[Intel-gfx] [PATCH] drm/i915: add cherryview specfic forcewake in execlists_elsp_write

[Jesse Barnes]

On Tue, 09 Sep 2014 21:45:08 +0530
Deepak S <deepak.s at intel.com> wrote:

> 
> On Monday 08 September 2014 08:10 PM, Daniel Vetter wrote:
> > On Mon, Sep 08, 2014 at 05:14:23PM +0300, Ville Syrjälä wrote:
> >> On Mon, Sep 08, 2014 at 05:02:43PM +0300, Ville Syrjälä wrote:
> >>> On Tue, Sep 09, 2014 at 07:14:16PM +0530,
> >>> deepak.s at linux.intel.com wrote:
> >>>> From: Deepak S <deepak.s at linux.intel.com>
> >>>>
> >>>> In chv, we have two power wells Render & Media. We need to use
> >>>> corresponsing forcewake count. If we dont follow this we are
> >>>> getting error "*ERROR*: Timed out waiting for forcewake old ack
> >>>> to clear" due to multiple entry into __vlv_force_wake_get.
> >>>>
> >>>> Signed-off-by: Deepak S <deepak.s at linux.intel.com>
> >>>> ---
> >>>>   drivers/gpu/drm/i915/intel_lrc.c | 29
> >>>> +++++++++++++++++++++++++---- 1 file changed, 25 insertions(+),
> >>>> 4 deletions(-)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/i915/intel_lrc.c
> >>>> b/drivers/gpu/drm/i915/intel_lrc.c index bd1b28d..bafd38b 100644
> >>>> --- a/drivers/gpu/drm/i915/intel_lrc.c
> >>>> +++ b/drivers/gpu/drm/i915/intel_lrc.c
> >>>> @@ -300,8 +300,18 @@ static void execlists_elsp_write(struct
> >>>> intel_engine_cs *ring,
> >>>>   	 * Instead, we do the runtime_pm_get/put when
> >>>> creating/destroying requests. */
> >>>>   	spin_lock_irqsave(&dev_priv->uncore.lock, flags);
> >>>> -	if (dev_priv->uncore.forcewake_count++ == 0)
> >>>> -		dev_priv->uncore.funcs.force_wake_get(dev_priv,
> >>>> FORCEWAKE_ALL);
> >>>> +	if (IS_CHERRYVIEW(dev_priv->dev)) {
> >>>> +		if (dev_priv->uncore.fw_rendercount++ == 0)
> >>>> +
> >>>> dev_priv->uncore.funcs.force_wake_get(dev_priv,
> >>>> +
> >>>> FORCEWAKE_RENDER);
> >>>> +		if (dev_priv->uncore.fw_mediacount++ == 0)
> >>>> +
> >>>> dev_priv->uncore.funcs.force_wake_get(dev_priv,
> >>>> +
> >>>> FORCEWAKE_MEDIA);
> >>> This will wake both wells. Is that needed or should we just pick
> >>> one based on the ring?
> >> Also unlike the comment says runtime_pm_get() can't sleep since
> >> someone must already be holding a reference, othwewise we surely
> >> can't go writing any registers. So in theory we should be able to
> >> call gen6_gt_force_wake_get() here, but maybe that would trigger a
> >> might_sleep() warning. the current force wake code duplication
> >> (esp. outside intel_uncore.c) is rather unfortunate and I'd like
> >> to see it killed off. Maybe we just need to pull the rpm get/put
> >> outside gen6_gt_force_wake_get()? I never really liked hiding it
> >> there anyway.
> > Yeah this is just broken design. And if you look at the other wheel
> > to track outstanding gpu work (requests) then it's not needed at
> > all.
> >
> > But I'm not sure what's the priority of the "rework execlists to use
> > requests" task is and when (if ever that will happen). Jesse is the
> > arbiter for this stuff anyway, so adding him.
> > -Daniel
> 
> hmm , agreed do we have a reworked execlist? The reason why added
> this, on chv when i enable execlist, due to incorrect forcewake count
> is causing multiple entries to forcewake_get resulting in "*ERROR*:
> Timed out waiting for forcewake old ack to clear" "and Hang.

I'm hoping we can get execlists reworked on top of the request/seqno
stuff shortly after it lands, but I don't think that's a reason to
block this fix, since Chris is still busy fixing up the request
changes.

Jesse