(In reply to Norbert Pfeiler from comment #115)
> I don’t think that’s it because I tried that with earlier kernel versions (I
> also hardcoded the dpcd in several places and that didn’t work). 
> I suspected it had to do with the ›dp_aux_ch timeout‹ and interrupts
> occurring around the time of link training. 
> 
> As I took a significant amount of time digging through the code I would like
> to get convinced about which commit fixed this.

You can use git to bisect what fixed it (just reverse the good/bad flagging).