Re: Race in rcar-v4l2.c

["Niklas Söderlund" <niklas.soderlund@ragnatech.se>]

Hi Tomi,

Thanks for your report.

I have an on-going series trying to clean this all up [1]. The one 
change in the v4l-async core I proposed was however rejected and I have 
yet to circle back to figure out a different solution.

Could you give it a try and see if it also solves this issue?

1. [PATCH 0/6] media: rcar-vin: Make use of multiple connections in v4l-async
   https://lore.kernel.org/linux-renesas-soc/20240129202254.1126012-1-niklas.soderlund+renesas@ragnatech.se/

On 2024-09-06 12:57:50 +0300, Tomi Valkeinen wrote:
> Hi Niklas,
> 
> There seems to be a race in rcar-v4l2.c, causing
> WARN_ON(entity->use_count < 0) in pipeline_pm_power_one().
> 
> If my understanding is correct, the VIN v4l2 nodes are being created
> (rvin_v4l2_register), meaning they are userspace accessible, but the media
> pipeline as a whole is not ready yet (e.g. media links).
> 
> So what happens is that after some video nodes have been created, the
> userspace opens them (I think it's udevd checking the new device nodes),
> causing rvin_open(). rvin_open() goes through the media graph and does some
> PM enabling (I'm not familiar with the legacy v4l2_pipeline_pm_get()).
> However, as the links are not there, it doesn't really enable much at all.
> 
> Then the driver goes forward and finishes with the media graph.
> 
> Then the userspace closes the opened video nodes, rvin_release() gets called
> and it goes through the media graph, which now contains all the entities,
> and powers them down. As the entities were never powered up, we hit the
> use_count warning.
> 
> This happens quite often to me when loading the modules, but I think it can
> be made to happen more often by adding msleep(1000) to the beginning of
> rvin_release(), thus ensuring that the graph setup is finished before the
> rvin_release() proceeds (and hoping that the graph setup was not ready when
> rvin_open() was called).
> 
>  Tomi
> 

-- 
Kind Regards,
Niklas Söderlund