[Intel-gfx] [PATCH] Revert "drm/i915/bdw: BDW Software Turbo"

[Jesse Barnes]

On Mon, 29 Sep 2014 09:52:38 -0700
Rodrigo Vivi <rodrigo.vivi at gmail.com> wrote:

> On Mon, Sep 29, 2014 at 9:38 AM, Daniel Vetter <daniel at ffwll.ch> wrote:
> > On Mon, Sep 29, 2014 at 08:48:53AM -0700, Jesse Barnes wrote:
> >> On Mon, 29 Sep 2014 15:11:51 +0200
> >> Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> >>
> >> > This reverts commit c76bb61a71083b2d90504cc6d0dda2047c5d63ca.
> >> >
> >> > It's apparently too broken so that Rodrigo submitted a patch to add a
> >> > config option for it. Given that the design is also ... suboptimal and
> >> > that I've only merged this to get lead engineers and managers off my
> >> > back for one second let's just revert this.
> >> >
> >> > /me puts on combat gear again
> >> >
> >> > It was worth a shot ...
> >>
> >> I thought we had a fix for the runtime PM issue this created?  And
> >> Rodrigo's fix is just a simple "only BDW needs this" patch, so I guess
> >> I don't see the big issue?
> 
> Well, actually the issue is to stay with high busted gt frequency even on idle.
> Or this or an incompatibility with the actual rps interface...
> 
> So my patch was actually a protect a feture under development with parameter.
> 
> >>
> >> Or is there another bug you didn't mention in the s-o-b section you're
> >> worried about?
> 
> The only issue I know is:
> References: https://bugs.freedesktop.org/show_bug.cgi?id=77869

Ah yeah, misread it thinking it was accidentally being enabled on
non-bdw too or something.

But yeah I expect turbo to be broken even with your patch, since we
won't ramp up/down the freq (or at least not properly).  The only thing
Daisy's patch addresses is making sure we at least try to ramp if we're
getting regular page flips.  The more complete fix (which should
address the test failure too) would be to peroidically poll the hw busy
state and adjust the freq based on that, independent of the flip
frequency.  That should allow GPGPU apps, offlien media encode/decode,
and these tests to work as they should, but it's a little more work.
We need to make sure the timer is shut down when we go into RC6 and
only fires frequently when the GPU is busy.

-- 
Jesse Barnes, Intel Open Source Technology Center