Re: [PATCH] ACPI: PM: postpone bringing devices to D0 unless we need them

From: Rafael J. Wysocki
Date: Tue Jun 22 2021 - 15:06:45 EST


Hi,

On Tue, Jun 22, 2021 at 8:13 PM Dmitry Torokhov
<dmitry.torokhov@xxxxxxxxx> wrote:
>
> Hi Rafael,
>
> On Tue, Jun 22, 2021 at 03:40:05PM +0200, Rafael J. Wysocki wrote:
> > On Tue, Jun 22, 2021 at 12:20 AM Dmitry Torokhov
> > <dmitry.torokhov@xxxxxxxxx> wrote:
> > >
> > > Currently ACPI power domain brings devices into D0 state in the "resume
> > > early" phase. Normally this does not cause any issues, as powering up
> > > happens quickly. However there are peripherals that have certain timing
> > > requirements for powering on, for example some models of Elan
> > > touchscreens need 300msec after powering up/releasing reset line before
> > > they can accept commands from the host. Such devices will dominate
> > > the time spent in early resume phase and cause increase in overall
> > > resume time as we wait for early resume to complete before we can
> > > proceed to the normal resume stage.
> > >
> > > There are ways for a driver to indicate that it can tolerate device
> > > being in the low power mode and that it knows how to power the device
> > > back up when resuming, bit that requires changes to individual drivers
> > > that may not really care about details of ACPI controlled power
> > > management.
> > >
> > > This change attempts to solve this issue at ACPI power domain level, by
> > > postponing powering up device until we get to the normal resume stage,
> > > unless there is early resume handler defined for the device, or device
> > > does not declare any resume handlers, in which case we continue powering
> > > up such devices early. This allows us to shave off several hundred
> > > milliseconds of resume time on affected systems.
> > >
> > > Signed-off-by: Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>
> > > ---
> > > drivers/acpi/device_pm.c | 46 +++++++++++++++++++++++++++++++++++-----
> > > 1 file changed, 41 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/drivers/acpi/device_pm.c b/drivers/acpi/device_pm.c
> > > index 096153761ebc..00b412ccb2e0 100644
> > > --- a/drivers/acpi/device_pm.c
> > > +++ b/drivers/acpi/device_pm.c
> > > @@ -1131,17 +1131,52 @@ static int acpi_subsys_resume_noirq(struct device *dev)
> > > *
> > > * Use ACPI to put the given device into the full-power state and carry out the
> > > * generic early resume procedure for it during system transition into the
> > > - * working state.
> > > + * working state, but only do that if device either defines early resume
> > > + * handler, or does not define power operations at all. Otherwise powering up
> > > + * of the device is postponed to the normal resume phase.
> > > */
> > > static int acpi_subsys_resume_early(struct device *dev)
> > > {
> > > + const struct dev_pm_ops *pm = dev->driver ? dev->driver->pm : NULL;
> > > + struct acpi_device *adev = ACPI_COMPANION(dev);
> > > int ret;
> > >
> > > - if (dev_pm_skip_resume(dev))
> > > - return 0;
> >
> > The above doesn't need to be changed AFAICS.
>
> I was trying to have if string if/else if/else, but I can keep it as it
> was.
>
> >
> > > + if (dev_pm_skip_resume(dev)) {
> > > + ret = 0;
> > > + } else if (!pm || pm->resume_early) {
> >
> > This is rather tricky, but I don't see a particular reason why it wouldn't work.
> >
> > > + ret = acpi_dev_resume(dev);
> > > + if (!ret)
> > > + ret = pm_generic_resume_early(dev);
> > > + } else {
> > > + if (adev)
> > > + acpi_device_wakeup_disable(adev);
> >
> > This isn't necessary here.
>
> Just to confirm - you are saying that disabling the device as a wakeup
> source can be safely postponed till the normal resume stage?

Yes, it should be safe. Moreover, it may be unsafe to change the
ordering between acpi_dev_pm_full_power() and
acpi_device_wakeup_disable().

> I was trying to keep as much of the original behavior as possible and this is
> a part of acpi_dev_resume() that we are now postponing.

I would postpone the whole thing.

Thanks!