1 789 ahrens /* 2 789 ahrens * CDDL HEADER START 3 789 ahrens * 4 789 ahrens * The contents of this file are subject to the terms of the 5 1472 perrin * Common Development and Distribution License (the "License"). 6 1472 perrin * You may not use this file except in compliance with the License. 7 789 ahrens * 8 789 ahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 789 ahrens * or http://www.opensolaris.org/os/licensing. 10 789 ahrens * See the License for the specific language governing permissions 11 789 ahrens * and limitations under the License. 12 789 ahrens * 13 789 ahrens * When distributing Covered Code, include this CDDL HEADER in each 14 789 ahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 789 ahrens * If applicable, add the following below this CDDL HEADER, with the 16 789 ahrens * fields enclosed by brackets "[]" replaced with your own identifying 17 789 ahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18 789 ahrens * 19 789 ahrens * CDDL HEADER END 20 789 ahrens */ 21 789 ahrens /* 22 8746 Matthew * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 789 ahrens * Use is subject to license terms. 24 789 ahrens */ 25 789 ahrens 26 789 ahrens #include <sys/zfs_context.h> 27 789 ahrens #include <sys/spa.h> 28 789 ahrens #include <sys/dmu.h> 29 789 ahrens #include <sys/zap.h> 30 789 ahrens #include <sys/arc.h> 31 789 ahrens #include <sys/stat.h> 32 789 ahrens #include <sys/resource.h> 33 789 ahrens #include <sys/zil.h> 34 789 ahrens #include <sys/zil_impl.h> 35 789 ahrens #include <sys/dsl_dataset.h> 36 789 ahrens #include <sys/vdev.h> 37 3668 gw25295 #include <sys/dmu_tx.h> 38 789 ahrens 39 789 ahrens /* 40 789 ahrens * The zfs intent log (ZIL) saves transaction records of system calls 41 789 ahrens * that change the file system in memory with enough information 42 789 ahrens * to be able to replay them. These are stored in memory until 43 789 ahrens * either the DMU transaction group (txg) commits them to the stable pool 44 789 ahrens * and they can be discarded, or they are flushed to the stable log 45 789 ahrens * (also in the pool) due to a fsync, O_DSYNC or other synchronous 46 789 ahrens * requirement. In the event of a panic or power fail then those log 47 789 ahrens * records (transactions) are replayed. 48 789 ahrens * 49 789 ahrens * There is one ZIL per file system. Its on-disk (pool) format consists 50 789 ahrens * of 3 parts: 51 789 ahrens * 52 789 ahrens * - ZIL header 53 789 ahrens * - ZIL blocks 54 789 ahrens * - ZIL records 55 789 ahrens * 56 789 ahrens * A log record holds a system call transaction. Log blocks can 57 789 ahrens * hold many log records and the blocks are chained together. 58 789 ahrens * Each ZIL block contains a block pointer (blkptr_t) to the next 59 789 ahrens * ZIL block in the chain. The ZIL header points to the first 60 789 ahrens * block in the chain. Note there is not a fixed place in the pool 61 789 ahrens * to hold blocks. They are dynamically allocated and freed as 62 789 ahrens * needed from the blocks available. Figure X shows the ZIL structure: 63 789 ahrens */ 64 789 ahrens 65 789 ahrens /* 66 2986 ek110237 * This global ZIL switch affects all pools 67 789 ahrens */ 68 789 ahrens int zil_disable = 0; /* disable intent logging */ 69 2986 ek110237 70 2986 ek110237 /* 71 2986 ek110237 * Tunable parameter for debugging or performance analysis. Setting 72 2986 ek110237 * zfs_nocacheflush will cause corruption on power loss if a volatile 73 2986 ek110237 * out-of-order write cache is enabled. 74 2986 ek110237 */ 75 2986 ek110237 boolean_t zfs_nocacheflush = B_FALSE; 76 789 ahrens 77 789 ahrens static kmem_cache_t *zil_lwb_cache; 78 10685 George 79 10685 George static boolean_t zil_empty(zilog_t *zilog); 80 789 ahrens 81 789 ahrens static int 82 10922 Jeff zil_bp_compare(const void *x1, const void *x2) 83 789 ahrens { 84 10922 Jeff const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva; 85 10922 Jeff const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva; 86 789 ahrens 87 789 ahrens if (DVA_GET_VDEV(dva1) < DVA_GET_VDEV(dva2)) 88 789 ahrens return (-1); 89 789 ahrens if (DVA_GET_VDEV(dva1) > DVA_GET_VDEV(dva2)) 90 789 ahrens return (1); 91 789 ahrens 92 789 ahrens if (DVA_GET_OFFSET(dva1) < DVA_GET_OFFSET(dva2)) 93 789 ahrens return (-1); 94 789 ahrens if (DVA_GET_OFFSET(dva1) > DVA_GET_OFFSET(dva2)) 95 789 ahrens return (1); 96 789 ahrens 97 789 ahrens return (0); 98 789 ahrens } 99 789 ahrens 100 789 ahrens static void 101 10922 Jeff zil_bp_tree_init(zilog_t *zilog) 102 789 ahrens { 103 10922 Jeff avl_create(&zilog->zl_bp_tree, zil_bp_compare, 104 10922 Jeff sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node)); 105 789 ahrens } 106 789 ahrens 107 789 ahrens static void 108 10922 Jeff zil_bp_tree_fini(zilog_t *zilog) 109 789 ahrens { 110 10922 Jeff avl_tree_t *t = &zilog->zl_bp_tree; 111 10922 Jeff zil_bp_node_t *zn; 112 789 ahrens void *cookie = NULL; 113 789 ahrens 114 789 ahrens while ((zn = avl_destroy_nodes(t, &cookie)) != NULL) 115 10922 Jeff kmem_free(zn, sizeof (zil_bp_node_t)); 116 789 ahrens 117 789 ahrens avl_destroy(t); 118 789 ahrens } 119 789 ahrens 120 10922 Jeff int 121 10922 Jeff zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp) 122 789 ahrens { 123 10922 Jeff avl_tree_t *t = &zilog->zl_bp_tree; 124 10922 Jeff const dva_t *dva = BP_IDENTITY(bp); 125 10922 Jeff zil_bp_node_t *zn; 126 789 ahrens avl_index_t where; 127 789 ahrens 128 789 ahrens if (avl_find(t, dva, &where) != NULL) 129 789 ahrens return (EEXIST); 130 789 ahrens 131 10922 Jeff zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP); 132 789 ahrens zn->zn_dva = *dva; 133 789 ahrens avl_insert(t, zn, where); 134 789 ahrens 135 789 ahrens return (0); 136 789 ahrens } 137 789 ahrens 138 1807 bonwick static zil_header_t * 139 1807 bonwick zil_header_in_syncing_context(zilog_t *zilog) 140 1807 bonwick { 141 1807 bonwick return ((zil_header_t *)zilog->zl_header); 142 1807 bonwick } 143 1807 bonwick 144 1807 bonwick static void 145 1807 bonwick zil_init_log_chain(zilog_t *zilog, blkptr_t *bp) 146 1807 bonwick { 147 1807 bonwick zio_cksum_t *zc = &bp->blk_cksum; 148 1807 bonwick 149 1807 bonwick zc->zc_word[ZIL_ZC_GUID_0] = spa_get_random(-1ULL); 150 1807 bonwick zc->zc_word[ZIL_ZC_GUID_1] = spa_get_random(-1ULL); 151 1807 bonwick zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os); 152 1807 bonwick zc->zc_word[ZIL_ZC_SEQ] = 1ULL; 153 1807 bonwick } 154 1807 bonwick 155 789 ahrens /* 156 10922 Jeff * Read a log block and make sure it's valid. 157 789 ahrens */ 158 789 ahrens static int 159 10922 Jeff zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst) 160 789 ahrens { 161 10922 Jeff enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 162 10922 Jeff uint32_t aflags = ARC_WAIT; 163 10922 Jeff arc_buf_t *abuf = NULL; 164 1544 eschrock zbookmark_t zb; 165 789 ahrens int error; 166 1544 eschrock 167 10922 Jeff if (zilog->zl_header->zh_claim_txg == 0) 168 10922 Jeff zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 169 789 ahrens 170 10922 Jeff if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 171 10922 Jeff zio_flags |= ZIO_FLAG_SPECULATIVE; 172 1807 bonwick 173 10922 Jeff SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET], 174 10922 Jeff ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 175 10922 Jeff 176 10922 Jeff error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 177 10922 Jeff ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 178 1807 bonwick 179 1807 bonwick if (error == 0) { 180 10922 Jeff char *data = abuf->b_data; 181 10922 Jeff uint64_t size = BP_GET_LSIZE(bp); 182 10922 Jeff zil_trailer_t *ztp = (zil_trailer_t *)(data + size) - 1; 183 1807 bonwick zio_cksum_t cksum = bp->blk_cksum; 184 10922 Jeff 185 10922 Jeff bcopy(data, dst, size); 186 10922 Jeff *nbp = ztp->zit_next_blk; 187 1807 bonwick 188 1807 bonwick /* 189 7522 Neil * Validate the checksummed log block. 190 7522 Neil * 191 1807 bonwick * Sequence numbers should be... sequential. The checksum 192 1807 bonwick * verifier for the next block should be bp's checksum plus 1. 193 7522 Neil * 194 7522 Neil * Also check the log chain linkage and size used. 195 1807 bonwick */ 196 1807 bonwick cksum.zc_word[ZIL_ZC_SEQ]++; 197 1807 bonwick 198 7522 Neil if (bcmp(&cksum, &ztp->zit_next_blk.blk_cksum, 199 7522 Neil sizeof (cksum)) || BP_IS_HOLE(&ztp->zit_next_blk) || 200 10922 Jeff (ztp->zit_nused > (size - sizeof (zil_trailer_t)))) 201 7522 Neil error = ECKSUM; 202 1807 bonwick 203 10922 Jeff VERIFY(arc_buf_remove_ref(abuf, &abuf) == 1); 204 789 ahrens } 205 789 ahrens 206 10922 Jeff return (error); 207 10922 Jeff } 208 10922 Jeff 209 10922 Jeff /* 210 10922 Jeff * Read a TX_WRITE log data block. 211 10922 Jeff */ 212 10922 Jeff static int 213 10922 Jeff zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf) 214 10922 Jeff { 215 10922 Jeff enum zio_flag zio_flags = ZIO_FLAG_CANFAIL; 216 10922 Jeff const blkptr_t *bp = &lr->lr_blkptr; 217 10922 Jeff uint32_t aflags = ARC_WAIT; 218 10922 Jeff arc_buf_t *abuf = NULL; 219 10922 Jeff zbookmark_t zb; 220 10922 Jeff int error; 221 10922 Jeff 222 10922 Jeff if (BP_IS_HOLE(bp)) { 223 10922 Jeff if (wbuf != NULL) 224 10922 Jeff bzero(wbuf, MAX(BP_GET_LSIZE(bp), lr->lr_length)); 225 10922 Jeff return (0); 226 10922 Jeff } 227 10922 Jeff 228 10922 Jeff if (zilog->zl_header->zh_claim_txg == 0) 229 10922 Jeff zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB; 230 10922 Jeff 231 10922 Jeff SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid, 232 10922 Jeff ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp)); 233 10922 Jeff 234 10922 Jeff error = arc_read_nolock(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf, 235 10922 Jeff ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb); 236 10922 Jeff 237 10922 Jeff if (error == 0) { 238 10922 Jeff if (wbuf != NULL) 239 10922 Jeff bcopy(abuf->b_data, wbuf, arc_buf_size(abuf)); 240 10922 Jeff (void) arc_buf_remove_ref(abuf, &abuf); 241 10922 Jeff } 242 789 ahrens 243 1807 bonwick return (error); 244 789 ahrens } 245 789 ahrens 246 789 ahrens /* 247 789 ahrens * Parse the intent log, and call parse_func for each valid record within. 248 789 ahrens */ 249 10922 Jeff int 250 789 ahrens zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func, 251 789 ahrens zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg) 252 789 ahrens { 253 1807 bonwick const zil_header_t *zh = zilog->zl_header; 254 10922 Jeff boolean_t claimed = !!zh->zh_claim_txg; 255 10922 Jeff uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX; 256 10922 Jeff uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX; 257 10922 Jeff uint64_t max_blk_seq = 0; 258 10922 Jeff uint64_t max_lr_seq = 0; 259 10922 Jeff uint64_t blk_count = 0; 260 10922 Jeff uint64_t lr_count = 0; 261 10922 Jeff blkptr_t blk, next_blk; 262 789 ahrens char *lrbuf, *lrp; 263 10922 Jeff int error = 0; 264 789 ahrens 265 10922 Jeff /* 266 10922 Jeff * Old logs didn't record the maximum zh_claim_lr_seq. 267 10922 Jeff */ 268 10922 Jeff if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID)) 269 10922 Jeff claim_lr_seq = UINT64_MAX; 270 789 ahrens 271 789 ahrens /* 272 789 ahrens * Starting at the block pointed to by zh_log we read the log chain. 273 789 ahrens * For each block in the chain we strongly check that block to 274 789 ahrens * ensure its validity. We stop when an invalid block is found. 275 789 ahrens * For each block pointer in the chain we call parse_blk_func(). 276 789 ahrens * For each record in each valid block we call parse_lr_func(). 277 1807 bonwick * If the log has been claimed, stop if we encounter a sequence 278 1807 bonwick * number greater than the highest claimed sequence number. 279 789 ahrens */ 280 10922 Jeff lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE); 281 10922 Jeff zil_bp_tree_init(zilog); 282 1807 bonwick 283 10922 Jeff for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) { 284 10922 Jeff zil_trailer_t *ztp = 285 10922 Jeff (zil_trailer_t *)(lrbuf + BP_GET_LSIZE(&blk)) - 1; 286 10922 Jeff uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ]; 287 10922 Jeff int reclen; 288 10922 Jeff 289 10922 Jeff if (blk_seq > claim_blk_seq) 290 10922 Jeff break; 291 10922 Jeff if ((error = parse_blk_func(zilog, &blk, arg, txg)) != 0) 292 10922 Jeff break; 293 10922 Jeff ASSERT(max_blk_seq < blk_seq); 294 10922 Jeff max_blk_seq = blk_seq; 295 10922 Jeff blk_count++; 296 10922 Jeff 297 10922 Jeff if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq) 298 1807 bonwick break; 299 1807 bonwick 300 10922 Jeff error = zil_read_log_block(zilog, &blk, &next_blk, lrbuf); 301 789 ahrens if (error) 302 789 ahrens break; 303 789 ahrens 304 789 ahrens for (lrp = lrbuf; lrp < lrbuf + ztp->zit_nused; lrp += reclen) { 305 789 ahrens lr_t *lr = (lr_t *)lrp; 306 789 ahrens reclen = lr->lrc_reclen; 307 789 ahrens ASSERT3U(reclen, >=, sizeof (lr_t)); 308 10922 Jeff if (lr->lrc_seq > claim_lr_seq) 309 10922 Jeff goto done; 310 10922 Jeff if ((error = parse_lr_func(zilog, lr, arg, txg)) != 0) 311 10922 Jeff goto done; 312 10922 Jeff ASSERT(max_lr_seq < lr->lrc_seq); 313 10922 Jeff max_lr_seq = lr->lrc_seq; 314 10922 Jeff lr_count++; 315 789 ahrens } 316 789 ahrens } 317 10922 Jeff done: 318 10922 Jeff zilog->zl_parse_error = error; 319 10922 Jeff zilog->zl_parse_blk_seq = max_blk_seq; 320 10922 Jeff zilog->zl_parse_lr_seq = max_lr_seq; 321 10922 Jeff zilog->zl_parse_blk_count = blk_count; 322 10922 Jeff zilog->zl_parse_lr_count = lr_count; 323 1807 bonwick 324 10922 Jeff ASSERT(!claimed || !(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID) || 325 10922 Jeff (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq)); 326 10922 Jeff 327 10922 Jeff zil_bp_tree_fini(zilog); 328 10922 Jeff zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE); 329 10922 Jeff 330 10922 Jeff return (error); 331 10922 Jeff } 332 10922 Jeff 333 10922 Jeff static int 334 10922 Jeff zil_claim_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t first_txg) 335 10922 Jeff { 336 10922 Jeff /* 337 10922 Jeff * Claim log block if not already committed and not already claimed. 338 10922 Jeff * If tx == NULL, just verify that the block is claimable. 339 10922 Jeff */ 340 10922 Jeff if (bp->blk_birth < first_txg || zil_bp_tree_add(zilog, bp) != 0) 341 10922 Jeff return (0); 342 10922 Jeff 343 10922 Jeff return (zio_wait(zio_claim(NULL, zilog->zl_spa, 344 10922 Jeff tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL, 345 10922 Jeff ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB))); 346 10922 Jeff } 347 10922 Jeff 348 10922 Jeff static int 349 10922 Jeff zil_claim_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t first_txg) 350 10922 Jeff { 351 10922 Jeff lr_write_t *lr = (lr_write_t *)lrc; 352 10922 Jeff int error; 353 10922 Jeff 354 10922 Jeff if (lrc->lrc_txtype != TX_WRITE) 355 10922 Jeff return (0); 356 10922 Jeff 357 10922 Jeff /* 358 10922 Jeff * If the block is not readable, don't claim it. This can happen 359 10922 Jeff * in normal operation when a log block is written to disk before 360 10922 Jeff * some of the dmu_sync() blocks it points to. In this case, the 361 10922 Jeff * transaction cannot have been committed to anyone (we would have 362 10922 Jeff * waited for all writes to be stable first), so it is semantically 363 10922 Jeff * correct to declare this the end of the log. 364 10922 Jeff */ 365 10922 Jeff if (lr->lr_blkptr.blk_birth >= first_txg && 366 10922 Jeff (error = zil_read_log_data(zilog, lr, NULL)) != 0) 367 10922 Jeff return (error); 368 10922 Jeff 369 10922 Jeff return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg)); 370 789 ahrens } 371 789 ahrens 372 789 ahrens /* ARGSUSED */ 373 10922 Jeff static int 374 10922 Jeff zil_free_log_block(zilog_t *zilog, blkptr_t *bp, void *tx, uint64_t claim_txg) 375 789 ahrens { 376 10922 Jeff zio_free_zil(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 377 789 ahrens 378 10922 Jeff return (0); 379 789 ahrens } 380 789 ahrens 381 10922 Jeff static int 382 789 ahrens zil_free_log_record(zilog_t *zilog, lr_t *lrc, void *tx, uint64_t claim_txg) 383 789 ahrens { 384 10922 Jeff lr_write_t *lr = (lr_write_t *)lrc; 385 10922 Jeff blkptr_t *bp = &lr->lr_blkptr; 386 10922 Jeff 387 789 ahrens /* 388 789 ahrens * If we previously claimed it, we need to free it. 389 789 ahrens */ 390 10922 Jeff if (claim_txg != 0 && lrc->lrc_txtype == TX_WRITE && 391 10922 Jeff bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0) 392 10922 Jeff zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp); 393 10922 Jeff 394 10922 Jeff return (0); 395 789 ahrens } 396 789 ahrens 397 789 ahrens /* 398 789 ahrens * Create an on-disk intent log. 399 789 ahrens */ 400 789 ahrens static void 401 789 ahrens zil_create(zilog_t *zilog) 402 789 ahrens { 403 1807 bonwick const zil_header_t *zh = zilog->zl_header; 404 789 ahrens lwb_t *lwb; 405 1807 bonwick uint64_t txg = 0; 406 1807 bonwick dmu_tx_t *tx = NULL; 407 789 ahrens blkptr_t blk; 408 1807 bonwick int error = 0; 409 789 ahrens 410 789 ahrens /* 411 1807 bonwick * Wait for any previous destroy to complete. 412 789 ahrens */ 413 1807 bonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 414 1807 bonwick 415 1807 bonwick ASSERT(zh->zh_claim_txg == 0); 416 1807 bonwick ASSERT(zh->zh_replay_seq == 0); 417 1807 bonwick 418 1807 bonwick blk = zh->zh_log; 419 789 ahrens 420 789 ahrens /* 421 8109 Neil * If we don't already have an initial log block or we have one 422 8109 Neil * but it's the wrong endianness then allocate one. 423 789 ahrens */ 424 8109 Neil if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) { 425 1807 bonwick tx = dmu_tx_create(zilog->zl_os); 426 10922 Jeff VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 427 1807 bonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 428 1807 bonwick txg = dmu_tx_get_txg(tx); 429 8109 Neil 430 8109 Neil if (!BP_IS_HOLE(&blk)) { 431 10922 Jeff zio_free_zil(zilog->zl_spa, txg, &blk); 432 8109 Neil BP_ZERO(&blk); 433 8109 Neil } 434 1807 bonwick 435 10922 Jeff error = zio_alloc_zil(zilog->zl_spa, txg, &blk, NULL, 436 10922 Jeff ZIL_MIN_BLKSZ, zilog->zl_logbias == ZFS_LOGBIAS_LATENCY); 437 1807 bonwick 438 1807 bonwick if (error == 0) 439 1807 bonwick zil_init_log_chain(zilog, &blk); 440 1362 perrin } 441 1807 bonwick 442 1807 bonwick /* 443 1807 bonwick * Allocate a log write buffer (lwb) for the first log block. 444 1807 bonwick */ 445 789 ahrens if (error == 0) { 446 789 ahrens lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 447 789 ahrens lwb->lwb_zilog = zilog; 448 789 ahrens lwb->lwb_blk = blk; 449 789 ahrens lwb->lwb_nused = 0; 450 789 ahrens lwb->lwb_sz = BP_GET_LSIZE(&lwb->lwb_blk); 451 789 ahrens lwb->lwb_buf = zio_buf_alloc(lwb->lwb_sz); 452 789 ahrens lwb->lwb_max_txg = txg; 453 2237 maybee lwb->lwb_zio = NULL; 454 10922 Jeff lwb->lwb_tx = NULL; 455 2237 maybee 456 789 ahrens mutex_enter(&zilog->zl_lock); 457 789 ahrens list_insert_tail(&zilog->zl_lwb_list, lwb); 458 789 ahrens mutex_exit(&zilog->zl_lock); 459 789 ahrens } 460 789 ahrens 461 1807 bonwick /* 462 1807 bonwick * If we just allocated the first log block, commit our transaction 463 1807 bonwick * and wait for zil_sync() to stuff the block poiner into zh_log. 464 1807 bonwick * (zh is part of the MOS, so we cannot modify it in open context.) 465 1807 bonwick */ 466 1807 bonwick if (tx != NULL) { 467 1807 bonwick dmu_tx_commit(tx); 468 1362 perrin txg_wait_synced(zilog->zl_dmu_pool, txg); 469 1807 bonwick } 470 1807 bonwick 471 1807 bonwick ASSERT(bcmp(&blk, &zh->zh_log, sizeof (blk)) == 0); 472 789 ahrens } 473 789 ahrens 474 789 ahrens /* 475 789 ahrens * In one tx, free all log blocks and clear the log header. 476 1807 bonwick * If keep_first is set, then we're replaying a log with no content. 477 1807 bonwick * We want to keep the first block, however, so that the first 478 1807 bonwick * synchronous transaction doesn't require a txg_wait_synced() 479 1807 bonwick * in zil_create(). We don't need to txg_wait_synced() here either 480 1807 bonwick * when keep_first is set, because both zil_create() and zil_destroy() 481 1807 bonwick * will wait for any in-progress destroys to complete. 482 789 ahrens */ 483 789 ahrens void 484 1807 bonwick zil_destroy(zilog_t *zilog, boolean_t keep_first) 485 789 ahrens { 486 1807 bonwick const zil_header_t *zh = zilog->zl_header; 487 1807 bonwick lwb_t *lwb; 488 789 ahrens dmu_tx_t *tx; 489 789 ahrens uint64_t txg; 490 789 ahrens 491 1807 bonwick /* 492 1807 bonwick * Wait for any previous destroy to complete. 493 1807 bonwick */ 494 1807 bonwick txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 495 789 ahrens 496 10922 Jeff zilog->zl_old_header = *zh; /* debugging aid */ 497 10922 Jeff 498 1807 bonwick if (BP_IS_HOLE(&zh->zh_log)) 499 789 ahrens return; 500 789 ahrens 501 789 ahrens tx = dmu_tx_create(zilog->zl_os); 502 10922 Jeff VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 503 789 ahrens dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 504 789 ahrens txg = dmu_tx_get_txg(tx); 505 789 ahrens 506 1807 bonwick mutex_enter(&zilog->zl_lock); 507 5223 perrin 508 1807 bonwick ASSERT3U(zilog->zl_destroy_txg, <, txg); 509 789 ahrens zilog->zl_destroy_txg = txg; 510 10922 Jeff zilog->zl_keep_first = keep_first; 511 1807 bonwick 512 1807 bonwick if (!list_is_empty(&zilog->zl_lwb_list)) { 513 1807 bonwick ASSERT(zh->zh_claim_txg == 0); 514 10922 Jeff ASSERT(!keep_first); 515 1807 bonwick while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 516 1807 bonwick list_remove(&zilog->zl_lwb_list, lwb); 517 1807 bonwick if (lwb->lwb_buf != NULL) 518 1807 bonwick zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 519 10922 Jeff zio_free_zil(zilog->zl_spa, txg, &lwb->lwb_blk); 520 1807 bonwick kmem_cache_free(zil_lwb_cache, lwb); 521 1807 bonwick } 522 10922 Jeff } else if (!keep_first) { 523 10922 Jeff (void) zil_parse(zilog, zil_free_log_block, 524 10922 Jeff zil_free_log_record, tx, zh->zh_claim_txg); 525 1807 bonwick } 526 2638 perrin mutex_exit(&zilog->zl_lock); 527 789 ahrens 528 789 ahrens dmu_tx_commit(tx); 529 8989 Neil } 530 8989 Neil 531 2199 ahrens int 532 789 ahrens zil_claim(char *osname, void *txarg) 533 789 ahrens { 534 789 ahrens dmu_tx_t *tx = txarg; 535 789 ahrens uint64_t first_txg = dmu_tx_get_txg(tx); 536 789 ahrens zilog_t *zilog; 537 789 ahrens zil_header_t *zh; 538 789 ahrens objset_t *os; 539 789 ahrens int error; 540 789 ahrens 541 10298 Matthew error = dmu_objset_hold(osname, FTAG, &os); 542 789 ahrens if (error) { 543 7294 perrin cmn_err(CE_WARN, "can't open objset for %s", osname); 544 2199 ahrens return (0); 545 789 ahrens } 546 789 ahrens 547 789 ahrens zilog = dmu_objset_zil(os); 548 1807 bonwick zh = zil_header_in_syncing_context(zilog); 549 8989 Neil 550 10922 Jeff if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR) { 551 9701 George if (!BP_IS_HOLE(&zh->zh_log)) 552 10922 Jeff zio_free_zil(zilog->zl_spa, first_txg, &zh->zh_log); 553 9701 George BP_ZERO(&zh->zh_log); 554 9701 George dsl_dataset_dirty(dmu_objset_ds(os), tx); 555 10921 Tim dmu_objset_rele(os, FTAG); 556 10921 Tim return (0); 557 9701 George } 558 789 ahrens 559 789 ahrens /* 560 1807 bonwick * Claim all log blocks if we haven't already done so, and remember 561 1807 bonwick * the highest claimed sequence number. This ensures that if we can 562 1807 bonwick * read only part of the log now (e.g. due to a missing device), 563 1807 bonwick * but we can read the entire log later, we will not try to replay 564 1807 bonwick * or destroy beyond the last block we successfully claimed. 565 789 ahrens */ 566 789 ahrens ASSERT3U(zh->zh_claim_txg, <=, first_txg); 567 789 ahrens if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) { 568 10922 Jeff (void) zil_parse(zilog, zil_claim_log_block, 569 10922 Jeff zil_claim_log_record, tx, first_txg); 570 789 ahrens zh->zh_claim_txg = first_txg; 571 10922 Jeff zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq; 572 10922 Jeff zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq; 573 10922 Jeff if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1) 574 10922 Jeff zh->zh_flags |= ZIL_REPLAY_NEEDED; 575 10922 Jeff zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID; 576 789 ahrens dsl_dataset_dirty(dmu_objset_ds(os), tx); 577 789 ahrens } 578 1807 bonwick 579 789 ahrens ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1)); 580 10298 Matthew dmu_objset_rele(os, FTAG); 581 7294 perrin return (0); 582 7294 perrin } 583 7294 perrin 584 7294 perrin /* 585 7294 perrin * Check the log by walking the log chain. 586 7294 perrin * Checksum errors are ok as they indicate the end of the chain. 587 7294 perrin * Any other error (no device or read failure) returns an error. 588 7294 perrin */ 589 7294 perrin int 590 10922 Jeff zil_check_log_chain(char *osname, void *tx) 591 7294 perrin { 592 7294 perrin zilog_t *zilog; 593 7294 perrin objset_t *os; 594 7294 perrin int error; 595 10922 Jeff 596 10922 Jeff ASSERT(tx == NULL); 597 7294 perrin 598 10298 Matthew error = dmu_objset_hold(osname, FTAG, &os); 599 7294 perrin if (error) { 600 7294 perrin cmn_err(CE_WARN, "can't open objset for %s", osname); 601 7294 perrin return (0); 602 7294 perrin } 603 7294 perrin 604 7294 perrin zilog = dmu_objset_zil(os); 605 7294 perrin 606 10922 Jeff /* 607 10922 Jeff * Because tx == NULL, zil_claim_log_block() will not actually claim 608 10922 Jeff * any blocks, but just determine whether it is possible to do so. 609 10922 Jeff * In addition to checking the log chain, zil_claim_log_block() 610 10922 Jeff * will invoke zio_claim() with a done func of spa_claim_notify(), 611 10922 Jeff * which will update spa_max_claim_txg. See spa_load() for details. 612 10922 Jeff */ 613 10922 Jeff error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx, 614 10922 Jeff zilog->zl_header->zh_claim_txg ? -1ULL : spa_first_txg(os->os_spa)); 615 10922 Jeff 616 10298 Matthew dmu_objset_rele(os, FTAG); 617 10922 Jeff 618 10922 Jeff return ((error == ECKSUM || error == ENOENT) ? 0 : error); 619 789 ahrens } 620 789 ahrens 621 5688 bonwick static int 622 5688 bonwick zil_vdev_compare(const void *x1, const void *x2) 623 5688 bonwick { 624 5875 perrin uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev; 625 5875 perrin uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev; 626 5688 bonwick 627 5688 bonwick if (v1 < v2) 628 5688 bonwick return (-1); 629 5688 bonwick if (v1 > v2) 630 5688 bonwick return (1); 631 5688 bonwick 632 5688 bonwick return (0); 633 5688 bonwick } 634 5688 bonwick 635 789 ahrens void 636 10922 Jeff zil_add_block(zilog_t *zilog, const blkptr_t *bp) 637 789 ahrens { 638 5688 bonwick avl_tree_t *t = &zilog->zl_vdev_tree; 639 5688 bonwick avl_index_t where; 640 5688 bonwick zil_vdev_node_t *zv, zvsearch; 641 5688 bonwick int ndvas = BP_GET_NDVAS(bp); 642 5688 bonwick int i; 643 789 ahrens 644 2986 ek110237 if (zfs_nocacheflush) 645 789 ahrens return; 646 789 ahrens 647 5688 bonwick ASSERT(zilog->zl_writer); 648 5688 bonwick 649 5688 bonwick /* 650 5688 bonwick * Even though we're zl_writer, we still need a lock because the 651 5688 bonwick * zl_get_data() callbacks may have dmu_sync() done callbacks 652 5688 bonwick * that will run concurrently. 653 5688 bonwick */ 654 5688 bonwick mutex_enter(&zilog->zl_vdev_lock); 655 5688 bonwick for (i = 0; i < ndvas; i++) { 656 5688 bonwick zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]); 657 5688 bonwick if (avl_find(t, &zvsearch, &where) == NULL) { 658 5688 bonwick zv = kmem_alloc(sizeof (*zv), KM_SLEEP); 659 5688 bonwick zv->zv_vdev = zvsearch.zv_vdev; 660 5688 bonwick avl_insert(t, zv, where); 661 3063 perrin } 662 3063 perrin } 663 5688 bonwick mutex_exit(&zilog->zl_vdev_lock); 664 3063 perrin } 665 3063 perrin 666 789 ahrens void 667 2638 perrin zil_flush_vdevs(zilog_t *zilog) 668 789 ahrens { 669 3063 perrin spa_t *spa = zilog->zl_spa; 670 5688 bonwick avl_tree_t *t = &zilog->zl_vdev_tree; 671 5688 bonwick void *cookie = NULL; 672 5688 bonwick zil_vdev_node_t *zv; 673 5688 bonwick zio_t *zio; 674 789 ahrens 675 3063 perrin ASSERT(zilog->zl_writer); 676 789 ahrens 677 5688 bonwick /* 678 5688 bonwick * We don't need zl_vdev_lock here because we're the zl_writer, 679 5688 bonwick * and all zl_get_data() callbacks are done. 680 5688 bonwick */ 681 5688 bonwick if (avl_numnodes(t) == 0) 682 5688 bonwick return; 683 5688 bonwick 684 7754 Jeff spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 685 5688 bonwick 686 7754 Jeff zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 687 5688 bonwick 688 5688 bonwick while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) { 689 5688 bonwick vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev); 690 5688 bonwick if (vd != NULL) 691 5688 bonwick zio_flush(zio, vd); 692 5688 bonwick kmem_free(zv, sizeof (*zv)); 693 3063 perrin } 694 789 ahrens 695 789 ahrens /* 696 789 ahrens * Wait for all the flushes to complete. Not all devices actually 697 789 ahrens * support the DKIOCFLUSHWRITECACHE ioctl, so it's OK if it fails. 698 789 ahrens */ 699 5688 bonwick (void) zio_wait(zio); 700 5688 bonwick 701 7754 Jeff spa_config_exit(spa, SCL_STATE, FTAG); 702 789 ahrens } 703 789 ahrens 704 789 ahrens /* 705 789 ahrens * Function called when a log block write completes 706 789 ahrens */ 707 789 ahrens static void 708 789 ahrens zil_lwb_write_done(zio_t *zio) 709 789 ahrens { 710 789 ahrens lwb_t *lwb = zio->io_private; 711 789 ahrens zilog_t *zilog = lwb->lwb_zilog; 712 10922 Jeff dmu_tx_t *tx = lwb->lwb_tx; 713 7754 Jeff 714 7754 Jeff ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 715 7754 Jeff ASSERT(BP_GET_CHECKSUM(zio->io_bp) == ZIO_CHECKSUM_ZILOG); 716 7754 Jeff ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG); 717 7754 Jeff ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 718 7754 Jeff ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER); 719 7754 Jeff ASSERT(!BP_IS_GANG(zio->io_bp)); 720 7754 Jeff ASSERT(!BP_IS_HOLE(zio->io_bp)); 721 7754 Jeff ASSERT(zio->io_bp->blk_fill == 0); 722 789 ahrens 723 789 ahrens /* 724 9493 Neil * Ensure the lwb buffer pointer is cleared before releasing 725 9493 Neil * the txg. If we have had an allocation failure and 726 9493 Neil * the txg is waiting to sync then we want want zil_sync() 727 9493 Neil * to remove the lwb so that it's not picked up as the next new 728 9493 Neil * one in zil_commit_writer(). zil_sync() will only remove 729 9493 Neil * the lwb if lwb_buf is null. 730 789 ahrens */ 731 789 ahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 732 789 ahrens mutex_enter(&zilog->zl_lock); 733 789 ahrens lwb->lwb_buf = NULL; 734 10922 Jeff lwb->lwb_tx = NULL; 735 10922 Jeff mutex_exit(&zilog->zl_lock); 736 9493 Neil 737 9493 Neil /* 738 9493 Neil * Now that we've written this log block, we have a stable pointer 739 9493 Neil * to the next block in the chain, so it's OK to let the txg in 740 10922 Jeff * which we allocated the next block sync. 741 9493 Neil */ 742 10922 Jeff dmu_tx_commit(tx); 743 789 ahrens } 744 789 ahrens 745 789 ahrens /* 746 2237 maybee * Initialize the io for a log block. 747 2237 maybee */ 748 2237 maybee static void 749 2237 maybee zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb) 750 2237 maybee { 751 2237 maybee zbookmark_t zb; 752 2237 maybee 753 10922 Jeff SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET], 754 10922 Jeff ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, 755 10922 Jeff lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]); 756 2237 maybee 757 2638 perrin if (zilog->zl_root_zio == NULL) { 758 2638 perrin zilog->zl_root_zio = zio_root(zilog->zl_spa, NULL, NULL, 759 2638 perrin ZIO_FLAG_CANFAIL); 760 2638 perrin } 761 3063 perrin if (lwb->lwb_zio == NULL) { 762 3063 perrin lwb->lwb_zio = zio_rewrite(zilog->zl_root_zio, zilog->zl_spa, 763 9701 George 0, &lwb->lwb_blk, lwb->lwb_buf, lwb->lwb_sz, 764 9701 George zil_lwb_write_done, lwb, ZIO_PRIORITY_LOG_WRITE, 765 10685 George ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE, &zb); 766 3063 perrin } 767 2237 maybee } 768 2237 maybee 769 2237 maybee /* 770 10879 Neil * Use the slog as long as the logbias is 'latency' and the current commit size 771 10879 Neil * is less than the limit or the total list size is less than 2X the limit. 772 10879 Neil * Limit checking is disabled by setting zil_slog_limit to UINT64_MAX. 773 10879 Neil */ 774 10879 Neil uint64_t zil_slog_limit = 1024 * 1024; 775 10879 Neil #define USE_SLOG(zilog) (((zilog)->zl_logbias == ZFS_LOGBIAS_LATENCY) && \ 776 10879 Neil (((zilog)->zl_cur_used < zil_slog_limit) || \ 777 10879 Neil ((zilog)->zl_itx_list_sz < (zil_slog_limit << 1)))) 778 10879 Neil 779 10879 Neil /* 780 789 ahrens * Start a log block write and advance to the next log block. 781 789 ahrens * Calls are serialized. 782 789 ahrens */ 783 789 ahrens static lwb_t * 784 789 ahrens zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb) 785 789 ahrens { 786 789 ahrens lwb_t *nlwb; 787 789 ahrens zil_trailer_t *ztp = (zil_trailer_t *)(lwb->lwb_buf + lwb->lwb_sz) - 1; 788 1807 bonwick spa_t *spa = zilog->zl_spa; 789 1807 bonwick blkptr_t *bp = &ztp->zit_next_blk; 790 10922 Jeff dmu_tx_t *tx; 791 789 ahrens uint64_t txg; 792 789 ahrens uint64_t zil_blksz; 793 789 ahrens int error; 794 789 ahrens 795 789 ahrens ASSERT(lwb->lwb_nused <= ZIL_BLK_DATA_SZ(lwb)); 796 789 ahrens 797 789 ahrens /* 798 789 ahrens * Allocate the next block and save its address in this block 799 789 ahrens * before writing it in order to establish the log chain. 800 789 ahrens * Note that if the allocation of nlwb synced before we wrote 801 789 ahrens * the block that points at it (lwb), we'd leak it if we crashed. 802 10922 Jeff * Therefore, we don't do dmu_tx_commit() until zil_lwb_write_done(). 803 10922 Jeff * We dirty the dataset to ensure that zil_sync() will be called 804 10922 Jeff * to clean up in the event of allocation failure or I/O failure. 805 789 ahrens */ 806 10922 Jeff tx = dmu_tx_create(zilog->zl_os); 807 10922 Jeff VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 808 10922 Jeff dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 809 10922 Jeff txg = dmu_tx_get_txg(tx); 810 10922 Jeff 811 10922 Jeff lwb->lwb_tx = tx; 812 789 ahrens 813 789 ahrens /* 814 1141 perrin * Pick a ZIL blocksize. We request a size that is the 815 1141 perrin * maximum of the previous used size, the current used size and 816 1141 perrin * the amount waiting in the queue. 817 789 ahrens */ 818 2237 maybee zil_blksz = MAX(zilog->zl_prev_used, 819 2237 maybee zilog->zl_cur_used + sizeof (*ztp)); 820 1141 perrin zil_blksz = MAX(zil_blksz, zilog->zl_itx_list_sz + sizeof (*ztp)); 821 1842 perrin zil_blksz = P2ROUNDUP_TYPED(zil_blksz, ZIL_MIN_BLKSZ, uint64_t); 822 1141 perrin if (zil_blksz > ZIL_MAX_BLKSZ) 823 1141 perrin zil_blksz = ZIL_MAX_BLKSZ; 824 789 ahrens 825 3063 perrin BP_ZERO(bp); 826 3063 perrin /* pass the old blkptr in order to spread log blocks across devs */ 827 10922 Jeff error = zio_alloc_zil(spa, txg, bp, &lwb->lwb_blk, zil_blksz, 828 10879 Neil USE_SLOG(zilog)); 829 789 ahrens if (error) { 830 1544 eschrock /* 831 10922 Jeff * Since we've just experienced an allocation failure, 832 3668 gw25295 * terminate the current lwb and send it on its way. 833 3668 gw25295 */ 834 3668 gw25295 ztp->zit_pad = 0; 835 3668 gw25295 ztp->zit_nused = lwb->lwb_nused; 836 3668 gw25295 ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 837 3668 gw25295 zio_nowait(lwb->lwb_zio); 838 3668 gw25295 839 3668 gw25295 /* 840 1544 eschrock * By returning NULL the caller will call tx_wait_synced() 841 1544 eschrock */ 842 789 ahrens return (NULL); 843 789 ahrens } 844 789 ahrens 845 1807 bonwick ASSERT3U(bp->blk_birth, ==, txg); 846 1544 eschrock ztp->zit_pad = 0; 847 789 ahrens ztp->zit_nused = lwb->lwb_nused; 848 789 ahrens ztp->zit_bt.zbt_cksum = lwb->lwb_blk.blk_cksum; 849 1807 bonwick bp->blk_cksum = lwb->lwb_blk.blk_cksum; 850 1807 bonwick bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++; 851 789 ahrens 852 789 ahrens /* 853 789 ahrens * Allocate a new log write buffer (lwb). 854 789 ahrens */ 855 789 ahrens nlwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP); 856 789 ahrens nlwb->lwb_zilog = zilog; 857 1807 bonwick nlwb->lwb_blk = *bp; 858 789 ahrens nlwb->lwb_nused = 0; 859 789 ahrens nlwb->lwb_sz = BP_GET_LSIZE(&nlwb->lwb_blk); 860 789 ahrens nlwb->lwb_buf = zio_buf_alloc(nlwb->lwb_sz); 861 789 ahrens nlwb->lwb_max_txg = txg; 862 2237 maybee nlwb->lwb_zio = NULL; 863 10922 Jeff nlwb->lwb_tx = NULL; 864 789 ahrens 865 789 ahrens /* 866 3063 perrin * Put new lwb at the end of the log chain 867 789 ahrens */ 868 789 ahrens mutex_enter(&zilog->zl_lock); 869 789 ahrens list_insert_tail(&zilog->zl_lwb_list, nlwb); 870 3063 perrin mutex_exit(&zilog->zl_lock); 871 3063 perrin 872 5688 bonwick /* Record the block for later vdev flushing */ 873 5688 bonwick zil_add_block(zilog, &lwb->lwb_blk); 874 789 ahrens 875 789 ahrens /* 876 2237 maybee * kick off the write for the old log block 877 789 ahrens */ 878 3063 perrin ASSERT(lwb->lwb_zio); 879 2237 maybee zio_nowait(lwb->lwb_zio); 880 789 ahrens 881 789 ahrens return (nlwb); 882 789 ahrens } 883 789 ahrens 884 789 ahrens static lwb_t * 885 789 ahrens zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb) 886 789 ahrens { 887 789 ahrens lr_t *lrc = &itx->itx_lr; /* common log record */ 888 10922 Jeff lr_write_t *lrw = (lr_write_t *)lrc; 889 10922 Jeff char *lr_buf; 890 789 ahrens uint64_t txg = lrc->lrc_txg; 891 789 ahrens uint64_t reclen = lrc->lrc_reclen; 892 10922 Jeff uint64_t dlen = 0; 893 789 ahrens 894 789 ahrens if (lwb == NULL) 895 789 ahrens return (NULL); 896 10922 Jeff 897 789 ahrens ASSERT(lwb->lwb_buf != NULL); 898 789 ahrens 899 2237 maybee if (lrc->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) 900 2237 maybee dlen = P2ROUNDUP_TYPED( 901 10922 Jeff lrw->lr_length, sizeof (uint64_t), uint64_t); 902 1669 perrin 903 1669 perrin zilog->zl_cur_used += (reclen + dlen); 904 1669 perrin 905 3063 perrin zil_lwb_write_init(zilog, lwb); 906 3063 perrin 907 1669 perrin /* 908 1669 perrin * If this record won't fit in the current log block, start a new one. 909 1669 perrin */ 910 1669 perrin if (lwb->lwb_nused + reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 911 1669 perrin lwb = zil_lwb_write_start(zilog, lwb); 912 2237 maybee if (lwb == NULL) 913 1669 perrin return (NULL); 914 3063 perrin zil_lwb_write_init(zilog, lwb); 915 1669 perrin ASSERT(lwb->lwb_nused == 0); 916 1669 perrin if (reclen + dlen > ZIL_BLK_DATA_SZ(lwb)) { 917 1669 perrin txg_wait_synced(zilog->zl_dmu_pool, txg); 918 789 ahrens return (lwb); 919 789 ahrens } 920 789 ahrens } 921 1141 perrin 922 10922 Jeff lr_buf = lwb->lwb_buf + lwb->lwb_nused; 923 10922 Jeff bcopy(lrc, lr_buf, reclen); 924 10922 Jeff lrc = (lr_t *)lr_buf; 925 10922 Jeff lrw = (lr_write_t *)lrc; 926 2237 maybee 927 2237 maybee /* 928 2237 maybee * If it's a write, fetch the data or get its blkptr as appropriate. 929 2237 maybee */ 930 2237 maybee if (lrc->lrc_txtype == TX_WRITE) { 931 2237 maybee if (txg > spa_freeze_txg(zilog->zl_spa)) 932 2237 maybee txg_wait_synced(zilog->zl_dmu_pool, txg); 933 2237 maybee if (itx->itx_wr_state != WR_COPIED) { 934 2237 maybee char *dbuf; 935 2237 maybee int error; 936 2237 maybee 937 2237 maybee if (dlen) { 938 2237 maybee ASSERT(itx->itx_wr_state == WR_NEED_COPY); 939 10922 Jeff dbuf = lr_buf + reclen; 940 10922 Jeff lrw->lr_common.lrc_reclen += dlen; 941 2237 maybee } else { 942 2237 maybee ASSERT(itx->itx_wr_state == WR_INDIRECT); 943 2237 maybee dbuf = NULL; 944 2237 maybee } 945 2237 maybee error = zilog->zl_get_data( 946 10922 Jeff itx->itx_private, lrw, dbuf, lwb->lwb_zio); 947 10209 Mark if (error == EIO) { 948 10209 Mark txg_wait_synced(zilog->zl_dmu_pool, txg); 949 10209 Mark return (lwb); 950 10209 Mark } 951 2237 maybee if (error) { 952 2237 maybee ASSERT(error == ENOENT || error == EEXIST || 953 2237 maybee error == EALREADY); 954 2237 maybee return (lwb); 955 2237 maybee } 956 2237 maybee } 957 1669 perrin } 958 2237 maybee 959 10922 Jeff /* 960 10922 Jeff * We're actually making an entry, so update lrc_seq to be the 961 10922 Jeff * log record sequence number. Note that this is generally not 962 10922 Jeff * equal to the itx sequence number because not all transactions 963 10922 Jeff * are synchronous, and sometimes spa_sync() gets there first. 964 10922 Jeff */ 965 10922 Jeff lrc->lrc_seq = ++zilog->zl_lr_seq; /* we are single threaded */ 966 2237 maybee lwb->lwb_nused += reclen + dlen; 967 789 ahrens lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg); 968 789 ahrens ASSERT3U(lwb->lwb_nused, <=, ZIL_BLK_DATA_SZ(lwb)); 969 789 ahrens ASSERT3U(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)), ==, 0); 970 789 ahrens 971 789 ahrens return (lwb); 972 789 ahrens } 973 789 ahrens 974 789 ahrens itx_t * 975 5331 amw zil_itx_create(uint64_t txtype, size_t lrsize) 976 789 ahrens { 977 789 ahrens itx_t *itx; 978 789 ahrens 979 1842 perrin lrsize = P2ROUNDUP_TYPED(lrsize, sizeof (uint64_t), size_t); 980 789 ahrens 981 789 ahrens itx = kmem_alloc(offsetof(itx_t, itx_lr) + lrsize, KM_SLEEP); 982 789 ahrens itx->itx_lr.lrc_txtype = txtype; 983 789 ahrens itx->itx_lr.lrc_reclen = lrsize; 984 6101 perrin itx->itx_sod = lrsize; /* if write & WR_NEED_COPY will be increased */ 985 789 ahrens itx->itx_lr.lrc_seq = 0; /* defensive */ 986 789 ahrens 987 789 ahrens return (itx); 988 789 ahrens } 989 789 ahrens 990 10922 Jeff void 991 10922 Jeff zil_itx_destroy(itx_t *itx) 992 10922 Jeff { 993 10922 Jeff kmem_free(itx, offsetof(itx_t, itx_lr) + itx->itx_lr.lrc_reclen); 994 10922 Jeff } 995 10922 Jeff 996 789 ahrens uint64_t 997 789 ahrens zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx) 998 789 ahrens { 999 789 ahrens uint64_t seq; 1000 789 ahrens 1001 789 ahrens ASSERT(itx->itx_lr.lrc_seq == 0); 1002 10922 Jeff ASSERT(!zilog->zl_replay); 1003 789 ahrens 1004 789 ahrens mutex_enter(&zilog->zl_lock); 1005 789 ahrens list_insert_tail(&zilog->zl_itx_list, itx); 1006 6101 perrin zilog->zl_itx_list_sz += itx->itx_sod; 1007 789 ahrens itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx); 1008 789 ahrens itx->itx_lr.lrc_seq = seq = ++zilog->zl_itx_seq; 1009 789 ahrens mutex_exit(&zilog->zl_lock); 1010 789 ahrens 1011 789 ahrens return (seq); 1012 789 ahrens } 1013 789 ahrens 1014 789 ahrens /* 1015 789 ahrens * Free up all in-memory intent log transactions that have now been synced. 1016 789 ahrens */ 1017 789 ahrens static void 1018 789 ahrens zil_itx_clean(zilog_t *zilog) 1019 789 ahrens { 1020 789 ahrens uint64_t synced_txg = spa_last_synced_txg(zilog->zl_spa); 1021 789 ahrens uint64_t freeze_txg = spa_freeze_txg(zilog->zl_spa); 1022 3778 johansen list_t clean_list; 1023 789 ahrens itx_t *itx; 1024 3778 johansen 1025 3778 johansen list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node)); 1026 789 ahrens 1027 789 ahrens mutex_enter(&zilog->zl_lock); 1028 2638 perrin /* wait for a log writer to finish walking list */ 1029 2638 perrin while (zilog->zl_writer) { 1030 2638 perrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1031 2638 perrin } 1032 3778 johansen 1033 3778 johansen /* 1034 3778 johansen * Move the sync'd log transactions to a separate list so we can call 1035 3778 johansen * kmem_free without holding the zl_lock. 1036 3778 johansen * 1037 3778 johansen * There is no need to set zl_writer as we don't drop zl_lock here 1038 3778 johansen */ 1039 789 ahrens while ((itx = list_head(&zilog->zl_itx_list)) != NULL && 1040 789 ahrens itx->itx_lr.lrc_txg <= MIN(synced_txg, freeze_txg)) { 1041 789 ahrens list_remove(&zilog->zl_itx_list, itx); 1042 6101 perrin zilog->zl_itx_list_sz -= itx->itx_sod; 1043 3778 johansen list_insert_tail(&clean_list, itx); 1044 3778 johansen } 1045 3778 johansen cv_broadcast(&zilog->zl_cv_writer); 1046 3778 johansen mutex_exit(&zilog->zl_lock); 1047 3778 johansen 1048 3778 johansen /* destroy sync'd log transactions */ 1049 3778 johansen while ((itx = list_head(&clean_list)) != NULL) { 1050 3778 johansen list_remove(&clean_list, itx); 1051 10922 Jeff zil_itx_destroy(itx); 1052 789 ahrens } 1053 3778 johansen list_destroy(&clean_list); 1054 789 ahrens } 1055 789 ahrens 1056 2638 perrin /* 1057 3063 perrin * If there are any in-memory intent log transactions which have now been 1058 3063 perrin * synced then start up a taskq to free them. 1059 2638 perrin */ 1060 789 ahrens void 1061 789 ahrens zil_clean(zilog_t *zilog) 1062 789 ahrens { 1063 3063 perrin itx_t *itx; 1064 3063 perrin 1065 789 ahrens mutex_enter(&zilog->zl_lock); 1066 3063 perrin itx = list_head(&zilog->zl_itx_list); 1067 3063 perrin if ((itx != NULL) && 1068 3063 perrin (itx->itx_lr.lrc_txg <= spa_last_synced_txg(zilog->zl_spa))) { 1069 789 ahrens (void) taskq_dispatch(zilog->zl_clean_taskq, 1070 10879 Neil (task_func_t *)zil_itx_clean, zilog, TQ_NOSLEEP); 1071 3063 perrin } 1072 789 ahrens mutex_exit(&zilog->zl_lock); 1073 789 ahrens } 1074 789 ahrens 1075 7754 Jeff static void 1076 2638 perrin zil_commit_writer(zilog_t *zilog, uint64_t seq, uint64_t foid) 1077 789 ahrens { 1078 789 ahrens uint64_t txg; 1079 3063 perrin uint64_t commit_seq = 0; 1080 10922 Jeff itx_t *itx, *itx_next; 1081 789 ahrens lwb_t *lwb; 1082 789 ahrens spa_t *spa; 1083 10922 Jeff int error = 0; 1084 789 ahrens 1085 2638 perrin zilog->zl_writer = B_TRUE; 1086 7754 Jeff ASSERT(zilog->zl_root_zio == NULL); 1087 789 ahrens spa = zilog->zl_spa; 1088 789 ahrens 1089 789 ahrens if (zilog->zl_suspend) { 1090 789 ahrens lwb = NULL; 1091 789 ahrens } else { 1092 789 ahrens lwb = list_tail(&zilog->zl_lwb_list); 1093 789 ahrens if (lwb == NULL) { 1094 2638 perrin /* 1095 2638 perrin * Return if there's nothing to flush before we 1096 2638 perrin * dirty the fs by calling zil_create() 1097 2638 perrin */ 1098 2638 perrin if (list_is_empty(&zilog->zl_itx_list)) { 1099 2638 perrin zilog->zl_writer = B_FALSE; 1100 2638 perrin return; 1101 2638 perrin } 1102 789 ahrens mutex_exit(&zilog->zl_lock); 1103 789 ahrens zil_create(zilog); 1104 789 ahrens mutex_enter(&zilog->zl_lock); 1105 789 ahrens lwb = list_tail(&zilog->zl_lwb_list); 1106 789 ahrens } 1107 789 ahrens } 1108 789 ahrens 1109 3063 perrin /* Loop through in-memory log transactions filling log blocks. */ 1110 2638 perrin DTRACE_PROBE1(zil__cw1, zilog_t *, zilog); 1111 10922 Jeff 1112 10922 Jeff for (itx = list_head(&zilog->zl_itx_list); itx; itx = itx_next) { 1113 2638 perrin /* 1114 10922 Jeff * Save the next pointer. Even though we drop zl_lock below, 1115 10922 Jeff * all threads that can remove itx list entries (other writers 1116 10922 Jeff * and zil_itx_clean()) can't do so until they have zl_writer. 1117 2638 perrin */ 1118 10922 Jeff itx_next = list_next(&zilog->zl_itx_list, itx); 1119 10922 Jeff 1120 10922 Jeff /* 1121 10922 Jeff * Determine whether to push this itx. 1122 10922 Jeff * Push all transactions related to specified foid and 1123 10922 Jeff * all other transactions except those that can be logged 1124 10922 Jeff * out of order (TX_WRITE, TX_TRUNCATE, TX_SETATTR, TX_ACL) 1125 10922 Jeff * for all other files. 1126 10922 Jeff * 1127 10922 Jeff * If foid == 0 (meaning "push all foids") or 1128 10922 Jeff * itx->itx_sync is set (meaning O_[D]SYNC), push regardless. 1129 10922 Jeff */ 1130 10922 Jeff if (foid != 0 && !itx->itx_sync && 1131 10922 Jeff TX_OOO(itx->itx_lr.lrc_txtype) && 1132 10922 Jeff ((lr_ooo_t *)&itx->itx_lr)->lr_foid != foid) 1133 10922 Jeff continue; /* skip this record */ 1134 789 ahrens 1135 789 ahrens if ((itx->itx_lr.lrc_seq > seq) && 1136 2638 perrin ((lwb == NULL) || (lwb->lwb_nused == 0) || 1137 10922 Jeff (lwb->lwb_nused + itx->itx_sod > ZIL_BLK_DATA_SZ(lwb)))) 1138 789 ahrens break; 1139 789 ahrens 1140 789 ahrens list_remove(&zilog->zl_itx_list, itx); 1141 6101 perrin zilog->zl_itx_list_sz -= itx->itx_sod; 1142 10922 Jeff 1143 3063 perrin mutex_exit(&zilog->zl_lock); 1144 10922 Jeff 1145 789 ahrens txg = itx->itx_lr.lrc_txg; 1146 789 ahrens ASSERT(txg); 1147 789 ahrens 1148 789 ahrens if (txg > spa_last_synced_txg(spa) || 1149 789 ahrens txg > spa_freeze_txg(spa)) 1150 789 ahrens lwb = zil_lwb_commit(zilog, itx, lwb); 1151 10922 Jeff 1152 10922 Jeff zil_itx_destroy(itx); 1153 10922 Jeff 1154 789 ahrens mutex_enter(&zilog->zl_lock); 1155 789 ahrens } 1156 2638 perrin DTRACE_PROBE1(zil__cw2, zilog_t *, zilog); 1157 3063 perrin /* determine commit sequence number */ 1158 3063 perrin itx = list_head(&zilog->zl_itx_list); 1159 3063 perrin if (itx) 1160 10922 Jeff commit_seq = itx->itx_lr.lrc_seq - 1; 1161 3063 perrin else 1162 3063 perrin commit_seq = zilog->zl_itx_seq; 1163 789 ahrens mutex_exit(&zilog->zl_lock); 1164 789 ahrens 1165 789 ahrens /* write the last block out */ 1166 3063 perrin if (lwb != NULL && lwb->lwb_zio != NULL) 1167 789 ahrens lwb = zil_lwb_write_start(zilog, lwb); 1168 789 ahrens 1169 1141 perrin zilog->zl_prev_used = zilog->zl_cur_used; 1170 1141 perrin zilog->zl_cur_used = 0; 1171 1141 perrin 1172 2638 perrin /* 1173 2638 perrin * Wait if necessary for the log blocks to be on stable storage. 1174 2638 perrin */ 1175 2638 perrin if (zilog->zl_root_zio) { 1176 2638 perrin DTRACE_PROBE1(zil__cw3, zilog_t *, zilog); 1177 10922 Jeff error = zio_wait(zilog->zl_root_zio); 1178 7754 Jeff zilog->zl_root_zio = NULL; 1179 2638 perrin DTRACE_PROBE1(zil__cw4, zilog_t *, zilog); 1180 5688 bonwick zil_flush_vdevs(zilog); 1181 789 ahrens } 1182 1141 perrin 1183 10922 Jeff if (error || lwb == NULL) 1184 789 ahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1185 3063 perrin 1186 3063 perrin mutex_enter(&zilog->zl_lock); 1187 1141 perrin zilog->zl_writer = B_FALSE; 1188 3063 perrin 1189 3063 perrin ASSERT3U(commit_seq, >=, zilog->zl_commit_seq); 1190 3063 perrin zilog->zl_commit_seq = commit_seq; 1191 10922 Jeff 1192 10922 Jeff /* 1193 10922 Jeff * Remember the highest committed log sequence number for ztest. 1194 10922 Jeff * We only update this value when all the log writes succeeded, 1195 10922 Jeff * because ztest wants to ASSERT that it got the whole log chain. 1196 10922 Jeff */ 1197 10922 Jeff if (error == 0 && lwb != NULL) 1198 10922 Jeff zilog->zl_commit_lr_seq = zilog->zl_lr_seq; 1199 2638 perrin } 1200 2638 perrin 1201 2638 perrin /* 1202 2638 perrin * Push zfs transactions to stable storage up to the supplied sequence number. 1203 2638 perrin * If foid is 0 push out all transactions, otherwise push only those 1204 2638 perrin * for that file or might have been used to create that file. 1205 2638 perrin */ 1206 2638 perrin void 1207 2638 perrin zil_commit(zilog_t *zilog, uint64_t seq, uint64_t foid) 1208 2638 perrin { 1209 2638 perrin if (zilog == NULL || seq == 0) 1210 2638 perrin return; 1211 2638 perrin 1212 2638 perrin mutex_enter(&zilog->zl_lock); 1213 2638 perrin 1214 2638 perrin seq = MIN(seq, zilog->zl_itx_seq); /* cap seq at largest itx seq */ 1215 2638 perrin 1216 3063 perrin while (zilog->zl_writer) { 1217 2638 perrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1218 10922 Jeff if (seq <= zilog->zl_commit_seq) { 1219 3063 perrin mutex_exit(&zilog->zl_lock); 1220 3063 perrin return; 1221 3063 perrin } 1222 3063 perrin } 1223 2638 perrin zil_commit_writer(zilog, seq, foid); /* drops zl_lock */ 1224 3063 perrin /* wake up others waiting on the commit */ 1225 3063 perrin cv_broadcast(&zilog->zl_cv_writer); 1226 3063 perrin mutex_exit(&zilog->zl_lock); 1227 789 ahrens } 1228 789 ahrens 1229 789 ahrens /* 1230 10922 Jeff * Report whether all transactions are committed. 1231 10922 Jeff */ 1232 10922 Jeff static boolean_t 1233 10922 Jeff zil_is_committed(zilog_t *zilog) 1234 10922 Jeff { 1235 10922 Jeff lwb_t *lwb; 1236 10922 Jeff boolean_t committed; 1237 10922 Jeff 1238 10922 Jeff mutex_enter(&zilog->zl_lock); 1239 10922 Jeff 1240 10922 Jeff while (zilog->zl_writer) 1241 10922 Jeff cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1242 10922 Jeff 1243 10922 Jeff if (!list_is_empty(&zilog->zl_itx_list)) 1244 10922 Jeff committed = B_FALSE; /* unpushed transactions */ 1245 10922 Jeff else if ((lwb = list_head(&zilog->zl_lwb_list)) == NULL) 1246 10922 Jeff committed = B_TRUE; /* intent log never used */ 1247 10922 Jeff else if (list_next(&zilog->zl_lwb_list, lwb) != NULL) 1248 10922 Jeff committed = B_FALSE; /* zil_sync() not done yet */ 1249 10922 Jeff else 1250 10922 Jeff committed = B_TRUE; /* everything synced */ 1251 10922 Jeff 1252 10922 Jeff mutex_exit(&zilog->zl_lock); 1253 10922 Jeff return (committed); 1254 10922 Jeff } 1255 10922 Jeff 1256 10922 Jeff /* 1257 789 ahrens * Called in syncing context to free committed log blocks and update log header. 1258 789 ahrens */ 1259 789 ahrens void 1260 789 ahrens zil_sync(zilog_t *zilog, dmu_tx_t *tx) 1261 789 ahrens { 1262 1807 bonwick zil_header_t *zh = zil_header_in_syncing_context(zilog); 1263 789 ahrens uint64_t txg = dmu_tx_get_txg(tx); 1264 789 ahrens spa_t *spa = zilog->zl_spa; 1265 10922 Jeff uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK]; 1266 789 ahrens lwb_t *lwb; 1267 789 ahrens 1268 9396 Matthew /* 1269 9396 Matthew * We don't zero out zl_destroy_txg, so make sure we don't try 1270 9396 Matthew * to destroy it twice. 1271 9396 Matthew */ 1272 9396 Matthew if (spa_sync_pass(spa) != 1) 1273 9396 Matthew return; 1274 9396 Matthew 1275 1807 bonwick mutex_enter(&zilog->zl_lock); 1276 1807 bonwick 1277 789 ahrens ASSERT(zilog->zl_stop_sync == 0); 1278 789 ahrens 1279 10922 Jeff if (*replayed_seq != 0) { 1280 10922 Jeff ASSERT(zh->zh_replay_seq < *replayed_seq); 1281 10922 Jeff zh->zh_replay_seq = *replayed_seq; 1282 10922 Jeff *replayed_seq = 0; 1283 10922 Jeff } 1284 789 ahrens 1285 789 ahrens if (zilog->zl_destroy_txg == txg) { 1286 1807 bonwick blkptr_t blk = zh->zh_log; 1287 1807 bonwick 1288 1807 bonwick ASSERT(list_head(&zilog->zl_lwb_list) == NULL); 1289 1807 bonwick 1290 1807 bonwick bzero(zh, sizeof (zil_header_t)); 1291 8227 Neil bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq)); 1292 1807 bonwick 1293 1807 bonwick if (zilog->zl_keep_first) { 1294 1807 bonwick /* 1295 1807 bonwick * If this block was part of log chain that couldn't 1296 1807 bonwick * be claimed because a device was missing during 1297 1807 bonwick * zil_claim(), but that device later returns, 1298 1807 bonwick * then this block could erroneously appear valid. 1299 1807 bonwick * To guard against this, assign a new GUID to the new 1300 1807 bonwick * log chain so it doesn't matter what blk points to. 1301 1807 bonwick */ 1302 1807 bonwick zil_init_log_chain(zilog, &blk); 1303 1807 bonwick zh->zh_log = blk; 1304 1807 bonwick } 1305 789 ahrens } 1306 789 ahrens 1307 9701 George while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1308 2638 perrin zh->zh_log = lwb->lwb_blk; 1309 789 ahrens if (lwb->lwb_buf != NULL || lwb->lwb_max_txg > txg) 1310 789 ahrens break; 1311 789 ahrens list_remove(&zilog->zl_lwb_list, lwb); 1312 10922 Jeff zio_free_zil(spa, txg, &lwb->lwb_blk); 1313 789 ahrens kmem_cache_free(zil_lwb_cache, lwb); 1314 3668 gw25295 1315 3668 gw25295 /* 1316 3668 gw25295 * If we don't have anything left in the lwb list then 1317 3668 gw25295 * we've had an allocation failure and we need to zero 1318 3668 gw25295 * out the zil_header blkptr so that we don't end 1319 3668 gw25295 * up freeing the same block twice. 1320 3668 gw25295 */ 1321 3668 gw25295 if (list_head(&zilog->zl_lwb_list) == NULL) 1322 3668 gw25295 BP_ZERO(&zh->zh_log); 1323 789 ahrens } 1324 789 ahrens mutex_exit(&zilog->zl_lock); 1325 789 ahrens } 1326 789 ahrens 1327 789 ahrens void 1328 789 ahrens zil_init(void) 1329 789 ahrens { 1330 789 ahrens zil_lwb_cache = kmem_cache_create("zil_lwb_cache", 1331 2856 nd150628 sizeof (struct lwb), 0, NULL, NULL, NULL, NULL, NULL, 0); 1332 789 ahrens } 1333 789 ahrens 1334 789 ahrens void 1335 789 ahrens zil_fini(void) 1336 789 ahrens { 1337 789 ahrens kmem_cache_destroy(zil_lwb_cache); 1338 789 ahrens } 1339 789 ahrens 1340 10310 Neil void 1341 10310 Neil zil_set_logbias(zilog_t *zilog, uint64_t logbias) 1342 10310 Neil { 1343 10310 Neil zilog->zl_logbias = logbias; 1344 10310 Neil } 1345 10310 Neil 1346 789 ahrens zilog_t * 1347 789 ahrens zil_alloc(objset_t *os, zil_header_t *zh_phys) 1348 789 ahrens { 1349 789 ahrens zilog_t *zilog; 1350 789 ahrens 1351 789 ahrens zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP); 1352 789 ahrens 1353 789 ahrens zilog->zl_header = zh_phys; 1354 789 ahrens zilog->zl_os = os; 1355 789 ahrens zilog->zl_spa = dmu_objset_spa(os); 1356 789 ahrens zilog->zl_dmu_pool = dmu_objset_pool(os); 1357 1807 bonwick zilog->zl_destroy_txg = TXG_INITIAL - 1; 1358 10310 Neil zilog->zl_logbias = dmu_objset_logbias(os); 1359 2856 nd150628 1360 2856 nd150628 mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL); 1361 789 ahrens 1362 789 ahrens list_create(&zilog->zl_itx_list, sizeof (itx_t), 1363 789 ahrens offsetof(itx_t, itx_node)); 1364 789 ahrens 1365 789 ahrens list_create(&zilog->zl_lwb_list, sizeof (lwb_t), 1366 789 ahrens offsetof(lwb_t, lwb_node)); 1367 789 ahrens 1368 5688 bonwick mutex_init(&zilog->zl_vdev_lock, NULL, MUTEX_DEFAULT, NULL); 1369 5688 bonwick 1370 5688 bonwick avl_create(&zilog->zl_vdev_tree, zil_vdev_compare, 1371 5688 bonwick sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node)); 1372 789 ahrens 1373 5913 perrin cv_init(&zilog->zl_cv_writer, NULL, CV_DEFAULT, NULL); 1374 5913 perrin cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL); 1375 5913 perrin 1376 789 ahrens return (zilog); 1377 789 ahrens } 1378 789 ahrens 1379 789 ahrens void 1380 789 ahrens zil_free(zilog_t *zilog) 1381 789 ahrens { 1382 789 ahrens lwb_t *lwb; 1383 789 ahrens 1384 789 ahrens zilog->zl_stop_sync = 1; 1385 789 ahrens 1386 789 ahrens while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) { 1387 789 ahrens list_remove(&zilog->zl_lwb_list, lwb); 1388 789 ahrens if (lwb->lwb_buf != NULL) 1389 789 ahrens zio_buf_free(lwb->lwb_buf, lwb->lwb_sz); 1390 789 ahrens kmem_cache_free(zil_lwb_cache, lwb); 1391 789 ahrens } 1392 789 ahrens list_destroy(&zilog->zl_lwb_list); 1393 789 ahrens 1394 5688 bonwick avl_destroy(&zilog->zl_vdev_tree); 1395 5688 bonwick mutex_destroy(&zilog->zl_vdev_lock); 1396 789 ahrens 1397 789 ahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1398 789 ahrens list_destroy(&zilog->zl_itx_list); 1399 2856 nd150628 mutex_destroy(&zilog->zl_lock); 1400 5913 perrin 1401 5913 perrin cv_destroy(&zilog->zl_cv_writer); 1402 5913 perrin cv_destroy(&zilog->zl_cv_suspend); 1403 789 ahrens 1404 789 ahrens kmem_free(zilog, sizeof (zilog_t)); 1405 789 ahrens } 1406 789 ahrens 1407 789 ahrens /* 1408 789 ahrens * Open an intent log. 1409 789 ahrens */ 1410 789 ahrens zilog_t * 1411 789 ahrens zil_open(objset_t *os, zil_get_data_t *get_data) 1412 789 ahrens { 1413 789 ahrens zilog_t *zilog = dmu_objset_zil(os); 1414 789 ahrens 1415 789 ahrens zilog->zl_get_data = get_data; 1416 789 ahrens zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri, 1417 789 ahrens 2, 2, TASKQ_PREPOPULATE); 1418 789 ahrens 1419 789 ahrens return (zilog); 1420 789 ahrens } 1421 789 ahrens 1422 789 ahrens /* 1423 789 ahrens * Close an intent log. 1424 789 ahrens */ 1425 789 ahrens void 1426 789 ahrens zil_close(zilog_t *zilog) 1427 789 ahrens { 1428 1807 bonwick /* 1429 1807 bonwick * If the log isn't already committed, mark the objset dirty 1430 1807 bonwick * (so zil_sync() will be called) and wait for that txg to sync. 1431 1807 bonwick */ 1432 1807 bonwick if (!zil_is_committed(zilog)) { 1433 1807 bonwick uint64_t txg; 1434 1807 bonwick dmu_tx_t *tx = dmu_tx_create(zilog->zl_os); 1435 10922 Jeff VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0); 1436 1807 bonwick dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1437 1807 bonwick txg = dmu_tx_get_txg(tx); 1438 1807 bonwick dmu_tx_commit(tx); 1439 1807 bonwick txg_wait_synced(zilog->zl_dmu_pool, txg); 1440 1807 bonwick } 1441 1807 bonwick 1442 789 ahrens taskq_destroy(zilog->zl_clean_taskq); 1443 789 ahrens zilog->zl_clean_taskq = NULL; 1444 789 ahrens zilog->zl_get_data = NULL; 1445 789 ahrens 1446 789 ahrens zil_itx_clean(zilog); 1447 789 ahrens ASSERT(list_head(&zilog->zl_itx_list) == NULL); 1448 789 ahrens } 1449 789 ahrens 1450 789 ahrens /* 1451 789 ahrens * Suspend an intent log. While in suspended mode, we still honor 1452 789 ahrens * synchronous semantics, but we rely on txg_wait_synced() to do it. 1453 789 ahrens * We suspend the log briefly when taking a snapshot so that the snapshot 1454 789 ahrens * contains all the data it's supposed to, and has an empty intent log. 1455 789 ahrens */ 1456 789 ahrens int 1457 789 ahrens zil_suspend(zilog_t *zilog) 1458 789 ahrens { 1459 1807 bonwick const zil_header_t *zh = zilog->zl_header; 1460 789 ahrens 1461 789 ahrens mutex_enter(&zilog->zl_lock); 1462 8989 Neil if (zh->zh_flags & ZIL_REPLAY_NEEDED) { /* unplayed log */ 1463 789 ahrens mutex_exit(&zilog->zl_lock); 1464 789 ahrens return (EBUSY); 1465 789 ahrens } 1466 1807 bonwick if (zilog->zl_suspend++ != 0) { 1467 1807 bonwick /* 1468 1807 bonwick * Someone else already began a suspend. 1469 1807 bonwick * Just wait for them to finish. 1470 1807 bonwick */ 1471 1807 bonwick while (zilog->zl_suspending) 1472 1807 bonwick cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock); 1473 1807 bonwick mutex_exit(&zilog->zl_lock); 1474 1807 bonwick return (0); 1475 1807 bonwick } 1476 1807 bonwick zilog->zl_suspending = B_TRUE; 1477 789 ahrens mutex_exit(&zilog->zl_lock); 1478 789 ahrens 1479 2638 perrin zil_commit(zilog, UINT64_MAX, 0); 1480 789 ahrens 1481 2638 perrin /* 1482 2638 perrin * Wait for any in-flight log writes to complete. 1483 2638 perrin */ 1484 789 ahrens mutex_enter(&zilog->zl_lock); 1485 2638 perrin while (zilog->zl_writer) 1486 2638 perrin cv_wait(&zilog->zl_cv_writer, &zilog->zl_lock); 1487 789 ahrens mutex_exit(&zilog->zl_lock); 1488 789 ahrens 1489 1807 bonwick zil_destroy(zilog, B_FALSE); 1490 1807 bonwick 1491 1807 bonwick mutex_enter(&zilog->zl_lock); 1492 1807 bonwick zilog->zl_suspending = B_FALSE; 1493 1807 bonwick cv_broadcast(&zilog->zl_cv_suspend); 1494 1807 bonwick mutex_exit(&zilog->zl_lock); 1495 789 ahrens 1496 789 ahrens return (0); 1497 789 ahrens } 1498 789 ahrens 1499 789 ahrens void 1500 789 ahrens zil_resume(zilog_t *zilog) 1501 789 ahrens { 1502 789 ahrens mutex_enter(&zilog->zl_lock); 1503 789 ahrens ASSERT(zilog->zl_suspend != 0); 1504 789 ahrens zilog->zl_suspend--; 1505 789 ahrens mutex_exit(&zilog->zl_lock); 1506 789 ahrens } 1507 789 ahrens 1508 789 ahrens typedef struct zil_replay_arg { 1509 789 ahrens zil_replay_func_t **zr_replay; 1510 789 ahrens void *zr_arg; 1511 789 ahrens boolean_t zr_byteswap; 1512 10922 Jeff char *zr_lr; 1513 789 ahrens } zil_replay_arg_t; 1514 789 ahrens 1515 10922 Jeff static int 1516 10922 Jeff zil_replay_error(zilog_t *zilog, lr_t *lr, int error) 1517 10922 Jeff { 1518 10922 Jeff char name[MAXNAMELEN]; 1519 10922 Jeff 1520 10922 Jeff zilog->zl_replaying_seq--; /* didn't actually replay this one */ 1521 10922 Jeff 1522 10922 Jeff dmu_objset_name(zilog->zl_os, name); 1523 10922 Jeff 1524 10922 Jeff cmn_err(CE_WARN, "ZFS replay transaction error %d, " 1525 10922 Jeff "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name, 1526 10922 Jeff (u_longlong_t)lr->lrc_seq, 1527 10922 Jeff (u_longlong_t)(lr->lrc_txtype & ~TX_CI), 1528 10922 Jeff (lr->lrc_txtype & TX_CI) ? "CI" : ""); 1529 10922 Jeff 1530 10922 Jeff return (error); 1531 10922 Jeff } 1532 10922 Jeff 1533 10922 Jeff static int 1534 789 ahrens zil_replay_log_record(zilog_t *zilog, lr_t *lr, void *zra, uint64_t claim_txg) 1535 789 ahrens { 1536 789 ahrens zil_replay_arg_t *zr = zra; 1537 1807 bonwick const zil_header_t *zh = zilog->zl_header; 1538 789 ahrens uint64_t reclen = lr->lrc_reclen; 1539 789 ahrens uint64_t txtype = lr->lrc_txtype; 1540 10922 Jeff int error = 0; 1541 789 ahrens 1542 10922 Jeff zilog->zl_replaying_seq = lr->lrc_seq; 1543 10922 Jeff 1544 10922 Jeff if (lr->lrc_seq <= zh->zh_replay_seq) /* already replayed */ 1545 10922 Jeff return (0); 1546 789 ahrens 1547 789 ahrens if (lr->lrc_txg < claim_txg) /* already committed */ 1548 10922 Jeff return (0); 1549 789 ahrens 1550 5331 amw /* Strip case-insensitive bit, still present in log record */ 1551 5331 amw txtype &= ~TX_CI; 1552 8227 Neil 1553 10922 Jeff if (txtype == 0 || txtype >= TX_MAX_TYPE) 1554 10922 Jeff return (zil_replay_error(zilog, lr, EINVAL)); 1555 10922 Jeff 1556 10922 Jeff /* 1557 10922 Jeff * If this record type can be logged out of order, the object 1558 10922 Jeff * (lr_foid) may no longer exist. That's legitimate, not an error. 1559 10922 Jeff */ 1560 10922 Jeff if (TX_OOO(txtype)) { 1561 10922 Jeff error = dmu_object_info(zilog->zl_os, 1562 10922 Jeff ((lr_ooo_t *)lr)->lr_foid, NULL); 1563 10922 Jeff if (error == ENOENT || error == EEXIST) 1564 10922 Jeff return (0); 1565 8227 Neil } 1566 5331 amw 1567 789 ahrens /* 1568 789 ahrens * Make a copy of the data so we can revise and extend it. 1569 789 ahrens */ 1570 10922 Jeff bcopy(lr, zr->zr_lr, reclen); 1571 10922 Jeff 1572 10922 Jeff /* 1573 10922 Jeff * If this is a TX_WRITE with a blkptr, suck in the data. 1574 10922 Jeff */ 1575 10922 Jeff if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) { 1576 10922 Jeff error = zil_read_log_data(zilog, (lr_write_t *)lr, 1577 10922 Jeff zr->zr_lr + reclen); 1578 10922 Jeff if (error) 1579 10922 Jeff return (zil_replay_error(zilog, lr, error)); 1580 10922 Jeff } 1581 789 ahrens 1582 789 ahrens /* 1583 789 ahrens * The log block containing this lr may have been byteswapped 1584 789 ahrens * so that we can easily examine common fields like lrc_txtype. 1585 10922 Jeff * However, the log is a mix of different record types, and only the 1586 789 ahrens * replay vectors know how to byteswap their records. Therefore, if 1587 789 ahrens * the lr was byteswapped, undo it before invoking the replay vector. 1588 789 ahrens */ 1589 789 ahrens if (zr->zr_byteswap) 1590 10922 Jeff byteswap_uint64_array(zr->zr_lr, reclen); 1591 789 ahrens 1592 789 ahrens /* 1593 8227 Neil * We must now do two things atomically: replay this log record, 1594 8227 Neil * and update the log header sequence number to reflect the fact that 1595 8227 Neil * we did so. At the end of each replay function the sequence number 1596 8227 Neil * is updated if we are in replay mode. 1597 7904 Neil */ 1598 10922 Jeff error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap); 1599 10922 Jeff if (error) { 1600 3063 perrin /* 1601 3063 perrin * The DMU's dnode layer doesn't see removes until the txg 1602 3063 perrin * commits, so a subsequent claim can spuriously fail with 1603 8227 Neil * EEXIST. So if we receive any error we try syncing out 1604 10922 Jeff * any removes then retry the transaction. Note that we 1605 10922 Jeff * specify B_FALSE for byteswap now, so we don't do it twice. 1606 3063 perrin */ 1607 10922 Jeff txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0); 1608 10922 Jeff error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE); 1609 10922 Jeff if (error) 1610 10922 Jeff return (zil_replay_error(zilog, lr, error)); 1611 789 ahrens } 1612 10922 Jeff return (0); 1613 3063 perrin } 1614 789 ahrens 1615 3063 perrin /* ARGSUSED */ 1616 10922 Jeff static int 1617 3063 perrin zil_incr_blks(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 1618 3063 perrin { 1619 3063 perrin zilog->zl_replay_blks++; 1620 10922 Jeff 1621 10922 Jeff return (0); 1622 789 ahrens } 1623 789 ahrens 1624 789 ahrens /* 1625 1362 perrin * If this dataset has a non-empty intent log, replay it and destroy it. 1626 789 ahrens */ 1627 789 ahrens void 1628 8227 Neil zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE]) 1629 789 ahrens { 1630 789 ahrens zilog_t *zilog = dmu_objset_zil(os); 1631 1807 bonwick const zil_header_t *zh = zilog->zl_header; 1632 1807 bonwick zil_replay_arg_t zr; 1633 1362 perrin 1634 8989 Neil if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) { 1635 1807 bonwick zil_destroy(zilog, B_TRUE); 1636 1362 perrin return; 1637 1362 perrin } 1638 789 ahrens 1639 789 ahrens zr.zr_replay = replay_func; 1640 789 ahrens zr.zr_arg = arg; 1641 1807 bonwick zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log); 1642 10922 Jeff zr.zr_lr = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP); 1643 789 ahrens 1644 789 ahrens /* 1645 789 ahrens * Wait for in-progress removes to sync before starting replay. 1646 789 ahrens */ 1647 789 ahrens txg_wait_synced(zilog->zl_dmu_pool, 0); 1648 789 ahrens 1649 8227 Neil zilog->zl_replay = B_TRUE; 1650 11066 rafael zilog->zl_replay_time = ddi_get_lbolt(); 1651 3063 perrin ASSERT(zilog->zl_replay_blks == 0); 1652 3063 perrin (void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr, 1653 1807 bonwick zh->zh_claim_txg); 1654 10922 Jeff kmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE); 1655 789 ahrens 1656 1807 bonwick zil_destroy(zilog, B_FALSE); 1657 5712 ahrens txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg); 1658 8227 Neil zilog->zl_replay = B_FALSE; 1659 789 ahrens } 1660 1646 perrin 1661 10922 Jeff boolean_t 1662 10922 Jeff zil_replaying(zilog_t *zilog, dmu_tx_t *tx) 1663 1646 perrin { 1664 10922 Jeff if (zilog == NULL) 1665 10922 Jeff return (B_TRUE); 1666 1646 perrin 1667 10922 Jeff if (zilog->zl_replay) { 1668 10922 Jeff dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx); 1669 10922 Jeff zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] = 1670 10922 Jeff zilog->zl_replaying_seq; 1671 10922 Jeff return (B_TRUE); 1672 2638 perrin } 1673 2638 perrin 1674 10922 Jeff return (B_FALSE); 1675 1646 perrin } 1676 9701 George 1677 9701 George /* ARGSUSED */ 1678 9701 George int 1679 9701 George zil_vdev_offline(char *osname, void *arg) 1680 9701 George { 1681 9701 George objset_t *os; 1682 9701 George zilog_t *zilog; 1683 9701 George int error; 1684 9701 George 1685 10298 Matthew error = dmu_objset_hold(osname, FTAG, &os); 1686 9701 George if (error) 1687 9701 George return (error); 1688 9701 George 1689 9701 George zilog = dmu_objset_zil(os); 1690 9701 George if (zil_suspend(zilog) != 0) 1691 9701 George error = EEXIST; 1692 9701 George else 1693 9701 George zil_resume(zilog); 1694 10298 Matthew dmu_objset_rele(os, FTAG); 1695 9701 George return (error); 1696 9701 George } 1697
Indexes created Thu Nov 26 09:40:41 UTC 2009
Terms of Use |
Privacy |
Trademarks |
Copyright Policy |
Site Guidelines |
Help
Your use of this web site or any of its content or software indicates your agreement to be bound by these Terms of Use.
Copyright © 1995-2008 Sun Microsystems, Inc.