[Intel-gfx] [RFC v3 1/3] PM/runtime: Add a new interface to get accounted time
Vincent Guittot
vincent.guittot at linaro.org
Thu Dec 20 08:44:26 UTC 2018
On Thu, 20 Dec 2018 at 09:16, Ulf Hansson <ulf.hansson at linaro.org> wrote:
>
> On Wed, 19 Dec 2018 at 17:52, Vincent Guittot
> <vincent.guittot at linaro.org> wrote:
> >
> > On Wed, 19 Dec 2018 at 17:36, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> > >
> > > On Wed, 19 Dec 2018 at 14:26, Vincent Guittot
> > > <vincent.guittot at linaro.org> wrote:
> > > >
> > > > On Wed, 19 Dec 2018 at 11:43, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> > > > >
> > > > > On Wed, 19 Dec 2018 at 11:34, Vincent Guittot
> > > > > <vincent.guittot at linaro.org> wrote:
> > > > > >
> > > > > > On Wed, 19 Dec 2018 at 11:21, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> > > > > > >
> > > >
> > > > > > > > > >
> > > > > > > > > > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> > > > > > > > > > index 7062469..6461469 100644
> > > > > > > > > > --- a/drivers/base/power/runtime.c
> > > > > > > > > > +++ b/drivers/base/power/runtime.c
> > > > > > > > > > @@ -88,6 +88,32 @@ static void __update_runtime_status(struct device *dev, enum rpm_status status)
> > > > > > > > > > dev->power.runtime_status = status;
> > > > > > > > > > }
> > > > > > > > > >
> > > > > > > > > > +u64 pm_runtime_accounted_time_get(struct device *dev, enum rpm_status status, bool update)
> > > > > > > > > > +{
> > > > > > > > > > + u64 now = ktime_to_ns(ktime_get());
> > > > > > > > >
> > > > > > > > > I think you should stay on jiffies here - and then switch to ktime in patch 3.
> > > > > > > > >
> > > > > > > > > > + u64 delta = 0, time = 0;
> > > > > > > > > > + unsigned long flags;
> > > > > > > > > > +
> > > > > > > > > > + spin_lock_irqsave(&dev->power.lock, flags);
> > > > > > > > > > +
> > > > > > > > > > + if (dev->power.disable_depth > 0)
> > > > > > > > > > + goto unlock;
> > > > > > > > > > +
> > > > > > > > > > + /* Add ongoing state if requested */
> > > > > > > > > > + if (update && dev->power.runtime_status == status)
> > > > > > > > > > + delta = now - dev->power.accounting_timestamp;
> > > > > > > > > > +
> > > > > > > > >
> > > > > > > > > Hmm. Do we really need to update the accounting timestamp? I would
> > > > > > > > > rather avoid it if possible.
> > > > > > > >
> > > > > > > > i915/drm uses this to track ongoing suspended state. In fact they are
> > > > > > > > mainly interested by this part
> > > > > > >
> > > > > > > Again, sorry I don't follow.
> > > > > >
> > > > > > In fact we don't update dev->power.accounting_timestamp but only use
> > > > > > it to get how much time has elapsed in the current state.
> > > > > >
> > > > > > >
> > > > > > > My suggested changes below, would do exactly that; track the ongoing
> > > > > > > suspended state.
> > > > > > >
> > > > > > > The user can call the function several times while the device remains
> > > > > > > RPM_SUSPENDED, and if needed the user could then compute the delta
> > > > > > > in-between the calls, for whatever reason that may be needed.
> > > > > >
> > > > > > So I'm not sure to catch your question:
> > > > > > Is your problem linked to status != RPM_SUSPENDED or the update
> > > > > > parameter that compute delta ?
> > > > >
> > > > > My intent was to keep things simple.
> > > > >
> > > > > 1. Only expose last suspended time, which means tracking the ongoing
> > > > > suspended state. In other words, you can also remove "enum rpm_status
> > > > > status" as the in-parameter to pm_runtime_accounted_time_get().
> > > >
> > > > Ok for this point if Rafael doesn't see any benefit of keeping the
> > > > generic interface
> > > >
> > > > > 2. Don't allow the user of pm_runtime_accounted_time_get() to update
> > > > > the current timestamp, in "dev->power.accounting_timestamp".
> > > >
> > > > But pm_runtime_accounted_time_get doesn't update
> > > > dev->power.accounting_timestamp, it only reads it to know when when
> > > > the last state transition happened
> > >
> > > I understand, sorry for not being clear enough.
> > >
> > > My point is, you must not update dev->power.suspended_time, without
> > > also setting a new value for dev->power.accounting_timestamp.
> > > Otherwise, the next time the core calls
> > > update_pm_runtime_accounting(), it will end up adding a wrong delta to
> > > dev->power.suspended_time.
> >
> > I fully agree on that and that's why dev->power.accounting_timestamp
> > is not and has never been modified
>
> Huh, I have miss-read your patch. What a mess, my apologies.
>
> >
> > >
> > > BTW, it seems like you have based this on top of some patch converting
> > > from jiffies to ktime, as I can't fine dev->power.suspended_time, but
> > > instead I have dev->power.suspended_jiffies.
> > >
> > > On Wed, 19 Dec 2018 at 14:26, Vincent Guittot
> > > <vincent.guittot at linaro.org> wrote:
> > > >
> > > > On Wed, 19 Dec 2018 at 11:43, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> > > > >
> > > > > On Wed, 19 Dec 2018 at 11:34, Vincent Guittot
> > > > > <vincent.guittot at linaro.org> wrote:
> > > > > >
> > > > > > On Wed, 19 Dec 2018 at 11:21, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> > > > > > >
> > > >
> > > > > > > > > >
> > > > > > > > > > diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c
> > > > > > > > > > index 7062469..6461469 100644
> > > > > > > > > > --- a/drivers/base/power/runtime.c
> > > > > > > > > > +++ b/drivers/base/power/runtime.c
> > > > > > > > > > @@ -88,6 +88,32 @@ static void __update_runtime_status(struct device *dev, enum rpm_status status)
> > > > > > > > > > dev->power.runtime_status = status;
> > > > > > > > > > }
> > > > > > > > > >
> > > > > > > > > > +u64 pm_runtime_accounted_time_get(struct device *dev, enum rpm_status status, bool update)
> > > > > > > > > > +{
> > > > > > > > > > + u64 now = ktime_to_ns(ktime_get());
> > > > > > > > >
> > > > > > > > > I think you should stay on jiffies here - and then switch to ktime in patch 3.
> > > > > > > > >
> > > > > > > > > > + u64 delta = 0, time = 0;
> > > > > > > > > > + unsigned long flags;
> > > > > > > > > > +
> > > > > > > > > > + spin_lock_irqsave(&dev->power.lock, flags);
> > > > > > > > > > +
> > > > > > > > > > + if (dev->power.disable_depth > 0)
> > > > > > > > > > + goto unlock;
> > > > > > > > > > +
> > > > > > > > > > + /* Add ongoing state if requested */
> > > > > > > > > > + if (update && dev->power.runtime_status == status)
> > > > > > > > > > + delta = now - dev->power.accounting_timestamp;
> > > > > > > > > > +
> > > > > > > > >
> > > > > > > > > Hmm. Do we really need to update the accounting timestamp? I would
> > > > > > > > > rather avoid it if possible.
> > > > > > > >
> > > > > > > > i915/drm uses this to track ongoing suspended state. In fact they are
> > > > > > > > mainly interested by this part
> > > > > > >
> > > > > > > Again, sorry I don't follow.
> > > > > >
> > > > > > In fact we don't update dev->power.accounting_timestamp but only use
> > > > > > it to get how much time has elapsed in the current state.
> > > > > >
> > > > > > >
> > > > > > > My suggested changes below, would do exactly that; track the ongoing
> > > > > > > suspended state.
> > > > > > >
> > > > > > > The user can call the function several times while the device remains
> > > > > > > RPM_SUSPENDED, and if needed the user could then compute the delta
> > > > > > > in-between the calls, for whatever reason that may be needed.
> > > > > >
> > > > > > So I'm not sure to catch your question:
> > > > > > Is your problem linked to status != RPM_SUSPENDED or the update
> > > > > > parameter that compute delta ?
> > > > >
> > > > > My intent was to keep things simple.
> > > > >
> > > > > 1. Only expose last suspended time, which means tracking the ongoing
> > > > > suspended state. In other words, you can also remove "enum rpm_status
> > > > > status" as the in-parameter to pm_runtime_accounted_time_get().
> > > >
> > > > Ok for this point if Rafael doesn't see any benefit of keeping the
> > > > generic interface
> > > >
> > > > > 2. Don't allow the user of pm_runtime_accounted_time_get() to update
> > > > > the current timestamp, in "dev->power.accounting_timestamp".
> > > >
> > > > But pm_runtime_accounted_time_get doesn't update
> > > > dev->power.accounting_timestamp, it only reads it to know when when
> > > > the last state transition happened
> > > >
> > > > >
> > > > > Is that okay for the drm driver, to do what it does today?
> > > >
> > > > drm driver needs 2 things: the accounted suspended time since the
> > > > last transition
> > >
> > > The core keeps tracks of the "total suspended time". Each time
> > > update_pm_runtime_accounting() is called, and the state is
> > > RPM_SUSPENDED it adds a delta to the total suspended time. Just to be
> > > clear, this may even happen when userspace reads the
> > > "runtime_suspended_time" sysfs node.
> > >
> > > My point is, the core doesn't track the "total suspended time since
> > > the last transition", which seems to be what the drm driver tries to
> > > figure out.
> > >
> > > Just to be clear, I don't think we should update the core to provide
> > > the data reflecting that time, as it would add overhead from
> > > additional time computations. I think it's better to push this down to
> >
> > Which kind of overhead are you referring ? This is done only when
> > pm_runtime_accounted_time_get') is called and doesn't modify
> > pm core metrics
>
> I was talking hypothetically.
>
> Having a function that performs some computation when actually called
> by the user, along the lines of what you propose in $subject patch, is
> in principle fine by me.
>
> The important part, is that we don't make core to perform *additional*
> unnecessary time computations, each time it calls
> update_pm_runtime_accounting().
>
> >
> > > those drivers that needs it, as it seems like a highly unusual thing.
> > >
> > > Instead, perhaps we should provide an API
> > > (pm_runtime_suspended_time()) that simply returns the value of
> > > dev->power.suspended_jiffies. The driver/subsystem could then call
> > > this API from its ->runtime_suspend|resume() callbacks, for example,
> > > to store values from it locally and later compute the deltas from it
> > > that it needs.
> >
> > not sure that i915/drm has such call back
> >
> > >
> > > Do note that, the core updates the status of the device to
> > > RPM_SUSPENDED, after the ->runtime_suspend() callback has returned a
> > > successful error code. Hence, calling the API from a
> > > ->runtime_suspend() callback would fetch the total suspended time, up
> > > until the last time the device became runtime resumed. That should be
> > > helpful, right?
> >
> > TBH, I don't know if this would help or not. i915/drm driver developer
> > should have the answer
> >
> > AFAICT, all this code is not driver in itself but some perf monitoring
> > stuff that estimate a events when it is not accessible anymore because
> > devices is suspended
> > >
> > > > and the time elapse in the current state when suspened
> > >
> > > Re-thinking this a bit from my earlier comments - and by following the
> > > above reasoning, it sounds like this better belongs in the
> > > driver/subsystem, without requiring any data from the core.
> > >
> > > The driver/subsystem could just store a timestamp in it's
> > > ->runtime_suspend() callback and whenever needed, it could compute a
> > > delta towards it. That should work, right?
> >
> > I don't know i915/drm enough to know all that details
>
> Okay, so let me re-summarize the main issue I see with your approach
> in $subject patch.
>
> dev->power.accounting_timestamp can't be used to know when last
> transition was made. If I understand correctly, that is how you use
> it. No?
Yes. At least that how I have interpreted the current code
>
> Anyway, as stated, that's because the timestamp becomes updated, if
> update_pm_runtime_accounting() is called via the sysfs nobs, which
> means there is no state transition happening, but only accounting data
> is updated.
Yes I have not realized that the update also happens there which makes
me think that i have
may be over interpreted the code and the initialization of
i915->pmu.suspended_jiffies_last
>
> So, what I think we can do from the core perspective, if it helps
> (which I am not sure of):
> 1. Export a function, which returns the value of dev->power.suspended_jiffies.
> 2. Export a wrapper function (to deal with locking) which calls
> update_pm_runtime_accounting(). This wrapper function allows the user
> the update the total suspended time, also taking into account the time
> spent in the current state.
Having now in mind that suspended_jiffies can be updated outside state
transition like via sysfs call,
we can maybe just implements 2 and return dev->power.suspended_jiffies
something like below
unsigned long pm_runtime_get_suspended_time(struct device *dev)
{
unsigned long time;
unsigned long flags;
spin_lock_irqsave(&dev->power.lock, flags);
update_pm_runtime_accounting(dev);
time = dev->power.suspended_time;
spin_unlock_irqrestore(&dev->power.lock, flags);
return time;
}
>
> Other than that, I think the rest should be managed in the drm driver itself.
>
> Kind regards
> Uffe
More information about the Intel-gfx
mailing list