[Intel-gfx] [TERRIBLE PATCH] Re: [regression?] i915 generating wakeups even when idle

[Andrew Lutomirski]

On Thu, Dec 9, 2010 at 12:20 PM, Jesse Barnes <jbarnes at virtuousgeek.org> wrote:
> On Wed, 8 Dec 2010 13:33:20 -0500
> Andrew Lutomirski <luto at mit.edu> wrote:
>
>> On Wed, Dec 8, 2010 at 1:17 PM, Andrew Lutomirski <luto at mit.edu> wrote:
>> > On Tue, Dec 7, 2010 at 5:13 PM, Chris Wilson <chris at chris-wilson.co.uk> wrote:
>> >> On Tue, 7 Dec 2010 16:31:24 -0500, Andrew Lutomirski <luto at mit.edu> wrote:
>> >>> On Tue, Dec 7, 2010 at 4:03 PM, Andrew Lutomirski <luto at mit.edu> wrote:
>> >>> > Hi all-
>> >>> >
>> >>> > Somewhere between Fedora 13 (with 2.6.35, I think) and Fedora 14 +
>> >>> > 2.6.36.1, i915 started generating exactly 50 interrupts per second
>> >>> > (suspiciously equal to my refresh rate) when X is running.  I have the
>> >>> > Xorg driver 2.12.0 (specifically
>> >>> > xorg-x11-drv-intel-2.12.0-6.fc14.1.x86_64).  When I switch to a text
>> >>> > console but leave X running, the interrupts stop.
>> >>> >
>> >>> > Any ideas what to look at?
>> >>>
>> >>> Quitting compiz fixes it.  Suspending compiz also fixes it.
>> >>
>> >> So it is the vblank interrupt. The vblank interrupt is get alive for a few
>> >> seconds after the last use. If it keeps going, then either the system is
>> >> as idle as you believe or we lost track of the last user and forget to
>> >> switch off the interrupt.
>> >>
>> >> drm.debug=0xf (echo 0xf > /sys/module/drm/parameters/debug) will have the
>> >> gory details of who/when triggers the vblank interrupt.
>> >> -Chris
>> >
>> > It's the seconds on the clock.  That causes activity once per second,
>> > which looks like this:
>> >
>>
>> [...]
>>
>> >
>> > Maybe that "several seconds" (5, according to the source) timer is way
>> > too long.  Is there any reason that drm_vblank_put doesn't turn off
>> > interrupts immediately (or, at the latest, on the very next vblank
>> > interrupt)?  After all, preventing deep sleep whenever there is
>> > display activity every five seconds seems like a waste of power.
>>
>> This patch fixes it.  It's obviously not ready for prime time, but if
>> you're OK with the idea I can fix it up and submit it.
>>
>> Signed-off-by: Andy Lutomirski <luto at mit.edu>
>>
>> Diff --git a/drivers/gpu/drm/drm_irq.c b/drivers/gpu/drm/drm_irq.c
>> index 9d3a503..49eca3f 100644
>> --- a/drivers/gpu/drm/drm_irq.c
>> +++ b/drivers/gpu/drm/drm_irq.c
>> @@ -471,7 +471,7 @@ void drm_vblank_put(struct drm_device *dev, int crtc)
>>
>>       /* Last user schedules interrupt disable */
>>       if (atomic_dec_and_test(&dev->vblank_refcount[crtc]))
>> -             mod_timer(&dev->vblank_disable_timer, jiffies + 5*DRM_HZ);
>> +             mod_timer(&dev->vblank_disable_timer, jiffies);
>>  }
>>  EXPORT_SYMBOL(drm_vblank_put);
>
> This will just move the problem around a bit; 5s is arguably too long
> to wait before disabling interrupts, but having a second hand or
> blinking : in your clock is the real issue here.  Why don't you disable
> that instead?

I did.

But I might also scroll a webpage every second or so (or have a
webpage loaded at some point that updates itself every second) and I
see no reason to prevent the system from going into its maximum sleep
state in between updates.

IOW, I think we should optimize for mostly-idle machines in addition
to completely-idle machines.

--Andy

>
> --
> Jesse Barnes, Intel Open Source Technology Center
>