[Intel-gfx] [PATCH] drm/i915: Apply the GTT write flush for all !llc machines

[Chris Wilson]

Quoting Ville Syrjälä (2017-09-07 19:23:27)
> On Thu, Sep 07, 2017 at 03:31:08PM +0100, Chris Wilson wrote:
> > We also see the delayed GTT write issue on i915g/i915gm, so let's
> > presume that it is a universal problem for all llc machines, and that we
> > just haven't yet noticed on g33, gen4 and gen5 machines.
> > 
> > Testcase: igt/gem_mmap_gtt/coherency # i915gm
> 
> That fails on my i945gm as well. mmio read before the clfush does 
> seem to cure it. Doing the mmio after the clflush still fails. I
> wonder if there's some prefetching going on or why exactly it
> behaves that way...

Hmm, I thought UC read was also a barrier for speculative prefetching.

> > References: https://bugs.freedesktop.org/show_bug.cgi?id=102577
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  drivers/gpu/drm/i915/i915_gem.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> > index 4dffebae5601..350b761b9e91 100644
> > --- a/drivers/gpu/drm/i915/i915_gem.c
> > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > @@ -694,7 +694,7 @@ flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains)
> >  
> >       switch (obj->base.write_domain) {
> >       case I915_GEM_DOMAIN_GTT:
> > -             if (INTEL_GEN(dev_priv) >= 6 && !HAS_LLC(dev_priv)) {
> > +             if (!HAS_LLC(dev_priv)) {
> >                       intel_runtime_pm_get(dev_priv);
> >                       spin_lock_irq(&dev_priv->uncore.lock);
> >                       POSTING_READ_FW(RING_ACTHD(dev_priv->engine[RCS]->mmio_base));
> 
> There's no documented register at that offset on gen2/3.
> Might be less risky to read something we know to actually exist.

Darn, those registers walking around.

i915_reg_t reg;

if (INTEL_GEN(dev_priv) >= 4)
	reg = RING_ACTHD(dev_priv->engine[RCS]->mmio_base);
else
	reg = ACTHD;
POSTING_READ_FW(reg);

Or we experiment with using RING_HEAD as that hasn't moved.
-Chris