[...]

quoted

What about the pm_clk_* API which was built for tracking clocks
associated with devices for runtime PM.

IOW, you could pm_clk_add(pg->pmc->dev, pg->clks[i]) and then your
_enable_clocks() would become pm_clk_suspend() an dyour
_disable_clocks() would become pm_clk_resume().

Very interesting, I was not aware of this.

quoted

I might not be following the mapping between PMC and PGs though so not
sure pg->pmc->dev is the right struct device, but you get the idea.

Yes, so this will not work here as-is, because the pmc->dev is common to
all pm-domains (it is the device that creates all the pm-domains). So to
make this work, I would need to create a device for each pm-domain and
add the clocks to that.

I see that this works very well for normal drivers, but it does not feel
so natural for pm-domains where we don't have a device struct today. By
the way, the rockchip pm-domains implementation is very much in the same
boat as tegra, where there are multiple clocks per pm-domain and it is
handled by a simple list. So I am not sure if you think that we should
be turning all pm-domains registered by pm_genpd_init() into a device
and then we can make use of these pm_clk_XXXX() APIs?

I have implemented the generic clk APIs that Ulf and I discussed for
handling multiple clocks, but if we think that this is a better way,
then I will hold off for now.

I think Kevin has a point that we already have PM clocks to build upon.
Could we perhaps try to extend that API instead to suite this needs as well?

I do realize that it will make this patchset more complicated. As I
stated earlier, this was just an idea I had, so to be clear I won't
hold back an ack for this patchset, if you decide to deal with this in
separate "improvement" step.

Kind regards
Uffe