zil.c revision 258632
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. 23249195Smm * Copyright (c) 2013 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; 152219089Spjd const dva_t *dva = BP_IDENTITY(bp); 153219089Spjd zil_bp_node_t *zn; 154168404Spjd avl_index_t where; 155168404Spjd 156168404Spjd if (avl_find(t, dva, &where) != NULL) 157249195Smm return (SET_ERROR(EEXIST)); 158168404Spjd 159219089Spjd zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 160168404Spjd zn->zn_dva = *dva; 161168404Spjd avl_insert(t, zn, where); 162168404Spjd 163168404Spjd return (0); 164168404Spjd} 165168404Spjd 166168404Spjdstatic zil_header_t * 167168404Spjdzil_header_in_syncing_context(zilog_t *zilog) 168168404Spjd{ 169168404Spjd return ((zil_header_t *)zilog->zl_header); 170168404Spjd} 171168404Spjd 172168404Spjdstatic void 173168404Spjdzil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 174168404Spjd{ 175168404Spjd zio_cksum_t *zc = &bp->blk_cksum; 176168404Spjd 177168404Spjd zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 178168404Spjd zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 179168404Spjd zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 180168404Spjd zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 181168404Spjd} 182168404Spjd 183168404Spjd/* 184219089Spjd * Read a log block and make sure it's valid. 185168404Spjd */ 186168404Spjdstatic int 187219089Spjdzil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 188219089Spjd char **end) 189168404Spjd{ 190219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 191219089Spjd uint32_t aflags = ARC_WAIT; 192219089Spjd arc_buf_t *abuf = NULL; 193168404Spjd zbookmark_t zb; 194168404Spjd int error; 195168404Spjd 196219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 197219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 198168404Spjd 199219089Spjd if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 200219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE; 201168404Spjd 202219089Spjd SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 203219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 204168404Spjd 205246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 206219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 207219089Spjd 208168404Spjd if (error == 0) { 209168404Spjd zio_cksum_t cksum = bp->blk_cksum; 210168404Spjd 211168404Spjd /* 212185029Spjd * Validate the checksummed log block. 213185029Spjd * 214168404Spjd * Sequence numbers should be... sequential. The checksum 215168404Spjd * verifier for the next block should be bp's checksum plus 1. 216185029Spjd * 217185029Spjd * Also check the log chain linkage and size used. 218168404Spjd */ 219168404Spjd cksum.zc_word[ZIL_ZC_SEQ]++; 220168404Spjd 221219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 222219089Spjd zil_chain_t *zilc = abuf->b_data; 223219089Spjd char *lr = (char *)(zilc + 1); 224219089Spjd uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 225219089Spjd 226219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 227219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 228249195Smm error = SET_ERROR(ECKSUM); 229219089Spjd } else { 230219089Spjd bcopy(lr, dst, len); 231219089Spjd *end = (char *)dst + len; 232219089Spjd *nbp = zilc->zc_next_blk; 233219089Spjd } 234219089Spjd } else { 235219089Spjd char *lr = abuf->b_data; 236219089Spjd uint64_t size = BP_GET_LSIZE(bp); 237219089Spjd zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 238219089Spjd 239219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 240219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 241219089Spjd (zilc->zc_nused > (size - sizeof (*zilc)))) { 242249195Smm error = SET_ERROR(ECKSUM); 243219089Spjd } else { 244219089Spjd bcopy(lr, dst, zilc->zc_nused); 245219089Spjd *end = (char *)dst + zilc->zc_nused; 246219089Spjd *nbp = zilc->zc_next_blk; 247219089Spjd } 248185029Spjd } 249168404Spjd 250248571Smm VERIFY(arc_buf_remove_ref(abuf, &abuf)); 251168404Spjd } 252168404Spjd 253219089Spjd return (error); 254219089Spjd} 255168404Spjd 256219089Spjd/* 257219089Spjd * Read a TX_WRITE log data block. 258219089Spjd */ 259219089Spjdstatic int 260219089Spjdzil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 261219089Spjd{ 262219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 263219089Spjd const blkptr_t *bp = &lr->lr_blkptr; 264219089Spjd uint32_t aflags = ARC_WAIT; 265219089Spjd arc_buf_t *abuf = NULL; 266219089Spjd zbookmark_t zb; 267219089Spjd int error; 268219089Spjd 269219089Spjd if (BP_IS_HOLE(bp)) { 270219089Spjd if (wbuf != NULL) 271219089Spjd bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 272219089Spjd return (0); 273219089Spjd } 274219089Spjd 275219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 276219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 277219089Spjd 278219089Spjd SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 279219089Spjd ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 280219089Spjd 281246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 282219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 283219089Spjd 284219089Spjd if (error == 0) { 285219089Spjd if (wbuf != NULL) 286219089Spjd bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 287219089Spjd (void) arc_buf_remove_ref(abuf, &abuf); 288219089Spjd } 289219089Spjd 290168404Spjd return (error); 291168404Spjd} 292168404Spjd 293168404Spjd/* 294168404Spjd * Parse the intent log, and call parse_func for each valid record within. 295168404Spjd */ 296219089Spjdint 297168404Spjdzil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 298168404Spjd zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 299168404Spjd{ 300168404Spjd const zil_header_t *zh = zilog->zl_header; 301219089Spjd boolean_t claimed = !!zh->zh_claim_txg; 302219089Spjd uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 303219089Spjd uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 304219089Spjd uint64_t max_blk_seq = 0; 305219089Spjd uint64_t max_lr_seq = 0; 306219089Spjd uint64_t blk_count = 0; 307219089Spjd uint64_t lr_count = 0; 308219089Spjd blkptr_t blk, next_blk; 309168404Spjd char *lrbuf, *lrp; 310219089Spjd int error = 0; 311168404Spjd 312219089Spjd /* 313219089Spjd * Old logs didn't record the maximum zh_claim_lr_seq. 314219089Spjd */ 315219089Spjd if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 316219089Spjd claim_lr_seq = UINT64_MAX; 317168404Spjd 318168404Spjd /* 319168404Spjd * Starting at the block pointed to by zh_log we read the log chain. 320168404Spjd * For each block in the chain we strongly check that block to 321168404Spjd * ensure its validity. We stop when an invalid block is found. 322168404Spjd * For each block pointer in the chain we call parse_blk_func(). 323168404Spjd * For each record in each valid block we call parse_lr_func(). 324168404Spjd * If the log has been claimed, stop if we encounter a sequence 325168404Spjd * number greater than the highest claimed sequence number. 326168404Spjd */ 327219089Spjd lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 328219089Spjd zil_bp_tree_init(zilog); 329168404Spjd 330219089Spjd for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 331219089Spjd uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 332219089Spjd int reclen; 333219089Spjd char *end; 334219089Spjd 335219089Spjd if (blk_seq > claim_blk_seq) 336168404Spjd break; 337219089Spjd if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 338219089Spjd break; 339219089Spjd ASSERT3U(max_blk_seq, <, blk_seq); 340219089Spjd max_blk_seq = blk_seq; 341219089Spjd blk_count++; 342168404Spjd 343219089Spjd if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 344219089Spjd break; 345168404Spjd 346219089Spjd error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 347248571Smm if (error != 0) 348168404Spjd break; 349168404Spjd 350219089Spjd for (lrp = lrbuf; lrp < end; lrp += reclen) { 351168404Spjd lr_t *lr = (lr_t *)lrp; 352168404Spjd reclen = lr->lrc_reclen; 353168404Spjd ASSERT3U(reclen, >=, sizeof (lr_t)); 354219089Spjd if (lr->lrc_seq > claim_lr_seq) 355219089Spjd goto done; 356219089Spjd if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 357219089Spjd goto done; 358219089Spjd ASSERT3U(max_lr_seq, <, lr->lrc_seq); 359219089Spjd max_lr_seq = lr->lrc_seq; 360219089Spjd lr_count++; 361168404Spjd } 362168404Spjd } 363219089Spjddone: 364219089Spjd zilog->zl_parse_error = error; 365219089Spjd zilog->zl_parse_blk_seq = max_blk_seq; 366219089Spjd zilog->zl_parse_lr_seq = max_lr_seq; 367219089Spjd zilog->zl_parse_blk_count = blk_count; 368219089Spjd zilog->zl_parse_lr_count = lr_count; 369168404Spjd 370219089Spjd ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 371219089Spjd (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 372219089Spjd 373219089Spjd zil_bp_tree_fini(zilog); 374219089Spjd zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 375219089Spjd 376219089Spjd return (error); 377168404Spjd} 378168404Spjd 379219089Spjdstatic int 380168404Spjdzil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 381168404Spjd{ 382168404Spjd /* 383168404Spjd * Claim log block if not already committed and not already claimed. 384219089Spjd * If tx == NULL, just verify that the block is claimable. 385168404Spjd */ 386219089Spjd if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 387219089Spjd return (0); 388219089Spjd 389219089Spjd return (zio_wait(zio_claim(NULL, zilog->zl_spa, 390219089Spjd tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 391219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 392168404Spjd} 393168404Spjd 394219089Spjdstatic int 395168404Spjdzil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 396168404Spjd{ 397219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 398219089Spjd int error; 399219089Spjd 400219089Spjd if (lrc->lrc_txtype != TX_WRITE) 401219089Spjd return (0); 402219089Spjd 403219089Spjd /* 404219089Spjd * If the block is not readable, don't claim it. This can happen 405219089Spjd * in normal operation when a log block is written to disk before 406219089Spjd * some of the dmu_sync() blocks it points to. In this case, the 407219089Spjd * transaction cannot have been committed to anyone (we would have 408219089Spjd * waited for all writes to be stable first), so it is semantically 409219089Spjd * correct to declare this the end of the log. 410219089Spjd */ 411219089Spjd if (lr->lr_blkptr.blk_birth >= first_txg && 412219089Spjd (error = zil_read_log_data(zilog, lr, NULL)) != 0) 413219089Spjd return (error); 414219089Spjd return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 415168404Spjd} 416168404Spjd 417168404Spjd/* ARGSUSED */ 418219089Spjdstatic int 419168404Spjdzil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 420168404Spjd{ 421219089Spjd zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 422219089Spjd 423219089Spjd return (0); 424168404Spjd} 425168404Spjd 426219089Spjdstatic int 427168404Spjdzil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 428168404Spjd{ 429219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 430219089Spjd blkptr_t *bp = &lr->lr_blkptr; 431219089Spjd 432168404Spjd /* 433168404Spjd * If we previously claimed it, we need to free it. 434168404Spjd */ 435219089Spjd if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 436219089Spjd bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 437219089Spjd zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 438219089Spjd 439219089Spjd return (0); 440219089Spjd} 441219089Spjd 442219089Spjdstatic lwb_t * 443219089Spjdzil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 444219089Spjd{ 445219089Spjd lwb_t *lwb; 446219089Spjd 447219089Spjd lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 448219089Spjd lwb->lwb_zilog = zilog; 449219089Spjd lwb->lwb_blk = *bp; 450219089Spjd lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 451219089Spjd lwb->lwb_max_txg = txg; 452219089Spjd lwb->lwb_zio = NULL; 453219089Spjd lwb->lwb_tx = NULL; 454219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 455219089Spjd lwb->lwb_nused = sizeof (zil_chain_t); 456219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp); 457219089Spjd } else { 458219089Spjd lwb->lwb_nused = 0; 459219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 460168404Spjd } 461219089Spjd 462219089Spjd mutex_enter(&zilog->zl_lock); 463219089Spjd list_insert_tail(&zilog->zl_lwb_list, lwb); 464219089Spjd mutex_exit(&zilog->zl_lock); 465219089Spjd 466219089Spjd return (lwb); 467168404Spjd} 468168404Spjd 469168404Spjd/* 470239620Smm * Called when we create in-memory log transactions so that we know 471239620Smm * to cleanup the itxs at the end of spa_sync(). 472239620Smm */ 473239620Smmvoid 474239620Smmzilog_dirty(zilog_t *zilog, uint64_t txg) 475239620Smm{ 476239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 477239620Smm dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 478239620Smm 479239620Smm if (dsl_dataset_is_snapshot(ds)) 480239620Smm panic("dirtying snapshot!"); 481239620Smm 482248571Smm if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 483239620Smm /* up the hold count until we can be written out */ 484239620Smm dmu_buf_add_ref(ds->ds_dbuf, zilog); 485239620Smm } 486239620Smm} 487239620Smm 488239620Smmboolean_t 489239620Smmzilog_is_dirty(zilog_t *zilog) 490239620Smm{ 491239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 492239620Smm 493239620Smm for (int t = 0; t < TXG_SIZE; t++) { 494239620Smm if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 495239620Smm return (B_TRUE); 496239620Smm } 497239620Smm return (B_FALSE); 498239620Smm} 499239620Smm 500239620Smm/* 501168404Spjd * Create an on-disk intent log. 502168404Spjd */ 503219089Spjdstatic lwb_t * 504168404Spjdzil_create(zilog_t *zilog) 505168404Spjd{ 506168404Spjd const zil_header_t *zh = zilog->zl_header; 507219089Spjd lwb_t *lwb = NULL; 508168404Spjd uint64_t txg = 0; 509168404Spjd dmu_tx_t *tx = NULL; 510168404Spjd blkptr_t blk; 511168404Spjd int error = 0; 512168404Spjd 513168404Spjd /* 514168404Spjd * Wait for any previous destroy to complete. 515168404Spjd */ 516168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 517168404Spjd 518168404Spjd ASSERT(zh->zh_claim_txg == 0); 519168404Spjd ASSERT(zh->zh_replay_seq == 0); 520168404Spjd 521168404Spjd blk = zh->zh_log; 522168404Spjd 523168404Spjd /* 524219089Spjd * Allocate an initial log block if: 525219089Spjd * - there isn't one already 526219089Spjd * - the existing block is the wrong endianess 527168404Spjd */ 528207908Smm if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 529168404Spjd tx = dmu_tx_create(zilog->zl_os); 530219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 531168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 532168404Spjd txg = dmu_tx_get_txg(tx); 533168404Spjd 534207908Smm if (!BP_IS_HOLE(&blk)) { 535219089Spjd zio_free_zil(zilog->zl_spa, txg, &blk); 536207908Smm BP_ZERO(&blk); 537207908Smm } 538207908Smm 539219089Spjd error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 540219089Spjd ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 541168404Spjd 542168404Spjd if (error == 0) 543168404Spjd zil_init_log_chain(zilog, &blk); 544168404Spjd } 545168404Spjd 546168404Spjd /* 547168404Spjd * Allocate a log write buffer (lwb) for the first log block. 548168404Spjd */ 549219089Spjd if (error == 0) 550219089Spjd lwb = zil_alloc_lwb(zilog, &blk, txg); 551168404Spjd 552168404Spjd /* 553168404Spjd * If we just allocated the first log block, commit our transaction 554168404Spjd * and wait for zil_sync() to stuff the block poiner into zh_log. 555168404Spjd * (zh is part of the MOS, so we cannot modify it in open context.) 556168404Spjd */ 557168404Spjd if (tx != NULL) { 558168404Spjd dmu_tx_commit(tx); 559168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 560168404Spjd } 561168404Spjd 562168404Spjd ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 563219089Spjd 564219089Spjd return (lwb); 565168404Spjd} 566168404Spjd 567168404Spjd/* 568168404Spjd * In one tx, free all log blocks and clear the log header. 569168404Spjd * If keep_first is set, then we're replaying a log with no content. 570168404Spjd * We want to keep the first block, however, so that the first 571168404Spjd * synchronous transaction doesn't require a txg_wait_synced() 572168404Spjd * in zil_create(). We don't need to txg_wait_synced() here either 573168404Spjd * when keep_first is set, because both zil_create() and zil_destroy() 574168404Spjd * will wait for any in-progress destroys to complete. 575168404Spjd */ 576168404Spjdvoid 577168404Spjdzil_destroy(zilog_t *zilog, boolean_t keep_first) 578168404Spjd{ 579168404Spjd const zil_header_t *zh = zilog->zl_header; 580168404Spjd lwb_t *lwb; 581168404Spjd dmu_tx_t *tx; 582168404Spjd uint64_t txg; 583168404Spjd 584168404Spjd /* 585168404Spjd * Wait for any previous destroy to complete. 586168404Spjd */ 587168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 588168404Spjd 589219089Spjd zilog->zl_old_header = *zh; /* debugging aid */ 590219089Spjd 591168404Spjd if (BP_IS_HOLE(&zh->zh_log)) 592168404Spjd return; 593168404Spjd 594168404Spjd tx = dmu_tx_create(zilog->zl_os); 595219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 596168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 597168404Spjd txg = dmu_tx_get_txg(tx); 598168404Spjd 599168404Spjd mutex_enter(&zilog->zl_lock); 600168404Spjd 601168404Spjd ASSERT3U(zilog->zl_destroy_txg, <, txg); 602168404Spjd zilog->zl_destroy_txg = txg; 603168404Spjd zilog->zl_keep_first = keep_first; 604168404Spjd 605168404Spjd if (!list_is_empty(&zilog->zl_lwb_list)) { 606168404Spjd ASSERT(zh->zh_claim_txg == 0); 607224526Smm VERIFY(!keep_first); 608168404Spjd while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 609168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 610168404Spjd if (lwb->lwb_buf != NULL) 611168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 612219089Spjd zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 613168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 614168404Spjd } 615219089Spjd } else if (!keep_first) { 616239620Smm zil_destroy_sync(zilog, tx); 617168404Spjd } 618168404Spjd mutex_exit(&zilog->zl_lock); 619168404Spjd 620168404Spjd dmu_tx_commit(tx); 621185029Spjd} 622168404Spjd 623239620Smmvoid 624239620Smmzil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 625239620Smm{ 626239620Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 627239620Smm (void) zil_parse(zilog, zil_free_log_block, 628239620Smm zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 629239620Smm} 630239620Smm 631168404Spjdint 632219089Spjdzil_claim(const char *osname, void *txarg) 633168404Spjd{ 634168404Spjd dmu_tx_t *tx = txarg; 635168404Spjd uint64_t first_txg = dmu_tx_get_txg(tx); 636168404Spjd zilog_t *zilog; 637168404Spjd zil_header_t *zh; 638168404Spjd objset_t *os; 639168404Spjd int error; 640168404Spjd 641248571Smm error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); 642248571Smm if (error != 0) { 643185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 644168404Spjd return (0); 645168404Spjd } 646168404Spjd 647168404Spjd zilog = dmu_objset_zil(os); 648168404Spjd zh = zil_header_in_syncing_context(zilog); 649168404Spjd 650219089Spjd if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 651213197Smm if (!BP_IS_HOLE(&zh->zh_log)) 652219089Spjd zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 653213197Smm BP_ZERO(&zh->zh_log); 654213197Smm dsl_dataset_dirty(dmu_objset_ds(os), tx); 655248571Smm dmu_objset_disown(os, FTAG); 656219089Spjd return (0); 657213197Smm } 658213197Smm 659168404Spjd /* 660168404Spjd * Claim all log blocks if we haven't already done so, and remember 661168404Spjd * the highest claimed sequence number. This ensures that if we can 662168404Spjd * read only part of the log now (e.g. due to a missing device), 663168404Spjd * but we can read the entire log later, we will not try to replay 664168404Spjd * or destroy beyond the last block we successfully claimed. 665168404Spjd */ 666168404Spjd ASSERT3U(zh->zh_claim_txg, <=, first_txg); 667168404Spjd if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 668219089Spjd (void) zil_parse(zilog, zil_claim_log_block, 669219089Spjd zil_claim_log_record, tx, first_txg); 670168404Spjd zh->zh_claim_txg = first_txg; 671219089Spjd zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 672219089Spjd zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 673219089Spjd if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 674219089Spjd zh->zh_flags |= ZIL_REPLAY_NEEDED; 675219089Spjd zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 676168404Spjd dsl_dataset_dirty(dmu_objset_ds(os), tx); 677168404Spjd } 678168404Spjd 679168404Spjd ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 680248571Smm dmu_objset_disown(os, FTAG); 681168404Spjd return (0); 682168404Spjd} 683168404Spjd 684185029Spjd/* 685185029Spjd * Check the log by walking the log chain. 686185029Spjd * Checksum errors are ok as they indicate the end of the chain. 687185029Spjd * Any other error (no device or read failure) returns an error. 688185029Spjd */ 689185029Spjdint 690219089Spjdzil_check_log_chain(const char *osname, void *tx) 691168404Spjd{ 692185029Spjd zilog_t *zilog; 693185029Spjd objset_t *os; 694219089Spjd blkptr_t *bp; 695185029Spjd int error; 696168404Spjd 697219089Spjd ASSERT(tx == NULL); 698219089Spjd 699219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 700248571Smm if (error != 0) { 701185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 702185029Spjd return (0); 703185029Spjd } 704168404Spjd 705185029Spjd zilog = dmu_objset_zil(os); 706219089Spjd bp = (blkptr_t *)&zilog->zl_header->zh_log; 707219089Spjd 708219089Spjd /* 709219089Spjd * Check the first block and determine if it's on a log device 710219089Spjd * which may have been removed or faulted prior to loading this 711219089Spjd * pool. If so, there's no point in checking the rest of the log 712219089Spjd * as its content should have already been synced to the pool. 713219089Spjd */ 714219089Spjd if (!BP_IS_HOLE(bp)) { 715219089Spjd vdev_t *vd; 716219089Spjd boolean_t valid = B_TRUE; 717219089Spjd 718219089Spjd spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 719219089Spjd vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 720219089Spjd if (vd->vdev_islog && vdev_is_dead(vd)) 721219089Spjd valid = vdev_log_state_valid(vd); 722219089Spjd spa_config_exit(os->os_spa, SCL_STATE, FTAG); 723219089Spjd 724219089Spjd if (!valid) { 725219089Spjd dmu_objset_rele(os, FTAG); 726219089Spjd return (0); 727219089Spjd } 728168404Spjd } 729185029Spjd 730219089Spjd /* 731219089Spjd * Because tx == NULL, zil_claim_log_block() will not actually claim 732219089Spjd * any blocks, but just determine whether it is possible to do so. 733219089Spjd * In addition to checking the log chain, zil_claim_log_block() 734219089Spjd * will invoke zio_claim() with a done func of spa_claim_notify(), 735219089Spjd * which will update spa_max_claim_txg. See spa_load() for details. 736219089Spjd */ 737219089Spjd error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 738219089Spjd zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 739219089Spjd 740219089Spjd dmu_objset_rele(os, FTAG); 741219089Spjd 742219089Spjd return ((error == ECKSUM || error == ENOENT) ? 0 : error); 743168404Spjd} 744168404Spjd 745185029Spjdstatic int 746185029Spjdzil_vdev_compare(const void *x1, const void *x2) 747185029Spjd{ 748219089Spjd const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 749219089Spjd const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 750185029Spjd 751185029Spjd if (v1 < v2) 752185029Spjd return (-1); 753185029Spjd if (v1 > v2) 754185029Spjd return (1); 755185029Spjd 756185029Spjd return (0); 757185029Spjd} 758185029Spjd 759168404Spjdvoid 760219089Spjdzil_add_block(zilog_t *zilog, const blkptr_t *bp) 761168404Spjd{ 762185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 763185029Spjd avl_index_t where; 764185029Spjd zil_vdev_node_t *zv, zvsearch; 765185029Spjd int ndvas = BP_GET_NDVAS(bp); 766185029Spjd int i; 767168404Spjd 768185029Spjd if (zfs_nocacheflush) 769185029Spjd return; 770168404Spjd 771185029Spjd ASSERT(zilog->zl_writer); 772168404Spjd 773185029Spjd /* 774185029Spjd * Even though we're zl_writer, we still need a lock because the 775185029Spjd * zl_get_data() callbacks may have dmu_sync() done callbacks 776185029Spjd * that will run concurrently. 777185029Spjd */ 778185029Spjd mutex_enter(&zilog->zl_vdev_lock); 779185029Spjd for (i = 0; i < ndvas; i++) { 780185029Spjd zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 781185029Spjd if (avl_find(t, &zvsearch, &where) == NULL) { 782185029Spjd zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 783185029Spjd zv->zv_vdev = zvsearch.zv_vdev; 784185029Spjd avl_insert(t, zv, where); 785185029Spjd } 786185029Spjd } 787185029Spjd mutex_exit(&zilog->zl_vdev_lock); 788168404Spjd} 789168404Spjd 790219089Spjdstatic void 791168404Spjdzil_flush_vdevs(zilog_t *zilog) 792168404Spjd{ 793168404Spjd spa_t *spa = zilog->zl_spa; 794185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 795185029Spjd void *cookie = NULL; 796185029Spjd zil_vdev_node_t *zv; 797185029Spjd zio_t *zio; 798168404Spjd 799168404Spjd ASSERT(zilog->zl_writer); 800168404Spjd 801185029Spjd /* 802185029Spjd * We don't need zl_vdev_lock here because we're the zl_writer, 803185029Spjd * and all zl_get_data() callbacks are done. 804185029Spjd */ 805185029Spjd if (avl_numnodes(t) == 0) 806185029Spjd return; 807185029Spjd 808185029Spjd spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 809185029Spjd 810185029Spjd zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 811185029Spjd 812185029Spjd while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 813185029Spjd vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 814185029Spjd if (vd != NULL) 815185029Spjd zio_flush(zio, vd); 816185029Spjd kmem_free(zv, sizeof (*zv)); 817168404Spjd } 818168404Spjd 819168404Spjd /* 820168404Spjd * Wait for all the flushes to complete. Not all devices actually 821168404Spjd * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 822168404Spjd */ 823185029Spjd (void) zio_wait(zio); 824185029Spjd 825185029Spjd spa_config_exit(spa, SCL_STATE, FTAG); 826168404Spjd} 827168404Spjd 828168404Spjd/* 829168404Spjd * Function called when a log block write completes 830168404Spjd */ 831168404Spjdstatic void 832168404Spjdzil_lwb_write_done(zio_t *zio) 833168404Spjd{ 834168404Spjd lwb_t *lwb = zio->io_private; 835168404Spjd zilog_t *zilog = lwb->lwb_zilog; 836219089Spjd dmu_tx_t *tx = lwb->lwb_tx; 837168404Spjd 838185029Spjd ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 839185029Spjd ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 840185029Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 841185029Spjd ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 842185029Spjd ASSERT(!BP_IS_GANG(zio->io_bp)); 843185029Spjd ASSERT(!BP_IS_HOLE(zio->io_bp)); 844185029Spjd ASSERT(zio->io_bp->blk_fill == 0); 845185029Spjd 846168404Spjd /* 847209962Smm * Ensure the lwb buffer pointer is cleared before releasing 848209962Smm * the txg. If we have had an allocation failure and 849209962Smm * the txg is waiting to sync then we want want zil_sync() 850209962Smm * to remove the lwb so that it's not picked up as the next new 851209962Smm * one in zil_commit_writer(). zil_sync() will only remove 852209962Smm * the lwb if lwb_buf is null. 853168404Spjd */ 854168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 855168404Spjd mutex_enter(&zilog->zl_lock); 856168404Spjd lwb->lwb_buf = NULL; 857219089Spjd lwb->lwb_tx = NULL; 858219089Spjd mutex_exit(&zilog->zl_lock); 859209962Smm 860209962Smm /* 861209962Smm * Now that we've written this log block, we have a stable pointer 862209962Smm * to the next block in the chain, so it's OK to let the txg in 863219089Spjd * which we allocated the next block sync. 864209962Smm */ 865219089Spjd dmu_tx_commit(tx); 866168404Spjd} 867168404Spjd 868168404Spjd/* 869168404Spjd * Initialize the io for a log block. 870168404Spjd */ 871168404Spjdstatic void 872168404Spjdzil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 873168404Spjd{ 874168404Spjd zbookmark_t zb; 875168404Spjd 876219089Spjd SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 877219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 878219089Spjd lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 879168404Spjd 880168404Spjd if (zilog->zl_root_zio == NULL) { 881168404Spjd zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 882168404Spjd ZIO_FLAG_CANFAIL); 883168404Spjd } 884168404Spjd if (lwb->lwb_zio == NULL) { 885168404Spjd lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 886219089Spjd 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 887258632Savg zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 888219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 889168404Spjd } 890168404Spjd} 891168404Spjd 892168404Spjd/* 893219089Spjd * Define a limited set of intent log block sizes. 894251631Sdelphij * 895219089Spjd * These must be a multiple of 4KB. Note only the amount used (again 896219089Spjd * aligned to 4KB) actually gets written. However, we can't always just 897219089Spjd * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 898219089Spjd */ 899219089Spjduint64_t zil_block_buckets[] = { 900219089Spjd 4096, /* non TX_WRITE */ 901219089Spjd 8192+4096, /* data base */ 902219089Spjd 32*1024 + 4096, /* NFS writes */ 903219089Spjd UINT64_MAX 904219089Spjd}; 905219089Spjd 906219089Spjd/* 907219089Spjd * Use the slog as long as the logbias is 'latency' and the current commit size 908219089Spjd * is less than the limit or the total list size is less than 2X the limit. 909219089Spjd * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 910219089Spjd */ 911219089Spjduint64_t zil_slog_limit = 1024 * 1024; 912219089Spjd#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 913219089Spjd (((zilog)->zl_cur_used < zil_slog_limit) || \ 914219089Spjd ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 915219089Spjd 916219089Spjd/* 917168404Spjd * Start a log block write and advance to the next log block. 918168404Spjd * Calls are serialized. 919168404Spjd */ 920168404Spjdstatic lwb_t * 921168404Spjdzil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 922168404Spjd{ 923219089Spjd lwb_t *nlwb = NULL; 924219089Spjd zil_chain_t *zilc; 925168404Spjd spa_t *spa = zilog->zl_spa; 926219089Spjd blkptr_t *bp; 927219089Spjd dmu_tx_t *tx; 928168404Spjd uint64_t txg; 929219089Spjd uint64_t zil_blksz, wsz; 930219089Spjd int i, error; 931168404Spjd 932219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 933219089Spjd zilc = (zil_chain_t *)lwb->lwb_buf; 934219089Spjd bp = &zilc->zc_next_blk; 935219089Spjd } else { 936219089Spjd zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 937219089Spjd bp = &zilc->zc_next_blk; 938219089Spjd } 939168404Spjd 940219089Spjd ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 941219089Spjd 942168404Spjd /* 943168404Spjd * Allocate the next block and save its address in this block 944168404Spjd * before writing it in order to establish the log chain. 945168404Spjd * Note that if the allocation of nlwb synced before we wrote 946168404Spjd * the block that points at it (lwb), we'd leak it if we crashed. 947219089Spjd * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 948219089Spjd * We dirty the dataset to ensure that zil_sync() will be called 949219089Spjd * to clean up in the event of allocation failure or I/O failure. 950168404Spjd */ 951219089Spjd tx = dmu_tx_create(zilog->zl_os); 952219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 953219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 954219089Spjd txg = dmu_tx_get_txg(tx); 955168404Spjd 956219089Spjd lwb->lwb_tx = tx; 957219089Spjd 958168404Spjd /* 959219089Spjd * Log blocks are pre-allocated. Here we select the size of the next 960219089Spjd * block, based on size used in the last block. 961219089Spjd * - first find the smallest bucket that will fit the block from a 962219089Spjd * limited set of block sizes. This is because it's faster to write 963219089Spjd * blocks allocated from the same metaslab as they are adjacent or 964219089Spjd * close. 965219089Spjd * - next find the maximum from the new suggested size and an array of 966219089Spjd * previous sizes. This lessens a picket fence effect of wrongly 967219089Spjd * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 968219089Spjd * requests. 969219089Spjd * 970219089Spjd * Note we only write what is used, but we can't just allocate 971219089Spjd * the maximum block size because we can exhaust the available 972219089Spjd * pool log space. 973168404Spjd */ 974219089Spjd zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 975219089Spjd for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 976219089Spjd continue; 977219089Spjd zil_blksz = zil_block_buckets[i]; 978219089Spjd if (zil_blksz == UINT64_MAX) 979219089Spjd zil_blksz = SPA_MAXBLOCKSIZE; 980219089Spjd zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 981219089Spjd for (i = 0; i < ZIL_PREV_BLKS; i++) 982219089Spjd zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 983219089Spjd zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 984168404Spjd 985168404Spjd BP_ZERO(bp); 986168404Spjd /* pass the old blkptr in order to spread log blocks across devs */ 987219089Spjd error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 988219089Spjd USE_SLOG(zilog)); 989248571Smm if (error == 0) { 990219089Spjd ASSERT3U(bp->blk_birth, ==, txg); 991219089Spjd bp->blk_cksum = lwb->lwb_blk.blk_cksum; 992219089Spjd bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 993168404Spjd 994168404Spjd /* 995219089Spjd * Allocate a new log write buffer (lwb). 996168404Spjd */ 997219089Spjd nlwb = zil_alloc_lwb(zilog, bp, txg); 998168404Spjd 999219089Spjd /* Record the block for later vdev flushing */ 1000219089Spjd zil_add_block(zilog, &lwb->lwb_blk); 1001168404Spjd } 1002168404Spjd 1003219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1004219089Spjd /* For Slim ZIL only write what is used. */ 1005219089Spjd wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1006219089Spjd ASSERT3U(wsz, <=, lwb->lwb_sz); 1007219089Spjd zio_shrink(lwb->lwb_zio, wsz); 1008168404Spjd 1009219089Spjd } else { 1010219089Spjd wsz = lwb->lwb_sz; 1011219089Spjd } 1012168404Spjd 1013219089Spjd zilc->zc_pad = 0; 1014219089Spjd zilc->zc_nused = lwb->lwb_nused; 1015219089Spjd zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1016168404Spjd 1017168404Spjd /* 1018219089Spjd * clear unused data for security 1019168404Spjd */ 1020219089Spjd bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1021168404Spjd 1022219089Spjd zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1023168404Spjd 1024168404Spjd /* 1025219089Spjd * If there was an allocation failure then nlwb will be null which 1026219089Spjd * forces a txg_wait_synced(). 1027168404Spjd */ 1028168404Spjd return (nlwb); 1029168404Spjd} 1030168404Spjd 1031168404Spjdstatic lwb_t * 1032168404Spjdzil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1033168404Spjd{ 1034168404Spjd lr_t *lrc = &itx->itx_lr; /* common log record */ 1035219089Spjd lr_write_t *lrw = (lr_write_t *)lrc; 1036219089Spjd char *lr_buf; 1037168404Spjd uint64_t txg = lrc->lrc_txg; 1038168404Spjd uint64_t reclen = lrc->lrc_reclen; 1039219089Spjd uint64_t dlen = 0; 1040168404Spjd 1041168404Spjd if (lwb == NULL) 1042168404Spjd return (NULL); 1043219089Spjd 1044168404Spjd ASSERT(lwb->lwb_buf != NULL); 1045239620Smm ASSERT(zilog_is_dirty(zilog) || 1046239620Smm spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1047168404Spjd 1048168404Spjd if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1049168404Spjd dlen = P2ROUNDUP_TYPED( 1050219089Spjd lrw->lr_length, sizeof (uint64_t), uint64_t); 1051168404Spjd 1052168404Spjd zilog->zl_cur_used += (reclen + dlen); 1053168404Spjd 1054168404Spjd zil_lwb_write_init(zilog, lwb); 1055168404Spjd 1056168404Spjd /* 1057168404Spjd * If this record won't fit in the current log block, start a new one. 1058168404Spjd */ 1059219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1060168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1061168404Spjd if (lwb == NULL) 1062168404Spjd return (NULL); 1063168404Spjd zil_lwb_write_init(zilog, lwb); 1064219089Spjd ASSERT(LWB_EMPTY(lwb)); 1065219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1066168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1067168404Spjd return (lwb); 1068168404Spjd } 1069168404Spjd } 1070168404Spjd 1071219089Spjd lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1072219089Spjd bcopy(lrc, lr_buf, reclen); 1073219089Spjd lrc = (lr_t *)lr_buf; 1074219089Spjd lrw = (lr_write_t *)lrc; 1075168404Spjd 1076168404Spjd /* 1077168404Spjd * If it's a write, fetch the data or get its blkptr as appropriate. 1078168404Spjd */ 1079168404Spjd if (lrc->lrc_txtype == TX_WRITE) { 1080168404Spjd if (txg > spa_freeze_txg(zilog->zl_spa)) 1081168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1082168404Spjd if (itx->itx_wr_state != WR_COPIED) { 1083168404Spjd char *dbuf; 1084168404Spjd int error; 1085168404Spjd 1086168404Spjd if (dlen) { 1087168404Spjd ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1088219089Spjd dbuf = lr_buf + reclen; 1089219089Spjd lrw->lr_common.lrc_reclen += dlen; 1090168404Spjd } else { 1091168404Spjd ASSERT(itx->itx_wr_state == WR_INDIRECT); 1092168404Spjd dbuf = NULL; 1093168404Spjd } 1094168404Spjd error = zilog->zl_get_data( 1095219089Spjd itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1096214378Smm if (error == EIO) { 1097214378Smm txg_wait_synced(zilog->zl_dmu_pool, txg); 1098214378Smm return (lwb); 1099214378Smm } 1100248571Smm if (error != 0) { 1101168404Spjd ASSERT(error == ENOENT || error == EEXIST || 1102168404Spjd error == EALREADY); 1103168404Spjd return (lwb); 1104168404Spjd } 1105168404Spjd } 1106168404Spjd } 1107168404Spjd 1108219089Spjd /* 1109219089Spjd * We're actually making an entry, so update lrc_seq to be the 1110219089Spjd * log record sequence number. Note that this is generally not 1111219089Spjd * equal to the itx sequence number because not all transactions 1112219089Spjd * are synchronous, and sometimes spa_sync() gets there first. 1113219089Spjd */ 1114219089Spjd lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1115168404Spjd lwb->lwb_nused += reclen + dlen; 1116168404Spjd lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1117219089Spjd ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1118240415Smm ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1119168404Spjd 1120168404Spjd return (lwb); 1121168404Spjd} 1122168404Spjd 1123168404Spjditx_t * 1124185029Spjdzil_itx_create(uint64_t txtype, size_t lrsize) 1125168404Spjd{ 1126168404Spjd itx_t *itx; 1127168404Spjd 1128168404Spjd lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1129168404Spjd 1130168404Spjd itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1131168404Spjd itx->itx_lr.lrc_txtype = txtype; 1132168404Spjd itx->itx_lr.lrc_reclen = lrsize; 1133185029Spjd itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1134168404Spjd itx->itx_lr.lrc_seq = 0; /* defensive */ 1135219089Spjd itx->itx_sync = B_TRUE; /* default is synchronous */ 1136168404Spjd 1137168404Spjd return (itx); 1138168404Spjd} 1139168404Spjd 1140219089Spjdvoid 1141219089Spjdzil_itx_destroy(itx_t *itx) 1142168404Spjd{ 1143219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1144219089Spjd} 1145168404Spjd 1146219089Spjd/* 1147219089Spjd * Free up the sync and async itxs. The itxs_t has already been detached 1148219089Spjd * so no locks are needed. 1149219089Spjd */ 1150219089Spjdstatic void 1151219089Spjdzil_itxg_clean(itxs_t *itxs) 1152219089Spjd{ 1153219089Spjd itx_t *itx; 1154219089Spjd list_t *list; 1155219089Spjd avl_tree_t *t; 1156219089Spjd void *cookie; 1157219089Spjd itx_async_node_t *ian; 1158168404Spjd 1159219089Spjd list = &itxs->i_sync_list; 1160219089Spjd while ((itx = list_head(list)) != NULL) { 1161219089Spjd list_remove(list, itx); 1162219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1163219089Spjd itx->itx_lr.lrc_reclen); 1164219089Spjd } 1165168404Spjd 1166219089Spjd cookie = NULL; 1167219089Spjd t = &itxs->i_async_tree; 1168219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1169219089Spjd list = &ian->ia_list; 1170219089Spjd while ((itx = list_head(list)) != NULL) { 1171219089Spjd list_remove(list, itx); 1172219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1173219089Spjd itx->itx_lr.lrc_reclen); 1174219089Spjd } 1175219089Spjd list_destroy(list); 1176219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1177219089Spjd } 1178219089Spjd avl_destroy(t); 1179219089Spjd 1180219089Spjd kmem_free(itxs, sizeof (itxs_t)); 1181168404Spjd} 1182168404Spjd 1183219089Spjdstatic int 1184219089Spjdzil_aitx_compare(const void *x1, const void *x2) 1185219089Spjd{ 1186219089Spjd const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1187219089Spjd const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1188219089Spjd 1189219089Spjd if (o1 < o2) 1190219089Spjd return (-1); 1191219089Spjd if (o1 > o2) 1192219089Spjd return (1); 1193219089Spjd 1194219089Spjd return (0); 1195219089Spjd} 1196219089Spjd 1197168404Spjd/* 1198219089Spjd * Remove all async itx with the given oid. 1199168404Spjd */ 1200168404Spjdstatic void 1201219089Spjdzil_remove_async(zilog_t *zilog, uint64_t oid) 1202168404Spjd{ 1203219089Spjd uint64_t otxg, txg; 1204219089Spjd itx_async_node_t *ian; 1205219089Spjd avl_tree_t *t; 1206219089Spjd avl_index_t where; 1207168404Spjd list_t clean_list; 1208168404Spjd itx_t *itx; 1209168404Spjd 1210219089Spjd ASSERT(oid != 0); 1211168404Spjd list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1212168404Spjd 1213219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1214219089Spjd otxg = ZILTEST_TXG; 1215219089Spjd else 1216219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1217219089Spjd 1218219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1219219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1220219089Spjd 1221219089Spjd mutex_enter(&itxg->itxg_lock); 1222219089Spjd if (itxg->itxg_txg != txg) { 1223219089Spjd mutex_exit(&itxg->itxg_lock); 1224219089Spjd continue; 1225219089Spjd } 1226219089Spjd 1227219089Spjd /* 1228219089Spjd * Locate the object node and append its list. 1229219089Spjd */ 1230219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1231219089Spjd ian = avl_find(t, &oid, &where); 1232219089Spjd if (ian != NULL) 1233219089Spjd list_move_tail(&clean_list, &ian->ia_list); 1234219089Spjd mutex_exit(&itxg->itxg_lock); 1235168404Spjd } 1236219089Spjd while ((itx = list_head(&clean_list)) != NULL) { 1237219089Spjd list_remove(&clean_list, itx); 1238219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1239219089Spjd itx->itx_lr.lrc_reclen); 1240219089Spjd } 1241219089Spjd list_destroy(&clean_list); 1242219089Spjd} 1243168404Spjd 1244219089Spjdvoid 1245219089Spjdzil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1246219089Spjd{ 1247219089Spjd uint64_t txg; 1248219089Spjd itxg_t *itxg; 1249219089Spjd itxs_t *itxs, *clean = NULL; 1250219089Spjd 1251168404Spjd /* 1252219089Spjd * Object ids can be re-instantiated in the next txg so 1253219089Spjd * remove any async transactions to avoid future leaks. 1254219089Spjd * This can happen if a fsync occurs on the re-instantiated 1255219089Spjd * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1256219089Spjd * the new file data and flushes a write record for the old object. 1257168404Spjd */ 1258219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1259219089Spjd zil_remove_async(zilog, itx->itx_oid); 1260219089Spjd 1261219089Spjd /* 1262219089Spjd * Ensure the data of a renamed file is committed before the rename. 1263219089Spjd */ 1264219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1265219089Spjd zil_async_to_sync(zilog, itx->itx_oid); 1266219089Spjd 1267239620Smm if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1268219089Spjd txg = ZILTEST_TXG; 1269219089Spjd else 1270219089Spjd txg = dmu_tx_get_txg(tx); 1271219089Spjd 1272219089Spjd itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1273219089Spjd mutex_enter(&itxg->itxg_lock); 1274219089Spjd itxs = itxg->itxg_itxs; 1275219089Spjd if (itxg->itxg_txg != txg) { 1276219089Spjd if (itxs != NULL) { 1277219089Spjd /* 1278219089Spjd * The zil_clean callback hasn't got around to cleaning 1279219089Spjd * this itxg. Save the itxs for release below. 1280219089Spjd * This should be rare. 1281219089Spjd */ 1282219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1283219089Spjd itxg->itxg_sod = 0; 1284219089Spjd clean = itxg->itxg_itxs; 1285219089Spjd } 1286219089Spjd ASSERT(itxg->itxg_sod == 0); 1287219089Spjd itxg->itxg_txg = txg; 1288219089Spjd itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1289219089Spjd 1290219089Spjd list_create(&itxs->i_sync_list, sizeof (itx_t), 1291219089Spjd offsetof(itx_t, itx_node)); 1292219089Spjd avl_create(&itxs->i_async_tree, zil_aitx_compare, 1293219089Spjd sizeof (itx_async_node_t), 1294219089Spjd offsetof(itx_async_node_t, ia_node)); 1295168404Spjd } 1296219089Spjd if (itx->itx_sync) { 1297219089Spjd list_insert_tail(&itxs->i_sync_list, itx); 1298219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1299219089Spjd itxg->itxg_sod += itx->itx_sod; 1300219089Spjd } else { 1301219089Spjd avl_tree_t *t = &itxs->i_async_tree; 1302219089Spjd uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1303219089Spjd itx_async_node_t *ian; 1304219089Spjd avl_index_t where; 1305168404Spjd 1306219089Spjd ian = avl_find(t, &foid, &where); 1307219089Spjd if (ian == NULL) { 1308219089Spjd ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1309219089Spjd list_create(&ian->ia_list, sizeof (itx_t), 1310219089Spjd offsetof(itx_t, itx_node)); 1311219089Spjd ian->ia_foid = foid; 1312219089Spjd avl_insert(t, ian, where); 1313219089Spjd } 1314219089Spjd list_insert_tail(&ian->ia_list, itx); 1315168404Spjd } 1316219089Spjd 1317219089Spjd itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1318239620Smm zilog_dirty(zilog, txg); 1319219089Spjd mutex_exit(&itxg->itxg_lock); 1320219089Spjd 1321219089Spjd /* Release the old itxs now we've dropped the lock */ 1322219089Spjd if (clean != NULL) 1323219089Spjd zil_itxg_clean(clean); 1324168404Spjd} 1325168404Spjd 1326168404Spjd/* 1327168404Spjd * If there are any in-memory intent log transactions which have now been 1328239620Smm * synced then start up a taskq to free them. We should only do this after we 1329239620Smm * have written out the uberblocks (i.e. txg has been comitted) so that 1330239620Smm * don't inadvertently clean out in-memory log records that would be required 1331239620Smm * by zil_commit(). 1332168404Spjd */ 1333168404Spjdvoid 1334219089Spjdzil_clean(zilog_t *zilog, uint64_t synced_txg) 1335168404Spjd{ 1336219089Spjd itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1337219089Spjd itxs_t *clean_me; 1338168404Spjd 1339219089Spjd mutex_enter(&itxg->itxg_lock); 1340219089Spjd if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1341219089Spjd mutex_exit(&itxg->itxg_lock); 1342219089Spjd return; 1343168404Spjd } 1344219089Spjd ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1345219089Spjd ASSERT(itxg->itxg_txg != 0); 1346219089Spjd ASSERT(zilog->zl_clean_taskq != NULL); 1347219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1348219089Spjd itxg->itxg_sod = 0; 1349219089Spjd clean_me = itxg->itxg_itxs; 1350219089Spjd itxg->itxg_itxs = NULL; 1351219089Spjd itxg->itxg_txg = 0; 1352219089Spjd mutex_exit(&itxg->itxg_lock); 1353219089Spjd /* 1354219089Spjd * Preferably start a task queue to free up the old itxs but 1355219089Spjd * if taskq_dispatch can't allocate resources to do that then 1356219089Spjd * free it in-line. This should be rare. Note, using TQ_SLEEP 1357219089Spjd * created a bad performance problem. 1358219089Spjd */ 1359219089Spjd if (taskq_dispatch(zilog->zl_clean_taskq, 1360219089Spjd (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1361219089Spjd zil_itxg_clean(clean_me); 1362168404Spjd} 1363168404Spjd 1364219089Spjd/* 1365219089Spjd * Get the list of itxs to commit into zl_itx_commit_list. 1366219089Spjd */ 1367185029Spjdstatic void 1368219089Spjdzil_get_commit_list(zilog_t *zilog) 1369168404Spjd{ 1370219089Spjd uint64_t otxg, txg; 1371219089Spjd list_t *commit_list = &zilog->zl_itx_commit_list; 1372219089Spjd uint64_t push_sod = 0; 1373219089Spjd 1374219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1375219089Spjd otxg = ZILTEST_TXG; 1376219089Spjd else 1377219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1378219089Spjd 1379219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1380219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1381219089Spjd 1382219089Spjd mutex_enter(&itxg->itxg_lock); 1383219089Spjd if (itxg->itxg_txg != txg) { 1384219089Spjd mutex_exit(&itxg->itxg_lock); 1385219089Spjd continue; 1386219089Spjd } 1387219089Spjd 1388219089Spjd list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1389219089Spjd push_sod += itxg->itxg_sod; 1390219089Spjd itxg->itxg_sod = 0; 1391219089Spjd 1392219089Spjd mutex_exit(&itxg->itxg_lock); 1393219089Spjd } 1394219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1395219089Spjd} 1396219089Spjd 1397219089Spjd/* 1398219089Spjd * Move the async itxs for a specified object to commit into sync lists. 1399219089Spjd */ 1400219089Spjdstatic void 1401219089Spjdzil_async_to_sync(zilog_t *zilog, uint64_t foid) 1402219089Spjd{ 1403219089Spjd uint64_t otxg, txg; 1404219089Spjd itx_async_node_t *ian; 1405219089Spjd avl_tree_t *t; 1406219089Spjd avl_index_t where; 1407219089Spjd 1408219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1409219089Spjd otxg = ZILTEST_TXG; 1410219089Spjd else 1411219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1412219089Spjd 1413219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1414219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1415219089Spjd 1416219089Spjd mutex_enter(&itxg->itxg_lock); 1417219089Spjd if (itxg->itxg_txg != txg) { 1418219089Spjd mutex_exit(&itxg->itxg_lock); 1419219089Spjd continue; 1420219089Spjd } 1421219089Spjd 1422219089Spjd /* 1423219089Spjd * If a foid is specified then find that node and append its 1424219089Spjd * list. Otherwise walk the tree appending all the lists 1425219089Spjd * to the sync list. We add to the end rather than the 1426219089Spjd * beginning to ensure the create has happened. 1427219089Spjd */ 1428219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1429219089Spjd if (foid != 0) { 1430219089Spjd ian = avl_find(t, &foid, &where); 1431219089Spjd if (ian != NULL) { 1432219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1433219089Spjd &ian->ia_list); 1434219089Spjd } 1435219089Spjd } else { 1436219089Spjd void *cookie = NULL; 1437219089Spjd 1438219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1439219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1440219089Spjd &ian->ia_list); 1441219089Spjd list_destroy(&ian->ia_list); 1442219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1443219089Spjd } 1444219089Spjd } 1445219089Spjd mutex_exit(&itxg->itxg_lock); 1446219089Spjd } 1447219089Spjd} 1448219089Spjd 1449219089Spjdstatic void 1450219089Spjdzil_commit_writer(zilog_t *zilog) 1451219089Spjd{ 1452168404Spjd uint64_t txg; 1453219089Spjd itx_t *itx; 1454168404Spjd lwb_t *lwb; 1455219089Spjd spa_t *spa = zilog->zl_spa; 1456219089Spjd int error = 0; 1457168404Spjd 1458185029Spjd ASSERT(zilog->zl_root_zio == NULL); 1459168404Spjd 1460219089Spjd mutex_exit(&zilog->zl_lock); 1461219089Spjd 1462219089Spjd zil_get_commit_list(zilog); 1463219089Spjd 1464219089Spjd /* 1465219089Spjd * Return if there's nothing to commit before we dirty the fs by 1466219089Spjd * calling zil_create(). 1467219089Spjd */ 1468219089Spjd if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1469219089Spjd mutex_enter(&zilog->zl_lock); 1470219089Spjd return; 1471219089Spjd } 1472219089Spjd 1473168404Spjd if (zilog->zl_suspend) { 1474168404Spjd lwb = NULL; 1475168404Spjd } else { 1476168404Spjd lwb = list_tail(&zilog->zl_lwb_list); 1477219089Spjd if (lwb == NULL) 1478219089Spjd lwb = zil_create(zilog); 1479168404Spjd } 1480168404Spjd 1481168404Spjd DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1482219089Spjd while (itx = list_head(&zilog->zl_itx_commit_list)) { 1483168404Spjd txg = itx->itx_lr.lrc_txg; 1484168404Spjd ASSERT(txg); 1485168404Spjd 1486219089Spjd if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1487168404Spjd lwb = zil_lwb_commit(zilog, itx, lwb); 1488219089Spjd list_remove(&zilog->zl_itx_commit_list, itx); 1489168404Spjd kmem_free(itx, offsetof(itx_t, itx_lr) 1490168404Spjd + itx->itx_lr.lrc_reclen); 1491168404Spjd } 1492168404Spjd DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1493168404Spjd 1494168404Spjd /* write the last block out */ 1495168404Spjd if (lwb != NULL && lwb->lwb_zio != NULL) 1496168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1497168404Spjd 1498168404Spjd zilog->zl_cur_used = 0; 1499168404Spjd 1500168404Spjd /* 1501168404Spjd * Wait if necessary for the log blocks to be on stable storage. 1502168404Spjd */ 1503168404Spjd if (zilog->zl_root_zio) { 1504219089Spjd error = zio_wait(zilog->zl_root_zio); 1505185029Spjd zilog->zl_root_zio = NULL; 1506185029Spjd zil_flush_vdevs(zilog); 1507168404Spjd } 1508168404Spjd 1509219089Spjd if (error || lwb == NULL) 1510168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 1511168404Spjd 1512168404Spjd mutex_enter(&zilog->zl_lock); 1513168404Spjd 1514219089Spjd /* 1515219089Spjd * Remember the highest committed log sequence number for ztest. 1516219089Spjd * We only update this value when all the log writes succeeded, 1517219089Spjd * because ztest wants to ASSERT that it got the whole log chain. 1518219089Spjd */ 1519219089Spjd if (error == 0 && lwb != NULL) 1520219089Spjd zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1521168404Spjd} 1522168404Spjd 1523168404Spjd/* 1524219089Spjd * Commit zfs transactions to stable storage. 1525168404Spjd * If foid is 0 push out all transactions, otherwise push only those 1526219089Spjd * for that object or might reference that object. 1527219089Spjd * 1528219089Spjd * itxs are committed in batches. In a heavily stressed zil there will be 1529219089Spjd * a commit writer thread who is writing out a bunch of itxs to the log 1530219089Spjd * for a set of committing threads (cthreads) in the same batch as the writer. 1531219089Spjd * Those cthreads are all waiting on the same cv for that batch. 1532219089Spjd * 1533219089Spjd * There will also be a different and growing batch of threads that are 1534219089Spjd * waiting to commit (qthreads). When the committing batch completes 1535219089Spjd * a transition occurs such that the cthreads exit and the qthreads become 1536219089Spjd * cthreads. One of the new cthreads becomes the writer thread for the 1537219089Spjd * batch. Any new threads arriving become new qthreads. 1538219089Spjd * 1539219089Spjd * Only 2 condition variables are needed and there's no transition 1540219089Spjd * between the two cvs needed. They just flip-flop between qthreads 1541219089Spjd * and cthreads. 1542219089Spjd * 1543219089Spjd * Using this scheme we can efficiently wakeup up only those threads 1544219089Spjd * that have been committed. 1545168404Spjd */ 1546168404Spjdvoid 1547219089Spjdzil_commit(zilog_t *zilog, uint64_t foid) 1548168404Spjd{ 1549219089Spjd uint64_t mybatch; 1550219089Spjd 1551219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1552168404Spjd return; 1553168404Spjd 1554219089Spjd /* move the async itxs for the foid to the sync queues */ 1555219089Spjd zil_async_to_sync(zilog, foid); 1556219089Spjd 1557168404Spjd mutex_enter(&zilog->zl_lock); 1558219089Spjd mybatch = zilog->zl_next_batch; 1559168404Spjd while (zilog->zl_writer) { 1560219089Spjd cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1561219089Spjd if (mybatch <= zilog->zl_com_batch) { 1562168404Spjd mutex_exit(&zilog->zl_lock); 1563168404Spjd return; 1564168404Spjd } 1565168404Spjd } 1566219089Spjd 1567219089Spjd zilog->zl_next_batch++; 1568219089Spjd zilog->zl_writer = B_TRUE; 1569219089Spjd zil_commit_writer(zilog); 1570219089Spjd zilog->zl_com_batch = mybatch; 1571219089Spjd zilog->zl_writer = B_FALSE; 1572168404Spjd mutex_exit(&zilog->zl_lock); 1573219089Spjd 1574219089Spjd /* wake up one thread to become the next writer */ 1575219089Spjd cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1576219089Spjd 1577219089Spjd /* wake up all threads waiting for this batch to be committed */ 1578219089Spjd cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1579168404Spjd} 1580168404Spjd 1581168404Spjd/* 1582168404Spjd * Called in syncing context to free committed log blocks and update log header. 1583168404Spjd */ 1584168404Spjdvoid 1585168404Spjdzil_sync(zilog_t *zilog, dmu_tx_t *tx) 1586168404Spjd{ 1587168404Spjd zil_header_t *zh = zil_header_in_syncing_context(zilog); 1588168404Spjd uint64_t txg = dmu_tx_get_txg(tx); 1589168404Spjd spa_t *spa = zilog->zl_spa; 1590219089Spjd uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1591168404Spjd lwb_t *lwb; 1592168404Spjd 1593209962Smm /* 1594209962Smm * We don't zero out zl_destroy_txg, so make sure we don't try 1595209962Smm * to destroy it twice. 1596209962Smm */ 1597209962Smm if (spa_sync_pass(spa) != 1) 1598209962Smm return; 1599209962Smm 1600168404Spjd mutex_enter(&zilog->zl_lock); 1601168404Spjd 1602168404Spjd ASSERT(zilog->zl_stop_sync == 0); 1603168404Spjd 1604219089Spjd if (*replayed_seq != 0) { 1605219089Spjd ASSERT(zh->zh_replay_seq < *replayed_seq); 1606219089Spjd zh->zh_replay_seq = *replayed_seq; 1607219089Spjd *replayed_seq = 0; 1608219089Spjd } 1609168404Spjd 1610168404Spjd if (zilog->zl_destroy_txg == txg) { 1611168404Spjd blkptr_t blk = zh->zh_log; 1612168404Spjd 1613168404Spjd ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1614168404Spjd 1615168404Spjd bzero(zh, sizeof (zil_header_t)); 1616209962Smm bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1617168404Spjd 1618168404Spjd if (zilog->zl_keep_first) { 1619168404Spjd /* 1620168404Spjd * If this block was part of log chain that couldn't 1621168404Spjd * be claimed because a device was missing during 1622168404Spjd * zil_claim(), but that device later returns, 1623168404Spjd * then this block could erroneously appear valid. 1624168404Spjd * To guard against this, assign a new GUID to the new 1625168404Spjd * log chain so it doesn't matter what blk points to. 1626168404Spjd */ 1627168404Spjd zil_init_log_chain(zilog, &blk); 1628168404Spjd zh->zh_log = blk; 1629168404Spjd } 1630168404Spjd } 1631168404Spjd 1632213197Smm while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1633168404Spjd zh->zh_log = lwb->lwb_blk; 1634168404Spjd if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1635168404Spjd break; 1636168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 1637219089Spjd zio_free_zil(spa, txg, &lwb->lwb_blk); 1638168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 1639168404Spjd 1640168404Spjd /* 1641168404Spjd * If we don't have anything left in the lwb list then 1642168404Spjd * we've had an allocation failure and we need to zero 1643168404Spjd * out the zil_header blkptr so that we don't end 1644168404Spjd * up freeing the same block twice. 1645168404Spjd */ 1646168404Spjd if (list_head(&zilog->zl_lwb_list) == NULL) 1647168404Spjd BP_ZERO(&zh->zh_log); 1648168404Spjd } 1649168404Spjd mutex_exit(&zilog->zl_lock); 1650168404Spjd} 1651168404Spjd 1652168404Spjdvoid 1653168404Spjdzil_init(void) 1654168404Spjd{ 1655168404Spjd zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1656168404Spjd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1657168404Spjd} 1658168404Spjd 1659168404Spjdvoid 1660168404Spjdzil_fini(void) 1661168404Spjd{ 1662168404Spjd kmem_cache_destroy(zil_lwb_cache); 1663168404Spjd} 1664168404Spjd 1665219089Spjdvoid 1666219089Spjdzil_set_sync(zilog_t *zilog, uint64_t sync) 1667219089Spjd{ 1668219089Spjd zilog->zl_sync = sync; 1669219089Spjd} 1670219089Spjd 1671219089Spjdvoid 1672219089Spjdzil_set_logbias(zilog_t *zilog, uint64_t logbias) 1673219089Spjd{ 1674219089Spjd zilog->zl_logbias = logbias; 1675219089Spjd} 1676219089Spjd 1677168404Spjdzilog_t * 1678168404Spjdzil_alloc(objset_t *os, zil_header_t *zh_phys) 1679168404Spjd{ 1680168404Spjd zilog_t *zilog; 1681168404Spjd 1682168404Spjd zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1683168404Spjd 1684168404Spjd zilog->zl_header = zh_phys; 1685168404Spjd zilog->zl_os = os; 1686168404Spjd zilog->zl_spa = dmu_objset_spa(os); 1687168404Spjd zilog->zl_dmu_pool = dmu_objset_pool(os); 1688168404Spjd zilog->zl_destroy_txg = TXG_INITIAL - 1; 1689219089Spjd zilog->zl_logbias = dmu_objset_logbias(os); 1690219089Spjd zilog->zl_sync = dmu_objset_syncprop(os); 1691219089Spjd zilog->zl_next_batch = 1; 1692168404Spjd 1693168404Spjd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1694168404Spjd 1695219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1696219089Spjd mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1697219089Spjd MUTEX_DEFAULT, NULL); 1698219089Spjd } 1699168404Spjd 1700168404Spjd list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1701168404Spjd offsetof(lwb_t, lwb_node)); 1702168404Spjd 1703219089Spjd list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1704219089Spjd offsetof(itx_t, itx_node)); 1705219089Spjd 1706185029Spjd mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1707168404Spjd 1708185029Spjd avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1709185029Spjd sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1710185029Spjd 1711185029Spjd cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1712185029Spjd cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1713219089Spjd cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1714219089Spjd cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1715185029Spjd 1716168404Spjd return (zilog); 1717168404Spjd} 1718168404Spjd 1719168404Spjdvoid 1720168404Spjdzil_free(zilog_t *zilog) 1721168404Spjd{ 1722168404Spjd zilog->zl_stop_sync = 1; 1723168404Spjd 1724248571Smm ASSERT0(zilog->zl_suspend); 1725248571Smm ASSERT0(zilog->zl_suspending); 1726248571Smm 1727224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1728168404Spjd list_destroy(&zilog->zl_lwb_list); 1729168404Spjd 1730185029Spjd avl_destroy(&zilog->zl_vdev_tree); 1731185029Spjd mutex_destroy(&zilog->zl_vdev_lock); 1732168404Spjd 1733219089Spjd ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1734219089Spjd list_destroy(&zilog->zl_itx_commit_list); 1735219089Spjd 1736219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1737219089Spjd /* 1738219089Spjd * It's possible for an itx to be generated that doesn't dirty 1739219089Spjd * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1740219089Spjd * callback to remove the entry. We remove those here. 1741219089Spjd * 1742219089Spjd * Also free up the ziltest itxs. 1743219089Spjd */ 1744219089Spjd if (zilog->zl_itxg[i].itxg_itxs) 1745219089Spjd zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1746219089Spjd mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1747219089Spjd } 1748219089Spjd 1749168404Spjd mutex_destroy(&zilog->zl_lock); 1750168404Spjd 1751185029Spjd cv_destroy(&zilog->zl_cv_writer); 1752185029Spjd cv_destroy(&zilog->zl_cv_suspend); 1753219089Spjd cv_destroy(&zilog->zl_cv_batch[0]); 1754219089Spjd cv_destroy(&zilog->zl_cv_batch[1]); 1755185029Spjd 1756168404Spjd kmem_free(zilog, sizeof (zilog_t)); 1757168404Spjd} 1758168404Spjd 1759168404Spjd/* 1760168404Spjd * Open an intent log. 1761168404Spjd */ 1762168404Spjdzilog_t * 1763168404Spjdzil_open(objset_t *os, zil_get_data_t *get_data) 1764168404Spjd{ 1765168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1766168404Spjd 1767224526Smm ASSERT(zilog->zl_clean_taskq == NULL); 1768224526Smm ASSERT(zilog->zl_get_data == NULL); 1769224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1770224526Smm 1771168404Spjd zilog->zl_get_data = get_data; 1772168404Spjd zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1773168404Spjd 2, 2, TASKQ_PREPOPULATE); 1774168404Spjd 1775168404Spjd return (zilog); 1776168404Spjd} 1777168404Spjd 1778168404Spjd/* 1779168404Spjd * Close an intent log. 1780168404Spjd */ 1781168404Spjdvoid 1782168404Spjdzil_close(zilog_t *zilog) 1783168404Spjd{ 1784224526Smm lwb_t *lwb; 1785219089Spjd uint64_t txg = 0; 1786219089Spjd 1787219089Spjd zil_commit(zilog, 0); /* commit all itx */ 1788219089Spjd 1789168404Spjd /* 1790219089Spjd * The lwb_max_txg for the stubby lwb will reflect the last activity 1791219089Spjd * for the zil. After a txg_wait_synced() on the txg we know all the 1792219089Spjd * callbacks have occurred that may clean the zil. Only then can we 1793219089Spjd * destroy the zl_clean_taskq. 1794168404Spjd */ 1795219089Spjd mutex_enter(&zilog->zl_lock); 1796224526Smm lwb = list_tail(&zilog->zl_lwb_list); 1797224526Smm if (lwb != NULL) 1798224526Smm txg = lwb->lwb_max_txg; 1799219089Spjd mutex_exit(&zilog->zl_lock); 1800219089Spjd if (txg) 1801168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1802239620Smm ASSERT(!zilog_is_dirty(zilog)); 1803168404Spjd 1804168404Spjd taskq_destroy(zilog->zl_clean_taskq); 1805168404Spjd zilog->zl_clean_taskq = NULL; 1806168404Spjd zilog->zl_get_data = NULL; 1807224526Smm 1808224526Smm /* 1809224526Smm * We should have only one LWB left on the list; remove it now. 1810224526Smm */ 1811224526Smm mutex_enter(&zilog->zl_lock); 1812224526Smm lwb = list_head(&zilog->zl_lwb_list); 1813224526Smm if (lwb != NULL) { 1814224526Smm ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1815224526Smm list_remove(&zilog->zl_lwb_list, lwb); 1816224526Smm zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1817224526Smm kmem_cache_free(zil_lwb_cache, lwb); 1818224526Smm } 1819224526Smm mutex_exit(&zilog->zl_lock); 1820168404Spjd} 1821168404Spjd 1822248571Smmstatic char *suspend_tag = "zil suspending"; 1823248571Smm 1824168404Spjd/* 1825168404Spjd * Suspend an intent log. While in suspended mode, we still honor 1826168404Spjd * synchronous semantics, but we rely on txg_wait_synced() to do it. 1827248571Smm * On old version pools, we suspend the log briefly when taking a 1828248571Smm * snapshot so that it will have an empty intent log. 1829248571Smm * 1830248571Smm * Long holds are not really intended to be used the way we do here -- 1831248571Smm * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1832248571Smm * could fail. Therefore we take pains to only put a long hold if it is 1833248571Smm * actually necessary. Fortunately, it will only be necessary if the 1834248571Smm * objset is currently mounted (or the ZVOL equivalent). In that case it 1835248571Smm * will already have a long hold, so we are not really making things any worse. 1836248571Smm * 1837248571Smm * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1838248571Smm * zvol_state_t), and use their mechanism to prevent their hold from being 1839248571Smm * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1840248571Smm * very little gain. 1841248571Smm * 1842248571Smm * if cookiep == NULL, this does both the suspend & resume. 1843248571Smm * Otherwise, it returns with the dataset "long held", and the cookie 1844248571Smm * should be passed into zil_resume(). 1845168404Spjd */ 1846168404Spjdint 1847248571Smmzil_suspend(const char *osname, void **cookiep) 1848168404Spjd{ 1849248571Smm objset_t *os; 1850248571Smm zilog_t *zilog; 1851248571Smm const zil_header_t *zh; 1852248571Smm int error; 1853168404Spjd 1854248571Smm error = dmu_objset_hold(osname, suspend_tag, &os); 1855248571Smm if (error != 0) 1856248571Smm return (error); 1857248571Smm zilog = dmu_objset_zil(os); 1858248571Smm 1859168404Spjd mutex_enter(&zilog->zl_lock); 1860248571Smm zh = zilog->zl_header; 1861248571Smm 1862200724Sdelphij if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1863168404Spjd mutex_exit(&zilog->zl_lock); 1864248571Smm dmu_objset_rele(os, suspend_tag); 1865249195Smm return (SET_ERROR(EBUSY)); 1866168404Spjd } 1867248571Smm 1868248571Smm /* 1869248571Smm * Don't put a long hold in the cases where we can avoid it. This 1870248571Smm * is when there is no cookie so we are doing a suspend & resume 1871248571Smm * (i.e. called from zil_vdev_offline()), and there's nothing to do 1872248571Smm * for the suspend because it's already suspended, or there's no ZIL. 1873248571Smm */ 1874248571Smm if (cookiep == NULL && !zilog->zl_suspending && 1875248571Smm (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1876248571Smm mutex_exit(&zilog->zl_lock); 1877248571Smm dmu_objset_rele(os, suspend_tag); 1878248571Smm return (0); 1879248571Smm } 1880248571Smm 1881248571Smm dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1882248571Smm dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1883248571Smm 1884248571Smm zilog->zl_suspend++; 1885248571Smm 1886248571Smm if (zilog->zl_suspend > 1) { 1887168404Spjd /* 1888248571Smm * Someone else is already suspending it. 1889168404Spjd * Just wait for them to finish. 1890168404Spjd */ 1891248571Smm 1892168404Spjd while (zilog->zl_suspending) 1893168404Spjd cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1894168404Spjd mutex_exit(&zilog->zl_lock); 1895248571Smm 1896248571Smm if (cookiep == NULL) 1897248571Smm zil_resume(os); 1898248571Smm else 1899248571Smm *cookiep = os; 1900168404Spjd return (0); 1901168404Spjd } 1902248571Smm 1903248571Smm /* 1904248571Smm * If there is no pointer to an on-disk block, this ZIL must not 1905248571Smm * be active (e.g. filesystem not mounted), so there's nothing 1906248571Smm * to clean up. 1907248571Smm */ 1908248571Smm if (BP_IS_HOLE(&zh->zh_log)) { 1909248571Smm ASSERT(cookiep != NULL); /* fast path already handled */ 1910248571Smm 1911248571Smm *cookiep = os; 1912248571Smm mutex_exit(&zilog->zl_lock); 1913248571Smm return (0); 1914248571Smm } 1915248571Smm 1916168404Spjd zilog->zl_suspending = B_TRUE; 1917168404Spjd mutex_exit(&zilog->zl_lock); 1918168404Spjd 1919219089Spjd zil_commit(zilog, 0); 1920168404Spjd 1921168404Spjd zil_destroy(zilog, B_FALSE); 1922168404Spjd 1923168404Spjd mutex_enter(&zilog->zl_lock); 1924168404Spjd zilog->zl_suspending = B_FALSE; 1925168404Spjd cv_broadcast(&zilog->zl_cv_suspend); 1926168404Spjd mutex_exit(&zilog->zl_lock); 1927168404Spjd 1928248571Smm if (cookiep == NULL) 1929248571Smm zil_resume(os); 1930248571Smm else 1931248571Smm *cookiep = os; 1932168404Spjd return (0); 1933168404Spjd} 1934168404Spjd 1935168404Spjdvoid 1936248571Smmzil_resume(void *cookie) 1937168404Spjd{ 1938248571Smm objset_t *os = cookie; 1939248571Smm zilog_t *zilog = dmu_objset_zil(os); 1940248571Smm 1941168404Spjd mutex_enter(&zilog->zl_lock); 1942168404Spjd ASSERT(zilog->zl_suspend != 0); 1943168404Spjd zilog->zl_suspend--; 1944168404Spjd mutex_exit(&zilog->zl_lock); 1945248571Smm dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1946248571Smm dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1947168404Spjd} 1948168404Spjd 1949219089Spjdtypedef struct zil_replay_arg { 1950219089Spjd zil_replay_func_t **zr_replay; 1951219089Spjd void *zr_arg; 1952219089Spjd boolean_t zr_byteswap; 1953219089Spjd char *zr_lr; 1954219089Spjd} zil_replay_arg_t; 1955219089Spjd 1956219089Spjdstatic int 1957219089Spjdzil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1958209962Smm{ 1959219089Spjd char name[MAXNAMELEN]; 1960209962Smm 1961219089Spjd zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1962209962Smm 1963219089Spjd dmu_objset_name(zilog->zl_os, name); 1964209962Smm 1965219089Spjd cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1966219089Spjd "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1967219089Spjd (u_longlong_t)lr->lrc_seq, 1968219089Spjd (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1969219089Spjd (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1970219089Spjd 1971219089Spjd return (error); 1972209962Smm} 1973209962Smm 1974219089Spjdstatic int 1975168404Spjdzil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1976168404Spjd{ 1977168404Spjd zil_replay_arg_t *zr = zra; 1978168404Spjd const zil_header_t *zh = zilog->zl_header; 1979168404Spjd uint64_t reclen = lr->lrc_reclen; 1980168404Spjd uint64_t txtype = lr->lrc_txtype; 1981219089Spjd int error = 0; 1982168404Spjd 1983219089Spjd zilog->zl_replaying_seq = lr->lrc_seq; 1984168404Spjd 1985219089Spjd if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1986219089Spjd return (0); 1987219089Spjd 1988168404Spjd if (lr->lrc_txg < claim_txg) /* already committed */ 1989219089Spjd return (0); 1990168404Spjd 1991185029Spjd /* Strip case-insensitive bit, still present in log record */ 1992185029Spjd txtype &= ~TX_CI; 1993185029Spjd 1994219089Spjd if (txtype == 0 || txtype >= TX_MAX_TYPE) 1995219089Spjd return (zil_replay_error(zilog, lr, EINVAL)); 1996219089Spjd 1997219089Spjd /* 1998219089Spjd * If this record type can be logged out of order, the object 1999219089Spjd * (lr_foid) may no longer exist. That's legitimate, not an error. 2000219089Spjd */ 2001219089Spjd if (TX_OOO(txtype)) { 2002219089Spjd error = dmu_object_info(zilog->zl_os, 2003219089Spjd ((lr_ooo_t *)lr)->lr_foid, NULL); 2004219089Spjd if (error == ENOENT || error == EEXIST) 2005219089Spjd return (0); 2006209962Smm } 2007209962Smm 2008168404Spjd /* 2009168404Spjd * Make a copy of the data so we can revise and extend it. 2010168404Spjd */ 2011219089Spjd bcopy(lr, zr->zr_lr, reclen); 2012168404Spjd 2013168404Spjd /* 2014219089Spjd * If this is a TX_WRITE with a blkptr, suck in the data. 2015219089Spjd */ 2016219089Spjd if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2017219089Spjd error = zil_read_log_data(zilog, (lr_write_t *)lr, 2018219089Spjd zr->zr_lr + reclen); 2019248571Smm if (error != 0) 2020219089Spjd return (zil_replay_error(zilog, lr, error)); 2021219089Spjd } 2022219089Spjd 2023219089Spjd /* 2024168404Spjd * The log block containing this lr may have been byteswapped 2025168404Spjd * so that we can easily examine common fields like lrc_txtype. 2026219089Spjd * However, the log is a mix of different record types, and only the 2027168404Spjd * replay vectors know how to byteswap their records. Therefore, if 2028168404Spjd * the lr was byteswapped, undo it before invoking the replay vector. 2029168404Spjd */ 2030168404Spjd if (zr->zr_byteswap) 2031219089Spjd byteswap_uint64_array(zr->zr_lr, reclen); 2032168404Spjd 2033168404Spjd /* 2034168404Spjd * We must now do two things atomically: replay this log record, 2035209962Smm * and update the log header sequence number to reflect the fact that 2036209962Smm * we did so. At the end of each replay function the sequence number 2037209962Smm * is updated if we are in replay mode. 2038168404Spjd */ 2039219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2040248571Smm if (error != 0) { 2041168404Spjd /* 2042168404Spjd * The DMU's dnode layer doesn't see removes until the txg 2043168404Spjd * commits, so a subsequent claim can spuriously fail with 2044209962Smm * EEXIST. So if we receive any error we try syncing out 2045219089Spjd * any removes then retry the transaction. Note that we 2046219089Spjd * specify B_FALSE for byteswap now, so we don't do it twice. 2047168404Spjd */ 2048219089Spjd txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2049219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2050248571Smm if (error != 0) 2051219089Spjd return (zil_replay_error(zilog, lr, error)); 2052168404Spjd } 2053219089Spjd return (0); 2054168404Spjd} 2055168404Spjd 2056168404Spjd/* ARGSUSED */ 2057219089Spjdstatic int 2058168404Spjdzil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2059168404Spjd{ 2060168404Spjd zilog->zl_replay_blks++; 2061219089Spjd 2062219089Spjd return (0); 2063168404Spjd} 2064168404Spjd 2065168404Spjd/* 2066168404Spjd * If this dataset has a non-empty intent log, replay it and destroy it. 2067168404Spjd */ 2068168404Spjdvoid 2069209962Smmzil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2070168404Spjd{ 2071168404Spjd zilog_t *zilog = dmu_objset_zil(os); 2072168404Spjd const zil_header_t *zh = zilog->zl_header; 2073168404Spjd zil_replay_arg_t zr; 2074168404Spjd 2075200724Sdelphij if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2076168404Spjd zil_destroy(zilog, B_TRUE); 2077168404Spjd return; 2078168404Spjd } 2079168404Spjd //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 2080168404Spjd 2081168404Spjd zr.zr_replay = replay_func; 2082168404Spjd zr.zr_arg = arg; 2083168404Spjd zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2084219089Spjd zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2085168404Spjd 2086168404Spjd /* 2087168404Spjd * Wait for in-progress removes to sync before starting replay. 2088168404Spjd */ 2089168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 2090168404Spjd 2091209962Smm zilog->zl_replay = B_TRUE; 2092219089Spjd zilog->zl_replay_time = ddi_get_lbolt(); 2093168404Spjd ASSERT(zilog->zl_replay_blks == 0); 2094168404Spjd (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2095168404Spjd zh->zh_claim_txg); 2096219089Spjd kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2097168404Spjd 2098168404Spjd zil_destroy(zilog, B_FALSE); 2099185029Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2100209962Smm zilog->zl_replay = B_FALSE; 2101168404Spjd //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2102168404Spjd} 2103168404Spjd 2104219089Spjdboolean_t 2105219089Spjdzil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2106168404Spjd{ 2107219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2108219089Spjd return (B_TRUE); 2109168404Spjd 2110219089Spjd if (zilog->zl_replay) { 2111219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2112219089Spjd zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2113219089Spjd zilog->zl_replaying_seq; 2114219089Spjd return (B_TRUE); 2115168404Spjd } 2116168404Spjd 2117219089Spjd return (B_FALSE); 2118168404Spjd} 2119213197Smm 2120213197Smm/* ARGSUSED */ 2121213197Smmint 2122219089Spjdzil_vdev_offline(const char *osname, void *arg) 2123213197Smm{ 2124213197Smm int error; 2125213197Smm 2126248571Smm error = zil_suspend(osname, NULL); 2127248571Smm if (error != 0) 2128249195Smm return (SET_ERROR(EEXIST)); 2129248571Smm return (0); 2130213197Smm} 2131