1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 */ 25 26 #include <sys/dsl_scan.h> 27 #include <sys/dsl_pool.h> 28 #include <sys/dsl_dataset.h> 29 #include <sys/dsl_prop.h> 30 #include <sys/dsl_dir.h> 31 #include <sys/dsl_synctask.h> 32 #include <sys/dnode.h> 33 #include <sys/dmu_tx.h> 34 #include <sys/dmu_objset.h> 35 #include <sys/arc.h> 36 #include <sys/zap.h> 37 #include <sys/zio.h> 38 #include <sys/zfs_context.h> 39 #include <sys/fs/zfs.h> 40 #include <sys/zfs_znode.h> 41 #include <sys/spa_impl.h> 42 #include <sys/vdev_impl.h> 43 #include <sys/zil_impl.h> 44 #include <sys/zio_checksum.h> 45 #include <sys/ddt.h> 46 #include <sys/sa.h> 47 #include <sys/sa_impl.h> 48 #include <sys/zfeature.h> 49 #ifdef _KERNEL 50 #include <sys/zfs_vfsops.h> 51 #endif 52 53 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, 54 const zbookmark_phys_t *); 55 56 static scan_cb_t dsl_scan_scrub_cb; 57 static void dsl_scan_cancel_sync(void *, dmu_tx_t *); 58 static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx); 59 60 int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */ 61 int zfs_resilver_delay = 2; /* number of ticks to delay resilver */ 62 int zfs_scrub_delay = 4; /* number of ticks to delay scrub */ 63 int zfs_scan_idle = 50; /* idle window in clock ticks */ 64 65 int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */ 66 int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */ 67 int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */ 68 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */ 69 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */ 70 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE; 71 int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */ 72 73 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \ 74 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \ 75 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER) 76 77 extern int zfs_txg_timeout; 78 79 /* the order has to match pool_scan_type */ 80 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = { 81 NULL, 82 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */ 83 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */ 84 }; 85 86 int 87 dsl_scan_init(dsl_pool_t *dp, uint64_t txg) 88 { 89 int err; 90 dsl_scan_t *scn; 91 spa_t *spa = dp->dp_spa; 92 uint64_t f; 93 94 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP); 95 scn->scn_dp = dp; 96 97 /* 98 * It's possible that we're resuming a scan after a reboot so 99 * make sure that the scan_async_destroying flag is initialized 100 * appropriately. 101 */ 102 ASSERT(!scn->scn_async_destroying); 103 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa, 104 SPA_FEATURE_ASYNC_DESTROY); 105 106 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 107 "scrub_func", sizeof (uint64_t), 1, &f); 108 if (err == 0) { 109 /* 110 * There was an old-style scrub in progress. Restart a 111 * new-style scrub from the beginning. 112 */ 113 scn->scn_restart_txg = txg; 114 zfs_dbgmsg("old-style scrub was in progress; " 115 "restarting new-style scrub in txg %llu", 116 scn->scn_restart_txg); 117 118 /* 119 * Load the queue obj from the old location so that it 120 * can be freed by dsl_scan_done(). 121 */ 122 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 123 "scrub_queue", sizeof (uint64_t), 1, 124 &scn->scn_phys.scn_queue_obj); 125 } else { 126 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 127 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 128 &scn->scn_phys); 129 if (err == ENOENT) 130 return (0); 131 else if (err) 132 return (err); 133 134 if (scn->scn_phys.scn_state == DSS_SCANNING && 135 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) { 136 /* 137 * A new-type scrub was in progress on an old 138 * pool, and the pool was accessed by old 139 * software. Restart from the beginning, since 140 * the old software may have changed the pool in 141 * the meantime. 142 */ 143 scn->scn_restart_txg = txg; 144 zfs_dbgmsg("new-style scrub was modified " 145 "by old software; restarting in txg %llu", 146 scn->scn_restart_txg); 147 } 148 } 149 150 spa_scan_stat_init(spa); 151 return (0); 152 } 153 154 void 155 dsl_scan_fini(dsl_pool_t *dp) 156 { 157 if (dp->dp_scan) { 158 kmem_free(dp->dp_scan, sizeof (dsl_scan_t)); 159 dp->dp_scan = NULL; 160 } 161 } 162 163 /* ARGSUSED */ 164 static int 165 dsl_scan_setup_check(void *arg, dmu_tx_t *tx) 166 { 167 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 168 169 if (scn->scn_phys.scn_state == DSS_SCANNING) 170 return (SET_ERROR(EBUSY)); 171 172 return (0); 173 } 174 175 static void 176 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx) 177 { 178 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 179 pool_scan_func_t *funcp = arg; 180 dmu_object_type_t ot = 0; 181 dsl_pool_t *dp = scn->scn_dp; 182 spa_t *spa = dp->dp_spa; 183 184 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING); 185 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS); 186 bzero(&scn->scn_phys, sizeof (scn->scn_phys)); 187 scn->scn_phys.scn_func = *funcp; 188 scn->scn_phys.scn_state = DSS_SCANNING; 189 scn->scn_phys.scn_min_txg = 0; 190 scn->scn_phys.scn_max_txg = tx->tx_txg; 191 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */ 192 scn->scn_phys.scn_start_time = gethrestime_sec(); 193 scn->scn_phys.scn_errors = 0; 194 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc; 195 scn->scn_restart_txg = 0; 196 scn->scn_done_txg = 0; 197 spa_scan_stat_init(spa); 198 199 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 200 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max; 201 202 /* rewrite all disk labels */ 203 vdev_config_dirty(spa->spa_root_vdev); 204 205 if (vdev_resilver_needed(spa->spa_root_vdev, 206 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) { 207 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START); 208 } else { 209 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START); 210 } 211 212 spa->spa_scrub_started = B_TRUE; 213 /* 214 * If this is an incremental scrub, limit the DDT scrub phase 215 * to just the auto-ditto class (for correctness); the rest 216 * of the scrub should go faster using top-down pruning. 217 */ 218 if (scn->scn_phys.scn_min_txg > TXG_INITIAL) 219 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO; 220 221 } 222 223 /* back to the generic stuff */ 224 225 if (dp->dp_blkstats == NULL) { 226 dp->dp_blkstats = 227 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP); 228 } 229 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 230 231 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) 232 ot = DMU_OT_ZAP_OTHER; 233 234 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, 235 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx); 236 237 dsl_scan_sync_state(scn, tx); 238 239 spa_history_log_internal(spa, "scan setup", tx, 240 "func=%u mintxg=%llu maxtxg=%llu", 241 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg); 242 } 243 244 /* ARGSUSED */ 245 static void 246 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx) 247 { 248 static const char *old_names[] = { 249 "scrub_bookmark", 250 "scrub_ddt_bookmark", 251 "scrub_ddt_class_max", 252 "scrub_queue", 253 "scrub_min_txg", 254 "scrub_max_txg", 255 "scrub_func", 256 "scrub_errors", 257 NULL 258 }; 259 260 dsl_pool_t *dp = scn->scn_dp; 261 spa_t *spa = dp->dp_spa; 262 int i; 263 264 /* Remove any remnants of an old-style scrub. */ 265 for (i = 0; old_names[i]; i++) { 266 (void) zap_remove(dp->dp_meta_objset, 267 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx); 268 } 269 270 if (scn->scn_phys.scn_queue_obj != 0) { 271 VERIFY(0 == dmu_object_free(dp->dp_meta_objset, 272 scn->scn_phys.scn_queue_obj, tx)); 273 scn->scn_phys.scn_queue_obj = 0; 274 } 275 276 /* 277 * If we were "restarted" from a stopped state, don't bother 278 * with anything else. 279 */ 280 if (scn->scn_phys.scn_state != DSS_SCANNING) 281 return; 282 283 if (complete) 284 scn->scn_phys.scn_state = DSS_FINISHED; 285 else 286 scn->scn_phys.scn_state = DSS_CANCELED; 287 288 spa_history_log_internal(spa, "scan done", tx, 289 "complete=%u", complete); 290 291 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 292 mutex_enter(&spa->spa_scrub_lock); 293 while (spa->spa_scrub_inflight > 0) { 294 cv_wait(&spa->spa_scrub_io_cv, 295 &spa->spa_scrub_lock); 296 } 297 mutex_exit(&spa->spa_scrub_lock); 298 spa->spa_scrub_started = B_FALSE; 299 spa->spa_scrub_active = B_FALSE; 300 301 /* 302 * If the scrub/resilver completed, update all DTLs to 303 * reflect this. Whether it succeeded or not, vacate 304 * all temporary scrub DTLs. 305 */ 306 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg, 307 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE); 308 if (complete) { 309 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ? 310 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH); 311 } 312 spa_errlog_rotate(spa); 313 314 /* 315 * We may have finished replacing a device. 316 * Let the async thread assess this and handle the detach. 317 */ 318 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); 319 } 320 321 scn->scn_phys.scn_end_time = gethrestime_sec(); 322 } 323 324 /* ARGSUSED */ 325 static int 326 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx) 327 { 328 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 329 330 if (scn->scn_phys.scn_state != DSS_SCANNING) 331 return (SET_ERROR(ENOENT)); 332 return (0); 333 } 334 335 /* ARGSUSED */ 336 static void 337 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx) 338 { 339 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 340 341 dsl_scan_done(scn, B_FALSE, tx); 342 dsl_scan_sync_state(scn, tx); 343 } 344 345 int 346 dsl_scan_cancel(dsl_pool_t *dp) 347 { 348 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check, 349 dsl_scan_cancel_sync, NULL, 3)); 350 } 351 352 static void dsl_scan_visitbp(blkptr_t *bp, 353 const zbookmark_phys_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf, 354 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype, 355 dmu_tx_t *tx); 356 static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds, 357 dmu_objset_type_t ostype, 358 dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx); 359 360 void 361 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp) 362 { 363 zio_free(dp->dp_spa, txg, bp); 364 } 365 366 void 367 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp) 368 { 369 ASSERT(dsl_pool_sync_context(dp)); 370 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags)); 371 } 372 373 static uint64_t 374 dsl_scan_ds_maxtxg(dsl_dataset_t *ds) 375 { 376 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg; 377 if (dsl_dataset_is_snapshot(ds)) 378 return (MIN(smt, ds->ds_phys->ds_creation_txg)); 379 return (smt); 380 } 381 382 static void 383 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx) 384 { 385 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset, 386 DMU_POOL_DIRECTORY_OBJECT, 387 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 388 &scn->scn_phys, tx)); 389 } 390 391 static boolean_t 392 dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb) 393 { 394 uint64_t elapsed_nanosecs; 395 int mintime; 396 397 /* we never skip user/group accounting objects */ 398 if (zb && (int64_t)zb->zb_object < 0) 399 return (B_FALSE); 400 401 if (scn->scn_pausing) 402 return (B_TRUE); /* we're already pausing */ 403 404 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) 405 return (B_FALSE); /* we're resuming */ 406 407 /* We only know how to resume from level-0 blocks. */ 408 if (zb && zb->zb_level != 0) 409 return (B_FALSE); 410 411 mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ? 412 zfs_resilver_min_time_ms : zfs_scan_min_time_ms; 413 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 414 if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 415 (NSEC2MSEC(elapsed_nanosecs) > mintime && 416 txg_sync_waiting(scn->scn_dp)) || 417 spa_shutting_down(scn->scn_dp->dp_spa)) { 418 if (zb) { 419 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n", 420 (longlong_t)zb->zb_objset, 421 (longlong_t)zb->zb_object, 422 (longlong_t)zb->zb_level, 423 (longlong_t)zb->zb_blkid); 424 scn->scn_phys.scn_bookmark = *zb; 425 } 426 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n", 427 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class, 428 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type, 429 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum, 430 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor); 431 scn->scn_pausing = B_TRUE; 432 return (B_TRUE); 433 } 434 return (B_FALSE); 435 } 436 437 typedef struct zil_scan_arg { 438 dsl_pool_t *zsa_dp; 439 zil_header_t *zsa_zh; 440 } zil_scan_arg_t; 441 442 /* ARGSUSED */ 443 static int 444 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 445 { 446 zil_scan_arg_t *zsa = arg; 447 dsl_pool_t *dp = zsa->zsa_dp; 448 dsl_scan_t *scn = dp->dp_scan; 449 zil_header_t *zh = zsa->zsa_zh; 450 zbookmark_phys_t zb; 451 452 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 453 return (0); 454 455 /* 456 * One block ("stubby") can be allocated a long time ago; we 457 * want to visit that one because it has been allocated 458 * (on-disk) even if it hasn't been claimed (even though for 459 * scrub there's nothing to do to it). 460 */ 461 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa)) 462 return (0); 463 464 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 465 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 466 467 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 468 return (0); 469 } 470 471 /* ARGSUSED */ 472 static int 473 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) 474 { 475 if (lrc->lrc_txtype == TX_WRITE) { 476 zil_scan_arg_t *zsa = arg; 477 dsl_pool_t *dp = zsa->zsa_dp; 478 dsl_scan_t *scn = dp->dp_scan; 479 zil_header_t *zh = zsa->zsa_zh; 480 lr_write_t *lr = (lr_write_t *)lrc; 481 blkptr_t *bp = &lr->lr_blkptr; 482 zbookmark_phys_t zb; 483 484 if (BP_IS_HOLE(bp) || 485 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 486 return (0); 487 488 /* 489 * birth can be < claim_txg if this record's txg is 490 * already txg sync'ed (but this log block contains 491 * other records that are not synced) 492 */ 493 if (claim_txg == 0 || bp->blk_birth < claim_txg) 494 return (0); 495 496 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 497 lr->lr_foid, ZB_ZIL_LEVEL, 498 lr->lr_offset / BP_GET_LSIZE(bp)); 499 500 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 501 } 502 return (0); 503 } 504 505 static void 506 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh) 507 { 508 uint64_t claim_txg = zh->zh_claim_txg; 509 zil_scan_arg_t zsa = { dp, zh }; 510 zilog_t *zilog; 511 512 /* 513 * We only want to visit blocks that have been claimed but not yet 514 * replayed (or, in read-only mode, blocks that *would* be claimed). 515 */ 516 if (claim_txg == 0 && spa_writeable(dp->dp_spa)) 517 return; 518 519 zilog = zil_alloc(dp->dp_meta_objset, zh); 520 521 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa, 522 claim_txg); 523 524 zil_free(zilog); 525 } 526 527 /* ARGSUSED */ 528 static void 529 dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp, 530 uint64_t objset, uint64_t object, uint64_t blkid) 531 { 532 zbookmark_phys_t czb; 533 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; 534 535 if (zfs_no_scrub_prefetch) 536 return; 537 538 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg || 539 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)) 540 return; 541 542 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid); 543 544 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp, 545 NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 546 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb); 547 } 548 549 static boolean_t 550 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp, 551 const zbookmark_phys_t *zb) 552 { 553 /* 554 * We never skip over user/group accounting objects (obj<0) 555 */ 556 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) && 557 (int64_t)zb->zb_object >= 0) { 558 /* 559 * If we already visited this bp & everything below (in 560 * a prior txg sync), don't bother doing it again. 561 */ 562 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark)) 563 return (B_TRUE); 564 565 /* 566 * If we found the block we're trying to resume from, or 567 * we went past it to a different object, zero it out to 568 * indicate that it's OK to start checking for pausing 569 * again. 570 */ 571 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 || 572 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) { 573 dprintf("resuming at %llx/%llx/%llx/%llx\n", 574 (longlong_t)zb->zb_objset, 575 (longlong_t)zb->zb_object, 576 (longlong_t)zb->zb_level, 577 (longlong_t)zb->zb_blkid); 578 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb)); 579 } 580 } 581 return (B_FALSE); 582 } 583 584 /* 585 * Return nonzero on i/o error. 586 * Return new buf to write out in *bufp. 587 */ 588 static int 589 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype, 590 dnode_phys_t *dnp, const blkptr_t *bp, 591 const zbookmark_phys_t *zb, dmu_tx_t *tx, arc_buf_t **bufp) 592 { 593 dsl_pool_t *dp = scn->scn_dp; 594 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD; 595 int err; 596 597 if (BP_GET_LEVEL(bp) > 0) { 598 uint32_t flags = ARC_WAIT; 599 int i; 600 blkptr_t *cbp; 601 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; 602 603 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 604 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 605 if (err) { 606 scn->scn_phys.scn_errors++; 607 return (err); 608 } 609 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) { 610 dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset, 611 zb->zb_object, zb->zb_blkid * epb + i); 612 } 613 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) { 614 zbookmark_phys_t czb; 615 616 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, 617 zb->zb_level - 1, 618 zb->zb_blkid * epb + i); 619 dsl_scan_visitbp(cbp, &czb, dnp, 620 *bufp, ds, scn, ostype, tx); 621 } 622 } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) { 623 uint32_t flags = ARC_WAIT; 624 625 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 626 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 627 if (err) { 628 scn->scn_phys.scn_errors++; 629 return (err); 630 } 631 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { 632 uint32_t flags = ARC_WAIT; 633 dnode_phys_t *cdnp; 634 int i, j; 635 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; 636 637 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 638 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 639 if (err) { 640 scn->scn_phys.scn_errors++; 641 return (err); 642 } 643 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) { 644 for (j = 0; j < cdnp->dn_nblkptr; j++) { 645 blkptr_t *cbp = &cdnp->dn_blkptr[j]; 646 dsl_scan_prefetch(scn, *bufp, cbp, 647 zb->zb_objset, zb->zb_blkid * epb + i, j); 648 } 649 } 650 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) { 651 dsl_scan_visitdnode(scn, ds, ostype, 652 cdnp, *bufp, zb->zb_blkid * epb + i, tx); 653 } 654 655 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { 656 uint32_t flags = ARC_WAIT; 657 objset_phys_t *osp; 658 659 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 660 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 661 if (err) { 662 scn->scn_phys.scn_errors++; 663 return (err); 664 } 665 666 osp = (*bufp)->b_data; 667 668 dsl_scan_visitdnode(scn, ds, osp->os_type, 669 &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx); 670 671 if (OBJSET_BUF_HAS_USERUSED(*bufp)) { 672 /* 673 * We also always visit user/group accounting 674 * objects, and never skip them, even if we are 675 * pausing. This is necessary so that the space 676 * deltas from this txg get integrated. 677 */ 678 dsl_scan_visitdnode(scn, ds, osp->os_type, 679 &osp->os_groupused_dnode, *bufp, 680 DMU_GROUPUSED_OBJECT, tx); 681 dsl_scan_visitdnode(scn, ds, osp->os_type, 682 &osp->os_userused_dnode, *bufp, 683 DMU_USERUSED_OBJECT, tx); 684 } 685 } 686 687 return (0); 688 } 689 690 static void 691 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds, 692 dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf, 693 uint64_t object, dmu_tx_t *tx) 694 { 695 int j; 696 697 for (j = 0; j < dnp->dn_nblkptr; j++) { 698 zbookmark_phys_t czb; 699 700 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 701 dnp->dn_nlevels - 1, j); 702 dsl_scan_visitbp(&dnp->dn_blkptr[j], 703 &czb, dnp, buf, ds, scn, ostype, tx); 704 } 705 706 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 707 zbookmark_phys_t czb; 708 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 709 0, DMU_SPILL_BLKID); 710 dsl_scan_visitbp(&dnp->dn_spill, 711 &czb, dnp, buf, ds, scn, ostype, tx); 712 } 713 } 714 715 /* 716 * The arguments are in this order because mdb can only print the 717 * first 5; we want them to be useful. 718 */ 719 static void 720 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb, 721 dnode_phys_t *dnp, arc_buf_t *pbuf, 722 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype, 723 dmu_tx_t *tx) 724 { 725 dsl_pool_t *dp = scn->scn_dp; 726 arc_buf_t *buf = NULL; 727 blkptr_t bp_toread = *bp; 728 729 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */ 730 731 if (dsl_scan_check_pause(scn, zb)) 732 return; 733 734 if (dsl_scan_check_resume(scn, dnp, zb)) 735 return; 736 737 if (BP_IS_HOLE(bp)) 738 return; 739 740 scn->scn_visited_this_txg++; 741 742 dprintf_bp(bp, 743 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p", 744 ds, ds ? ds->ds_object : 0, 745 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, 746 pbuf, bp); 747 748 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 749 return; 750 751 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx, 752 &buf) != 0) 753 return; 754 755 /* 756 * If dsl_scan_ddt() has aready visited this block, it will have 757 * already done any translations or scrubbing, so don't call the 758 * callback again. 759 */ 760 if (ddt_class_contains(dp->dp_spa, 761 scn->scn_phys.scn_ddt_class_max, bp)) { 762 ASSERT(buf == NULL); 763 return; 764 } 765 766 /* 767 * If this block is from the future (after cur_max_txg), then we 768 * are doing this on behalf of a deleted snapshot, and we will 769 * revisit the future block on the next pass of this dataset. 770 * Don't scan it now unless we need to because something 771 * under it was modified. 772 */ 773 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) { 774 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb); 775 } 776 if (buf) 777 (void) arc_buf_remove_ref(buf, &buf); 778 } 779 780 static void 781 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp, 782 dmu_tx_t *tx) 783 { 784 zbookmark_phys_t zb; 785 786 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 787 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); 788 dsl_scan_visitbp(bp, &zb, NULL, NULL, 789 ds, scn, DMU_OST_NONE, tx); 790 791 dprintf_ds(ds, "finished scan%s", ""); 792 } 793 794 void 795 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx) 796 { 797 dsl_pool_t *dp = ds->ds_dir->dd_pool; 798 dsl_scan_t *scn = dp->dp_scan; 799 uint64_t mintxg; 800 801 if (scn->scn_phys.scn_state != DSS_SCANNING) 802 return; 803 804 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) { 805 if (dsl_dataset_is_snapshot(ds)) { 806 /* Note, scn_cur_{min,max}_txg stays the same. */ 807 scn->scn_phys.scn_bookmark.zb_objset = 808 ds->ds_phys->ds_next_snap_obj; 809 zfs_dbgmsg("destroying ds %llu; currently traversing; " 810 "reset zb_objset to %llu", 811 (u_longlong_t)ds->ds_object, 812 (u_longlong_t)ds->ds_phys->ds_next_snap_obj); 813 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN; 814 } else { 815 SET_BOOKMARK(&scn->scn_phys.scn_bookmark, 816 ZB_DESTROYED_OBJSET, 0, 0, 0); 817 zfs_dbgmsg("destroying ds %llu; currently traversing; " 818 "reset bookmark to -1,0,0,0", 819 (u_longlong_t)ds->ds_object); 820 } 821 } else if (zap_lookup_int_key(dp->dp_meta_objset, 822 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) { 823 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1); 824 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 825 scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 826 if (dsl_dataset_is_snapshot(ds)) { 827 /* 828 * We keep the same mintxg; it could be > 829 * ds_creation_txg if the previous snapshot was 830 * deleted too. 831 */ 832 VERIFY(zap_add_int_key(dp->dp_meta_objset, 833 scn->scn_phys.scn_queue_obj, 834 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0); 835 zfs_dbgmsg("destroying ds %llu; in queue; " 836 "replacing with %llu", 837 (u_longlong_t)ds->ds_object, 838 (u_longlong_t)ds->ds_phys->ds_next_snap_obj); 839 } else { 840 zfs_dbgmsg("destroying ds %llu; in queue; removing", 841 (u_longlong_t)ds->ds_object); 842 } 843 } else { 844 zfs_dbgmsg("destroying ds %llu; ignoring", 845 (u_longlong_t)ds->ds_object); 846 } 847 848 /* 849 * dsl_scan_sync() should be called after this, and should sync 850 * out our changed state, but just to be safe, do it here. 851 */ 852 dsl_scan_sync_state(scn, tx); 853 } 854 855 void 856 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx) 857 { 858 dsl_pool_t *dp = ds->ds_dir->dd_pool; 859 dsl_scan_t *scn = dp->dp_scan; 860 uint64_t mintxg; 861 862 if (scn->scn_phys.scn_state != DSS_SCANNING) 863 return; 864 865 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0); 866 867 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) { 868 scn->scn_phys.scn_bookmark.zb_objset = 869 ds->ds_phys->ds_prev_snap_obj; 870 zfs_dbgmsg("snapshotting ds %llu; currently traversing; " 871 "reset zb_objset to %llu", 872 (u_longlong_t)ds->ds_object, 873 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj); 874 } else if (zap_lookup_int_key(dp->dp_meta_objset, 875 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) { 876 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 877 scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 878 VERIFY(zap_add_int_key(dp->dp_meta_objset, 879 scn->scn_phys.scn_queue_obj, 880 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0); 881 zfs_dbgmsg("snapshotting ds %llu; in queue; " 882 "replacing with %llu", 883 (u_longlong_t)ds->ds_object, 884 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj); 885 } 886 dsl_scan_sync_state(scn, tx); 887 } 888 889 void 890 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx) 891 { 892 dsl_pool_t *dp = ds1->ds_dir->dd_pool; 893 dsl_scan_t *scn = dp->dp_scan; 894 uint64_t mintxg; 895 896 if (scn->scn_phys.scn_state != DSS_SCANNING) 897 return; 898 899 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) { 900 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object; 901 zfs_dbgmsg("clone_swap ds %llu; currently traversing; " 902 "reset zb_objset to %llu", 903 (u_longlong_t)ds1->ds_object, 904 (u_longlong_t)ds2->ds_object); 905 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) { 906 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object; 907 zfs_dbgmsg("clone_swap ds %llu; currently traversing; " 908 "reset zb_objset to %llu", 909 (u_longlong_t)ds2->ds_object, 910 (u_longlong_t)ds1->ds_object); 911 } 912 913 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 914 ds1->ds_object, &mintxg) == 0) { 915 int err; 916 917 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg); 918 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg); 919 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 920 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx)); 921 err = zap_add_int_key(dp->dp_meta_objset, 922 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx); 923 VERIFY(err == 0 || err == EEXIST); 924 if (err == EEXIST) { 925 /* Both were there to begin with */ 926 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset, 927 scn->scn_phys.scn_queue_obj, 928 ds1->ds_object, mintxg, tx)); 929 } 930 zfs_dbgmsg("clone_swap ds %llu; in queue; " 931 "replacing with %llu", 932 (u_longlong_t)ds1->ds_object, 933 (u_longlong_t)ds2->ds_object); 934 } else if (zap_lookup_int_key(dp->dp_meta_objset, 935 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) { 936 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg); 937 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg); 938 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 939 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx)); 940 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset, 941 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx)); 942 zfs_dbgmsg("clone_swap ds %llu; in queue; " 943 "replacing with %llu", 944 (u_longlong_t)ds2->ds_object, 945 (u_longlong_t)ds1->ds_object); 946 } 947 948 dsl_scan_sync_state(scn, tx); 949 } 950 951 struct enqueue_clones_arg { 952 dmu_tx_t *tx; 953 uint64_t originobj; 954 }; 955 956 /* ARGSUSED */ 957 static int 958 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 959 { 960 struct enqueue_clones_arg *eca = arg; 961 dsl_dataset_t *ds; 962 int err; 963 dsl_scan_t *scn = dp->dp_scan; 964 965 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj) 966 return (0); 967 968 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 969 if (err) 970 return (err); 971 972 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) { 973 dsl_dataset_t *prev; 974 err = dsl_dataset_hold_obj(dp, 975 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev); 976 977 dsl_dataset_rele(ds, FTAG); 978 if (err) 979 return (err); 980 ds = prev; 981 } 982 VERIFY(zap_add_int_key(dp->dp_meta_objset, 983 scn->scn_phys.scn_queue_obj, ds->ds_object, 984 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0); 985 dsl_dataset_rele(ds, FTAG); 986 return (0); 987 } 988 989 static void 990 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx) 991 { 992 dsl_pool_t *dp = scn->scn_dp; 993 dsl_dataset_t *ds; 994 objset_t *os; 995 996 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 997 998 if (dmu_objset_from_ds(ds, &os)) 999 goto out; 1000 1001 /* 1002 * Only the ZIL in the head (non-snapshot) is valid. Even though 1003 * snapshots can have ZIL block pointers (which may be the same 1004 * BP as in the head), they must be ignored. So we traverse the 1005 * ZIL here, rather than in scan_recurse(), because the regular 1006 * snapshot block-sharing rules don't apply to it. 1007 */ 1008 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds)) 1009 dsl_scan_zil(dp, &os->os_zil_header); 1010 1011 /* 1012 * Iterate over the bps in this ds. 1013 */ 1014 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1015 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx); 1016 1017 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP); 1018 dsl_dataset_name(ds, dsname); 1019 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; " 1020 "pausing=%u", 1021 (longlong_t)dsobj, dsname, 1022 (longlong_t)scn->scn_phys.scn_cur_min_txg, 1023 (longlong_t)scn->scn_phys.scn_cur_max_txg, 1024 (int)scn->scn_pausing); 1025 kmem_free(dsname, ZFS_MAXNAMELEN); 1026 1027 if (scn->scn_pausing) 1028 goto out; 1029 1030 /* 1031 * We've finished this pass over this dataset. 1032 */ 1033 1034 /* 1035 * If we did not completely visit this dataset, do another pass. 1036 */ 1037 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) { 1038 zfs_dbgmsg("incomplete pass; visiting again"); 1039 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN; 1040 VERIFY(zap_add_int_key(dp->dp_meta_objset, 1041 scn->scn_phys.scn_queue_obj, ds->ds_object, 1042 scn->scn_phys.scn_cur_max_txg, tx) == 0); 1043 goto out; 1044 } 1045 1046 /* 1047 * Add descendent datasets to work queue. 1048 */ 1049 if (ds->ds_phys->ds_next_snap_obj != 0) { 1050 VERIFY(zap_add_int_key(dp->dp_meta_objset, 1051 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj, 1052 ds->ds_phys->ds_creation_txg, tx) == 0); 1053 } 1054 if (ds->ds_phys->ds_num_children > 1) { 1055 boolean_t usenext = B_FALSE; 1056 if (ds->ds_phys->ds_next_clones_obj != 0) { 1057 uint64_t count; 1058 /* 1059 * A bug in a previous version of the code could 1060 * cause upgrade_clones_cb() to not set 1061 * ds_next_snap_obj when it should, leading to a 1062 * missing entry. Therefore we can only use the 1063 * next_clones_obj when its count is correct. 1064 */ 1065 int err = zap_count(dp->dp_meta_objset, 1066 ds->ds_phys->ds_next_clones_obj, &count); 1067 if (err == 0 && 1068 count == ds->ds_phys->ds_num_children - 1) 1069 usenext = B_TRUE; 1070 } 1071 1072 if (usenext) { 1073 VERIFY0(zap_join_key(dp->dp_meta_objset, 1074 ds->ds_phys->ds_next_clones_obj, 1075 scn->scn_phys.scn_queue_obj, 1076 ds->ds_phys->ds_creation_txg, tx)); 1077 } else { 1078 struct enqueue_clones_arg eca; 1079 eca.tx = tx; 1080 eca.originobj = ds->ds_object; 1081 1082 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1083 enqueue_clones_cb, &eca, DS_FIND_CHILDREN)); 1084 } 1085 } 1086 1087 out: 1088 dsl_dataset_rele(ds, FTAG); 1089 } 1090 1091 /* ARGSUSED */ 1092 static int 1093 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 1094 { 1095 dmu_tx_t *tx = arg; 1096 dsl_dataset_t *ds; 1097 int err; 1098 dsl_scan_t *scn = dp->dp_scan; 1099 1100 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 1101 if (err) 1102 return (err); 1103 1104 while (ds->ds_phys->ds_prev_snap_obj != 0) { 1105 dsl_dataset_t *prev; 1106 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 1107 FTAG, &prev); 1108 if (err) { 1109 dsl_dataset_rele(ds, FTAG); 1110 return (err); 1111 } 1112 1113 /* 1114 * If this is a clone, we don't need to worry about it for now. 1115 */ 1116 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) { 1117 dsl_dataset_rele(ds, FTAG); 1118 dsl_dataset_rele(prev, FTAG); 1119 return (0); 1120 } 1121 dsl_dataset_rele(ds, FTAG); 1122 ds = prev; 1123 } 1124 1125 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 1126 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0); 1127 dsl_dataset_rele(ds, FTAG); 1128 return (0); 1129 } 1130 1131 /* 1132 * Scrub/dedup interaction. 1133 * 1134 * If there are N references to a deduped block, we don't want to scrub it 1135 * N times -- ideally, we should scrub it exactly once. 1136 * 1137 * We leverage the fact that the dde's replication class (enum ddt_class) 1138 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest 1139 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order. 1140 * 1141 * To prevent excess scrubbing, the scrub begins by walking the DDT 1142 * to find all blocks with refcnt > 1, and scrubs each of these once. 1143 * Since there are two replication classes which contain blocks with 1144 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first. 1145 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1. 1146 * 1147 * There would be nothing more to say if a block's refcnt couldn't change 1148 * during a scrub, but of course it can so we must account for changes 1149 * in a block's replication class. 1150 * 1151 * Here's an example of what can occur: 1152 * 1153 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1 1154 * when visited during the top-down scrub phase, it will be scrubbed twice. 1155 * This negates our scrub optimization, but is otherwise harmless. 1156 * 1157 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1 1158 * on each visit during the top-down scrub phase, it will never be scrubbed. 1159 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's 1160 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to 1161 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1 1162 * while a scrub is in progress, it scrubs the block right then. 1163 */ 1164 static void 1165 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx) 1166 { 1167 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark; 1168 ddt_entry_t dde = { 0 }; 1169 int error; 1170 uint64_t n = 0; 1171 1172 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) { 1173 ddt_t *ddt; 1174 1175 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max) 1176 break; 1177 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n", 1178 (longlong_t)ddb->ddb_class, 1179 (longlong_t)ddb->ddb_type, 1180 (longlong_t)ddb->ddb_checksum, 1181 (longlong_t)ddb->ddb_cursor); 1182 1183 /* There should be no pending changes to the dedup table */ 1184 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum]; 1185 ASSERT(avl_first(&ddt->ddt_tree) == NULL); 1186 1187 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx); 1188 n++; 1189 1190 if (dsl_scan_check_pause(scn, NULL)) 1191 break; 1192 } 1193 1194 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u", 1195 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max, 1196 (int)scn->scn_pausing); 1197 1198 ASSERT(error == 0 || error == ENOENT); 1199 ASSERT(error != ENOENT || 1200 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max); 1201 } 1202 1203 /* ARGSUSED */ 1204 void 1205 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum, 1206 ddt_entry_t *dde, dmu_tx_t *tx) 1207 { 1208 const ddt_key_t *ddk = &dde->dde_key; 1209 ddt_phys_t *ddp = dde->dde_phys; 1210 blkptr_t bp; 1211 zbookmark_phys_t zb = { 0 }; 1212 1213 if (scn->scn_phys.scn_state != DSS_SCANNING) 1214 return; 1215 1216 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 1217 if (ddp->ddp_phys_birth == 0 || 1218 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg) 1219 continue; 1220 ddt_bp_create(checksum, ddk, ddp, &bp); 1221 1222 scn->scn_visited_this_txg++; 1223 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb); 1224 } 1225 } 1226 1227 static void 1228 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx) 1229 { 1230 dsl_pool_t *dp = scn->scn_dp; 1231 zap_cursor_t zc; 1232 zap_attribute_t za; 1233 1234 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 1235 scn->scn_phys.scn_ddt_class_max) { 1236 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 1237 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 1238 dsl_scan_ddt(scn, tx); 1239 if (scn->scn_pausing) 1240 return; 1241 } 1242 1243 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) { 1244 /* First do the MOS & ORIGIN */ 1245 1246 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 1247 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 1248 dsl_scan_visit_rootbp(scn, NULL, 1249 &dp->dp_meta_rootbp, tx); 1250 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 1251 if (scn->scn_pausing) 1252 return; 1253 1254 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) { 1255 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1256 enqueue_cb, tx, DS_FIND_CHILDREN)); 1257 } else { 1258 dsl_scan_visitds(scn, 1259 dp->dp_origin_snap->ds_object, tx); 1260 } 1261 ASSERT(!scn->scn_pausing); 1262 } else if (scn->scn_phys.scn_bookmark.zb_objset != 1263 ZB_DESTROYED_OBJSET) { 1264 /* 1265 * If we were paused, continue from here. Note if the 1266 * ds we were paused on was deleted, the zb_objset may 1267 * be -1, so we will skip this and find a new objset 1268 * below. 1269 */ 1270 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx); 1271 if (scn->scn_pausing) 1272 return; 1273 } 1274 1275 /* 1276 * In case we were paused right at the end of the ds, zero the 1277 * bookmark so we don't think that we're still trying to resume. 1278 */ 1279 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t)); 1280 1281 /* keep pulling things out of the zap-object-as-queue */ 1282 while (zap_cursor_init(&zc, dp->dp_meta_objset, 1283 scn->scn_phys.scn_queue_obj), 1284 zap_cursor_retrieve(&zc, &za) == 0) { 1285 dsl_dataset_t *ds; 1286 uint64_t dsobj; 1287 1288 dsobj = strtonum(za.za_name, NULL); 1289 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 1290 scn->scn_phys.scn_queue_obj, dsobj, tx)); 1291 1292 /* Set up min/max txg */ 1293 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 1294 if (za.za_first_integer != 0) { 1295 scn->scn_phys.scn_cur_min_txg = 1296 MAX(scn->scn_phys.scn_min_txg, 1297 za.za_first_integer); 1298 } else { 1299 scn->scn_phys.scn_cur_min_txg = 1300 MAX(scn->scn_phys.scn_min_txg, 1301 ds->ds_phys->ds_prev_snap_txg); 1302 } 1303 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds); 1304 dsl_dataset_rele(ds, FTAG); 1305 1306 dsl_scan_visitds(scn, dsobj, tx); 1307 zap_cursor_fini(&zc); 1308 if (scn->scn_pausing) 1309 return; 1310 } 1311 zap_cursor_fini(&zc); 1312 } 1313 1314 static boolean_t 1315 dsl_scan_free_should_pause(dsl_scan_t *scn) 1316 { 1317 uint64_t elapsed_nanosecs; 1318 1319 if (zfs_recover) 1320 return (B_FALSE); 1321 1322 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 1323 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 1324 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms && 1325 txg_sync_waiting(scn->scn_dp)) || 1326 spa_shutting_down(scn->scn_dp->dp_spa)); 1327 } 1328 1329 static int 1330 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 1331 { 1332 dsl_scan_t *scn = arg; 1333 1334 if (!scn->scn_is_bptree || 1335 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) { 1336 if (dsl_scan_free_should_pause(scn)) 1337 return (SET_ERROR(ERESTART)); 1338 } 1339 1340 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa, 1341 dmu_tx_get_txg(tx), bp, 0)); 1342 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, 1343 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp), 1344 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); 1345 scn->scn_visited_this_txg++; 1346 return (0); 1347 } 1348 1349 boolean_t 1350 dsl_scan_active(dsl_scan_t *scn) 1351 { 1352 spa_t *spa = scn->scn_dp->dp_spa; 1353 uint64_t used = 0, comp, uncomp; 1354 1355 if (spa->spa_load_state != SPA_LOAD_NONE) 1356 return (B_FALSE); 1357 if (spa_shutting_down(spa)) 1358 return (B_FALSE); 1359 if (scn->scn_phys.scn_state == DSS_SCANNING || 1360 (scn->scn_async_destroying && !scn->scn_async_stalled)) 1361 return (B_TRUE); 1362 1363 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 1364 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj, 1365 &used, &comp, &uncomp); 1366 } 1367 return (used != 0); 1368 } 1369 1370 void 1371 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx) 1372 { 1373 dsl_scan_t *scn = dp->dp_scan; 1374 spa_t *spa = dp->dp_spa; 1375 int err = 0; 1376 1377 /* 1378 * Check for scn_restart_txg before checking spa_load_state, so 1379 * that we can restart an old-style scan while the pool is being 1380 * imported (see dsl_scan_init). 1381 */ 1382 if (scn->scn_restart_txg != 0 && 1383 scn->scn_restart_txg <= tx->tx_txg) { 1384 pool_scan_func_t func = POOL_SCAN_SCRUB; 1385 dsl_scan_done(scn, B_FALSE, tx); 1386 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) 1387 func = POOL_SCAN_RESILVER; 1388 zfs_dbgmsg("restarting scan func=%u txg=%llu", 1389 func, tx->tx_txg); 1390 dsl_scan_setup_sync(&func, tx); 1391 } 1392 1393 /* 1394 * If the scan is inactive due to a stalled async destroy, try again. 1395 */ 1396 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) || 1397 spa_sync_pass(dp->dp_spa) > 1) 1398 return; 1399 1400 scn->scn_visited_this_txg = 0; 1401 scn->scn_pausing = B_FALSE; 1402 scn->scn_sync_start_time = gethrtime(); 1403 spa->spa_scrub_active = B_TRUE; 1404 1405 /* 1406 * First process the async destroys. If we pause, don't do 1407 * any scrubbing or resilvering. This ensures that there are no 1408 * async destroys while we are scanning, so the scan code doesn't 1409 * have to worry about traversing it. It is also faster to free the 1410 * blocks than to scrub them. 1411 */ 1412 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 1413 scn->scn_is_bptree = B_FALSE; 1414 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1415 NULL, ZIO_FLAG_MUSTSUCCEED); 1416 err = bpobj_iterate(&dp->dp_free_bpobj, 1417 dsl_scan_free_block_cb, scn, tx); 1418 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root)); 1419 1420 if (err != 0 && err != ERESTART) 1421 zfs_panic_recover("error %u from bpobj_iterate()", err); 1422 } 1423 1424 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) { 1425 ASSERT(scn->scn_async_destroying); 1426 scn->scn_is_bptree = B_TRUE; 1427 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1428 NULL, ZIO_FLAG_MUSTSUCCEED); 1429 err = bptree_iterate(dp->dp_meta_objset, 1430 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx); 1431 VERIFY0(zio_wait(scn->scn_zio_root)); 1432 1433 if (err == EIO || err == ECKSUM) { 1434 err = 0; 1435 } else if (err != 0 && err != ERESTART) { 1436 zfs_panic_recover("error %u from " 1437 "traverse_dataset_destroyed()", err); 1438 } 1439 1440 /* 1441 * If we didn't make progress, mark the async destroy as 1442 * stalled, so that we will not initiate a spa_sync() on 1443 * its behalf. 1444 */ 1445 scn->scn_async_stalled = (scn->scn_visited_this_txg == 0); 1446 1447 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) { 1448 /* finished; deactivate async destroy feature */ 1449 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx); 1450 ASSERT(!spa_feature_is_active(spa, 1451 SPA_FEATURE_ASYNC_DESTROY)); 1452 VERIFY0(zap_remove(dp->dp_meta_objset, 1453 DMU_POOL_DIRECTORY_OBJECT, 1454 DMU_POOL_BPTREE_OBJ, tx)); 1455 VERIFY0(bptree_free(dp->dp_meta_objset, 1456 dp->dp_bptree_obj, tx)); 1457 dp->dp_bptree_obj = 0; 1458 scn->scn_async_destroying = B_FALSE; 1459 } 1460 } 1461 if (scn->scn_visited_this_txg) { 1462 zfs_dbgmsg("freed %llu blocks in %llums from " 1463 "free_bpobj/bptree txg %llu; err=%u", 1464 (longlong_t)scn->scn_visited_this_txg, 1465 (longlong_t) 1466 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time), 1467 (longlong_t)tx->tx_txg, err); 1468 scn->scn_visited_this_txg = 0; 1469 1470 /* 1471 * Write out changes to the DDT that may be required as a 1472 * result of the blocks freed. This ensures that the DDT 1473 * is clean when a scrub/resilver runs. 1474 */ 1475 ddt_sync(spa, tx->tx_txg); 1476 } 1477 if (err != 0) 1478 return; 1479 if (!scn->scn_async_destroying && zfs_free_leak_on_eio && 1480 (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 || 1481 dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 || 1482 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) { 1483 /* 1484 * We have finished background destroying, but there is still 1485 * some space left in the dp_free_dir. Transfer this leaked 1486 * space to the dp_leak_dir. 1487 */ 1488 if (dp->dp_leak_dir == NULL) { 1489 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 1490 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 1491 LEAK_DIR_NAME, tx); 1492 VERIFY0(dsl_pool_open_special_dir(dp, 1493 LEAK_DIR_NAME, &dp->dp_leak_dir)); 1494 rrw_exit(&dp->dp_config_rwlock, FTAG); 1495 } 1496 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD, 1497 dp->dp_free_dir->dd_phys->dd_used_bytes, 1498 dp->dp_free_dir->dd_phys->dd_compressed_bytes, 1499 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx); 1500 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD, 1501 -dp->dp_free_dir->dd_phys->dd_used_bytes, 1502 -dp->dp_free_dir->dd_phys->dd_compressed_bytes, 1503 -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx); 1504 } 1505 if (!scn->scn_async_destroying) { 1506 /* finished; verify that space accounting went to zero */ 1507 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes); 1508 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes); 1509 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes); 1510 } 1511 1512 if (scn->scn_phys.scn_state != DSS_SCANNING) 1513 return; 1514 1515 if (scn->scn_done_txg == tx->tx_txg) { 1516 ASSERT(!scn->scn_pausing); 1517 /* finished with scan. */ 1518 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg); 1519 dsl_scan_done(scn, B_TRUE, tx); 1520 ASSERT3U(spa->spa_scrub_inflight, ==, 0); 1521 dsl_scan_sync_state(scn, tx); 1522 return; 1523 } 1524 1525 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 1526 scn->scn_phys.scn_ddt_class_max) { 1527 zfs_dbgmsg("doing scan sync txg %llu; " 1528 "ddt bm=%llu/%llu/%llu/%llx", 1529 (longlong_t)tx->tx_txg, 1530 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class, 1531 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type, 1532 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum, 1533 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor); 1534 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0); 1535 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0); 1536 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0); 1537 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0); 1538 } else { 1539 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu", 1540 (longlong_t)tx->tx_txg, 1541 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset, 1542 (longlong_t)scn->scn_phys.scn_bookmark.zb_object, 1543 (longlong_t)scn->scn_phys.scn_bookmark.zb_level, 1544 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid); 1545 } 1546 1547 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1548 NULL, ZIO_FLAG_CANFAIL); 1549 dsl_pool_config_enter(dp, FTAG); 1550 dsl_scan_visit(scn, tx); 1551 dsl_pool_config_exit(dp, FTAG); 1552 (void) zio_wait(scn->scn_zio_root); 1553 scn->scn_zio_root = NULL; 1554 1555 zfs_dbgmsg("visited %llu blocks in %llums", 1556 (longlong_t)scn->scn_visited_this_txg, 1557 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time)); 1558 1559 if (!scn->scn_pausing) { 1560 scn->scn_done_txg = tx->tx_txg + 1; 1561 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu", 1562 tx->tx_txg, scn->scn_done_txg); 1563 } 1564 1565 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 1566 mutex_enter(&spa->spa_scrub_lock); 1567 while (spa->spa_scrub_inflight > 0) { 1568 cv_wait(&spa->spa_scrub_io_cv, 1569 &spa->spa_scrub_lock); 1570 } 1571 mutex_exit(&spa->spa_scrub_lock); 1572 } 1573 1574 dsl_scan_sync_state(scn, tx); 1575 } 1576 1577 /* 1578 * This will start a new scan, or restart an existing one. 1579 */ 1580 void 1581 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg) 1582 { 1583 if (txg == 0) { 1584 dmu_tx_t *tx; 1585 tx = dmu_tx_create_dd(dp->dp_mos_dir); 1586 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT)); 1587 1588 txg = dmu_tx_get_txg(tx); 1589 dp->dp_scan->scn_restart_txg = txg; 1590 dmu_tx_commit(tx); 1591 } else { 1592 dp->dp_scan->scn_restart_txg = txg; 1593 } 1594 zfs_dbgmsg("restarting resilver txg=%llu", txg); 1595 } 1596 1597 boolean_t 1598 dsl_scan_resilvering(dsl_pool_t *dp) 1599 { 1600 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING && 1601 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER); 1602 } 1603 1604 /* 1605 * scrub consumers 1606 */ 1607 1608 static void 1609 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp) 1610 { 1611 int i; 1612 1613 /* 1614 * If we resume after a reboot, zab will be NULL; don't record 1615 * incomplete stats in that case. 1616 */ 1617 if (zab == NULL) 1618 return; 1619 1620 for (i = 0; i < 4; i++) { 1621 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS; 1622 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL; 1623 if (t & DMU_OT_NEWTYPE) 1624 t = DMU_OT_OTHER; 1625 zfs_blkstat_t *zb = &zab->zab_type[l][t]; 1626 int equal; 1627 1628 zb->zb_count++; 1629 zb->zb_asize += BP_GET_ASIZE(bp); 1630 zb->zb_lsize += BP_GET_LSIZE(bp); 1631 zb->zb_psize += BP_GET_PSIZE(bp); 1632 zb->zb_gangs += BP_COUNT_GANG(bp); 1633 1634 switch (BP_GET_NDVAS(bp)) { 1635 case 2: 1636 if (DVA_GET_VDEV(&bp->blk_dva[0]) == 1637 DVA_GET_VDEV(&bp->blk_dva[1])) 1638 zb->zb_ditto_2_of_2_samevdev++; 1639 break; 1640 case 3: 1641 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) == 1642 DVA_GET_VDEV(&bp->blk_dva[1])) + 1643 (DVA_GET_VDEV(&bp->blk_dva[0]) == 1644 DVA_GET_VDEV(&bp->blk_dva[2])) + 1645 (DVA_GET_VDEV(&bp->blk_dva[1]) == 1646 DVA_GET_VDEV(&bp->blk_dva[2])); 1647 if (equal == 1) 1648 zb->zb_ditto_2_of_3_samevdev++; 1649 else if (equal == 3) 1650 zb->zb_ditto_3_of_3_samevdev++; 1651 break; 1652 } 1653 } 1654 } 1655 1656 static void 1657 dsl_scan_scrub_done(zio_t *zio) 1658 { 1659 spa_t *spa = zio->io_spa; 1660 1661 zio_data_buf_free(zio->io_data, zio->io_size); 1662 1663 mutex_enter(&spa->spa_scrub_lock); 1664 spa->spa_scrub_inflight--; 1665 cv_broadcast(&spa->spa_scrub_io_cv); 1666 1667 if (zio->io_error && (zio->io_error != ECKSUM || 1668 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) { 1669 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++; 1670 } 1671 mutex_exit(&spa->spa_scrub_lock); 1672 } 1673 1674 static int 1675 dsl_scan_scrub_cb(dsl_pool_t *dp, 1676 const blkptr_t *bp, const zbookmark_phys_t *zb) 1677 { 1678 dsl_scan_t *scn = dp->dp_scan; 1679 size_t size = BP_GET_PSIZE(bp); 1680 spa_t *spa = dp->dp_spa; 1681 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp); 1682 boolean_t needs_io; 1683 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL; 1684 int scan_delay = 0; 1685 1686 if (phys_birth <= scn->scn_phys.scn_min_txg || 1687 phys_birth >= scn->scn_phys.scn_max_txg) 1688 return (0); 1689 1690 count_block(dp->dp_blkstats, bp); 1691 1692 if (BP_IS_EMBEDDED(bp)) 1693 return (0); 1694 1695 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn)); 1696 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) { 1697 zio_flags |= ZIO_FLAG_SCRUB; 1698 needs_io = B_TRUE; 1699 scan_delay = zfs_scrub_delay; 1700 } else { 1701 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER); 1702 zio_flags |= ZIO_FLAG_RESILVER; 1703 needs_io = B_FALSE; 1704 scan_delay = zfs_resilver_delay; 1705 } 1706 1707 /* If it's an intent log block, failure is expected. */ 1708 if (zb->zb_level == ZB_ZIL_LEVEL) 1709 zio_flags |= ZIO_FLAG_SPECULATIVE; 1710 1711 for (int d = 0; d < BP_GET_NDVAS(bp); d++) { 1712 vdev_t *vd = vdev_lookup_top(spa, 1713 DVA_GET_VDEV(&bp->blk_dva[d])); 1714 1715 /* 1716 * Keep track of how much data we've examined so that 1717 * zpool(1M) status can make useful progress reports. 1718 */ 1719 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]); 1720 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]); 1721 1722 /* if it's a resilver, this may not be in the target range */ 1723 if (!needs_io) { 1724 if (DVA_GET_GANG(&bp->blk_dva[d])) { 1725 /* 1726 * Gang members may be spread across multiple 1727 * vdevs, so the best estimate we have is the 1728 * scrub range, which has already been checked. 1729 * XXX -- it would be better to change our 1730 * allocation policy to ensure that all 1731 * gang members reside on the same vdev. 1732 */ 1733 needs_io = B_TRUE; 1734 } else { 1735 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL, 1736 phys_birth, 1); 1737 } 1738 } 1739 } 1740 1741 if (needs_io && !zfs_no_scrub_io) { 1742 vdev_t *rvd = spa->spa_root_vdev; 1743 uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight; 1744 void *data = zio_data_buf_alloc(size); 1745 1746 mutex_enter(&spa->spa_scrub_lock); 1747 while (spa->spa_scrub_inflight >= maxinflight) 1748 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 1749 spa->spa_scrub_inflight++; 1750 mutex_exit(&spa->spa_scrub_lock); 1751 1752 /* 1753 * If we're seeing recent (zfs_scan_idle) "important" I/Os 1754 * then throttle our workload to limit the impact of a scan. 1755 */ 1756 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle) 1757 delay(scan_delay); 1758 1759 zio_nowait(zio_read(NULL, spa, bp, data, size, 1760 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB, 1761 zio_flags, zb)); 1762 } 1763 1764 /* do not relocate this block */ 1765 return (0); 1766 } 1767 1768 int 1769 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func) 1770 { 1771 spa_t *spa = dp->dp_spa; 1772 1773 /* 1774 * Purge all vdev caches and probe all devices. We do this here 1775 * rather than in sync context because this requires a writer lock 1776 * on the spa_config lock, which we can't do from sync context. The 1777 * spa_scrub_reopen flag indicates that vdev_open() should not 1778 * attempt to start another scrub. 1779 */ 1780 spa_vdev_state_enter(spa, SCL_NONE); 1781 spa->spa_scrub_reopen = B_TRUE; 1782 vdev_reopen(spa->spa_root_vdev); 1783 spa->spa_scrub_reopen = B_FALSE; 1784 (void) spa_vdev_state_exit(spa, NULL, 0); 1785 1786 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check, 1787 dsl_scan_setup_sync, &func, 0)); 1788 } 1789