[Intel-gfx] [PATCH v2 1/3] drm/i915/hda: Add audio component stub

[Daniel Vetter]

On Tue, May 17, 2016 at 3:20 PM, Takashi Iwai <tiwai at suse.de> wrote:
>> Ok, I looked at patch 3, and that indeed would lead to trouble without
>> patch 1. But the real trouble is the unconditional wait_completion in
>> there - blocking for another driver to complete loading from a driver load
>> function is a no-go. The correct way to do this is to bail out with
>> EPROBE_DEFER if not all the parts are available there. Also, throw out
>> that request_module.
>>
>> By bailing out with EPROBE_DEFER you avoid deadlocks, and the driver core
>> also knows what's going on. Which is incidentally what's used to
>> implicitly order suspend/resume. Driver core will restart your probe as
>> soon as some new devices/drivers have registers (assuming that hopefully
>> then you're unblocking), but if you're unlucky your driver can go through
>> that loop a few times.
>>
>> But that was just a very quick look, we definitely shouldn't need any
>> wait_completion in driver load to handle cross-module depencies.
>
> Yeah, I admit that wait_completion() is hackish.  OTOH, EPROBE_DEFER
> doesn't work in the case of HD-audio because we want to give up
> binding and continue without i915 but only with onboard audio, instead
> of endlessly reprobing for the never-appearing component.  The i915
> binding is no hard dependency; i.e. it isn't (always) mandatory, and
> EPROBE_DEFER can't handle such a fallback, AFAIK.
>
> If there is a good way to deal with it, please let me know.  I'd love
> to rewrite to a cleaner way.

The only way to deal with that is to split the driver into two, and
hotplug them individually. Fundamentally any approach where you need
to know whether i915 shows up or not and act accordingly is just plain
flawed, there's no way around it. That's also why EPROBE_DEFER doesn't
bother dealing with it.

Imo if you have the sound side of hdmi/dp audio, then just
EPROBE_DEFER until i915 is loaded (assuming it's not disabled through
nomodeset or Kconfig). If it's not there then continue without it (and
without hdmi/dp audio ofc). Trying to be clever just means we need to
hand roll things all over the place all the time. We have some code on
earlier platforms for runtime clock adjustements (on ironlake) in
i915.ko, and I really don't want that kind of hacks any more.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch