Re: 2.6.17-rc6-mm1/pktcdvd - BUG: possible circular locking

[Peter Osterlund]

Arjan van de Ven <arjan@xxxxxxxxxxxxxxx> writes:

> On Fri, 2006-07-14 at 13:22 +0200, Peter Osterlund wrote:
> > > and what locking prevents this? And via multiple opens?
> > 
> > You are right that my reasoning was incorrect. If someone is doing
> > "pktsetup ; pktsetup -d" quickly in a loop while someone else is
> > trying to open the device, one thread could be at the start of
> > pkt_open() at the same time as another thread is in pkt_new_dev().
> > 
> > However, I added a 5s delay in pkt_open() to enlarge the race window.
> > I still couldn't make the driver lock up though. The explanation is
> > that pkt_new_dev() calls blkdev_get() with the CD device (eg /dev/hdc)
> > as bdev parameter, while do_open() locks the bd_mutex for the pktcdvd
> > device (eg /dev/pktcdvd/0).
> > 
> > Do you still think this could deadlock? If not, how should the code be
> > annotated to make this warning go away?
> 
> unless we KNOW it won't deadlock (eg we have a "this cannot deadlock
> BECAUSE of X, Y and Z") I don't think annotations are the right idea. In
> addition, the "how to annotate" really depends on what X, Y and Z
> are....

In the first call chain, do_open -> pkt_open, the bd_mutex object that
is being locked corresponds to a pktcdvd block device, because those
are the only devices that have their open method set to pkt_open.

In the second call chain, pkt_ctl_ioctl -> pkt_new_dev -> do_open, the
bd_mutex object that is being locked *does not* correspond to a
pktcdvd block device, because pkt_new_dev will bail out with a "Can't
chain pktcdvd devices" error if you call it with "dev" set to a
pktcdvd device.

Therefore, there is no AB-BA deadlock possibility. The locking
dependencies are A -> B and B -> A', where it is known that A, B and
A' are all different.

So the claim from the lockdep code, "BUG: possible circular locking
deadlock detected!", is a false alarm.

-- 
Peter Osterlund - petero2@xxxxxxxxx
http://web.telia.com/~u89404340
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/