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 /* 2255da60b9SMark J Musante * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23*c9ba2a43SEric Schrock * Copyright (c) 2011 by Delphix. All rights reserved. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 2655da60b9SMark J Musante /* Portions Copyright 2010 Robert Milkowski */ 2755da60b9SMark J Musante 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> 384b964adaSGeorge Wilson #include <sys/vdev_impl.h> 39d63d470bSgw #include <sys/dmu_tx.h> 403f9d6ad7SLin Ling #include <sys/dsl_pool.h> 41fa9e4066Sahrens 42fa9e4066Sahrens /* 43fa9e4066Sahrens * The zfs intent log (ZIL) saves transaction records of system calls 44fa9e4066Sahrens * that change the file system in memory with enough information 45fa9e4066Sahrens * to be able to replay them. These are stored in memory until 46fa9e4066Sahrens * either the DMU transaction group (txg) commits them to the stable pool 47fa9e4066Sahrens * and they can be discarded, or they are flushed to the stable log 48fa9e4066Sahrens * (also in the pool) due to a fsync, O_DSYNC or other synchronous 49fa9e4066Sahrens * requirement. In the event of a panic or power fail then those log 50fa9e4066Sahrens * records (transactions) are replayed. 51fa9e4066Sahrens * 52fa9e4066Sahrens * There is one ZIL per file system. Its on-disk (pool) format consists 53fa9e4066Sahrens * of 3 parts: 54fa9e4066Sahrens * 55fa9e4066Sahrens * - ZIL header 56fa9e4066Sahrens * - ZIL blocks 57fa9e4066Sahrens * - ZIL records 58fa9e4066Sahrens * 59fa9e4066Sahrens * A log record holds a system call transaction. Log blocks can 60fa9e4066Sahrens * hold many log records and the blocks are chained together. 61fa9e4066Sahrens * Each ZIL block contains a block pointer (blkptr_t) to the next 62fa9e4066Sahrens * ZIL block in the chain. The ZIL header points to the first 63fa9e4066Sahrens * block in the chain. Note there is not a fixed place in the pool 64fa9e4066Sahrens * to hold blocks. They are dynamically allocated and freed as 65fa9e4066Sahrens * needed from the blocks available. Figure X shows the ZIL structure: 66fa9e4066Sahrens */ 67fa9e4066Sahrens 68fa9e4066Sahrens /* 69416e0cd8Sek * This global ZIL switch affects all pools 70fa9e4066Sahrens */ 7155da60b9SMark J Musante int zil_replay_disable = 0; /* disable intent logging replay */ 72416e0cd8Sek 73416e0cd8Sek /* 74416e0cd8Sek * Tunable parameter for debugging or performance analysis. Setting 75416e0cd8Sek * zfs_nocacheflush will cause corruption on power loss if a volatile 76416e0cd8Sek * out-of-order write cache is enabled. 77416e0cd8Sek */ 78416e0cd8Sek boolean_t zfs_nocacheflush = B_FALSE; 79fa9e4066Sahrens 80fa9e4066Sahrens static kmem_cache_t *zil_lwb_cache; 81fa9e4066Sahrens 8291de656bSNeil Perrin static void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 838f18d1faSGeorge Wilson 846e1f5caaSNeil Perrin #define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 856e1f5caaSNeil Perrin sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 866e1f5caaSNeil Perrin 876e1f5caaSNeil Perrin 885002558fSNeil Perrin /* 895002558fSNeil Perrin * ziltest is by and large an ugly hack, but very useful in 905002558fSNeil Perrin * checking replay without tedious work. 915002558fSNeil Perrin * When running ziltest we want to keep all itx's and so maintain 925002558fSNeil Perrin * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 935002558fSNeil Perrin * We subtract TXG_CONCURRENT_STATES to allow for common code. 945002558fSNeil Perrin */ 955002558fSNeil Perrin #define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 965002558fSNeil Perrin 97fa9e4066Sahrens static int 98b24ab676SJeff Bonwick zil_bp_compare(const void *x1, const void *x2) 99fa9e4066Sahrens { 100b24ab676SJeff Bonwick const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 101b24ab676SJeff Bonwick const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 102fa9e4066Sahrens 103fa9e4066Sahrens if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 104fa9e4066Sahrens return (-1); 105fa9e4066Sahrens if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 106fa9e4066Sahrens return (1); 107fa9e4066Sahrens 108fa9e4066Sahrens if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 109fa9e4066Sahrens return (-1); 110fa9e4066Sahrens if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 111fa9e4066Sahrens return (1); 112fa9e4066Sahrens 113fa9e4066Sahrens return (0); 114fa9e4066Sahrens } 115fa9e4066Sahrens 116fa9e4066Sahrens static void 117b24ab676SJeff Bonwick zil_bp_tree_init(zilog_t *zilog) 118fa9e4066Sahrens { 119b24ab676SJeff Bonwick avl_create(&zilog->zl_bp_tree, zil_bp_compare, 120b24ab676SJeff Bonwick sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 121fa9e4066Sahrens } 122fa9e4066Sahrens 123fa9e4066Sahrens static void 124b24ab676SJeff Bonwick zil_bp_tree_fini(zilog_t *zilog) 125fa9e4066Sahrens { 126b24ab676SJeff Bonwick avl_tree_t *t = &zilog->zl_bp_tree; 127b24ab676SJeff Bonwick zil_bp_node_t *zn; 128fa9e4066Sahrens void *cookie = NULL; 129fa9e4066Sahrens 130fa9e4066Sahrens while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 131b24ab676SJeff Bonwick kmem_free(zn, sizeof (zil_bp_node_t)); 132fa9e4066Sahrens 133fa9e4066Sahrens avl_destroy(t); 134fa9e4066Sahrens } 135fa9e4066Sahrens 136b24ab676SJeff Bonwick int 137b24ab676SJeff Bonwick zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 138fa9e4066Sahrens { 139b24ab676SJeff Bonwick avl_tree_t *t = &zilog->zl_bp_tree; 140b24ab676SJeff Bonwick const dva_t *dva = BP_IDENTITY(bp); 141b24ab676SJeff Bonwick zil_bp_node_t *zn; 142fa9e4066Sahrens avl_index_t where; 143fa9e4066Sahrens 144fa9e4066Sahrens if (avl_find(t, dva, &where) != NULL) 145fa9e4066Sahrens return (EEXIST); 146fa9e4066Sahrens 147b24ab676SJeff Bonwick zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 148fa9e4066Sahrens zn->zn_dva = *dva; 149fa9e4066Sahrens avl_insert(t, zn, where); 150fa9e4066Sahrens 151fa9e4066Sahrens return (0); 152fa9e4066Sahrens } 153fa9e4066Sahrens 154d80c45e0Sbonwick static zil_header_t * 155d80c45e0Sbonwick zil_header_in_syncing_context(zilog_t *zilog) 156d80c45e0Sbonwick { 157d80c45e0Sbonwick return ((zil_header_t *)zilog->zl_header); 158d80c45e0Sbonwick } 159d80c45e0Sbonwick 160d80c45e0Sbonwick static void 161d80c45e0Sbonwick zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 162d80c45e0Sbonwick { 163d80c45e0Sbonwick zio_cksum_t *zc = &bp->blk_cksum; 164d80c45e0Sbonwick 165d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 166d80c45e0Sbonwick zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 167d80c45e0Sbonwick zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 168d80c45e0Sbonwick zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 169d80c45e0Sbonwick } 170d80c45e0Sbonwick 171fa9e4066Sahrens /* 172b24ab676SJeff Bonwick * Read a log block and make sure it's valid. 173fa9e4066Sahrens */ 174fa9e4066Sahrens static int 1756e1f5caaSNeil Perrin zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 1766e1f5caaSNeil Perrin char **end) 177fa9e4066Sahrens { 178b24ab676SJeff Bonwick enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 17913506d1eSmaybee uint32_t aflags = ARC_WAIT; 180b24ab676SJeff Bonwick arc_buf_t *abuf = NULL; 181b24ab676SJeff Bonwick zbookmark_t zb; 182fa9e4066Sahrens int error; 183fa9e4066Sahrens 184b24ab676SJeff Bonwick if (zilog->zl_header->zh_claim_txg == 0) 185b24ab676SJeff Bonwick zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 186ea8dc4b6Seschrock 187b24ab676SJeff Bonwick if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 188b24ab676SJeff Bonwick zio_flags |= ZIO_FLAG_SPECULATIVE; 189fa9e4066Sahrens 190b24ab676SJeff Bonwick SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 191b24ab676SJeff Bonwick ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 192b24ab676SJeff Bonwick 1933f9d6ad7SLin Ling error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 194b24ab676SJeff Bonwick ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 195fa9e4066Sahrens 196d80c45e0Sbonwick if (error == 0) { 197d80c45e0Sbonwick zio_cksum_t cksum = bp->blk_cksum; 198fa9e4066Sahrens 199d80c45e0Sbonwick /* 200f5e6e722SNeil Perrin * Validate the checksummed log block. 201f5e6e722SNeil Perrin * 202d80c45e0Sbonwick * Sequence numbers should be... sequential. The checksum 203d80c45e0Sbonwick * verifier for the next block should be bp's checksum plus 1. 204f5e6e722SNeil Perrin * 205f5e6e722SNeil Perrin * Also check the log chain linkage and size used. 206d80c45e0Sbonwick */ 207d80c45e0Sbonwick cksum.zc_word[ZIL_ZC_SEQ]++; 208d80c45e0Sbonwick 2096e1f5caaSNeil Perrin if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 2106e1f5caaSNeil Perrin zil_chain_t *zilc = abuf->b_data; 2116e1f5caaSNeil Perrin char *lr = (char *)(zilc + 1); 2126e1f5caaSNeil Perrin uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 2136e1f5caaSNeil Perrin 2146e1f5caaSNeil Perrin if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 2156e1f5caaSNeil Perrin sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 2166e1f5caaSNeil Perrin error = ECKSUM; 2176e1f5caaSNeil Perrin } else { 2186e1f5caaSNeil Perrin bcopy(lr, dst, len); 2196e1f5caaSNeil Perrin *end = (char *)dst + len; 2206e1f5caaSNeil Perrin *nbp = zilc->zc_next_blk; 2216e1f5caaSNeil Perrin } 2226e1f5caaSNeil Perrin } else { 2236e1f5caaSNeil Perrin char *lr = abuf->b_data; 2246e1f5caaSNeil Perrin uint64_t size = BP_GET_LSIZE(bp); 2256e1f5caaSNeil Perrin zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 2266e1f5caaSNeil Perrin 2276e1f5caaSNeil Perrin if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 2286e1f5caaSNeil Perrin sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 2296e1f5caaSNeil Perrin (zilc->zc_nused > (size - sizeof (*zilc)))) { 2306e1f5caaSNeil Perrin error = ECKSUM; 2316e1f5caaSNeil Perrin } else { 2326e1f5caaSNeil Perrin bcopy(lr, dst, zilc->zc_nused); 2336e1f5caaSNeil Perrin *end = (char *)dst + zilc->zc_nused; 2346e1f5caaSNeil Perrin *nbp = zilc->zc_next_blk; 2356e1f5caaSNeil Perrin } 2366e1f5caaSNeil Perrin } 237fa9e4066Sahrens 238b24ab676SJeff Bonwick VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 239fa9e4066Sahrens } 240fa9e4066Sahrens 241b24ab676SJeff Bonwick return (error); 242b24ab676SJeff Bonwick } 243b24ab676SJeff Bonwick 244b24ab676SJeff Bonwick /* 245b24ab676SJeff Bonwick * Read a TX_WRITE log data block. 246b24ab676SJeff Bonwick */ 247b24ab676SJeff Bonwick static int 248b24ab676SJeff Bonwick zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 249b24ab676SJeff Bonwick { 250b24ab676SJeff Bonwick enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 251b24ab676SJeff Bonwick const blkptr_t *bp = &lr->lr_blkptr; 252b24ab676SJeff Bonwick uint32_t aflags = ARC_WAIT; 253b24ab676SJeff Bonwick arc_buf_t *abuf = NULL; 254b24ab676SJeff Bonwick zbookmark_t zb; 255b24ab676SJeff Bonwick int error; 256b24ab676SJeff Bonwick 257b24ab676SJeff Bonwick if (BP_IS_HOLE(bp)) { 258b24ab676SJeff Bonwick if (wbuf != NULL) 259b24ab676SJeff Bonwick bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 260b24ab676SJeff Bonwick return (0); 261b24ab676SJeff Bonwick } 262b24ab676SJeff Bonwick 263b24ab676SJeff Bonwick if (zilog->zl_header->zh_claim_txg == 0) 264b24ab676SJeff Bonwick zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 265b24ab676SJeff Bonwick 266b24ab676SJeff Bonwick SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 267b24ab676SJeff Bonwick ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 268b24ab676SJeff Bonwick 269b24ab676SJeff Bonwick error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 270b24ab676SJeff Bonwick ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 271b24ab676SJeff Bonwick 272b24ab676SJeff Bonwick if (error == 0) { 273b24ab676SJeff Bonwick if (wbuf != NULL) 274b24ab676SJeff Bonwick bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 275b24ab676SJeff Bonwick (void) arc_buf_remove_ref(abuf, &abuf); 276b24ab676SJeff Bonwick } 277fa9e4066Sahrens 278d80c45e0Sbonwick return (error); 279fa9e4066Sahrens } 280fa9e4066Sahrens 281fa9e4066Sahrens /* 282fa9e4066Sahrens * Parse the intent log, and call parse_func for each valid record within. 283fa9e4066Sahrens */ 284b24ab676SJeff Bonwick int 285fa9e4066Sahrens zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 286fa9e4066Sahrens zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 287fa9e4066Sahrens { 288d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 289b24ab676SJeff Bonwick boolean_t claimed = !!zh->zh_claim_txg; 290b24ab676SJeff Bonwick uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 291b24ab676SJeff Bonwick uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 292b24ab676SJeff Bonwick uint64_t max_blk_seq = 0; 293b24ab676SJeff Bonwick uint64_t max_lr_seq = 0; 294b24ab676SJeff Bonwick uint64_t blk_count = 0; 295b24ab676SJeff Bonwick uint64_t lr_count = 0; 296b24ab676SJeff Bonwick blkptr_t blk, next_blk; 297fa9e4066Sahrens char *lrbuf, *lrp; 298b24ab676SJeff Bonwick int error = 0; 299fa9e4066Sahrens 300b24ab676SJeff Bonwick /* 301b24ab676SJeff Bonwick * Old logs didn't record the maximum zh_claim_lr_seq. 302b24ab676SJeff Bonwick */ 303b24ab676SJeff Bonwick if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 304b24ab676SJeff Bonwick claim_lr_seq = UINT64_MAX; 305fa9e4066Sahrens 306fa9e4066Sahrens /* 307fa9e4066Sahrens * Starting at the block pointed to by zh_log we read the log chain. 308fa9e4066Sahrens * For each block in the chain we strongly check that block to 309fa9e4066Sahrens * ensure its validity. We stop when an invalid block is found. 310fa9e4066Sahrens * For each block pointer in the chain we call parse_blk_func(). 311fa9e4066Sahrens * For each record in each valid block we call parse_lr_func(). 312d80c45e0Sbonwick * If the log has been claimed, stop if we encounter a sequence 313d80c45e0Sbonwick * number greater than the highest claimed sequence number. 314fa9e4066Sahrens */ 315b24ab676SJeff Bonwick lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 316b24ab676SJeff Bonwick zil_bp_tree_init(zilog); 317d80c45e0Sbonwick 318b24ab676SJeff Bonwick for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 319b24ab676SJeff Bonwick uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 320b24ab676SJeff Bonwick int reclen; 3216e1f5caaSNeil Perrin char *end; 322d80c45e0Sbonwick 323b24ab676SJeff Bonwick if (blk_seq > claim_blk_seq) 324b24ab676SJeff Bonwick break; 325b24ab676SJeff Bonwick if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 326b24ab676SJeff Bonwick break; 3276e1f5caaSNeil Perrin ASSERT3U(max_blk_seq, <, blk_seq); 328b24ab676SJeff Bonwick max_blk_seq = blk_seq; 329b24ab676SJeff Bonwick blk_count++; 330fa9e4066Sahrens 331b24ab676SJeff Bonwick if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 332b24ab676SJeff Bonwick break; 333fa9e4066Sahrens 3346e1f5caaSNeil Perrin error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 335fa9e4066Sahrens if (error) 336fa9e4066Sahrens break; 337fa9e4066Sahrens 3386e1f5caaSNeil Perrin for (lrp = lrbuf; lrp < end; lrp += reclen) { 339fa9e4066Sahrens lr_t *lr = (lr_t *)lrp; 340fa9e4066Sahrens reclen = lr->lrc_reclen; 341fa9e4066Sahrens ASSERT3U(reclen, >=, sizeof (lr_t)); 342b24ab676SJeff Bonwick if (lr->lrc_seq > claim_lr_seq) 343b24ab676SJeff Bonwick goto done; 344b24ab676SJeff Bonwick if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 345b24ab676SJeff Bonwick goto done; 3466e1f5caaSNeil Perrin ASSERT3U(max_lr_seq, <, lr->lrc_seq); 347b24ab676SJeff Bonwick max_lr_seq = lr->lrc_seq; 348b24ab676SJeff Bonwick lr_count++; 349fa9e4066Sahrens } 350fa9e4066Sahrens } 351b24ab676SJeff Bonwick done: 352b24ab676SJeff Bonwick zilog->zl_parse_error = error; 353b24ab676SJeff Bonwick zilog->zl_parse_blk_seq = max_blk_seq; 354b24ab676SJeff Bonwick zilog->zl_parse_lr_seq = max_lr_seq; 355b24ab676SJeff Bonwick zilog->zl_parse_blk_count = blk_count; 356b24ab676SJeff Bonwick zilog->zl_parse_lr_count = lr_count; 357b24ab676SJeff Bonwick 358b24ab676SJeff Bonwick ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 359b24ab676SJeff Bonwick (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 360d80c45e0Sbonwick 361b24ab676SJeff Bonwick zil_bp_tree_fini(zilog); 362b24ab676SJeff Bonwick zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 363b24ab676SJeff Bonwick 364b24ab676SJeff Bonwick return (error); 365fa9e4066Sahrens } 366fa9e4066Sahrens 367b24ab676SJeff Bonwick static int 368fa9e4066Sahrens zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 369fa9e4066Sahrens { 370fa9e4066Sahrens /* 371fa9e4066Sahrens * Claim log block if not already committed and not already claimed. 372b24ab676SJeff Bonwick * If tx == NULL, just verify that the block is claimable. 373fa9e4066Sahrens */ 374b24ab676SJeff Bonwick if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 375b24ab676SJeff Bonwick return (0); 376b24ab676SJeff Bonwick 377b24ab676SJeff Bonwick return (zio_wait(zio_claim(NULL, zilog->zl_spa, 378b24ab676SJeff Bonwick tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 379b24ab676SJeff Bonwick ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 380fa9e4066Sahrens } 381fa9e4066Sahrens 382b24ab676SJeff Bonwick static int 383fa9e4066Sahrens zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 384fa9e4066Sahrens { 385b24ab676SJeff Bonwick lr_write_t *lr = (lr_write_t *)lrc; 386b24ab676SJeff Bonwick int error; 387b24ab676SJeff Bonwick 388b24ab676SJeff Bonwick if (lrc->lrc_txtype != TX_WRITE) 389b24ab676SJeff Bonwick return (0); 390b24ab676SJeff Bonwick 391b24ab676SJeff Bonwick /* 392b24ab676SJeff Bonwick * If the block is not readable, don't claim it. This can happen 393b24ab676SJeff Bonwick * in normal operation when a log block is written to disk before 394b24ab676SJeff Bonwick * some of the dmu_sync() blocks it points to. In this case, the 395b24ab676SJeff Bonwick * transaction cannot have been committed to anyone (we would have 396b24ab676SJeff Bonwick * waited for all writes to be stable first), so it is semantically 397b24ab676SJeff Bonwick * correct to declare this the end of the log. 398b24ab676SJeff Bonwick */ 399b24ab676SJeff Bonwick if (lr->lr_blkptr.blk_birth >= first_txg && 400b24ab676SJeff Bonwick (error = zil_read_log_data(zilog, lr, NULL)) != 0) 401b24ab676SJeff Bonwick return (error); 402b24ab676SJeff Bonwick return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 403fa9e4066Sahrens } 404fa9e4066Sahrens 405fa9e4066Sahrens /* ARGSUSED */ 406b24ab676SJeff Bonwick static int 407fa9e4066Sahrens zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 408fa9e4066Sahrens { 409b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 410b24ab676SJeff Bonwick 411b24ab676SJeff Bonwick return (0); 412fa9e4066Sahrens } 413fa9e4066Sahrens 414b24ab676SJeff Bonwick static int 415fa9e4066Sahrens zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 416fa9e4066Sahrens { 417b24ab676SJeff Bonwick lr_write_t *lr = (lr_write_t *)lrc; 418b24ab676SJeff Bonwick blkptr_t *bp = &lr->lr_blkptr; 419b24ab676SJeff Bonwick 420fa9e4066Sahrens /* 421fa9e4066Sahrens * If we previously claimed it, we need to free it. 422fa9e4066Sahrens */ 423b24ab676SJeff Bonwick if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 424b24ab676SJeff Bonwick bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 425b24ab676SJeff Bonwick zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 426b24ab676SJeff Bonwick 427b24ab676SJeff Bonwick return (0); 428fa9e4066Sahrens } 429fa9e4066Sahrens 4306e1f5caaSNeil Perrin static lwb_t * 4316e1f5caaSNeil Perrin zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 4326e1f5caaSNeil Perrin { 4336e1f5caaSNeil Perrin lwb_t *lwb; 4346e1f5caaSNeil Perrin 4356e1f5caaSNeil Perrin lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 4366e1f5caaSNeil Perrin lwb->lwb_zilog = zilog; 4376e1f5caaSNeil Perrin lwb->lwb_blk = *bp; 4386e1f5caaSNeil Perrin lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 4396e1f5caaSNeil Perrin lwb->lwb_max_txg = txg; 4406e1f5caaSNeil Perrin lwb->lwb_zio = NULL; 4416e1f5caaSNeil Perrin lwb->lwb_tx = NULL; 4426e1f5caaSNeil Perrin if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 4436e1f5caaSNeil Perrin lwb->lwb_nused = sizeof (zil_chain_t); 4446e1f5caaSNeil Perrin lwb->lwb_sz = BP_GET_LSIZE(bp); 4456e1f5caaSNeil Perrin } else { 4466e1f5caaSNeil Perrin lwb->lwb_nused = 0; 4476e1f5caaSNeil Perrin lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 4486e1f5caaSNeil Perrin } 4496e1f5caaSNeil Perrin 4506e1f5caaSNeil Perrin mutex_enter(&zilog->zl_lock); 4516e1f5caaSNeil Perrin list_insert_tail(&zilog->zl_lwb_list, lwb); 4526e1f5caaSNeil Perrin mutex_exit(&zilog->zl_lock); 4536e1f5caaSNeil Perrin 4546e1f5caaSNeil Perrin return (lwb); 4556e1f5caaSNeil Perrin } 4566e1f5caaSNeil Perrin 457fa9e4066Sahrens /* 458fa9e4066Sahrens * Create an on-disk intent log. 459fa9e4066Sahrens */ 4606e1f5caaSNeil Perrin static lwb_t * 461fa9e4066Sahrens zil_create(zilog_t *zilog) 462fa9e4066Sahrens { 463d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 4646e1f5caaSNeil Perrin lwb_t *lwb = NULL; 465d80c45e0Sbonwick uint64_t txg = 0; 466d80c45e0Sbonwick dmu_tx_t *tx = NULL; 467fa9e4066Sahrens blkptr_t blk; 468d80c45e0Sbonwick int error = 0; 469fa9e4066Sahrens 470fa9e4066Sahrens /* 471d80c45e0Sbonwick * Wait for any previous destroy to complete. 472fa9e4066Sahrens */ 473d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 474d80c45e0Sbonwick 475d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 476d80c45e0Sbonwick ASSERT(zh->zh_replay_seq == 0); 477d80c45e0Sbonwick 478d80c45e0Sbonwick blk = zh->zh_log; 479fa9e4066Sahrens 480fa9e4066Sahrens /* 4816e1f5caaSNeil Perrin * Allocate an initial log block if: 4826e1f5caaSNeil Perrin * - there isn't one already 4836e1f5caaSNeil Perrin * - the existing block is the wrong endianess 484fa9e4066Sahrens */ 485899217ddSNeil Perrin if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 486d80c45e0Sbonwick tx = dmu_tx_create(zilog->zl_os); 487b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 488d80c45e0Sbonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 489d80c45e0Sbonwick txg = dmu_tx_get_txg(tx); 490d80c45e0Sbonwick 491899217ddSNeil Perrin if (!BP_IS_HOLE(&blk)) { 492b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, txg, &blk); 493899217ddSNeil Perrin BP_ZERO(&blk); 494899217ddSNeil Perrin } 495899217ddSNeil Perrin 496b24ab676SJeff Bonwick error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 497b24ab676SJeff Bonwick ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 498d80c45e0Sbonwick 499d80c45e0Sbonwick if (error == 0) 500d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 50113f5297eSperrin } 502fa9e4066Sahrens 503d80c45e0Sbonwick /* 504d80c45e0Sbonwick * Allocate a log write buffer (lwb) for the first log block. 505d80c45e0Sbonwick */ 5066e1f5caaSNeil Perrin if (error == 0) 5076e1f5caaSNeil Perrin lwb = zil_alloc_lwb(zilog, &blk, txg); 508fa9e4066Sahrens 509d80c45e0Sbonwick /* 510d80c45e0Sbonwick * If we just allocated the first log block, commit our transaction 511d80c45e0Sbonwick * and wait for zil_sync() to stuff the block poiner into zh_log. 512d80c45e0Sbonwick * (zh is part of the MOS, so we cannot modify it in open context.) 513d80c45e0Sbonwick */ 514d80c45e0Sbonwick if (tx != NULL) { 515d80c45e0Sbonwick dmu_tx_commit(tx); 51613f5297eSperrin txg_wait_synced(zilog->zl_dmu_pool, txg); 517d80c45e0Sbonwick } 518d80c45e0Sbonwick 519d80c45e0Sbonwick ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 5206e1f5caaSNeil Perrin 5216e1f5caaSNeil Perrin return (lwb); 522fa9e4066Sahrens } 523fa9e4066Sahrens 524fa9e4066Sahrens /* 525fa9e4066Sahrens * In one tx, free all log blocks and clear the log header. 526d80c45e0Sbonwick * If keep_first is set, then we're replaying a log with no content. 527d80c45e0Sbonwick * We want to keep the first block, however, so that the first 528d80c45e0Sbonwick * synchronous transaction doesn't require a txg_wait_synced() 529d80c45e0Sbonwick * in zil_create(). We don't need to txg_wait_synced() here either 530d80c45e0Sbonwick * when keep_first is set, because both zil_create() and zil_destroy() 531d80c45e0Sbonwick * will wait for any in-progress destroys to complete. 532fa9e4066Sahrens */ 533fa9e4066Sahrens void 534d80c45e0Sbonwick zil_destroy(zilog_t *zilog, boolean_t keep_first) 535fa9e4066Sahrens { 536d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 537d80c45e0Sbonwick lwb_t *lwb; 538fa9e4066Sahrens dmu_tx_t *tx; 539fa9e4066Sahrens uint64_t txg; 540fa9e4066Sahrens 541d80c45e0Sbonwick /* 542d80c45e0Sbonwick * Wait for any previous destroy to complete. 543d80c45e0Sbonwick */ 544d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 545fa9e4066Sahrens 546b24ab676SJeff Bonwick zilog->zl_old_header = *zh; /* debugging aid */ 547b24ab676SJeff Bonwick 548d80c45e0Sbonwick if (BP_IS_HOLE(&zh->zh_log)) 549fa9e4066Sahrens return; 550fa9e4066Sahrens 551fa9e4066Sahrens tx = dmu_tx_create(zilog->zl_os); 552b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 553fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 554fa9e4066Sahrens txg = dmu_tx_get_txg(tx); 555fa9e4066Sahrens 556d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 557d80c45e0Sbonwick 558d80c45e0Sbonwick ASSERT3U(zilog->zl_destroy_txg, <, txg); 559fa9e4066Sahrens zilog->zl_destroy_txg = txg; 560b24ab676SJeff Bonwick zilog->zl_keep_first = keep_first; 561d80c45e0Sbonwick 562d80c45e0Sbonwick if (!list_is_empty(&zilog->zl_lwb_list)) { 563d80c45e0Sbonwick ASSERT(zh->zh_claim_txg == 0); 564*c9ba2a43SEric Schrock VERIFY(!keep_first); 565d80c45e0Sbonwick while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 566d80c45e0Sbonwick list_remove(&zilog->zl_lwb_list, lwb); 567d80c45e0Sbonwick if (lwb->lwb_buf != NULL) 568d80c45e0Sbonwick zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 569b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 570d80c45e0Sbonwick kmem_cache_free(zil_lwb_cache, lwb); 571d80c45e0Sbonwick } 572b24ab676SJeff Bonwick } else if (!keep_first) { 573b24ab676SJeff Bonwick (void) zil_parse(zilog, zil_free_log_block, 574b24ab676SJeff Bonwick zil_free_log_record, tx, zh->zh_claim_txg); 575d80c45e0Sbonwick } 576b19a79ecSperrin mutex_exit(&zilog->zl_lock); 577fa9e4066Sahrens 578fa9e4066Sahrens dmu_tx_commit(tx); 579fa9e4066Sahrens } 580fa9e4066Sahrens 5811d452cf5Sahrens int 582fd136879SMatthew Ahrens zil_claim(const char *osname, void *txarg) 583fa9e4066Sahrens { 584fa9e4066Sahrens dmu_tx_t *tx = txarg; 585fa9e4066Sahrens uint64_t first_txg = dmu_tx_get_txg(tx); 586fa9e4066Sahrens zilog_t *zilog; 587fa9e4066Sahrens zil_header_t *zh; 588fa9e4066Sahrens objset_t *os; 589fa9e4066Sahrens int error; 590fa9e4066Sahrens 591503ad85cSMatthew Ahrens error = dmu_objset_hold(osname, FTAG, &os); 592fa9e4066Sahrens if (error) { 593b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 5941d452cf5Sahrens return (0); 595fa9e4066Sahrens } 596fa9e4066Sahrens 597fa9e4066Sahrens zilog = dmu_objset_zil(os); 598d80c45e0Sbonwick zh = zil_header_in_syncing_context(zilog); 599fa9e4066Sahrens 600b24ab676SJeff Bonwick if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 601e6ca193dSGeorge Wilson if (!BP_IS_HOLE(&zh->zh_log)) 602b24ab676SJeff Bonwick zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 603e6ca193dSGeorge Wilson BP_ZERO(&zh->zh_log); 604e6ca193dSGeorge Wilson dsl_dataset_dirty(dmu_objset_ds(os), tx); 605468c413aSTim Haley dmu_objset_rele(os, FTAG); 606468c413aSTim Haley return (0); 607e6ca193dSGeorge Wilson } 608e6ca193dSGeorge Wilson 609fa9e4066Sahrens /* 610d80c45e0Sbonwick * Claim all log blocks if we haven't already done so, and remember 611d80c45e0Sbonwick * the highest claimed sequence number. This ensures that if we can 612d80c45e0Sbonwick * read only part of the log now (e.g. due to a missing device), 613d80c45e0Sbonwick * but we can read the entire log later, we will not try to replay 614d80c45e0Sbonwick * or destroy beyond the last block we successfully claimed. 615fa9e4066Sahrens */ 616fa9e4066Sahrens ASSERT3U(zh->zh_claim_txg, <=, first_txg); 617fa9e4066Sahrens if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 618b24ab676SJeff Bonwick (void) zil_parse(zilog, zil_claim_log_block, 619d80c45e0Sbonwick zil_claim_log_record, tx, first_txg); 620b24ab676SJeff Bonwick zh->zh_claim_txg = first_txg; 621b24ab676SJeff Bonwick zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 622b24ab676SJeff Bonwick zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 623b24ab676SJeff Bonwick if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 624b24ab676SJeff Bonwick zh->zh_flags |= ZIL_REPLAY_NEEDED; 625b24ab676SJeff Bonwick zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 626fa9e4066Sahrens dsl_dataset_dirty(dmu_objset_ds(os), tx); 627fa9e4066Sahrens } 628d80c45e0Sbonwick 629fa9e4066Sahrens ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 630503ad85cSMatthew Ahrens dmu_objset_rele(os, FTAG); 6311d452cf5Sahrens return (0); 632b87f3af3Sperrin } 633b87f3af3Sperrin 634b87f3af3Sperrin /* 635b87f3af3Sperrin * Check the log by walking the log chain. 636b87f3af3Sperrin * Checksum errors are ok as they indicate the end of the chain. 637b87f3af3Sperrin * Any other error (no device or read failure) returns an error. 638b87f3af3Sperrin */ 639b87f3af3Sperrin int 640fd136879SMatthew Ahrens zil_check_log_chain(const char *osname, void *tx) 641b87f3af3Sperrin { 642b87f3af3Sperrin zilog_t *zilog; 643b87f3af3Sperrin objset_t *os; 6444b964adaSGeorge Wilson blkptr_t *bp; 645b87f3af3Sperrin int error; 646b87f3af3Sperrin 647b24ab676SJeff Bonwick ASSERT(tx == NULL); 648b24ab676SJeff Bonwick 649503ad85cSMatthew Ahrens error = dmu_objset_hold(osname, FTAG, &os); 650b87f3af3Sperrin if (error) { 651b87f3af3Sperrin cmn_err(CE_WARN, "can't open objset for %s", osname); 652b87f3af3Sperrin return (0); 653b87f3af3Sperrin } 654b87f3af3Sperrin 655b87f3af3Sperrin zilog = dmu_objset_zil(os); 6564b964adaSGeorge Wilson bp = (blkptr_t *)&zilog->zl_header->zh_log; 6574b964adaSGeorge Wilson 6584b964adaSGeorge Wilson /* 6594b964adaSGeorge Wilson * Check the first block and determine if it's on a log device 6604b964adaSGeorge Wilson * which may have been removed or faulted prior to loading this 6614b964adaSGeorge Wilson * pool. If so, there's no point in checking the rest of the log 6624b964adaSGeorge Wilson * as its content should have already been synced to the pool. 6634b964adaSGeorge Wilson */ 6644b964adaSGeorge Wilson if (!BP_IS_HOLE(bp)) { 6654b964adaSGeorge Wilson vdev_t *vd; 6664b964adaSGeorge Wilson boolean_t valid = B_TRUE; 6674b964adaSGeorge Wilson 6684b964adaSGeorge Wilson spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 6694b964adaSGeorge Wilson vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 6704b964adaSGeorge Wilson if (vd->vdev_islog && vdev_is_dead(vd)) 6714b964adaSGeorge Wilson valid = vdev_log_state_valid(vd); 6724b964adaSGeorge Wilson spa_config_exit(os->os_spa, SCL_STATE, FTAG); 6734b964adaSGeorge Wilson 6744b964adaSGeorge Wilson if (!valid) { 6754b964adaSGeorge Wilson dmu_objset_rele(os, FTAG); 6764b964adaSGeorge Wilson return (0); 6774b964adaSGeorge Wilson } 6784b964adaSGeorge Wilson } 679b87f3af3Sperrin 680b24ab676SJeff Bonwick /* 681b24ab676SJeff Bonwick * Because tx == NULL, zil_claim_log_block() will not actually claim 682b24ab676SJeff Bonwick * any blocks, but just determine whether it is possible to do so. 683b24ab676SJeff Bonwick * In addition to checking the log chain, zil_claim_log_block() 684b24ab676SJeff Bonwick * will invoke zio_claim() with a done func of spa_claim_notify(), 685b24ab676SJeff Bonwick * which will update spa_max_claim_txg. See spa_load() for details. 686b24ab676SJeff Bonwick */ 687b24ab676SJeff Bonwick error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 688b24ab676SJeff Bonwick zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 689b24ab676SJeff Bonwick 690503ad85cSMatthew Ahrens dmu_objset_rele(os, FTAG); 691b24ab676SJeff Bonwick 692b24ab676SJeff Bonwick return ((error == ECKSUM || error == ENOENT) ? 0 : error); 693b87f3af3Sperrin } 694b87f3af3Sperrin 69517f17c2dSbonwick static int 69617f17c2dSbonwick zil_vdev_compare(const void *x1, const void *x2) 69717f17c2dSbonwick { 6985002558fSNeil Perrin const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 6995002558fSNeil Perrin const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 70017f17c2dSbonwick 70117f17c2dSbonwick if (v1 < v2) 70217f17c2dSbonwick return (-1); 70317f17c2dSbonwick if (v1 > v2) 70417f17c2dSbonwick return (1); 70517f17c2dSbonwick 70617f17c2dSbonwick return (0); 70717f17c2dSbonwick } 70817f17c2dSbonwick 709fa9e4066Sahrens void 710b24ab676SJeff Bonwick zil_add_block(zilog_t *zilog, const blkptr_t *bp) 711fa9e4066Sahrens { 71217f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 71317f17c2dSbonwick avl_index_t where; 71417f17c2dSbonwick zil_vdev_node_t *zv, zvsearch; 71517f17c2dSbonwick int ndvas = BP_GET_NDVAS(bp); 71617f17c2dSbonwick int i; 717fa9e4066Sahrens 718416e0cd8Sek if (zfs_nocacheflush) 719fa9e4066Sahrens return; 720fa9e4066Sahrens 72117f17c2dSbonwick ASSERT(zilog->zl_writer); 72217f17c2dSbonwick 72317f17c2dSbonwick /* 72417f17c2dSbonwick * Even though we're zl_writer, we still need a lock because the 72517f17c2dSbonwick * zl_get_data() callbacks may have dmu_sync() done callbacks 72617f17c2dSbonwick * that will run concurrently. 72717f17c2dSbonwick */ 72817f17c2dSbonwick mutex_enter(&zilog->zl_vdev_lock); 72917f17c2dSbonwick for (i = 0; i < ndvas; i++) { 73017f17c2dSbonwick zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 73117f17c2dSbonwick if (avl_find(t, &zvsearch, &where) == NULL) { 73217f17c2dSbonwick zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 73317f17c2dSbonwick zv->zv_vdev = zvsearch.zv_vdev; 73417f17c2dSbonwick avl_insert(t, zv, where); 73567bd71c6Sperrin } 73667bd71c6Sperrin } 73717f17c2dSbonwick mutex_exit(&zilog->zl_vdev_lock); 738fa9e4066Sahrens } 739fa9e4066Sahrens 74091de656bSNeil Perrin static void 74167bd71c6Sperrin zil_flush_vdevs(zilog_t *zilog) 74267bd71c6Sperrin { 74367bd71c6Sperrin spa_t *spa = zilog->zl_spa; 74417f17c2dSbonwick avl_tree_t *t = &zilog->zl_vdev_tree; 74517f17c2dSbonwick void *cookie = NULL; 74617f17c2dSbonwick zil_vdev_node_t *zv; 74717f17c2dSbonwick zio_t *zio; 748fa9e4066Sahrens 74967bd71c6Sperrin ASSERT(zilog->zl_writer); 75067bd71c6Sperrin 75117f17c2dSbonwick /* 75217f17c2dSbonwick * We don't need zl_vdev_lock here because we're the zl_writer, 75317f17c2dSbonwick * and all zl_get_data() callbacks are done. 75417f17c2dSbonwick */ 75517f17c2dSbonwick if (avl_numnodes(t) == 0) 75617f17c2dSbonwick return; 75717f17c2dSbonwick 758e14bb325SJeff Bonwick spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 75917f17c2dSbonwick 760e14bb325SJeff Bonwick zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 761fa9e4066Sahrens 76217f17c2dSbonwick while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 76317f17c2dSbonwick vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 76417f17c2dSbonwick if (vd != NULL) 76517f17c2dSbonwick zio_flush(zio, vd); 76617f17c2dSbonwick kmem_free(zv, sizeof (*zv)); 76767bd71c6Sperrin } 76817f17c2dSbonwick 769fa9e4066Sahrens /* 770fa9e4066Sahrens * Wait for all the flushes to complete. Not all devices actually 771fa9e4066Sahrens * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 772fa9e4066Sahrens */ 77317f17c2dSbonwick (void) zio_wait(zio); 77417f17c2dSbonwick 775e14bb325SJeff Bonwick spa_config_exit(spa, SCL_STATE, FTAG); 776fa9e4066Sahrens } 777fa9e4066Sahrens 778fa9e4066Sahrens /* 779fa9e4066Sahrens * Function called when a log block write completes 780fa9e4066Sahrens */ 781fa9e4066Sahrens static void 782fa9e4066Sahrens zil_lwb_write_done(zio_t *zio) 783fa9e4066Sahrens { 784fa9e4066Sahrens lwb_t *lwb = zio->io_private; 785fa9e4066Sahrens zilog_t *zilog = lwb->lwb_zilog; 786b24ab676SJeff Bonwick dmu_tx_t *tx = lwb->lwb_tx; 787fa9e4066Sahrens 788e14bb325SJeff Bonwick ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 789e14bb325SJeff Bonwick ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 790e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 791e14bb325SJeff Bonwick ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 792e14bb325SJeff Bonwick ASSERT(!BP_IS_GANG(zio->io_bp)); 793e14bb325SJeff Bonwick ASSERT(!BP_IS_HOLE(zio->io_bp)); 794e14bb325SJeff Bonwick ASSERT(zio->io_bp->blk_fill == 0); 795e14bb325SJeff Bonwick 796fa9e4066Sahrens /* 797ef0d8e11SNeil Perrin * Ensure the lwb buffer pointer is cleared before releasing 798ef0d8e11SNeil Perrin * the txg. If we have had an allocation failure and 799ef0d8e11SNeil Perrin * the txg is waiting to sync then we want want zil_sync() 800ef0d8e11SNeil Perrin * to remove the lwb so that it's not picked up as the next new 801ef0d8e11SNeil Perrin * one in zil_commit_writer(). zil_sync() will only remove 802ef0d8e11SNeil Perrin * the lwb if lwb_buf is null. 803fa9e4066Sahrens */ 804fa9e4066Sahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 805fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 806fa9e4066Sahrens lwb->lwb_buf = NULL; 807b24ab676SJeff Bonwick lwb->lwb_tx = NULL; 808b24ab676SJeff Bonwick mutex_exit(&zilog->zl_lock); 809ef0d8e11SNeil Perrin 810ef0d8e11SNeil Perrin /* 811ef0d8e11SNeil Perrin * Now that we've written this log block, we have a stable pointer 812ef0d8e11SNeil Perrin * to the next block in the chain, so it's OK to let the txg in 813b24ab676SJeff Bonwick * which we allocated the next block sync. 814ef0d8e11SNeil Perrin */ 815b24ab676SJeff Bonwick dmu_tx_commit(tx); 816fa9e4066Sahrens } 817fa9e4066Sahrens 818c5c6ffa0Smaybee /* 819c5c6ffa0Smaybee * Initialize the io for a log block. 820c5c6ffa0Smaybee */ 821c5c6ffa0Smaybee static void 822c5c6ffa0Smaybee zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 823c5c6ffa0Smaybee { 824c5c6ffa0Smaybee zbookmark_t zb; 825c5c6ffa0Smaybee 826b24ab676SJeff Bonwick SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 827b24ab676SJeff Bonwick ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 828b24ab676SJeff Bonwick lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 829c5c6ffa0Smaybee 830b19a79ecSperrin if (zilog->zl_root_zio == NULL) { 831b19a79ecSperrin zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 832b19a79ecSperrin ZIO_FLAG_CANFAIL); 833b19a79ecSperrin } 83467bd71c6Sperrin if (lwb->lwb_zio == NULL) { 83567bd71c6Sperrin lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 8366e1f5caaSNeil Perrin 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 837e6ca193dSGeorge Wilson zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 8388f18d1faSGeorge Wilson ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 83967bd71c6Sperrin } 840c5c6ffa0Smaybee } 841c5c6ffa0Smaybee 8426e1f5caaSNeil Perrin /* 8436e1f5caaSNeil Perrin * Define a limited set of intent log block sizes. 8446e1f5caaSNeil Perrin * These must be a multiple of 4KB. Note only the amount used (again 8456e1f5caaSNeil Perrin * aligned to 4KB) actually gets written. However, we can't always just 8466e1f5caaSNeil Perrin * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 8476e1f5caaSNeil Perrin */ 8486e1f5caaSNeil Perrin uint64_t zil_block_buckets[] = { 8496e1f5caaSNeil Perrin 4096, /* non TX_WRITE */ 8506e1f5caaSNeil Perrin 8192+4096, /* data base */ 8516e1f5caaSNeil Perrin 32*1024 + 4096, /* NFS writes */ 8526e1f5caaSNeil Perrin UINT64_MAX 8536e1f5caaSNeil Perrin }; 8546e1f5caaSNeil Perrin 855d48e086fSNeil Perrin /* 856d48e086fSNeil Perrin * Use the slog as long as the logbias is 'latency' and the current commit size 857d48e086fSNeil Perrin * is less than the limit or the total list size is less than 2X the limit. 858d48e086fSNeil Perrin * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 859d48e086fSNeil Perrin */ 860d48e086fSNeil Perrin uint64_t zil_slog_limit = 1024 * 1024; 861d48e086fSNeil Perrin #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 862d48e086fSNeil Perrin (((zilog)->zl_cur_used < zil_slog_limit) || \ 863d48e086fSNeil Perrin ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 864d48e086fSNeil Perrin 865fa9e4066Sahrens /* 866fa9e4066Sahrens * Start a log block write and advance to the next log block. 867fa9e4066Sahrens * Calls are serialized. 868fa9e4066Sahrens */ 869fa9e4066Sahrens static lwb_t * 870fa9e4066Sahrens zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 871fa9e4066Sahrens { 8726e1f5caaSNeil Perrin lwb_t *nlwb = NULL; 8736e1f5caaSNeil Perrin zil_chain_t *zilc; 874d80c45e0Sbonwick spa_t *spa = zilog->zl_spa; 8756e1f5caaSNeil Perrin blkptr_t *bp; 876b24ab676SJeff Bonwick dmu_tx_t *tx; 877fa9e4066Sahrens uint64_t txg; 878ada693c4SNeil Perrin uint64_t zil_blksz, wsz; 8796e1f5caaSNeil Perrin int i, error; 8806e1f5caaSNeil Perrin 8816e1f5caaSNeil Perrin if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 8826e1f5caaSNeil Perrin zilc = (zil_chain_t *)lwb->lwb_buf; 8836e1f5caaSNeil Perrin bp = &zilc->zc_next_blk; 8846e1f5caaSNeil Perrin } else { 8856e1f5caaSNeil Perrin zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 8866e1f5caaSNeil Perrin bp = &zilc->zc_next_blk; 8876e1f5caaSNeil Perrin } 888fa9e4066Sahrens 8896e1f5caaSNeil Perrin ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 890fa9e4066Sahrens 891fa9e4066Sahrens /* 892fa9e4066Sahrens * Allocate the next block and save its address in this block 893fa9e4066Sahrens * before writing it in order to establish the log chain. 894fa9e4066Sahrens * Note that if the allocation of nlwb synced before we wrote 895fa9e4066Sahrens * the block that points at it (lwb), we'd leak it if we crashed. 896b24ab676SJeff Bonwick * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 897b24ab676SJeff Bonwick * We dirty the dataset to ensure that zil_sync() will be called 898b24ab676SJeff Bonwick * to clean up in the event of allocation failure or I/O failure. 899fa9e4066Sahrens */ 900b24ab676SJeff Bonwick tx = dmu_tx_create(zilog->zl_os); 901b24ab676SJeff Bonwick VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 902b24ab676SJeff Bonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 903b24ab676SJeff Bonwick txg = dmu_tx_get_txg(tx); 904b24ab676SJeff Bonwick 905b24ab676SJeff Bonwick lwb->lwb_tx = tx; 906fa9e4066Sahrens 907fa9e4066Sahrens /* 9086e1f5caaSNeil Perrin * Log blocks are pre-allocated. Here we select the size of the next 9096e1f5caaSNeil Perrin * block, based on size used in the last block. 9106e1f5caaSNeil Perrin * - first find the smallest bucket that will fit the block from a 9116e1f5caaSNeil Perrin * limited set of block sizes. This is because it's faster to write 9126e1f5caaSNeil Perrin * blocks allocated from the same metaslab as they are adjacent or 9136e1f5caaSNeil Perrin * close. 9146e1f5caaSNeil Perrin * - next find the maximum from the new suggested size and an array of 9156e1f5caaSNeil Perrin * previous sizes. This lessens a picket fence effect of wrongly 9166e1f5caaSNeil Perrin * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 9176e1f5caaSNeil Perrin * requests. 9186e1f5caaSNeil Perrin * 9196e1f5caaSNeil Perrin * Note we only write what is used, but we can't just allocate 9206e1f5caaSNeil Perrin * the maximum block size because we can exhaust the available 9216e1f5caaSNeil Perrin * pool log space. 922fa9e4066Sahrens */ 9236e1f5caaSNeil Perrin zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 9246e1f5caaSNeil Perrin for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 9256e1f5caaSNeil Perrin continue; 9266e1f5caaSNeil Perrin zil_blksz = zil_block_buckets[i]; 9276e1f5caaSNeil Perrin if (zil_blksz == UINT64_MAX) 9286e1f5caaSNeil Perrin zil_blksz = SPA_MAXBLOCKSIZE; 9296e1f5caaSNeil Perrin zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 9306e1f5caaSNeil Perrin for (i = 0; i < ZIL_PREV_BLKS; i++) 9316e1f5caaSNeil Perrin zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 9326e1f5caaSNeil Perrin zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 933fa9e4066Sahrens 93467bd71c6Sperrin BP_ZERO(bp); 93567bd71c6Sperrin /* pass the old blkptr in order to spread log blocks across devs */ 936b24ab676SJeff Bonwick error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 937d48e086fSNeil Perrin USE_SLOG(zilog)); 9386e1f5caaSNeil Perrin if (!error) { 9396e1f5caaSNeil Perrin ASSERT3U(bp->blk_birth, ==, txg); 9406e1f5caaSNeil Perrin bp->blk_cksum = lwb->lwb_blk.blk_cksum; 9416e1f5caaSNeil Perrin bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 942d63d470bSgw 943ea8dc4b6Seschrock /* 9446e1f5caaSNeil Perrin * Allocate a new log write buffer (lwb). 945ea8dc4b6Seschrock */ 9466e1f5caaSNeil Perrin nlwb = zil_alloc_lwb(zilog, bp, txg); 9476e1f5caaSNeil Perrin 9486e1f5caaSNeil Perrin /* Record the block for later vdev flushing */ 9496e1f5caaSNeil Perrin zil_add_block(zilog, &lwb->lwb_blk); 950fa9e4066Sahrens } 951fa9e4066Sahrens 9526e1f5caaSNeil Perrin if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 9536e1f5caaSNeil Perrin /* For Slim ZIL only write what is used. */ 954ada693c4SNeil Perrin wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 955ada693c4SNeil Perrin ASSERT3U(wsz, <=, lwb->lwb_sz); 956ada693c4SNeil Perrin zio_shrink(lwb->lwb_zio, wsz); 957fa9e4066Sahrens 958ada693c4SNeil Perrin } else { 959ada693c4SNeil Perrin wsz = lwb->lwb_sz; 9606e1f5caaSNeil Perrin } 961ada693c4SNeil Perrin 9626e1f5caaSNeil Perrin zilc->zc_pad = 0; 9636e1f5caaSNeil Perrin zilc->zc_nused = lwb->lwb_nused; 9646e1f5caaSNeil Perrin zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 965fa9e4066Sahrens 966ada693c4SNeil Perrin /* 967ada693c4SNeil Perrin * clear unused data for security 968ada693c4SNeil Perrin */ 969ada693c4SNeil Perrin bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 970ada693c4SNeil Perrin 9716e1f5caaSNeil Perrin zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 97267bd71c6Sperrin 973fa9e4066Sahrens /* 9746e1f5caaSNeil Perrin * If there was an allocation failure then nlwb will be null which 9756e1f5caaSNeil Perrin * forces a txg_wait_synced(). 976fa9e4066Sahrens */ 977fa9e4066Sahrens return (nlwb); 978fa9e4066Sahrens } 979fa9e4066Sahrens 980fa9e4066Sahrens static lwb_t * 981fa9e4066Sahrens zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 982fa9e4066Sahrens { 983fa9e4066Sahrens lr_t *lrc = &itx->itx_lr; /* common log record */ 984b24ab676SJeff Bonwick lr_write_t *lrw = (lr_write_t *)lrc; 985b24ab676SJeff Bonwick char *lr_buf; 986fa9e4066Sahrens uint64_t txg = lrc->lrc_txg; 987fa9e4066Sahrens uint64_t reclen = lrc->lrc_reclen; 988b24ab676SJeff Bonwick uint64_t dlen = 0; 989fa9e4066Sahrens 990fa9e4066Sahrens if (lwb == NULL) 991fa9e4066Sahrens return (NULL); 992b24ab676SJeff Bonwick 993fa9e4066Sahrens ASSERT(lwb->lwb_buf != NULL); 994fa9e4066Sahrens 995c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 996c5c6ffa0Smaybee dlen = P2ROUNDUP_TYPED( 997b24ab676SJeff Bonwick lrw->lr_length, sizeof (uint64_t), uint64_t); 998fa9e4066Sahrens 999104e2ed7Sperrin zilog->zl_cur_used += (reclen + dlen); 100022ac5be4Sperrin 100167bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 100267bd71c6Sperrin 1003fa9e4066Sahrens /* 1004fa9e4066Sahrens * If this record won't fit in the current log block, start a new one. 1005fa9e4066Sahrens */ 10066e1f5caaSNeil Perrin if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1007fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 1008c5c6ffa0Smaybee if (lwb == NULL) 1009fa9e4066Sahrens return (NULL); 101067bd71c6Sperrin zil_lwb_write_init(zilog, lwb); 10116e1f5caaSNeil Perrin ASSERT(LWB_EMPTY(lwb)); 10126e1f5caaSNeil Perrin if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1013fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, txg); 1014fa9e4066Sahrens return (lwb); 1015fa9e4066Sahrens } 1016fa9e4066Sahrens } 1017fa9e4066Sahrens 1018b24ab676SJeff Bonwick lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1019b24ab676SJeff Bonwick bcopy(lrc, lr_buf, reclen); 1020b24ab676SJeff Bonwick lrc = (lr_t *)lr_buf; 1021b24ab676SJeff Bonwick lrw = (lr_write_t *)lrc; 1022c5c6ffa0Smaybee 1023c5c6ffa0Smaybee /* 1024c5c6ffa0Smaybee * If it's a write, fetch the data or get its blkptr as appropriate. 1025c5c6ffa0Smaybee */ 1026c5c6ffa0Smaybee if (lrc->lrc_txtype == TX_WRITE) { 1027c5c6ffa0Smaybee if (txg > spa_freeze_txg(zilog->zl_spa)) 1028c5c6ffa0Smaybee txg_wait_synced(zilog->zl_dmu_pool, txg); 1029c5c6ffa0Smaybee if (itx->itx_wr_state != WR_COPIED) { 1030c5c6ffa0Smaybee char *dbuf; 1031c5c6ffa0Smaybee int error; 1032c5c6ffa0Smaybee 1033c5c6ffa0Smaybee if (dlen) { 1034c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1035b24ab676SJeff Bonwick dbuf = lr_buf + reclen; 1036b24ab676SJeff Bonwick lrw->lr_common.lrc_reclen += dlen; 1037c5c6ffa0Smaybee } else { 1038c5c6ffa0Smaybee ASSERT(itx->itx_wr_state == WR_INDIRECT); 1039c5c6ffa0Smaybee dbuf = NULL; 1040c5c6ffa0Smaybee } 1041c5c6ffa0Smaybee error = zilog->zl_get_data( 1042b24ab676SJeff Bonwick itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1043c87b8fc5SMark J Musante if (error == EIO) { 1044c87b8fc5SMark J Musante txg_wait_synced(zilog->zl_dmu_pool, txg); 1045c87b8fc5SMark J Musante return (lwb); 1046c87b8fc5SMark J Musante } 1047c5c6ffa0Smaybee if (error) { 1048c5c6ffa0Smaybee ASSERT(error == ENOENT || error == EEXIST || 1049c5c6ffa0Smaybee error == EALREADY); 1050c5c6ffa0Smaybee return (lwb); 1051c5c6ffa0Smaybee } 1052c5c6ffa0Smaybee } 1053104e2ed7Sperrin } 1054c5c6ffa0Smaybee 1055b24ab676SJeff Bonwick /* 1056b24ab676SJeff Bonwick * We're actually making an entry, so update lrc_seq to be the 1057b24ab676SJeff Bonwick * log record sequence number. Note that this is generally not 1058b24ab676SJeff Bonwick * equal to the itx sequence number because not all transactions 1059b24ab676SJeff Bonwick * are synchronous, and sometimes spa_sync() gets there first. 1060b24ab676SJeff Bonwick */ 1061b24ab676SJeff Bonwick lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1062c5c6ffa0Smaybee lwb->lwb_nused += reclen + dlen; 1063fa9e4066Sahrens lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 10646e1f5caaSNeil Perrin ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1065fa9e4066Sahrens ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 1066fa9e4066Sahrens 1067fa9e4066Sahrens return (lwb); 1068fa9e4066Sahrens } 1069fa9e4066Sahrens 1070fa9e4066Sahrens itx_t * 1071da6c28aaSamw zil_itx_create(uint64_t txtype, size_t lrsize) 1072fa9e4066Sahrens { 1073fa9e4066Sahrens itx_t *itx; 1074fa9e4066Sahrens 1075b4d654b0Sperrin lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1076fa9e4066Sahrens 1077fa9e4066Sahrens itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1078fa9e4066Sahrens itx->itx_lr.lrc_txtype = txtype; 1079fa9e4066Sahrens itx->itx_lr.lrc_reclen = lrsize; 1080abf76b6eSperrin itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1081fa9e4066Sahrens itx->itx_lr.lrc_seq = 0; /* defensive */ 10825002558fSNeil Perrin itx->itx_sync = B_TRUE; /* default is synchronous */ 1083fa9e4066Sahrens 1084fa9e4066Sahrens return (itx); 1085fa9e4066Sahrens } 1086fa9e4066Sahrens 1087b24ab676SJeff Bonwick void 1088b24ab676SJeff Bonwick zil_itx_destroy(itx_t *itx) 1089b24ab676SJeff Bonwick { 1090b24ab676SJeff Bonwick kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1091b24ab676SJeff Bonwick } 1092b24ab676SJeff Bonwick 10935002558fSNeil Perrin /* 10945002558fSNeil Perrin * Free up the sync and async itxs. The itxs_t has already been detached 10955002558fSNeil Perrin * so no locks are needed. 10965002558fSNeil Perrin */ 10975002558fSNeil Perrin static void 10985002558fSNeil Perrin zil_itxg_clean(itxs_t *itxs) 1099fa9e4066Sahrens { 11005002558fSNeil Perrin itx_t *itx; 11015002558fSNeil Perrin list_t *list; 11025002558fSNeil Perrin avl_tree_t *t; 11035002558fSNeil Perrin void *cookie; 11045002558fSNeil Perrin itx_async_node_t *ian; 11055002558fSNeil Perrin 11065002558fSNeil Perrin list = &itxs->i_sync_list; 11075002558fSNeil Perrin while ((itx = list_head(list)) != NULL) { 11085002558fSNeil Perrin list_remove(list, itx); 11095002558fSNeil Perrin kmem_free(itx, offsetof(itx_t, itx_lr) + 11105002558fSNeil Perrin itx->itx_lr.lrc_reclen); 11115002558fSNeil Perrin } 1112fa9e4066Sahrens 11135002558fSNeil Perrin cookie = NULL; 11145002558fSNeil Perrin t = &itxs->i_async_tree; 11155002558fSNeil Perrin while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 11165002558fSNeil Perrin list = &ian->ia_list; 11175002558fSNeil Perrin while ((itx = list_head(list)) != NULL) { 11185002558fSNeil Perrin list_remove(list, itx); 11195002558fSNeil Perrin kmem_free(itx, offsetof(itx_t, itx_lr) + 11205002558fSNeil Perrin itx->itx_lr.lrc_reclen); 11215002558fSNeil Perrin } 11225002558fSNeil Perrin list_destroy(list); 11235002558fSNeil Perrin kmem_free(ian, sizeof (itx_async_node_t)); 11245002558fSNeil Perrin } 11255002558fSNeil Perrin avl_destroy(t); 1126fa9e4066Sahrens 11275002558fSNeil Perrin kmem_free(itxs, sizeof (itxs_t)); 11285002558fSNeil Perrin } 11295002558fSNeil Perrin 11305002558fSNeil Perrin static int 11315002558fSNeil Perrin zil_aitx_compare(const void *x1, const void *x2) 11325002558fSNeil Perrin { 11335002558fSNeil Perrin const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 11345002558fSNeil Perrin const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1135fa9e4066Sahrens 11365002558fSNeil Perrin if (o1 < o2) 11375002558fSNeil Perrin return (-1); 11385002558fSNeil Perrin if (o1 > o2) 11395002558fSNeil Perrin return (1); 11405002558fSNeil Perrin 11415002558fSNeil Perrin return (0); 1142fa9e4066Sahrens } 1143fa9e4066Sahrens 1144fa9e4066Sahrens /* 11455002558fSNeil Perrin * Remove all async itx with the given oid. 1146fa9e4066Sahrens */ 114791de656bSNeil Perrin static void 11485002558fSNeil Perrin zil_remove_async(zilog_t *zilog, uint64_t oid) 1149fa9e4066Sahrens { 11505002558fSNeil Perrin uint64_t otxg, txg; 11515002558fSNeil Perrin itx_async_node_t *ian; 11525002558fSNeil Perrin avl_tree_t *t; 11535002558fSNeil Perrin avl_index_t where; 1154a584ef65Sjohansen list_t clean_list; 1155fa9e4066Sahrens itx_t *itx; 1156fa9e4066Sahrens 11575002558fSNeil Perrin ASSERT(oid != 0); 1158a584ef65Sjohansen list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1159a584ef65Sjohansen 11605002558fSNeil Perrin if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 11615002558fSNeil Perrin otxg = ZILTEST_TXG; 11625002558fSNeil Perrin else 11635002558fSNeil Perrin otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1164a584ef65Sjohansen 11655002558fSNeil Perrin for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 11665002558fSNeil Perrin itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 11675002558fSNeil Perrin 11685002558fSNeil Perrin mutex_enter(&itxg->itxg_lock); 11695002558fSNeil Perrin if (itxg->itxg_txg != txg) { 11705002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 11715002558fSNeil Perrin continue; 11725002558fSNeil Perrin } 1173a584ef65Sjohansen 11745002558fSNeil Perrin /* 11755002558fSNeil Perrin * Locate the object node and append its list. 11765002558fSNeil Perrin */ 11775002558fSNeil Perrin t = &itxg->itxg_itxs->i_async_tree; 11785002558fSNeil Perrin ian = avl_find(t, &oid, &where); 11795002558fSNeil Perrin if (ian != NULL) 11805002558fSNeil Perrin list_move_tail(&clean_list, &ian->ia_list); 11815002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 11825002558fSNeil Perrin } 1183a584ef65Sjohansen while ((itx = list_head(&clean_list)) != NULL) { 1184a584ef65Sjohansen list_remove(&clean_list, itx); 11855002558fSNeil Perrin kmem_free(itx, offsetof(itx_t, itx_lr) + 11865002558fSNeil Perrin itx->itx_lr.lrc_reclen); 1187a584ef65Sjohansen } 1188a584ef65Sjohansen list_destroy(&clean_list); 1189fa9e4066Sahrens } 1190fa9e4066Sahrens 11915002558fSNeil Perrin void 11925002558fSNeil Perrin zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 11935002558fSNeil Perrin { 11945002558fSNeil Perrin uint64_t txg; 11955002558fSNeil Perrin itxg_t *itxg; 11965002558fSNeil Perrin itxs_t *itxs, *clean = NULL; 11975002558fSNeil Perrin 11985002558fSNeil Perrin /* 119991de656bSNeil Perrin * Object ids can be re-instantiated in the next txg so 12005002558fSNeil Perrin * remove any async transactions to avoid future leaks. 12015002558fSNeil Perrin * This can happen if a fsync occurs on the re-instantiated 12025002558fSNeil Perrin * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 12035002558fSNeil Perrin * the new file data and flushes a write record for the old object. 12045002558fSNeil Perrin */ 12055002558fSNeil Perrin if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 120651bd2f97SNeil Perrin zil_remove_async(zilog, itx->itx_oid); 12075002558fSNeil Perrin 120891de656bSNeil Perrin /* 120991de656bSNeil Perrin * Ensure the data of a renamed file is committed before the rename. 121091de656bSNeil Perrin */ 121191de656bSNeil Perrin if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 121291de656bSNeil Perrin zil_async_to_sync(zilog, itx->itx_oid); 121391de656bSNeil Perrin 12145002558fSNeil Perrin if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 12155002558fSNeil Perrin txg = ZILTEST_TXG; 12165002558fSNeil Perrin else 12175002558fSNeil Perrin txg = dmu_tx_get_txg(tx); 12185002558fSNeil Perrin 12195002558fSNeil Perrin itxg = &zilog->zl_itxg[txg & TXG_MASK]; 12205002558fSNeil Perrin mutex_enter(&itxg->itxg_lock); 12215002558fSNeil Perrin itxs = itxg->itxg_itxs; 12225002558fSNeil Perrin if (itxg->itxg_txg != txg) { 12235002558fSNeil Perrin if (itxs != NULL) { 12245002558fSNeil Perrin /* 12255002558fSNeil Perrin * The zil_clean callback hasn't got around to cleaning 12265002558fSNeil Perrin * this itxg. Save the itxs for release below. 12275002558fSNeil Perrin * This should be rare. 12285002558fSNeil Perrin */ 12295002558fSNeil Perrin atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 12305002558fSNeil Perrin itxg->itxg_sod = 0; 12315002558fSNeil Perrin clean = itxg->itxg_itxs; 12325002558fSNeil Perrin } 12335002558fSNeil Perrin ASSERT(itxg->itxg_sod == 0); 12345002558fSNeil Perrin itxg->itxg_txg = txg; 12355002558fSNeil Perrin itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 12365002558fSNeil Perrin 12375002558fSNeil Perrin list_create(&itxs->i_sync_list, sizeof (itx_t), 12385002558fSNeil Perrin offsetof(itx_t, itx_node)); 12395002558fSNeil Perrin avl_create(&itxs->i_async_tree, zil_aitx_compare, 12405002558fSNeil Perrin sizeof (itx_async_node_t), 12415002558fSNeil Perrin offsetof(itx_async_node_t, ia_node)); 12425002558fSNeil Perrin } 12435002558fSNeil Perrin if (itx->itx_sync) { 12445002558fSNeil Perrin list_insert_tail(&itxs->i_sync_list, itx); 12455002558fSNeil Perrin atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 12465002558fSNeil Perrin itxg->itxg_sod += itx->itx_sod; 12475002558fSNeil Perrin } else { 12485002558fSNeil Perrin avl_tree_t *t = &itxs->i_async_tree; 12495002558fSNeil Perrin uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 12505002558fSNeil Perrin itx_async_node_t *ian; 12515002558fSNeil Perrin avl_index_t where; 12525002558fSNeil Perrin 12535002558fSNeil Perrin ian = avl_find(t, &foid, &where); 12545002558fSNeil Perrin if (ian == NULL) { 12555002558fSNeil Perrin ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 12565002558fSNeil Perrin list_create(&ian->ia_list, sizeof (itx_t), 12575002558fSNeil Perrin offsetof(itx_t, itx_node)); 12585002558fSNeil Perrin ian->ia_foid = foid; 12595002558fSNeil Perrin avl_insert(t, ian, where); 12605002558fSNeil Perrin } 12615002558fSNeil Perrin list_insert_tail(&ian->ia_list, itx); 12625002558fSNeil Perrin } 12635002558fSNeil Perrin 12645002558fSNeil Perrin itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 12655002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 12665002558fSNeil Perrin 12675002558fSNeil Perrin /* Release the old itxs now we've dropped the lock */ 12685002558fSNeil Perrin if (clean != NULL) 12695002558fSNeil Perrin zil_itxg_clean(clean); 12705002558fSNeil Perrin } 12715002558fSNeil Perrin 1272b19a79ecSperrin /* 127367bd71c6Sperrin * If there are any in-memory intent log transactions which have now been 127467bd71c6Sperrin * synced then start up a taskq to free them. 1275b19a79ecSperrin */ 1276fa9e4066Sahrens void 12775002558fSNeil Perrin zil_clean(zilog_t *zilog, uint64_t synced_txg) 1278fa9e4066Sahrens { 12795002558fSNeil Perrin itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 12805002558fSNeil Perrin itxs_t *clean_me; 128167bd71c6Sperrin 12825002558fSNeil Perrin mutex_enter(&itxg->itxg_lock); 12835002558fSNeil Perrin if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 12845002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 12855002558fSNeil Perrin return; 12865002558fSNeil Perrin } 12875002558fSNeil Perrin ASSERT3U(itxg->itxg_txg, <=, synced_txg); 12885002558fSNeil Perrin ASSERT(itxg->itxg_txg != 0); 12895002558fSNeil Perrin ASSERT(zilog->zl_clean_taskq != NULL); 12905002558fSNeil Perrin atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 12915002558fSNeil Perrin itxg->itxg_sod = 0; 12925002558fSNeil Perrin clean_me = itxg->itxg_itxs; 12935002558fSNeil Perrin itxg->itxg_itxs = NULL; 12945002558fSNeil Perrin itxg->itxg_txg = 0; 12955002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 12965002558fSNeil Perrin /* 12975002558fSNeil Perrin * Preferably start a task queue to free up the old itxs but 12985002558fSNeil Perrin * if taskq_dispatch can't allocate resources to do that then 12995002558fSNeil Perrin * free it in-line. This should be rare. Note, using TQ_SLEEP 13005002558fSNeil Perrin * created a bad performance problem. 13015002558fSNeil Perrin */ 13025002558fSNeil Perrin if (taskq_dispatch(zilog->zl_clean_taskq, 13035002558fSNeil Perrin (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == NULL) 13045002558fSNeil Perrin zil_itxg_clean(clean_me); 13055002558fSNeil Perrin } 13065002558fSNeil Perrin 13075002558fSNeil Perrin /* 13085002558fSNeil Perrin * Get the list of itxs to commit into zl_itx_commit_list. 13095002558fSNeil Perrin */ 131091de656bSNeil Perrin static void 13115002558fSNeil Perrin zil_get_commit_list(zilog_t *zilog) 13125002558fSNeil Perrin { 13135002558fSNeil Perrin uint64_t otxg, txg; 13145002558fSNeil Perrin list_t *commit_list = &zilog->zl_itx_commit_list; 13155002558fSNeil Perrin uint64_t push_sod = 0; 13165002558fSNeil Perrin 13175002558fSNeil Perrin if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 13185002558fSNeil Perrin otxg = ZILTEST_TXG; 13195002558fSNeil Perrin else 13205002558fSNeil Perrin otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 13215002558fSNeil Perrin 13225002558fSNeil Perrin for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 13235002558fSNeil Perrin itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 13245002558fSNeil Perrin 13255002558fSNeil Perrin mutex_enter(&itxg->itxg_lock); 13265002558fSNeil Perrin if (itxg->itxg_txg != txg) { 13275002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 13285002558fSNeil Perrin continue; 13295002558fSNeil Perrin } 13305002558fSNeil Perrin 13315002558fSNeil Perrin list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 13325002558fSNeil Perrin push_sod += itxg->itxg_sod; 13335002558fSNeil Perrin itxg->itxg_sod = 0; 13345002558fSNeil Perrin 13355002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 13365002558fSNeil Perrin } 13375002558fSNeil Perrin atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 13385002558fSNeil Perrin } 13395002558fSNeil Perrin 13405002558fSNeil Perrin /* 13415002558fSNeil Perrin * Move the async itxs for a specified object to commit into sync lists. 13425002558fSNeil Perrin */ 134391de656bSNeil Perrin static void 13445002558fSNeil Perrin zil_async_to_sync(zilog_t *zilog, uint64_t foid) 13455002558fSNeil Perrin { 13465002558fSNeil Perrin uint64_t otxg, txg; 13475002558fSNeil Perrin itx_async_node_t *ian; 13485002558fSNeil Perrin avl_tree_t *t; 13495002558fSNeil Perrin avl_index_t where; 13505002558fSNeil Perrin 13515002558fSNeil Perrin if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 13525002558fSNeil Perrin otxg = ZILTEST_TXG; 13535002558fSNeil Perrin else 13545002558fSNeil Perrin otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 13555002558fSNeil Perrin 13565002558fSNeil Perrin for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 13575002558fSNeil Perrin itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 13585002558fSNeil Perrin 13595002558fSNeil Perrin mutex_enter(&itxg->itxg_lock); 13605002558fSNeil Perrin if (itxg->itxg_txg != txg) { 13615002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 13625002558fSNeil Perrin continue; 13635002558fSNeil Perrin } 13645002558fSNeil Perrin 13655002558fSNeil Perrin /* 13665002558fSNeil Perrin * If a foid is specified then find that node and append its 13675002558fSNeil Perrin * list. Otherwise walk the tree appending all the lists 13685002558fSNeil Perrin * to the sync list. We add to the end rather than the 13695002558fSNeil Perrin * beginning to ensure the create has happened. 13705002558fSNeil Perrin */ 13715002558fSNeil Perrin t = &itxg->itxg_itxs->i_async_tree; 13725002558fSNeil Perrin if (foid != 0) { 13735002558fSNeil Perrin ian = avl_find(t, &foid, &where); 13745002558fSNeil Perrin if (ian != NULL) { 13755002558fSNeil Perrin list_move_tail(&itxg->itxg_itxs->i_sync_list, 13765002558fSNeil Perrin &ian->ia_list); 13775002558fSNeil Perrin } 13785002558fSNeil Perrin } else { 13795002558fSNeil Perrin void *cookie = NULL; 13805002558fSNeil Perrin 13815002558fSNeil Perrin while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 13825002558fSNeil Perrin list_move_tail(&itxg->itxg_itxs->i_sync_list, 13835002558fSNeil Perrin &ian->ia_list); 13845002558fSNeil Perrin list_destroy(&ian->ia_list); 13855002558fSNeil Perrin kmem_free(ian, sizeof (itx_async_node_t)); 13865002558fSNeil Perrin } 13875002558fSNeil Perrin } 13885002558fSNeil Perrin mutex_exit(&itxg->itxg_lock); 138967bd71c6Sperrin } 1390fa9e4066Sahrens } 1391fa9e4066Sahrens 1392e14bb325SJeff Bonwick static void 13935002558fSNeil Perrin zil_commit_writer(zilog_t *zilog) 1394fa9e4066Sahrens { 1395fa9e4066Sahrens uint64_t txg; 13965002558fSNeil Perrin itx_t *itx; 1397fa9e4066Sahrens lwb_t *lwb; 13985002558fSNeil Perrin spa_t *spa = zilog->zl_spa; 1399b24ab676SJeff Bonwick int error = 0; 1400fa9e4066Sahrens 1401e14bb325SJeff Bonwick ASSERT(zilog->zl_root_zio == NULL); 14025002558fSNeil Perrin 14035002558fSNeil Perrin mutex_exit(&zilog->zl_lock); 14045002558fSNeil Perrin 14055002558fSNeil Perrin zil_get_commit_list(zilog); 14065002558fSNeil Perrin 14075002558fSNeil Perrin /* 14085002558fSNeil Perrin * Return if there's nothing to commit before we dirty the fs by 14095002558fSNeil Perrin * calling zil_create(). 14105002558fSNeil Perrin */ 14115002558fSNeil Perrin if (list_head(&zilog->zl_itx_commit_list) == NULL) { 14125002558fSNeil Perrin mutex_enter(&zilog->zl_lock); 14135002558fSNeil Perrin return; 14145002558fSNeil Perrin } 1415fa9e4066Sahrens 1416fa9e4066Sahrens if (zilog->zl_suspend) { 1417fa9e4066Sahrens lwb = NULL; 1418fa9e4066Sahrens } else { 1419fa9e4066Sahrens lwb = list_tail(&zilog->zl_lwb_list); 14205002558fSNeil Perrin if (lwb == NULL) 14216e1f5caaSNeil Perrin lwb = zil_create(zilog); 1422fa9e4066Sahrens } 1423fa9e4066Sahrens 1424b19a79ecSperrin DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 14255002558fSNeil Perrin while (itx = list_head(&zilog->zl_itx_commit_list)) { 1426fa9e4066Sahrens txg = itx->itx_lr.lrc_txg; 1427fa9e4066Sahrens ASSERT(txg); 1428fa9e4066Sahrens 14295002558fSNeil Perrin if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1430fa9e4066Sahrens lwb = zil_lwb_commit(zilog, itx, lwb); 14315002558fSNeil Perrin list_remove(&zilog->zl_itx_commit_list, itx); 14325002558fSNeil Perrin kmem_free(itx, offsetof(itx_t, itx_lr) 14335002558fSNeil Perrin + itx->itx_lr.lrc_reclen); 1434fa9e4066Sahrens } 1435b19a79ecSperrin DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1436fa9e4066Sahrens 1437fa9e4066Sahrens /* write the last block out */ 143867bd71c6Sperrin if (lwb != NULL && lwb->lwb_zio != NULL) 1439fa9e4066Sahrens lwb = zil_lwb_write_start(zilog, lwb); 1440fa9e4066Sahrens 144122ac5be4Sperrin zilog->zl_cur_used = 0; 1442fa9e4066Sahrens 1443fa9e4066Sahrens /* 1444b19a79ecSperrin * Wait if necessary for the log blocks to be on stable storage. 1445fa9e4066Sahrens */ 1446b19a79ecSperrin if (zilog->zl_root_zio) { 1447b24ab676SJeff Bonwick error = zio_wait(zilog->zl_root_zio); 1448e14bb325SJeff Bonwick zilog->zl_root_zio = NULL; 144917f17c2dSbonwick zil_flush_vdevs(zilog); 1450fa9e4066Sahrens } 145122ac5be4Sperrin 1452b24ab676SJeff Bonwick if (error || lwb == NULL) 1453fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 145467bd71c6Sperrin 145567bd71c6Sperrin mutex_enter(&zilog->zl_lock); 1456b24ab676SJeff Bonwick 1457b24ab676SJeff Bonwick /* 1458b24ab676SJeff Bonwick * Remember the highest committed log sequence number for ztest. 1459b24ab676SJeff Bonwick * We only update this value when all the log writes succeeded, 1460b24ab676SJeff Bonwick * because ztest wants to ASSERT that it got the whole log chain. 1461b24ab676SJeff Bonwick */ 1462b24ab676SJeff Bonwick if (error == 0 && lwb != NULL) 1463b24ab676SJeff Bonwick zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1464b19a79ecSperrin } 1465b19a79ecSperrin 1466b19a79ecSperrin /* 14675002558fSNeil Perrin * Commit zfs transactions to stable storage. 1468b19a79ecSperrin * If foid is 0 push out all transactions, otherwise push only those 14695002558fSNeil Perrin * for that object or might reference that object. 14705002558fSNeil Perrin * 14715002558fSNeil Perrin * itxs are committed in batches. In a heavily stressed zil there will be 14725002558fSNeil Perrin * a commit writer thread who is writing out a bunch of itxs to the log 14735002558fSNeil Perrin * for a set of committing threads (cthreads) in the same batch as the writer. 14745002558fSNeil Perrin * Those cthreads are all waiting on the same cv for that batch. 14755002558fSNeil Perrin * 14765002558fSNeil Perrin * There will also be a different and growing batch of threads that are 14775002558fSNeil Perrin * waiting to commit (qthreads). When the committing batch completes 14785002558fSNeil Perrin * a transition occurs such that the cthreads exit and the qthreads become 14795002558fSNeil Perrin * cthreads. One of the new cthreads becomes the writer thread for the 14805002558fSNeil Perrin * batch. Any new threads arriving become new qthreads. 14815002558fSNeil Perrin * 14825002558fSNeil Perrin * Only 2 condition variables are needed and there's no transition 14835002558fSNeil Perrin * between the two cvs needed. They just flip-flop between qthreads 14845002558fSNeil Perrin * and cthreads. 14855002558fSNeil Perrin * 14865002558fSNeil Perrin * Using this scheme we can efficiently wakeup up only those threads 14875002558fSNeil Perrin * that have been committed. 1488b19a79ecSperrin */ 1489b19a79ecSperrin void 14905002558fSNeil Perrin zil_commit(zilog_t *zilog, uint64_t foid) 1491b19a79ecSperrin { 14925002558fSNeil Perrin uint64_t mybatch; 1493b19a79ecSperrin 14945002558fSNeil Perrin if (zilog->zl_sync == ZFS_SYNC_DISABLED) 14955002558fSNeil Perrin return; 1496b19a79ecSperrin 14975002558fSNeil Perrin /* move the async itxs for the foid to the sync queues */ 14985002558fSNeil Perrin zil_async_to_sync(zilog, foid); 1499b19a79ecSperrin 15005002558fSNeil Perrin mutex_enter(&zilog->zl_lock); 15015002558fSNeil Perrin mybatch = zilog->zl_next_batch; 150267bd71c6Sperrin while (zilog->zl_writer) { 15035002558fSNeil Perrin cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 15045002558fSNeil Perrin if (mybatch <= zilog->zl_com_batch) { 150567bd71c6Sperrin mutex_exit(&zilog->zl_lock); 150667bd71c6Sperrin return; 150767bd71c6Sperrin } 150867bd71c6Sperrin } 1509b24ab676SJeff Bonwick 15105002558fSNeil Perrin zilog->zl_next_batch++; 15115002558fSNeil Perrin zilog->zl_writer = B_TRUE; 15125002558fSNeil Perrin zil_commit_writer(zilog); 15135002558fSNeil Perrin zilog->zl_com_batch = mybatch; 15145002558fSNeil Perrin zilog->zl_writer = B_FALSE; 15155002558fSNeil Perrin mutex_exit(&zilog->zl_lock); 1516b24ab676SJeff Bonwick 15175002558fSNeil Perrin /* wake up one thread to become the next writer */ 15185002558fSNeil Perrin cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1519b24ab676SJeff Bonwick 15205002558fSNeil Perrin /* wake up all threads waiting for this batch to be committed */ 15215002558fSNeil Perrin cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1522b24ab676SJeff Bonwick } 1523b24ab676SJeff Bonwick 1524fa9e4066Sahrens /* 1525fa9e4066Sahrens * Called in syncing context to free committed log blocks and update log header. 1526fa9e4066Sahrens */ 1527fa9e4066Sahrens void 1528fa9e4066Sahrens zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1529fa9e4066Sahrens { 1530d80c45e0Sbonwick zil_header_t *zh = zil_header_in_syncing_context(zilog); 1531fa9e4066Sahrens uint64_t txg = dmu_tx_get_txg(tx); 1532fa9e4066Sahrens spa_t *spa = zilog->zl_spa; 1533b24ab676SJeff Bonwick uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1534fa9e4066Sahrens lwb_t *lwb; 1535fa9e4066Sahrens 153614843421SMatthew Ahrens /* 153714843421SMatthew Ahrens * We don't zero out zl_destroy_txg, so make sure we don't try 153814843421SMatthew Ahrens * to destroy it twice. 153914843421SMatthew Ahrens */ 154014843421SMatthew Ahrens if (spa_sync_pass(spa) != 1) 154114843421SMatthew Ahrens return; 154214843421SMatthew Ahrens 1543d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1544d80c45e0Sbonwick 1545fa9e4066Sahrens ASSERT(zilog->zl_stop_sync == 0); 1546fa9e4066Sahrens 1547b24ab676SJeff Bonwick if (*replayed_seq != 0) { 1548b24ab676SJeff Bonwick ASSERT(zh->zh_replay_seq < *replayed_seq); 1549b24ab676SJeff Bonwick zh->zh_replay_seq = *replayed_seq; 1550b24ab676SJeff Bonwick *replayed_seq = 0; 1551b24ab676SJeff Bonwick } 1552fa9e4066Sahrens 1553fa9e4066Sahrens if (zilog->zl_destroy_txg == txg) { 1554d80c45e0Sbonwick blkptr_t blk = zh->zh_log; 1555d80c45e0Sbonwick 1556d80c45e0Sbonwick ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1557d80c45e0Sbonwick 1558d80c45e0Sbonwick bzero(zh, sizeof (zil_header_t)); 15591209a471SNeil Perrin bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1560d80c45e0Sbonwick 1561d80c45e0Sbonwick if (zilog->zl_keep_first) { 1562d80c45e0Sbonwick /* 1563d80c45e0Sbonwick * If this block was part of log chain that couldn't 1564d80c45e0Sbonwick * be claimed because a device was missing during 1565d80c45e0Sbonwick * zil_claim(), but that device later returns, 1566d80c45e0Sbonwick * then this block could erroneously appear valid. 1567d80c45e0Sbonwick * To guard against this, assign a new GUID to the new 1568d80c45e0Sbonwick * log chain so it doesn't matter what blk points to. 1569d80c45e0Sbonwick */ 1570d80c45e0Sbonwick zil_init_log_chain(zilog, &blk); 1571d80c45e0Sbonwick zh->zh_log = blk; 1572d80c45e0Sbonwick } 1573fa9e4066Sahrens } 1574fa9e4066Sahrens 1575e6ca193dSGeorge Wilson while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1576b19a79ecSperrin zh->zh_log = lwb->lwb_blk; 1577fa9e4066Sahrens if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1578fa9e4066Sahrens break; 1579fa9e4066Sahrens list_remove(&zilog->zl_lwb_list, lwb); 1580b24ab676SJeff Bonwick zio_free_zil(spa, txg, &lwb->lwb_blk); 1581fa9e4066Sahrens kmem_cache_free(zil_lwb_cache, lwb); 1582d63d470bSgw 1583d63d470bSgw /* 1584d63d470bSgw * If we don't have anything left in the lwb list then 1585d63d470bSgw * we've had an allocation failure and we need to zero 1586d63d470bSgw * out the zil_header blkptr so that we don't end 1587d63d470bSgw * up freeing the same block twice. 1588d63d470bSgw */ 1589d63d470bSgw if (list_head(&zilog->zl_lwb_list) == NULL) 1590d63d470bSgw BP_ZERO(&zh->zh_log); 1591fa9e4066Sahrens } 1592fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1593fa9e4066Sahrens } 1594fa9e4066Sahrens 1595fa9e4066Sahrens void 1596fa9e4066Sahrens zil_init(void) 1597fa9e4066Sahrens { 1598fa9e4066Sahrens zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 15995ad82045Snd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1600fa9e4066Sahrens } 1601fa9e4066Sahrens 1602fa9e4066Sahrens void 1603fa9e4066Sahrens zil_fini(void) 1604fa9e4066Sahrens { 1605fa9e4066Sahrens kmem_cache_destroy(zil_lwb_cache); 1606fa9e4066Sahrens } 1607fa9e4066Sahrens 160855da60b9SMark J Musante void 160955da60b9SMark J Musante zil_set_sync(zilog_t *zilog, uint64_t sync) 161055da60b9SMark J Musante { 161155da60b9SMark J Musante zilog->zl_sync = sync; 161255da60b9SMark J Musante } 161355da60b9SMark J Musante 1614e09fa4daSNeil Perrin void 1615e09fa4daSNeil Perrin zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1616e09fa4daSNeil Perrin { 1617e09fa4daSNeil Perrin zilog->zl_logbias = logbias; 1618e09fa4daSNeil Perrin } 1619e09fa4daSNeil Perrin 1620fa9e4066Sahrens zilog_t * 1621fa9e4066Sahrens zil_alloc(objset_t *os, zil_header_t *zh_phys) 1622fa9e4066Sahrens { 1623fa9e4066Sahrens zilog_t *zilog; 1624fa9e4066Sahrens 1625fa9e4066Sahrens zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1626fa9e4066Sahrens 1627fa9e4066Sahrens zilog->zl_header = zh_phys; 1628fa9e4066Sahrens zilog->zl_os = os; 1629fa9e4066Sahrens zilog->zl_spa = dmu_objset_spa(os); 1630fa9e4066Sahrens zilog->zl_dmu_pool = dmu_objset_pool(os); 1631d80c45e0Sbonwick zilog->zl_destroy_txg = TXG_INITIAL - 1; 1632e09fa4daSNeil Perrin zilog->zl_logbias = dmu_objset_logbias(os); 163355da60b9SMark J Musante zilog->zl_sync = dmu_objset_syncprop(os); 16345002558fSNeil Perrin zilog->zl_next_batch = 1; 1635fa9e4066Sahrens 16365ad82045Snd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 16375ad82045Snd 16385002558fSNeil Perrin for (int i = 0; i < TXG_SIZE; i++) { 16395002558fSNeil Perrin mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 16405002558fSNeil Perrin MUTEX_DEFAULT, NULL); 16415002558fSNeil Perrin } 1642fa9e4066Sahrens 1643fa9e4066Sahrens list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1644fa9e4066Sahrens offsetof(lwb_t, lwb_node)); 1645fa9e4066Sahrens 16465002558fSNeil Perrin list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 16475002558fSNeil Perrin offsetof(itx_t, itx_node)); 16485002558fSNeil Perrin 164917f17c2dSbonwick mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 165017f17c2dSbonwick 165117f17c2dSbonwick avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 165217f17c2dSbonwick sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1653fa9e4066Sahrens 1654b7b97454Sperrin cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1655b7b97454Sperrin cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 16565002558fSNeil Perrin cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 16575002558fSNeil Perrin cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1658b7b97454Sperrin 1659fa9e4066Sahrens return (zilog); 1660fa9e4066Sahrens } 1661fa9e4066Sahrens 1662fa9e4066Sahrens void 1663fa9e4066Sahrens zil_free(zilog_t *zilog) 1664fa9e4066Sahrens { 1665fa9e4066Sahrens zilog->zl_stop_sync = 1; 1666fa9e4066Sahrens 1667*c9ba2a43SEric Schrock ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1668fa9e4066Sahrens list_destroy(&zilog->zl_lwb_list); 1669fa9e4066Sahrens 167017f17c2dSbonwick avl_destroy(&zilog->zl_vdev_tree); 167117f17c2dSbonwick mutex_destroy(&zilog->zl_vdev_lock); 1672fa9e4066Sahrens 16735002558fSNeil Perrin ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 16745002558fSNeil Perrin list_destroy(&zilog->zl_itx_commit_list); 16755002558fSNeil Perrin 16765002558fSNeil Perrin for (int i = 0; i < TXG_SIZE; i++) { 16775002558fSNeil Perrin /* 16785002558fSNeil Perrin * It's possible for an itx to be generated that doesn't dirty 16795002558fSNeil Perrin * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 16805002558fSNeil Perrin * callback to remove the entry. We remove those here. 16815002558fSNeil Perrin * 16825002558fSNeil Perrin * Also free up the ziltest itxs. 16835002558fSNeil Perrin */ 16845002558fSNeil Perrin if (zilog->zl_itxg[i].itxg_itxs) 16855002558fSNeil Perrin zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 16865002558fSNeil Perrin mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 16875002558fSNeil Perrin } 16885002558fSNeil Perrin 16895ad82045Snd mutex_destroy(&zilog->zl_lock); 1690fa9e4066Sahrens 1691b7b97454Sperrin cv_destroy(&zilog->zl_cv_writer); 1692b7b97454Sperrin cv_destroy(&zilog->zl_cv_suspend); 16935002558fSNeil Perrin cv_destroy(&zilog->zl_cv_batch[0]); 16945002558fSNeil Perrin cv_destroy(&zilog->zl_cv_batch[1]); 1695b7b97454Sperrin 1696fa9e4066Sahrens kmem_free(zilog, sizeof (zilog_t)); 1697fa9e4066Sahrens } 1698fa9e4066Sahrens 1699fa9e4066Sahrens /* 1700fa9e4066Sahrens * Open an intent log. 1701fa9e4066Sahrens */ 1702fa9e4066Sahrens zilog_t * 1703fa9e4066Sahrens zil_open(objset_t *os, zil_get_data_t *get_data) 1704fa9e4066Sahrens { 1705fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1706fa9e4066Sahrens 1707*c9ba2a43SEric Schrock ASSERT(zilog->zl_clean_taskq == NULL); 1708*c9ba2a43SEric Schrock ASSERT(zilog->zl_get_data == NULL); 1709*c9ba2a43SEric Schrock ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1710*c9ba2a43SEric Schrock 1711fa9e4066Sahrens zilog->zl_get_data = get_data; 1712fa9e4066Sahrens zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1713fa9e4066Sahrens 2, 2, TASKQ_PREPOPULATE); 1714fa9e4066Sahrens 1715fa9e4066Sahrens return (zilog); 1716fa9e4066Sahrens } 1717fa9e4066Sahrens 1718fa9e4066Sahrens /* 1719fa9e4066Sahrens * Close an intent log. 1720fa9e4066Sahrens */ 1721fa9e4066Sahrens void 1722fa9e4066Sahrens zil_close(zilog_t *zilog) 1723fa9e4066Sahrens { 1724*c9ba2a43SEric Schrock lwb_t *lwb; 17255002558fSNeil Perrin uint64_t txg = 0; 17265002558fSNeil Perrin 17275002558fSNeil Perrin zil_commit(zilog, 0); /* commit all itx */ 17285002558fSNeil Perrin 1729d80c45e0Sbonwick /* 17305002558fSNeil Perrin * The lwb_max_txg for the stubby lwb will reflect the last activity 17315002558fSNeil Perrin * for the zil. After a txg_wait_synced() on the txg we know all the 17325002558fSNeil Perrin * callbacks have occurred that may clean the zil. Only then can we 17335002558fSNeil Perrin * destroy the zl_clean_taskq. 1734d80c45e0Sbonwick */ 17355002558fSNeil Perrin mutex_enter(&zilog->zl_lock); 1736*c9ba2a43SEric Schrock lwb = list_tail(&zilog->zl_lwb_list); 1737*c9ba2a43SEric Schrock if (lwb != NULL) 1738*c9ba2a43SEric Schrock txg = lwb->lwb_max_txg; 17395002558fSNeil Perrin mutex_exit(&zilog->zl_lock); 17405002558fSNeil Perrin if (txg) 1741d80c45e0Sbonwick txg_wait_synced(zilog->zl_dmu_pool, txg); 1742d80c45e0Sbonwick 1743fa9e4066Sahrens taskq_destroy(zilog->zl_clean_taskq); 1744fa9e4066Sahrens zilog->zl_clean_taskq = NULL; 1745fa9e4066Sahrens zilog->zl_get_data = NULL; 1746*c9ba2a43SEric Schrock 1747*c9ba2a43SEric Schrock /* 1748*c9ba2a43SEric Schrock * We should have only one LWB left on the list; remove it now. 1749*c9ba2a43SEric Schrock */ 1750*c9ba2a43SEric Schrock mutex_enter(&zilog->zl_lock); 1751*c9ba2a43SEric Schrock lwb = list_head(&zilog->zl_lwb_list); 1752*c9ba2a43SEric Schrock if (lwb != NULL) { 1753*c9ba2a43SEric Schrock ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1754*c9ba2a43SEric Schrock list_remove(&zilog->zl_lwb_list, lwb); 1755*c9ba2a43SEric Schrock zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1756*c9ba2a43SEric Schrock kmem_cache_free(zil_lwb_cache, lwb); 1757*c9ba2a43SEric Schrock } 1758*c9ba2a43SEric Schrock mutex_exit(&zilog->zl_lock); 1759fa9e4066Sahrens } 1760fa9e4066Sahrens 1761fa9e4066Sahrens /* 1762fa9e4066Sahrens * Suspend an intent log. While in suspended mode, we still honor 1763fa9e4066Sahrens * synchronous semantics, but we rely on txg_wait_synced() to do it. 1764fa9e4066Sahrens * We suspend the log briefly when taking a snapshot so that the snapshot 1765fa9e4066Sahrens * contains all the data it's supposed to, and has an empty intent log. 1766fa9e4066Sahrens */ 1767fa9e4066Sahrens int 1768fa9e4066Sahrens zil_suspend(zilog_t *zilog) 1769fa9e4066Sahrens { 1770d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1771fa9e4066Sahrens 1772fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 17733589c4f0SNeil Perrin if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1774fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1775fa9e4066Sahrens return (EBUSY); 1776fa9e4066Sahrens } 1777d80c45e0Sbonwick if (zilog->zl_suspend++ != 0) { 1778d80c45e0Sbonwick /* 1779d80c45e0Sbonwick * Someone else already began a suspend. 1780d80c45e0Sbonwick * Just wait for them to finish. 1781d80c45e0Sbonwick */ 1782d80c45e0Sbonwick while (zilog->zl_suspending) 1783d80c45e0Sbonwick cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1784d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1785d80c45e0Sbonwick return (0); 1786d80c45e0Sbonwick } 1787d80c45e0Sbonwick zilog->zl_suspending = B_TRUE; 1788fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1789fa9e4066Sahrens 17905002558fSNeil Perrin zil_commit(zilog, 0); 1791fa9e4066Sahrens 1792d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1793d80c45e0Sbonwick 1794d80c45e0Sbonwick mutex_enter(&zilog->zl_lock); 1795d80c45e0Sbonwick zilog->zl_suspending = B_FALSE; 1796d80c45e0Sbonwick cv_broadcast(&zilog->zl_cv_suspend); 1797d80c45e0Sbonwick mutex_exit(&zilog->zl_lock); 1798fa9e4066Sahrens 1799fa9e4066Sahrens return (0); 1800fa9e4066Sahrens } 1801fa9e4066Sahrens 1802fa9e4066Sahrens void 1803fa9e4066Sahrens zil_resume(zilog_t *zilog) 1804fa9e4066Sahrens { 1805fa9e4066Sahrens mutex_enter(&zilog->zl_lock); 1806fa9e4066Sahrens ASSERT(zilog->zl_suspend != 0); 1807fa9e4066Sahrens zilog->zl_suspend--; 1808fa9e4066Sahrens mutex_exit(&zilog->zl_lock); 1809fa9e4066Sahrens } 1810fa9e4066Sahrens 1811fa9e4066Sahrens typedef struct zil_replay_arg { 1812fa9e4066Sahrens zil_replay_func_t **zr_replay; 1813fa9e4066Sahrens void *zr_arg; 1814fa9e4066Sahrens boolean_t zr_byteswap; 1815b24ab676SJeff Bonwick char *zr_lr; 1816fa9e4066Sahrens } zil_replay_arg_t; 1817fa9e4066Sahrens 1818b24ab676SJeff Bonwick static int 1819b24ab676SJeff Bonwick zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1820b24ab676SJeff Bonwick { 1821b24ab676SJeff Bonwick char name[MAXNAMELEN]; 1822b24ab676SJeff Bonwick 1823b24ab676SJeff Bonwick zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1824b24ab676SJeff Bonwick 1825b24ab676SJeff Bonwick dmu_objset_name(zilog->zl_os, name); 1826b24ab676SJeff Bonwick 1827b24ab676SJeff Bonwick cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1828b24ab676SJeff Bonwick "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1829b24ab676SJeff Bonwick (u_longlong_t)lr->lrc_seq, 1830b24ab676SJeff Bonwick (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1831b24ab676SJeff Bonwick (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1832b24ab676SJeff Bonwick 1833b24ab676SJeff Bonwick return (error); 1834b24ab676SJeff Bonwick } 1835b24ab676SJeff Bonwick 1836b24ab676SJeff Bonwick static int 1837fa9e4066Sahrens zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1838fa9e4066Sahrens { 1839fa9e4066Sahrens zil_replay_arg_t *zr = zra; 1840d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1841fa9e4066Sahrens uint64_t reclen = lr->lrc_reclen; 1842fa9e4066Sahrens uint64_t txtype = lr->lrc_txtype; 1843b24ab676SJeff Bonwick int error = 0; 1844fa9e4066Sahrens 1845b24ab676SJeff Bonwick zilog->zl_replaying_seq = lr->lrc_seq; 1846fa9e4066Sahrens 1847fa9e4066Sahrens if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1848b24ab676SJeff Bonwick return (0); 1849b24ab676SJeff Bonwick 1850b24ab676SJeff Bonwick if (lr->lrc_txg < claim_txg) /* already committed */ 1851b24ab676SJeff Bonwick return (0); 1852fa9e4066Sahrens 1853da6c28aaSamw /* Strip case-insensitive bit, still present in log record */ 1854da6c28aaSamw txtype &= ~TX_CI; 1855da6c28aaSamw 1856b24ab676SJeff Bonwick if (txtype == 0 || txtype >= TX_MAX_TYPE) 1857b24ab676SJeff Bonwick return (zil_replay_error(zilog, lr, EINVAL)); 1858b24ab676SJeff Bonwick 1859b24ab676SJeff Bonwick /* 1860b24ab676SJeff Bonwick * If this record type can be logged out of order, the object 1861b24ab676SJeff Bonwick * (lr_foid) may no longer exist. That's legitimate, not an error. 1862b24ab676SJeff Bonwick */ 1863b24ab676SJeff Bonwick if (TX_OOO(txtype)) { 1864b24ab676SJeff Bonwick error = dmu_object_info(zilog->zl_os, 1865b24ab676SJeff Bonwick ((lr_ooo_t *)lr)->lr_foid, NULL); 1866b24ab676SJeff Bonwick if (error == ENOENT || error == EEXIST) 1867b24ab676SJeff Bonwick return (0); 18681209a471SNeil Perrin } 18691209a471SNeil Perrin 1870fa9e4066Sahrens /* 1871fa9e4066Sahrens * Make a copy of the data so we can revise and extend it. 1872fa9e4066Sahrens */ 1873b24ab676SJeff Bonwick bcopy(lr, zr->zr_lr, reclen); 1874b24ab676SJeff Bonwick 1875b24ab676SJeff Bonwick /* 1876b24ab676SJeff Bonwick * If this is a TX_WRITE with a blkptr, suck in the data. 1877b24ab676SJeff Bonwick */ 1878b24ab676SJeff Bonwick if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1879b24ab676SJeff Bonwick error = zil_read_log_data(zilog, (lr_write_t *)lr, 1880b24ab676SJeff Bonwick zr->zr_lr + reclen); 1881b24ab676SJeff Bonwick if (error) 1882b24ab676SJeff Bonwick return (zil_replay_error(zilog, lr, error)); 1883b24ab676SJeff Bonwick } 1884fa9e4066Sahrens 1885fa9e4066Sahrens /* 1886fa9e4066Sahrens * The log block containing this lr may have been byteswapped 1887fa9e4066Sahrens * so that we can easily examine common fields like lrc_txtype. 1888b24ab676SJeff Bonwick * However, the log is a mix of different record types, and only the 1889fa9e4066Sahrens * replay vectors know how to byteswap their records. Therefore, if 1890fa9e4066Sahrens * the lr was byteswapped, undo it before invoking the replay vector. 1891fa9e4066Sahrens */ 1892fa9e4066Sahrens if (zr->zr_byteswap) 1893b24ab676SJeff Bonwick byteswap_uint64_array(zr->zr_lr, reclen); 1894fa9e4066Sahrens 1895fa9e4066Sahrens /* 1896fa9e4066Sahrens * We must now do two things atomically: replay this log record, 18971209a471SNeil Perrin * and update the log header sequence number to reflect the fact that 18981209a471SNeil Perrin * we did so. At the end of each replay function the sequence number 18991209a471SNeil Perrin * is updated if we are in replay mode. 1900fa9e4066Sahrens */ 1901b24ab676SJeff Bonwick error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 1902b24ab676SJeff Bonwick if (error) { 190367bd71c6Sperrin /* 190467bd71c6Sperrin * The DMU's dnode layer doesn't see removes until the txg 190567bd71c6Sperrin * commits, so a subsequent claim can spuriously fail with 19061209a471SNeil Perrin * EEXIST. So if we receive any error we try syncing out 1907b24ab676SJeff Bonwick * any removes then retry the transaction. Note that we 1908b24ab676SJeff Bonwick * specify B_FALSE for byteswap now, so we don't do it twice. 190967bd71c6Sperrin */ 1910b24ab676SJeff Bonwick txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1911b24ab676SJeff Bonwick error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 1912b24ab676SJeff Bonwick if (error) 1913b24ab676SJeff Bonwick return (zil_replay_error(zilog, lr, error)); 1914fa9e4066Sahrens } 1915b24ab676SJeff Bonwick return (0); 191667bd71c6Sperrin } 1917fa9e4066Sahrens 191867bd71c6Sperrin /* ARGSUSED */ 1919b24ab676SJeff Bonwick static int 192067bd71c6Sperrin zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 192167bd71c6Sperrin { 192267bd71c6Sperrin zilog->zl_replay_blks++; 1923b24ab676SJeff Bonwick 1924b24ab676SJeff Bonwick return (0); 1925fa9e4066Sahrens } 1926fa9e4066Sahrens 1927fa9e4066Sahrens /* 192813f5297eSperrin * If this dataset has a non-empty intent log, replay it and destroy it. 1929fa9e4066Sahrens */ 1930fa9e4066Sahrens void 19311209a471SNeil Perrin zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1932fa9e4066Sahrens { 1933fa9e4066Sahrens zilog_t *zilog = dmu_objset_zil(os); 1934d80c45e0Sbonwick const zil_header_t *zh = zilog->zl_header; 1935d80c45e0Sbonwick zil_replay_arg_t zr; 193613f5297eSperrin 19373589c4f0SNeil Perrin if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1938d80c45e0Sbonwick zil_destroy(zilog, B_TRUE); 193913f5297eSperrin return; 194013f5297eSperrin } 1941fa9e4066Sahrens 1942fa9e4066Sahrens zr.zr_replay = replay_func; 1943fa9e4066Sahrens zr.zr_arg = arg; 1944d80c45e0Sbonwick zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1945b24ab676SJeff Bonwick zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1946fa9e4066Sahrens 1947fa9e4066Sahrens /* 1948fa9e4066Sahrens * Wait for in-progress removes to sync before starting replay. 1949fa9e4066Sahrens */ 1950fa9e4066Sahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1951fa9e4066Sahrens 19521209a471SNeil Perrin zilog->zl_replay = B_TRUE; 1953d3d50737SRafael Vanoni zilog->zl_replay_time = ddi_get_lbolt(); 195467bd71c6Sperrin ASSERT(zilog->zl_replay_blks == 0); 195567bd71c6Sperrin (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1956d80c45e0Sbonwick zh->zh_claim_txg); 1957b24ab676SJeff Bonwick kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 1958fa9e4066Sahrens 1959d80c45e0Sbonwick zil_destroy(zilog, B_FALSE); 1960a4611edeSahrens txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 19611209a471SNeil Perrin zilog->zl_replay = B_FALSE; 1962fa9e4066Sahrens } 1963436b2950Sperrin 1964b24ab676SJeff Bonwick boolean_t 1965b24ab676SJeff Bonwick zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 1966436b2950Sperrin { 196755da60b9SMark J Musante if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1968b24ab676SJeff Bonwick return (B_TRUE); 1969436b2950Sperrin 1970b24ab676SJeff Bonwick if (zilog->zl_replay) { 1971b24ab676SJeff Bonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1972b24ab676SJeff Bonwick zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 1973b24ab676SJeff Bonwick zilog->zl_replaying_seq; 1974b24ab676SJeff Bonwick return (B_TRUE); 1975b19a79ecSperrin } 1976b19a79ecSperrin 1977b24ab676SJeff Bonwick return (B_FALSE); 1978436b2950Sperrin } 1979e6ca193dSGeorge Wilson 1980e6ca193dSGeorge Wilson /* ARGSUSED */ 1981e6ca193dSGeorge Wilson int 1982fd136879SMatthew Ahrens zil_vdev_offline(const char *osname, void *arg) 1983e6ca193dSGeorge Wilson { 1984e6ca193dSGeorge Wilson objset_t *os; 1985e6ca193dSGeorge Wilson zilog_t *zilog; 1986e6ca193dSGeorge Wilson int error; 1987e6ca193dSGeorge Wilson 1988503ad85cSMatthew Ahrens error = dmu_objset_hold(osname, FTAG, &os); 1989e6ca193dSGeorge Wilson if (error) 1990e6ca193dSGeorge Wilson return (error); 1991e6ca193dSGeorge Wilson 1992e6ca193dSGeorge Wilson zilog = dmu_objset_zil(os); 1993e6ca193dSGeorge Wilson if (zil_suspend(zilog) != 0) 1994e6ca193dSGeorge Wilson error = EEXIST; 1995e6ca193dSGeorge Wilson else 1996e6ca193dSGeorge Wilson zil_resume(zilog); 1997503ad85cSMatthew Ahrens dmu_objset_rele(os, FTAG); 1998e6ca193dSGeorge Wilson return (error); 1999e6ca193dSGeorge Wilson } 2000