Re: [PATCH v5 1/2] pm: runtime: Add new devm functions

From: Csókás Bence
Date: Thu Mar 27 2025 - 05:02:46 EST


Hi,

On 2025. 03. 26. 18:38, Rafael J. Wysocki wrote:
I said I didn't like it and I'm still not liking it.

You didn't really elaborate further, but now I'm glad I could understand your dislike.

The problem is that the primary role of pm_runtime_set_active() is to
prepare the device for enabling runtime PM, so in the majority of
cases it should be followed by pm_runtime_enable(). It is also not
always necessary to call pm_runtime_set_suspended() after disabling
runtime PM for a device, like when the device has been
runtime-suspended before disabling runtime PM for it. This is not
like releasing a resource that has been allocated and using devm for
it in the above way is at least questionable.

Now, there is a reason why calling pm_runtime_set_suspended() on a
device after disabling runtime PM for it is a good idea at all.
Namely, disabling runtime PM alone does not release the device's
suppliers or its parent, so if you want to release them after
disabling runtime PM for the device, you need to do something more.
I'm thinking that this is a mistake in the design of the runtime PM
core.

Well, this is the order in which the original driver worked before anyways. As a quick fix, would it work if we created a devm function that would pm_runtime_set_active(), immediately followed by pm_runtime_enable(), and on cleanup it would pm_runtime_set_suspended() followed by pm_runtime_disable_action() (i.e. pm_runtime_dont_use_autosuspend() and pm_runtime_disable())?

If there were functions like pm_runtime_enable_in_state() (taking an
additional state argument and acquiring all of the necessary
references on the parent and suppliers of the target device) and
pm_runtime_disable_and_forget() (that in addition to disabling runtime
PM would drop the references acquired by the former), then it would
make a perfect sense to provide a devm variant of
pm_runtime_enable_in_state() with the cleanup action pointing to
pm_runtime_disable_and_forget().

If this helps, I can do some work on providing
pm_runtime_enable_in_state() and pm_runtime_disable_and_forget() or
equivalent.

I mean sure, if that's something you want to work on, but it sounds like it would entail much more work, plus it wouldn't be easy to backport it to 6.14.y stable later.

Bence