[Intel-gfx] [PATCH v3] drm/i915/dp: Give up link training clock recovery after 10 failed attempts

[Rodrigo Vivi]

On Tue, Jul 17, 2018 at 06:05:51PM -0700, Nathan Ciobanu wrote:
> On Tue, Jul 17, 2018 at 03:21:17PM -0700, Dhinakaran Pandiyan wrote:
> > On Mon, 2018-07-16 at 16:51 -0700, Marc Herbert wrote:
> > > > 
> > > > > 
> > > > > I think the bug is with this infinite loop which is at the mercy
> > > > > of an external device
> > > > > and in my case I have this MST hub which appears to be DP 1.2
> > > > > that triggers the
> > > > > infinite loop case. We have to limit the number of iterations and
> > > > > DP 1.4 spec happened to fix this issue. I'm a newbie in this area
> > > > > but in this case
> > > > > shouldn't we apply the same counter to all <= "DP 1.4" devices?
> > > > ok, the infinite loop situation is really bad... but I don't
> > > > believe
> > > > we are going with the right fix...
> > > > and a good indication is that your fix is for DP-1.4 while your bug
> > > > is seeing on DP-1.2
> > > I thought the only reason the infinite loop fix isn't in 1.2 is just
> > > because there's
> > > no 1.2.1 spec... (plus the naive assumption that buggy sinks don't
> > > exist)
> > > 
> > Irrespective of whether the sink is DP1.2 or 1.4, if there are sinks
> > out there that keep toggling between two values there should be an
> > overall limit to how many times this loop gets executed. Even if this
> > isn't right fix for the issue at hand, the loop has to break.
> > 
> > Then there's the question of what the limit should be. We could use the
> > DP1.4 limit as a reference and apply it widely. But there's a problem
> > here, we have 4 vswing values and 4 pre-emphasis values. If the sink
> > progressively changes one variable at a time, we'll need at least 16
> > iterations. Nathan's patch here will prematurely error out and that
> > doesn't sound reasonable to impose on non DP1.4 sinks.
> 
> I think it would be a max of 13 iterations since one of the vswing
> values will be max_vswing and the loop will terminate at that point
> without trying the other 3 preemph values for that voltage, correct?

I was talking to DK yesterday and he suggested that we should go with
a huge number for DP_1.2 and with the spec for DP_1.4. According to DK
80 was covering all combinations few times.

So you get your patch and create some max_cr_tries = dp_1.4 ? spec : 80;

or something like that.

I think I like that because infinite loop situation here is so bad
and we should avoid even if it was something else that got really wrong.

Thanks,
Rodrigo.

> 
> -Nathan
> > 
> > -DK
> > 
> > 
> > > 
> > > 
> > > _______________________________________________
> > > Intel-gfx mailing list
> > > Intel-gfx at lists.freedesktop.org
> > > https://lists.freedesktop.org/mailman/listinfo/intel-gfx
> > 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx