Re: i915 observations

[Mateusz Poszwa <old4%o2.pl@localhost>]

On Thu, Jul 21, 2022 at 09:29:36PM +0200, Tobias Nygren wrote:
> > Eventu?lly, the image tends to converge to the intended one (except
> > maybe for the clock itself, which tends to lag behind all the time).
> 
> I believe the effect you see is framebuffer contents written to
> the CPU cache but not flushed out to memory. This is why the artefacts
> lazily repair themselves when background jobs that cause frequent cache
> line misses are run. (Note how the artefacts are always 1 pixel high,
> 64-bytes wide and seemingly randomly positioned.)

Why, yes, that sounds like a probable cause. I remember that the
repairs were quicker when the system was under load, but I initially
thought it was due to the copious output being displayed. Your
explanation fits my observation that the repairs are still quick
even if the terminal running the demanding command is hidden, or
when the output is redirected to /dev/null. My conjecture was
that the problem is connected with some power-saving states.