Re: [PATCH 15/22] media: sunxi: Remove the sun6i-csi driver implementation

Hi,

On Mon 13 Sep 21, 10:17, Maxime Ripard wrote:

On Fri, Sep 10, 2021 at 08:41:40PM +0200, Paul Kocialkowski wrote:

quoted

As described in the commit adding support for the new sun6i-csi driver,
a complete rewrite was necessary to support the Allwinner A31 ISP as
well as fix a number of issues with the current implementation.

Farewell and thanks for all the pixels!

Signed-off-by: Paul Kocialkowski <redacted>

For completeness, this is what the other commit log mentions:

quoted

While adapting the sun6i-csi driver for MIPI CSI-2 support was
possible, it became clear that adding support for the ISP required
very heavy changes to the driver which were quite hard to break down
into a series of subsequent changes.

quoted

The first major difficulty comes from the lack of v4l2 subdev that
acts a bridge, separate from the video node representing the DMA
engine. To support the ISP, only parts of the hardware must be
configured (excluding aspects related to the DMA output), which made
the separation a hard requirement.

quoted

Another significant difficulty was the specific dance that is required
to have both the ISP and CSI device be part of the same media device.
Because the ISP and CSI are two different hardware blocks, they have
two distinct drivers that will each try to register their own v4l2
and media devices, resulting in two distinct pipelines. When the ISP
is in use, we actually want the CSI driver to register with the ISP's
v4l2 and media devices while keeping the ability to register its own
when the ISP is not in use. This is done by:
1. Having the CSI driver check whether the ISP is available, using
   sun6i_csi_isp_detect();
2. If not, it can register when its own async subdevs are ready, using
   sun6i_csi_v4l2_complete();
3. If so, it will register its bridge as an async subdev which will
   be picked-up by the ISP driver (from the fwnode graph link);
4. When the subdev becomes bound to the ISP's v4l2 device, we can
   then access that device (and the associated media device) to
   complete registration of the capture video node, using
   sun6i_csi_isp_complete();
Besides the logic rework, other issues were identified and resolved:
- The sync mechanism for buffer flipping was based on the frame done
  interrupt, which is too late (next frame is already being processed).
  This lead to requiring 3 buffers to start and writing two addresses
  when starting. Using vsync as a sync point seems to be the correct
  approach and allows using only two buffers without tearing;
- Using devm_regmap_init_mmio_clk was incorrect since the reset also
  comes into play;
- Some register definitions were inverted compared to their actual
  effect (which was inherited from the Allwinner documentation and
  code): comments were added where relevant;
- The deprecated v4l2_async_notifier_parse_fwnode_endpoints() helper
  is no longer used by the driver;

With that being said, NAK.

Having heavy changes to a driver is completely fine, and is kind of
expected really with such a big change. Breaking all possibility of
bisection and throwing away years of stabilization and maintenance
isn't.

And all those small bug fixes you mention at the end are just that:
small bug fixes that can be done on the current driver just fine too.

I understand that this looks like we're trashing all the work that was
done previously by removing the current driver and adding the new one
but the logic for deciding what to write into registers was carefully
preserved from the original driver to make sure that the works of
stabilization and maintenance are not lost.

However I would understand that my good promise on this is not enough,
so perhaps I could provide a combinatory verification that the same set
of mbus/pixel formats end up with the same thing being written into
registers.

In addition I understand that it will be necessary to split the changes
up into small commits to clarify the transition path between the two
drivers. So I will do my best to split things up.

Does that seem like an agreeable plan or do you see other things that
would be blockers?

My initial thought was that it would be much easier to review the driver as a
rewrite, but I'm not too surprised I was wrong. To be honest it was nearly
impossible to actually have the initial development happen as sequential steps
and I preferred to allocate my time on other tasks than splitting the changes
into these sequential steps.

Cheers,

Paul

-- 
Paul Kocialkowski, Bootlin
Embedded Linux and kernel engineering
https://bootlin.com