xref: /illumos-gate/usr/src/uts/common/fs/zfs/dsl_scan.c (revision 7802d7bf98dec568dadf72286893b1fe5abd8602)
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.
23*7802d7bfSMatthew Ahrens  * Copyright (c) 2011, 2014 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 
53*7802d7bfSMatthew Ahrens typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
54*7802d7bfSMatthew Ahrens     const zbookmark_phys_t *);
553f9d6ad7SLin Ling 
563f9d6ad7SLin Ling static scan_cb_t dsl_scan_scrub_cb;
573b2aab18SMatthew Ahrens static void dsl_scan_cancel_sync(void *, dmu_tx_t *);
583f9d6ad7SLin Ling static void dsl_scan_sync_state(dsl_scan_t *, dmu_tx_t *tx);
593f9d6ad7SLin Ling 
6044ecc532SGeorge Wilson int zfs_top_maxinflight = 32;		/* maximum I/Os per top-level */
6144ecc532SGeorge Wilson int zfs_resilver_delay = 2;		/* number of ticks to delay resilver */
6244ecc532SGeorge Wilson int zfs_scrub_delay = 4;		/* number of ticks to delay scrub */
6344ecc532SGeorge Wilson int zfs_scan_idle = 50;			/* idle window in clock ticks */
6444ecc532SGeorge Wilson 
653f9d6ad7SLin Ling int zfs_scan_min_time_ms = 1000; /* min millisecs to scrub per txg */
66cde58dbcSMatthew Ahrens int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
673f9d6ad7SLin Ling int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
683f9d6ad7SLin Ling boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
697fd05ac4SMatthew Ahrens boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
703f9d6ad7SLin Ling enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
713f9d6ad7SLin Ling int dsl_scan_delay_completion = B_FALSE; /* set to delay scan completion */
723f9d6ad7SLin Ling 
733f9d6ad7SLin Ling #define	DSL_SCAN_IS_SCRUB_RESILVER(scn) \
743f9d6ad7SLin Ling 	((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
753f9d6ad7SLin Ling 	(scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
763f9d6ad7SLin Ling 
773f9d6ad7SLin Ling extern int zfs_txg_timeout;
783f9d6ad7SLin Ling 
793f9d6ad7SLin Ling /* the order has to match pool_scan_type */
803f9d6ad7SLin Ling static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
813f9d6ad7SLin Ling 	NULL,
823f9d6ad7SLin Ling 	dsl_scan_scrub_cb,	/* POOL_SCAN_SCRUB */
833f9d6ad7SLin Ling 	dsl_scan_scrub_cb,	/* POOL_SCAN_RESILVER */
843f9d6ad7SLin Ling };
853f9d6ad7SLin Ling 
863f9d6ad7SLin Ling int
873f9d6ad7SLin Ling dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
883f9d6ad7SLin Ling {
893f9d6ad7SLin Ling 	int err;
903f9d6ad7SLin Ling 	dsl_scan_t *scn;
913f9d6ad7SLin Ling 	spa_t *spa = dp->dp_spa;
923f9d6ad7SLin Ling 	uint64_t f;
933f9d6ad7SLin Ling 
943f9d6ad7SLin Ling 	scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
953f9d6ad7SLin Ling 	scn->scn_dp = dp;
963f9d6ad7SLin Ling 
974a923759SGeorge Wilson 	/*
984a923759SGeorge Wilson 	 * It's possible that we're resuming a scan after a reboot so
994a923759SGeorge Wilson 	 * make sure that the scan_async_destroying flag is initialized
1004a923759SGeorge Wilson 	 * appropriately.
1014a923759SGeorge Wilson 	 */
1024a923759SGeorge Wilson 	ASSERT(!scn->scn_async_destroying);
1034a923759SGeorge Wilson 	scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
1042acef22dSMatthew Ahrens 	    SPA_FEATURE_ASYNC_DESTROY);
1054a923759SGeorge Wilson 
1063f9d6ad7SLin Ling 	err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1073f9d6ad7SLin Ling 	    "scrub_func", sizeof (uint64_t), 1, &f);
1083f9d6ad7SLin Ling 	if (err == 0) {
1093f9d6ad7SLin Ling 		/*
1103f9d6ad7SLin Ling 		 * There was an old-style scrub in progress.  Restart a
1113f9d6ad7SLin Ling 		 * new-style scrub from the beginning.
1123f9d6ad7SLin Ling 		 */
1133f9d6ad7SLin Ling 		scn->scn_restart_txg = txg;
1143f9d6ad7SLin Ling 		zfs_dbgmsg("old-style scrub was in progress; "
1153f9d6ad7SLin Ling 		    "restarting new-style scrub in txg %llu",
1163f9d6ad7SLin Ling 		    scn->scn_restart_txg);
1173f9d6ad7SLin Ling 
1183f9d6ad7SLin Ling 		/*
1193f9d6ad7SLin Ling 		 * Load the queue obj from the old location so that it
1203f9d6ad7SLin Ling 		 * can be freed by dsl_scan_done().
1213f9d6ad7SLin Ling 		 */
1223f9d6ad7SLin Ling 		(void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1233f9d6ad7SLin Ling 		    "scrub_queue", sizeof (uint64_t), 1,
1243f9d6ad7SLin Ling 		    &scn->scn_phys.scn_queue_obj);
1253f9d6ad7SLin Ling 	} else {
1263f9d6ad7SLin Ling 		err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
1273f9d6ad7SLin Ling 		    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
1283f9d6ad7SLin Ling 		    &scn->scn_phys);
1293f9d6ad7SLin Ling 		if (err == ENOENT)
1303f9d6ad7SLin Ling 			return (0);
1313f9d6ad7SLin Ling 		else if (err)
1323f9d6ad7SLin Ling 			return (err);
1333f9d6ad7SLin Ling 
1343f9d6ad7SLin Ling 		if (scn->scn_phys.scn_state == DSS_SCANNING &&
1353f9d6ad7SLin Ling 		    spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
1363f9d6ad7SLin Ling 			/*
1373f9d6ad7SLin Ling 			 * A new-type scrub was in progress on an old
1383f9d6ad7SLin Ling 			 * pool, and the pool was accessed by old
1393f9d6ad7SLin Ling 			 * software.  Restart from the beginning, since
1403f9d6ad7SLin Ling 			 * the old software may have changed the pool in
1413f9d6ad7SLin Ling 			 * the meantime.
1423f9d6ad7SLin Ling 			 */
1433f9d6ad7SLin Ling 			scn->scn_restart_txg = txg;
1443f9d6ad7SLin Ling 			zfs_dbgmsg("new-style scrub was modified "
1453f9d6ad7SLin Ling 			    "by old software; restarting in txg %llu",
1463f9d6ad7SLin Ling 			    scn->scn_restart_txg);
1473f9d6ad7SLin Ling 		}
1483f9d6ad7SLin Ling 	}
1493f9d6ad7SLin Ling 
1503f9d6ad7SLin Ling 	spa_scan_stat_init(spa);
1513f9d6ad7SLin Ling 	return (0);
1523f9d6ad7SLin Ling }
1533f9d6ad7SLin Ling 
1543f9d6ad7SLin Ling void
1553f9d6ad7SLin Ling dsl_scan_fini(dsl_pool_t *dp)
1563f9d6ad7SLin Ling {
1573f9d6ad7SLin Ling 	if (dp->dp_scan) {
1583f9d6ad7SLin Ling 		kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
1593f9d6ad7SLin Ling 		dp->dp_scan = NULL;
1603f9d6ad7SLin Ling 	}
1613f9d6ad7SLin Ling }
1623f9d6ad7SLin Ling 
1633f9d6ad7SLin Ling /* ARGSUSED */
1643f9d6ad7SLin Ling static int
1653b2aab18SMatthew Ahrens dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
1663f9d6ad7SLin Ling {
1673b2aab18SMatthew Ahrens 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1683f9d6ad7SLin Ling 
1693f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state == DSS_SCANNING)
170be6fd75aSMatthew Ahrens 		return (SET_ERROR(EBUSY));
1713f9d6ad7SLin Ling 
1723f9d6ad7SLin Ling 	return (0);
1733f9d6ad7SLin Ling }
1743f9d6ad7SLin Ling 
1753f9d6ad7SLin Ling static void
1763b2aab18SMatthew Ahrens dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
1773f9d6ad7SLin Ling {
1783b2aab18SMatthew Ahrens 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
1793b2aab18SMatthew Ahrens 	pool_scan_func_t *funcp = arg;
1803f9d6ad7SLin Ling 	dmu_object_type_t ot = 0;
1813f9d6ad7SLin Ling 	dsl_pool_t *dp = scn->scn_dp;
1823f9d6ad7SLin Ling 	spa_t *spa = dp->dp_spa;
1833f9d6ad7SLin Ling 
1843f9d6ad7SLin Ling 	ASSERT(scn->scn_phys.scn_state != DSS_SCANNING);
1853f9d6ad7SLin Ling 	ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
1863f9d6ad7SLin Ling 	bzero(&scn->scn_phys, sizeof (scn->scn_phys));
1873f9d6ad7SLin Ling 	scn->scn_phys.scn_func = *funcp;
1883f9d6ad7SLin Ling 	scn->scn_phys.scn_state = DSS_SCANNING;
1893f9d6ad7SLin Ling 	scn->scn_phys.scn_min_txg = 0;
1903f9d6ad7SLin Ling 	scn->scn_phys.scn_max_txg = tx->tx_txg;
1913f9d6ad7SLin Ling 	scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
1923f9d6ad7SLin Ling 	scn->scn_phys.scn_start_time = gethrestime_sec();
1933f9d6ad7SLin Ling 	scn->scn_phys.scn_errors = 0;
1943f9d6ad7SLin Ling 	scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
1953f9d6ad7SLin Ling 	scn->scn_restart_txg = 0;
196b4952e17SGeorge Wilson 	scn->scn_done_txg = 0;
1973f9d6ad7SLin Ling 	spa_scan_stat_init(spa);
1983f9d6ad7SLin Ling 
1993f9d6ad7SLin Ling 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
2003f9d6ad7SLin Ling 		scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
2013f9d6ad7SLin Ling 
2023f9d6ad7SLin Ling 		/* rewrite all disk labels */
2033f9d6ad7SLin Ling 		vdev_config_dirty(spa->spa_root_vdev);
2043f9d6ad7SLin Ling 
2053f9d6ad7SLin Ling 		if (vdev_resilver_needed(spa->spa_root_vdev,
2063f9d6ad7SLin Ling 		    &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
2073f9d6ad7SLin Ling 			spa_event_notify(spa, NULL, ESC_ZFS_RESILVER_START);
2083f9d6ad7SLin Ling 		} else {
2093f9d6ad7SLin Ling 			spa_event_notify(spa, NULL, ESC_ZFS_SCRUB_START);
2103f9d6ad7SLin Ling 		}
2113f9d6ad7SLin Ling 
2123f9d6ad7SLin Ling 		spa->spa_scrub_started = B_TRUE;
2133f9d6ad7SLin Ling 		/*
2143f9d6ad7SLin Ling 		 * If this is an incremental scrub, limit the DDT scrub phase
2153f9d6ad7SLin Ling 		 * to just the auto-ditto class (for correctness); the rest
2163f9d6ad7SLin Ling 		 * of the scrub should go faster using top-down pruning.
2173f9d6ad7SLin Ling 		 */
2183f9d6ad7SLin Ling 		if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
2193f9d6ad7SLin Ling 			scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
2203f9d6ad7SLin Ling 
2213f9d6ad7SLin Ling 	}
2223f9d6ad7SLin Ling 
2233f9d6ad7SLin Ling 	/* back to the generic stuff */
2243f9d6ad7SLin Ling 
2253f9d6ad7SLin Ling 	if (dp->dp_blkstats == NULL) {
2263f9d6ad7SLin Ling 		dp->dp_blkstats =
2273f9d6ad7SLin Ling 		    kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
2283f9d6ad7SLin Ling 	}
2293f9d6ad7SLin Ling 	bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
2303f9d6ad7SLin Ling 
2313f9d6ad7SLin Ling 	if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
2323f9d6ad7SLin Ling 		ot = DMU_OT_ZAP_OTHER;
2333f9d6ad7SLin Ling 
2343f9d6ad7SLin Ling 	scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
2353f9d6ad7SLin Ling 	    ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
2363f9d6ad7SLin Ling 
2373f9d6ad7SLin Ling 	dsl_scan_sync_state(scn, tx);
2383f9d6ad7SLin Ling 
2394445fffbSMatthew Ahrens 	spa_history_log_internal(spa, "scan setup", tx,
2403f9d6ad7SLin Ling 	    "func=%u mintxg=%llu maxtxg=%llu",
2413f9d6ad7SLin Ling 	    *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
2423f9d6ad7SLin Ling }
2433f9d6ad7SLin Ling 
2443f9d6ad7SLin Ling /* ARGSUSED */
2453f9d6ad7SLin Ling static void
2463f9d6ad7SLin Ling dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
2473f9d6ad7SLin Ling {
2483f9d6ad7SLin Ling 	static const char *old_names[] = {
2493f9d6ad7SLin Ling 		"scrub_bookmark",
2503f9d6ad7SLin Ling 		"scrub_ddt_bookmark",
2513f9d6ad7SLin Ling 		"scrub_ddt_class_max",
2523f9d6ad7SLin Ling 		"scrub_queue",
2533f9d6ad7SLin Ling 		"scrub_min_txg",
2543f9d6ad7SLin Ling 		"scrub_max_txg",
2553f9d6ad7SLin Ling 		"scrub_func",
2563f9d6ad7SLin Ling 		"scrub_errors",
2573f9d6ad7SLin Ling 		NULL
2583f9d6ad7SLin Ling 	};
2593f9d6ad7SLin Ling 
2603f9d6ad7SLin Ling 	dsl_pool_t *dp = scn->scn_dp;
2613f9d6ad7SLin Ling 	spa_t *spa = dp->dp_spa;
2623f9d6ad7SLin Ling 	int i;
2633f9d6ad7SLin Ling 
2643f9d6ad7SLin Ling 	/* Remove any remnants of an old-style scrub. */
2653f9d6ad7SLin Ling 	for (i = 0; old_names[i]; i++) {
2663f9d6ad7SLin Ling 		(void) zap_remove(dp->dp_meta_objset,
2673f9d6ad7SLin Ling 		    DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
2683f9d6ad7SLin Ling 	}
2693f9d6ad7SLin Ling 
2703f9d6ad7SLin Ling 	if (scn->scn_phys.scn_queue_obj != 0) {
2713f9d6ad7SLin Ling 		VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
2723f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, tx));
2733f9d6ad7SLin Ling 		scn->scn_phys.scn_queue_obj = 0;
2743f9d6ad7SLin Ling 	}
2753f9d6ad7SLin Ling 
2763f9d6ad7SLin Ling 	/*
2773f9d6ad7SLin Ling 	 * If we were "restarted" from a stopped state, don't bother
2783f9d6ad7SLin Ling 	 * with anything else.
2793f9d6ad7SLin Ling 	 */
2803f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state != DSS_SCANNING)
2813f9d6ad7SLin Ling 		return;
2823f9d6ad7SLin Ling 
2833f9d6ad7SLin Ling 	if (complete)
2843f9d6ad7SLin Ling 		scn->scn_phys.scn_state = DSS_FINISHED;
2853f9d6ad7SLin Ling 	else
2863f9d6ad7SLin Ling 		scn->scn_phys.scn_state = DSS_CANCELED;
2873f9d6ad7SLin Ling 
2884445fffbSMatthew Ahrens 	spa_history_log_internal(spa, "scan done", tx,
2893f9d6ad7SLin Ling 	    "complete=%u", complete);
2903f9d6ad7SLin Ling 
2913f9d6ad7SLin Ling 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
2923f9d6ad7SLin Ling 		mutex_enter(&spa->spa_scrub_lock);
2933f9d6ad7SLin Ling 		while (spa->spa_scrub_inflight > 0) {
2943f9d6ad7SLin Ling 			cv_wait(&spa->spa_scrub_io_cv,
2953f9d6ad7SLin Ling 			    &spa->spa_scrub_lock);
2963f9d6ad7SLin Ling 		}
2973f9d6ad7SLin Ling 		mutex_exit(&spa->spa_scrub_lock);
2983f9d6ad7SLin Ling 		spa->spa_scrub_started = B_FALSE;
2993f9d6ad7SLin Ling 		spa->spa_scrub_active = B_FALSE;
3003f9d6ad7SLin Ling 
3013f9d6ad7SLin Ling 		/*
3023f9d6ad7SLin Ling 		 * If the scrub/resilver completed, update all DTLs to
3033f9d6ad7SLin Ling 		 * reflect this.  Whether it succeeded or not, vacate
3043f9d6ad7SLin Ling 		 * all temporary scrub DTLs.
3053f9d6ad7SLin Ling 		 */
3063f9d6ad7SLin Ling 		vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
3073f9d6ad7SLin Ling 		    complete ? scn->scn_phys.scn_max_txg : 0, B_TRUE);
3083f9d6ad7SLin Ling 		if (complete) {
3093f9d6ad7SLin Ling 			spa_event_notify(spa, NULL, scn->scn_phys.scn_min_txg ?
3103f9d6ad7SLin Ling 			    ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
3113f9d6ad7SLin Ling 		}
3123f9d6ad7SLin Ling 		spa_errlog_rotate(spa);
3133f9d6ad7SLin Ling 
3143f9d6ad7SLin Ling 		/*
3153f9d6ad7SLin Ling 		 * We may have finished replacing a device.
3163f9d6ad7SLin Ling 		 * Let the async thread assess this and handle the detach.
3173f9d6ad7SLin Ling 		 */
3183f9d6ad7SLin Ling 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
3193f9d6ad7SLin Ling 	}
3203f9d6ad7SLin Ling 
3213f9d6ad7SLin Ling 	scn->scn_phys.scn_end_time = gethrestime_sec();
3223f9d6ad7SLin Ling }
3233f9d6ad7SLin Ling 
3243f9d6ad7SLin Ling /* ARGSUSED */
3253f9d6ad7SLin Ling static int
3263b2aab18SMatthew Ahrens dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
3273f9d6ad7SLin Ling {
3283b2aab18SMatthew Ahrens 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
3293f9d6ad7SLin Ling 
3303f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state != DSS_SCANNING)
331be6fd75aSMatthew Ahrens 		return (SET_ERROR(ENOENT));
3323f9d6ad7SLin Ling 	return (0);
3333f9d6ad7SLin Ling }
3343f9d6ad7SLin Ling 
3353f9d6ad7SLin Ling /* ARGSUSED */
3363f9d6ad7SLin Ling static void
3373b2aab18SMatthew Ahrens dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
3383f9d6ad7SLin Ling {
3393b2aab18SMatthew Ahrens 	dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
3403f9d6ad7SLin Ling 
3413f9d6ad7SLin Ling 	dsl_scan_done(scn, B_FALSE, tx);
3423f9d6ad7SLin Ling 	dsl_scan_sync_state(scn, tx);
3433f9d6ad7SLin Ling }
3443f9d6ad7SLin Ling 
3453f9d6ad7SLin Ling int
3463f9d6ad7SLin Ling dsl_scan_cancel(dsl_pool_t *dp)
3473f9d6ad7SLin Ling {
3483b2aab18SMatthew Ahrens 	return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
3493b2aab18SMatthew Ahrens 	    dsl_scan_cancel_sync, NULL, 3));
3503f9d6ad7SLin Ling }
3513f9d6ad7SLin Ling 
3523f9d6ad7SLin Ling static void dsl_scan_visitbp(blkptr_t *bp,
353*7802d7bfSMatthew Ahrens     const zbookmark_phys_t *zb, dnode_phys_t *dnp, arc_buf_t *pbuf,
3543f9d6ad7SLin Ling     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
3553f9d6ad7SLin Ling     dmu_tx_t *tx);
3563f9d6ad7SLin Ling static void dsl_scan_visitdnode(dsl_scan_t *, dsl_dataset_t *ds,
3573f9d6ad7SLin Ling     dmu_objset_type_t ostype,
3583f9d6ad7SLin Ling     dnode_phys_t *dnp, arc_buf_t *buf, uint64_t object, dmu_tx_t *tx);
3593f9d6ad7SLin Ling 
3603f9d6ad7SLin Ling void
3613f9d6ad7SLin Ling dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
3623f9d6ad7SLin Ling {
3633f9d6ad7SLin Ling 	zio_free(dp->dp_spa, txg, bp);
3643f9d6ad7SLin Ling }
3653f9d6ad7SLin Ling 
3663f9d6ad7SLin Ling void
3673f9d6ad7SLin Ling dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
3683f9d6ad7SLin Ling {
3693f9d6ad7SLin Ling 	ASSERT(dsl_pool_sync_context(dp));
3703f9d6ad7SLin Ling 	zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, pio->io_flags));
3713f9d6ad7SLin Ling }
3723f9d6ad7SLin Ling 
3733f9d6ad7SLin Ling static uint64_t
3743f9d6ad7SLin Ling dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
3753f9d6ad7SLin Ling {
3763f9d6ad7SLin Ling 	uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
3773f9d6ad7SLin Ling 	if (dsl_dataset_is_snapshot(ds))
3783f9d6ad7SLin Ling 		return (MIN(smt, ds->ds_phys->ds_creation_txg));
3793f9d6ad7SLin Ling 	return (smt);
3803f9d6ad7SLin Ling }
3813f9d6ad7SLin Ling 
3823f9d6ad7SLin Ling static void
3833f9d6ad7SLin Ling dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx)
3843f9d6ad7SLin Ling {
3853b2aab18SMatthew Ahrens 	VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
3863f9d6ad7SLin Ling 	    DMU_POOL_DIRECTORY_OBJECT,
3873f9d6ad7SLin Ling 	    DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
3883f9d6ad7SLin Ling 	    &scn->scn_phys, tx));
3893f9d6ad7SLin Ling }
3903f9d6ad7SLin Ling 
3913f9d6ad7SLin Ling static boolean_t
392*7802d7bfSMatthew Ahrens dsl_scan_check_pause(dsl_scan_t *scn, const zbookmark_phys_t *zb)
3933f9d6ad7SLin Ling {
3943f9d6ad7SLin Ling 	uint64_t elapsed_nanosecs;
3953f9d6ad7SLin Ling 	int mintime;
3963f9d6ad7SLin Ling 
3973f9d6ad7SLin Ling 	/* we never skip user/group accounting objects */
3983f9d6ad7SLin Ling 	if (zb && (int64_t)zb->zb_object < 0)
3993f9d6ad7SLin Ling 		return (B_FALSE);
4003f9d6ad7SLin Ling 
4013f9d6ad7SLin Ling 	if (scn->scn_pausing)
4023f9d6ad7SLin Ling 		return (B_TRUE); /* we're already pausing */
4033f9d6ad7SLin Ling 
404ad135b5dSChristopher Siden 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
4053f9d6ad7SLin Ling 		return (B_FALSE); /* we're resuming */
4063f9d6ad7SLin Ling 
4073f9d6ad7SLin Ling 	/* We only know how to resume from level-0 blocks. */
4083f9d6ad7SLin Ling 	if (zb && zb->zb_level != 0)
4093f9d6ad7SLin Ling 		return (B_FALSE);
4103f9d6ad7SLin Ling 
4113f9d6ad7SLin Ling 	mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
4123f9d6ad7SLin Ling 	    zfs_resilver_min_time_ms : zfs_scan_min_time_ms;
4133f9d6ad7SLin Ling 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
4143f9d6ad7SLin Ling 	if (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
4150689f76cSAdam Leventhal 	    (NSEC2MSEC(elapsed_nanosecs) > mintime &&
4163f9d6ad7SLin Ling 	    txg_sync_waiting(scn->scn_dp)) ||
4173f9d6ad7SLin Ling 	    spa_shutting_down(scn->scn_dp->dp_spa)) {
4183f9d6ad7SLin Ling 		if (zb) {
4193f9d6ad7SLin Ling 			dprintf("pausing at bookmark %llx/%llx/%llx/%llx\n",
4203f9d6ad7SLin Ling 			    (longlong_t)zb->zb_objset,
4213f9d6ad7SLin Ling 			    (longlong_t)zb->zb_object,
4223f9d6ad7SLin Ling 			    (longlong_t)zb->zb_level,
4233f9d6ad7SLin Ling 			    (longlong_t)zb->zb_blkid);
4243f9d6ad7SLin Ling 			scn->scn_phys.scn_bookmark = *zb;
4253f9d6ad7SLin Ling 		}
4263f9d6ad7SLin Ling 		dprintf("pausing at DDT bookmark %llx/%llx/%llx/%llx\n",
4273f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
4283f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
4293f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
4303f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
4313f9d6ad7SLin Ling 		scn->scn_pausing = B_TRUE;
4323f9d6ad7SLin Ling 		return (B_TRUE);
4333f9d6ad7SLin Ling 	}
4343f9d6ad7SLin Ling 	return (B_FALSE);
4353f9d6ad7SLin Ling }
4363f9d6ad7SLin Ling 
4373f9d6ad7SLin Ling typedef struct zil_scan_arg {
4383f9d6ad7SLin Ling 	dsl_pool_t	*zsa_dp;
4393f9d6ad7SLin Ling 	zil_header_t	*zsa_zh;
4403f9d6ad7SLin Ling } zil_scan_arg_t;
4413f9d6ad7SLin Ling 
4423f9d6ad7SLin Ling /* ARGSUSED */
4433f9d6ad7SLin Ling static int
4443f9d6ad7SLin Ling dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
4453f9d6ad7SLin Ling {
4463f9d6ad7SLin Ling 	zil_scan_arg_t *zsa = arg;
4473f9d6ad7SLin Ling 	dsl_pool_t *dp = zsa->zsa_dp;
4483f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
4493f9d6ad7SLin Ling 	zil_header_t *zh = zsa->zsa_zh;
450*7802d7bfSMatthew Ahrens 	zbookmark_phys_t zb;
4513f9d6ad7SLin Ling 
45243466aaeSMax Grossman 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
4533f9d6ad7SLin Ling 		return (0);
4543f9d6ad7SLin Ling 
4553f9d6ad7SLin Ling 	/*
4563f9d6ad7SLin Ling 	 * One block ("stubby") can be allocated a long time ago; we
4573f9d6ad7SLin Ling 	 * want to visit that one because it has been allocated
4583f9d6ad7SLin Ling 	 * (on-disk) even if it hasn't been claimed (even though for
4593f9d6ad7SLin Ling 	 * scrub there's nothing to do to it).
4603f9d6ad7SLin Ling 	 */
4613f9d6ad7SLin Ling 	if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
4623f9d6ad7SLin Ling 		return (0);
4633f9d6ad7SLin Ling 
4643f9d6ad7SLin Ling 	SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
4653f9d6ad7SLin Ling 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
4663f9d6ad7SLin Ling 
4673f9d6ad7SLin Ling 	VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
4683f9d6ad7SLin Ling 	return (0);
4693f9d6ad7SLin Ling }
4703f9d6ad7SLin Ling 
4713f9d6ad7SLin Ling /* ARGSUSED */
4723f9d6ad7SLin Ling static int
4733f9d6ad7SLin Ling dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
4743f9d6ad7SLin Ling {
4753f9d6ad7SLin Ling 	if (lrc->lrc_txtype == TX_WRITE) {
4763f9d6ad7SLin Ling 		zil_scan_arg_t *zsa = arg;
4773f9d6ad7SLin Ling 		dsl_pool_t *dp = zsa->zsa_dp;
4783f9d6ad7SLin Ling 		dsl_scan_t *scn = dp->dp_scan;
4793f9d6ad7SLin Ling 		zil_header_t *zh = zsa->zsa_zh;
4803f9d6ad7SLin Ling 		lr_write_t *lr = (lr_write_t *)lrc;
4813f9d6ad7SLin Ling 		blkptr_t *bp = &lr->lr_blkptr;
482*7802d7bfSMatthew Ahrens 		zbookmark_phys_t zb;
4833f9d6ad7SLin Ling 
48443466aaeSMax Grossman 		if (BP_IS_HOLE(bp) ||
48543466aaeSMax Grossman 		    bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
4863f9d6ad7SLin Ling 			return (0);
4873f9d6ad7SLin Ling 
4883f9d6ad7SLin Ling 		/*
4893f9d6ad7SLin Ling 		 * birth can be < claim_txg if this record's txg is
4903f9d6ad7SLin Ling 		 * already txg sync'ed (but this log block contains
4913f9d6ad7SLin Ling 		 * other records that are not synced)
4923f9d6ad7SLin Ling 		 */
4933f9d6ad7SLin Ling 		if (claim_txg == 0 || bp->blk_birth < claim_txg)
4943f9d6ad7SLin Ling 			return (0);
4953f9d6ad7SLin Ling 
4963f9d6ad7SLin Ling 		SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
4973f9d6ad7SLin Ling 		    lr->lr_foid, ZB_ZIL_LEVEL,
4983f9d6ad7SLin Ling 		    lr->lr_offset / BP_GET_LSIZE(bp));
4993f9d6ad7SLin Ling 
5003f9d6ad7SLin Ling 		VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
5013f9d6ad7SLin Ling 	}
5023f9d6ad7SLin Ling 	return (0);
5033f9d6ad7SLin Ling }
5043f9d6ad7SLin Ling 
5053f9d6ad7SLin Ling static void
5063f9d6ad7SLin Ling dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
5073f9d6ad7SLin Ling {
5083f9d6ad7SLin Ling 	uint64_t claim_txg = zh->zh_claim_txg;
5093f9d6ad7SLin Ling 	zil_scan_arg_t zsa = { dp, zh };
5103f9d6ad7SLin Ling 	zilog_t *zilog;
5113f9d6ad7SLin Ling 
5123f9d6ad7SLin Ling 	/*
5133f9d6ad7SLin Ling 	 * We only want to visit blocks that have been claimed but not yet
5143f9d6ad7SLin Ling 	 * replayed (or, in read-only mode, blocks that *would* be claimed).
5153f9d6ad7SLin Ling 	 */
5163f9d6ad7SLin Ling 	if (claim_txg == 0 && spa_writeable(dp->dp_spa))
5173f9d6ad7SLin Ling 		return;
5183f9d6ad7SLin Ling 
5193f9d6ad7SLin Ling 	zilog = zil_alloc(dp->dp_meta_objset, zh);
5203f9d6ad7SLin Ling 
5213f9d6ad7SLin Ling 	(void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
5223f9d6ad7SLin Ling 	    claim_txg);
5233f9d6ad7SLin Ling 
5243f9d6ad7SLin Ling 	zil_free(zilog);
5253f9d6ad7SLin Ling }
5263f9d6ad7SLin Ling 
5273f9d6ad7SLin Ling /* ARGSUSED */
5283f9d6ad7SLin Ling static void
5293f9d6ad7SLin Ling dsl_scan_prefetch(dsl_scan_t *scn, arc_buf_t *buf, blkptr_t *bp,
5303f9d6ad7SLin Ling     uint64_t objset, uint64_t object, uint64_t blkid)
5313f9d6ad7SLin Ling {
532*7802d7bfSMatthew Ahrens 	zbookmark_phys_t czb;
5333f9d6ad7SLin Ling 	uint32_t flags = ARC_NOWAIT | ARC_PREFETCH;
5343f9d6ad7SLin Ling 
5353f9d6ad7SLin Ling 	if (zfs_no_scrub_prefetch)
5363f9d6ad7SLin Ling 		return;
5373f9d6ad7SLin Ling 
5383f9d6ad7SLin Ling 	if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_min_txg ||
5393f9d6ad7SLin Ling 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE))
5403f9d6ad7SLin Ling 		return;
5413f9d6ad7SLin Ling 
5423f9d6ad7SLin Ling 	SET_BOOKMARK(&czb, objset, object, BP_GET_LEVEL(bp), blkid);
5433f9d6ad7SLin Ling 
544cde58dbcSMatthew Ahrens 	(void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa, bp,
5451b912ec7SGeorge Wilson 	    NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
54644ecc532SGeorge Wilson 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD, &flags, &czb);
5473f9d6ad7SLin Ling }
5483f9d6ad7SLin Ling 
5493f9d6ad7SLin Ling static boolean_t
5503f9d6ad7SLin Ling dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
551*7802d7bfSMatthew Ahrens     const zbookmark_phys_t *zb)
5523f9d6ad7SLin Ling {
5533f9d6ad7SLin Ling 	/*
5543f9d6ad7SLin Ling 	 * We never skip over user/group accounting objects (obj<0)
5553f9d6ad7SLin Ling 	 */
556ad135b5dSChristopher Siden 	if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
5573f9d6ad7SLin Ling 	    (int64_t)zb->zb_object >= 0) {
5583f9d6ad7SLin Ling 		/*
5593f9d6ad7SLin Ling 		 * If we already visited this bp & everything below (in
5603f9d6ad7SLin Ling 		 * a prior txg sync), don't bother doing it again.
5613f9d6ad7SLin Ling 		 */
562ad135b5dSChristopher Siden 		if (zbookmark_is_before(dnp, zb, &scn->scn_phys.scn_bookmark))
5633f9d6ad7SLin Ling 			return (B_TRUE);
5643f9d6ad7SLin Ling 
5653f9d6ad7SLin Ling 		/*
5663f9d6ad7SLin Ling 		 * If we found the block we're trying to resume from, or
5673f9d6ad7SLin Ling 		 * we went past it to a different object, zero it out to
5683f9d6ad7SLin Ling 		 * indicate that it's OK to start checking for pausing
5693f9d6ad7SLin Ling 		 * again.
5703f9d6ad7SLin Ling 		 */
5713f9d6ad7SLin Ling 		if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
5723f9d6ad7SLin Ling 		    zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
5733f9d6ad7SLin Ling 			dprintf("resuming at %llx/%llx/%llx/%llx\n",
5743f9d6ad7SLin Ling 			    (longlong_t)zb->zb_objset,
5753f9d6ad7SLin Ling 			    (longlong_t)zb->zb_object,
5763f9d6ad7SLin Ling 			    (longlong_t)zb->zb_level,
5773f9d6ad7SLin Ling 			    (longlong_t)zb->zb_blkid);
5783f9d6ad7SLin Ling 			bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
5793f9d6ad7SLin Ling 		}
5803f9d6ad7SLin Ling 	}
5813f9d6ad7SLin Ling 	return (B_FALSE);
5823f9d6ad7SLin Ling }
5833f9d6ad7SLin Ling 
5843f9d6ad7SLin Ling /*
5853f9d6ad7SLin Ling  * Return nonzero on i/o error.
5863f9d6ad7SLin Ling  * Return new buf to write out in *bufp.
5873f9d6ad7SLin Ling  */
5883f9d6ad7SLin Ling static int
5893f9d6ad7SLin Ling dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
5903f9d6ad7SLin Ling     dnode_phys_t *dnp, const blkptr_t *bp,
591*7802d7bfSMatthew Ahrens     const zbookmark_phys_t *zb, dmu_tx_t *tx, arc_buf_t **bufp)
5923f9d6ad7SLin Ling {
5933f9d6ad7SLin Ling 	dsl_pool_t *dp = scn->scn_dp;
59444ecc532SGeorge Wilson 	int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
5953f9d6ad7SLin Ling 	int err;
5963f9d6ad7SLin Ling 
5973f9d6ad7SLin Ling 	if (BP_GET_LEVEL(bp) > 0) {
5983f9d6ad7SLin Ling 		uint32_t flags = ARC_WAIT;
5993f9d6ad7SLin Ling 		int i;
6003f9d6ad7SLin Ling 		blkptr_t *cbp;
6013f9d6ad7SLin Ling 		int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
6023f9d6ad7SLin Ling 
6031b912ec7SGeorge Wilson 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
60444ecc532SGeorge Wilson 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
6053f9d6ad7SLin Ling 		if (err) {
6063f9d6ad7SLin Ling 			scn->scn_phys.scn_errors++;
6073f9d6ad7SLin Ling 			return (err);
6083f9d6ad7SLin Ling 		}
6093f9d6ad7SLin Ling 		for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
6103f9d6ad7SLin Ling 			dsl_scan_prefetch(scn, *bufp, cbp, zb->zb_objset,
6113f9d6ad7SLin Ling 			    zb->zb_object, zb->zb_blkid * epb + i);
6123f9d6ad7SLin Ling 		}
6133f9d6ad7SLin Ling 		for (i = 0, cbp = (*bufp)->b_data; i < epb; i++, cbp++) {
614*7802d7bfSMatthew Ahrens 			zbookmark_phys_t czb;
6153f9d6ad7SLin Ling 
6163f9d6ad7SLin Ling 			SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
6173f9d6ad7SLin Ling 			    zb->zb_level - 1,
6183f9d6ad7SLin Ling 			    zb->zb_blkid * epb + i);
6193f9d6ad7SLin Ling 			dsl_scan_visitbp(cbp, &czb, dnp,
6203f9d6ad7SLin Ling 			    *bufp, ds, scn, ostype, tx);
6213f9d6ad7SLin Ling 		}
6223f9d6ad7SLin Ling 	} else if (BP_GET_TYPE(bp) == DMU_OT_USERGROUP_USED) {
6233f9d6ad7SLin Ling 		uint32_t flags = ARC_WAIT;
6243f9d6ad7SLin Ling 
6251b912ec7SGeorge Wilson 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
62644ecc532SGeorge Wilson 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
6273f9d6ad7SLin Ling 		if (err) {
6283f9d6ad7SLin Ling 			scn->scn_phys.scn_errors++;
6293f9d6ad7SLin Ling 			return (err);
6303f9d6ad7SLin Ling 		}
6313f9d6ad7SLin Ling 	} else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
6323f9d6ad7SLin Ling 		uint32_t flags = ARC_WAIT;
6333f9d6ad7SLin Ling 		dnode_phys_t *cdnp;
6343f9d6ad7SLin Ling 		int i, j;
6353f9d6ad7SLin Ling 		int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
6363f9d6ad7SLin Ling 
6371b912ec7SGeorge Wilson 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
63844ecc532SGeorge Wilson 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
6393f9d6ad7SLin Ling 		if (err) {
6403f9d6ad7SLin Ling 			scn->scn_phys.scn_errors++;
6413f9d6ad7SLin Ling 			return (err);
6423f9d6ad7SLin Ling 		}
6433f9d6ad7SLin Ling 		for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
6443f9d6ad7SLin Ling 			for (j = 0; j < cdnp->dn_nblkptr; j++) {
6453f9d6ad7SLin Ling 				blkptr_t *cbp = &cdnp->dn_blkptr[j];
6463f9d6ad7SLin Ling 				dsl_scan_prefetch(scn, *bufp, cbp,
6473f9d6ad7SLin Ling 				    zb->zb_objset, zb->zb_blkid * epb + i, j);
6483f9d6ad7SLin Ling 			}
6493f9d6ad7SLin Ling 		}
6503f9d6ad7SLin Ling 		for (i = 0, cdnp = (*bufp)->b_data; i < epb; i++, cdnp++) {
6513f9d6ad7SLin Ling 			dsl_scan_visitdnode(scn, ds, ostype,
6523f9d6ad7SLin Ling 			    cdnp, *bufp, zb->zb_blkid * epb + i, tx);
6533f9d6ad7SLin Ling 		}
6543f9d6ad7SLin Ling 
6553f9d6ad7SLin Ling 	} else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
6563f9d6ad7SLin Ling 		uint32_t flags = ARC_WAIT;
6573f9d6ad7SLin Ling 		objset_phys_t *osp;
6583f9d6ad7SLin Ling 
6591b912ec7SGeorge Wilson 		err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, bufp,
66044ecc532SGeorge Wilson 		    ZIO_PRIORITY_ASYNC_READ, zio_flags, &flags, zb);
6613f9d6ad7SLin Ling 		if (err) {
6623f9d6ad7SLin Ling 			scn->scn_phys.scn_errors++;
6633f9d6ad7SLin Ling 			return (err);
6643f9d6ad7SLin Ling 		}
6653f9d6ad7SLin Ling 
6663f9d6ad7SLin Ling 		osp = (*bufp)->b_data;
6673f9d6ad7SLin Ling 
6683f9d6ad7SLin Ling 		dsl_scan_visitdnode(scn, ds, osp->os_type,
6693f9d6ad7SLin Ling 		    &osp->os_meta_dnode, *bufp, DMU_META_DNODE_OBJECT, tx);
6703f9d6ad7SLin Ling 
6713f9d6ad7SLin Ling 		if (OBJSET_BUF_HAS_USERUSED(*bufp)) {
6723f9d6ad7SLin Ling 			/*
6733f9d6ad7SLin Ling 			 * We also always visit user/group accounting
6743f9d6ad7SLin Ling 			 * objects, and never skip them, even if we are
6753f9d6ad7SLin Ling 			 * pausing.  This is necessary so that the space
6763f9d6ad7SLin Ling 			 * deltas from this txg get integrated.
6773f9d6ad7SLin Ling 			 */
6783f9d6ad7SLin Ling 			dsl_scan_visitdnode(scn, ds, osp->os_type,
6793f9d6ad7SLin Ling 			    &osp->os_groupused_dnode, *bufp,
6803f9d6ad7SLin Ling 			    DMU_GROUPUSED_OBJECT, tx);
6813f9d6ad7SLin Ling 			dsl_scan_visitdnode(scn, ds, osp->os_type,
6823f9d6ad7SLin Ling 			    &osp->os_userused_dnode, *bufp,
6833f9d6ad7SLin Ling 			    DMU_USERUSED_OBJECT, tx);
6843f9d6ad7SLin Ling 		}
6853f9d6ad7SLin Ling 	}
6863f9d6ad7SLin Ling 
6873f9d6ad7SLin Ling 	return (0);
6883f9d6ad7SLin Ling }
6893f9d6ad7SLin Ling 
6903f9d6ad7SLin Ling static void
6913f9d6ad7SLin Ling dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
6923f9d6ad7SLin Ling     dmu_objset_type_t ostype, dnode_phys_t *dnp, arc_buf_t *buf,
6933f9d6ad7SLin Ling     uint64_t object, dmu_tx_t *tx)
6943f9d6ad7SLin Ling {
6953f9d6ad7SLin Ling 	int j;
6963f9d6ad7SLin Ling 
6973f9d6ad7SLin Ling 	for (j = 0; j < dnp->dn_nblkptr; j++) {
698*7802d7bfSMatthew Ahrens 		zbookmark_phys_t czb;
6993f9d6ad7SLin Ling 
7003f9d6ad7SLin Ling 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
7013f9d6ad7SLin Ling 		    dnp->dn_nlevels - 1, j);
7023f9d6ad7SLin Ling 		dsl_scan_visitbp(&dnp->dn_blkptr[j],
7033f9d6ad7SLin Ling 		    &czb, dnp, buf, ds, scn, ostype, tx);
7043f9d6ad7SLin Ling 	}
7053f9d6ad7SLin Ling 
7063f9d6ad7SLin Ling 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
707*7802d7bfSMatthew Ahrens 		zbookmark_phys_t czb;
7083f9d6ad7SLin Ling 		SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
7093f9d6ad7SLin Ling 		    0, DMU_SPILL_BLKID);
7103f9d6ad7SLin Ling 		dsl_scan_visitbp(&dnp->dn_spill,
7113f9d6ad7SLin Ling 		    &czb, dnp, buf, ds, scn, ostype, tx);
7123f9d6ad7SLin Ling 	}
7133f9d6ad7SLin Ling }
7143f9d6ad7SLin Ling 
7153f9d6ad7SLin Ling /*
7163f9d6ad7SLin Ling  * The arguments are in this order because mdb can only print the
7173f9d6ad7SLin Ling  * first 5; we want them to be useful.
7183f9d6ad7SLin Ling  */
7193f9d6ad7SLin Ling static void
720*7802d7bfSMatthew Ahrens dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
7213f9d6ad7SLin Ling     dnode_phys_t *dnp, arc_buf_t *pbuf,
7223f9d6ad7SLin Ling     dsl_dataset_t *ds, dsl_scan_t *scn, dmu_objset_type_t ostype,
7233f9d6ad7SLin Ling     dmu_tx_t *tx)
7243f9d6ad7SLin Ling {
7253f9d6ad7SLin Ling 	dsl_pool_t *dp = scn->scn_dp;
7263f9d6ad7SLin Ling 	arc_buf_t *buf = NULL;
7273f9d6ad7SLin Ling 	blkptr_t bp_toread = *bp;
7283f9d6ad7SLin Ling 
7293f9d6ad7SLin Ling 	/* ASSERT(pbuf == NULL || arc_released(pbuf)); */
7303f9d6ad7SLin Ling 
7313f9d6ad7SLin Ling 	if (dsl_scan_check_pause(scn, zb))
7323f9d6ad7SLin Ling 		return;
7333f9d6ad7SLin Ling 
7343f9d6ad7SLin Ling 	if (dsl_scan_check_resume(scn, dnp, zb))
7353f9d6ad7SLin Ling 		return;
7363f9d6ad7SLin Ling 
73743466aaeSMax Grossman 	if (BP_IS_HOLE(bp))
7383f9d6ad7SLin Ling 		return;
7393f9d6ad7SLin Ling 
7403f9d6ad7SLin Ling 	scn->scn_visited_this_txg++;
7413f9d6ad7SLin Ling 
7423f9d6ad7SLin Ling 	dprintf_bp(bp,
7433f9d6ad7SLin Ling 	    "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx buf=%p bp=%p",
7443f9d6ad7SLin Ling 	    ds, ds ? ds->ds_object : 0,
7453f9d6ad7SLin Ling 	    zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
7463f9d6ad7SLin Ling 	    pbuf, bp);
7473f9d6ad7SLin Ling 
7483f9d6ad7SLin Ling 	if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
7493f9d6ad7SLin Ling 		return;
7503f9d6ad7SLin Ling 
7513f9d6ad7SLin Ling 	if (dsl_scan_recurse(scn, ds, ostype, dnp, &bp_toread, zb, tx,
7523f9d6ad7SLin Ling 	    &buf) != 0)
7533f9d6ad7SLin Ling 		return;
7543f9d6ad7SLin Ling 
7553f9d6ad7SLin Ling 	/*
7563f9d6ad7SLin Ling 	 * If dsl_scan_ddt() has aready visited this block, it will have
7573f9d6ad7SLin Ling 	 * already done any translations or scrubbing, so don't call the
7583f9d6ad7SLin Ling 	 * callback again.
7593f9d6ad7SLin Ling 	 */
7603f9d6ad7SLin Ling 	if (ddt_class_contains(dp->dp_spa,
7613f9d6ad7SLin Ling 	    scn->scn_phys.scn_ddt_class_max, bp)) {
7623f9d6ad7SLin Ling 		ASSERT(buf == NULL);
7633f9d6ad7SLin Ling 		return;
7643f9d6ad7SLin Ling 	}
7653f9d6ad7SLin Ling 
7663f9d6ad7SLin Ling 	/*
7673f9d6ad7SLin Ling 	 * If this block is from the future (after cur_max_txg), then we
7683f9d6ad7SLin Ling 	 * are doing this on behalf of a deleted snapshot, and we will
7693f9d6ad7SLin Ling 	 * revisit the future block on the next pass of this dataset.
7703f9d6ad7SLin Ling 	 * Don't scan it now unless we need to because something
7713f9d6ad7SLin Ling 	 * under it was modified.
7723f9d6ad7SLin Ling 	 */
773b4952e17SGeorge Wilson 	if (BP_PHYSICAL_BIRTH(bp) <= scn->scn_phys.scn_cur_max_txg) {
7743f9d6ad7SLin Ling 		scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
7753f9d6ad7SLin Ling 	}
7763f9d6ad7SLin Ling 	if (buf)
7773f9d6ad7SLin Ling 		(void) arc_buf_remove_ref(buf, &buf);
7783f9d6ad7SLin Ling }
7793f9d6ad7SLin Ling 
7803f9d6ad7SLin Ling static void
7813f9d6ad7SLin Ling dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
7823f9d6ad7SLin Ling     dmu_tx_t *tx)
7833f9d6ad7SLin Ling {
784*7802d7bfSMatthew Ahrens 	zbookmark_phys_t zb;
7853f9d6ad7SLin Ling 
7863f9d6ad7SLin Ling 	SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
7873f9d6ad7SLin Ling 	    ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
7883f9d6ad7SLin Ling 	dsl_scan_visitbp(bp, &zb, NULL, NULL,
7893f9d6ad7SLin Ling 	    ds, scn, DMU_OST_NONE, tx);
7903f9d6ad7SLin Ling 
7913f9d6ad7SLin Ling 	dprintf_ds(ds, "finished scan%s", "");
7923f9d6ad7SLin Ling }
7933f9d6ad7SLin Ling 
7943f9d6ad7SLin Ling void
7953f9d6ad7SLin Ling dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
7963f9d6ad7SLin Ling {
7973f9d6ad7SLin Ling 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
7983f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
7993f9d6ad7SLin Ling 	uint64_t mintxg;
8003f9d6ad7SLin Ling 
8013f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state != DSS_SCANNING)
8023f9d6ad7SLin Ling 		return;
8033f9d6ad7SLin Ling 
8043f9d6ad7SLin Ling 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
8053f9d6ad7SLin Ling 		if (dsl_dataset_is_snapshot(ds)) {
8063f9d6ad7SLin Ling 			/* Note, scn_cur_{min,max}_txg stays the same. */
8073f9d6ad7SLin Ling 			scn->scn_phys.scn_bookmark.zb_objset =
8083f9d6ad7SLin Ling 			    ds->ds_phys->ds_next_snap_obj;
8093f9d6ad7SLin Ling 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
8103f9d6ad7SLin Ling 			    "reset zb_objset to %llu",
8113f9d6ad7SLin Ling 			    (u_longlong_t)ds->ds_object,
8123f9d6ad7SLin Ling 			    (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
8133f9d6ad7SLin Ling 			scn->scn_phys.scn_flags |= DSF_VISIT_DS_AGAIN;
8143f9d6ad7SLin Ling 		} else {
8153f9d6ad7SLin Ling 			SET_BOOKMARK(&scn->scn_phys.scn_bookmark,
8163f9d6ad7SLin Ling 			    ZB_DESTROYED_OBJSET, 0, 0, 0);
8173f9d6ad7SLin Ling 			zfs_dbgmsg("destroying ds %llu; currently traversing; "
8183f9d6ad7SLin Ling 			    "reset bookmark to -1,0,0,0",
8193f9d6ad7SLin Ling 			    (u_longlong_t)ds->ds_object);
8203f9d6ad7SLin Ling 		}
8213f9d6ad7SLin Ling 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
8223f9d6ad7SLin Ling 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
8233f9d6ad7SLin Ling 		ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
824b420f3adSRichard Lowe 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
8253f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
8263f9d6ad7SLin Ling 		if (dsl_dataset_is_snapshot(ds)) {
8273f9d6ad7SLin Ling 			/*
8283f9d6ad7SLin Ling 			 * We keep the same mintxg; it could be >
8293f9d6ad7SLin Ling 			 * ds_creation_txg if the previous snapshot was
8303f9d6ad7SLin Ling 			 * deleted too.
8313f9d6ad7SLin Ling 			 */
8323f9d6ad7SLin Ling 			VERIFY(zap_add_int_key(dp->dp_meta_objset,
8333f9d6ad7SLin Ling 			    scn->scn_phys.scn_queue_obj,
8343f9d6ad7SLin Ling 			    ds->ds_phys->ds_next_snap_obj, mintxg, tx) == 0);
8353f9d6ad7SLin Ling 			zfs_dbgmsg("destroying ds %llu; in queue; "
8363f9d6ad7SLin Ling 			    "replacing with %llu",
8373f9d6ad7SLin Ling 			    (u_longlong_t)ds->ds_object,
8383f9d6ad7SLin Ling 			    (u_longlong_t)ds->ds_phys->ds_next_snap_obj);
8393f9d6ad7SLin Ling 		} else {
8403f9d6ad7SLin Ling 			zfs_dbgmsg("destroying ds %llu; in queue; removing",
8413f9d6ad7SLin Ling 			    (u_longlong_t)ds->ds_object);
8423f9d6ad7SLin Ling 		}
8433f9d6ad7SLin Ling 	} else {
8443f9d6ad7SLin Ling 		zfs_dbgmsg("destroying ds %llu; ignoring",
8453f9d6ad7SLin Ling 		    (u_longlong_t)ds->ds_object);
8463f9d6ad7SLin Ling 	}
8473f9d6ad7SLin Ling 
8483f9d6ad7SLin Ling 	/*
8493f9d6ad7SLin Ling 	 * dsl_scan_sync() should be called after this, and should sync
8503f9d6ad7SLin Ling 	 * out our changed state, but just to be safe, do it here.
8513f9d6ad7SLin Ling 	 */
8523f9d6ad7SLin Ling 	dsl_scan_sync_state(scn, tx);
8533f9d6ad7SLin Ling }
8543f9d6ad7SLin Ling 
8553f9d6ad7SLin Ling void
8563f9d6ad7SLin Ling dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
8573f9d6ad7SLin Ling {
8583f9d6ad7SLin Ling 	dsl_pool_t *dp = ds->ds_dir->dd_pool;
8593f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
8603f9d6ad7SLin Ling 	uint64_t mintxg;
8613f9d6ad7SLin Ling 
8623f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state != DSS_SCANNING)
8633f9d6ad7SLin Ling 		return;
8643f9d6ad7SLin Ling 
8653f9d6ad7SLin Ling 	ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
8663f9d6ad7SLin Ling 
8673f9d6ad7SLin Ling 	if (scn->scn_phys.scn_bookmark.zb_objset == ds->ds_object) {
8683f9d6ad7SLin Ling 		scn->scn_phys.scn_bookmark.zb_objset =
8693f9d6ad7SLin Ling 		    ds->ds_phys->ds_prev_snap_obj;
8703f9d6ad7SLin Ling 		zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
8713f9d6ad7SLin Ling 		    "reset zb_objset to %llu",
8723f9d6ad7SLin Ling 		    (u_longlong_t)ds->ds_object,
8733f9d6ad7SLin Ling 		    (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
8743f9d6ad7SLin Ling 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
8753f9d6ad7SLin Ling 	    scn->scn_phys.scn_queue_obj, ds->ds_object, &mintxg) == 0) {
876b420f3adSRichard Lowe 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
8773f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
8783f9d6ad7SLin Ling 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
8793f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj,
8803f9d6ad7SLin Ling 		    ds->ds_phys->ds_prev_snap_obj, mintxg, tx) == 0);
8813f9d6ad7SLin Ling 		zfs_dbgmsg("snapshotting ds %llu; in queue; "
8823f9d6ad7SLin Ling 		    "replacing with %llu",
8833f9d6ad7SLin Ling 		    (u_longlong_t)ds->ds_object,
8843f9d6ad7SLin Ling 		    (u_longlong_t)ds->ds_phys->ds_prev_snap_obj);
8853f9d6ad7SLin Ling 	}
8863f9d6ad7SLin Ling 	dsl_scan_sync_state(scn, tx);
8873f9d6ad7SLin Ling }
8883f9d6ad7SLin Ling 
8893f9d6ad7SLin Ling void
8903f9d6ad7SLin Ling dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
8913f9d6ad7SLin Ling {
8923f9d6ad7SLin Ling 	dsl_pool_t *dp = ds1->ds_dir->dd_pool;
8933f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
8943f9d6ad7SLin Ling 	uint64_t mintxg;
8953f9d6ad7SLin Ling 
8963f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state != DSS_SCANNING)
8973f9d6ad7SLin Ling 		return;
8983f9d6ad7SLin Ling 
8993f9d6ad7SLin Ling 	if (scn->scn_phys.scn_bookmark.zb_objset == ds1->ds_object) {
9003f9d6ad7SLin Ling 		scn->scn_phys.scn_bookmark.zb_objset = ds2->ds_object;
9013f9d6ad7SLin Ling 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
9023f9d6ad7SLin Ling 		    "reset zb_objset to %llu",
9033f9d6ad7SLin Ling 		    (u_longlong_t)ds1->ds_object,
9043f9d6ad7SLin Ling 		    (u_longlong_t)ds2->ds_object);
9053f9d6ad7SLin Ling 	} else if (scn->scn_phys.scn_bookmark.zb_objset == ds2->ds_object) {
9063f9d6ad7SLin Ling 		scn->scn_phys.scn_bookmark.zb_objset = ds1->ds_object;
9073f9d6ad7SLin Ling 		zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
9083f9d6ad7SLin Ling 		    "reset zb_objset to %llu",
9093f9d6ad7SLin Ling 		    (u_longlong_t)ds2->ds_object,
9103f9d6ad7SLin Ling 		    (u_longlong_t)ds1->ds_object);
9113f9d6ad7SLin Ling 	}
9123f9d6ad7SLin Ling 
9133f9d6ad7SLin Ling 	if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
9143f9d6ad7SLin Ling 	    ds1->ds_object, &mintxg) == 0) {
9153f9d6ad7SLin Ling 		int err;
9163f9d6ad7SLin Ling 
9173f9d6ad7SLin Ling 		ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
9183f9d6ad7SLin Ling 		ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
919b420f3adSRichard Lowe 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
9203f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
9213f9d6ad7SLin Ling 		err = zap_add_int_key(dp->dp_meta_objset,
9223f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
9233f9d6ad7SLin Ling 		VERIFY(err == 0 || err == EEXIST);
9243f9d6ad7SLin Ling 		if (err == EEXIST) {
9253f9d6ad7SLin Ling 			/* Both were there to begin with */
9263f9d6ad7SLin Ling 			VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
9273f9d6ad7SLin Ling 			    scn->scn_phys.scn_queue_obj,
9283f9d6ad7SLin Ling 			    ds1->ds_object, mintxg, tx));
9293f9d6ad7SLin Ling 		}
9303f9d6ad7SLin Ling 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
9313f9d6ad7SLin Ling 		    "replacing with %llu",
9323f9d6ad7SLin Ling 		    (u_longlong_t)ds1->ds_object,
9333f9d6ad7SLin Ling 		    (u_longlong_t)ds2->ds_object);
9343f9d6ad7SLin Ling 	} else if (zap_lookup_int_key(dp->dp_meta_objset,
9353f9d6ad7SLin Ling 	    scn->scn_phys.scn_queue_obj, ds2->ds_object, &mintxg) == 0) {
9363f9d6ad7SLin Ling 		ASSERT3U(mintxg, ==, ds1->ds_phys->ds_prev_snap_txg);
9373f9d6ad7SLin Ling 		ASSERT3U(mintxg, ==, ds2->ds_phys->ds_prev_snap_txg);
938b420f3adSRichard Lowe 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
9393f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
9403f9d6ad7SLin Ling 		VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
9413f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
9423f9d6ad7SLin Ling 		zfs_dbgmsg("clone_swap ds %llu; in queue; "
9433f9d6ad7SLin Ling 		    "replacing with %llu",
9443f9d6ad7SLin Ling 		    (u_longlong_t)ds2->ds_object,
9453f9d6ad7SLin Ling 		    (u_longlong_t)ds1->ds_object);
9463f9d6ad7SLin Ling 	}
9473f9d6ad7SLin Ling 
9483f9d6ad7SLin Ling 	dsl_scan_sync_state(scn, tx);
9493f9d6ad7SLin Ling }
9503f9d6ad7SLin Ling 
9513f9d6ad7SLin Ling struct enqueue_clones_arg {
9523f9d6ad7SLin Ling 	dmu_tx_t *tx;
9533f9d6ad7SLin Ling 	uint64_t originobj;
9543f9d6ad7SLin Ling };
9553f9d6ad7SLin Ling 
9563f9d6ad7SLin Ling /* ARGSUSED */
9573f9d6ad7SLin Ling static int
9583b2aab18SMatthew Ahrens enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
9593f9d6ad7SLin Ling {
9603f9d6ad7SLin Ling 	struct enqueue_clones_arg *eca = arg;
9613f9d6ad7SLin Ling 	dsl_dataset_t *ds;
9623f9d6ad7SLin Ling 	int err;
9633f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
9643f9d6ad7SLin Ling 
9653b2aab18SMatthew Ahrens 	if (hds->ds_dir->dd_phys->dd_origin_obj != eca->originobj)
9663b2aab18SMatthew Ahrens 		return (0);
9673b2aab18SMatthew Ahrens 
9683b2aab18SMatthew Ahrens 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
9693f9d6ad7SLin Ling 	if (err)
9703f9d6ad7SLin Ling 		return (err);
9713f9d6ad7SLin Ling 
9723b2aab18SMatthew Ahrens 	while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
9733b2aab18SMatthew Ahrens 		dsl_dataset_t *prev;
9743b2aab18SMatthew Ahrens 		err = dsl_dataset_hold_obj(dp,
9753b2aab18SMatthew Ahrens 		    ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
9763f9d6ad7SLin Ling 
9773b2aab18SMatthew Ahrens 		dsl_dataset_rele(ds, FTAG);
9783b2aab18SMatthew Ahrens 		if (err)
9793b2aab18SMatthew Ahrens 			return (err);
9803b2aab18SMatthew Ahrens 		ds = prev;
9813f9d6ad7SLin Ling 	}
9823b2aab18SMatthew Ahrens 	VERIFY(zap_add_int_key(dp->dp_meta_objset,
9833b2aab18SMatthew Ahrens 	    scn->scn_phys.scn_queue_obj, ds->ds_object,
9843b2aab18SMatthew Ahrens 	    ds->ds_phys->ds_prev_snap_txg, eca->tx) == 0);
9853f9d6ad7SLin Ling 	dsl_dataset_rele(ds, FTAG);
9863f9d6ad7SLin Ling 	return (0);
9873f9d6ad7SLin Ling }
9883f9d6ad7SLin Ling 
9893f9d6ad7SLin Ling static void
9903f9d6ad7SLin Ling dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
9913f9d6ad7SLin Ling {
9923f9d6ad7SLin Ling 	dsl_pool_t *dp = scn->scn_dp;
9933f9d6ad7SLin Ling 	dsl_dataset_t *ds;
9946e0cbcaaSMatthew Ahrens 	objset_t *os;
9953f9d6ad7SLin Ling 
996b420f3adSRichard Lowe 	VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
9973f9d6ad7SLin Ling 
9986e0cbcaaSMatthew Ahrens 	if (dmu_objset_from_ds(ds, &os))
9996e0cbcaaSMatthew Ahrens 		goto out;
10006e0cbcaaSMatthew Ahrens 
10016e0cbcaaSMatthew Ahrens 	/*
10026e0cbcaaSMatthew Ahrens 	 * Only the ZIL in the head (non-snapshot) is valid.  Even though
10036e0cbcaaSMatthew Ahrens 	 * snapshots can have ZIL block pointers (which may be the same
10046e0cbcaaSMatthew Ahrens 	 * BP as in the head), they must be ignored.  So we traverse the
10056e0cbcaaSMatthew Ahrens 	 * ZIL here, rather than in scan_recurse(), because the regular
10066e0cbcaaSMatthew Ahrens 	 * snapshot block-sharing rules don't apply to it.
10076e0cbcaaSMatthew Ahrens 	 */
10086e0cbcaaSMatthew Ahrens 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds))
10096e0cbcaaSMatthew Ahrens 		dsl_scan_zil(dp, &os->os_zil_header);
10106e0cbcaaSMatthew Ahrens 
10113f9d6ad7SLin Ling 	/*
10123f9d6ad7SLin Ling 	 * Iterate over the bps in this ds.
10133f9d6ad7SLin Ling 	 */
10143f9d6ad7SLin Ling 	dmu_buf_will_dirty(ds->ds_dbuf, tx);
10153f9d6ad7SLin Ling 	dsl_scan_visit_rootbp(scn, ds, &ds->ds_phys->ds_bp, tx);
10163f9d6ad7SLin Ling 
10173f9d6ad7SLin Ling 	char *dsname = kmem_alloc(ZFS_MAXNAMELEN, KM_SLEEP);
10183f9d6ad7SLin Ling 	dsl_dataset_name(ds, dsname);
10193f9d6ad7SLin Ling 	zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
10203f9d6ad7SLin Ling 	    "pausing=%u",
10213f9d6ad7SLin Ling 	    (longlong_t)dsobj, dsname,
10223f9d6ad7SLin Ling 	    (longlong_t)scn->scn_phys.scn_cur_min_txg,
10233f9d6ad7SLin Ling 	    (longlong_t)scn->scn_phys.scn_cur_max_txg,
10243f9d6ad7SLin Ling 	    (int)scn->scn_pausing);
10253f9d6ad7SLin Ling 	kmem_free(dsname, ZFS_MAXNAMELEN);
10263f9d6ad7SLin Ling 
10273f9d6ad7SLin Ling 	if (scn->scn_pausing)
10283f9d6ad7SLin Ling 		goto out;
10293f9d6ad7SLin Ling 
10303f9d6ad7SLin Ling 	/*
10313f9d6ad7SLin Ling 	 * We've finished this pass over this dataset.
10323f9d6ad7SLin Ling 	 */
10333f9d6ad7SLin Ling 
10343f9d6ad7SLin Ling 	/*
10353f9d6ad7SLin Ling 	 * If we did not completely visit this dataset, do another pass.
10363f9d6ad7SLin Ling 	 */
10373f9d6ad7SLin Ling 	if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
10383f9d6ad7SLin Ling 		zfs_dbgmsg("incomplete pass; visiting again");
10393f9d6ad7SLin Ling 		scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
10403f9d6ad7SLin Ling 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
10413f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds->ds_object,
10423f9d6ad7SLin Ling 		    scn->scn_phys.scn_cur_max_txg, tx) == 0);
10433f9d6ad7SLin Ling 		goto out;
10443f9d6ad7SLin Ling 	}
10453f9d6ad7SLin Ling 
10463f9d6ad7SLin Ling 	/*
10473f9d6ad7SLin Ling 	 * Add descendent datasets to work queue.
10483f9d6ad7SLin Ling 	 */
10493f9d6ad7SLin Ling 	if (ds->ds_phys->ds_next_snap_obj != 0) {
10503f9d6ad7SLin Ling 		VERIFY(zap_add_int_key(dp->dp_meta_objset,
10513f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, ds->ds_phys->ds_next_snap_obj,
10523f9d6ad7SLin Ling 		    ds->ds_phys->ds_creation_txg, tx) == 0);
10533f9d6ad7SLin Ling 	}
10543f9d6ad7SLin Ling 	if (ds->ds_phys->ds_num_children > 1) {
10553f9d6ad7SLin Ling 		boolean_t usenext = B_FALSE;
10563f9d6ad7SLin Ling 		if (ds->ds_phys->ds_next_clones_obj != 0) {
10573f9d6ad7SLin Ling 			uint64_t count;
10583f9d6ad7SLin Ling 			/*
10593f9d6ad7SLin Ling 			 * A bug in a previous version of the code could
10603f9d6ad7SLin Ling 			 * cause upgrade_clones_cb() to not set
10613f9d6ad7SLin Ling 			 * ds_next_snap_obj when it should, leading to a
10623f9d6ad7SLin Ling 			 * missing entry.  Therefore we can only use the
10633f9d6ad7SLin Ling 			 * next_clones_obj when its count is correct.
10643f9d6ad7SLin Ling 			 */
10653f9d6ad7SLin Ling 			int err = zap_count(dp->dp_meta_objset,
10663f9d6ad7SLin Ling 			    ds->ds_phys->ds_next_clones_obj, &count);
10673f9d6ad7SLin Ling 			if (err == 0 &&
10683f9d6ad7SLin Ling 			    count == ds->ds_phys->ds_num_children - 1)
10693f9d6ad7SLin Ling 				usenext = B_TRUE;
10703f9d6ad7SLin Ling 		}
10713f9d6ad7SLin Ling 
10723f9d6ad7SLin Ling 		if (usenext) {
10733b2aab18SMatthew Ahrens 			VERIFY0(zap_join_key(dp->dp_meta_objset,
10743f9d6ad7SLin Ling 			    ds->ds_phys->ds_next_clones_obj,
10753f9d6ad7SLin Ling 			    scn->scn_phys.scn_queue_obj,
10763b2aab18SMatthew Ahrens 			    ds->ds_phys->ds_creation_txg, tx));
10773f9d6ad7SLin Ling 		} else {
10783f9d6ad7SLin Ling 			struct enqueue_clones_arg eca;
10793f9d6ad7SLin Ling 			eca.tx = tx;
10803f9d6ad7SLin Ling 			eca.originobj = ds->ds_object;
10813f9d6ad7SLin Ling 
10823b2aab18SMatthew Ahrens 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
10833b2aab18SMatthew Ahrens 			    enqueue_clones_cb, &eca, DS_FIND_CHILDREN));
10843f9d6ad7SLin Ling 		}
10853f9d6ad7SLin Ling 	}
10863f9d6ad7SLin Ling 
10873f9d6ad7SLin Ling out:
10883f9d6ad7SLin Ling 	dsl_dataset_rele(ds, FTAG);
10893f9d6ad7SLin Ling }
10903f9d6ad7SLin Ling 
10913f9d6ad7SLin Ling /* ARGSUSED */
10923f9d6ad7SLin Ling static int
10933b2aab18SMatthew Ahrens enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
10943f9d6ad7SLin Ling {
10953f9d6ad7SLin Ling 	dmu_tx_t *tx = arg;
10963f9d6ad7SLin Ling 	dsl_dataset_t *ds;
10973f9d6ad7SLin Ling 	int err;
10983f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
10993f9d6ad7SLin Ling 
11003b2aab18SMatthew Ahrens 	err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
11013f9d6ad7SLin Ling 	if (err)
11023f9d6ad7SLin Ling 		return (err);
11033f9d6ad7SLin Ling 
11043f9d6ad7SLin Ling 	while (ds->ds_phys->ds_prev_snap_obj != 0) {
11053f9d6ad7SLin Ling 		dsl_dataset_t *prev;
11063f9d6ad7SLin Ling 		err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
11073f9d6ad7SLin Ling 		    FTAG, &prev);
11083f9d6ad7SLin Ling 		if (err) {
11093f9d6ad7SLin Ling 			dsl_dataset_rele(ds, FTAG);
11103f9d6ad7SLin Ling 			return (err);
11113f9d6ad7SLin Ling 		}
11123f9d6ad7SLin Ling 
11133f9d6ad7SLin Ling 		/*
11143f9d6ad7SLin Ling 		 * If this is a clone, we don't need to worry about it for now.
11153f9d6ad7SLin Ling 		 */
11163f9d6ad7SLin Ling 		if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
11173f9d6ad7SLin Ling 			dsl_dataset_rele(ds, FTAG);
11183f9d6ad7SLin Ling 			dsl_dataset_rele(prev, FTAG);
11193f9d6ad7SLin Ling 			return (0);
11203f9d6ad7SLin Ling 		}
11213f9d6ad7SLin Ling 		dsl_dataset_rele(ds, FTAG);
11223f9d6ad7SLin Ling 		ds = prev;
11233f9d6ad7SLin Ling 	}
11243f9d6ad7SLin Ling 
11253f9d6ad7SLin Ling 	VERIFY(zap_add_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
11263f9d6ad7SLin Ling 	    ds->ds_object, ds->ds_phys->ds_prev_snap_txg, tx) == 0);
11273f9d6ad7SLin Ling 	dsl_dataset_rele(ds, FTAG);
11283f9d6ad7SLin Ling 	return (0);
11293f9d6ad7SLin Ling }
11303f9d6ad7SLin Ling 
11313f9d6ad7SLin Ling /*
11323f9d6ad7SLin Ling  * Scrub/dedup interaction.
11333f9d6ad7SLin Ling  *
11343f9d6ad7SLin Ling  * If there are N references to a deduped block, we don't want to scrub it
11353f9d6ad7SLin Ling  * N times -- ideally, we should scrub it exactly once.
11363f9d6ad7SLin Ling  *
11373f9d6ad7SLin Ling  * We leverage the fact that the dde's replication class (enum ddt_class)
11383f9d6ad7SLin Ling  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
11393f9d6ad7SLin Ling  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
11403f9d6ad7SLin Ling  *
11413f9d6ad7SLin Ling  * To prevent excess scrubbing, the scrub begins by walking the DDT
11423f9d6ad7SLin Ling  * to find all blocks with refcnt > 1, and scrubs each of these once.
11433f9d6ad7SLin Ling  * Since there are two replication classes which contain blocks with
11443f9d6ad7SLin Ling  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
11453f9d6ad7SLin Ling  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
11463f9d6ad7SLin Ling  *
11473f9d6ad7SLin Ling  * There would be nothing more to say if a block's refcnt couldn't change
11483f9d6ad7SLin Ling  * during a scrub, but of course it can so we must account for changes
11493f9d6ad7SLin Ling  * in a block's replication class.
11503f9d6ad7SLin Ling  *
11513f9d6ad7SLin Ling  * Here's an example of what can occur:
11523f9d6ad7SLin Ling  *
11533f9d6ad7SLin Ling  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
11543f9d6ad7SLin Ling  * when visited during the top-down scrub phase, it will be scrubbed twice.
11553f9d6ad7SLin Ling  * This negates our scrub optimization, but is otherwise harmless.
11563f9d6ad7SLin Ling  *
11573f9d6ad7SLin Ling  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
11583f9d6ad7SLin Ling  * on each visit during the top-down scrub phase, it will never be scrubbed.
11593f9d6ad7SLin Ling  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
11603f9d6ad7SLin Ling  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
11613f9d6ad7SLin Ling  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
11623f9d6ad7SLin Ling  * while a scrub is in progress, it scrubs the block right then.
11633f9d6ad7SLin Ling  */
11643f9d6ad7SLin Ling static void
11653f9d6ad7SLin Ling dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
11663f9d6ad7SLin Ling {
11673f9d6ad7SLin Ling 	ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
11683f9d6ad7SLin Ling 	ddt_entry_t dde = { 0 };
11693f9d6ad7SLin Ling 	int error;
11703f9d6ad7SLin Ling 	uint64_t n = 0;
11713f9d6ad7SLin Ling 
11723f9d6ad7SLin Ling 	while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
11733f9d6ad7SLin Ling 		ddt_t *ddt;
11743f9d6ad7SLin Ling 
11753f9d6ad7SLin Ling 		if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
11763f9d6ad7SLin Ling 			break;
11773f9d6ad7SLin Ling 		dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
11783f9d6ad7SLin Ling 		    (longlong_t)ddb->ddb_class,
11793f9d6ad7SLin Ling 		    (longlong_t)ddb->ddb_type,
11803f9d6ad7SLin Ling 		    (longlong_t)ddb->ddb_checksum,
11813f9d6ad7SLin Ling 		    (longlong_t)ddb->ddb_cursor);
11823f9d6ad7SLin Ling 
11833f9d6ad7SLin Ling 		/* There should be no pending changes to the dedup table */
11843f9d6ad7SLin Ling 		ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
11853f9d6ad7SLin Ling 		ASSERT(avl_first(&ddt->ddt_tree) == NULL);
11863f9d6ad7SLin Ling 
11873f9d6ad7SLin Ling 		dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
11883f9d6ad7SLin Ling 		n++;
11893f9d6ad7SLin Ling 
11903f9d6ad7SLin Ling 		if (dsl_scan_check_pause(scn, NULL))
11913f9d6ad7SLin Ling 			break;
11923f9d6ad7SLin Ling 	}
11933f9d6ad7SLin Ling 
11943f9d6ad7SLin Ling 	zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; pausing=%u",
11953f9d6ad7SLin Ling 	    (longlong_t)n, (int)scn->scn_phys.scn_ddt_class_max,
11963f9d6ad7SLin Ling 	    (int)scn->scn_pausing);
11973f9d6ad7SLin Ling 
11983f9d6ad7SLin Ling 	ASSERT(error == 0 || error == ENOENT);
11993f9d6ad7SLin Ling 	ASSERT(error != ENOENT ||
12003f9d6ad7SLin Ling 	    ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
12013f9d6ad7SLin Ling }
12023f9d6ad7SLin Ling 
12033f9d6ad7SLin Ling /* ARGSUSED */
12043f9d6ad7SLin Ling void
12053f9d6ad7SLin Ling dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
12063f9d6ad7SLin Ling     ddt_entry_t *dde, dmu_tx_t *tx)
12073f9d6ad7SLin Ling {
12083f9d6ad7SLin Ling 	const ddt_key_t *ddk = &dde->dde_key;
12093f9d6ad7SLin Ling 	ddt_phys_t *ddp = dde->dde_phys;
12103f9d6ad7SLin Ling 	blkptr_t bp;
1211*7802d7bfSMatthew Ahrens 	zbookmark_phys_t zb = { 0 };
12123f9d6ad7SLin Ling 
12133f9d6ad7SLin Ling 	if (scn->scn_phys.scn_state != DSS_SCANNING)
12143f9d6ad7SLin Ling 		return;
12153f9d6ad7SLin Ling 
12163f9d6ad7SLin Ling 	for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
12173f9d6ad7SLin Ling 		if (ddp->ddp_phys_birth == 0 ||
1218b4952e17SGeorge Wilson 		    ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
12193f9d6ad7SLin Ling 			continue;
12203f9d6ad7SLin Ling 		ddt_bp_create(checksum, ddk, ddp, &bp);
12213f9d6ad7SLin Ling 
12223f9d6ad7SLin Ling 		scn->scn_visited_this_txg++;
12233f9d6ad7SLin Ling 		scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
12243f9d6ad7SLin Ling 	}
12253f9d6ad7SLin Ling }
12263f9d6ad7SLin Ling 
12273f9d6ad7SLin Ling static void
12283f9d6ad7SLin Ling dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
12293f9d6ad7SLin Ling {
12303f9d6ad7SLin Ling 	dsl_pool_t *dp = scn->scn_dp;
12313f9d6ad7SLin Ling 	zap_cursor_t zc;
12323f9d6ad7SLin Ling 	zap_attribute_t za;
12333f9d6ad7SLin Ling 
12343f9d6ad7SLin Ling 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
12353f9d6ad7SLin Ling 	    scn->scn_phys.scn_ddt_class_max) {
12363f9d6ad7SLin Ling 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
12373f9d6ad7SLin Ling 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
12383f9d6ad7SLin Ling 		dsl_scan_ddt(scn, tx);
12393f9d6ad7SLin Ling 		if (scn->scn_pausing)
12403f9d6ad7SLin Ling 			return;
12413f9d6ad7SLin Ling 	}
12423f9d6ad7SLin Ling 
12433f9d6ad7SLin Ling 	if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
12443f9d6ad7SLin Ling 		/* First do the MOS & ORIGIN */
12453f9d6ad7SLin Ling 
12463f9d6ad7SLin Ling 		scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
12473f9d6ad7SLin Ling 		scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
12483f9d6ad7SLin Ling 		dsl_scan_visit_rootbp(scn, NULL,
12493f9d6ad7SLin Ling 		    &dp->dp_meta_rootbp, tx);
12503f9d6ad7SLin Ling 		spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
12513f9d6ad7SLin Ling 		if (scn->scn_pausing)
12523f9d6ad7SLin Ling 			return;
12533f9d6ad7SLin Ling 
12543f9d6ad7SLin Ling 		if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
12553b2aab18SMatthew Ahrens 			VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
12563b2aab18SMatthew Ahrens 			    enqueue_cb, tx, DS_FIND_CHILDREN));
12573f9d6ad7SLin Ling 		} else {
12583f9d6ad7SLin Ling 			dsl_scan_visitds(scn,
12593f9d6ad7SLin Ling 			    dp->dp_origin_snap->ds_object, tx);
12603f9d6ad7SLin Ling 		}
12613f9d6ad7SLin Ling 		ASSERT(!scn->scn_pausing);
12623f9d6ad7SLin Ling 	} else if (scn->scn_phys.scn_bookmark.zb_objset !=
12633f9d6ad7SLin Ling 	    ZB_DESTROYED_OBJSET) {
12643f9d6ad7SLin Ling 		/*
12653f9d6ad7SLin Ling 		 * If we were paused, continue from here.  Note if the
12663f9d6ad7SLin Ling 		 * ds we were paused on was deleted, the zb_objset may
12673f9d6ad7SLin Ling 		 * be -1, so we will skip this and find a new objset
12683f9d6ad7SLin Ling 		 * below.
12693f9d6ad7SLin Ling 		 */
12703f9d6ad7SLin Ling 		dsl_scan_visitds(scn, scn->scn_phys.scn_bookmark.zb_objset, tx);
12713f9d6ad7SLin Ling 		if (scn->scn_pausing)
12723f9d6ad7SLin Ling 			return;
12733f9d6ad7SLin Ling 	}
12743f9d6ad7SLin Ling 
12753f9d6ad7SLin Ling 	/*
12763f9d6ad7SLin Ling 	 * In case we were paused right at the end of the ds, zero the
12773f9d6ad7SLin Ling 	 * bookmark so we don't think that we're still trying to resume.
12783f9d6ad7SLin Ling 	 */
1279*7802d7bfSMatthew Ahrens 	bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
12803f9d6ad7SLin Ling 
12813f9d6ad7SLin Ling 	/* keep pulling things out of the zap-object-as-queue */
12823f9d6ad7SLin Ling 	while (zap_cursor_init(&zc, dp->dp_meta_objset,
12833f9d6ad7SLin Ling 	    scn->scn_phys.scn_queue_obj),
12843f9d6ad7SLin Ling 	    zap_cursor_retrieve(&zc, &za) == 0) {
12853f9d6ad7SLin Ling 		dsl_dataset_t *ds;
12863f9d6ad7SLin Ling 		uint64_t dsobj;
12873f9d6ad7SLin Ling 
12883f9d6ad7SLin Ling 		dsobj = strtonum(za.za_name, NULL);
1289b420f3adSRichard Lowe 		VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
12903f9d6ad7SLin Ling 		    scn->scn_phys.scn_queue_obj, dsobj, tx));
12913f9d6ad7SLin Ling 
12923f9d6ad7SLin Ling 		/* Set up min/max txg */
1293b420f3adSRichard Lowe 		VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
12943f9d6ad7SLin Ling 		if (za.za_first_integer != 0) {
12953f9d6ad7SLin Ling 			scn->scn_phys.scn_cur_min_txg =
12963f9d6ad7SLin Ling 			    MAX(scn->scn_phys.scn_min_txg,
12973f9d6ad7SLin Ling 			    za.za_first_integer);
12983f9d6ad7SLin Ling 		} else {
12993f9d6ad7SLin Ling 			scn->scn_phys.scn_cur_min_txg =
13003f9d6ad7SLin Ling 			    MAX(scn->scn_phys.scn_min_txg,
13013f9d6ad7SLin Ling 			    ds->ds_phys->ds_prev_snap_txg);
13023f9d6ad7SLin Ling 		}
13033f9d6ad7SLin Ling 		scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
13043f9d6ad7SLin Ling 		dsl_dataset_rele(ds, FTAG);
13053f9d6ad7SLin Ling 
13063f9d6ad7SLin Ling 		dsl_scan_visitds(scn, dsobj, tx);
13073f9d6ad7SLin Ling 		zap_cursor_fini(&zc);
13083f9d6ad7SLin Ling 		if (scn->scn_pausing)
13093f9d6ad7SLin Ling 			return;
13103f9d6ad7SLin Ling 	}
13113f9d6ad7SLin Ling 	zap_cursor_fini(&zc);
13123f9d6ad7SLin Ling }
13133f9d6ad7SLin Ling 
1314ad135b5dSChristopher Siden static boolean_t
1315ad135b5dSChristopher Siden dsl_scan_free_should_pause(dsl_scan_t *scn)
1316cde58dbcSMatthew Ahrens {
1317cde58dbcSMatthew Ahrens 	uint64_t elapsed_nanosecs;
1318cde58dbcSMatthew Ahrens 
13198b36997aSMatthew Ahrens 	if (zfs_recover)
13208b36997aSMatthew Ahrens 		return (B_FALSE);
13218b36997aSMatthew Ahrens 
1322cde58dbcSMatthew Ahrens 	elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
1323ad135b5dSChristopher Siden 	return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
13240689f76cSAdam Leventhal 	    (NSEC2MSEC(elapsed_nanosecs) > zfs_free_min_time_ms &&
1325cde58dbcSMatthew Ahrens 	    txg_sync_waiting(scn->scn_dp)) ||
1326ad135b5dSChristopher Siden 	    spa_shutting_down(scn->scn_dp->dp_spa));
1327ad135b5dSChristopher Siden }
1328ad135b5dSChristopher Siden 
1329ad135b5dSChristopher Siden static int
1330ad135b5dSChristopher Siden dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
1331ad135b5dSChristopher Siden {
1332ad135b5dSChristopher Siden 	dsl_scan_t *scn = arg;
1333ad135b5dSChristopher Siden 
1334ad135b5dSChristopher Siden 	if (!scn->scn_is_bptree ||
1335ad135b5dSChristopher Siden 	    (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
1336ad135b5dSChristopher Siden 		if (dsl_scan_free_should_pause(scn))
1337be6fd75aSMatthew Ahrens 			return (SET_ERROR(ERESTART));
1338ad135b5dSChristopher Siden 	}
1339cde58dbcSMatthew Ahrens 
1340cde58dbcSMatthew Ahrens 	zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
1341cde58dbcSMatthew Ahrens 	    dmu_tx_get_txg(tx), bp, 0));
1342cde58dbcSMatthew Ahrens 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
1343cde58dbcSMatthew Ahrens 	    -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
1344cde58dbcSMatthew Ahrens 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
1345cde58dbcSMatthew Ahrens 	scn->scn_visited_this_txg++;
1346cde58dbcSMatthew Ahrens 	return (0);
1347cde58dbcSMatthew Ahrens }
1348cde58dbcSMatthew Ahrens 
1349cde58dbcSMatthew Ahrens boolean_t
1350cde58dbcSMatthew Ahrens dsl_scan_active(dsl_scan_t *scn)
1351cde58dbcSMatthew Ahrens {
1352cde58dbcSMatthew Ahrens 	spa_t *spa = scn->scn_dp->dp_spa;
1353cde58dbcSMatthew Ahrens 	uint64_t used = 0, comp, uncomp;
1354cde58dbcSMatthew Ahrens 
1355cde58dbcSMatthew Ahrens 	if (spa->spa_load_state != SPA_LOAD_NONE)
1356cde58dbcSMatthew Ahrens 		return (B_FALSE);
1357cde58dbcSMatthew Ahrens 	if (spa_shutting_down(spa))
1358cde58dbcSMatthew Ahrens 		return (B_FALSE);
13594a923759SGeorge Wilson 	if (scn->scn_phys.scn_state == DSS_SCANNING ||
13607fd05ac4SMatthew Ahrens 	    (scn->scn_async_destroying && !scn->scn_async_stalled))
1361cde58dbcSMatthew Ahrens 		return (B_TRUE);
1362cde58dbcSMatthew Ahrens 
1363cde58dbcSMatthew Ahrens 	if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1364cde58dbcSMatthew Ahrens 		(void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
1365cde58dbcSMatthew Ahrens 		    &used, &comp, &uncomp);
1366cde58dbcSMatthew Ahrens 	}
1367cde58dbcSMatthew Ahrens 	return (used != 0);
1368cde58dbcSMatthew Ahrens }
1369cde58dbcSMatthew Ahrens 
13703f9d6ad7SLin Ling void
13713f9d6ad7SLin Ling dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
13723f9d6ad7SLin Ling {
13733f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
13743f9d6ad7SLin Ling 	spa_t *spa = dp->dp_spa;
13757fd05ac4SMatthew Ahrens 	int err = 0;
13763f9d6ad7SLin Ling 
13773f9d6ad7SLin Ling 	/*
13783f9d6ad7SLin Ling 	 * Check for scn_restart_txg before checking spa_load_state, so
13793f9d6ad7SLin Ling 	 * that we can restart an old-style scan while the pool is being
13803f9d6ad7SLin Ling 	 * imported (see dsl_scan_init).
13813f9d6ad7SLin Ling 	 */
13823f9d6ad7SLin Ling 	if (scn->scn_restart_txg != 0 &&
13833f9d6ad7SLin Ling 	    scn->scn_restart_txg <= tx->tx_txg) {
13843f9d6ad7SLin Ling 		pool_scan_func_t func = POOL_SCAN_SCRUB;
13853f9d6ad7SLin Ling 		dsl_scan_done(scn, B_FALSE, tx);
13863f9d6ad7SLin Ling 		if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
13873f9d6ad7SLin Ling 			func = POOL_SCAN_RESILVER;
13883f9d6ad7SLin Ling 		zfs_dbgmsg("restarting scan func=%u txg=%llu",
13893f9d6ad7SLin Ling 		    func, tx->tx_txg);
13903b2aab18SMatthew Ahrens 		dsl_scan_setup_sync(&func, tx);
13913f9d6ad7SLin Ling 	}
13923f9d6ad7SLin Ling 
13937fd05ac4SMatthew Ahrens 	/*
13947fd05ac4SMatthew Ahrens 	 * If the scan is inactive due to a stalled async destroy, try again.
13957fd05ac4SMatthew Ahrens 	 */
13967fd05ac4SMatthew Ahrens 	if ((!scn->scn_async_stalled && !dsl_scan_active(scn)) ||
1397cde58dbcSMatthew Ahrens 	    spa_sync_pass(dp->dp_spa) > 1)
13983f9d6ad7SLin Ling 		return;
13993f9d6ad7SLin Ling 
14003f9d6ad7SLin Ling 	scn->scn_visited_this_txg = 0;
14013f9d6ad7SLin Ling 	scn->scn_pausing = B_FALSE;
14023f9d6ad7SLin Ling 	scn->scn_sync_start_time = gethrtime();
14033f9d6ad7SLin Ling 	spa->spa_scrub_active = B_TRUE;
14043f9d6ad7SLin Ling 
1405cde58dbcSMatthew Ahrens 	/*
14067fd05ac4SMatthew Ahrens 	 * First process the async destroys.  If we pause, don't do
14077fd05ac4SMatthew Ahrens 	 * any scrubbing or resilvering.  This ensures that there are no
14087fd05ac4SMatthew Ahrens 	 * async destroys while we are scanning, so the scan code doesn't
14097fd05ac4SMatthew Ahrens 	 * have to worry about traversing it.  It is also faster to free the
14107fd05ac4SMatthew Ahrens 	 * blocks than to scrub them.
1411cde58dbcSMatthew Ahrens 	 */
1412cde58dbcSMatthew Ahrens 	if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
1413ad135b5dSChristopher Siden 		scn->scn_is_bptree = B_FALSE;
1414cde58dbcSMatthew Ahrens 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
1415cde58dbcSMatthew Ahrens 		    NULL, ZIO_FLAG_MUSTSUCCEED);
1416cde58dbcSMatthew Ahrens 		err = bpobj_iterate(&dp->dp_free_bpobj,
1417ad135b5dSChristopher Siden 		    dsl_scan_free_block_cb, scn, tx);
1418b420f3adSRichard Lowe 		VERIFY3U(0, ==, zio_wait(scn->scn_zio_root));
1419ad135b5dSChristopher Siden 
14207fd05ac4SMatthew Ahrens 		if (err != 0 && err != ERESTART)
14217fd05ac4SMatthew Ahrens 			zfs_panic_recover("error %u from bpobj_iterate()", err);
14227fd05ac4SMatthew Ahrens 	}
14237fd05ac4SMatthew Ahrens 
14247fd05ac4SMatthew Ahrens 	if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
14257fd05ac4SMatthew Ahrens 		ASSERT(scn->scn_async_destroying);
14267fd05ac4SMatthew Ahrens 		scn->scn_is_bptree = B_TRUE;
14277fd05ac4SMatthew Ahrens 		scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
14287fd05ac4SMatthew Ahrens 		    NULL, ZIO_FLAG_MUSTSUCCEED);
14297fd05ac4SMatthew Ahrens 		err = bptree_iterate(dp->dp_meta_objset,
14307fd05ac4SMatthew Ahrens 		    dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
14317fd05ac4SMatthew Ahrens 		VERIFY0(zio_wait(scn->scn_zio_root));
14327fd05ac4SMatthew Ahrens 
14337fd05ac4SMatthew Ahrens 		if (err == EIO || err == ECKSUM) {
14347fd05ac4SMatthew Ahrens 			err = 0;
14357fd05ac4SMatthew Ahrens 		} else if (err != 0 && err != ERESTART) {
14367fd05ac4SMatthew Ahrens 			zfs_panic_recover("error %u from "
14377fd05ac4SMatthew Ahrens 			    "traverse_dataset_destroyed()", err);
1438ad135b5dSChristopher Siden 		}
14397fd05ac4SMatthew Ahrens 
14407fd05ac4SMatthew Ahrens 		/*
14417fd05ac4SMatthew Ahrens 		 * If we didn't make progress, mark the async destroy as
14427fd05ac4SMatthew Ahrens 		 * stalled, so that we will not initiate a spa_sync() on
14437fd05ac4SMatthew Ahrens 		 * its behalf.
14447fd05ac4SMatthew Ahrens 		 */
14457fd05ac4SMatthew Ahrens 		scn->scn_async_stalled = (scn->scn_visited_this_txg == 0);
14467fd05ac4SMatthew Ahrens 
14477fd05ac4SMatthew Ahrens 		if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
14487fd05ac4SMatthew Ahrens 			/* finished; deactivate async destroy feature */
14497fd05ac4SMatthew Ahrens 			spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
14507fd05ac4SMatthew Ahrens 			ASSERT(!spa_feature_is_active(spa,
14517fd05ac4SMatthew Ahrens 			    SPA_FEATURE_ASYNC_DESTROY));
14527fd05ac4SMatthew Ahrens 			VERIFY0(zap_remove(dp->dp_meta_objset,
14537fd05ac4SMatthew Ahrens 			    DMU_POOL_DIRECTORY_OBJECT,
14547fd05ac4SMatthew Ahrens 			    DMU_POOL_BPTREE_OBJ, tx));
14557fd05ac4SMatthew Ahrens 			VERIFY0(bptree_free(dp->dp_meta_objset,
14567fd05ac4SMatthew Ahrens 			    dp->dp_bptree_obj, tx));
14577fd05ac4SMatthew Ahrens 			dp->dp_bptree_obj = 0;
14587fd05ac4SMatthew Ahrens 			scn->scn_async_destroying = B_FALSE;
1459cde58dbcSMatthew Ahrens 		}
14607fd05ac4SMatthew Ahrens 	}
14617fd05ac4SMatthew Ahrens 	if (scn->scn_visited_this_txg) {
14627fd05ac4SMatthew Ahrens 		zfs_dbgmsg("freed %llu blocks in %llums from "
14637fd05ac4SMatthew Ahrens 		    "free_bpobj/bptree txg %llu; err=%u",
14647fd05ac4SMatthew Ahrens 		    (longlong_t)scn->scn_visited_this_txg,
14657fd05ac4SMatthew Ahrens 		    (longlong_t)
14667fd05ac4SMatthew Ahrens 		    NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
14677fd05ac4SMatthew Ahrens 		    (longlong_t)tx->tx_txg, err);
14687fd05ac4SMatthew Ahrens 		scn->scn_visited_this_txg = 0;
14697fd05ac4SMatthew Ahrens 
14707fd05ac4SMatthew Ahrens 		/*
14717fd05ac4SMatthew Ahrens 		 * Write out changes to the DDT that may be required as a
14727fd05ac4SMatthew Ahrens 		 * result of the blocks freed.  This ensures that the DDT
14737fd05ac4SMatthew Ahrens 		 * is clean when a scrub/resilver runs.
14747fd05ac4SMatthew Ahrens 		 */
14757fd05ac4SMatthew Ahrens 		ddt_sync(spa, tx->tx_txg);
14767fd05ac4SMatthew Ahrens 	}
14777fd05ac4SMatthew Ahrens 	if (err != 0)
14787fd05ac4SMatthew Ahrens 		return;
14797fd05ac4SMatthew Ahrens 	if (!scn->scn_async_destroying && zfs_free_leak_on_eio &&
14807fd05ac4SMatthew Ahrens 	    (dp->dp_free_dir->dd_phys->dd_used_bytes != 0 ||
14817fd05ac4SMatthew Ahrens 	    dp->dp_free_dir->dd_phys->dd_compressed_bytes != 0 ||
14827fd05ac4SMatthew Ahrens 	    dp->dp_free_dir->dd_phys->dd_uncompressed_bytes != 0)) {
14837fd05ac4SMatthew Ahrens 		/*
14847fd05ac4SMatthew Ahrens 		 * We have finished background destroying, but there is still
14857fd05ac4SMatthew Ahrens 		 * some space left in the dp_free_dir. Transfer this leaked
14867fd05ac4SMatthew Ahrens 		 * space to the dp_leak_dir.
14877fd05ac4SMatthew Ahrens 		 */
14887fd05ac4SMatthew Ahrens 		if (dp->dp_leak_dir == NULL) {
14897fd05ac4SMatthew Ahrens 			rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
14907fd05ac4SMatthew Ahrens 			(void) dsl_dir_create_sync(dp, dp->dp_root_dir,
14917fd05ac4SMatthew Ahrens 			    LEAK_DIR_NAME, tx);
14927fd05ac4SMatthew Ahrens 			VERIFY0(dsl_pool_open_special_dir(dp,
14937fd05ac4SMatthew Ahrens 			    LEAK_DIR_NAME, &dp->dp_leak_dir));
14947fd05ac4SMatthew Ahrens 			rrw_exit(&dp->dp_config_rwlock, FTAG);
14957fd05ac4SMatthew Ahrens 		}
14967fd05ac4SMatthew Ahrens 		dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
14977fd05ac4SMatthew Ahrens 		    dp->dp_free_dir->dd_phys->dd_used_bytes,
14987fd05ac4SMatthew Ahrens 		    dp->dp_free_dir->dd_phys->dd_compressed_bytes,
14997fd05ac4SMatthew Ahrens 		    dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
15007fd05ac4SMatthew Ahrens 		dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
15017fd05ac4SMatthew Ahrens 		    -dp->dp_free_dir->dd_phys->dd_used_bytes,
15027fd05ac4SMatthew Ahrens 		    -dp->dp_free_dir->dd_phys->dd_compressed_bytes,
15037fd05ac4SMatthew Ahrens 		    -dp->dp_free_dir->dd_phys->dd_uncompressed_bytes, tx);
15047fd05ac4SMatthew Ahrens 	}
15057fd05ac4SMatthew Ahrens 	if (!scn->scn_async_destroying) {
15065d7b4d43SMatthew Ahrens 		/* finished; verify that space accounting went to zero */
15075d7b4d43SMatthew Ahrens 		ASSERT0(dp->dp_free_dir->dd_phys->dd_used_bytes);
15085d7b4d43SMatthew Ahrens 		ASSERT0(dp->dp_free_dir->dd_phys->dd_compressed_bytes);
15095d7b4d43SMatthew Ahrens 		ASSERT0(dp->dp_free_dir->dd_phys->dd_uncompressed_bytes);
1510cde58dbcSMatthew Ahrens 	}
1511cde58dbcSMatthew Ahrens 
1512cde58dbcSMatthew Ahrens 	if (scn->scn_phys.scn_state != DSS_SCANNING)
1513cde58dbcSMatthew Ahrens 		return;
1514cde58dbcSMatthew Ahrens 
1515b4952e17SGeorge Wilson 	if (scn->scn_done_txg == tx->tx_txg) {
1516b4952e17SGeorge Wilson 		ASSERT(!scn->scn_pausing);
1517b4952e17SGeorge Wilson 		/* finished with scan. */
1518b4952e17SGeorge Wilson 		zfs_dbgmsg("txg %llu scan complete", tx->tx_txg);
1519b4952e17SGeorge Wilson 		dsl_scan_done(scn, B_TRUE, tx);
1520b4952e17SGeorge Wilson 		ASSERT3U(spa->spa_scrub_inflight, ==, 0);
1521b4952e17SGeorge Wilson 		dsl_scan_sync_state(scn, tx);
1522b4952e17SGeorge Wilson 		return;
1523b4952e17SGeorge Wilson 	}
1524b4952e17SGeorge Wilson 
15253f9d6ad7SLin Ling 	if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
15263f9d6ad7SLin Ling 	    scn->scn_phys.scn_ddt_class_max) {
15273f9d6ad7SLin Ling 		zfs_dbgmsg("doing scan sync txg %llu; "
15283f9d6ad7SLin Ling 		    "ddt bm=%llu/%llu/%llu/%llx",
15293f9d6ad7SLin Ling 		    (longlong_t)tx->tx_txg,
15303f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_class,
15313f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_type,
15323f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_checksum,
15333f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_ddt_bookmark.ddb_cursor);
15343f9d6ad7SLin Ling 		ASSERT(scn->scn_phys.scn_bookmark.zb_objset == 0);
15353f9d6ad7SLin Ling 		ASSERT(scn->scn_phys.scn_bookmark.zb_object == 0);
15363f9d6ad7SLin Ling 		ASSERT(scn->scn_phys.scn_bookmark.zb_level == 0);
15373f9d6ad7SLin Ling 		ASSERT(scn->scn_phys.scn_bookmark.zb_blkid == 0);
15383f9d6ad7SLin Ling 	} else {
15393f9d6ad7SLin Ling 		zfs_dbgmsg("doing scan sync txg %llu; bm=%llu/%llu/%llu/%llu",
15403f9d6ad7SLin Ling 		    (longlong_t)tx->tx_txg,
15413f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_objset,
15423f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_object,
15433f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_level,
15443f9d6ad7SLin Ling 		    (longlong_t)scn->scn_phys.scn_bookmark.zb_blkid);
15453f9d6ad7SLin Ling 	}
15463f9d6ad7SLin Ling 
1547cde58dbcSMatthew Ahrens 	scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
15483f9d6ad7SLin Ling 	    NULL, ZIO_FLAG_CANFAIL);
15493b2aab18SMatthew Ahrens 	dsl_pool_config_enter(dp, FTAG);
15503f9d6ad7SLin Ling 	dsl_scan_visit(scn, tx);
15513b2aab18SMatthew Ahrens 	dsl_pool_config_exit(dp, FTAG);
1552cde58dbcSMatthew Ahrens 	(void) zio_wait(scn->scn_zio_root);
1553cde58dbcSMatthew Ahrens 	scn->scn_zio_root = NULL;
15543f9d6ad7SLin Ling 
15553f9d6ad7SLin Ling 	zfs_dbgmsg("visited %llu blocks in %llums",
15563f9d6ad7SLin Ling 	    (longlong_t)scn->scn_visited_this_txg,
15570689f76cSAdam Leventhal 	    (longlong_t)NSEC2MSEC(gethrtime() - scn->scn_sync_start_time));
15583f9d6ad7SLin Ling 
15593f9d6ad7SLin Ling 	if (!scn->scn_pausing) {
1560b4952e17SGeorge Wilson 		scn->scn_done_txg = tx->tx_txg + 1;
1561b4952e17SGeorge Wilson 		zfs_dbgmsg("txg %llu traversal complete, waiting till txg %llu",
1562b4952e17SGeorge Wilson 		    tx->tx_txg, scn->scn_done_txg);
15633f9d6ad7SLin Ling 	}
15643f9d6ad7SLin Ling 
15653f9d6ad7SLin Ling 	if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
15663f9d6ad7SLin Ling 		mutex_enter(&spa->spa_scrub_lock);
15673f9d6ad7SLin Ling 		while (spa->spa_scrub_inflight > 0) {
15683f9d6ad7SLin Ling 			cv_wait(&spa->spa_scrub_io_cv,
15693f9d6ad7SLin Ling 			    &spa->spa_scrub_lock);
15703f9d6ad7SLin Ling 		}
15713f9d6ad7SLin Ling 		mutex_exit(&spa->spa_scrub_lock);
15723f9d6ad7SLin Ling 	}
15733f9d6ad7SLin Ling 
15743f9d6ad7SLin Ling 	dsl_scan_sync_state(scn, tx);
15753f9d6ad7SLin Ling }
15763f9d6ad7SLin Ling 
15773f9d6ad7SLin Ling /*
15783f9d6ad7SLin Ling  * This will start a new scan, or restart an existing one.
15793f9d6ad7SLin Ling  */
15803f9d6ad7SLin Ling void
15813f9d6ad7SLin Ling dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
15823f9d6ad7SLin Ling {
15833f9d6ad7SLin Ling 	if (txg == 0) {
15843f9d6ad7SLin Ling 		dmu_tx_t *tx;
15853f9d6ad7SLin Ling 		tx = dmu_tx_create_dd(dp->dp_mos_dir);
15863f9d6ad7SLin Ling 		VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
15873f9d6ad7SLin Ling 
15883f9d6ad7SLin Ling 		txg = dmu_tx_get_txg(tx);
15893f9d6ad7SLin Ling 		dp->dp_scan->scn_restart_txg = txg;
15903f9d6ad7SLin Ling 		dmu_tx_commit(tx);
15913f9d6ad7SLin Ling 	} else {
15923f9d6ad7SLin Ling 		dp->dp_scan->scn_restart_txg = txg;
15933f9d6ad7SLin Ling 	}
15943f9d6ad7SLin Ling 	zfs_dbgmsg("restarting resilver txg=%llu", txg);
15953f9d6ad7SLin Ling }
15963f9d6ad7SLin Ling 
15973f9d6ad7SLin Ling boolean_t
15983f9d6ad7SLin Ling dsl_scan_resilvering(dsl_pool_t *dp)
15993f9d6ad7SLin Ling {
16003f9d6ad7SLin Ling 	return (dp->dp_scan->scn_phys.scn_state == DSS_SCANNING &&
16013f9d6ad7SLin Ling 	    dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
16023f9d6ad7SLin Ling }
16033f9d6ad7SLin Ling 
16043f9d6ad7SLin Ling /*
16053f9d6ad7SLin Ling  * scrub consumers
16063f9d6ad7SLin Ling  */
16073f9d6ad7SLin Ling 
16083f9d6ad7SLin Ling static void
16093f9d6ad7SLin Ling count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
16103f9d6ad7SLin Ling {
16113f9d6ad7SLin Ling 	int i;
16123f9d6ad7SLin Ling 
16133f9d6ad7SLin Ling 	/*
16143f9d6ad7SLin Ling 	 * If we resume after a reboot, zab will be NULL; don't record
16153f9d6ad7SLin Ling 	 * incomplete stats in that case.
16163f9d6ad7SLin Ling 	 */
16173f9d6ad7SLin Ling 	if (zab == NULL)
16183f9d6ad7SLin Ling 		return;
16193f9d6ad7SLin Ling 
16203f9d6ad7SLin Ling 	for (i = 0; i < 4; i++) {
16213f9d6ad7SLin Ling 		int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
16223f9d6ad7SLin Ling 		int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
1623ad135b5dSChristopher Siden 		if (t & DMU_OT_NEWTYPE)
1624ad135b5dSChristopher Siden 			t = DMU_OT_OTHER;
16253f9d6ad7SLin Ling 		zfs_blkstat_t *zb = &zab->zab_type[l][t];
16263f9d6ad7SLin Ling 		int equal;
16273f9d6ad7SLin Ling 
16283f9d6ad7SLin Ling 		zb->zb_count++;
16293f9d6ad7SLin Ling 		zb->zb_asize += BP_GET_ASIZE(bp);
16303f9d6ad7SLin Ling 		zb->zb_lsize += BP_GET_LSIZE(bp);
16313f9d6ad7SLin Ling 		zb->zb_psize += BP_GET_PSIZE(bp);
16323f9d6ad7SLin Ling 		zb->zb_gangs += BP_COUNT_GANG(bp);
16333f9d6ad7SLin Ling 
16343f9d6ad7SLin Ling 		switch (BP_GET_NDVAS(bp)) {
16353f9d6ad7SLin Ling 		case 2:
16363f9d6ad7SLin Ling 			if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
16373f9d6ad7SLin Ling 			    DVA_GET_VDEV(&bp->blk_dva[1]))
16383f9d6ad7SLin Ling 				zb->zb_ditto_2_of_2_samevdev++;
16393f9d6ad7SLin Ling 			break;
16403f9d6ad7SLin Ling 		case 3:
16413f9d6ad7SLin Ling 			equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
16423f9d6ad7SLin Ling 			    DVA_GET_VDEV(&bp->blk_dva[1])) +
16433f9d6ad7SLin Ling 			    (DVA_GET_VDEV(&bp->blk_dva[0]) ==
16443f9d6ad7SLin Ling 			    DVA_GET_VDEV(&bp->blk_dva[2])) +
16453f9d6ad7SLin Ling 			    (DVA_GET_VDEV(&bp->blk_dva[1]) ==
16463f9d6ad7SLin Ling 			    DVA_GET_VDEV(&bp->blk_dva[2]));
16473f9d6ad7SLin Ling 			if (equal == 1)
16483f9d6ad7SLin Ling 				zb->zb_ditto_2_of_3_samevdev++;
16493f9d6ad7SLin Ling 			else if (equal == 3)
16503f9d6ad7SLin Ling 				zb->zb_ditto_3_of_3_samevdev++;
16513f9d6ad7SLin Ling 			break;
16523f9d6ad7SLin Ling 		}
16533f9d6ad7SLin Ling 	}
16543f9d6ad7SLin Ling }
16553f9d6ad7SLin Ling 
16563f9d6ad7SLin Ling static void
16573f9d6ad7SLin Ling dsl_scan_scrub_done(zio_t *zio)
16583f9d6ad7SLin Ling {
16593f9d6ad7SLin Ling 	spa_t *spa = zio->io_spa;
16603f9d6ad7SLin Ling 
16613f9d6ad7SLin Ling 	zio_data_buf_free(zio->io_data, zio->io_size);
16623f9d6ad7SLin Ling 
16633f9d6ad7SLin Ling 	mutex_enter(&spa->spa_scrub_lock);
16643f9d6ad7SLin Ling 	spa->spa_scrub_inflight--;
16653f9d6ad7SLin Ling 	cv_broadcast(&spa->spa_scrub_io_cv);
16663f9d6ad7SLin Ling 
16673f9d6ad7SLin Ling 	if (zio->io_error && (zio->io_error != ECKSUM ||
16683f9d6ad7SLin Ling 	    !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
16693f9d6ad7SLin Ling 		spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors++;
16703f9d6ad7SLin Ling 	}
16713f9d6ad7SLin Ling 	mutex_exit(&spa->spa_scrub_lock);
16723f9d6ad7SLin Ling }
16733f9d6ad7SLin Ling 
16743f9d6ad7SLin Ling static int
16753f9d6ad7SLin Ling dsl_scan_scrub_cb(dsl_pool_t *dp,
1676*7802d7bfSMatthew Ahrens     const blkptr_t *bp, const zbookmark_phys_t *zb)
16773f9d6ad7SLin Ling {
16783f9d6ad7SLin Ling 	dsl_scan_t *scn = dp->dp_scan;
16793f9d6ad7SLin Ling 	size_t size = BP_GET_PSIZE(bp);
16803f9d6ad7SLin Ling 	spa_t *spa = dp->dp_spa;
16813f9d6ad7SLin Ling 	uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
16823f9d6ad7SLin Ling 	boolean_t needs_io;
168344ecc532SGeorge Wilson 	int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
168444ecc532SGeorge Wilson 	int scan_delay = 0;
16853f9d6ad7SLin Ling 
16863f9d6ad7SLin Ling 	if (phys_birth <= scn->scn_phys.scn_min_txg ||
16873f9d6ad7SLin Ling 	    phys_birth >= scn->scn_phys.scn_max_txg)
16883f9d6ad7SLin Ling 		return (0);
16893f9d6ad7SLin Ling 
16903f9d6ad7SLin Ling 	count_block(dp->dp_blkstats, bp);
16913f9d6ad7SLin Ling 
16925d7b4d43SMatthew Ahrens 	if (BP_IS_EMBEDDED(bp))
16935d7b4d43SMatthew Ahrens 		return (0);
16945d7b4d43SMatthew Ahrens 
16953f9d6ad7SLin Ling 	ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
16963f9d6ad7SLin Ling 	if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
16973f9d6ad7SLin Ling 		zio_flags |= ZIO_FLAG_SCRUB;
16983f9d6ad7SLin Ling 		needs_io = B_TRUE;
169944ecc532SGeorge Wilson 		scan_delay = zfs_scrub_delay;
1700d5285caeSGeorge Wilson 	} else {
1701d5285caeSGeorge Wilson 		ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
17023f9d6ad7SLin Ling 		zio_flags |= ZIO_FLAG_RESILVER;
17033f9d6ad7SLin Ling 		needs_io = B_FALSE;
170444ecc532SGeorge Wilson 		scan_delay = zfs_resilver_delay;
17053f9d6ad7SLin Ling 	}
17063f9d6ad7SLin Ling 
17073f9d6ad7SLin Ling 	/* If it's an intent log block, failure is expected. */
17083f9d6ad7SLin Ling 	if (zb->zb_level == ZB_ZIL_LEVEL)
17093f9d6ad7SLin Ling 		zio_flags |= ZIO_FLAG_SPECULATIVE;
17103f9d6ad7SLin Ling 
17113f9d6ad7SLin Ling 	for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
17123f9d6ad7SLin Ling 		vdev_t *vd = vdev_lookup_top(spa,
17133f9d6ad7SLin Ling 		    DVA_GET_VDEV(&bp->blk_dva[d]));
17143f9d6ad7SLin Ling 
17153f9d6ad7SLin Ling 		/*
17163f9d6ad7SLin Ling 		 * Keep track of how much data we've examined so that
17173f9d6ad7SLin Ling 		 * zpool(1M) status can make useful progress reports.
17183f9d6ad7SLin Ling 		 */
17193f9d6ad7SLin Ling 		scn->scn_phys.scn_examined += DVA_GET_ASIZE(&bp->blk_dva[d]);
17203f9d6ad7SLin Ling 		spa->spa_scan_pass_exam += DVA_GET_ASIZE(&bp->blk_dva[d]);
17213f9d6ad7SLin Ling 
17223f9d6ad7SLin Ling 		/* if it's a resilver, this may not be in the target range */
17233f9d6ad7SLin Ling 		if (!needs_io) {
17243f9d6ad7SLin Ling 			if (DVA_GET_GANG(&bp->blk_dva[d])) {
17253f9d6ad7SLin Ling 				/*
17263f9d6ad7SLin Ling 				 * Gang members may be spread across multiple
17273f9d6ad7SLin Ling 				 * vdevs, so the best estimate we have is the
17283f9d6ad7SLin Ling 				 * scrub range, which has already been checked.
17293f9d6ad7SLin Ling 				 * XXX -- it would be better to change our
17303f9d6ad7SLin Ling 				 * allocation policy to ensure that all
17313f9d6ad7SLin Ling 				 * gang members reside on the same vdev.
17323f9d6ad7SLin Ling 				 */
17333f9d6ad7SLin Ling 				needs_io = B_TRUE;
17343f9d6ad7SLin Ling 			} else {
17353f9d6ad7SLin Ling 				needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
17363f9d6ad7SLin Ling 				    phys_birth, 1);
17373f9d6ad7SLin Ling 			}
17383f9d6ad7SLin Ling 		}
17393f9d6ad7SLin Ling 	}
17403f9d6ad7SLin Ling 
17413f9d6ad7SLin Ling 	if (needs_io && !zfs_no_scrub_io) {
174244ecc532SGeorge Wilson 		vdev_t *rvd = spa->spa_root_vdev;
174344ecc532SGeorge Wilson 		uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
17443f9d6ad7SLin Ling 		void *data = zio_data_buf_alloc(size);
17453f9d6ad7SLin Ling 
17463f9d6ad7SLin Ling 		mutex_enter(&spa->spa_scrub_lock);
174744ecc532SGeorge Wilson 		while (spa->spa_scrub_inflight >= maxinflight)
17483f9d6ad7SLin Ling 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
17493f9d6ad7SLin Ling 		spa->spa_scrub_inflight++;
17503f9d6ad7SLin Ling 		mutex_exit(&spa->spa_scrub_lock);
17513f9d6ad7SLin Ling 
175244ecc532SGeorge Wilson 		/*
175344ecc532SGeorge Wilson 		 * If we're seeing recent (zfs_scan_idle) "important" I/Os
175444ecc532SGeorge Wilson 		 * then throttle our workload to limit the impact of a scan.
175544ecc532SGeorge Wilson 		 */
175644ecc532SGeorge Wilson 		if (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle)
175744ecc532SGeorge Wilson 			delay(scan_delay);
175844ecc532SGeorge Wilson 
17593f9d6ad7SLin Ling 		zio_nowait(zio_read(NULL, spa, bp, data, size,
176069962b56SMatthew Ahrens 		    dsl_scan_scrub_done, NULL, ZIO_PRIORITY_SCRUB,
17613f9d6ad7SLin Ling 		    zio_flags, zb));
17623f9d6ad7SLin Ling 	}
17633f9d6ad7SLin Ling 
17643f9d6ad7SLin Ling 	/* do not relocate this block */
17653f9d6ad7SLin Ling 	return (0);
17663f9d6ad7SLin Ling }
17673f9d6ad7SLin Ling 
17683f9d6ad7SLin Ling int
17693f9d6ad7SLin Ling dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
17703f9d6ad7SLin Ling {
17713f9d6ad7SLin Ling 	spa_t *spa = dp->dp_spa;
17723f9d6ad7SLin Ling 
17733f9d6ad7SLin Ling 	/*
17743f9d6ad7SLin Ling 	 * Purge all vdev caches and probe all devices.  We do this here
17753f9d6ad7SLin Ling 	 * rather than in sync context because this requires a writer lock
17763f9d6ad7SLin Ling 	 * on the spa_config lock, which we can't do from sync context.  The
17773f9d6ad7SLin Ling 	 * spa_scrub_reopen flag indicates that vdev_open() should not
17783f9d6ad7SLin Ling 	 * attempt to start another scrub.
17793f9d6ad7SLin Ling 	 */
17803f9d6ad7SLin Ling 	spa_vdev_state_enter(spa, SCL_NONE);
17813f9d6ad7SLin Ling 	spa->spa_scrub_reopen = B_TRUE;
17823f9d6ad7SLin Ling 	vdev_reopen(spa->spa_root_vdev);
17833f9d6ad7SLin Ling 	spa->spa_scrub_reopen = B_FALSE;
17843f9d6ad7SLin Ling 	(void) spa_vdev_state_exit(spa, NULL, 0);
17853f9d6ad7SLin Ling 
17863b2aab18SMatthew Ahrens 	return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
17873b2aab18SMatthew Ahrens 	    dsl_scan_setup_sync, &func, 0));
17883f9d6ad7SLin Ling }
1789