| Message ID | 20170706094948.8779-3-dietmar.eggemann@arm.com (mailing list archive) |
|---|---|
| State | Changes Requested, archived |
| Headers | show |
On 06-07-17, 10:49, Dietmar Eggemann wrote: > A frequency-invariant load-tracking solution based on cpufreq transition > notifier will not work for future fast frequency switching policies. > That is why a different solution is presented with this patch. > > Let cpufreq call the function arch_set_freq_scale() to pass the current > frequency, the max supported frequency and the cpumask of the related > cpus to a consumer (an arch) which defines arch_set_freq_scale(). > > The consumer has to associate arch_set_freq_scale with the name of its > own implementation foo_set_freq_scale() to overwrite the empty standard > definition in drivers/cpufreq/cpufreq.c. > An arch could do this in one of its arch-specific header files > (e.g. arch/$ARCH/include/asm/topology.h) which gets included in > drivers/cpufreq/cpufreq.c. > > In case arch_set_freq_scale() is not defined (and because of the > pr_debug() drivers/cpufreq/cpufreq.c is not compiled with -DDEBUG) The line within () needs to be improved to convey a clear message. > the > function cpufreq_set_freq_scale() gets compiled out. > > Cc: Rafael J. Wysocki <rjw@rjwysocki.net> > Cc: Viresh Kumar <viresh.kumar@linaro.org> > Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com> > --- > drivers/cpufreq/cpufreq.c | 26 ++++++++++++++++++++++++++ > 1 file changed, 26 insertions(+) > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > index 9bf97a366029..a04c5886a5ce 100644 > --- a/drivers/cpufreq/cpufreq.c > +++ b/drivers/cpufreq/cpufreq.c > @@ -347,6 +347,28 @@ static void __cpufreq_notify_transition(struct cpufreq_policy *policy, > } > } > > +/********************************************************************* > + * FREQUENCY INVARIANT CPU CAPACITY SUPPORT * > + *********************************************************************/ > + > +#ifndef arch_set_freq_scale > +static void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq, > + unsigned long max_freq) > +{} > +#endif > + > +static void cpufreq_set_freq_scale(struct cpufreq_policy *policy, > + struct cpufreq_freqs *freqs) > +{ > + unsigned long cur_freq = freqs ? freqs->new : policy->cur; > + unsigned long max_freq = policy->cpuinfo.max_freq; > + > + pr_debug("cpus %*pbl cur/cur max freq %lu/%lu kHz\n", > + cpumask_pr_args(policy->related_cpus), cur_freq, max_freq); > + > + arch_set_freq_scale(policy->related_cpus, cur_freq, max_freq); I am not sure why all these are required to be sent here and will come back to it later on after going through other patches. > +} > + > /** > * cpufreq_notify_transition - call notifier chain and adjust_jiffies > * on frequency transition. > @@ -405,6 +427,8 @@ void cpufreq_freq_transition_begin(struct cpufreq_policy *policy, > > spin_unlock(&policy->transition_lock); > > + cpufreq_set_freq_scale(policy, freqs); > + Why do this before even changing the frequency ? We may fail while changing it. IMHO, you should call this routine whenever we update policy->cur and that happens regularly in __cpufreq_notify_transition() and few other places.. > cpufreq_notify_transition(policy, freqs, CPUFREQ_PRECHANGE); > } > EXPORT_SYMBOL_GPL(cpufreq_freq_transition_begin); > @@ -2203,6 +2227,8 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, > blocking_notifier_call_chain(&cpufreq_policy_notifier_list, > CPUFREQ_NOTIFY, new_policy); > > + cpufreq_set_freq_scale(new_policy, NULL); Why added it here ? To get it initialized ? If yes, then we should do that in cpufreq_online() where we first initialize policy->cur. Apart from this, you also need to update this in the schedutil governor (if you haven't done that in this series later) as that also updates policy->cur in the fast path.
On Thursday, July 06, 2017 04:10:27 PM Viresh Kumar wrote: > On 06-07-17, 10:49, Dietmar Eggemann wrote: > > A frequency-invariant load-tracking solution based on cpufreq transition > > notifier will not work for future fast frequency switching policies. > > That is why a different solution is presented with this patch. > > > > Let cpufreq call the function arch_set_freq_scale() to pass the current > > frequency, the max supported frequency and the cpumask of the related > > cpus to a consumer (an arch) which defines arch_set_freq_scale(). > > > > The consumer has to associate arch_set_freq_scale with the name of its > > own implementation foo_set_freq_scale() to overwrite the empty standard > > definition in drivers/cpufreq/cpufreq.c. > > An arch could do this in one of its arch-specific header files > > (e.g. arch/$ARCH/include/asm/topology.h) which gets included in > > drivers/cpufreq/cpufreq.c. > > > > In case arch_set_freq_scale() is not defined (and because of the > > pr_debug() drivers/cpufreq/cpufreq.c is not compiled with -DDEBUG) > > The line within () needs to be improved to convey a clear message. > > > the > > function cpufreq_set_freq_scale() gets compiled out. > > > > Cc: Rafael J. Wysocki <rjw@rjwysocki.net> > > Cc: Viresh Kumar <viresh.kumar@linaro.org> > > Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com> > > --- > > drivers/cpufreq/cpufreq.c | 26 ++++++++++++++++++++++++++ > > 1 file changed, 26 insertions(+) > > > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > > index 9bf97a366029..a04c5886a5ce 100644 > > --- a/drivers/cpufreq/cpufreq.c > > +++ b/drivers/cpufreq/cpufreq.c > > @@ -347,6 +347,28 @@ static void __cpufreq_notify_transition(struct cpufreq_policy *policy, > > } > > } > > > > +/********************************************************************* > > + * FREQUENCY INVARIANT CPU CAPACITY SUPPORT * > > + *********************************************************************/ > > + > > +#ifndef arch_set_freq_scale > > +static void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq, > > + unsigned long max_freq) > > +{} > > +#endif > > + > > +static void cpufreq_set_freq_scale(struct cpufreq_policy *policy, > > + struct cpufreq_freqs *freqs) > > +{ > > + unsigned long cur_freq = freqs ? freqs->new : policy->cur; > > + unsigned long max_freq = policy->cpuinfo.max_freq; > > + > > + pr_debug("cpus %*pbl cur/cur max freq %lu/%lu kHz\n", > > + cpumask_pr_args(policy->related_cpus), cur_freq, max_freq); > > + > > + arch_set_freq_scale(policy->related_cpus, cur_freq, max_freq); > > I am not sure why all these are required to be sent here and will come back to > it later on after going through other patches. > > > +} > > + > > /** > > * cpufreq_notify_transition - call notifier chain and adjust_jiffies > > * on frequency transition. > > @@ -405,6 +427,8 @@ void cpufreq_freq_transition_begin(struct cpufreq_policy *policy, > > > > spin_unlock(&policy->transition_lock); > > > > + cpufreq_set_freq_scale(policy, freqs); > > + > > Why do this before even changing the frequency ? We may fail while changing it. > > IMHO, you should call this routine whenever we update policy->cur and that > happens regularly in __cpufreq_notify_transition() and few other places.. There seems to be a general problem with doing this in the core with respect to things like intel_pstate that use their own governor callbacks and don't invoke cpufreq_freq_transition_begin() then. Thanks, Rafael
diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c index 9bf97a366029..a04c5886a5ce 100644 --- a/drivers/cpufreq/cpufreq.c +++ b/drivers/cpufreq/cpufreq.c @@ -347,6 +347,28 @@ static void __cpufreq_notify_transition(struct cpufreq_policy *policy, } } +/********************************************************************* + * FREQUENCY INVARIANT CPU CAPACITY SUPPORT * + *********************************************************************/ + +#ifndef arch_set_freq_scale +static void arch_set_freq_scale(struct cpumask *cpus, unsigned long cur_freq, + unsigned long max_freq) +{} +#endif + +static void cpufreq_set_freq_scale(struct cpufreq_policy *policy, + struct cpufreq_freqs *freqs) +{ + unsigned long cur_freq = freqs ? freqs->new : policy->cur; + unsigned long max_freq = policy->cpuinfo.max_freq; + + pr_debug("cpus %*pbl cur/cur max freq %lu/%lu kHz\n", + cpumask_pr_args(policy->related_cpus), cur_freq, max_freq); + + arch_set_freq_scale(policy->related_cpus, cur_freq, max_freq); +} + /** * cpufreq_notify_transition - call notifier chain and adjust_jiffies * on frequency transition. @@ -405,6 +427,8 @@ void cpufreq_freq_transition_begin(struct cpufreq_policy *policy, spin_unlock(&policy->transition_lock); + cpufreq_set_freq_scale(policy, freqs); + cpufreq_notify_transition(policy, freqs, CPUFREQ_PRECHANGE); } EXPORT_SYMBOL_GPL(cpufreq_freq_transition_begin); @@ -2203,6 +2227,8 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, blocking_notifier_call_chain(&cpufreq_policy_notifier_list, CPUFREQ_NOTIFY, new_policy); + cpufreq_set_freq_scale(new_policy, NULL); + policy->min = new_policy->min; policy->max = new_policy->max;
A frequency-invariant load-tracking solution based on cpufreq transition notifier will not work for future fast frequency switching policies. That is why a different solution is presented with this patch. Let cpufreq call the function arch_set_freq_scale() to pass the current frequency, the max supported frequency and the cpumask of the related cpus to a consumer (an arch) which defines arch_set_freq_scale(). The consumer has to associate arch_set_freq_scale with the name of its own implementation foo_set_freq_scale() to overwrite the empty standard definition in drivers/cpufreq/cpufreq.c. An arch could do this in one of its arch-specific header files (e.g. arch/$ARCH/include/asm/topology.h) which gets included in drivers/cpufreq/cpufreq.c. In case arch_set_freq_scale() is not defined (and because of the pr_debug() drivers/cpufreq/cpufreq.c is not compiled with -DDEBUG) the function cpufreq_set_freq_scale() gets compiled out. Cc: Rafael J. Wysocki <rjw@rjwysocki.net> Cc: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com> --- drivers/cpufreq/cpufreq.c | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)