Re: [PATCH] r8152: Allocate interrupt buffer as part of struct r8152

[Guenter Roeck <linux@roeck-us.net>]

On Tue, Jan 31, 2017 at 02:53:47PM -0500, Alan Stern wrote:
> On Tue, 31 Jan 2017, Guenter Roeck wrote:
> 
> > When unloading the r8152 driver using the 'unbind' sysfs attribute
> > in a system with KASAN enabled, the following error message is seen
> > on a regular basis.
> 
> ...
> 
> > The two-byte allocation in conjunction with code analysis suggests that
> > the interrupt buffer has been overwritten. Added instrumentation in the
> > driver shows that the interrupt handler is called after RTL8152_UNPLUG
> > was set, and that this event is associated with the error message above.
> > This suggests that there are situations where the interrupt buffer is used
> > after it has been freed.
> > 
> > To avoid the problem, allocate the interrupt buffer as part of struct
> > r8152.
> > 
> > Cc: Hayes Wang <hayeswang@realtek.com>
> > Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> > ---
> > The problem is seen in chromeos-4.4, but there is not reason to believe
> > that it does not occur with the upstream kernel. It is still seen in
> > chromeos-4.4 after all patches from upstream and linux-next have been
> > applied to the driver.
> > 
> > While relatively simple, I am not really convinced that this is the best
> > (or even an acceptable) solution for this problem. I am open to suggestions
> > for a better fix.
> 
> The proper approach is to keep the allocation as it is, but _before_
> deallocating the buffer, make sure that the interrupt buffer won't be
> accessed any more.  This may involve calling usb_kill_urb(), or
> synchronize_irq(), or something similar.
> 

Just to keep everyone up to date, the problem was that the usb subsystem,
due to bad platform code in chromeos-4.4, did not properly stop DMA from
the hardware when the driver was removed. This resulted in a DMA transfer
into the freed buffer. The r8152 driver is completely innocent.

Sorry for the noise.

Guenter