[PATCH V2 4/9] drm/exynos: add exynos_dp_panel driver registration to drm driver

[Thierry Reding]

Adding Boris and Ludovic since this topic came up in a different thread
as well.

On Thu, Aug 21, 2014 at 01:25:07AM -0700, Stéphane Marchesin wrote:
> On Thu, Aug 21, 2014 at 12:36 AM, Thierry Reding
> <thierry.reding at gmail.com> wrote:
> > On Tue, Aug 19, 2014 at 09:02:39PM -0700, Stéphane Marchesin wrote:
> >> On Tue, Apr 22, 2014 at 8:26 AM, Thierry Reding
> >> <thierry.reding at gmail.com> wrote:
> >> > On Tue, Apr 22, 2014 at 08:33:23PM +0530, Ajay kumar wrote:
> >> >> Hi Thierry,
> >> >>
> >> >> On Tue, Apr 22, 2014 at 2:03 PM, Thierry Reding
> >> >> <thierry.reding at gmail.com> wrote:
> >> >> > On Tue, Apr 22, 2014 at 04:09:13AM +0530, Ajay Kumar wrote:
> >> >> >> Register exynos_dp_panel before the list of exynos crtcs and
> >> >> >> connectors are probed.
> >> >> >>
> >> >> >> This is needed because exynos_dp_panel should be registered to
> >> >> >> the drm_panel list via panel-exynos-dp probe, i.e much before
> >> >> >> exynos_dp_bind calls of_drm_find_panel().
> >> >> >>
> >> >> >> Signed-off-by: Ajay Kumar <ajaykumar.rs at samsung.com>
> >> >> >> ---
> >> >> >> Changes since V1:
> >> >> >>       Added platform_driver_unregister(&exynos_dp_panel_driver) to
> >> >> >>       exynos_drm_platform_remove as per Jingoo Han's correction
> >> >> >>
> >> >> >>  drivers/gpu/drm/exynos/exynos_drm_drv.c |   15 +++++++++++++++
> >> >> >>  drivers/gpu/drm/exynos/exynos_drm_drv.h |    1 +
> >> >> >>  2 files changed, 16 insertions(+)
> >> >> >>
> >> >> >> diff --git a/drivers/gpu/drm/exynos/exynos_drm_drv.c b/drivers/gpu/drm/exynos/exynos_drm_drv.c
> >> >> >> index 1d653f8..2db7f67 100644
> >> >> >> --- a/drivers/gpu/drm/exynos/exynos_drm_drv.c
> >> >> >> +++ b/drivers/gpu/drm/exynos/exynos_drm_drv.c
> >> >> >> @@ -530,12 +530,23 @@ static int exynos_drm_platform_probe(struct platform_device *pdev)
> >> >> >>               goto err_unregister_ipp_drv;
> >> >> >>  #endif
> >> >> >>
> >> >> >> +#ifdef CONFIG_DRM_PANEL_EXYNOS_DP
> >> >> >> +     ret = platform_driver_register(&exynos_dp_panel_driver);
> >> >> >> +     if (ret < 0)
> >> >> >> +             goto err_unregister_dp_panel;
> >> >> >> +#endif
> >> >> >
> >> >> > No, this is not how you're supposed to use DRM panel drivers. The idea
> >> >> > is that you write a standalone driver for a given panel.
> >> >> >
> >> >> > What you do here has a number of problems. For one it's a driver that's
> >> >> > tightly coupled to Exynos SoCs. But if I have a different SoC that uses
> >> >> > the same panel I want to be able to use the same driver, and not have to
> >> >> > rewrite the driver for my SoC.
> >> >> >
> >> >> > Another problem is that you're assuming here that the driver is built in
> >> >> > and it will break if you try to build either Exynos DRM or the panel
> >> >> > driver as a module. This is perhaps nothing you care about right now,
> >> >> > but eventually people will want to ship a single kernel that can run on
> >> >> > a number of SoCs. But if we keep adding things like this, that kernel
> >> >> > will keep growing in size until it no longer fits in any kind of memory.
> >> >> >
> >> >> > Thierry
> >> >>
> >> >> I completely agree with you in this!
> >> >>
> >> >> Yes, this is not acceptable, but I want to know an "acceptable"
> >> >> workaround for the situation below:
> >> >> I register the driver using module_init().
> >> >> And, exynos_drm gets probed much before the panel driver probe happens.
> >> >> So, the panel driver hasn't probed yet, but exynos_dp via exynos_drm
> >> >> tries to call
> >> >> "of_drm_find_panel" which always returns NULL.
> >> >
> >> > That's a situation that your driver needs to be able to deal with. The
> >> > driver registration order doesn't matter one bit. It may happen to work
> >> > most of the time, but as soon as one of the resources that your panel
> >> > driver needs isn't there when the panel is probed, then it won't be
> >> > registered and of_drm_find_panel() will still return NULL.
> >> >
> >> > Usually the right thing to do in that case would be to return (and
> >> > propagate) -EPROBE_DEFER so that your driver's probe is deferred and
> >> > retried when other drivers have been probed. That way it should
> >> > eventually get a non-NULL panel.
> >>
> >> So I just gave this (drm_panel + probe deferring) a shot on exynos,
> >> and correctly reacting to -EPROBE_DEFER postpones DP initialization by
> >> approximately 1.5 second. Is there a good way to handle that? As it
> >> stands, this isn't usable.
> >
> > How much is 1.5 seconds compared to the overall boot time of the device?
> 
> 1.5s is 15-20% of my boot time (if you count the boot time from
> firmware start to login prompt, otherwise it's more). Note that on
> other platforms, we've seen this take as much as 5 or 6s, but for the
> exynos case it is "only" 1.5s.
> 
> > What exactly is causing this 1.5 second delay?
> 
> A regulator isn't ready, and then drm_panel returns defer. Then the
> whole drm driver init is deferred.

That is the correct way to do this currently. It's unfortunate that it
causes such long delays.

> > This really is a fundamental issue with deferred probing and the issue
> > has come up several times in the past. A couple of possible solutions
> > have been proposed, with the latest being here[0] I think. That ended in
> > a bit of a debacle, unfortunately, but on of the outcomes was that a lot
> > of the ordering problems could be fixed by using phandle references to
> > track dependencies. I'm not aware of anyone working on that right now,
> > presumably because everyone is busy getting features merged rather than
> > optimizing boot speed.
> 
> Yes, I don't believe boot time ordering will ever happen upstream,

Maybe explicit boot time ordering won't happen upstream, but I don't
think anyone would object if you posted patches that decrease boot time
significantly, provided that they're still a generic solution. Tracking
dependencies parse from a device tree sounds like an option that would
be acceptable.

> but then the current implementation with EPROBE_DEFER isn't usable either.

"Isn't usable" is pretty strong. It does get you a working panel, right?
It's perhaps not optimized, but it gives you something that's guaranteed
to work (unless you have an actual bug somewhere), and something that
works in the general case, for everyone.

> Any ideas? ATM it seems like the only way out is to just write my
> own dt format for the panel and ignore drm_panel.

Well, since probe order is currently dependent on device instantiation
order, one possibility would be to rearrange the DT nodes in a way that
dependencies are listed first. So in your case I'd expect the regulator
provider to be listed first, followed by the panel and finally the "DRM"
device. Providing none of them trigger deferred probing themselves they
should probe in exactly that order.

Of course there are no guarantees that the probing order will stay that
way forever and you will only be solving this problem for one device at
a time rather than in general (which is the primary reason for deferred
probing in the first place).

One other possibility I guess would be not to rely on deferred probing
but to enable DRM/KMS to hotplug panels. Theoretically this should be
possible already but it comes at a cost. If I'm not mistaken that's what
Boris implemented for the Atmel HLCDC driver. However if the panel isn't
available at DRM driver ->load() time, then it relies on the hotplugging
mechanism to detect the panel later on, which can itself introduce more
delays. I suspect this could be improved in various ways, but one of the
more complicated issues is that it causes problems when you want to use
the framebuffer console. The issue is that the fbdev helper code is set
up very early and if no outputs with valid modes are found it will
default to a 1024x768 resolution for the framebuffer console. At the
same time it will prevent any outputs that are hotplugged later on to
increase the resolution again. This works fine for monitors which
usually support a number of different modes (1024x768 being standard
enough that probably every monitor in existence supports it). But for
panels this won't work, since they usually support only a single mode.

There are a number of ways that this could be solved I guess. One would
be to defer fbdev creation itself until all outputs have been detected.
This could prove difficult because it may not always be possible to
determine which outputs to wait for. Another solution would be to allow
the framebuffer console to be resized, but I've been told in the past
that it's impractical if not impossible.

Of course if you don't need a framebuffer console (you could use kmscon
or none at all), then things should work for you if you simply consider
the panel as hotpluggable monitor.

Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/dri-devel/attachments/20140821/aa7bdfa1/attachment.sig>