zil.c revision 288549
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. 23268657Sdelphij * Copyright (c) 2011, 2014 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/* 69251631Sdelphij * Disable intent logging replay. This global ZIL switch affects all pools. 70168404Spjd */ 71251631Sdelphijint zil_replay_disable = 0; 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"); 86249921Ssmhboolean_t zfs_trim_enabled = B_TRUE; 87249921SsmhSYSCTL_DECL(_vfs_zfs_trim); 88249921SsmhTUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 89249921SsmhSYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 90249921Ssmh "Enable ZFS TRIM"); 91168404Spjd 92168404Spjdstatic kmem_cache_t *zil_lwb_cache; 93168404Spjd 94219089Spjdstatic void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 95219089Spjd 96219089Spjd#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 97219089Spjd sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 98219089Spjd 99219089Spjd 100219089Spjd/* 101219089Spjd * ziltest is by and large an ugly hack, but very useful in 102219089Spjd * checking replay without tedious work. 103219089Spjd * When running ziltest we want to keep all itx's and so maintain 104219089Spjd * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 105219089Spjd * We subtract TXG_CONCURRENT_STATES to allow for common code. 106219089Spjd */ 107219089Spjd#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 108219089Spjd 109168404Spjdstatic int 110219089Spjdzil_bp_compare(const void *x1, const void *x2) 111168404Spjd{ 112219089Spjd const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 113219089Spjd const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 114168404Spjd 115168404Spjd if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 116168404Spjd return (-1); 117168404Spjd if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 118168404Spjd return (1); 119168404Spjd 120168404Spjd if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 121168404Spjd return (-1); 122168404Spjd if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 123168404Spjd return (1); 124168404Spjd 125168404Spjd return (0); 126168404Spjd} 127168404Spjd 128168404Spjdstatic void 129219089Spjdzil_bp_tree_init(zilog_t *zilog) 130168404Spjd{ 131219089Spjd avl_create(&zilog->zl_bp_tree, zil_bp_compare, 132219089Spjd sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 133168404Spjd} 134168404Spjd 135168404Spjdstatic void 136219089Spjdzil_bp_tree_fini(zilog_t *zilog) 137168404Spjd{ 138219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 139219089Spjd zil_bp_node_t *zn; 140168404Spjd void *cookie = NULL; 141168404Spjd 142168404Spjd while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 143219089Spjd kmem_free(zn, sizeof (zil_bp_node_t)); 144168404Spjd 145168404Spjd avl_destroy(t); 146168404Spjd} 147168404Spjd 148219089Spjdint 149219089Spjdzil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 150168404Spjd{ 151219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 152268649Sdelphij const dva_t *dva; 153219089Spjd zil_bp_node_t *zn; 154168404Spjd avl_index_t where; 155168404Spjd 156268649Sdelphij if (BP_IS_EMBEDDED(bp)) 157268649Sdelphij return (0); 158268649Sdelphij 159268649Sdelphij dva = BP_IDENTITY(bp); 160268649Sdelphij 161168404Spjd if (avl_find(t, dva, &where) != NULL) 162249195Smm return (SET_ERROR(EEXIST)); 163168404Spjd 164219089Spjd zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 165168404Spjd zn->zn_dva = *dva; 166168404Spjd avl_insert(t, zn, where); 167168404Spjd 168168404Spjd return (0); 169168404Spjd} 170168404Spjd 171168404Spjdstatic zil_header_t * 172168404Spjdzil_header_in_syncing_context(zilog_t *zilog) 173168404Spjd{ 174168404Spjd return ((zil_header_t *)zilog->zl_header); 175168404Spjd} 176168404Spjd 177168404Spjdstatic void 178168404Spjdzil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 179168404Spjd{ 180168404Spjd zio_cksum_t *zc = &bp->blk_cksum; 181168404Spjd 182168404Spjd zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 183168404Spjd zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 184168404Spjd zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 185168404Spjd zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 186168404Spjd} 187168404Spjd 188168404Spjd/* 189219089Spjd * Read a log block and make sure it's valid. 190168404Spjd */ 191168404Spjdstatic int 192219089Spjdzil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 193219089Spjd char **end) 194168404Spjd{ 195219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 196277586Sdelphij arc_flags_t aflags = ARC_FLAG_WAIT; 197219089Spjd arc_buf_t *abuf = NULL; 198268657Sdelphij zbookmark_phys_t zb; 199168404Spjd int error; 200168404Spjd 201219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 202219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 203168404Spjd 204219089Spjd if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 205219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE; 206168404Spjd 207219089Spjd SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 208219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 209168404Spjd 210246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 211219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 212219089Spjd 213168404Spjd if (error == 0) { 214168404Spjd zio_cksum_t cksum = bp->blk_cksum; 215168404Spjd 216168404Spjd /* 217185029Spjd * Validate the checksummed log block. 218185029Spjd * 219168404Spjd * Sequence numbers should be... sequential. The checksum 220168404Spjd * verifier for the next block should be bp's checksum plus 1. 221185029Spjd * 222185029Spjd * Also check the log chain linkage and size used. 223168404Spjd */ 224168404Spjd cksum.zc_word[ZIL_ZC_SEQ]++; 225168404Spjd 226219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 227219089Spjd zil_chain_t *zilc = abuf->b_data; 228219089Spjd char *lr = (char *)(zilc + 1); 229219089Spjd uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 230219089Spjd 231219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 232219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 233249195Smm error = SET_ERROR(ECKSUM); 234219089Spjd } else { 235276081Sdelphij ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 236219089Spjd bcopy(lr, dst, len); 237219089Spjd *end = (char *)dst + len; 238219089Spjd *nbp = zilc->zc_next_blk; 239219089Spjd } 240219089Spjd } else { 241219089Spjd char *lr = abuf->b_data; 242219089Spjd uint64_t size = BP_GET_LSIZE(bp); 243219089Spjd zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 244219089Spjd 245219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 246219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 247219089Spjd (zilc->zc_nused > (size - sizeof (*zilc)))) { 248249195Smm error = SET_ERROR(ECKSUM); 249219089Spjd } else { 250276081Sdelphij ASSERT3U(zilc->zc_nused, <=, 251276081Sdelphij SPA_OLD_MAXBLOCKSIZE); 252219089Spjd bcopy(lr, dst, zilc->zc_nused); 253219089Spjd *end = (char *)dst + zilc->zc_nused; 254219089Spjd *nbp = zilc->zc_next_blk; 255219089Spjd } 256185029Spjd } 257168404Spjd 258248571Smm VERIFY(arc_buf_remove_ref(abuf, &abuf)); 259168404Spjd } 260168404Spjd 261219089Spjd return (error); 262219089Spjd} 263168404Spjd 264219089Spjd/* 265219089Spjd * Read a TX_WRITE log data block. 266219089Spjd */ 267219089Spjdstatic int 268219089Spjdzil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 269219089Spjd{ 270219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 271219089Spjd const blkptr_t *bp = &lr->lr_blkptr; 272277586Sdelphij arc_flags_t aflags = ARC_FLAG_WAIT; 273219089Spjd arc_buf_t *abuf = NULL; 274268657Sdelphij zbookmark_phys_t zb; 275219089Spjd int error; 276219089Spjd 277219089Spjd if (BP_IS_HOLE(bp)) { 278219089Spjd if (wbuf != NULL) 279219089Spjd bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 280219089Spjd return (0); 281219089Spjd } 282219089Spjd 283219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 284219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 285219089Spjd 286219089Spjd SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 287219089Spjd ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 288219089Spjd 289246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 290219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 291219089Spjd 292219089Spjd if (error == 0) { 293219089Spjd if (wbuf != NULL) 294219089Spjd bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 295219089Spjd (void) arc_buf_remove_ref(abuf, &abuf); 296219089Spjd } 297219089Spjd 298168404Spjd return (error); 299168404Spjd} 300168404Spjd 301168404Spjd/* 302168404Spjd * Parse the intent log, and call parse_func for each valid record within. 303168404Spjd */ 304219089Spjdint 305168404Spjdzil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 306168404Spjd zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 307168404Spjd{ 308168404Spjd const zil_header_t *zh = zilog->zl_header; 309219089Spjd boolean_t claimed = !!zh->zh_claim_txg; 310219089Spjd uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 311219089Spjd uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 312219089Spjd uint64_t max_blk_seq = 0; 313219089Spjd uint64_t max_lr_seq = 0; 314219089Spjd uint64_t blk_count = 0; 315219089Spjd uint64_t lr_count = 0; 316219089Spjd blkptr_t blk, next_blk; 317168404Spjd char *lrbuf, *lrp; 318219089Spjd int error = 0; 319168404Spjd 320219089Spjd /* 321219089Spjd * Old logs didn't record the maximum zh_claim_lr_seq. 322219089Spjd */ 323219089Spjd if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 324219089Spjd claim_lr_seq = UINT64_MAX; 325168404Spjd 326168404Spjd /* 327168404Spjd * Starting at the block pointed to by zh_log we read the log chain. 328168404Spjd * For each block in the chain we strongly check that block to 329168404Spjd * ensure its validity. We stop when an invalid block is found. 330168404Spjd * For each block pointer in the chain we call parse_blk_func(). 331168404Spjd * For each record in each valid block we call parse_lr_func(). 332168404Spjd * If the log has been claimed, stop if we encounter a sequence 333168404Spjd * number greater than the highest claimed sequence number. 334168404Spjd */ 335276081Sdelphij lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 336219089Spjd zil_bp_tree_init(zilog); 337168404Spjd 338219089Spjd for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 339219089Spjd uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 340219089Spjd int reclen; 341219089Spjd char *end; 342219089Spjd 343219089Spjd if (blk_seq > claim_blk_seq) 344168404Spjd break; 345219089Spjd if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 346219089Spjd break; 347219089Spjd ASSERT3U(max_blk_seq, <, blk_seq); 348219089Spjd max_blk_seq = blk_seq; 349219089Spjd blk_count++; 350168404Spjd 351219089Spjd if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 352219089Spjd break; 353168404Spjd 354219089Spjd error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 355248571Smm if (error != 0) 356168404Spjd break; 357168404Spjd 358219089Spjd for (lrp = lrbuf; lrp < end; lrp += reclen) { 359168404Spjd lr_t *lr = (lr_t *)lrp; 360168404Spjd reclen = lr->lrc_reclen; 361168404Spjd ASSERT3U(reclen, >=, sizeof (lr_t)); 362219089Spjd if (lr->lrc_seq > claim_lr_seq) 363219089Spjd goto done; 364219089Spjd if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 365219089Spjd goto done; 366219089Spjd ASSERT3U(max_lr_seq, <, lr->lrc_seq); 367219089Spjd max_lr_seq = lr->lrc_seq; 368219089Spjd lr_count++; 369168404Spjd } 370168404Spjd } 371219089Spjddone: 372219089Spjd zilog->zl_parse_error = error; 373219089Spjd zilog->zl_parse_blk_seq = max_blk_seq; 374219089Spjd zilog->zl_parse_lr_seq = max_lr_seq; 375219089Spjd zilog->zl_parse_blk_count = blk_count; 376219089Spjd zilog->zl_parse_lr_count = lr_count; 377168404Spjd 378219089Spjd ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 379219089Spjd (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 380219089Spjd 381219089Spjd zil_bp_tree_fini(zilog); 382276081Sdelphij zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 383219089Spjd 384219089Spjd return (error); 385168404Spjd} 386168404Spjd 387219089Spjdstatic int 388168404Spjdzil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 389168404Spjd{ 390168404Spjd /* 391168404Spjd * Claim log block if not already committed and not already claimed. 392219089Spjd * If tx == NULL, just verify that the block is claimable. 393168404Spjd */ 394263397Sdelphij if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 395263397Sdelphij zil_bp_tree_add(zilog, bp) != 0) 396219089Spjd return (0); 397219089Spjd 398219089Spjd return (zio_wait(zio_claim(NULL, zilog->zl_spa, 399219089Spjd tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 400219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 401168404Spjd} 402168404Spjd 403219089Spjdstatic int 404168404Spjdzil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 405168404Spjd{ 406219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 407219089Spjd int error; 408219089Spjd 409219089Spjd if (lrc->lrc_txtype != TX_WRITE) 410219089Spjd return (0); 411219089Spjd 412219089Spjd /* 413219089Spjd * If the block is not readable, don't claim it. This can happen 414219089Spjd * in normal operation when a log block is written to disk before 415219089Spjd * some of the dmu_sync() blocks it points to. In this case, the 416219089Spjd * transaction cannot have been committed to anyone (we would have 417219089Spjd * waited for all writes to be stable first), so it is semantically 418219089Spjd * correct to declare this the end of the log. 419219089Spjd */ 420219089Spjd if (lr->lr_blkptr.blk_birth >= first_txg && 421219089Spjd (error = zil_read_log_data(zilog, lr, NULL)) != 0) 422219089Spjd return (error); 423219089Spjd return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 424168404Spjd} 425168404Spjd 426168404Spjd/* ARGSUSED */ 427219089Spjdstatic int 428168404Spjdzil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 429168404Spjd{ 430219089Spjd zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 431219089Spjd 432219089Spjd return (0); 433168404Spjd} 434168404Spjd 435219089Spjdstatic int 436168404Spjdzil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 437168404Spjd{ 438219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 439219089Spjd blkptr_t *bp = &lr->lr_blkptr; 440219089Spjd 441168404Spjd /* 442168404Spjd * If we previously claimed it, we need to free it. 443168404Spjd */ 444219089Spjd if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 445263397Sdelphij bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 446263397Sdelphij !BP_IS_HOLE(bp)) 447219089Spjd zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 448219089Spjd 449219089Spjd return (0); 450219089Spjd} 451219089Spjd 452219089Spjdstatic lwb_t * 453219089Spjdzil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 454219089Spjd{ 455219089Spjd lwb_t *lwb; 456219089Spjd 457219089Spjd lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 458219089Spjd lwb->lwb_zilog = zilog; 459219089Spjd lwb->lwb_blk = *bp; 460219089Spjd lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 461219089Spjd lwb->lwb_max_txg = txg; 462219089Spjd lwb->lwb_zio = NULL; 463219089Spjd lwb->lwb_tx = NULL; 464219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 465219089Spjd lwb->lwb_nused = sizeof (zil_chain_t); 466219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp); 467219089Spjd } else { 468219089Spjd lwb->lwb_nused = 0; 469219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 470168404Spjd } 471219089Spjd 472219089Spjd mutex_enter(&zilog->zl_lock); 473219089Spjd list_insert_tail(&zilog->zl_lwb_list, lwb); 474219089Spjd mutex_exit(&zilog->zl_lock); 475219089Spjd 476219089Spjd return (lwb); 477168404Spjd} 478168404Spjd 479168404Spjd/* 480239620Smm * Called when we create in-memory log transactions so that we know 481239620Smm * to cleanup the itxs at the end of spa_sync(). 482239620Smm */ 483239620Smmvoid 484239620Smmzilog_dirty(zilog_t *zilog, uint64_t txg) 485239620Smm{ 486239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 487239620Smm dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 488239620Smm 489288549Smav if (ds->ds_is_snapshot) 490239620Smm panic("dirtying snapshot!"); 491239620Smm 492248571Smm if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 493239620Smm /* up the hold count until we can be written out */ 494239620Smm dmu_buf_add_ref(ds->ds_dbuf, zilog); 495239620Smm } 496239620Smm} 497239620Smm 498239620Smmboolean_t 499239620Smmzilog_is_dirty(zilog_t *zilog) 500239620Smm{ 501239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 502239620Smm 503239620Smm for (int t = 0; t < TXG_SIZE; t++) { 504239620Smm if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 505239620Smm return (B_TRUE); 506239620Smm } 507239620Smm return (B_FALSE); 508239620Smm} 509239620Smm 510239620Smm/* 511168404Spjd * Create an on-disk intent log. 512168404Spjd */ 513219089Spjdstatic lwb_t * 514168404Spjdzil_create(zilog_t *zilog) 515168404Spjd{ 516168404Spjd const zil_header_t *zh = zilog->zl_header; 517219089Spjd lwb_t *lwb = NULL; 518168404Spjd uint64_t txg = 0; 519168404Spjd dmu_tx_t *tx = NULL; 520168404Spjd blkptr_t blk; 521168404Spjd int error = 0; 522168404Spjd 523168404Spjd /* 524168404Spjd * Wait for any previous destroy to complete. 525168404Spjd */ 526168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 527168404Spjd 528168404Spjd ASSERT(zh->zh_claim_txg == 0); 529168404Spjd ASSERT(zh->zh_replay_seq == 0); 530168404Spjd 531168404Spjd blk = zh->zh_log; 532168404Spjd 533168404Spjd /* 534219089Spjd * Allocate an initial log block if: 535219089Spjd * - there isn't one already 536219089Spjd * - the existing block is the wrong endianess 537168404Spjd */ 538207908Smm if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 539168404Spjd tx = dmu_tx_create(zilog->zl_os); 540219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 541168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 542168404Spjd txg = dmu_tx_get_txg(tx); 543168404Spjd 544207908Smm if (!BP_IS_HOLE(&blk)) { 545219089Spjd zio_free_zil(zilog->zl_spa, txg, &blk); 546207908Smm BP_ZERO(&blk); 547207908Smm } 548207908Smm 549219089Spjd error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 550219089Spjd ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 551168404Spjd 552168404Spjd if (error == 0) 553168404Spjd zil_init_log_chain(zilog, &blk); 554168404Spjd } 555168404Spjd 556168404Spjd /* 557168404Spjd * Allocate a log write buffer (lwb) for the first log block. 558168404Spjd */ 559219089Spjd if (error == 0) 560219089Spjd lwb = zil_alloc_lwb(zilog, &blk, txg); 561168404Spjd 562168404Spjd /* 563168404Spjd * If we just allocated the first log block, commit our transaction 564168404Spjd * and wait for zil_sync() to stuff the block poiner into zh_log. 565168404Spjd * (zh is part of the MOS, so we cannot modify it in open context.) 566168404Spjd */ 567168404Spjd if (tx != NULL) { 568168404Spjd dmu_tx_commit(tx); 569168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 570168404Spjd } 571168404Spjd 572168404Spjd ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 573219089Spjd 574219089Spjd return (lwb); 575168404Spjd} 576168404Spjd 577168404Spjd/* 578168404Spjd * In one tx, free all log blocks and clear the log header. 579168404Spjd * If keep_first is set, then we're replaying a log with no content. 580168404Spjd * We want to keep the first block, however, so that the first 581168404Spjd * synchronous transaction doesn't require a txg_wait_synced() 582168404Spjd * in zil_create(). We don't need to txg_wait_synced() here either 583168404Spjd * when keep_first is set, because both zil_create() and zil_destroy() 584168404Spjd * will wait for any in-progress destroys to complete. 585168404Spjd */ 586168404Spjdvoid 587168404Spjdzil_destroy(zilog_t *zilog, boolean_t keep_first) 588168404Spjd{ 589168404Spjd const zil_header_t *zh = zilog->zl_header; 590168404Spjd lwb_t *lwb; 591168404Spjd dmu_tx_t *tx; 592168404Spjd uint64_t txg; 593168404Spjd 594168404Spjd /* 595168404Spjd * Wait for any previous destroy to complete. 596168404Spjd */ 597168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 598168404Spjd 599219089Spjd zilog->zl_old_header = *zh; /* debugging aid */ 600219089Spjd 601168404Spjd if (BP_IS_HOLE(&zh->zh_log)) 602168404Spjd return; 603168404Spjd 604168404Spjd tx = dmu_tx_create(zilog->zl_os); 605219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 606168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 607168404Spjd txg = dmu_tx_get_txg(tx); 608168404Spjd 609168404Spjd mutex_enter(&zilog->zl_lock); 610168404Spjd 611168404Spjd ASSERT3U(zilog->zl_destroy_txg, <, txg); 612168404Spjd zilog->zl_destroy_txg = txg; 613168404Spjd zilog->zl_keep_first = keep_first; 614168404Spjd 615168404Spjd if (!list_is_empty(&zilog->zl_lwb_list)) { 616168404Spjd ASSERT(zh->zh_claim_txg == 0); 617224526Smm VERIFY(!keep_first); 618168404Spjd while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 619168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 620168404Spjd if (lwb->lwb_buf != NULL) 621168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 622219089Spjd zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 623168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 624168404Spjd } 625219089Spjd } else if (!keep_first) { 626239620Smm zil_destroy_sync(zilog, tx); 627168404Spjd } 628168404Spjd mutex_exit(&zilog->zl_lock); 629168404Spjd 630168404Spjd dmu_tx_commit(tx); 631185029Spjd} 632168404Spjd 633239620Smmvoid 634239620Smmzil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 635239620Smm{ 636239620Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 637239620Smm (void) zil_parse(zilog, zil_free_log_block, 638239620Smm zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 639239620Smm} 640239620Smm 641168404Spjdint 642219089Spjdzil_claim(const char *osname, void *txarg) 643168404Spjd{ 644168404Spjd dmu_tx_t *tx = txarg; 645168404Spjd uint64_t first_txg = dmu_tx_get_txg(tx); 646168404Spjd zilog_t *zilog; 647168404Spjd zil_header_t *zh; 648168404Spjd objset_t *os; 649168404Spjd int error; 650168404Spjd 651248571Smm error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); 652248571Smm if (error != 0) { 653272133Sdelphij /* 654272133Sdelphij * EBUSY indicates that the objset is inconsistent, in which 655272133Sdelphij * case it can not have a ZIL. 656272133Sdelphij */ 657272133Sdelphij if (error != EBUSY) { 658272133Sdelphij cmn_err(CE_WARN, "can't open objset for %s, error %u", 659272133Sdelphij osname, error); 660272133Sdelphij } 661168404Spjd return (0); 662168404Spjd } 663168404Spjd 664168404Spjd zilog = dmu_objset_zil(os); 665168404Spjd zh = zil_header_in_syncing_context(zilog); 666168404Spjd 667219089Spjd if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 668213197Smm if (!BP_IS_HOLE(&zh->zh_log)) 669219089Spjd zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 670213197Smm BP_ZERO(&zh->zh_log); 671213197Smm dsl_dataset_dirty(dmu_objset_ds(os), tx); 672248571Smm dmu_objset_disown(os, FTAG); 673219089Spjd return (0); 674213197Smm } 675213197Smm 676168404Spjd /* 677168404Spjd * Claim all log blocks if we haven't already done so, and remember 678168404Spjd * the highest claimed sequence number. This ensures that if we can 679168404Spjd * read only part of the log now (e.g. due to a missing device), 680168404Spjd * but we can read the entire log later, we will not try to replay 681168404Spjd * or destroy beyond the last block we successfully claimed. 682168404Spjd */ 683168404Spjd ASSERT3U(zh->zh_claim_txg, <=, first_txg); 684168404Spjd if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 685219089Spjd (void) zil_parse(zilog, zil_claim_log_block, 686219089Spjd zil_claim_log_record, tx, first_txg); 687168404Spjd zh->zh_claim_txg = first_txg; 688219089Spjd zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 689219089Spjd zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 690219089Spjd if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 691219089Spjd zh->zh_flags |= ZIL_REPLAY_NEEDED; 692219089Spjd zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 693168404Spjd dsl_dataset_dirty(dmu_objset_ds(os), tx); 694168404Spjd } 695168404Spjd 696168404Spjd ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 697248571Smm dmu_objset_disown(os, FTAG); 698168404Spjd return (0); 699168404Spjd} 700168404Spjd 701185029Spjd/* 702185029Spjd * Check the log by walking the log chain. 703185029Spjd * Checksum errors are ok as they indicate the end of the chain. 704185029Spjd * Any other error (no device or read failure) returns an error. 705185029Spjd */ 706185029Spjdint 707219089Spjdzil_check_log_chain(const char *osname, void *tx) 708168404Spjd{ 709185029Spjd zilog_t *zilog; 710185029Spjd objset_t *os; 711219089Spjd blkptr_t *bp; 712185029Spjd int error; 713168404Spjd 714219089Spjd ASSERT(tx == NULL); 715219089Spjd 716219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 717248571Smm if (error != 0) { 718185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 719185029Spjd return (0); 720185029Spjd } 721168404Spjd 722185029Spjd zilog = dmu_objset_zil(os); 723219089Spjd bp = (blkptr_t *)&zilog->zl_header->zh_log; 724219089Spjd 725219089Spjd /* 726219089Spjd * Check the first block and determine if it's on a log device 727219089Spjd * which may have been removed or faulted prior to loading this 728219089Spjd * pool. If so, there's no point in checking the rest of the log 729219089Spjd * as its content should have already been synced to the pool. 730219089Spjd */ 731219089Spjd if (!BP_IS_HOLE(bp)) { 732219089Spjd vdev_t *vd; 733219089Spjd boolean_t valid = B_TRUE; 734219089Spjd 735219089Spjd spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 736219089Spjd vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 737219089Spjd if (vd->vdev_islog && vdev_is_dead(vd)) 738219089Spjd valid = vdev_log_state_valid(vd); 739219089Spjd spa_config_exit(os->os_spa, SCL_STATE, FTAG); 740219089Spjd 741219089Spjd if (!valid) { 742219089Spjd dmu_objset_rele(os, FTAG); 743219089Spjd return (0); 744219089Spjd } 745168404Spjd } 746185029Spjd 747219089Spjd /* 748219089Spjd * Because tx == NULL, zil_claim_log_block() will not actually claim 749219089Spjd * any blocks, but just determine whether it is possible to do so. 750219089Spjd * In addition to checking the log chain, zil_claim_log_block() 751219089Spjd * will invoke zio_claim() with a done func of spa_claim_notify(), 752219089Spjd * which will update spa_max_claim_txg. See spa_load() for details. 753219089Spjd */ 754219089Spjd error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 755219089Spjd zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 756219089Spjd 757219089Spjd dmu_objset_rele(os, FTAG); 758219089Spjd 759219089Spjd return ((error == ECKSUM || error == ENOENT) ? 0 : error); 760168404Spjd} 761168404Spjd 762185029Spjdstatic int 763185029Spjdzil_vdev_compare(const void *x1, const void *x2) 764185029Spjd{ 765219089Spjd const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 766219089Spjd const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 767185029Spjd 768185029Spjd if (v1 < v2) 769185029Spjd return (-1); 770185029Spjd if (v1 > v2) 771185029Spjd return (1); 772185029Spjd 773185029Spjd return (0); 774185029Spjd} 775185029Spjd 776168404Spjdvoid 777219089Spjdzil_add_block(zilog_t *zilog, const blkptr_t *bp) 778168404Spjd{ 779185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 780185029Spjd avl_index_t where; 781185029Spjd zil_vdev_node_t *zv, zvsearch; 782185029Spjd int ndvas = BP_GET_NDVAS(bp); 783185029Spjd int i; 784168404Spjd 785185029Spjd if (zfs_nocacheflush) 786185029Spjd return; 787168404Spjd 788185029Spjd ASSERT(zilog->zl_writer); 789168404Spjd 790185029Spjd /* 791185029Spjd * Even though we're zl_writer, we still need a lock because the 792185029Spjd * zl_get_data() callbacks may have dmu_sync() done callbacks 793185029Spjd * that will run concurrently. 794185029Spjd */ 795185029Spjd mutex_enter(&zilog->zl_vdev_lock); 796185029Spjd for (i = 0; i < ndvas; i++) { 797185029Spjd zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 798185029Spjd if (avl_find(t, &zvsearch, &where) == NULL) { 799185029Spjd zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 800185029Spjd zv->zv_vdev = zvsearch.zv_vdev; 801185029Spjd avl_insert(t, zv, where); 802185029Spjd } 803185029Spjd } 804185029Spjd mutex_exit(&zilog->zl_vdev_lock); 805168404Spjd} 806168404Spjd 807219089Spjdstatic void 808168404Spjdzil_flush_vdevs(zilog_t *zilog) 809168404Spjd{ 810168404Spjd spa_t *spa = zilog->zl_spa; 811185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 812185029Spjd void *cookie = NULL; 813185029Spjd zil_vdev_node_t *zv; 814185029Spjd zio_t *zio; 815168404Spjd 816168404Spjd ASSERT(zilog->zl_writer); 817168404Spjd 818185029Spjd /* 819185029Spjd * We don't need zl_vdev_lock here because we're the zl_writer, 820185029Spjd * and all zl_get_data() callbacks are done. 821185029Spjd */ 822185029Spjd if (avl_numnodes(t) == 0) 823185029Spjd return; 824185029Spjd 825185029Spjd spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 826185029Spjd 827185029Spjd zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 828185029Spjd 829185029Spjd while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 830185029Spjd vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 831185029Spjd if (vd != NULL) 832185029Spjd zio_flush(zio, vd); 833185029Spjd kmem_free(zv, sizeof (*zv)); 834168404Spjd } 835168404Spjd 836168404Spjd /* 837168404Spjd * Wait for all the flushes to complete. Not all devices actually 838168404Spjd * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 839168404Spjd */ 840185029Spjd (void) zio_wait(zio); 841185029Spjd 842185029Spjd spa_config_exit(spa, SCL_STATE, FTAG); 843168404Spjd} 844168404Spjd 845168404Spjd/* 846168404Spjd * Function called when a log block write completes 847168404Spjd */ 848168404Spjdstatic void 849168404Spjdzil_lwb_write_done(zio_t *zio) 850168404Spjd{ 851168404Spjd lwb_t *lwb = zio->io_private; 852168404Spjd zilog_t *zilog = lwb->lwb_zilog; 853219089Spjd dmu_tx_t *tx = lwb->lwb_tx; 854168404Spjd 855185029Spjd ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 856185029Spjd ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 857185029Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 858185029Spjd ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 859185029Spjd ASSERT(!BP_IS_GANG(zio->io_bp)); 860185029Spjd ASSERT(!BP_IS_HOLE(zio->io_bp)); 861268649Sdelphij ASSERT(BP_GET_FILL(zio->io_bp) == 0); 862185029Spjd 863168404Spjd /* 864209962Smm * Ensure the lwb buffer pointer is cleared before releasing 865209962Smm * the txg. If we have had an allocation failure and 866209962Smm * the txg is waiting to sync then we want want zil_sync() 867209962Smm * to remove the lwb so that it's not picked up as the next new 868209962Smm * one in zil_commit_writer(). zil_sync() will only remove 869209962Smm * the lwb if lwb_buf is null. 870168404Spjd */ 871168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 872168404Spjd mutex_enter(&zilog->zl_lock); 873168404Spjd lwb->lwb_buf = NULL; 874219089Spjd lwb->lwb_tx = NULL; 875219089Spjd mutex_exit(&zilog->zl_lock); 876209962Smm 877209962Smm /* 878209962Smm * Now that we've written this log block, we have a stable pointer 879209962Smm * to the next block in the chain, so it's OK to let the txg in 880219089Spjd * which we allocated the next block sync. 881209962Smm */ 882219089Spjd dmu_tx_commit(tx); 883168404Spjd} 884168404Spjd 885168404Spjd/* 886168404Spjd * Initialize the io for a log block. 887168404Spjd */ 888168404Spjdstatic void 889168404Spjdzil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 890168404Spjd{ 891268657Sdelphij zbookmark_phys_t zb; 892168404Spjd 893219089Spjd SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 894219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 895219089Spjd lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 896168404Spjd 897168404Spjd if (zilog->zl_root_zio == NULL) { 898168404Spjd zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 899168404Spjd ZIO_FLAG_CANFAIL); 900168404Spjd } 901168404Spjd if (lwb->lwb_zio == NULL) { 902168404Spjd lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 903219089Spjd 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 904260763Savg zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 905219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 906168404Spjd } 907168404Spjd} 908168404Spjd 909168404Spjd/* 910219089Spjd * Define a limited set of intent log block sizes. 911251631Sdelphij * 912219089Spjd * These must be a multiple of 4KB. Note only the amount used (again 913219089Spjd * aligned to 4KB) actually gets written. However, we can't always just 914276081Sdelphij * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 915219089Spjd */ 916219089Spjduint64_t zil_block_buckets[] = { 917219089Spjd 4096, /* non TX_WRITE */ 918219089Spjd 8192+4096, /* data base */ 919219089Spjd 32*1024 + 4096, /* NFS writes */ 920219089Spjd UINT64_MAX 921219089Spjd}; 922219089Spjd 923219089Spjd/* 924219089Spjd * Use the slog as long as the logbias is 'latency' and the current commit size 925219089Spjd * is less than the limit or the total list size is less than 2X the limit. 926219089Spjd * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 927219089Spjd */ 928219089Spjduint64_t zil_slog_limit = 1024 * 1024; 929219089Spjd#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 930219089Spjd (((zilog)->zl_cur_used < zil_slog_limit) || \ 931219089Spjd ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 932219089Spjd 933219089Spjd/* 934168404Spjd * Start a log block write and advance to the next log block. 935168404Spjd * Calls are serialized. 936168404Spjd */ 937168404Spjdstatic lwb_t * 938168404Spjdzil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 939168404Spjd{ 940219089Spjd lwb_t *nlwb = NULL; 941219089Spjd zil_chain_t *zilc; 942168404Spjd spa_t *spa = zilog->zl_spa; 943219089Spjd blkptr_t *bp; 944219089Spjd dmu_tx_t *tx; 945168404Spjd uint64_t txg; 946219089Spjd uint64_t zil_blksz, wsz; 947219089Spjd int i, error; 948168404Spjd 949219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 950219089Spjd zilc = (zil_chain_t *)lwb->lwb_buf; 951219089Spjd bp = &zilc->zc_next_blk; 952219089Spjd } else { 953219089Spjd zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 954219089Spjd bp = &zilc->zc_next_blk; 955219089Spjd } 956168404Spjd 957219089Spjd ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 958219089Spjd 959168404Spjd /* 960168404Spjd * Allocate the next block and save its address in this block 961168404Spjd * before writing it in order to establish the log chain. 962168404Spjd * Note that if the allocation of nlwb synced before we wrote 963168404Spjd * the block that points at it (lwb), we'd leak it if we crashed. 964219089Spjd * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 965219089Spjd * We dirty the dataset to ensure that zil_sync() will be called 966219089Spjd * to clean up in the event of allocation failure or I/O failure. 967168404Spjd */ 968219089Spjd tx = dmu_tx_create(zilog->zl_os); 969219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 970219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 971219089Spjd txg = dmu_tx_get_txg(tx); 972168404Spjd 973219089Spjd lwb->lwb_tx = tx; 974219089Spjd 975168404Spjd /* 976219089Spjd * Log blocks are pre-allocated. Here we select the size of the next 977219089Spjd * block, based on size used in the last block. 978219089Spjd * - first find the smallest bucket that will fit the block from a 979219089Spjd * limited set of block sizes. This is because it's faster to write 980219089Spjd * blocks allocated from the same metaslab as they are adjacent or 981219089Spjd * close. 982219089Spjd * - next find the maximum from the new suggested size and an array of 983219089Spjd * previous sizes. This lessens a picket fence effect of wrongly 984219089Spjd * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 985219089Spjd * requests. 986219089Spjd * 987219089Spjd * Note we only write what is used, but we can't just allocate 988219089Spjd * the maximum block size because we can exhaust the available 989219089Spjd * pool log space. 990168404Spjd */ 991219089Spjd zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 992219089Spjd for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 993219089Spjd continue; 994219089Spjd zil_blksz = zil_block_buckets[i]; 995219089Spjd if (zil_blksz == UINT64_MAX) 996276081Sdelphij zil_blksz = SPA_OLD_MAXBLOCKSIZE; 997219089Spjd zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 998219089Spjd for (i = 0; i < ZIL_PREV_BLKS; i++) 999219089Spjd zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1000219089Spjd zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1001168404Spjd 1002168404Spjd BP_ZERO(bp); 1003168404Spjd /* pass the old blkptr in order to spread log blocks across devs */ 1004219089Spjd error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 1005219089Spjd USE_SLOG(zilog)); 1006248571Smm if (error == 0) { 1007219089Spjd ASSERT3U(bp->blk_birth, ==, txg); 1008219089Spjd bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1009219089Spjd bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1010168404Spjd 1011168404Spjd /* 1012219089Spjd * Allocate a new log write buffer (lwb). 1013168404Spjd */ 1014219089Spjd nlwb = zil_alloc_lwb(zilog, bp, txg); 1015168404Spjd 1016219089Spjd /* Record the block for later vdev flushing */ 1017219089Spjd zil_add_block(zilog, &lwb->lwb_blk); 1018168404Spjd } 1019168404Spjd 1020219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1021219089Spjd /* For Slim ZIL only write what is used. */ 1022219089Spjd wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1023219089Spjd ASSERT3U(wsz, <=, lwb->lwb_sz); 1024219089Spjd zio_shrink(lwb->lwb_zio, wsz); 1025168404Spjd 1026219089Spjd } else { 1027219089Spjd wsz = lwb->lwb_sz; 1028219089Spjd } 1029168404Spjd 1030219089Spjd zilc->zc_pad = 0; 1031219089Spjd zilc->zc_nused = lwb->lwb_nused; 1032219089Spjd zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1033168404Spjd 1034168404Spjd /* 1035219089Spjd * clear unused data for security 1036168404Spjd */ 1037219089Spjd bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1038168404Spjd 1039219089Spjd zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1040168404Spjd 1041168404Spjd /* 1042219089Spjd * If there was an allocation failure then nlwb will be null which 1043219089Spjd * forces a txg_wait_synced(). 1044168404Spjd */ 1045168404Spjd return (nlwb); 1046168404Spjd} 1047168404Spjd 1048168404Spjdstatic lwb_t * 1049168404Spjdzil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1050168404Spjd{ 1051168404Spjd lr_t *lrc = &itx->itx_lr; /* common log record */ 1052219089Spjd lr_write_t *lrw = (lr_write_t *)lrc; 1053219089Spjd char *lr_buf; 1054168404Spjd uint64_t txg = lrc->lrc_txg; 1055168404Spjd uint64_t reclen = lrc->lrc_reclen; 1056219089Spjd uint64_t dlen = 0; 1057168404Spjd 1058168404Spjd if (lwb == NULL) 1059168404Spjd return (NULL); 1060219089Spjd 1061168404Spjd ASSERT(lwb->lwb_buf != NULL); 1062239620Smm ASSERT(zilog_is_dirty(zilog) || 1063239620Smm spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1064168404Spjd 1065168404Spjd if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1066168404Spjd dlen = P2ROUNDUP_TYPED( 1067219089Spjd lrw->lr_length, sizeof (uint64_t), uint64_t); 1068168404Spjd 1069168404Spjd zilog->zl_cur_used += (reclen + dlen); 1070168404Spjd 1071168404Spjd zil_lwb_write_init(zilog, lwb); 1072168404Spjd 1073168404Spjd /* 1074168404Spjd * If this record won't fit in the current log block, start a new one. 1075168404Spjd */ 1076219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1077168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1078168404Spjd if (lwb == NULL) 1079168404Spjd return (NULL); 1080168404Spjd zil_lwb_write_init(zilog, lwb); 1081219089Spjd ASSERT(LWB_EMPTY(lwb)); 1082219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1083168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1084168404Spjd return (lwb); 1085168404Spjd } 1086168404Spjd } 1087168404Spjd 1088219089Spjd lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1089219089Spjd bcopy(lrc, lr_buf, reclen); 1090219089Spjd lrc = (lr_t *)lr_buf; 1091219089Spjd lrw = (lr_write_t *)lrc; 1092168404Spjd 1093168404Spjd /* 1094168404Spjd * If it's a write, fetch the data or get its blkptr as appropriate. 1095168404Spjd */ 1096168404Spjd if (lrc->lrc_txtype == TX_WRITE) { 1097168404Spjd if (txg > spa_freeze_txg(zilog->zl_spa)) 1098168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1099168404Spjd if (itx->itx_wr_state != WR_COPIED) { 1100168404Spjd char *dbuf; 1101168404Spjd int error; 1102168404Spjd 1103168404Spjd if (dlen) { 1104168404Spjd ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1105219089Spjd dbuf = lr_buf + reclen; 1106219089Spjd lrw->lr_common.lrc_reclen += dlen; 1107168404Spjd } else { 1108168404Spjd ASSERT(itx->itx_wr_state == WR_INDIRECT); 1109168404Spjd dbuf = NULL; 1110168404Spjd } 1111168404Spjd error = zilog->zl_get_data( 1112219089Spjd itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1113214378Smm if (error == EIO) { 1114214378Smm txg_wait_synced(zilog->zl_dmu_pool, txg); 1115214378Smm return (lwb); 1116214378Smm } 1117248571Smm if (error != 0) { 1118168404Spjd ASSERT(error == ENOENT || error == EEXIST || 1119168404Spjd error == EALREADY); 1120168404Spjd return (lwb); 1121168404Spjd } 1122168404Spjd } 1123168404Spjd } 1124168404Spjd 1125219089Spjd /* 1126219089Spjd * We're actually making an entry, so update lrc_seq to be the 1127219089Spjd * log record sequence number. Note that this is generally not 1128219089Spjd * equal to the itx sequence number because not all transactions 1129219089Spjd * are synchronous, and sometimes spa_sync() gets there first. 1130219089Spjd */ 1131219089Spjd lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1132168404Spjd lwb->lwb_nused += reclen + dlen; 1133168404Spjd lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1134219089Spjd ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1135240415Smm ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1136168404Spjd 1137168404Spjd return (lwb); 1138168404Spjd} 1139168404Spjd 1140168404Spjditx_t * 1141185029Spjdzil_itx_create(uint64_t txtype, size_t lrsize) 1142168404Spjd{ 1143168404Spjd itx_t *itx; 1144168404Spjd 1145168404Spjd lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1146168404Spjd 1147168404Spjd itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1148168404Spjd itx->itx_lr.lrc_txtype = txtype; 1149168404Spjd itx->itx_lr.lrc_reclen = lrsize; 1150185029Spjd itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1151168404Spjd itx->itx_lr.lrc_seq = 0; /* defensive */ 1152219089Spjd itx->itx_sync = B_TRUE; /* default is synchronous */ 1153168404Spjd 1154168404Spjd return (itx); 1155168404Spjd} 1156168404Spjd 1157219089Spjdvoid 1158219089Spjdzil_itx_destroy(itx_t *itx) 1159168404Spjd{ 1160219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1161219089Spjd} 1162168404Spjd 1163219089Spjd/* 1164219089Spjd * Free up the sync and async itxs. The itxs_t has already been detached 1165219089Spjd * so no locks are needed. 1166219089Spjd */ 1167219089Spjdstatic void 1168219089Spjdzil_itxg_clean(itxs_t *itxs) 1169219089Spjd{ 1170219089Spjd itx_t *itx; 1171219089Spjd list_t *list; 1172219089Spjd avl_tree_t *t; 1173219089Spjd void *cookie; 1174219089Spjd itx_async_node_t *ian; 1175168404Spjd 1176219089Spjd list = &itxs->i_sync_list; 1177219089Spjd while ((itx = list_head(list)) != NULL) { 1178219089Spjd list_remove(list, itx); 1179219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1180219089Spjd itx->itx_lr.lrc_reclen); 1181219089Spjd } 1182168404Spjd 1183219089Spjd cookie = NULL; 1184219089Spjd t = &itxs->i_async_tree; 1185219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1186219089Spjd list = &ian->ia_list; 1187219089Spjd while ((itx = list_head(list)) != NULL) { 1188219089Spjd list_remove(list, itx); 1189219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1190219089Spjd itx->itx_lr.lrc_reclen); 1191219089Spjd } 1192219089Spjd list_destroy(list); 1193219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1194219089Spjd } 1195219089Spjd avl_destroy(t); 1196219089Spjd 1197219089Spjd kmem_free(itxs, sizeof (itxs_t)); 1198168404Spjd} 1199168404Spjd 1200219089Spjdstatic int 1201219089Spjdzil_aitx_compare(const void *x1, const void *x2) 1202219089Spjd{ 1203219089Spjd const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1204219089Spjd const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1205219089Spjd 1206219089Spjd if (o1 < o2) 1207219089Spjd return (-1); 1208219089Spjd if (o1 > o2) 1209219089Spjd return (1); 1210219089Spjd 1211219089Spjd return (0); 1212219089Spjd} 1213219089Spjd 1214168404Spjd/* 1215219089Spjd * Remove all async itx with the given oid. 1216168404Spjd */ 1217168404Spjdstatic void 1218219089Spjdzil_remove_async(zilog_t *zilog, uint64_t oid) 1219168404Spjd{ 1220219089Spjd uint64_t otxg, txg; 1221219089Spjd itx_async_node_t *ian; 1222219089Spjd avl_tree_t *t; 1223219089Spjd avl_index_t where; 1224168404Spjd list_t clean_list; 1225168404Spjd itx_t *itx; 1226168404Spjd 1227219089Spjd ASSERT(oid != 0); 1228168404Spjd list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1229168404Spjd 1230219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1231219089Spjd otxg = ZILTEST_TXG; 1232219089Spjd else 1233219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1234219089Spjd 1235219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1236219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1237219089Spjd 1238219089Spjd mutex_enter(&itxg->itxg_lock); 1239219089Spjd if (itxg->itxg_txg != txg) { 1240219089Spjd mutex_exit(&itxg->itxg_lock); 1241219089Spjd continue; 1242219089Spjd } 1243219089Spjd 1244219089Spjd /* 1245219089Spjd * Locate the object node and append its list. 1246219089Spjd */ 1247219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1248219089Spjd ian = avl_find(t, &oid, &where); 1249219089Spjd if (ian != NULL) 1250219089Spjd list_move_tail(&clean_list, &ian->ia_list); 1251219089Spjd mutex_exit(&itxg->itxg_lock); 1252168404Spjd } 1253219089Spjd while ((itx = list_head(&clean_list)) != NULL) { 1254219089Spjd list_remove(&clean_list, itx); 1255219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1256219089Spjd itx->itx_lr.lrc_reclen); 1257219089Spjd } 1258219089Spjd list_destroy(&clean_list); 1259219089Spjd} 1260168404Spjd 1261219089Spjdvoid 1262219089Spjdzil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1263219089Spjd{ 1264219089Spjd uint64_t txg; 1265219089Spjd itxg_t *itxg; 1266219089Spjd itxs_t *itxs, *clean = NULL; 1267219089Spjd 1268168404Spjd /* 1269219089Spjd * Object ids can be re-instantiated in the next txg so 1270219089Spjd * remove any async transactions to avoid future leaks. 1271219089Spjd * This can happen if a fsync occurs on the re-instantiated 1272219089Spjd * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1273219089Spjd * the new file data and flushes a write record for the old object. 1274168404Spjd */ 1275219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1276219089Spjd zil_remove_async(zilog, itx->itx_oid); 1277219089Spjd 1278219089Spjd /* 1279219089Spjd * Ensure the data of a renamed file is committed before the rename. 1280219089Spjd */ 1281219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1282219089Spjd zil_async_to_sync(zilog, itx->itx_oid); 1283219089Spjd 1284239620Smm if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1285219089Spjd txg = ZILTEST_TXG; 1286219089Spjd else 1287219089Spjd txg = dmu_tx_get_txg(tx); 1288219089Spjd 1289219089Spjd itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1290219089Spjd mutex_enter(&itxg->itxg_lock); 1291219089Spjd itxs = itxg->itxg_itxs; 1292219089Spjd if (itxg->itxg_txg != txg) { 1293219089Spjd if (itxs != NULL) { 1294219089Spjd /* 1295219089Spjd * The zil_clean callback hasn't got around to cleaning 1296219089Spjd * this itxg. Save the itxs for release below. 1297219089Spjd * This should be rare. 1298219089Spjd */ 1299219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1300219089Spjd itxg->itxg_sod = 0; 1301219089Spjd clean = itxg->itxg_itxs; 1302219089Spjd } 1303219089Spjd ASSERT(itxg->itxg_sod == 0); 1304219089Spjd itxg->itxg_txg = txg; 1305219089Spjd itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1306219089Spjd 1307219089Spjd list_create(&itxs->i_sync_list, sizeof (itx_t), 1308219089Spjd offsetof(itx_t, itx_node)); 1309219089Spjd avl_create(&itxs->i_async_tree, zil_aitx_compare, 1310219089Spjd sizeof (itx_async_node_t), 1311219089Spjd offsetof(itx_async_node_t, ia_node)); 1312168404Spjd } 1313219089Spjd if (itx->itx_sync) { 1314219089Spjd list_insert_tail(&itxs->i_sync_list, itx); 1315219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1316219089Spjd itxg->itxg_sod += itx->itx_sod; 1317219089Spjd } else { 1318219089Spjd avl_tree_t *t = &itxs->i_async_tree; 1319219089Spjd uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1320219089Spjd itx_async_node_t *ian; 1321219089Spjd avl_index_t where; 1322168404Spjd 1323219089Spjd ian = avl_find(t, &foid, &where); 1324219089Spjd if (ian == NULL) { 1325219089Spjd ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1326219089Spjd list_create(&ian->ia_list, sizeof (itx_t), 1327219089Spjd offsetof(itx_t, itx_node)); 1328219089Spjd ian->ia_foid = foid; 1329219089Spjd avl_insert(t, ian, where); 1330219089Spjd } 1331219089Spjd list_insert_tail(&ian->ia_list, itx); 1332168404Spjd } 1333219089Spjd 1334219089Spjd itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1335239620Smm zilog_dirty(zilog, txg); 1336219089Spjd mutex_exit(&itxg->itxg_lock); 1337219089Spjd 1338219089Spjd /* Release the old itxs now we've dropped the lock */ 1339219089Spjd if (clean != NULL) 1340219089Spjd zil_itxg_clean(clean); 1341168404Spjd} 1342168404Spjd 1343168404Spjd/* 1344168404Spjd * If there are any in-memory intent log transactions which have now been 1345239620Smm * synced then start up a taskq to free them. We should only do this after we 1346239620Smm * have written out the uberblocks (i.e. txg has been comitted) so that 1347239620Smm * don't inadvertently clean out in-memory log records that would be required 1348239620Smm * by zil_commit(). 1349168404Spjd */ 1350168404Spjdvoid 1351219089Spjdzil_clean(zilog_t *zilog, uint64_t synced_txg) 1352168404Spjd{ 1353219089Spjd itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1354219089Spjd itxs_t *clean_me; 1355168404Spjd 1356219089Spjd mutex_enter(&itxg->itxg_lock); 1357219089Spjd if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1358219089Spjd mutex_exit(&itxg->itxg_lock); 1359219089Spjd return; 1360168404Spjd } 1361219089Spjd ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1362219089Spjd ASSERT(itxg->itxg_txg != 0); 1363219089Spjd ASSERT(zilog->zl_clean_taskq != NULL); 1364219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1365219089Spjd itxg->itxg_sod = 0; 1366219089Spjd clean_me = itxg->itxg_itxs; 1367219089Spjd itxg->itxg_itxs = NULL; 1368219089Spjd itxg->itxg_txg = 0; 1369219089Spjd mutex_exit(&itxg->itxg_lock); 1370219089Spjd /* 1371219089Spjd * Preferably start a task queue to free up the old itxs but 1372219089Spjd * if taskq_dispatch can't allocate resources to do that then 1373219089Spjd * free it in-line. This should be rare. Note, using TQ_SLEEP 1374219089Spjd * created a bad performance problem. 1375219089Spjd */ 1376219089Spjd if (taskq_dispatch(zilog->zl_clean_taskq, 1377219089Spjd (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1378219089Spjd zil_itxg_clean(clean_me); 1379168404Spjd} 1380168404Spjd 1381219089Spjd/* 1382219089Spjd * Get the list of itxs to commit into zl_itx_commit_list. 1383219089Spjd */ 1384185029Spjdstatic void 1385219089Spjdzil_get_commit_list(zilog_t *zilog) 1386168404Spjd{ 1387219089Spjd uint64_t otxg, txg; 1388219089Spjd list_t *commit_list = &zilog->zl_itx_commit_list; 1389219089Spjd uint64_t push_sod = 0; 1390219089Spjd 1391219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1392219089Spjd otxg = ZILTEST_TXG; 1393219089Spjd else 1394219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1395219089Spjd 1396219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1397219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1398219089Spjd 1399219089Spjd mutex_enter(&itxg->itxg_lock); 1400219089Spjd if (itxg->itxg_txg != txg) { 1401219089Spjd mutex_exit(&itxg->itxg_lock); 1402219089Spjd continue; 1403219089Spjd } 1404219089Spjd 1405219089Spjd list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1406219089Spjd push_sod += itxg->itxg_sod; 1407219089Spjd itxg->itxg_sod = 0; 1408219089Spjd 1409219089Spjd mutex_exit(&itxg->itxg_lock); 1410219089Spjd } 1411219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1412219089Spjd} 1413219089Spjd 1414219089Spjd/* 1415219089Spjd * Move the async itxs for a specified object to commit into sync lists. 1416219089Spjd */ 1417219089Spjdstatic void 1418219089Spjdzil_async_to_sync(zilog_t *zilog, uint64_t foid) 1419219089Spjd{ 1420219089Spjd uint64_t otxg, txg; 1421219089Spjd itx_async_node_t *ian; 1422219089Spjd avl_tree_t *t; 1423219089Spjd avl_index_t where; 1424219089Spjd 1425219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1426219089Spjd otxg = ZILTEST_TXG; 1427219089Spjd else 1428219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1429219089Spjd 1430219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1431219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1432219089Spjd 1433219089Spjd mutex_enter(&itxg->itxg_lock); 1434219089Spjd if (itxg->itxg_txg != txg) { 1435219089Spjd mutex_exit(&itxg->itxg_lock); 1436219089Spjd continue; 1437219089Spjd } 1438219089Spjd 1439219089Spjd /* 1440219089Spjd * If a foid is specified then find that node and append its 1441219089Spjd * list. Otherwise walk the tree appending all the lists 1442219089Spjd * to the sync list. We add to the end rather than the 1443219089Spjd * beginning to ensure the create has happened. 1444219089Spjd */ 1445219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1446219089Spjd if (foid != 0) { 1447219089Spjd ian = avl_find(t, &foid, &where); 1448219089Spjd if (ian != NULL) { 1449219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1450219089Spjd &ian->ia_list); 1451219089Spjd } 1452219089Spjd } else { 1453219089Spjd void *cookie = NULL; 1454219089Spjd 1455219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1456219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1457219089Spjd &ian->ia_list); 1458219089Spjd list_destroy(&ian->ia_list); 1459219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1460219089Spjd } 1461219089Spjd } 1462219089Spjd mutex_exit(&itxg->itxg_lock); 1463219089Spjd } 1464219089Spjd} 1465219089Spjd 1466219089Spjdstatic void 1467219089Spjdzil_commit_writer(zilog_t *zilog) 1468219089Spjd{ 1469168404Spjd uint64_t txg; 1470219089Spjd itx_t *itx; 1471168404Spjd lwb_t *lwb; 1472219089Spjd spa_t *spa = zilog->zl_spa; 1473219089Spjd int error = 0; 1474168404Spjd 1475185029Spjd ASSERT(zilog->zl_root_zio == NULL); 1476168404Spjd 1477219089Spjd mutex_exit(&zilog->zl_lock); 1478219089Spjd 1479219089Spjd zil_get_commit_list(zilog); 1480219089Spjd 1481219089Spjd /* 1482219089Spjd * Return if there's nothing to commit before we dirty the fs by 1483219089Spjd * calling zil_create(). 1484219089Spjd */ 1485219089Spjd if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1486219089Spjd mutex_enter(&zilog->zl_lock); 1487219089Spjd return; 1488219089Spjd } 1489219089Spjd 1490168404Spjd if (zilog->zl_suspend) { 1491168404Spjd lwb = NULL; 1492168404Spjd } else { 1493168404Spjd lwb = list_tail(&zilog->zl_lwb_list); 1494219089Spjd if (lwb == NULL) 1495219089Spjd lwb = zil_create(zilog); 1496168404Spjd } 1497168404Spjd 1498168404Spjd DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1499219089Spjd while (itx = list_head(&zilog->zl_itx_commit_list)) { 1500168404Spjd txg = itx->itx_lr.lrc_txg; 1501168404Spjd ASSERT(txg); 1502168404Spjd 1503219089Spjd if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1504168404Spjd lwb = zil_lwb_commit(zilog, itx, lwb); 1505219089Spjd list_remove(&zilog->zl_itx_commit_list, itx); 1506168404Spjd kmem_free(itx, offsetof(itx_t, itx_lr) 1507168404Spjd + itx->itx_lr.lrc_reclen); 1508168404Spjd } 1509168404Spjd DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1510168404Spjd 1511168404Spjd /* write the last block out */ 1512168404Spjd if (lwb != NULL && lwb->lwb_zio != NULL) 1513168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1514168404Spjd 1515168404Spjd zilog->zl_cur_used = 0; 1516168404Spjd 1517168404Spjd /* 1518168404Spjd * Wait if necessary for the log blocks to be on stable storage. 1519168404Spjd */ 1520168404Spjd if (zilog->zl_root_zio) { 1521219089Spjd error = zio_wait(zilog->zl_root_zio); 1522185029Spjd zilog->zl_root_zio = NULL; 1523185029Spjd zil_flush_vdevs(zilog); 1524168404Spjd } 1525168404Spjd 1526219089Spjd if (error || lwb == NULL) 1527168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 1528168404Spjd 1529168404Spjd mutex_enter(&zilog->zl_lock); 1530168404Spjd 1531219089Spjd /* 1532219089Spjd * Remember the highest committed log sequence number for ztest. 1533219089Spjd * We only update this value when all the log writes succeeded, 1534219089Spjd * because ztest wants to ASSERT that it got the whole log chain. 1535219089Spjd */ 1536219089Spjd if (error == 0 && lwb != NULL) 1537219089Spjd zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1538168404Spjd} 1539168404Spjd 1540168404Spjd/* 1541219089Spjd * Commit zfs transactions to stable storage. 1542168404Spjd * If foid is 0 push out all transactions, otherwise push only those 1543219089Spjd * for that object or might reference that object. 1544219089Spjd * 1545219089Spjd * itxs are committed in batches. In a heavily stressed zil there will be 1546219089Spjd * a commit writer thread who is writing out a bunch of itxs to the log 1547219089Spjd * for a set of committing threads (cthreads) in the same batch as the writer. 1548219089Spjd * Those cthreads are all waiting on the same cv for that batch. 1549219089Spjd * 1550219089Spjd * There will also be a different and growing batch of threads that are 1551219089Spjd * waiting to commit (qthreads). When the committing batch completes 1552219089Spjd * a transition occurs such that the cthreads exit and the qthreads become 1553219089Spjd * cthreads. One of the new cthreads becomes the writer thread for the 1554219089Spjd * batch. Any new threads arriving become new qthreads. 1555219089Spjd * 1556219089Spjd * Only 2 condition variables are needed and there's no transition 1557219089Spjd * between the two cvs needed. They just flip-flop between qthreads 1558219089Spjd * and cthreads. 1559219089Spjd * 1560219089Spjd * Using this scheme we can efficiently wakeup up only those threads 1561219089Spjd * that have been committed. 1562168404Spjd */ 1563168404Spjdvoid 1564219089Spjdzil_commit(zilog_t *zilog, uint64_t foid) 1565168404Spjd{ 1566219089Spjd uint64_t mybatch; 1567219089Spjd 1568219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1569168404Spjd return; 1570168404Spjd 1571219089Spjd /* move the async itxs for the foid to the sync queues */ 1572219089Spjd zil_async_to_sync(zilog, foid); 1573219089Spjd 1574168404Spjd mutex_enter(&zilog->zl_lock); 1575219089Spjd mybatch = zilog->zl_next_batch; 1576168404Spjd while (zilog->zl_writer) { 1577219089Spjd cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1578219089Spjd if (mybatch <= zilog->zl_com_batch) { 1579168404Spjd mutex_exit(&zilog->zl_lock); 1580168404Spjd return; 1581168404Spjd } 1582168404Spjd } 1583219089Spjd 1584219089Spjd zilog->zl_next_batch++; 1585219089Spjd zilog->zl_writer = B_TRUE; 1586219089Spjd zil_commit_writer(zilog); 1587219089Spjd zilog->zl_com_batch = mybatch; 1588219089Spjd zilog->zl_writer = B_FALSE; 1589168404Spjd mutex_exit(&zilog->zl_lock); 1590219089Spjd 1591219089Spjd /* wake up one thread to become the next writer */ 1592219089Spjd cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1593219089Spjd 1594219089Spjd /* wake up all threads waiting for this batch to be committed */ 1595219089Spjd cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1596168404Spjd} 1597168404Spjd 1598168404Spjd/* 1599168404Spjd * Called in syncing context to free committed log blocks and update log header. 1600168404Spjd */ 1601168404Spjdvoid 1602168404Spjdzil_sync(zilog_t *zilog, dmu_tx_t *tx) 1603168404Spjd{ 1604168404Spjd zil_header_t *zh = zil_header_in_syncing_context(zilog); 1605168404Spjd uint64_t txg = dmu_tx_get_txg(tx); 1606168404Spjd spa_t *spa = zilog->zl_spa; 1607219089Spjd uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1608168404Spjd lwb_t *lwb; 1609168404Spjd 1610209962Smm /* 1611209962Smm * We don't zero out zl_destroy_txg, so make sure we don't try 1612209962Smm * to destroy it twice. 1613209962Smm */ 1614209962Smm if (spa_sync_pass(spa) != 1) 1615209962Smm return; 1616209962Smm 1617168404Spjd mutex_enter(&zilog->zl_lock); 1618168404Spjd 1619168404Spjd ASSERT(zilog->zl_stop_sync == 0); 1620168404Spjd 1621219089Spjd if (*replayed_seq != 0) { 1622219089Spjd ASSERT(zh->zh_replay_seq < *replayed_seq); 1623219089Spjd zh->zh_replay_seq = *replayed_seq; 1624219089Spjd *replayed_seq = 0; 1625219089Spjd } 1626168404Spjd 1627168404Spjd if (zilog->zl_destroy_txg == txg) { 1628168404Spjd blkptr_t blk = zh->zh_log; 1629168404Spjd 1630168404Spjd ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1631168404Spjd 1632168404Spjd bzero(zh, sizeof (zil_header_t)); 1633209962Smm bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1634168404Spjd 1635168404Spjd if (zilog->zl_keep_first) { 1636168404Spjd /* 1637168404Spjd * If this block was part of log chain that couldn't 1638168404Spjd * be claimed because a device was missing during 1639168404Spjd * zil_claim(), but that device later returns, 1640168404Spjd * then this block could erroneously appear valid. 1641168404Spjd * To guard against this, assign a new GUID to the new 1642168404Spjd * log chain so it doesn't matter what blk points to. 1643168404Spjd */ 1644168404Spjd zil_init_log_chain(zilog, &blk); 1645168404Spjd zh->zh_log = blk; 1646168404Spjd } 1647168404Spjd } 1648168404Spjd 1649213197Smm while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1650168404Spjd zh->zh_log = lwb->lwb_blk; 1651168404Spjd if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1652168404Spjd break; 1653168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 1654219089Spjd zio_free_zil(spa, txg, &lwb->lwb_blk); 1655168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 1656168404Spjd 1657168404Spjd /* 1658168404Spjd * If we don't have anything left in the lwb list then 1659168404Spjd * we've had an allocation failure and we need to zero 1660168404Spjd * out the zil_header blkptr so that we don't end 1661168404Spjd * up freeing the same block twice. 1662168404Spjd */ 1663168404Spjd if (list_head(&zilog->zl_lwb_list) == NULL) 1664168404Spjd BP_ZERO(&zh->zh_log); 1665168404Spjd } 1666168404Spjd mutex_exit(&zilog->zl_lock); 1667168404Spjd} 1668168404Spjd 1669168404Spjdvoid 1670168404Spjdzil_init(void) 1671168404Spjd{ 1672168404Spjd zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1673168404Spjd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1674168404Spjd} 1675168404Spjd 1676168404Spjdvoid 1677168404Spjdzil_fini(void) 1678168404Spjd{ 1679168404Spjd kmem_cache_destroy(zil_lwb_cache); 1680168404Spjd} 1681168404Spjd 1682219089Spjdvoid 1683219089Spjdzil_set_sync(zilog_t *zilog, uint64_t sync) 1684219089Spjd{ 1685219089Spjd zilog->zl_sync = sync; 1686219089Spjd} 1687219089Spjd 1688219089Spjdvoid 1689219089Spjdzil_set_logbias(zilog_t *zilog, uint64_t logbias) 1690219089Spjd{ 1691219089Spjd zilog->zl_logbias = logbias; 1692219089Spjd} 1693219089Spjd 1694168404Spjdzilog_t * 1695168404Spjdzil_alloc(objset_t *os, zil_header_t *zh_phys) 1696168404Spjd{ 1697168404Spjd zilog_t *zilog; 1698168404Spjd 1699168404Spjd zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1700168404Spjd 1701168404Spjd zilog->zl_header = zh_phys; 1702168404Spjd zilog->zl_os = os; 1703168404Spjd zilog->zl_spa = dmu_objset_spa(os); 1704168404Spjd zilog->zl_dmu_pool = dmu_objset_pool(os); 1705168404Spjd zilog->zl_destroy_txg = TXG_INITIAL - 1; 1706219089Spjd zilog->zl_logbias = dmu_objset_logbias(os); 1707219089Spjd zilog->zl_sync = dmu_objset_syncprop(os); 1708219089Spjd zilog->zl_next_batch = 1; 1709168404Spjd 1710168404Spjd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1711168404Spjd 1712219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1713219089Spjd mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1714219089Spjd MUTEX_DEFAULT, NULL); 1715219089Spjd } 1716168404Spjd 1717168404Spjd list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1718168404Spjd offsetof(lwb_t, lwb_node)); 1719168404Spjd 1720219089Spjd list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1721219089Spjd offsetof(itx_t, itx_node)); 1722219089Spjd 1723185029Spjd mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1724168404Spjd 1725185029Spjd avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1726185029Spjd sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1727185029Spjd 1728185029Spjd cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1729185029Spjd cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1730219089Spjd cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1731219089Spjd cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1732185029Spjd 1733168404Spjd return (zilog); 1734168404Spjd} 1735168404Spjd 1736168404Spjdvoid 1737168404Spjdzil_free(zilog_t *zilog) 1738168404Spjd{ 1739168404Spjd zilog->zl_stop_sync = 1; 1740168404Spjd 1741248571Smm ASSERT0(zilog->zl_suspend); 1742248571Smm ASSERT0(zilog->zl_suspending); 1743248571Smm 1744224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1745168404Spjd list_destroy(&zilog->zl_lwb_list); 1746168404Spjd 1747185029Spjd avl_destroy(&zilog->zl_vdev_tree); 1748185029Spjd mutex_destroy(&zilog->zl_vdev_lock); 1749168404Spjd 1750219089Spjd ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1751219089Spjd list_destroy(&zilog->zl_itx_commit_list); 1752219089Spjd 1753219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1754219089Spjd /* 1755219089Spjd * It's possible for an itx to be generated that doesn't dirty 1756219089Spjd * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1757219089Spjd * callback to remove the entry. We remove those here. 1758219089Spjd * 1759219089Spjd * Also free up the ziltest itxs. 1760219089Spjd */ 1761219089Spjd if (zilog->zl_itxg[i].itxg_itxs) 1762219089Spjd zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1763219089Spjd mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1764219089Spjd } 1765219089Spjd 1766168404Spjd mutex_destroy(&zilog->zl_lock); 1767168404Spjd 1768185029Spjd cv_destroy(&zilog->zl_cv_writer); 1769185029Spjd cv_destroy(&zilog->zl_cv_suspend); 1770219089Spjd cv_destroy(&zilog->zl_cv_batch[0]); 1771219089Spjd cv_destroy(&zilog->zl_cv_batch[1]); 1772185029Spjd 1773168404Spjd kmem_free(zilog, sizeof (zilog_t)); 1774168404Spjd} 1775168404Spjd 1776168404Spjd/* 1777168404Spjd * Open an intent log. 1778168404Spjd */ 1779168404Spjdzilog_t * 1780168404Spjdzil_open(objset_t *os, zil_get_data_t *get_data) 1781168404Spjd{ 1782168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1783168404Spjd 1784224526Smm ASSERT(zilog->zl_clean_taskq == NULL); 1785224526Smm ASSERT(zilog->zl_get_data == NULL); 1786224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1787224526Smm 1788168404Spjd zilog->zl_get_data = get_data; 1789168404Spjd zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1790168404Spjd 2, 2, TASKQ_PREPOPULATE); 1791168404Spjd 1792168404Spjd return (zilog); 1793168404Spjd} 1794168404Spjd 1795168404Spjd/* 1796168404Spjd * Close an intent log. 1797168404Spjd */ 1798168404Spjdvoid 1799168404Spjdzil_close(zilog_t *zilog) 1800168404Spjd{ 1801224526Smm lwb_t *lwb; 1802219089Spjd uint64_t txg = 0; 1803219089Spjd 1804219089Spjd zil_commit(zilog, 0); /* commit all itx */ 1805219089Spjd 1806168404Spjd /* 1807219089Spjd * The lwb_max_txg for the stubby lwb will reflect the last activity 1808219089Spjd * for the zil. After a txg_wait_synced() on the txg we know all the 1809219089Spjd * callbacks have occurred that may clean the zil. Only then can we 1810219089Spjd * destroy the zl_clean_taskq. 1811168404Spjd */ 1812219089Spjd mutex_enter(&zilog->zl_lock); 1813224526Smm lwb = list_tail(&zilog->zl_lwb_list); 1814224526Smm if (lwb != NULL) 1815224526Smm txg = lwb->lwb_max_txg; 1816219089Spjd mutex_exit(&zilog->zl_lock); 1817219089Spjd if (txg) 1818168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1819239620Smm ASSERT(!zilog_is_dirty(zilog)); 1820168404Spjd 1821168404Spjd taskq_destroy(zilog->zl_clean_taskq); 1822168404Spjd zilog->zl_clean_taskq = NULL; 1823168404Spjd zilog->zl_get_data = NULL; 1824224526Smm 1825224526Smm /* 1826224526Smm * We should have only one LWB left on the list; remove it now. 1827224526Smm */ 1828224526Smm mutex_enter(&zilog->zl_lock); 1829224526Smm lwb = list_head(&zilog->zl_lwb_list); 1830224526Smm if (lwb != NULL) { 1831224526Smm ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1832224526Smm list_remove(&zilog->zl_lwb_list, lwb); 1833224526Smm zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1834224526Smm kmem_cache_free(zil_lwb_cache, lwb); 1835224526Smm } 1836224526Smm mutex_exit(&zilog->zl_lock); 1837168404Spjd} 1838168404Spjd 1839248571Smmstatic char *suspend_tag = "zil suspending"; 1840248571Smm 1841168404Spjd/* 1842168404Spjd * Suspend an intent log. While in suspended mode, we still honor 1843168404Spjd * synchronous semantics, but we rely on txg_wait_synced() to do it. 1844248571Smm * On old version pools, we suspend the log briefly when taking a 1845248571Smm * snapshot so that it will have an empty intent log. 1846248571Smm * 1847248571Smm * Long holds are not really intended to be used the way we do here -- 1848248571Smm * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1849248571Smm * could fail. Therefore we take pains to only put a long hold if it is 1850248571Smm * actually necessary. Fortunately, it will only be necessary if the 1851248571Smm * objset is currently mounted (or the ZVOL equivalent). In that case it 1852248571Smm * will already have a long hold, so we are not really making things any worse. 1853248571Smm * 1854248571Smm * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1855248571Smm * zvol_state_t), and use their mechanism to prevent their hold from being 1856248571Smm * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1857248571Smm * very little gain. 1858248571Smm * 1859248571Smm * if cookiep == NULL, this does both the suspend & resume. 1860248571Smm * Otherwise, it returns with the dataset "long held", and the cookie 1861248571Smm * should be passed into zil_resume(). 1862168404Spjd */ 1863168404Spjdint 1864248571Smmzil_suspend(const char *osname, void **cookiep) 1865168404Spjd{ 1866248571Smm objset_t *os; 1867248571Smm zilog_t *zilog; 1868248571Smm const zil_header_t *zh; 1869248571Smm int error; 1870168404Spjd 1871248571Smm error = dmu_objset_hold(osname, suspend_tag, &os); 1872248571Smm if (error != 0) 1873248571Smm return (error); 1874248571Smm zilog = dmu_objset_zil(os); 1875248571Smm 1876168404Spjd mutex_enter(&zilog->zl_lock); 1877248571Smm zh = zilog->zl_header; 1878248571Smm 1879200724Sdelphij if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1880168404Spjd mutex_exit(&zilog->zl_lock); 1881248571Smm dmu_objset_rele(os, suspend_tag); 1882249195Smm return (SET_ERROR(EBUSY)); 1883168404Spjd } 1884248571Smm 1885248571Smm /* 1886248571Smm * Don't put a long hold in the cases where we can avoid it. This 1887248571Smm * is when there is no cookie so we are doing a suspend & resume 1888248571Smm * (i.e. called from zil_vdev_offline()), and there's nothing to do 1889248571Smm * for the suspend because it's already suspended, or there's no ZIL. 1890248571Smm */ 1891248571Smm if (cookiep == NULL && !zilog->zl_suspending && 1892248571Smm (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1893248571Smm mutex_exit(&zilog->zl_lock); 1894248571Smm dmu_objset_rele(os, suspend_tag); 1895248571Smm return (0); 1896248571Smm } 1897248571Smm 1898248571Smm dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1899248571Smm dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1900248571Smm 1901248571Smm zilog->zl_suspend++; 1902248571Smm 1903248571Smm if (zilog->zl_suspend > 1) { 1904168404Spjd /* 1905248571Smm * Someone else is already suspending it. 1906168404Spjd * Just wait for them to finish. 1907168404Spjd */ 1908248571Smm 1909168404Spjd while (zilog->zl_suspending) 1910168404Spjd cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1911168404Spjd mutex_exit(&zilog->zl_lock); 1912248571Smm 1913248571Smm if (cookiep == NULL) 1914248571Smm zil_resume(os); 1915248571Smm else 1916248571Smm *cookiep = os; 1917168404Spjd return (0); 1918168404Spjd } 1919248571Smm 1920248571Smm /* 1921248571Smm * If there is no pointer to an on-disk block, this ZIL must not 1922248571Smm * be active (e.g. filesystem not mounted), so there's nothing 1923248571Smm * to clean up. 1924248571Smm */ 1925248571Smm if (BP_IS_HOLE(&zh->zh_log)) { 1926248571Smm ASSERT(cookiep != NULL); /* fast path already handled */ 1927248571Smm 1928248571Smm *cookiep = os; 1929248571Smm mutex_exit(&zilog->zl_lock); 1930248571Smm return (0); 1931248571Smm } 1932248571Smm 1933168404Spjd zilog->zl_suspending = B_TRUE; 1934168404Spjd mutex_exit(&zilog->zl_lock); 1935168404Spjd 1936219089Spjd zil_commit(zilog, 0); 1937168404Spjd 1938168404Spjd zil_destroy(zilog, B_FALSE); 1939168404Spjd 1940168404Spjd mutex_enter(&zilog->zl_lock); 1941168404Spjd zilog->zl_suspending = B_FALSE; 1942168404Spjd cv_broadcast(&zilog->zl_cv_suspend); 1943168404Spjd mutex_exit(&zilog->zl_lock); 1944168404Spjd 1945248571Smm if (cookiep == NULL) 1946248571Smm zil_resume(os); 1947248571Smm else 1948248571Smm *cookiep = os; 1949168404Spjd return (0); 1950168404Spjd} 1951168404Spjd 1952168404Spjdvoid 1953248571Smmzil_resume(void *cookie) 1954168404Spjd{ 1955248571Smm objset_t *os = cookie; 1956248571Smm zilog_t *zilog = dmu_objset_zil(os); 1957248571Smm 1958168404Spjd mutex_enter(&zilog->zl_lock); 1959168404Spjd ASSERT(zilog->zl_suspend != 0); 1960168404Spjd zilog->zl_suspend--; 1961168404Spjd mutex_exit(&zilog->zl_lock); 1962248571Smm dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1963248571Smm dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1964168404Spjd} 1965168404Spjd 1966219089Spjdtypedef struct zil_replay_arg { 1967219089Spjd zil_replay_func_t **zr_replay; 1968219089Spjd void *zr_arg; 1969219089Spjd boolean_t zr_byteswap; 1970219089Spjd char *zr_lr; 1971219089Spjd} zil_replay_arg_t; 1972219089Spjd 1973219089Spjdstatic int 1974219089Spjdzil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1975209962Smm{ 1976219089Spjd char name[MAXNAMELEN]; 1977209962Smm 1978219089Spjd zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1979209962Smm 1980219089Spjd dmu_objset_name(zilog->zl_os, name); 1981209962Smm 1982219089Spjd cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1983219089Spjd "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1984219089Spjd (u_longlong_t)lr->lrc_seq, 1985219089Spjd (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1986219089Spjd (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1987219089Spjd 1988219089Spjd return (error); 1989209962Smm} 1990209962Smm 1991219089Spjdstatic int 1992168404Spjdzil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1993168404Spjd{ 1994168404Spjd zil_replay_arg_t *zr = zra; 1995168404Spjd const zil_header_t *zh = zilog->zl_header; 1996168404Spjd uint64_t reclen = lr->lrc_reclen; 1997168404Spjd uint64_t txtype = lr->lrc_txtype; 1998219089Spjd int error = 0; 1999168404Spjd 2000219089Spjd zilog->zl_replaying_seq = lr->lrc_seq; 2001168404Spjd 2002219089Spjd if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2003219089Spjd return (0); 2004219089Spjd 2005168404Spjd if (lr->lrc_txg < claim_txg) /* already committed */ 2006219089Spjd return (0); 2007168404Spjd 2008185029Spjd /* Strip case-insensitive bit, still present in log record */ 2009185029Spjd txtype &= ~TX_CI; 2010185029Spjd 2011219089Spjd if (txtype == 0 || txtype >= TX_MAX_TYPE) 2012219089Spjd return (zil_replay_error(zilog, lr, EINVAL)); 2013219089Spjd 2014219089Spjd /* 2015219089Spjd * If this record type can be logged out of order, the object 2016219089Spjd * (lr_foid) may no longer exist. That's legitimate, not an error. 2017219089Spjd */ 2018219089Spjd if (TX_OOO(txtype)) { 2019219089Spjd error = dmu_object_info(zilog->zl_os, 2020219089Spjd ((lr_ooo_t *)lr)->lr_foid, NULL); 2021219089Spjd if (error == ENOENT || error == EEXIST) 2022219089Spjd return (0); 2023209962Smm } 2024209962Smm 2025168404Spjd /* 2026168404Spjd * Make a copy of the data so we can revise and extend it. 2027168404Spjd */ 2028219089Spjd bcopy(lr, zr->zr_lr, reclen); 2029168404Spjd 2030168404Spjd /* 2031219089Spjd * If this is a TX_WRITE with a blkptr, suck in the data. 2032219089Spjd */ 2033219089Spjd if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2034219089Spjd error = zil_read_log_data(zilog, (lr_write_t *)lr, 2035219089Spjd zr->zr_lr + reclen); 2036248571Smm if (error != 0) 2037219089Spjd return (zil_replay_error(zilog, lr, error)); 2038219089Spjd } 2039219089Spjd 2040219089Spjd /* 2041168404Spjd * The log block containing this lr may have been byteswapped 2042168404Spjd * so that we can easily examine common fields like lrc_txtype. 2043219089Spjd * However, the log is a mix of different record types, and only the 2044168404Spjd * replay vectors know how to byteswap their records. Therefore, if 2045168404Spjd * the lr was byteswapped, undo it before invoking the replay vector. 2046168404Spjd */ 2047168404Spjd if (zr->zr_byteswap) 2048219089Spjd byteswap_uint64_array(zr->zr_lr, reclen); 2049168404Spjd 2050168404Spjd /* 2051168404Spjd * We must now do two things atomically: replay this log record, 2052209962Smm * and update the log header sequence number to reflect the fact that 2053209962Smm * we did so. At the end of each replay function the sequence number 2054209962Smm * is updated if we are in replay mode. 2055168404Spjd */ 2056219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2057248571Smm if (error != 0) { 2058168404Spjd /* 2059168404Spjd * The DMU's dnode layer doesn't see removes until the txg 2060168404Spjd * commits, so a subsequent claim can spuriously fail with 2061209962Smm * EEXIST. So if we receive any error we try syncing out 2062219089Spjd * any removes then retry the transaction. Note that we 2063219089Spjd * specify B_FALSE for byteswap now, so we don't do it twice. 2064168404Spjd */ 2065219089Spjd txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2066219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2067248571Smm if (error != 0) 2068219089Spjd return (zil_replay_error(zilog, lr, error)); 2069168404Spjd } 2070219089Spjd return (0); 2071168404Spjd} 2072168404Spjd 2073168404Spjd/* ARGSUSED */ 2074219089Spjdstatic int 2075168404Spjdzil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2076168404Spjd{ 2077168404Spjd zilog->zl_replay_blks++; 2078219089Spjd 2079219089Spjd return (0); 2080168404Spjd} 2081168404Spjd 2082168404Spjd/* 2083168404Spjd * If this dataset has a non-empty intent log, replay it and destroy it. 2084168404Spjd */ 2085168404Spjdvoid 2086209962Smmzil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2087168404Spjd{ 2088168404Spjd zilog_t *zilog = dmu_objset_zil(os); 2089168404Spjd const zil_header_t *zh = zilog->zl_header; 2090168404Spjd zil_replay_arg_t zr; 2091168404Spjd 2092200724Sdelphij if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2093168404Spjd zil_destroy(zilog, B_TRUE); 2094168404Spjd return; 2095168404Spjd } 2096168404Spjd //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 2097168404Spjd 2098168404Spjd zr.zr_replay = replay_func; 2099168404Spjd zr.zr_arg = arg; 2100168404Spjd zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2101219089Spjd zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2102168404Spjd 2103168404Spjd /* 2104168404Spjd * Wait for in-progress removes to sync before starting replay. 2105168404Spjd */ 2106168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 2107168404Spjd 2108209962Smm zilog->zl_replay = B_TRUE; 2109219089Spjd zilog->zl_replay_time = ddi_get_lbolt(); 2110168404Spjd ASSERT(zilog->zl_replay_blks == 0); 2111168404Spjd (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2112168404Spjd zh->zh_claim_txg); 2113219089Spjd kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2114168404Spjd 2115168404Spjd zil_destroy(zilog, B_FALSE); 2116185029Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2117209962Smm zilog->zl_replay = B_FALSE; 2118168404Spjd //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2119168404Spjd} 2120168404Spjd 2121219089Spjdboolean_t 2122219089Spjdzil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2123168404Spjd{ 2124219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2125219089Spjd return (B_TRUE); 2126168404Spjd 2127219089Spjd if (zilog->zl_replay) { 2128219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2129219089Spjd zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2130219089Spjd zilog->zl_replaying_seq; 2131219089Spjd return (B_TRUE); 2132168404Spjd } 2133168404Spjd 2134219089Spjd return (B_FALSE); 2135168404Spjd} 2136213197Smm 2137213197Smm/* ARGSUSED */ 2138213197Smmint 2139219089Spjdzil_vdev_offline(const char *osname, void *arg) 2140213197Smm{ 2141213197Smm int error; 2142213197Smm 2143248571Smm error = zil_suspend(osname, NULL); 2144248571Smm if (error != 0) 2145249195Smm return (SET_ERROR(EEXIST)); 2146248571Smm return (0); 2147213197Smm} 2148