Hi Shivam,

Thanks for the v2. Hugo posted v2 of two series on Nov 13 that do the
same thing, so it would be good to work together on one of them:

- "Leave self-refresh before an ABM level update" [1] takes the panel
  out of PSR or Panel Replay around the DC update that carries the new
  level.
- "Commit panel_power_savings writes from the kernel" [2]. Its 2/5 is
  the same atomic check change as in this patch and David's 2/2 [3],
  and its 4/5 commits the write from the store.

The part this patch is missing is [1]. Once ABM changes no longer force
a modeset, the level goes out as a fast update while the panel may
still be in self-refresh. On DCN 2.1 that hangs DMUB: Hugo hit it on
Cezanne, and we've seen it on Renoir. On Strix (DCN 3.5), with Hugo's
v1 series but without [1], a write on an idle desktop left PSR stuck
and the ABM ramp stalled until some unrelated commit. The IGT run
doesn't catch this, since the panel has to already be in self-refresh
when the level changes.

Hugo's 4/5 also skips the commit when the level doesn't change (as
Harry asked), sets allow_modeset to false so a write never turns into
a full modeset, and takes a runtime PM reference.

Hamza, agreed on running this through promotion first. We're putting
Hugo's two series through this week's promotion testing in place of
this patch, since they include [1].

Could you test Hugo's two v2 series on your Rembrandt and review them?
That's DCN 3.1.2, where PSR is on by default, so it covers a case
neither Hugo (DCN 2.1) nor we (DCN 3.5) have. It would help to:

- let psr_state reach 6 before each write, then check that it leaves
  and comes back
- rerun amd_abm, which Hugo hasn't done on v2

Hugo, please Cc Shivam on v3.

[1] 
https://lore.kernel.org/r/[email protected]
[2] 
https://lore.kernel.org/r/[email protected]
[3] 
https://lore.kernel.org/amd-gfx/[email protected]/

Thanks,
Chenyu

Reply via email to