Hello, On Friday 04 March 2016 23:04:46 Laurent Pinchart wrote: > On Friday 04 March 2016 10:24:10 Alan Stern wrote: > > On Fri, 4 Mar 2016, Ulf Hansson wrote: > >> On 3 March 2016 at 21:16, Laurent Pinchart wrote: > >>> The pm_runtime_force_suspend() and pm_runtime_force_resume() helpers > >>> are designed to help driver being RPM-centric by offering an easy way to > >>> manager runtime PM state during system suspend and resume. The first > >>> function will force the device into runtime suspend at system suspend > >>> time, while the second one will perform the reverse operation at system > >>> resume time. > >>> > >>> However, the pm_runtime_force_resume() really forces resume, regarding > >>> of whether the device was running or already suspended before the call > >>> to pm_runtime_force_suspend(). This results in devices being runtime > >>> resumed at system resume time when they shouldn't. > >>> > >>> Fix this by recording whether the device has been forcefully suspended > >>> in pm_runtime_force_suspend() and condition resume in > >>> pm_runtime_force_resume() to that state. > >>> > >>> All current users of pm_runtime_force_resume() call the function > >>> uncontionally in their system resume handler (some actually set it as > >>> the resume handler), all after calling pm_runtime_force_suspend() at > >>> system suspend time. The change in behaviour should thus be safe. > >>> > >>> Signed-off-by: Laurent Pinchart > >>> <laurent.pinchart+rene...@ideasonboard.com> > >>> > >>> @@ -1475,6 +1476,7 @@ int pm_runtime_force_suspend(struct device *dev) > >>> goto err; > >>> > >>> pm_runtime_set_suspended(dev); > >>> + dev->power.is_force_suspended = true; > >>> return 0; > >>> err: > >>> pm_runtime_enable(dev); > >>> @@ -1510,6 +1515,7 @@ int pm_runtime_force_resume(struct device *dev) > >>> if (ret) > >>> goto out; > >>> > >>> + dev->power.is_force_suspended = false; > >>> pm_runtime_set_active(dev); > >>> pm_runtime_mark_last_busy(dev); > >>> out: > > > > Setting a bitflag is not SMP-safe. When you write to one of the > > runtime-PM bits under dev->power, it is necessary to hold > > dev->power.lock. > > > >> Overall I have no objections to this change, as I think it's improving > >> the behaviour! > >> > >> What I was thinking though, but it might be a bit controversial. :-)... > >> Instead of relying on whether we actually forced runtime suspend > >> earlier, why couldn't we instead check the runtime PM usage count of > >> the device? > >> > >> Only when it's greater than zero, we shall do the forced resume of the > >> device, otherwise just re-enable runtime PM. > >> > >> This would have the affect of leaving devices in runtime suspend, > >> until they really needs to be used again. It would thus decrease the > >> total system PM resume time. > >> > >> Do you think this could work? > > > > If you do this then there would be no need for is_force_suspended. It > > seems like a good idea to me. > > I agree, that's a better idea. Drivers shouldn't call > pm_runtime_force_resume() if they haven't called pm_runtime_force_suspend(), > so checking the PM use count should be fine. I'll modify the patch, test it > and resubmit.
I gave it an unfortunately unsuccessful try. The problem I ran into is that device_prepare() calls pm_runtime_get_noresume() calls pm_runtime_get_noresume(), with the corresponding pm_runtime_put() call being performed in device_complete(). The device power usage_count is thus always non-zero in the system resume handler, so I can't base the decision on that. I also noticed that pm_genpd_prepare() runtime-resumes the device (when the power domain is in the GPD_STATE_ACTIVE state). I don't know why that is, but it means that in practice my device gets runtime-resumed when suspending the system while it could stay runtime-suspended in practice. -- Regards, Laurent Pinchart