Re: [PATCH] usbnet: Fix use-after-free on disconnect

[Lukas Wunner <lukas@wunner.de>]

On Thu, Apr 14, 2022 at 01:07:35PM +0200, Greg Kroah-Hartman wrote:
> On Thu, Apr 14, 2022 at 12:58:58PM +0200, Lukas Wunner wrote:
> > On Wed, Apr 13, 2022 at 08:59:48PM +0200, Oliver Neukum wrote:
> > > On 13.04.22 16:16, Lukas Wunner wrote:
> > > > --- a/drivers/net/usb/usbnet.c
> > > > +++ b/drivers/net/usb/usbnet.c
> > > > @@ -469,6 +469,9 @@ static enum skb_state defer_bh(struct usbnet *dev, struct sk_buff *skb,
> > > >   */
> > > >  void usbnet_defer_kevent (struct usbnet *dev, int work)
> > > >  {
> > > > +	if (dev->intf->condition == USB_INTERFACE_UNBINDING)
> > > > +		return;
> > > 
> > > But, no, you cannot do this. This is a very blatant layering violation.
> > > You cannot use states internal to usb core like that in a driver.
> > 
> > Why do you think it's internal?
> > 
> > enum usb_interface_condition is defined in include/linux/usb.h
> > for everyone to see and use.  If it was meant to be private,
> > I'd expect it to be marked as such or live in drivers/usb/core/usb.h.
> 
> Because we didn't think people would do crazy things like this.

I assume "crazy things" encompasses reading and writing intf->condition
without any locking or explicit memory barriers.  However many drivers
do that through the exported functions:

  usb_reset_device()
  usb_lock_device_for_reset()
  usb_driver_claim_interface()
  usb_driver_release_interface()

In any case, I've decided to pursue a different approach which fixes the
issue in core networking code rather than usbnet.  USB Ethernet may not
be the only culprit after all.  A replacement patch superseding this one
was just submitted:

https://lore.kernel.org/netdev/18b3541e5372bc9b9fc733d422f4e698c089077c.1650177997.git.lukas@wunner.de

Thanks,

Lukas