Em Sex, 2015-11-13 às 23:26 +0200, Ville Syrjälä escreveu:
> On Fri, Nov 13, 2015 at 11:20:19PM +0200, Ville Syrjälä wrote:
> > On Fri, Nov 13, 2015 at 09:03:43PM +0000, Chris Wilson wrote:
> > > On Fri, Nov 13, 2015 at 05:53:41PM -0200, Paulo Zanoni wrote:
> > > > Instead of waiting for 50ms, just wait until the next vblank,
> > > > since
> > > > it's the minimum requirement.
> > > >
> > > > This moves PC7 residency on my specific BDW machine running
> > > > Cinnamon
> > > > from 60-70% to 84-89%. Without FBC, I get 20-25%. I'm using a
> > > > 3200x1800 eDP panel. Notice: this was the case when the patch
> > > > was
> > > > originally proposed, the order of the FBC patches changed since
> > > > then,
> > > > so the actual numbers might be slightly different now.
> > > >
> > > > v2:
> > > > - Rebase after changing the patch order.
> > > > - Update the commit message.
> > > >
> > > > Signed-off-by: Paulo Zanoni <paulo.r.zanoni@xxxxxxxxx>
> > > > ---
> > > > drivers/gpu/drm/i915/i915_drv.h | 2 +-
> > > > drivers/gpu/drm/i915/intel_fbc.c | 12 +++---------
> > > > 2 files changed, 4 insertions(+), 10 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/i915/i915_drv.h
> > > > b/drivers/gpu/drm/i915/i915_drv.h
> > > > index 9418bd5..ea08714 100644
> > > > --- a/drivers/gpu/drm/i915/i915_drv.h
> > > > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > > > @@ -919,9 +919,9 @@ struct i915_fbc {
> > > >
> > > > struct intel_fbc_work {
> > > > bool scheduled;
> > > > + u32 scheduled_vblank;
> > > > struct work_struct work;
> > > > struct drm_framebuffer *fb;
> > > > - unsigned long enable_jiffies;
> > > > } work;
> > > >
> > > > const char *no_fbc_reason;
> > > > diff --git a/drivers/gpu/drm/i915/intel_fbc.c
> > > > b/drivers/gpu/drm/i915/intel_fbc.c
> > > > index aa82075..72de8a1 100644
> > > > --- a/drivers/gpu/drm/i915/intel_fbc.c
> > > > +++ b/drivers/gpu/drm/i915/intel_fbc.c
> > > > @@ -391,7 +391,6 @@ static void intel_fbc_work_fn(struct
> > > > work_struct *__work)
> > > > container_of(__work, struct drm_i915_private,
> > > > fbc.work.work);
> > > > struct intel_fbc_work *work = &dev_priv->fbc.work;
> > > > struct intel_crtc *crtc = dev_priv->fbc.crtc;
> > > > - unsigned long delay_jiffies = msecs_to_jiffies(50);
> > > >
> > > > retry:
> > > > /* Delay the actual enabling to let pageflipping cease
> > > > and the
> > > > @@ -400,14 +399,9 @@ retry:
> > > > * vblank to pass after disabling the FBC before we
> > > > attempt
> > > > * to modify the control registers.
> > > > *
> > > > - * A more complicated solution would involve tracking
> > > > vblanks
> > > > - * following the termination of the page-flipping
> > > > sequence
> > > > - * and indeed performing the enable as a co-routine
> > > > and not
> > > > - * waiting synchronously upon the vblank.
> > > > - *
> > > > * WaFbcWaitForVBlankBeforeEnable:ilk,snb
> > > > */
> > > > - wait_remaining_ms_from_jiffies(work->enable_jiffies,
> > > > delay_jiffies);
> > > > + intel_wait_for_vblank(dev_priv->dev, crtc->pipe);
> > > >
> > > > mutex_lock(&dev_priv->fbc.lock);
> > > >
> > > > @@ -416,7 +410,7 @@ retry:
> > > > goto out;
> > > >
> > > > /* Were we delayed again while this function was
> > > > sleeping? */
> > > > - if (time_after(work->enable_jiffies + delay_jiffies,
> > > > jiffies)) {
> > > > + if (drm_crtc_vblank_get(&crtc->base) == work-
> > > > >scheduled_vblank) {
> > > > mutex_unlock(&dev_priv->fbc.lock);
> > > > goto retry;
> > > > }
> > > > @@ -449,7 +443,7 @@ static void
> > > > intel_fbc_schedule_activation(struct intel_crtc *crtc)
> > > > * jiffy count. */
> > > > work->fb = crtc->base.primary->fb;
> > > > work->scheduled = true;
> > > > - work->enable_jiffies = jiffies;
> > > > + work->scheduled_vblank = drm_crtc_vblank_count(&crtc-
> > > > >base);
> > >
> > > Isn't the frame counter only incrementing whilst the vblank IRQ
> > > is
> > > enabled? Ville?
> >
> > I see a "+ if (drm_crtc_vblank_get(" earlier.
>
> Hmm. Actually it's doing
> "drm_crtc_vblank_get(&crtc->base) == work->scheduled_vblank)"
> which looks rather like nonsense.
>
> Not sure what the intention here was...
Ouch. The intent was for that to be another call for
drm_crtc_vblank_count().
The code in discussion is completely based on the drm_wait_one_vblank()
code: call drm_vblank_count(), then call it again until it returns
something different. The difference is that we actually call
drm_wait_one_vblank() in the middle of the process, and that
scheduled_vblank may also be updated in the meantime, so so may have to
call drm_wait_one_vblank() again.
>
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
