zil.c revision 246666
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. 23239620Smm * Copyright (c) 2012 by Delphix. All rights reserved. 24168404Spjd */ 25168404Spjd 26219089Spjd/* Portions Copyright 2010 Robert Milkowski */ 27219089Spjd 28168404Spjd#include <sys/zfs_context.h> 29168404Spjd#include <sys/spa.h> 30168404Spjd#include <sys/dmu.h> 31168404Spjd#include <sys/zap.h> 32168404Spjd#include <sys/arc.h> 33168404Spjd#include <sys/stat.h> 34168404Spjd#include <sys/resource.h> 35168404Spjd#include <sys/zil.h> 36168404Spjd#include <sys/zil_impl.h> 37168404Spjd#include <sys/dsl_dataset.h> 38219089Spjd#include <sys/vdev_impl.h> 39168404Spjd#include <sys/dmu_tx.h> 40219089Spjd#include <sys/dsl_pool.h> 41168404Spjd 42168404Spjd/* 43168404Spjd * The zfs intent log (ZIL) saves transaction records of system calls 44168404Spjd * that change the file system in memory with enough information 45168404Spjd * to be able to replay them. These are stored in memory until 46168404Spjd * either the DMU transaction group (txg) commits them to the stable pool 47168404Spjd * and they can be discarded, or they are flushed to the stable log 48168404Spjd * (also in the pool) due to a fsync, O_DSYNC or other synchronous 49168404Spjd * requirement. In the event of a panic or power fail then those log 50168404Spjd * records (transactions) are replayed. 51168404Spjd * 52168404Spjd * There is one ZIL per file system. Its on-disk (pool) format consists 53168404Spjd * of 3 parts: 54168404Spjd * 55168404Spjd * - ZIL header 56168404Spjd * - ZIL blocks 57168404Spjd * - ZIL records 58168404Spjd * 59168404Spjd * A log record holds a system call transaction. Log blocks can 60168404Spjd * hold many log records and the blocks are chained together. 61168404Spjd * Each ZIL block contains a block pointer (blkptr_t) to the next 62168404Spjd * ZIL block in the chain. The ZIL header points to the first 63168404Spjd * block in the chain. Note there is not a fixed place in the pool 64168404Spjd * to hold blocks. They are dynamically allocated and freed as 65168404Spjd * needed from the blocks available. Figure X shows the ZIL structure: 66168404Spjd */ 67168404Spjd 68168404Spjd/* 69168404Spjd * This global ZIL switch affects all pools 70168404Spjd */ 71219089Spjdint zil_replay_disable = 0; /* disable intent logging replay */ 72168404SpjdSYSCTL_DECL(_vfs_zfs); 73219089SpjdTUNABLE_INT("vfs.zfs.zil_replay_disable", &zil_replay_disable); 74219089SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, zil_replay_disable, CTLFLAG_RW, 75219089Spjd &zil_replay_disable, 0, "Disable intent logging replay"); 76168404Spjd 77168404Spjd/* 78168404Spjd * Tunable parameter for debugging or performance analysis. Setting 79168404Spjd * zfs_nocacheflush will cause corruption on power loss if a volatile 80168404Spjd * out-of-order write cache is enabled. 81168404Spjd */ 82168404Spjdboolean_t zfs_nocacheflush = B_FALSE; 83168404SpjdTUNABLE_INT("vfs.zfs.cache_flush_disable", &zfs_nocacheflush); 84168404SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, cache_flush_disable, CTLFLAG_RDTUN, 85168404Spjd &zfs_nocacheflush, 0, "Disable cache flush"); 86240868Spjdboolean_t zfs_notrim = B_TRUE; 87240868SpjdTUNABLE_INT("vfs.zfs.trim_disable", &zfs_notrim); 88240868SpjdSYSCTL_INT(_vfs_zfs, OID_AUTO, trim_disable, CTLFLAG_RDTUN, &zfs_notrim, 0, 89240868Spjd "Disable trim"); 90168404Spjd 91168404Spjdstatic kmem_cache_t *zil_lwb_cache; 92168404Spjd 93219089Spjdstatic void zil_async_to_sync(zilog_t *zilog, uint64_t foid); 94219089Spjd 95219089Spjd#define LWB_EMPTY(lwb) ((BP_GET_LSIZE(&lwb->lwb_blk) - \ 96219089Spjd sizeof (zil_chain_t)) == (lwb->lwb_sz - lwb->lwb_nused)) 97219089Spjd 98219089Spjd 99219089Spjd/* 100219089Spjd * ziltest is by and large an ugly hack, but very useful in 101219089Spjd * checking replay without tedious work. 102219089Spjd * When running ziltest we want to keep all itx's and so maintain 103219089Spjd * a single list in the zl_itxg[] that uses a high txg: ZILTEST_TXG 104219089Spjd * We subtract TXG_CONCURRENT_STATES to allow for common code. 105219089Spjd */ 106219089Spjd#define ZILTEST_TXG (UINT64_MAX - TXG_CONCURRENT_STATES) 107219089Spjd 108168404Spjdstatic int 109219089Spjdzil_bp_compare(const void *x1, const void *x2) 110168404Spjd{ 111219089Spjd const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 112219089Spjd const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 113168404Spjd 114168404Spjd if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 115168404Spjd return (-1); 116168404Spjd if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 117168404Spjd return (1); 118168404Spjd 119168404Spjd if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 120168404Spjd return (-1); 121168404Spjd if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 122168404Spjd return (1); 123168404Spjd 124168404Spjd return (0); 125168404Spjd} 126168404Spjd 127168404Spjdstatic void 128219089Spjdzil_bp_tree_init(zilog_t *zilog) 129168404Spjd{ 130219089Spjd avl_create(&zilog->zl_bp_tree, zil_bp_compare, 131219089Spjd sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 132168404Spjd} 133168404Spjd 134168404Spjdstatic void 135219089Spjdzil_bp_tree_fini(zilog_t *zilog) 136168404Spjd{ 137219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 138219089Spjd zil_bp_node_t *zn; 139168404Spjd void *cookie = NULL; 140168404Spjd 141168404Spjd while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 142219089Spjd kmem_free(zn, sizeof (zil_bp_node_t)); 143168404Spjd 144168404Spjd avl_destroy(t); 145168404Spjd} 146168404Spjd 147219089Spjdint 148219089Spjdzil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 149168404Spjd{ 150219089Spjd avl_tree_t *t = &zilog->zl_bp_tree; 151219089Spjd const dva_t *dva = BP_IDENTITY(bp); 152219089Spjd zil_bp_node_t *zn; 153168404Spjd avl_index_t where; 154168404Spjd 155168404Spjd if (avl_find(t, dva, &where) != NULL) 156168404Spjd return (EEXIST); 157168404Spjd 158219089Spjd zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 159168404Spjd zn->zn_dva = *dva; 160168404Spjd avl_insert(t, zn, where); 161168404Spjd 162168404Spjd return (0); 163168404Spjd} 164168404Spjd 165168404Spjdstatic zil_header_t * 166168404Spjdzil_header_in_syncing_context(zilog_t *zilog) 167168404Spjd{ 168168404Spjd return ((zil_header_t *)zilog->zl_header); 169168404Spjd} 170168404Spjd 171168404Spjdstatic void 172168404Spjdzil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 173168404Spjd{ 174168404Spjd zio_cksum_t *zc = &bp->blk_cksum; 175168404Spjd 176168404Spjd zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 177168404Spjd zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 178168404Spjd zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 179168404Spjd zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 180168404Spjd} 181168404Spjd 182168404Spjd/* 183219089Spjd * Read a log block and make sure it's valid. 184168404Spjd */ 185168404Spjdstatic int 186219089Spjdzil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst, 187219089Spjd char **end) 188168404Spjd{ 189219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 190219089Spjd uint32_t aflags = ARC_WAIT; 191219089Spjd arc_buf_t *abuf = NULL; 192168404Spjd zbookmark_t zb; 193168404Spjd int error; 194168404Spjd 195219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 196219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 197168404Spjd 198219089Spjd if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 199219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE; 200168404Spjd 201219089Spjd SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 202219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 203168404Spjd 204246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 205219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 206219089Spjd 207168404Spjd if (error == 0) { 208168404Spjd zio_cksum_t cksum = bp->blk_cksum; 209168404Spjd 210168404Spjd /* 211185029Spjd * Validate the checksummed log block. 212185029Spjd * 213168404Spjd * Sequence numbers should be... sequential. The checksum 214168404Spjd * verifier for the next block should be bp's checksum plus 1. 215185029Spjd * 216185029Spjd * Also check the log chain linkage and size used. 217168404Spjd */ 218168404Spjd cksum.zc_word[ZIL_ZC_SEQ]++; 219168404Spjd 220219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 221219089Spjd zil_chain_t *zilc = abuf->b_data; 222219089Spjd char *lr = (char *)(zilc + 1); 223219089Spjd uint64_t len = zilc->zc_nused - sizeof (zil_chain_t); 224219089Spjd 225219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 226219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) { 227219089Spjd error = ECKSUM; 228219089Spjd } else { 229219089Spjd bcopy(lr, dst, len); 230219089Spjd *end = (char *)dst + len; 231219089Spjd *nbp = zilc->zc_next_blk; 232219089Spjd } 233219089Spjd } else { 234219089Spjd char *lr = abuf->b_data; 235219089Spjd uint64_t size = BP_GET_LSIZE(bp); 236219089Spjd zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1; 237219089Spjd 238219089Spjd if (bcmp(&cksum, &zilc->zc_next_blk.blk_cksum, 239219089Spjd sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) || 240219089Spjd (zilc->zc_nused > (size - sizeof (*zilc)))) { 241219089Spjd error = ECKSUM; 242219089Spjd } else { 243219089Spjd bcopy(lr, dst, zilc->zc_nused); 244219089Spjd *end = (char *)dst + zilc->zc_nused; 245219089Spjd *nbp = zilc->zc_next_blk; 246219089Spjd } 247185029Spjd } 248168404Spjd 249219089Spjd VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 250168404Spjd } 251168404Spjd 252219089Spjd return (error); 253219089Spjd} 254168404Spjd 255219089Spjd/* 256219089Spjd * Read a TX_WRITE log data block. 257219089Spjd */ 258219089Spjdstatic int 259219089Spjdzil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 260219089Spjd{ 261219089Spjd enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 262219089Spjd const blkptr_t *bp = &lr->lr_blkptr; 263219089Spjd uint32_t aflags = ARC_WAIT; 264219089Spjd arc_buf_t *abuf = NULL; 265219089Spjd zbookmark_t zb; 266219089Spjd int error; 267219089Spjd 268219089Spjd if (BP_IS_HOLE(bp)) { 269219089Spjd if (wbuf != NULL) 270219089Spjd bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 271219089Spjd return (0); 272219089Spjd } 273219089Spjd 274219089Spjd if (zilog->zl_header->zh_claim_txg == 0) 275219089Spjd zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 276219089Spjd 277219089Spjd SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 278219089Spjd ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 279219089Spjd 280246666Smm error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 281219089Spjd ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 282219089Spjd 283219089Spjd if (error == 0) { 284219089Spjd if (wbuf != NULL) 285219089Spjd bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 286219089Spjd (void) arc_buf_remove_ref(abuf, &abuf); 287219089Spjd } 288219089Spjd 289168404Spjd return (error); 290168404Spjd} 291168404Spjd 292168404Spjd/* 293168404Spjd * Parse the intent log, and call parse_func for each valid record within. 294168404Spjd */ 295219089Spjdint 296168404Spjdzil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 297168404Spjd zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 298168404Spjd{ 299168404Spjd const zil_header_t *zh = zilog->zl_header; 300219089Spjd boolean_t claimed = !!zh->zh_claim_txg; 301219089Spjd uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 302219089Spjd uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 303219089Spjd uint64_t max_blk_seq = 0; 304219089Spjd uint64_t max_lr_seq = 0; 305219089Spjd uint64_t blk_count = 0; 306219089Spjd uint64_t lr_count = 0; 307219089Spjd blkptr_t blk, next_blk; 308168404Spjd char *lrbuf, *lrp; 309219089Spjd int error = 0; 310168404Spjd 311219089Spjd /* 312219089Spjd * Old logs didn't record the maximum zh_claim_lr_seq. 313219089Spjd */ 314219089Spjd if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 315219089Spjd claim_lr_seq = UINT64_MAX; 316168404Spjd 317168404Spjd /* 318168404Spjd * Starting at the block pointed to by zh_log we read the log chain. 319168404Spjd * For each block in the chain we strongly check that block to 320168404Spjd * ensure its validity. We stop when an invalid block is found. 321168404Spjd * For each block pointer in the chain we call parse_blk_func(). 322168404Spjd * For each record in each valid block we call parse_lr_func(). 323168404Spjd * If the log has been claimed, stop if we encounter a sequence 324168404Spjd * number greater than the highest claimed sequence number. 325168404Spjd */ 326219089Spjd lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 327219089Spjd zil_bp_tree_init(zilog); 328168404Spjd 329219089Spjd for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 330219089Spjd uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 331219089Spjd int reclen; 332219089Spjd char *end; 333219089Spjd 334219089Spjd if (blk_seq > claim_blk_seq) 335168404Spjd break; 336219089Spjd if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 337219089Spjd break; 338219089Spjd ASSERT3U(max_blk_seq, <, blk_seq); 339219089Spjd max_blk_seq = blk_seq; 340219089Spjd blk_count++; 341168404Spjd 342219089Spjd if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 343219089Spjd break; 344168404Spjd 345219089Spjd error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf, &end); 346168404Spjd if (error) 347168404Spjd break; 348168404Spjd 349219089Spjd for (lrp = lrbuf; lrp < end; lrp += reclen) { 350168404Spjd lr_t *lr = (lr_t *)lrp; 351168404Spjd reclen = lr->lrc_reclen; 352168404Spjd ASSERT3U(reclen, >=, sizeof (lr_t)); 353219089Spjd if (lr->lrc_seq > claim_lr_seq) 354219089Spjd goto done; 355219089Spjd if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 356219089Spjd goto done; 357219089Spjd ASSERT3U(max_lr_seq, <, lr->lrc_seq); 358219089Spjd max_lr_seq = lr->lrc_seq; 359219089Spjd lr_count++; 360168404Spjd } 361168404Spjd } 362219089Spjddone: 363219089Spjd zilog->zl_parse_error = error; 364219089Spjd zilog->zl_parse_blk_seq = max_blk_seq; 365219089Spjd zilog->zl_parse_lr_seq = max_lr_seq; 366219089Spjd zilog->zl_parse_blk_count = blk_count; 367219089Spjd zilog->zl_parse_lr_count = lr_count; 368168404Spjd 369219089Spjd ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 370219089Spjd (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 371219089Spjd 372219089Spjd zil_bp_tree_fini(zilog); 373219089Spjd zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 374219089Spjd 375219089Spjd return (error); 376168404Spjd} 377168404Spjd 378219089Spjdstatic int 379168404Spjdzil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 380168404Spjd{ 381168404Spjd /* 382168404Spjd * Claim log block if not already committed and not already claimed. 383219089Spjd * If tx == NULL, just verify that the block is claimable. 384168404Spjd */ 385219089Spjd if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 386219089Spjd return (0); 387219089Spjd 388219089Spjd return (zio_wait(zio_claim(NULL, zilog->zl_spa, 389219089Spjd tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 390219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 391168404Spjd} 392168404Spjd 393219089Spjdstatic int 394168404Spjdzil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 395168404Spjd{ 396219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 397219089Spjd int error; 398219089Spjd 399219089Spjd if (lrc->lrc_txtype != TX_WRITE) 400219089Spjd return (0); 401219089Spjd 402219089Spjd /* 403219089Spjd * If the block is not readable, don't claim it. This can happen 404219089Spjd * in normal operation when a log block is written to disk before 405219089Spjd * some of the dmu_sync() blocks it points to. In this case, the 406219089Spjd * transaction cannot have been committed to anyone (we would have 407219089Spjd * waited for all writes to be stable first), so it is semantically 408219089Spjd * correct to declare this the end of the log. 409219089Spjd */ 410219089Spjd if (lr->lr_blkptr.blk_birth >= first_txg && 411219089Spjd (error = zil_read_log_data(zilog, lr, NULL)) != 0) 412219089Spjd return (error); 413219089Spjd return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 414168404Spjd} 415168404Spjd 416168404Spjd/* ARGSUSED */ 417219089Spjdstatic int 418168404Spjdzil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 419168404Spjd{ 420219089Spjd zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 421219089Spjd 422219089Spjd return (0); 423168404Spjd} 424168404Spjd 425219089Spjdstatic int 426168404Spjdzil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 427168404Spjd{ 428219089Spjd lr_write_t *lr = (lr_write_t *)lrc; 429219089Spjd blkptr_t *bp = &lr->lr_blkptr; 430219089Spjd 431168404Spjd /* 432168404Spjd * If we previously claimed it, we need to free it. 433168404Spjd */ 434219089Spjd if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 435219089Spjd bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 436219089Spjd zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 437219089Spjd 438219089Spjd return (0); 439219089Spjd} 440219089Spjd 441219089Spjdstatic lwb_t * 442219089Spjdzil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, uint64_t txg) 443219089Spjd{ 444219089Spjd lwb_t *lwb; 445219089Spjd 446219089Spjd lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 447219089Spjd lwb->lwb_zilog = zilog; 448219089Spjd lwb->lwb_blk = *bp; 449219089Spjd lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp)); 450219089Spjd lwb->lwb_max_txg = txg; 451219089Spjd lwb->lwb_zio = NULL; 452219089Spjd lwb->lwb_tx = NULL; 453219089Spjd if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) { 454219089Spjd lwb->lwb_nused = sizeof (zil_chain_t); 455219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp); 456219089Spjd } else { 457219089Spjd lwb->lwb_nused = 0; 458219089Spjd lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t); 459168404Spjd } 460219089Spjd 461219089Spjd mutex_enter(&zilog->zl_lock); 462219089Spjd list_insert_tail(&zilog->zl_lwb_list, lwb); 463219089Spjd mutex_exit(&zilog->zl_lock); 464219089Spjd 465219089Spjd return (lwb); 466168404Spjd} 467168404Spjd 468168404Spjd/* 469239620Smm * Called when we create in-memory log transactions so that we know 470239620Smm * to cleanup the itxs at the end of spa_sync(). 471239620Smm */ 472239620Smmvoid 473239620Smmzilog_dirty(zilog_t *zilog, uint64_t txg) 474239620Smm{ 475239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 476239620Smm dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os); 477239620Smm 478239620Smm if (dsl_dataset_is_snapshot(ds)) 479239620Smm panic("dirtying snapshot!"); 480239620Smm 481239620Smm if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg) == 0) { 482239620Smm /* up the hold count until we can be written out */ 483239620Smm dmu_buf_add_ref(ds->ds_dbuf, zilog); 484239620Smm } 485239620Smm} 486239620Smm 487239620Smmboolean_t 488239620Smmzilog_is_dirty(zilog_t *zilog) 489239620Smm{ 490239620Smm dsl_pool_t *dp = zilog->zl_dmu_pool; 491239620Smm 492239620Smm for (int t = 0; t < TXG_SIZE; t++) { 493239620Smm if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t)) 494239620Smm return (B_TRUE); 495239620Smm } 496239620Smm return (B_FALSE); 497239620Smm} 498239620Smm 499239620Smm/* 500168404Spjd * Create an on-disk intent log. 501168404Spjd */ 502219089Spjdstatic lwb_t * 503168404Spjdzil_create(zilog_t *zilog) 504168404Spjd{ 505168404Spjd const zil_header_t *zh = zilog->zl_header; 506219089Spjd lwb_t *lwb = NULL; 507168404Spjd uint64_t txg = 0; 508168404Spjd dmu_tx_t *tx = NULL; 509168404Spjd blkptr_t blk; 510168404Spjd int error = 0; 511168404Spjd 512168404Spjd /* 513168404Spjd * Wait for any previous destroy to complete. 514168404Spjd */ 515168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 516168404Spjd 517168404Spjd ASSERT(zh->zh_claim_txg == 0); 518168404Spjd ASSERT(zh->zh_replay_seq == 0); 519168404Spjd 520168404Spjd blk = zh->zh_log; 521168404Spjd 522168404Spjd /* 523219089Spjd * Allocate an initial log block if: 524219089Spjd * - there isn't one already 525219089Spjd * - the existing block is the wrong endianess 526168404Spjd */ 527207908Smm if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 528168404Spjd tx = dmu_tx_create(zilog->zl_os); 529219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 530168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 531168404Spjd txg = dmu_tx_get_txg(tx); 532168404Spjd 533207908Smm if (!BP_IS_HOLE(&blk)) { 534219089Spjd zio_free_zil(zilog->zl_spa, txg, &blk); 535207908Smm BP_ZERO(&blk); 536207908Smm } 537207908Smm 538219089Spjd error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 539219089Spjd ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 540168404Spjd 541168404Spjd if (error == 0) 542168404Spjd zil_init_log_chain(zilog, &blk); 543168404Spjd } 544168404Spjd 545168404Spjd /* 546168404Spjd * Allocate a log write buffer (lwb) for the first log block. 547168404Spjd */ 548219089Spjd if (error == 0) 549219089Spjd lwb = zil_alloc_lwb(zilog, &blk, txg); 550168404Spjd 551168404Spjd /* 552168404Spjd * If we just allocated the first log block, commit our transaction 553168404Spjd * and wait for zil_sync() to stuff the block poiner into zh_log. 554168404Spjd * (zh is part of the MOS, so we cannot modify it in open context.) 555168404Spjd */ 556168404Spjd if (tx != NULL) { 557168404Spjd dmu_tx_commit(tx); 558168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 559168404Spjd } 560168404Spjd 561168404Spjd ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 562219089Spjd 563219089Spjd return (lwb); 564168404Spjd} 565168404Spjd 566168404Spjd/* 567168404Spjd * In one tx, free all log blocks and clear the log header. 568168404Spjd * If keep_first is set, then we're replaying a log with no content. 569168404Spjd * We want to keep the first block, however, so that the first 570168404Spjd * synchronous transaction doesn't require a txg_wait_synced() 571168404Spjd * in zil_create(). We don't need to txg_wait_synced() here either 572168404Spjd * when keep_first is set, because both zil_create() and zil_destroy() 573168404Spjd * will wait for any in-progress destroys to complete. 574168404Spjd */ 575168404Spjdvoid 576168404Spjdzil_destroy(zilog_t *zilog, boolean_t keep_first) 577168404Spjd{ 578168404Spjd const zil_header_t *zh = zilog->zl_header; 579168404Spjd lwb_t *lwb; 580168404Spjd dmu_tx_t *tx; 581168404Spjd uint64_t txg; 582168404Spjd 583168404Spjd /* 584168404Spjd * Wait for any previous destroy to complete. 585168404Spjd */ 586168404Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 587168404Spjd 588219089Spjd zilog->zl_old_header = *zh; /* debugging aid */ 589219089Spjd 590168404Spjd if (BP_IS_HOLE(&zh->zh_log)) 591168404Spjd return; 592168404Spjd 593168404Spjd tx = dmu_tx_create(zilog->zl_os); 594219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 595168404Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 596168404Spjd txg = dmu_tx_get_txg(tx); 597168404Spjd 598168404Spjd mutex_enter(&zilog->zl_lock); 599168404Spjd 600168404Spjd ASSERT3U(zilog->zl_destroy_txg, <, txg); 601168404Spjd zilog->zl_destroy_txg = txg; 602168404Spjd zilog->zl_keep_first = keep_first; 603168404Spjd 604168404Spjd if (!list_is_empty(&zilog->zl_lwb_list)) { 605168404Spjd ASSERT(zh->zh_claim_txg == 0); 606224526Smm VERIFY(!keep_first); 607168404Spjd while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 608168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 609168404Spjd if (lwb->lwb_buf != NULL) 610168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 611219089Spjd zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 612168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 613168404Spjd } 614219089Spjd } else if (!keep_first) { 615239620Smm zil_destroy_sync(zilog, tx); 616168404Spjd } 617168404Spjd mutex_exit(&zilog->zl_lock); 618168404Spjd 619168404Spjd dmu_tx_commit(tx); 620185029Spjd} 621168404Spjd 622239620Smmvoid 623239620Smmzil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx) 624239620Smm{ 625239620Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 626239620Smm (void) zil_parse(zilog, zil_free_log_block, 627239620Smm zil_free_log_record, tx, zilog->zl_header->zh_claim_txg); 628239620Smm} 629239620Smm 630168404Spjdint 631219089Spjdzil_claim(const char *osname, void *txarg) 632168404Spjd{ 633168404Spjd dmu_tx_t *tx = txarg; 634168404Spjd uint64_t first_txg = dmu_tx_get_txg(tx); 635168404Spjd zilog_t *zilog; 636168404Spjd zil_header_t *zh; 637168404Spjd objset_t *os; 638168404Spjd int error; 639168404Spjd 640219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 641168404Spjd if (error) { 642185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 643168404Spjd return (0); 644168404Spjd } 645168404Spjd 646168404Spjd zilog = dmu_objset_zil(os); 647168404Spjd zh = zil_header_in_syncing_context(zilog); 648168404Spjd 649219089Spjd if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 650213197Smm if (!BP_IS_HOLE(&zh->zh_log)) 651219089Spjd zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 652213197Smm BP_ZERO(&zh->zh_log); 653213197Smm dsl_dataset_dirty(dmu_objset_ds(os), tx); 654219089Spjd dmu_objset_rele(os, FTAG); 655219089Spjd return (0); 656213197Smm } 657213197Smm 658168404Spjd /* 659168404Spjd * Claim all log blocks if we haven't already done so, and remember 660168404Spjd * the highest claimed sequence number. This ensures that if we can 661168404Spjd * read only part of the log now (e.g. due to a missing device), 662168404Spjd * but we can read the entire log later, we will not try to replay 663168404Spjd * or destroy beyond the last block we successfully claimed. 664168404Spjd */ 665168404Spjd ASSERT3U(zh->zh_claim_txg, <=, first_txg); 666168404Spjd if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 667219089Spjd (void) zil_parse(zilog, zil_claim_log_block, 668219089Spjd zil_claim_log_record, tx, first_txg); 669168404Spjd zh->zh_claim_txg = first_txg; 670219089Spjd zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 671219089Spjd zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 672219089Spjd if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 673219089Spjd zh->zh_flags |= ZIL_REPLAY_NEEDED; 674219089Spjd zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 675168404Spjd dsl_dataset_dirty(dmu_objset_ds(os), tx); 676168404Spjd } 677168404Spjd 678168404Spjd ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 679219089Spjd dmu_objset_rele(os, FTAG); 680168404Spjd return (0); 681168404Spjd} 682168404Spjd 683185029Spjd/* 684185029Spjd * Check the log by walking the log chain. 685185029Spjd * Checksum errors are ok as they indicate the end of the chain. 686185029Spjd * Any other error (no device or read failure) returns an error. 687185029Spjd */ 688185029Spjdint 689219089Spjdzil_check_log_chain(const char *osname, void *tx) 690168404Spjd{ 691185029Spjd zilog_t *zilog; 692185029Spjd objset_t *os; 693219089Spjd blkptr_t *bp; 694185029Spjd int error; 695168404Spjd 696219089Spjd ASSERT(tx == NULL); 697219089Spjd 698219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 699185029Spjd if (error) { 700185029Spjd cmn_err(CE_WARN, "can't open objset for %s", osname); 701185029Spjd return (0); 702185029Spjd } 703168404Spjd 704185029Spjd zilog = dmu_objset_zil(os); 705219089Spjd bp = (blkptr_t *)&zilog->zl_header->zh_log; 706219089Spjd 707219089Spjd /* 708219089Spjd * Check the first block and determine if it's on a log device 709219089Spjd * which may have been removed or faulted prior to loading this 710219089Spjd * pool. If so, there's no point in checking the rest of the log 711219089Spjd * as its content should have already been synced to the pool. 712219089Spjd */ 713219089Spjd if (!BP_IS_HOLE(bp)) { 714219089Spjd vdev_t *vd; 715219089Spjd boolean_t valid = B_TRUE; 716219089Spjd 717219089Spjd spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER); 718219089Spjd vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0])); 719219089Spjd if (vd->vdev_islog && vdev_is_dead(vd)) 720219089Spjd valid = vdev_log_state_valid(vd); 721219089Spjd spa_config_exit(os->os_spa, SCL_STATE, FTAG); 722219089Spjd 723219089Spjd if (!valid) { 724219089Spjd dmu_objset_rele(os, FTAG); 725219089Spjd return (0); 726219089Spjd } 727168404Spjd } 728185029Spjd 729219089Spjd /* 730219089Spjd * Because tx == NULL, zil_claim_log_block() will not actually claim 731219089Spjd * any blocks, but just determine whether it is possible to do so. 732219089Spjd * In addition to checking the log chain, zil_claim_log_block() 733219089Spjd * will invoke zio_claim() with a done func of spa_claim_notify(), 734219089Spjd * which will update spa_max_claim_txg. See spa_load() for details. 735219089Spjd */ 736219089Spjd error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 737219089Spjd zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 738219089Spjd 739219089Spjd dmu_objset_rele(os, FTAG); 740219089Spjd 741219089Spjd return ((error == ECKSUM || error == ENOENT) ? 0 : error); 742168404Spjd} 743168404Spjd 744185029Spjdstatic int 745185029Spjdzil_vdev_compare(const void *x1, const void *x2) 746185029Spjd{ 747219089Spjd const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 748219089Spjd const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 749185029Spjd 750185029Spjd if (v1 < v2) 751185029Spjd return (-1); 752185029Spjd if (v1 > v2) 753185029Spjd return (1); 754185029Spjd 755185029Spjd return (0); 756185029Spjd} 757185029Spjd 758168404Spjdvoid 759219089Spjdzil_add_block(zilog_t *zilog, const blkptr_t *bp) 760168404Spjd{ 761185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 762185029Spjd avl_index_t where; 763185029Spjd zil_vdev_node_t *zv, zvsearch; 764185029Spjd int ndvas = BP_GET_NDVAS(bp); 765185029Spjd int i; 766168404Spjd 767185029Spjd if (zfs_nocacheflush) 768185029Spjd return; 769168404Spjd 770185029Spjd ASSERT(zilog->zl_writer); 771168404Spjd 772185029Spjd /* 773185029Spjd * Even though we're zl_writer, we still need a lock because the 774185029Spjd * zl_get_data() callbacks may have dmu_sync() done callbacks 775185029Spjd * that will run concurrently. 776185029Spjd */ 777185029Spjd mutex_enter(&zilog->zl_vdev_lock); 778185029Spjd for (i = 0; i < ndvas; i++) { 779185029Spjd zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 780185029Spjd if (avl_find(t, &zvsearch, &where) == NULL) { 781185029Spjd zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 782185029Spjd zv->zv_vdev = zvsearch.zv_vdev; 783185029Spjd avl_insert(t, zv, where); 784185029Spjd } 785185029Spjd } 786185029Spjd mutex_exit(&zilog->zl_vdev_lock); 787168404Spjd} 788168404Spjd 789219089Spjdstatic void 790168404Spjdzil_flush_vdevs(zilog_t *zilog) 791168404Spjd{ 792168404Spjd spa_t *spa = zilog->zl_spa; 793185029Spjd avl_tree_t *t = &zilog->zl_vdev_tree; 794185029Spjd void *cookie = NULL; 795185029Spjd zil_vdev_node_t *zv; 796185029Spjd zio_t *zio; 797168404Spjd 798168404Spjd ASSERT(zilog->zl_writer); 799168404Spjd 800185029Spjd /* 801185029Spjd * We don't need zl_vdev_lock here because we're the zl_writer, 802185029Spjd * and all zl_get_data() callbacks are done. 803185029Spjd */ 804185029Spjd if (avl_numnodes(t) == 0) 805185029Spjd return; 806185029Spjd 807185029Spjd spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 808185029Spjd 809185029Spjd zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 810185029Spjd 811185029Spjd while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 812185029Spjd vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 813185029Spjd if (vd != NULL) 814185029Spjd zio_flush(zio, vd); 815185029Spjd kmem_free(zv, sizeof (*zv)); 816168404Spjd } 817168404Spjd 818168404Spjd /* 819168404Spjd * Wait for all the flushes to complete. Not all devices actually 820168404Spjd * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 821168404Spjd */ 822185029Spjd (void) zio_wait(zio); 823185029Spjd 824185029Spjd spa_config_exit(spa, SCL_STATE, FTAG); 825168404Spjd} 826168404Spjd 827168404Spjd/* 828168404Spjd * Function called when a log block write completes 829168404Spjd */ 830168404Spjdstatic void 831168404Spjdzil_lwb_write_done(zio_t *zio) 832168404Spjd{ 833168404Spjd lwb_t *lwb = zio->io_private; 834168404Spjd zilog_t *zilog = lwb->lwb_zilog; 835219089Spjd dmu_tx_t *tx = lwb->lwb_tx; 836168404Spjd 837185029Spjd ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 838185029Spjd ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 839185029Spjd ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 840185029Spjd ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 841185029Spjd ASSERT(!BP_IS_GANG(zio->io_bp)); 842185029Spjd ASSERT(!BP_IS_HOLE(zio->io_bp)); 843185029Spjd ASSERT(zio->io_bp->blk_fill == 0); 844185029Spjd 845168404Spjd /* 846209962Smm * Ensure the lwb buffer pointer is cleared before releasing 847209962Smm * the txg. If we have had an allocation failure and 848209962Smm * the txg is waiting to sync then we want want zil_sync() 849209962Smm * to remove the lwb so that it's not picked up as the next new 850209962Smm * one in zil_commit_writer(). zil_sync() will only remove 851209962Smm * the lwb if lwb_buf is null. 852168404Spjd */ 853168404Spjd zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 854168404Spjd mutex_enter(&zilog->zl_lock); 855168404Spjd lwb->lwb_buf = NULL; 856219089Spjd lwb->lwb_tx = NULL; 857219089Spjd mutex_exit(&zilog->zl_lock); 858209962Smm 859209962Smm /* 860209962Smm * Now that we've written this log block, we have a stable pointer 861209962Smm * to the next block in the chain, so it's OK to let the txg in 862219089Spjd * which we allocated the next block sync. 863209962Smm */ 864219089Spjd dmu_tx_commit(tx); 865168404Spjd} 866168404Spjd 867168404Spjd/* 868168404Spjd * Initialize the io for a log block. 869168404Spjd */ 870168404Spjdstatic void 871168404Spjdzil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 872168404Spjd{ 873168404Spjd zbookmark_t zb; 874168404Spjd 875219089Spjd SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 876219089Spjd ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 877219089Spjd lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 878168404Spjd 879168404Spjd if (zilog->zl_root_zio == NULL) { 880168404Spjd zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 881168404Spjd ZIO_FLAG_CANFAIL); 882168404Spjd } 883168404Spjd if (lwb->lwb_zio == NULL) { 884168404Spjd lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 885219089Spjd 0, &lwb->lwb_blk, lwb->lwb_buf, BP_GET_LSIZE(&lwb->lwb_blk), 886213197Smm zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 887219089Spjd ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 888168404Spjd } 889168404Spjd} 890168404Spjd 891168404Spjd/* 892219089Spjd * Define a limited set of intent log block sizes. 893219089Spjd * These must be a multiple of 4KB. Note only the amount used (again 894219089Spjd * aligned to 4KB) actually gets written. However, we can't always just 895219089Spjd * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted. 896219089Spjd */ 897219089Spjduint64_t zil_block_buckets[] = { 898219089Spjd 4096, /* non TX_WRITE */ 899219089Spjd 8192+4096, /* data base */ 900219089Spjd 32*1024 + 4096, /* NFS writes */ 901219089Spjd UINT64_MAX 902219089Spjd}; 903219089Spjd 904219089Spjd/* 905219089Spjd * Use the slog as long as the logbias is 'latency' and the current commit size 906219089Spjd * is less than the limit or the total list size is less than 2X the limit. 907219089Spjd * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 908219089Spjd */ 909219089Spjduint64_t zil_slog_limit = 1024 * 1024; 910219089Spjd#define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 911219089Spjd (((zilog)->zl_cur_used < zil_slog_limit) || \ 912219089Spjd ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 913219089Spjd 914219089Spjd/* 915168404Spjd * Start a log block write and advance to the next log block. 916168404Spjd * Calls are serialized. 917168404Spjd */ 918168404Spjdstatic lwb_t * 919168404Spjdzil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 920168404Spjd{ 921219089Spjd lwb_t *nlwb = NULL; 922219089Spjd zil_chain_t *zilc; 923168404Spjd spa_t *spa = zilog->zl_spa; 924219089Spjd blkptr_t *bp; 925219089Spjd dmu_tx_t *tx; 926168404Spjd uint64_t txg; 927219089Spjd uint64_t zil_blksz, wsz; 928219089Spjd int i, error; 929168404Spjd 930219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 931219089Spjd zilc = (zil_chain_t *)lwb->lwb_buf; 932219089Spjd bp = &zilc->zc_next_blk; 933219089Spjd } else { 934219089Spjd zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz); 935219089Spjd bp = &zilc->zc_next_blk; 936219089Spjd } 937168404Spjd 938219089Spjd ASSERT(lwb->lwb_nused <= lwb->lwb_sz); 939219089Spjd 940168404Spjd /* 941168404Spjd * Allocate the next block and save its address in this block 942168404Spjd * before writing it in order to establish the log chain. 943168404Spjd * Note that if the allocation of nlwb synced before we wrote 944168404Spjd * the block that points at it (lwb), we'd leak it if we crashed. 945219089Spjd * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 946219089Spjd * We dirty the dataset to ensure that zil_sync() will be called 947219089Spjd * to clean up in the event of allocation failure or I/O failure. 948168404Spjd */ 949219089Spjd tx = dmu_tx_create(zilog->zl_os); 950219089Spjd VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 951219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 952219089Spjd txg = dmu_tx_get_txg(tx); 953168404Spjd 954219089Spjd lwb->lwb_tx = tx; 955219089Spjd 956168404Spjd /* 957219089Spjd * Log blocks are pre-allocated. Here we select the size of the next 958219089Spjd * block, based on size used in the last block. 959219089Spjd * - first find the smallest bucket that will fit the block from a 960219089Spjd * limited set of block sizes. This is because it's faster to write 961219089Spjd * blocks allocated from the same metaslab as they are adjacent or 962219089Spjd * close. 963219089Spjd * - next find the maximum from the new suggested size and an array of 964219089Spjd * previous sizes. This lessens a picket fence effect of wrongly 965219089Spjd * guesssing the size if we have a stream of say 2k, 64k, 2k, 64k 966219089Spjd * requests. 967219089Spjd * 968219089Spjd * Note we only write what is used, but we can't just allocate 969219089Spjd * the maximum block size because we can exhaust the available 970219089Spjd * pool log space. 971168404Spjd */ 972219089Spjd zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t); 973219089Spjd for (i = 0; zil_blksz > zil_block_buckets[i]; i++) 974219089Spjd continue; 975219089Spjd zil_blksz = zil_block_buckets[i]; 976219089Spjd if (zil_blksz == UINT64_MAX) 977219089Spjd zil_blksz = SPA_MAXBLOCKSIZE; 978219089Spjd zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz; 979219089Spjd for (i = 0; i < ZIL_PREV_BLKS; i++) 980219089Spjd zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]); 981219089Spjd zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1); 982168404Spjd 983168404Spjd BP_ZERO(bp); 984168404Spjd /* pass the old blkptr in order to spread log blocks across devs */ 985219089Spjd error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 986219089Spjd USE_SLOG(zilog)); 987219089Spjd if (!error) { 988219089Spjd ASSERT3U(bp->blk_birth, ==, txg); 989219089Spjd bp->blk_cksum = lwb->lwb_blk.blk_cksum; 990219089Spjd bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 991168404Spjd 992168404Spjd /* 993219089Spjd * Allocate a new log write buffer (lwb). 994168404Spjd */ 995219089Spjd nlwb = zil_alloc_lwb(zilog, bp, txg); 996168404Spjd 997219089Spjd /* Record the block for later vdev flushing */ 998219089Spjd zil_add_block(zilog, &lwb->lwb_blk); 999168404Spjd } 1000168404Spjd 1001219089Spjd if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) { 1002219089Spjd /* For Slim ZIL only write what is used. */ 1003219089Spjd wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, uint64_t); 1004219089Spjd ASSERT3U(wsz, <=, lwb->lwb_sz); 1005219089Spjd zio_shrink(lwb->lwb_zio, wsz); 1006168404Spjd 1007219089Spjd } else { 1008219089Spjd wsz = lwb->lwb_sz; 1009219089Spjd } 1010168404Spjd 1011219089Spjd zilc->zc_pad = 0; 1012219089Spjd zilc->zc_nused = lwb->lwb_nused; 1013219089Spjd zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum; 1014168404Spjd 1015168404Spjd /* 1016219089Spjd * clear unused data for security 1017168404Spjd */ 1018219089Spjd bzero(lwb->lwb_buf + lwb->lwb_nused, wsz - lwb->lwb_nused); 1019168404Spjd 1020219089Spjd zio_nowait(lwb->lwb_zio); /* Kick off the write for the old log block */ 1021168404Spjd 1022168404Spjd /* 1023219089Spjd * If there was an allocation failure then nlwb will be null which 1024219089Spjd * forces a txg_wait_synced(). 1025168404Spjd */ 1026168404Spjd return (nlwb); 1027168404Spjd} 1028168404Spjd 1029168404Spjdstatic lwb_t * 1030168404Spjdzil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 1031168404Spjd{ 1032168404Spjd lr_t *lrc = &itx->itx_lr; /* common log record */ 1033219089Spjd lr_write_t *lrw = (lr_write_t *)lrc; 1034219089Spjd char *lr_buf; 1035168404Spjd uint64_t txg = lrc->lrc_txg; 1036168404Spjd uint64_t reclen = lrc->lrc_reclen; 1037219089Spjd uint64_t dlen = 0; 1038168404Spjd 1039168404Spjd if (lwb == NULL) 1040168404Spjd return (NULL); 1041219089Spjd 1042168404Spjd ASSERT(lwb->lwb_buf != NULL); 1043239620Smm ASSERT(zilog_is_dirty(zilog) || 1044239620Smm spa_freeze_txg(zilog->zl_spa) != UINT64_MAX); 1045168404Spjd 1046168404Spjd if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 1047168404Spjd dlen = P2ROUNDUP_TYPED( 1048219089Spjd lrw->lr_length, sizeof (uint64_t), uint64_t); 1049168404Spjd 1050168404Spjd zilog->zl_cur_used += (reclen + dlen); 1051168404Spjd 1052168404Spjd zil_lwb_write_init(zilog, lwb); 1053168404Spjd 1054168404Spjd /* 1055168404Spjd * If this record won't fit in the current log block, start a new one. 1056168404Spjd */ 1057219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1058168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1059168404Spjd if (lwb == NULL) 1060168404Spjd return (NULL); 1061168404Spjd zil_lwb_write_init(zilog, lwb); 1062219089Spjd ASSERT(LWB_EMPTY(lwb)); 1063219089Spjd if (lwb->lwb_nused + reclen + dlen > lwb->lwb_sz) { 1064168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1065168404Spjd return (lwb); 1066168404Spjd } 1067168404Spjd } 1068168404Spjd 1069219089Spjd lr_buf = lwb->lwb_buf + lwb->lwb_nused; 1070219089Spjd bcopy(lrc, lr_buf, reclen); 1071219089Spjd lrc = (lr_t *)lr_buf; 1072219089Spjd lrw = (lr_write_t *)lrc; 1073168404Spjd 1074168404Spjd /* 1075168404Spjd * If it's a write, fetch the data or get its blkptr as appropriate. 1076168404Spjd */ 1077168404Spjd if (lrc->lrc_txtype == TX_WRITE) { 1078168404Spjd if (txg > spa_freeze_txg(zilog->zl_spa)) 1079168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1080168404Spjd if (itx->itx_wr_state != WR_COPIED) { 1081168404Spjd char *dbuf; 1082168404Spjd int error; 1083168404Spjd 1084168404Spjd if (dlen) { 1085168404Spjd ASSERT(itx->itx_wr_state == WR_NEED_COPY); 1086219089Spjd dbuf = lr_buf + reclen; 1087219089Spjd lrw->lr_common.lrc_reclen += dlen; 1088168404Spjd } else { 1089168404Spjd ASSERT(itx->itx_wr_state == WR_INDIRECT); 1090168404Spjd dbuf = NULL; 1091168404Spjd } 1092168404Spjd error = zilog->zl_get_data( 1093219089Spjd itx->itx_private, lrw, dbuf, lwb->lwb_zio); 1094214378Smm if (error == EIO) { 1095214378Smm txg_wait_synced(zilog->zl_dmu_pool, txg); 1096214378Smm return (lwb); 1097214378Smm } 1098168404Spjd if (error) { 1099168404Spjd ASSERT(error == ENOENT || error == EEXIST || 1100168404Spjd error == EALREADY); 1101168404Spjd return (lwb); 1102168404Spjd } 1103168404Spjd } 1104168404Spjd } 1105168404Spjd 1106219089Spjd /* 1107219089Spjd * We're actually making an entry, so update lrc_seq to be the 1108219089Spjd * log record sequence number. Note that this is generally not 1109219089Spjd * equal to the itx sequence number because not all transactions 1110219089Spjd * are synchronous, and sometimes spa_sync() gets there first. 1111219089Spjd */ 1112219089Spjd lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 1113168404Spjd lwb->lwb_nused += reclen + dlen; 1114168404Spjd lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 1115219089Spjd ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz); 1116240415Smm ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t))); 1117168404Spjd 1118168404Spjd return (lwb); 1119168404Spjd} 1120168404Spjd 1121168404Spjditx_t * 1122185029Spjdzil_itx_create(uint64_t txtype, size_t lrsize) 1123168404Spjd{ 1124168404Spjd itx_t *itx; 1125168404Spjd 1126168404Spjd lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 1127168404Spjd 1128168404Spjd itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 1129168404Spjd itx->itx_lr.lrc_txtype = txtype; 1130168404Spjd itx->itx_lr.lrc_reclen = lrsize; 1131185029Spjd itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 1132168404Spjd itx->itx_lr.lrc_seq = 0; /* defensive */ 1133219089Spjd itx->itx_sync = B_TRUE; /* default is synchronous */ 1134168404Spjd 1135168404Spjd return (itx); 1136168404Spjd} 1137168404Spjd 1138219089Spjdvoid 1139219089Spjdzil_itx_destroy(itx_t *itx) 1140168404Spjd{ 1141219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 1142219089Spjd} 1143168404Spjd 1144219089Spjd/* 1145219089Spjd * Free up the sync and async itxs. The itxs_t has already been detached 1146219089Spjd * so no locks are needed. 1147219089Spjd */ 1148219089Spjdstatic void 1149219089Spjdzil_itxg_clean(itxs_t *itxs) 1150219089Spjd{ 1151219089Spjd itx_t *itx; 1152219089Spjd list_t *list; 1153219089Spjd avl_tree_t *t; 1154219089Spjd void *cookie; 1155219089Spjd itx_async_node_t *ian; 1156168404Spjd 1157219089Spjd list = &itxs->i_sync_list; 1158219089Spjd while ((itx = list_head(list)) != NULL) { 1159219089Spjd list_remove(list, itx); 1160219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1161219089Spjd itx->itx_lr.lrc_reclen); 1162219089Spjd } 1163168404Spjd 1164219089Spjd cookie = NULL; 1165219089Spjd t = &itxs->i_async_tree; 1166219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1167219089Spjd list = &ian->ia_list; 1168219089Spjd while ((itx = list_head(list)) != NULL) { 1169219089Spjd list_remove(list, itx); 1170219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1171219089Spjd itx->itx_lr.lrc_reclen); 1172219089Spjd } 1173219089Spjd list_destroy(list); 1174219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1175219089Spjd } 1176219089Spjd avl_destroy(t); 1177219089Spjd 1178219089Spjd kmem_free(itxs, sizeof (itxs_t)); 1179168404Spjd} 1180168404Spjd 1181219089Spjdstatic int 1182219089Spjdzil_aitx_compare(const void *x1, const void *x2) 1183219089Spjd{ 1184219089Spjd const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid; 1185219089Spjd const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid; 1186219089Spjd 1187219089Spjd if (o1 < o2) 1188219089Spjd return (-1); 1189219089Spjd if (o1 > o2) 1190219089Spjd return (1); 1191219089Spjd 1192219089Spjd return (0); 1193219089Spjd} 1194219089Spjd 1195168404Spjd/* 1196219089Spjd * Remove all async itx with the given oid. 1197168404Spjd */ 1198168404Spjdstatic void 1199219089Spjdzil_remove_async(zilog_t *zilog, uint64_t oid) 1200168404Spjd{ 1201219089Spjd uint64_t otxg, txg; 1202219089Spjd itx_async_node_t *ian; 1203219089Spjd avl_tree_t *t; 1204219089Spjd avl_index_t where; 1205168404Spjd list_t clean_list; 1206168404Spjd itx_t *itx; 1207168404Spjd 1208219089Spjd ASSERT(oid != 0); 1209168404Spjd list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1210168404Spjd 1211219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1212219089Spjd otxg = ZILTEST_TXG; 1213219089Spjd else 1214219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1215219089Spjd 1216219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1217219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1218219089Spjd 1219219089Spjd mutex_enter(&itxg->itxg_lock); 1220219089Spjd if (itxg->itxg_txg != txg) { 1221219089Spjd mutex_exit(&itxg->itxg_lock); 1222219089Spjd continue; 1223219089Spjd } 1224219089Spjd 1225219089Spjd /* 1226219089Spjd * Locate the object node and append its list. 1227219089Spjd */ 1228219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1229219089Spjd ian = avl_find(t, &oid, &where); 1230219089Spjd if (ian != NULL) 1231219089Spjd list_move_tail(&clean_list, &ian->ia_list); 1232219089Spjd mutex_exit(&itxg->itxg_lock); 1233168404Spjd } 1234219089Spjd while ((itx = list_head(&clean_list)) != NULL) { 1235219089Spjd list_remove(&clean_list, itx); 1236219089Spjd kmem_free(itx, offsetof(itx_t, itx_lr) + 1237219089Spjd itx->itx_lr.lrc_reclen); 1238219089Spjd } 1239219089Spjd list_destroy(&clean_list); 1240219089Spjd} 1241168404Spjd 1242219089Spjdvoid 1243219089Spjdzil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 1244219089Spjd{ 1245219089Spjd uint64_t txg; 1246219089Spjd itxg_t *itxg; 1247219089Spjd itxs_t *itxs, *clean = NULL; 1248219089Spjd 1249168404Spjd /* 1250219089Spjd * Object ids can be re-instantiated in the next txg so 1251219089Spjd * remove any async transactions to avoid future leaks. 1252219089Spjd * This can happen if a fsync occurs on the re-instantiated 1253219089Spjd * object for a WR_INDIRECT or WR_NEED_COPY write, which gets 1254219089Spjd * the new file data and flushes a write record for the old object. 1255168404Spjd */ 1256219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_REMOVE) 1257219089Spjd zil_remove_async(zilog, itx->itx_oid); 1258219089Spjd 1259219089Spjd /* 1260219089Spjd * Ensure the data of a renamed file is committed before the rename. 1261219089Spjd */ 1262219089Spjd if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME) 1263219089Spjd zil_async_to_sync(zilog, itx->itx_oid); 1264219089Spjd 1265239620Smm if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) 1266219089Spjd txg = ZILTEST_TXG; 1267219089Spjd else 1268219089Spjd txg = dmu_tx_get_txg(tx); 1269219089Spjd 1270219089Spjd itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1271219089Spjd mutex_enter(&itxg->itxg_lock); 1272219089Spjd itxs = itxg->itxg_itxs; 1273219089Spjd if (itxg->itxg_txg != txg) { 1274219089Spjd if (itxs != NULL) { 1275219089Spjd /* 1276219089Spjd * The zil_clean callback hasn't got around to cleaning 1277219089Spjd * this itxg. Save the itxs for release below. 1278219089Spjd * This should be rare. 1279219089Spjd */ 1280219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1281219089Spjd itxg->itxg_sod = 0; 1282219089Spjd clean = itxg->itxg_itxs; 1283219089Spjd } 1284219089Spjd ASSERT(itxg->itxg_sod == 0); 1285219089Spjd itxg->itxg_txg = txg; 1286219089Spjd itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t), KM_SLEEP); 1287219089Spjd 1288219089Spjd list_create(&itxs->i_sync_list, sizeof (itx_t), 1289219089Spjd offsetof(itx_t, itx_node)); 1290219089Spjd avl_create(&itxs->i_async_tree, zil_aitx_compare, 1291219089Spjd sizeof (itx_async_node_t), 1292219089Spjd offsetof(itx_async_node_t, ia_node)); 1293168404Spjd } 1294219089Spjd if (itx->itx_sync) { 1295219089Spjd list_insert_tail(&itxs->i_sync_list, itx); 1296219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, itx->itx_sod); 1297219089Spjd itxg->itxg_sod += itx->itx_sod; 1298219089Spjd } else { 1299219089Spjd avl_tree_t *t = &itxs->i_async_tree; 1300219089Spjd uint64_t foid = ((lr_ooo_t *)&itx->itx_lr)->lr_foid; 1301219089Spjd itx_async_node_t *ian; 1302219089Spjd avl_index_t where; 1303168404Spjd 1304219089Spjd ian = avl_find(t, &foid, &where); 1305219089Spjd if (ian == NULL) { 1306219089Spjd ian = kmem_alloc(sizeof (itx_async_node_t), KM_SLEEP); 1307219089Spjd list_create(&ian->ia_list, sizeof (itx_t), 1308219089Spjd offsetof(itx_t, itx_node)); 1309219089Spjd ian->ia_foid = foid; 1310219089Spjd avl_insert(t, ian, where); 1311219089Spjd } 1312219089Spjd list_insert_tail(&ian->ia_list, itx); 1313168404Spjd } 1314219089Spjd 1315219089Spjd itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1316239620Smm zilog_dirty(zilog, txg); 1317219089Spjd mutex_exit(&itxg->itxg_lock); 1318219089Spjd 1319219089Spjd /* Release the old itxs now we've dropped the lock */ 1320219089Spjd if (clean != NULL) 1321219089Spjd zil_itxg_clean(clean); 1322168404Spjd} 1323168404Spjd 1324168404Spjd/* 1325168404Spjd * If there are any in-memory intent log transactions which have now been 1326239620Smm * synced then start up a taskq to free them. We should only do this after we 1327239620Smm * have written out the uberblocks (i.e. txg has been comitted) so that 1328239620Smm * don't inadvertently clean out in-memory log records that would be required 1329239620Smm * by zil_commit(). 1330168404Spjd */ 1331168404Spjdvoid 1332219089Spjdzil_clean(zilog_t *zilog, uint64_t synced_txg) 1333168404Spjd{ 1334219089Spjd itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK]; 1335219089Spjd itxs_t *clean_me; 1336168404Spjd 1337219089Spjd mutex_enter(&itxg->itxg_lock); 1338219089Spjd if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) { 1339219089Spjd mutex_exit(&itxg->itxg_lock); 1340219089Spjd return; 1341168404Spjd } 1342219089Spjd ASSERT3U(itxg->itxg_txg, <=, synced_txg); 1343219089Spjd ASSERT(itxg->itxg_txg != 0); 1344219089Spjd ASSERT(zilog->zl_clean_taskq != NULL); 1345219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -itxg->itxg_sod); 1346219089Spjd itxg->itxg_sod = 0; 1347219089Spjd clean_me = itxg->itxg_itxs; 1348219089Spjd itxg->itxg_itxs = NULL; 1349219089Spjd itxg->itxg_txg = 0; 1350219089Spjd mutex_exit(&itxg->itxg_lock); 1351219089Spjd /* 1352219089Spjd * Preferably start a task queue to free up the old itxs but 1353219089Spjd * if taskq_dispatch can't allocate resources to do that then 1354219089Spjd * free it in-line. This should be rare. Note, using TQ_SLEEP 1355219089Spjd * created a bad performance problem. 1356219089Spjd */ 1357219089Spjd if (taskq_dispatch(zilog->zl_clean_taskq, 1358219089Spjd (void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0) 1359219089Spjd zil_itxg_clean(clean_me); 1360168404Spjd} 1361168404Spjd 1362219089Spjd/* 1363219089Spjd * Get the list of itxs to commit into zl_itx_commit_list. 1364219089Spjd */ 1365185029Spjdstatic void 1366219089Spjdzil_get_commit_list(zilog_t *zilog) 1367168404Spjd{ 1368219089Spjd uint64_t otxg, txg; 1369219089Spjd list_t *commit_list = &zilog->zl_itx_commit_list; 1370219089Spjd uint64_t push_sod = 0; 1371219089Spjd 1372219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1373219089Spjd otxg = ZILTEST_TXG; 1374219089Spjd else 1375219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1376219089Spjd 1377219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1378219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1379219089Spjd 1380219089Spjd mutex_enter(&itxg->itxg_lock); 1381219089Spjd if (itxg->itxg_txg != txg) { 1382219089Spjd mutex_exit(&itxg->itxg_lock); 1383219089Spjd continue; 1384219089Spjd } 1385219089Spjd 1386219089Spjd list_move_tail(commit_list, &itxg->itxg_itxs->i_sync_list); 1387219089Spjd push_sod += itxg->itxg_sod; 1388219089Spjd itxg->itxg_sod = 0; 1389219089Spjd 1390219089Spjd mutex_exit(&itxg->itxg_lock); 1391219089Spjd } 1392219089Spjd atomic_add_64(&zilog->zl_itx_list_sz, -push_sod); 1393219089Spjd} 1394219089Spjd 1395219089Spjd/* 1396219089Spjd * Move the async itxs for a specified object to commit into sync lists. 1397219089Spjd */ 1398219089Spjdstatic void 1399219089Spjdzil_async_to_sync(zilog_t *zilog, uint64_t foid) 1400219089Spjd{ 1401219089Spjd uint64_t otxg, txg; 1402219089Spjd itx_async_node_t *ian; 1403219089Spjd avl_tree_t *t; 1404219089Spjd avl_index_t where; 1405219089Spjd 1406219089Spjd if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */ 1407219089Spjd otxg = ZILTEST_TXG; 1408219089Spjd else 1409219089Spjd otxg = spa_last_synced_txg(zilog->zl_spa) + 1; 1410219089Spjd 1411219089Spjd for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) { 1412219089Spjd itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK]; 1413219089Spjd 1414219089Spjd mutex_enter(&itxg->itxg_lock); 1415219089Spjd if (itxg->itxg_txg != txg) { 1416219089Spjd mutex_exit(&itxg->itxg_lock); 1417219089Spjd continue; 1418219089Spjd } 1419219089Spjd 1420219089Spjd /* 1421219089Spjd * If a foid is specified then find that node and append its 1422219089Spjd * list. Otherwise walk the tree appending all the lists 1423219089Spjd * to the sync list. We add to the end rather than the 1424219089Spjd * beginning to ensure the create has happened. 1425219089Spjd */ 1426219089Spjd t = &itxg->itxg_itxs->i_async_tree; 1427219089Spjd if (foid != 0) { 1428219089Spjd ian = avl_find(t, &foid, &where); 1429219089Spjd if (ian != NULL) { 1430219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1431219089Spjd &ian->ia_list); 1432219089Spjd } 1433219089Spjd } else { 1434219089Spjd void *cookie = NULL; 1435219089Spjd 1436219089Spjd while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) { 1437219089Spjd list_move_tail(&itxg->itxg_itxs->i_sync_list, 1438219089Spjd &ian->ia_list); 1439219089Spjd list_destroy(&ian->ia_list); 1440219089Spjd kmem_free(ian, sizeof (itx_async_node_t)); 1441219089Spjd } 1442219089Spjd } 1443219089Spjd mutex_exit(&itxg->itxg_lock); 1444219089Spjd } 1445219089Spjd} 1446219089Spjd 1447219089Spjdstatic void 1448219089Spjdzil_commit_writer(zilog_t *zilog) 1449219089Spjd{ 1450168404Spjd uint64_t txg; 1451219089Spjd itx_t *itx; 1452168404Spjd lwb_t *lwb; 1453219089Spjd spa_t *spa = zilog->zl_spa; 1454219089Spjd int error = 0; 1455168404Spjd 1456185029Spjd ASSERT(zilog->zl_root_zio == NULL); 1457168404Spjd 1458219089Spjd mutex_exit(&zilog->zl_lock); 1459219089Spjd 1460219089Spjd zil_get_commit_list(zilog); 1461219089Spjd 1462219089Spjd /* 1463219089Spjd * Return if there's nothing to commit before we dirty the fs by 1464219089Spjd * calling zil_create(). 1465219089Spjd */ 1466219089Spjd if (list_head(&zilog->zl_itx_commit_list) == NULL) { 1467219089Spjd mutex_enter(&zilog->zl_lock); 1468219089Spjd return; 1469219089Spjd } 1470219089Spjd 1471168404Spjd if (zilog->zl_suspend) { 1472168404Spjd lwb = NULL; 1473168404Spjd } else { 1474168404Spjd lwb = list_tail(&zilog->zl_lwb_list); 1475219089Spjd if (lwb == NULL) 1476219089Spjd lwb = zil_create(zilog); 1477168404Spjd } 1478168404Spjd 1479168404Spjd DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1480219089Spjd while (itx = list_head(&zilog->zl_itx_commit_list)) { 1481168404Spjd txg = itx->itx_lr.lrc_txg; 1482168404Spjd ASSERT(txg); 1483168404Spjd 1484219089Spjd if (txg > spa_last_synced_txg(spa) || txg > spa_freeze_txg(spa)) 1485168404Spjd lwb = zil_lwb_commit(zilog, itx, lwb); 1486219089Spjd list_remove(&zilog->zl_itx_commit_list, itx); 1487168404Spjd kmem_free(itx, offsetof(itx_t, itx_lr) 1488168404Spjd + itx->itx_lr.lrc_reclen); 1489168404Spjd } 1490168404Spjd DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1491168404Spjd 1492168404Spjd /* write the last block out */ 1493168404Spjd if (lwb != NULL && lwb->lwb_zio != NULL) 1494168404Spjd lwb = zil_lwb_write_start(zilog, lwb); 1495168404Spjd 1496168404Spjd zilog->zl_cur_used = 0; 1497168404Spjd 1498168404Spjd /* 1499168404Spjd * Wait if necessary for the log blocks to be on stable storage. 1500168404Spjd */ 1501168404Spjd if (zilog->zl_root_zio) { 1502219089Spjd error = zio_wait(zilog->zl_root_zio); 1503185029Spjd zilog->zl_root_zio = NULL; 1504185029Spjd zil_flush_vdevs(zilog); 1505168404Spjd } 1506168404Spjd 1507219089Spjd if (error || lwb == NULL) 1508168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 1509168404Spjd 1510168404Spjd mutex_enter(&zilog->zl_lock); 1511168404Spjd 1512219089Spjd /* 1513219089Spjd * Remember the highest committed log sequence number for ztest. 1514219089Spjd * We only update this value when all the log writes succeeded, 1515219089Spjd * because ztest wants to ASSERT that it got the whole log chain. 1516219089Spjd */ 1517219089Spjd if (error == 0 && lwb != NULL) 1518219089Spjd zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1519168404Spjd} 1520168404Spjd 1521168404Spjd/* 1522219089Spjd * Commit zfs transactions to stable storage. 1523168404Spjd * If foid is 0 push out all transactions, otherwise push only those 1524219089Spjd * for that object or might reference that object. 1525219089Spjd * 1526219089Spjd * itxs are committed in batches. In a heavily stressed zil there will be 1527219089Spjd * a commit writer thread who is writing out a bunch of itxs to the log 1528219089Spjd * for a set of committing threads (cthreads) in the same batch as the writer. 1529219089Spjd * Those cthreads are all waiting on the same cv for that batch. 1530219089Spjd * 1531219089Spjd * There will also be a different and growing batch of threads that are 1532219089Spjd * waiting to commit (qthreads). When the committing batch completes 1533219089Spjd * a transition occurs such that the cthreads exit and the qthreads become 1534219089Spjd * cthreads. One of the new cthreads becomes the writer thread for the 1535219089Spjd * batch. Any new threads arriving become new qthreads. 1536219089Spjd * 1537219089Spjd * Only 2 condition variables are needed and there's no transition 1538219089Spjd * between the two cvs needed. They just flip-flop between qthreads 1539219089Spjd * and cthreads. 1540219089Spjd * 1541219089Spjd * Using this scheme we can efficiently wakeup up only those threads 1542219089Spjd * that have been committed. 1543168404Spjd */ 1544168404Spjdvoid 1545219089Spjdzil_commit(zilog_t *zilog, uint64_t foid) 1546168404Spjd{ 1547219089Spjd uint64_t mybatch; 1548219089Spjd 1549219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 1550168404Spjd return; 1551168404Spjd 1552219089Spjd /* move the async itxs for the foid to the sync queues */ 1553219089Spjd zil_async_to_sync(zilog, foid); 1554219089Spjd 1555168404Spjd mutex_enter(&zilog->zl_lock); 1556219089Spjd mybatch = zilog->zl_next_batch; 1557168404Spjd while (zilog->zl_writer) { 1558219089Spjd cv_wait(&zilog->zl_cv_batch[mybatch & 1], &zilog->zl_lock); 1559219089Spjd if (mybatch <= zilog->zl_com_batch) { 1560168404Spjd mutex_exit(&zilog->zl_lock); 1561168404Spjd return; 1562168404Spjd } 1563168404Spjd } 1564219089Spjd 1565219089Spjd zilog->zl_next_batch++; 1566219089Spjd zilog->zl_writer = B_TRUE; 1567219089Spjd zil_commit_writer(zilog); 1568219089Spjd zilog->zl_com_batch = mybatch; 1569219089Spjd zilog->zl_writer = B_FALSE; 1570168404Spjd mutex_exit(&zilog->zl_lock); 1571219089Spjd 1572219089Spjd /* wake up one thread to become the next writer */ 1573219089Spjd cv_signal(&zilog->zl_cv_batch[(mybatch+1) & 1]); 1574219089Spjd 1575219089Spjd /* wake up all threads waiting for this batch to be committed */ 1576219089Spjd cv_broadcast(&zilog->zl_cv_batch[mybatch & 1]); 1577168404Spjd} 1578168404Spjd 1579168404Spjd/* 1580168404Spjd * Called in syncing context to free committed log blocks and update log header. 1581168404Spjd */ 1582168404Spjdvoid 1583168404Spjdzil_sync(zilog_t *zilog, dmu_tx_t *tx) 1584168404Spjd{ 1585168404Spjd zil_header_t *zh = zil_header_in_syncing_context(zilog); 1586168404Spjd uint64_t txg = dmu_tx_get_txg(tx); 1587168404Spjd spa_t *spa = zilog->zl_spa; 1588219089Spjd uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1589168404Spjd lwb_t *lwb; 1590168404Spjd 1591209962Smm /* 1592209962Smm * We don't zero out zl_destroy_txg, so make sure we don't try 1593209962Smm * to destroy it twice. 1594209962Smm */ 1595209962Smm if (spa_sync_pass(spa) != 1) 1596209962Smm return; 1597209962Smm 1598168404Spjd mutex_enter(&zilog->zl_lock); 1599168404Spjd 1600168404Spjd ASSERT(zilog->zl_stop_sync == 0); 1601168404Spjd 1602219089Spjd if (*replayed_seq != 0) { 1603219089Spjd ASSERT(zh->zh_replay_seq < *replayed_seq); 1604219089Spjd zh->zh_replay_seq = *replayed_seq; 1605219089Spjd *replayed_seq = 0; 1606219089Spjd } 1607168404Spjd 1608168404Spjd if (zilog->zl_destroy_txg == txg) { 1609168404Spjd blkptr_t blk = zh->zh_log; 1610168404Spjd 1611168404Spjd ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1612168404Spjd 1613168404Spjd bzero(zh, sizeof (zil_header_t)); 1614209962Smm bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1615168404Spjd 1616168404Spjd if (zilog->zl_keep_first) { 1617168404Spjd /* 1618168404Spjd * If this block was part of log chain that couldn't 1619168404Spjd * be claimed because a device was missing during 1620168404Spjd * zil_claim(), but that device later returns, 1621168404Spjd * then this block could erroneously appear valid. 1622168404Spjd * To guard against this, assign a new GUID to the new 1623168404Spjd * log chain so it doesn't matter what blk points to. 1624168404Spjd */ 1625168404Spjd zil_init_log_chain(zilog, &blk); 1626168404Spjd zh->zh_log = blk; 1627168404Spjd } 1628168404Spjd } 1629168404Spjd 1630213197Smm while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1631168404Spjd zh->zh_log = lwb->lwb_blk; 1632168404Spjd if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1633168404Spjd break; 1634168404Spjd list_remove(&zilog->zl_lwb_list, lwb); 1635219089Spjd zio_free_zil(spa, txg, &lwb->lwb_blk); 1636168404Spjd kmem_cache_free(zil_lwb_cache, lwb); 1637168404Spjd 1638168404Spjd /* 1639168404Spjd * If we don't have anything left in the lwb list then 1640168404Spjd * we've had an allocation failure and we need to zero 1641168404Spjd * out the zil_header blkptr so that we don't end 1642168404Spjd * up freeing the same block twice. 1643168404Spjd */ 1644168404Spjd if (list_head(&zilog->zl_lwb_list) == NULL) 1645168404Spjd BP_ZERO(&zh->zh_log); 1646168404Spjd } 1647168404Spjd mutex_exit(&zilog->zl_lock); 1648168404Spjd} 1649168404Spjd 1650168404Spjdvoid 1651168404Spjdzil_init(void) 1652168404Spjd{ 1653168404Spjd zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1654168404Spjd sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1655168404Spjd} 1656168404Spjd 1657168404Spjdvoid 1658168404Spjdzil_fini(void) 1659168404Spjd{ 1660168404Spjd kmem_cache_destroy(zil_lwb_cache); 1661168404Spjd} 1662168404Spjd 1663219089Spjdvoid 1664219089Spjdzil_set_sync(zilog_t *zilog, uint64_t sync) 1665219089Spjd{ 1666219089Spjd zilog->zl_sync = sync; 1667219089Spjd} 1668219089Spjd 1669219089Spjdvoid 1670219089Spjdzil_set_logbias(zilog_t *zilog, uint64_t logbias) 1671219089Spjd{ 1672219089Spjd zilog->zl_logbias = logbias; 1673219089Spjd} 1674219089Spjd 1675168404Spjdzilog_t * 1676168404Spjdzil_alloc(objset_t *os, zil_header_t *zh_phys) 1677168404Spjd{ 1678168404Spjd zilog_t *zilog; 1679168404Spjd 1680168404Spjd zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1681168404Spjd 1682168404Spjd zilog->zl_header = zh_phys; 1683168404Spjd zilog->zl_os = os; 1684168404Spjd zilog->zl_spa = dmu_objset_spa(os); 1685168404Spjd zilog->zl_dmu_pool = dmu_objset_pool(os); 1686168404Spjd zilog->zl_destroy_txg = TXG_INITIAL - 1; 1687219089Spjd zilog->zl_logbias = dmu_objset_logbias(os); 1688219089Spjd zilog->zl_sync = dmu_objset_syncprop(os); 1689219089Spjd zilog->zl_next_batch = 1; 1690168404Spjd 1691168404Spjd mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1692168404Spjd 1693219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1694219089Spjd mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL, 1695219089Spjd MUTEX_DEFAULT, NULL); 1696219089Spjd } 1697168404Spjd 1698168404Spjd list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1699168404Spjd offsetof(lwb_t, lwb_node)); 1700168404Spjd 1701219089Spjd list_create(&zilog->zl_itx_commit_list, sizeof (itx_t), 1702219089Spjd offsetof(itx_t, itx_node)); 1703219089Spjd 1704185029Spjd mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1705168404Spjd 1706185029Spjd avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1707185029Spjd sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1708185029Spjd 1709185029Spjd cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1710185029Spjd cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1711219089Spjd cv_init(&zilog->zl_cv_batch[0], NULL, CV_DEFAULT, NULL); 1712219089Spjd cv_init(&zilog->zl_cv_batch[1], NULL, CV_DEFAULT, NULL); 1713185029Spjd 1714168404Spjd return (zilog); 1715168404Spjd} 1716168404Spjd 1717168404Spjdvoid 1718168404Spjdzil_free(zilog_t *zilog) 1719168404Spjd{ 1720168404Spjd zilog->zl_stop_sync = 1; 1721168404Spjd 1722224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1723168404Spjd list_destroy(&zilog->zl_lwb_list); 1724168404Spjd 1725185029Spjd avl_destroy(&zilog->zl_vdev_tree); 1726185029Spjd mutex_destroy(&zilog->zl_vdev_lock); 1727168404Spjd 1728219089Spjd ASSERT(list_is_empty(&zilog->zl_itx_commit_list)); 1729219089Spjd list_destroy(&zilog->zl_itx_commit_list); 1730219089Spjd 1731219089Spjd for (int i = 0; i < TXG_SIZE; i++) { 1732219089Spjd /* 1733219089Spjd * It's possible for an itx to be generated that doesn't dirty 1734219089Spjd * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean() 1735219089Spjd * callback to remove the entry. We remove those here. 1736219089Spjd * 1737219089Spjd * Also free up the ziltest itxs. 1738219089Spjd */ 1739219089Spjd if (zilog->zl_itxg[i].itxg_itxs) 1740219089Spjd zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs); 1741219089Spjd mutex_destroy(&zilog->zl_itxg[i].itxg_lock); 1742219089Spjd } 1743219089Spjd 1744168404Spjd mutex_destroy(&zilog->zl_lock); 1745168404Spjd 1746185029Spjd cv_destroy(&zilog->zl_cv_writer); 1747185029Spjd cv_destroy(&zilog->zl_cv_suspend); 1748219089Spjd cv_destroy(&zilog->zl_cv_batch[0]); 1749219089Spjd cv_destroy(&zilog->zl_cv_batch[1]); 1750185029Spjd 1751168404Spjd kmem_free(zilog, sizeof (zilog_t)); 1752168404Spjd} 1753168404Spjd 1754168404Spjd/* 1755168404Spjd * Open an intent log. 1756168404Spjd */ 1757168404Spjdzilog_t * 1758168404Spjdzil_open(objset_t *os, zil_get_data_t *get_data) 1759168404Spjd{ 1760168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1761168404Spjd 1762224526Smm ASSERT(zilog->zl_clean_taskq == NULL); 1763224526Smm ASSERT(zilog->zl_get_data == NULL); 1764224526Smm ASSERT(list_is_empty(&zilog->zl_lwb_list)); 1765224526Smm 1766168404Spjd zilog->zl_get_data = get_data; 1767168404Spjd zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1768168404Spjd 2, 2, TASKQ_PREPOPULATE); 1769168404Spjd 1770168404Spjd return (zilog); 1771168404Spjd} 1772168404Spjd 1773168404Spjd/* 1774168404Spjd * Close an intent log. 1775168404Spjd */ 1776168404Spjdvoid 1777168404Spjdzil_close(zilog_t *zilog) 1778168404Spjd{ 1779224526Smm lwb_t *lwb; 1780219089Spjd uint64_t txg = 0; 1781219089Spjd 1782219089Spjd zil_commit(zilog, 0); /* commit all itx */ 1783219089Spjd 1784168404Spjd /* 1785219089Spjd * The lwb_max_txg for the stubby lwb will reflect the last activity 1786219089Spjd * for the zil. After a txg_wait_synced() on the txg we know all the 1787219089Spjd * callbacks have occurred that may clean the zil. Only then can we 1788219089Spjd * destroy the zl_clean_taskq. 1789168404Spjd */ 1790219089Spjd mutex_enter(&zilog->zl_lock); 1791224526Smm lwb = list_tail(&zilog->zl_lwb_list); 1792224526Smm if (lwb != NULL) 1793224526Smm txg = lwb->lwb_max_txg; 1794219089Spjd mutex_exit(&zilog->zl_lock); 1795219089Spjd if (txg) 1796168404Spjd txg_wait_synced(zilog->zl_dmu_pool, txg); 1797239620Smm ASSERT(!zilog_is_dirty(zilog)); 1798168404Spjd 1799168404Spjd taskq_destroy(zilog->zl_clean_taskq); 1800168404Spjd zilog->zl_clean_taskq = NULL; 1801168404Spjd zilog->zl_get_data = NULL; 1802224526Smm 1803224526Smm /* 1804224526Smm * We should have only one LWB left on the list; remove it now. 1805224526Smm */ 1806224526Smm mutex_enter(&zilog->zl_lock); 1807224526Smm lwb = list_head(&zilog->zl_lwb_list); 1808224526Smm if (lwb != NULL) { 1809224526Smm ASSERT(lwb == list_tail(&zilog->zl_lwb_list)); 1810224526Smm list_remove(&zilog->zl_lwb_list, lwb); 1811224526Smm zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1812224526Smm kmem_cache_free(zil_lwb_cache, lwb); 1813224526Smm } 1814224526Smm mutex_exit(&zilog->zl_lock); 1815168404Spjd} 1816168404Spjd 1817168404Spjd/* 1818168404Spjd * Suspend an intent log. While in suspended mode, we still honor 1819168404Spjd * synchronous semantics, but we rely on txg_wait_synced() to do it. 1820168404Spjd * We suspend the log briefly when taking a snapshot so that the snapshot 1821168404Spjd * contains all the data it's supposed to, and has an empty intent log. 1822168404Spjd */ 1823168404Spjdint 1824168404Spjdzil_suspend(zilog_t *zilog) 1825168404Spjd{ 1826168404Spjd const zil_header_t *zh = zilog->zl_header; 1827168404Spjd 1828168404Spjd mutex_enter(&zilog->zl_lock); 1829200724Sdelphij if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1830168404Spjd mutex_exit(&zilog->zl_lock); 1831168404Spjd return (EBUSY); 1832168404Spjd } 1833168404Spjd if (zilog->zl_suspend++ != 0) { 1834168404Spjd /* 1835168404Spjd * Someone else already began a suspend. 1836168404Spjd * Just wait for them to finish. 1837168404Spjd */ 1838168404Spjd while (zilog->zl_suspending) 1839168404Spjd cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1840168404Spjd mutex_exit(&zilog->zl_lock); 1841168404Spjd return (0); 1842168404Spjd } 1843168404Spjd zilog->zl_suspending = B_TRUE; 1844168404Spjd mutex_exit(&zilog->zl_lock); 1845168404Spjd 1846219089Spjd zil_commit(zilog, 0); 1847168404Spjd 1848168404Spjd zil_destroy(zilog, B_FALSE); 1849168404Spjd 1850168404Spjd mutex_enter(&zilog->zl_lock); 1851168404Spjd zilog->zl_suspending = B_FALSE; 1852168404Spjd cv_broadcast(&zilog->zl_cv_suspend); 1853168404Spjd mutex_exit(&zilog->zl_lock); 1854168404Spjd 1855168404Spjd return (0); 1856168404Spjd} 1857168404Spjd 1858168404Spjdvoid 1859168404Spjdzil_resume(zilog_t *zilog) 1860168404Spjd{ 1861168404Spjd mutex_enter(&zilog->zl_lock); 1862168404Spjd ASSERT(zilog->zl_suspend != 0); 1863168404Spjd zilog->zl_suspend--; 1864168404Spjd mutex_exit(&zilog->zl_lock); 1865168404Spjd} 1866168404Spjd 1867219089Spjdtypedef struct zil_replay_arg { 1868219089Spjd zil_replay_func_t **zr_replay; 1869219089Spjd void *zr_arg; 1870219089Spjd boolean_t zr_byteswap; 1871219089Spjd char *zr_lr; 1872219089Spjd} zil_replay_arg_t; 1873219089Spjd 1874219089Spjdstatic int 1875219089Spjdzil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1876209962Smm{ 1877219089Spjd char name[MAXNAMELEN]; 1878209962Smm 1879219089Spjd zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1880209962Smm 1881219089Spjd dmu_objset_name(zilog->zl_os, name); 1882209962Smm 1883219089Spjd cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1884219089Spjd "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1885219089Spjd (u_longlong_t)lr->lrc_seq, 1886219089Spjd (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1887219089Spjd (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1888219089Spjd 1889219089Spjd return (error); 1890209962Smm} 1891209962Smm 1892219089Spjdstatic int 1893168404Spjdzil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1894168404Spjd{ 1895168404Spjd zil_replay_arg_t *zr = zra; 1896168404Spjd const zil_header_t *zh = zilog->zl_header; 1897168404Spjd uint64_t reclen = lr->lrc_reclen; 1898168404Spjd uint64_t txtype = lr->lrc_txtype; 1899219089Spjd int error = 0; 1900168404Spjd 1901219089Spjd zilog->zl_replaying_seq = lr->lrc_seq; 1902168404Spjd 1903219089Spjd if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1904219089Spjd return (0); 1905219089Spjd 1906168404Spjd if (lr->lrc_txg < claim_txg) /* already committed */ 1907219089Spjd return (0); 1908168404Spjd 1909185029Spjd /* Strip case-insensitive bit, still present in log record */ 1910185029Spjd txtype &= ~TX_CI; 1911185029Spjd 1912219089Spjd if (txtype == 0 || txtype >= TX_MAX_TYPE) 1913219089Spjd return (zil_replay_error(zilog, lr, EINVAL)); 1914219089Spjd 1915219089Spjd /* 1916219089Spjd * If this record type can be logged out of order, the object 1917219089Spjd * (lr_foid) may no longer exist. That's legitimate, not an error. 1918219089Spjd */ 1919219089Spjd if (TX_OOO(txtype)) { 1920219089Spjd error = dmu_object_info(zilog->zl_os, 1921219089Spjd ((lr_ooo_t *)lr)->lr_foid, NULL); 1922219089Spjd if (error == ENOENT || error == EEXIST) 1923219089Spjd return (0); 1924209962Smm } 1925209962Smm 1926168404Spjd /* 1927168404Spjd * Make a copy of the data so we can revise and extend it. 1928168404Spjd */ 1929219089Spjd bcopy(lr, zr->zr_lr, reclen); 1930168404Spjd 1931168404Spjd /* 1932219089Spjd * If this is a TX_WRITE with a blkptr, suck in the data. 1933219089Spjd */ 1934219089Spjd if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1935219089Spjd error = zil_read_log_data(zilog, (lr_write_t *)lr, 1936219089Spjd zr->zr_lr + reclen); 1937219089Spjd if (error) 1938219089Spjd return (zil_replay_error(zilog, lr, error)); 1939219089Spjd } 1940219089Spjd 1941219089Spjd /* 1942168404Spjd * The log block containing this lr may have been byteswapped 1943168404Spjd * so that we can easily examine common fields like lrc_txtype. 1944219089Spjd * However, the log is a mix of different record types, and only the 1945168404Spjd * replay vectors know how to byteswap their records. Therefore, if 1946168404Spjd * the lr was byteswapped, undo it before invoking the replay vector. 1947168404Spjd */ 1948168404Spjd if (zr->zr_byteswap) 1949219089Spjd byteswap_uint64_array(zr->zr_lr, reclen); 1950168404Spjd 1951168404Spjd /* 1952168404Spjd * We must now do two things atomically: replay this log record, 1953209962Smm * and update the log header sequence number to reflect the fact that 1954209962Smm * we did so. At the end of each replay function the sequence number 1955209962Smm * is updated if we are in replay mode. 1956168404Spjd */ 1957219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 1958219089Spjd if (error) { 1959168404Spjd /* 1960168404Spjd * The DMU's dnode layer doesn't see removes until the txg 1961168404Spjd * commits, so a subsequent claim can spuriously fail with 1962209962Smm * EEXIST. So if we receive any error we try syncing out 1963219089Spjd * any removes then retry the transaction. Note that we 1964219089Spjd * specify B_FALSE for byteswap now, so we don't do it twice. 1965168404Spjd */ 1966219089Spjd txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1967219089Spjd error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 1968219089Spjd if (error) 1969219089Spjd return (zil_replay_error(zilog, lr, error)); 1970168404Spjd } 1971219089Spjd return (0); 1972168404Spjd} 1973168404Spjd 1974168404Spjd/* ARGSUSED */ 1975219089Spjdstatic int 1976168404Spjdzil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 1977168404Spjd{ 1978168404Spjd zilog->zl_replay_blks++; 1979219089Spjd 1980219089Spjd return (0); 1981168404Spjd} 1982168404Spjd 1983168404Spjd/* 1984168404Spjd * If this dataset has a non-empty intent log, replay it and destroy it. 1985168404Spjd */ 1986168404Spjdvoid 1987209962Smmzil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1988168404Spjd{ 1989168404Spjd zilog_t *zilog = dmu_objset_zil(os); 1990168404Spjd const zil_header_t *zh = zilog->zl_header; 1991168404Spjd zil_replay_arg_t zr; 1992168404Spjd 1993200724Sdelphij if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1994168404Spjd zil_destroy(zilog, B_TRUE); 1995168404Spjd return; 1996168404Spjd } 1997168404Spjd //printf("ZFS: Replaying ZIL on %s...\n", os->os->os_spa->spa_name); 1998168404Spjd 1999168404Spjd zr.zr_replay = replay_func; 2000168404Spjd zr.zr_arg = arg; 2001168404Spjd zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 2002219089Spjd zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 2003168404Spjd 2004168404Spjd /* 2005168404Spjd * Wait for in-progress removes to sync before starting replay. 2006168404Spjd */ 2007168404Spjd txg_wait_synced(zilog->zl_dmu_pool, 0); 2008168404Spjd 2009209962Smm zilog->zl_replay = B_TRUE; 2010219089Spjd zilog->zl_replay_time = ddi_get_lbolt(); 2011168404Spjd ASSERT(zilog->zl_replay_blks == 0); 2012168404Spjd (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 2013168404Spjd zh->zh_claim_txg); 2014219089Spjd kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 2015168404Spjd 2016168404Spjd zil_destroy(zilog, B_FALSE); 2017185029Spjd txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 2018209962Smm zilog->zl_replay = B_FALSE; 2019168404Spjd //printf("ZFS: Replay of ZIL on %s finished.\n", os->os->os_spa->spa_name); 2020168404Spjd} 2021168404Spjd 2022219089Spjdboolean_t 2023219089Spjdzil_replaying(zilog_t *zilog, dmu_tx_t *tx) 2024168404Spjd{ 2025219089Spjd if (zilog->zl_sync == ZFS_SYNC_DISABLED) 2026219089Spjd return (B_TRUE); 2027168404Spjd 2028219089Spjd if (zilog->zl_replay) { 2029219089Spjd dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 2030219089Spjd zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 2031219089Spjd zilog->zl_replaying_seq; 2032219089Spjd return (B_TRUE); 2033168404Spjd } 2034168404Spjd 2035219089Spjd return (B_FALSE); 2036168404Spjd} 2037213197Smm 2038213197Smm/* ARGSUSED */ 2039213197Smmint 2040219089Spjdzil_vdev_offline(const char *osname, void *arg) 2041213197Smm{ 2042213197Smm objset_t *os; 2043213197Smm zilog_t *zilog; 2044213197Smm int error; 2045213197Smm 2046219089Spjd error = dmu_objset_hold(osname, FTAG, &os); 2047213197Smm if (error) 2048213197Smm return (error); 2049213197Smm 2050213197Smm zilog = dmu_objset_zil(os); 2051213197Smm if (zil_suspend(zilog) != 0) 2052213197Smm error = EEXIST; 2053213197Smm else 2054213197Smm zil_resume(zilog); 2055219089Spjd dmu_objset_rele(os, FTAG); 2056213197Smm return (error); 2057213197Smm} 2058