On Wed, Dec 22, 2021 at 12:54:06PM +0000, Robin Murphy wrote:

Certainly RK3399 (and I guess RK3288 too) suffers from the same
priority-inversion issue of being unaware that VD_GPU needs power before
PD_GPU can be successfully turned on to probe the GPU to claim and enable
the regulator (via "mali-supply" for DVFS purposes) that needed to be on in
the first place. Currently all the boards are bodging around this with
"regulator-always-on" (e.g. commit 06b2818678d9).

Does the SoC actually support the supply being completely off in
operation?  A lot of devices want everything powered even if idle during
full operation since keeping voltage differentials smaller makes it a
lot easier to design to avoid leakage current type issues at the points
where the different power domains connect.