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. 23249643Smm * 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/* 69252751Sdelphij * Disable intent logging replay. This global ZIL switch affects all pools. 70168404Spjd */ 71252751Sdelphijint 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"); 86251419Ssmhboolean_t zfs_trim_enabled = B_TRUE; 87251419SsmhSYSCTL_DECL(_vfs_zfs_trim); 88251419SsmhTUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 89251419SsmhSYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 90251419Ssmh "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) 157249643Smm 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 205247406Smm 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)) { 228249643Smm 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)))) { 242249643Smm 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 250249643Smm 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 281247406Smm 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); 347249643Smm 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 */ 386263398Sdelphij if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 387263398Sdelphij zil_bp_tree_add(zilog, bp) != 0) 388219089Spjd return (0); 389219089Spjd 390219089Spjd return (zio_wait(zio_claim(NULL, zilog->zl_spa, 391219089Spjd tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 392219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 393168404Spjd} 394168404Spjd 395219089Spjdstatic int 396168404Spjdzil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 397168404Spjd{ 398219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 399219089Spjd int error; 400219089Spjd 401219089Spjd if (lrc->lrc_txtype != TX_WRITE) 402219089Spjd return (0); 403219089Spjd 404219089Spjd /* 405219089Spjd * If the block is not readable, don't claim it. This can happen 406219089Spjd * in normal operation when a log block is written to disk before 407219089Spjd * some of the dmu_sync() blocks it points to. In this case, the 408219089Spjd * transaction cannot have been committed to anyone (we would have 409219089Spjd * waited for all writes to be stable first), so it is semantically 410219089Spjd * correct to declare this the end of the log. 411219089Spjd */ 412219089Spjd if (lr->lr_blkptr.blk_birth >= first_txg && 413219089Spjd (error = zil_read_log_data(zilog, lr, NULL)) != 0) 414219089Spjd return (error); 415219089Spjd return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 416168404Spjd} 417168404Spjd 418168404Spjd/* ARGSUSED */ 419219089Spjdstatic int 420168404Spjdzil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 421168404Spjd{ 422219089Spjd zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 423219089Spjd 424219089Spjd return (0); 425168404Spjd} 426168404Spjd 427219089Spjdstatic int 428168404Spjdzil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 429168404Spjd{ 430219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 431219089Spjd blkptr_t *bp = &lr->lr_blkptr; 432219089Spjd 433168404Spjd /* 434168404Spjd * If we previously claimed it, we need to free it. 435168404Spjd */ 436219089Spjd if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 437263398Sdelphij bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 438263398Sdelphij !BP_IS_HOLE(bp)) 439219089Spjd zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 440219089Spjd 441219089Spjd return (0); 442219089Spjd} 443219089Spjd 444219089Spjdstatic lwb_t * 445219089Spjdzil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 446219089Spjd{ 447219089Spjd lwb_t *lwb; 448219089Spjd 449219089Spjd lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 450219089Spjd lwb->lwb_zilog = zilog; 451219089Spjd lwb->lwb_blk = *bp; 452219089Spjd lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 453219089Spjd lwb->lwb_max_txg = txg; 454219089Spjd lwb->lwb_zio = NULL; 455219089Spjd lwb->lwb_tx = NULL; 456219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 457219089Spjd lwb->lwb_nused = sizeof (zil_chain_t); 458219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp); 459219089Spjd } else { 460219089Spjd lwb->lwb_nused = 0; 461219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 462168404Spjd } 463219089Spjd 464219089Spjd mutex_enter(&zilog->zl_lock); 465219089Spjd list_insert_tail(&zilog->zl_lwb_list, lwb); 466219089Spjd mutex_exit(&zilog->zl_lock); 467219089Spjd 468219089Spjd return (lwb); 469168404Spjd} 470168404Spjd 471168404Spjd/* 472243674Smm * Called when we create in-memory log transactions so that we know 473243674Smm * to cleanup the itxs at the end of spa_sync(). 474243674Smm */ 475243674Smmvoid 476243674Smmzilog_dirty(zilog_t *zilog, uint64_t txg) 477243674Smm{ 478243674Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 479243674Smm dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 480243674Smm 481243674Smm if (dsl_dataset_is_snapshot(ds)) 482243674Smm panic("dirtying snapshot!"); 483243674Smm 484249643Smm if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 485243674Smm /* up the hold count until we can be written out */ 486243674Smm dmu_buf_add_ref(ds->ds_dbuf, zilog); 487243674Smm } 488243674Smm} 489243674Smm 490243674Smmboolean_t 491243674Smmzilog_is_dirty(zilog_t *zilog) 492243674Smm{ 493243674Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 494243674Smm 495243674Smm for (int t = 0; t < TXG_SIZE; t++) { 496243674Smm if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 497243674Smm return (B_TRUE); 498243674Smm } 499243674Smm return (B_FALSE); 500243674Smm} 501243674Smm 502243674Smm/* 503168404Spjd * Create an on-disk intent log. 504168404Spjd */ 505219089Spjdstatic lwb_t * 506168404Spjdzil_create(zilog_t *zilog) 507168404Spjd{ 508168404Spjd const zil_header_t *zh = zilog->zl_header; 509219089Spjd lwb_t *lwb = NULL; 510168404Spjd uint64_t txg = 0; 511168404Spjd dmu_tx_t *tx = NULL; 512168404Spjd blkptr_t blk; 513168404Spjd int error = 0; 514168404Spjd 515168404Spjd /* 516168404Spjd * Wait for any previous destroy to complete. 517168404Spjd */ 518168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 519168404Spjd 520168404Spjd ASSERT(zh->zh_claim_txg == 0); 521168404Spjd ASSERT(zh->zh_replay_seq == 0); 522168404Spjd 523168404Spjd blk = zh->zh_log; 524168404Spjd 525168404Spjd /* 526219089Spjd * Allocate an initial log block if: 527219089Spjd * - there isn't one already 528219089Spjd * - the existing block is the wrong endianess 529168404Spjd */ 530207908Smm if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 531168404Spjd tx = dmu_tx_create(zilog->zl_os); 532219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 533168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 534168404Spjd txg = dmu_tx_get_txg(tx); 535168404Spjd 536207908Smm if (!BP_IS_HOLE(&blk)) { 537219089Spjd zio_free_zil(zilog->zl_spa, txg, &blk); 538207908Smm BP_ZERO(&blk); 539207908Smm } 540207908Smm 541219089Spjd error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 542219089Spjd ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 543168404Spjd 544168404Spjd if (error == 0) 545168404Spjd zil_init_log_chain(zilog, &blk); 546168404Spjd } 547168404Spjd 548168404Spjd /* 549168404Spjd * Allocate a log write buffer (lwb) for the first log block. 550168404Spjd */ 551219089Spjd if (error == 0) 552219089Spjd lwb = zil_alloc_lwb(zilog, &blk, txg); 553168404Spjd 554168404Spjd /* 555168404Spjd * If we just allocated the first log block, commit our transaction 556168404Spjd * and wait for zil_sync() to stuff the block poiner into zh_log. 557168404Spjd * (zh is part of the MOS, so we cannot modify it in open context.) 558168404Spjd */ 559168404Spjd if (tx != NULL) { 560168404Spjd dmu_tx_commit(tx); 561168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 562168404Spjd } 563168404Spjd 564168404Spjd ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 565219089Spjd 566219089Spjd return (lwb); 567168404Spjd} 568168404Spjd 569168404Spjd/* 570168404Spjd * In one tx, free all log blocks and clear the log header. 571168404Spjd * If keep_first is set, then we're replaying a log with no content. 572168404Spjd * We want to keep the first block, however, so that the first 573168404Spjd * synchronous transaction doesn't require a txg_wait_synced() 574168404Spjd * in zil_create(). We don't need to txg_wait_synced() here either 575168404Spjd * when keep_first is set, because both zil_create() and zil_destroy() 576168404Spjd * will wait for any in-progress destroys to complete. 577168404Spjd */ 578168404Spjdvoid 579168404Spjdzil_destroy(zilog_t *zilog, boolean_t keep_first) 580168404Spjd{ 581168404Spjd const zil_header_t *zh = zilog->zl_header; 582168404Spjd lwb_t *lwb; 583168404Spjd dmu_tx_t *tx; 584168404Spjd uint64_t txg; 585168404Spjd 586168404Spjd /* 587168404Spjd * Wait for any previous destroy to complete. 588168404Spjd */ 589168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 590168404Spjd 591219089Spjd zilog->zl_old_header = *zh; /* debugging aid */ 592219089Spjd 593168404Spjd if (BP_IS_HOLE(&zh->zh_log)) 594168404Spjd return; 595168404Spjd 596168404Spjd tx = dmu_tx_create(zilog->zl_os); 597219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 598168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 599168404Spjd txg = dmu_tx_get_txg(tx); 600168404Spjd 601168404Spjd mutex_enter(&zilog->zl_lock); 602168404Spjd 603168404Spjd ASSERT3U(zilog->zl_destroy_txg, <, txg); 604168404Spjd zilog->zl_destroy_txg = txg; 605168404Spjd zilog->zl_keep_first = keep_first; 606168404Spjd 607168404Spjd if (!list_is_empty(&zilog->zl_lwb_list)) { 608168404Spjd ASSERT(zh->zh_claim_txg == 0); 609224526Smm VERIFY(!keep_first); 610168404Spjd while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 611168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 612168404Spjd if (lwb->lwb_buf != NULL) 613168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 614219089Spjd zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 615168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 616168404Spjd } 617219089Spjd } else if (!keep_first) { 618243674Smm zil_destroy_sync(zilog, tx); 619168404Spjd } 620168404Spjd mutex_exit(&zilog->zl_lock); 621168404Spjd 622168404Spjd dmu_tx_commit(tx); 623185029Spjd} 624168404Spjd 625243674Smmvoid 626243674Smmzil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 627243674Smm{ 628243674Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 629243674Smm (void) zil_parse(zilog, zil_free_log_block, 630243674Smm zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 631243674Smm} 632243674Smm 633168404Spjdint 634219089Spjdzil_claim(const char *osname, void *txarg) 635168404Spjd{ 636168404Spjd dmu_tx_t *tx = txarg; 637168404Spjd uint64_t first_txg = dmu_tx_get_txg(tx); 638168404Spjd zilog_t *zilog; 639168404Spjd zil_header_t *zh; 640168404Spjd objset_t *os; 641168404Spjd int error; 642168404Spjd 643249643Smm error = dmu_objset_own(osname, DMU_OST_ANY, B_FALSE, FTAG, &os); 644249643Smm if (error != 0) { 645185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 646168404Spjd return (0); 647168404Spjd } 648168404Spjd 649168404Spjd zilog = dmu_objset_zil(os); 650168404Spjd zh = zil_header_in_syncing_context(zilog); 651168404Spjd 652219089Spjd if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 653213197Smm if (!BP_IS_HOLE(&zh->zh_log)) 654219089Spjd zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 655213197Smm BP_ZERO(&zh->zh_log); 656213197Smm dsl_dataset_dirty(dmu_objset_ds(os), tx); 657249643Smm dmu_objset_disown(os, FTAG); 658219089Spjd return (0); 659213197Smm } 660213197Smm 661168404Spjd /* 662168404Spjd * Claim all log blocks if we haven't already done so, and remember 663168404Spjd * the highest claimed sequence number. This ensures that if we can 664168404Spjd * read only part of the log now (e.g. due to a missing device), 665168404Spjd * but we can read the entire log later, we will not try to replay 666168404Spjd * or destroy beyond the last block we successfully claimed. 667168404Spjd */ 668168404Spjd ASSERT3U(zh->zh_claim_txg, <=, first_txg); 669168404Spjd if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 670219089Spjd (void) zil_parse(zilog, zil_claim_log_block, 671219089Spjd zil_claim_log_record, tx, first_txg); 672168404Spjd zh->zh_claim_txg = first_txg; 673219089Spjd zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 674219089Spjd zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 675219089Spjd if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 676219089Spjd zh->zh_flags |= ZIL_REPLAY_NEEDED; 677219089Spjd zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 678168404Spjd dsl_dataset_dirty(dmu_objset_ds(os), tx); 679168404Spjd } 680168404Spjd 681168404Spjd ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 682249643Smm dmu_objset_disown(os, FTAG); 683168404Spjd return (0); 684168404Spjd} 685168404Spjd 686185029Spjd/* 687185029Spjd * Check the log by walking the log chain. 688185029Spjd * Checksum errors are ok as they indicate the end of the chain. 689185029Spjd * Any other error (no device or read failure) returns an error. 690185029Spjd */ 691185029Spjdint 692219089Spjdzil_check_log_chain(const char *osname, void *tx) 693168404Spjd{ 694185029Spjd zilog_t *zilog; 695185029Spjd objset_t *os; 696219089Spjd blkptr_t *bp; 697185029Spjd int error; 698168404Spjd 699219089Spjd ASSERT(tx == NULL); 700219089Spjd 701219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 702249643Smm if (error != 0) { 703185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 704185029Spjd return (0); 705185029Spjd } 706168404Spjd 707185029Spjd zilog = dmu_objset_zil(os); 708219089Spjd bp = (blkptr_t *)&zilog->zl_header->zh_log; 709219089Spjd 710219089Spjd /* 711219089Spjd * Check the first block and determine if it's on a log device 712219089Spjd * which may have been removed or faulted prior to loading this 713219089Spjd * pool. If so, there's no point in checking the rest of the log 714219089Spjd * as its content should have already been synced to the pool. 715219089Spjd */ 716219089Spjd if (!BP_IS_HOLE(bp)) { 717219089Spjd vdev_t *vd; 718219089Spjd boolean_t valid = B_TRUE; 719219089Spjd 720219089Spjd spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 721219089Spjd vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 722219089Spjd if (vd->vdev_islog && vdev_is_dead(vd)) 723219089Spjd valid = vdev_log_state_valid(vd); 724219089Spjd spa_config_exit(os->os_spa, SCL_STATE, FTAG); 725219089Spjd 726219089Spjd if (!valid) { 727219089Spjd dmu_objset_rele(os, FTAG); 728219089Spjd return (0); 729219089Spjd } 730168404Spjd } 731185029Spjd 732219089Spjd /* 733219089Spjd * Because tx == NULL, zil_claim_log_block() will not actually claim 734219089Spjd * any blocks, but just determine whether it is possible to do so. 735219089Spjd * In addition to checking the log chain, zil_claim_log_block() 736219089Spjd * will invoke zio_claim() with a done func of spa_claim_notify(), 737219089Spjd * which will update spa_max_claim_txg. See spa_load() for details. 738219089Spjd */ 739219089Spjd error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 740219089Spjd zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 741219089Spjd 742219089Spjd dmu_objset_rele(os, FTAG); 743219089Spjd 744219089Spjd return ((error == ECKSUM || error == ENOENT) ? 0 : error); 745168404Spjd} 746168404Spjd 747185029Spjdstatic int 748185029Spjdzil_vdev_compare(const void *x1, const void *x2) 749185029Spjd{ 750219089Spjd const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 751219089Spjd const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 752185029Spjd 753185029Spjd if (v1 < v2) 754185029Spjd return (-1); 755185029Spjd if (v1 > v2) 756185029Spjd return (1); 757185029Spjd 758185029Spjd return (0); 759185029Spjd} 760185029Spjd 761168404Spjdvoid 762219089Spjdzil_add_block(zilog_t *zilog, const blkptr_t *bp) 763168404Spjd{ 764185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 765185029Spjd avl_index_t where; 766185029Spjd zil_vdev_node_t *zv, zvsearch; 767185029Spjd int ndvas = BP_GET_NDVAS(bp); 768185029Spjd int i; 769168404Spjd 770185029Spjd if (zfs_nocacheflush) 771185029Spjd return; 772168404Spjd 773185029Spjd ASSERT(zilog->zl_writer); 774168404Spjd 775185029Spjd /* 776185029Spjd * Even though we're zl_writer, we still need a lock because the 777185029Spjd * zl_get_data() callbacks may have dmu_sync() done callbacks 778185029Spjd * that will run concurrently. 779185029Spjd */ 780185029Spjd mutex_enter(&zilog->zl_vdev_lock); 781185029Spjd for (i = 0; i < ndvas; i++) { 782185029Spjd zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 783185029Spjd if (avl_find(t, &zvsearch, &where) == NULL) { 784185029Spjd zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 785185029Spjd zv->zv_vdev = zvsearch.zv_vdev; 786185029Spjd avl_insert(t, zv, where); 787185029Spjd } 788185029Spjd } 789185029Spjd mutex_exit(&zilog->zl_vdev_lock); 790168404Spjd} 791168404Spjd 792219089Spjdstatic void 793168404Spjdzil_flush_vdevs(zilog_t *zilog) 794168404Spjd{ 795168404Spjd spa_t *spa = zilog->zl_spa; 796185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 797185029Spjd void *cookie = NULL; 798185029Spjd zil_vdev_node_t *zv; 799185029Spjd zio_t *zio; 800168404Spjd 801168404Spjd ASSERT(zilog->zl_writer); 802168404Spjd 803185029Spjd /* 804185029Spjd * We don't need zl_vdev_lock here because we're the zl_writer, 805185029Spjd * and all zl_get_data() callbacks are done. 806185029Spjd */ 807185029Spjd if (avl_numnodes(t) == 0) 808185029Spjd return; 809185029Spjd 810185029Spjd spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 811185029Spjd 812185029Spjd zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 813185029Spjd 814185029Spjd while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 815185029Spjd vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 816185029Spjd if (vd != NULL) 817185029Spjd zio_flush(zio, vd); 818185029Spjd kmem_free(zv, sizeof (*zv)); 819168404Spjd } 820168404Spjd 821168404Spjd /* 822168404Spjd * Wait for all the flushes to complete. Not all devices actually 823168404Spjd * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 824168404Spjd */ 825185029Spjd (void) zio_wait(zio); 826185029Spjd 827185029Spjd spa_config_exit(spa, SCL_STATE, FTAG); 828168404Spjd} 829168404Spjd 830168404Spjd/* 831168404Spjd * Function called when a log block write completes 832168404Spjd */ 833168404Spjdstatic void 834168404Spjdzil_lwb_write_done(zio_t *zio) 835168404Spjd{ 836168404Spjd lwb_t *lwb = zio->io_private; 837168404Spjd zilog_t *zilog = lwb->lwb_zilog; 838219089Spjd dmu_tx_t *tx = lwb->lwb_tx; 839168404Spjd 840185029Spjd ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 841185029Spjd ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 842185029Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 843185029Spjd ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 844185029Spjd ASSERT(!BP_IS_GANG(zio->io_bp)); 845185029Spjd ASSERT(!BP_IS_HOLE(zio->io_bp)); 846185029Spjd ASSERT(zio->io_bp->blk_fill == 0); 847185029Spjd 848168404Spjd /* 849209962Smm * Ensure the lwb buffer pointer is cleared before releasing 850209962Smm * the txg. If we have had an allocation failure and 851209962Smm * the txg is waiting to sync then we want want zil_sync() 852209962Smm * to remove the lwb so that it's not picked up as the next new 853209962Smm * one in zil_commit_writer(). zil_sync() will only remove 854209962Smm * the lwb if lwb_buf is null. 855168404Spjd */ 856168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 857168404Spjd mutex_enter(&zilog->zl_lock); 858168404Spjd lwb->lwb_buf = NULL; 859219089Spjd lwb->lwb_tx = NULL; 860219089Spjd mutex_exit(&zilog->zl_lock); 861209962Smm 862209962Smm /* 863209962Smm * Now that we've written this log block, we have a stable pointer 864209962Smm * to the next block in the chain, so it's OK to let the txg in 865219089Spjd * which we allocated the next block sync. 866209962Smm */ 867219089Spjd dmu_tx_commit(tx); 868168404Spjd} 869168404Spjd 870168404Spjd/* 871168404Spjd * Initialize the io for a log block. 872168404Spjd */ 873168404Spjdstatic void 874168404Spjdzil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 875168404Spjd{ 876168404Spjd zbookmark_t zb; 877168404Spjd 878219089Spjd SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 879219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 880219089Spjd lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 881168404Spjd 882168404Spjd if (zilog->zl_root_zio == NULL) { 883168404Spjd zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 884168404Spjd ZIO_FLAG_CANFAIL); 885168404Spjd } 886168404Spjd if (lwb->lwb_zio == NULL) { 887168404Spjd lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 888219089Spjd 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 889260764Savg zil_lwb_write_done, lwb, ZIO_PRIORITY_SYNC_WRITE, 890219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 891168404Spjd } 892168404Spjd} 893168404Spjd 894168404Spjd/* 895219089Spjd * Define a limited set of intent log block sizes. 896252751Sdelphij * 897219089Spjd * These must be a multiple of 4KB. Note only the amount used (again 898219089Spjd * aligned to 4KB) actually gets written. However, we can't always just 899219089Spjd * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 900219089Spjd */ 901219089Spjduint64_t zil_block_buckets[] = { 902219089Spjd 4096, /* non TX_WRITE */ 903219089Spjd 8192+4096, /* data base */ 904219089Spjd 32*1024 + 4096, /* NFS writes */ 905219089Spjd UINT64_MAX 906219089Spjd}; 907219089Spjd 908219089Spjd/* 909219089Spjd * Use the slog as long as the logbias is 'latency' and the current commit size 910219089Spjd * is less than the limit or the total list size is less than 2X the limit. 911219089Spjd * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 912219089Spjd */ 913219089Spjduint64_t zil_slog_limit = 1024 * 1024; 914219089Spjd#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 915219089Spjd (((zilog)->zl_cur_used < zil_slog_limit) || \ 916219089Spjd ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 917219089Spjd 918219089Spjd/* 919168404Spjd * Start a log block write and advance to the next log block. 920168404Spjd * Calls are serialized. 921168404Spjd */ 922168404Spjdstatic lwb_t * 923168404Spjdzil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 924168404Spjd{ 925219089Spjd lwb_t *nlwb = NULL; 926219089Spjd zil_chain_t *zilc; 927168404Spjd spa_t *spa = zilog->zl_spa; 928219089Spjd blkptr_t *bp; 929219089Spjd dmu_tx_t *tx; 930168404Spjd uint64_t txg; 931219089Spjd uint64_t zil_blksz, wsz; 932219089Spjd int i, error; 933168404Spjd 934219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 935219089Spjd zilc = (zil_chain_t *)lwb->lwb_buf; 936219089Spjd bp = &zilc->zc_next_blk; 937219089Spjd } else { 938219089Spjd zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 939219089Spjd bp = &zilc->zc_next_blk; 940219089Spjd } 941168404Spjd 942219089Spjd ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 943219089Spjd 944168404Spjd /* 945168404Spjd * Allocate the next block and save its address in this block 946168404Spjd * before writing it in order to establish the log chain. 947168404Spjd * Note that if the allocation of nlwb synced before we wrote 948168404Spjd * the block that points at it (lwb), we'd leak it if we crashed. 949219089Spjd * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 950219089Spjd * We dirty the dataset to ensure that zil_sync() will be called 951219089Spjd * to clean up in the event of allocation failure or I/O failure. 952168404Spjd */ 953219089Spjd tx = dmu_tx_create(zilog->zl_os); 954219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 955219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 956219089Spjd txg = dmu_tx_get_txg(tx); 957168404Spjd 958219089Spjd lwb->lwb_tx = tx; 959219089Spjd 960168404Spjd /* 961219089Spjd * Log blocks are pre-allocated. Here we select the size of the next 962219089Spjd * block, based on size used in the last block. 963219089Spjd * - first find the smallest bucket that will fit the block from a 964219089Spjd * limited set of block sizes. This is because it's faster to write 965219089Spjd * blocks allocated from the same metaslab as they are adjacent or 966219089Spjd * close. 967219089Spjd * - next find the maximum from the new suggested size and an array of 968219089Spjd * previous sizes. This lessens a picket fence effect of wrongly 969219089Spjd * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 970219089Spjd * requests. 971219089Spjd * 972219089Spjd * Note we only write what is used, but we can't just allocate 973219089Spjd * the maximum block size because we can exhaust the available 974219089Spjd * pool log space. 975168404Spjd */ 976219089Spjd zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 977219089Spjd for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 978219089Spjd continue; 979219089Spjd zil_blksz = zil_block_buckets[i]; 980219089Spjd if (zil_blksz == UINT64_MAX) 981219089Spjd zil_blksz = SPA_MAXBLOCKSIZE; 982219089Spjd zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 983219089Spjd for (i = 0; i < ZIL_PREV_BLKS; i++) 984219089Spjd zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 985219089Spjd zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 986168404Spjd 987168404Spjd BP_ZERO(bp); 988168404Spjd /* pass the old blkptr in order to spread log blocks across devs */ 989219089Spjd error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 990219089Spjd USE_SLOG(zilog)); 991249643Smm if (error == 0) { 992219089Spjd ASSERT3U(bp->blk_birth, ==, txg); 993219089Spjd bp->blk_cksum = lwb->lwb_blk.blk_cksum; 994219089Spjd bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 995168404Spjd 996168404Spjd /* 997219089Spjd * Allocate a new log write buffer (lwb). 998168404Spjd */ 999219089Spjd nlwb = zil_alloc_lwb(zilog, bp, txg); 1000168404Spjd 1001219089Spjd /* Record the block for later vdev flushing */ 1002219089Spjd zil_add_block(zilog, &lwb->lwb_blk); 1003168404Spjd } 1004168404Spjd 1005219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1006219089Spjd /* For Slim ZIL only write what is used. */ 1007219089Spjd wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1008219089Spjd ASSERT3U(wsz, <=, lwb->lwb_sz); 1009219089Spjd zio_shrink(lwb->lwb_zio, wsz); 1010168404Spjd 1011219089Spjd } else { 1012219089Spjd wsz = lwb->lwb_sz; 1013219089Spjd } 1014168404Spjd 1015219089Spjd zilc->zc_pad = 0; 1016219089Spjd zilc->zc_nused = lwb->lwb_nused; 1017219089Spjd zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1018168404Spjd 1019168404Spjd /* 1020219089Spjd * clear unused data for security 1021168404Spjd */ 1022219089Spjd bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1023168404Spjd 1024219089Spjd zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1025168404Spjd 1026168404Spjd /* 1027219089Spjd * If there was an allocation failure then nlwb will be null which 1028219089Spjd * forces a txg_wait_synced(). 1029168404Spjd */ 1030168404Spjd return (nlwb); 1031168404Spjd} 1032168404Spjd 1033168404Spjdstatic lwb_t * 1034168404Spjdzil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1035168404Spjd{ 1036168404Spjd lr_t *lrc = &itx->itx_lr; /* common log record */ 1037219089Spjd lr_write_t *lrw = (lr_write_t *)lrc; 1038219089Spjd char *lr_buf; 1039168404Spjd uint64_t txg = lrc->lrc_txg; 1040168404Spjd uint64_t reclen = lrc->lrc_reclen; 1041219089Spjd uint64_t dlen = 0; 1042168404Spjd 1043168404Spjd if (lwb == NULL) 1044168404Spjd return (NULL); 1045219089Spjd 1046168404Spjd ASSERT(lwb->lwb_buf != NULL); 1047243674Smm ASSERT(zilog_is_dirty(zilog) || 1048243674Smm spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1049168404Spjd 1050168404Spjd if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1051168404Spjd dlen = P2ROUNDUP_TYPED( 1052219089Spjd lrw->lr_length, sizeof (uint64_t), uint64_t); 1053168404Spjd 1054168404Spjd zilog->zl_cur_used += (reclen + dlen); 1055168404Spjd 1056168404Spjd zil_lwb_write_init(zilog, lwb); 1057168404Spjd 1058168404Spjd /* 1059168404Spjd * If this record won't fit in the current log block, start a new one. 1060168404Spjd */ 1061219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1062168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1063168404Spjd if (lwb == NULL) 1064168404Spjd return (NULL); 1065168404Spjd zil_lwb_write_init(zilog, lwb); 1066219089Spjd ASSERT(LWB_EMPTY(lwb)); 1067219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1068168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1069168404Spjd return (lwb); 1070168404Spjd } 1071168404Spjd } 1072168404Spjd 1073219089Spjd lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1074219089Spjd bcopy(lrc, lr_buf, reclen); 1075219089Spjd lrc = (lr_t *)lr_buf; 1076219089Spjd lrw = (lr_write_t *)lrc; 1077168404Spjd 1078168404Spjd /* 1079168404Spjd * If it's a write, fetch the data or get its blkptr as appropriate. 1080168404Spjd */ 1081168404Spjd if (lrc->lrc_txtype == TX_WRITE) { 1082168404Spjd if (txg > spa_freeze_txg(zilog->zl_spa)) 1083168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1084168404Spjd if (itx->itx_wr_state != WR_COPIED) { 1085168404Spjd char *dbuf; 1086168404Spjd int error; 1087168404Spjd 1088168404Spjd if (dlen) { 1089168404Spjd ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1090219089Spjd dbuf = lr_buf + reclen; 1091219089Spjd lrw->lr_common.lrc_reclen += dlen; 1092168404Spjd } else { 1093168404Spjd ASSERT(itx->itx_wr_state == WR_INDIRECT); 1094168404Spjd dbuf = NULL; 1095168404Spjd } 1096168404Spjd error = zilog->zl_get_data( 1097219089Spjd itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1098214378Smm if (error == EIO) { 1099214378Smm txg_wait_synced(zilog->zl_dmu_pool, txg); 1100214378Smm return (lwb); 1101214378Smm } 1102249643Smm if (error != 0) { 1103168404Spjd ASSERT(error == ENOENT || error == EEXIST || 1104168404Spjd error == EALREADY); 1105168404Spjd return (lwb); 1106168404Spjd } 1107168404Spjd } 1108168404Spjd } 1109168404Spjd 1110219089Spjd /* 1111219089Spjd * We're actually making an entry, so update lrc_seq to be the 1112219089Spjd * log record sequence number. Note that this is generally not 1113219089Spjd * equal to the itx sequence number because not all transactions 1114219089Spjd * are synchronous, and sometimes spa_sync() gets there first. 1115219089Spjd */ 1116219089Spjd lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1117168404Spjd lwb->lwb_nused += reclen + dlen; 1118168404Spjd lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1119219089Spjd ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1120243674Smm ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1121168404Spjd 1122168404Spjd return (lwb); 1123168404Spjd} 1124168404Spjd 1125168404Spjditx_t * 1126185029Spjdzil_itx_create(uint64_t txtype, size_t lrsize) 1127168404Spjd{ 1128168404Spjd itx_t *itx; 1129168404Spjd 1130168404Spjd lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1131168404Spjd 1132168404Spjd itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1133168404Spjd itx->itx_lr.lrc_txtype = txtype; 1134168404Spjd itx->itx_lr.lrc_reclen = lrsize; 1135185029Spjd itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1136168404Spjd itx->itx_lr.lrc_seq = 0; /* defensive */ 1137219089Spjd itx->itx_sync = B_TRUE; /* default is synchronous */ 1138168404Spjd 1139168404Spjd return (itx); 1140168404Spjd} 1141168404Spjd 1142219089Spjdvoid 1143219089Spjdzil_itx_destroy(itx_t *itx) 1144168404Spjd{ 1145219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1146219089Spjd} 1147168404Spjd 1148219089Spjd/* 1149219089Spjd * Free up the sync and async itxs. The itxs_t has already been detached 1150219089Spjd * so no locks are needed. 1151219089Spjd */ 1152219089Spjdstatic void 1153219089Spjdzil_itxg_clean(itxs_t *itxs) 1154219089Spjd{ 1155219089Spjd itx_t *itx; 1156219089Spjd list_t *list; 1157219089Spjd avl_tree_t *t; 1158219089Spjd void *cookie; 1159219089Spjd itx_async_node_t *ian; 1160168404Spjd 1161219089Spjd list = &itxs->i_sync_list; 1162219089Spjd while ((itx = list_head(list)) != NULL) { 1163219089Spjd list_remove(list, itx); 1164219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1165219089Spjd itx->itx_lr.lrc_reclen); 1166219089Spjd } 1167168404Spjd 1168219089Spjd cookie = NULL; 1169219089Spjd t = &itxs->i_async_tree; 1170219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1171219089Spjd list = &ian->ia_list; 1172219089Spjd while ((itx = list_head(list)) != NULL) { 1173219089Spjd list_remove(list, itx); 1174219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1175219089Spjd itx->itx_lr.lrc_reclen); 1176219089Spjd } 1177219089Spjd list_destroy(list); 1178219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1179219089Spjd } 1180219089Spjd avl_destroy(t); 1181219089Spjd 1182219089Spjd kmem_free(itxs, sizeof (itxs_t)); 1183168404Spjd} 1184168404Spjd 1185219089Spjdstatic int 1186219089Spjdzil_aitx_compare(const void *x1, const void *x2) 1187219089Spjd{ 1188219089Spjd const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1189219089Spjd const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1190219089Spjd 1191219089Spjd if (o1 < o2) 1192219089Spjd return (-1); 1193219089Spjd if (o1 > o2) 1194219089Spjd return (1); 1195219089Spjd 1196219089Spjd return (0); 1197219089Spjd} 1198219089Spjd 1199168404Spjd/* 1200219089Spjd * Remove all async itx with the given oid. 1201168404Spjd */ 1202168404Spjdstatic void 1203219089Spjdzil_remove_async(zilog_t *zilog, uint64_t oid) 1204168404Spjd{ 1205219089Spjd uint64_t otxg, txg; 1206219089Spjd itx_async_node_t *ian; 1207219089Spjd avl_tree_t *t; 1208219089Spjd avl_index_t where; 1209168404Spjd list_t clean_list; 1210168404Spjd itx_t *itx; 1211168404Spjd 1212219089Spjd ASSERT(oid != 0); 1213168404Spjd list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1214168404Spjd 1215219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1216219089Spjd otxg = ZILTEST_TXG; 1217219089Spjd else 1218219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1219219089Spjd 1220219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1221219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1222219089Spjd 1223219089Spjd mutex_enter(&itxg->itxg_lock); 1224219089Spjd if (itxg->itxg_txg != txg) { 1225219089Spjd mutex_exit(&itxg->itxg_lock); 1226219089Spjd continue; 1227219089Spjd } 1228219089Spjd 1229219089Spjd /* 1230219089Spjd * Locate the object node and append its list. 1231219089Spjd */ 1232219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1233219089Spjd ian = avl_find(t, &oid, &where); 1234219089Spjd if (ian != NULL) 1235219089Spjd list_move_tail(&clean_list, &ian->ia_list); 1236219089Spjd mutex_exit(&itxg->itxg_lock); 1237168404Spjd } 1238219089Spjd while ((itx = list_head(&clean_list)) != NULL) { 1239219089Spjd list_remove(&clean_list, itx); 1240219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1241219089Spjd itx->itx_lr.lrc_reclen); 1242219089Spjd } 1243219089Spjd list_destroy(&clean_list); 1244219089Spjd} 1245168404Spjd 1246219089Spjdvoid 1247219089Spjdzil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1248219089Spjd{ 1249219089Spjd uint64_t txg; 1250219089Spjd itxg_t *itxg; 1251219089Spjd itxs_t *itxs, *clean = NULL; 1252219089Spjd 1253168404Spjd /* 1254219089Spjd * Object ids can be re-instantiated in the next txg so 1255219089Spjd * remove any async transactions to avoid future leaks. 1256219089Spjd * This can happen if a fsync occurs on the re-instantiated 1257219089Spjd * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1258219089Spjd * the new file data and flushes a write record for the old object. 1259168404Spjd */ 1260219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1261219089Spjd zil_remove_async(zilog, itx->itx_oid); 1262219089Spjd 1263219089Spjd /* 1264219089Spjd * Ensure the data of a renamed file is committed before the rename. 1265219089Spjd */ 1266219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1267219089Spjd zil_async_to_sync(zilog, itx->itx_oid); 1268219089Spjd 1269243674Smm if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1270219089Spjd txg = ZILTEST_TXG; 1271219089Spjd else 1272219089Spjd txg = dmu_tx_get_txg(tx); 1273219089Spjd 1274219089Spjd itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1275219089Spjd mutex_enter(&itxg->itxg_lock); 1276219089Spjd itxs = itxg->itxg_itxs; 1277219089Spjd if (itxg->itxg_txg != txg) { 1278219089Spjd if (itxs != NULL) { 1279219089Spjd /* 1280219089Spjd * The zil_clean callback hasn't got around to cleaning 1281219089Spjd * this itxg. Save the itxs for release below. 1282219089Spjd * This should be rare. 1283219089Spjd */ 1284219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1285219089Spjd itxg->itxg_sod = 0; 1286219089Spjd clean = itxg->itxg_itxs; 1287219089Spjd } 1288219089Spjd ASSERT(itxg->itxg_sod == 0); 1289219089Spjd itxg->itxg_txg = txg; 1290219089Spjd itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1291219089Spjd 1292219089Spjd list_create(&itxs->i_sync_list, sizeof (itx_t), 1293219089Spjd offsetof(itx_t, itx_node)); 1294219089Spjd avl_create(&itxs->i_async_tree, zil_aitx_compare, 1295219089Spjd sizeof (itx_async_node_t), 1296219089Spjd offsetof(itx_async_node_t, ia_node)); 1297168404Spjd } 1298219089Spjd if (itx->itx_sync) { 1299219089Spjd list_insert_tail(&itxs->i_sync_list, itx); 1300219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1301219089Spjd itxg->itxg_sod += itx->itx_sod; 1302219089Spjd } else { 1303219089Spjd avl_tree_t *t = &itxs->i_async_tree; 1304219089Spjd uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1305219089Spjd itx_async_node_t *ian; 1306219089Spjd avl_index_t where; 1307168404Spjd 1308219089Spjd ian = avl_find(t, &foid, &where); 1309219089Spjd if (ian == NULL) { 1310219089Spjd ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1311219089Spjd list_create(&ian->ia_list, sizeof (itx_t), 1312219089Spjd offsetof(itx_t, itx_node)); 1313219089Spjd ian->ia_foid = foid; 1314219089Spjd avl_insert(t, ian, where); 1315219089Spjd } 1316219089Spjd list_insert_tail(&ian->ia_list, itx); 1317168404Spjd } 1318219089Spjd 1319219089Spjd itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1320243674Smm zilog_dirty(zilog, txg); 1321219089Spjd mutex_exit(&itxg->itxg_lock); 1322219089Spjd 1323219089Spjd /* Release the old itxs now we've dropped the lock */ 1324219089Spjd if (clean != NULL) 1325219089Spjd zil_itxg_clean(clean); 1326168404Spjd} 1327168404Spjd 1328168404Spjd/* 1329168404Spjd * If there are any in-memory intent log transactions which have now been 1330243674Smm * synced then start up a taskq to free them. We should only do this after we 1331243674Smm * have written out the uberblocks (i.e. txg has been comitted) so that 1332243674Smm * don't inadvertently clean out in-memory log records that would be required 1333243674Smm * by zil_commit(). 1334168404Spjd */ 1335168404Spjdvoid 1336219089Spjdzil_clean(zilog_t *zilog, uint64_t synced_txg) 1337168404Spjd{ 1338219089Spjd itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1339219089Spjd itxs_t *clean_me; 1340168404Spjd 1341219089Spjd mutex_enter(&itxg->itxg_lock); 1342219089Spjd if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1343219089Spjd mutex_exit(&itxg->itxg_lock); 1344219089Spjd return; 1345168404Spjd } 1346219089Spjd ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1347219089Spjd ASSERT(itxg->itxg_txg != 0); 1348219089Spjd ASSERT(zilog->zl_clean_taskq != NULL); 1349219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1350219089Spjd itxg->itxg_sod = 0; 1351219089Spjd clean_me = itxg->itxg_itxs; 1352219089Spjd itxg->itxg_itxs = NULL; 1353219089Spjd itxg->itxg_txg = 0; 1354219089Spjd mutex_exit(&itxg->itxg_lock); 1355219089Spjd /* 1356219089Spjd * Preferably start a task queue to free up the old itxs but 1357219089Spjd * if taskq_dispatch can't allocate resources to do that then 1358219089Spjd * free it in-line. This should be rare. Note, using TQ_SLEEP 1359219089Spjd * created a bad performance problem. 1360219089Spjd */ 1361219089Spjd if (taskq_dispatch(zilog->zl_clean_taskq, 1362219089Spjd (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1363219089Spjd zil_itxg_clean(clean_me); 1364168404Spjd} 1365168404Spjd 1366219089Spjd/* 1367219089Spjd * Get the list of itxs to commit into zl_itx_commit_list. 1368219089Spjd */ 1369185029Spjdstatic void 1370219089Spjdzil_get_commit_list(zilog_t *zilog) 1371168404Spjd{ 1372219089Spjd uint64_t otxg, txg; 1373219089Spjd list_t *commit_list = &zilog->zl_itx_commit_list; 1374219089Spjd uint64_t push_sod = 0; 1375219089Spjd 1376219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1377219089Spjd otxg = ZILTEST_TXG; 1378219089Spjd else 1379219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1380219089Spjd 1381219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1382219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1383219089Spjd 1384219089Spjd mutex_enter(&itxg->itxg_lock); 1385219089Spjd if (itxg->itxg_txg != txg) { 1386219089Spjd mutex_exit(&itxg->itxg_lock); 1387219089Spjd continue; 1388219089Spjd } 1389219089Spjd 1390219089Spjd list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1391219089Spjd push_sod += itxg->itxg_sod; 1392219089Spjd itxg->itxg_sod = 0; 1393219089Spjd 1394219089Spjd mutex_exit(&itxg->itxg_lock); 1395219089Spjd } 1396219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1397219089Spjd} 1398219089Spjd 1399219089Spjd/* 1400219089Spjd * Move the async itxs for a specified object to commit into sync lists. 1401219089Spjd */ 1402219089Spjdstatic void 1403219089Spjdzil_async_to_sync(zilog_t *zilog, uint64_t foid) 1404219089Spjd{ 1405219089Spjd uint64_t otxg, txg; 1406219089Spjd itx_async_node_t *ian; 1407219089Spjd avl_tree_t *t; 1408219089Spjd avl_index_t where; 1409219089Spjd 1410219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1411219089Spjd otxg = ZILTEST_TXG; 1412219089Spjd else 1413219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1414219089Spjd 1415219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1416219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1417219089Spjd 1418219089Spjd mutex_enter(&itxg->itxg_lock); 1419219089Spjd if (itxg->itxg_txg != txg) { 1420219089Spjd mutex_exit(&itxg->itxg_lock); 1421219089Spjd continue; 1422219089Spjd } 1423219089Spjd 1424219089Spjd /* 1425219089Spjd * If a foid is specified then find that node and append its 1426219089Spjd * list. Otherwise walk the tree appending all the lists 1427219089Spjd * to the sync list. We add to the end rather than the 1428219089Spjd * beginning to ensure the create has happened. 1429219089Spjd */ 1430219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1431219089Spjd if (foid != 0) { 1432219089Spjd ian = avl_find(t, &foid, &where); 1433219089Spjd if (ian != NULL) { 1434219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1435219089Spjd &ian->ia_list); 1436219089Spjd } 1437219089Spjd } else { 1438219089Spjd void *cookie = NULL; 1439219089Spjd 1440219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1441219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1442219089Spjd &ian->ia_list); 1443219089Spjd list_destroy(&ian->ia_list); 1444219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1445219089Spjd } 1446219089Spjd } 1447219089Spjd mutex_exit(&itxg->itxg_lock); 1448219089Spjd } 1449219089Spjd} 1450219089Spjd 1451219089Spjdstatic void 1452219089Spjdzil_commit_writer(zilog_t *zilog) 1453219089Spjd{ 1454168404Spjd uint64_t txg; 1455219089Spjd itx_t *itx; 1456168404Spjd lwb_t *lwb; 1457219089Spjd spa_t *spa = zilog->zl_spa; 1458219089Spjd int error = 0; 1459168404Spjd 1460185029Spjd ASSERT(zilog->zl_root_zio == NULL); 1461168404Spjd 1462219089Spjd mutex_exit(&zilog->zl_lock); 1463219089Spjd 1464219089Spjd zil_get_commit_list(zilog); 1465219089Spjd 1466219089Spjd /* 1467219089Spjd * Return if there's nothing to commit before we dirty the fs by 1468219089Spjd * calling zil_create(). 1469219089Spjd */ 1470219089Spjd if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1471219089Spjd mutex_enter(&zilog->zl_lock); 1472219089Spjd return; 1473219089Spjd } 1474219089Spjd 1475168404Spjd if (zilog->zl_suspend) { 1476168404Spjd lwb = NULL; 1477168404Spjd } else { 1478168404Spjd lwb = list_tail(&zilog->zl_lwb_list); 1479219089Spjd if (lwb == NULL) 1480219089Spjd lwb = zil_create(zilog); 1481168404Spjd } 1482168404Spjd 1483168404Spjd DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1484219089Spjd while (itx = list_head(&zilog->zl_itx_commit_list)) { 1485168404Spjd txg = itx->itx_lr.lrc_txg; 1486168404Spjd ASSERT(txg); 1487168404Spjd 1488219089Spjd if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1489168404Spjd lwb = zil_lwb_commit(zilog, itx, lwb); 1490219089Spjd list_remove(&zilog->zl_itx_commit_list, itx); 1491168404Spjd kmem_free(itx, offsetof(itx_t, itx_lr) 1492168404Spjd + itx->itx_lr.lrc_reclen); 1493168404Spjd } 1494168404Spjd DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1495168404Spjd 1496168404Spjd /* write the last block out */ 1497168404Spjd if (lwb != NULL && lwb->lwb_zio != NULL) 1498168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1499168404Spjd 1500168404Spjd zilog->zl_cur_used = 0; 1501168404Spjd 1502168404Spjd /* 1503168404Spjd * Wait if necessary for the log blocks to be on stable storage. 1504168404Spjd */ 1505168404Spjd if (zilog->zl_root_zio) { 1506219089Spjd error = zio_wait(zilog->zl_root_zio); 1507185029Spjd zilog->zl_root_zio = NULL; 1508185029Spjd zil_flush_vdevs(zilog); 1509168404Spjd } 1510168404Spjd 1511219089Spjd if (error || lwb == NULL) 1512168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 1513168404Spjd 1514168404Spjd mutex_enter(&zilog->zl_lock); 1515168404Spjd 1516219089Spjd /* 1517219089Spjd * Remember the highest committed log sequence number for ztest. 1518219089Spjd * We only update this value when all the log writes succeeded, 1519219089Spjd * because ztest wants to ASSERT that it got the whole log chain. 1520219089Spjd */ 1521219089Spjd if (error == 0 && lwb != NULL) 1522219089Spjd zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1523168404Spjd} 1524168404Spjd 1525168404Spjd/* 1526219089Spjd * Commit zfs transactions to stable storage. 1527168404Spjd * If foid is 0 push out all transactions, otherwise push only those 1528219089Spjd * for that object or might reference that object. 1529219089Spjd * 1530219089Spjd * itxs are committed in batches. In a heavily stressed zil there will be 1531219089Spjd * a commit writer thread who is writing out a bunch of itxs to the log 1532219089Spjd * for a set of committing threads (cthreads) in the same batch as the writer. 1533219089Spjd * Those cthreads are all waiting on the same cv for that batch. 1534219089Spjd * 1535219089Spjd * There will also be a different and growing batch of threads that are 1536219089Spjd * waiting to commit (qthreads). When the committing batch completes 1537219089Spjd * a transition occurs such that the cthreads exit and the qthreads become 1538219089Spjd * cthreads. One of the new cthreads becomes the writer thread for the 1539219089Spjd * batch. Any new threads arriving become new qthreads. 1540219089Spjd * 1541219089Spjd * Only 2 condition variables are needed and there's no transition 1542219089Spjd * between the two cvs needed. They just flip-flop between qthreads 1543219089Spjd * and cthreads. 1544219089Spjd * 1545219089Spjd * Using this scheme we can efficiently wakeup up only those threads 1546219089Spjd * that have been committed. 1547168404Spjd */ 1548168404Spjdvoid 1549219089Spjdzil_commit(zilog_t *zilog, uint64_t foid) 1550168404Spjd{ 1551219089Spjd uint64_t mybatch; 1552219089Spjd 1553219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1554168404Spjd return; 1555168404Spjd 1556219089Spjd /* move the async itxs for the foid to the sync queues */ 1557219089Spjd zil_async_to_sync(zilog, foid); 1558219089Spjd 1559168404Spjd mutex_enter(&zilog->zl_lock); 1560219089Spjd mybatch = zilog->zl_next_batch; 1561168404Spjd while (zilog->zl_writer) { 1562219089Spjd cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1563219089Spjd if (mybatch <= zilog->zl_com_batch) { 1564168404Spjd mutex_exit(&zilog->zl_lock); 1565168404Spjd return; 1566168404Spjd } 1567168404Spjd } 1568219089Spjd 1569219089Spjd zilog->zl_next_batch++; 1570219089Spjd zilog->zl_writer = B_TRUE; 1571219089Spjd zil_commit_writer(zilog); 1572219089Spjd zilog->zl_com_batch = mybatch; 1573219089Spjd zilog->zl_writer = B_FALSE; 1574168404Spjd mutex_exit(&zilog->zl_lock); 1575219089Spjd 1576219089Spjd /* wake up one thread to become the next writer */ 1577219089Spjd cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1578219089Spjd 1579219089Spjd /* wake up all threads waiting for this batch to be committed */ 1580219089Spjd cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1581168404Spjd} 1582168404Spjd 1583168404Spjd/* 1584168404Spjd * Called in syncing context to free committed log blocks and update log header. 1585168404Spjd */ 1586168404Spjdvoid 1587168404Spjdzil_sync(zilog_t *zilog, dmu_tx_t *tx) 1588168404Spjd{ 1589168404Spjd zil_header_t *zh = zil_header_in_syncing_context(zilog); 1590168404Spjd uint64_t txg = dmu_tx_get_txg(tx); 1591168404Spjd spa_t *spa = zilog->zl_spa; 1592219089Spjd uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1593168404Spjd lwb_t *lwb; 1594168404Spjd 1595209962Smm /* 1596209962Smm * We don't zero out zl_destroy_txg, so make sure we don't try 1597209962Smm * to destroy it twice. 1598209962Smm */ 1599209962Smm if (spa_sync_pass(spa) != 1) 1600209962Smm return; 1601209962Smm 1602168404Spjd mutex_enter(&zilog->zl_lock); 1603168404Spjd 1604168404Spjd ASSERT(zilog->zl_stop_sync == 0); 1605168404Spjd 1606219089Spjd if (*replayed_seq != 0) { 1607219089Spjd ASSERT(zh->zh_replay_seq < *replayed_seq); 1608219089Spjd zh->zh_replay_seq = *replayed_seq; 1609219089Spjd *replayed_seq = 0; 1610219089Spjd } 1611168404Spjd 1612168404Spjd if (zilog->zl_destroy_txg == txg) { 1613168404Spjd blkptr_t blk = zh->zh_log; 1614168404Spjd 1615168404Spjd ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1616168404Spjd 1617168404Spjd bzero(zh, sizeof (zil_header_t)); 1618209962Smm bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1619168404Spjd 1620168404Spjd if (zilog->zl_keep_first) { 1621168404Spjd /* 1622168404Spjd * If this block was part of log chain that couldn't 1623168404Spjd * be claimed because a device was missing during 1624168404Spjd * zil_claim(), but that device later returns, 1625168404Spjd * then this block could erroneously appear valid. 1626168404Spjd * To guard against this, assign a new GUID to the new 1627168404Spjd * log chain so it doesn't matter what blk points to. 1628168404Spjd */ 1629168404Spjd zil_init_log_chain(zilog, &blk); 1630168404Spjd zh->zh_log = blk; 1631168404Spjd } 1632168404Spjd } 1633168404Spjd 1634213197Smm while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1635168404Spjd zh->zh_log = lwb->lwb_blk; 1636168404Spjd if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1637168404Spjd break; 1638168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 1639219089Spjd zio_free_zil(spa, txg, &lwb->lwb_blk); 1640168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 1641168404Spjd 1642168404Spjd /* 1643168404Spjd * If we don't have anything left in the lwb list then 1644168404Spjd * we've had an allocation failure and we need to zero 1645168404Spjd * out the zil_header blkptr so that we don't end 1646168404Spjd * up freeing the same block twice. 1647168404Spjd */ 1648168404Spjd if (list_head(&zilog->zl_lwb_list) == NULL) 1649168404Spjd BP_ZERO(&zh->zh_log); 1650168404Spjd } 1651168404Spjd mutex_exit(&zilog->zl_lock); 1652168404Spjd} 1653168404Spjd 1654168404Spjdvoid 1655168404Spjdzil_init(void) 1656168404Spjd{ 1657168404Spjd zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1658168404Spjd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1659168404Spjd} 1660168404Spjd 1661168404Spjdvoid 1662168404Spjdzil_fini(void) 1663168404Spjd{ 1664168404Spjd kmem_cache_destroy(zil_lwb_cache); 1665168404Spjd} 1666168404Spjd 1667219089Spjdvoid 1668219089Spjdzil_set_sync(zilog_t *zilog, uint64_t sync) 1669219089Spjd{ 1670219089Spjd zilog->zl_sync = sync; 1671219089Spjd} 1672219089Spjd 1673219089Spjdvoid 1674219089Spjdzil_set_logbias(zilog_t *zilog, uint64_t logbias) 1675219089Spjd{ 1676219089Spjd zilog->zl_logbias = logbias; 1677219089Spjd} 1678219089Spjd 1679168404Spjdzilog_t * 1680168404Spjdzil_alloc(objset_t *os, zil_header_t *zh_phys) 1681168404Spjd{ 1682168404Spjd zilog_t *zilog; 1683168404Spjd 1684168404Spjd zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1685168404Spjd 1686168404Spjd zilog->zl_header = zh_phys; 1687168404Spjd zilog->zl_os = os; 1688168404Spjd zilog->zl_spa = dmu_objset_spa(os); 1689168404Spjd zilog->zl_dmu_pool = dmu_objset_pool(os); 1690168404Spjd zilog->zl_destroy_txg = TXG_INITIAL - 1; 1691219089Spjd zilog->zl_logbias = dmu_objset_logbias(os); 1692219089Spjd zilog->zl_sync = dmu_objset_syncprop(os); 1693219089Spjd zilog->zl_next_batch = 1; 1694168404Spjd 1695168404Spjd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1696168404Spjd 1697219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1698219089Spjd mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1699219089Spjd MUTEX_DEFAULT, NULL); 1700219089Spjd } 1701168404Spjd 1702168404Spjd list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1703168404Spjd offsetof(lwb_t, lwb_node)); 1704168404Spjd 1705219089Spjd list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1706219089Spjd offsetof(itx_t, itx_node)); 1707219089Spjd 1708185029Spjd mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1709168404Spjd 1710185029Spjd avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1711185029Spjd sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1712185029Spjd 1713185029Spjd cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1714185029Spjd cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1715219089Spjd cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1716219089Spjd cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1717185029Spjd 1718168404Spjd return (zilog); 1719168404Spjd} 1720168404Spjd 1721168404Spjdvoid 1722168404Spjdzil_free(zilog_t *zilog) 1723168404Spjd{ 1724168404Spjd zilog->zl_stop_sync = 1; 1725168404Spjd 1726249643Smm ASSERT0(zilog->zl_suspend); 1727249643Smm ASSERT0(zilog->zl_suspending); 1728249643Smm 1729224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1730168404Spjd list_destroy(&zilog->zl_lwb_list); 1731168404Spjd 1732185029Spjd avl_destroy(&zilog->zl_vdev_tree); 1733185029Spjd mutex_destroy(&zilog->zl_vdev_lock); 1734168404Spjd 1735219089Spjd ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1736219089Spjd list_destroy(&zilog->zl_itx_commit_list); 1737219089Spjd 1738219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1739219089Spjd /* 1740219089Spjd * It's possible for an itx to be generated that doesn't dirty 1741219089Spjd * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1742219089Spjd * callback to remove the entry. We remove those here. 1743219089Spjd * 1744219089Spjd * Also free up the ziltest itxs. 1745219089Spjd */ 1746219089Spjd if (zilog->zl_itxg[i].itxg_itxs) 1747219089Spjd zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1748219089Spjd mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1749219089Spjd } 1750219089Spjd 1751168404Spjd mutex_destroy(&zilog->zl_lock); 1752168404Spjd 1753185029Spjd cv_destroy(&zilog->zl_cv_writer); 1754185029Spjd cv_destroy(&zilog->zl_cv_suspend); 1755219089Spjd cv_destroy(&zilog->zl_cv_batch[0]); 1756219089Spjd cv_destroy(&zilog->zl_cv_batch[1]); 1757185029Spjd 1758168404Spjd kmem_free(zilog, sizeof (zilog_t)); 1759168404Spjd} 1760168404Spjd 1761168404Spjd/* 1762168404Spjd * Open an intent log. 1763168404Spjd */ 1764168404Spjdzilog_t * 1765168404Spjdzil_open(objset_t *os, zil_get_data_t *get_data) 1766168404Spjd{ 1767168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1768168404Spjd 1769224526Smm ASSERT(zilog->zl_clean_taskq == NULL); 1770224526Smm ASSERT(zilog->zl_get_data == NULL); 1771224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1772224526Smm 1773168404Spjd zilog->zl_get_data = get_data; 1774168404Spjd zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1775168404Spjd 2, 2, TASKQ_PREPOPULATE); 1776168404Spjd 1777168404Spjd return (zilog); 1778168404Spjd} 1779168404Spjd 1780168404Spjd/* 1781168404Spjd * Close an intent log. 1782168404Spjd */ 1783168404Spjdvoid 1784168404Spjdzil_close(zilog_t *zilog) 1785168404Spjd{ 1786224526Smm lwb_t *lwb; 1787219089Spjd uint64_t txg = 0; 1788219089Spjd 1789219089Spjd zil_commit(zilog, 0); /* commit all itx */ 1790219089Spjd 1791168404Spjd /* 1792219089Spjd * The lwb_max_txg for the stubby lwb will reflect the last activity 1793219089Spjd * for the zil. After a txg_wait_synced() on the txg we know all the 1794219089Spjd * callbacks have occurred that may clean the zil. Only then can we 1795219089Spjd * destroy the zl_clean_taskq. 1796168404Spjd */ 1797219089Spjd mutex_enter(&zilog->zl_lock); 1798224526Smm lwb = list_tail(&zilog->zl_lwb_list); 1799224526Smm if (lwb != NULL) 1800224526Smm txg = lwb->lwb_max_txg; 1801219089Spjd mutex_exit(&zilog->zl_lock); 1802219089Spjd if (txg) 1803168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1804243674Smm ASSERT(!zilog_is_dirty(zilog)); 1805168404Spjd 1806168404Spjd taskq_destroy(zilog->zl_clean_taskq); 1807168404Spjd zilog->zl_clean_taskq = NULL; 1808168404Spjd zilog->zl_get_data = NULL; 1809224526Smm 1810224526Smm /* 1811224526Smm * We should have only one LWB left on the list; remove it now. 1812224526Smm */ 1813224526Smm mutex_enter(&zilog->zl_lock); 1814224526Smm lwb = list_head(&zilog->zl_lwb_list); 1815224526Smm if (lwb != NULL) { 1816224526Smm ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1817224526Smm list_remove(&zilog->zl_lwb_list, lwb); 1818224526Smm zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1819224526Smm kmem_cache_free(zil_lwb_cache, lwb); 1820224526Smm } 1821224526Smm mutex_exit(&zilog->zl_lock); 1822168404Spjd} 1823168404Spjd 1824249643Smmstatic char *suspend_tag = "zil suspending"; 1825249643Smm 1826168404Spjd/* 1827168404Spjd * Suspend an intent log. While in suspended mode, we still honor 1828168404Spjd * synchronous semantics, but we rely on txg_wait_synced() to do it. 1829249643Smm * On old version pools, we suspend the log briefly when taking a 1830249643Smm * snapshot so that it will have an empty intent log. 1831249643Smm * 1832249643Smm * Long holds are not really intended to be used the way we do here -- 1833249643Smm * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1834249643Smm * could fail. Therefore we take pains to only put a long hold if it is 1835249643Smm * actually necessary. Fortunately, it will only be necessary if the 1836249643Smm * objset is currently mounted (or the ZVOL equivalent). In that case it 1837249643Smm * will already have a long hold, so we are not really making things any worse. 1838249643Smm * 1839249643Smm * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1840249643Smm * zvol_state_t), and use their mechanism to prevent their hold from being 1841249643Smm * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1842249643Smm * very little gain. 1843249643Smm * 1844249643Smm * if cookiep == NULL, this does both the suspend & resume. 1845249643Smm * Otherwise, it returns with the dataset "long held", and the cookie 1846249643Smm * should be passed into zil_resume(). 1847168404Spjd */ 1848168404Spjdint 1849249643Smmzil_suspend(const char *osname, void **cookiep) 1850168404Spjd{ 1851249643Smm objset_t *os; 1852249643Smm zilog_t *zilog; 1853249643Smm const zil_header_t *zh; 1854249643Smm int error; 1855168404Spjd 1856249643Smm error = dmu_objset_hold(osname, suspend_tag, &os); 1857249643Smm if (error != 0) 1858249643Smm return (error); 1859249643Smm zilog = dmu_objset_zil(os); 1860249643Smm 1861168404Spjd mutex_enter(&zilog->zl_lock); 1862249643Smm zh = zilog->zl_header; 1863249643Smm 1864200724Sdelphij if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1865168404Spjd mutex_exit(&zilog->zl_lock); 1866249643Smm dmu_objset_rele(os, suspend_tag); 1867249643Smm return (SET_ERROR(EBUSY)); 1868168404Spjd } 1869249643Smm 1870249643Smm /* 1871249643Smm * Don't put a long hold in the cases where we can avoid it. This 1872249643Smm * is when there is no cookie so we are doing a suspend & resume 1873249643Smm * (i.e. called from zil_vdev_offline()), and there's nothing to do 1874249643Smm * for the suspend because it's already suspended, or there's no ZIL. 1875249643Smm */ 1876249643Smm if (cookiep == NULL && !zilog->zl_suspending && 1877249643Smm (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1878249643Smm mutex_exit(&zilog->zl_lock); 1879249643Smm dmu_objset_rele(os, suspend_tag); 1880249643Smm return (0); 1881249643Smm } 1882249643Smm 1883249643Smm dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1884249643Smm dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1885249643Smm 1886249643Smm zilog->zl_suspend++; 1887249643Smm 1888249643Smm if (zilog->zl_suspend > 1) { 1889168404Spjd /* 1890249643Smm * Someone else is already suspending it. 1891168404Spjd * Just wait for them to finish. 1892168404Spjd */ 1893249643Smm 1894168404Spjd while (zilog->zl_suspending) 1895168404Spjd cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1896168404Spjd mutex_exit(&zilog->zl_lock); 1897249643Smm 1898249643Smm if (cookiep == NULL) 1899249643Smm zil_resume(os); 1900249643Smm else 1901249643Smm *cookiep = os; 1902168404Spjd return (0); 1903168404Spjd } 1904249643Smm 1905249643Smm /* 1906249643Smm * If there is no pointer to an on-disk block, this ZIL must not 1907249643Smm * be active (e.g. filesystem not mounted), so there's nothing 1908249643Smm * to clean up. 1909249643Smm */ 1910249643Smm if (BP_IS_HOLE(&zh->zh_log)) { 1911249643Smm ASSERT(cookiep != NULL); /* fast path already handled */ 1912249643Smm 1913249643Smm *cookiep = os; 1914249643Smm mutex_exit(&zilog->zl_lock); 1915249643Smm return (0); 1916249643Smm } 1917249643Smm 1918168404Spjd zilog->zl_suspending = B_TRUE; 1919168404Spjd mutex_exit(&zilog->zl_lock); 1920168404Spjd 1921219089Spjd zil_commit(zilog, 0); 1922168404Spjd 1923168404Spjd zil_destroy(zilog, B_FALSE); 1924168404Spjd 1925168404Spjd mutex_enter(&zilog->zl_lock); 1926168404Spjd zilog->zl_suspending = B_FALSE; 1927168404Spjd cv_broadcast(&zilog->zl_cv_suspend); 1928168404Spjd mutex_exit(&zilog->zl_lock); 1929168404Spjd 1930249643Smm if (cookiep == NULL) 1931249643Smm zil_resume(os); 1932249643Smm else 1933249643Smm *cookiep = os; 1934168404Spjd return (0); 1935168404Spjd} 1936168404Spjd 1937168404Spjdvoid 1938249643Smmzil_resume(void *cookie) 1939168404Spjd{ 1940249643Smm objset_t *os = cookie; 1941249643Smm zilog_t *zilog = dmu_objset_zil(os); 1942249643Smm 1943168404Spjd mutex_enter(&zilog->zl_lock); 1944168404Spjd ASSERT(zilog->zl_suspend != 0); 1945168404Spjd zilog->zl_suspend--; 1946168404Spjd mutex_exit(&zilog->zl_lock); 1947249643Smm dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 1948249643Smm dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 1949168404Spjd} 1950168404Spjd 1951219089Spjdtypedef struct zil_replay_arg { 1952219089Spjd zil_replay_func_t **zr_replay; 1953219089Spjd void *zr_arg; 1954219089Spjd boolean_t zr_byteswap; 1955219089Spjd char *zr_lr; 1956219089Spjd} zil_replay_arg_t; 1957219089Spjd 1958219089Spjdstatic int 1959219089Spjdzil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1960209962Smm{ 1961219089Spjd char name[MAXNAMELEN]; 1962209962Smm 1963219089Spjd zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1964209962Smm 1965219089Spjd dmu_objset_name(zilog->zl_os, name); 1966209962Smm 1967219089Spjd cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1968219089Spjd "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1969219089Spjd (u_longlong_t)lr->lrc_seq, 1970219089Spjd (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1971219089Spjd (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1972219089Spjd 1973219089Spjd return (error); 1974209962Smm} 1975209962Smm 1976219089Spjdstatic int 1977168404Spjdzil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1978168404Spjd{ 1979168404Spjd zil_replay_arg_t *zr = zra; 1980168404Spjd const zil_header_t *zh = zilog->zl_header; 1981168404Spjd uint64_t reclen = lr->lrc_reclen; 1982168404Spjd uint64_t txtype = lr->lrc_txtype; 1983219089Spjd int error = 0; 1984168404Spjd 1985219089Spjd zilog->zl_replaying_seq = lr->lrc_seq; 1986168404Spjd 1987219089Spjd if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1988219089Spjd return (0); 1989219089Spjd 1990168404Spjd if (lr->lrc_txg < claim_txg) /* already committed */ 1991219089Spjd return (0); 1992168404Spjd 1993185029Spjd /* Strip case-insensitive bit, still present in log record */ 1994185029Spjd txtype &= ~TX_CI; 1995185029Spjd 1996219089Spjd if (txtype == 0 || txtype >= TX_MAX_TYPE) 1997219089Spjd return (zil_replay_error(zilog, lr, EINVAL)); 1998219089Spjd 1999219089Spjd /* 2000219089Spjd * If this record type can be logged out of order, the object 2001219089Spjd * (lr_foid) may no longer exist. That's legitimate, not an error. 2002219089Spjd */ 2003219089Spjd if (TX_OOO(txtype)) { 2004219089Spjd error = dmu_object_info(zilog->zl_os, 2005219089Spjd ((lr_ooo_t *)lr)->lr_foid, NULL); 2006219089Spjd if (error == ENOENT || error == EEXIST) 2007219089Spjd return (0); 2008209962Smm } 2009209962Smm 2010168404Spjd /* 2011168404Spjd * Make a copy of the data so we can revise and extend it. 2012168404Spjd */ 2013219089Spjd bcopy(lr, zr->zr_lr, reclen); 2014168404Spjd 2015168404Spjd /* 2016219089Spjd * If this is a TX_WRITE with a blkptr, suck in the data. 2017219089Spjd */ 2018219089Spjd if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2019219089Spjd error = zil_read_log_data(zilog, (lr_write_t *)lr, 2020219089Spjd zr->zr_lr + reclen); 2021249643Smm if (error != 0) 2022219089Spjd return (zil_replay_error(zilog, lr, error)); 2023219089Spjd } 2024219089Spjd 2025219089Spjd /* 2026168404Spjd * The log block containing this lr may have been byteswapped 2027168404Spjd * so that we can easily examine common fields like lrc_txtype. 2028219089Spjd * However, the log is a mix of different record types, and only the 2029168404Spjd * replay vectors know how to byteswap their records. Therefore, if 2030168404Spjd * the lr was byteswapped, undo it before invoking the replay vector. 2031168404Spjd */ 2032168404Spjd if (zr->zr_byteswap) 2033219089Spjd byteswap_uint64_array(zr->zr_lr, reclen); 2034168404Spjd 2035168404Spjd /* 2036168404Spjd * We must now do two things atomically: replay this log record, 2037209962Smm * and update the log header sequence number to reflect the fact that 2038209962Smm * we did so. At the end of each replay function the sequence number 2039209962Smm * is updated if we are in replay mode. 2040168404Spjd */ 2041219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2042249643Smm if (error != 0) { 2043168404Spjd /* 2044168404Spjd * The DMU's dnode layer doesn't see removes until the txg 2045168404Spjd * commits, so a subsequent claim can spuriously fail with 2046209962Smm * EEXIST. So if we receive any error we try syncing out 2047219089Spjd * any removes then retry the transaction. Note that we 2048219089Spjd * specify B_FALSE for byteswap now, so we don't do it twice. 2049168404Spjd */ 2050219089Spjd txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2051219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2052249643Smm if (error != 0) 2053219089Spjd return (zil_replay_error(zilog, lr, error)); 2054168404Spjd } 2055219089Spjd return (0); 2056168404Spjd} 2057168404Spjd 2058168404Spjd/* ARGSUSED */ 2059219089Spjdstatic int 2060168404Spjdzil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2061168404Spjd{ 2062168404Spjd zilog->zl_replay_blks++; 2063219089Spjd 2064219089Spjd return (0); 2065168404Spjd} 2066168404Spjd 2067168404Spjd/* 2068168404Spjd * If this dataset has a non-empty intent log, replay it and destroy it. 2069168404Spjd */ 2070168404Spjdvoid 2071209962Smmzil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2072168404Spjd{ 2073168404Spjd zilog_t *zilog = dmu_objset_zil(os); 2074168404Spjd const zil_header_t *zh = zilog->zl_header; 2075168404Spjd zil_replay_arg_t zr; 2076168404Spjd 2077200724Sdelphij if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2078168404Spjd zil_destroy(zilog, B_TRUE); 2079168404Spjd return; 2080168404Spjd } 2081168404Spjd //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 2082168404Spjd 2083168404Spjd zr.zr_replay = replay_func; 2084168404Spjd zr.zr_arg = arg; 2085168404Spjd zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2086219089Spjd zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2087168404Spjd 2088168404Spjd /* 2089168404Spjd * Wait for in-progress removes to sync before starting replay. 2090168404Spjd */ 2091168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 2092168404Spjd 2093209962Smm zilog->zl_replay = B_TRUE; 2094219089Spjd zilog->zl_replay_time = ddi_get_lbolt(); 2095168404Spjd ASSERT(zilog->zl_replay_blks == 0); 2096168404Spjd (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2097168404Spjd zh->zh_claim_txg); 2098219089Spjd kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2099168404Spjd 2100168404Spjd zil_destroy(zilog, B_FALSE); 2101185029Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2102209962Smm zilog->zl_replay = B_FALSE; 2103168404Spjd //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2104168404Spjd} 2105168404Spjd 2106219089Spjdboolean_t 2107219089Spjdzil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2108168404Spjd{ 2109219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2110219089Spjd return (B_TRUE); 2111168404Spjd 2112219089Spjd if (zilog->zl_replay) { 2113219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2114219089Spjd zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2115219089Spjd zilog->zl_replaying_seq; 2116219089Spjd return (B_TRUE); 2117168404Spjd } 2118168404Spjd 2119219089Spjd return (B_FALSE); 2120168404Spjd} 2121213197Smm 2122213197Smm/* ARGSUSED */ 2123213197Smmint 2124219089Spjdzil_vdev_offline(const char *osname, void *arg) 2125213197Smm{ 2126213197Smm int error; 2127213197Smm 2128249643Smm error = zil_suspend(osname, NULL); 2129249643Smm if (error != 0) 2130249643Smm return (SET_ERROR(EEXIST)); 2131249643Smm return (0); 2132213197Smm} 2133