On Mon, 4 Oct 2021 at 17:57, Dmitry Osipenko <digetx@xxxxxxxxx> wrote: > > 04.10.2021 14:01, Ulf Hansson пишет: > > On Fri, 1 Oct 2021 at 21:00, Dmitry Osipenko <digetx@xxxxxxxxx> wrote: > >> > >> 01.10.2021 17:55, Ulf Hansson пишет: > >>> On Fri, 1 Oct 2021 at 16:29, Dmitry Osipenko <digetx@xxxxxxxxx> wrote: > >>>> > >>>> 01.10.2021 16:39, Ulf Hansson пишет: > >>>>> On Mon, 27 Sept 2021 at 00:42, Dmitry Osipenko <digetx@xxxxxxxxx> wrote: > >>>>>> > >>>>>> Add runtime power management and support generic power domains. > >>>>>> > >>>>>> Tested-by: Peter Geis <pgwipeout@xxxxxxxxx> # Ouya T30 > >>>>>> Tested-by: Paul Fertser <fercerpav@xxxxxxxxx> # PAZ00 T20 > >>>>>> Tested-by: Nicolas Chauvet <kwizart@xxxxxxxxx> # PAZ00 T20 and TK1 T124 > >>>>>> Tested-by: Matt Merhar <mattmerhar@xxxxxxxxxxxxxx> # Ouya T30 > >>>>>> Signed-off-by: Dmitry Osipenko <digetx@xxxxxxxxx> > >>>>>> --- > >>>>>> drivers/gpu/drm/tegra/gr2d.c | 155 +++++++++++++++++++++++++++++++++-- > >>>>> > >>>>> [...] > >>>>> > >>>>>> static int gr2d_remove(struct platform_device *pdev) > >>>>>> @@ -259,15 +312,101 @@ static int gr2d_remove(struct platform_device *pdev) > >>>>>> return err; > >>>>>> } > >>>>>> > >>>>>> + pm_runtime_dont_use_autosuspend(&pdev->dev); > >>>>>> + pm_runtime_disable(&pdev->dev); > >>>>> > >>>>> There is no guarantee that the ->runtime_suspend() has been invoked > >>>>> here, which means that clock may be left prepared/enabled beyond this > >>>>> point. > >>>>> > >>>>> I suggest you call pm_runtime_force_suspend(), instead of > >>>>> pm_runtime_disable(), to make sure that gets done. > >>>> > >>>> The pm_runtime_disable() performs the final synchronization, please see [1]. > >>>> > >>>> [1] > >>>> https://elixir.bootlin.com/linux/v5.15-rc3/source/drivers/base/power/runtime.c#L1412 > >>> > >>> pm_runtime_disable() end up calling _pm_runtime_barrier(), which calls > >>> cancel_work_sync() if dev->power.request_pending has been set. > >>> > >>> If the work that was punted to the pm_wq in rpm_idle() has not been > >>> started yet, we end up just canceling it. In other words, there are no > >>> guarantees it runs to completion. > >> > >> You're right. Although, in a case of this particular patch, the syncing > >> is actually implicitly done by pm_runtime_dont_use_autosuspend(). > >> > >> But for drivers which don't use auto-suspend, there is no sync. This > >> looks like a disaster, it's a very common pattern for drivers to > >> 'put+disable'. > >> > >>> Moreover, use space may have bumped the usage count via sysfs for the > >>> device (pm_runtime_forbid()) to keep the device runtime resumed. > >> > >> Right, this is also a disaster in a case of driver removal. > >> > >>>> Calling pm_runtime_force_suspend() isn't correct because each 'enable' > >>>> must have the corresponding 'disable'. Hence there is no problem here. > >>> > >>> pm_runtime_force_suspend() calls pm_runtime_disable(), so I think that > >>> should be fine. No? > >> > >> [adding Rafael] > >> > >> Rafael, could you please explain how drivers are supposed to properly > >> suspend and disable RPM to cut off power and reset state that was > >> altered by the driver's resume callback? What we're missing? Is Ulf's > >> suggestion acceptable? > >> > >> The RPM state of a device is getting reset on driver's removal, hence > >> all refcounts that were bumped by the rpm-resume callback of the device > >> driver will be screwed up if device is kept resumed after removal. I > >> just verified that it's true in practice. > > > > Note that, what makes the Tegra drivers a bit special is that they are > > always built with CONFIG_PM being set (selected from the "SoC" > > Kconfig). > > > > Therefore, pm_runtime_force_suspend() can work for some of these > > cases. Using this, would potentially avoid the driver from having to > > runtime resume the device in ->remove(), according to the below > > generic sequence, which is used in many drivers. > > > > pm_runtime_get_sync() > > clk_disable_unprepare() (+ additional things to turn off the device) > > pm_runtime_disable() > > pm_runtime_put_noidle() > > It's not a problem to change this patchset. The problem is that if > you'll grep mainline for 'pm_runtime_disable', you will find that there > are a lot of drivers in a potential trouble. Let's start by fixing this patchset, please - then we can consider what to do with the other cases separately. > > I'm proposing that we should change pm_runtime_disable() to perform the > syncing with this oneliner: > > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c > index ec94049442b9..5c9f28165824 100644 > --- a/drivers/base/power/runtime.c > +++ b/drivers/base/power/runtime.c > @@ -1380,6 +1380,8 @@ EXPORT_SYMBOL_GPL(pm_runtime_barrier); > */ > void __pm_runtime_disable(struct device *dev, bool check_resume) > { > + flush_work(&dev->power.work); > + What about the latency this may introduce? I am not sure that is acceptable here!? > spin_lock_irq(&dev->power.lock); > > if (dev->power.disable_depth > 0) { > > Objections? > > The sysfs rpm-forbid is a separate problem and it's less troublesome > since it requires root privileges. It's also not something that > userspace touches casually. For now I don't know what could be done > about it. As I said, the common method to address this problem is to run the following sequence: pm_runtime_get_sync() "power off the device" pm_runtime_disable() pm_runtime_put_noidle() This works even if user space, via sysfs, has triggered a call to pm_runtime_forbid(). Or doesn't it? If you don't like it, pm_runtime_force_suspend() should work too, at least for your cases, I believe. Kind regards Uffe