Message-ID: <CAPDyKFpSY+FeKh7ocjQ_nGNZA5+3tWAL8e7ZNKXKNFP-yoiu_g@mail.gmail.com>
Date: Fri, 5 Sep 2025 16:27:27 +0200
From: Ulf Hansson <ulf.hansson@...aro.org>
To: Nicolas Frattaroli <nicolas.frattaroli@...labora.com>
Cc: linux-pm@...r.kernel.org, linux-kernel@...r.kernel.org, 
	kernel@...labora.com, linux-rockchip@...ts.infradead.org, 
	Cristian Ciocaltea <cristian.ciocaltea@...labora.com>, 
	Sebastian Reichel <sebastian.reichel@...labora.com>, Heiko Stuebner <heiko@...ech.de>
Subject: Re: [PATCH] pmdomian: core: don't unset stay_on during sync_state

[...]

>
> Okay so I believe I have found the root cause of the regression. UFS is
> innocent, disabling UFS just happens to avoid it due to how the timing of
> things works out.
>
> The real issue is that the NPU power domains on the RK3576, which are
> currently unused, have an undeclared dependency on vdd_npu_s0.
>
> Declaring this dependency with a `domain-supply` and adding the
> necessary flag in the rockchip PD controller to use it does not solve
> he problem. This is because the rockchip PD controller cannot acquire
> those supplies during probe, as they're not available yet and their
> availability depends on the PD controller finishing probe.
>
> That's why it acquires them in the PD enable callback, but the NPU
> PDs are never enabled because they're unused.
>
> This worked fine when unused PDs were still turned off quite early, as
> this meant they were turned off before regulators. Now the unused
> regulators are turned off before turning off the unused PDs happens.

I see, thanks for sharing these details. What a mess.

>
> I don't really see an easy way to fix this with a patch that's fit for
> an rc cycle. We can't request the regulator early or even just add a
> device link, as the regulator is not around yet.

Right, I will work on a patch or two that allows rockchip
power-domains to opt-out from genpds new behavior and to keep using
the old one.

I think we prefer to do it like this (should be quite a limited amount
of code and okay for an rc), rather than reverting for everyone.

>
> Marking vdd_npu_s0 as always-on would be abusing DT to work around a
> Linux kernel shortcoming, which is a no-no.
>
> What we need is either a way to register with pmdomain core that
> certain PDs need a late init for additional supplies, which is then
> called before any of the unused regulator power off functionality is
> invoked by the regulator core.
>
> Any ideas?

Yes :-)

I would suggest implementing an auxiliary driver, along with the
rockchip_pm_domain_driver. The main job for the auxiliary driver would
be to get the regulator in its ->probe() - and if it fails because the
regulator isn't available yet, it should keep trying by returning
-EPROBE_DEFER. See more about the auxiliary bus/device/driver in
include/linux/auxiliary_bus.h and module_auxiliary_driver().

Moreover, when the rockchip_pm_domain_driver probes, it becomes
responsible for pre-parsing the OF nodes for the domain-supply DT
property, for each of the specified power-domains. If it finds a
domain-supply, it should register an auxiliary device that corresponds
to that particular power-domain. This can be done by using
platform-data that is shared with the auxiliary device/driver. See
devm_auxiliary_device_create().

Furthermore we would need some kind of synchronization mechanism
between the rockchip_pm_domain_driver and the auxiliary driver, to
manage the regulator get/enable/disable. I think that should be rather
easy to work out.

Do you think this can work?

Kind regards
Uffe

Powered by blists - more mailing lists