On ke, 2015-06-10 at 15:21 +0100, Chris Wilson wrote:
> On Wed, Jun 10, 2015 at 05:07:46PM +0300, Imre Deak wrote:
> > On ti, 2015-06-09 at 11:21 +0300, Jani Nikula wrote:
> > > On Mon, 08 Jun 2015, Imre Deak <imre.deak@xxxxxxxxx> wrote:
> > > > By running igt/store_dword_loop_render on BXT we can hit a coherency
> > > > problem where the seqno written at GPU command completion time is not
> > > > seen by the CPU. This results in __i915_wait_request seeing the stale
> > > > seqno and not completing the request (not considering the lost
> > > > interrupt/GPU reset mechanism). I also verified that this isn't a case
> > > > of a lost interrupt, or that the command didn't complete somehow: when
> > > > the coherency issue occured I read the seqno via an uncached GTT mapping
> > > > too. While the cached version of the seqno still showed the stale value
> > > > the one read via the uncached mapping was the correct one.
> > > >
> > > > Work around this issue by clflushing the corresponding CPU cacheline
> > > > following any store of the seqno and preceding any reading of it. When
> > > > reading it do this only when the caller expects a coherent view.
> > > >
> > > > Testcase: igt/store_dword_loop_render
> > > > Signed-off-by: Imre Deak <imre.deak@xxxxxxxxx>
> > > > ---
> > > > drivers/gpu/drm/i915/intel_lrc.c | 17 +++++++++++++++++
> > > > drivers/gpu/drm/i915/intel_ringbuffer.h | 7 +++++++
> > > > 2 files changed, 24 insertions(+)
> > > >
> > > > diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> > > > index 9f5485d..88bc5525 100644
> > > > --- a/drivers/gpu/drm/i915/intel_lrc.c
> > > > +++ b/drivers/gpu/drm/i915/intel_lrc.c
> > > > @@ -1288,12 +1288,29 @@ static int gen8_emit_flush_render(struct intel_ringbuffer *ringbuf,
> > > >
> > > > static u32 gen8_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
> > > > {
> > > > + /*
> > > > + * On BXT-A1 there is a coherency issue whereby the MI_STORE_DATA_IMM
> > > > + * storing the completed request's seqno occasionally doesn't
> > > > + * invalidate the CPU cache. Work around this by clflushing the
> > > > + * corresponding cacheline whenever the caller wants the coherency to
> > > > + * be guaranteed. Note that this cacheline is known to be
> > > > + * clean at this point, since we only write it in gen8_set_seqno(),
> > > > + * where we also do a clflush after the write. So this clflush in
> > > > + * practice becomes an invalidate operation.
>
> Did you compare and contrast with the gen6+ w/a? A clflush may just work
> out quicker considering that the posting read would involve a spinlock
> and fw dance.
Actually, I did, but only saw that it only works, didn't benchmark it.
I'd also think that clflush would be faster, since it's only a cache
invalidate at this point. But I will compare the two things now.
> -Chris
>
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
