Re: camss NULL-deref on power on with 6.12-rc2

[Johan Hovold <johan@kernel.org>]

On Mon, Apr 07, 2025 at 12:38:56PM +0200, Johan Hovold wrote:
> On Mon, Apr 07, 2025 at 10:58:52AM +0100, Bryan O'Donoghue wrote:
> > On 07/04/2025 10:12, Johan Hovold wrote:

> > > [    5.740833] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000030
> 
> > > [    5.744704] Call trace:
> > > [    5.744706]  camss_find_sensor_pad+0x20/0x74 [qcom_camss] (P)
> > > [    5.744711]  camss_get_pixel_clock+0x18/0x64 [qcom_camss]
> > > [    5.744716]  vfe_get+0xb8/0x504 [qcom_camss]
> > > [    5.744724]  vfe_set_power+0x30/0x58 [qcom_camss]
> > > [    5.744731]  pipeline_pm_power_one+0x13c/0x150 [videodev]
> > > [    5.744745]  pipeline_pm_power.part.0+0x58/0xf4 [videodev]
> > > [    5.744754]  v4l2_pipeline_pm_use+0x58/0x94 [videodev]
> > > [    5.744762]  v4l2_pipeline_pm_get+0x14/0x20 [videodev]
> > > [    5.744771]  video_open+0x78/0xf4 [qcom_camss]
> > > [    5.744776]  v4l2_open+0x80/0x120 [videodev]

> I've only seen it twice myself (that I've noticed, at least this time it
> prevented the display from probing so I knew something was wrong).

Just hit this again with 6.15-rc1 after the third reboot so timing has
likely changed slightly which now makes it easier to hit this.

> Since it's obviously a race condition I think you'll need to analyse the
> code to try to figure out where the bug is. With an hypothesis you may
> be able to instrument a reliable reproducer (e.g. by adding appropriate
> delays to extend the race window).

It's apparently udev which powers up the camera when running v4l_id:

[    5.859741] CPU: 4 UID: 0 PID: 420 Comm: v4l_id Not tainted 6.15.0-rc1 #106 PREEMPT

So this looks like the classic bug of drivers registering their devices
before they have been fully set up.

> The fact that the sensor driver is probe deferring may also be relevant
> here.

Johan