zil.c revision 243674
1168404Spjd/* 2168404Spjd * CDDL HEADER START 3168404Spjd * 4168404Spjd * The contents of this file are subject to the terms of the 5168404Spjd * Common Development and Distribution License (the "License"). 6168404Spjd * You may not use this file except in compliance with the License. 7168404Spjd * 8168404Spjd * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9168404Spjd * or http://www.opensolaris.org/os/licensing. 10168404Spjd * See the License for the specific language governing permissions 11168404Spjd * and limitations under the License. 12168404Spjd * 13168404Spjd * When distributing Covered Code, include this CDDL HEADER in each 14168404Spjd * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15168404Spjd * If applicable, add the following below this CDDL HEADER, with the 16168404Spjd * fields enclosed by brackets "[]" replaced with your own identifying 17168404Spjd * information: Portions Copyright [yyyy] [name of copyright owner] 18168404Spjd * 19168404Spjd * CDDL HEADER END 20168404Spjd */ 21168404Spjd/* 22219089Spjd * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23243674Smm * Copyright (c) 2012 by Delphix. All rights reserved. 24168404Spjd */ 25168404Spjd 26219089Spjd/* Portions Copyright 2010 Robert Milkowski */ 27219089Spjd 28168404Spjd#include <sys/zfs_context.h> 29168404Spjd#include <sys/spa.h> 30168404Spjd#include <sys/dmu.h> 31168404Spjd#include <sys/zap.h> 32168404Spjd#include <sys/arc.h> 33168404Spjd#include <sys/stat.h> 34168404Spjd#include <sys/resource.h> 35168404Spjd#include <sys/zil.h> 36168404Spjd#include <sys/zil_impl.h> 37168404Spjd#include <sys/dsl_dataset.h> 38219089Spjd#include <sys/vdev_impl.h> 39168404Spjd#include <sys/dmu_tx.h> 40219089Spjd#include <sys/dsl_pool.h> 41168404Spjd 42168404Spjd/* 43168404Spjd * The zfs intent log (ZIL) saves transaction records of system calls 44168404Spjd * that change the file system in memory with enough information 45168404Spjd * to be able to replay them. These are stored in memory until 46168404Spjd * either the DMU transaction group (txg) commits them to the stable pool 47168404Spjd * and they can be discarded, or they are flushed to the stable log 48168404Spjd * (also in the pool) due to a fsync, O_DSYNC or other synchronous 49168404Spjd * requirement. In the event of a panic or power fail then those log 50168404Spjd * records (transactions) are replayed. 51168404Spjd * 52168404Spjd * There is one ZIL per file system. Its on-disk (pool) format consists 53168404Spjd * of 3 parts: 54168404Spjd * 55168404Spjd * - ZIL header 56168404Spjd * - ZIL blocks 57168404Spjd * - ZIL records 58168404Spjd * 59168404Spjd * A log record holds a system call transaction. Log blocks can 60168404Spjd * hold many log records and the blocks are chained together. 61168404Spjd * Each ZIL block contains a block pointer (blkptr_t) to the next 62168404Spjd * ZIL block in the chain. The ZIL header points to the first 63168404Spjd * block in the chain. Note there is not a fixed place in the pool 64168404Spjd * to hold blocks. They are dynamically allocated and freed as 65168404Spjd * needed from the blocks available. Figure X shows the ZIL structure: 66168404Spjd */ 67168404Spjd 68168404Spjd/* 69168404Spjd * This global ZIL switch affects all pools 70168404Spjd */ 71219089Spjdint zil_replay_disable = 0; /* disable intent logging replay */ 72168404SpjdSYSCTL_DECL(_vfs_zfs); 73219089SpjdTUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 74219089SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 75219089Spjd &zil_replay_disable, 0, "Disable intent logging replay"); 76168404Spjd 77168404Spjd/* 78168404Spjd * Tunable parameter for debugging or performance analysis. Setting 79168404Spjd * zfs_nocacheflush will cause corruption on power loss if a volatile 80168404Spjd * out-of-order write cache is enabled. 81168404Spjd */ 82168404Spjdboolean_t zfs_nocacheflush = B_FALSE; 83168404SpjdTUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 84168404SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 85168404Spjd &zfs_nocacheflush, 0, "Disable cache flush"); 86168404Spjd 87168404Spjdstatic kmem_cache_t *zil_lwb_cache; 88168404Spjd 89219089Spjdstatic void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 90219089Spjd 91219089Spjd#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 92219089Spjd sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 93219089Spjd 94219089Spjd 95219089Spjd/* 96219089Spjd * ziltest is by and large an ugly hack, but very useful in 97219089Spjd * checking replay without tedious work. 98219089Spjd * When running ziltest we want to keep all itx's and so maintain 99219089Spjd * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 100219089Spjd * We subtract TXG_CONCURRENT_STATES to allow for common code. 101219089Spjd */ 102219089Spjd#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 103219089Spjd 104168404Spjdstatic int 105219089Spjdzil_bp_compare(const void *x1, const void *x2) 106168404Spjd{ 107219089Spjd const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 108219089Spjd const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 109168404Spjd 110168404Spjd if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 111168404Spjd return (-1); 112168404Spjd if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 113168404Spjd return (1); 114168404Spjd 115168404Spjd if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 116168404Spjd return (-1); 117168404Spjd if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 118168404Spjd return (1); 119168404Spjd 120168404Spjd return (0); 121168404Spjd} 122168404Spjd 123168404Spjdstatic void 124219089Spjdzil_bp_tree_init(zilog_t *zilog) 125168404Spjd{ 126219089Spjd avl_create(&zilog->zl_bp_tree, zil_bp_compare, 127219089Spjd sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 128168404Spjd} 129168404Spjd 130168404Spjdstatic void 131219089Spjdzil_bp_tree_fini(zilog_t *zilog) 132168404Spjd{ 133219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 134219089Spjd zil_bp_node_t *zn; 135168404Spjd void *cookie = NULL; 136168404Spjd 137168404Spjd while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 138219089Spjd kmem_free(zn, sizeof (zil_bp_node_t)); 139168404Spjd 140168404Spjd avl_destroy(t); 141168404Spjd} 142168404Spjd 143219089Spjdint 144219089Spjdzil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 145168404Spjd{ 146219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 147219089Spjd const dva_t *dva = BP_IDENTITY(bp); 148219089Spjd zil_bp_node_t *zn; 149168404Spjd avl_index_t where; 150168404Spjd 151168404Spjd if (avl_find(t, dva, &where) != NULL) 152168404Spjd return (EEXIST); 153168404Spjd 154219089Spjd zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 155168404Spjd zn->zn_dva = *dva; 156168404Spjd avl_insert(t, zn, where); 157168404Spjd 158168404Spjd return (0); 159168404Spjd} 160168404Spjd 161168404Spjdstatic zil_header_t * 162168404Spjdzil_header_in_syncing_context(zilog_t *zilog) 163168404Spjd{ 164168404Spjd return ((zil_header_t *)zilog->zl_header); 165168404Spjd} 166168404Spjd 167168404Spjdstatic void 168168404Spjdzil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 169168404Spjd{ 170168404Spjd zio_cksum_t *zc = &bp->blk_cksum; 171168404Spjd 172168404Spjd zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 173168404Spjd zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 174168404Spjd zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 175168404Spjd zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 176168404Spjd} 177168404Spjd 178168404Spjd/* 179219089Spjd * Read a log block and make sure it's valid. 180168404Spjd */ 181168404Spjdstatic int 182219089Spjdzil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 183219089Spjd char **end) 184168404Spjd{ 185219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 186219089Spjd uint32_t aflags = ARC_WAIT; 187219089Spjd arc_buf_t *abuf = NULL; 188168404Spjd zbookmark_t zb; 189168404Spjd int error; 190168404Spjd 191219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 192219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 193168404Spjd 194219089Spjd if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 195219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE; 196168404Spjd 197219089Spjd SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 198219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 199168404Spjd 200219089Spjd error = dsl_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 201219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 202219089Spjd 203168404Spjd if (error == 0) { 204168404Spjd zio_cksum_t cksum = bp->blk_cksum; 205168404Spjd 206168404Spjd /* 207185029Spjd * Validate the checksummed log block. 208185029Spjd * 209168404Spjd * Sequence numbers should be... sequential. The checksum 210168404Spjd * verifier for the next block should be bp's checksum plus 1. 211185029Spjd * 212185029Spjd * Also check the log chain linkage and size used. 213168404Spjd */ 214168404Spjd cksum.zc_word[ZIL_ZC_SEQ]++; 215168404Spjd 216219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 217219089Spjd zil_chain_t *zilc = abuf->b_data; 218219089Spjd char *lr = (char *)(zilc + 1); 219219089Spjd uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 220219089Spjd 221219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 222219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 223219089Spjd error = ECKSUM; 224219089Spjd } else { 225219089Spjd bcopy(lr, dst, len); 226219089Spjd *end = (char *)dst + len; 227219089Spjd *nbp = zilc->zc_next_blk; 228219089Spjd } 229219089Spjd } else { 230219089Spjd char *lr = abuf->b_data; 231219089Spjd uint64_t size = BP_GET_LSIZE(bp); 232219089Spjd zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 233219089Spjd 234219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 235219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 236219089Spjd (zilc->zc_nused > (size - sizeof (*zilc)))) { 237219089Spjd error = ECKSUM; 238219089Spjd } else { 239219089Spjd bcopy(lr, dst, zilc->zc_nused); 240219089Spjd *end = (char *)dst + zilc->zc_nused; 241219089Spjd *nbp = zilc->zc_next_blk; 242219089Spjd } 243185029Spjd } 244168404Spjd 245219089Spjd VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 246168404Spjd } 247168404Spjd 248219089Spjd return (error); 249219089Spjd} 250168404Spjd 251219089Spjd/* 252219089Spjd * Read a TX_WRITE log data block. 253219089Spjd */ 254219089Spjdstatic int 255219089Spjdzil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 256219089Spjd{ 257219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 258219089Spjd const blkptr_t *bp = &lr->lr_blkptr; 259219089Spjd uint32_t aflags = ARC_WAIT; 260219089Spjd arc_buf_t *abuf = NULL; 261219089Spjd zbookmark_t zb; 262219089Spjd int error; 263219089Spjd 264219089Spjd if (BP_IS_HOLE(bp)) { 265219089Spjd if (wbuf != NULL) 266219089Spjd bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 267219089Spjd return (0); 268219089Spjd } 269219089Spjd 270219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 271219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 272219089Spjd 273219089Spjd SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 274219089Spjd ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 275219089Spjd 276219089Spjd error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 277219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 278219089Spjd 279219089Spjd if (error == 0) { 280219089Spjd if (wbuf != NULL) 281219089Spjd bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 282219089Spjd (void) arc_buf_remove_ref(abuf, &abuf); 283219089Spjd } 284219089Spjd 285168404Spjd return (error); 286168404Spjd} 287168404Spjd 288168404Spjd/* 289168404Spjd * Parse the intent log, and call parse_func for each valid record within. 290168404Spjd */ 291219089Spjdint 292168404Spjdzil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 293168404Spjd zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 294168404Spjd{ 295168404Spjd const zil_header_t *zh = zilog->zl_header; 296219089Spjd boolean_t claimed = !!zh->zh_claim_txg; 297219089Spjd uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 298219089Spjd uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 299219089Spjd uint64_t max_blk_seq = 0; 300219089Spjd uint64_t max_lr_seq = 0; 301219089Spjd uint64_t blk_count = 0; 302219089Spjd uint64_t lr_count = 0; 303219089Spjd blkptr_t blk, next_blk; 304168404Spjd char *lrbuf, *lrp; 305219089Spjd int error = 0; 306168404Spjd 307219089Spjd /* 308219089Spjd * Old logs didn't record the maximum zh_claim_lr_seq. 309219089Spjd */ 310219089Spjd if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 311219089Spjd claim_lr_seq = UINT64_MAX; 312168404Spjd 313168404Spjd /* 314168404Spjd * Starting at the block pointed to by zh_log we read the log chain. 315168404Spjd * For each block in the chain we strongly check that block to 316168404Spjd * ensure its validity. We stop when an invalid block is found. 317168404Spjd * For each block pointer in the chain we call parse_blk_func(). 318168404Spjd * For each record in each valid block we call parse_lr_func(). 319168404Spjd * If the log has been claimed, stop if we encounter a sequence 320168404Spjd * number greater than the highest claimed sequence number. 321168404Spjd */ 322219089Spjd lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 323219089Spjd zil_bp_tree_init(zilog); 324168404Spjd 325219089Spjd for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 326219089Spjd uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 327219089Spjd int reclen; 328219089Spjd char *end; 329219089Spjd 330219089Spjd if (blk_seq > claim_blk_seq) 331168404Spjd break; 332219089Spjd if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 333219089Spjd break; 334219089Spjd ASSERT3U(max_blk_seq, <, blk_seq); 335219089Spjd max_blk_seq = blk_seq; 336219089Spjd blk_count++; 337168404Spjd 338219089Spjd if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 339219089Spjd break; 340168404Spjd 341219089Spjd error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 342168404Spjd if (error) 343168404Spjd break; 344168404Spjd 345219089Spjd for (lrp = lrbuf; lrp < end; lrp += reclen) { 346168404Spjd lr_t *lr = (lr_t *)lrp; 347168404Spjd reclen = lr->lrc_reclen; 348168404Spjd ASSERT3U(reclen, >=, sizeof (lr_t)); 349219089Spjd if (lr->lrc_seq > claim_lr_seq) 350219089Spjd goto done; 351219089Spjd if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 352219089Spjd goto done; 353219089Spjd ASSERT3U(max_lr_seq, <, lr->lrc_seq); 354219089Spjd max_lr_seq = lr->lrc_seq; 355219089Spjd lr_count++; 356168404Spjd } 357168404Spjd } 358219089Spjddone: 359219089Spjd zilog->zl_parse_error = error; 360219089Spjd zilog->zl_parse_blk_seq = max_blk_seq; 361219089Spjd zilog->zl_parse_lr_seq = max_lr_seq; 362219089Spjd zilog->zl_parse_blk_count = blk_count; 363219089Spjd zilog->zl_parse_lr_count = lr_count; 364168404Spjd 365219089Spjd ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 366219089Spjd (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 367219089Spjd 368219089Spjd zil_bp_tree_fini(zilog); 369219089Spjd zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 370219089Spjd 371219089Spjd return (error); 372168404Spjd} 373168404Spjd 374219089Spjdstatic int 375168404Spjdzil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 376168404Spjd{ 377168404Spjd /* 378168404Spjd * Claim log block if not already committed and not already claimed. 379219089Spjd * If tx == NULL, just verify that the block is claimable. 380168404Spjd */ 381219089Spjd if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 382219089Spjd return (0); 383219089Spjd 384219089Spjd return (zio_wait(zio_claim(NULL, zilog->zl_spa, 385219089Spjd tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 386219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 387168404Spjd} 388168404Spjd 389219089Spjdstatic int 390168404Spjdzil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 391168404Spjd{ 392219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 393219089Spjd int error; 394219089Spjd 395219089Spjd if (lrc->lrc_txtype != TX_WRITE) 396219089Spjd return (0); 397219089Spjd 398219089Spjd /* 399219089Spjd * If the block is not readable, don't claim it. This can happen 400219089Spjd * in normal operation when a log block is written to disk before 401219089Spjd * some of the dmu_sync() blocks it points to. In this case, the 402219089Spjd * transaction cannot have been committed to anyone (we would have 403219089Spjd * waited for all writes to be stable first), so it is semantically 404219089Spjd * correct to declare this the end of the log. 405219089Spjd */ 406219089Spjd if (lr->lr_blkptr.blk_birth >= first_txg && 407219089Spjd (error = zil_read_log_data(zilog, lr, NULL)) != 0) 408219089Spjd return (error); 409219089Spjd return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 410168404Spjd} 411168404Spjd 412168404Spjd/* ARGSUSED */ 413219089Spjdstatic int 414168404Spjdzil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 415168404Spjd{ 416219089Spjd zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 417219089Spjd 418219089Spjd return (0); 419168404Spjd} 420168404Spjd 421219089Spjdstatic int 422168404Spjdzil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 423168404Spjd{ 424219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 425219089Spjd blkptr_t *bp = &lr->lr_blkptr; 426219089Spjd 427168404Spjd /* 428168404Spjd * If we previously claimed it, we need to free it. 429168404Spjd */ 430219089Spjd if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 431219089Spjd bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 432219089Spjd zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 433219089Spjd 434219089Spjd return (0); 435219089Spjd} 436219089Spjd 437219089Spjdstatic lwb_t * 438219089Spjdzil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 439219089Spjd{ 440219089Spjd lwb_t *lwb; 441219089Spjd 442219089Spjd lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 443219089Spjd lwb->lwb_zilog = zilog; 444219089Spjd lwb->lwb_blk = *bp; 445219089Spjd lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 446219089Spjd lwb->lwb_max_txg = txg; 447219089Spjd lwb->lwb_zio = NULL; 448219089Spjd lwb->lwb_tx = NULL; 449219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 450219089Spjd lwb->lwb_nused = sizeof (zil_chain_t); 451219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp); 452219089Spjd } else { 453219089Spjd lwb->lwb_nused = 0; 454219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 455168404Spjd } 456219089Spjd 457219089Spjd mutex_enter(&zilog->zl_lock); 458219089Spjd list_insert_tail(&zilog->zl_lwb_list, lwb); 459219089Spjd mutex_exit(&zilog->zl_lock); 460219089Spjd 461219089Spjd return (lwb); 462168404Spjd} 463168404Spjd 464168404Spjd/* 465243674Smm * Called when we create in-memory log transactions so that we know 466243674Smm * to cleanup the itxs at the end of spa_sync(). 467243674Smm */ 468243674Smmvoid 469243674Smmzilog_dirty(zilog_t *zilog, uint64_t txg) 470243674Smm{ 471243674Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 472243674Smm dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 473243674Smm 474243674Smm if (dsl_dataset_is_snapshot(ds)) 475243674Smm panic("dirtying snapshot!"); 476243674Smm 477243674Smm if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg) == 0) { 478243674Smm /* up the hold count until we can be written out */ 479243674Smm dmu_buf_add_ref(ds->ds_dbuf, zilog); 480243674Smm } 481243674Smm} 482243674Smm 483243674Smmboolean_t 484243674Smmzilog_is_dirty(zilog_t *zilog) 485243674Smm{ 486243674Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 487243674Smm 488243674Smm for (int t = 0; t < TXG_SIZE; t++) { 489243674Smm if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 490243674Smm return (B_TRUE); 491243674Smm } 492243674Smm return (B_FALSE); 493243674Smm} 494243674Smm 495243674Smm/* 496168404Spjd * Create an on-disk intent log. 497168404Spjd */ 498219089Spjdstatic lwb_t * 499168404Spjdzil_create(zilog_t *zilog) 500168404Spjd{ 501168404Spjd const zil_header_t *zh = zilog->zl_header; 502219089Spjd lwb_t *lwb = NULL; 503168404Spjd uint64_t txg = 0; 504168404Spjd dmu_tx_t *tx = NULL; 505168404Spjd blkptr_t blk; 506168404Spjd int error = 0; 507168404Spjd 508168404Spjd /* 509168404Spjd * Wait for any previous destroy to complete. 510168404Spjd */ 511168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 512168404Spjd 513168404Spjd ASSERT(zh->zh_claim_txg == 0); 514168404Spjd ASSERT(zh->zh_replay_seq == 0); 515168404Spjd 516168404Spjd blk = zh->zh_log; 517168404Spjd 518168404Spjd /* 519219089Spjd * Allocate an initial log block if: 520219089Spjd * - there isn't one already 521219089Spjd * - the existing block is the wrong endianess 522168404Spjd */ 523207908Smm if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 524168404Spjd tx = dmu_tx_create(zilog->zl_os); 525219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 526168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 527168404Spjd txg = dmu_tx_get_txg(tx); 528168404Spjd 529207908Smm if (!BP_IS_HOLE(&blk)) { 530219089Spjd zio_free_zil(zilog->zl_spa, txg, &blk); 531207908Smm BP_ZERO(&blk); 532207908Smm } 533207908Smm 534219089Spjd error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 535219089Spjd ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 536168404Spjd 537168404Spjd if (error == 0) 538168404Spjd zil_init_log_chain(zilog, &blk); 539168404Spjd } 540168404Spjd 541168404Spjd /* 542168404Spjd * Allocate a log write buffer (lwb) for the first log block. 543168404Spjd */ 544219089Spjd if (error == 0) 545219089Spjd lwb = zil_alloc_lwb(zilog, &blk, txg); 546168404Spjd 547168404Spjd /* 548168404Spjd * If we just allocated the first log block, commit our transaction 549168404Spjd * and wait for zil_sync() to stuff the block poiner into zh_log. 550168404Spjd * (zh is part of the MOS, so we cannot modify it in open context.) 551168404Spjd */ 552168404Spjd if (tx != NULL) { 553168404Spjd dmu_tx_commit(tx); 554168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 555168404Spjd } 556168404Spjd 557168404Spjd ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 558219089Spjd 559219089Spjd return (lwb); 560168404Spjd} 561168404Spjd 562168404Spjd/* 563168404Spjd * In one tx, free all log blocks and clear the log header. 564168404Spjd * If keep_first is set, then we're replaying a log with no content. 565168404Spjd * We want to keep the first block, however, so that the first 566168404Spjd * synchronous transaction doesn't require a txg_wait_synced() 567168404Spjd * in zil_create(). We don't need to txg_wait_synced() here either 568168404Spjd * when keep_first is set, because both zil_create() and zil_destroy() 569168404Spjd * will wait for any in-progress destroys to complete. 570168404Spjd */ 571168404Spjdvoid 572168404Spjdzil_destroy(zilog_t *zilog, boolean_t keep_first) 573168404Spjd{ 574168404Spjd const zil_header_t *zh = zilog->zl_header; 575168404Spjd lwb_t *lwb; 576168404Spjd dmu_tx_t *tx; 577168404Spjd uint64_t txg; 578168404Spjd 579168404Spjd /* 580168404Spjd * Wait for any previous destroy to complete. 581168404Spjd */ 582168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 583168404Spjd 584219089Spjd zilog->zl_old_header = *zh; /* debugging aid */ 585219089Spjd 586168404Spjd if (BP_IS_HOLE(&zh->zh_log)) 587168404Spjd return; 588168404Spjd 589168404Spjd tx = dmu_tx_create(zilog->zl_os); 590219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 591168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 592168404Spjd txg = dmu_tx_get_txg(tx); 593168404Spjd 594168404Spjd mutex_enter(&zilog->zl_lock); 595168404Spjd 596168404Spjd ASSERT3U(zilog->zl_destroy_txg, <, txg); 597168404Spjd zilog->zl_destroy_txg = txg; 598168404Spjd zilog->zl_keep_first = keep_first; 599168404Spjd 600168404Spjd if (!list_is_empty(&zilog->zl_lwb_list)) { 601168404Spjd ASSERT(zh->zh_claim_txg == 0); 602224526Smm VERIFY(!keep_first); 603168404Spjd while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 604168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 605168404Spjd if (lwb->lwb_buf != NULL) 606168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 607219089Spjd zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 608168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 609168404Spjd } 610219089Spjd } else if (!keep_first) { 611243674Smm zil_destroy_sync(zilog, tx); 612168404Spjd } 613168404Spjd mutex_exit(&zilog->zl_lock); 614168404Spjd 615168404Spjd dmu_tx_commit(tx); 616185029Spjd} 617168404Spjd 618243674Smmvoid 619243674Smmzil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 620243674Smm{ 621243674Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 622243674Smm (void) zil_parse(zilog, zil_free_log_block, 623243674Smm zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 624243674Smm} 625243674Smm 626168404Spjdint 627219089Spjdzil_claim(const char *osname, void *txarg) 628168404Spjd{ 629168404Spjd dmu_tx_t *tx = txarg; 630168404Spjd uint64_t first_txg = dmu_tx_get_txg(tx); 631168404Spjd zilog_t *zilog; 632168404Spjd zil_header_t *zh; 633168404Spjd objset_t *os; 634168404Spjd int error; 635168404Spjd 636219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 637168404Spjd if (error) { 638185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 639168404Spjd return (0); 640168404Spjd } 641168404Spjd 642168404Spjd zilog = dmu_objset_zil(os); 643168404Spjd zh = zil_header_in_syncing_context(zilog); 644168404Spjd 645219089Spjd if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 646213197Smm if (!BP_IS_HOLE(&zh->zh_log)) 647219089Spjd zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 648213197Smm BP_ZERO(&zh->zh_log); 649213197Smm dsl_dataset_dirty(dmu_objset_ds(os), tx); 650219089Spjd dmu_objset_rele(os, FTAG); 651219089Spjd return (0); 652213197Smm } 653213197Smm 654168404Spjd /* 655168404Spjd * Claim all log blocks if we haven't already done so, and remember 656168404Spjd * the highest claimed sequence number. This ensures that if we can 657168404Spjd * read only part of the log now (e.g. due to a missing device), 658168404Spjd * but we can read the entire log later, we will not try to replay 659168404Spjd * or destroy beyond the last block we successfully claimed. 660168404Spjd */ 661168404Spjd ASSERT3U(zh->zh_claim_txg, <=, first_txg); 662168404Spjd if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 663219089Spjd (void) zil_parse(zilog, zil_claim_log_block, 664219089Spjd zil_claim_log_record, tx, first_txg); 665168404Spjd zh->zh_claim_txg = first_txg; 666219089Spjd zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 667219089Spjd zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 668219089Spjd if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 669219089Spjd zh->zh_flags |= ZIL_REPLAY_NEEDED; 670219089Spjd zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 671168404Spjd dsl_dataset_dirty(dmu_objset_ds(os), tx); 672168404Spjd } 673168404Spjd 674168404Spjd ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 675219089Spjd dmu_objset_rele(os, FTAG); 676168404Spjd return (0); 677168404Spjd} 678168404Spjd 679185029Spjd/* 680185029Spjd * Check the log by walking the log chain. 681185029Spjd * Checksum errors are ok as they indicate the end of the chain. 682185029Spjd * Any other error (no device or read failure) returns an error. 683185029Spjd */ 684185029Spjdint 685219089Spjdzil_check_log_chain(const char *osname, void *tx) 686168404Spjd{ 687185029Spjd zilog_t *zilog; 688185029Spjd objset_t *os; 689219089Spjd blkptr_t *bp; 690185029Spjd int error; 691168404Spjd 692219089Spjd ASSERT(tx == NULL); 693219089Spjd 694219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 695185029Spjd if (error) { 696185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 697185029Spjd return (0); 698185029Spjd } 699168404Spjd 700185029Spjd zilog = dmu_objset_zil(os); 701219089Spjd bp = (blkptr_t *)&zilog->zl_header->zh_log; 702219089Spjd 703219089Spjd /* 704219089Spjd * Check the first block and determine if it's on a log device 705219089Spjd * which may have been removed or faulted prior to loading this 706219089Spjd * pool. If so, there's no point in checking the rest of the log 707219089Spjd * as its content should have already been synced to the pool. 708219089Spjd */ 709219089Spjd if (!BP_IS_HOLE(bp)) { 710219089Spjd vdev_t *vd; 711219089Spjd boolean_t valid = B_TRUE; 712219089Spjd 713219089Spjd spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 714219089Spjd vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 715219089Spjd if (vd->vdev_islog && vdev_is_dead(vd)) 716219089Spjd valid = vdev_log_state_valid(vd); 717219089Spjd spa_config_exit(os->os_spa, SCL_STATE, FTAG); 718219089Spjd 719219089Spjd if (!valid) { 720219089Spjd dmu_objset_rele(os, FTAG); 721219089Spjd return (0); 722219089Spjd } 723168404Spjd } 724185029Spjd 725219089Spjd /* 726219089Spjd * Because tx == NULL, zil_claim_log_block() will not actually claim 727219089Spjd * any blocks, but just determine whether it is possible to do so. 728219089Spjd * In addition to checking the log chain, zil_claim_log_block() 729219089Spjd * will invoke zio_claim() with a done func of spa_claim_notify(), 730219089Spjd * which will update spa_max_claim_txg. See spa_load() for details. 731219089Spjd */ 732219089Spjd error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 733219089Spjd zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 734219089Spjd 735219089Spjd dmu_objset_rele(os, FTAG); 736219089Spjd 737219089Spjd return ((error == ECKSUM || error == ENOENT) ? 0 : error); 738168404Spjd} 739168404Spjd 740185029Spjdstatic int 741185029Spjdzil_vdev_compare(const void *x1, const void *x2) 742185029Spjd{ 743219089Spjd const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 744219089Spjd const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 745185029Spjd 746185029Spjd if (v1 < v2) 747185029Spjd return (-1); 748185029Spjd if (v1 > v2) 749185029Spjd return (1); 750185029Spjd 751185029Spjd return (0); 752185029Spjd} 753185029Spjd 754168404Spjdvoid 755219089Spjdzil_add_block(zilog_t *zilog, const blkptr_t *bp) 756168404Spjd{ 757185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 758185029Spjd avl_index_t where; 759185029Spjd zil_vdev_node_t *zv, zvsearch; 760185029Spjd int ndvas = BP_GET_NDVAS(bp); 761185029Spjd int i; 762168404Spjd 763185029Spjd if (zfs_nocacheflush) 764185029Spjd return; 765168404Spjd 766185029Spjd ASSERT(zilog->zl_writer); 767168404Spjd 768185029Spjd /* 769185029Spjd * Even though we're zl_writer, we still need a lock because the 770185029Spjd * zl_get_data() callbacks may have dmu_sync() done callbacks 771185029Spjd * that will run concurrently. 772185029Spjd */ 773185029Spjd mutex_enter(&zilog->zl_vdev_lock); 774185029Spjd for (i = 0; i < ndvas; i++) { 775185029Spjd zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 776185029Spjd if (avl_find(t, &zvsearch, &where) == NULL) { 777185029Spjd zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 778185029Spjd zv->zv_vdev = zvsearch.zv_vdev; 779185029Spjd avl_insert(t, zv, where); 780185029Spjd } 781185029Spjd } 782185029Spjd mutex_exit(&zilog->zl_vdev_lock); 783168404Spjd} 784168404Spjd 785219089Spjdstatic void 786168404Spjdzil_flush_vdevs(zilog_t *zilog) 787168404Spjd{ 788168404Spjd spa_t *spa = zilog->zl_spa; 789185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 790185029Spjd void *cookie = NULL; 791185029Spjd zil_vdev_node_t *zv; 792185029Spjd zio_t *zio; 793168404Spjd 794168404Spjd ASSERT(zilog->zl_writer); 795168404Spjd 796185029Spjd /* 797185029Spjd * We don't need zl_vdev_lock here because we're the zl_writer, 798185029Spjd * and all zl_get_data() callbacks are done. 799185029Spjd */ 800185029Spjd if (avl_numnodes(t) == 0) 801185029Spjd return; 802185029Spjd 803185029Spjd spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 804185029Spjd 805185029Spjd zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 806185029Spjd 807185029Spjd while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 808185029Spjd vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 809185029Spjd if (vd != NULL) 810185029Spjd zio_flush(zio, vd); 811185029Spjd kmem_free(zv, sizeof (*zv)); 812168404Spjd } 813168404Spjd 814168404Spjd /* 815168404Spjd * Wait for all the flushes to complete. Not all devices actually 816168404Spjd * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 817168404Spjd */ 818185029Spjd (void) zio_wait(zio); 819185029Spjd 820185029Spjd spa_config_exit(spa, SCL_STATE, FTAG); 821168404Spjd} 822168404Spjd 823168404Spjd/* 824168404Spjd * Function called when a log block write completes 825168404Spjd */ 826168404Spjdstatic void 827168404Spjdzil_lwb_write_done(zio_t *zio) 828168404Spjd{ 829168404Spjd lwb_t *lwb = zio->io_private; 830168404Spjd zilog_t *zilog = lwb->lwb_zilog; 831219089Spjd dmu_tx_t *tx = lwb->lwb_tx; 832168404Spjd 833185029Spjd ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 834185029Spjd ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 835185029Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 836185029Spjd ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 837185029Spjd ASSERT(!BP_IS_GANG(zio->io_bp)); 838185029Spjd ASSERT(!BP_IS_HOLE(zio->io_bp)); 839185029Spjd ASSERT(zio->io_bp->blk_fill == 0); 840185029Spjd 841168404Spjd /* 842209962Smm * Ensure the lwb buffer pointer is cleared before releasing 843209962Smm * the txg. If we have had an allocation failure and 844209962Smm * the txg is waiting to sync then we want want zil_sync() 845209962Smm * to remove the lwb so that it's not picked up as the next new 846209962Smm * one in zil_commit_writer(). zil_sync() will only remove 847209962Smm * the lwb if lwb_buf is null. 848168404Spjd */ 849168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 850168404Spjd mutex_enter(&zilog->zl_lock); 851168404Spjd lwb->lwb_buf = NULL; 852219089Spjd lwb->lwb_tx = NULL; 853219089Spjd mutex_exit(&zilog->zl_lock); 854209962Smm 855209962Smm /* 856209962Smm * Now that we've written this log block, we have a stable pointer 857209962Smm * to the next block in the chain, so it's OK to let the txg in 858219089Spjd * which we allocated the next block sync. 859209962Smm */ 860219089Spjd dmu_tx_commit(tx); 861168404Spjd} 862168404Spjd 863168404Spjd/* 864168404Spjd * Initialize the io for a log block. 865168404Spjd */ 866168404Spjdstatic void 867168404Spjdzil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 868168404Spjd{ 869168404Spjd zbookmark_t zb; 870168404Spjd 871219089Spjd SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 872219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 873219089Spjd lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 874168404Spjd 875168404Spjd if (zilog->zl_root_zio == NULL) { 876168404Spjd zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 877168404Spjd ZIO_FLAG_CANFAIL); 878168404Spjd } 879168404Spjd if (lwb->lwb_zio == NULL) { 880168404Spjd lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 881219089Spjd 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 882213197Smm zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 883219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 884168404Spjd } 885168404Spjd} 886168404Spjd 887168404Spjd/* 888219089Spjd * Define a limited set of intent log block sizes. 889219089Spjd * These must be a multiple of 4KB. Note only the amount used (again 890219089Spjd * aligned to 4KB) actually gets written. However, we can't always just 891219089Spjd * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 892219089Spjd */ 893219089Spjduint64_t zil_block_buckets[] = { 894219089Spjd 4096, /* non TX_WRITE */ 895219089Spjd 8192+4096, /* data base */ 896219089Spjd 32*1024 + 4096, /* NFS writes */ 897219089Spjd UINT64_MAX 898219089Spjd}; 899219089Spjd 900219089Spjd/* 901219089Spjd * Use the slog as long as the logbias is 'latency' and the current commit size 902219089Spjd * is less than the limit or the total list size is less than 2X the limit. 903219089Spjd * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 904219089Spjd */ 905219089Spjduint64_t zil_slog_limit = 1024 * 1024; 906219089Spjd#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 907219089Spjd (((zilog)->zl_cur_used < zil_slog_limit) || \ 908219089Spjd ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 909219089Spjd 910219089Spjd/* 911168404Spjd * Start a log block write and advance to the next log block. 912168404Spjd * Calls are serialized. 913168404Spjd */ 914168404Spjdstatic lwb_t * 915168404Spjdzil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 916168404Spjd{ 917219089Spjd lwb_t *nlwb = NULL; 918219089Spjd zil_chain_t *zilc; 919168404Spjd spa_t *spa = zilog->zl_spa; 920219089Spjd blkptr_t *bp; 921219089Spjd dmu_tx_t *tx; 922168404Spjd uint64_t txg; 923219089Spjd uint64_t zil_blksz, wsz; 924219089Spjd int i, error; 925168404Spjd 926219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 927219089Spjd zilc = (zil_chain_t *)lwb->lwb_buf; 928219089Spjd bp = &zilc->zc_next_blk; 929219089Spjd } else { 930219089Spjd zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 931219089Spjd bp = &zilc->zc_next_blk; 932219089Spjd } 933168404Spjd 934219089Spjd ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 935219089Spjd 936168404Spjd /* 937168404Spjd * Allocate the next block and save its address in this block 938168404Spjd * before writing it in order to establish the log chain. 939168404Spjd * Note that if the allocation of nlwb synced before we wrote 940168404Spjd * the block that points at it (lwb), we'd leak it if we crashed. 941219089Spjd * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 942219089Spjd * We dirty the dataset to ensure that zil_sync() will be called 943219089Spjd * to clean up in the event of allocation failure or I/O failure. 944168404Spjd */ 945219089Spjd tx = dmu_tx_create(zilog->zl_os); 946219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 947219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 948219089Spjd txg = dmu_tx_get_txg(tx); 949168404Spjd 950219089Spjd lwb->lwb_tx = tx; 951219089Spjd 952168404Spjd /* 953219089Spjd * Log blocks are pre-allocated. Here we select the size of the next 954219089Spjd * block, based on size used in the last block. 955219089Spjd * - first find the smallest bucket that will fit the block from a 956219089Spjd * limited set of block sizes. This is because it's faster to write 957219089Spjd * blocks allocated from the same metaslab as they are adjacent or 958219089Spjd * close. 959219089Spjd * - next find the maximum from the new suggested size and an array of 960219089Spjd * previous sizes. This lessens a picket fence effect of wrongly 961219089Spjd * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 962219089Spjd * requests. 963219089Spjd * 964219089Spjd * Note we only write what is used, but we can't just allocate 965219089Spjd * the maximum block size because we can exhaust the available 966219089Spjd * pool log space. 967168404Spjd */ 968219089Spjd zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 969219089Spjd for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 970219089Spjd continue; 971219089Spjd zil_blksz = zil_block_buckets[i]; 972219089Spjd if (zil_blksz == UINT64_MAX) 973219089Spjd zil_blksz = SPA_MAXBLOCKSIZE; 974219089Spjd zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 975219089Spjd for (i = 0; i < ZIL_PREV_BLKS; i++) 976219089Spjd zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 977219089Spjd zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 978168404Spjd 979168404Spjd BP_ZERO(bp); 980168404Spjd /* pass the old blkptr in order to spread log blocks across devs */ 981219089Spjd error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 982219089Spjd USE_SLOG(zilog)); 983219089Spjd if (!error) { 984219089Spjd ASSERT3U(bp->blk_birth, ==, txg); 985219089Spjd bp->blk_cksum = lwb->lwb_blk.blk_cksum; 986219089Spjd bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 987168404Spjd 988168404Spjd /* 989219089Spjd * Allocate a new log write buffer (lwb). 990168404Spjd */ 991219089Spjd nlwb = zil_alloc_lwb(zilog, bp, txg); 992168404Spjd 993219089Spjd /* Record the block for later vdev flushing */ 994219089Spjd zil_add_block(zilog, &lwb->lwb_blk); 995168404Spjd } 996168404Spjd 997219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 998219089Spjd /* For Slim ZIL only write what is used. */ 999219089Spjd wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1000219089Spjd ASSERT3U(wsz, <=, lwb->lwb_sz); 1001219089Spjd zio_shrink(lwb->lwb_zio, wsz); 1002168404Spjd 1003219089Spjd } else { 1004219089Spjd wsz = lwb->lwb_sz; 1005219089Spjd } 1006168404Spjd 1007219089Spjd zilc->zc_pad = 0; 1008219089Spjd zilc->zc_nused = lwb->lwb_nused; 1009219089Spjd zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1010168404Spjd 1011168404Spjd /* 1012219089Spjd * clear unused data for security 1013168404Spjd */ 1014219089Spjd bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1015168404Spjd 1016219089Spjd zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1017168404Spjd 1018168404Spjd /* 1019219089Spjd * If there was an allocation failure then nlwb will be null which 1020219089Spjd * forces a txg_wait_synced(). 1021168404Spjd */ 1022168404Spjd return (nlwb); 1023168404Spjd} 1024168404Spjd 1025168404Spjdstatic lwb_t * 1026168404Spjdzil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1027168404Spjd{ 1028168404Spjd lr_t *lrc = &itx->itx_lr; /* common log record */ 1029219089Spjd lr_write_t *lrw = (lr_write_t *)lrc; 1030219089Spjd char *lr_buf; 1031168404Spjd uint64_t txg = lrc->lrc_txg; 1032168404Spjd uint64_t reclen = lrc->lrc_reclen; 1033219089Spjd uint64_t dlen = 0; 1034168404Spjd 1035168404Spjd if (lwb == NULL) 1036168404Spjd return (NULL); 1037219089Spjd 1038168404Spjd ASSERT(lwb->lwb_buf != NULL); 1039243674Smm ASSERT(zilog_is_dirty(zilog) || 1040243674Smm spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1041168404Spjd 1042168404Spjd if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1043168404Spjd dlen = P2ROUNDUP_TYPED( 1044219089Spjd lrw->lr_length, sizeof (uint64_t), uint64_t); 1045168404Spjd 1046168404Spjd zilog->zl_cur_used += (reclen + dlen); 1047168404Spjd 1048168404Spjd zil_lwb_write_init(zilog, lwb); 1049168404Spjd 1050168404Spjd /* 1051168404Spjd * If this record won't fit in the current log block, start a new one. 1052168404Spjd */ 1053219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1054168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1055168404Spjd if (lwb == NULL) 1056168404Spjd return (NULL); 1057168404Spjd zil_lwb_write_init(zilog, lwb); 1058219089Spjd ASSERT(LWB_EMPTY(lwb)); 1059219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1060168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1061168404Spjd return (lwb); 1062168404Spjd } 1063168404Spjd } 1064168404Spjd 1065219089Spjd lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1066219089Spjd bcopy(lrc, lr_buf, reclen); 1067219089Spjd lrc = (lr_t *)lr_buf; 1068219089Spjd lrw = (lr_write_t *)lrc; 1069168404Spjd 1070168404Spjd /* 1071168404Spjd * If it's a write, fetch the data or get its blkptr as appropriate. 1072168404Spjd */ 1073168404Spjd if (lrc->lrc_txtype == TX_WRITE) { 1074168404Spjd if (txg > spa_freeze_txg(zilog->zl_spa)) 1075168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1076168404Spjd if (itx->itx_wr_state != WR_COPIED) { 1077168404Spjd char *dbuf; 1078168404Spjd int error; 1079168404Spjd 1080168404Spjd if (dlen) { 1081168404Spjd ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1082219089Spjd dbuf = lr_buf + reclen; 1083219089Spjd lrw->lr_common.lrc_reclen += dlen; 1084168404Spjd } else { 1085168404Spjd ASSERT(itx->itx_wr_state == WR_INDIRECT); 1086168404Spjd dbuf = NULL; 1087168404Spjd } 1088168404Spjd error = zilog->zl_get_data( 1089219089Spjd itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1090214378Smm if (error == EIO) { 1091214378Smm txg_wait_synced(zilog->zl_dmu_pool, txg); 1092214378Smm return (lwb); 1093214378Smm } 1094168404Spjd if (error) { 1095168404Spjd ASSERT(error == ENOENT || error == EEXIST || 1096168404Spjd error == EALREADY); 1097168404Spjd return (lwb); 1098168404Spjd } 1099168404Spjd } 1100168404Spjd } 1101168404Spjd 1102219089Spjd /* 1103219089Spjd * We're actually making an entry, so update lrc_seq to be the 1104219089Spjd * log record sequence number. Note that this is generally not 1105219089Spjd * equal to the itx sequence number because not all transactions 1106219089Spjd * are synchronous, and sometimes spa_sync() gets there first. 1107219089Spjd */ 1108219089Spjd lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1109168404Spjd lwb->lwb_nused += reclen + dlen; 1110168404Spjd lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1111219089Spjd ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1112243674Smm ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1113168404Spjd 1114168404Spjd return (lwb); 1115168404Spjd} 1116168404Spjd 1117168404Spjditx_t * 1118185029Spjdzil_itx_create(uint64_t txtype, size_t lrsize) 1119168404Spjd{ 1120168404Spjd itx_t *itx; 1121168404Spjd 1122168404Spjd lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1123168404Spjd 1124168404Spjd itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1125168404Spjd itx->itx_lr.lrc_txtype = txtype; 1126168404Spjd itx->itx_lr.lrc_reclen = lrsize; 1127185029Spjd itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1128168404Spjd itx->itx_lr.lrc_seq = 0; /* defensive */ 1129219089Spjd itx->itx_sync = B_TRUE; /* default is synchronous */ 1130168404Spjd 1131168404Spjd return (itx); 1132168404Spjd} 1133168404Spjd 1134219089Spjdvoid 1135219089Spjdzil_itx_destroy(itx_t *itx) 1136168404Spjd{ 1137219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1138219089Spjd} 1139168404Spjd 1140219089Spjd/* 1141219089Spjd * Free up the sync and async itxs. The itxs_t has already been detached 1142219089Spjd * so no locks are needed. 1143219089Spjd */ 1144219089Spjdstatic void 1145219089Spjdzil_itxg_clean(itxs_t *itxs) 1146219089Spjd{ 1147219089Spjd itx_t *itx; 1148219089Spjd list_t *list; 1149219089Spjd avl_tree_t *t; 1150219089Spjd void *cookie; 1151219089Spjd itx_async_node_t *ian; 1152168404Spjd 1153219089Spjd list = &itxs->i_sync_list; 1154219089Spjd while ((itx = list_head(list)) != NULL) { 1155219089Spjd list_remove(list, itx); 1156219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1157219089Spjd itx->itx_lr.lrc_reclen); 1158219089Spjd } 1159168404Spjd 1160219089Spjd cookie = NULL; 1161219089Spjd t = &itxs->i_async_tree; 1162219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1163219089Spjd list = &ian->ia_list; 1164219089Spjd while ((itx = list_head(list)) != NULL) { 1165219089Spjd list_remove(list, itx); 1166219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1167219089Spjd itx->itx_lr.lrc_reclen); 1168219089Spjd } 1169219089Spjd list_destroy(list); 1170219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1171219089Spjd } 1172219089Spjd avl_destroy(t); 1173219089Spjd 1174219089Spjd kmem_free(itxs, sizeof (itxs_t)); 1175168404Spjd} 1176168404Spjd 1177219089Spjdstatic int 1178219089Spjdzil_aitx_compare(const void *x1, const void *x2) 1179219089Spjd{ 1180219089Spjd const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1181219089Spjd const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1182219089Spjd 1183219089Spjd if (o1 < o2) 1184219089Spjd return (-1); 1185219089Spjd if (o1 > o2) 1186219089Spjd return (1); 1187219089Spjd 1188219089Spjd return (0); 1189219089Spjd} 1190219089Spjd 1191168404Spjd/* 1192219089Spjd * Remove all async itx with the given oid. 1193168404Spjd */ 1194168404Spjdstatic void 1195219089Spjdzil_remove_async(zilog_t *zilog, uint64_t oid) 1196168404Spjd{ 1197219089Spjd uint64_t otxg, txg; 1198219089Spjd itx_async_node_t *ian; 1199219089Spjd avl_tree_t *t; 1200219089Spjd avl_index_t where; 1201168404Spjd list_t clean_list; 1202168404Spjd itx_t *itx; 1203168404Spjd 1204219089Spjd ASSERT(oid != 0); 1205168404Spjd list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1206168404Spjd 1207219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1208219089Spjd otxg = ZILTEST_TXG; 1209219089Spjd else 1210219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1211219089Spjd 1212219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1213219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1214219089Spjd 1215219089Spjd mutex_enter(&itxg->itxg_lock); 1216219089Spjd if (itxg->itxg_txg != txg) { 1217219089Spjd mutex_exit(&itxg->itxg_lock); 1218219089Spjd continue; 1219219089Spjd } 1220219089Spjd 1221219089Spjd /* 1222219089Spjd * Locate the object node and append its list. 1223219089Spjd */ 1224219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1225219089Spjd ian = avl_find(t, &oid, &where); 1226219089Spjd if (ian != NULL) 1227219089Spjd list_move_tail(&clean_list, &ian->ia_list); 1228219089Spjd mutex_exit(&itxg->itxg_lock); 1229168404Spjd } 1230219089Spjd while ((itx = list_head(&clean_list)) != NULL) { 1231219089Spjd list_remove(&clean_list, itx); 1232219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1233219089Spjd itx->itx_lr.lrc_reclen); 1234219089Spjd } 1235219089Spjd list_destroy(&clean_list); 1236219089Spjd} 1237168404Spjd 1238219089Spjdvoid 1239219089Spjdzil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1240219089Spjd{ 1241219089Spjd uint64_t txg; 1242219089Spjd itxg_t *itxg; 1243219089Spjd itxs_t *itxs, *clean = NULL; 1244219089Spjd 1245168404Spjd /* 1246219089Spjd * Object ids can be re-instantiated in the next txg so 1247219089Spjd * remove any async transactions to avoid future leaks. 1248219089Spjd * This can happen if a fsync occurs on the re-instantiated 1249219089Spjd * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1250219089Spjd * the new file data and flushes a write record for the old object. 1251168404Spjd */ 1252219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1253219089Spjd zil_remove_async(zilog, itx->itx_oid); 1254219089Spjd 1255219089Spjd /* 1256219089Spjd * Ensure the data of a renamed file is committed before the rename. 1257219089Spjd */ 1258219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1259219089Spjd zil_async_to_sync(zilog, itx->itx_oid); 1260219089Spjd 1261243674Smm if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1262219089Spjd txg = ZILTEST_TXG; 1263219089Spjd else 1264219089Spjd txg = dmu_tx_get_txg(tx); 1265219089Spjd 1266219089Spjd itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1267219089Spjd mutex_enter(&itxg->itxg_lock); 1268219089Spjd itxs = itxg->itxg_itxs; 1269219089Spjd if (itxg->itxg_txg != txg) { 1270219089Spjd if (itxs != NULL) { 1271219089Spjd /* 1272219089Spjd * The zil_clean callback hasn't got around to cleaning 1273219089Spjd * this itxg. Save the itxs for release below. 1274219089Spjd * This should be rare. 1275219089Spjd */ 1276219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1277219089Spjd itxg->itxg_sod = 0; 1278219089Spjd clean = itxg->itxg_itxs; 1279219089Spjd } 1280219089Spjd ASSERT(itxg->itxg_sod == 0); 1281219089Spjd itxg->itxg_txg = txg; 1282219089Spjd itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1283219089Spjd 1284219089Spjd list_create(&itxs->i_sync_list, sizeof (itx_t), 1285219089Spjd offsetof(itx_t, itx_node)); 1286219089Spjd avl_create(&itxs->i_async_tree, zil_aitx_compare, 1287219089Spjd sizeof (itx_async_node_t), 1288219089Spjd offsetof(itx_async_node_t, ia_node)); 1289168404Spjd } 1290219089Spjd if (itx->itx_sync) { 1291219089Spjd list_insert_tail(&itxs->i_sync_list, itx); 1292219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1293219089Spjd itxg->itxg_sod += itx->itx_sod; 1294219089Spjd } else { 1295219089Spjd avl_tree_t *t = &itxs->i_async_tree; 1296219089Spjd uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1297219089Spjd itx_async_node_t *ian; 1298219089Spjd avl_index_t where; 1299168404Spjd 1300219089Spjd ian = avl_find(t, &foid, &where); 1301219089Spjd if (ian == NULL) { 1302219089Spjd ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1303219089Spjd list_create(&ian->ia_list, sizeof (itx_t), 1304219089Spjd offsetof(itx_t, itx_node)); 1305219089Spjd ian->ia_foid = foid; 1306219089Spjd avl_insert(t, ian, where); 1307219089Spjd } 1308219089Spjd list_insert_tail(&ian->ia_list, itx); 1309168404Spjd } 1310219089Spjd 1311219089Spjd itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1312243674Smm zilog_dirty(zilog, txg); 1313219089Spjd mutex_exit(&itxg->itxg_lock); 1314219089Spjd 1315219089Spjd /* Release the old itxs now we've dropped the lock */ 1316219089Spjd if (clean != NULL) 1317219089Spjd zil_itxg_clean(clean); 1318168404Spjd} 1319168404Spjd 1320168404Spjd/* 1321168404Spjd * If there are any in-memory intent log transactions which have now been 1322243674Smm * synced then start up a taskq to free them. We should only do this after we 1323243674Smm * have written out the uberblocks (i.e. txg has been comitted) so that 1324243674Smm * don't inadvertently clean out in-memory log records that would be required 1325243674Smm * by zil_commit(). 1326168404Spjd */ 1327168404Spjdvoid 1328219089Spjdzil_clean(zilog_t *zilog, uint64_t synced_txg) 1329168404Spjd{ 1330219089Spjd itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1331219089Spjd itxs_t *clean_me; 1332168404Spjd 1333219089Spjd mutex_enter(&itxg->itxg_lock); 1334219089Spjd if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1335219089Spjd mutex_exit(&itxg->itxg_lock); 1336219089Spjd return; 1337168404Spjd } 1338219089Spjd ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1339219089Spjd ASSERT(itxg->itxg_txg != 0); 1340219089Spjd ASSERT(zilog->zl_clean_taskq != NULL); 1341219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1342219089Spjd itxg->itxg_sod = 0; 1343219089Spjd clean_me = itxg->itxg_itxs; 1344219089Spjd itxg->itxg_itxs = NULL; 1345219089Spjd itxg->itxg_txg = 0; 1346219089Spjd mutex_exit(&itxg->itxg_lock); 1347219089Spjd /* 1348219089Spjd * Preferably start a task queue to free up the old itxs but 1349219089Spjd * if taskq_dispatch can't allocate resources to do that then 1350219089Spjd * free it in-line. This should be rare. Note, using TQ_SLEEP 1351219089Spjd * created a bad performance problem. 1352219089Spjd */ 1353219089Spjd if (taskq_dispatch(zilog->zl_clean_taskq, 1354219089Spjd (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1355219089Spjd zil_itxg_clean(clean_me); 1356168404Spjd} 1357168404Spjd 1358219089Spjd/* 1359219089Spjd * Get the list of itxs to commit into zl_itx_commit_list. 1360219089Spjd */ 1361185029Spjdstatic void 1362219089Spjdzil_get_commit_list(zilog_t *zilog) 1363168404Spjd{ 1364219089Spjd uint64_t otxg, txg; 1365219089Spjd list_t *commit_list = &zilog->zl_itx_commit_list; 1366219089Spjd uint64_t push_sod = 0; 1367219089Spjd 1368219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1369219089Spjd otxg = ZILTEST_TXG; 1370219089Spjd else 1371219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1372219089Spjd 1373219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1374219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1375219089Spjd 1376219089Spjd mutex_enter(&itxg->itxg_lock); 1377219089Spjd if (itxg->itxg_txg != txg) { 1378219089Spjd mutex_exit(&itxg->itxg_lock); 1379219089Spjd continue; 1380219089Spjd } 1381219089Spjd 1382219089Spjd list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1383219089Spjd push_sod += itxg->itxg_sod; 1384219089Spjd itxg->itxg_sod = 0; 1385219089Spjd 1386219089Spjd mutex_exit(&itxg->itxg_lock); 1387219089Spjd } 1388219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1389219089Spjd} 1390219089Spjd 1391219089Spjd/* 1392219089Spjd * Move the async itxs for a specified object to commit into sync lists. 1393219089Spjd */ 1394219089Spjdstatic void 1395219089Spjdzil_async_to_sync(zilog_t *zilog, uint64_t foid) 1396219089Spjd{ 1397219089Spjd uint64_t otxg, txg; 1398219089Spjd itx_async_node_t *ian; 1399219089Spjd avl_tree_t *t; 1400219089Spjd avl_index_t where; 1401219089Spjd 1402219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1403219089Spjd otxg = ZILTEST_TXG; 1404219089Spjd else 1405219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1406219089Spjd 1407219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1408219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1409219089Spjd 1410219089Spjd mutex_enter(&itxg->itxg_lock); 1411219089Spjd if (itxg->itxg_txg != txg) { 1412219089Spjd mutex_exit(&itxg->itxg_lock); 1413219089Spjd continue; 1414219089Spjd } 1415219089Spjd 1416219089Spjd /* 1417219089Spjd * If a foid is specified then find that node and append its 1418219089Spjd * list. Otherwise walk the tree appending all the lists 1419219089Spjd * to the sync list. We add to the end rather than the 1420219089Spjd * beginning to ensure the create has happened. 1421219089Spjd */ 1422219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1423219089Spjd if (foid != 0) { 1424219089Spjd ian = avl_find(t, &foid, &where); 1425219089Spjd if (ian != NULL) { 1426219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1427219089Spjd &ian->ia_list); 1428219089Spjd } 1429219089Spjd } else { 1430219089Spjd void *cookie = NULL; 1431219089Spjd 1432219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1433219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1434219089Spjd &ian->ia_list); 1435219089Spjd list_destroy(&ian->ia_list); 1436219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1437219089Spjd } 1438219089Spjd } 1439219089Spjd mutex_exit(&itxg->itxg_lock); 1440219089Spjd } 1441219089Spjd} 1442219089Spjd 1443219089Spjdstatic void 1444219089Spjdzil_commit_writer(zilog_t *zilog) 1445219089Spjd{ 1446168404Spjd uint64_t txg; 1447219089Spjd itx_t *itx; 1448168404Spjd lwb_t *lwb; 1449219089Spjd spa_t *spa = zilog->zl_spa; 1450219089Spjd int error = 0; 1451168404Spjd 1452185029Spjd ASSERT(zilog->zl_root_zio == NULL); 1453168404Spjd 1454219089Spjd mutex_exit(&zilog->zl_lock); 1455219089Spjd 1456219089Spjd zil_get_commit_list(zilog); 1457219089Spjd 1458219089Spjd /* 1459219089Spjd * Return if there's nothing to commit before we dirty the fs by 1460219089Spjd * calling zil_create(). 1461219089Spjd */ 1462219089Spjd if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1463219089Spjd mutex_enter(&zilog->zl_lock); 1464219089Spjd return; 1465219089Spjd } 1466219089Spjd 1467168404Spjd if (zilog->zl_suspend) { 1468168404Spjd lwb = NULL; 1469168404Spjd } else { 1470168404Spjd lwb = list_tail(&zilog->zl_lwb_list); 1471219089Spjd if (lwb == NULL) 1472219089Spjd lwb = zil_create(zilog); 1473168404Spjd } 1474168404Spjd 1475168404Spjd DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1476219089Spjd while (itx = list_head(&zilog->zl_itx_commit_list)) { 1477168404Spjd txg = itx->itx_lr.lrc_txg; 1478168404Spjd ASSERT(txg); 1479168404Spjd 1480219089Spjd if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1481168404Spjd lwb = zil_lwb_commit(zilog, itx, lwb); 1482219089Spjd list_remove(&zilog->zl_itx_commit_list, itx); 1483168404Spjd kmem_free(itx, offsetof(itx_t, itx_lr) 1484168404Spjd + itx->itx_lr.lrc_reclen); 1485168404Spjd } 1486168404Spjd DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1487168404Spjd 1488168404Spjd /* write the last block out */ 1489168404Spjd if (lwb != NULL && lwb->lwb_zio != NULL) 1490168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1491168404Spjd 1492168404Spjd zilog->zl_cur_used = 0; 1493168404Spjd 1494168404Spjd /* 1495168404Spjd * Wait if necessary for the log blocks to be on stable storage. 1496168404Spjd */ 1497168404Spjd if (zilog->zl_root_zio) { 1498219089Spjd error = zio_wait(zilog->zl_root_zio); 1499185029Spjd zilog->zl_root_zio = NULL; 1500185029Spjd zil_flush_vdevs(zilog); 1501168404Spjd } 1502168404Spjd 1503219089Spjd if (error || lwb == NULL) 1504168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 1505168404Spjd 1506168404Spjd mutex_enter(&zilog->zl_lock); 1507168404Spjd 1508219089Spjd /* 1509219089Spjd * Remember the highest committed log sequence number for ztest. 1510219089Spjd * We only update this value when all the log writes succeeded, 1511219089Spjd * because ztest wants to ASSERT that it got the whole log chain. 1512219089Spjd */ 1513219089Spjd if (error == 0 && lwb != NULL) 1514219089Spjd zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1515168404Spjd} 1516168404Spjd 1517168404Spjd/* 1518219089Spjd * Commit zfs transactions to stable storage. 1519168404Spjd * If foid is 0 push out all transactions, otherwise push only those 1520219089Spjd * for that object or might reference that object. 1521219089Spjd * 1522219089Spjd * itxs are committed in batches. In a heavily stressed zil there will be 1523219089Spjd * a commit writer thread who is writing out a bunch of itxs to the log 1524219089Spjd * for a set of committing threads (cthreads) in the same batch as the writer. 1525219089Spjd * Those cthreads are all waiting on the same cv for that batch. 1526219089Spjd * 1527219089Spjd * There will also be a different and growing batch of threads that are 1528219089Spjd * waiting to commit (qthreads). When the committing batch completes 1529219089Spjd * a transition occurs such that the cthreads exit and the qthreads become 1530219089Spjd * cthreads. One of the new cthreads becomes the writer thread for the 1531219089Spjd * batch. Any new threads arriving become new qthreads. 1532219089Spjd * 1533219089Spjd * Only 2 condition variables are needed and there's no transition 1534219089Spjd * between the two cvs needed. They just flip-flop between qthreads 1535219089Spjd * and cthreads. 1536219089Spjd * 1537219089Spjd * Using this scheme we can efficiently wakeup up only those threads 1538219089Spjd * that have been committed. 1539168404Spjd */ 1540168404Spjdvoid 1541219089Spjdzil_commit(zilog_t *zilog, uint64_t foid) 1542168404Spjd{ 1543219089Spjd uint64_t mybatch; 1544219089Spjd 1545219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1546168404Spjd return; 1547168404Spjd 1548219089Spjd /* move the async itxs for the foid to the sync queues */ 1549219089Spjd zil_async_to_sync(zilog, foid); 1550219089Spjd 1551168404Spjd mutex_enter(&zilog->zl_lock); 1552219089Spjd mybatch = zilog->zl_next_batch; 1553168404Spjd while (zilog->zl_writer) { 1554219089Spjd cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1555219089Spjd if (mybatch <= zilog->zl_com_batch) { 1556168404Spjd mutex_exit(&zilog->zl_lock); 1557168404Spjd return; 1558168404Spjd } 1559168404Spjd } 1560219089Spjd 1561219089Spjd zilog->zl_next_batch++; 1562219089Spjd zilog->zl_writer = B_TRUE; 1563219089Spjd zil_commit_writer(zilog); 1564219089Spjd zilog->zl_com_batch = mybatch; 1565219089Spjd zilog->zl_writer = B_FALSE; 1566168404Spjd mutex_exit(&zilog->zl_lock); 1567219089Spjd 1568219089Spjd /* wake up one thread to become the next writer */ 1569219089Spjd cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1570219089Spjd 1571219089Spjd /* wake up all threads waiting for this batch to be committed */ 1572219089Spjd cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1573168404Spjd} 1574168404Spjd 1575168404Spjd/* 1576168404Spjd * Called in syncing context to free committed log blocks and update log header. 1577168404Spjd */ 1578168404Spjdvoid 1579168404Spjdzil_sync(zilog_t *zilog, dmu_tx_t *tx) 1580168404Spjd{ 1581168404Spjd zil_header_t *zh = zil_header_in_syncing_context(zilog); 1582168404Spjd uint64_t txg = dmu_tx_get_txg(tx); 1583168404Spjd spa_t *spa = zilog->zl_spa; 1584219089Spjd uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1585168404Spjd lwb_t *lwb; 1586168404Spjd 1587209962Smm /* 1588209962Smm * We don't zero out zl_destroy_txg, so make sure we don't try 1589209962Smm * to destroy it twice. 1590209962Smm */ 1591209962Smm if (spa_sync_pass(spa) != 1) 1592209962Smm return; 1593209962Smm 1594168404Spjd mutex_enter(&zilog->zl_lock); 1595168404Spjd 1596168404Spjd ASSERT(zilog->zl_stop_sync == 0); 1597168404Spjd 1598219089Spjd if (*replayed_seq != 0) { 1599219089Spjd ASSERT(zh->zh_replay_seq < *replayed_seq); 1600219089Spjd zh->zh_replay_seq = *replayed_seq; 1601219089Spjd *replayed_seq = 0; 1602219089Spjd } 1603168404Spjd 1604168404Spjd if (zilog->zl_destroy_txg == txg) { 1605168404Spjd blkptr_t blk = zh->zh_log; 1606168404Spjd 1607168404Spjd ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1608168404Spjd 1609168404Spjd bzero(zh, sizeof (zil_header_t)); 1610209962Smm bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1611168404Spjd 1612168404Spjd if (zilog->zl_keep_first) { 1613168404Spjd /* 1614168404Spjd * If this block was part of log chain that couldn't 1615168404Spjd * be claimed because a device was missing during 1616168404Spjd * zil_claim(), but that device later returns, 1617168404Spjd * then this block could erroneously appear valid. 1618168404Spjd * To guard against this, assign a new GUID to the new 1619168404Spjd * log chain so it doesn't matter what blk points to. 1620168404Spjd */ 1621168404Spjd zil_init_log_chain(zilog, &blk); 1622168404Spjd zh->zh_log = blk; 1623168404Spjd } 1624168404Spjd } 1625168404Spjd 1626213197Smm while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1627168404Spjd zh->zh_log = lwb->lwb_blk; 1628168404Spjd if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1629168404Spjd break; 1630168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 1631219089Spjd zio_free_zil(spa, txg, &lwb->lwb_blk); 1632168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 1633168404Spjd 1634168404Spjd /* 1635168404Spjd * If we don't have anything left in the lwb list then 1636168404Spjd * we've had an allocation failure and we need to zero 1637168404Spjd * out the zil_header blkptr so that we don't end 1638168404Spjd * up freeing the same block twice. 1639168404Spjd */ 1640168404Spjd if (list_head(&zilog->zl_lwb_list) == NULL) 1641168404Spjd BP_ZERO(&zh->zh_log); 1642168404Spjd } 1643168404Spjd mutex_exit(&zilog->zl_lock); 1644168404Spjd} 1645168404Spjd 1646168404Spjdvoid 1647168404Spjdzil_init(void) 1648168404Spjd{ 1649168404Spjd zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1650168404Spjd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1651168404Spjd} 1652168404Spjd 1653168404Spjdvoid 1654168404Spjdzil_fini(void) 1655168404Spjd{ 1656168404Spjd kmem_cache_destroy(zil_lwb_cache); 1657168404Spjd} 1658168404Spjd 1659219089Spjdvoid 1660219089Spjdzil_set_sync(zilog_t *zilog, uint64_t sync) 1661219089Spjd{ 1662219089Spjd zilog->zl_sync = sync; 1663219089Spjd} 1664219089Spjd 1665219089Spjdvoid 1666219089Spjdzil_set_logbias(zilog_t *zilog, uint64_t logbias) 1667219089Spjd{ 1668219089Spjd zilog->zl_logbias = logbias; 1669219089Spjd} 1670219089Spjd 1671168404Spjdzilog_t * 1672168404Spjdzil_alloc(objset_t *os, zil_header_t *zh_phys) 1673168404Spjd{ 1674168404Spjd zilog_t *zilog; 1675168404Spjd 1676168404Spjd zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1677168404Spjd 1678168404Spjd zilog->zl_header = zh_phys; 1679168404Spjd zilog->zl_os = os; 1680168404Spjd zilog->zl_spa = dmu_objset_spa(os); 1681168404Spjd zilog->zl_dmu_pool = dmu_objset_pool(os); 1682168404Spjd zilog->zl_destroy_txg = TXG_INITIAL - 1; 1683219089Spjd zilog->zl_logbias = dmu_objset_logbias(os); 1684219089Spjd zilog->zl_sync = dmu_objset_syncprop(os); 1685219089Spjd zilog->zl_next_batch = 1; 1686168404Spjd 1687168404Spjd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1688168404Spjd 1689219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1690219089Spjd mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1691219089Spjd MUTEX_DEFAULT, NULL); 1692219089Spjd } 1693168404Spjd 1694168404Spjd list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1695168404Spjd offsetof(lwb_t, lwb_node)); 1696168404Spjd 1697219089Spjd list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1698219089Spjd offsetof(itx_t, itx_node)); 1699219089Spjd 1700185029Spjd mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1701168404Spjd 1702185029Spjd avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1703185029Spjd sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1704185029Spjd 1705185029Spjd cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1706185029Spjd cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1707219089Spjd cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1708219089Spjd cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1709185029Spjd 1710168404Spjd return (zilog); 1711168404Spjd} 1712168404Spjd 1713168404Spjdvoid 1714168404Spjdzil_free(zilog_t *zilog) 1715168404Spjd{ 1716168404Spjd zilog->zl_stop_sync = 1; 1717168404Spjd 1718224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1719168404Spjd list_destroy(&zilog->zl_lwb_list); 1720168404Spjd 1721185029Spjd avl_destroy(&zilog->zl_vdev_tree); 1722185029Spjd mutex_destroy(&zilog->zl_vdev_lock); 1723168404Spjd 1724219089Spjd ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1725219089Spjd list_destroy(&zilog->zl_itx_commit_list); 1726219089Spjd 1727219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1728219089Spjd /* 1729219089Spjd * It's possible for an itx to be generated that doesn't dirty 1730219089Spjd * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1731219089Spjd * callback to remove the entry. We remove those here. 1732219089Spjd * 1733219089Spjd * Also free up the ziltest itxs. 1734219089Spjd */ 1735219089Spjd if (zilog->zl_itxg[i].itxg_itxs) 1736219089Spjd zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1737219089Spjd mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1738219089Spjd } 1739219089Spjd 1740168404Spjd mutex_destroy(&zilog->zl_lock); 1741168404Spjd 1742185029Spjd cv_destroy(&zilog->zl_cv_writer); 1743185029Spjd cv_destroy(&zilog->zl_cv_suspend); 1744219089Spjd cv_destroy(&zilog->zl_cv_batch[0]); 1745219089Spjd cv_destroy(&zilog->zl_cv_batch[1]); 1746185029Spjd 1747168404Spjd kmem_free(zilog, sizeof (zilog_t)); 1748168404Spjd} 1749168404Spjd 1750168404Spjd/* 1751168404Spjd * Open an intent log. 1752168404Spjd */ 1753168404Spjdzilog_t * 1754168404Spjdzil_open(objset_t *os, zil_get_data_t *get_data) 1755168404Spjd{ 1756168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1757168404Spjd 1758224526Smm ASSERT(zilog->zl_clean_taskq == NULL); 1759224526Smm ASSERT(zilog->zl_get_data == NULL); 1760224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1761224526Smm 1762168404Spjd zilog->zl_get_data = get_data; 1763168404Spjd zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1764168404Spjd 2, 2, TASKQ_PREPOPULATE); 1765168404Spjd 1766168404Spjd return (zilog); 1767168404Spjd} 1768168404Spjd 1769168404Spjd/* 1770168404Spjd * Close an intent log. 1771168404Spjd */ 1772168404Spjdvoid 1773168404Spjdzil_close(zilog_t *zilog) 1774168404Spjd{ 1775224526Smm lwb_t *lwb; 1776219089Spjd uint64_t txg = 0; 1777219089Spjd 1778219089Spjd zil_commit(zilog, 0); /* commit all itx */ 1779219089Spjd 1780168404Spjd /* 1781219089Spjd * The lwb_max_txg for the stubby lwb will reflect the last activity 1782219089Spjd * for the zil. After a txg_wait_synced() on the txg we know all the 1783219089Spjd * callbacks have occurred that may clean the zil. Only then can we 1784219089Spjd * destroy the zl_clean_taskq. 1785168404Spjd */ 1786219089Spjd mutex_enter(&zilog->zl_lock); 1787224526Smm lwb = list_tail(&zilog->zl_lwb_list); 1788224526Smm if (lwb != NULL) 1789224526Smm txg = lwb->lwb_max_txg; 1790219089Spjd mutex_exit(&zilog->zl_lock); 1791219089Spjd if (txg) 1792168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1793243674Smm ASSERT(!zilog_is_dirty(zilog)); 1794168404Spjd 1795168404Spjd taskq_destroy(zilog->zl_clean_taskq); 1796168404Spjd zilog->zl_clean_taskq = NULL; 1797168404Spjd zilog->zl_get_data = NULL; 1798224526Smm 1799224526Smm /* 1800224526Smm * We should have only one LWB left on the list; remove it now. 1801224526Smm */ 1802224526Smm mutex_enter(&zilog->zl_lock); 1803224526Smm lwb = list_head(&zilog->zl_lwb_list); 1804224526Smm if (lwb != NULL) { 1805224526Smm ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1806224526Smm list_remove(&zilog->zl_lwb_list, lwb); 1807224526Smm zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1808224526Smm kmem_cache_free(zil_lwb_cache, lwb); 1809224526Smm } 1810224526Smm mutex_exit(&zilog->zl_lock); 1811168404Spjd} 1812168404Spjd 1813168404Spjd/* 1814168404Spjd * Suspend an intent log. While in suspended mode, we still honor 1815168404Spjd * synchronous semantics, but we rely on txg_wait_synced() to do it. 1816168404Spjd * We suspend the log briefly when taking a snapshot so that the snapshot 1817168404Spjd * contains all the data it's supposed to, and has an empty intent log. 1818168404Spjd */ 1819168404Spjdint 1820168404Spjdzil_suspend(zilog_t *zilog) 1821168404Spjd{ 1822168404Spjd const zil_header_t *zh = zilog->zl_header; 1823168404Spjd 1824168404Spjd mutex_enter(&zilog->zl_lock); 1825200724Sdelphij if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1826168404Spjd mutex_exit(&zilog->zl_lock); 1827168404Spjd return (EBUSY); 1828168404Spjd } 1829168404Spjd if (zilog->zl_suspend++ != 0) { 1830168404Spjd /* 1831168404Spjd * Someone else already began a suspend. 1832168404Spjd * Just wait for them to finish. 1833168404Spjd */ 1834168404Spjd while (zilog->zl_suspending) 1835168404Spjd cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1836168404Spjd mutex_exit(&zilog->zl_lock); 1837168404Spjd return (0); 1838168404Spjd } 1839168404Spjd zilog->zl_suspending = B_TRUE; 1840168404Spjd mutex_exit(&zilog->zl_lock); 1841168404Spjd 1842219089Spjd zil_commit(zilog, 0); 1843168404Spjd 1844168404Spjd zil_destroy(zilog, B_FALSE); 1845168404Spjd 1846168404Spjd mutex_enter(&zilog->zl_lock); 1847168404Spjd zilog->zl_suspending = B_FALSE; 1848168404Spjd cv_broadcast(&zilog->zl_cv_suspend); 1849168404Spjd mutex_exit(&zilog->zl_lock); 1850168404Spjd 1851168404Spjd return (0); 1852168404Spjd} 1853168404Spjd 1854168404Spjdvoid 1855168404Spjdzil_resume(zilog_t *zilog) 1856168404Spjd{ 1857168404Spjd mutex_enter(&zilog->zl_lock); 1858168404Spjd ASSERT(zilog->zl_suspend != 0); 1859168404Spjd zilog->zl_suspend--; 1860168404Spjd mutex_exit(&zilog->zl_lock); 1861168404Spjd} 1862168404Spjd 1863219089Spjdtypedef struct zil_replay_arg { 1864219089Spjd zil_replay_func_t **zr_replay; 1865219089Spjd void *zr_arg; 1866219089Spjd boolean_t zr_byteswap; 1867219089Spjd char *zr_lr; 1868219089Spjd} zil_replay_arg_t; 1869219089Spjd 1870219089Spjdstatic int 1871219089Spjdzil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1872209962Smm{ 1873219089Spjd char name[MAXNAMELEN]; 1874209962Smm 1875219089Spjd zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1876209962Smm 1877219089Spjd dmu_objset_name(zilog->zl_os, name); 1878209962Smm 1879219089Spjd cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1880219089Spjd "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1881219089Spjd (u_longlong_t)lr->lrc_seq, 1882219089Spjd (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1883219089Spjd (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1884219089Spjd 1885219089Spjd return (error); 1886209962Smm} 1887209962Smm 1888219089Spjdstatic int 1889168404Spjdzil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1890168404Spjd{ 1891168404Spjd zil_replay_arg_t *zr = zra; 1892168404Spjd const zil_header_t *zh = zilog->zl_header; 1893168404Spjd uint64_t reclen = lr->lrc_reclen; 1894168404Spjd uint64_t txtype = lr->lrc_txtype; 1895219089Spjd int error = 0; 1896168404Spjd 1897219089Spjd zilog->zl_replaying_seq = lr->lrc_seq; 1898168404Spjd 1899219089Spjd if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1900219089Spjd return (0); 1901219089Spjd 1902168404Spjd if (lr->lrc_txg < claim_txg) /* already committed */ 1903219089Spjd return (0); 1904168404Spjd 1905185029Spjd /* Strip case-insensitive bit, still present in log record */ 1906185029Spjd txtype &= ~TX_CI; 1907185029Spjd 1908219089Spjd if (txtype == 0 || txtype >= TX_MAX_TYPE) 1909219089Spjd return (zil_replay_error(zilog, lr, EINVAL)); 1910219089Spjd 1911219089Spjd /* 1912219089Spjd * If this record type can be logged out of order, the object 1913219089Spjd * (lr_foid) may no longer exist. That's legitimate, not an error. 1914219089Spjd */ 1915219089Spjd if (TX_OOO(txtype)) { 1916219089Spjd error = dmu_object_info(zilog->zl_os, 1917219089Spjd ((lr_ooo_t *)lr)->lr_foid, NULL); 1918219089Spjd if (error == ENOENT || error == EEXIST) 1919219089Spjd return (0); 1920209962Smm } 1921209962Smm 1922168404Spjd /* 1923168404Spjd * Make a copy of the data so we can revise and extend it. 1924168404Spjd */ 1925219089Spjd bcopy(lr, zr->zr_lr, reclen); 1926168404Spjd 1927168404Spjd /* 1928219089Spjd * If this is a TX_WRITE with a blkptr, suck in the data. 1929219089Spjd */ 1930219089Spjd if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1931219089Spjd error = zil_read_log_data(zilog, (lr_write_t *)lr, 1932219089Spjd zr->zr_lr + reclen); 1933219089Spjd if (error) 1934219089Spjd return (zil_replay_error(zilog, lr, error)); 1935219089Spjd } 1936219089Spjd 1937219089Spjd /* 1938168404Spjd * The log block containing this lr may have been byteswapped 1939168404Spjd * so that we can easily examine common fields like lrc_txtype. 1940219089Spjd * However, the log is a mix of different record types, and only the 1941168404Spjd * replay vectors know how to byteswap their records. Therefore, if 1942168404Spjd * the lr was byteswapped, undo it before invoking the replay vector. 1943168404Spjd */ 1944168404Spjd if (zr->zr_byteswap) 1945219089Spjd byteswap_uint64_array(zr->zr_lr, reclen); 1946168404Spjd 1947168404Spjd /* 1948168404Spjd * We must now do two things atomically: replay this log record, 1949209962Smm * and update the log header sequence number to reflect the fact that 1950209962Smm * we did so. At the end of each replay function the sequence number 1951209962Smm * is updated if we are in replay mode. 1952168404Spjd */ 1953219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 1954219089Spjd if (error) { 1955168404Spjd /* 1956168404Spjd * The DMU's dnode layer doesn't see removes until the txg 1957168404Spjd * commits, so a subsequent claim can spuriously fail with 1958209962Smm * EEXIST. So if we receive any error we try syncing out 1959219089Spjd * any removes then retry the transaction. Note that we 1960219089Spjd * specify B_FALSE for byteswap now, so we don't do it twice. 1961168404Spjd */ 1962219089Spjd txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1963219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 1964219089Spjd if (error) 1965219089Spjd return (zil_replay_error(zilog, lr, error)); 1966168404Spjd } 1967219089Spjd return (0); 1968168404Spjd} 1969168404Spjd 1970168404Spjd/* ARGSUSED */ 1971219089Spjdstatic int 1972168404Spjdzil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 1973168404Spjd{ 1974168404Spjd zilog->zl_replay_blks++; 1975219089Spjd 1976219089Spjd return (0); 1977168404Spjd} 1978168404Spjd 1979168404Spjd/* 1980168404Spjd * If this dataset has a non-empty intent log, replay it and destroy it. 1981168404Spjd */ 1982168404Spjdvoid 1983209962Smmzil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1984168404Spjd{ 1985168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1986168404Spjd const zil_header_t *zh = zilog->zl_header; 1987168404Spjd zil_replay_arg_t zr; 1988168404Spjd 1989200724Sdelphij if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1990168404Spjd zil_destroy(zilog, B_TRUE); 1991168404Spjd return; 1992168404Spjd } 1993168404Spjd //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 1994168404Spjd 1995168404Spjd zr.zr_replay = replay_func; 1996168404Spjd zr.zr_arg = arg; 1997168404Spjd zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1998219089Spjd zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1999168404Spjd 2000168404Spjd /* 2001168404Spjd * Wait for in-progress removes to sync before starting replay. 2002168404Spjd */ 2003168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 2004168404Spjd 2005209962Smm zilog->zl_replay = B_TRUE; 2006219089Spjd zilog->zl_replay_time = ddi_get_lbolt(); 2007168404Spjd ASSERT(zilog->zl_replay_blks == 0); 2008168404Spjd (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2009168404Spjd zh->zh_claim_txg); 2010219089Spjd kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2011168404Spjd 2012168404Spjd zil_destroy(zilog, B_FALSE); 2013185029Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2014209962Smm zilog->zl_replay = B_FALSE; 2015168404Spjd //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2016168404Spjd} 2017168404Spjd 2018219089Spjdboolean_t 2019219089Spjdzil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2020168404Spjd{ 2021219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2022219089Spjd return (B_TRUE); 2023168404Spjd 2024219089Spjd if (zilog->zl_replay) { 2025219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2026219089Spjd zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2027219089Spjd zilog->zl_replaying_seq; 2028219089Spjd return (B_TRUE); 2029168404Spjd } 2030168404Spjd 2031219089Spjd return (B_FALSE); 2032168404Spjd} 2033213197Smm 2034213197Smm/* ARGSUSED */ 2035213197Smmint 2036219089Spjdzil_vdev_offline(const char *osname, void *arg) 2037213197Smm{ 2038213197Smm objset_t *os; 2039213197Smm zilog_t *zilog; 2040213197Smm int error; 2041213197Smm 2042219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 2043213197Smm if (error) 2044213197Smm return (error); 2045213197Smm 2046213197Smm zilog = dmu_objset_zil(os); 2047213197Smm if (zil_suspend(zilog) != 0) 2048213197Smm error = EEXIST; 2049213197Smm else 2050213197Smm zil_resume(zilog); 2051219089Spjd dmu_objset_rele(os, FTAG); 2052213197Smm return (error); 2053213197Smm} 2054