[Intel-gfx] [PATCH] Revert "drm/i915: Implement Link Rate fallback on Link training failure"

[Imre Deak]

On Thu, Mar 02, 2017 at 12:07:05PM +0200, Imre Deak wrote:
> On Wed, Mar 01, 2017 at 07:56:26PM +0200, Ville Syrjälä wrote:
> > On Wed, Mar 01, 2017 at 07:35:15PM +0200, Jani Nikula wrote:
> > > On Wed, 01 Mar 2017, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > > > On Wed, Mar 01, 2017 at 06:17:49PM +0100, Daniel Vetter wrote:
> > > >> This reverts commit 233ce881dd91fb13eb6b09deefae33168e6ead4c.
> > > >> 
> > > >> I assumed it's ok, but really should have double-checked - CI caught
> > > >> tons of fail :(
> > > 
> > > Considering the velocity of drm-tip, I think any CI results for patches
> > > have a rather limited best before date. The patch should've been resent
> > > and gone through testing again before merging.
> > > 
> > > > For the record, the failure comes from the error message in
> > > > intel_dp_get_link_train_fallback_values() as take the fallback path. As
> > > > userspace is informed, we don't need an *ERROR* at that point.
> > > >
> > > > The really interesting question is why we are seeing link-training
> > > > failures in CI at all, and whether igt should be checking and reporting
> > > > link-status=BAD.
> > > 
> > > It's possible (I didn't check the logs) this pertains to the failure
> > > mode I've sometimes seen, where clock recovery fails, but as we continue
> > > with channel equalization anyway (without this patch), everything
> > > succeeds there. At worst we need to root cause and fix that issue
> > > first. :(
> > 
> > The skl case seems pretty clear. We register DP for both port A and port
> > E even though we should register it only for port E (I think). They
> > both end up both using AUX A and so we think the same sink is connected
> > to both, and then we try to enable port A which fail for obvious reasons.
> > 
> > The culprit is init_vbt_defaults() which always sets .supports_dp=true
> > for port A unless later overridden by the VBT. In this case the VBT has
> > no port A, so we leave the .supports_dp flag set. So presumably we
> > should just nuke this stuff from init_vbt_defaults().
> > 
> > IIRC this was discussed at some point between Imre and Paulo, but I
> > can't remember what the conclusion was, or if in fact there was one.
> > 
> > The ilk failure case is a lot less clear. It's one of those cases
> > where the sink just keeps requesting the same vswing/preemph all
> > the time. I've seen it sometime in the past, but I've never been
> > able to figure out what has caused it to happen with any specific
> > sink.
> 
> Looks like starting with:
> [    3.621706] [drm:drm_dp_i2c_do_msg] I2C nack (result=0, size=1
> [    3.622794] [drm:drm_dp_i2c_do_msg] I2C nack (result=0, size=1
> [    3.623836] [drm:drm_dp_i2c_do_msg] I2C nack (result=0, size=1
> [    3.624903] [drm:drm_dp_i2c_do_msg] I2C nack (result=0, size=1
> [    3.625983] [drm:drm_dp_i2c_do_msg] I2C nack (result=0, size=1
> 
> and so failing to read out EDID. As a result we'll use the default 1024x768
> which fails. In the working case we use the native 1920x1200 which succeeds.
> Could be that 1024x768 didn't work even before?

Hm no. The EDID read actually succeeds. The difference in mode is due to
a different monitor being attached. In any case I saw the monitor with
1024x768 mode already in earlier logs too where the modeset succeeded, so the
problem is somewhere else.

--Imre