Date:   Mon, 19 Jul 2021 12:53:11 +0200
From:   Dennis Rachui <drachui@...adit-jv.com>
To:     Niklas Söderlund <niklas.soderlund@...natech.se>
CC:     Laurent Pinchart <laurent.pinchart@...asonboard.com>,
        Steve Longerbeam <slongerbeam@...il.com>,
        Mauro Carvalho Chehab <mchehab@...nel.org>,
        Maxime Ripard <mripard@...nel.org>,
        Sakari Ailus <sakari.ailus@...ux.intel.com>,
        <linux-media@...r.kernel.org>, <linux-renesas-soc@...r.kernel.org>,
        <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] media: rcar-csi2: do not update format while streaming

Hi Niklas,

> On Fri, Jul 16, 2021 17:02:41 +0200, Niklas Söderlund wrote:
> > On 2021-07-16 16:09:21 +0200, Dennis Rachui wrote:
> > > On Thu, Jul 15, 2021 at 12:57:37PM +0200, Niklas Söderlund wrote:
> > > > On 2021-07-14 21:40:51 +0300, Laurent Pinchart wrote:
> > > > > On Tue, Jul 13, 2021 at 11:42:17AM +0200, Niklas Söderlund wrote:
> > > > > > On 2021-07-11 01:42:44 +0300, Laurent Pinchart wrote:
> > > > > > > On Fri, Jul 09, 2021 at 04:20:40PM +0200, Niklas Söderlund wrote:
> > > > > > > > On 2021-07-08 15:22:58 +0200, Dennis Rachui wrote:
> > > > > > > > > Verify that streaming is not active before setting the pad format.
> > > > > > > > > 
> > > > > > > > > According to the VIDIOC documentation [1] changes to the 
> > > > > > > > > active format of a media pad via the VIDIOC_SUBDEV_S_FMT 
> > > > > > > > > ioctl are applied to the underlying hardware.
> > > > > > > > > In rcar-csi2 a format change only applies to hardware, when 
> > > > > > > > > the pipeline is started. While the device is not in use, it 
> > > > > > > > > is therefore okay to update the format.
> > > > > > > > > 
> > > > > > > > > However, when the pipeline is active, this leads to a format 
> > > > > > > > > mismatch between driver and device.
> > > > > > > > > Other applications can query the format with 
> > > > > > > > > VIDIOC_SUBDEV_G_FMT at any time and would be reported a 
> > > > > > > > > format that does not fit the current stream.
> > > > > > > > > 
> > > > > > > > > This commit prevents format update while streaming is active 
> > > > > > > > > and returns -EBUSY to user space, as suggested by [1].
> > > > > > > > > 
> > > > > > > > > [1] 
> > > > > > > > > Documentation/userspace-api/media/v4l/vidioc-subdev-g-fmt.rs
> > > > > > > > > t
> > > > > > > > 
> > > > > > > > I like that this is addressed, but I wonder is this not 
> > > > > > > > something that should be fixed in the V4L2 core and not in drivers?
> > > > > > > 
> > > > > > > Some drivers may support format changes during streaming (that's 
> > > > > > > allowed by the V4L2 API, I'm not sure if it's used anywhere 
> > > > > > > though). While I'd favour not duplicating the same logic in 
> > > > > > > different (and differently
> > > > > > > buggy) ways in drivers, I'm not sure how this could be 
> > > > > > > implemented in a sane way in the V4L2 core in its current state.
> > > > > > 
> > > > > > I understand it's possible from some devices to support to format 
> > > > > > changes during streaming, but as you point out it's the exception 
> > > > > > and not the rule, if used at all.
> > > > > > 
> > > > > > So my point is if we start to enforce this in drivers we are 
> > > > > > headed down a road where this will be messier to clean up. Would 
> > > > > > it not make more sens to default the V4L2 core to disallow format 
> > > > > > changes while streaming and add a new flag to V4L2_SUBDEV_CAP_ to 
> > > > > > signal that the subdevice supports format changes while streaming?
> > > > > > 
> > > > > > We already have V4L2_SUBDEV_CAP_RO_SUBDEV to signal that a 
> > > > > > subdevice only supports read-only operations so I think it would 
> > > > > > not be too hard to move this functionality into the core?
> > > > > 
> > > > > Yes, that's something we could try. The subdev core will then need 
> > > > > to track the streaming state, which may require wrapping the 
> > > > > .s_stream() call. Locking should then also likely be handled by the 
> > > > > core. Probably nothing impossible, but quite a bit of work. Any 
> > > > > volunteer ? :-)
> > > > 
> > > > We already track the stream count in struct media_entity and it's 
> > > > incremented/decremented under the media device lock by
> > > > media_pipeline_start() and media_pipeline_stop(). So I don't think 
> > > > it's such a hard feature to add.
> > > > 
> > > > The large task IMHO is to figure out if we have any subdevice in tree 
> > > > that allows format changes while streaming and that would need to set 
> > > > this new V4L2_SUBDEV_CAP_ flag.
> > >
> > > Many subdevs allow format changes during streaming. The question is whether any of them do so knowingly, or if they're all buggy :-) I'd be surprised if there
> > > were more than a couple of drivers that actually support this correctly.
> > 
> > From my perspective, the current stream_count from struct media_entity
> > would not be sufficient. References should be counted per struct media_pad.
> > Otherwise, devices that allow multiple parallel streams would block user
> > space from updating media-pads that are not part of an active media-pipeline.
> > E.g. in rcar-csi2 this could effect a source pad connected to currently
> > unused VIN device.
>
> I understand your reasoning, but with the current V4L2 design is this 
> really a concern? As s_stream() is not pad/stream aware it acts more or 
> less as a big start / stop button.
>
> When starting all enabled media links formats needs to be evaluated 
> (something we can't do yet with multiplexed streams..) and all 
> subdivides that are part of the media graph are started. We can not 
> enable or disable any media links while the pipeline is streaming so I'm 
> not sure configuring the format of pads not part of the active capture 
> is such a big concern, am I missing something?

I missed the fact that media links can not be enabled / disabled while
the pipeline is streaming. So I agree that preventing format updates while
the entity is busy is no big concern. Userspace applications should be aware
of this just as they must be aware that link enable / disable only works
while the pipeline is not running.

> 
> > > > > > > > > Note: after creation of this commit, it was noticed that 
> > > > > > > > > Steve Longerbeam has a very similar solution in his fork.
> > > > > > > > > 
> > > > > > > > > Fixes: 769afd212b16 ("media: rcar-csi2: add Renesas R-Car 
> > > > > > > > > MIPI CSI-2 receiver driver")
> > > > > > > > > Cc: Steve Longerbeam <slongerbeam@...il.com>
> > > > > > > > > Signed-off-by: Dennis Rachui <drachui@...adit-jv.com>
> > > > > > > > > ---
> > > > > > > > >  drivers/media/platform/rcar-vin/rcar-csi2.c | 21 
> > > > > > > > > ++++++++++++++++++++-
> > > > > > > > >  1 file changed, 20 insertions(+), 1 deletion(-)
> > > > > > > > > 
> > > > > > > > > diff --git a/drivers/media/platform/rcar-vin/rcar-csi2.c 
> > > > > > > > > b/drivers/media/platform/rcar-vin/rcar-csi2.c
> > > > > > > > > index e28eff0..98152e1 100644
> > > > > > > > > --- a/drivers/media/platform/rcar-vin/rcar-csi2.c
> > > > > > > > > +++ b/drivers/media/platform/rcar-vin/rcar-csi2.c
> > > > > > > > > @@ -724,18 +724,37 @@ static int rcsi2_set_pad_format(struct 
> > > > > > > > > v4l2_subdev *sd,  {
> > > > > > > > >  	struct rcar_csi2 *priv = sd_to_csi2(sd);
> > > > > > > > >  	struct v4l2_mbus_framefmt *framefmt;
> > > > > > > > > +	int ret = 0;
> > > > > > > > > +
> > > > > > > > > +	mutex_lock(&priv->lock);
> > > > > > > > >  
> > > > > > > > >  	if (!rcsi2_code_to_fmt(format->format.code))
> > > > > > > > >  		format->format.code = rcar_csi2_formats[0].code;
> > > > > > > > >  
> > > > > > > > >  	if (format->which == V4L2_SUBDEV_FORMAT_ACTIVE) {
> > > > > > > > > +
> > > > > > > > > +		/*
> > > > > > > > > +		 * Do not apply changes to active format while streaming.
> > > > > > > > > +		 *
> > > > > > > > > +		 * Since video streams could be forwarded from sink pad to any
> > > > > > > > > +		 * source pad (depending on CSI-2 channel routing), all
> > > > > > > > > +		 * media pads are effected by this rule.
> > > > > > > > > +		 */
> > > > > > > > > +		if (priv->stream_count > 0) {
> > > > > > > > > +			ret = -EBUSY;
> > > > > > > > > +			goto out;
> > > > > > > > > +		}
> > > > > > > > > +
> > > > > > > > >  		priv->mf = format->format;
> > > > > > > > >  	} else {
> > > > > > > > >  		framefmt = v4l2_subdev_get_try_format(sd, sd_state, 0);
> > > > > > > > >  		*framefmt = format->format;
> > > > > > > > >  	}
> > > > > > > > >  
> > > > > > > > > -	return 0;
> > > > > > > > > +out:
> > > > > > > > > +	mutex_unlock(&priv->lock);
> > > > > > > > > +
> > > > > > > > > +	return ret;
> > > > > > > > >  }
> > > > > > > > >  
> > > > > > > > >  static int rcsi2_get_pad_format(struct v4l2_subdev *sd,

--
Regards,
Dennis Rachui

Powered by blists - more mailing lists