Re: [PATCH v3 20/24] pmdomain: core: Default to use of_genpd_sync_state() for genpd providers

[Saravana Kannan <saravanak@google.com>]

On Wed, Sep 24, 2025 at 4:41 AM Jon Hunter <jonathanh@nvidia.com> wrote:
>
> Hi Ulf,
>
> On 03/09/2025 13:33, Jon Hunter wrote:
>
> ...
>
> >>> Following this change I am seeing the following warning on our Tegra194
> >>> devices ...
> >>>
> >>>    WARNING KERN tegra-bpmp bpmp: sync_state() pending due to
> >>> 17000000.gpu
> >>>    WARNING KERN tegra-bpmp bpmp: sync_state() pending due to 3960000.cec
> >>>    WARNING KERN tegra-bpmp bpmp: sync_state() pending due to
> >>> 15380000.nvjpg
> >>>    WARNING KERN tegra-bpmp bpmp: sync_state() pending due to
> >>> 154c0000.nvenc
> >>>    WARNING KERN tegra-bpmp bpmp: sync_state() pending due to
> >>> 15a80000.nvenc
> >>>
> >>> Per your change [0], the 'GENPD_FLAG_NO_SYNC_STATE' is set for Tegra
> >>> and so should Tegra be using of_genpd_sync_state() by default?
> >>
> >> This is a different power-domain provider (bpmp) in
> >> drivers/firmware/tegra/bpmp.c and
> >> drivers/pmdomain/tegra/powergate-bpmp.c.
> >>
> >> For the bpmp we don't need GENPD_FLAG_NO_SYNC_STATE, as the
> >> power-domain provider is described along with the
> >> "nvidia,tegra186-bpmp" compatible string. In the other case
> >> (drivers/soc/tegra/pmc.c) the "core-domain" and "powergates" are
> >> described through child-nodes, while ->sync_state() is managed by the
> >> parent-device-node.
> >>
> >> In the bpmp case there is no ->sync_state() callback assigned, which
> >> means genpd decides to assign a default one.
> >>
> >> The reason for the warnings above is because we are still waiting for
> >> those devices to be probed, hence the ->sync_state() callback is still
> >> waiting to be invoked. Enforcing ->sync_state() callback to be invoked
> >> can be done via user-space if that is needed.
> >>
> >> Did that make sense?
> >
> > Sorry for the delay, I was on vacation. Yes makes sense and drivers for
> > some of the above drivers are not yet upstreamed to mainline and so this
> > would be expected for now.
>
>
> I have been doing more testing and do see a lot of "tegra-bpmp bpmp:
> sync_state() pending due to" on our platforms for basically are driver
> that is built as a module.

It being "built as a module" is not reason enough for this warning to
happen though. One of the main points of fw_devlink is for things to
work just as well with modules.

In this particular system, do you never plan to load the modules? Or
is the module load just missing this timeout by a few seconds or
something?

If these can be turned off, why not turn these off using the sysfs
file or the timeout commandline option to turn them off? You are
burning power by leaving these on. A warning seems appropriate to me.

-Saravana