Re: uvcvideo: Race on dev->state between uvc_disconnect() and uvc_v4l2_open()

[Eugene Shatokhin <eugene.shatokhin@rosalab.ru>]

25.05.2015 01:32, Laurent Pinchart пишет:
> Hi Eugene,
>
> On Wednesday 20 May 2015 17:48:41 Eugene Shatokhin wrote:
>> Hi,
>>
>> There is a race in uvcvideo module between uvc_disconnect() and
>> uvc_v4l2_open() on dev->state. Checked and reproduced that with kernel
>> 4.1-rc1.
>>
>> drivers/media/usb/uvc/uvc_driver.c, uvc_disconnect():
>>
>> 	dev->state |= UVC_DEV_DISCONNECTED;
>>
>> drivers/media/usb/uvc/uvc_v4l2.c, uvc_v4l2_open():
>>
>> 	if (stream->dev->state & UVC_DEV_DISCONNECTED)
>> 		return -ENODEV;
>>
>> I checked that the race does happen by introducing a delay in
>> uvc_disconnect() right before that assignment and armed a hardware
>> breakpoint to detect the access to stream->dev->state from
>> uvc_v4l2_open(). When I disconnected the webcam while Google Hangout was
>> running, the hardware breakpoint triggered several times for that read
>> in uvc_v4l2_open (uvc_v4l2.c:484). uvc_v4l2_open() was called in the
>> context of GoogleTalkPlugin processes.
>>
>> Not sure if the race is intentional but I guess, better to report it
>> anyway. Nothing has crashed during my (brief) testing yet, but still.
>
> The race condition between disconnect and open is unavoidable. What is
> avoidable, though, is the crashes and other ill side-effects that could result
> from it. The following commit handles that.
>
> commit ca9afe6f87b569cdf8e797395381f18ae23a2905
> Author: Hans Verkuil <hverkuil@xs4all.nl>
> Date:   Fri Nov 26 06:54:53 2010 -0300
>
>      [media] v4l2-dev: fix race condition
>
>      The unregister function had a race condition with the v4l2_open
>      function. Ensure that both functions test and clear the REGISTER
>      flag from within a critical section.
>
>      Thanks to Laurent Pinchart for finding this race.
>
>      Signed-off-by: Hans Verkuil <hverkuil@xs4all.nl>
>      Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>
>
>
> The race was previously handled by the uvcvideo driver, and the code got
> removed in the following commit after the above commit got merged.
>
> commit 716fdee110ceb816cca8c46c0890d08c5a1addb9
> Author: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> Date:   Tue Sep 29 21:07:19 2009 -0300
>
>      V4L/DVB (13152): uvcvideo: Rely on videodev to reference-count the device
>
>      The uvcvideo driver has a driver-wide lock and a reference count to
> protect
>      against a disconnect/open race. Now that videodev handles the race itself,
>      reference-counting in the driver can be removed.
>
>      Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
>      Signed-off-by: Mauro Carvalho Chehab <mchehab@redhat.com>
>
>
> Setting the UVC_DEV_DISCONNECTED flag seems unneeded nowadays. I'll have to
> carefully think about it though, and it's too late right now to do so :-)
>

I see. Thanks for the explanation!

Regards,

Eugene