Re: Race condition between driver_probe_device and device_shutdownâ

[Alan Stern]

On Wed, 23 May 2012, Greg KH wrote:

> On Thu, May 24, 2012 at 09:39:46AM +0800, Ming Lei wrote:
> > On Wed, May 23, 2012 at 11:06 PM, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:
> > > On Wed, 23 May 2012, Ming Lei wrote:
> > >> The .shutdown callback pointer is got from device->driver, which is
> > >> changed in probe and release path. Also runtime PM thing has been
> > >> involved into shutting down recently, so looks not only hardware parts
> > >> are involved now.
> > >
> > > This is a tricky question.  Overall I think you're probably right.
> > >
> > > It's certainly true that holding the device lock across the shutdown
> > > callback is the easiest and most reliable way to prevent these races.
> > 
> > But holding device lock across .shutdown is very inefficient because
> > most of devices' driver have not shutdown callback, so I think it is better

The code there is racy already.  It does:

		} else if (dev->driver && dev->driver->shutdown) {

without any locking protection.  If the driver is unbound while this 
statement runs then dev->driver could be non-NULL for the first test 
and NULL for the second.

> > to fix the race by prevent driver core from probing or releasing once
> > shutdown is started.
> > 
> > How about the below patch?
> 
> How about waiting for the original poster to respond as to exactly how
> they are hitting this race before doing anything?

In addition, the patch is too complicated.  For this type of
synchronization you should use SRCU.  See
Documentation/RCU/whatisRCU.txt and related files.

Alan Stern

--
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/