Re: xfstests: standard way of handling loop devices

[Tomas Racek <tracek@redhat.com>]

> On Thu, Aug 09, 2012 at 04:30:21AM -0400, Tomas Racek wrote:
> > Hi,
> > 
> > I am currently working on tests that check FITRIM implementation
> > (251, 260 and one new I'm writing now) and I want to use loopback
> > device as fallback if $SCRATCH_DEV doesn't support discard. Has
> > anybody been working on some xfstests' standard way of
> > creating/destroying loop devices?
> > 
> > I could do with something as simple as this (in common.rc):
> 
> Probably a good idea given the random failures we get with loopback
> device unmounting due to the racy unmount-based automatic device
> destruction.
> 
> > 
> > _create_loop_device()
> > {
> >         size=${1}
> >         dev=`losetup -f`
> >         file="$TEST_DIR/$(basename $dev).fs"
> 
> That won't work - we create loop devices with files on the scratch
> device, too, and some tests create more than one. This is also racy

I've missed that...

> in that two threads could both get then same next free loopback
> device, but I'm not sure we care about that case very much.
> 
> >         truncate -s $size $file || _fail "Cannot create image file
> >         $file"
> 
> It's better to use xfs_io that introduce new external tool
> dependencies.

OK.

> 
> >         losetup $dev $file || _fail "Cannot associate $file with
> >         $dev"
> >         echo $dev
> > }
> > 
> > _destroy_loop_device()
> > {
> >         dev=${1}
> >         umount $dev 2>&1
> 
> If unmount fails, what then?
> 
> >         file=`losetup -a | grep $dev | sed -n "s/.*(\(.*\))$/\1/p"`
> >         losetup -d $dev && rm -f $file || _fail "Cannot destroy
> >         loop device"
> 
> And if unmount destroys the loop device automatically? That will fail
> the test, right?

I wasn't aware of that. I've always used the two-step approach:

losetup /dev/loopX file
mount /dev/loopX mntpoint

and subsequent umount never destroyed loop device in my case. I tried to use only

mount file mntpoint

which then resulted in behaviour you described. Is this the rule or is some other magic in that?

> Also, what happens if we unmount the filesystem first so we can run
> consistency checks on the image before we destroy it?
> 
> I'd suggest that it is the test's responsibility to create, mount,
> unmount, check and destroy the image file as those vary from test to
> test. Hence a better idea is to just use an image path/device API.
> i.e:

Thanks for useful comments, I appreciate that. 

Tomas

> 
> _create_loop_device()
> {
> 	file=$1
>         dev=`losetup -f`
>         losetup $dev $file || _fail "Cannot associate $file with
>         $dev"
>         echo $dev
> }
> 
> _destroy_loop_device()
> {
> 	dev=$1
> 	losetup -d $dev || _fail "Cannot destroy loop device"
> }
> 
> Cheers,
> 
> Dave.
> --
> Dave Chinner
> david@fromorbit.com
> 

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