Re: xfstests - SCRATCH_DIR mounted/unmounted twice when testing nfs

[Boris Ranto <branto@redhat.com>]

On Thu, 2011-05-19 at 10:02 +1000, Dave Chinner wrote:
> On Wed, May 18, 2011 at 12:41:38PM +0000, Boris Ranto wrote:
> > Tests 124, 127 and 128 unmount their SCRATCH_DEV manually while using _cleanup_testdir in trapped cleanup function.
> > This can lead to test fails due to double unmount on nfs where _cleanup_testdir unmounts SCRATCH_DEV.
> > 
> > Tests 129 and 130 use _setup_testdir and _scratch_mount that can lead to double mount on nfs where _setup_testdir mounts SCRATCH_DEV.
> > 
> > The least invasive patch (only nfs shall be affected by this patch) I could come up with that fixed this issue used conditional umounts in _cleanup_testdir and conditional mount/remount in _scratch_mount (remount was used so that mount flags do not get lost).
> > 
> From the code:
> 
> #
> # Warning for UDF and NFS:
> # this function calls _setup_udf_scratchdir and _setup_udf_scratchdir
> # which actually uses the scratch dir for the test dir.
> #
> # This was done because testdir was intended to be a persistent
> # XFS only partition.  This should eventually change, and treat
> # at least local filesystems all the same.
> #
> 
> I think that this means that the way _setup_testdir uses the scratch
> device was a nasty hack and was intended to be fixed at some point.
> Perhaps we should actually understand the issue that lead to this
> hack first, and then determine what is needed to fix it rather than
> just layering more hacks on top of it?
> 
> > Alternatively, _setup_testdir could stop mounting SCRATCH_DEV but
> > that would be probably too invasive.
> 
> I don't really care about how invasive fixing the problem
> properly is if it's the better long term solution....

I also prefer that solution. The fix could be relatively simple: just
treat nfs as regular local filesystem. The question is whether there are
any test cases that depends on the old _setup_testdir behaviour.

Looking through the tests I found two tests that have nfs in their
_supported_fs field: 062 and 192.

062 expects scratch dev to be regular device and hence fails for nfs
either way. Therefore restricting the test to xfs (maybe +udf) shouldn't
cause any problems.

192 uses TEST_DIR directly so it probably expects TEST_DIR to be mounted
xfs filesystem, therefore just changing _supported_fs to xfs (maybe
+udf) should be sufficient.


> Cheers,
> 
> Dave.


So far I could test the following patch (nfs only, udf unchanged) on few
tests with success. I'll let it run for all the tests and report back
later if it still makes sense.

diff --git a/062 b/062
index 5cb6f92..83644be 100755
--- a/062
+++ b/062
@@ -71,7 +71,7 @@ _create_test_bed()
 }
 
 # real QA test starts here
-_supported_fs xfs udf nfs
+_supported_fs xfs
 _supported_os Linux
 
 _require_scratch
 
diff --git a/192 b/192
index d8301d5..1e582d1 100755
--- a/192
+++ b/192
@@ -45,7 +45,7 @@ _access_time()
 
 # real QA test starts here
 
-_supported_fs xfs udf nfs
+_supported_fs xfs
 _supported_os Linux
 #delay=150
 #delay=75
diff --git a/common.rc b/common.rc
index e634fbb..3d1a17b 100644
--- a/common.rc
+++ b/common.rc
@@ -302,7 +302,10 @@ _scratch_mkfs()
         _scratch_mkfs_xfs $*
 	;;
     nfs*)
-	# do nothing for nfs
+       # clean scratch dir
+       _scratch_mount
+       rm -rf $SCRATCH_MNT/* $SCRATCH_MNT/.*
+       _scratch_unmount
 	;;
     udf)
         $MKFS_UDF_PROG $MKFS_OPTIONS $* $SCRATCH_DEV > /dev/null
@@ -703,6 +706,10 @@ _require_scratch()
 		 then
 		     _notrun "this test requires a valid \$SCRATCH_DEV"
 		 fi
+		 if [ ! -d "$SCRATCH_MNT" ]
+		 then
+		     _notrun "this test requires a valid \$SCRATCH_MNT"
+		 fi
 		 ;;
 	*)
 		 if [ -z "$SCRATCH_DEV" -o "`_is_block_dev $SCRATCH_DEV`" = "" ]
@@ -1320,35 +1327,9 @@ _setup_udf_scratchdir()
     testdir=$SCRATCH_MNT
 }
 
-_setup_nfs_scratchdir()
-{
-    [ "$FSTYP" != "nfs" ] \
-	&& _fail "setup_nfs_testdir: \$FSTYP is not nfs"
-    [ -z "$SCRATCH_DEV" ] \
-	&& _notrun "this test requires a valid host fs for \$SCRATCH_DEV"
-    [ -z "$SCRATCH_MNT" ] \
-	&& _notrun "this test requires a valid \$SCRATCH_MNT"
-
-    # mounted?
-    if _mount | grep -q $SCRATCH_DEV
-    then
-        # if it's mounted, make sure its on $SCRATCH_MNT
-        if ! _mount | grep $SCRATCH_DEV | grep -q $SCRATCH_MNT
-        then
-            _fail "\$SCRATCH_DEV is mounted but not on \$SCRATCH_MNT -
aborting"
-        fi
-    	$UMOUNT_PROG $SCRATCH_DEV
-    fi
-
-    _scratch_mkfs
-    _scratch_mount
-
-    testdir=$SCRATCH_MNT
-}
-
 #
-# Warning for UDF and NFS:
-# this function calls _setup_udf_scratchdir and _setup_udf_scratchdir
+# Warning for UDF:
+# this function calls _setup_udf_scratchdir
 # which actually uses the scratch dir for the test dir.
 #
 # This was done because testdir was intended to be a persistent
@@ -1361,9 +1342,6 @@ _setup_testdir()
     udf)
 	_setup_udf_scratchdir
 	;;
-    nfs*)
-	_setup_nfs_scratchdir
-	;;
     *)
 	testdir=$TEST_DIR
 	;;
@@ -1377,10 +1355,6 @@ _cleanup_testdir()
 	# umount testdir as it is $SCRATCH_MNT which could be used by xfs next
 	[ -n "$testdir" ] && $UMOUNT_PROG $testdir
 	;;
-    nfs*)
-	# umount testdir as it is $SCRATCH_MNT which could be used by xfs next
-	[ -n "$testdir" ] && $UMOUNT_PROG $testdir
-	;;
     *)
 	# do nothing, testdir is $TEST_DIR
 	:

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs