[Intel-gfx] [PATCH 3/4] drm/i915: Fix random aux transactions failures.

[Daniel Vetter]

On Wed, Oct 21, 2015 at 09:18:06AM +0200, Daniel Vetter wrote:
> On Wed, Oct 21, 2015 at 10:28:53AM -0700, Rodrigo Vivi wrote:
> > Mainly aux communications on sink_crc
> > were failing a lot randomly on recent platforms.
> > The first solution was to try to use intel_dp_dpcd_read_wake, but then
> > it was suggested to move retries to drm level.
> > 
> > Since drm level was already taking care of retries and didn't want
> > to through random retries on that level the second solution was to
> > put the retries at aux_transfer layer what was nacked.
> > 
> > So I realized we had so many retries in different places and
> > started to organize that a bit. During this organization I noticed
> > that we weren't handing at all the case were the message size was
> > zeroed. And this was exactly the case that was affecting sink_crc.
> > 
> > Also we weren't respect BSPec who says this size message = 0 or > 20
> > are forbidden.
> > 
> > It is a fact that we still have no clue why we are getting this
> > forbidden value there. But anyway we need to handle that for now
> > so we return -EBUSY and drm level takes care of the retries that
> > are already in place.
> > 
> > Cc: Jani Nikula <jani.nikula at intel.com>
> > Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> > Signed-off-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > ---
> >  drivers/gpu/drm/i915/intel_dp.c | 11 +++++++++++
> >  1 file changed, 11 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/i915/intel_dp.c b/drivers/gpu/drm/i915/intel_dp.c
> > index aa3d8f6..80850d6 100644
> > --- a/drivers/gpu/drm/i915/intel_dp.c
> > +++ b/drivers/gpu/drm/i915/intel_dp.c
> > @@ -911,6 +911,17 @@ done:
> >  	/* Unload any bytes sent back from the other side */
> >  	recv_bytes = ((status & DP_AUX_CH_CTL_MESSAGE_SIZE_MASK) >>
> >  		      DP_AUX_CH_CTL_MESSAGE_SIZE_SHIFT);
> > +
> > +	/*
> > +	 * By BSpec: "Message sizes of 0 or >20 are not allowed."
> > +	 * We have no idea of what happened so we return -EBUSY so
> > +	 * drm layer takes care for the necessary retries.
> > +	 */
> > +	if (recv_bytes == 0 || recv_bytes > 20) {
> > +		ret = -EBUSY;
> > +		goto out;
> > +	}
> 
> Hm, this should be caught be the dp aux helper library. Both callers for
> ->transfer should check for this and reject with -EINVAL (since such a
> transaction is simply not allowed by dp aux). In the case of
> drm_dp_i2c_do_msg maybe even with a WARN_ON since the i2c logic should
> split things up correctly.

Meh, totally misread what's going on here, this is from the hardware. How
does this even happen? Do you get some kind of garbage value? Should we
maybe clear this register field first? It certainly would explain a lot of
our random dp aux retry fun ...
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch