On Sun, 19 Apr 2020, Dmitry Torokhov wrote:
> On Sun, Apr 19, 2020 at 10:07:34AM -0400, Alan Stern wrote:
> > On Sat, 18 Apr 2020, Dmitry Torokhov wrote:
> >
> > > On Sat, Apr 18, 2020 at 09:09:44PM -0700, Dmitry Torokhov wrote:
> > > > Hi Alan,
> > > >
> > > > On Sat, Apr 18, 2020 at 10:16:32PM -0400, Alan Stern wrote:
> > > > > linux-input people:
> > > > >
> > > > > syzbot has found a bug related to USB/HID/input, and I have narrowed it
> > > > > down to the wacom driver. As far as I can tell, the problem is caused
> > > > > the fact that drivers/hid/wacom_sys.c calls input_register_device()
> > > > > in several places, but it never calls input_unregister_device().
> > > > >
> > > > > I know very little about the input subsystem, but this certainly seems
> > > > > like a bug.
> > > >
> > > > Wacom driver uses devm_input_allocate_device(), so unregister should
> > > > happen automatically on device removal once we exit wacom_probe().
> > > >
> > > > >
> > > > > When the device is unplugged, the disconnect pathway doesn't call
> > > > > hid_hw_close(). That routine doesn't get called until the user closes
> > > > > the device file (which can be long after the device is gone and
> > > > > hid_hw_stop() has run). Then usbhid_close() gets a use-after-free
> > > > > error when it tries to access data structures that were deallocated by
> > > > > usbhid_stop(). No doubt there are other problems too, but this is
> > > > > the one that syzbot found.
> > > >
> > > > Unregistering the input device should result in calling wacom_close()
> > > > (if device was previously opened), which, as far as I can tell, calls
> > > > hid_hw_close().
> > > >
> > > > I wonder if it is valid to call hid_hw_stop() before hid_hw_close()?
> >
> > No, it isn't. If it were, for example, why would evdev_disconnect() ->
> > evdev_cleanup() need to call input_close_device()?
>
> Because input and HID are not the same. For input, when we attempt to
> unregister an input device we will go through all attached input
> handlers (like evdev) and if they believe they have the device open they
> will attempt to close it. How close is implemented is up to particular
> driver.
>
> I am not sure about HID implementation details, but I could envision
> transports where you can tell the transport that you no longer want
> events to be delivered to you ("close") vs you want to disable hardware
> ("stop") and support any order of them.
Jiri, you should know: Are HID drivers supposed to work okay when the
->close callback is issued after (or concurrently with) the ->stop
callback?
The actual bug found by syzbot was a race between those two routines in
usbhid.
> > And why would
> > usbhid_disconnect() deallocate the usbhid structure which usbhid_stop()
> > accesses?
>
> This happens only after we return from hid_destroy_device(), so
> even in the presence of devm I'd expect that all devm-related stuff
> instantiated by hid-wacom would have been completed before we get back
> to usbhid_disconnect().
>
> Can we validate that calls to wacom_close() happen?
I could find out if you think it's important. In the syzbot tests, the
crash occurs before wacom_close() is called.
Alan Stern
