dsl_scan.c revision 268650
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) 2013 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 53typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *); 54 55static scan_cb_t dsl_scan_scrub_cb; 56static void dsl_scan_cancel_sync(void *, dmu_tx_t *); 57static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx); 58 59unsigned int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */ 60unsigned int zfs_resilver_delay = 2; /* number of ticks to delay resilver */ 61unsigned int zfs_scrub_delay = 4; /* number of ticks to delay scrub */ 62unsigned int zfs_scan_idle = 50; /* idle window in clock ticks */ 63 64unsigned int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */ 65unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */ 66unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver 67 per txg */ 68boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */ 69boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */ 70 71SYSCTL_DECL(_vfs_zfs); 72TUNABLE_INT("vfs.zfs.top_maxinflight", &zfs_top_maxinflight); 73SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RW, 74 &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev"); 75TUNABLE_INT("vfs.zfs.resilver_delay", &zfs_resilver_delay); 76SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RW, 77 &zfs_resilver_delay, 0, "Number of ticks to delay resilver"); 78TUNABLE_INT("vfs.zfs.scrub_delay", &zfs_scrub_delay); 79SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RW, 80 &zfs_scrub_delay, 0, "Number of ticks to delay scrub"); 81TUNABLE_INT("vfs.zfs.scan_idle", &zfs_scan_idle); 82SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RW, 83 &zfs_scan_idle, 0, "Idle scan window in clock ticks"); 84TUNABLE_INT("vfs.zfs.scan_min_time_ms", &zfs_scan_min_time_ms); 85SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RW, 86 &zfs_scan_min_time_ms, 0, "Min millisecs to scrub per txg"); 87TUNABLE_INT("vfs.zfs.free_min_time_ms", &zfs_free_min_time_ms); 88SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RW, 89 &zfs_free_min_time_ms, 0, "Min millisecs to free per txg"); 90TUNABLE_INT("vfs.zfs.resilver_min_time_ms", &zfs_resilver_min_time_ms); 91SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RW, 92 &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg"); 93TUNABLE_INT("vfs.zfs.no_scrub_io", &zfs_no_scrub_io); 94SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RW, 95 &zfs_no_scrub_io, 0, "Disable scrub I/O"); 96TUNABLE_INT("vfs.zfs.no_scrub_prefetch", &zfs_no_scrub_prefetch); 97SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RW, 98 &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching"); 99 100enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE; 101 102#define DSL_SCAN_IS_SCRUB_RESILVER(scn) \ 103 ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \ 104 (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER) 105 106extern int zfs_txg_timeout; 107 108/* the order has to match pool_scan_type */ 109static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = { 110 NULL, 111 dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */ 112 dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */ 113}; 114 115int 116dsl_scan_init(dsl_pool_t *dp, uint64_t txg) 117{ 118 int err; 119 dsl_scan_t *scn; 120 spa_t *spa = dp->dp_spa; 121 uint64_t f; 122 123 scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP); 124 scn->scn_dp = dp; 125 126 /* 127 * It's possible that we're resuming a scan after a reboot so 128 * make sure that the scan_async_destroying flag is initialized 129 * appropriately. 130 */ 131 ASSERT(!scn->scn_async_destroying); 132 scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa, 133 SPA_FEATURE_ASYNC_DESTROY); 134 135 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 136 "scrub_func", sizeof (uint64_t), 1, &f); 137 if (err == 0) { 138 /* 139 * There was an old-style scrub in progress. Restart a 140 * new-style scrub from the beginning. 141 */ 142 scn->scn_restart_txg = txg; 143 zfs_dbgmsg("old-style scrub was in progress; " 144 "restarting new-style scrub in txg %llu", 145 scn->scn_restart_txg); 146 147 /* 148 * Load the queue obj from the old location so that it 149 * can be freed by dsl_scan_done(). 150 */ 151 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 152 "scrub_queue", sizeof (uint64_t), 1, 153 &scn->scn_phys.scn_queue_obj); 154 } else { 155 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 156 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 157 &scn->scn_phys); 158 if (err == ENOENT) 159 return (0); 160 else if (err) 161 return (err); 162 163 if (scn->scn_phys.scn_state == DSS_SCANNING && 164 spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) { 165 /* 166 * A new-type scrub was in progress on an old 167 * pool, and the pool was accessed by old 168 * software. Restart from the beginning, since 169 * the old software may have changed the pool in 170 * the meantime. 171 */ 172 scn->scn_restart_txg = txg; 173 zfs_dbgmsg("new-style scrub was modified " 174 "by old software; restarting in txg %llu", 175 scn->scn_restart_txg); 176 } 177 } 178 179 spa_scan_stat_init(spa); 180 return (0); 181} 182 183void 184dsl_scan_fini(dsl_pool_t *dp) 185{ 186 if (dp->dp_scan) { 187 kmem_free(dp->dp_scan, sizeof (dsl_scan_t)); 188 dp->dp_scan = NULL; 189 } 190} 191 192/* ARGSUSED */ 193static int 194dsl_scan_setup_check(void *arg, dmu_tx_t *tx) 195{ 196 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 197 198 if (scn->scn_phys.scn_state == DSS_SCANNING) 199 return (SET_ERROR(EBUSY)); 200 201 return (0); 202} 203 204static void 205dsl_scan_setup_sync(void *arg, dmu_tx_t *tx) 206{ 207 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 208 pool_scan_func_t *funcp = arg; 209 dmu_object_type_t ot = 0; 210 dsl_pool_t *dp = scn->scn_dp; 211 spa_t *spa = dp->dp_spa; 212 213 ASSERT(scn->scn_phys.scn_state != DSS_SCANNING); 214 ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS); 215 bzero(&scn->scn_phys, sizeof (scn->scn_phys)); 216 scn->scn_phys.scn_func = *funcp; 217 scn->scn_phys.scn_state = DSS_SCANNING; 218 scn->scn_phys.scn_min_txg = 0; 219 scn->scn_phys.scn_max_txg = tx->tx_txg; 220 scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */ 221 scn->scn_phys.scn_start_time = gethrestime_sec(); 222 scn->scn_phys.scn_errors = 0; 223 scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc; 224 scn->scn_restart_txg = 0; 225 scn->scn_done_txg = 0; 226 spa_scan_stat_init(spa); 227 228 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 229 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max; 230 231 /* rewrite all disk labels */ 232 vdev_config_dirty(spa->spa_root_vdev); 233 234 if (vdev_resilver_needed(spa->spa_root_vdev, 235 &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) { 236 spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START); 237 } else { 238 spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START); 239 } 240 241 spa->spa_scrub_started = B_TRUE; 242 /* 243 * If this is an incremental scrub, limit the DDT scrub phase 244 * to just the auto-ditto class (for correctness); the rest 245 * of the scrub should go faster using top-down pruning. 246 */ 247 if (scn->scn_phys.scn_min_txg > TXG_INITIAL) 248 scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO; 249 250 } 251 252 /* back to the generic stuff */ 253 254 if (dp->dp_blkstats == NULL) { 255 dp->dp_blkstats = 256 kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP); 257 } 258 bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 259 260 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) 261 ot = DMU_OT_ZAP_OTHER; 262 263 scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, 264 ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx); 265 266 dsl_scan_sync_state(scn, tx); 267 268 spa_history_log_internal(spa, "scan setup", tx, 269 "func=%u mintxg=%llu maxtxg=%llu", 270 *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg); 271} 272 273/* ARGSUSED */ 274static void 275dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx) 276{ 277 static const char *old_names[] = { 278 "scrub_bookmark", 279 "scrub_ddt_bookmark", 280 "scrub_ddt_class_max", 281 "scrub_queue", 282 "scrub_min_txg", 283 "scrub_max_txg", 284 "scrub_func", 285 "scrub_errors", 286 NULL 287 }; 288 289 dsl_pool_t *dp = scn->scn_dp; 290 spa_t *spa = dp->dp_spa; 291 int i; 292 293 /* Remove any remnants of an old-style scrub. */ 294 for (i = 0; old_names[i]; i++) { 295 (void) zap_remove(dp->dp_meta_objset, 296 DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx); 297 } 298 299 if (scn->scn_phys.scn_queue_obj != 0) { 300 VERIFY(0 == dmu_object_free(dp->dp_meta_objset, 301 scn->scn_phys.scn_queue_obj, tx)); 302 scn->scn_phys.scn_queue_obj = 0; 303 } 304 305 /* 306 * If we were "restarted" from a stopped state, don't bother 307 * with anything else. 308 */ 309 if (scn->scn_phys.scn_state != DSS_SCANNING) 310 return; 311 312 if (complete) 313 scn->scn_phys.scn_state = DSS_FINISHED; 314 else 315 scn->scn_phys.scn_state = DSS_CANCELED; 316 317 spa_history_log_internal(spa, "scan done", tx, 318 "complete=%u", complete); 319 320 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 321 mutex_enter(&spa->spa_scrub_lock); 322 while (spa->spa_scrub_inflight > 0) { 323 cv_wait(&spa->spa_scrub_io_cv, 324 &spa->spa_scrub_lock); 325 } 326 mutex_exit(&spa->spa_scrub_lock); 327 spa->spa_scrub_started = B_FALSE; 328 spa->spa_scrub_active = B_FALSE; 329 330 /* 331 * If the scrub/resilver completed, update all DTLs to 332 * reflect this. Whether it succeeded or not, vacate 333 * all temporary scrub DTLs. 334 */ 335 vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg, 336 complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE); 337 if (complete) { 338 spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ? 339 ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH); 340 } 341 spa_errlog_rotate(spa); 342 343 /* 344 * We may have finished replacing a device. 345 * Let the async thread assess this and handle the detach. 346 */ 347 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); 348 } 349 350 scn->scn_phys.scn_end_time = gethrestime_sec(); 351} 352 353/* ARGSUSED */ 354static int 355dsl_scan_cancel_check(void *arg, dmu_tx_t *tx) 356{ 357 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 358 359 if (scn->scn_phys.scn_state != DSS_SCANNING) 360 return (SET_ERROR(ENOENT)); 361 return (0); 362} 363 364/* ARGSUSED */ 365static void 366dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx) 367{ 368 dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 369 370 dsl_scan_done(scn, B_FALSE, tx); 371 dsl_scan_sync_state(scn, tx); 372} 373 374int 375dsl_scan_cancel(dsl_pool_t *dp) 376{ 377 return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check, 378 dsl_scan_cancel_sync, NULL, 3)); 379} 380 381static void dsl_scan_visitbp(blkptr_t *bp, 382 const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf, 383 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype, 384 dmu_tx_t *tx); 385static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds, 386 dmu_objset_type_t ostype, 387 dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx); 388 389void 390dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp) 391{ 392 zio_free(dp->dp_spa, txg, bp); 393} 394 395void 396dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp) 397{ 398 ASSERT(dsl_pool_sync_context(dp)); 399 zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp), 400 pio->io_flags)); 401} 402 403static uint64_t 404dsl_scan_ds_maxtxg(dsl_dataset_t *ds) 405{ 406 uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg; 407 if (dsl_dataset_is_snapshot(ds)) 408 return (MIN(smt, ds->ds_phys->ds_creation_txg)); 409 return (smt); 410} 411 412static void 413dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx) 414{ 415 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset, 416 DMU_POOL_DIRECTORY_OBJECT, 417 DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 418 &scn->scn_phys, tx)); 419} 420 421static boolean_t 422dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb) 423{ 424 uint64_t elapsed_nanosecs; 425 unsigned int mintime; 426 427 /* we never skip user/group accounting objects */ 428 if (zb && (int64_t)zb->zb_object < 0) 429 return (B_FALSE); 430 431 if (scn->scn_pausing) 432 return (B_TRUE); /* we're already pausing */ 433 434 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) 435 return (B_FALSE); /* we're resuming */ 436 437 /* We only know how to resume from level-0 blocks. */ 438 if (zb && zb->zb_level != 0) 439 return (B_FALSE); 440 441 mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ? 442 zfs_resilver_min_time_ms : zfs_scan_min_time_ms; 443 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 444 if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 445 (NSEC2MSEC(elapsed_nanosecs) > mintime && 446 txg_sync_waiting(scn->scn_dp)) || 447 spa_shutting_down(scn->scn_dp->dp_spa)) { 448 if (zb) { 449 dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n", 450 (longlong_t)zb->zb_objset, 451 (longlong_t)zb->zb_object, 452 (longlong_t)zb->zb_level, 453 (longlong_t)zb->zb_blkid); 454 scn->scn_phys.scn_bookmark = *zb; 455 } 456 dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n", 457 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class, 458 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type, 459 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum, 460 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor); 461 scn->scn_pausing = B_TRUE; 462 return (B_TRUE); 463 } 464 return (B_FALSE); 465} 466 467typedef struct zil_scan_arg { 468 dsl_pool_t *zsa_dp; 469 zil_header_t *zsa_zh; 470} zil_scan_arg_t; 471 472/* ARGSUSED */ 473static int 474dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 475{ 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 zbookmark_t zb; 481 482 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 483 return (0); 484 485 /* 486 * One block ("stubby") can be allocated a long time ago; we 487 * want to visit that one because it has been allocated 488 * (on-disk) even if it hasn't been claimed (even though for 489 * scrub there's nothing to do to it). 490 */ 491 if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa)) 492 return (0); 493 494 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 495 ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 496 497 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 498 return (0); 499} 500 501/* ARGSUSED */ 502static int 503dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) 504{ 505 if (lrc->lrc_txtype == TX_WRITE) { 506 zil_scan_arg_t *zsa = arg; 507 dsl_pool_t *dp = zsa->zsa_dp; 508 dsl_scan_t *scn = dp->dp_scan; 509 zil_header_t *zh = zsa->zsa_zh; 510 lr_write_t *lr = (lr_write_t *)lrc; 511 blkptr_t *bp = &lr->lr_blkptr; 512 zbookmark_t zb; 513 514 if (BP_IS_HOLE(bp) || 515 bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 516 return (0); 517 518 /* 519 * birth can be < claim_txg if this record's txg is 520 * already txg sync'ed (but this log block contains 521 * other records that are not synced) 522 */ 523 if (claim_txg == 0 || bp->blk_birth < claim_txg) 524 return (0); 525 526 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 527 lr->lr_foid, ZB_ZIL_LEVEL, 528 lr->lr_offset / BP_GET_LSIZE(bp)); 529 530 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 531 } 532 return (0); 533} 534 535static void 536dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh) 537{ 538 uint64_t claim_txg = zh->zh_claim_txg; 539 zil_scan_arg_t zsa = { dp, zh }; 540 zilog_t *zilog; 541 542 /* 543 * We only want to visit blocks that have been claimed but not yet 544 * replayed (or, in read-only mode, blocks that *would* be claimed). 545 */ 546 if (claim_txg == 0 && spa_writeable(dp->dp_spa)) 547 return; 548 549 zilog = zil_alloc(dp->dp_meta_objset, zh); 550 551 (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa, 552 claim_txg); 553 554 zil_free(zilog); 555} 556 557/* ARGSUSED */ 558static void 559dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp, 560 uint64_t objset, uint64_t object, uint64_t blkid) 561{ 562 zbookmark_t czb; 563 uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; 564 565 if (zfs_no_scrub_prefetch) 566 return; 567 568 if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg || 569 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)) 570 return; 571 572 SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid); 573 574 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp, 575 NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 576 ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb); 577} 578 579static boolean_t 580dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp, 581 const zbookmark_t *zb) 582{ 583 /* 584 * We never skip over user/group accounting objects (obj<0) 585 */ 586 if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) && 587 (int64_t)zb->zb_object >= 0) { 588 /* 589 * If we already visited this bp & everything below (in 590 * a prior txg sync), don't bother doing it again. 591 */ 592 if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark)) 593 return (B_TRUE); 594 595 /* 596 * If we found the block we're trying to resume from, or 597 * we went past it to a different object, zero it out to 598 * indicate that it's OK to start checking for pausing 599 * again. 600 */ 601 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 || 602 zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) { 603 dprintf("resuming at %llx/%llx/%llx/%llx\n", 604 (longlong_t)zb->zb_objset, 605 (longlong_t)zb->zb_object, 606 (longlong_t)zb->zb_level, 607 (longlong_t)zb->zb_blkid); 608 bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb)); 609 } 610 } 611 return (B_FALSE); 612} 613 614/* 615 * Return nonzero on i/o error. 616 * Return new buf to write out in *bufp. 617 */ 618static int 619dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype, 620 dnode_phys_t *dnp, const blkptr_t *bp, 621 const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp) 622{ 623 dsl_pool_t *dp = scn->scn_dp; 624 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD; 625 int err; 626 627 if (BP_GET_LEVEL(bp) > 0) { 628 uint32_t flags = ARC_WAIT; 629 int i; 630 blkptr_t *cbp; 631 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; 632 633 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 634 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 635 if (err) { 636 scn->scn_phys.scn_errors++; 637 return (err); 638 } 639 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) { 640 dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset, 641 zb->zb_object, zb->zb_blkid * epb + i); 642 } 643 for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) { 644 zbookmark_t czb; 645 646 SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, 647 zb->zb_level - 1, 648 zb->zb_blkid * epb + i); 649 dsl_scan_visitbp(cbp, &czb, dnp, 650 *bufp, ds, scn, ostype, tx); 651 } 652 } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) { 653 uint32_t flags = ARC_WAIT; 654 655 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 656 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 657 if (err) { 658 scn->scn_phys.scn_errors++; 659 return (err); 660 } 661 } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { 662 uint32_t flags = ARC_WAIT; 663 dnode_phys_t *cdnp; 664 int i, j; 665 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; 666 667 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 668 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 669 if (err) { 670 scn->scn_phys.scn_errors++; 671 return (err); 672 } 673 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) { 674 for (j = 0; j < cdnp->dn_nblkptr; j++) { 675 blkptr_t *cbp = &cdnp->dn_blkptr[j]; 676 dsl_scan_prefetch(scn, *bufp, cbp, 677 zb->zb_objset, zb->zb_blkid * epb + i, j); 678 } 679 } 680 for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) { 681 dsl_scan_visitdnode(scn, ds, ostype, 682 cdnp, *bufp, zb->zb_blkid * epb + i, tx); 683 } 684 685 } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { 686 uint32_t flags = ARC_WAIT; 687 objset_phys_t *osp; 688 689 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 690 ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 691 if (err) { 692 scn->scn_phys.scn_errors++; 693 return (err); 694 } 695 696 osp = (*bufp)->b_data; 697 698 dsl_scan_visitdnode(scn, ds, osp->os_type, 699 &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx); 700 701 if (OBJSET_BUF_HAS_USERUSED(*bufp)) { 702 /* 703 * We also always visit user/group accounting 704 * objects, and never skip them, even if we are 705 * pausing. This is necessary so that the space 706 * deltas from this txg get integrated. 707 */ 708 dsl_scan_visitdnode(scn, ds, osp->os_type, 709 &osp->os_groupused_dnode, *bufp, 710 DMU_GROUPUSED_OBJECT, tx); 711 dsl_scan_visitdnode(scn, ds, osp->os_type, 712 &osp->os_userused_dnode, *bufp, 713 DMU_USERUSED_OBJECT, tx); 714 } 715 } 716 717 return (0); 718} 719 720static void 721dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds, 722 dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf, 723 uint64_t object, dmu_tx_t *tx) 724{ 725 int j; 726 727 for (j = 0; j < dnp->dn_nblkptr; j++) { 728 zbookmark_t czb; 729 730 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 731 dnp->dn_nlevels - 1, j); 732 dsl_scan_visitbp(&dnp->dn_blkptr[j], 733 &czb, dnp, buf, ds, scn, ostype, tx); 734 } 735 736 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 737 zbookmark_t czb; 738 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 739 0, DMU_SPILL_BLKID); 740 dsl_scan_visitbp(&dnp->dn_spill, 741 &czb, dnp, buf, ds, scn, ostype, tx); 742 } 743} 744 745/* 746 * The arguments are in this order because mdb can only print the 747 * first 5; we want them to be useful. 748 */ 749static void 750dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb, 751 dnode_phys_t *dnp, arc_buf_t *pbuf, 752 dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype, 753 dmu_tx_t *tx) 754{ 755 dsl_pool_t *dp = scn->scn_dp; 756 arc_buf_t *buf = NULL; 757 blkptr_t bp_toread = *bp; 758 759 /* ASSERT(pbuf == NULL || arc_released(pbuf)); */ 760 761 if (dsl_scan_check_pause(scn, zb)) 762 return; 763 764 if (dsl_scan_check_resume(scn, dnp, zb)) 765 return; 766 767 if (BP_IS_HOLE(bp)) 768 return; 769 770 scn->scn_visited_this_txg++; 771 772 dprintf_bp(bp, 773 "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p", 774 ds, ds ? ds->ds_object : 0, 775 zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, 776 pbuf, bp); 777 778 if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 779 return; 780 781 if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx, 782 &buf) != 0) 783 return; 784 785 /* 786 * If dsl_scan_ddt() has aready visited this block, it will have 787 * already done any translations or scrubbing, so don't call the 788 * callback again. 789 */ 790 if (ddt_class_contains(dp->dp_spa, 791 scn->scn_phys.scn_ddt_class_max, bp)) { 792 ASSERT(buf == NULL); 793 return; 794 } 795 796 /* 797 * If this block is from the future (after cur_max_txg), then we 798 * are doing this on behalf of a deleted snapshot, and we will 799 * revisit the future block on the next pass of this dataset. 800 * Don't scan it now unless we need to because something 801 * under it was modified. 802 */ 803 if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) { 804 scan_funcs[scn->scn_phys.scn_func](dp, bp, zb); 805 } 806 if (buf) 807 (void) arc_buf_remove_ref(buf, &buf); 808} 809 810static void 811dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp, 812 dmu_tx_t *tx) 813{ 814 zbookmark_t zb; 815 816 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 817 ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); 818 dsl_scan_visitbp(bp, &zb, NULL, NULL, 819 ds, scn, DMU_OST_NONE, tx); 820 821 dprintf_ds(ds, "finished scan%s", ""); 822} 823 824void 825dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx) 826{ 827 dsl_pool_t *dp = ds->ds_dir->dd_pool; 828 dsl_scan_t *scn = dp->dp_scan; 829 uint64_t mintxg; 830 831 if (scn->scn_phys.scn_state != DSS_SCANNING) 832 return; 833 834 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) { 835 if (dsl_dataset_is_snapshot(ds)) { 836 /* Note, scn_cur_{min,max}_txg stays the same. */ 837 scn->scn_phys.scn_bookmark.zb_objset = 838 ds->ds_phys->ds_next_snap_obj; 839 zfs_dbgmsg("destroying ds %llu; currently traversing; " 840 "reset zb_objset to %llu", 841 (u_longlong_t)ds->ds_object, 842 (u_longlong_t)ds->ds_phys->ds_next_snap_obj); 843 scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN; 844 } else { 845 SET_BOOKMARK(&scn->scn_phys.scn_bookmark, 846 ZB_DESTROYED_OBJSET, 0, 0, 0); 847 zfs_dbgmsg("destroying ds %llu; currently traversing; " 848 "reset bookmark to -1,0,0,0", 849 (u_longlong_t)ds->ds_object); 850 } 851 } else if (zap_lookup_int_key(dp->dp_meta_objset, 852 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) { 853 ASSERT3U(ds->ds_phys->ds_num_children, <=, 1); 854 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 855 scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 856 if (dsl_dataset_is_snapshot(ds)) { 857 /* 858 * We keep the same mintxg; it could be > 859 * ds_creation_txg if the previous snapshot was 860 * deleted too. 861 */ 862 VERIFY(zap_add_int_key(dp->dp_meta_objset, 863 scn->scn_phys.scn_queue_obj, 864 ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0); 865 zfs_dbgmsg("destroying ds %llu; in queue; " 866 "replacing with %llu", 867 (u_longlong_t)ds->ds_object, 868 (u_longlong_t)ds->ds_phys->ds_next_snap_obj); 869 } else { 870 zfs_dbgmsg("destroying ds %llu; in queue; removing", 871 (u_longlong_t)ds->ds_object); 872 } 873 } else { 874 zfs_dbgmsg("destroying ds %llu; ignoring", 875 (u_longlong_t)ds->ds_object); 876 } 877 878 /* 879 * dsl_scan_sync() should be called after this, and should sync 880 * out our changed state, but just to be safe, do it here. 881 */ 882 dsl_scan_sync_state(scn, tx); 883} 884 885void 886dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx) 887{ 888 dsl_pool_t *dp = ds->ds_dir->dd_pool; 889 dsl_scan_t *scn = dp->dp_scan; 890 uint64_t mintxg; 891 892 if (scn->scn_phys.scn_state != DSS_SCANNING) 893 return; 894 895 ASSERT(ds->ds_phys->ds_prev_snap_obj != 0); 896 897 if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) { 898 scn->scn_phys.scn_bookmark.zb_objset = 899 ds->ds_phys->ds_prev_snap_obj; 900 zfs_dbgmsg("snapshotting ds %llu; currently traversing; " 901 "reset zb_objset to %llu", 902 (u_longlong_t)ds->ds_object, 903 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj); 904 } else if (zap_lookup_int_key(dp->dp_meta_objset, 905 scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) { 906 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 907 scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 908 VERIFY(zap_add_int_key(dp->dp_meta_objset, 909 scn->scn_phys.scn_queue_obj, 910 ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0); 911 zfs_dbgmsg("snapshotting ds %llu; in queue; " 912 "replacing with %llu", 913 (u_longlong_t)ds->ds_object, 914 (u_longlong_t)ds->ds_phys->ds_prev_snap_obj); 915 } 916 dsl_scan_sync_state(scn, tx); 917} 918 919void 920dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx) 921{ 922 dsl_pool_t *dp = ds1->ds_dir->dd_pool; 923 dsl_scan_t *scn = dp->dp_scan; 924 uint64_t mintxg; 925 926 if (scn->scn_phys.scn_state != DSS_SCANNING) 927 return; 928 929 if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) { 930 scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object; 931 zfs_dbgmsg("clone_swap ds %llu; currently traversing; " 932 "reset zb_objset to %llu", 933 (u_longlong_t)ds1->ds_object, 934 (u_longlong_t)ds2->ds_object); 935 } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) { 936 scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object; 937 zfs_dbgmsg("clone_swap ds %llu; currently traversing; " 938 "reset zb_objset to %llu", 939 (u_longlong_t)ds2->ds_object, 940 (u_longlong_t)ds1->ds_object); 941 } 942 943 if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 944 ds1->ds_object, &mintxg) == 0) { 945 int err; 946 947 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg); 948 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg); 949 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 950 scn->scn_phys.scn_queue_obj, ds1->ds_object, tx)); 951 err = zap_add_int_key(dp->dp_meta_objset, 952 scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx); 953 VERIFY(err == 0 || err == EEXIST); 954 if (err == EEXIST) { 955 /* Both were there to begin with */ 956 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset, 957 scn->scn_phys.scn_queue_obj, 958 ds1->ds_object, mintxg, tx)); 959 } 960 zfs_dbgmsg("clone_swap ds %llu; in queue; " 961 "replacing with %llu", 962 (u_longlong_t)ds1->ds_object, 963 (u_longlong_t)ds2->ds_object); 964 } else if (zap_lookup_int_key(dp->dp_meta_objset, 965 scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) { 966 ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg); 967 ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg); 968 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 969 scn->scn_phys.scn_queue_obj, ds2->ds_object, tx)); 970 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset, 971 scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx)); 972 zfs_dbgmsg("clone_swap ds %llu; in queue; " 973 "replacing with %llu", 974 (u_longlong_t)ds2->ds_object, 975 (u_longlong_t)ds1->ds_object); 976 } 977 978 dsl_scan_sync_state(scn, tx); 979} 980 981struct enqueue_clones_arg { 982 dmu_tx_t *tx; 983 uint64_t originobj; 984}; 985 986/* ARGSUSED */ 987static int 988enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 989{ 990 struct enqueue_clones_arg *eca = arg; 991 dsl_dataset_t *ds; 992 int err; 993 dsl_scan_t *scn = dp->dp_scan; 994 995 if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj) 996 return (0); 997 998 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 999 if (err) 1000 return (err); 1001 1002 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) { 1003 dsl_dataset_t *prev; 1004 err = dsl_dataset_hold_obj(dp, 1005 ds->ds_phys->ds_prev_snap_obj, FTAG, &prev); 1006 1007 dsl_dataset_rele(ds, FTAG); 1008 if (err) 1009 return (err); 1010 ds = prev; 1011 } 1012 VERIFY(zap_add_int_key(dp->dp_meta_objset, 1013 scn->scn_phys.scn_queue_obj, ds->ds_object, 1014 ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0); 1015 dsl_dataset_rele(ds, FTAG); 1016 return (0); 1017} 1018 1019static void 1020dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx) 1021{ 1022 dsl_pool_t *dp = scn->scn_dp; 1023 dsl_dataset_t *ds; 1024 objset_t *os; 1025 1026 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 1027 1028 if (dmu_objset_from_ds(ds, &os)) 1029 goto out; 1030 1031 /* 1032 * Only the ZIL in the head (non-snapshot) is valid. Even though 1033 * snapshots can have ZIL block pointers (which may be the same 1034 * BP as in the head), they must be ignored. So we traverse the 1035 * ZIL here, rather than in scan_recurse(), because the regular 1036 * snapshot block-sharing rules don't apply to it. 1037 */ 1038 if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds)) 1039 dsl_scan_zil(dp, &os->os_zil_header); 1040 1041 /* 1042 * Iterate over the bps in this ds. 1043 */ 1044 dmu_buf_will_dirty(ds->ds_dbuf, tx); 1045 dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx); 1046 1047 char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP); 1048 dsl_dataset_name(ds, dsname); 1049 zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; " 1050 "pausing=%u", 1051 (longlong_t)dsobj, dsname, 1052 (longlong_t)scn->scn_phys.scn_cur_min_txg, 1053 (longlong_t)scn->scn_phys.scn_cur_max_txg, 1054 (int)scn->scn_pausing); 1055 kmem_free(dsname, ZFS_MAXNAMELEN); 1056 1057 if (scn->scn_pausing) 1058 goto out; 1059 1060 /* 1061 * We've finished this pass over this dataset. 1062 */ 1063 1064 /* 1065 * If we did not completely visit this dataset, do another pass. 1066 */ 1067 if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) { 1068 zfs_dbgmsg("incomplete pass; visiting again"); 1069 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN; 1070 VERIFY(zap_add_int_key(dp->dp_meta_objset, 1071 scn->scn_phys.scn_queue_obj, ds->ds_object, 1072 scn->scn_phys.scn_cur_max_txg, tx) == 0); 1073 goto out; 1074 } 1075 1076 /* 1077 * Add descendent datasets to work queue. 1078 */ 1079 if (ds->ds_phys->ds_next_snap_obj != 0) { 1080 VERIFY(zap_add_int_key(dp->dp_meta_objset, 1081 scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj, 1082 ds->ds_phys->ds_creation_txg, tx) == 0); 1083 } 1084 if (ds->ds_phys->ds_num_children > 1) { 1085 boolean_t usenext = B_FALSE; 1086 if (ds->ds_phys->ds_next_clones_obj != 0) { 1087 uint64_t count; 1088 /* 1089 * A bug in a previous version of the code could 1090 * cause upgrade_clones_cb() to not set 1091 * ds_next_snap_obj when it should, leading to a 1092 * missing entry. Therefore we can only use the 1093 * next_clones_obj when its count is correct. 1094 */ 1095 int err = zap_count(dp->dp_meta_objset, 1096 ds->ds_phys->ds_next_clones_obj, &count); 1097 if (err == 0 && 1098 count == ds->ds_phys->ds_num_children - 1) 1099 usenext = B_TRUE; 1100 } 1101 1102 if (usenext) { 1103 VERIFY0(zap_join_key(dp->dp_meta_objset, 1104 ds->ds_phys->ds_next_clones_obj, 1105 scn->scn_phys.scn_queue_obj, 1106 ds->ds_phys->ds_creation_txg, tx)); 1107 } else { 1108 struct enqueue_clones_arg eca; 1109 eca.tx = tx; 1110 eca.originobj = ds->ds_object; 1111 1112 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1113 enqueue_clones_cb, &eca, DS_FIND_CHILDREN)); 1114 } 1115 } 1116 1117out: 1118 dsl_dataset_rele(ds, FTAG); 1119} 1120 1121/* ARGSUSED */ 1122static int 1123enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 1124{ 1125 dmu_tx_t *tx = arg; 1126 dsl_dataset_t *ds; 1127 int err; 1128 dsl_scan_t *scn = dp->dp_scan; 1129 1130 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 1131 if (err) 1132 return (err); 1133 1134 while (ds->ds_phys->ds_prev_snap_obj != 0) { 1135 dsl_dataset_t *prev; 1136 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 1137 FTAG, &prev); 1138 if (err) { 1139 dsl_dataset_rele(ds, FTAG); 1140 return (err); 1141 } 1142 1143 /* 1144 * If this is a clone, we don't need to worry about it for now. 1145 */ 1146 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) { 1147 dsl_dataset_rele(ds, FTAG); 1148 dsl_dataset_rele(prev, FTAG); 1149 return (0); 1150 } 1151 dsl_dataset_rele(ds, FTAG); 1152 ds = prev; 1153 } 1154 1155 VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 1156 ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0); 1157 dsl_dataset_rele(ds, FTAG); 1158 return (0); 1159} 1160 1161/* 1162 * Scrub/dedup interaction. 1163 * 1164 * If there are N references to a deduped block, we don't want to scrub it 1165 * N times -- ideally, we should scrub it exactly once. 1166 * 1167 * We leverage the fact that the dde's replication class (enum ddt_class) 1168 * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest 1169 * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order. 1170 * 1171 * To prevent excess scrubbing, the scrub begins by walking the DDT 1172 * to find all blocks with refcnt > 1, and scrubs each of these once. 1173 * Since there are two replication classes which contain blocks with 1174 * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first. 1175 * Finally the top-down scrub begins, only visiting blocks with refcnt == 1. 1176 * 1177 * There would be nothing more to say if a block's refcnt couldn't change 1178 * during a scrub, but of course it can so we must account for changes 1179 * in a block's replication class. 1180 * 1181 * Here's an example of what can occur: 1182 * 1183 * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1 1184 * when visited during the top-down scrub phase, it will be scrubbed twice. 1185 * This negates our scrub optimization, but is otherwise harmless. 1186 * 1187 * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1 1188 * on each visit during the top-down scrub phase, it will never be scrubbed. 1189 * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's 1190 * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to 1191 * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1 1192 * while a scrub is in progress, it scrubs the block right then. 1193 */ 1194static void 1195dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx) 1196{ 1197 ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark; 1198 ddt_entry_t dde = { 0 }; 1199 int error; 1200 uint64_t n = 0; 1201 1202 while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) { 1203 ddt_t *ddt; 1204 1205 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max) 1206 break; 1207 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n", 1208 (longlong_t)ddb->ddb_class, 1209 (longlong_t)ddb->ddb_type, 1210 (longlong_t)ddb->ddb_checksum, 1211 (longlong_t)ddb->ddb_cursor); 1212 1213 /* There should be no pending changes to the dedup table */ 1214 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum]; 1215 ASSERT(avl_first(&ddt->ddt_tree) == NULL); 1216 1217 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx); 1218 n++; 1219 1220 if (dsl_scan_check_pause(scn, NULL)) 1221 break; 1222 } 1223 1224 zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u", 1225 (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max, 1226 (int)scn->scn_pausing); 1227 1228 ASSERT(error == 0 || error == ENOENT); 1229 ASSERT(error != ENOENT || 1230 ddb->ddb_class > scn->scn_phys.scn_ddt_class_max); 1231} 1232 1233/* ARGSUSED */ 1234void 1235dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum, 1236 ddt_entry_t *dde, dmu_tx_t *tx) 1237{ 1238 const ddt_key_t *ddk = &dde->dde_key; 1239 ddt_phys_t *ddp = dde->dde_phys; 1240 blkptr_t bp; 1241 zbookmark_t zb = { 0 }; 1242 1243 if (scn->scn_phys.scn_state != DSS_SCANNING) 1244 return; 1245 1246 for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 1247 if (ddp->ddp_phys_birth == 0 || 1248 ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg) 1249 continue; 1250 ddt_bp_create(checksum, ddk, ddp, &bp); 1251 1252 scn->scn_visited_this_txg++; 1253 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb); 1254 } 1255} 1256 1257static void 1258dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx) 1259{ 1260 dsl_pool_t *dp = scn->scn_dp; 1261 zap_cursor_t zc; 1262 zap_attribute_t za; 1263 1264 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 1265 scn->scn_phys.scn_ddt_class_max) { 1266 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 1267 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 1268 dsl_scan_ddt(scn, tx); 1269 if (scn->scn_pausing) 1270 return; 1271 } 1272 1273 if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) { 1274 /* First do the MOS & ORIGIN */ 1275 1276 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 1277 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 1278 dsl_scan_visit_rootbp(scn, NULL, 1279 &dp->dp_meta_rootbp, tx); 1280 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 1281 if (scn->scn_pausing) 1282 return; 1283 1284 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) { 1285 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 1286 enqueue_cb, tx, DS_FIND_CHILDREN)); 1287 } else { 1288 dsl_scan_visitds(scn, 1289 dp->dp_origin_snap->ds_object, tx); 1290 } 1291 ASSERT(!scn->scn_pausing); 1292 } else if (scn->scn_phys.scn_bookmark.zb_objset != 1293 ZB_DESTROYED_OBJSET) { 1294 /* 1295 * If we were paused, continue from here. Note if the 1296 * ds we were paused on was deleted, the zb_objset may 1297 * be -1, so we will skip this and find a new objset 1298 * below. 1299 */ 1300 dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx); 1301 if (scn->scn_pausing) 1302 return; 1303 } 1304 1305 /* 1306 * In case we were paused right at the end of the ds, zero the 1307 * bookmark so we don't think that we're still trying to resume. 1308 */ 1309 bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t)); 1310 1311 /* keep pulling things out of the zap-object-as-queue */ 1312 while (zap_cursor_init(&zc, dp->dp_meta_objset, 1313 scn->scn_phys.scn_queue_obj), 1314 zap_cursor_retrieve(&zc, &za) == 0) { 1315 dsl_dataset_t *ds; 1316 uint64_t dsobj; 1317 1318 dsobj = strtonum(za.za_name, NULL); 1319 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 1320 scn->scn_phys.scn_queue_obj, dsobj, tx)); 1321 1322 /* Set up min/max txg */ 1323 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 1324 if (za.za_first_integer != 0) { 1325 scn->scn_phys.scn_cur_min_txg = 1326 MAX(scn->scn_phys.scn_min_txg, 1327 za.za_first_integer); 1328 } else { 1329 scn->scn_phys.scn_cur_min_txg = 1330 MAX(scn->scn_phys.scn_min_txg, 1331 ds->ds_phys->ds_prev_snap_txg); 1332 } 1333 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds); 1334 dsl_dataset_rele(ds, FTAG); 1335 1336 dsl_scan_visitds(scn, dsobj, tx); 1337 zap_cursor_fini(&zc); 1338 if (scn->scn_pausing) 1339 return; 1340 } 1341 zap_cursor_fini(&zc); 1342} 1343 1344static boolean_t 1345dsl_scan_free_should_pause(dsl_scan_t *scn) 1346{ 1347 uint64_t elapsed_nanosecs; 1348 1349 if (zfs_recover) 1350 return (B_FALSE); 1351 1352 elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 1353 return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 1354 (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms && 1355 txg_sync_waiting(scn->scn_dp)) || 1356 spa_shutting_down(scn->scn_dp->dp_spa)); 1357} 1358 1359static int 1360dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 1361{ 1362 dsl_scan_t *scn = arg; 1363 1364 if (!scn->scn_is_bptree || 1365 (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) { 1366 if (dsl_scan_free_should_pause(scn)) 1367 return (SET_ERROR(ERESTART)); 1368 } 1369 1370 zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa, 1371 dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0)); 1372 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, 1373 -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp), 1374 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); 1375 scn->scn_visited_this_txg++; 1376 return (0); 1377} 1378 1379boolean_t 1380dsl_scan_active(dsl_scan_t *scn) 1381{ 1382 spa_t *spa = scn->scn_dp->dp_spa; 1383 uint64_t used = 0, comp, uncomp; 1384 1385 if (spa->spa_load_state != SPA_LOAD_NONE) 1386 return (B_FALSE); 1387 if (spa_shutting_down(spa)) 1388 return (B_FALSE); 1389 if (scn->scn_phys.scn_state == DSS_SCANNING || 1390 (scn->scn_async_destroying && !scn->scn_async_stalled)) 1391 return (B_TRUE); 1392 1393 if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 1394 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj, 1395 &used, &comp, &uncomp); 1396 } 1397 return (used != 0); 1398} 1399 1400void 1401dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx) 1402{ 1403 dsl_scan_t *scn = dp->dp_scan; 1404 spa_t *spa = dp->dp_spa; 1405 int err = 0; 1406 1407 /* 1408 * Check for scn_restart_txg before checking spa_load_state, so 1409 * that we can restart an old-style scan while the pool is being 1410 * imported (see dsl_scan_init). 1411 */ 1412 if (scn->scn_restart_txg != 0 && 1413 scn->scn_restart_txg <= tx->tx_txg) { 1414 pool_scan_func_t func = POOL_SCAN_SCRUB; 1415 dsl_scan_done(scn, B_FALSE, tx); 1416 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) 1417 func = POOL_SCAN_RESILVER; 1418 zfs_dbgmsg("restarting scan func=%u txg=%llu", 1419 func, tx->tx_txg); 1420 dsl_scan_setup_sync(&func, tx); 1421 } 1422 1423 /* 1424 * If the scan is inactive due to a stalled async destroy, try again. 1425 */ 1426 if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) || 1427 spa_sync_pass(dp->dp_spa) > 1) 1428 return; 1429 1430 scn->scn_visited_this_txg = 0; 1431 scn->scn_pausing = B_FALSE; 1432 scn->scn_sync_start_time = gethrtime(); 1433 spa->spa_scrub_active = B_TRUE; 1434 1435 /* 1436 * First process the async destroys. If we pause, don't do 1437 * any scrubbing or resilvering. This ensures that there are no 1438 * async destroys while we are scanning, so the scan code doesn't 1439 * have to worry about traversing it. It is also faster to free the 1440 * blocks than to scrub them. 1441 */ 1442 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 1443 scn->scn_is_bptree = B_FALSE; 1444 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1445 NULL, ZIO_FLAG_MUSTSUCCEED); 1446 err = bpobj_iterate(&dp->dp_free_bpobj, 1447 dsl_scan_free_block_cb, scn, tx); 1448 VERIFY3U(0, ==, zio_wait(scn->scn_zio_root)); 1449 1450 if (err != 0 && err != ERESTART) 1451 zfs_panic_recover("error %u from bpobj_iterate()", err); 1452 } 1453 1454 if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) { 1455 ASSERT(scn->scn_async_destroying); 1456 scn->scn_is_bptree = B_TRUE; 1457 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1458 NULL, ZIO_FLAG_MUSTSUCCEED); 1459 err = bptree_iterate(dp->dp_meta_objset, 1460 dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx); 1461 VERIFY0(zio_wait(scn->scn_zio_root)); 1462 1463 if (err == EIO || err == ECKSUM) { 1464 err = 0; 1465 } else if (err != 0 && err != ERESTART) { 1466 zfs_panic_recover("error %u from " 1467 "traverse_dataset_destroyed()", err); 1468 } 1469 1470 /* 1471 * If we didn't make progress, mark the async destroy as 1472 * stalled, so that we will not initiate a spa_sync() on 1473 * its behalf. 1474 */ 1475 scn->scn_async_stalled = (scn->scn_visited_this_txg == 0); 1476 1477 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) { 1478 /* finished; deactivate async destroy feature */ 1479 spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx); 1480 ASSERT(!spa_feature_is_active(spa, 1481 SPA_FEATURE_ASYNC_DESTROY)); 1482 VERIFY0(zap_remove(dp->dp_meta_objset, 1483 DMU_POOL_DIRECTORY_OBJECT, 1484 DMU_POOL_BPTREE_OBJ, tx)); 1485 VERIFY0(bptree_free(dp->dp_meta_objset, 1486 dp->dp_bptree_obj, tx)); 1487 dp->dp_bptree_obj = 0; 1488 scn->scn_async_destroying = B_FALSE; 1489 } 1490 } 1491 if (scn->scn_visited_this_txg) { 1492 zfs_dbgmsg("freed %llu blocks in %llums from " 1493 "free_bpobj/bptree txg %llu; err=%u", 1494 (longlong_t)scn->scn_visited_this_txg, 1495 (longlong_t) 1496 NSEC2MSEC(gethrtime() - scn->scn_sync_start_time), 1497 (longlong_t)tx->tx_txg, err); 1498 scn->scn_visited_this_txg = 0; 1499 1500 /* 1501 * Write out changes to the DDT that may be required as a 1502 * result of the blocks freed. This ensures that the DDT 1503 * is clean when a scrub/resilver runs. 1504 */ 1505 ddt_sync(spa, tx->tx_txg); 1506 } 1507 if (err != 0) 1508 return; 1509 if (!scn->scn_async_destroying && zfs_free_leak_on_eio && 1510 (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 || 1511 dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 || 1512 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) { 1513 /* 1514 * We have finished background destroying, but there is still 1515 * some space left in the dp_free_dir. Transfer this leaked 1516 * space to the dp_leak_dir. 1517 */ 1518 if (dp->dp_leak_dir == NULL) { 1519 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 1520 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, 1521 LEAK_DIR_NAME, tx); 1522 VERIFY0(dsl_pool_open_special_dir(dp, 1523 LEAK_DIR_NAME, &dp->dp_leak_dir)); 1524 rrw_exit(&dp->dp_config_rwlock, FTAG); 1525 } 1526 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD, 1527 dp->dp_free_dir->dd_phys->dd_used_bytes, 1528 dp->dp_free_dir->dd_phys->dd_compressed_bytes, 1529 dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx); 1530 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD, 1531 -dp->dp_free_dir->dd_phys->dd_used_bytes, 1532 -dp->dp_free_dir->dd_phys->dd_compressed_bytes, 1533 -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx); 1534 } 1535 if (!scn->scn_async_destroying) { 1536 /* finished; verify that space accounting went to zero */ 1537 ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes); 1538 ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes); 1539 ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes); 1540 } 1541 1542 if (scn->scn_phys.scn_state != DSS_SCANNING) 1543 return; 1544 1545 if (scn->scn_done_txg == tx->tx_txg) { 1546 ASSERT(!scn->scn_pausing); 1547 /* finished with scan. */ 1548 zfs_dbgmsg("txg %llu scan complete", tx->tx_txg); 1549 dsl_scan_done(scn, B_TRUE, tx); 1550 ASSERT3U(spa->spa_scrub_inflight, ==, 0); 1551 dsl_scan_sync_state(scn, tx); 1552 return; 1553 } 1554 1555 if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 1556 scn->scn_phys.scn_ddt_class_max) { 1557 zfs_dbgmsg("doing scan sync txg %llu; " 1558 "ddt bm=%llu/%llu/%llu/%llx", 1559 (longlong_t)tx->tx_txg, 1560 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class, 1561 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type, 1562 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum, 1563 (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor); 1564 ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0); 1565 ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0); 1566 ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0); 1567 ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0); 1568 } else { 1569 zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu", 1570 (longlong_t)tx->tx_txg, 1571 (longlong_t)scn->scn_phys.scn_bookmark.zb_objset, 1572 (longlong_t)scn->scn_phys.scn_bookmark.zb_object, 1573 (longlong_t)scn->scn_phys.scn_bookmark.zb_level, 1574 (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid); 1575 } 1576 1577 scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1578 NULL, ZIO_FLAG_CANFAIL); 1579 dsl_pool_config_enter(dp, FTAG); 1580 dsl_scan_visit(scn, tx); 1581 dsl_pool_config_exit(dp, FTAG); 1582 (void) zio_wait(scn->scn_zio_root); 1583 scn->scn_zio_root = NULL; 1584 1585 zfs_dbgmsg("visited %llu blocks in %llums", 1586 (longlong_t)scn->scn_visited_this_txg, 1587 (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time)); 1588 1589 if (!scn->scn_pausing) { 1590 scn->scn_done_txg = tx->tx_txg + 1; 1591 zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu", 1592 tx->tx_txg, scn->scn_done_txg); 1593 } 1594 1595 if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 1596 mutex_enter(&spa->spa_scrub_lock); 1597 while (spa->spa_scrub_inflight > 0) { 1598 cv_wait(&spa->spa_scrub_io_cv, 1599 &spa->spa_scrub_lock); 1600 } 1601 mutex_exit(&spa->spa_scrub_lock); 1602 } 1603 1604 dsl_scan_sync_state(scn, tx); 1605} 1606 1607/* 1608 * This will start a new scan, or restart an existing one. 1609 */ 1610void 1611dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg) 1612{ 1613 if (txg == 0) { 1614 dmu_tx_t *tx; 1615 tx = dmu_tx_create_dd(dp->dp_mos_dir); 1616 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT)); 1617 1618 txg = dmu_tx_get_txg(tx); 1619 dp->dp_scan->scn_restart_txg = txg; 1620 dmu_tx_commit(tx); 1621 } else { 1622 dp->dp_scan->scn_restart_txg = txg; 1623 } 1624 zfs_dbgmsg("restarting resilver txg=%llu", txg); 1625} 1626 1627boolean_t 1628dsl_scan_resilvering(dsl_pool_t *dp) 1629{ 1630 return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING && 1631 dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER); 1632} 1633 1634/* 1635 * scrub consumers 1636 */ 1637 1638static void 1639count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp) 1640{ 1641 int i; 1642 1643 /* 1644 * If we resume after a reboot, zab will be NULL; don't record 1645 * incomplete stats in that case. 1646 */ 1647 if (zab == NULL) 1648 return; 1649 1650 for (i = 0; i < 4; i++) { 1651 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS; 1652 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL; 1653 if (t & DMU_OT_NEWTYPE) 1654 t = DMU_OT_OTHER; 1655 zfs_blkstat_t *zb = &zab->zab_type[l][t]; 1656 int equal; 1657 1658 zb->zb_count++; 1659 zb->zb_asize += BP_GET_ASIZE(bp); 1660 zb->zb_lsize += BP_GET_LSIZE(bp); 1661 zb->zb_psize += BP_GET_PSIZE(bp); 1662 zb->zb_gangs += BP_COUNT_GANG(bp); 1663 1664 switch (BP_GET_NDVAS(bp)) { 1665 case 2: 1666 if (DVA_GET_VDEV(&bp->blk_dva[0]) == 1667 DVA_GET_VDEV(&bp->blk_dva[1])) 1668 zb->zb_ditto_2_of_2_samevdev++; 1669 break; 1670 case 3: 1671 equal = (DVA_GET_VDEV(&bp->blk_dva[0]) == 1672 DVA_GET_VDEV(&bp->blk_dva[1])) + 1673 (DVA_GET_VDEV(&bp->blk_dva[0]) == 1674 DVA_GET_VDEV(&bp->blk_dva[2])) + 1675 (DVA_GET_VDEV(&bp->blk_dva[1]) == 1676 DVA_GET_VDEV(&bp->blk_dva[2])); 1677 if (equal == 1) 1678 zb->zb_ditto_2_of_3_samevdev++; 1679 else if (equal == 3) 1680 zb->zb_ditto_3_of_3_samevdev++; 1681 break; 1682 } 1683 } 1684} 1685 1686static void 1687dsl_scan_scrub_done(zio_t *zio) 1688{ 1689 spa_t *spa = zio->io_spa; 1690 1691 zio_data_buf_free(zio->io_data, zio->io_size); 1692 1693 mutex_enter(&spa->spa_scrub_lock); 1694 spa->spa_scrub_inflight--; 1695 cv_broadcast(&spa->spa_scrub_io_cv); 1696 1697 if (zio->io_error && (zio->io_error != ECKSUM || 1698 !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) { 1699 spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++; 1700 } 1701 mutex_exit(&spa->spa_scrub_lock); 1702} 1703 1704static int 1705dsl_scan_scrub_cb(dsl_pool_t *dp, 1706 const blkptr_t *bp, const zbookmark_t *zb) 1707{ 1708 dsl_scan_t *scn = dp->dp_scan; 1709 size_t size = BP_GET_PSIZE(bp); 1710 spa_t *spa = dp->dp_spa; 1711 uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp); 1712 boolean_t needs_io; 1713 int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL; 1714 unsigned int scan_delay = 0; 1715 1716 if (phys_birth <= scn->scn_phys.scn_min_txg || 1717 phys_birth >= scn->scn_phys.scn_max_txg) 1718 return (0); 1719 1720 count_block(dp->dp_blkstats, bp); 1721 1722 if (BP_IS_EMBEDDED(bp)) 1723 return (0); 1724 1725 ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn)); 1726 if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) { 1727 zio_flags |= ZIO_FLAG_SCRUB; 1728 needs_io = B_TRUE; 1729 scan_delay = zfs_scrub_delay; 1730 } else { 1731 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER); 1732 zio_flags |= ZIO_FLAG_RESILVER; 1733 needs_io = B_FALSE; 1734 scan_delay = zfs_resilver_delay; 1735 } 1736 1737 /* If it's an intent log block, failure is expected. */ 1738 if (zb->zb_level == ZB_ZIL_LEVEL) 1739 zio_flags |= ZIO_FLAG_SPECULATIVE; 1740 1741 for (int d = 0; d < BP_GET_NDVAS(bp); d++) { 1742 vdev_t *vd = vdev_lookup_top(spa, 1743 DVA_GET_VDEV(&bp->blk_dva[d])); 1744 1745 /* 1746 * Keep track of how much data we've examined so that 1747 * zpool(1M) status can make useful progress reports. 1748 */ 1749 scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]); 1750 spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]); 1751 1752 /* if it's a resilver, this may not be in the target range */ 1753 if (!needs_io) { 1754 if (DVA_GET_GANG(&bp->blk_dva[d])) { 1755 /* 1756 * Gang members may be spread across multiple 1757 * vdevs, so the best estimate we have is the 1758 * scrub range, which has already been checked. 1759 * XXX -- it would be better to change our 1760 * allocation policy to ensure that all 1761 * gang members reside on the same vdev. 1762 */ 1763 needs_io = B_TRUE; 1764 } else { 1765 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL, 1766 phys_birth, 1); 1767 } 1768 } 1769 } 1770 1771 if (needs_io && !zfs_no_scrub_io) { 1772 vdev_t *rvd = spa->spa_root_vdev; 1773 uint64_t maxinflight = rvd->vdev_children * 1774 MAX(zfs_top_maxinflight, 1); 1775 void *data = zio_data_buf_alloc(size); 1776 1777 mutex_enter(&spa->spa_scrub_lock); 1778 while (spa->spa_scrub_inflight >= maxinflight) 1779 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 1780 spa->spa_scrub_inflight++; 1781 mutex_exit(&spa->spa_scrub_lock); 1782 1783 /* 1784 * If we're seeing recent (zfs_scan_idle) "important" I/Os 1785 * then throttle our workload to limit the impact of a scan. 1786 */ 1787 if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle) 1788 delay(MAX((int)scan_delay, 0)); 1789 1790 zio_nowait(zio_read(NULL, spa, bp, data, size, 1791 dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB, 1792 zio_flags, zb)); 1793 } 1794 1795 /* do not relocate this block */ 1796 return (0); 1797} 1798 1799int 1800dsl_scan(dsl_pool_t *dp, pool_scan_func_t func) 1801{ 1802 spa_t *spa = dp->dp_spa; 1803 1804 /* 1805 * Purge all vdev caches and probe all devices. We do this here 1806 * rather than in sync context because this requires a writer lock 1807 * on the spa_config lock, which we can't do from sync context. The 1808 * spa_scrub_reopen flag indicates that vdev_open() should not 1809 * attempt to start another scrub. 1810 */ 1811 spa_vdev_state_enter(spa, SCL_NONE); 1812 spa->spa_scrub_reopen = B_TRUE; 1813 vdev_reopen(spa->spa_root_vdev); 1814 spa->spa_scrub_reopen = B_FALSE; 1815 (void) spa_vdev_state_exit(spa, NULL, 0); 1816 1817 return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check, 1818 dsl_scan_setup_sync, &func, 0)); 1819} 1820