Re: [PATCH 00/51] treewide: Switch to __pm_runtime_put_autosuspend()

[Ulf Hansson]

On Tue, 8 Oct 2024 at 20:25, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>
> On Tue, Oct 8, 2024 at 12:35 AM Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> >
> > On Tue, 8 Oct 2024 at 00:25, Laurent Pinchart
> > <laurent.pinchart@xxxxxxxxxxxxxxxx> wrote:
> > >
> > > Hi Ulf,
> > >
> > > On Tue, Oct 08, 2024 at 12:08:24AM +0200, Ulf Hansson wrote:
> > > > On Mon, 7 Oct 2024 at 20:49, Laurent Pinchart wrote:
> > > > > On Fri, Oct 04, 2024 at 04:38:36PM +0200, Ulf Hansson wrote:
> > > > > > On Fri, 4 Oct 2024 at 11:41, Sakari Ailus wrote:
> > > > > > >
> > > > > > > Hello everyone,
> > > > > > >
> > > > > > > This set will switch the users of pm_runtime_put_autosuspend() to
> > > > > > > __pm_runtime_put_autosuspend() while the former will soon be re-purposed
> > > > > > > to include a call to pm_runtime_mark_last_busy(). The two are almost
> > > > > > > always used together, apart from bugs which are likely common. Going
> > > > > > > forward, most new users should be using pm_runtime_put_autosuspend().
> > > > > > >
> > > > > > > Once this conversion is done and pm_runtime_put_autosuspend() re-purposed,
> > > > > > > I'll post another set to merge the calls to __pm_runtime_put_autosuspend()
> > > > > > > and pm_runtime_mark_last_busy().
> > > > > >
> > > > > > That sounds like it could cause a lot of churns.
> > > > > >
> > > > > > Why not add a new helper function that does the
> > > > > > pm_runtime_put_autosuspend() and the pm_runtime_mark_last_busy()
> > > > > > things? Then we can start moving users over to this new interface,
> > > > > > rather than having this intermediate step?
> > > > >
> > > > > I think the API would be nicer if we used the shortest and simplest
> > > > > function names for the most common use cases. Following
> > > > > pm_runtime_put_autosuspend() with pm_runtime_mark_last_busy() is that
> > > > > most common use case. That's why I like Sakari's approach of repurposing
> > > > > pm_runtime_put_autosuspend(), and introducing
> > > > > __pm_runtime_put_autosuspend() for the odd cases where
> > > > > pm_runtime_mark_last_busy() shouldn't be called.
> > > >
> > > > Okay, so the reason for this approach is because we couldn't find a
> > > > short and descriptive name that could be used in favor of
> > > > pm_runtime_put_autosuspend(). Let me throw some ideas at it and maybe
> > > > you like it - or not. :-)
> > >
> > > I like the idea at least :-)
> > >
> > > > I don't know what options you guys discussed, but to me the entire
> > > > "autosuspend"-suffix isn't really that necessary in my opinion. There
> > > > are more ways than calling pm_runtime_put_autosuspend() that triggers
> > > > us to use the RPM_AUTO flag for rpm_suspend(). For example, just
> > > > calling pm_runtime_put() has the similar effect.
> > >
> > > To be honest, I'm lost there. pm_runtime_put() calls
> > > __pm_runtime_idle(RPM_GET_PUT | RPM_ASYNC), while
> > > pm_runtime_put_autosuspend() calls __pm_runtime_suspend(RPM_GET_PUT |
> > > RPM_ASYNC | RPM_AUTO).
> >
> > __pm_runtime_idle() ends up calling rpm_idle(), which may call
> > rpm_suspend() - if it succeeds to idle the device. In that case, it
> > tags on the RPM_AUTO flag in the call to rpm_suspend(). Quite similar
> > to what is happening when calling pm_runtime_put_autosuspend().
>
> Right.
>
> For almost everybody, except for a small bunch of drivers that
> actually have a .runtime_idle() callback, pm_runtime_put() is
> literally equivalent to pm_runtime_put_autosuspend().
>
> So really the question is why anyone who doesn't provide a
> .runtime_idle() callback bothers with using this special
> pm_runtime_put_autosuspend() thing, which really means "do a
> runtime_put(), but skip my .runtime_idle() callback".

My guess is that it's in most cases a legacy pattern that is being
followed. Also note that rpm_idle() didn't "always" tag on the
RPM_AUTO flag, even if it's quite a while ago (2013) since we added
it.

Unless there is some actual optimization involved, as it also allows
us to skip calling rpm_idle() and go directly for rpm_suspend().

>
> > >
> > > >
> > > > Moreover, it's similar for pm_runtime_mark_last_busy(), it's called
> > > > during rpm_resume() too, for example. So why bother about having
> > > > "mark_last_busy" in the new name too.
> > > >
> > > > That said, my suggestion is simply "pm_runtime_put_suspend".
> > >
> > > Can we do even better, and make pm_runtime_put() to handle autosuspend
> > > automatically when autosuspend is enabled ?
> >
> > As stated above, this is already the case.
>
> What really is needed appears to be a combination of
> pm_runtime_mark_last_busy() with pm_runtime_put().

This makes sense to me too, but I don't think we should limit it to this.

Making pm_runtime_put_autosuspend (or if the name
"pm_runtime_put_suspend" is better?) to do the similar thing, is
probably a good idea too. At least in my opinion.

>
> Granted, pm_runtime_put() could do the pm_runtime_mark_last_busy()
> thing automatically if autosuspend is enabled and the only consequence
> of it might be delaying a suspend of the device until its autosuspend
> timer expires, which should not be a problem in the vast majority of
> cases.

Right.

I guess we should expect the *sync* variants to be used, if the timer
really needs to be overridden.

Kind regards
Uffe