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. 23310516Savg * Copyright (c) 2011, 2016 by Delphix. All rights reserved. 24297112Smav * Copyright (c) 2014 Integros [integros.com] 25168404Spjd */ 26168404Spjd 27219089Spjd/* Portions Copyright 2010 Robert Milkowski */ 28219089Spjd 29168404Spjd#include <sys/zfs_context.h> 30168404Spjd#include <sys/spa.h> 31168404Spjd#include <sys/dmu.h> 32168404Spjd#include <sys/zap.h> 33168404Spjd#include <sys/arc.h> 34168404Spjd#include <sys/stat.h> 35168404Spjd#include <sys/resource.h> 36168404Spjd#include <sys/zil.h> 37168404Spjd#include <sys/zil_impl.h> 38168404Spjd#include <sys/dsl_dataset.h> 39219089Spjd#include <sys/vdev_impl.h> 40168404Spjd#include <sys/dmu_tx.h> 41219089Spjd#include <sys/dsl_pool.h> 42168404Spjd 43168404Spjd/* 44168404Spjd * The zfs intent log (ZIL) saves transaction records of system calls 45168404Spjd * that change the file system in memory with enough information 46168404Spjd * to be able to replay them. These are stored in memory until 47168404Spjd * either the DMU transaction group (txg) commits them to the stable pool 48168404Spjd * and they can be discarded, or they are flushed to the stable log 49168404Spjd * (also in the pool) due to a fsync, O_DSYNC or other synchronous 50168404Spjd * requirement. In the event of a panic or power fail then those log 51168404Spjd * records (transactions) are replayed. 52168404Spjd * 53168404Spjd * There is one ZIL per file system. Its on-disk (pool) format consists 54168404Spjd * of 3 parts: 55168404Spjd * 56168404Spjd * - ZIL header 57168404Spjd * - ZIL blocks 58168404Spjd * - ZIL records 59168404Spjd * 60168404Spjd * A log record holds a system call transaction. Log blocks can 61168404Spjd * hold many log records and the blocks are chained together. 62168404Spjd * Each ZIL block contains a block pointer (blkptr_t) to the next 63168404Spjd * ZIL block in the chain. The ZIL header points to the first 64168404Spjd * block in the chain. Note there is not a fixed place in the pool 65168404Spjd * to hold blocks. They are dynamically allocated and freed as 66168404Spjd * needed from the blocks available. Figure X shows the ZIL structure: 67168404Spjd */ 68168404Spjd 69168404Spjd/* 70251631Sdelphij * Disable intent logging replay. This global ZIL switch affects all pools. 71168404Spjd */ 72251631Sdelphijint zil_replay_disable = 0; 73168404SpjdSYSCTL_DECL(_vfs_zfs); 74219089SpjdTUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 75219089SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 76219089Spjd &zil_replay_disable, 0, "Disable intent logging replay"); 77168404Spjd 78168404Spjd/* 79168404Spjd * Tunable parameter for debugging or performance analysis. Setting 80168404Spjd * zfs_nocacheflush will cause corruption on power loss if a volatile 81168404Spjd * out-of-order write cache is enabled. 82168404Spjd */ 83168404Spjdboolean_t zfs_nocacheflush = B_FALSE; 84168404SpjdTUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 85168404SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 86168404Spjd &zfs_nocacheflush, 0, "Disable cache flush"); 87249921Ssmhboolean_t zfs_trim_enabled = B_TRUE; 88249921SsmhSYSCTL_DECL(_vfs_zfs_trim); 89249921SsmhTUNABLE_INT("vfs.zfs.trim.enabled", &zfs_trim_enabled); 90249921SsmhSYSCTL_INT(_vfs_zfs_trim, OID_AUTO, enabled, CTLFLAG_RDTUN, &zfs_trim_enabled, 0, 91249921Ssmh "Enable ZFS TRIM"); 92168404Spjd 93320496Savg/* 94320496Savg * Limit SLOG write size per commit executed with synchronous priority. 95320496Savg * Any writes above that executed with lower (asynchronous) priority to 96320496Savg * limit potential SLOG device abuse by single active ZIL writer. 97320496Savg */ 98320496Savguint64_t zil_slog_limit = 768 * 1024; 99320496SavgSYSCTL_QUAD(_vfs_zfs, OID_AUTO, zil_slog_limit, CTLFLAG_RWTUN, 100320496Savg &zil_slog_limit, 0, "Maximal SLOG commit size with sync priority"); 101320496Savg 102168404Spjdstatic kmem_cache_t *zil_lwb_cache; 103168404Spjd 104219089Spjd#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 105219089Spjd sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 106219089Spjd 107219089Spjd 108219089Spjd/* 109219089Spjd * ziltest is by and large an ugly hack, but very useful in 110219089Spjd * checking replay without tedious work. 111219089Spjd * When running ziltest we want to keep all itx's and so maintain 112219089Spjd * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 113219089Spjd * We subtract TXG_CONCURRENT_STATES to allow for common code. 114219089Spjd */ 115219089Spjd#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 116219089Spjd 117168404Spjdstatic int 118219089Spjdzil_bp_compare(const void *x1, const void *x2) 119168404Spjd{ 120219089Spjd const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 121219089Spjd const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 122168404Spjd 123168404Spjd if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 124168404Spjd return (-1); 125168404Spjd if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 126168404Spjd return (1); 127168404Spjd 128168404Spjd if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 129168404Spjd return (-1); 130168404Spjd if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 131168404Spjd return (1); 132168404Spjd 133168404Spjd return (0); 134168404Spjd} 135168404Spjd 136168404Spjdstatic void 137219089Spjdzil_bp_tree_init(zilog_t *zilog) 138168404Spjd{ 139219089Spjd avl_create(&zilog->zl_bp_tree, zil_bp_compare, 140219089Spjd sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 141168404Spjd} 142168404Spjd 143168404Spjdstatic void 144219089Spjdzil_bp_tree_fini(zilog_t *zilog) 145168404Spjd{ 146219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 147219089Spjd zil_bp_node_t *zn; 148168404Spjd void *cookie = NULL; 149168404Spjd 150168404Spjd while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 151219089Spjd kmem_free(zn, sizeof (zil_bp_node_t)); 152168404Spjd 153168404Spjd avl_destroy(t); 154168404Spjd} 155168404Spjd 156219089Spjdint 157219089Spjdzil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 158168404Spjd{ 159219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 160268649Sdelphij const dva_t *dva; 161219089Spjd zil_bp_node_t *zn; 162168404Spjd avl_index_t where; 163168404Spjd 164268649Sdelphij if (BP_IS_EMBEDDED(bp)) 165268649Sdelphij return (0); 166268649Sdelphij 167268649Sdelphij dva = BP_IDENTITY(bp); 168268649Sdelphij 169168404Spjd if (avl_find(t, dva, &where) != NULL) 170249195Smm return (SET_ERROR(EEXIST)); 171168404Spjd 172219089Spjd zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 173168404Spjd zn->zn_dva = *dva; 174168404Spjd avl_insert(t, zn, where); 175168404Spjd 176168404Spjd return (0); 177168404Spjd} 178168404Spjd 179168404Spjdstatic zil_header_t * 180168404Spjdzil_header_in_syncing_context(zilog_t *zilog) 181168404Spjd{ 182168404Spjd return ((zil_header_t *)zilog->zl_header); 183168404Spjd} 184168404Spjd 185168404Spjdstatic void 186168404Spjdzil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 187168404Spjd{ 188168404Spjd zio_cksum_t *zc = &bp->blk_cksum; 189168404Spjd 190168404Spjd zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 191168404Spjd zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 192168404Spjd zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 193168404Spjd zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 194168404Spjd} 195168404Spjd 196168404Spjd/* 197219089Spjd * Read a log block and make sure it's valid. 198168404Spjd */ 199168404Spjdstatic int 200219089Spjdzil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 201219089Spjd char **end) 202168404Spjd{ 203219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 204277586Sdelphij arc_flags_t aflags = ARC_FLAG_WAIT; 205219089Spjd arc_buf_t *abuf = NULL; 206268657Sdelphij zbookmark_phys_t zb; 207168404Spjd int error; 208168404Spjd 209219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 210219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 211168404Spjd 212219089Spjd if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 213219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE; 214168404Spjd 215219089Spjd SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 216219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 217168404Spjd 218246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 219219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 220219089Spjd 221168404Spjd if (error == 0) { 222168404Spjd zio_cksum_t cksum = bp->blk_cksum; 223168404Spjd 224168404Spjd /* 225185029Spjd * Validate the checksummed log block. 226185029Spjd * 227168404Spjd * Sequence numbers should be... sequential. The checksum 228168404Spjd * verifier for the next block should be bp's checksum plus 1. 229185029Spjd * 230185029Spjd * Also check the log chain linkage and size used. 231168404Spjd */ 232168404Spjd cksum.zc_word[ZIL_ZC_SEQ]++; 233168404Spjd 234219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 235219089Spjd zil_chain_t *zilc = abuf->b_data; 236219089Spjd char *lr = (char *)(zilc + 1); 237219089Spjd uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 238219089Spjd 239219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 240219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 241249195Smm error = SET_ERROR(ECKSUM); 242219089Spjd } else { 243276081Sdelphij ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE); 244219089Spjd bcopy(lr, dst, len); 245219089Spjd *end = (char *)dst + len; 246219089Spjd *nbp = zilc->zc_next_blk; 247219089Spjd } 248219089Spjd } else { 249219089Spjd char *lr = abuf->b_data; 250219089Spjd uint64_t size = BP_GET_LSIZE(bp); 251219089Spjd zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 252219089Spjd 253219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 254219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 255219089Spjd (zilc->zc_nused > (size - sizeof (*zilc)))) { 256249195Smm error = SET_ERROR(ECKSUM); 257219089Spjd } else { 258276081Sdelphij ASSERT3U(zilc->zc_nused, <=, 259276081Sdelphij SPA_OLD_MAXBLOCKSIZE); 260219089Spjd bcopy(lr, dst, zilc->zc_nused); 261219089Spjd *end = (char *)dst + zilc->zc_nused; 262219089Spjd *nbp = zilc->zc_next_blk; 263219089Spjd } 264185029Spjd } 265168404Spjd 266307266Smav arc_buf_destroy(abuf, &abuf); 267168404Spjd } 268168404Spjd 269219089Spjd return (error); 270219089Spjd} 271168404Spjd 272219089Spjd/* 273219089Spjd * Read a TX_WRITE log data block. 274219089Spjd */ 275219089Spjdstatic int 276219089Spjdzil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 277219089Spjd{ 278219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 279219089Spjd const blkptr_t *bp = &lr->lr_blkptr; 280277586Sdelphij arc_flags_t aflags = ARC_FLAG_WAIT; 281219089Spjd arc_buf_t *abuf = NULL; 282268657Sdelphij zbookmark_phys_t zb; 283219089Spjd int error; 284219089Spjd 285219089Spjd if (BP_IS_HOLE(bp)) { 286219089Spjd if (wbuf != NULL) 287219089Spjd bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 288219089Spjd return (0); 289219089Spjd } 290219089Spjd 291219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 292219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 293219089Spjd 294219089Spjd SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 295219089Spjd ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 296219089Spjd 297246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 298219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 299219089Spjd 300219089Spjd if (error == 0) { 301219089Spjd if (wbuf != NULL) 302219089Spjd bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 303307266Smav arc_buf_destroy(abuf, &abuf); 304219089Spjd } 305219089Spjd 306168404Spjd return (error); 307168404Spjd} 308168404Spjd 309168404Spjd/* 310168404Spjd * Parse the intent log, and call parse_func for each valid record within. 311168404Spjd */ 312219089Spjdint 313168404Spjdzil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 314168404Spjd zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 315168404Spjd{ 316168404Spjd const zil_header_t *zh = zilog->zl_header; 317219089Spjd boolean_t claimed = !!zh->zh_claim_txg; 318219089Spjd uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 319219089Spjd uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 320219089Spjd uint64_t max_blk_seq = 0; 321219089Spjd uint64_t max_lr_seq = 0; 322219089Spjd uint64_t blk_count = 0; 323219089Spjd uint64_t lr_count = 0; 324219089Spjd blkptr_t blk, next_blk; 325168404Spjd char *lrbuf, *lrp; 326219089Spjd int error = 0; 327168404Spjd 328219089Spjd /* 329219089Spjd * Old logs didn't record the maximum zh_claim_lr_seq. 330219089Spjd */ 331219089Spjd if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 332219089Spjd claim_lr_seq = UINT64_MAX; 333168404Spjd 334168404Spjd /* 335168404Spjd * Starting at the block pointed to by zh_log we read the log chain. 336168404Spjd * For each block in the chain we strongly check that block to 337168404Spjd * ensure its validity. We stop when an invalid block is found. 338168404Spjd * For each block pointer in the chain we call parse_blk_func(). 339168404Spjd * For each record in each valid block we call parse_lr_func(). 340168404Spjd * If the log has been claimed, stop if we encounter a sequence 341168404Spjd * number greater than the highest claimed sequence number. 342168404Spjd */ 343276081Sdelphij lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE); 344219089Spjd zil_bp_tree_init(zilog); 345168404Spjd 346219089Spjd for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 347219089Spjd uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 348219089Spjd int reclen; 349219089Spjd char *end; 350219089Spjd 351219089Spjd if (blk_seq > claim_blk_seq) 352168404Spjd break; 353219089Spjd if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 354219089Spjd break; 355219089Spjd ASSERT3U(max_blk_seq, <, blk_seq); 356219089Spjd max_blk_seq = blk_seq; 357219089Spjd blk_count++; 358168404Spjd 359219089Spjd if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 360219089Spjd break; 361168404Spjd 362219089Spjd error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 363248571Smm if (error != 0) 364168404Spjd break; 365168404Spjd 366219089Spjd for (lrp = lrbuf; lrp < end; lrp += reclen) { 367168404Spjd lr_t *lr = (lr_t *)lrp; 368168404Spjd reclen = lr->lrc_reclen; 369168404Spjd ASSERT3U(reclen, >=, sizeof (lr_t)); 370219089Spjd if (lr->lrc_seq > claim_lr_seq) 371219089Spjd goto done; 372219089Spjd if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 373219089Spjd goto done; 374219089Spjd ASSERT3U(max_lr_seq, <, lr->lrc_seq); 375219089Spjd max_lr_seq = lr->lrc_seq; 376219089Spjd lr_count++; 377168404Spjd } 378168404Spjd } 379219089Spjddone: 380219089Spjd zilog->zl_parse_error = error; 381219089Spjd zilog->zl_parse_blk_seq = max_blk_seq; 382219089Spjd zilog->zl_parse_lr_seq = max_lr_seq; 383219089Spjd zilog->zl_parse_blk_count = blk_count; 384219089Spjd zilog->zl_parse_lr_count = lr_count; 385168404Spjd 386219089Spjd ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 387219089Spjd (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 388219089Spjd 389219089Spjd zil_bp_tree_fini(zilog); 390276081Sdelphij zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE); 391219089Spjd 392219089Spjd return (error); 393168404Spjd} 394168404Spjd 395219089Spjdstatic int 396168404Spjdzil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 397168404Spjd{ 398168404Spjd /* 399168404Spjd * Claim log block if not already committed and not already claimed. 400219089Spjd * If tx == NULL, just verify that the block is claimable. 401168404Spjd */ 402263397Sdelphij if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg || 403263397Sdelphij zil_bp_tree_add(zilog, bp) != 0) 404219089Spjd return (0); 405219089Spjd 406219089Spjd return (zio_wait(zio_claim(NULL, zilog->zl_spa, 407219089Spjd tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 408219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 409168404Spjd} 410168404Spjd 411219089Spjdstatic int 412168404Spjdzil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 413168404Spjd{ 414219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 415219089Spjd int error; 416219089Spjd 417219089Spjd if (lrc->lrc_txtype != TX_WRITE) 418219089Spjd return (0); 419219089Spjd 420219089Spjd /* 421219089Spjd * If the block is not readable, don't claim it. This can happen 422219089Spjd * in normal operation when a log block is written to disk before 423219089Spjd * some of the dmu_sync() blocks it points to. In this case, the 424219089Spjd * transaction cannot have been committed to anyone (we would have 425219089Spjd * waited for all writes to be stable first), so it is semantically 426219089Spjd * correct to declare this the end of the log. 427219089Spjd */ 428219089Spjd if (lr->lr_blkptr.blk_birth >= first_txg && 429219089Spjd (error = zil_read_log_data(zilog, lr, NULL)) != 0) 430219089Spjd return (error); 431219089Spjd return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 432168404Spjd} 433168404Spjd 434168404Spjd/* ARGSUSED */ 435219089Spjdstatic int 436168404Spjdzil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 437168404Spjd{ 438219089Spjd zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 439219089Spjd 440219089Spjd return (0); 441168404Spjd} 442168404Spjd 443219089Spjdstatic int 444168404Spjdzil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 445168404Spjd{ 446219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 447219089Spjd blkptr_t *bp = &lr->lr_blkptr; 448219089Spjd 449168404Spjd /* 450168404Spjd * If we previously claimed it, we need to free it. 451168404Spjd */ 452219089Spjd if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 453263397Sdelphij bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 && 454263397Sdelphij !BP_IS_HOLE(bp)) 455219089Spjd zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 456219089Spjd 457219089Spjd return (0); 458219089Spjd} 459219089Spjd 460219089Spjdstatic lwb_t * 461320496Savgzil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg) 462219089Spjd{ 463219089Spjd lwb_t *lwb; 464219089Spjd 465219089Spjd lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 466219089Spjd lwb->lwb_zilog = zilog; 467219089Spjd lwb->lwb_blk = *bp; 468320496Savg lwb->lwb_slog = slog; 469219089Spjd lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 470219089Spjd lwb->lwb_max_txg = txg; 471219089Spjd lwb->lwb_zio = NULL; 472219089Spjd lwb->lwb_tx = NULL; 473219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 474219089Spjd lwb->lwb_nused = sizeof (zil_chain_t); 475219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp); 476219089Spjd } else { 477219089Spjd lwb->lwb_nused = 0; 478219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 479168404Spjd } 480219089Spjd 481219089Spjd mutex_enter(&zilog->zl_lock); 482219089Spjd list_insert_tail(&zilog->zl_lwb_list, lwb); 483219089Spjd mutex_exit(&zilog->zl_lock); 484219089Spjd 485219089Spjd return (lwb); 486168404Spjd} 487168404Spjd 488168404Spjd/* 489239620Smm * Called when we create in-memory log transactions so that we know 490239620Smm * to cleanup the itxs at the end of spa_sync(). 491239620Smm */ 492239620Smmvoid 493239620Smmzilog_dirty(zilog_t *zilog, uint64_t txg) 494239620Smm{ 495239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 496239620Smm dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 497239620Smm 498288549Smav if (ds->ds_is_snapshot) 499239620Smm panic("dirtying snapshot!"); 500239620Smm 501248571Smm if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) { 502239620Smm /* up the hold count until we can be written out */ 503239620Smm dmu_buf_add_ref(ds->ds_dbuf, zilog); 504239620Smm } 505239620Smm} 506239620Smm 507310516Savg/* 508310516Savg * Determine if the zil is dirty in the specified txg. Callers wanting to 509310516Savg * ensure that the dirty state does not change must hold the itxg_lock for 510310516Savg * the specified txg. Holding the lock will ensure that the zil cannot be 511310516Savg * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current 512310516Savg * state. 513310516Savg */ 514239620Smmboolean_t 515310516Savgzilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg) 516310516Savg{ 517310516Savg dsl_pool_t *dp = zilog->zl_dmu_pool; 518310516Savg 519310516Savg if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK)) 520310516Savg return (B_TRUE); 521310516Savg return (B_FALSE); 522310516Savg} 523310516Savg 524310516Savg/* 525310516Savg * Determine if the zil is dirty. The zil is considered dirty if it has 526310516Savg * any pending itx records that have not been cleaned by zil_clean(). 527310516Savg */ 528310516Savgboolean_t 529239620Smmzilog_is_dirty(zilog_t *zilog) 530239620Smm{ 531239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 532239620Smm 533239620Smm for (int t = 0; t < TXG_SIZE; t++) { 534239620Smm if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 535239620Smm return (B_TRUE); 536239620Smm } 537239620Smm return (B_FALSE); 538239620Smm} 539239620Smm 540239620Smm/* 541168404Spjd * Create an on-disk intent log. 542168404Spjd */ 543219089Spjdstatic lwb_t * 544168404Spjdzil_create(zilog_t *zilog) 545168404Spjd{ 546168404Spjd const zil_header_t *zh = zilog->zl_header; 547219089Spjd lwb_t *lwb = NULL; 548168404Spjd uint64_t txg = 0; 549168404Spjd dmu_tx_t *tx = NULL; 550168404Spjd blkptr_t blk; 551168404Spjd int error = 0; 552320496Savg boolean_t slog = FALSE; 553168404Spjd 554168404Spjd /* 555168404Spjd * Wait for any previous destroy to complete. 556168404Spjd */ 557168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 558168404Spjd 559168404Spjd ASSERT(zh->zh_claim_txg == 0); 560168404Spjd ASSERT(zh->zh_replay_seq == 0); 561168404Spjd 562168404Spjd blk = zh->zh_log; 563168404Spjd 564168404Spjd /* 565219089Spjd * Allocate an initial log block if: 566219089Spjd * - there isn't one already 567219089Spjd * - the existing block is the wrong endianess 568168404Spjd */ 569207908Smm if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 570168404Spjd tx = dmu_tx_create(zilog->zl_os); 571219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 572168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 573168404Spjd txg = dmu_tx_get_txg(tx); 574168404Spjd 575207908Smm if (!BP_IS_HOLE(&blk)) { 576219089Spjd zio_free_zil(zilog->zl_spa, txg, &blk); 577207908Smm BP_ZERO(&blk); 578207908Smm } 579207908Smm 580219089Spjd error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 581320496Savg ZIL_MIN_BLKSZ, &slog); 582168404Spjd 583168404Spjd if (error == 0) 584168404Spjd zil_init_log_chain(zilog, &blk); 585168404Spjd } 586168404Spjd 587168404Spjd /* 588168404Spjd * Allocate a log write buffer (lwb) for the first log block. 589168404Spjd */ 590219089Spjd if (error == 0) 591320496Savg lwb = zil_alloc_lwb(zilog, &blk, slog, txg); 592168404Spjd 593168404Spjd /* 594168404Spjd * If we just allocated the first log block, commit our transaction 595168404Spjd * and wait for zil_sync() to stuff the block poiner into zh_log. 596168404Spjd * (zh is part of the MOS, so we cannot modify it in open context.) 597168404Spjd */ 598168404Spjd if (tx != NULL) { 599168404Spjd dmu_tx_commit(tx); 600168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 601168404Spjd } 602168404Spjd 603168404Spjd ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 604219089Spjd 605219089Spjd return (lwb); 606168404Spjd} 607168404Spjd 608168404Spjd/* 609168404Spjd * In one tx, free all log blocks and clear the log header. 610168404Spjd * If keep_first is set, then we're replaying a log with no content. 611168404Spjd * We want to keep the first block, however, so that the first 612168404Spjd * synchronous transaction doesn't require a txg_wait_synced() 613168404Spjd * in zil_create(). We don't need to txg_wait_synced() here either 614168404Spjd * when keep_first is set, because both zil_create() and zil_destroy() 615168404Spjd * will wait for any in-progress destroys to complete. 616168404Spjd */ 617168404Spjdvoid 618168404Spjdzil_destroy(zilog_t *zilog, boolean_t keep_first) 619168404Spjd{ 620168404Spjd const zil_header_t *zh = zilog->zl_header; 621168404Spjd lwb_t *lwb; 622168404Spjd dmu_tx_t *tx; 623168404Spjd uint64_t txg; 624168404Spjd 625168404Spjd /* 626168404Spjd * Wait for any previous destroy to complete. 627168404Spjd */ 628168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 629168404Spjd 630219089Spjd zilog->zl_old_header = *zh; /* debugging aid */ 631219089Spjd 632168404Spjd if (BP_IS_HOLE(&zh->zh_log)) 633168404Spjd return; 634168404Spjd 635168404Spjd tx = dmu_tx_create(zilog->zl_os); 636219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 637168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 638168404Spjd txg = dmu_tx_get_txg(tx); 639168404Spjd 640168404Spjd mutex_enter(&zilog->zl_lock); 641168404Spjd 642168404Spjd ASSERT3U(zilog->zl_destroy_txg, <, txg); 643168404Spjd zilog->zl_destroy_txg = txg; 644168404Spjd zilog->zl_keep_first = keep_first; 645168404Spjd 646168404Spjd if (!list_is_empty(&zilog->zl_lwb_list)) { 647168404Spjd ASSERT(zh->zh_claim_txg == 0); 648224526Smm VERIFY(!keep_first); 649168404Spjd while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 650168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 651168404Spjd if (lwb->lwb_buf != NULL) 652168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 653219089Spjd zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 654168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 655168404Spjd } 656219089Spjd } else if (!keep_first) { 657239620Smm zil_destroy_sync(zilog, tx); 658168404Spjd } 659168404Spjd mutex_exit(&zilog->zl_lock); 660168404Spjd 661168404Spjd dmu_tx_commit(tx); 662185029Spjd} 663168404Spjd 664239620Smmvoid 665239620Smmzil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 666239620Smm{ 667239620Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 668239620Smm (void) zil_parse(zilog, zil_free_log_block, 669239620Smm zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 670239620Smm} 671239620Smm 672168404Spjdint 673288569Smavzil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg) 674168404Spjd{ 675168404Spjd dmu_tx_t *tx = txarg; 676168404Spjd uint64_t first_txg = dmu_tx_get_txg(tx); 677168404Spjd zilog_t *zilog; 678168404Spjd zil_header_t *zh; 679168404Spjd objset_t *os; 680168404Spjd int error; 681168404Spjd 682288569Smav error = dmu_objset_own_obj(dp, ds->ds_object, 683288569Smav DMU_OST_ANY, B_FALSE, FTAG, &os); 684248571Smm if (error != 0) { 685272133Sdelphij /* 686272133Sdelphij * EBUSY indicates that the objset is inconsistent, in which 687272133Sdelphij * case it can not have a ZIL. 688272133Sdelphij */ 689272133Sdelphij if (error != EBUSY) { 690288569Smav cmn_err(CE_WARN, "can't open objset for %llu, error %u", 691288569Smav (unsigned long long)ds->ds_object, error); 692272133Sdelphij } 693168404Spjd return (0); 694168404Spjd } 695168404Spjd 696168404Spjd zilog = dmu_objset_zil(os); 697168404Spjd zh = zil_header_in_syncing_context(zilog); 698168404Spjd 699219089Spjd if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 700213197Smm if (!BP_IS_HOLE(&zh->zh_log)) 701219089Spjd zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 702213197Smm BP_ZERO(&zh->zh_log); 703213197Smm dsl_dataset_dirty(dmu_objset_ds(os), tx); 704248571Smm dmu_objset_disown(os, FTAG); 705219089Spjd return (0); 706213197Smm } 707213197Smm 708168404Spjd /* 709168404Spjd * Claim all log blocks if we haven't already done so, and remember 710168404Spjd * the highest claimed sequence number. This ensures that if we can 711168404Spjd * read only part of the log now (e.g. due to a missing device), 712168404Spjd * but we can read the entire log later, we will not try to replay 713168404Spjd * or destroy beyond the last block we successfully claimed. 714168404Spjd */ 715168404Spjd ASSERT3U(zh->zh_claim_txg, <=, first_txg); 716168404Spjd if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 717219089Spjd (void) zil_parse(zilog, zil_claim_log_block, 718219089Spjd zil_claim_log_record, tx, first_txg); 719168404Spjd zh->zh_claim_txg = first_txg; 720219089Spjd zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 721219089Spjd zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 722219089Spjd if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 723219089Spjd zh->zh_flags |= ZIL_REPLAY_NEEDED; 724219089Spjd zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 725168404Spjd dsl_dataset_dirty(dmu_objset_ds(os), tx); 726168404Spjd } 727168404Spjd 728168404Spjd ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 729248571Smm dmu_objset_disown(os, FTAG); 730168404Spjd return (0); 731168404Spjd} 732168404Spjd 733185029Spjd/* 734185029Spjd * Check the log by walking the log chain. 735185029Spjd * Checksum errors are ok as they indicate the end of the chain. 736185029Spjd * Any other error (no device or read failure) returns an error. 737185029Spjd */ 738288569Smav/* ARGSUSED */ 739185029Spjdint 740288569Smavzil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx) 741168404Spjd{ 742185029Spjd zilog_t *zilog; 743185029Spjd objset_t *os; 744219089Spjd blkptr_t *bp; 745185029Spjd int error; 746168404Spjd 747219089Spjd ASSERT(tx == NULL); 748219089Spjd 749288569Smav error = dmu_objset_from_ds(ds, &os); 750248571Smm if (error != 0) { 751288569Smav cmn_err(CE_WARN, "can't open objset %llu, error %d", 752288569Smav (unsigned long long)ds->ds_object, error); 753185029Spjd return (0); 754185029Spjd } 755168404Spjd 756185029Spjd zilog = dmu_objset_zil(os); 757219089Spjd bp = (blkptr_t *)&zilog->zl_header->zh_log; 758219089Spjd 759219089Spjd /* 760219089Spjd * Check the first block and determine if it's on a log device 761219089Spjd * which may have been removed or faulted prior to loading this 762219089Spjd * pool. If so, there's no point in checking the rest of the log 763219089Spjd * as its content should have already been synced to the pool. 764219089Spjd */ 765219089Spjd if (!BP_IS_HOLE(bp)) { 766219089Spjd vdev_t *vd; 767219089Spjd boolean_t valid = B_TRUE; 768219089Spjd 769219089Spjd spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 770219089Spjd vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 771219089Spjd if (vd->vdev_islog && vdev_is_dead(vd)) 772219089Spjd valid = vdev_log_state_valid(vd); 773219089Spjd spa_config_exit(os->os_spa, SCL_STATE, FTAG); 774219089Spjd 775288569Smav if (!valid) 776219089Spjd return (0); 777168404Spjd } 778185029Spjd 779219089Spjd /* 780219089Spjd * Because tx == NULL, zil_claim_log_block() will not actually claim 781219089Spjd * any blocks, but just determine whether it is possible to do so. 782219089Spjd * In addition to checking the log chain, zil_claim_log_block() 783219089Spjd * will invoke zio_claim() with a done func of spa_claim_notify(), 784219089Spjd * which will update spa_max_claim_txg. See spa_load() for details. 785219089Spjd */ 786219089Spjd error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 787219089Spjd zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 788219089Spjd 789219089Spjd return ((error == ECKSUM || error == ENOENT) ? 0 : error); 790168404Spjd} 791168404Spjd 792185029Spjdstatic int 793185029Spjdzil_vdev_compare(const void *x1, const void *x2) 794185029Spjd{ 795219089Spjd const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 796219089Spjd const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 797185029Spjd 798185029Spjd if (v1 < v2) 799185029Spjd return (-1); 800185029Spjd if (v1 > v2) 801185029Spjd return (1); 802185029Spjd 803185029Spjd return (0); 804185029Spjd} 805185029Spjd 806168404Spjdvoid 807219089Spjdzil_add_block(zilog_t *zilog, const blkptr_t *bp) 808168404Spjd{ 809185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 810185029Spjd avl_index_t where; 811185029Spjd zil_vdev_node_t *zv, zvsearch; 812185029Spjd int ndvas = BP_GET_NDVAS(bp); 813185029Spjd int i; 814168404Spjd 815185029Spjd if (zfs_nocacheflush) 816185029Spjd return; 817168404Spjd 818185029Spjd ASSERT(zilog->zl_writer); 819168404Spjd 820185029Spjd /* 821185029Spjd * Even though we're zl_writer, we still need a lock because the 822185029Spjd * zl_get_data() callbacks may have dmu_sync() done callbacks 823185029Spjd * that will run concurrently. 824185029Spjd */ 825185029Spjd mutex_enter(&zilog->zl_vdev_lock); 826185029Spjd for (i = 0; i < ndvas; i++) { 827185029Spjd zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 828185029Spjd if (avl_find(t, &zvsearch, &where) == NULL) { 829185029Spjd zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 830185029Spjd zv->zv_vdev = zvsearch.zv_vdev; 831185029Spjd avl_insert(t, zv, where); 832185029Spjd } 833185029Spjd } 834185029Spjd mutex_exit(&zilog->zl_vdev_lock); 835168404Spjd} 836168404Spjd 837219089Spjdstatic void 838168404Spjdzil_flush_vdevs(zilog_t *zilog) 839168404Spjd{ 840168404Spjd spa_t *spa = zilog->zl_spa; 841185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 842185029Spjd void *cookie = NULL; 843185029Spjd zil_vdev_node_t *zv; 844315385Smav zio_t *zio = NULL; 845168404Spjd 846168404Spjd ASSERT(zilog->zl_writer); 847168404Spjd 848185029Spjd /* 849185029Spjd * We don't need zl_vdev_lock here because we're the zl_writer, 850185029Spjd * and all zl_get_data() callbacks are done. 851185029Spjd */ 852185029Spjd if (avl_numnodes(t) == 0) 853185029Spjd return; 854185029Spjd 855185029Spjd spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 856185029Spjd 857185029Spjd while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 858185029Spjd vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 859315385Smav if (vd != NULL && !vd->vdev_nowritecache) { 860315385Smav if (zio == NULL) 861315385Smav zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 862185029Spjd zio_flush(zio, vd); 863315385Smav } 864185029Spjd kmem_free(zv, sizeof (*zv)); 865168404Spjd } 866168404Spjd 867168404Spjd /* 868168404Spjd * Wait for all the flushes to complete. Not all devices actually 869168404Spjd * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 870168404Spjd */ 871315385Smav if (zio) 872315385Smav (void) zio_wait(zio); 873185029Spjd 874185029Spjd spa_config_exit(spa, SCL_STATE, FTAG); 875168404Spjd} 876168404Spjd 877168404Spjd/* 878168404Spjd * Function called when a log block write completes 879168404Spjd */ 880168404Spjdstatic void 881168404Spjdzil_lwb_write_done(zio_t *zio) 882168404Spjd{ 883168404Spjd lwb_t *lwb = zio->io_private; 884168404Spjd zilog_t *zilog = lwb->lwb_zilog; 885219089Spjd dmu_tx_t *tx = lwb->lwb_tx; 886168404Spjd 887185029Spjd ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 888185029Spjd ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 889185029Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 890185029Spjd ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 891185029Spjd ASSERT(!BP_IS_GANG(zio->io_bp)); 892185029Spjd ASSERT(!BP_IS_HOLE(zio->io_bp)); 893268649Sdelphij ASSERT(BP_GET_FILL(zio->io_bp) == 0); 894185029Spjd 895168404Spjd /* 896209962Smm * Ensure the lwb buffer pointer is cleared before releasing 897209962Smm * the txg. If we have had an allocation failure and 898209962Smm * the txg is waiting to sync then we want want zil_sync() 899209962Smm * to remove the lwb so that it's not picked up as the next new 900209962Smm * one in zil_commit_writer(). zil_sync() will only remove 901209962Smm * the lwb if lwb_buf is null. 902168404Spjd */ 903168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 904168404Spjd mutex_enter(&zilog->zl_lock); 905168404Spjd lwb->lwb_buf = NULL; 906219089Spjd lwb->lwb_tx = NULL; 907219089Spjd mutex_exit(&zilog->zl_lock); 908209962Smm 909209962Smm /* 910209962Smm * Now that we've written this log block, we have a stable pointer 911209962Smm * to the next block in the chain, so it's OK to let the txg in 912219089Spjd * which we allocated the next block sync. 913209962Smm */ 914219089Spjd dmu_tx_commit(tx); 915168404Spjd} 916168404Spjd 917168404Spjd/* 918168404Spjd * Initialize the io for a log block. 919168404Spjd */ 920168404Spjdstatic void 921168404Spjdzil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 922168404Spjd{ 923268657Sdelphij zbookmark_phys_t zb; 924320496Savg zio_priority_t prio; 925168404Spjd 926219089Spjd SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 927219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 928219089Spjd lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 929168404Spjd 930168404Spjd if (zilog->zl_root_zio == NULL) { 931168404Spjd zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 932168404Spjd ZIO_FLAG_CANFAIL); 933168404Spjd } 934168404Spjd if (lwb->lwb_zio == NULL) { 935320496Savg if (zilog->zl_cur_used <= zil_slog_limit || !lwb->lwb_slog) 936320496Savg prio = ZIO_PRIORITY_SYNC_WRITE; 937320496Savg else 938320496Savg prio = ZIO_PRIORITY_ASYNC_WRITE; 939168404Spjd lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 940219089Spjd 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 941320496Savg zil_lwb_write_done, lwb, prio, 942219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 943168404Spjd } 944168404Spjd} 945168404Spjd 946168404Spjd/* 947219089Spjd * Define a limited set of intent log block sizes. 948251631Sdelphij * 949219089Spjd * These must be a multiple of 4KB. Note only the amount used (again 950219089Spjd * aligned to 4KB) actually gets written. However, we can't always just 951276081Sdelphij * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted. 952219089Spjd */ 953219089Spjduint64_t zil_block_buckets[] = { 954219089Spjd 4096, /* non TX_WRITE */ 955219089Spjd 8192+4096, /* data base */ 956219089Spjd 32*1024 + 4096, /* NFS writes */ 957219089Spjd UINT64_MAX 958219089Spjd}; 959219089Spjd 960219089Spjd/* 961168404Spjd * Start a log block write and advance to the next log block. 962168404Spjd * Calls are serialized. 963168404Spjd */ 964168404Spjdstatic lwb_t * 965315388Smavzil_lwb_write_start(zilog_t *zilog, lwb_t *lwb, boolean_t last) 966168404Spjd{ 967219089Spjd lwb_t *nlwb = NULL; 968219089Spjd zil_chain_t *zilc; 969168404Spjd spa_t *spa = zilog->zl_spa; 970219089Spjd blkptr_t *bp; 971219089Spjd dmu_tx_t *tx; 972168404Spjd uint64_t txg; 973219089Spjd uint64_t zil_blksz, wsz; 974219089Spjd int i, error; 975320496Savg boolean_t slog; 976168404Spjd 977219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 978219089Spjd zilc = (zil_chain_t *)lwb->lwb_buf; 979219089Spjd bp = &zilc->zc_next_blk; 980219089Spjd } else { 981219089Spjd zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 982219089Spjd bp = &zilc->zc_next_blk; 983219089Spjd } 984168404Spjd 985219089Spjd ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 986219089Spjd 987168404Spjd /* 988168404Spjd * Allocate the next block and save its address in this block 989168404Spjd * before writing it in order to establish the log chain. 990168404Spjd * Note that if the allocation of nlwb synced before we wrote 991168404Spjd * the block that points at it (lwb), we'd leak it if we crashed. 992219089Spjd * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 993219089Spjd * We dirty the dataset to ensure that zil_sync() will be called 994219089Spjd * to clean up in the event of allocation failure or I/O failure. 995168404Spjd */ 996219089Spjd tx = dmu_tx_create(zilog->zl_os); 997330987Savg 998330987Savg /* 999330987Savg * Since we are not going to create any new dirty data, and we 1000330987Savg * can even help with clearing the existing dirty data, we 1001330987Savg * should not be subject to the dirty data based delays. We 1002330987Savg * use TXG_NOTHROTTLE to bypass the delay mechanism. 1003330987Savg */ 1004330987Savg VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE)); 1005330987Savg 1006219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1007219089Spjd txg = dmu_tx_get_txg(tx); 1008168404Spjd 1009219089Spjd lwb->lwb_tx = tx; 1010219089Spjd 1011168404Spjd /* 1012219089Spjd * Log blocks are pre-allocated. Here we select the size of the next 1013219089Spjd * block, based on size used in the last block. 1014219089Spjd * - first find the smallest bucket that will fit the block from a 1015219089Spjd * limited set of block sizes. This is because it's faster to write 1016219089Spjd * blocks allocated from the same metaslab as they are adjacent or 1017219089Spjd * close. 1018219089Spjd * - next find the maximum from the new suggested size and an array of 1019219089Spjd * previous sizes. This lessens a picket fence effect of wrongly 1020219089Spjd * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 1021219089Spjd * requests. 1022219089Spjd * 1023219089Spjd * Note we only write what is used, but we can't just allocate 1024219089Spjd * the maximum block size because we can exhaust the available 1025219089Spjd * pool log space. 1026168404Spjd */ 1027219089Spjd zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 1028219089Spjd for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 1029219089Spjd continue; 1030219089Spjd zil_blksz = zil_block_buckets[i]; 1031219089Spjd if (zil_blksz == UINT64_MAX) 1032276081Sdelphij zil_blksz = SPA_OLD_MAXBLOCKSIZE; 1033219089Spjd zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 1034219089Spjd for (i = 0; i < ZIL_PREV_BLKS; i++) 1035219089Spjd zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 1036219089Spjd zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 1037168404Spjd 1038168404Spjd BP_ZERO(bp); 1039168404Spjd /* pass the old blkptr in order to spread log blocks across devs */ 1040320496Savg error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, &slog); 1041248571Smm if (error == 0) { 1042219089Spjd ASSERT3U(bp->blk_birth, ==, txg); 1043219089Spjd bp->blk_cksum = lwb->lwb_blk.blk_cksum; 1044219089Spjd bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 1045168404Spjd 1046168404Spjd /* 1047219089Spjd * Allocate a new log write buffer (lwb). 1048168404Spjd */ 1049320496Savg nlwb = zil_alloc_lwb(zilog, bp, slog, txg); 1050168404Spjd 1051219089Spjd /* Record the block for later vdev flushing */ 1052219089Spjd zil_add_block(zilog, &lwb->lwb_blk); 1053168404Spjd } 1054168404Spjd 1055219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1056219089Spjd /* For Slim ZIL only write what is used. */ 1057219089Spjd wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1058219089Spjd ASSERT3U(wsz, <=, lwb->lwb_sz); 1059219089Spjd zio_shrink(lwb->lwb_zio, wsz); 1060168404Spjd 1061219089Spjd } else { 1062219089Spjd wsz = lwb->lwb_sz; 1063219089Spjd } 1064168404Spjd 1065219089Spjd zilc->zc_pad = 0; 1066219089Spjd zilc->zc_nused = lwb->lwb_nused; 1067219089Spjd zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1068168404Spjd 1069168404Spjd /* 1070219089Spjd * clear unused data for security 1071168404Spjd */ 1072219089Spjd bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1073168404Spjd 1074315388Smav if (last) 1075315388Smav lwb->lwb_zio->io_pipeline &= ~ZIO_STAGE_ISSUE_ASYNC; 1076219089Spjd zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1077168404Spjd 1078168404Spjd /* 1079219089Spjd * If there was an allocation failure then nlwb will be null which 1080219089Spjd * forces a txg_wait_synced(). 1081168404Spjd */ 1082168404Spjd return (nlwb); 1083168404Spjd} 1084168404Spjd 1085168404Spjdstatic lwb_t * 1086168404Spjdzil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1087168404Spjd{ 1088320496Savg lr_t *lrcb, *lrc = &itx->itx_lr; /* common log record */ 1089320496Savg lr_write_t *lrwb, *lrw = (lr_write_t *)lrc; 1090219089Spjd char *lr_buf; 1091168404Spjd uint64_t txg = lrc->lrc_txg; 1092168404Spjd uint64_t reclen = lrc->lrc_reclen; 1093219089Spjd uint64_t dlen = 0; 1094320496Savg uint64_t dnow, lwb_sp; 1095168404Spjd 1096168404Spjd if (lwb == NULL) 1097168404Spjd return (NULL); 1098219089Spjd 1099168404Spjd ASSERT(lwb->lwb_buf != NULL); 1100168404Spjd 1101168404Spjd if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1102168404Spjd dlen = P2ROUNDUP_TYPED( 1103219089Spjd lrw->lr_length, sizeof (uint64_t), uint64_t); 1104168404Spjd 1105168404Spjd zilog->zl_cur_used += (reclen + dlen); 1106168404Spjd 1107168404Spjd zil_lwb_write_init(zilog, lwb); 1108168404Spjd 1109320496Savgcont: 1110168404Spjd /* 1111168404Spjd * If this record won't fit in the current log block, start a new one. 1112320496Savg * For WR_NEED_COPY optimize layout for minimal number of chunks, but 1113320496Savg * try to keep wasted space withing reasonable range (12%). 1114168404Spjd */ 1115320496Savg lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1116320496Savg if (reclen > lwb_sp || (reclen + dlen > lwb_sp && 1117320496Savg lwb_sp < ZIL_MAX_LOG_DATA / 8 && (dlen % ZIL_MAX_LOG_DATA == 0 || 1118320496Savg lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) { 1119315388Smav lwb = zil_lwb_write_start(zilog, lwb, B_FALSE); 1120168404Spjd if (lwb == NULL) 1121168404Spjd return (NULL); 1122168404Spjd zil_lwb_write_init(zilog, lwb); 1123219089Spjd ASSERT(LWB_EMPTY(lwb)); 1124320496Savg lwb_sp = lwb->lwb_sz - lwb->lwb_nused; 1125320496Savg ASSERT3U(reclen + MIN(dlen, sizeof(uint64_t)), <=, lwb_sp); 1126168404Spjd } 1127168404Spjd 1128320496Savg dnow = MIN(dlen, lwb_sp - reclen); 1129219089Spjd lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1130219089Spjd bcopy(lrc, lr_buf, reclen); 1131320496Savg lrcb = (lr_t *)lr_buf; 1132320496Savg lrwb = (lr_write_t *)lrcb; 1133168404Spjd 1134168404Spjd /* 1135168404Spjd * If it's a write, fetch the data or get its blkptr as appropriate. 1136168404Spjd */ 1137168404Spjd if (lrc->lrc_txtype == TX_WRITE) { 1138168404Spjd if (txg > spa_freeze_txg(zilog->zl_spa)) 1139168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1140168404Spjd if (itx->itx_wr_state != WR_COPIED) { 1141168404Spjd char *dbuf; 1142168404Spjd int error; 1143168404Spjd 1144320496Savg if (itx->itx_wr_state == WR_NEED_COPY) { 1145219089Spjd dbuf = lr_buf + reclen; 1146320496Savg lrcb->lrc_reclen += dnow; 1147320496Savg if (lrwb->lr_length > dnow) 1148320496Savg lrwb->lr_length = dnow; 1149320496Savg lrw->lr_offset += dnow; 1150320496Savg lrw->lr_length -= dnow; 1151168404Spjd } else { 1152168404Spjd ASSERT(itx->itx_wr_state == WR_INDIRECT); 1153168404Spjd dbuf = NULL; 1154168404Spjd } 1155168404Spjd error = zilog->zl_get_data( 1156320496Savg itx->itx_private, lrwb, dbuf, lwb->lwb_zio); 1157214378Smm if (error == EIO) { 1158214378Smm txg_wait_synced(zilog->zl_dmu_pool, txg); 1159214378Smm return (lwb); 1160214378Smm } 1161248571Smm if (error != 0) { 1162168404Spjd ASSERT(error == ENOENT || error == EEXIST || 1163168404Spjd error == EALREADY); 1164168404Spjd return (lwb); 1165168404Spjd } 1166168404Spjd } 1167168404Spjd } 1168168404Spjd 1169219089Spjd /* 1170219089Spjd * We're actually making an entry, so update lrc_seq to be the 1171219089Spjd * log record sequence number. Note that this is generally not 1172219089Spjd * equal to the itx sequence number because not all transactions 1173219089Spjd * are synchronous, and sometimes spa_sync() gets there first. 1174219089Spjd */ 1175320496Savg lrcb->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1176320496Savg lwb->lwb_nused += reclen + dnow; 1177168404Spjd lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1178219089Spjd ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1179240415Smm ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1180168404Spjd 1181320496Savg dlen -= dnow; 1182320496Savg if (dlen > 0) { 1183320496Savg zilog->zl_cur_used += reclen; 1184320496Savg goto cont; 1185320496Savg } 1186320496Savg 1187168404Spjd return (lwb); 1188168404Spjd} 1189168404Spjd 1190168404Spjditx_t * 1191185029Spjdzil_itx_create(uint64_t txtype, size_t lrsize) 1192168404Spjd{ 1193168404Spjd itx_t *itx; 1194168404Spjd 1195168404Spjd lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1196168404Spjd 1197168404Spjd itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1198168404Spjd itx->itx_lr.lrc_txtype = txtype; 1199168404Spjd itx->itx_lr.lrc_reclen = lrsize; 1200168404Spjd itx->itx_lr.lrc_seq = 0; /* defensive */ 1201219089Spjd itx->itx_sync = B_TRUE; /* default is synchronous */ 1202168404Spjd 1203168404Spjd return (itx); 1204168404Spjd} 1205168404Spjd 1206219089Spjdvoid 1207219089Spjdzil_itx_destroy(itx_t *itx) 1208168404Spjd{ 1209219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1210219089Spjd} 1211168404Spjd 1212219089Spjd/* 1213219089Spjd * Free up the sync and async itxs. The itxs_t has already been detached 1214219089Spjd * so no locks are needed. 1215219089Spjd */ 1216219089Spjdstatic void 1217219089Spjdzil_itxg_clean(itxs_t *itxs) 1218219089Spjd{ 1219219089Spjd itx_t *itx; 1220219089Spjd list_t *list; 1221219089Spjd avl_tree_t *t; 1222219089Spjd void *cookie; 1223219089Spjd itx_async_node_t *ian; 1224168404Spjd 1225219089Spjd list = &itxs->i_sync_list; 1226219089Spjd while ((itx = list_head(list)) != NULL) { 1227219089Spjd list_remove(list, itx); 1228219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1229219089Spjd itx->itx_lr.lrc_reclen); 1230219089Spjd } 1231168404Spjd 1232219089Spjd cookie = NULL; 1233219089Spjd t = &itxs->i_async_tree; 1234219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1235219089Spjd list = &ian->ia_list; 1236219089Spjd while ((itx = list_head(list)) != NULL) { 1237219089Spjd list_remove(list, itx); 1238219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1239219089Spjd itx->itx_lr.lrc_reclen); 1240219089Spjd } 1241219089Spjd list_destroy(list); 1242219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1243219089Spjd } 1244219089Spjd avl_destroy(t); 1245219089Spjd 1246219089Spjd kmem_free(itxs, sizeof (itxs_t)); 1247168404Spjd} 1248168404Spjd 1249219089Spjdstatic int 1250219089Spjdzil_aitx_compare(const void *x1, const void *x2) 1251219089Spjd{ 1252219089Spjd const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1253219089Spjd const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1254219089Spjd 1255219089Spjd if (o1 < o2) 1256219089Spjd return (-1); 1257219089Spjd if (o1 > o2) 1258219089Spjd return (1); 1259219089Spjd 1260219089Spjd return (0); 1261219089Spjd} 1262219089Spjd 1263168404Spjd/* 1264219089Spjd * Remove all async itx with the given oid. 1265168404Spjd */ 1266168404Spjdstatic void 1267219089Spjdzil_remove_async(zilog_t *zilog, uint64_t oid) 1268168404Spjd{ 1269219089Spjd uint64_t otxg, txg; 1270219089Spjd itx_async_node_t *ian; 1271219089Spjd avl_tree_t *t; 1272219089Spjd avl_index_t where; 1273168404Spjd list_t clean_list; 1274168404Spjd itx_t *itx; 1275168404Spjd 1276219089Spjd ASSERT(oid != 0); 1277168404Spjd list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1278168404Spjd 1279219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1280219089Spjd otxg = ZILTEST_TXG; 1281219089Spjd else 1282219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1283219089Spjd 1284219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1285219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1286219089Spjd 1287219089Spjd mutex_enter(&itxg->itxg_lock); 1288219089Spjd if (itxg->itxg_txg != txg) { 1289219089Spjd mutex_exit(&itxg->itxg_lock); 1290219089Spjd continue; 1291219089Spjd } 1292219089Spjd 1293219089Spjd /* 1294219089Spjd * Locate the object node and append its list. 1295219089Spjd */ 1296219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1297219089Spjd ian = avl_find(t, &oid, &where); 1298219089Spjd if (ian != NULL) 1299219089Spjd list_move_tail(&clean_list, &ian->ia_list); 1300219089Spjd mutex_exit(&itxg->itxg_lock); 1301168404Spjd } 1302219089Spjd while ((itx = list_head(&clean_list)) != NULL) { 1303219089Spjd list_remove(&clean_list, itx); 1304219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1305219089Spjd itx->itx_lr.lrc_reclen); 1306219089Spjd } 1307219089Spjd list_destroy(&clean_list); 1308219089Spjd} 1309168404Spjd 1310219089Spjdvoid 1311219089Spjdzil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1312219089Spjd{ 1313219089Spjd uint64_t txg; 1314219089Spjd itxg_t *itxg; 1315219089Spjd itxs_t *itxs, *clean = NULL; 1316219089Spjd 1317168404Spjd /* 1318219089Spjd * Object ids can be re-instantiated in the next txg so 1319219089Spjd * remove any async transactions to avoid future leaks. 1320219089Spjd * This can happen if a fsync occurs on the re-instantiated 1321219089Spjd * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1322219089Spjd * the new file data and flushes a write record for the old object. 1323168404Spjd */ 1324219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1325219089Spjd zil_remove_async(zilog, itx->itx_oid); 1326219089Spjd 1327219089Spjd /* 1328219089Spjd * Ensure the data of a renamed file is committed before the rename. 1329219089Spjd */ 1330219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1331219089Spjd zil_async_to_sync(zilog, itx->itx_oid); 1332219089Spjd 1333239620Smm if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1334219089Spjd txg = ZILTEST_TXG; 1335219089Spjd else 1336219089Spjd txg = dmu_tx_get_txg(tx); 1337219089Spjd 1338219089Spjd itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1339219089Spjd mutex_enter(&itxg->itxg_lock); 1340219089Spjd itxs = itxg->itxg_itxs; 1341219089Spjd if (itxg->itxg_txg != txg) { 1342219089Spjd if (itxs != NULL) { 1343219089Spjd /* 1344219089Spjd * The zil_clean callback hasn't got around to cleaning 1345219089Spjd * this itxg. Save the itxs for release below. 1346219089Spjd * This should be rare. 1347219089Spjd */ 1348219089Spjd clean = itxg->itxg_itxs; 1349219089Spjd } 1350219089Spjd itxg->itxg_txg = txg; 1351219089Spjd itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1352219089Spjd 1353219089Spjd list_create(&itxs->i_sync_list, sizeof (itx_t), 1354219089Spjd offsetof(itx_t, itx_node)); 1355219089Spjd avl_create(&itxs->i_async_tree, zil_aitx_compare, 1356219089Spjd sizeof (itx_async_node_t), 1357219089Spjd offsetof(itx_async_node_t, ia_node)); 1358168404Spjd } 1359219089Spjd if (itx->itx_sync) { 1360219089Spjd list_insert_tail(&itxs->i_sync_list, itx); 1361219089Spjd } else { 1362219089Spjd avl_tree_t *t = &itxs->i_async_tree; 1363219089Spjd uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1364219089Spjd itx_async_node_t *ian; 1365219089Spjd avl_index_t where; 1366168404Spjd 1367219089Spjd ian = avl_find(t, &foid, &where); 1368219089Spjd if (ian == NULL) { 1369219089Spjd ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1370219089Spjd list_create(&ian->ia_list, sizeof (itx_t), 1371219089Spjd offsetof(itx_t, itx_node)); 1372219089Spjd ian->ia_foid = foid; 1373219089Spjd avl_insert(t, ian, where); 1374219089Spjd } 1375219089Spjd list_insert_tail(&ian->ia_list, itx); 1376168404Spjd } 1377219089Spjd 1378219089Spjd itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1379239620Smm zilog_dirty(zilog, txg); 1380219089Spjd mutex_exit(&itxg->itxg_lock); 1381219089Spjd 1382219089Spjd /* Release the old itxs now we've dropped the lock */ 1383219089Spjd if (clean != NULL) 1384219089Spjd zil_itxg_clean(clean); 1385168404Spjd} 1386168404Spjd 1387168404Spjd/* 1388168404Spjd * If there are any in-memory intent log transactions which have now been 1389239620Smm * synced then start up a taskq to free them. We should only do this after we 1390239620Smm * have written out the uberblocks (i.e. txg has been comitted) so that 1391239620Smm * don't inadvertently clean out in-memory log records that would be required 1392239620Smm * by zil_commit(). 1393168404Spjd */ 1394168404Spjdvoid 1395219089Spjdzil_clean(zilog_t *zilog, uint64_t synced_txg) 1396168404Spjd{ 1397219089Spjd itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1398219089Spjd itxs_t *clean_me; 1399168404Spjd 1400219089Spjd mutex_enter(&itxg->itxg_lock); 1401219089Spjd if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1402219089Spjd mutex_exit(&itxg->itxg_lock); 1403219089Spjd return; 1404168404Spjd } 1405219089Spjd ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1406219089Spjd ASSERT(itxg->itxg_txg != 0); 1407219089Spjd ASSERT(zilog->zl_clean_taskq != NULL); 1408219089Spjd clean_me = itxg->itxg_itxs; 1409219089Spjd itxg->itxg_itxs = NULL; 1410219089Spjd itxg->itxg_txg = 0; 1411219089Spjd mutex_exit(&itxg->itxg_lock); 1412219089Spjd /* 1413219089Spjd * Preferably start a task queue to free up the old itxs but 1414219089Spjd * if taskq_dispatch can't allocate resources to do that then 1415219089Spjd * free it in-line. This should be rare. Note, using TQ_SLEEP 1416219089Spjd * created a bad performance problem. 1417219089Spjd */ 1418219089Spjd if (taskq_dispatch(zilog->zl_clean_taskq, 1419219089Spjd (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1420219089Spjd zil_itxg_clean(clean_me); 1421168404Spjd} 1422168404Spjd 1423219089Spjd/* 1424219089Spjd * Get the list of itxs to commit into zl_itx_commit_list. 1425219089Spjd */ 1426185029Spjdstatic void 1427219089Spjdzil_get_commit_list(zilog_t *zilog) 1428168404Spjd{ 1429219089Spjd uint64_t otxg, txg; 1430219089Spjd list_t *commit_list = &zilog->zl_itx_commit_list; 1431219089Spjd 1432219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1433219089Spjd otxg = ZILTEST_TXG; 1434219089Spjd else 1435219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1436219089Spjd 1437310516Savg /* 1438310516Savg * This is inherently racy, since there is nothing to prevent 1439310516Savg * the last synced txg from changing. That's okay since we'll 1440310516Savg * only commit things in the future. 1441310516Savg */ 1442219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1443219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1444219089Spjd 1445219089Spjd mutex_enter(&itxg->itxg_lock); 1446219089Spjd if (itxg->itxg_txg != txg) { 1447219089Spjd mutex_exit(&itxg->itxg_lock); 1448219089Spjd continue; 1449219089Spjd } 1450219089Spjd 1451310516Savg /* 1452310516Savg * If we're adding itx records to the zl_itx_commit_list, 1453310516Savg * then the zil better be dirty in this "txg". We can assert 1454310516Savg * that here since we're holding the itxg_lock which will 1455310516Savg * prevent spa_sync from cleaning it. Once we add the itxs 1456310516Savg * to the zl_itx_commit_list we must commit it to disk even 1457310516Savg * if it's unnecessary (i.e. the txg was synced). 1458310516Savg */ 1459310516Savg ASSERT(zilog_is_dirty_in_txg(zilog, txg) || 1460310516Savg spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1461219089Spjd list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1462219089Spjd 1463219089Spjd mutex_exit(&itxg->itxg_lock); 1464219089Spjd } 1465219089Spjd} 1466219089Spjd 1467219089Spjd/* 1468219089Spjd * Move the async itxs for a specified object to commit into sync lists. 1469219089Spjd */ 1470308596Smavvoid 1471219089Spjdzil_async_to_sync(zilog_t *zilog, uint64_t foid) 1472219089Spjd{ 1473219089Spjd uint64_t otxg, txg; 1474219089Spjd itx_async_node_t *ian; 1475219089Spjd avl_tree_t *t; 1476219089Spjd avl_index_t where; 1477219089Spjd 1478219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1479219089Spjd otxg = ZILTEST_TXG; 1480219089Spjd else 1481219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1482219089Spjd 1483310516Savg /* 1484310516Savg * This is inherently racy, since there is nothing to prevent 1485310516Savg * the last synced txg from changing. 1486310516Savg */ 1487219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1488219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1489219089Spjd 1490219089Spjd mutex_enter(&itxg->itxg_lock); 1491219089Spjd if (itxg->itxg_txg != txg) { 1492219089Spjd mutex_exit(&itxg->itxg_lock); 1493219089Spjd continue; 1494219089Spjd } 1495219089Spjd 1496219089Spjd /* 1497219089Spjd * If a foid is specified then find that node and append its 1498219089Spjd * list. Otherwise walk the tree appending all the lists 1499219089Spjd * to the sync list. We add to the end rather than the 1500219089Spjd * beginning to ensure the create has happened. 1501219089Spjd */ 1502219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1503219089Spjd if (foid != 0) { 1504219089Spjd ian = avl_find(t, &foid, &where); 1505219089Spjd if (ian != NULL) { 1506219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1507219089Spjd &ian->ia_list); 1508219089Spjd } 1509219089Spjd } else { 1510219089Spjd void *cookie = NULL; 1511219089Spjd 1512219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1513219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1514219089Spjd &ian->ia_list); 1515219089Spjd list_destroy(&ian->ia_list); 1516219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1517219089Spjd } 1518219089Spjd } 1519219089Spjd mutex_exit(&itxg->itxg_lock); 1520219089Spjd } 1521219089Spjd} 1522219089Spjd 1523219089Spjdstatic void 1524219089Spjdzil_commit_writer(zilog_t *zilog) 1525219089Spjd{ 1526168404Spjd uint64_t txg; 1527219089Spjd itx_t *itx; 1528168404Spjd lwb_t *lwb; 1529219089Spjd spa_t *spa = zilog->zl_spa; 1530219089Spjd int error = 0; 1531168404Spjd 1532185029Spjd ASSERT(zilog->zl_root_zio == NULL); 1533168404Spjd 1534219089Spjd mutex_exit(&zilog->zl_lock); 1535219089Spjd 1536219089Spjd zil_get_commit_list(zilog); 1537219089Spjd 1538219089Spjd /* 1539219089Spjd * Return if there's nothing to commit before we dirty the fs by 1540219089Spjd * calling zil_create(). 1541219089Spjd */ 1542219089Spjd if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1543219089Spjd mutex_enter(&zilog->zl_lock); 1544219089Spjd return; 1545219089Spjd } 1546219089Spjd 1547168404Spjd if (zilog->zl_suspend) { 1548168404Spjd lwb = NULL; 1549168404Spjd } else { 1550168404Spjd lwb = list_tail(&zilog->zl_lwb_list); 1551219089Spjd if (lwb == NULL) 1552219089Spjd lwb = zil_create(zilog); 1553168404Spjd } 1554168404Spjd 1555168404Spjd DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1556219089Spjd while (itx = list_head(&zilog->zl_itx_commit_list)) { 1557168404Spjd txg = itx->itx_lr.lrc_txg; 1558310516Savg ASSERT3U(txg, !=, 0); 1559168404Spjd 1560310516Savg /* 1561310516Savg * This is inherently racy and may result in us writing 1562310516Savg * out a log block for a txg that was just synced. This is 1563310516Savg * ok since we'll end cleaning up that log block the next 1564310516Savg * time we call zil_sync(). 1565310516Savg */ 1566219089Spjd if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1567168404Spjd lwb = zil_lwb_commit(zilog, itx, lwb); 1568219089Spjd list_remove(&zilog->zl_itx_commit_list, itx); 1569168404Spjd kmem_free(itx, offsetof(itx_t, itx_lr) 1570168404Spjd + itx->itx_lr.lrc_reclen); 1571168404Spjd } 1572168404Spjd DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1573168404Spjd 1574168404Spjd /* write the last block out */ 1575168404Spjd if (lwb != NULL && lwb->lwb_zio != NULL) 1576315388Smav lwb = zil_lwb_write_start(zilog, lwb, B_TRUE); 1577168404Spjd 1578168404Spjd zilog->zl_cur_used = 0; 1579168404Spjd 1580168404Spjd /* 1581168404Spjd * Wait if necessary for the log blocks to be on stable storage. 1582168404Spjd */ 1583168404Spjd if (zilog->zl_root_zio) { 1584219089Spjd error = zio_wait(zilog->zl_root_zio); 1585185029Spjd zilog->zl_root_zio = NULL; 1586185029Spjd zil_flush_vdevs(zilog); 1587168404Spjd } 1588168404Spjd 1589219089Spjd if (error || lwb == NULL) 1590168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 1591168404Spjd 1592168404Spjd mutex_enter(&zilog->zl_lock); 1593168404Spjd 1594219089Spjd /* 1595219089Spjd * Remember the highest committed log sequence number for ztest. 1596219089Spjd * We only update this value when all the log writes succeeded, 1597219089Spjd * because ztest wants to ASSERT that it got the whole log chain. 1598219089Spjd */ 1599219089Spjd if (error == 0 && lwb != NULL) 1600219089Spjd zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1601168404Spjd} 1602168404Spjd 1603168404Spjd/* 1604219089Spjd * Commit zfs transactions to stable storage. 1605168404Spjd * If foid is 0 push out all transactions, otherwise push only those 1606219089Spjd * for that object or might reference that object. 1607219089Spjd * 1608219089Spjd * itxs are committed in batches. In a heavily stressed zil there will be 1609219089Spjd * a commit writer thread who is writing out a bunch of itxs to the log 1610219089Spjd * for a set of committing threads (cthreads) in the same batch as the writer. 1611219089Spjd * Those cthreads are all waiting on the same cv for that batch. 1612219089Spjd * 1613219089Spjd * There will also be a different and growing batch of threads that are 1614219089Spjd * waiting to commit (qthreads). When the committing batch completes 1615219089Spjd * a transition occurs such that the cthreads exit and the qthreads become 1616219089Spjd * cthreads. One of the new cthreads becomes the writer thread for the 1617219089Spjd * batch. Any new threads arriving become new qthreads. 1618219089Spjd * 1619219089Spjd * Only 2 condition variables are needed and there's no transition 1620219089Spjd * between the two cvs needed. They just flip-flop between qthreads 1621219089Spjd * and cthreads. 1622219089Spjd * 1623219089Spjd * Using this scheme we can efficiently wakeup up only those threads 1624219089Spjd * that have been committed. 1625168404Spjd */ 1626168404Spjdvoid 1627219089Spjdzil_commit(zilog_t *zilog, uint64_t foid) 1628168404Spjd{ 1629219089Spjd uint64_t mybatch; 1630219089Spjd 1631219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1632168404Spjd return; 1633168404Spjd 1634219089Spjd /* move the async itxs for the foid to the sync queues */ 1635219089Spjd zil_async_to_sync(zilog, foid); 1636219089Spjd 1637168404Spjd mutex_enter(&zilog->zl_lock); 1638219089Spjd mybatch = zilog->zl_next_batch; 1639168404Spjd while (zilog->zl_writer) { 1640219089Spjd cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1641219089Spjd if (mybatch <= zilog->zl_com_batch) { 1642168404Spjd mutex_exit(&zilog->zl_lock); 1643168404Spjd return; 1644168404Spjd } 1645168404Spjd } 1646219089Spjd 1647219089Spjd zilog->zl_next_batch++; 1648219089Spjd zilog->zl_writer = B_TRUE; 1649219089Spjd zil_commit_writer(zilog); 1650219089Spjd zilog->zl_com_batch = mybatch; 1651219089Spjd zilog->zl_writer = B_FALSE; 1652168404Spjd mutex_exit(&zilog->zl_lock); 1653219089Spjd 1654219089Spjd /* wake up one thread to become the next writer */ 1655219089Spjd cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1656219089Spjd 1657219089Spjd /* wake up all threads waiting for this batch to be committed */ 1658219089Spjd cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1659168404Spjd} 1660168404Spjd 1661168404Spjd/* 1662168404Spjd * Called in syncing context to free committed log blocks and update log header. 1663168404Spjd */ 1664168404Spjdvoid 1665168404Spjdzil_sync(zilog_t *zilog, dmu_tx_t *tx) 1666168404Spjd{ 1667168404Spjd zil_header_t *zh = zil_header_in_syncing_context(zilog); 1668168404Spjd uint64_t txg = dmu_tx_get_txg(tx); 1669168404Spjd spa_t *spa = zilog->zl_spa; 1670219089Spjd uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1671168404Spjd lwb_t *lwb; 1672168404Spjd 1673209962Smm /* 1674209962Smm * We don't zero out zl_destroy_txg, so make sure we don't try 1675209962Smm * to destroy it twice. 1676209962Smm */ 1677209962Smm if (spa_sync_pass(spa) != 1) 1678209962Smm return; 1679209962Smm 1680168404Spjd mutex_enter(&zilog->zl_lock); 1681168404Spjd 1682168404Spjd ASSERT(zilog->zl_stop_sync == 0); 1683168404Spjd 1684219089Spjd if (*replayed_seq != 0) { 1685219089Spjd ASSERT(zh->zh_replay_seq < *replayed_seq); 1686219089Spjd zh->zh_replay_seq = *replayed_seq; 1687219089Spjd *replayed_seq = 0; 1688219089Spjd } 1689168404Spjd 1690168404Spjd if (zilog->zl_destroy_txg == txg) { 1691168404Spjd blkptr_t blk = zh->zh_log; 1692168404Spjd 1693168404Spjd ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1694168404Spjd 1695168404Spjd bzero(zh, sizeof (zil_header_t)); 1696209962Smm bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1697168404Spjd 1698168404Spjd if (zilog->zl_keep_first) { 1699168404Spjd /* 1700168404Spjd * If this block was part of log chain that couldn't 1701168404Spjd * be claimed because a device was missing during 1702168404Spjd * zil_claim(), but that device later returns, 1703168404Spjd * then this block could erroneously appear valid. 1704168404Spjd * To guard against this, assign a new GUID to the new 1705168404Spjd * log chain so it doesn't matter what blk points to. 1706168404Spjd */ 1707168404Spjd zil_init_log_chain(zilog, &blk); 1708168404Spjd zh->zh_log = blk; 1709168404Spjd } 1710168404Spjd } 1711168404Spjd 1712213197Smm while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1713168404Spjd zh->zh_log = lwb->lwb_blk; 1714168404Spjd if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1715168404Spjd break; 1716168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 1717219089Spjd zio_free_zil(spa, txg, &lwb->lwb_blk); 1718168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 1719168404Spjd 1720168404Spjd /* 1721168404Spjd * If we don't have anything left in the lwb list then 1722168404Spjd * we've had an allocation failure and we need to zero 1723168404Spjd * out the zil_header blkptr so that we don't end 1724168404Spjd * up freeing the same block twice. 1725168404Spjd */ 1726168404Spjd if (list_head(&zilog->zl_lwb_list) == NULL) 1727168404Spjd BP_ZERO(&zh->zh_log); 1728168404Spjd } 1729168404Spjd mutex_exit(&zilog->zl_lock); 1730168404Spjd} 1731168404Spjd 1732168404Spjdvoid 1733168404Spjdzil_init(void) 1734168404Spjd{ 1735168404Spjd zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1736168404Spjd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1737168404Spjd} 1738168404Spjd 1739168404Spjdvoid 1740168404Spjdzil_fini(void) 1741168404Spjd{ 1742168404Spjd kmem_cache_destroy(zil_lwb_cache); 1743168404Spjd} 1744168404Spjd 1745219089Spjdvoid 1746219089Spjdzil_set_sync(zilog_t *zilog, uint64_t sync) 1747219089Spjd{ 1748219089Spjd zilog->zl_sync = sync; 1749219089Spjd} 1750219089Spjd 1751219089Spjdvoid 1752219089Spjdzil_set_logbias(zilog_t *zilog, uint64_t logbias) 1753219089Spjd{ 1754219089Spjd zilog->zl_logbias = logbias; 1755219089Spjd} 1756219089Spjd 1757168404Spjdzilog_t * 1758168404Spjdzil_alloc(objset_t *os, zil_header_t *zh_phys) 1759168404Spjd{ 1760168404Spjd zilog_t *zilog; 1761168404Spjd 1762168404Spjd zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1763168404Spjd 1764168404Spjd zilog->zl_header = zh_phys; 1765168404Spjd zilog->zl_os = os; 1766168404Spjd zilog->zl_spa = dmu_objset_spa(os); 1767168404Spjd zilog->zl_dmu_pool = dmu_objset_pool(os); 1768168404Spjd zilog->zl_destroy_txg = TXG_INITIAL - 1; 1769219089Spjd zilog->zl_logbias = dmu_objset_logbias(os); 1770219089Spjd zilog->zl_sync = dmu_objset_syncprop(os); 1771219089Spjd zilog->zl_next_batch = 1; 1772168404Spjd 1773168404Spjd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1774168404Spjd 1775219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1776219089Spjd mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1777219089Spjd MUTEX_DEFAULT, NULL); 1778219089Spjd } 1779168404Spjd 1780168404Spjd list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1781168404Spjd offsetof(lwb_t, lwb_node)); 1782168404Spjd 1783219089Spjd list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1784219089Spjd offsetof(itx_t, itx_node)); 1785219089Spjd 1786185029Spjd mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1787168404Spjd 1788185029Spjd avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1789185029Spjd sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1790185029Spjd 1791185029Spjd cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1792185029Spjd cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1793219089Spjd cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1794219089Spjd cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1795185029Spjd 1796168404Spjd return (zilog); 1797168404Spjd} 1798168404Spjd 1799168404Spjdvoid 1800168404Spjdzil_free(zilog_t *zilog) 1801168404Spjd{ 1802168404Spjd zilog->zl_stop_sync = 1; 1803168404Spjd 1804248571Smm ASSERT0(zilog->zl_suspend); 1805248571Smm ASSERT0(zilog->zl_suspending); 1806248571Smm 1807224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1808168404Spjd list_destroy(&zilog->zl_lwb_list); 1809168404Spjd 1810185029Spjd avl_destroy(&zilog->zl_vdev_tree); 1811185029Spjd mutex_destroy(&zilog->zl_vdev_lock); 1812168404Spjd 1813219089Spjd ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1814219089Spjd list_destroy(&zilog->zl_itx_commit_list); 1815219089Spjd 1816219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1817219089Spjd /* 1818219089Spjd * It's possible for an itx to be generated that doesn't dirty 1819219089Spjd * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1820219089Spjd * callback to remove the entry. We remove those here. 1821219089Spjd * 1822219089Spjd * Also free up the ziltest itxs. 1823219089Spjd */ 1824219089Spjd if (zilog->zl_itxg[i].itxg_itxs) 1825219089Spjd zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1826219089Spjd mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1827219089Spjd } 1828219089Spjd 1829168404Spjd mutex_destroy(&zilog->zl_lock); 1830168404Spjd 1831185029Spjd cv_destroy(&zilog->zl_cv_writer); 1832185029Spjd cv_destroy(&zilog->zl_cv_suspend); 1833219089Spjd cv_destroy(&zilog->zl_cv_batch[0]); 1834219089Spjd cv_destroy(&zilog->zl_cv_batch[1]); 1835185029Spjd 1836168404Spjd kmem_free(zilog, sizeof (zilog_t)); 1837168404Spjd} 1838168404Spjd 1839168404Spjd/* 1840168404Spjd * Open an intent log. 1841168404Spjd */ 1842168404Spjdzilog_t * 1843168404Spjdzil_open(objset_t *os, zil_get_data_t *get_data) 1844168404Spjd{ 1845168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1846168404Spjd 1847224526Smm ASSERT(zilog->zl_clean_taskq == NULL); 1848224526Smm ASSERT(zilog->zl_get_data == NULL); 1849224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1850224526Smm 1851168404Spjd zilog->zl_get_data = get_data; 1852168404Spjd zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1853168404Spjd 2, 2, TASKQ_PREPOPULATE); 1854168404Spjd 1855168404Spjd return (zilog); 1856168404Spjd} 1857168404Spjd 1858168404Spjd/* 1859168404Spjd * Close an intent log. 1860168404Spjd */ 1861168404Spjdvoid 1862168404Spjdzil_close(zilog_t *zilog) 1863168404Spjd{ 1864224526Smm lwb_t *lwb; 1865219089Spjd uint64_t txg = 0; 1866219089Spjd 1867219089Spjd zil_commit(zilog, 0); /* commit all itx */ 1868219089Spjd 1869168404Spjd /* 1870219089Spjd * The lwb_max_txg for the stubby lwb will reflect the last activity 1871219089Spjd * for the zil. After a txg_wait_synced() on the txg we know all the 1872219089Spjd * callbacks have occurred that may clean the zil. Only then can we 1873219089Spjd * destroy the zl_clean_taskq. 1874168404Spjd */ 1875219089Spjd mutex_enter(&zilog->zl_lock); 1876224526Smm lwb = list_tail(&zilog->zl_lwb_list); 1877224526Smm if (lwb != NULL) 1878224526Smm txg = lwb->lwb_max_txg; 1879219089Spjd mutex_exit(&zilog->zl_lock); 1880219089Spjd if (txg) 1881168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1882168404Spjd 1883310516Savg if (zilog_is_dirty(zilog)) 1884310516Savg zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg); 1885310516Savg VERIFY(!zilog_is_dirty(zilog)); 1886310516Savg 1887168404Spjd taskq_destroy(zilog->zl_clean_taskq); 1888168404Spjd zilog->zl_clean_taskq = NULL; 1889168404Spjd zilog->zl_get_data = NULL; 1890224526Smm 1891224526Smm /* 1892224526Smm * We should have only one LWB left on the list; remove it now. 1893224526Smm */ 1894224526Smm mutex_enter(&zilog->zl_lock); 1895224526Smm lwb = list_head(&zilog->zl_lwb_list); 1896224526Smm if (lwb != NULL) { 1897224526Smm ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1898224526Smm list_remove(&zilog->zl_lwb_list, lwb); 1899224526Smm zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1900224526Smm kmem_cache_free(zil_lwb_cache, lwb); 1901224526Smm } 1902224526Smm mutex_exit(&zilog->zl_lock); 1903168404Spjd} 1904168404Spjd 1905248571Smmstatic char *suspend_tag = "zil suspending"; 1906248571Smm 1907168404Spjd/* 1908168404Spjd * Suspend an intent log. While in suspended mode, we still honor 1909168404Spjd * synchronous semantics, but we rely on txg_wait_synced() to do it. 1910248571Smm * On old version pools, we suspend the log briefly when taking a 1911248571Smm * snapshot so that it will have an empty intent log. 1912248571Smm * 1913248571Smm * Long holds are not really intended to be used the way we do here -- 1914248571Smm * held for such a short time. A concurrent caller of dsl_dataset_long_held() 1915248571Smm * could fail. Therefore we take pains to only put a long hold if it is 1916248571Smm * actually necessary. Fortunately, it will only be necessary if the 1917248571Smm * objset is currently mounted (or the ZVOL equivalent). In that case it 1918248571Smm * will already have a long hold, so we are not really making things any worse. 1919248571Smm * 1920248571Smm * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or 1921248571Smm * zvol_state_t), and use their mechanism to prevent their hold from being 1922248571Smm * dropped (e.g. VFS_HOLD()). However, that would be even more pain for 1923248571Smm * very little gain. 1924248571Smm * 1925248571Smm * if cookiep == NULL, this does both the suspend & resume. 1926248571Smm * Otherwise, it returns with the dataset "long held", and the cookie 1927248571Smm * should be passed into zil_resume(). 1928168404Spjd */ 1929168404Spjdint 1930248571Smmzil_suspend(const char *osname, void **cookiep) 1931168404Spjd{ 1932248571Smm objset_t *os; 1933248571Smm zilog_t *zilog; 1934248571Smm const zil_header_t *zh; 1935248571Smm int error; 1936168404Spjd 1937248571Smm error = dmu_objset_hold(osname, suspend_tag, &os); 1938248571Smm if (error != 0) 1939248571Smm return (error); 1940248571Smm zilog = dmu_objset_zil(os); 1941248571Smm 1942168404Spjd mutex_enter(&zilog->zl_lock); 1943248571Smm zh = zilog->zl_header; 1944248571Smm 1945200724Sdelphij if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1946168404Spjd mutex_exit(&zilog->zl_lock); 1947248571Smm dmu_objset_rele(os, suspend_tag); 1948249195Smm return (SET_ERROR(EBUSY)); 1949168404Spjd } 1950248571Smm 1951248571Smm /* 1952248571Smm * Don't put a long hold in the cases where we can avoid it. This 1953248571Smm * is when there is no cookie so we are doing a suspend & resume 1954248571Smm * (i.e. called from zil_vdev_offline()), and there's nothing to do 1955248571Smm * for the suspend because it's already suspended, or there's no ZIL. 1956248571Smm */ 1957248571Smm if (cookiep == NULL && !zilog->zl_suspending && 1958248571Smm (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) { 1959248571Smm mutex_exit(&zilog->zl_lock); 1960248571Smm dmu_objset_rele(os, suspend_tag); 1961248571Smm return (0); 1962248571Smm } 1963248571Smm 1964248571Smm dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag); 1965248571Smm dsl_pool_rele(dmu_objset_pool(os), suspend_tag); 1966248571Smm 1967248571Smm zilog->zl_suspend++; 1968248571Smm 1969248571Smm if (zilog->zl_suspend > 1) { 1970168404Spjd /* 1971248571Smm * Someone else is already suspending it. 1972168404Spjd * Just wait for them to finish. 1973168404Spjd */ 1974248571Smm 1975168404Spjd while (zilog->zl_suspending) 1976168404Spjd cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1977168404Spjd mutex_exit(&zilog->zl_lock); 1978248571Smm 1979248571Smm if (cookiep == NULL) 1980248571Smm zil_resume(os); 1981248571Smm else 1982248571Smm *cookiep = os; 1983168404Spjd return (0); 1984168404Spjd } 1985248571Smm 1986248571Smm /* 1987248571Smm * If there is no pointer to an on-disk block, this ZIL must not 1988248571Smm * be active (e.g. filesystem not mounted), so there's nothing 1989248571Smm * to clean up. 1990248571Smm */ 1991248571Smm if (BP_IS_HOLE(&zh->zh_log)) { 1992248571Smm ASSERT(cookiep != NULL); /* fast path already handled */ 1993248571Smm 1994248571Smm *cookiep = os; 1995248571Smm mutex_exit(&zilog->zl_lock); 1996248571Smm return (0); 1997248571Smm } 1998248571Smm 1999168404Spjd zilog->zl_suspending = B_TRUE; 2000168404Spjd mutex_exit(&zilog->zl_lock); 2001168404Spjd 2002219089Spjd zil_commit(zilog, 0); 2003168404Spjd 2004168404Spjd zil_destroy(zilog, B_FALSE); 2005168404Spjd 2006168404Spjd mutex_enter(&zilog->zl_lock); 2007168404Spjd zilog->zl_suspending = B_FALSE; 2008168404Spjd cv_broadcast(&zilog->zl_cv_suspend); 2009168404Spjd mutex_exit(&zilog->zl_lock); 2010168404Spjd 2011248571Smm if (cookiep == NULL) 2012248571Smm zil_resume(os); 2013248571Smm else 2014248571Smm *cookiep = os; 2015168404Spjd return (0); 2016168404Spjd} 2017168404Spjd 2018168404Spjdvoid 2019248571Smmzil_resume(void *cookie) 2020168404Spjd{ 2021248571Smm objset_t *os = cookie; 2022248571Smm zilog_t *zilog = dmu_objset_zil(os); 2023248571Smm 2024168404Spjd mutex_enter(&zilog->zl_lock); 2025168404Spjd ASSERT(zilog->zl_suspend != 0); 2026168404Spjd zilog->zl_suspend--; 2027168404Spjd mutex_exit(&zilog->zl_lock); 2028248571Smm dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag); 2029248571Smm dsl_dataset_rele(dmu_objset_ds(os), suspend_tag); 2030168404Spjd} 2031168404Spjd 2032219089Spjdtypedef struct zil_replay_arg { 2033219089Spjd zil_replay_func_t **zr_replay; 2034219089Spjd void *zr_arg; 2035219089Spjd boolean_t zr_byteswap; 2036219089Spjd char *zr_lr; 2037219089Spjd} zil_replay_arg_t; 2038219089Spjd 2039219089Spjdstatic int 2040219089Spjdzil_replay_error(zilog_t *zilog, lr_t *lr, int error) 2041209962Smm{ 2042307122Smav char name[ZFS_MAX_DATASET_NAME_LEN]; 2043209962Smm 2044219089Spjd zilog->zl_replaying_seq--; /* didn't actually replay this one */ 2045209962Smm 2046219089Spjd dmu_objset_name(zilog->zl_os, name); 2047209962Smm 2048219089Spjd cmn_err(CE_WARN, "ZFS replay transaction error %d, " 2049219089Spjd "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 2050219089Spjd (u_longlong_t)lr->lrc_seq, 2051219089Spjd (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 2052219089Spjd (lr->lrc_txtype & TX_CI) ? "CI" : ""); 2053219089Spjd 2054219089Spjd return (error); 2055209962Smm} 2056209962Smm 2057219089Spjdstatic int 2058168404Spjdzil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 2059168404Spjd{ 2060168404Spjd zil_replay_arg_t *zr = zra; 2061168404Spjd const zil_header_t *zh = zilog->zl_header; 2062168404Spjd uint64_t reclen = lr->lrc_reclen; 2063168404Spjd uint64_t txtype = lr->lrc_txtype; 2064219089Spjd int error = 0; 2065168404Spjd 2066219089Spjd zilog->zl_replaying_seq = lr->lrc_seq; 2067168404Spjd 2068219089Spjd if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 2069219089Spjd return (0); 2070219089Spjd 2071168404Spjd if (lr->lrc_txg < claim_txg) /* already committed */ 2072219089Spjd return (0); 2073168404Spjd 2074185029Spjd /* Strip case-insensitive bit, still present in log record */ 2075185029Spjd txtype &= ~TX_CI; 2076185029Spjd 2077219089Spjd if (txtype == 0 || txtype >= TX_MAX_TYPE) 2078219089Spjd return (zil_replay_error(zilog, lr, EINVAL)); 2079219089Spjd 2080219089Spjd /* 2081219089Spjd * If this record type can be logged out of order, the object 2082219089Spjd * (lr_foid) may no longer exist. That's legitimate, not an error. 2083219089Spjd */ 2084219089Spjd if (TX_OOO(txtype)) { 2085219089Spjd error = dmu_object_info(zilog->zl_os, 2086219089Spjd ((lr_ooo_t *)lr)->lr_foid, NULL); 2087219089Spjd if (error == ENOENT || error == EEXIST) 2088219089Spjd return (0); 2089209962Smm } 2090209962Smm 2091168404Spjd /* 2092168404Spjd * Make a copy of the data so we can revise and extend it. 2093168404Spjd */ 2094219089Spjd bcopy(lr, zr->zr_lr, reclen); 2095168404Spjd 2096168404Spjd /* 2097219089Spjd * If this is a TX_WRITE with a blkptr, suck in the data. 2098219089Spjd */ 2099219089Spjd if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 2100219089Spjd error = zil_read_log_data(zilog, (lr_write_t *)lr, 2101219089Spjd zr->zr_lr + reclen); 2102248571Smm if (error != 0) 2103219089Spjd return (zil_replay_error(zilog, lr, error)); 2104219089Spjd } 2105219089Spjd 2106219089Spjd /* 2107168404Spjd * The log block containing this lr may have been byteswapped 2108168404Spjd * so that we can easily examine common fields like lrc_txtype. 2109219089Spjd * However, the log is a mix of different record types, and only the 2110168404Spjd * replay vectors know how to byteswap their records. Therefore, if 2111168404Spjd * the lr was byteswapped, undo it before invoking the replay vector. 2112168404Spjd */ 2113168404Spjd if (zr->zr_byteswap) 2114219089Spjd byteswap_uint64_array(zr->zr_lr, reclen); 2115168404Spjd 2116168404Spjd /* 2117168404Spjd * We must now do two things atomically: replay this log record, 2118209962Smm * and update the log header sequence number to reflect the fact that 2119209962Smm * we did so. At the end of each replay function the sequence number 2120209962Smm * is updated if we are in replay mode. 2121168404Spjd */ 2122219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 2123248571Smm if (error != 0) { 2124168404Spjd /* 2125168404Spjd * The DMU's dnode layer doesn't see removes until the txg 2126168404Spjd * commits, so a subsequent claim can spuriously fail with 2127209962Smm * EEXIST. So if we receive any error we try syncing out 2128219089Spjd * any removes then retry the transaction. Note that we 2129219089Spjd * specify B_FALSE for byteswap now, so we don't do it twice. 2130168404Spjd */ 2131219089Spjd txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 2132219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 2133248571Smm if (error != 0) 2134219089Spjd return (zil_replay_error(zilog, lr, error)); 2135168404Spjd } 2136219089Spjd return (0); 2137168404Spjd} 2138168404Spjd 2139168404Spjd/* ARGSUSED */ 2140219089Spjdstatic int 2141168404Spjdzil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 2142168404Spjd{ 2143168404Spjd zilog->zl_replay_blks++; 2144219089Spjd 2145219089Spjd return (0); 2146168404Spjd} 2147168404Spjd 2148168404Spjd/* 2149168404Spjd * If this dataset has a non-empty intent log, replay it and destroy it. 2150168404Spjd */ 2151168404Spjdvoid 2152209962Smmzil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 2153168404Spjd{ 2154168404Spjd zilog_t *zilog = dmu_objset_zil(os); 2155168404Spjd const zil_header_t *zh = zilog->zl_header; 2156168404Spjd zil_replay_arg_t zr; 2157168404Spjd 2158200724Sdelphij if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 2159168404Spjd zil_destroy(zilog, B_TRUE); 2160168404Spjd return; 2161168404Spjd } 2162168404Spjd 2163168404Spjd zr.zr_replay = replay_func; 2164168404Spjd zr.zr_arg = arg; 2165168404Spjd zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2166219089Spjd zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2167168404Spjd 2168168404Spjd /* 2169168404Spjd * Wait for in-progress removes to sync before starting replay. 2170168404Spjd */ 2171168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 2172168404Spjd 2173209962Smm zilog->zl_replay = B_TRUE; 2174219089Spjd zilog->zl_replay_time = ddi_get_lbolt(); 2175168404Spjd ASSERT(zilog->zl_replay_blks == 0); 2176168404Spjd (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2177168404Spjd zh->zh_claim_txg); 2178219089Spjd kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2179168404Spjd 2180168404Spjd zil_destroy(zilog, B_FALSE); 2181185029Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2182209962Smm zilog->zl_replay = B_FALSE; 2183168404Spjd} 2184168404Spjd 2185219089Spjdboolean_t 2186219089Spjdzil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2187168404Spjd{ 2188219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2189219089Spjd return (B_TRUE); 2190168404Spjd 2191219089Spjd if (zilog->zl_replay) { 2192219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2193219089Spjd zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2194219089Spjd zilog->zl_replaying_seq; 2195219089Spjd return (B_TRUE); 2196168404Spjd } 2197168404Spjd 2198219089Spjd return (B_FALSE); 2199168404Spjd} 2200213197Smm 2201213197Smm/* ARGSUSED */ 2202213197Smmint 2203219089Spjdzil_vdev_offline(const char *osname, void *arg) 2204213197Smm{ 2205213197Smm int error; 2206213197Smm 2207248571Smm error = zil_suspend(osname, NULL); 2208248571Smm if (error != 0) 2209249195Smm return (SET_ERROR(EEXIST)); 2210248571Smm return (0); 2211213197Smm} 2212