A potential race

Hi!

There is a potential race condition between usbvision_v4l2_close
<http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L403> and usbvision_disconnect
<http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1569>. The possible scenario may be the following. usbvision_disconnect <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1569> starts execution, assigns usbvision->remove_pending = 1 <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1587>, and is interrupted (rescheduled) after mutex_unlock <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1592>. After that usbvision_v4l2_close <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L403> is executed, decrease usbvision->user-- <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L419>, checks usbvision->remove_pending <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L422>, executes usbvision_release <http://lxr.free-electrons.com/ident?i=usbvision_release> and finishes. Then usbvision_disconnect
<http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1569> continues its execution. It checks usbversion->user <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1594> (it is already 0) and also execute usbvision_release <http://lxr.free-electrons.com/ident?i=usbvision_release>. Thus, release is executed twice. The same situation may occur if usbvision_v4l2_close <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L403> is interrupted by

usbvision_disconnect <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1569>. Moreover, the same problem is in usbvision_radio_close <http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1135>. In all these cases the check before call usbvision_release <http://lxr.free-electrons.com/ident?i=usbvision_release> under mutex_lock protection does not solve the problem, because there may occur an open() after the check and the race takes place again. The question is: why the usbvision_release
<http://lxr.free-electrons.com/ident?i=usbvision_release> is called from close() (usbvision_v4l2_close
<http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L403> and usbvision_radio_close
<http://lxr.free-electrons.com/source/drivers/media/usb/usbvision/usbvision-video.c#L1135>)? Usually release functions are called from disconnect.

Quoted 8 lines

Hi!

There is a potential race condition between usbvision_v4l2_close and usbvision_disconnect. The possible scenario may be the following. usbvision_disconnect starts execution, assigns usbvision->remove_pending = 1, and is interrupted
(rescheduled) after mutex_unlock. After that usbvision_v4l2_close is executed, decrease usbvision->user-- , checks usbvision->remove_pending, executes usbvision_release and finishes. Then usbvision_disconnect continues its execution. It checks
usbversion->user (it is already 0) and also execute usbvision_release. Thus, release is executed twice. The same situation may
occur if usbvision_v4l2_close is interrupted by usbvision_disconnect. Moreover, the same problem is in usbvision_radio_close. In all these cases the check before call usbvision_release under mutex_lock protection does not solve the problem, because there may occur an open() after the check and the race takes place again. The question is: why the usbvision_release is called from close() (usbvision_v4l2_close and usbvision_radio_close)? Usually release functions are called from
disconnect.

Please don't use html mail, mailinglists will silently reject this.

The usbvision driver is old and unloved and known to be very bad code. It needs a huge amount of work to make all this work correctly.

I don't see anyone picking this up...

Regards,

Hans

01.07.2016 19:53, Hans Verkuil пишет:

Quoted 13 lines · Pavel Andrianov

Quoted 8 lines

Hi!

There is a potential race condition between usbvision_v4l2_close and usbvision_disconnect. The possible scenario may be the following. usbvision_disconnect starts execution, assigns usbvision->remove_pending = 1, and is interrupted
(rescheduled) after mutex_unlock. After that usbvision_v4l2_close is executed, decrease usbvision->user-- , checks usbvision->remove_pending, executes usbvision_release and finishes. Then usbvision_disconnect continues its execution. It checks
usbversion->user (it is already 0) and also execute usbvision_release. Thus, release is executed twice. The same situation may
occur if usbvision_v4l2_close is interrupted by usbvision_disconnect. Moreover, the same problem is in usbvision_radio_close. In all these cases the check before call usbvision_release under mutex_lock protection does not solve the problem, because there may occur an open() after the check and the race takes place again. The question is: why the usbvision_release is called from close() (usbvision_v4l2_close and usbvision_radio_close)? Usually release functions are called from
disconnect.

Please don't use html mail, mailinglists will silently reject this.

The usbvision driver is old and unloved and known to be very bad code. It needs a huge amount of work to make all this work correctly.

Quoted 5 lines

I don't see anyone picking this up...

Regards,

Hans

If you know the driver, could you, please, explain me, why usbvision_release is called from close functions (usbvision_v4l2_close and usbvision_radio_close) and not only from disconnect? Thanks!

Quoted 14 lines

01.07.2016 19:53, Hans Verkuil пишет:

Quoted 13 lines · Pavel Andrianov

Quoted 8 lines

   Hi!

There is a potential race condition between usbvision_v4l2_close and usbvision_disconnect. The possible scenario may be the following. usbvision_disconnect starts execution, assigns usbvision->remove_pending = 1, and is interrupted
(rescheduled) after mutex_unlock. After that usbvision_v4l2_close is executed, decrease usbvision->user-- , checks usbvision->remove_pending, executes usbvision_release and finishes. Then usbvision_disconnect continues its execution. It checks
usbversion->user (it is already 0) and also execute usbvision_release. Thus, release is executed twice. The same situation may
occur if usbvision_v4l2_close is interrupted by usbvision_disconnect. Moreover, the same problem is in usbvision_radio_close. In all these cases the check before call usbvision_release under mutex_lock protection does not solve the problem, because there may occur an open() after the check and the race takes place again. The question is: why the usbvision_release is called from close() (usbvision_v4l2_close and usbvision_radio_close)? Usually release functions are called from
disconnect.

Please don't use html mail, mailinglists will silently reject this.

The usbvision driver is old and unloved and known to be very bad code. It needs a huge amount of work to make all this work correctly.

Quoted 9 lines

Quoted 5 lines

I don't see anyone picking this up...

Regards,

Hans

If you know the driver, could you, please, explain me, why usbvision_release is called from close functions (usbvision_v4l2_close and usbvision_radio_close) and not only from disconnect? Thanks!

Because the author didn't know what he was doing. Although, to be fair, we have much better solutions for this. But who is willing to put in the time to fix this properly?

The basic idea was: if someone still has a video/radio node open when disconnect happens, then we leave it to the last close to call release, otherwise we can call release right away.

It needs to be rewritten.

If you're volunteering to clean this up, then I can give pointers.

Regards,

Hans