OpenZFS 9290 - device removal reduces redundancy of mirrors

Mirrors are supposed to provide redundancy in the face of whole-disk
failure and silent damage (e.g. some data on disk is not right, but ZFS
hasn't detected the whole device as being broken). However, the current
device removal implementation bypasses some of the mirror's redundancy.
Note that in no case is incorrect data returned, but we might get a
checksum error when we should have been able to find the right data.

There are two underlying problems:

1. When we remove a mirror device, we only read one side of the mirror.
Since we can't verify the checksum, this side may be silently bad, but
the good data is on the other side of the mirror (which we didn't read).
This can cause the removal to "bake in" the busted data – all copies of
the data in the new location are the same, busted version, while we left
the good version behind.

The fix for this is to read and copy both sides of the mirror. If the
old and new vdevs are mirrors, we will read both sides of the old
mirror, and write each copy to the corresponding side of the new mirror.
(If the old and new vdevs have a different number of children, we will
do this as best as possible.) Even though we aren't verifying checksums,
this ensures that as long as there's a good copy of the data, we'll have
a good copy after the removal, even if there's silent damage to one side
of the mirror. If we're removing a mirror that has some silent damage,
we'll have exactly the same damage in the new location (assuming that
the new location is also a mirror).

2. When we read from an indirect vdev that points to a mirror vdev, we
only consider one copy of the data. This can lead to reduced effective
redundancy, because we might read a bad copy of the data from one side
of the mirror, and not retry the other, good side of the mirror.

Note that the problem is not with the removal process, but rather after
the removal has completed (having copied correct data to both sides of
the mirror), if one side of the new mirror is silently damaged, we
encounter the problem when reading the relocated data via the indirect
vdev. Also note that the problem doesn't occur when ZFS knows that one
side of the mirror is bad, e.g. when a disk entirely fails or is
offlined.

The impact is that reads (from indirect vdevs that point to mirrors) may
return a checksum error even though the good data exists on one side of
the mirror, and scrub doesn't repair all data on the mirror (if some of
it is pointed to via an indirect vdev).

The fix for this is complicated by "split blocks" - one logical block
may be split into two (or more) pieces with each piece moved to a
different new location. In this case we need to read all versions of
each split (one from each side of the mirror), and figure out which
combination of versions results in the correct checksum, and then repair
the incorrect versions.

This ensures that we supply the same redundancy whether you use device
removal or not. For example, if a mirror has small silent errors on all
of its children, we can still reconstruct the correct data, as long as
those errors are at sufficiently-separated offsets (specifically,
separated by the largest block size - default of 128KB, but up to 16MB).

Porting notes:

* A new indirect vdev check was moved from dsl_scan_needs_resilver_cb()
  to dsl_scan_needs_resilver(), which was added to ZoL as part of the
  sequential scrub work.

* Passed NULL for zfs_ereport_post_checksum()'s zbookmark_phys_t
  parameter.  The extra parameter is unique to ZoL.

* When posting indirect checksum errors the ABD can be passed directly,
  zfs_ereport_post_checksum() is not yet ABD-aware in OpenZFS.

Authored by: Matthew Ahrens <[email protected]>
Reviewed by: Tim Chase <[email protected]>
Reviewed by: Brian Behlendorf <[email protected]>
Ported-by: Tim Chase <[email protected]>

OpenZFS-issue: https://illumos.org/issues/9290
OpenZFS-commit: https://github.com/openzfs/openzfs/pull/591
Closes #6900

9 files changed, 829 insertions, 218 deletions

diff --git a/module/zfs/dsl_scan.c b/module/zfs/dsl_scan.c
index 53953a6c5..b87b4d555 100644
--- a/module/zfs/dsl_scan.c
+++ b/module/zfs/dsl_scan.c
@@ -2959,6 +2959,19 @@ dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
 {
 	vdev_t *vd;
 
+	vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
+
+	if (vd->vdev_ops == &vdev_indirect_ops) {
+		/*
+		 * The indirect vdev can point to multiple
+		 * vdevs.  For simplicity, always create
+		 * the resilver zio_t. zio_vdev_io_start()
+		 * will bypass the child resilver i/o's if
+		 * they are on vdevs that don't have DTL's.
+		 */
+		return (B_TRUE);
+	}
+
 	if (DVA_GET_GANG(dva)) {
 		/*
 		 * Gang members may be spread across multiple
@@ -2971,8 +2984,6 @@ dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
 		return (B_TRUE);
 	}
 
-	vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
-
 	/*
 	 * Check if the txg falls within the range which must be
 	 * resilvered.  DVAs outside this range can always be skipped.
diff --git a/module/zfs/metaslab.c b/module/zfs/metaslab.c
index 25090f089..52a60cc5e 100644
--- a/module/zfs/metaslab.c
+++ b/module/zfs/metaslab.c
@@ -3234,7 +3234,7 @@ metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
 		return;
 
 	if (spa->spa_vdev_removal != NULL &&
-	    spa->spa_vdev_removal->svr_vdev == vd &&
+	    spa->spa_vdev_removal->svr_vdev_id == vd->vdev_id &&
 	    vdev_is_concrete(vd)) {
 		/*
 		 * Note: we check if the vdev is concrete because when
diff --git a/module/zfs/spa.c b/module/zfs/spa.c
index 08fc7bbda..9a5346b42 100644
--- a/module/zfs/spa.c
+++ b/module/zfs/spa.c
@@ -4860,8 +4860,7 @@ spa_vdev_add(spa_t *spa, nvlist_t *nvroot)
 		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) {
+			    tvd->vdev_ashift != spa->spa_max_ashift) {
 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
 			}
 			/* Fail if top level vdev is raidz */
@@ -4970,10 +4969,8 @@ 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) {
+	if (spa->spa_vdev_removal != NULL)
 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
-	}
 
 	if (oldvd == NULL)
 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
diff --git a/module/zfs/spa_misc.c b/module/zfs/spa_misc.c
index 5c6e82567..54299a7e4 100644
--- a/module/zfs/spa_misc.c
+++ b/module/zfs/spa_misc.c
@@ -1708,9 +1708,12 @@ spa_update_dspace(spa_t *spa)
 		 * allocated twice (on the old device and the new
 		 * device).
 		 */
-		vdev_t *vd = spa->spa_vdev_removal->svr_vdev;
+		spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
+		vdev_t *vd =
+		    vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id);
 		spa->spa_dspace -= spa_deflate(spa) ?
 		    vd->vdev_stat.vs_dspace : vd->vdev_stat.vs_space;
+		spa_config_exit(spa, SCL_VDEV, FTAG);
 	}
 }
 
diff --git a/module/zfs/vdev.c b/module/zfs/vdev.c
index f3e3f90fa..0dea0b4e4 100644
--- a/module/zfs/vdev.c
+++ b/module/zfs/vdev.c
@@ -857,6 +857,32 @@ vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
 	svd->vdev_stat.vs_space = 0;
 	svd->vdev_stat.vs_dspace = 0;
 
+	/*
+	 * State which may be set on a top-level vdev that's in the
+	 * process of being removed.
+	 */
+	ASSERT0(tvd->vdev_indirect_config.vic_births_object);
+	ASSERT0(tvd->vdev_indirect_config.vic_mapping_object);
+	ASSERT3U(tvd->vdev_indirect_config.vic_prev_indirect_vdev, ==, -1ULL);
+	ASSERT3P(tvd->vdev_indirect_mapping, ==, NULL);
+	ASSERT3P(tvd->vdev_indirect_births, ==, NULL);
+	ASSERT3P(tvd->vdev_obsolete_sm, ==, NULL);
+	ASSERT0(tvd->vdev_removing);
+	tvd->vdev_removing = svd->vdev_removing;
+	tvd->vdev_indirect_config = svd->vdev_indirect_config;
+	tvd->vdev_indirect_mapping = svd->vdev_indirect_mapping;
+	tvd->vdev_indirect_births = svd->vdev_indirect_births;
+	range_tree_swap(&svd->vdev_obsolete_segments,
+	    &tvd->vdev_obsolete_segments);
+	tvd->vdev_obsolete_sm = svd->vdev_obsolete_sm;
+	svd->vdev_indirect_config.vic_mapping_object = 0;
+	svd->vdev_indirect_config.vic_births_object = 0;
+	svd->vdev_indirect_config.vic_prev_indirect_vdev = -1ULL;
+	svd->vdev_indirect_mapping = NULL;
+	svd->vdev_indirect_births = NULL;
+	svd->vdev_obsolete_sm = NULL;
+	svd->vdev_removing = 0;
+
 	for (t = 0; t < TXG_SIZE; t++) {
 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
diff --git a/module/zfs/vdev_indirect.c b/module/zfs/vdev_indirect.c
index 86a05daa8..3ccdfee3b 100644
--- a/module/zfs/vdev_indirect.c
+++ b/module/zfs/vdev_indirect.c
@@ -23,6 +23,7 @@
 #include <sys/vdev_impl.h>
 #include <sys/fs/zfs.h>
 #include <sys/zio.h>
+#include <sys/zio_checksum.h>
 #include <sys/metaslab.h>
 #include <sys/refcount.h>
 #include <sys/dmu.h>
@@ -44,10 +45,11 @@
  * "vdev_remap" operation that executes a callback on each contiguous
  * segment of the new location.  This function is used in multiple ways:
  *
- *  - reads and repair writes to this device use the callback to create
- *    a child io for each mapped segment.
+ *  - i/os to this vdev use the callback to determine where the
+ *    data is now located, and issue child i/os for each segment's new
+ *    location.
  *
- *  - frees and claims to this device use the callback to free or claim
+ *  - frees and claims to this vdev use the callback to free or claim
  *    each mapped segment.  (Note that we don't actually need to claim
  *    log blocks on indirect vdevs, because we don't allocate to
  *    removing vdevs.  However, zdb uses zio_claim() for its leak
@@ -202,6 +204,95 @@ unsigned long zfs_condense_min_mapping_bytes = 128 * 1024;
 int zfs_condense_indirect_commit_entry_delay_ms = 0;
 
 /*
+ * If a split block contains more than this many segments, consider it too
+ * computationally expensive to check all (2^num_segments) possible
+ * combinations. Instead, try at most 2^_segments_max randomly-selected
+ * combinations.
+ *
+ * This is reasonable if only a few segment copies are damaged and the
+ * majority of segment copies are good. It allows all segment copies to
+ * participate fairly in the reconstruction and prevents repeated use of
+ * one bad copy.
+ */
+int zfs_reconstruct_indirect_segments_max = 10;
+
+/*
+ * The indirect_child_t represents the vdev that we will read from, when we
+ * need to read all copies of the data (e.g. for scrub or reconstruction).
+ * For plain (non-mirror) top-level vdevs (i.e. is_vdev is not a mirror),
+ * ic_vdev is the same as is_vdev.  However, for mirror top-level vdevs,
+ * ic_vdev is a child of the mirror.
+ */
+typedef struct indirect_child {
+	abd_t *ic_data;
+	vdev_t *ic_vdev;
+} indirect_child_t;
+
+/*
+ * The indirect_split_t represents one mapped segment of an i/o to the
+ * indirect vdev. For non-split (contiguously-mapped) blocks, there will be
+ * only one indirect_split_t, with is_split_offset==0 and is_size==io_size.
+ * For split blocks, there will be several of these.
+ */
+typedef struct indirect_split {
+	list_node_t is_node; /* link on iv_splits */
+
+	/*
+	 * is_split_offset is the offset into the i/o.
+	 * This is the sum of the previous splits' is_size's.
+	 */
+	uint64_t is_split_offset;
+
+	vdev_t *is_vdev; /* top-level vdev */
+	uint64_t is_target_offset; /* offset on is_vdev */
+	uint64_t is_size;
+	int is_children; /* number of entries in is_child[] */
+
+	/*
+	 * is_good_child is the child that we are currently using to
+	 * attempt reconstruction.
+	 */
+	int is_good_child;
+
+	indirect_child_t is_child[1]; /* variable-length */
+} indirect_split_t;
+
+/*
+ * The indirect_vsd_t is associated with each i/o to the indirect vdev.
+ * It is the "Vdev-Specific Data" in the zio_t's io_vsd.
+ */
+typedef struct indirect_vsd {
+	boolean_t iv_split_block;
+	boolean_t iv_reconstruct;
+
+	list_t iv_splits; /* list of indirect_split_t's */
+} indirect_vsd_t;
+
+static void
+vdev_indirect_map_free(zio_t *zio)
+{
+	indirect_vsd_t *iv = zio->io_vsd;
+
+	indirect_split_t *is;
+	while ((is = list_head(&iv->iv_splits)) != NULL) {
+		for (int c = 0; c < is->is_children; c++) {
+			indirect_child_t *ic = &is->is_child[c];
+			if (ic->ic_data != NULL)
+				abd_free(ic->ic_data);
+		}
+		list_remove(&iv->iv_splits, is);
+		kmem_free(is,
+		    offsetof(indirect_split_t, is_child[is->is_children]));
+	}
+	kmem_free(iv, sizeof (*iv));
+}
+
+static const zio_vsd_ops_t vdev_indirect_vsd_ops = {
+	vdev_indirect_map_free,
+	zio_vsd_default_cksum_report
+};
+
+/*
  * Mark the given offset and size as being obsolete in the given txg.
  */
 void
@@ -814,12 +905,6 @@ vdev_indirect_close(vdev_t *vd)
 }
 
 /* ARGSUSED */
-static void
-vdev_indirect_io_done(zio_t *zio)
-{
-}
-
-/* ARGSUSED */
 static int
 vdev_indirect_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
     uint64_t *ashift)
@@ -990,41 +1075,471 @@ vdev_indirect_child_io_done(zio_t *zio)
 	abd_put(zio->io_abd);
 }
 
+/*
+ * This is a callback for vdev_indirect_remap() which allocates an
+ * indirect_split_t for each split segment and adds it to iv_splits.
+ */
 static void
-vdev_indirect_io_start_cb(uint64_t split_offset, vdev_t *vd, uint64_t offset,
+vdev_indirect_gather_splits(uint64_t split_offset, vdev_t *vd, uint64_t offset,
     uint64_t size, void *arg)
 {
 	zio_t *zio = arg;
+	indirect_vsd_t *iv = zio->io_vsd;
 
 	ASSERT3P(vd, !=, NULL);
 
 	if (vd->vdev_ops == &vdev_indirect_ops)
 		return;
 
-	zio_nowait(zio_vdev_child_io(zio, NULL, vd, offset,
-	    abd_get_offset(zio->io_abd, split_offset),
-	    size, zio->io_type, zio->io_priority,
-	    0, vdev_indirect_child_io_done, zio));
+	int n = 1;
+	if (vd->vdev_ops == &vdev_mirror_ops)
+		n = vd->vdev_children;
+
+	indirect_split_t *is =
+	    kmem_zalloc(offsetof(indirect_split_t, is_child[n]), KM_SLEEP);
+
+	is->is_children = n;
+	is->is_size = size;
+	is->is_split_offset = split_offset;
+	is->is_target_offset = offset;
+	is->is_vdev = vd;
+
+	/*
+	 * Note that we only consider multiple copies of the data for
+	 * *mirror* vdevs.  We don't for "replacing" or "spare" vdevs, even
+	 * though they use the same ops as mirror, because there's only one
+	 * "good" copy under the replacing/spare.
+	 */
+	if (vd->vdev_ops == &vdev_mirror_ops) {
+		for (int i = 0; i < n; i++) {
+			is->is_child[i].ic_vdev = vd->vdev_child[i];
+		}
+	} else {
+		is->is_child[0].ic_vdev = vd;
+	}
+
+	list_insert_tail(&iv->iv_splits, is);
+}
+
+static void
+vdev_indirect_read_split_done(zio_t *zio)
+{
+	indirect_child_t *ic = zio->io_private;
+
+	if (zio->io_error != 0) {
+		/*
+		 * Clear ic_data to indicate that we do not have data for this
+		 * child.
+		 */
+		abd_free(ic->ic_data);
+		ic->ic_data = NULL;
+	}
+}
+
+/*
+ * Issue reads for all copies (mirror children) of all splits.
+ */
+static void
+vdev_indirect_read_all(zio_t *zio)
+{
+	indirect_vsd_t *iv = zio->io_vsd;
+
+	for (indirect_split_t *is = list_head(&iv->iv_splits);
+	    is != NULL; is = list_next(&iv->iv_splits, is)) {
+		for (int i = 0; i < is->is_children; i++) {
+			indirect_child_t *ic = &is->is_child[i];
+
+			if (!vdev_readable(ic->ic_vdev))
+				continue;
+
+			/*
+			 * Note, we may read from a child whose DTL
+			 * indicates that the data may not be present here.
+			 * While this might result in a few i/os that will
+			 * likely return incorrect data, it simplifies the
+			 * code since we can treat scrub and resilver
+			 * identically.  (The incorrect data will be
+			 * detected and ignored when we verify the
+			 * checksum.)
+			 */
+
+			ic->ic_data = abd_alloc_sametype(zio->io_abd,
+			    is->is_size);
+
+			zio_nowait(zio_vdev_child_io(zio, NULL,
+			    ic->ic_vdev, is->is_target_offset, ic->ic_data,
+			    is->is_size, zio->io_type, zio->io_priority, 0,
+			    vdev_indirect_read_split_done, ic));
+		}
+	}
+	iv->iv_reconstruct = B_TRUE;
 }
 
 static void
 vdev_indirect_io_start(zio_t *zio)
 {
 	ASSERTV(spa_t *spa = zio->io_spa);
+	indirect_vsd_t *iv = kmem_zalloc(sizeof (*iv), KM_SLEEP);
+	list_create(&iv->iv_splits,
+	    sizeof (indirect_split_t), offsetof(indirect_split_t, is_node));
+
+	zio->io_vsd = iv;
+	zio->io_vsd_ops = &vdev_indirect_vsd_ops;
 
 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
 	if (zio->io_type != ZIO_TYPE_READ) {
 		ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
-		ASSERT((zio->io_flags &
-		    (ZIO_FLAG_SELF_HEAL | ZIO_FLAG_INDUCE_DAMAGE)) != 0);
+		/*
+		 * Note: this code can handle other kinds of writes,
+		 * but we don't expect them.
+		 */
+		ASSERT((zio->io_flags & (ZIO_FLAG_SELF_HEAL |
+		    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)) != 0);
 	}
 
 	vdev_indirect_remap(zio->io_vd, zio->io_offset, zio->io_size,
-	    vdev_indirect_io_start_cb, zio);
+	    vdev_indirect_gather_splits, zio);
+
+	indirect_split_t *first = list_head(&iv->iv_splits);
+	if (first->is_size == zio->io_size) {
+		/*
+		 * This is not a split block; we are pointing to the entire
+		 * data, which will checksum the same as the original data.
+		 * Pass the BP down so that the child i/o can verify the
+		 * checksum, and try a different location if available
+		 * (e.g. on a mirror).
+		 *
+		 * While this special case could be handled the same as the
+		 * general (split block) case, doing it this way ensures
+		 * that the vast majority of blocks on indirect vdevs
+		 * (which are not split) are handled identically to blocks
+		 * on non-indirect vdevs.  This allows us to be less strict
+		 * about performance in the general (but rare) case.
+		 */
+		ASSERT0(first->is_split_offset);
+		ASSERT3P(list_next(&iv->iv_splits, first), ==, NULL);
+		zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
+		    first->is_vdev, first->is_target_offset,
+		    abd_get_offset(zio->io_abd, 0),
+		    zio->io_size, zio->io_type, zio->io_priority, 0,
+		    vdev_indirect_child_io_done, zio));
+	} else {
+		iv->iv_split_block = B_TRUE;
+		if (zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER)) {
+			/*
+			 * Read all copies.  Note that for simplicity,
+			 * we don't bother consulting the DTL in the
+			 * resilver case.
+			 */
+			vdev_indirect_read_all(zio);
+		} else {
+			/*
+			 * Read one copy of each split segment, from the
+			 * top-level vdev.  Since we don't know the
+			 * checksum of each split individually, the child
+			 * zio can't ensure that we get the right data.
+			 * E.g. if it's a mirror, it will just read from a
+			 * random (healthy) leaf vdev.  We have to verify
+			 * the checksum in vdev_indirect_io_done().
+			 */
+			for (indirect_split_t *is = list_head(&iv->iv_splits);
+			    is != NULL; is = list_next(&iv->iv_splits, is)) {
+				zio_nowait(zio_vdev_child_io(zio, NULL,
+				    is->is_vdev, is->is_target_offset,
+				    abd_get_offset(zio->io_abd,
+				    is->is_split_offset), is->is_size,
+				    zio->io_type, zio->io_priority, 0,
+				    vdev_indirect_child_io_done, zio));
+			}
+
+		}
+	}
 
 	zio_execute(zio);
 }
 
+/*
+ * Report a checksum error for a child.
+ */
+static void
+vdev_indirect_checksum_error(zio_t *zio,
+    indirect_split_t *is, indirect_child_t *ic)
+{
+	vdev_t *vd = ic->ic_vdev;
+
+	if (zio->io_flags & ZIO_FLAG_SPECULATIVE)
+		return;
+
+	mutex_enter(&vd->vdev_stat_lock);
+	vd->vdev_stat.vs_checksum_errors++;
+	mutex_exit(&vd->vdev_stat_lock);
+
+	zio_bad_cksum_t zbc = {{{ 0 }}};
+	abd_t *bad_abd = ic->ic_data;
+	abd_t *good_abd = is->is_child[is->is_good_child].ic_data;
+	zfs_ereport_post_checksum(zio->io_spa, vd, NULL, zio,
+	    is->is_target_offset, is->is_size, good_abd, bad_abd, &zbc);
+}
+
+/*
+ * Issue repair i/os for any incorrect copies.  We do this by comparing
+ * each split segment's correct data (is_good_child's ic_data) with each
+ * other copy of the data.  If they differ, then we overwrite the bad data
+ * with the good copy.  Note that we do this without regard for the DTL's,
+ * which simplifies this code and also issues the optimal number of writes
+ * (based on which copies actually read bad data, as opposed to which we
+ * think might be wrong).  For the same reason, we always use
+ * ZIO_FLAG_SELF_HEAL, to bypass the DTL check in zio_vdev_io_start().
+ */
+static void
+vdev_indirect_repair(zio_t *zio)
+{
+	indirect_vsd_t *iv = zio->io_vsd;
+
+	enum zio_flag flags = ZIO_FLAG_IO_REPAIR;
+
+	if (!(zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER)))
+		flags |= ZIO_FLAG_SELF_HEAL;
+
+	if (!spa_writeable(zio->io_spa))
+		return;
+
+	for (indirect_split_t *is = list_head(&iv->iv_splits);
+	    is != NULL; is = list_next(&iv->iv_splits, is)) {
+		indirect_child_t *good_child = &is->is_child[is->is_good_child];
+
+		for (int c = 0; c < is->is_children; c++) {
+			indirect_child_t *ic = &is->is_child[c];
+			if (ic == good_child)
+				continue;
+			if (ic->ic_data == NULL)
+				continue;
+			if (abd_cmp(good_child->ic_data, ic->ic_data) == 0)
+				continue;
+
+			zio_nowait(zio_vdev_child_io(zio, NULL,
+			    ic->ic_vdev, is->is_target_offset,
+			    good_child->ic_data, is->is_size,
+			    ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
+			    ZIO_FLAG_IO_REPAIR | ZIO_FLAG_SELF_HEAL,
+			    NULL, NULL));
+
+			vdev_indirect_checksum_error(zio, is, ic);
+		}
+	}
+}
+
+/*
+ * Report checksum errors on all children that we read from.
+ */
+static void
+vdev_indirect_all_checksum_errors(zio_t *zio)
+{
+	indirect_vsd_t *iv = zio->io_vsd;
+
+	if (zio->io_flags & ZIO_FLAG_SPECULATIVE)
+		return;
+
+	for (indirect_split_t *is = list_head(&iv->iv_splits);
+	    is != NULL; is = list_next(&iv->iv_splits, is)) {
+		for (int c = 0; c < is->is_children; c++) {
+			indirect_child_t *ic = &is->is_child[c];
+
+			if (ic->ic_data == NULL)
+				continue;
+
+			vdev_t *vd = ic->ic_vdev;
+
+			mutex_enter(&vd->vdev_stat_lock);
+			vd->vdev_stat.vs_checksum_errors++;
+			mutex_exit(&vd->vdev_stat_lock);
+
+			zfs_ereport_post_checksum(zio->io_spa, vd, NULL, zio,
+			    is->is_target_offset, is->is_size,
+			    NULL, NULL, NULL);
+		}
+	}
+}
+
+/*
+ * This function is called when we have read all copies of the data and need
+ * to try to find a combination of copies that gives us the right checksum.
+ *
+ * If we pointed to any mirror vdevs, this effectively does the job of the
+ * mirror.  The mirror vdev code can't do its own job because we don't know
+ * the checksum of each split segment individually.  We have to try every
+ * combination of copies of split segments, until we find one that checksums
+ * correctly.  (Or until we have tried all combinations, or have tried
+ * 2^zfs_reconstruct_indirect_segments_max combinations.  In these cases we
+ * set io_error to ECKSUM to propagate the error up to the user.)
+ *
+ * For example, if we have 3 segments in the split,
+ * and each points to a 2-way mirror, we will have the following pieces of
+ * data:
+ *
+ *       |     mirror child
+ * split |     [0]        [1]
+ * ======|=====================
+ *   A   |  data_A_0   data_A_1
+ *   B   |  data_B_0   data_B_1
+ *   C   |  data_C_0   data_C_1
+ *
+ * We will try the following (mirror children)^(number of splits) (2^3=8)
+ * combinations, which is similar to bitwise-little-endian counting in
+ * binary.  In general each "digit" corresponds to a split segment, and the
+ * base of each digit is is_children, which can be different for each
+ * digit.
+ *
+ * "low bit"        "high bit"
+ *        v                 v
+ * data_A_0 data_B_0 data_C_0
+ * data_A_1 data_B_0 data_C_0
+ * data_A_0 data_B_1 data_C_0
+ * data_A_1 data_B_1 data_C_0
+ * data_A_0 data_B_0 data_C_1
+ * data_A_1 data_B_0 data_C_1
+ * data_A_0 data_B_1 data_C_1
+ * data_A_1 data_B_1 data_C_1
+ *
+ * Note that the split segments may be on the same or different top-level
+ * vdevs. In either case, we try lots of combinations (see
+ * zfs_reconstruct_indirect_segments_max).  This ensures that if a mirror has
+ * small silent errors on all of its children, we can still reconstruct the
+ * correct data, as long as those errors are at sufficiently-separated
+ * offsets (specifically, separated by the largest block size - default of
+ * 128KB, but up to 16MB).
+ */
+static void
+vdev_indirect_reconstruct_io_done(zio_t *zio)
+{
+	indirect_vsd_t *iv = zio->io_vsd;
+	uint64_t attempts = 0;
+	uint64_t attempts_max = 1ULL << zfs_reconstruct_indirect_segments_max;
+	int segments = 0;
+
+	for (indirect_split_t *is = list_head(&iv->iv_splits);
+	    is != NULL; is = list_next(&iv->iv_splits, is))
+		segments++;
+
+	for (;;) {
+		/* copy data from splits to main zio */
+		int ret;
+		for (indirect_split_t *is = list_head(&iv->iv_splits);
+		    is != NULL; is = list_next(&iv->iv_splits, is)) {
+
+			/*
+			 * If this child failed, its ic_data will be NULL.
+			 * Skip this combination.
+			 */
+			if (is->is_child[is->is_good_child].ic_data == NULL) {
+				ret = EIO;
+				goto next;
+			}
+
+			abd_copy_off(zio->io_abd,
+			    is->is_child[is->is_good_child].ic_data,
+			    is->is_split_offset, 0, is->is_size);
+		}
+
+		/* See if this checksum matches. */
+		zio_bad_cksum_t zbc;
+		ret = zio_checksum_error(zio, &zbc);
+		if (ret == 0) {
+			/* Found a matching checksum.  Issue repair i/os. */
+			vdev_indirect_repair(zio);
+			zio_checksum_verified(zio);
+			return;
+		}
+
+		/*
+		 * Checksum failed; try a different combination of split
+		 * children.
+		 */
+		boolean_t more;
+next:
+		more = B_FALSE;
+		if (segments <= zfs_reconstruct_indirect_segments_max) {
+			/*
+			 * There are relatively few segments, so
+			 * deterministically check all combinations.  We do
+			 * this by by adding one to the first split's
+			 * good_child.  If it overflows, then "carry over" to
+			 * the next split (like counting in base is_children,
+			 * but each digit can have a different base).
+			 */
+			for (indirect_split_t *is = list_head(&iv->iv_splits);
+			    is != NULL; is = list_next(&iv->iv_splits, is)) {
+				is->is_good_child++;
+				if (is->is_good_child < is->is_children) {
+					more = B_TRUE;
+					break;
+				}
+				is->is_good_child = 0;
+			}
+		} else if (++attempts < attempts_max) {
+			/*
+			 * There are too many combinations to try all of them
+			 * in a reasonable amount of time, so try a fixed
+			 * number of random combinations, after which we'll
+			 * consider the block unrecoverable.
+			 */
+			for (indirect_split_t *is = list_head(&iv->iv_splits);
+			    is != NULL; is = list_next(&iv->iv_splits, is)) {
+				is->is_good_child =
+				    spa_get_random(is->is_children);
+			}
+			more = B_TRUE;
+		}
+		if (!more) {
+			/* All combinations failed. */
+			zio->io_error = ret;
+			vdev_indirect_all_checksum_errors(zio);
+			zio_checksum_verified(zio);
+			return;
+		}
+	}
+}
+
+static void
+vdev_indirect_io_done(zio_t *zio)
+{
+	indirect_vsd_t *iv = zio->io_vsd;
+
+	if (iv->iv_reconstruct) {
+		/*
+		 * We have read all copies of the data (e.g. from mirrors),
+		 * either because this was a scrub/resilver, or because the
+		 * one-copy read didn't checksum correctly.
+		 */
+		vdev_indirect_reconstruct_io_done(zio);
+		return;
+	}
+
+	if (!iv->iv_split_block) {
+		/*
+		 * This was not a split block, so we passed the BP down,
+		 * and the checksum was handled by the (one) child zio.
+		 */
+		return;
+	}
+
+	zio_bad_cksum_t zbc;
+	int ret = zio_checksum_error(zio, &zbc);
+	if (ret == 0) {
+		zio_checksum_verified(zio);
+		return;
+	}
+
+	/*
+	 * The checksum didn't match.  Read all copies of all splits, and
+	 * then we will try to reconstruct.  The next time
+	 * vdev_indirect_io_done() is called, iv_reconstruct will be set.
+	 */
+	vdev_indirect_read_all(zio);
+
+	zio_vdev_io_redone(zio);
+}
+
 vdev_ops_t vdev_indirect_ops = {
 	vdev_indirect_open,
 	vdev_indirect_close,
@@ -1061,4 +1576,8 @@ MODULE_PARM_DESC(zfs_condense_min_mapping_bytes,
 module_param(zfs_condense_indirect_commit_entry_delay_ms, int, 0644);
 MODULE_PARM_DESC(zfs_condense_indirect_commit_entry_delay_ms,
 	"Delay while condensing vdev mapping");
+
+module_param(zfs_reconstruct_indirect_segments_max, int, 0644);
+MODULE_PARM_DESC(zfs_reconstruct_indirect_segments_max,
+	"Maximum number of split segments check all combinations");
 #endif
diff --git a/module/zfs/vdev_mirror.c b/module/zfs/vdev_mirror.c
index b56f955eb..4b01f317b 100644
--- a/module/zfs/vdev_mirror.c
+++ b/module/zfs/vdev_mirror.c
@@ -486,12 +486,15 @@ vdev_mirror_io_start(zio_t *zio)
 	mm = vdev_mirror_map_init(zio);
 
 	if (zio->io_type == ZIO_TYPE_READ) {
-		if ((zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
+		if (zio->io_bp != NULL &&
+		    (zio->io_flags & ZIO_FLAG_SCRUB) && !mm->mm_replacing) {
 			/*
-			 * For scrubbing reads we need to allocate a read
-			 * buffer for each child and issue reads to all
-			 * children.  If any child succeeds, it will copy its
-			 * data into zio->io_data in vdev_mirror_scrub_done.
+			 * For scrubbing reads (if we can verify the
+			 * checksum here, as indicated by io_bp being
+			 * non-NULL) we need to allocate a read buffer for
+			 * each child and issue reads to all children.  If
+			 * any child succeeds, it will copy its data into
+			 * zio->io_data in vdev_mirror_scrub_done.
 			 */
 			for (c = 0; c < mm->mm_children; c++) {
 				mc = &mm->mm_child[c];
@@ -640,7 +643,21 @@ vdev_mirror_io_done(zio_t *zio)
 			if (mc->mc_error == 0) {
 				if (mc->mc_tried)
 					continue;
+				/*
+				 * We didn't try this child.  We need to
+				 * repair it if:
+				 * 1. it's a scrub (in which case we have
+				 * tried everything that was healthy)
+				 *  - or -
+				 * 2. it's an indirect vdev (in which case
+				 * it could point to any other vdev, which
+				 * might have a bad DTL)
+				 *  - or -
+				 * 3. the DTL indicates that this data is
+				 * missing from this vdev
+				 */
 				if (!(zio->io_flags & ZIO_FLAG_SCRUB) &&
+				    mc->mc_vd->vdev_ops != &vdev_indirect_ops &&
 				    !vdev_dtl_contains(mc->mc_vd, DTL_PARTIAL,
 				    zio->io_txg, 1))
 					continue;
diff --git a/module/zfs/vdev_removal.c b/module/zfs/vdev_removal.c
index 6e81bf014..0fca8fb03 100644
--- a/module/zfs/vdev_removal.c
+++ b/module/zfs/vdev_removal.c
@@ -84,18 +84,12 @@ typedef struct vdev_copy_arg {
 	kmutex_t	vca_lock;
 } vdev_copy_arg_t;
 
-typedef struct vdev_copy_seg_arg {
-	vdev_copy_arg_t	*vcsa_copy_arg;
-	uint64_t	vcsa_txg;
-	dva_t		*vcsa_dest_dva;
-	blkptr_t	*vcsa_dest_bp;
-} vdev_copy_seg_arg_t;
-
 /*
- * The maximum amount of allowed data we're allowed to copy from a device
- * at a time when removing it.
+ * The maximum amount of memory we can use for outstanding i/o while
+ * doing a device removal.  This determines how much i/o we can have
+ * in flight concurrently.
  */
-int zfs_remove_max_copy_bytes = 8 * 1024 * 1024;
+int zfs_remove_max_copy_bytes = 64 * 1024 * 1024;
 
 /*
  * The largest contiguous segment that we will attempt to allocate when
@@ -165,7 +159,7 @@ spa_vdev_removal_create(vdev_t *vd)
 	mutex_init(&svr->svr_lock, NULL, MUTEX_DEFAULT, NULL);
 	cv_init(&svr->svr_cv, NULL, CV_DEFAULT, NULL);
 	svr->svr_allocd_segs = range_tree_create(NULL, NULL);
-	svr->svr_vdev = vd;
+	svr->svr_vdev_id = vd->vdev_id;
 
 	for (int i = 0; i < TXG_SIZE; i++) {
 		svr->svr_frees[i] = range_tree_create(NULL, NULL);
@@ -207,9 +201,10 @@ spa_vdev_removal_destroy(spa_vdev_removal_t *svr)
 static void
 vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
 {
-	vdev_t *vd = arg;
+	int vdev_id = (uintptr_t)arg;
+	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
+	vdev_t *vd = vdev_lookup_top(spa, vdev_id);
 	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
-	spa_t *spa = vd->vdev_spa;
 	objset_t *mos = spa->spa_dsl_pool->dp_meta_objset;
 	spa_vdev_removal_t *svr = NULL;
 	ASSERTV(uint64_t txg = dmu_tx_get_txg(tx));
@@ -331,7 +326,7 @@ vdev_remove_initiate_sync(void *arg, dmu_tx_t *tx)
 	ASSERT3P(spa->spa_vdev_removal, ==, NULL);
 	spa->spa_vdev_removal = svr;
 	svr->svr_thread = thread_create(NULL, 0,
-	    spa_vdev_remove_thread, vd, 0, &p0, TS_RUN, minclsyspri);
+	    spa_vdev_remove_thread, spa, 0, &p0, TS_RUN, minclsyspri);
 }
 
 /*
@@ -372,21 +367,24 @@ spa_remove_init(spa_t *spa)
 		spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
 		vdev_t *vd = vdev_lookup_top(spa,
 		    spa->spa_removing_phys.sr_removing_vdev);
-		spa_config_exit(spa, SCL_STATE, FTAG);
 
-		if (vd == NULL)
+		if (vd == NULL) {
+			spa_config_exit(spa, SCL_STATE, FTAG);
 			return (EINVAL);
+		}
 
 		vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
 
 		ASSERT(vdev_is_concrete(vd));
 		spa_vdev_removal_t *svr = spa_vdev_removal_create(vd);
-		ASSERT(svr->svr_vdev->vdev_removing);
+		ASSERT3U(svr->svr_vdev_id, ==, vd->vdev_id);
+		ASSERT(vd->vdev_removing);
 
 		vd->vdev_indirect_mapping = vdev_indirect_mapping_open(
 		    spa->spa_meta_objset, vic->vic_mapping_object);
 		vd->vdev_indirect_births = vdev_indirect_births_open(
 		    spa->spa_meta_objset, vic->vic_births_object);
+		spa_config_exit(spa, SCL_STATE, FTAG);
 
 		spa->spa_vdev_removal = svr;
 	}
@@ -439,15 +437,8 @@ spa_restart_removal(spa_t *spa)
 	if (!spa_writeable(spa))
 		return;
 
-	vdev_t *vd = svr->svr_vdev;
-	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
-
-	ASSERT3P(vd, !=, NULL);
-	ASSERT(vd->vdev_removing);
-
-	zfs_dbgmsg("restarting removal of %llu at count=%llu",
-	    vd->vdev_id, vdev_indirect_mapping_num_entries(vim));
-	svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, vd,
+	zfs_dbgmsg("restarting removal of %llu", svr->svr_vdev_id);
+	svr->svr_thread = thread_create(NULL, 0, spa_vdev_remove_thread, spa,
 	    0, &p0, TS_RUN, minclsyspri);
 }
 
@@ -468,7 +459,7 @@ free_from_removing_vdev(vdev_t *vd, uint64_t offset, uint64_t size,
 	ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
 	ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, ==,
 	    vdev_indirect_mapping_object(vim));
-	ASSERT3P(vd, ==, svr->svr_vdev);
+	ASSERT3U(vd->vdev_id, ==, svr->svr_vdev_id);
 	ASSERT3U(spa_syncing_txg(spa), ==, txg);
 
 	mutex_enter(&svr->svr_lock);
@@ -646,7 +637,7 @@ spa_finish_removal(spa_t *spa, dsl_scan_state_t state, dmu_tx_t *tx)
 
 	if (state == DSS_FINISHED) {
 		spa_removing_phys_t *srp = &spa->spa_removing_phys;
-		vdev_t *vd = svr->svr_vdev;
+		vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
 		vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
 
 		if (srp->sr_prev_indirect_vdev != UINT64_MAX) {
@@ -690,7 +681,7 @@ vdev_mapping_sync(void *arg, dmu_tx_t *tx)
 {
 	spa_vdev_removal_t *svr = arg;
 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
-	vdev_t *vd = svr->svr_vdev;
+	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
 	ASSERTV(vdev_indirect_config_t *vic = &vd->vdev_indirect_config);
 	uint64_t txg = dmu_tx_get_txg(tx);
 	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
@@ -718,64 +709,128 @@ vdev_mapping_sync(void *arg, dmu_tx_t *tx)
 	spa_sync_removing_state(spa, tx);
 }
 
+/*
+ * All reads and writes associated with a call to spa_vdev_copy_segment()
+ * are done.
+ */
+static void
+spa_vdev_copy_nullzio_done(zio_t *zio)
+{
+	spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa);
+}
+
+/*
+ * The write of the new location is done.
+ */
 static void
 spa_vdev_copy_segment_write_done(zio_t *zio)
 {
-	vdev_copy_seg_arg_t *vcsa = zio->io_private;
-	vdev_copy_arg_t *vca = vcsa->vcsa_copy_arg;
-	spa_config_exit(zio->io_spa, SCL_STATE, FTAG);
+	vdev_copy_arg_t *vca = zio->io_private;
+
 	abd_free(zio->io_abd);
 
 	mutex_enter(&vca->vca_lock);
 	vca->vca_outstanding_bytes -= zio->io_size;
 	cv_signal(&vca->vca_cv);
 	mutex_exit(&vca->vca_lock);
-
-	ASSERT0(zio->io_error);
-	kmem_free(vcsa->vcsa_dest_bp, sizeof (blkptr_t));
-	kmem_free(vcsa, sizeof (vdev_copy_seg_arg_t));
 }
 
+/*
+ * The read of the old location is done.  The parent zio is the write to
+ * the new location.  Allow it to start.
+ */
 static void
 spa_vdev_copy_segment_read_done(zio_t *zio)
 {
-	vdev_copy_seg_arg_t *vcsa = zio->io_private;
-	dva_t *dest_dva = vcsa->vcsa_dest_dva;
-	uint64_t txg = vcsa->vcsa_txg;
-	spa_t *spa = zio->io_spa;
-	ASSERTV(vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(dest_dva)));
-	blkptr_t *bp = NULL;
-	dva_t *dva = NULL;
-	uint64_t size = zio->io_size;
-
-	ASSERT3P(dest_vd, !=, NULL);
-	ASSERT0(zio->io_error);
-
-	vcsa->vcsa_dest_bp = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
-	bp = vcsa->vcsa_dest_bp;
-	dva = bp->blk_dva;
-
-	BP_ZERO(bp);
-
-	/* initialize with dest_dva */
-	bcopy(dest_dva, dva, sizeof (dva_t));
-	BP_SET_BIRTH(bp, TXG_INITIAL, TXG_INITIAL);
-
-	BP_SET_LSIZE(bp, size);
-	BP_SET_PSIZE(bp, size);
-	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
-	BP_SET_CHECKSUM(bp, ZIO_CHECKSUM_OFF);
-	BP_SET_TYPE(bp, DMU_OT_NONE);
-	BP_SET_LEVEL(bp, 0);
-	BP_SET_DEDUP(bp, 0);
-	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
-
-	zio_nowait(zio_rewrite(spa->spa_txg_zio[txg & TXG_MASK], spa,
-	    txg, bp, zio->io_abd, size,
-	    spa_vdev_copy_segment_write_done, vcsa,
-	    ZIO_PRIORITY_REMOVAL, 0, NULL));
+	zio_nowait(zio_unique_parent(zio));
+}
+
+/*
+ * If the old and new vdevs are mirrors, we will read both sides of the old
+ * mirror, and write each copy to the corresponding side of the new mirror.
+ * If the old and new vdevs have a different number of children, we will do
+ * this as best as possible.  Since we aren't verifying checksums, this
+ * ensures that as long as there's a good copy of the data, we'll have a
+ * good copy after the removal, even if there's silent damage to one side
+ * of the mirror. If we're removing a mirror that has some silent damage,
+ * we'll have exactly the same damage in the new location (assuming that
+ * the new location is also a mirror).
+ *
+ * We accomplish this by creating a tree of zio_t's, with as many writes as
+ * there are "children" of the new vdev (a non-redundant vdev counts as one
+ * child, a 2-way mirror has 2 children, etc). Each write has an associated
+ * read from a child of the old vdev. Typically there will be the same
+ * number of children of the old and new vdevs.  However, if there are more
+ * children of the new vdev, some child(ren) of the old vdev will be issued
+ * multiple reads.  If there are more children of the old vdev, some copies
+ * will be dropped.
+ *
+ * For example, the tree of zio_t's for a 2-way mirror is:
+ *
+ *                            null
+ *                           /    \
+ *    write(new vdev, child 0)      write(new vdev, child 1)
+ *      |                             |
+ *    read(old vdev, child 0)       read(old vdev, child 1)
+ *
+ * Child zio's complete before their parents complete.  However, zio's
+ * created with zio_vdev_child_io() may be issued before their children
+ * complete.  In this case we need to make sure that the children (reads)
+ * complete before the parents (writes) are *issued*.  We do this by not
+ * calling zio_nowait() on each write until its corresponding read has
+ * completed.
+ *
+ * The spa_config_lock must be held while zio's created by
+ * zio_vdev_child_io() are in progress, to ensure that the vdev tree does
+ * not change (e.g. due to a concurrent "zpool attach/detach"). The "null"
+ * zio is needed to release the spa_config_lock after all the reads and
+ * writes complete. (Note that we can't grab the config lock for each read,
+ * because it is not reentrant - we could deadlock with a thread waiting
+ * for a write lock.)
+ */
+static void
+spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio,
+    vdev_t *source_vd, uint64_t source_offset,
+    vdev_t *dest_child_vd, uint64_t dest_offset, int dest_id, uint64_t size)
+{
+	ASSERT3U(spa_config_held(nzio->io_spa, SCL_ALL, RW_READER), !=, 0);
+
+	mutex_enter(&vca->vca_lock);
+	vca->vca_outstanding_bytes += size;
+	mutex_exit(&vca->vca_lock);
+
+	abd_t *abd = abd_alloc_for_io(size, B_FALSE);
+
+	vdev_t *source_child_vd;
+	if (source_vd->vdev_ops == &vdev_mirror_ops && dest_id != -1) {
+		/*
+		 * Source and dest are both mirrors.  Copy from the same
+		 * child id as we are copying to (wrapping around if there
+		 * are more dest children than source children).
+		 */
+		source_child_vd =
+		    source_vd->vdev_child[dest_id % source_vd->vdev_children];
+	} else {
+		source_child_vd = source_vd;
+	}
+
+	zio_t *write_zio = zio_vdev_child_io(nzio, NULL,
+	    dest_child_vd, dest_offset, abd, size,
+	    ZIO_TYPE_WRITE, ZIO_PRIORITY_REMOVAL,
+	    ZIO_FLAG_CANFAIL,
+	    spa_vdev_copy_segment_write_done, vca);
+
+	zio_nowait(zio_vdev_child_io(write_zio, NULL,
+	    source_child_vd, source_offset, abd, size,
+	    ZIO_TYPE_READ, ZIO_PRIORITY_REMOVAL,
+	    ZIO_FLAG_CANFAIL,
+	    spa_vdev_copy_segment_read_done, vca));
 }
 
+/*
+ * Allocate a new location for this segment, and create the zio_t's to
+ * read from the old location and write to the new location.
+ */
 static int
 spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
     vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
@@ -784,10 +839,7 @@ spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
 	spa_t *spa = vd->vdev_spa;
 	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
 	vdev_indirect_mapping_entry_t *entry;
-	vdev_copy_seg_arg_t *private;
 	dva_t dst = {{ 0 }};
-	blkptr_t blk, *bp = &blk;
-	dva_t *dva = bp->blk_dva;
 
 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
 
@@ -804,51 +856,28 @@ spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
 	 */
 	ASSERT3U(DVA_GET_ASIZE(&dst), ==, size);
 
-	mutex_enter(&vca->vca_lock);
-	vca->vca_outstanding_bytes += size;
-	mutex_exit(&vca->vca_lock);
-
 	entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
 	DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
 	entry->vime_mapping.vimep_dst = dst;
 
-	private = kmem_alloc(sizeof (vdev_copy_seg_arg_t), KM_SLEEP);
-	private->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
-	private->vcsa_txg = txg;
-	private->vcsa_copy_arg = vca;
-
 	/*
-	 * This lock is eventually released by the donefunc for the
-	 * zio_write_phys that finishes copying the data.
+	 * See comment before spa_vdev_copy_one_child().
 	 */
-	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
-
-	/*
-	 * Do logical I/O, letting the redundancy vdevs (like mirror)
-	 * handle their own I/O instead of duplicating that code here.
-	 */
-	BP_ZERO(bp);
-
-	DVA_SET_VDEV(&dva[0], vd->vdev_id);
-	DVA_SET_OFFSET(&dva[0], start);
-	DVA_SET_GANG(&dva[0], 0);
-	DVA_SET_ASIZE(&dva[0], vdev_psize_to_asize(vd, size));
-
-	BP_SET_BIRTH(bp, TXG_INITIAL, TXG_INITIAL);
-
-	BP_SET_LSIZE(bp, size);
-	BP_SET_PSIZE(bp, size);
-	BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF);
-	BP_SET_CHECKSUM(bp, ZIO_CHECKSUM_OFF);
-	BP_SET_TYPE(bp, DMU_OT_NONE);
-	BP_SET_LEVEL(bp, 0);
-	BP_SET_DEDUP(bp, 0);
-	BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
-
-	zio_nowait(zio_read(spa->spa_txg_zio[txg & TXG_MASK], spa,
-	    bp, abd_alloc_for_io(size, B_FALSE), size,
-	    spa_vdev_copy_segment_read_done, private,
-	    ZIO_PRIORITY_REMOVAL, 0, NULL));
+	spa_config_enter(spa, SCL_STATE, spa, RW_READER);
+	zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL,
+	    spa_vdev_copy_nullzio_done, NULL, 0);
+	vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst));
+	if (dest_vd->vdev_ops == &vdev_mirror_ops) {
+		for (int i = 0; i < dest_vd->vdev_children; i++) {
+			vdev_t *child = dest_vd->vdev_child[i];
+			spa_vdev_copy_one_child(vca, nzio, vd, start,
+			    child, DVA_GET_OFFSET(&dst), i, size);
+		}
+	} else {
+		spa_vdev_copy_one_child(vca, nzio, vd, start,
+		    dest_vd, DVA_GET_OFFSET(&dst), -1, size);
+	}
+	zio_nowait(nzio);
 
 	list_insert_tail(&svr->svr_new_segments[txg & TXG_MASK], entry);
 	ASSERT3U(start + size, <=, vd->vdev_ms_count << vd->vdev_ms_shift);
@@ -866,8 +895,8 @@ static void
 vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
 {
 	spa_vdev_removal_t *svr = arg;
-	vdev_t *vd = svr->svr_vdev;
-	spa_t *spa = vd->vdev_spa;
+	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
+	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
 
 	ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
 
@@ -896,37 +925,6 @@ vdev_remove_complete_sync(void *arg, dmu_tx_t *tx)
 }
 
 static void
-vdev_indirect_state_transfer(vdev_t *ivd, vdev_t *vd)
-{
-	ivd->vdev_indirect_config = vd->vdev_indirect_config;
-
-	ASSERT3P(ivd->vdev_indirect_mapping, ==, NULL);
-	ASSERT(vd->vdev_indirect_mapping != NULL);
-	ivd->vdev_indirect_mapping = vd->vdev_indirect_mapping;
-	vd->vdev_indirect_mapping = NULL;
-
-	ASSERT3P(ivd->vdev_indirect_births, ==, NULL);
-	ASSERT(vd->vdev_indirect_births != NULL);
-	ivd->vdev_indirect_births = vd->vdev_indirect_births;
-	vd->vdev_indirect_births = NULL;
-
-	ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
-	ASSERT0(range_tree_space(ivd->vdev_obsolete_segments));
-
-	if (vd->vdev_obsolete_sm != NULL) {
-		ASSERT3U(ivd->vdev_asize, ==, vd->vdev_asize);
-
-		/*
-		 * We cannot use space_map_{open,close} because we hold all
-		 * the config locks as writer.
-		 */
-		ASSERT3P(ivd->vdev_obsolete_sm, ==, NULL);
-		ivd->vdev_obsolete_sm = vd->vdev_obsolete_sm;
-		vd->vdev_obsolete_sm = NULL;
-	}
-}
-
-static void
 vdev_remove_enlist_zaps(vdev_t *vd, nvlist_t *zlist)
 {
 	ASSERT3P(zlist, !=, NULL);
@@ -961,17 +959,13 @@ vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
 	vdev_remove_enlist_zaps(vd, svr->svr_zaplist);
 
 	ivd = vdev_add_parent(vd, &vdev_indirect_ops);
+	ivd->vdev_removing = 0;
 
 	vd->vdev_leaf_zap = 0;
 
 	vdev_remove_child(ivd, vd);
 	vdev_compact_children(ivd);
 
-	vdev_indirect_state_transfer(ivd, vd);
-
-	svr->svr_vdev = ivd;
-
-	ASSERT(!ivd->vdev_removing);
 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
 
 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
@@ -994,9 +988,8 @@ vdev_remove_replace_with_indirect(vdev_t *vd, uint64_t txg)
  * context by the removal thread after we have copied all vdev's data.
  */
 static void
-vdev_remove_complete(vdev_t *vd)
+vdev_remove_complete(spa_t *spa)
 {
-	spa_t *spa = vd->vdev_spa;
 	uint64_t txg;
 
 	/*
@@ -1004,8 +997,12 @@ vdev_remove_complete(vdev_t *vd)
 	 * vdev_metaslab_fini()
 	 */
 	txg_wait_synced(spa->spa_dsl_pool, 0);
-
 	txg = spa_vdev_enter(spa);
+	vdev_t *vd = vdev_lookup_top(spa, spa->spa_vdev_removal->svr_vdev_id);
+
+	sysevent_t *ev = spa_event_create(spa, vd, NULL,
+	    ESC_ZFS_VDEV_REMOVE_DEV);
+
 	zfs_dbgmsg("finishing device removal for vdev %llu in txg %llu",
 	    vd->vdev_id, txg);
 
@@ -1025,6 +1022,10 @@ vdev_remove_complete(vdev_t *vd)
 	/*
 	 * We now release the locks, allowing spa_sync to run and finish the
 	 * removal via vdev_remove_complete_sync in syncing context.
+	 *
+	 * Note that we hold on to the vdev_t that has been replaced.  Since
+	 * it isn't part of the vdev tree any longer, it can't be concurrently
+	 * manipulated, even while we don't have the config lock.
 	 */
 	(void) spa_vdev_exit(spa, NULL, txg, 0);
 
@@ -1046,6 +1047,9 @@ vdev_remove_complete(vdev_t *vd)
 	 */
 	vdev_config_dirty(spa->spa_root_vdev);
 	(void) spa_vdev_exit(spa, vd, txg, 0);
+
+	if (ev != NULL)
+		spa_event_post(ev);
 }
 
 /*
@@ -1056,7 +1060,7 @@ vdev_remove_complete(vdev_t *vd)
  * this size again this txg.
  */
 static void
-spa_vdev_copy_impl(spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
+spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
     uint64_t *max_alloc, dmu_tx_t *tx)
 {
 	uint64_t txg = dmu_tx_get_txg(tx);
@@ -1095,7 +1099,7 @@ spa_vdev_copy_impl(spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
 	while (length > 0) {
 		uint64_t mylen = MIN(length, thismax);
 
-		int error = spa_vdev_copy_segment(svr->svr_vdev,
+		int error = spa_vdev_copy_segment(vd,
 		    offset, mylen, txg, vca, &zal);
 
 		if (error == ENOSPC) {
@@ -1153,12 +1157,14 @@ spa_vdev_copy_impl(spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
 static void
 spa_vdev_remove_thread(void *arg)
 {
-	vdev_t *vd = arg;
-	spa_t *spa = vd->vdev_spa;
+	spa_t *spa = arg;
 	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
 	vdev_copy_arg_t vca;
 	uint64_t max_alloc = zfs_remove_max_segment;
 	uint64_t last_txg = 0;
+
+	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
+	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
 	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
 	uint64_t start_offset = vdev_indirect_mapping_max_offset(vim);
 
@@ -1166,7 +1172,6 @@ spa_vdev_remove_thread(void *arg)
 	ASSERT(vdev_is_concrete(vd));
 	ASSERT(vd->vdev_removing);
 	ASSERT(vd->vdev_indirect_config.vic_mapping_object != 0);
-	ASSERT3P(svr->svr_vdev, ==, vd);
 	ASSERT(vim != NULL);
 
 	mutex_init(&vca.vca_lock, NULL, MUTEX_DEFAULT, NULL);
@@ -1247,6 +1252,17 @@ spa_vdev_remove_thread(void *arg)
 
 			mutex_exit(&svr->svr_lock);
 
+			/*
+			 * We need to periodically drop the config lock so that
+			 * writers can get in.  Additionally, we can't wait
+			 * for a txg to sync while holding a config lock
+			 * (since a waiting writer could cause a 3-way deadlock
+			 * with the sync thread, which also gets a config
+			 * lock for reader).  So we can't hold the config lock
+			 * while calling dmu_tx_assign().
+			 */
+			spa_config_exit(spa, SCL_CONFIG, FTAG);
+
 			mutex_enter(&vca.vca_lock);
 			while (vca.vca_outstanding_bytes >
 			    zfs_remove_max_copy_bytes) {
@@ -1260,11 +1276,19 @@ spa_vdev_remove_thread(void *arg)
 			VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
 			uint64_t txg = dmu_tx_get_txg(tx);
 
+			/*
+			 * Reacquire the vdev_config lock.  The vdev_t
+			 * that we're removing may have changed, e.g. due
+			 * to a vdev_attach or vdev_detach.
+			 */
+			spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
+			vd = vdev_lookup_top(spa, svr->svr_vdev_id);
+
 			if (txg != last_txg)
 				max_alloc = zfs_remove_max_segment;
 			last_txg = txg;
 
-			spa_vdev_copy_impl(svr, &vca, &max_alloc, tx);
+			spa_vdev_copy_impl(vd, svr, &vca, &max_alloc, tx);
 
 			dmu_tx_commit(tx);
 			mutex_enter(&svr->svr_lock);
@@ -1272,6 +1296,9 @@ spa_vdev_remove_thread(void *arg)
 	}
 
 	mutex_exit(&svr->svr_lock);
+
+	spa_config_exit(spa, SCL_CONFIG, FTAG);
+
 	/*
 	 * Wait for all copies to finish before cleaning up the vca.
 	 */
@@ -1289,7 +1316,7 @@ spa_vdev_remove_thread(void *arg)
 		mutex_exit(&svr->svr_lock);
 	} else {
 		ASSERT0(range_tree_space(svr->svr_allocd_segs));
-		vdev_remove_complete(vd);
+		vdev_remove_complete(spa);
 	}
 }
 
@@ -1330,7 +1357,7 @@ spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
 {
 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
 	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
-	vdev_t *vd = svr->svr_vdev;
+	vdev_t *vd = vdev_lookup_top(spa, svr->svr_vdev_id);
 	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
 	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
 	objset_t *mos = spa->spa_meta_objset;
@@ -1403,8 +1430,11 @@ spa_vdev_remove_cancel_sync(void *arg, dmu_tx_t *tx)
 			 * because we have not allocated mappings for it yet.
 			 */
 			uint64_t syncd = vdev_indirect_mapping_max_offset(vim);
-			range_tree_clear(svr->svr_allocd_segs, syncd,
-			    msp->ms_sm->sm_start + msp->ms_sm->sm_size - syncd);
+			uint64_t sm_end = msp->ms_sm->sm_start +
+			    msp->ms_sm->sm_size;
+			if (sm_end > syncd)
+				range_tree_clear(svr->svr_allocd_segs,
+				    syncd, sm_end - syncd);
 
 			mutex_exit(&svr->svr_lock);
 		}
@@ -1465,7 +1495,7 @@ spa_vdev_remove_cancel(spa_t *spa)
 	if (spa->spa_vdev_removal == NULL)
 		return (ENOTACTIVE);
 
-	uint64_t vdid = spa->spa_vdev_removal->svr_vdev->vdev_id;
+	uint64_t vdid = spa->spa_vdev_removal->svr_vdev_id;
 
 	int error = dsl_sync_task(spa->spa_name, spa_vdev_remove_cancel_check,
 	    spa_vdev_remove_cancel_sync, NULL, 0, ZFS_SPACE_CHECK_NONE);
@@ -1774,7 +1804,7 @@ spa_vdev_remove_top(vdev_t *vd, uint64_t *txg)
 	dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, *txg);
 	dsl_sync_task_nowait(spa->spa_dsl_pool,
 	    vdev_remove_initiate_sync,
-	    vd, 0, ZFS_SPACE_CHECK_NONE, tx);
+	    (void *)(uintptr_t)vd->vdev_id, 0, ZFS_SPACE_CHECK_NONE, tx);
 	dmu_tx_commit(tx);
 
 	return (0);
diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 5259a0c6f..9bb2459bb 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -1212,17 +1212,6 @@ zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
 	ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) ||
 	    done != NULL);
 
-	/*
-	 * In the common case, where the parent zio was to a normal vdev,
-	 * the child zio must be to a child vdev of that vdev.  Otherwise,
-	 * the child zio must be to a top-level vdev.
-	 */
-	if (pio->io_vd != NULL && pio->io_vd->vdev_ops != &vdev_indirect_ops) {
-		ASSERT3P(vd->vdev_parent, ==, pio->io_vd);
-	} else {
-		ASSERT3P(vd, ==, vd->vdev_top);
-	}
-
 	if (type == ZIO_TYPE_READ && bp != NULL) {
 		/*
 		 * If we have the bp, then the child should perform the
@@ -1283,7 +1272,7 @@ zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
 
 zio_t *
 zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size,
-    int type, zio_priority_t priority, enum zio_flag flags,
+    zio_type_t type, zio_priority_t priority, enum zio_flag flags,
     zio_done_func_t *done, void *private)
 {
 	zio_t *zio;
@@ -3481,7 +3470,7 @@ zio_vdev_io_start(zio_t *zio)
 		 */
 		ASSERT(zio->io_flags &
 		    (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL |
-		    ZIO_FLAG_INDUCE_DAMAGE));
+		    ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE));
 	}
 
 	align = 1ULL << vd->vdev_top->vdev_ashift;
@@ -3521,18 +3510,37 @@ zio_vdev_io_start(zio_t *zio)
 	 * If this is a repair I/O, and there's no self-healing involved --
 	 * that is, we're just resilvering what we expect to resilver --
 	 * then don't do the I/O unless zio's txg is actually in vd's DTL.
-	 * This prevents spurious resilvering with nested replication.
-	 * For example, given a mirror of mirrors, (A+B)+(C+D), if only
-	 * A is out of date, we'll read from C+D, then use the data to
-	 * resilver A+B -- but we don't actually want to resilver B, just A.
-	 * The top-level mirror has no way to know this, so instead we just
-	 * discard unnecessary repairs as we work our way down the vdev tree.
-	 * The same logic applies to any form of nested replication:
-	 * ditto + mirror, RAID-Z + replacing, etc.  This covers them all.
+	 * This prevents spurious resilvering.
+	 *
+	 * There are a few ways that we can end up creating these spurious
+	 * resilver i/os:
+	 *
+	 * 1. A resilver i/o will be issued if any DVA in the BP has a
+	 * dirty DTL.  The mirror code will issue resilver writes to
+	 * each DVA, including the one(s) that are not on vdevs with dirty
+	 * DTLs.
+	 *
+	 * 2. With nested replication, which happens when we have a
+	 * "replacing" or "spare" vdev that's a child of a mirror or raidz.
+	 * For example, given mirror(replacing(A+B), C), it's likely that
+	 * only A is out of date (it's the new device). In this case, we'll
+	 * read from C, then use the data to resilver A+B -- but we don't
+	 * actually want to resilver B, just A. The top-level mirror has no
+	 * way to know this, so instead we just discard unnecessary repairs
+	 * as we work our way down the vdev tree.
+	 *
+	 * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc.
+	 * The same logic applies to any form of nested replication: ditto
+	 * + mirror, RAID-Z + replacing, etc.
+	 *
+	 * However, indirect vdevs point off to other vdevs which may have
+	 * DTL's, so we never bypass them.  The child i/os on concrete vdevs
+	 * will be properly bypassed instead.
 	 */
 	if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
 	    !(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
 	    zio->io_txg != 0 &&	/* not a delegated i/o */
+	    vd->vdev_ops != &vdev_indirect_ops &&
 	    !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
 		ASSERT(zio->io_type == ZIO_TYPE_WRITE);
 		zio_vdev_io_bypass(zio);