13f9d6ad7SLin Ling /* 23f9d6ad7SLin Ling * CDDL HEADER START 33f9d6ad7SLin Ling * 43f9d6ad7SLin Ling * The contents of this file are subject to the terms of the 53f9d6ad7SLin Ling * Common Development and Distribution License (the "License"). 63f9d6ad7SLin Ling * You may not use this file except in compliance with the License. 73f9d6ad7SLin Ling * 83f9d6ad7SLin Ling * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 93f9d6ad7SLin Ling * or http://www.opensolaris.org/os/licensing. 103f9d6ad7SLin Ling * See the License for the specific language governing permissions 113f9d6ad7SLin Ling * and limitations under the License. 123f9d6ad7SLin Ling * 133f9d6ad7SLin Ling * When distributing Covered Code, include this CDDL HEADER in each 143f9d6ad7SLin Ling * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 153f9d6ad7SLin Ling * If applicable, add the following below this CDDL HEADER, with the 163f9d6ad7SLin Ling * fields enclosed by brackets "[]" replaced with your own identifying 173f9d6ad7SLin Ling * information: Portions Copyright [yyyy] [name of copyright owner] 183f9d6ad7SLin Ling * 193f9d6ad7SLin Ling * CDDL HEADER END 203f9d6ad7SLin Ling */ 213f9d6ad7SLin Ling /* 223f9d6ad7SLin Ling * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. 23be6fd75aSMatthew Ahrens * Copyright (c) 2013 by Delphix. All rights reserved. 243f9d6ad7SLin Ling */ 253f9d6ad7SLin Ling 263f9d6ad7SLin Ling #include <sys/dsl_scan.h> 273f9d6ad7SLin Ling #include <sys/dsl_pool.h> 283f9d6ad7SLin Ling #include <sys/dsl_dataset.h> 293f9d6ad7SLin Ling #include <sys/dsl_prop.h> 303f9d6ad7SLin Ling #include <sys/dsl_dir.h> 313f9d6ad7SLin Ling #include <sys/dsl_synctask.h> 323f9d6ad7SLin Ling #include <sys/dnode.h> 333f9d6ad7SLin Ling #include <sys/dmu_tx.h> 343f9d6ad7SLin Ling #include <sys/dmu_objset.h> 353f9d6ad7SLin Ling #include <sys/arc.h> 363f9d6ad7SLin Ling #include <sys/zap.h> 373f9d6ad7SLin Ling #include <sys/zio.h> 383f9d6ad7SLin Ling #include <sys/zfs_context.h> 393f9d6ad7SLin Ling #include <sys/fs/zfs.h> 403f9d6ad7SLin Ling #include <sys/zfs_znode.h> 413f9d6ad7SLin Ling #include <sys/spa_impl.h> 423f9d6ad7SLin Ling #include <sys/vdev_impl.h> 433f9d6ad7SLin Ling #include <sys/zil_impl.h> 443f9d6ad7SLin Ling #include <sys/zio_checksum.h> 453f9d6ad7SLin Ling #include <sys/ddt.h> 463f9d6ad7SLin Ling #include <sys/sa.h> 473f9d6ad7SLin Ling #include <sys/sa_impl.h> 48ad135b5dSChristopher Siden #include <sys/zfeature.h> 493f9d6ad7SLin Ling #ifdef _KERNEL 503f9d6ad7SLin Ling #include <sys/zfs_vfsops.h> 513f9d6ad7SLin Ling #endif 523f9d6ad7SLin Ling 533f9d6ad7SLin Ling typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *); 543f9d6ad7SLin Ling 553f9d6ad7SLin Ling static scan_cb_t dsl_scan_defrag_cb; 563f9d6ad7SLin Ling static scan_cb_t dsl_scan_scrub_cb; 573f9d6ad7SLin Ling static scan_cb_t dsl_scan_remove_cb; 583b2aab18SMatthew Ahrens static void dsl_scan_cancel_sync(void *, dmu_tx_t *); 593f9d6ad7SLin Ling static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx); 603f9d6ad7SLin Ling 6144ecc532SGeorge Wilson int zfs_top_maxinflight = 32; /* maximum I/Os per top-level */ 6244ecc532SGeorge Wilson int zfs_resilver_delay = 2; /* number of ticks to delay resilver */ 6344ecc532SGeorge Wilson int zfs_scrub_delay = 4; /* number of ticks to delay scrub */ 6444ecc532SGeorge Wilson int zfs_scan_idle = 50; /* idle window in clock ticks */ 6544ecc532SGeorge Wilson 663f9d6ad7SLin Ling int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */ 67cde58dbcSMatthew Ahrens int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */ 683f9d6ad7SLin Ling int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */ 693f9d6ad7SLin Ling boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */ 703f9d6ad7SLin Ling boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable srub prefetching */ 713f9d6ad7SLin Ling enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE; 723f9d6ad7SLin Ling int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */ 733f9d6ad7SLin Ling 743f9d6ad7SLin Ling #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \ 753f9d6ad7SLin Ling ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \ 763f9d6ad7SLin Ling (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER) 773f9d6ad7SLin Ling 783f9d6ad7SLin Ling extern int zfs_txg_timeout; 793f9d6ad7SLin Ling 803f9d6ad7SLin Ling /* the order has to match pool_scan_type */ 813f9d6ad7SLin Ling static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = { 823f9d6ad7SLin Ling NULL, 833f9d6ad7SLin Ling dsl_scan_scrub_cb, /* POOL_SCAN_SCRUB */ 843f9d6ad7SLin Ling dsl_scan_scrub_cb, /* POOL_SCAN_RESILVER */ 853f9d6ad7SLin Ling }; 863f9d6ad7SLin Ling 873f9d6ad7SLin Ling int 883f9d6ad7SLin Ling dsl_scan_init(dsl_pool_t *dp, uint64_t txg) 893f9d6ad7SLin Ling { 903f9d6ad7SLin Ling int err; 913f9d6ad7SLin Ling dsl_scan_t *scn; 923f9d6ad7SLin Ling spa_t *spa = dp->dp_spa; 933f9d6ad7SLin Ling uint64_t f; 943f9d6ad7SLin Ling 953f9d6ad7SLin Ling scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP); 963f9d6ad7SLin Ling scn->scn_dp = dp; 973f9d6ad7SLin Ling 984a923759SGeorge Wilson /* 994a923759SGeorge Wilson * It's possible that we're resuming a scan after a reboot so 1004a923759SGeorge Wilson * make sure that the scan_async_destroying flag is initialized 1014a923759SGeorge Wilson * appropriately. 1024a923759SGeorge Wilson */ 1034a923759SGeorge Wilson ASSERT(!scn->scn_async_destroying); 1044a923759SGeorge Wilson scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa, 1052acef22dSMatthew Ahrens SPA_FEATURE_ASYNC_DESTROY); 1064a923759SGeorge Wilson 1073f9d6ad7SLin Ling err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 1083f9d6ad7SLin Ling "scrub_func", sizeof (uint64_t), 1, &f); 1093f9d6ad7SLin Ling if (err == 0) { 1103f9d6ad7SLin Ling /* 1113f9d6ad7SLin Ling * There was an old-style scrub in progress. Restart a 1123f9d6ad7SLin Ling * new-style scrub from the beginning. 1133f9d6ad7SLin Ling */ 1143f9d6ad7SLin Ling scn->scn_restart_txg = txg; 1153f9d6ad7SLin Ling zfs_dbgmsg("old-style scrub was in progress; " 1163f9d6ad7SLin Ling "restarting new-style scrub in txg %llu", 1173f9d6ad7SLin Ling scn->scn_restart_txg); 1183f9d6ad7SLin Ling 1193f9d6ad7SLin Ling /* 1203f9d6ad7SLin Ling * Load the queue obj from the old location so that it 1213f9d6ad7SLin Ling * can be freed by dsl_scan_done(). 1223f9d6ad7SLin Ling */ 1233f9d6ad7SLin Ling (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 1243f9d6ad7SLin Ling "scrub_queue", sizeof (uint64_t), 1, 1253f9d6ad7SLin Ling &scn->scn_phys.scn_queue_obj); 1263f9d6ad7SLin Ling } else { 1273f9d6ad7SLin Ling err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 1283f9d6ad7SLin Ling DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 1293f9d6ad7SLin Ling &scn->scn_phys); 1303f9d6ad7SLin Ling if (err == ENOENT) 1313f9d6ad7SLin Ling return (0); 1323f9d6ad7SLin Ling else if (err) 1333f9d6ad7SLin Ling return (err); 1343f9d6ad7SLin Ling 1353f9d6ad7SLin Ling if (scn->scn_phys.scn_state == DSS_SCANNING && 1363f9d6ad7SLin Ling spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) { 1373f9d6ad7SLin Ling /* 1383f9d6ad7SLin Ling * A new-type scrub was in progress on an old 1393f9d6ad7SLin Ling * pool, and the pool was accessed by old 1403f9d6ad7SLin Ling * software. Restart from the beginning, since 1413f9d6ad7SLin Ling * the old software may have changed the pool in 1423f9d6ad7SLin Ling * the meantime. 1433f9d6ad7SLin Ling */ 1443f9d6ad7SLin Ling scn->scn_restart_txg = txg; 1453f9d6ad7SLin Ling zfs_dbgmsg("new-style scrub was modified " 1463f9d6ad7SLin Ling "by old software; restarting in txg %llu", 1473f9d6ad7SLin Ling scn->scn_restart_txg); 1483f9d6ad7SLin Ling } 1493f9d6ad7SLin Ling } 1503f9d6ad7SLin Ling 1513f9d6ad7SLin Ling spa_scan_stat_init(spa); 1523f9d6ad7SLin Ling return (0); 1533f9d6ad7SLin Ling } 1543f9d6ad7SLin Ling 1553f9d6ad7SLin Ling void 1563f9d6ad7SLin Ling dsl_scan_fini(dsl_pool_t *dp) 1573f9d6ad7SLin Ling { 1583f9d6ad7SLin Ling if (dp->dp_scan) { 1593f9d6ad7SLin Ling kmem_free(dp->dp_scan, sizeof (dsl_scan_t)); 1603f9d6ad7SLin Ling dp->dp_scan = NULL; 1613f9d6ad7SLin Ling } 1623f9d6ad7SLin Ling } 1633f9d6ad7SLin Ling 1643f9d6ad7SLin Ling /* ARGSUSED */ 1653f9d6ad7SLin Ling static int 1663b2aab18SMatthew Ahrens dsl_scan_setup_check(void *arg, dmu_tx_t *tx) 1673f9d6ad7SLin Ling { 1683b2aab18SMatthew Ahrens dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 1693f9d6ad7SLin Ling 1703f9d6ad7SLin Ling if (scn->scn_phys.scn_state == DSS_SCANNING) 171be6fd75aSMatthew Ahrens return (SET_ERROR(EBUSY)); 1723f9d6ad7SLin Ling 1733f9d6ad7SLin Ling return (0); 1743f9d6ad7SLin Ling } 1753f9d6ad7SLin Ling 1763f9d6ad7SLin Ling static void 1773b2aab18SMatthew Ahrens dsl_scan_setup_sync(void *arg, dmu_tx_t *tx) 1783f9d6ad7SLin Ling { 1793b2aab18SMatthew Ahrens dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 1803b2aab18SMatthew Ahrens pool_scan_func_t *funcp = arg; 1813f9d6ad7SLin Ling dmu_object_type_t ot = 0; 1823f9d6ad7SLin Ling dsl_pool_t *dp = scn->scn_dp; 1833f9d6ad7SLin Ling spa_t *spa = dp->dp_spa; 1843f9d6ad7SLin Ling 1853f9d6ad7SLin Ling ASSERT(scn->scn_phys.scn_state != DSS_SCANNING); 1863f9d6ad7SLin Ling ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS); 1873f9d6ad7SLin Ling bzero(&scn->scn_phys, sizeof (scn->scn_phys)); 1883f9d6ad7SLin Ling scn->scn_phys.scn_func = *funcp; 1893f9d6ad7SLin Ling scn->scn_phys.scn_state = DSS_SCANNING; 1903f9d6ad7SLin Ling scn->scn_phys.scn_min_txg = 0; 1913f9d6ad7SLin Ling scn->scn_phys.scn_max_txg = tx->tx_txg; 1923f9d6ad7SLin Ling scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */ 1933f9d6ad7SLin Ling scn->scn_phys.scn_start_time = gethrestime_sec(); 1943f9d6ad7SLin Ling scn->scn_phys.scn_errors = 0; 1953f9d6ad7SLin Ling scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc; 1963f9d6ad7SLin Ling scn->scn_restart_txg = 0; 197b4952e17SGeorge Wilson scn->scn_done_txg = 0; 1983f9d6ad7SLin Ling spa_scan_stat_init(spa); 1993f9d6ad7SLin Ling 2003f9d6ad7SLin Ling if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 2013f9d6ad7SLin Ling scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max; 2023f9d6ad7SLin Ling 2033f9d6ad7SLin Ling /* rewrite all disk labels */ 2043f9d6ad7SLin Ling vdev_config_dirty(spa->spa_root_vdev); 2053f9d6ad7SLin Ling 2063f9d6ad7SLin Ling if (vdev_resilver_needed(spa->spa_root_vdev, 2073f9d6ad7SLin Ling &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) { 2083f9d6ad7SLin Ling spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START); 2093f9d6ad7SLin Ling } else { 2103f9d6ad7SLin Ling spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START); 2113f9d6ad7SLin Ling } 2123f9d6ad7SLin Ling 2133f9d6ad7SLin Ling spa->spa_scrub_started = B_TRUE; 2143f9d6ad7SLin Ling /* 2153f9d6ad7SLin Ling * If this is an incremental scrub, limit the DDT scrub phase 2163f9d6ad7SLin Ling * to just the auto-ditto class (for correctness); the rest 2173f9d6ad7SLin Ling * of the scrub should go faster using top-down pruning. 2183f9d6ad7SLin Ling */ 2193f9d6ad7SLin Ling if (scn->scn_phys.scn_min_txg > TXG_INITIAL) 2203f9d6ad7SLin Ling scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO; 2213f9d6ad7SLin Ling 2223f9d6ad7SLin Ling } 2233f9d6ad7SLin Ling 2243f9d6ad7SLin Ling /* back to the generic stuff */ 2253f9d6ad7SLin Ling 2263f9d6ad7SLin Ling if (dp->dp_blkstats == NULL) { 2273f9d6ad7SLin Ling dp->dp_blkstats = 2283f9d6ad7SLin Ling kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP); 2293f9d6ad7SLin Ling } 2303f9d6ad7SLin Ling bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t)); 2313f9d6ad7SLin Ling 2323f9d6ad7SLin Ling if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) 2333f9d6ad7SLin Ling ot = DMU_OT_ZAP_OTHER; 2343f9d6ad7SLin Ling 2353f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, 2363f9d6ad7SLin Ling ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx); 2373f9d6ad7SLin Ling 2383f9d6ad7SLin Ling dsl_scan_sync_state(scn, tx); 2393f9d6ad7SLin Ling 2404445fffbSMatthew Ahrens spa_history_log_internal(spa, "scan setup", tx, 2413f9d6ad7SLin Ling "func=%u mintxg=%llu maxtxg=%llu", 2423f9d6ad7SLin Ling *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg); 2433f9d6ad7SLin Ling } 2443f9d6ad7SLin Ling 2453f9d6ad7SLin Ling /* ARGSUSED */ 2463f9d6ad7SLin Ling static void 2473f9d6ad7SLin Ling dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx) 2483f9d6ad7SLin Ling { 2493f9d6ad7SLin Ling static const char *old_names[] = { 2503f9d6ad7SLin Ling "scrub_bookmark", 2513f9d6ad7SLin Ling "scrub_ddt_bookmark", 2523f9d6ad7SLin Ling "scrub_ddt_class_max", 2533f9d6ad7SLin Ling "scrub_queue", 2543f9d6ad7SLin Ling "scrub_min_txg", 2553f9d6ad7SLin Ling "scrub_max_txg", 2563f9d6ad7SLin Ling "scrub_func", 2573f9d6ad7SLin Ling "scrub_errors", 2583f9d6ad7SLin Ling NULL 2593f9d6ad7SLin Ling }; 2603f9d6ad7SLin Ling 2613f9d6ad7SLin Ling dsl_pool_t *dp = scn->scn_dp; 2623f9d6ad7SLin Ling spa_t *spa = dp->dp_spa; 2633f9d6ad7SLin Ling int i; 2643f9d6ad7SLin Ling 2653f9d6ad7SLin Ling /* Remove any remnants of an old-style scrub. */ 2663f9d6ad7SLin Ling for (i = 0; old_names[i]; i++) { 2673f9d6ad7SLin Ling (void) zap_remove(dp->dp_meta_objset, 2683f9d6ad7SLin Ling DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx); 2693f9d6ad7SLin Ling } 2703f9d6ad7SLin Ling 2713f9d6ad7SLin Ling if (scn->scn_phys.scn_queue_obj != 0) { 2723f9d6ad7SLin Ling VERIFY(0 == dmu_object_free(dp->dp_meta_objset, 2733f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, tx)); 2743f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj = 0; 2753f9d6ad7SLin Ling } 2763f9d6ad7SLin Ling 2773f9d6ad7SLin Ling /* 2783f9d6ad7SLin Ling * If we were "restarted" from a stopped state, don't bother 2793f9d6ad7SLin Ling * with anything else. 2803f9d6ad7SLin Ling */ 2813f9d6ad7SLin Ling if (scn->scn_phys.scn_state != DSS_SCANNING) 2823f9d6ad7SLin Ling return; 2833f9d6ad7SLin Ling 2843f9d6ad7SLin Ling if (complete) 2853f9d6ad7SLin Ling scn->scn_phys.scn_state = DSS_FINISHED; 2863f9d6ad7SLin Ling else 2873f9d6ad7SLin Ling scn->scn_phys.scn_state = DSS_CANCELED; 2883f9d6ad7SLin Ling 2894445fffbSMatthew Ahrens spa_history_log_internal(spa, "scan done", tx, 2903f9d6ad7SLin Ling "complete=%u", complete); 2913f9d6ad7SLin Ling 2923f9d6ad7SLin Ling if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 2933f9d6ad7SLin Ling mutex_enter(&spa->spa_scrub_lock); 2943f9d6ad7SLin Ling while (spa->spa_scrub_inflight > 0) { 2953f9d6ad7SLin Ling cv_wait(&spa->spa_scrub_io_cv, 2963f9d6ad7SLin Ling &spa->spa_scrub_lock); 2973f9d6ad7SLin Ling } 2983f9d6ad7SLin Ling mutex_exit(&spa->spa_scrub_lock); 2993f9d6ad7SLin Ling spa->spa_scrub_started = B_FALSE; 3003f9d6ad7SLin Ling spa->spa_scrub_active = B_FALSE; 3013f9d6ad7SLin Ling 3023f9d6ad7SLin Ling /* 3033f9d6ad7SLin Ling * If the scrub/resilver completed, update all DTLs to 3043f9d6ad7SLin Ling * reflect this. Whether it succeeded or not, vacate 3053f9d6ad7SLin Ling * all temporary scrub DTLs. 3063f9d6ad7SLin Ling */ 3073f9d6ad7SLin Ling vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg, 3083f9d6ad7SLin Ling complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE); 3093f9d6ad7SLin Ling if (complete) { 3103f9d6ad7SLin Ling spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ? 3113f9d6ad7SLin Ling ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH); 3123f9d6ad7SLin Ling } 3133f9d6ad7SLin Ling spa_errlog_rotate(spa); 3143f9d6ad7SLin Ling 3153f9d6ad7SLin Ling /* 3163f9d6ad7SLin Ling * We may have finished replacing a device. 3173f9d6ad7SLin Ling * Let the async thread assess this and handle the detach. 3183f9d6ad7SLin Ling */ 3193f9d6ad7SLin Ling spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); 3203f9d6ad7SLin Ling } 3213f9d6ad7SLin Ling 3223f9d6ad7SLin Ling scn->scn_phys.scn_end_time = gethrestime_sec(); 3233f9d6ad7SLin Ling } 3243f9d6ad7SLin Ling 3253f9d6ad7SLin Ling /* ARGSUSED */ 3263f9d6ad7SLin Ling static int 3273b2aab18SMatthew Ahrens dsl_scan_cancel_check(void *arg, dmu_tx_t *tx) 3283f9d6ad7SLin Ling { 3293b2aab18SMatthew Ahrens dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 3303f9d6ad7SLin Ling 3313f9d6ad7SLin Ling if (scn->scn_phys.scn_state != DSS_SCANNING) 332be6fd75aSMatthew Ahrens return (SET_ERROR(ENOENT)); 3333f9d6ad7SLin Ling return (0); 3343f9d6ad7SLin Ling } 3353f9d6ad7SLin Ling 3363f9d6ad7SLin Ling /* ARGSUSED */ 3373f9d6ad7SLin Ling static void 3383b2aab18SMatthew Ahrens dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx) 3393f9d6ad7SLin Ling { 3403b2aab18SMatthew Ahrens dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan; 3413f9d6ad7SLin Ling 3423f9d6ad7SLin Ling dsl_scan_done(scn, B_FALSE, tx); 3433f9d6ad7SLin Ling dsl_scan_sync_state(scn, tx); 3443f9d6ad7SLin Ling } 3453f9d6ad7SLin Ling 3463f9d6ad7SLin Ling int 3473f9d6ad7SLin Ling dsl_scan_cancel(dsl_pool_t *dp) 3483f9d6ad7SLin Ling { 3493b2aab18SMatthew Ahrens return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check, 3503b2aab18SMatthew Ahrens dsl_scan_cancel_sync, NULL, 3)); 3513f9d6ad7SLin Ling } 3523f9d6ad7SLin Ling 3533f9d6ad7SLin Ling static void dsl_scan_visitbp(blkptr_t *bp, 3543f9d6ad7SLin Ling const zbookmark_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf, 3553f9d6ad7SLin Ling dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype, 3563f9d6ad7SLin Ling dmu_tx_t *tx); 3573f9d6ad7SLin Ling static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds, 3583f9d6ad7SLin Ling dmu_objset_type_t ostype, 3593f9d6ad7SLin Ling dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx); 3603f9d6ad7SLin Ling 3613f9d6ad7SLin Ling void 3623f9d6ad7SLin Ling dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp) 3633f9d6ad7SLin Ling { 3643f9d6ad7SLin Ling zio_free(dp->dp_spa, txg, bp); 3653f9d6ad7SLin Ling } 3663f9d6ad7SLin Ling 3673f9d6ad7SLin Ling void 3683f9d6ad7SLin Ling dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp) 3693f9d6ad7SLin Ling { 3703f9d6ad7SLin Ling ASSERT(dsl_pool_sync_context(dp)); 3713f9d6ad7SLin Ling zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags)); 3723f9d6ad7SLin Ling } 3733f9d6ad7SLin Ling 3743f9d6ad7SLin Ling static uint64_t 3753f9d6ad7SLin Ling dsl_scan_ds_maxtxg(dsl_dataset_t *ds) 3763f9d6ad7SLin Ling { 3773f9d6ad7SLin Ling uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg; 3783f9d6ad7SLin Ling if (dsl_dataset_is_snapshot(ds)) 3793f9d6ad7SLin Ling return (MIN(smt, ds->ds_phys->ds_creation_txg)); 3803f9d6ad7SLin Ling return (smt); 3813f9d6ad7SLin Ling } 3823f9d6ad7SLin Ling 3833f9d6ad7SLin Ling static void 3843f9d6ad7SLin Ling dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx) 3853f9d6ad7SLin Ling { 3863b2aab18SMatthew Ahrens VERIFY0(zap_update(scn->scn_dp->dp_meta_objset, 3873f9d6ad7SLin Ling DMU_POOL_DIRECTORY_OBJECT, 3883f9d6ad7SLin Ling DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS, 3893f9d6ad7SLin Ling &scn->scn_phys, tx)); 3903f9d6ad7SLin Ling } 3913f9d6ad7SLin Ling 3923f9d6ad7SLin Ling static boolean_t 3933f9d6ad7SLin Ling dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_t *zb) 3943f9d6ad7SLin Ling { 3953f9d6ad7SLin Ling uint64_t elapsed_nanosecs; 3963f9d6ad7SLin Ling int mintime; 3973f9d6ad7SLin Ling 3983f9d6ad7SLin Ling /* we never skip user/group accounting objects */ 3993f9d6ad7SLin Ling if (zb && (int64_t)zb->zb_object < 0) 4003f9d6ad7SLin Ling return (B_FALSE); 4013f9d6ad7SLin Ling 4023f9d6ad7SLin Ling if (scn->scn_pausing) 4033f9d6ad7SLin Ling return (B_TRUE); /* we're already pausing */ 4043f9d6ad7SLin Ling 405ad135b5dSChristopher Siden if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) 4063f9d6ad7SLin Ling return (B_FALSE); /* we're resuming */ 4073f9d6ad7SLin Ling 4083f9d6ad7SLin Ling /* We only know how to resume from level-0 blocks. */ 4093f9d6ad7SLin Ling if (zb && zb->zb_level != 0) 4103f9d6ad7SLin Ling return (B_FALSE); 4113f9d6ad7SLin Ling 4123f9d6ad7SLin Ling mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ? 4133f9d6ad7SLin Ling zfs_resilver_min_time_ms : zfs_scan_min_time_ms; 4143f9d6ad7SLin Ling elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 4153f9d6ad7SLin Ling if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 4160689f76cSAdam Leventhal (NSEC2MSEC(elapsed_nanosecs) > mintime && 4173f9d6ad7SLin Ling txg_sync_waiting(scn->scn_dp)) || 4183f9d6ad7SLin Ling spa_shutting_down(scn->scn_dp->dp_spa)) { 4193f9d6ad7SLin Ling if (zb) { 4203f9d6ad7SLin Ling dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n", 4213f9d6ad7SLin Ling (longlong_t)zb->zb_objset, 4223f9d6ad7SLin Ling (longlong_t)zb->zb_object, 4233f9d6ad7SLin Ling (longlong_t)zb->zb_level, 4243f9d6ad7SLin Ling (longlong_t)zb->zb_blkid); 4253f9d6ad7SLin Ling scn->scn_phys.scn_bookmark = *zb; 4263f9d6ad7SLin Ling } 4273f9d6ad7SLin Ling dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n", 4283f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class, 4293f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type, 4303f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum, 4313f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor); 4323f9d6ad7SLin Ling scn->scn_pausing = B_TRUE; 4333f9d6ad7SLin Ling return (B_TRUE); 4343f9d6ad7SLin Ling } 4353f9d6ad7SLin Ling return (B_FALSE); 4363f9d6ad7SLin Ling } 4373f9d6ad7SLin Ling 4383f9d6ad7SLin Ling typedef struct zil_scan_arg { 4393f9d6ad7SLin Ling dsl_pool_t *zsa_dp; 4403f9d6ad7SLin Ling zil_header_t *zsa_zh; 4413f9d6ad7SLin Ling } zil_scan_arg_t; 4423f9d6ad7SLin Ling 4433f9d6ad7SLin Ling /* ARGSUSED */ 4443f9d6ad7SLin Ling static int 4453f9d6ad7SLin Ling dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg) 4463f9d6ad7SLin Ling { 4473f9d6ad7SLin Ling zil_scan_arg_t *zsa = arg; 4483f9d6ad7SLin Ling dsl_pool_t *dp = zsa->zsa_dp; 4493f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 4503f9d6ad7SLin Ling zil_header_t *zh = zsa->zsa_zh; 4513f9d6ad7SLin Ling zbookmark_t zb; 4523f9d6ad7SLin Ling 45343466aaeSMax Grossman if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 4543f9d6ad7SLin Ling return (0); 4553f9d6ad7SLin Ling 4563f9d6ad7SLin Ling /* 4573f9d6ad7SLin Ling * One block ("stubby") can be allocated a long time ago; we 4583f9d6ad7SLin Ling * want to visit that one because it has been allocated 4593f9d6ad7SLin Ling * (on-disk) even if it hasn't been claimed (even though for 4603f9d6ad7SLin Ling * scrub there's nothing to do to it). 4613f9d6ad7SLin Ling */ 4623f9d6ad7SLin Ling if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa)) 4633f9d6ad7SLin Ling return (0); 4643f9d6ad7SLin Ling 4653f9d6ad7SLin Ling SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 4663f9d6ad7SLin Ling ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]); 4673f9d6ad7SLin Ling 4683f9d6ad7SLin Ling VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 4693f9d6ad7SLin Ling return (0); 4703f9d6ad7SLin Ling } 4713f9d6ad7SLin Ling 4723f9d6ad7SLin Ling /* ARGSUSED */ 4733f9d6ad7SLin Ling static int 4743f9d6ad7SLin Ling dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg) 4753f9d6ad7SLin Ling { 4763f9d6ad7SLin Ling if (lrc->lrc_txtype == TX_WRITE) { 4773f9d6ad7SLin Ling zil_scan_arg_t *zsa = arg; 4783f9d6ad7SLin Ling dsl_pool_t *dp = zsa->zsa_dp; 4793f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 4803f9d6ad7SLin Ling zil_header_t *zh = zsa->zsa_zh; 4813f9d6ad7SLin Ling lr_write_t *lr = (lr_write_t *)lrc; 4823f9d6ad7SLin Ling blkptr_t *bp = &lr->lr_blkptr; 4833f9d6ad7SLin Ling zbookmark_t zb; 4843f9d6ad7SLin Ling 48543466aaeSMax Grossman if (BP_IS_HOLE(bp) || 48643466aaeSMax Grossman bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 4873f9d6ad7SLin Ling return (0); 4883f9d6ad7SLin Ling 4893f9d6ad7SLin Ling /* 4903f9d6ad7SLin Ling * birth can be < claim_txg if this record's txg is 4913f9d6ad7SLin Ling * already txg sync'ed (but this log block contains 4923f9d6ad7SLin Ling * other records that are not synced) 4933f9d6ad7SLin Ling */ 4943f9d6ad7SLin Ling if (claim_txg == 0 || bp->blk_birth < claim_txg) 4953f9d6ad7SLin Ling return (0); 4963f9d6ad7SLin Ling 4973f9d6ad7SLin Ling SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET], 4983f9d6ad7SLin Ling lr->lr_foid, ZB_ZIL_LEVEL, 4993f9d6ad7SLin Ling lr->lr_offset / BP_GET_LSIZE(bp)); 5003f9d6ad7SLin Ling 5013f9d6ad7SLin Ling VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb)); 5023f9d6ad7SLin Ling } 5033f9d6ad7SLin Ling return (0); 5043f9d6ad7SLin Ling } 5053f9d6ad7SLin Ling 5063f9d6ad7SLin Ling static void 5073f9d6ad7SLin Ling dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh) 5083f9d6ad7SLin Ling { 5093f9d6ad7SLin Ling uint64_t claim_txg = zh->zh_claim_txg; 5103f9d6ad7SLin Ling zil_scan_arg_t zsa = { dp, zh }; 5113f9d6ad7SLin Ling zilog_t *zilog; 5123f9d6ad7SLin Ling 5133f9d6ad7SLin Ling /* 5143f9d6ad7SLin Ling * We only want to visit blocks that have been claimed but not yet 5153f9d6ad7SLin Ling * replayed (or, in read-only mode, blocks that *would* be claimed). 5163f9d6ad7SLin Ling */ 5173f9d6ad7SLin Ling if (claim_txg == 0 && spa_writeable(dp->dp_spa)) 5183f9d6ad7SLin Ling return; 5193f9d6ad7SLin Ling 5203f9d6ad7SLin Ling zilog = zil_alloc(dp->dp_meta_objset, zh); 5213f9d6ad7SLin Ling 5223f9d6ad7SLin Ling (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa, 5233f9d6ad7SLin Ling claim_txg); 5243f9d6ad7SLin Ling 5253f9d6ad7SLin Ling zil_free(zilog); 5263f9d6ad7SLin Ling } 5273f9d6ad7SLin Ling 5283f9d6ad7SLin Ling /* ARGSUSED */ 5293f9d6ad7SLin Ling static void 5303f9d6ad7SLin Ling dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp, 5313f9d6ad7SLin Ling uint64_t objset, uint64_t object, uint64_t blkid) 5323f9d6ad7SLin Ling { 5333f9d6ad7SLin Ling zbookmark_t czb; 5343f9d6ad7SLin Ling uint32_t flags = ARC_NOWAIT | ARC_PREFETCH; 5353f9d6ad7SLin Ling 5363f9d6ad7SLin Ling if (zfs_no_scrub_prefetch) 5373f9d6ad7SLin Ling return; 5383f9d6ad7SLin Ling 5393f9d6ad7SLin Ling if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg || 5403f9d6ad7SLin Ling (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE)) 5413f9d6ad7SLin Ling return; 5423f9d6ad7SLin Ling 5433f9d6ad7SLin Ling SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid); 5443f9d6ad7SLin Ling 545cde58dbcSMatthew Ahrens (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp, 5461b912ec7SGeorge Wilson NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 54744ecc532SGeorge Wilson ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb); 5483f9d6ad7SLin Ling } 5493f9d6ad7SLin Ling 5503f9d6ad7SLin Ling static boolean_t 5513f9d6ad7SLin Ling dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp, 5523f9d6ad7SLin Ling const zbookmark_t *zb) 5533f9d6ad7SLin Ling { 5543f9d6ad7SLin Ling /* 5553f9d6ad7SLin Ling * We never skip over user/group accounting objects (obj<0) 5563f9d6ad7SLin Ling */ 557ad135b5dSChristopher Siden if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) && 5583f9d6ad7SLin Ling (int64_t)zb->zb_object >= 0) { 5593f9d6ad7SLin Ling /* 5603f9d6ad7SLin Ling * If we already visited this bp & everything below (in 5613f9d6ad7SLin Ling * a prior txg sync), don't bother doing it again. 5623f9d6ad7SLin Ling */ 563ad135b5dSChristopher Siden if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark)) 5643f9d6ad7SLin Ling return (B_TRUE); 5653f9d6ad7SLin Ling 5663f9d6ad7SLin Ling /* 5673f9d6ad7SLin Ling * If we found the block we're trying to resume from, or 5683f9d6ad7SLin Ling * we went past it to a different object, zero it out to 5693f9d6ad7SLin Ling * indicate that it's OK to start checking for pausing 5703f9d6ad7SLin Ling * again. 5713f9d6ad7SLin Ling */ 5723f9d6ad7SLin Ling if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 || 5733f9d6ad7SLin Ling zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) { 5743f9d6ad7SLin Ling dprintf("resuming at %llx/%llx/%llx/%llx\n", 5753f9d6ad7SLin Ling (longlong_t)zb->zb_objset, 5763f9d6ad7SLin Ling (longlong_t)zb->zb_object, 5773f9d6ad7SLin Ling (longlong_t)zb->zb_level, 5783f9d6ad7SLin Ling (longlong_t)zb->zb_blkid); 5793f9d6ad7SLin Ling bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb)); 5803f9d6ad7SLin Ling } 5813f9d6ad7SLin Ling } 5823f9d6ad7SLin Ling return (B_FALSE); 5833f9d6ad7SLin Ling } 5843f9d6ad7SLin Ling 5853f9d6ad7SLin Ling /* 5863f9d6ad7SLin Ling * Return nonzero on i/o error. 5873f9d6ad7SLin Ling * Return new buf to write out in *bufp. 5883f9d6ad7SLin Ling */ 5893f9d6ad7SLin Ling static int 5903f9d6ad7SLin Ling dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype, 5913f9d6ad7SLin Ling dnode_phys_t *dnp, const blkptr_t *bp, 5923f9d6ad7SLin Ling const zbookmark_t *zb, dmu_tx_t *tx, arc_buf_t **bufp) 5933f9d6ad7SLin Ling { 5943f9d6ad7SLin Ling dsl_pool_t *dp = scn->scn_dp; 59544ecc532SGeorge Wilson int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD; 5963f9d6ad7SLin Ling int err; 5973f9d6ad7SLin Ling 5983f9d6ad7SLin Ling if (BP_GET_LEVEL(bp) > 0) { 5993f9d6ad7SLin Ling uint32_t flags = ARC_WAIT; 6003f9d6ad7SLin Ling int i; 6013f9d6ad7SLin Ling blkptr_t *cbp; 6023f9d6ad7SLin Ling int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT; 6033f9d6ad7SLin Ling 6041b912ec7SGeorge Wilson err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 60544ecc532SGeorge Wilson ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 6063f9d6ad7SLin Ling if (err) { 6073f9d6ad7SLin Ling scn->scn_phys.scn_errors++; 6083f9d6ad7SLin Ling return (err); 6093f9d6ad7SLin Ling } 6103f9d6ad7SLin Ling for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) { 6113f9d6ad7SLin Ling dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset, 6123f9d6ad7SLin Ling zb->zb_object, zb->zb_blkid * epb + i); 6133f9d6ad7SLin Ling } 6143f9d6ad7SLin Ling for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) { 6153f9d6ad7SLin Ling zbookmark_t czb; 6163f9d6ad7SLin Ling 6173f9d6ad7SLin Ling SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object, 6183f9d6ad7SLin Ling zb->zb_level - 1, 6193f9d6ad7SLin Ling zb->zb_blkid * epb + i); 6203f9d6ad7SLin Ling dsl_scan_visitbp(cbp, &czb, dnp, 6213f9d6ad7SLin Ling *bufp, ds, scn, ostype, tx); 6223f9d6ad7SLin Ling } 6233f9d6ad7SLin Ling } else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) { 6243f9d6ad7SLin Ling uint32_t flags = ARC_WAIT; 6253f9d6ad7SLin Ling 6261b912ec7SGeorge Wilson err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 62744ecc532SGeorge Wilson ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 6283f9d6ad7SLin Ling if (err) { 6293f9d6ad7SLin Ling scn->scn_phys.scn_errors++; 6303f9d6ad7SLin Ling return (err); 6313f9d6ad7SLin Ling } 6323f9d6ad7SLin Ling } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) { 6333f9d6ad7SLin Ling uint32_t flags = ARC_WAIT; 6343f9d6ad7SLin Ling dnode_phys_t *cdnp; 6353f9d6ad7SLin Ling int i, j; 6363f9d6ad7SLin Ling int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT; 6373f9d6ad7SLin Ling 6381b912ec7SGeorge Wilson err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 63944ecc532SGeorge Wilson ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 6403f9d6ad7SLin Ling if (err) { 6413f9d6ad7SLin Ling scn->scn_phys.scn_errors++; 6423f9d6ad7SLin Ling return (err); 6433f9d6ad7SLin Ling } 6443f9d6ad7SLin Ling for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) { 6453f9d6ad7SLin Ling for (j = 0; j < cdnp->dn_nblkptr; j++) { 6463f9d6ad7SLin Ling blkptr_t *cbp = &cdnp->dn_blkptr[j]; 6473f9d6ad7SLin Ling dsl_scan_prefetch(scn, *bufp, cbp, 6483f9d6ad7SLin Ling zb->zb_objset, zb->zb_blkid * epb + i, j); 6493f9d6ad7SLin Ling } 6503f9d6ad7SLin Ling } 6513f9d6ad7SLin Ling for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) { 6523f9d6ad7SLin Ling dsl_scan_visitdnode(scn, ds, ostype, 6533f9d6ad7SLin Ling cdnp, *bufp, zb->zb_blkid * epb + i, tx); 6543f9d6ad7SLin Ling } 6553f9d6ad7SLin Ling 6563f9d6ad7SLin Ling } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) { 6573f9d6ad7SLin Ling uint32_t flags = ARC_WAIT; 6583f9d6ad7SLin Ling objset_phys_t *osp; 6593f9d6ad7SLin Ling 6601b912ec7SGeorge Wilson err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp, 66144ecc532SGeorge Wilson ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb); 6623f9d6ad7SLin Ling if (err) { 6633f9d6ad7SLin Ling scn->scn_phys.scn_errors++; 6643f9d6ad7SLin Ling return (err); 6653f9d6ad7SLin Ling } 6663f9d6ad7SLin Ling 6673f9d6ad7SLin Ling osp = (*bufp)->b_data; 6683f9d6ad7SLin Ling 6693f9d6ad7SLin Ling dsl_scan_visitdnode(scn, ds, osp->os_type, 6703f9d6ad7SLin Ling &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx); 6713f9d6ad7SLin Ling 6723f9d6ad7SLin Ling if (OBJSET_BUF_HAS_USERUSED(*bufp)) { 6733f9d6ad7SLin Ling /* 6743f9d6ad7SLin Ling * We also always visit user/group accounting 6753f9d6ad7SLin Ling * objects, and never skip them, even if we are 6763f9d6ad7SLin Ling * pausing. This is necessary so that the space 6773f9d6ad7SLin Ling * deltas from this txg get integrated. 6783f9d6ad7SLin Ling */ 6793f9d6ad7SLin Ling dsl_scan_visitdnode(scn, ds, osp->os_type, 6803f9d6ad7SLin Ling &osp->os_groupused_dnode, *bufp, 6813f9d6ad7SLin Ling DMU_GROUPUSED_OBJECT, tx); 6823f9d6ad7SLin Ling dsl_scan_visitdnode(scn, ds, osp->os_type, 6833f9d6ad7SLin Ling &osp->os_userused_dnode, *bufp, 6843f9d6ad7SLin Ling DMU_USERUSED_OBJECT, tx); 6853f9d6ad7SLin Ling } 6863f9d6ad7SLin Ling } 6873f9d6ad7SLin Ling 6883f9d6ad7SLin Ling return (0); 6893f9d6ad7SLin Ling } 6903f9d6ad7SLin Ling 6913f9d6ad7SLin Ling static void 6923f9d6ad7SLin Ling dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds, 6933f9d6ad7SLin Ling dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf, 6943f9d6ad7SLin Ling uint64_t object, dmu_tx_t *tx) 6953f9d6ad7SLin Ling { 6963f9d6ad7SLin Ling int j; 6973f9d6ad7SLin Ling 6983f9d6ad7SLin Ling for (j = 0; j < dnp->dn_nblkptr; j++) { 6993f9d6ad7SLin Ling zbookmark_t czb; 7003f9d6ad7SLin Ling 7013f9d6ad7SLin Ling SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 7023f9d6ad7SLin Ling dnp->dn_nlevels - 1, j); 7033f9d6ad7SLin Ling dsl_scan_visitbp(&dnp->dn_blkptr[j], 7043f9d6ad7SLin Ling &czb, dnp, buf, ds, scn, ostype, tx); 7053f9d6ad7SLin Ling } 7063f9d6ad7SLin Ling 7073f9d6ad7SLin Ling if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 7083f9d6ad7SLin Ling zbookmark_t czb; 7093f9d6ad7SLin Ling SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object, 7103f9d6ad7SLin Ling 0, DMU_SPILL_BLKID); 7113f9d6ad7SLin Ling dsl_scan_visitbp(&dnp->dn_spill, 7123f9d6ad7SLin Ling &czb, dnp, buf, ds, scn, ostype, tx); 7133f9d6ad7SLin Ling } 7143f9d6ad7SLin Ling } 7153f9d6ad7SLin Ling 7163f9d6ad7SLin Ling /* 7173f9d6ad7SLin Ling * The arguments are in this order because mdb can only print the 7183f9d6ad7SLin Ling * first 5; we want them to be useful. 7193f9d6ad7SLin Ling */ 7203f9d6ad7SLin Ling static void 7213f9d6ad7SLin Ling dsl_scan_visitbp(blkptr_t *bp, const zbookmark_t *zb, 7223f9d6ad7SLin Ling dnode_phys_t *dnp, arc_buf_t *pbuf, 7233f9d6ad7SLin Ling dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype, 7243f9d6ad7SLin Ling dmu_tx_t *tx) 7253f9d6ad7SLin Ling { 7263f9d6ad7SLin Ling dsl_pool_t *dp = scn->scn_dp; 7273f9d6ad7SLin Ling arc_buf_t *buf = NULL; 7283f9d6ad7SLin Ling blkptr_t bp_toread = *bp; 7293f9d6ad7SLin Ling 7303f9d6ad7SLin Ling /* ASSERT(pbuf == NULL || arc_released(pbuf)); */ 7313f9d6ad7SLin Ling 7323f9d6ad7SLin Ling if (dsl_scan_check_pause(scn, zb)) 7333f9d6ad7SLin Ling return; 7343f9d6ad7SLin Ling 7353f9d6ad7SLin Ling if (dsl_scan_check_resume(scn, dnp, zb)) 7363f9d6ad7SLin Ling return; 7373f9d6ad7SLin Ling 73843466aaeSMax Grossman if (BP_IS_HOLE(bp)) 7393f9d6ad7SLin Ling return; 7403f9d6ad7SLin Ling 7413f9d6ad7SLin Ling scn->scn_visited_this_txg++; 7423f9d6ad7SLin Ling 7433f9d6ad7SLin Ling dprintf_bp(bp, 7443f9d6ad7SLin Ling "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p", 7453f9d6ad7SLin Ling ds, ds ? ds->ds_object : 0, 7463f9d6ad7SLin Ling zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, 7473f9d6ad7SLin Ling pbuf, bp); 7483f9d6ad7SLin Ling 7493f9d6ad7SLin Ling if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) 7503f9d6ad7SLin Ling return; 7513f9d6ad7SLin Ling 7523f9d6ad7SLin Ling if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx, 7533f9d6ad7SLin Ling &buf) != 0) 7543f9d6ad7SLin Ling return; 7553f9d6ad7SLin Ling 7563f9d6ad7SLin Ling /* 7573f9d6ad7SLin Ling * If dsl_scan_ddt() has aready visited this block, it will have 7583f9d6ad7SLin Ling * already done any translations or scrubbing, so don't call the 7593f9d6ad7SLin Ling * callback again. 7603f9d6ad7SLin Ling */ 7613f9d6ad7SLin Ling if (ddt_class_contains(dp->dp_spa, 7623f9d6ad7SLin Ling scn->scn_phys.scn_ddt_class_max, bp)) { 7633f9d6ad7SLin Ling ASSERT(buf == NULL); 7643f9d6ad7SLin Ling return; 7653f9d6ad7SLin Ling } 7663f9d6ad7SLin Ling 7673f9d6ad7SLin Ling /* 7683f9d6ad7SLin Ling * If this block is from the future (after cur_max_txg), then we 7693f9d6ad7SLin Ling * are doing this on behalf of a deleted snapshot, and we will 7703f9d6ad7SLin Ling * revisit the future block on the next pass of this dataset. 7713f9d6ad7SLin Ling * Don't scan it now unless we need to because something 7723f9d6ad7SLin Ling * under it was modified. 7733f9d6ad7SLin Ling */ 774b4952e17SGeorge Wilson if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) { 7753f9d6ad7SLin Ling scan_funcs[scn->scn_phys.scn_func](dp, bp, zb); 7763f9d6ad7SLin Ling } 7773f9d6ad7SLin Ling if (buf) 7783f9d6ad7SLin Ling (void) arc_buf_remove_ref(buf, &buf); 7793f9d6ad7SLin Ling } 7803f9d6ad7SLin Ling 7813f9d6ad7SLin Ling static void 7823f9d6ad7SLin Ling dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp, 7833f9d6ad7SLin Ling dmu_tx_t *tx) 7843f9d6ad7SLin Ling { 7853f9d6ad7SLin Ling zbookmark_t zb; 7863f9d6ad7SLin Ling 7873f9d6ad7SLin Ling SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 7883f9d6ad7SLin Ling ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID); 7893f9d6ad7SLin Ling dsl_scan_visitbp(bp, &zb, NULL, NULL, 7903f9d6ad7SLin Ling ds, scn, DMU_OST_NONE, tx); 7913f9d6ad7SLin Ling 7923f9d6ad7SLin Ling dprintf_ds(ds, "finished scan%s", ""); 7933f9d6ad7SLin Ling } 7943f9d6ad7SLin Ling 7953f9d6ad7SLin Ling void 7963f9d6ad7SLin Ling dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx) 7973f9d6ad7SLin Ling { 7983f9d6ad7SLin Ling dsl_pool_t *dp = ds->ds_dir->dd_pool; 7993f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 8003f9d6ad7SLin Ling uint64_t mintxg; 8013f9d6ad7SLin Ling 8023f9d6ad7SLin Ling if (scn->scn_phys.scn_state != DSS_SCANNING) 8033f9d6ad7SLin Ling return; 8043f9d6ad7SLin Ling 8053f9d6ad7SLin Ling if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) { 8063f9d6ad7SLin Ling if (dsl_dataset_is_snapshot(ds)) { 8073f9d6ad7SLin Ling /* Note, scn_cur_{min,max}_txg stays the same. */ 8083f9d6ad7SLin Ling scn->scn_phys.scn_bookmark.zb_objset = 8093f9d6ad7SLin Ling ds->ds_phys->ds_next_snap_obj; 8103f9d6ad7SLin Ling zfs_dbgmsg("destroying ds %llu; currently traversing; " 8113f9d6ad7SLin Ling "reset zb_objset to %llu", 8123f9d6ad7SLin Ling (u_longlong_t)ds->ds_object, 8133f9d6ad7SLin Ling (u_longlong_t)ds->ds_phys->ds_next_snap_obj); 8143f9d6ad7SLin Ling scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN; 8153f9d6ad7SLin Ling } else { 8163f9d6ad7SLin Ling SET_BOOKMARK(&scn->scn_phys.scn_bookmark, 8173f9d6ad7SLin Ling ZB_DESTROYED_OBJSET, 0, 0, 0); 8183f9d6ad7SLin Ling zfs_dbgmsg("destroying ds %llu; currently traversing; " 8193f9d6ad7SLin Ling "reset bookmark to -1,0,0,0", 8203f9d6ad7SLin Ling (u_longlong_t)ds->ds_object); 8213f9d6ad7SLin Ling } 8223f9d6ad7SLin Ling } else if (zap_lookup_int_key(dp->dp_meta_objset, 8233f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) { 8243f9d6ad7SLin Ling ASSERT3U(ds->ds_phys->ds_num_children, <=, 1); 825b420f3adSRichard Lowe VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 8263f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 8273f9d6ad7SLin Ling if (dsl_dataset_is_snapshot(ds)) { 8283f9d6ad7SLin Ling /* 8293f9d6ad7SLin Ling * We keep the same mintxg; it could be > 8303f9d6ad7SLin Ling * ds_creation_txg if the previous snapshot was 8313f9d6ad7SLin Ling * deleted too. 8323f9d6ad7SLin Ling */ 8333f9d6ad7SLin Ling VERIFY(zap_add_int_key(dp->dp_meta_objset, 8343f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, 8353f9d6ad7SLin Ling ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0); 8363f9d6ad7SLin Ling zfs_dbgmsg("destroying ds %llu; in queue; " 8373f9d6ad7SLin Ling "replacing with %llu", 8383f9d6ad7SLin Ling (u_longlong_t)ds->ds_object, 8393f9d6ad7SLin Ling (u_longlong_t)ds->ds_phys->ds_next_snap_obj); 8403f9d6ad7SLin Ling } else { 8413f9d6ad7SLin Ling zfs_dbgmsg("destroying ds %llu; in queue; removing", 8423f9d6ad7SLin Ling (u_longlong_t)ds->ds_object); 8433f9d6ad7SLin Ling } 8443f9d6ad7SLin Ling } else { 8453f9d6ad7SLin Ling zfs_dbgmsg("destroying ds %llu; ignoring", 8463f9d6ad7SLin Ling (u_longlong_t)ds->ds_object); 8473f9d6ad7SLin Ling } 8483f9d6ad7SLin Ling 8493f9d6ad7SLin Ling /* 8503f9d6ad7SLin Ling * dsl_scan_sync() should be called after this, and should sync 8513f9d6ad7SLin Ling * out our changed state, but just to be safe, do it here. 8523f9d6ad7SLin Ling */ 8533f9d6ad7SLin Ling dsl_scan_sync_state(scn, tx); 8543f9d6ad7SLin Ling } 8553f9d6ad7SLin Ling 8563f9d6ad7SLin Ling void 8573f9d6ad7SLin Ling dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx) 8583f9d6ad7SLin Ling { 8593f9d6ad7SLin Ling dsl_pool_t *dp = ds->ds_dir->dd_pool; 8603f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 8613f9d6ad7SLin Ling uint64_t mintxg; 8623f9d6ad7SLin Ling 8633f9d6ad7SLin Ling if (scn->scn_phys.scn_state != DSS_SCANNING) 8643f9d6ad7SLin Ling return; 8653f9d6ad7SLin Ling 8663f9d6ad7SLin Ling ASSERT(ds->ds_phys->ds_prev_snap_obj != 0); 8673f9d6ad7SLin Ling 8683f9d6ad7SLin Ling if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) { 8693f9d6ad7SLin Ling scn->scn_phys.scn_bookmark.zb_objset = 8703f9d6ad7SLin Ling ds->ds_phys->ds_prev_snap_obj; 8713f9d6ad7SLin Ling zfs_dbgmsg("snapshotting ds %llu; currently traversing; " 8723f9d6ad7SLin Ling "reset zb_objset to %llu", 8733f9d6ad7SLin Ling (u_longlong_t)ds->ds_object, 8743f9d6ad7SLin Ling (u_longlong_t)ds->ds_phys->ds_prev_snap_obj); 8753f9d6ad7SLin Ling } else if (zap_lookup_int_key(dp->dp_meta_objset, 8763f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) { 877b420f3adSRichard Lowe VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 8783f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds->ds_object, tx)); 8793f9d6ad7SLin Ling VERIFY(zap_add_int_key(dp->dp_meta_objset, 8803f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, 8813f9d6ad7SLin Ling ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0); 8823f9d6ad7SLin Ling zfs_dbgmsg("snapshotting ds %llu; in queue; " 8833f9d6ad7SLin Ling "replacing with %llu", 8843f9d6ad7SLin Ling (u_longlong_t)ds->ds_object, 8853f9d6ad7SLin Ling (u_longlong_t)ds->ds_phys->ds_prev_snap_obj); 8863f9d6ad7SLin Ling } 8873f9d6ad7SLin Ling dsl_scan_sync_state(scn, tx); 8883f9d6ad7SLin Ling } 8893f9d6ad7SLin Ling 8903f9d6ad7SLin Ling void 8913f9d6ad7SLin Ling dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx) 8923f9d6ad7SLin Ling { 8933f9d6ad7SLin Ling dsl_pool_t *dp = ds1->ds_dir->dd_pool; 8943f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 8953f9d6ad7SLin Ling uint64_t mintxg; 8963f9d6ad7SLin Ling 8973f9d6ad7SLin Ling if (scn->scn_phys.scn_state != DSS_SCANNING) 8983f9d6ad7SLin Ling return; 8993f9d6ad7SLin Ling 9003f9d6ad7SLin Ling if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) { 9013f9d6ad7SLin Ling scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object; 9023f9d6ad7SLin Ling zfs_dbgmsg("clone_swap ds %llu; currently traversing; " 9033f9d6ad7SLin Ling "reset zb_objset to %llu", 9043f9d6ad7SLin Ling (u_longlong_t)ds1->ds_object, 9053f9d6ad7SLin Ling (u_longlong_t)ds2->ds_object); 9063f9d6ad7SLin Ling } else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) { 9073f9d6ad7SLin Ling scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object; 9083f9d6ad7SLin Ling zfs_dbgmsg("clone_swap ds %llu; currently traversing; " 9093f9d6ad7SLin Ling "reset zb_objset to %llu", 9103f9d6ad7SLin Ling (u_longlong_t)ds2->ds_object, 9113f9d6ad7SLin Ling (u_longlong_t)ds1->ds_object); 9123f9d6ad7SLin Ling } 9133f9d6ad7SLin Ling 9143f9d6ad7SLin Ling if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 9153f9d6ad7SLin Ling ds1->ds_object, &mintxg) == 0) { 9163f9d6ad7SLin Ling int err; 9173f9d6ad7SLin Ling 9183f9d6ad7SLin Ling ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg); 9193f9d6ad7SLin Ling ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg); 920b420f3adSRichard Lowe VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 9213f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds1->ds_object, tx)); 9223f9d6ad7SLin Ling err = zap_add_int_key(dp->dp_meta_objset, 9233f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx); 9243f9d6ad7SLin Ling VERIFY(err == 0 || err == EEXIST); 9253f9d6ad7SLin Ling if (err == EEXIST) { 9263f9d6ad7SLin Ling /* Both were there to begin with */ 9273f9d6ad7SLin Ling VERIFY(0 == zap_add_int_key(dp->dp_meta_objset, 9283f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, 9293f9d6ad7SLin Ling ds1->ds_object, mintxg, tx)); 9303f9d6ad7SLin Ling } 9313f9d6ad7SLin Ling zfs_dbgmsg("clone_swap ds %llu; in queue; " 9323f9d6ad7SLin Ling "replacing with %llu", 9333f9d6ad7SLin Ling (u_longlong_t)ds1->ds_object, 9343f9d6ad7SLin Ling (u_longlong_t)ds2->ds_object); 9353f9d6ad7SLin Ling } else if (zap_lookup_int_key(dp->dp_meta_objset, 9363f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) { 9373f9d6ad7SLin Ling ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg); 9383f9d6ad7SLin Ling ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg); 939b420f3adSRichard Lowe VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 9403f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds2->ds_object, tx)); 9413f9d6ad7SLin Ling VERIFY(0 == zap_add_int_key(dp->dp_meta_objset, 9423f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx)); 9433f9d6ad7SLin Ling zfs_dbgmsg("clone_swap ds %llu; in queue; " 9443f9d6ad7SLin Ling "replacing with %llu", 9453f9d6ad7SLin Ling (u_longlong_t)ds2->ds_object, 9463f9d6ad7SLin Ling (u_longlong_t)ds1->ds_object); 9473f9d6ad7SLin Ling } 9483f9d6ad7SLin Ling 9493f9d6ad7SLin Ling dsl_scan_sync_state(scn, tx); 9503f9d6ad7SLin Ling } 9513f9d6ad7SLin Ling 9523f9d6ad7SLin Ling struct enqueue_clones_arg { 9533f9d6ad7SLin Ling dmu_tx_t *tx; 9543f9d6ad7SLin Ling uint64_t originobj; 9553f9d6ad7SLin Ling }; 9563f9d6ad7SLin Ling 9573f9d6ad7SLin Ling /* ARGSUSED */ 9583f9d6ad7SLin Ling static int 9593b2aab18SMatthew Ahrens enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 9603f9d6ad7SLin Ling { 9613f9d6ad7SLin Ling struct enqueue_clones_arg *eca = arg; 9623f9d6ad7SLin Ling dsl_dataset_t *ds; 9633f9d6ad7SLin Ling int err; 9643f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 9653f9d6ad7SLin Ling 9663b2aab18SMatthew Ahrens if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj) 9673b2aab18SMatthew Ahrens return (0); 9683b2aab18SMatthew Ahrens 9693b2aab18SMatthew Ahrens err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 9703f9d6ad7SLin Ling if (err) 9713f9d6ad7SLin Ling return (err); 9723f9d6ad7SLin Ling 9733b2aab18SMatthew Ahrens while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) { 9743b2aab18SMatthew Ahrens dsl_dataset_t *prev; 9753b2aab18SMatthew Ahrens err = dsl_dataset_hold_obj(dp, 9763b2aab18SMatthew Ahrens ds->ds_phys->ds_prev_snap_obj, FTAG, &prev); 9773f9d6ad7SLin Ling 9783b2aab18SMatthew Ahrens dsl_dataset_rele(ds, FTAG); 9793b2aab18SMatthew Ahrens if (err) 9803b2aab18SMatthew Ahrens return (err); 9813b2aab18SMatthew Ahrens ds = prev; 9823f9d6ad7SLin Ling } 9833b2aab18SMatthew Ahrens VERIFY(zap_add_int_key(dp->dp_meta_objset, 9843b2aab18SMatthew Ahrens scn->scn_phys.scn_queue_obj, ds->ds_object, 9853b2aab18SMatthew Ahrens ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0); 9863f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 9873f9d6ad7SLin Ling return (0); 9883f9d6ad7SLin Ling } 9893f9d6ad7SLin Ling 9903f9d6ad7SLin Ling static void 9913f9d6ad7SLin Ling dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx) 9923f9d6ad7SLin Ling { 9933f9d6ad7SLin Ling dsl_pool_t *dp = scn->scn_dp; 9943f9d6ad7SLin Ling dsl_dataset_t *ds; 9956e0cbcaaSMatthew Ahrens objset_t *os; 9963f9d6ad7SLin Ling 997b420f3adSRichard Lowe VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 9983f9d6ad7SLin Ling 9996e0cbcaaSMatthew Ahrens if (dmu_objset_from_ds(ds, &os)) 10006e0cbcaaSMatthew Ahrens goto out; 10016e0cbcaaSMatthew Ahrens 10026e0cbcaaSMatthew Ahrens /* 10036e0cbcaaSMatthew Ahrens * Only the ZIL in the head (non-snapshot) is valid. Even though 10046e0cbcaaSMatthew Ahrens * snapshots can have ZIL block pointers (which may be the same 10056e0cbcaaSMatthew Ahrens * BP as in the head), they must be ignored. So we traverse the 10066e0cbcaaSMatthew Ahrens * ZIL here, rather than in scan_recurse(), because the regular 10076e0cbcaaSMatthew Ahrens * snapshot block-sharing rules don't apply to it. 10086e0cbcaaSMatthew Ahrens */ 10096e0cbcaaSMatthew Ahrens if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds)) 10106e0cbcaaSMatthew Ahrens dsl_scan_zil(dp, &os->os_zil_header); 10116e0cbcaaSMatthew Ahrens 10123f9d6ad7SLin Ling /* 10133f9d6ad7SLin Ling * Iterate over the bps in this ds. 10143f9d6ad7SLin Ling */ 10153f9d6ad7SLin Ling dmu_buf_will_dirty(ds->ds_dbuf, tx); 10163f9d6ad7SLin Ling dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx); 10173f9d6ad7SLin Ling 10183f9d6ad7SLin Ling char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP); 10193f9d6ad7SLin Ling dsl_dataset_name(ds, dsname); 10203f9d6ad7SLin Ling zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; " 10213f9d6ad7SLin Ling "pausing=%u", 10223f9d6ad7SLin Ling (longlong_t)dsobj, dsname, 10233f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_cur_min_txg, 10243f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_cur_max_txg, 10253f9d6ad7SLin Ling (int)scn->scn_pausing); 10263f9d6ad7SLin Ling kmem_free(dsname, ZFS_MAXNAMELEN); 10273f9d6ad7SLin Ling 10283f9d6ad7SLin Ling if (scn->scn_pausing) 10293f9d6ad7SLin Ling goto out; 10303f9d6ad7SLin Ling 10313f9d6ad7SLin Ling /* 10323f9d6ad7SLin Ling * We've finished this pass over this dataset. 10333f9d6ad7SLin Ling */ 10343f9d6ad7SLin Ling 10353f9d6ad7SLin Ling /* 10363f9d6ad7SLin Ling * If we did not completely visit this dataset, do another pass. 10373f9d6ad7SLin Ling */ 10383f9d6ad7SLin Ling if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) { 10393f9d6ad7SLin Ling zfs_dbgmsg("incomplete pass; visiting again"); 10403f9d6ad7SLin Ling scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN; 10413f9d6ad7SLin Ling VERIFY(zap_add_int_key(dp->dp_meta_objset, 10423f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds->ds_object, 10433f9d6ad7SLin Ling scn->scn_phys.scn_cur_max_txg, tx) == 0); 10443f9d6ad7SLin Ling goto out; 10453f9d6ad7SLin Ling } 10463f9d6ad7SLin Ling 10473f9d6ad7SLin Ling /* 10483f9d6ad7SLin Ling * Add descendent datasets to work queue. 10493f9d6ad7SLin Ling */ 10503f9d6ad7SLin Ling if (ds->ds_phys->ds_next_snap_obj != 0) { 10513f9d6ad7SLin Ling VERIFY(zap_add_int_key(dp->dp_meta_objset, 10523f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj, 10533f9d6ad7SLin Ling ds->ds_phys->ds_creation_txg, tx) == 0); 10543f9d6ad7SLin Ling } 10553f9d6ad7SLin Ling if (ds->ds_phys->ds_num_children > 1) { 10563f9d6ad7SLin Ling boolean_t usenext = B_FALSE; 10573f9d6ad7SLin Ling if (ds->ds_phys->ds_next_clones_obj != 0) { 10583f9d6ad7SLin Ling uint64_t count; 10593f9d6ad7SLin Ling /* 10603f9d6ad7SLin Ling * A bug in a previous version of the code could 10613f9d6ad7SLin Ling * cause upgrade_clones_cb() to not set 10623f9d6ad7SLin Ling * ds_next_snap_obj when it should, leading to a 10633f9d6ad7SLin Ling * missing entry. Therefore we can only use the 10643f9d6ad7SLin Ling * next_clones_obj when its count is correct. 10653f9d6ad7SLin Ling */ 10663f9d6ad7SLin Ling int err = zap_count(dp->dp_meta_objset, 10673f9d6ad7SLin Ling ds->ds_phys->ds_next_clones_obj, &count); 10683f9d6ad7SLin Ling if (err == 0 && 10693f9d6ad7SLin Ling count == ds->ds_phys->ds_num_children - 1) 10703f9d6ad7SLin Ling usenext = B_TRUE; 10713f9d6ad7SLin Ling } 10723f9d6ad7SLin Ling 10733f9d6ad7SLin Ling if (usenext) { 10743b2aab18SMatthew Ahrens VERIFY0(zap_join_key(dp->dp_meta_objset, 10753f9d6ad7SLin Ling ds->ds_phys->ds_next_clones_obj, 10763f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, 10773b2aab18SMatthew Ahrens ds->ds_phys->ds_creation_txg, tx)); 10783f9d6ad7SLin Ling } else { 10793f9d6ad7SLin Ling struct enqueue_clones_arg eca; 10803f9d6ad7SLin Ling eca.tx = tx; 10813f9d6ad7SLin Ling eca.originobj = ds->ds_object; 10823f9d6ad7SLin Ling 10833b2aab18SMatthew Ahrens VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 10843b2aab18SMatthew Ahrens enqueue_clones_cb, &eca, DS_FIND_CHILDREN)); 10853f9d6ad7SLin Ling } 10863f9d6ad7SLin Ling } 10873f9d6ad7SLin Ling 10883f9d6ad7SLin Ling out: 10893f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 10903f9d6ad7SLin Ling } 10913f9d6ad7SLin Ling 10923f9d6ad7SLin Ling /* ARGSUSED */ 10933f9d6ad7SLin Ling static int 10943b2aab18SMatthew Ahrens enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg) 10953f9d6ad7SLin Ling { 10963f9d6ad7SLin Ling dmu_tx_t *tx = arg; 10973f9d6ad7SLin Ling dsl_dataset_t *ds; 10983f9d6ad7SLin Ling int err; 10993f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 11003f9d6ad7SLin Ling 11013b2aab18SMatthew Ahrens err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds); 11023f9d6ad7SLin Ling if (err) 11033f9d6ad7SLin Ling return (err); 11043f9d6ad7SLin Ling 11053f9d6ad7SLin Ling while (ds->ds_phys->ds_prev_snap_obj != 0) { 11063f9d6ad7SLin Ling dsl_dataset_t *prev; 11073f9d6ad7SLin Ling err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj, 11083f9d6ad7SLin Ling FTAG, &prev); 11093f9d6ad7SLin Ling if (err) { 11103f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 11113f9d6ad7SLin Ling return (err); 11123f9d6ad7SLin Ling } 11133f9d6ad7SLin Ling 11143f9d6ad7SLin Ling /* 11153f9d6ad7SLin Ling * If this is a clone, we don't need to worry about it for now. 11163f9d6ad7SLin Ling */ 11173f9d6ad7SLin Ling if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) { 11183f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 11193f9d6ad7SLin Ling dsl_dataset_rele(prev, FTAG); 11203f9d6ad7SLin Ling return (0); 11213f9d6ad7SLin Ling } 11223f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 11233f9d6ad7SLin Ling ds = prev; 11243f9d6ad7SLin Ling } 11253f9d6ad7SLin Ling 11263f9d6ad7SLin Ling VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj, 11273f9d6ad7SLin Ling ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0); 11283f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 11293f9d6ad7SLin Ling return (0); 11303f9d6ad7SLin Ling } 11313f9d6ad7SLin Ling 11323f9d6ad7SLin Ling /* 11333f9d6ad7SLin Ling * Scrub/dedup interaction. 11343f9d6ad7SLin Ling * 11353f9d6ad7SLin Ling * If there are N references to a deduped block, we don't want to scrub it 11363f9d6ad7SLin Ling * N times -- ideally, we should scrub it exactly once. 11373f9d6ad7SLin Ling * 11383f9d6ad7SLin Ling * We leverage the fact that the dde's replication class (enum ddt_class) 11393f9d6ad7SLin Ling * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest 11403f9d6ad7SLin Ling * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order. 11413f9d6ad7SLin Ling * 11423f9d6ad7SLin Ling * To prevent excess scrubbing, the scrub begins by walking the DDT 11433f9d6ad7SLin Ling * to find all blocks with refcnt > 1, and scrubs each of these once. 11443f9d6ad7SLin Ling * Since there are two replication classes which contain blocks with 11453f9d6ad7SLin Ling * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first. 11463f9d6ad7SLin Ling * Finally the top-down scrub begins, only visiting blocks with refcnt == 1. 11473f9d6ad7SLin Ling * 11483f9d6ad7SLin Ling * There would be nothing more to say if a block's refcnt couldn't change 11493f9d6ad7SLin Ling * during a scrub, but of course it can so we must account for changes 11503f9d6ad7SLin Ling * in a block's replication class. 11513f9d6ad7SLin Ling * 11523f9d6ad7SLin Ling * Here's an example of what can occur: 11533f9d6ad7SLin Ling * 11543f9d6ad7SLin Ling * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1 11553f9d6ad7SLin Ling * when visited during the top-down scrub phase, it will be scrubbed twice. 11563f9d6ad7SLin Ling * This negates our scrub optimization, but is otherwise harmless. 11573f9d6ad7SLin Ling * 11583f9d6ad7SLin Ling * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1 11593f9d6ad7SLin Ling * on each visit during the top-down scrub phase, it will never be scrubbed. 11603f9d6ad7SLin Ling * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's 11613f9d6ad7SLin Ling * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to 11623f9d6ad7SLin Ling * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1 11633f9d6ad7SLin Ling * while a scrub is in progress, it scrubs the block right then. 11643f9d6ad7SLin Ling */ 11653f9d6ad7SLin Ling static void 11663f9d6ad7SLin Ling dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx) 11673f9d6ad7SLin Ling { 11683f9d6ad7SLin Ling ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark; 11693f9d6ad7SLin Ling ddt_entry_t dde = { 0 }; 11703f9d6ad7SLin Ling int error; 11713f9d6ad7SLin Ling uint64_t n = 0; 11723f9d6ad7SLin Ling 11733f9d6ad7SLin Ling while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) { 11743f9d6ad7SLin Ling ddt_t *ddt; 11753f9d6ad7SLin Ling 11763f9d6ad7SLin Ling if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max) 11773f9d6ad7SLin Ling break; 11783f9d6ad7SLin Ling dprintf("visiting ddb=%llu/%llu/%llu/%llx\n", 11793f9d6ad7SLin Ling (longlong_t)ddb->ddb_class, 11803f9d6ad7SLin Ling (longlong_t)ddb->ddb_type, 11813f9d6ad7SLin Ling (longlong_t)ddb->ddb_checksum, 11823f9d6ad7SLin Ling (longlong_t)ddb->ddb_cursor); 11833f9d6ad7SLin Ling 11843f9d6ad7SLin Ling /* There should be no pending changes to the dedup table */ 11853f9d6ad7SLin Ling ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum]; 11863f9d6ad7SLin Ling ASSERT(avl_first(&ddt->ddt_tree) == NULL); 11873f9d6ad7SLin Ling 11883f9d6ad7SLin Ling dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx); 11893f9d6ad7SLin Ling n++; 11903f9d6ad7SLin Ling 11913f9d6ad7SLin Ling if (dsl_scan_check_pause(scn, NULL)) 11923f9d6ad7SLin Ling break; 11933f9d6ad7SLin Ling } 11943f9d6ad7SLin Ling 11953f9d6ad7SLin Ling zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u", 11963f9d6ad7SLin Ling (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max, 11973f9d6ad7SLin Ling (int)scn->scn_pausing); 11983f9d6ad7SLin Ling 11993f9d6ad7SLin Ling ASSERT(error == 0 || error == ENOENT); 12003f9d6ad7SLin Ling ASSERT(error != ENOENT || 12013f9d6ad7SLin Ling ddb->ddb_class > scn->scn_phys.scn_ddt_class_max); 12023f9d6ad7SLin Ling } 12033f9d6ad7SLin Ling 12043f9d6ad7SLin Ling /* ARGSUSED */ 12053f9d6ad7SLin Ling void 12063f9d6ad7SLin Ling dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum, 12073f9d6ad7SLin Ling ddt_entry_t *dde, dmu_tx_t *tx) 12083f9d6ad7SLin Ling { 12093f9d6ad7SLin Ling const ddt_key_t *ddk = &dde->dde_key; 12103f9d6ad7SLin Ling ddt_phys_t *ddp = dde->dde_phys; 12113f9d6ad7SLin Ling blkptr_t bp; 12123f9d6ad7SLin Ling zbookmark_t zb = { 0 }; 12133f9d6ad7SLin Ling 12143f9d6ad7SLin Ling if (scn->scn_phys.scn_state != DSS_SCANNING) 12153f9d6ad7SLin Ling return; 12163f9d6ad7SLin Ling 12173f9d6ad7SLin Ling for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 12183f9d6ad7SLin Ling if (ddp->ddp_phys_birth == 0 || 1219b4952e17SGeorge Wilson ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg) 12203f9d6ad7SLin Ling continue; 12213f9d6ad7SLin Ling ddt_bp_create(checksum, ddk, ddp, &bp); 12223f9d6ad7SLin Ling 12233f9d6ad7SLin Ling scn->scn_visited_this_txg++; 12243f9d6ad7SLin Ling scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb); 12253f9d6ad7SLin Ling } 12263f9d6ad7SLin Ling } 12273f9d6ad7SLin Ling 12283f9d6ad7SLin Ling static void 12293f9d6ad7SLin Ling dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx) 12303f9d6ad7SLin Ling { 12313f9d6ad7SLin Ling dsl_pool_t *dp = scn->scn_dp; 12323f9d6ad7SLin Ling zap_cursor_t zc; 12333f9d6ad7SLin Ling zap_attribute_t za; 12343f9d6ad7SLin Ling 12353f9d6ad7SLin Ling if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 12363f9d6ad7SLin Ling scn->scn_phys.scn_ddt_class_max) { 12373f9d6ad7SLin Ling scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 12383f9d6ad7SLin Ling scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 12393f9d6ad7SLin Ling dsl_scan_ddt(scn, tx); 12403f9d6ad7SLin Ling if (scn->scn_pausing) 12413f9d6ad7SLin Ling return; 12423f9d6ad7SLin Ling } 12433f9d6ad7SLin Ling 12443f9d6ad7SLin Ling if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) { 12453f9d6ad7SLin Ling /* First do the MOS & ORIGIN */ 12463f9d6ad7SLin Ling 12473f9d6ad7SLin Ling scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg; 12483f9d6ad7SLin Ling scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg; 12493f9d6ad7SLin Ling dsl_scan_visit_rootbp(scn, NULL, 12503f9d6ad7SLin Ling &dp->dp_meta_rootbp, tx); 12513f9d6ad7SLin Ling spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp); 12523f9d6ad7SLin Ling if (scn->scn_pausing) 12533f9d6ad7SLin Ling return; 12543f9d6ad7SLin Ling 12553f9d6ad7SLin Ling if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) { 12563b2aab18SMatthew Ahrens VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 12573b2aab18SMatthew Ahrens enqueue_cb, tx, DS_FIND_CHILDREN)); 12583f9d6ad7SLin Ling } else { 12593f9d6ad7SLin Ling dsl_scan_visitds(scn, 12603f9d6ad7SLin Ling dp->dp_origin_snap->ds_object, tx); 12613f9d6ad7SLin Ling } 12623f9d6ad7SLin Ling ASSERT(!scn->scn_pausing); 12633f9d6ad7SLin Ling } else if (scn->scn_phys.scn_bookmark.zb_objset != 12643f9d6ad7SLin Ling ZB_DESTROYED_OBJSET) { 12653f9d6ad7SLin Ling /* 12663f9d6ad7SLin Ling * If we were paused, continue from here. Note if the 12673f9d6ad7SLin Ling * ds we were paused on was deleted, the zb_objset may 12683f9d6ad7SLin Ling * be -1, so we will skip this and find a new objset 12693f9d6ad7SLin Ling * below. 12703f9d6ad7SLin Ling */ 12713f9d6ad7SLin Ling dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx); 12723f9d6ad7SLin Ling if (scn->scn_pausing) 12733f9d6ad7SLin Ling return; 12743f9d6ad7SLin Ling } 12753f9d6ad7SLin Ling 12763f9d6ad7SLin Ling /* 12773f9d6ad7SLin Ling * In case we were paused right at the end of the ds, zero the 12783f9d6ad7SLin Ling * bookmark so we don't think that we're still trying to resume. 12793f9d6ad7SLin Ling */ 12803f9d6ad7SLin Ling bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_t)); 12813f9d6ad7SLin Ling 12823f9d6ad7SLin Ling /* keep pulling things out of the zap-object-as-queue */ 12833f9d6ad7SLin Ling while (zap_cursor_init(&zc, dp->dp_meta_objset, 12843f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj), 12853f9d6ad7SLin Ling zap_cursor_retrieve(&zc, &za) == 0) { 12863f9d6ad7SLin Ling dsl_dataset_t *ds; 12873f9d6ad7SLin Ling uint64_t dsobj; 12883f9d6ad7SLin Ling 12893f9d6ad7SLin Ling dsobj = strtonum(za.za_name, NULL); 1290b420f3adSRichard Lowe VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset, 12913f9d6ad7SLin Ling scn->scn_phys.scn_queue_obj, dsobj, tx)); 12923f9d6ad7SLin Ling 12933f9d6ad7SLin Ling /* Set up min/max txg */ 1294b420f3adSRichard Lowe VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds)); 12953f9d6ad7SLin Ling if (za.za_first_integer != 0) { 12963f9d6ad7SLin Ling scn->scn_phys.scn_cur_min_txg = 12973f9d6ad7SLin Ling MAX(scn->scn_phys.scn_min_txg, 12983f9d6ad7SLin Ling za.za_first_integer); 12993f9d6ad7SLin Ling } else { 13003f9d6ad7SLin Ling scn->scn_phys.scn_cur_min_txg = 13013f9d6ad7SLin Ling MAX(scn->scn_phys.scn_min_txg, 13023f9d6ad7SLin Ling ds->ds_phys->ds_prev_snap_txg); 13033f9d6ad7SLin Ling } 13043f9d6ad7SLin Ling scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds); 13053f9d6ad7SLin Ling dsl_dataset_rele(ds, FTAG); 13063f9d6ad7SLin Ling 13073f9d6ad7SLin Ling dsl_scan_visitds(scn, dsobj, tx); 13083f9d6ad7SLin Ling zap_cursor_fini(&zc); 13093f9d6ad7SLin Ling if (scn->scn_pausing) 13103f9d6ad7SLin Ling return; 13113f9d6ad7SLin Ling } 13123f9d6ad7SLin Ling zap_cursor_fini(&zc); 13133f9d6ad7SLin Ling } 13143f9d6ad7SLin Ling 1315ad135b5dSChristopher Siden static boolean_t 1316ad135b5dSChristopher Siden dsl_scan_free_should_pause(dsl_scan_t *scn) 1317cde58dbcSMatthew Ahrens { 1318cde58dbcSMatthew Ahrens uint64_t elapsed_nanosecs; 1319cde58dbcSMatthew Ahrens 13208b36997aSMatthew Ahrens if (zfs_recover) 13218b36997aSMatthew Ahrens return (B_FALSE); 13228b36997aSMatthew Ahrens 1323cde58dbcSMatthew Ahrens elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time; 1324ad135b5dSChristopher Siden return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout || 13250689f76cSAdam Leventhal (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms && 1326cde58dbcSMatthew Ahrens txg_sync_waiting(scn->scn_dp)) || 1327ad135b5dSChristopher Siden spa_shutting_down(scn->scn_dp->dp_spa)); 1328ad135b5dSChristopher Siden } 1329ad135b5dSChristopher Siden 1330ad135b5dSChristopher Siden static int 1331ad135b5dSChristopher Siden dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 1332ad135b5dSChristopher Siden { 1333ad135b5dSChristopher Siden dsl_scan_t *scn = arg; 1334ad135b5dSChristopher Siden 1335ad135b5dSChristopher Siden if (!scn->scn_is_bptree || 1336ad135b5dSChristopher Siden (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) { 1337ad135b5dSChristopher Siden if (dsl_scan_free_should_pause(scn)) 1338be6fd75aSMatthew Ahrens return (SET_ERROR(ERESTART)); 1339ad135b5dSChristopher Siden } 1340cde58dbcSMatthew Ahrens 1341cde58dbcSMatthew Ahrens zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa, 1342cde58dbcSMatthew Ahrens dmu_tx_get_txg(tx), bp, 0)); 1343cde58dbcSMatthew Ahrens dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, 1344cde58dbcSMatthew Ahrens -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp), 1345cde58dbcSMatthew Ahrens -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); 1346cde58dbcSMatthew Ahrens scn->scn_visited_this_txg++; 1347cde58dbcSMatthew Ahrens return (0); 1348cde58dbcSMatthew Ahrens } 1349cde58dbcSMatthew Ahrens 1350cde58dbcSMatthew Ahrens boolean_t 1351cde58dbcSMatthew Ahrens dsl_scan_active(dsl_scan_t *scn) 1352cde58dbcSMatthew Ahrens { 1353cde58dbcSMatthew Ahrens spa_t *spa = scn->scn_dp->dp_spa; 1354cde58dbcSMatthew Ahrens uint64_t used = 0, comp, uncomp; 1355cde58dbcSMatthew Ahrens 1356cde58dbcSMatthew Ahrens if (spa->spa_load_state != SPA_LOAD_NONE) 1357cde58dbcSMatthew Ahrens return (B_FALSE); 1358cde58dbcSMatthew Ahrens if (spa_shutting_down(spa)) 1359cde58dbcSMatthew Ahrens return (B_FALSE); 13604a923759SGeorge Wilson if (scn->scn_phys.scn_state == DSS_SCANNING || 13614a923759SGeorge Wilson scn->scn_async_destroying) 1362cde58dbcSMatthew Ahrens return (B_TRUE); 1363cde58dbcSMatthew Ahrens 1364cde58dbcSMatthew Ahrens if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 1365cde58dbcSMatthew Ahrens (void) bpobj_space(&scn->scn_dp->dp_free_bpobj, 1366cde58dbcSMatthew Ahrens &used, &comp, &uncomp); 1367cde58dbcSMatthew Ahrens } 1368cde58dbcSMatthew Ahrens return (used != 0); 1369cde58dbcSMatthew Ahrens } 1370cde58dbcSMatthew Ahrens 13713f9d6ad7SLin Ling void 13723f9d6ad7SLin Ling dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx) 13733f9d6ad7SLin Ling { 13743f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 13753f9d6ad7SLin Ling spa_t *spa = dp->dp_spa; 1376cde58dbcSMatthew Ahrens int err; 13773f9d6ad7SLin Ling 13783f9d6ad7SLin Ling /* 13793f9d6ad7SLin Ling * Check for scn_restart_txg before checking spa_load_state, so 13803f9d6ad7SLin Ling * that we can restart an old-style scan while the pool is being 13813f9d6ad7SLin Ling * imported (see dsl_scan_init). 13823f9d6ad7SLin Ling */ 13833f9d6ad7SLin Ling if (scn->scn_restart_txg != 0 && 13843f9d6ad7SLin Ling scn->scn_restart_txg <= tx->tx_txg) { 13853f9d6ad7SLin Ling pool_scan_func_t func = POOL_SCAN_SCRUB; 13863f9d6ad7SLin Ling dsl_scan_done(scn, B_FALSE, tx); 13873f9d6ad7SLin Ling if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) 13883f9d6ad7SLin Ling func = POOL_SCAN_RESILVER; 13893f9d6ad7SLin Ling zfs_dbgmsg("restarting scan func=%u txg=%llu", 13903f9d6ad7SLin Ling func, tx->tx_txg); 13913b2aab18SMatthew Ahrens dsl_scan_setup_sync(&func, tx); 13923f9d6ad7SLin Ling } 13933f9d6ad7SLin Ling 1394cde58dbcSMatthew Ahrens if (!dsl_scan_active(scn) || 1395cde58dbcSMatthew Ahrens spa_sync_pass(dp->dp_spa) > 1) 13963f9d6ad7SLin Ling return; 13973f9d6ad7SLin Ling 13983f9d6ad7SLin Ling scn->scn_visited_this_txg = 0; 13993f9d6ad7SLin Ling scn->scn_pausing = B_FALSE; 14003f9d6ad7SLin Ling scn->scn_sync_start_time = gethrtime(); 14013f9d6ad7SLin Ling spa->spa_scrub_active = B_TRUE; 14023f9d6ad7SLin Ling 1403cde58dbcSMatthew Ahrens /* 1404cde58dbcSMatthew Ahrens * First process the free list. If we pause the free, don't do 1405cde58dbcSMatthew Ahrens * any scanning. This ensures that there is no free list when 1406cde58dbcSMatthew Ahrens * we are scanning, so the scan code doesn't have to worry about 1407cde58dbcSMatthew Ahrens * traversing it. 1408cde58dbcSMatthew Ahrens */ 1409cde58dbcSMatthew Ahrens if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) { 1410ad135b5dSChristopher Siden scn->scn_is_bptree = B_FALSE; 1411cde58dbcSMatthew Ahrens scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1412cde58dbcSMatthew Ahrens NULL, ZIO_FLAG_MUSTSUCCEED); 1413cde58dbcSMatthew Ahrens err = bpobj_iterate(&dp->dp_free_bpobj, 1414ad135b5dSChristopher Siden dsl_scan_free_block_cb, scn, tx); 1415b420f3adSRichard Lowe VERIFY3U(0, ==, zio_wait(scn->scn_zio_root)); 1416ad135b5dSChristopher Siden 1417ad135b5dSChristopher Siden if (err == 0 && spa_feature_is_active(spa, 14182acef22dSMatthew Ahrens SPA_FEATURE_ASYNC_DESTROY)) { 14194a923759SGeorge Wilson ASSERT(scn->scn_async_destroying); 1420ad135b5dSChristopher Siden scn->scn_is_bptree = B_TRUE; 1421ad135b5dSChristopher Siden scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 1422ad135b5dSChristopher Siden NULL, ZIO_FLAG_MUSTSUCCEED); 1423ad135b5dSChristopher Siden err = bptree_iterate(dp->dp_meta_objset, 1424ad135b5dSChristopher Siden dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, 1425ad135b5dSChristopher Siden scn, tx); 14263b2aab18SMatthew Ahrens VERIFY0(zio_wait(scn->scn_zio_root)); 14273b2aab18SMatthew Ahrens 14283b2aab18SMatthew Ahrens if (err == 0) { 14293b2aab18SMatthew Ahrens /* finished; deactivate async destroy feature */ 14302acef22dSMatthew Ahrens spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, 14312acef22dSMatthew Ahrens tx); 14322acef22dSMatthew Ahrens ASSERT(!spa_feature_is_active(spa, 14332acef22dSMatthew Ahrens SPA_FEATURE_ASYNC_DESTROY)); 14343b2aab18SMatthew Ahrens VERIFY0(zap_remove(dp->dp_meta_objset, 14353b2aab18SMatthew Ahrens DMU_POOL_DIRECTORY_OBJECT, 14363b2aab18SMatthew Ahrens DMU_POOL_BPTREE_OBJ, tx)); 14373b2aab18SMatthew Ahrens VERIFY0(bptree_free(dp->dp_meta_objset, 14383b2aab18SMatthew Ahrens dp->dp_bptree_obj, tx)); 14393b2aab18SMatthew Ahrens dp->dp_bptree_obj = 0; 14404a923759SGeorge Wilson scn->scn_async_destroying = B_FALSE; 14413b2aab18SMatthew Ahrens } 1442ad135b5dSChristopher Siden } 1443cde58dbcSMatthew Ahrens if (scn->scn_visited_this_txg) { 1444cde58dbcSMatthew Ahrens zfs_dbgmsg("freed %llu blocks in %llums from " 1445ad135b5dSChristopher Siden "free_bpobj/bptree txg %llu", 1446cde58dbcSMatthew Ahrens (longlong_t)scn->scn_visited_this_txg, 1447cde58dbcSMatthew Ahrens (longlong_t) 14480689f76cSAdam Leventhal NSEC2MSEC(gethrtime() - scn->scn_sync_start_time), 1449cde58dbcSMatthew Ahrens (longlong_t)tx->tx_txg); 1450cde58dbcSMatthew Ahrens scn->scn_visited_this_txg = 0; 1451cde58dbcSMatthew Ahrens /* 1452cde58dbcSMatthew Ahrens * Re-sync the ddt so that we can further modify 1453cde58dbcSMatthew Ahrens * it when doing bprewrite. 1454cde58dbcSMatthew Ahrens */ 1455cde58dbcSMatthew Ahrens ddt_sync(spa, tx->tx_txg); 1456cde58dbcSMatthew Ahrens } 1457cde58dbcSMatthew Ahrens if (err == ERESTART) 1458cde58dbcSMatthew Ahrens return; 1459*5d7b4d43SMatthew Ahrens /* finished; verify that space accounting went to zero */ 1460*5d7b4d43SMatthew Ahrens ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes); 1461*5d7b4d43SMatthew Ahrens ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes); 1462*5d7b4d43SMatthew Ahrens ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes); 1463cde58dbcSMatthew Ahrens } 1464cde58dbcSMatthew Ahrens 1465cde58dbcSMatthew Ahrens if (scn->scn_phys.scn_state != DSS_SCANNING) 1466cde58dbcSMatthew Ahrens return; 1467cde58dbcSMatthew Ahrens 1468b4952e17SGeorge Wilson if (scn->scn_done_txg == tx->tx_txg) { 1469b4952e17SGeorge Wilson ASSERT(!scn->scn_pausing); 1470b4952e17SGeorge Wilson /* finished with scan. */ 1471b4952e17SGeorge Wilson zfs_dbgmsg("txg %llu scan complete", tx->tx_txg); 1472b4952e17SGeorge Wilson dsl_scan_done(scn, B_TRUE, tx); 1473b4952e17SGeorge Wilson ASSERT3U(spa->spa_scrub_inflight, ==, 0); 1474b4952e17SGeorge Wilson dsl_scan_sync_state(scn, tx); 1475b4952e17SGeorge Wilson return; 1476b4952e17SGeorge Wilson } 1477b4952e17SGeorge Wilson 14783f9d6ad7SLin Ling if (scn->scn_phys.scn_ddt_bookmark.ddb_class <= 14793f9d6ad7SLin Ling scn->scn_phys.scn_ddt_class_max) { 14803f9d6ad7SLin Ling zfs_dbgmsg("doing scan sync txg %llu; " 14813f9d6ad7SLin Ling "ddt bm=%llu/%llu/%llu/%llx", 14823f9d6ad7SLin Ling (longlong_t)tx->tx_txg, 14833f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class, 14843f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type, 14853f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum, 14863f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor); 14873f9d6ad7SLin Ling ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0); 14883f9d6ad7SLin Ling ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0); 14893f9d6ad7SLin Ling ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0); 14903f9d6ad7SLin Ling ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0); 14913f9d6ad7SLin Ling } else { 14923f9d6ad7SLin Ling zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu", 14933f9d6ad7SLin Ling (longlong_t)tx->tx_txg, 14943f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_bookmark.zb_objset, 14953f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_bookmark.zb_object, 14963f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_bookmark.zb_level, 14973f9d6ad7SLin Ling (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid); 14983f9d6ad7SLin Ling } 14993f9d6ad7SLin Ling 1500cde58dbcSMatthew Ahrens scn->scn_zio_root = zio_root(dp->dp_spa, NULL, 15013f9d6ad7SLin Ling NULL, ZIO_FLAG_CANFAIL); 15023b2aab18SMatthew Ahrens dsl_pool_config_enter(dp, FTAG); 15033f9d6ad7SLin Ling dsl_scan_visit(scn, tx); 15043b2aab18SMatthew Ahrens dsl_pool_config_exit(dp, FTAG); 1505cde58dbcSMatthew Ahrens (void) zio_wait(scn->scn_zio_root); 1506cde58dbcSMatthew Ahrens scn->scn_zio_root = NULL; 15073f9d6ad7SLin Ling 15083f9d6ad7SLin Ling zfs_dbgmsg("visited %llu blocks in %llums", 15093f9d6ad7SLin Ling (longlong_t)scn->scn_visited_this_txg, 15100689f76cSAdam Leventhal (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time)); 15113f9d6ad7SLin Ling 15123f9d6ad7SLin Ling if (!scn->scn_pausing) { 1513b4952e17SGeorge Wilson scn->scn_done_txg = tx->tx_txg + 1; 1514b4952e17SGeorge Wilson zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu", 1515b4952e17SGeorge Wilson tx->tx_txg, scn->scn_done_txg); 15163f9d6ad7SLin Ling } 15173f9d6ad7SLin Ling 15183f9d6ad7SLin Ling if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) { 15193f9d6ad7SLin Ling mutex_enter(&spa->spa_scrub_lock); 15203f9d6ad7SLin Ling while (spa->spa_scrub_inflight > 0) { 15213f9d6ad7SLin Ling cv_wait(&spa->spa_scrub_io_cv, 15223f9d6ad7SLin Ling &spa->spa_scrub_lock); 15233f9d6ad7SLin Ling } 15243f9d6ad7SLin Ling mutex_exit(&spa->spa_scrub_lock); 15253f9d6ad7SLin Ling } 15263f9d6ad7SLin Ling 15273f9d6ad7SLin Ling dsl_scan_sync_state(scn, tx); 15283f9d6ad7SLin Ling } 15293f9d6ad7SLin Ling 15303f9d6ad7SLin Ling /* 15313f9d6ad7SLin Ling * This will start a new scan, or restart an existing one. 15323f9d6ad7SLin Ling */ 15333f9d6ad7SLin Ling void 15343f9d6ad7SLin Ling dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg) 15353f9d6ad7SLin Ling { 15363f9d6ad7SLin Ling if (txg == 0) { 15373f9d6ad7SLin Ling dmu_tx_t *tx; 15383f9d6ad7SLin Ling tx = dmu_tx_create_dd(dp->dp_mos_dir); 15393f9d6ad7SLin Ling VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT)); 15403f9d6ad7SLin Ling 15413f9d6ad7SLin Ling txg = dmu_tx_get_txg(tx); 15423f9d6ad7SLin Ling dp->dp_scan->scn_restart_txg = txg; 15433f9d6ad7SLin Ling dmu_tx_commit(tx); 15443f9d6ad7SLin Ling } else { 15453f9d6ad7SLin Ling dp->dp_scan->scn_restart_txg = txg; 15463f9d6ad7SLin Ling } 15473f9d6ad7SLin Ling zfs_dbgmsg("restarting resilver txg=%llu", txg); 15483f9d6ad7SLin Ling } 15493f9d6ad7SLin Ling 15503f9d6ad7SLin Ling boolean_t 15513f9d6ad7SLin Ling dsl_scan_resilvering(dsl_pool_t *dp) 15523f9d6ad7SLin Ling { 15533f9d6ad7SLin Ling return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING && 15543f9d6ad7SLin Ling dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER); 15553f9d6ad7SLin Ling } 15563f9d6ad7SLin Ling 15573f9d6ad7SLin Ling /* 15583f9d6ad7SLin Ling * scrub consumers 15593f9d6ad7SLin Ling */ 15603f9d6ad7SLin Ling 15613f9d6ad7SLin Ling static void 15623f9d6ad7SLin Ling count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp) 15633f9d6ad7SLin Ling { 15643f9d6ad7SLin Ling int i; 15653f9d6ad7SLin Ling 15663f9d6ad7SLin Ling /* 15673f9d6ad7SLin Ling * If we resume after a reboot, zab will be NULL; don't record 15683f9d6ad7SLin Ling * incomplete stats in that case. 15693f9d6ad7SLin Ling */ 15703f9d6ad7SLin Ling if (zab == NULL) 15713f9d6ad7SLin Ling return; 15723f9d6ad7SLin Ling 15733f9d6ad7SLin Ling for (i = 0; i < 4; i++) { 15743f9d6ad7SLin Ling int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS; 15753f9d6ad7SLin Ling int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL; 1576ad135b5dSChristopher Siden if (t & DMU_OT_NEWTYPE) 1577ad135b5dSChristopher Siden t = DMU_OT_OTHER; 15783f9d6ad7SLin Ling zfs_blkstat_t *zb = &zab->zab_type[l][t]; 15793f9d6ad7SLin Ling int equal; 15803f9d6ad7SLin Ling 15813f9d6ad7SLin Ling zb->zb_count++; 15823f9d6ad7SLin Ling zb->zb_asize += BP_GET_ASIZE(bp); 15833f9d6ad7SLin Ling zb->zb_lsize += BP_GET_LSIZE(bp); 15843f9d6ad7SLin Ling zb->zb_psize += BP_GET_PSIZE(bp); 15853f9d6ad7SLin Ling zb->zb_gangs += BP_COUNT_GANG(bp); 15863f9d6ad7SLin Ling 15873f9d6ad7SLin Ling switch (BP_GET_NDVAS(bp)) { 15883f9d6ad7SLin Ling case 2: 15893f9d6ad7SLin Ling if (DVA_GET_VDEV(&bp->blk_dva[0]) == 15903f9d6ad7SLin Ling DVA_GET_VDEV(&bp->blk_dva[1])) 15913f9d6ad7SLin Ling zb->zb_ditto_2_of_2_samevdev++; 15923f9d6ad7SLin Ling break; 15933f9d6ad7SLin Ling case 3: 15943f9d6ad7SLin Ling equal = (DVA_GET_VDEV(&bp->blk_dva[0]) == 15953f9d6ad7SLin Ling DVA_GET_VDEV(&bp->blk_dva[1])) + 15963f9d6ad7SLin Ling (DVA_GET_VDEV(&bp->blk_dva[0]) == 15973f9d6ad7SLin Ling DVA_GET_VDEV(&bp->blk_dva[2])) + 15983f9d6ad7SLin Ling (DVA_GET_VDEV(&bp->blk_dva[1]) == 15993f9d6ad7SLin Ling DVA_GET_VDEV(&bp->blk_dva[2])); 16003f9d6ad7SLin Ling if (equal == 1) 16013f9d6ad7SLin Ling zb->zb_ditto_2_of_3_samevdev++; 16023f9d6ad7SLin Ling else if (equal == 3) 16033f9d6ad7SLin Ling zb->zb_ditto_3_of_3_samevdev++; 16043f9d6ad7SLin Ling break; 16053f9d6ad7SLin Ling } 16063f9d6ad7SLin Ling } 16073f9d6ad7SLin Ling } 16083f9d6ad7SLin Ling 16093f9d6ad7SLin Ling static void 16103f9d6ad7SLin Ling dsl_scan_scrub_done(zio_t *zio) 16113f9d6ad7SLin Ling { 16123f9d6ad7SLin Ling spa_t *spa = zio->io_spa; 16133f9d6ad7SLin Ling 16143f9d6ad7SLin Ling zio_data_buf_free(zio->io_data, zio->io_size); 16153f9d6ad7SLin Ling 16163f9d6ad7SLin Ling mutex_enter(&spa->spa_scrub_lock); 16173f9d6ad7SLin Ling spa->spa_scrub_inflight--; 16183f9d6ad7SLin Ling cv_broadcast(&spa->spa_scrub_io_cv); 16193f9d6ad7SLin Ling 16203f9d6ad7SLin Ling if (zio->io_error && (zio->io_error != ECKSUM || 16213f9d6ad7SLin Ling !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) { 16223f9d6ad7SLin Ling spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++; 16233f9d6ad7SLin Ling } 16243f9d6ad7SLin Ling mutex_exit(&spa->spa_scrub_lock); 16253f9d6ad7SLin Ling } 16263f9d6ad7SLin Ling 16273f9d6ad7SLin Ling static int 16283f9d6ad7SLin Ling dsl_scan_scrub_cb(dsl_pool_t *dp, 16293f9d6ad7SLin Ling const blkptr_t *bp, const zbookmark_t *zb) 16303f9d6ad7SLin Ling { 16313f9d6ad7SLin Ling dsl_scan_t *scn = dp->dp_scan; 16323f9d6ad7SLin Ling size_t size = BP_GET_PSIZE(bp); 16333f9d6ad7SLin Ling spa_t *spa = dp->dp_spa; 16343f9d6ad7SLin Ling uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp); 16353f9d6ad7SLin Ling boolean_t needs_io; 163644ecc532SGeorge Wilson int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL; 163744ecc532SGeorge Wilson int scan_delay = 0; 16383f9d6ad7SLin Ling 16393f9d6ad7SLin Ling if (phys_birth <= scn->scn_phys.scn_min_txg || 16403f9d6ad7SLin Ling phys_birth >= scn->scn_phys.scn_max_txg) 16413f9d6ad7SLin Ling return (0); 16423f9d6ad7SLin Ling 16433f9d6ad7SLin Ling count_block(dp->dp_blkstats, bp); 16443f9d6ad7SLin Ling 1645*5d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 1646*5d7b4d43SMatthew Ahrens return (0); 1647*5d7b4d43SMatthew Ahrens 16483f9d6ad7SLin Ling ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn)); 16493f9d6ad7SLin Ling if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) { 16503f9d6ad7SLin Ling zio_flags |= ZIO_FLAG_SCRUB; 16513f9d6ad7SLin Ling needs_io = B_TRUE; 165244ecc532SGeorge Wilson scan_delay = zfs_scrub_delay; 1653d5285caeSGeorge Wilson } else { 1654d5285caeSGeorge Wilson ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER); 16553f9d6ad7SLin Ling zio_flags |= ZIO_FLAG_RESILVER; 16563f9d6ad7SLin Ling needs_io = B_FALSE; 165744ecc532SGeorge Wilson scan_delay = zfs_resilver_delay; 16583f9d6ad7SLin Ling } 16593f9d6ad7SLin Ling 16603f9d6ad7SLin Ling /* If it's an intent log block, failure is expected. */ 16613f9d6ad7SLin Ling if (zb->zb_level == ZB_ZIL_LEVEL) 16623f9d6ad7SLin Ling zio_flags |= ZIO_FLAG_SPECULATIVE; 16633f9d6ad7SLin Ling 16643f9d6ad7SLin Ling for (int d = 0; d < BP_GET_NDVAS(bp); d++) { 16653f9d6ad7SLin Ling vdev_t *vd = vdev_lookup_top(spa, 16663f9d6ad7SLin Ling DVA_GET_VDEV(&bp->blk_dva[d])); 16673f9d6ad7SLin Ling 16683f9d6ad7SLin Ling /* 16693f9d6ad7SLin Ling * Keep track of how much data we've examined so that 16703f9d6ad7SLin Ling * zpool(1M) status can make useful progress reports. 16713f9d6ad7SLin Ling */ 16723f9d6ad7SLin Ling scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]); 16733f9d6ad7SLin Ling spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]); 16743f9d6ad7SLin Ling 16753f9d6ad7SLin Ling /* if it's a resilver, this may not be in the target range */ 16763f9d6ad7SLin Ling if (!needs_io) { 16773f9d6ad7SLin Ling if (DVA_GET_GANG(&bp->blk_dva[d])) { 16783f9d6ad7SLin Ling /* 16793f9d6ad7SLin Ling * Gang members may be spread across multiple 16803f9d6ad7SLin Ling * vdevs, so the best estimate we have is the 16813f9d6ad7SLin Ling * scrub range, which has already been checked. 16823f9d6ad7SLin Ling * XXX -- it would be better to change our 16833f9d6ad7SLin Ling * allocation policy to ensure that all 16843f9d6ad7SLin Ling * gang members reside on the same vdev. 16853f9d6ad7SLin Ling */ 16863f9d6ad7SLin Ling needs_io = B_TRUE; 16873f9d6ad7SLin Ling } else { 16883f9d6ad7SLin Ling needs_io = vdev_dtl_contains(vd, DTL_PARTIAL, 16893f9d6ad7SLin Ling phys_birth, 1); 16903f9d6ad7SLin Ling } 16913f9d6ad7SLin Ling } 16923f9d6ad7SLin Ling } 16933f9d6ad7SLin Ling 16943f9d6ad7SLin Ling if (needs_io && !zfs_no_scrub_io) { 169544ecc532SGeorge Wilson vdev_t *rvd = spa->spa_root_vdev; 169644ecc532SGeorge Wilson uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight; 16973f9d6ad7SLin Ling void *data = zio_data_buf_alloc(size); 16983f9d6ad7SLin Ling 16993f9d6ad7SLin Ling mutex_enter(&spa->spa_scrub_lock); 170044ecc532SGeorge Wilson while (spa->spa_scrub_inflight >= maxinflight) 17013f9d6ad7SLin Ling cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 17023f9d6ad7SLin Ling spa->spa_scrub_inflight++; 17033f9d6ad7SLin Ling mutex_exit(&spa->spa_scrub_lock); 17043f9d6ad7SLin Ling 170544ecc532SGeorge Wilson /* 170644ecc532SGeorge Wilson * If we're seeing recent (zfs_scan_idle) "important" I/Os 170744ecc532SGeorge Wilson * then throttle our workload to limit the impact of a scan. 170844ecc532SGeorge Wilson */ 170944ecc532SGeorge Wilson if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle) 171044ecc532SGeorge Wilson delay(scan_delay); 171144ecc532SGeorge Wilson 17123f9d6ad7SLin Ling zio_nowait(zio_read(NULL, spa, bp, data, size, 171369962b56SMatthew Ahrens dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB, 17143f9d6ad7SLin Ling zio_flags, zb)); 17153f9d6ad7SLin Ling } 17163f9d6ad7SLin Ling 17173f9d6ad7SLin Ling /* do not relocate this block */ 17183f9d6ad7SLin Ling return (0); 17193f9d6ad7SLin Ling } 17203f9d6ad7SLin Ling 17213f9d6ad7SLin Ling int 17223f9d6ad7SLin Ling dsl_scan(dsl_pool_t *dp, pool_scan_func_t func) 17233f9d6ad7SLin Ling { 17243f9d6ad7SLin Ling spa_t *spa = dp->dp_spa; 17253f9d6ad7SLin Ling 17263f9d6ad7SLin Ling /* 17273f9d6ad7SLin Ling * Purge all vdev caches and probe all devices. We do this here 17283f9d6ad7SLin Ling * rather than in sync context because this requires a writer lock 17293f9d6ad7SLin Ling * on the spa_config lock, which we can't do from sync context. The 17303f9d6ad7SLin Ling * spa_scrub_reopen flag indicates that vdev_open() should not 17313f9d6ad7SLin Ling * attempt to start another scrub. 17323f9d6ad7SLin Ling */ 17333f9d6ad7SLin Ling spa_vdev_state_enter(spa, SCL_NONE); 17343f9d6ad7SLin Ling spa->spa_scrub_reopen = B_TRUE; 17353f9d6ad7SLin Ling vdev_reopen(spa->spa_root_vdev); 17363f9d6ad7SLin Ling spa->spa_scrub_reopen = B_FALSE; 17373f9d6ad7SLin Ling (void) spa_vdev_state_exit(spa, NULL, 0); 17383f9d6ad7SLin Ling 17393b2aab18SMatthew Ahrens return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check, 17403b2aab18SMatthew Ahrens dsl_scan_setup_sync, &func, 0)); 17413f9d6ad7SLin Ling } 1742