[Intel-gfx] [PATCH] BUG-REPORT: snd-hda: hacked-together EPROBE_DEFER support

[Daniel Vetter]

On Mon, Jun 26, 2017 at 7:47 PM, Takashi Iwai <tiwai at suse.de> wrote:
> On Mon, 26 Jun 2017 18:16:30 +0200,
> Daniel Vetter wrote:
>>
>> On Wed, Jun 21, 2017 at 05:30:10PM +0200, Takashi Iwai wrote:
>> > On Wed, 21 Jun 2017 17:23:57 +0200,
>> > Chris Wilson wrote:
>> > >
>> > > Quoting Daniel Vetter (2017-06-21 16:08:54)
>> > > > So back when the i915 power well support landed in
>> > > >
>> > > > commit 99a2008d0b32d72dfc2a54e7be1eb698dd2e3bd6
>> > > > Author: Wang Xingchao <xingchao.wang at linux.intel.com>
>> > > > Date:   Thu May 30 22:07:10 2013 +0800
>> > > >
>> > > >     ALSA: hda - Add power-welll support for haswell HDA
>> > > >
>> > > > the logic to handle the cross-module depencies was hand-rolled using a
>> > > > async work item, and that just doesn't work.
>> > > >
>> > > > The correct way to handle cross-module deps is either:
>> > > > - request_module + failing when the other module isn't there
>> > > >
>> > > > OR
>> > > >
>> > > > - failing the module load with EPROBE_DEFER.
>> > > >
>> > > > You can't mix them, if you do then the entire load path just
>> > > > busy-spins blowing through cpu cycles forever with no way to stop
>> > > > this.
>> > > >
>> > > > snd-hda-intel does mix it, because the hda codec drivers are loaded
>> > > > using request_module, but the i915 depency is handled using
>> > > > PROBE_DEFER (or well, should be, but I haven't found any code at all).
>> > > > This is a major pain when trying to debug i915 load failures.
>> > > >
>> > > > This patch here is a horrible hackish attempt at somewhat correctly
>> > > > wriing EPROBE_DEFER through. Stuff that's missing:
>> > > > - Check all the other places where load errors are conveniently
>> > > >   dropped on the floor.
>> > > > - Also fix up the firmware_cb path.
>> > > > - Drop the debug noise I've left in to make it clear this isn't
>> > > >   anything for merging.
>> > >
>> > > This tames "hdaudio hdaudioC0D0: Unable to bind the codec" which was
>> > > continuously spewing previously, and now the system is usable again.
>> >
>> > Could you give a failing scenario?  I'm not opposing to the suggested
>> > solution, we need to fix the mess in anyway, but I just would like to
>> > know how to trigger the problem easily.
>>
>> Disable i915 loading e.g. with i915.modeset=0. Watch how snd-hda*
>> collective blow through 100% of the cpu time spewing into dmesg (and make
>> the system completely unuseable for kernel work because you can't find
>> your own debug printk anymore).
>
> Ah, that's the case we discussed in the past.  We know that it's
> problematic for component binding, but we're ignoring this scenario
> because it's supposed to be no real use-case but only for some
> temporary workarounds.
>
> We had some bigger-hammer patchset, but it didn't justify for the
> further development of the reasoning above.
>
>> This is on a snb, where we don't even need the cross-module stuff ... But
>> I think it goes sideways in other cases too, if you simply build but don't
>> load i915. So every time an i915 breaks module load things become real
>> painful.
>
> Even on SNB, we still need i915 for the HDMI/DP ELD notification.  The
> hardware inquiry over HD-audio verb was so unstable, so we rather take
> a path directly inquiring to the gfx driver.

Ah right, forgot about that.

>> Unfortunately the patch is a bit too big for our fixup branch in drm-tip,
>> so plan B would be to stop building snd-hda (which will make the intel
>> audio team unhappy, but mea culpa if they don't fix this mess).
>
> OK, let me think and take a look for older patchset, too.

Yeah would be great if we can somehow address this, preferrably using
EPROBE_DEFER or something else that's standard. At least the component
stuff really doesn't work without wiring EPROBE_DEFER through.

And if that patch series requires some soaking I think I could easily
add it to our drm-tip CI branch for testing (and making our developers
lifes easier), we already pull in your -next/-fixes trees anyway.
Pulling in another topic branch would be simple.

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch