On Thu, 19 Aug 2021 at 21:35, Dmitry Osipenko <digetx@xxxxxxxxx> wrote:
>
> 19.08.2021 16:07, Ulf Hansson пишет:
> > On Wed, 18 Aug 2021 at 17:43, Dmitry Osipenko <digetx@xxxxxxxxx> wrote:
> >>
> >> 18.08.2021 13:08, Ulf Hansson пишет:
> >>> On Wed, 18 Aug 2021 at 11:50, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> >>>>
> >>>> On 18-08-21, 11:41, Ulf Hansson wrote:
> >>>>> On Wed, 18 Aug 2021 at 11:14, Viresh Kumar <viresh.kumar@xxxxxxxxxx> wrote:
> >>>>>> What we need here is just configure. So something like this then:
> >>>>>>
> >>>>>> - genpd->get_performance_state()
> >>>>>> -> dev_pm_opp_get_current_opp() //New API
> >>>>>> -> dev_pm_genpd_set_performance_state(dev, current_opp->pstate);
> >>>>>>
> >>>>>> This can be done just once from probe() then.
> >>>>>
> >>>>> How would dev_pm_opp_get_current_opp() work? Do you have a suggestion?
> >>>>
> >>>> The opp core already has a way of finding current OPP, that's what
> >>>> Dmitry is trying to use here. It finds it using clk_get_rate(), if
> >>>> that is zero, it picks the lowest freq possible.
> >>>>
> >>>>> I am sure I understand the problem. When a device is getting probed,
> >>>>> it needs to consume power, how else can the corresponding driver
> >>>>> successfully probe it?
> >>>>
> >>>> Dmitry can answer that better, but a device doesn't necessarily need
> >>>> to consume energy in probe. It can consume bus clock, like APB we
> >>>> have, but the more energy consuming stuff can be left disabled until
> >>>> the time a user comes up. Probe will just end up registering the
> >>>> driver and initializing it.
> >>>
> >>> That's perfectly fine, as then it's likely that it won't vote for an
> >>> OPP, but can postpone that as well.
> >>>
> >>> Perhaps the problem is rather that the HW may already carry a non-zero
> >>> vote made from a bootloader. If the consumer driver tries to clear
> >>> that vote (calling dev_pm_opp_set_rate(dev, 0), for example), it would
> >>> still not lead to any updates of the performance state in genpd,
> >>> because genpd internally has initialized the performance-state to
> >>> zero.
> >>
> >> We don't need to discover internal SoC devices because we use
> >> device-tree on ARM. For most devices power isn't required at a probe
> >> time because probe function doesn't touch h/w at all, thus devices are
> >> left in suspended state after probe.
> >>
> >> We have three components comprising PM on Tegra:
> >>
> >> 1. Power gate
> >> 2. Clock state
> >> 3. Voltage state
> >>
> >> GENPD on/off represents the 'power gate'.
> >>
> >> Clock and reset are controlled by device drivers using clk and rst APIs.
> >>
> >> Voltage state is represented by GENPD's performance level.
> >>
> >> GENPD core assumes that at a first rpm-resume of a consumer device, its
> >> genpd_performance=0. Not true for Tegra because h/w of the device is
> >> preconfigured to a non-zero perf level initially, h/w may not support
> >> zero level at all.
> >
> > I think you may be misunderstanding genpd's behaviour around this, but
> > let me elaborate.
> >
> > In genpd_runtime_resume(), we try to restore the performance state for
> > the device that genpd_runtime_suspend() *may* have dropped earlier.
> > That means, if genpd_runtime_resume() is called prior
> > genpd_runtime_suspend() for the first time, it means that
> > genpd_runtime_resume() will *not* restore a performance state, but
> > instead just leave the performance state as is for the device (see
> > genpd_restore_performance_state()).
> >
> > In other words, a consumer driver may use the following sequence to
> > set an initial performance state for the device during ->probe():
> >
> > ...
> > rate = clk_get_rate()
> > dev_pm_opp_set_rate(rate)
> >
> > pm_runtime_enable()
> > pm_runtime_resume_and_get()
> > ...
> >
> > Note that, it's the consumer driver's responsibility to manage device
> > specific resources, in its ->runtime_suspend|resume() callbacks.
> > Typically that means dealing with clock gating/ungating, for example.
> >
> > In the other scenario where a consumer driver prefers to *not* call
> > pm_runtime_resume_and_get() in its ->probe(), because it doesn't need
> > to power on the device to complete probing, then we don't want to vote
> > for an OPP at all - and we also want the performance state for the
> > device in genpd to be set to zero. Correct?
>
> Yes
>
> > Is this the main problem you are trying to solve, because I think this
> > doesn't work out of the box as of today?
>
> The main problem is that the restored performance state is zero for the
> first genpd_runtime_resume(), while it's not zero from the h/w perspective.
This should not be a problem, but can be handled by the consumer driver.
genpd_runtime_resume() calls genpd_restore_performance_state() to
restore a performance state for the device. However, in the scenario
you describe, "gpd_data->rpm_pstate" is zero, which makes
genpd_restore_performance_state() to just leave the device's
performance state as is - it will *not* restore the performance state
to zero.
To make the consumer driver deal with this, it would need to call
dev_pm_opp_set_rate() from within its ->runtime_resume() callback.
>
> > There is another concern though, but perhaps it's not a problem after
> > all. Viresh told us that dev_pm_opp_set_rate() may turn on resources
> > like clock/regulators. That could certainly be problematic, in
> > particular if the device and its genpd have OPP tables associated with
> > it and the consumer driver wants to follow the above sequence in
> > probe.
>
> dev_pm_opp_set_rate() won't enable clocks and regulators, but it may
> change the clock rate and voltage. This is also platform/driver specific
> because it's up to OPP user how to configure OPP table. On Tegra we only
> assign clock to OPP table, regulators are unused.
>
> > Viresh, can you please chime in here and elaborate on some of the
> > magic happening behind dev_pm_opp_set_rate() API - is there a problem
> > here or not?
> >
> >>
> >> GENPD core assumes that consumer devices can work at any performance
> >> level. Not true for Tegra because voltage needs to be set in accordance
> >> to the clock rate before clock is enabled, otherwise h/w won't work
> >> properly, perhaps clock may be unstable or h/w won't be latching.
> >
> > Correct. Genpd relies on the callers to use the OPP framework if there
> > are constraints like you describe above.
> >
> > That said, it's not forbidden for a consumer driver to call
> > dev_pm_genpd_set_performance_state() directly, but then it better
> > knows exactly what it's doing.
> >
> >>
> >> Performance level should be set to 0 while device is suspended.
> >
> > Do you mean system suspend or runtime suspend? Or both?
>
> Runtime suspend.
Alright. So that's already taken care of for us in genpd_runtime_suspend().
Or perhaps you have discovered some problem with this?
>
> >> Performance level needs to be bumped on rpm-resume of a device in
> >> accordance to h/w state before hardware is enabled.
> >
> > Assuming there was a performance state set for the device when
> > genpd_runtime_suspend() was called, genpd_runtime_resume() will
> > restore that state according to the sequence you described.
>
> What do you think about adding API that will allow drivers to explicitly
> set the restored performance state of a power domain?
>
> Another option could be to change the GENPD core, making it to set the
> rpm_pstate when dev_pm_genpd_set_performance_state(dev) is invoked and
> device is rpm-suspended, instead of calling the
> genpd->set_performance_state callback.
>
> Then drivers will be able to sync the perf state at a probe time.
>
> What do you think?
I don't think it's needed, see my reply earlier above. However your
change touches another problem though, see below.
>
> diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
> index a934c679e6ce..cc15ab9eacc9 100644
> --- a/drivers/base/power/domain.c
> +++ b/drivers/base/power/domain.c
> @@ -435,7 +435,7 @@ static void genpd_restore_performance_state(struct
> device *dev,
> int dev_pm_genpd_set_performance_state(struct device *dev, unsigned int
> state)
> {
> struct generic_pm_domain *genpd;
> - int ret;
> + int ret = 0;
>
> genpd = dev_to_genpd_safe(dev);
> if (!genpd)
> @@ -446,7 +446,10 @@ int dev_pm_genpd_set_performance_state(struct
> device *dev, unsigned int state)
> return -EINVAL;
>
> genpd_lock(genpd);
> - ret = genpd_set_performance_state(dev, state);
> + if (pm_runtime_suspended(dev))
> + dev_gpd_data(dev)->rpm_pstate = state;
> + else
> + ret = genpd_set_performance_state(dev, state);
> genpd_unlock(genpd);
This doesn't work for all cases. For example, when a consumer driver
deploys runtime PM support in its ->probe() according to the below
sequence:
...
dev_pm_opp_set_rate(rate)
pm_runtime_get_noresume()
pm_runtime_set_active()
pm_runtime_enable()
...
pm_runtime_put()
...
We need to call genpd_set_performance_state() independently of whether
the device is runtime suspended or not.
Although, it actually seems like good idea to update
dev_gpd_data(dev)->rpm_pstate = state here, as to make sure
genpd_runtime_resume() doesn't restore an old/invalid value that was
saved while dropping the performance state vote for the device in
genpd_runtime_suspend() earlier.
Let me send a patch for this shortly, to close this window of a possible error.
>
> return ret;
>
>
Kind regards
Uffe
