OpenZFS 7614, 9064 - zfs device evacuation/removal

OpenZFS 7614 - zfs device evacuation/removal
OpenZFS 9064 - remove_mirror should wait for device removal to complete

This project allows top-level vdevs to be removed from the storage pool
with "zpool remove", reducing the total amount of storage in the pool.
This operation copies all allocated regions of the device to be removed
onto other devices, recording the mapping from old to new location.
After the removal is complete, read and free operations to the removed
(now "indirect") vdev must be remapped and performed at the new location
on disk.  The indirect mapping table is kept in memory whenever the pool
is loaded, so there is minimal performance overhead when doing operations
on the indirect vdev.

The size of the in-memory mapping table will be reduced when its entries
become "obsolete" because they are no longer used by any block pointers
in the pool.  An entry becomes obsolete when all the blocks that use
it are freed.  An entry can also become obsolete when all the snapshots
that reference it are deleted, and the block pointers that reference it
have been "remapped" in all filesystems/zvols (and clones).  Whenever an
indirect block is written, all the block pointers in it will be "remapped"
to their new (concrete) locations if possible.  This process can be
accelerated by using the "zfs remap" command to proactively rewrite all
indirect blocks that reference indirect (removed) vdevs.

Note that when a device is removed, we do not verify the checksum of
the data that is copied.  This makes the process much faster, but if it
were used on redundant vdevs (i.e. mirror or raidz vdevs), it would be
possible to copy the wrong data, when we have the correct data on e.g.
the other side of the mirror.

At the moment, only mirrors and simple top-level vdevs can be removed
and no removal is allowed if any of the top-level vdevs are raidz.

Porting Notes:

* Avoid zero-sized kmem_alloc() in vdev_compact_children().

    The device evacuation code adds a dependency that
    vdev_compact_children() be able to properly empty the vdev_child
    array by setting it to NULL and zeroing vdev_children.  Under Linux,
    kmem_alloc() and related functions return a sentinel pointer rather
    than NULL for zero-sized allocations.

* Remove comment regarding "mpt" driver where zfs_remove_max_segment
  is initialized to SPA_MAXBLOCKSIZE.

  Change zfs_condense_indirect_commit_entry_delay_ticks to
  zfs_condense_indirect_commit_entry_delay_ms for consistency with
  most other tunables in which delays are specified in ms.

* ZTS changes:

    Use set_tunable rather than mdb
    Use zpool sync as appropriate
    Use sync_pool instead of sync
    Kill jobs during test_removal_with_operation to allow unmount/export
    Don't add non-disk names such as "mirror" or "raidz" to $DISKS
    Use $TEST_BASE_DIR instead of /tmp
    Increase HZ from 100 to 1000 which is more common on Linux

    removal_multiple_indirection.ksh
        Reduce iterations in order to not time out on the code
        coverage builders.

    removal_resume_export:
        Functionally, the test case is correct but there exists a race
        where the kernel thread hasn't been fully started yet and is
        not visible.  Wait for up to 1 second for the removal thread
        to be started before giving up on it.  Also, increase the
        amount of data copied in order that the removal not finish
        before the export has a chance to fail.

* MMP compatibility, the concept of concrete versus non-concrete devices
  has slightly changed the semantics of vdev_writeable().  Update
  mmp_random_leaf_impl() accordingly.

* Updated dbuf_remap() to handle the org.zfsonlinux:large_dnode pool
  feature which is not supported by OpenZFS.

* Added support for new vdev removal tracepoints.

* Test cases removal_with_zdb and removal_condense_export have been
  intentionally disabled.  When run manually they pass as intended,
  but when running in the automated test environment they produce
  unreliable results on the latest Fedora release.

  They may work better once the upstream pool import refectoring is
  merged into ZoL at which point they will be re-enabled.

Authored by: Matthew Ahrens <[email protected]>
Reviewed-by: Alex Reece <[email protected]>
Reviewed-by: George Wilson <[email protected]>
Reviewed-by: John Kennedy <[email protected]>
Reviewed-by: Prakash Surya <[email protected]>
Reviewed by: Richard Laager <[email protected]>
Reviewed by: Tim Chase <[email protected]>
Reviewed by: Brian Behlendorf <[email protected]>
Approved by: Garrett D'Amore <[email protected]>
Ported-by: Tim Chase <[email protected]>
Signed-off-by: Tim Chase <[email protected]>

OpenZFS-issue: https://www.illumos.org/issues/7614
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/f539f1eb
Closes #6900

1 files changed, 257 insertions, 338 deletions

diff --git a/module/zfs/spa.c b/module/zfs/spa.c
index 53b5aabf0..08fc7bbda 100644
--- a/module/zfs/spa.c
+++ b/module/zfs/spa.c
@@ -52,6 +52,9 @@
 #include <sys/zil.h>
 #include <sys/ddt.h>
 #include <sys/vdev_impl.h>
+#include <sys/vdev_removal.h>
+#include <sys/vdev_indirect_mapping.h>
+#include <sys/vdev_indirect_births.h>
 #include <sys/vdev_disk.h>
 #include <sys/metaslab.h>
 #include <sys/metaslab_impl.h>
@@ -59,6 +62,7 @@
 #include <sys/uberblock_impl.h>
 #include <sys/txg.h>
 #include <sys/avl.h>
+#include <sys/bpobj.h>
 #include <sys/dmu_traverse.h>
 #include <sys/dmu_objset.h>
 #include <sys/unique.h>
@@ -96,7 +100,7 @@
  * The interval, in seconds, at which failed configuration cache file writes
  * should be retried.
  */
-static int zfs_ccw_retry_interval = 300;
+int zfs_ccw_retry_interval = 300;
 
 typedef enum zti_modes {
 	ZTI_MODE_FIXED,			/* value is # of threads (min 1) */
@@ -150,14 +154,11 @@ const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = {
 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* IOCTL */
 };
 
-static sysevent_t *spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl,
-    const char *name);
-static void spa_event_post(sysevent_t *ev);
 static void spa_sync_version(void *arg, dmu_tx_t *tx);
 static void spa_sync_props(void *arg, dmu_tx_t *tx);
 static boolean_t spa_has_active_shared_spare(spa_t *spa);
 static inline int spa_load_impl(spa_t *spa, uint64_t, nvlist_t *config,
-    spa_load_state_t state, spa_import_type_t type, boolean_t mosconfig,
+    spa_load_state_t state, spa_import_type_t type, boolean_t trust_config,
     char **ereport);
 static void spa_vdev_resilver_done(spa_t *spa);
 
@@ -817,7 +818,7 @@ spa_change_guid(spa_t *spa)
 	    spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED);
 
 	if (error == 0) {
-		spa_config_sync(spa, B_FALSE, B_TRUE);
+		spa_write_cachefile(spa, B_FALSE, B_TRUE);
 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID);
 	}
 
@@ -1149,6 +1150,9 @@ spa_activate(spa_t *spa, int mode)
 		spa_create_zio_taskqs(spa);
 	}
 
+	for (size_t i = 0; i < TXG_SIZE; i++)
+		spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL, 0);
+
 	list_create(&spa->spa_config_dirty_list, sizeof (vdev_t),
 	    offsetof(vdev_t, vdev_config_dirty_node));
 	list_create(&spa->spa_evicting_os_list, sizeof (objset_t),
@@ -1245,6 +1249,12 @@ spa_deactivate(spa_t *spa)
 		}
 	}
 
+	for (size_t i = 0; i < TXG_SIZE; i++) {
+		ASSERT3P(spa->spa_txg_zio[i], !=, NULL);
+		VERIFY0(zio_wait(spa->spa_txg_zio[i]));
+		spa->spa_txg_zio[i] = NULL;
+	}
+
 	metaslab_class_destroy(spa->spa_normal_class);
 	spa->spa_normal_class = NULL;
 
@@ -1385,6 +1395,13 @@ spa_unload(spa_t *spa)
 		spa->spa_async_zio_root = NULL;
 	}
 
+	if (spa->spa_vdev_removal != NULL) {
+		spa_vdev_removal_destroy(spa->spa_vdev_removal);
+		spa->spa_vdev_removal = NULL;
+	}
+
+	spa_condense_fini(spa);
+
 	bpobj_close(&spa->spa_deferred_bpobj);
 
 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
@@ -1442,6 +1459,8 @@ spa_unload(spa_t *spa)
 
 	spa->spa_async_suspended = 0;
 
+	spa->spa_indirect_vdevs_loaded = B_FALSE;
+
 	if (spa->spa_comment != NULL) {
 		spa_strfree(spa->spa_comment);
 		spa->spa_comment = NULL;
@@ -1456,7 +1475,7 @@ spa_unload(spa_t *spa)
  * 'spa_spares.sav_config'.  We parse this into vdevs, try to open them, and
  * then re-generate a more complete list including status information.
  */
-static void
+void
 spa_load_spares(spa_t *spa)
 {
 	nvlist_t **spares;
@@ -1573,7 +1592,7 @@ spa_load_spares(spa_t *spa)
  * Devices which are already active have their details maintained, and are
  * not re-opened.
  */
-static void
+void
 spa_load_l2cache(spa_t *spa)
 {
 	nvlist_t **l2cache = NULL;
@@ -1734,7 +1753,7 @@ spa_check_removed(vdev_t *vd)
 		spa_check_removed(vd->vdev_child[c]);
 
 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) &&
-	    !vd->vdev_ishole) {
+	    vdev_is_concrete(vd)) {
 		zfs_post_autoreplace(vd->vdev_spa, vd);
 		spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK);
 	}
@@ -1817,27 +1836,26 @@ spa_config_valid(spa_t *spa, nvlist_t *config)
 
 		/*
 		 * Resolve any "missing" vdevs in the current configuration.
+		 * Also trust the MOS config about any "indirect" vdevs.
 		 * If we find that the MOS config has more accurate information
 		 * about the top-level vdev then use that vdev instead.
 		 */
-		if (tvd->vdev_ops == &vdev_missing_ops &&
-		    mtvd->vdev_ops != &vdev_missing_ops) {
-
-			if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG))
-				continue;
+		if ((tvd->vdev_ops == &vdev_missing_ops &&
+		    mtvd->vdev_ops != &vdev_missing_ops) ||
+		    (mtvd->vdev_ops == &vdev_indirect_ops &&
+		    tvd->vdev_ops != &vdev_indirect_ops)) {
 
 			/*
 			 * Device specific actions.
 			 */
 			if (mtvd->vdev_islog) {
+				if (!(spa->spa_import_flags &
+				    ZFS_IMPORT_MISSING_LOG)) {
+					continue;
+				}
+
 				spa_set_log_state(spa, SPA_LOG_CLEAR);
-			} else {
-				/*
-				 * XXX - once we have 'readonly' pool
-				 * support we should be able to handle
-				 * missing data devices by transitioning
-				 * the pool to readonly.
-				 */
+			} else if (mtvd->vdev_ops != &vdev_indirect_ops) {
 				continue;
 			}
 
@@ -1851,10 +1869,6 @@ spa_config_valid(spa_t *spa, nvlist_t *config)
 			vdev_add_child(rvd, mtvd);
 			vdev_add_child(mrvd, tvd);
 
-			spa_config_exit(spa, SCL_ALL, FTAG);
-			vdev_load(mtvd);
-			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
-
 			vdev_reopen(rvd);
 		} else {
 			if (mtvd->vdev_islog) {
@@ -1873,6 +1887,14 @@ spa_config_valid(spa_t *spa, nvlist_t *config)
 			 */
 			spa_config_valid_zaps(tvd, mtvd);
 		}
+
+		/*
+		 * Never trust this info from userland; always use what's
+		 * in the MOS.  This prevents it from getting out of sync
+		 * with the rest of the info in the MOS.
+		 */
+		tvd->vdev_removing = mtvd->vdev_removing;
+		tvd->vdev_indirect_config = mtvd->vdev_indirect_config;
 	}
 
 	vdev_free(mrvd);
@@ -1949,11 +1971,11 @@ spa_activate_log(spa_t *spa)
 }
 
 int
-spa_offline_log(spa_t *spa)
+spa_reset_logs(spa_t *spa)
 {
 	int error;
 
-	error = dmu_objset_find(spa_name(spa), zil_vdev_offline,
+	error = dmu_objset_find(spa_name(spa), zil_reset,
 	    NULL, DS_FIND_CHILDREN);
 	if (error == 0) {
 		/*
@@ -2155,7 +2177,7 @@ static int
 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err)
 {
 	vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux);
-	return (err);
+	return (SET_ERROR(err));
 }
 
 /*
@@ -2547,7 +2569,7 @@ out:
 __attribute__((always_inline))
 static inline int
 spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
-    spa_load_state_t state, spa_import_type_t type, boolean_t mosconfig,
+    spa_load_state_t state, spa_import_type_t type, boolean_t trust_config,
     char **ereport)
 {
 	int error = 0;
@@ -2566,7 +2588,7 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	 * If this is an untrusted config, access the pool in read-only mode.
 	 * This prevents things like resilvering recently removed devices.
 	 */
-	if (!mosconfig)
+	if (!trust_config)
 		spa->spa_mode = FREAD;
 
 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
@@ -2634,7 +2656,7 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	 */
 	if (type != SPA_IMPORT_ASSEMBLE) {
 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
-		error = vdev_validate(rvd, mosconfig);
+		error = vdev_validate(rvd, trust_config);
 		spa_config_exit(spa, SCL_ALL, FTAG);
 
 		if (error != 0)
@@ -2755,7 +2777,7 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	 * can handle missing vdevs.
 	 */
 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VDEV_CHILDREN,
-	    &children) != 0 && mosconfig && type != SPA_IMPORT_ASSEMBLE &&
+	    &children) != 0 && trust_config && type != SPA_IMPORT_ASSEMBLE &&
 	    rvd->vdev_guid_sum != ub->ub_guid_sum)
 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO));
 
@@ -2779,6 +2801,13 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	spa->spa_claim_max_txg = spa->spa_first_txg;
 	spa->spa_prev_software_version = ub->ub_software_version;
 
+	/*
+	 * Everything that we read before we do spa_remove_init() must
+	 * have been rewritten after the last device removal was initiated.
+	 * Otherwise we could be reading from indirect vdevs before
+	 * we have loaded their mappings.
+	 */
+
 	error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool);
 	if (error)
 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
@@ -2787,6 +2816,41 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object) != 0)
 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
 
+	/*
+	 * Validate the config, using the MOS config to fill in any
+	 * information which might be missing.  If we fail to validate
+	 * the config then declare the pool unfit for use. If we're
+	 * assembling a pool from a split, the log is not transferred
+	 * over.
+	 */
+	if (type != SPA_IMPORT_ASSEMBLE) {
+		nvlist_t *mos_config;
+		if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0)
+			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
+
+		if (!spa_config_valid(spa, mos_config)) {
+			nvlist_free(mos_config);
+			return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM,
+			    ENXIO));
+		}
+		nvlist_free(mos_config);
+
+		/*
+		 * Now that we've validated the config, check the state of the
+		 * root vdev.  If it can't be opened, it indicates one or
+		 * more toplevel vdevs are faulted.
+		 */
+		if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN)
+			return (SET_ERROR(ENXIO));
+	}
+
+	/*
+	 * Everything that we read before spa_remove_init() must be stored
+	 * on concreted vdevs.  Therefore we do this as early as possible.
+	 */
+	if (spa_remove_init(spa) != 0)
+		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
+
 	if (spa_version(spa) >= SPA_VERSION_FEATURES) {
 		boolean_t missing_feat_read = B_FALSE;
 		nvlist_t *unsup_feat, *enabled_feat;
@@ -2894,33 +2958,34 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	if (error != 0)
 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
 
-	if (!mosconfig) {
+	if (!trust_config) {
 		uint64_t hostid;
-		nvlist_t *policy = NULL, *nvconfig;
+		nvlist_t *policy = NULL;
+		nvlist_t *mos_config;
 
-		if (load_nvlist(spa, spa->spa_config_object, &nvconfig) != 0)
+		if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0)
 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
 
-		if (!spa_is_root(spa) && nvlist_lookup_uint64(nvconfig,
+		if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config,
 		    ZPOOL_CONFIG_HOSTID, &hostid) == 0) {
 			char *hostname;
 			unsigned long myhostid = 0;
 
-			VERIFY(nvlist_lookup_string(nvconfig,
+			VERIFY(nvlist_lookup_string(mos_config,
 			    ZPOOL_CONFIG_HOSTNAME, &hostname) == 0);
 
 			myhostid = spa_get_hostid();
 			if (hostid && myhostid && hostid != myhostid) {
-				nvlist_free(nvconfig);
+				nvlist_free(mos_config);
 				return (SET_ERROR(EBADF));
 			}
 		}
 		if (nvlist_lookup_nvlist(spa->spa_config,
 		    ZPOOL_REWIND_POLICY, &policy) == 0)
-			VERIFY(nvlist_add_nvlist(nvconfig,
+			VERIFY(nvlist_add_nvlist(mos_config,
 			    ZPOOL_REWIND_POLICY, policy) == 0);
 
-		spa_config_set(spa, nvconfig);
+		spa_config_set(spa, mos_config);
 		spa_unload(spa);
 		spa_deactivate(spa);
 		spa_activate(spa, orig_mode);
@@ -3120,7 +3185,15 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 	/*
 	 * Load the vdev state for all toplevel vdevs.
 	 */
-	vdev_load(rvd);
+	error = vdev_load(rvd);
+	if (error != 0) {
+		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
+	}
+
+	error = spa_condense_init(spa);
+	if (error != 0) {
+		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
+	}
 
 	/*
 	 * Propagate the leaf DTLs we just loaded all the way up the tree.
@@ -3138,38 +3211,10 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 
 	spa_update_dspace(spa);
 
-	/*
-	 * Validate the config, using the MOS config to fill in any
-	 * information which might be missing.  If we fail to validate
-	 * the config then declare the pool unfit for use. If we're
-	 * assembling a pool from a split, the log is not transferred
-	 * over.
-	 */
-	if (type != SPA_IMPORT_ASSEMBLE) {
-		nvlist_t *nvconfig;
-
-		if (load_nvlist(spa, spa->spa_config_object, &nvconfig) != 0)
-			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
-
-		if (!spa_config_valid(spa, nvconfig)) {
-			nvlist_free(nvconfig);
-			return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM,
-			    ENXIO));
-		}
-		nvlist_free(nvconfig);
-
-		/*
-		 * Now that we've validated the config, check the state of the
-		 * root vdev.  If it can't be opened, it indicates one or
-		 * more toplevel vdevs are faulted.
-		 */
-		if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN)
-			return (SET_ERROR(ENXIO));
-
-		if (spa_writeable(spa) && spa_check_logs(spa)) {
-			*ereport = FM_EREPORT_ZFS_LOG_REPLAY;
-			return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG, ENXIO));
-		}
+	if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa) &&
+	    spa_check_logs(spa)) {
+		*ereport = FM_EREPORT_ZFS_LOG_REPLAY;
+		return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG, ENXIO));
 	}
 
 	if (missing_feat_write) {
@@ -3199,6 +3244,18 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 		int need_update = B_FALSE;
 		dsl_pool_t *dp = spa_get_dsl(spa);
 
+		/*
+		 * We must check this before we start the sync thread, because
+		 * we only want to start a condense thread for condense
+		 * operations that were in progress when the pool was
+		 * imported.  Once we start syncing, spa_sync() could
+		 * initiate a condense (and start a thread for it).  In
+		 * that case it would be wrong to start a second
+		 * condense thread.
+		 */
+		boolean_t condense_in_progress =
+		    (spa->spa_condensing_indirect != NULL);
+
 		ASSERT(state != SPA_LOAD_TRYIMPORT);
 
 		/*
@@ -3278,6 +3335,16 @@ spa_load_impl(spa_t *spa, uint64_t pool_guid, nvlist_t *config,
 		 * Clean up any stale temporary dataset userrefs.
 		 */
 		dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool);
+
+		/*
+		 * Note: unlike condensing, we don't need an analogous
+		 * "removal_in_progress" dance because no other thread
+		 * can start a removal while we hold the spa_namespace_lock.
+		 */
+		spa_restart_removal(spa);
+
+		if (condense_in_progress)
+			spa_condense_indirect_restart(spa);
 	}
 
 	return (0);
@@ -3463,7 +3530,7 @@ spa_open_common(const char *pool, spa_t **spapp, void *tag, nvlist_t *nvpolicy,
 			 */
 			spa_unload(spa);
 			spa_deactivate(spa);
-			spa_config_sync(spa, B_TRUE, B_TRUE);
+			spa_write_cachefile(spa, B_TRUE, B_TRUE);
 			spa_remove(spa);
 			if (locked)
 				mutex_exit(&spa_namespace_lock);
@@ -4098,6 +4165,9 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 	spa->spa_uberblock.ub_version = version;
 	spa->spa_ubsync = spa->spa_uberblock;
 	spa->spa_load_state = SPA_LOAD_CREATE;
+	spa->spa_removing_phys.sr_state = DSS_NONE;
+	spa->spa_removing_phys.sr_removing_vdev = -1;
+	spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
 
 	/*
 	 * Create "The Godfather" zio to hold all async IOs
@@ -4283,7 +4353,7 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 	if (dp->dp_root_dir->dd_crypto_obj != 0)
 		VERIFY0(spa_keystore_remove_mapping(spa, root_dsobj, FTAG));
 
-	spa_config_sync(spa, B_FALSE, B_TRUE);
+	spa_write_cachefile(spa, B_FALSE, B_TRUE);
 
 	/*
 	 * Don't count references from objsets that are already closed
@@ -4344,7 +4414,7 @@ spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags)
 		if (props != NULL)
 			spa_configfile_set(spa, props, B_FALSE);
 
-		spa_config_sync(spa, B_FALSE, B_TRUE);
+		spa_write_cachefile(spa, B_FALSE, B_TRUE);
 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
 
 		mutex_exit(&spa_namespace_lock);
@@ -4688,7 +4758,7 @@ export_spa:
 
 	if (new_state != POOL_STATE_UNINITIALIZED) {
 		if (!hardforce)
-			spa_config_sync(spa, B_TRUE, B_TRUE);
+			spa_write_cachefile(spa, B_TRUE, B_TRUE);
 		spa_remove(spa);
 	}
 	mutex_exit(&spa_namespace_lock);
@@ -4780,8 +4850,41 @@ spa_vdev_add(spa_t *spa, nvlist_t *nvroot)
 		return (spa_vdev_exit(spa, vd, txg, error));
 
 	/*
-	 * Transfer each new top-level vdev from vd to rvd.
+	 * If we are in the middle of a device removal, we can only add
+	 * devices which match the existing devices in the pool.
+	 * If we are in the middle of a removal, or have some indirect
+	 * vdevs, we can not add raidz toplevels.
 	 */
+	if (spa->spa_vdev_removal != NULL ||
+	    spa->spa_removing_phys.sr_prev_indirect_vdev != -1) {
+		for (int c = 0; c < vd->vdev_children; c++) {
+			tvd = vd->vdev_child[c];
+			if (spa->spa_vdev_removal != NULL &&
+			    tvd->vdev_ashift !=
+			    spa->spa_vdev_removal->svr_vdev->vdev_ashift) {
+				return (spa_vdev_exit(spa, vd, txg, EINVAL));
+			}
+			/* Fail if top level vdev is raidz */
+			if (tvd->vdev_ops == &vdev_raidz_ops) {
+				return (spa_vdev_exit(spa, vd, txg, EINVAL));
+			}
+			/*
+			 * Need the top level mirror to be
+			 * a mirror of leaf vdevs only
+			 */
+			if (tvd->vdev_ops == &vdev_mirror_ops) {
+				for (uint64_t cid = 0;
+				    cid < tvd->vdev_children; cid++) {
+					vdev_t *cvd = tvd->vdev_child[cid];
+					if (!cvd->vdev_ops->vdev_op_leaf) {
+						return (spa_vdev_exit(spa, vd,
+						    txg, EINVAL));
+					}
+				}
+			}
+		}
+	}
+
 	for (int c = 0; c < vd->vdev_children; c++) {
 
 		/*
@@ -4867,6 +4970,11 @@ spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing)
 
 	oldvd = spa_lookup_by_guid(spa, guid, B_FALSE);
 
+	if (spa->spa_vdev_removal != NULL ||
+	    spa->spa_removing_phys.sr_prev_indirect_vdev != -1) {
+		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
+	}
+
 	if (oldvd == NULL)
 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
 
@@ -5317,7 +5425,7 @@ spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 	/* clear the log and flush everything up to now */
 	activate_slog = spa_passivate_log(spa);
 	(void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
-	error = spa_offline_log(spa);
+	error = spa_reset_logs(spa);
 	txg = spa_vdev_config_enter(spa);
 
 	if (activate_slog)
@@ -5345,7 +5453,7 @@ spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 		vdev_t *vd = rvd->vdev_child[c];
 
 		/* don't count the holes & logs as children */
-		if (vd->vdev_islog || vd->vdev_ishole) {
+		if (vd->vdev_islog || !vdev_is_concrete(vd)) {
 			if (lastlog == 0)
 				lastlog = c;
 			continue;
@@ -5398,7 +5506,7 @@ spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
 		/* make sure there's nothing stopping the split */
 		if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops ||
 		    vml[c]->vdev_islog ||
-		    vml[c]->vdev_ishole ||
+		    !vdev_is_concrete(vml[c]) ||
 		    vml[c]->vdev_isspare ||
 		    vml[c]->vdev_isl2cache ||
 		    !vdev_writeable(vml[c]) ||
@@ -5588,257 +5696,6 @@ out:
 	return (error);
 }
 
-static nvlist_t *
-spa_nvlist_lookup_by_guid(nvlist_t **nvpp, int count, uint64_t target_guid)
-{
-	for (int i = 0; i < count; i++) {
-		uint64_t guid;
-
-		VERIFY(nvlist_lookup_uint64(nvpp[i], ZPOOL_CONFIG_GUID,
-		    &guid) == 0);
-
-		if (guid == target_guid)
-			return (nvpp[i]);
-	}
-
-	return (NULL);
-}
-
-static void
-spa_vdev_remove_aux(nvlist_t *config, char *name, nvlist_t **dev, int count,
-    nvlist_t *dev_to_remove)
-{
-	nvlist_t **newdev = NULL;
-
-	if (count > 1)
-		newdev = kmem_alloc((count - 1) * sizeof (void *), KM_SLEEP);
-
-	for (int i = 0, j = 0; i < count; i++) {
-		if (dev[i] == dev_to_remove)
-			continue;
-		VERIFY(nvlist_dup(dev[i], &newdev[j++], KM_SLEEP) == 0);
-	}
-
-	VERIFY(nvlist_remove(config, name, DATA_TYPE_NVLIST_ARRAY) == 0);
-	VERIFY(nvlist_add_nvlist_array(config, name, newdev, count - 1) == 0);
-
-	for (int i = 0; i < count - 1; i++)
-		nvlist_free(newdev[i]);
-
-	if (count > 1)
-		kmem_free(newdev, (count - 1) * sizeof (void *));
-}
-
-/*
- * Evacuate the device.
- */
-static int
-spa_vdev_remove_evacuate(spa_t *spa, vdev_t *vd)
-{
-	uint64_t txg;
-	int error = 0;
-
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
-	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
-	ASSERT(vd == vd->vdev_top);
-
-	/*
-	 * Evacuate the device.  We don't hold the config lock as writer
-	 * since we need to do I/O but we do keep the
-	 * spa_namespace_lock held.  Once this completes the device
-	 * should no longer have any blocks allocated on it.
-	 */
-	if (vd->vdev_islog) {
-		if (vd->vdev_stat.vs_alloc != 0)
-			error = spa_offline_log(spa);
-	} else {
-		error = SET_ERROR(ENOTSUP);
-	}
-
-	if (error)
-		return (error);
-
-	/*
-	 * The evacuation succeeded.  Remove any remaining MOS metadata
-	 * associated with this vdev, and wait for these changes to sync.
-	 */
-	ASSERT0(vd->vdev_stat.vs_alloc);
-	txg = spa_vdev_config_enter(spa);
-	vd->vdev_removing = B_TRUE;
-	vdev_dirty_leaves(vd, VDD_DTL, txg);
-	vdev_config_dirty(vd);
-	spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
-
-	return (0);
-}
-
-/*
- * Complete the removal by cleaning up the namespace.
- */
-static void
-spa_vdev_remove_from_namespace(spa_t *spa, vdev_t *vd)
-{
-	vdev_t *rvd = spa->spa_root_vdev;
-	uint64_t id = vd->vdev_id;
-	boolean_t last_vdev = (id == (rvd->vdev_children - 1));
-
-	ASSERT(MUTEX_HELD(&spa_namespace_lock));
-	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
-	ASSERT(vd == vd->vdev_top);
-
-	/*
-	 * Only remove any devices which are empty.
-	 */
-	if (vd->vdev_stat.vs_alloc != 0)
-		return;
-
-	(void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
-
-	if (list_link_active(&vd->vdev_state_dirty_node))
-		vdev_state_clean(vd);
-	if (list_link_active(&vd->vdev_config_dirty_node))
-		vdev_config_clean(vd);
-
-	vdev_free(vd);
-
-	if (last_vdev) {
-		vdev_compact_children(rvd);
-	} else {
-		vd = vdev_alloc_common(spa, id, 0, &vdev_hole_ops);
-		vdev_add_child(rvd, vd);
-	}
-	vdev_config_dirty(rvd);
-
-	/*
-	 * Reassess the health of our root vdev.
-	 */
-	vdev_reopen(rvd);
-}
-
-/*
- * Remove a device from the pool -
- *
- * Removing a device from the vdev namespace requires several steps
- * and can take a significant amount of time.  As a result we use
- * the spa_vdev_config_[enter/exit] functions which allow us to
- * grab and release the spa_config_lock while still holding the namespace
- * lock.  During each step the configuration is synced out.
- *
- * Currently, this supports removing only hot spares, slogs, and level 2 ARC
- * devices.
- */
-int
-spa_vdev_remove(spa_t *spa, uint64_t guid, boolean_t unspare)
-{
-	vdev_t *vd;
-	sysevent_t *ev = NULL;
-	metaslab_group_t *mg;
-	nvlist_t **spares, **l2cache, *nv;
-	uint64_t txg = 0;
-	uint_t nspares, nl2cache;
-	int error = 0;
-	boolean_t locked = MUTEX_HELD(&spa_namespace_lock);
-
-	ASSERT(spa_writeable(spa));
-
-	if (!locked)
-		txg = spa_vdev_enter(spa);
-
-	vd = spa_lookup_by_guid(spa, guid, B_FALSE);
-
-	if (spa->spa_spares.sav_vdevs != NULL &&
-	    nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
-	    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0 &&
-	    (nv = spa_nvlist_lookup_by_guid(spares, nspares, guid)) != NULL) {
-		/*
-		 * Only remove the hot spare if it's not currently in use
-		 * in this pool.
-		 */
-		if (vd == NULL || unspare) {
-			if (vd == NULL)
-				vd = spa_lookup_by_guid(spa, guid, B_TRUE);
-			ev = spa_event_create(spa, vd, NULL,
-			    ESC_ZFS_VDEV_REMOVE_AUX);
-			spa_vdev_remove_aux(spa->spa_spares.sav_config,
-			    ZPOOL_CONFIG_SPARES, spares, nspares, nv);
-			spa_load_spares(spa);
-			spa->spa_spares.sav_sync = B_TRUE;
-		} else {
-			error = SET_ERROR(EBUSY);
-		}
-	} else if (spa->spa_l2cache.sav_vdevs != NULL &&
-	    nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
-	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0 &&
-	    (nv = spa_nvlist_lookup_by_guid(l2cache, nl2cache, guid)) != NULL) {
-		/*
-		 * Cache devices can always be removed.
-		 */
-		vd = spa_lookup_by_guid(spa, guid, B_TRUE);
-		ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_AUX);
-		spa_vdev_remove_aux(spa->spa_l2cache.sav_config,
-		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache, nv);
-		spa_load_l2cache(spa);
-		spa->spa_l2cache.sav_sync = B_TRUE;
-	} else if (vd != NULL && vd->vdev_islog) {
-		ASSERT(!locked);
-		ASSERT(vd == vd->vdev_top);
-
-		mg = vd->vdev_mg;
-
-		/*
-		 * Stop allocating from this vdev.
-		 */
-		metaslab_group_passivate(mg);
-
-		/*
-		 * Wait for the youngest allocations and frees to sync,
-		 * and then wait for the deferral of those frees to finish.
-		 */
-		spa_vdev_config_exit(spa, NULL,
-		    txg + TXG_CONCURRENT_STATES + TXG_DEFER_SIZE, 0, FTAG);
-
-		/*
-		 * Attempt to evacuate the vdev.
-		 */
-		error = spa_vdev_remove_evacuate(spa, vd);
-
-		txg = spa_vdev_config_enter(spa);
-
-		/*
-		 * If we couldn't evacuate the vdev, unwind.
-		 */
-		if (error) {
-			metaslab_group_activate(mg);
-			return (spa_vdev_exit(spa, NULL, txg, error));
-		}
-
-		/*
-		 * Clean up the vdev namespace.
-		 */
-		ev = spa_event_create(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE_DEV);
-		spa_vdev_remove_from_namespace(spa, vd);
-
-	} else if (vd != NULL) {
-		/*
-		 * Normal vdevs cannot be removed (yet).
-		 */
-		error = SET_ERROR(ENOTSUP);
-	} else {
-		/*
-		 * There is no vdev of any kind with the specified guid.
-		 */
-		error = SET_ERROR(ENOENT);
-	}
-
-	if (!locked)
-		error = spa_vdev_exit(spa, NULL, txg, error);
-
-	if (ev)
-		spa_event_post(ev);
-
-	return (error);
-}
-
 /*
  * Find any device that's done replacing, or a vdev marked 'unspare' that's
  * currently spared, so we can detach it.
@@ -6205,9 +6062,12 @@ spa_async_suspend(spa_t *spa)
 {
 	mutex_enter(&spa->spa_async_lock);
 	spa->spa_async_suspended++;
-	while (spa->spa_async_thread != NULL)
+	while (spa->spa_async_thread != NULL ||
+	    spa->spa_condense_thread != NULL)
 		cv_wait(&spa->spa_async_cv, &spa->spa_async_lock);
 	mutex_exit(&spa->spa_async_lock);
+
+	spa_vdev_remove_suspend(spa);
 }
 
 void
@@ -6217,6 +6077,7 @@ spa_async_resume(spa_t *spa)
 	ASSERT(spa->spa_async_suspended != 0);
 	spa->spa_async_suspended--;
 	mutex_exit(&spa->spa_async_lock);
+	spa_restart_removal(spa);
 }
 
 static boolean_t
@@ -6763,6 +6624,39 @@ spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx)
 	rrw_exit(&dp->dp_config_rwlock, FTAG);
 }
 
+static void
+vdev_indirect_state_sync_verify(vdev_t *vd)
+{
+	ASSERTV(vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping);
+	ASSERTV(vdev_indirect_births_t *vib = vd->vdev_indirect_births);
+
+	if (vd->vdev_ops == &vdev_indirect_ops) {
+		ASSERT(vim != NULL);
+		ASSERT(vib != NULL);
+	}
+
+	if (vdev_obsolete_sm_object(vd) != 0) {
+		ASSERT(vd->vdev_obsolete_sm != NULL);
+		ASSERT(vd->vdev_removing ||
+		    vd->vdev_ops == &vdev_indirect_ops);
+		ASSERT(vdev_indirect_mapping_num_entries(vim) > 0);
+		ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0);
+
+		ASSERT3U(vdev_obsolete_sm_object(vd), ==,
+		    space_map_object(vd->vdev_obsolete_sm));
+		ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=,
+		    space_map_allocated(vd->vdev_obsolete_sm));
+	}
+	ASSERT(vd->vdev_obsolete_segments != NULL);
+
+	/*
+	 * Since frees / remaps to an indirect vdev can only
+	 * happen in syncing context, the obsolete segments
+	 * tree must be empty when we start syncing.
+	 */
+	ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
+}
+
 /*
  * Sync the specified transaction group.  New blocks may be dirtied as
  * part of the process, so we iterate until it converges.
@@ -6783,6 +6677,13 @@ spa_sync(spa_t *spa, uint64_t txg)
 	VERIFY(spa_writeable(spa));
 
 	/*
+	 * Wait for i/os issued in open context that need to complete
+	 * before this txg syncs.
+	 */
+	VERIFY0(zio_wait(spa->spa_txg_zio[txg & TXG_MASK]));
+	spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL, 0);
+
+	/*
 	 * Lock out configuration changes.
 	 */
 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
@@ -6879,6 +6780,16 @@ spa_sync(spa_t *spa, uint64_t txg)
 	ASSERT3U(mc->mc_alloc_max_slots, <=,
 	    max_queue_depth * rvd->vdev_children);
 
+	for (int c = 0; c < rvd->vdev_children; c++) {
+		vdev_t *vd = rvd->vdev_child[c];
+		vdev_indirect_state_sync_verify(vd);
+
+		if (vdev_indirect_should_condense(vd)) {
+			spa_condense_indirect_start_sync(vd, tx);
+			break;
+		}
+	}
+
 	/*
 	 * Iterate to convergence.
 	 */
@@ -6908,7 +6819,11 @@ spa_sync(spa_t *spa, uint64_t txg)
 		ddt_sync(spa, txg);
 		dsl_scan_sync(dp, tx);
 
-		while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg)))
+		if (spa->spa_vdev_removal != NULL)
+			svr_sync(spa, tx);
+
+		while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg))
+		    != NULL)
 			vdev_sync(vd, txg);
 
 		if (pass == 1) {
@@ -6962,6 +6877,10 @@ spa_sync(spa_t *spa, uint64_t txg)
 	}
 #endif
 
+	if (spa->spa_vdev_removal != NULL) {
+		ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]);
+	}
+
 	/*
 	 * Rewrite the vdev configuration (which includes the uberblock)
 	 * to commit the transaction group.
@@ -6986,7 +6905,8 @@ spa_sync(spa_t *spa, uint64_t txg)
 
 			for (int c = 0; c < children; c++) {
 				vd = rvd->vdev_child[(c0 + c) % children];
-				if (vd->vdev_ms_array == 0 || vd->vdev_islog)
+				if (vd->vdev_ms_array == 0 || vd->vdev_islog ||
+				    !vdev_is_concrete(vd))
 					continue;
 				svd[svdcount++] = vd;
 				if (svdcount == SPA_DVAS_PER_BP)
@@ -7223,7 +7143,7 @@ spa_has_active_shared_spare(spa_t *spa)
 	return (B_FALSE);
 }
 
-static sysevent_t *
+sysevent_t *
 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
 {
 	sysevent_t *ev = NULL;
@@ -7239,7 +7159,7 @@ spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
 	return (ev);
 }
 
-static void
+void
 spa_event_post(sysevent_t *ev)
 {
 #ifdef _KERNEL
@@ -7286,7 +7206,6 @@ EXPORT_SYMBOL(spa_scan_get_stats);
 EXPORT_SYMBOL(spa_vdev_add);
 EXPORT_SYMBOL(spa_vdev_attach);
 EXPORT_SYMBOL(spa_vdev_detach);
-EXPORT_SYMBOL(spa_vdev_remove);
 EXPORT_SYMBOL(spa_vdev_setpath);
 EXPORT_SYMBOL(spa_vdev_setfru);
 EXPORT_SYMBOL(spa_vdev_split_mirror);