[Intel-gfx] [PATCH] drm/i915: flush system agent TLBs on SNB so we can WC map the PTEs

[Daniel Vetter]

On Fri, Oct 12, 2012 at 03:54:50AM -0400, Dave Airlie wrote:
> On Thu, Oct 11, 2012 at 7:54 PM, Jesse Barnes <jbarnes at virtuousgeek.org> wrote:
> > On Thu, 11 Oct 2012 20:29:47 -0300
> > Paulo Zanoni <przanoni at gmail.com> wrote:
> >
> >> Hi
> >>
> >> 2012/10/11 Jesse Barnes <jbarnes at virtuousgeek.org>:
> >> > I've only lightly tested this so far, but the corruption seems to be
> >> > gone if I write the GFX_FLSH_CNTL reg after binding an object.  This
> >> > register should control the TLB for the system agent, which is what CPU
> >> > mapped objects will go through.
> >> >
> >> > Signed-off-by: Jesse Barnes <jbarnes at virtuousgeek.org>
> >>
> >> I'm not sure if this is the patch you asked me to test on IRC, but,
> >> well this is the patch I tested :)
> >>
> >> I tested it on HSW on top of dinq + some other patches. Booted the
> >> machine, ran mostly xfterm4 under Xfce, but I also ran firefox,
> >> gnome-shell and restarted X a few times. No GPU hangs so far. I'm
> >> testing this for about 3-4 hours.
> >>
> >> I do have to say that I could not find this register on the HSW
> >> documentation, but I also did no see any "Unclaimed write" messages
> >> containing the 101008 address... I'm a little confused.
> >>
> >> And the interesting thing: intel_gpu_tools is useless now. I keep
> >> getting "Couldn't map MMIO region: Resource temporarily unavailable".
> >> Do you see this too?
> >
> > No I haven't seen that, I ran some tests today with the tip of the tree
> > too...  Dunno why the resource files would return EBUSY?  Maybe because
> > a driver is bound?
> 
> More likely because of some WB/WC/UC collision. We don't like aliases.

Yep, you need to upgrade, latest i-g-t should handle the wc/uc split
correctly. The important testcase is gem_gtt_cpu_tlb, if that one works
this patch is good.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch