Re: [PATCH] drm/bridge: ti-sn65dsi86: Properly undo autosuspend

[Laurent Pinchart]

Hi Doug,

On Wed, Feb 23, 2022 at 07:43:27AM -0800, Doug Anderson wrote:
> On Tue, Feb 22, 2022 at 9:08 PM Laurent Pinchart wrote:
> > On Tue, Feb 22, 2022 at 11:44:54PM +0100, Linus Walleij wrote:
> > > On Tue, Feb 22, 2022 at 11:19 PM Douglas Anderson wrote:
> > > >
> > > > The PM Runtime docs say:
> > > >   Drivers in ->remove() callback should undo the runtime PM changes done
> > > >   in ->probe(). Usually this means calling pm_runtime_disable(),
> > > >   pm_runtime_dont_use_autosuspend() etc.
> > > >
> > > > We weren't doing that for autosuspend. Let's do it.
> > > >
> > > > Fixes: 9bede63127c6 ("drm/bridge: ti-sn65dsi86: Use pm_runtime autosuspend")
> > > > Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
> > >
> > > Hm. I know a few places in drivers where I don't do this :/
> >
> > It seems to be a very common problem indeed, I haven't seen any driver
> > yet that uses pm_runtime_dont_use_autosuspend(). We could play a game of
> > whack-a-mole, but we'll never win. Could this be solved in the runtime
> > PM framework instead ? pm_runtime_disable() could disable auto-suspend.
> > If there are legitimate use cases for disabling runtime PM temporarily
> > without disabling auto-suspend, then a new function designed
> > specifically for remove() that would take care of cleaning everything up
> > could be another option.
> 
> Yeah, it would be good. It's probably not a yak I have time to shave
> right now, though. :(

I don't insist on shaving that yak right now :-) This patch is fine.

> I _suspect_ that there are legitimate reasons we can't just magically
> do it in pm_runtime_disable(). If nothing else I believe there are
> legitimate code paths during normal operation that look like this:
> 
>   pm_runtime_disable();
>   do_something_that_needs_pm_runtime_disabled();
>   pm_runtime_enable();
> 
> Also: if it were really a simple problem to solve one would have
> thought that it would have been solved initially instead of adding
> documentation particularly mentioning
> pm_runtime_dont_use_autosuspend()

I'm not sure, look at how long it took for us to get
pm_runtime_resume_and_get(). The problem has been considered for years
as a non-issue by the runtime PM developers. It feels like the API is
developed without considering its users.

> How about a middle ground, though: we could add a devm function that
> does all the magic. Somewhat recently devm_pm_runtime_enable() was
> added. What if we add a variant for those that use autosuspend, like:
> 
> devm_pm_runtime_enable_with_autosuspend(dev, initial_delay)
> 
> That would:
> * pm_runtime_enable()
> * pm_runtime_set_autosuspend_delay()
> * pm_runtime_use_autosuspend()
> * Use devm_add_action_or_reset() to undo everything.
> 
> Assuming that the pm_runtime folks are OK with that, we could
> transition things over to the new function once it rolls into
> mainline.
> 
> So this doesn't magically fix all existing code but provides a path to
> make this more discoverable.

Sounds like a good idea. I wonder if this could be handled by
devm_pm_runtime_enable() actually. If a driver calls
devm_pm_runtime_enable() and then enables auto-suspend, can't the
runtime PM core reasonably expect that auto-suspend should be disabled
at .remove() time ? The pm_runtime_disable_action() function could
disable auto-suspend unconditionally (assuming
pm_runtime_use_autosuspend() and pm_runtime_dont_use_autosuspend() don't
need to be balanced, if they do, then I'll just go cry in a corner).

-- 
Regards,

Laurent Pinchart