4066Sahrens list_insert_tail(&zilog->zl_lwb_list, lwb); 458fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 459fa9e4066Sahrens } 460fa9e4066Sahrens 461d80c45e0Sbonwick /* 462d80c45e0Sbonwick * If we just allocated the first log block, commit our transaction 463d80c45e0Sbonwick * and wait for zil_sync() to stuff the block poiner into zh_log. 464d80c45e0Sbonwick * (zh is part of the MOS, so we cannot modify it in open context.) 465d80c45e0Sbonwick */ 466d80c45e0Sbonwick if (tx != NULL) { 467d80c45e0Sbonwick dmu_tx_commit(tx); 46813f5297eSperrin txg_wait_synced(zilog->zl_dmu_pool, txg); 469d80c45e0Sbonwick } 470d80c45e0Sbonwick 471d80c45e0Sbonwick ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 472fa9e4066Sahrens } 473fa9e4066Sahrens 474fa9e4066Sahrens /* 475fa9e4066Sahrens * In one tx, free all log blocks and clear the log header. 476d80c45e0Sbonwick * If keep_first is set, then we're replaying a log with no content. 477d80c45e0Sbonwick * We want to keep the first block, however, so that the first 478d80c45e0Sbonwick * synchronous transaction doesn't require a txg_wait_synced() 479d80c45e0Sbonwick * in zil_create(). We don't need to txg_wait_synced() here either 480d80c45e0Sbonwick * when keep_first is set, because both zil_create() and zil_destroy() 481d80c45e0Sbonwick * will wait for any in-progress destroys to complete. 482fa9e4066Sahrens */ 483fa9e4066Sahrens void 484d80c45e0Sbonwick zil_destroy(zilog_t *zilog, boolean_t keep_first) 485fa9e4066Sahrens { 486d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 487d80c45e0Sbonwick lwb_t *lwb; 488fa9e4066Sahrens dmu_tx_t *tx; 489fa9e4066Sahrens uint64_t txg; 490fa9e4066Sahrens 491d80c45e0Sbonwick /* 492d80c45e0Sbonwick * Wait for any previous destroy to complete. 493d80c45e0Sbonwick */ 494d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 495fa9e4066Sahrens 496b24ab676SJeff Bonwick zilog->zl_old_header = *zh; /* debugging aid */ 497b24ab676SJeff Bonwick 498d80c45e0Sbonwick if (BP_IS_HOLE(&zh->zh_log)) 499fa9e4066Sahrens return; 500fa9e4066Sahrens 501fa9e4066Sahrens tx = dmu_tx_create(zilog->zl_os); 502b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 503fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 504fa9e4066Sahrens txg = dmu_tx_get_txg(tx); 505fa9e4066Sahrens 506d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 507d80c45e0Sbonwick 508d80c45e0Sbonwick ASSERT3U(zilog->zl_destroy_txg, <, txg); 509fa9e4066Sahrens zilog->zl_destroy_txg = txg; 510b24ab676SJeff Bonwick zilog->zl_keep_first = keep_first; 511d80c45e0Sbonwick 512d80c45e0Sbonwick if (!list_is_empty(&zilog->zl_lwb_list)) { 513d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 514b24ab676SJeff Bonwick ASSERT(!keep_first); 515d80c45e0Sbonwick while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 516d80c45e0Sbonwick list_remove(&zilog->zl_lwb_list, lwb); 517d80c45e0Sbonwick if (lwb->lwb_buf != NULL) 518d80c45e0Sbonwick zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 519b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 520d80c45e0Sbonwick kmem_cache_free(zil_lwb_cache, lwb); 521d80c45e0Sbonwick } 522b24ab676SJeff Bonwick } else if (!keep_first) { 523b24ab676SJeff Bonwick (void) zil_parse(zilog, zil_free_log_block, 524b24ab676SJeff Bonwick zil_free_log_record, tx, zh->zh_claim_txg); 525d80c45e0Sbonwick } 526b19a79ecSperrin mutex_exit(&zilog->zl_lock); 527fa9e4066Sahrens 528fa9e4066Sahrens dmu_tx_commit(tx); 529fa9e4066Sahrens } 530fa9e4066Sahrens 5311d452cf5Sahrens int 532*fd136879SMatthew Ahrens zil_claim(const char *osname, void *txarg) 533fa9e4066Sahrens { 534fa9e4066Sahrens dmu_tx_t *tx = txarg; 535fa9e4066Sahrens uint64_t first_txg = dmu_tx_get_txg(tx); 536fa9e4066Sahrens zilog_t *zilog; 537fa9e4066Sahrens zil_header_t *zh; 538fa9e4066Sahrens objset_t *os; 539fa9e4066Sahrens int error; 540fa9e4066Sahrens 541503ad85cSMatthew Ahrens error = dmu_objset_hold(osname, FTAG, &os); 542fa9e4066Sahrens if (error) { 543b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 5441d452cf5Sahrens return (0); 545fa9e4066Sahrens } 546fa9e4066Sahrens 547fa9e4066Sahrens zilog = dmu_objset_zil(os); 548d80c45e0Sbonwick zh = zil_header_in_syncing_context(zilog); 549fa9e4066Sahrens 550b24ab676SJeff Bonwick if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 551e6ca193dSGeorge Wilson if (!BP_IS_HOLE(&zh->zh_log)) 552b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 553e6ca193dSGeorge Wilson BP_ZERO(&zh->zh_log); 554e6ca193dSGeorge Wilson dsl_dataset_dirty(dmu_objset_ds(os), tx); 555468c413aSTim Haley dmu_objset_rele(os, FTAG); 556468c413aSTim Haley return (0); 557e6ca193dSGeorge Wilson } 558e6ca193dSGeorge Wilson 559fa9e4066Sahrens /* 560d80c45e0Sbonwick * Claim all log blocks if we haven't already done so, and remember 561d80c45e0Sbonwick * the highest claimed sequence number. This ensures that if we can 562d80c45e0Sbonwick * read only part of the log now (e.g. due to a missing device), 563d80c45e0Sbonwick * but we can read the entire log later, we will not try to replay 564d80c45e0Sbonwick * or destroy beyond the last block we successfully claimed. 565fa9e4066Sahrens */ 566fa9e4066Sahrens ASSERT3U(zh->zh_claim_txg, <=, first_txg); 567fa9e4066Sahrens if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 568b24ab676SJeff Bonwick (void) zil_parse(zilog, zil_claim_log_block, 569d80c45e0Sbonwick zil_claim_log_record, tx, first_txg); 570b24ab676SJeff Bonwick zh->zh_claim_txg = first_txg; 571b24ab676SJeff Bonwick zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 572b24ab676SJeff Bonwick zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 573b24ab676SJeff Bonwick if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 574b24ab676SJeff Bonwick zh->zh_flags |= ZIL_REPLAY_NEEDED; 575b24ab676SJeff Bonwick zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 576fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(os), tx); 577fa9e4066Sahrens } 578d80c45e0Sbonwick 579fa9e4066Sahrens ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 580503ad85cSMatthew Ahrens dmu_objset_rele(os, FTAG); 5811d452cf5Sahrens return (0); 582b87f3af3Sperrin } 583b87f3af3Sperrin 584b87f3af3Sperrin /* 585b87f3af3Sperrin * Check the log by walking the log chain. 586b87f3af3Sperrin * Checksum errors are ok as they indicate the end of the chain. 587b87f3af3Sperrin * Any other error (no device or read failure) returns an error. 588b87f3af3Sperrin */ 589b87f3af3Sperrin int 590*fd136879SMatthew Ahrens zil_check_log_chain(const char *osname, void *tx) 591b87f3af3Sperrin { 592b87f3af3Sperrin zilog_t *zilog; 593b87f3af3Sperrin objset_t *os; 594b87f3af3Sperrin int error; 595b87f3af3Sperrin 596b24ab676SJeff Bonwick ASSERT(tx == NULL); 597b24ab676SJeff Bonwick 598503ad85cSMatthew Ahrens error = dmu_objset_hold(osname, FTAG, &os); 599b87f3af3Sperrin if (error) { 600b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 601b87f3af3Sperrin return (0); 602b87f3af3Sperrin } 603b87f3af3Sperrin 604b87f3af3Sperrin zilog = dmu_objset_zil(os); 605b87f3af3Sperrin 606b24ab676SJeff Bonwick /* 607b24ab676SJeff Bonwick * Because tx == NULL, zil_claim_log_block() will not actually claim 608b24ab676SJeff Bonwick * any blocks, but just determine whether it is possible to do so. 609b24ab676SJeff Bonwick * In addition to checking the log chain, zil_claim_log_block() 610b24ab676SJeff Bonwick * will invoke zio_claim() with a done func of spa_claim_notify(), 611b24ab676SJeff Bonwick * which will update spa_max_claim_txg. See spa_load() for details. 612b24ab676SJeff Bonwick */ 613b24ab676SJeff Bonwick error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 614b24ab676SJeff Bonwick zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 615b24ab676SJeff Bonwick 616503ad85cSMatthew Ahrens dmu_objset_rele(os, FTAG); 617b24ab676SJeff Bonwick 618b24ab676SJeff Bonwick return ((error == ECKSUM || error == ENOENT) ? 0 : error); 619b87f3af3Sperrin } 620b87f3af3Sperrin 62117f17c2dSbonwick static int 62217f17c2dSbonwick zil_vdev_compare(const void *x1, const void *x2) 62317f17c2dSbonwick { 624c0d19b33Sperrin uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 625c0d19b33Sperrin uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 62617f17c2dSbonwick 62717f17c2dSbonwick if (v1 < v2) 62817f17c2dSbonwick return (-1); 62917f17c2dSbonwick if (v1 > v2) 63017f17c2dSbonwick return (1); 63117f17c2dSbonwick 63217f17c2dSbonwick return (0); 63317f17c2dSbonwick } 63417f17c2dSbonwick 635fa9e4066Sahrens void 636b24ab676SJeff Bonwick zil_add_block(zilog_t *zilog, const blkptr_t *bp) 637fa9e4066Sahrens { 63817f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 63917f17c2dSbonwick avl_index_t where; 64017f17c2dSbonwick zil_vdev_node_t *zv, zvsearch; 64117f17c2dSbonwick int ndvas = BP_GET_NDVAS(bp); 64217f17c2dSbonwick int i; 643fa9e4066Sahrens 644416e0cd8Sek if (zfs_nocacheflush) 645fa9e4066Sahrens return; 646fa9e4066Sahrens 64717f17c2dSbonwick ASSERT(zilog->zl_writer); 64817f17c2dSbonwick 64917f17c2dSbonwick /* 65017f17c2dSbonwick * Even though we're zl_writer, we still need a lock because the 65117f17c2dSbonwick * zl_get_data() callbacks may have dmu_sync() done callbacks 65217f17c2dSbonwick * that will run concurrently. 65317f17c2dSbonwick */ 65417f17c2dSbonwick mutex_enter(&zilog->zl_vdev_lock); 65517f17c2dSbonwick for (i = 0; i < ndvas; i++) { 65617f17c2dSbonwick zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 65717f17c2dSbonwick if (avl_find(t, &zvsearch, &where) == NULL) { 65817f17c2dSbonwick zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 65917f17c2dSbonwick zv->zv_vdev = zvsearch.zv_vdev; 66017f17c2dSbonwick avl_insert(t, zv, where); 66167bd71c6Sperrin } 66267bd71c6Sperrin } 66317f17c2dSbonwick mutex_exit(&zilog->zl_vdev_lock); 664fa9e4066Sahrens } 665fa9e4066Sahrens 66667bd71c6Sperrin void 66767bd71c6Sperrin zil_flush_vdevs(zilog_t *zilog) 66867bd71c6Sperrin { 66967bd71c6Sperrin spa_t *spa = zilog->zl_spa; 67017f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 67117f17c2dSbonwick void *cookie = NULL; 67217f17c2dSbonwick zil_vdev_node_t *zv; 67317f17c2dSbonwick zio_t *zio; 674fa9e4066Sahrens 67567bd71c6Sperrin ASSERT(zilog->zl_writer); 67667bd71c6Sperrin 67717f17c2dSbonwick /* 67817f17c2dSbonwick * We don't need zl_vdev_lock here because we're the zl_writer, 67917f17c2dSbonwick * and all zl_get_data() callbacks are done. 68017f17c2dSbonwick */ 68117f17c2dSbonwick if (avl_numnodes(t) == 0) 68217f17c2dSbonwick return; 68317f17c2dSbonwick 684e14bb325SJeff Bonwick spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 68517f17c2dSbonwick 686e14bb325SJeff Bonwick zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 687fa9e4066Sahrens 68817f17c2dSbonwick while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 68917f17c2dSbonwick vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 69017f17c2dSbonwick if (vd != NULL) 69117f17c2dSbonwick zio_flush(zio, vd); 69217f17c2dSbonwick kmem_free(zv, sizeof (*zv)); 69367bd71c6Sperrin } 69417f17c2dSbonwick 695fa9e4066Sahrens /* 696fa9e4066Sahrens * Wait for all the flushes to complete. Not all devices actually 697fa9e4066Sahrens * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 698fa9e4066Sahrens */ 69917f17c2dSbonwick (void) zio_wait(zio); 70017f17c2dSbonwick 701e14bb325SJeff Bonwick spa_config_exit(spa, SCL_STATE, FTAG); 702fa9e4066Sahrens } 703fa9e4066Sahrens 704fa9e4066Sahrens /* 705fa9e4066Sahrens * Function called when a log block write completes 706fa9e4066Sahrens */ 707fa9e4066Sahrens static void 708fa9e4066Sahrens zil_lwb_write_done(zio_t *zio) 709fa9e4066Sahrens { 710fa9e4066Sahrens lwb_t *lwb = zio->io_private; 711fa9e4066Sahrens zilog_t *zilog = lwb->lwb_zilog; 712b24ab676SJeff Bonwick dmu_tx_t *tx = lwb->lwb_tx; 713fa9e4066Sahrens 714e14bb325SJeff Bonwick ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 715e14bb325SJeff Bonwick ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG); 716e14bb325SJeff Bonwick ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 717e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 718e14bb325SJeff Bonwick ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 719e14bb325SJeff Bonwick ASSERT(!BP_IS_GANG(zio->io_bp)); 720e14bb325SJeff Bonwick ASSERT(!BP_IS_HOLE(zio->io_bp)); 721e14bb325SJeff Bonwick ASSERT(zio->io_bp->blk_fill == 0); 722e14bb325SJeff Bonwick 723fa9e4066Sahrens /* 724ef0d8e11SNeil Perrin * Ensure the lwb buffer pointer is cleared before releasing 725ef0d8e11SNeil Perrin * the txg. If we have had an allocation failure and 726ef0d8e11SNeil Perrin * the txg is waiting to sync then we want want zil_sync() 727ef0d8e11SNeil Perrin * to remove the lwb so that it's not picked up as the next new 728ef0d8e11SNeil Perrin * one in zil_commit_writer(). zil_sync() will only remove 729ef0d8e11SNeil Perrin * the lwb if lwb_buf is null. 730fa9e4066Sahrens */ 731fa9e4066Sahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 732fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 733fa9e4066Sahrens lwb->lwb_buf = NULL; 734b24ab676SJeff Bonwick lwb->lwb_tx = NULL; 735b24ab676SJeff Bonwick mutex_exit(&zilog->zl_lock); 736ef0d8e11SNeil Perrin 737ef0d8e11SNeil Perrin /* 738ef0d8e11SNeil Perrin * Now that we've written this log block, we have a stable pointer 739ef0d8e11SNeil Perrin * to the next block in the chain, so it's OK to let the txg in 740b24ab676SJeff Bonwick * which we allocated the next block sync. 741ef0d8e11SNeil Perrin */ 742b24ab676SJeff Bonwick dmu_tx_commit(tx); 743fa9e4066Sahrens } 744fa9e4066Sahrens 745c5c6ffa0Smaybee /* 746c5c6ffa0Smaybee * Initialize the io for a log block. 747c5c6ffa0Smaybee */ 748c5c6ffa0Smaybee static void 749c5c6ffa0Smaybee zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 750c5c6ffa0Smaybee { 751c5c6ffa0Smaybee zbookmark_t zb; 752c5c6ffa0Smaybee 753b24ab676SJeff Bonwick SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 754b24ab676SJeff Bonwick ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 755b24ab676SJeff Bonwick lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 756c5c6ffa0Smaybee 757b19a79ecSperrin if (zilog->zl_root_zio == NULL) { 758b19a79ecSperrin zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 759b19a79ecSperrin ZIO_FLAG_CANFAIL); 760b19a79ecSperrin } 76167bd71c6Sperrin if (lwb->lwb_zio == NULL) { 76267bd71c6Sperrin lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 763e6ca193dSGeorge Wilson 0, &lwb->lwb_blk, lwb->lwb_buf, lwb->lwb_sz, 764e6ca193dSGeorge Wilson zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 7658f18d1faSGeorge Wilson ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 76667bd71c6Sperrin } 767c5c6ffa0Smaybee } 768c5c6ffa0Smaybee 769d48e086fSNeil Perrin /* 770d48e086fSNeil Perrin * Use the slog as long as the logbias is 'latency' and the current commit size 771d48e086fSNeil Perrin * is less than the limit or the total list size is less than 2X the limit. 772d48e086fSNeil Perrin * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 773d48e086fSNeil Perrin */ 774d48e086fSNeil Perrin uint64_t zil_slog_limit = 1024 * 1024; 775d48e086fSNeil Perrin #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 776d48e086fSNeil Perrin (((zilog)->zl_cur_used < zil_slog_limit) || \ 777d48e086fSNeil Perrin ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 778d48e086fSNeil Perrin 779fa9e4066Sahrens /* 780fa9e4066Sahrens * Start a log block write and advance to the next log block. 781fa9e4066Sahrens * Calls are serialized. 782fa9e4066Sahrens */ 783fa9e4066Sahrens static lwb_t * 784fa9e4066Sahrens zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 785fa9e4066Sahrens { 786fa9e4066Sahrens lwb_t *nlwb; 787fa9e4066Sahrens zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; 788d80c45e0Sbonwick spa_t *spa = zilog->zl_spa; 789d80c45e0Sbonwick blkptr_t *bp = &ztp->zit_next_blk; 790b24ab676SJeff Bonwick dmu_tx_t *tx; 791fa9e4066Sahrens uint64_t txg; 792fa9e4066Sahrens uint64_t zil_blksz; 793fa9e4066Sahrens int error; 794fa9e4066Sahrens 795fa9e4066Sahrens ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); 796fa9e4066Sahrens 797fa9e4066Sahrens /* 798fa9e4066Sahrens * Allocate the next block and save its address in this block 799fa9e4066Sahrens * before writing it in order to establish the log chain. 800fa9e4066Sahrens * Note that if the allocation of nlwb synced before we wrote 801fa9e4066Sahrens * the block that points at it (lwb), we'd leak it if we crashed. 802b24ab676SJeff Bonwick * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 803b24ab676SJeff Bonwick * We dirty the dataset to ensure that zil_sync() will be called 804b24ab676SJeff Bonwick * to clean up in the event of allocation failure or I/O failure. 805fa9e4066Sahrens */ 806b24ab676SJeff Bonwick tx = dmu_tx_create(zilog->zl_os); 807b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 808b24ab676SJeff Bonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 809b24ab676SJeff Bonwick txg = dmu_tx_get_txg(tx); 810b24ab676SJeff Bonwick 811b24ab676SJeff Bonwick lwb->lwb_tx = tx; 812fa9e4066Sahrens 813fa9e4066Sahrens /* 81422ac5be4Sperrin * Pick a ZIL blocksize. We request a size that is the 81522ac5be4Sperrin * maximum of the previous used size, the current used size and 81622ac5be4Sperrin * the amount waiting in the queue. 817fa9e4066Sahrens */ 818c5c6ffa0Smaybee zil_blksz = MAX(zilog->zl_prev_used, 819c5c6ffa0Smaybee zilog->zl_cur_used + sizeof (*ztp)); 82022ac5be4Sperrin zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); 821b4d654b0Sperrin zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); 82222ac5be4Sperrin if (zil_blksz > ZIL_MAX_BLKSZ) 82322ac5be4Sperrin zil_blksz = ZIL_MAX_BLKSZ; 824fa9e4066Sahrens 82567bd71c6Sperrin BP_ZERO(bp); 82667bd71c6Sperrin /* pass the old blkptr in order to spread log blocks across devs */ 827b24ab676SJeff Bonwick error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 828d48e086fSNeil Perrin USE_SLOG(zilog)); 829fa9e4066Sahrens if (error) { 830d63d470bSgw /* 831b24ab676SJeff Bonwick * Since we've just experienced an allocation failure, 832d63d470bSgw * terminate the current lwb and send it on its way. 833d63d470bSgw */ 834d63d470bSgw ztp->zit_pad = 0; 835d63d470bSgw ztp->zit_nused = lwb->lwb_nused; 836d63d470bSgw ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 837d63d470bSgw zio_nowait(lwb->lwb_zio); 838d63d470bSgw 839ea8dc4b6Seschrock /* 840ea8dc4b6Seschrock * By returning NULL the caller will call tx_wait_synced() 841ea8dc4b6Seschrock */ 842fa9e4066Sahrens return (NULL); 843fa9e4066Sahrens } 844fa9e4066Sahrens 845d80c45e0Sbonwick ASSERT3U(bp->blk_birth, ==, txg); 846ea8dc4b6Seschrock ztp->zit_pad = 0; 847fa9e4066Sahrens ztp->zit_nused = lwb->lwb_nused; 848fa9e4066Sahrens ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 849d80c45e0Sbonwick bp->blk_cksum = lwb->lwb_blk.blk_cksum; 850d80c45e0Sbonwick bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 851fa9e4066Sahrens 852fa9e4066Sahrens /* 853fa9e4066Sahrens * Allocate a new log write buffer (lwb). 854fa9e4066Sahrens */ 855fa9e4066Sahrens nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 856fa9e4066Sahrens nlwb->lwb_zilog = zilog; 857d80c45e0Sbonwick nlwb->lwb_blk = *bp; 858fa9e4066Sahrens nlwb->lwb_nused = 0; 859fa9e4066Sahrens nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); 860fa9e4066Sahrens nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); 861fa9e4066Sahrens nlwb->lwb_max_txg = txg; 862c5c6ffa0Smaybee nlwb->lwb_zio = NULL; 863b24ab676SJeff Bonwick nlwb->lwb_tx = NULL; 864fa9e4066Sahrens 865fa9e4066Sahrens /* 86667bd71c6Sperrin * Put new lwb at the end of the log chain 867fa9e4066Sahrens */ 868fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 869fa9e4066Sahrens list_insert_tail(&zilog->zl_lwb_list, nlwb); 870fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 871fa9e4066Sahrens 87217f17c2dSbonwick /* Record the block for later vdev flushing */ 87317f17c2dSbonwick zil_add_block(zilog, &lwb->lwb_blk); 87467bd71c6Sperrin 875fa9e4066Sahrens /* 876c5c6ffa0Smaybee * kick off the write for the old log block 877fa9e4066Sahrens */ 87867bd71c6Sperrin ASSERT(lwb->lwb_zio); 879c5c6ffa0Smaybee zio_nowait(lwb->lwb_zio); 880fa9e4066Sahrens 881fa9e4066Sahrens return (nlwb); 882fa9e4066Sahrens } 883fa9e4066Sahrens 884fa9e4066Sahrens static lwb_t * 885fa9e4066Sahrens zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 886fa9e4066Sahrens { 887fa9e4066Sahrens lr_t *lrc = &itx->itx_lr; /* common log record */ 888b24ab676SJeff Bonwick lr_write_t *lrw = (lr_write_t *)lrc; 889b24ab676SJeff Bonwick char *lr_buf; 890fa9e4066Sahrens uint64_t txg = lrc->lrc_txg; 891fa9e4066Sahrens uint64_t reclen = lrc->lrc_reclen; 892b24ab676SJeff Bonwick uint64_t dlen = 0; 893fa9e4066Sahrens 894fa9e4066Sahrens if (lwb == NULL) 895fa9e4066Sahrens return (NULL); 896b24ab676SJeff Bonwick 897fa9e4066Sahrens ASSERT(lwb->lwb_buf != NULL); 898fa9e4066Sahrens 899c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 900c5c6ffa0Smaybee dlen = P2ROUNDUP_TYPED( 901b24ab676SJeff Bonwick lrw->lr_length, sizeof (uint64_t), uint64_t); 902fa9e4066Sahrens 903104e2ed7Sperrin zilog->zl_cur_used += (reclen + dlen); 90422ac5be4Sperrin 90567bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 90667bd71c6Sperrin 907fa9e4066Sahrens /* 908fa9e4066Sahrens * If this record won't fit in the current log block, start a new one. 909fa9e4066Sahrens */ 910104e2ed7Sperrin if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 911fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 912c5c6ffa0Smaybee if (lwb == NULL) 913fa9e4066Sahrens return (NULL); 91467bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 915ea8dc4b6Seschrock ASSERT(lwb->lwb_nused == 0); 916104e2ed7Sperrin if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 917fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, txg); 918fa9e4066Sahrens return (lwb); 919fa9e4066Sahrens } 920fa9e4066Sahrens } 921fa9e4066Sahrens 922b24ab676SJeff Bonwick lr_buf = lwb->lwb_buf + lwb->lwb_nused; 923b24ab676SJeff Bonwick bcopy(lrc, lr_buf, reclen); 924b24ab676SJeff Bonwick lrc = (lr_t *)lr_buf; 925b24ab676SJeff Bonwick lrw = (lr_write_t *)lrc; 926c5c6ffa0Smaybee 927c5c6ffa0Smaybee /* 928c5c6ffa0Smaybee * If it's a write, fetch the data or get its blkptr as appropriate. 929c5c6ffa0Smaybee */ 930c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE) { 931c5c6ffa0Smaybee if (txg > spa_freeze_txg(zilog->zl_spa)) 932c5c6ffa0Smaybee txg_wait_synced(zilog->zl_dmu_pool, txg); 933c5c6ffa0Smaybee if (itx->itx_wr_state != WR_COPIED) { 934c5c6ffa0Smaybee char *dbuf; 935c5c6ffa0Smaybee int error; 936c5c6ffa0Smaybee 937c5c6ffa0Smaybee if (dlen) { 938c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_NEED_COPY); 939b24ab676SJeff Bonwick dbuf = lr_buf + reclen; 940b24ab676SJeff Bonwick lrw->lr_common.lrc_reclen += dlen; 941c5c6ffa0Smaybee } else { 942c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_INDIRECT); 943c5c6ffa0Smaybee dbuf = NULL; 944c5c6ffa0Smaybee } 945c5c6ffa0Smaybee error = zilog->zl_get_data( 946b24ab676SJeff Bonwick itx->itx_private, lrw, dbuf, lwb->lwb_zio); 947c87b8fc5SMark J Musante if (error == EIO) { 948c87b8fc5SMark J Musante txg_wait_synced(zilog->zl_dmu_pool, txg); 949c87b8fc5SMark J Musante return (lwb); 950c87b8fc5SMark J Musante } 951c5c6ffa0Smaybee if (error) { 952c5c6ffa0Smaybee ASSERT(error == ENOENT || error == EEXIST || 953c5c6ffa0Smaybee error == EALREADY); 954c5c6ffa0Smaybee return (lwb); 955c5c6ffa0Smaybee } 956c5c6ffa0Smaybee } 957104e2ed7Sperrin } 958c5c6ffa0Smaybee 959b24ab676SJeff Bonwick /* 960b24ab676SJeff Bonwick * We're actually making an entry, so update lrc_seq to be the 961b24ab676SJeff Bonwick * log record sequence number. Note that this is generally not 962b24ab676SJeff Bonwick * equal to the itx sequence number because not all transactions 963b24ab676SJeff Bonwick * are synchronous, and sometimes spa_sync() gets there first. 964b24ab676SJeff Bonwick */ 965b24ab676SJeff Bonwick lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 966c5c6ffa0Smaybee lwb->lwb_nused += reclen + dlen; 967fa9e4066Sahrens lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 968fa9e4066Sahrens ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); 969fa9e4066Sahrens ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 970fa9e4066Sahrens 971fa9e4066Sahrens return (lwb); 972fa9e4066Sahrens } 973fa9e4066Sahrens 974fa9e4066Sahrens itx_t * 975da6c28aaSamw zil_itx_create(uint64_t txtype, size_t lrsize) 976fa9e4066Sahrens { 977fa9e4066Sahrens itx_t *itx; 978fa9e4066Sahrens 979b4d654b0Sperrin lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 980fa9e4066Sahrens 981fa9e4066Sahrens itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 982fa9e4066Sahrens itx->itx_lr.lrc_txtype = txtype; 983fa9e4066Sahrens itx->itx_lr.lrc_reclen = lrsize; 984abf76b6eSperrin itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 985fa9e4066Sahrens itx->itx_lr.lrc_seq = 0; /* defensive */ 986fa9e4066Sahrens 987fa9e4066Sahrens return (itx); 988fa9e4066Sahrens } 989fa9e4066Sahrens 990b24ab676SJeff Bonwick void 991b24ab676SJeff Bonwick zil_itx_destroy(itx_t *itx) 992b24ab676SJeff Bonwick { 993b24ab676SJeff Bonwick kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 994b24ab676SJeff Bonwick } 995b24ab676SJeff Bonwick 996fa9e4066Sahrens uint64_t 997fa9e4066Sahrens zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 998fa9e4066Sahrens { 999fa9e4066Sahrens uint64_t seq; 1000fa9e4066Sahrens 1001fa9e4066Sahrens ASSERT(itx->itx_lr.lrc_seq == 0); 1002b24ab676SJeff Bonwick ASSERT(!zilog->zl_replay); 1003fa9e4066Sahrens 1004fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1005fa9e4066Sahrens list_insert_tail(&zilog->zl_itx_list, itx); 1006abf76b6eSperrin zilog->zl_itx_list_sz += itx->itx_sod; 1007fa9e4066Sahrens itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1008fa9e4066Sahrens itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 1009fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1010fa9e4066Sahrens 1011fa9e4066Sahrens return (seq); 1012fa9e4066Sahrens } 1013fa9e4066Sahrens 1014fa9e4066Sahrens /* 1015fa9e4066Sahrens * Free up all in-memory intent log transactions that have now been synced. 1016fa9e4066Sahrens */ 1017fa9e4066Sahrens static void 1018fa9e4066Sahrens zil_itx_clean(zilog_t *zilog) 1019fa9e4066Sahrens { 1020fa9e4066Sahrens uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 1021fa9e4066Sahrens uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 1022a584ef65Sjohansen list_t clean_list; 1023fa9e4066Sahrens itx_t *itx; 1024fa9e4066Sahrens 1025a584ef65Sjohansen list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1026a584ef65Sjohansen 1027fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1028b19a79ecSperrin /* wait for a log writer to finish walking list */ 1029b19a79ecSperrin while (zilog->zl_writer) { 1030b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1031b19a79ecSperrin } 1032a584ef65Sjohansen 1033a584ef65Sjohansen /* 1034a584ef65Sjohansen * Move the sync'd log transactions to a separate list so we can call 1035a584ef65Sjohansen * kmem_free without holding the zl_lock. 1036a584ef65Sjohansen * 1037a584ef65Sjohansen * There is no need to set zl_writer as we don't drop zl_lock here 1038a584ef65Sjohansen */ 1039fa9e4066Sahrens while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 1040fa9e4066Sahrens itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 1041fa9e4066Sahrens list_remove(&zilog->zl_itx_list, itx); 1042abf76b6eSperrin zilog->zl_itx_list_sz -= itx->itx_sod; 1043a584ef65Sjohansen list_insert_tail(&clean_list, itx); 1044fa9e4066Sahrens } 104567bd71c6Sperrin cv_broadcast(&zilog->zl_cv_writer); 1046fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1047a584ef65Sjohansen 1048a584ef65Sjohansen /* destroy sync'd log transactions */ 1049a584ef65Sjohansen while ((itx = list_head(&clean_list)) != NULL) { 1050a584ef65Sjohansen list_remove(&clean_list, itx); 1051b24ab676SJeff Bonwick zil_itx_destroy(itx); 1052a584ef65Sjohansen } 1053a584ef65Sjohansen list_destroy(&clean_list); 1054fa9e4066Sahrens } 1055fa9e4066Sahrens 1056b19a79ecSperrin /* 105767bd71c6Sperrin * If there are any in-memory intent log transactions which have now been 105867bd71c6Sperrin * synced then start up a taskq to free them. 1059b19a79ecSperrin */ 1060fa9e4066Sahrens void 1061fa9e4066Sahrens zil_clean(zilog_t *zilog) 1062fa9e4066Sahrens { 106367bd71c6Sperrin itx_t *itx; 106467bd71c6Sperrin 1065fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 106667bd71c6Sperrin itx = list_head(&zilog->zl_itx_list); 106767bd71c6Sperrin if ((itx != NULL) && 106867bd71c6Sperrin (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 1069fa9e4066Sahrens (void) taskq_dispatch(zilog->zl_clean_taskq, 1070d48e086fSNeil Perrin (task_func_t *)zil_itx_clean, zilog, TQ_NOSLEEP); 107167bd71c6Sperrin } 1072fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1073fa9e4066Sahrens } 1074fa9e4066Sahrens 1075e14bb325SJeff Bonwick static void 1076b19a79ecSperrin zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 1077fa9e4066Sahrens { 1078fa9e4066Sahrens uint64_t txg; 107967bd71c6Sperrin uint64_t commit_seq = 0; 1080b24ab676SJeff Bonwick itx_t *itx, *itx_next; 1081fa9e4066Sahrens lwb_t *lwb; 1082fa9e4066Sahrens spa_t *spa; 1083b24ab676SJeff Bonwick int error = 0; 1084fa9e4066Sahrens 1085fa9e4066Sahrens zilog->zl_writer = B_TRUE; 1086e14bb325SJeff Bonwick ASSERT(zilog->zl_root_zio == NULL); 1087b19a79ecSperrin spa = zilog->zl_spa; 1088fa9e4066Sahrens 1089fa9e4066Sahrens if (zilog->zl_suspend) { 1090fa9e4066Sahrens lwb = NULL; 1091fa9e4066Sahrens } else { 1092fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 1093fa9e4066Sahrens if (lwb == NULL) { 1094b19a79ecSperrin /* 1095b19a79ecSperrin * Return if there's nothing to flush before we 1096b19a79ecSperrin * dirty the fs by calling zil_create() 1097b19a79ecSperrin */ 1098b19a79ecSperrin if (list_is_empty(&zilog->zl_itx_list)) { 1099b19a79ecSperrin zilog->zl_writer = B_FALSE; 1100b19a79ecSperrin return; 1101b19a79ecSperrin } 1102fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1103fa9e4066Sahrens zil_create(zilog); 1104fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1105fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 1106fa9e4066Sahrens } 1107fa9e4066Sahrens } 1108fa9e4066Sahrens 110967bd71c6Sperrin /* Loop through in-memory log transactions filling log blocks. */ 1110b19a79ecSperrin DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1111b24ab676SJeff Bonwick 1112b24ab676SJeff Bonwick for (itx = list_head(&zilog->zl_itx_list); itx; itx = itx_next) { 1113b19a79ecSperrin /* 1114b24ab676SJeff Bonwick * Save the next pointer. Even though we drop zl_lock below, 1115b24ab676SJeff Bonwick * all threads that can remove itx list entries (other writers 1116b24ab676SJeff Bonwick * and zil_itx_clean()) can't do so until they have zl_writer. 1117b19a79ecSperrin */ 1118b24ab676SJeff Bonwick itx_next = list_next(&zilog->zl_itx_list, itx); 1119b24ab676SJeff Bonwick 1120b24ab676SJeff Bonwick /* 1121b24ab676SJeff Bonwick * Determine whether to push this itx. 1122b24ab676SJeff Bonwick * Push all transactions related to specified foid and 1123b24ab676SJeff Bonwick * all other transactions except those that can be logged 1124b24ab676SJeff Bonwick * out of order (TX_WRITE, TX_TRUNCATE, TX_SETATTR, TX_ACL) 1125b24ab676SJeff Bonwick * for all other files. 1126b24ab676SJeff Bonwick * 1127b24ab676SJeff Bonwick * If foid == 0 (meaning "push all foids") or 1128b24ab676SJeff Bonwick * itx->itx_sync is set (meaning O_[D]SYNC), push regardless. 1129b24ab676SJeff Bonwick */ 1130b24ab676SJeff Bonwick if (foid != 0 && !itx->itx_sync && 1131b24ab676SJeff Bonwick TX_OOO(itx->itx_lr.lrc_txtype) && 1132b24ab676SJeff Bonwick ((lr_ooo_t *)&itx->itx_lr)->lr_foid != foid) 1133b24ab676SJeff Bonwick continue; /* skip this record */ 1134fa9e4066Sahrens 1135fa9e4066Sahrens if ((itx->itx_lr.lrc_seq > seq) && 1136b19a79ecSperrin ((lwb == NULL) || (lwb->lwb_nused == 0) || 1137b24ab676SJeff Bonwick (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) 1138fa9e4066Sahrens break; 1139fa9e4066Sahrens 1140fa9e4066Sahrens list_remove(&zilog->zl_itx_list, itx); 1141abf76b6eSperrin zilog->zl_itx_list_sz -= itx->itx_sod; 1142b24ab676SJeff Bonwick 114367bd71c6Sperrin mutex_exit(&zilog->zl_lock); 1144b24ab676SJeff Bonwick 1145fa9e4066Sahrens txg = itx->itx_lr.lrc_txg; 1146fa9e4066Sahrens ASSERT(txg); 1147fa9e4066Sahrens 1148fa9e4066Sahrens if (txg > spa_last_synced_txg(spa) || 1149fa9e4066Sahrens txg > spa_freeze_txg(spa)) 1150fa9e4066Sahrens lwb = zil_lwb_commit(zilog, itx, lwb); 1151b24ab676SJeff Bonwick 1152b24ab676SJeff Bonwick zil_itx_destroy(itx); 1153b24ab676SJeff Bonwick 1154fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1155fa9e4066Sahrens } 1156b19a79ecSperrin DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 115767bd71c6Sperrin /* determine commit sequence number */ 115867bd71c6Sperrin itx = list_head(&zilog->zl_itx_list); 115967bd71c6Sperrin if (itx) 1160b24ab676SJeff Bonwick commit_seq = itx->itx_lr.lrc_seq - 1; 116167bd71c6Sperrin else 116267bd71c6Sperrin commit_seq = zilog->zl_itx_seq; 1163fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1164fa9e4066Sahrens 1165fa9e4066Sahrens /* write the last block out */ 116667bd71c6Sperrin if (lwb != NULL && lwb->lwb_zio != NULL) 1167fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 1168fa9e4066Sahrens 116922ac5be4Sperrin zilog->zl_prev_used = zilog->zl_cur_used; 117022ac5be4Sperrin zilog->zl_cur_used = 0; 1171fa9e4066Sahrens 1172fa9e4066Sahrens /* 1173b19a79ecSperrin * Wait if necessary for the log blocks to be on stable storage. 1174fa9e4066Sahrens */ 1175b19a79ecSperrin if (zilog->zl_root_zio) { 1176b19a79ecSperrin DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1177b24ab676SJeff Bonwick error = zio_wait(zilog->zl_root_zio); 1178e14bb325SJeff Bonwick zilog->zl_root_zio = NULL; 1179b19a79ecSperrin DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 118017f17c2dSbonwick zil_flush_vdevs(zilog); 1181fa9e4066Sahrens } 118222ac5be4Sperrin 1183b24ab676SJeff Bonwick if (error || lwb == NULL) 1184fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 118567bd71c6Sperrin 118667bd71c6Sperrin mutex_enter(&zilog->zl_lock); 118722ac5be4Sperrin zilog->zl_writer = B_FALSE; 118867bd71c6Sperrin 118967bd71c6Sperrin ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 119067bd71c6Sperrin zilog->zl_commit_seq = commit_seq; 1191b24ab676SJeff Bonwick 1192b24ab676SJeff Bonwick /* 1193b24ab676SJeff Bonwick * Remember the highest committed log sequence number for ztest. 1194b24ab676SJeff Bonwick * We only update this value when all the log writes succeeded, 1195b24ab676SJeff Bonwick * because ztest wants to ASSERT that it got the whole log chain. 1196b24ab676SJeff Bonwick */ 1197b24ab676SJeff Bonwick if (error == 0 && lwb != NULL) 1198b24ab676SJeff Bonwick zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1199b19a79ecSperrin } 1200b19a79ecSperrin 1201b19a79ecSperrin /* 1202b19a79ecSperrin * Push zfs transactions to stable storage up to the supplied sequence number. 1203b19a79ecSperrin * If foid is 0 push out all transactions, otherwise push only those 1204b19a79ecSperrin * for that file or might have been used to create that file. 1205b19a79ecSperrin */ 1206b19a79ecSperrin void 1207b19a79ecSperrin zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1208b19a79ecSperrin { 1209b19a79ecSperrin if (zilog == NULL || seq == 0) 1210b19a79ecSperrin return; 1211b19a79ecSperrin 1212b19a79ecSperrin mutex_enter(&zilog->zl_lock); 1213b19a79ecSperrin 1214b19a79ecSperrin seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1215b19a79ecSperrin 121667bd71c6Sperrin while (zilog->zl_writer) { 1217b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1218b24ab676SJeff Bonwick if (seq <= zilog->zl_commit_seq) { 121967bd71c6Sperrin mutex_exit(&zilog->zl_lock); 122067bd71c6Sperrin return; 122167bd71c6Sperrin } 122267bd71c6Sperrin } 1223b19a79ecSperrin zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 122467bd71c6Sperrin /* wake up others waiting on the commit */ 122567bd71c6Sperrin cv_broadcast(&zilog->zl_cv_writer); 122667bd71c6Sperrin mutex_exit(&zilog->zl_lock); 1227fa9e4066Sahrens } 1228fa9e4066Sahrens 1229b24ab676SJeff Bonwick /* 1230b24ab676SJeff Bonwick * Report whether all transactions are committed. 1231b24ab676SJeff Bonwick */ 1232b24ab676SJeff Bonwick static boolean_t 1233b24ab676SJeff Bonwick zil_is_committed(zilog_t *zilog) 1234b24ab676SJeff Bonwick { 1235b24ab676SJeff Bonwick lwb_t *lwb; 1236b24ab676SJeff Bonwick boolean_t committed; 1237b24ab676SJeff Bonwick 1238b24ab676SJeff Bonwick mutex_enter(&zilog->zl_lock); 1239b24ab676SJeff Bonwick 1240b24ab676SJeff Bonwick while (zilog->zl_writer) 1241b24ab676SJeff Bonwick cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1242b24ab676SJeff Bonwick 1243b24ab676SJeff Bonwick if (!list_is_empty(&zilog->zl_itx_list)) 1244b24ab676SJeff Bonwick committed = B_FALSE; /* unpushed transactions */ 1245b24ab676SJeff Bonwick else if ((lwb = list_head(&zilog->zl_lwb_list)) == NULL) 1246b24ab676SJeff Bonwick committed = B_TRUE; /* intent log never used */ 1247b24ab676SJeff Bonwick else if (list_next(&zilog->zl_lwb_list, lwb) != NULL) 1248b24ab676SJeff Bonwick committed = B_FALSE; /* zil_sync() not done yet */ 1249b24ab676SJeff Bonwick else 1250b24ab676SJeff Bonwick committed = B_TRUE; /* everything synced */ 1251b24ab676SJeff Bonwick 1252b24ab676SJeff Bonwick mutex_exit(&zilog->zl_lock); 1253b24ab676SJeff Bonwick return (committed); 1254b24ab676SJeff Bonwick } 1255b24ab676SJeff Bonwick 1256fa9e4066Sahrens /* 1257fa9e4066Sahrens * Called in syncing context to free committed log blocks and update log header. 1258fa9e4066Sahrens */ 1259fa9e4066Sahrens void 1260fa9e4066Sahrens zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1261fa9e4066Sahrens { 1262d80c45e0Sbonwick zil_header_t *zh = zil_header_in_syncing_context(zilog); 1263fa9e4066Sahrens uint64_t txg = dmu_tx_get_txg(tx); 1264fa9e4066Sahrens spa_t *spa = zilog->zl_spa; 1265b24ab676SJeff Bonwick uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1266fa9e4066Sahrens lwb_t *lwb; 1267fa9e4066Sahrens 126814843421SMatthew Ahrens /* 126914843421SMatthew Ahrens * We don't zero out zl_destroy_txg, so make sure we don't try 127014843421SMatthew Ahrens * to destroy it twice. 127114843421SMatthew Ahrens */ 127214843421SMatthew Ahrens if (spa_sync_pass(spa) != 1) 127314843421SMatthew Ahrens return; 127414843421SMatthew Ahrens 1275d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1276d80c45e0Sbonwick 1277fa9e4066Sahrens ASSERT(zilog->zl_stop_sync == 0); 1278fa9e4066Sahrens 1279b24ab676SJeff Bonwick if (*replayed_seq != 0) { 1280b24ab676SJeff Bonwick ASSERT(zh->zh_replay_seq < *replayed_seq); 1281b24ab676SJeff Bonwick zh->zh_replay_seq = *replayed_seq; 1282b24ab676SJeff Bonwick *replayed_seq = 0; 1283b24ab676SJeff Bonwick } 1284fa9e4066Sahrens 1285fa9e4066Sahrens if (zilog->zl_destroy_txg == txg) { 1286d80c45e0Sbonwick blkptr_t blk = zh->zh_log; 1287d80c45e0Sbonwick 1288d80c45e0Sbonwick ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1289d80c45e0Sbonwick 1290d80c45e0Sbonwick bzero(zh, sizeof (zil_header_t)); 12911209a471SNeil Perrin bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1292d80c45e0Sbonwick 1293d80c45e0Sbonwick if (zilog->zl_keep_first) { 1294d80c45e0Sbonwick /* 1295d80c45e0Sbonwick * If this block was part of log chain that couldn't 1296d80c45e0Sbonwick * be claimed because a device was missing during 1297d80c45e0Sbonwick * zil_claim(), but that device later returns, 1298d80c45e0Sbonwick * then this block could erroneously appear valid. 1299d80c45e0Sbonwick * To guard against this, assign a new GUID to the new 1300d80c45e0Sbonwick * log chain so it doesn't matter what blk points to. 1301d80c45e0Sbonwick */ 1302d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 1303d80c45e0Sbonwick zh->zh_log = blk; 1304d80c45e0Sbonwick } 1305fa9e4066Sahrens } 1306fa9e4066Sahrens 1307e6ca193dSGeorge Wilson while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1308b19a79ecSperrin zh->zh_log = lwb->lwb_blk; 1309fa9e4066Sahrens if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1310fa9e4066Sahrens break; 1311fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1312b24ab676SJeff Bonwick zio_free_zil(spa, txg, &lwb->lwb_blk); 1313fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1314d63d470bSgw 1315d63d470bSgw /* 1316d63d470bSgw * If we don't have anything left in the lwb list then 1317d63d470bSgw * we've had an allocation failure and we need to zero 1318d63d470bSgw * out the zil_header blkptr so that we don't end 1319d63d470bSgw * up freeing the same block twice. 1320d63d470bSgw */ 1321d63d470bSgw if (list_head(&zilog->zl_lwb_list) == NULL) 1322d63d470bSgw BP_ZERO(&zh->zh_log); 1323fa9e4066Sahrens } 1324fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1325fa9e4066Sahrens } 1326fa9e4066Sahrens 1327fa9e4066Sahrens void 1328fa9e4066Sahrens zil_init(void) 1329fa9e4066Sahrens { 1330fa9e4066Sahrens zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 13315ad82045Snd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1332fa9e4066Sahrens } 1333fa9e4066Sahrens 1334fa9e4066Sahrens void 1335fa9e4066Sahrens zil_fini(void) 1336fa9e4066Sahrens { 1337fa9e4066Sahrens kmem_cache_destroy(zil_lwb_cache); 1338fa9e4066Sahrens } 1339fa9e4066Sahrens 1340e09fa4daSNeil Perrin void 1341e09fa4daSNeil Perrin zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1342e09fa4daSNeil Perrin { 1343e09fa4daSNeil Perrin zilog->zl_logbias = logbias; 1344e09fa4daSNeil Perrin } 1345e09fa4daSNeil Perrin 1346fa9e4066Sahrens zilog_t * 1347fa9e4066Sahrens zil_alloc(objset_t *os, zil_header_t *zh_phys) 1348fa9e4066Sahrens { 1349fa9e4066Sahrens zilog_t *zilog; 1350fa9e4066Sahrens 1351fa9e4066Sahrens zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1352fa9e4066Sahrens 1353fa9e4066Sahrens zilog->zl_header = zh_phys; 1354fa9e4066Sahrens zilog->zl_os = os; 1355fa9e4066Sahrens zilog->zl_spa = dmu_objset_spa(os); 1356fa9e4066Sahrens zilog->zl_dmu_pool = dmu_objset_pool(os); 1357d80c45e0Sbonwick zilog->zl_destroy_txg = TXG_INITIAL - 1; 1358e09fa4daSNeil Perrin zilog->zl_logbias = dmu_objset_logbias(os); 1359fa9e4066Sahrens 13605ad82045Snd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 13615ad82045Snd 1362fa9e4066Sahrens list_create(&zilog->zl_itx_list, sizeof (itx_t), 1363fa9e4066Sahrens offsetof(itx_t, itx_node)); 1364fa9e4066Sahrens 1365fa9e4066Sahrens list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1366fa9e4066Sahrens offsetof(lwb_t, lwb_node)); 1367fa9e4066Sahrens 136817f17c2dSbonwick mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 136917f17c2dSbonwick 137017f17c2dSbonwick avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 137117f17c2dSbonwick sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1372fa9e4066Sahrens 1373b7b97454Sperrin cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1374b7b97454Sperrin cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1375b7b97454Sperrin 1376fa9e4066Sahrens return (zilog); 1377fa9e4066Sahrens } 1378fa9e4066Sahrens 1379fa9e4066Sahrens void 1380fa9e4066Sahrens zil_free(zilog_t *zilog) 1381fa9e4066Sahrens { 1382fa9e4066Sahrens lwb_t *lwb; 1383fa9e4066Sahrens 1384fa9e4066Sahrens zilog->zl_stop_sync = 1; 1385fa9e4066Sahrens 1386fa9e4066Sahrens while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1387fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1388fa9e4066Sahrens if (lwb->lwb_buf != NULL) 1389fa9e4066Sahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1390fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1391fa9e4066Sahrens } 1392fa9e4066Sahrens list_destroy(&zilog->zl_lwb_list); 1393fa9e4066Sahrens 139417f17c2dSbonwick avl_destroy(&zilog->zl_vdev_tree); 139517f17c2dSbonwick mutex_destroy(&zilog->zl_vdev_lock); 1396fa9e4066Sahrens 1397fa9e4066Sahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1398fa9e4066Sahrens list_destroy(&zilog->zl_itx_list); 13995ad82045Snd mutex_destroy(&zilog->zl_lock); 1400fa9e4066Sahrens 1401b7b97454Sperrin cv_destroy(&zilog->zl_cv_writer); 1402b7b97454Sperrin cv_destroy(&zilog->zl_cv_suspend); 1403b7b97454Sperrin 1404fa9e4066Sahrens kmem_free(zilog, sizeof (zilog_t)); 1405fa9e4066Sahrens } 1406fa9e4066Sahrens 1407fa9e4066Sahrens /* 1408fa9e4066Sahrens * Open an intent log. 1409fa9e4066Sahrens */ 1410fa9e4066Sahrens zilog_t * 1411fa9e4066Sahrens zil_open(objset_t *os, zil_get_data_t *get_data) 1412fa9e4066Sahrens { 1413fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1414fa9e4066Sahrens 1415fa9e4066Sahrens zilog->zl_get_data = get_data; 1416fa9e4066Sahrens zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1417fa9e4066Sahrens 2, 2, TASKQ_PREPOPULATE); 1418fa9e4066Sahrens 1419fa9e4066Sahrens return (zilog); 1420fa9e4066Sahrens } 1421fa9e4066Sahrens 1422fa9e4066Sahrens /* 1423fa9e4066Sahrens * Close an intent log. 1424fa9e4066Sahrens */ 1425fa9e4066Sahrens void 1426fa9e4066Sahrens zil_close(zilog_t *zilog) 1427fa9e4066Sahrens { 1428d80c45e0Sbonwick /* 1429d80c45e0Sbonwick * If the log isn't already committed, mark the objset dirty 1430d80c45e0Sbonwick * (so zil_sync() will be called) and wait for that txg to sync. 1431d80c45e0Sbonwick */ 1432d80c45e0Sbonwick if (!zil_is_committed(zilog)) { 1433d80c45e0Sbonwick uint64_t txg; 1434d80c45e0Sbonwick dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1435b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 1436d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1437d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 1438d80c45e0Sbonwick dmu_tx_commit(tx); 1439d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, txg); 1440d80c45e0Sbonwick } 1441d80c45e0Sbonwick 1442fa9e4066Sahrens taskq_destroy(zilog->zl_clean_taskq); 1443fa9e4066Sahrens zilog->zl_clean_taskq = NULL; 1444fa9e4066Sahrens zilog->zl_get_data = NULL; 1445fa9e4066Sahrens 1446fa9e4066Sahrens zil_itx_clean(zilog); 1447fa9e4066Sahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1448fa9e4066Sahrens } 1449fa9e4066Sahrens 1450fa9e4066Sahrens /* 1451fa9e4066Sahrens * Suspend an intent log. While in suspended mode, we still honor 1452fa9e4066Sahrens * synchronous semantics, but we rely on txg_wait_synced() to do it. 1453fa9e4066Sahrens * We suspend the log briefly when taking a snapshot so that the snapshot 1454fa9e4066Sahrens * contains all the data it's supposed to, and has an empty intent log. 1455fa9e4066Sahrens */ 1456fa9e4066Sahrens int 1457fa9e4066Sahrens zil_suspend(zilog_t *zilog) 1458fa9e4066Sahrens { 1459d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1460fa9e4066Sahrens 1461fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 14623589c4f0SNeil Perrin if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1463fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1464fa9e4066Sahrens return (EBUSY); 1465fa9e4066Sahrens } 1466d80c45e0Sbonwick if (zilog->zl_suspend++ != 0) { 1467d80c45e0Sbonwick /* 1468d80c45e0Sbonwick * Someone else already began a suspend. 1469d80c45e0Sbonwick * Just wait for them to finish. 1470d80c45e0Sbonwick */ 1471d80c45e0Sbonwick while (zilog->zl_suspending) 1472d80c45e0Sbonwick cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1473d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1474d80c45e0Sbonwick return (0); 1475d80c45e0Sbonwick } 1476d80c45e0Sbonwick zilog->zl_suspending = B_TRUE; 1477fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1478fa9e4066Sahrens 1479b19a79ecSperrin zil_commit(zilog, UINT64_MAX, 0); 1480fa9e4066Sahrens 1481b19a79ecSperrin /* 1482b19a79ecSperrin * Wait for any in-flight log writes to complete. 1483b19a79ecSperrin */ 1484fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1485b19a79ecSperrin while (zilog->zl_writer) 1486b19a79ecSperrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1487fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1488fa9e4066Sahrens 1489d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1490d80c45e0Sbonwick 1491d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1492d80c45e0Sbonwick zilog->zl_suspending = B_FALSE; 1493d80c45e0Sbonwick cv_broadcast(&zilog->zl_cv_suspend); 1494d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1495fa9e4066Sahrens 1496fa9e4066Sahrens return (0); 1497fa9e4066Sahrens } 1498fa9e4066Sahrens 1499fa9e4066Sahrens void 1500fa9e4066Sahrens zil_resume(zilog_t *zilog) 1501fa9e4066Sahrens { 1502fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1503fa9e4066Sahrens ASSERT(zilog->zl_suspend != 0); 1504fa9e4066Sahrens zilog->zl_suspend--; 1505fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1506fa9e4066Sahrens } 1507fa9e4066Sahrens 1508fa9e4066Sahrens typedef struct zil_replay_arg { 1509fa9e4066Sahrens zil_replay_func_t **zr_replay; 1510fa9e4066Sahrens void *zr_arg; 1511fa9e4066Sahrens boolean_t zr_byteswap; 1512b24ab676SJeff Bonwick char *zr_lr; 1513fa9e4066Sahrens } zil_replay_arg_t; 1514fa9e4066Sahrens 1515b24ab676SJeff Bonwick static int 1516b24ab676SJeff Bonwick zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1517b24ab676SJeff Bonwick { 1518b24ab676SJeff Bonwick char name[MAXNAMELEN]; 1519b24ab676SJeff Bonwick 1520b24ab676SJeff Bonwick zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1521b24ab676SJeff Bonwick 1522b24ab676SJeff Bonwick dmu_objset_name(zilog->zl_os, name); 1523b24ab676SJeff Bonwick 1524b24ab676SJeff Bonwick cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1525b24ab676SJeff Bonwick "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1526b24ab676SJeff Bonwick (u_longlong_t)lr->lrc_seq, 1527b24ab676SJeff Bonwick (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1528b24ab676SJeff Bonwick (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1529b24ab676SJeff Bonwick 1530b24ab676SJeff Bonwick return (error); 1531b24ab676SJeff Bonwick } 1532b24ab676SJeff Bonwick 1533b24ab676SJeff Bonwick static int 1534fa9e4066Sahrens zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1535fa9e4066Sahrens { 1536fa9e4066Sahrens zil_replay_arg_t *zr = zra; 1537d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1538fa9e4066Sahrens uint64_t reclen = lr->lrc_reclen; 1539fa9e4066Sahrens uint64_t txtype = lr->lrc_txtype; 1540b24ab676SJeff Bonwick int error = 0; 1541fa9e4066Sahrens 1542b24ab676SJeff Bonwick zilog->zl_replaying_seq = lr->lrc_seq; 1543fa9e4066Sahrens 1544fa9e4066Sahrens if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1545b24ab676SJeff Bonwick return (0); 1546b24ab676SJeff Bonwick 1547b24ab676SJeff Bonwick if (lr->lrc_txg < claim_txg) /* already committed */ 1548b24ab676SJeff Bonwick return (0); 1549fa9e4066Sahrens 1550da6c28aaSamw /* Strip case-insensitive bit, still present in log record */ 1551da6c28aaSamw txtype &= ~TX_CI; 1552da6c28aaSamw 1553b24ab676SJeff Bonwick if (txtype == 0 || txtype >= TX_MAX_TYPE) 1554b24ab676SJeff Bonwick return (zil_replay_error(zilog, lr, EINVAL)); 1555b24ab676SJeff Bonwick 1556b24ab676SJeff Bonwick /* 1557b24ab676SJeff Bonwick * If this record type can be logged out of order, the object 1558b24ab676SJeff Bonwick * (lr_foid) may no longer exist. That's legitimate, not an error. 1559b24ab676SJeff Bonwick */ 1560b24ab676SJeff Bonwick if (TX_OOO(txtype)) { 1561b24ab676SJeff Bonwick error = dmu_object_info(zilog->zl_os, 1562b24ab676SJeff Bonwick ((lr_ooo_t *)lr)->lr_foid, NULL); 1563b24ab676SJeff Bonwick if (error == ENOENT || error == EEXIST) 1564b24ab676SJeff Bonwick return (0); 15651209a471SNeil Perrin } 15661209a471SNeil Perrin 1567fa9e4066Sahrens /* 1568fa9e4066Sahrens * Make a copy of the data so we can revise and extend it. 1569fa9e4066Sahrens */ 1570b24ab676SJeff Bonwick bcopy(lr, zr->zr_lr, reclen); 1571b24ab676SJeff Bonwick 1572b24ab676SJeff Bonwick /* 1573b24ab676SJeff Bonwick * If this is a TX_WRITE with a blkptr, suck in the data. 1574b24ab676SJeff Bonwick */ 1575b24ab676SJeff Bonwick if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1576b24ab676SJeff Bonwick error = zil_read_log_data(zilog, (lr_write_t *)lr, 1577b24ab676SJeff Bonwick zr->zr_lr + reclen); 1578b24ab676SJeff Bonwick if (error) 1579b24ab676SJeff Bonwick return (zil_replay_error(zilog, lr, error)); 1580b24ab676SJeff Bonwick } 1581fa9e4066Sahrens 1582fa9e4066Sahrens /* 1583fa9e4066Sahrens * The log block containing this lr may have been byteswapped 1584fa9e4066Sahrens * so that we can easily examine common fields like lrc_txtype. 1585b24ab676SJeff Bonwick * However, the log is a mix of different record types, and only the 1586fa9e4066Sahrens * replay vectors know how to byteswap their records. Therefore, if 1587fa9e4066Sahrens * the lr was byteswapped, undo it before invoking the replay vector. 1588fa9e4066Sahrens */ 1589fa9e4066Sahrens if (zr->zr_byteswap) 1590b24ab676SJeff Bonwick byteswap_uint64_array(zr->zr_lr, reclen); 1591fa9e4066Sahrens 1592fa9e4066Sahrens /* 1593fa9e4066Sahrens * We must now do two things atomically: replay this log record, 15941209a471SNeil Perrin * and update the log header sequence number to reflect the fact that 15951209a471SNeil Perrin * we did so. At the end of each replay function the sequence number 15961209a471SNeil Perrin * is updated if we are in replay mode. 1597fa9e4066Sahrens */ 1598b24ab676SJeff Bonwick error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 1599b24ab676SJeff Bonwick if (error) { 160067bd71c6Sperrin /* 160167bd71c6Sperrin * The DMU's dnode layer doesn't see removes until the txg 160267bd71c6Sperrin * commits, so a subsequent claim can spuriously fail with 16031209a471SNeil Perrin * EEXIST. So if we receive any error we try syncing out 1604b24ab676SJeff Bonwick * any removes then retry the transaction. Note that we 1605b24ab676SJeff Bonwick * specify B_FALSE for byteswap now, so we don't do it twice. 160667bd71c6Sperrin */ 1607b24ab676SJeff Bonwick txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1608b24ab676SJeff Bonwick error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 1609b24ab676SJeff Bonwick if (error) 1610b24ab676SJeff Bonwick return (zil_replay_error(zilog, lr, error)); 1611fa9e4066Sahrens } 1612b24ab676SJeff Bonwick return (0); 161367bd71c6Sperrin } 1614fa9e4066Sahrens 161567bd71c6Sperrin /* ARGSUSED */ 1616b24ab676SJeff Bonwick static int 161767bd71c6Sperrin zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 161867bd71c6Sperrin { 161967bd71c6Sperrin zilog->zl_replay_blks++; 1620b24ab676SJeff Bonwick 1621b24ab676SJeff Bonwick return (0); 1622fa9e4066Sahrens } 1623fa9e4066Sahrens 1624fa9e4066Sahrens /* 162513f5297eSperrin * If this dataset has a non-empty intent log, replay it and destroy it. 1626fa9e4066Sahrens */ 1627fa9e4066Sahrens void 16281209a471SNeil Perrin zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1629fa9e4066Sahrens { 1630fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1631d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1632d80c45e0Sbonwick zil_replay_arg_t zr; 163313f5297eSperrin 16343589c4f0SNeil Perrin if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1635d80c45e0Sbonwick zil_destroy(zilog, B_TRUE); 163613f5297eSperrin return; 163713f5297eSperrin } 1638fa9e4066Sahrens 1639fa9e4066Sahrens zr.zr_replay = replay_func; 1640fa9e4066Sahrens zr.zr_arg = arg; 1641d80c45e0Sbonwick zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1642b24ab676SJeff Bonwick zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1643fa9e4066Sahrens 1644fa9e4066Sahrens /* 1645fa9e4066Sahrens * Wait for in-progress removes to sync before starting replay. 1646fa9e4066Sahrens */ 1647fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1648fa9e4066Sahrens 16491209a471SNeil Perrin zilog->zl_replay = B_TRUE; 1650d3d50737SRafael Vanoni zilog->zl_replay_time = ddi_get_lbolt(); 165167bd71c6Sperrin ASSERT(zilog->zl_replay_blks == 0); 165267bd71c6Sperrin (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1653d80c45e0Sbonwick zh->zh_claim_txg); 1654b24ab676SJeff Bonwick kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 1655fa9e4066Sahrens 1656d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1657a4611edeSahrens txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 16581209a471SNeil Perrin zilog->zl_replay = B_FALSE; 1659fa9e4066Sahrens } 1660436b2950Sperrin 1661b24ab676SJeff Bonwick boolean_t 1662b24ab676SJeff Bonwick zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 1663436b2950Sperrin { 1664b24ab676SJeff Bonwick if (zilog == NULL) 1665b24ab676SJeff Bonwick return (B_TRUE); 1666436b2950Sperrin 1667b24ab676SJeff Bonwick if (zilog->zl_replay) { 1668b24ab676SJeff Bonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1669b24ab676SJeff Bonwick zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 1670b24ab676SJeff Bonwick zilog->zl_replaying_seq; 1671b24ab676SJeff Bonwick return (B_TRUE); 1672b19a79ecSperrin } 1673b19a79ecSperrin 1674b24ab676SJeff Bonwick return (B_FALSE); 1675436b2950Sperrin } 1676e6ca193dSGeorge Wilson 1677e6ca193dSGeorge Wilson /* ARGSUSED */ 1678e6ca193dSGeorge Wilson int 1679*fd136879SMatthew Ahrens zil_vdev_offline(const char *osname, void *arg) 1680e6ca193dSGeorge Wilson { 1681e6ca193dSGeorge Wilson objset_t *os; 1682e6ca193dSGeorge Wilson zilog_t *zilog; 1683e6ca193dSGeorge Wilson int error; 1684e6ca193dSGeorge Wilson 1685503ad85cSMatthew Ahrens error = dmu_objset_hold(osname, FTAG, &os); 1686e6ca193dSGeorge Wilson if (error) 1687e6ca193dSGeorge Wilson return (error); 1688e6ca193dSGeorge Wilson 1689e6ca193dSGeorge Wilson zilog = dmu_objset_zil(os); 1690e6ca193dSGeorge Wilson if (zil_suspend(zilog) != 0) 1691e6ca193dSGeorge Wilson error = EEXIST; 1692e6ca193dSGeorge Wilson else 1693e6ca193dSGeorge Wilson zil_resume(zilog); 1694503ad85cSMatthew Ahrens dmu_objset_rele(os, FTAG); 1695e6ca193dSGeorge Wilson return (error); 1696e6ca193dSGeorge Wilson } 1697