[1/3] generic/081: fix lvm config not being cleaned up properly

Commit Message

Darrick J. Wong Aug. 15, 2019, 3:18 p.m. UTC

From: Darrick J. Wong <[e-mail]>

Fix a race between _cleanup and dmeventd that causes the lvm
configuration not to be cleaned up and subsequent tests to fail.

Signed-off-by: Darrick J. Wong <[e-mail]>
---
 tests/generic/081 |   29 +++++++++++++++++++++++------
 1 file changed, 23 insertions(+), 6 deletions(-)

Patch

diff --git a/tests/generic/081 b/tests/generic/081
index 10f4a186..e8f4f5b5 100755
--- a/tests/generic/081
+++ b/tests/generic/081
@@ -19,12 +19,29 @@  _cleanup()
 {
 	cd /
 	rm -f $tmp.*
-	# lvm may have umounted it on I/O error, but in case it does not
-	# wait a bit for lvm to settle, sigh..
-	sleep 2
-	$UMOUNT_PROG $mnt >/dev/null 2>&1
-	$LVM_PROG vgremove -f $vgname >>$seqres.full 2>&1
-	$LVM_PROG pvremove -f $SCRATCH_DEV >>$seqres.full 2>&1
+
+	# Tear down the lvm vg and snapshot.
+	#
+	# NOTE: We do the unmount and {vg,pv}remove in a loop here because
+	# dmeventd could be configured to unmount the filesystem automatically
+	# after the IO errors.  That is racy with the umount we're trying to do
+	# here because there's a window in which the directory tree has been
+	# removed from the mount namespaces (so the umount call here sees no
+	# mount and exits) but the filesystem hasn't yet released the block
+	# device, which causes the vgremove here to fail.
+	#
+	# We "solve" the race by repeating the umount/lvm teardown until the
+	# block device goes away, because we cannot exit this test without
+	# removing the lvm devices from the scratch device -- this will cause
+	# other tests to fail.
+	while test -e /dev/mapper/$vgname-$snapname || \
+	      test -e /dev/mapper/$vgname-$lvname; do
+		$UMOUNT_PROG $mnt >> $seqres.full 2>&1
+		$LVM_PROG vgremove -f $vgname >>$seqres.full 2>&1
+		$LVM_PROG pvremove -f $SCRATCH_DEV >>$seqres.full 2>&1
+		test $? -eq 0 && break
+		sleep 2
+	done
 }
 
 # get standard environment, filters and checks