Re: [PULL] drm-intel-next

[Chris Wilson <chris@chris-wilson.co.uk>]

On Wed, Oct 22, 2014 at 09:09:43AM +1000, Dave Airlie wrote:
> On 21 October 2014 23:38, Daniel Vetter <daniel.vetter@ffwll.ch> wrote:
> > Hi Dave,
> >
> > drm-intel-next-2014-10-03:
> > - first batch of skl stage 1 enabling
> > - fixes from Rodrigo to the PSR, fbc and sink crc code
> > - kerneldoc for the frontbuffer tracking code, runtime pm code and the basic
> >   interrupt enable/disable functions
> > - smaller stuff all over
> > drm-intel-next-2014-09-19:
> > - bunch more i830M fixes from Ville
> > - full ppgtt now again enabled by default
> > - more ppgtt fixes from Michel Thierry and Chris Wilson
> > - plane config work from Gustavo Padovan
> > - spinlock clarifications
> > - piles of smaller improvements all over, as usual
> 
> Chris made some noises about PPGTT being broken somewhere on irc last week,
> 
> Chris, did you figure that out?

Nope. All full-ppgtt platforms (ivb/byt/hsw) suffer from spontaneously
loosing track of the right page directories and end up executing
garbage. It correlates with load, so frequently makes igt (and a few
tests in particular) die, along with piglit, webgl conformance,
benchmarking and even eventually light loads on composited desktops.

I've made the pd uncached, added copious flushing, forced switch_mm
every time, added noops, cacheline alignment, srm, forced the execution
of batches to be synchronous, and yet IPEHR != *ACTHD. The register and
command state looks valid. The gpu resets ok and runs fine until the
error strikes again.

[So I need to test whether switch_mm(aliasing_ppgtt) on every batch fails
as well, and whether i915.enable_rc6=0 masks it. There is the worrying
bits in bspec that talk of non-RCS as being part of the render context
state, but only the RCS pd registers are shown in the context diagrams.
I guess I should inspect the context state and see if I can spot the other
registers. If context restore (and with rc6 that could happen at any time)
switched the pd on the other rings, that would be a nice snafu.]

I would suggest that full-ppgtt be disabled unless someone else has had
better luck finding a hsd or figuring out the missing magic.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre