Message ID | 20231212142730.998913-2-vincent.guittot@linaro.org |
---|---|
State | New |
Headers | show |
Series | [1/4] cpufreq: Add a cpufreq pressure feedback for the scheduler | expand |
On 12-12-23, 15:27, Vincent Guittot wrote: > Provide to the scheduler a feedback about the temporary max available > capacity. Unlike arch_update_thermal_pressure, this doesn't need to be > filtered as the pressure will happen for dozens ms or more. > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> > --- > drivers/cpufreq/cpufreq.c | 48 +++++++++++++++++++++++++++++++++++++++ > include/linux/cpufreq.h | 10 ++++++++ > 2 files changed, 58 insertions(+) > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > index 44db4f59c4cc..7d5f71be8d29 100644 > --- a/drivers/cpufreq/cpufreq.c > +++ b/drivers/cpufreq/cpufreq.c > @@ -2563,6 +2563,50 @@ int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu) > } > EXPORT_SYMBOL(cpufreq_get_policy); > > +DEFINE_PER_CPU(unsigned long, cpufreq_pressure); > +EXPORT_PER_CPU_SYMBOL_GPL(cpufreq_pressure); > + > +/** > + * cpufreq_update_pressure() - Update cpufreq pressure for CPUs > + * @cpus : The related CPUs for which max capacity has been reduced > + * @capped_freq : The maximum allowed frequency that CPUs can run at > + * > + * Update the value of cpufreq pressure for all @cpus in the mask. The > + * cpumask should include all (online+offline) affected CPUs, to avoid > + * operating on stale data when hot-plug is used for some CPUs. The > + * @capped_freq reflects the currently allowed max CPUs frequency due to > + * freq_qos capping. It might be also a boost frequency value, which is bigger > + * than the internal 'capacity_freq_ref' max frequency. In such case the > + * pressure value should simply be removed, since this is an indication that > + * there is no capping. The @capped_freq must be provided in kHz. > + */ > +static void cpufreq_update_pressure(const struct cpumask *cpus, Since this is defined as 'static', why not just pass policy here ? > + unsigned long capped_freq) > +{ > + unsigned long max_capacity, capacity, pressure; > + u32 max_freq; > + int cpu; > + > + cpu = cpumask_first(cpus); > + max_capacity = arch_scale_cpu_capacity(cpu); This anyway expects all of them to be from the same policy .. > + max_freq = arch_scale_freq_ref(cpu); > + > + /* > + * Handle properly the boost frequencies, which should simply clean > + * the thermal pressure value. > + */ > + if (max_freq <= capped_freq) > + capacity = max_capacity; > + else > + capacity = mult_frac(max_capacity, capped_freq, max_freq); > + > + pressure = max_capacity - capacity; > + Extra blank line here. > + > + for_each_cpu(cpu, cpus) > + WRITE_ONCE(per_cpu(cpufreq_pressure, cpu), pressure); > +} > + > /** > * cpufreq_set_policy - Modify cpufreq policy parameters. > * @policy: Policy object to modify. > @@ -2584,6 +2628,7 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, > { > struct cpufreq_policy_data new_data; > struct cpufreq_governor *old_gov; > + struct cpumask *cpus; > int ret; > > memcpy(&new_data.cpuinfo, &policy->cpuinfo, sizeof(policy->cpuinfo)); > @@ -2618,6 +2663,9 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, > policy->max = __resolve_freq(policy, policy->max, CPUFREQ_RELATION_H); > trace_cpu_frequency_limits(policy); > > + cpus = policy->related_cpus; You don't need the extra variable anyway, but lets just pass policy instead to the routine. > + cpufreq_update_pressure(cpus, policy->max); > + > policy->cached_target_freq = UINT_MAX; > > pr_debug("new min and max freqs are %u - %u kHz\n", > diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h > index afda5f24d3dd..b1d97edd3253 100644 > --- a/include/linux/cpufreq.h > +++ b/include/linux/cpufreq.h > @@ -241,6 +241,12 @@ struct kobject *get_governor_parent_kobj(struct cpufreq_policy *policy); > void cpufreq_enable_fast_switch(struct cpufreq_policy *policy); > void cpufreq_disable_fast_switch(struct cpufreq_policy *policy); > bool has_target_index(void); > + > +DECLARE_PER_CPU(unsigned long, cpufreq_pressure); > +static inline unsigned long cpufreq_get_pressure(int cpu) > +{ > + return per_cpu(cpufreq_pressure, cpu); > +} > #else > static inline unsigned int cpufreq_get(unsigned int cpu) > { > @@ -263,6 +269,10 @@ static inline bool cpufreq_supports_freq_invariance(void) > return false; > } > static inline void disable_cpufreq(void) { } > +static inline unsigned long cpufreq_get_pressure(int cpu) > +{ > + return 0; > +} > #endif > > #ifdef CONFIG_CPU_FREQ_STAT > -- > 2.34.1
On Wed, 13 Dec 2023 at 08:17, Viresh Kumar <viresh.kumar@linaro.org> wrote: > > On 12-12-23, 15:27, Vincent Guittot wrote: > > Provide to the scheduler a feedback about the temporary max available > > capacity. Unlike arch_update_thermal_pressure, this doesn't need to be > > filtered as the pressure will happen for dozens ms or more. > > > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> > > --- > > drivers/cpufreq/cpufreq.c | 48 +++++++++++++++++++++++++++++++++++++++ > > include/linux/cpufreq.h | 10 ++++++++ > > 2 files changed, 58 insertions(+) > > > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > > index 44db4f59c4cc..7d5f71be8d29 100644 > > --- a/drivers/cpufreq/cpufreq.c > > +++ b/drivers/cpufreq/cpufreq.c > > @@ -2563,6 +2563,50 @@ int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu) > > } > > EXPORT_SYMBOL(cpufreq_get_policy); > > > > +DEFINE_PER_CPU(unsigned long, cpufreq_pressure); > > +EXPORT_PER_CPU_SYMBOL_GPL(cpufreq_pressure); > > + > > +/** > > + * cpufreq_update_pressure() - Update cpufreq pressure for CPUs > > + * @cpus : The related CPUs for which max capacity has been reduced > > + * @capped_freq : The maximum allowed frequency that CPUs can run at > > + * > > + * Update the value of cpufreq pressure for all @cpus in the mask. The > > + * cpumask should include all (online+offline) affected CPUs, to avoid > > + * operating on stale data when hot-plug is used for some CPUs. The > > + * @capped_freq reflects the currently allowed max CPUs frequency due to > > + * freq_qos capping. It might be also a boost frequency value, which is bigger > > + * than the internal 'capacity_freq_ref' max frequency. In such case the > > + * pressure value should simply be removed, since this is an indication that > > + * there is no capping. The @capped_freq must be provided in kHz. > > + */ > > +static void cpufreq_update_pressure(const struct cpumask *cpus, > > Since this is defined as 'static', why not just pass policy here ? Mainly because we only need the cpumask and also because this follows the same pattern as other place like arch_topology.c > > > + unsigned long capped_freq) > > +{ > > + unsigned long max_capacity, capacity, pressure; > > + u32 max_freq; > > + int cpu; > > + > > + cpu = cpumask_first(cpus); > > + max_capacity = arch_scale_cpu_capacity(cpu); > > This anyway expects all of them to be from the same policy .. > > > + max_freq = arch_scale_freq_ref(cpu); > > + > > + /* > > + * Handle properly the boost frequencies, which should simply clean > > + * the thermal pressure value. > > + */ > > + if (max_freq <= capped_freq) > > + capacity = max_capacity; > > + else > > + capacity = mult_frac(max_capacity, capped_freq, max_freq); > > + > > + pressure = max_capacity - capacity; > > + > > Extra blank line here. > > > + > > + for_each_cpu(cpu, cpus) > > + WRITE_ONCE(per_cpu(cpufreq_pressure, cpu), pressure); > > +} > > + > > /** > > * cpufreq_set_policy - Modify cpufreq policy parameters. > > * @policy: Policy object to modify. > > @@ -2584,6 +2628,7 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, > > { > > struct cpufreq_policy_data new_data; > > struct cpufreq_governor *old_gov; > > + struct cpumask *cpus; > > int ret; > > > > memcpy(&new_data.cpuinfo, &policy->cpuinfo, sizeof(policy->cpuinfo)); > > @@ -2618,6 +2663,9 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, > > policy->max = __resolve_freq(policy, policy->max, CPUFREQ_RELATION_H); > > trace_cpu_frequency_limits(policy); > > > > + cpus = policy->related_cpus; > > You don't need the extra variable anyway, but lets just pass policy > instead to the routine. In fact I have followed what was done in cpufreq_cooling.c with arch_update_thermal_pressure(). Will remove it > > > + cpufreq_update_pressure(cpus, policy->max); > > + > > policy->cached_target_freq = UINT_MAX; > > > > pr_debug("new min and max freqs are %u - %u kHz\n", > > diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h > > index afda5f24d3dd..b1d97edd3253 100644 > > --- a/include/linux/cpufreq.h > > +++ b/include/linux/cpufreq.h > > @@ -241,6 +241,12 @@ struct kobject *get_governor_parent_kobj(struct cpufreq_policy *policy); > > void cpufreq_enable_fast_switch(struct cpufreq_policy *policy); > > void cpufreq_disable_fast_switch(struct cpufreq_policy *policy); > > bool has_target_index(void); > > + > > +DECLARE_PER_CPU(unsigned long, cpufreq_pressure); > > +static inline unsigned long cpufreq_get_pressure(int cpu) > > +{ > > + return per_cpu(cpufreq_pressure, cpu); > > +} > > #else > > static inline unsigned int cpufreq_get(unsigned int cpu) > > { > > @@ -263,6 +269,10 @@ static inline bool cpufreq_supports_freq_invariance(void) > > return false; > > } > > static inline void disable_cpufreq(void) { } > > +static inline unsigned long cpufreq_get_pressure(int cpu) > > +{ > > + return 0; > > +} > > #endif > > > > #ifdef CONFIG_CPU_FREQ_STAT > > -- > > 2.34.1 > > -- > viresh
On 12/14/23 07:57, Vincent Guittot wrote: > On Thu, 14 Dec 2023 at 06:43, Viresh Kumar <viresh.kumar@linaro.org> wrote: >> >> On 12-12-23, 15:27, Vincent Guittot wrote: >>> @@ -2618,6 +2663,9 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, >>> policy->max = __resolve_freq(policy, policy->max, CPUFREQ_RELATION_H); >>> trace_cpu_frequency_limits(policy); >>> >>> + cpus = policy->related_cpus; >>> + cpufreq_update_pressure(cpus, policy->max); >>> + >>> policy->cached_target_freq = UINT_MAX; >> >> One more question, why are you doing this from cpufreq_set_policy ? If >> due to cpufreq cooling or from userspace, we end up limiting the >> maximum possible frequency, will this routine always get called ? > > Yes, any update of a FREQ_QOS_MAX ends up calling cpufreq_set_policy() > to update the policy->max > Agree, cpufreq sysfs scaling_max_freq is also important to handle in this new design. Currently we don't reflect that as reduced CPU capacity in the scheduler. There was discussion when I proposed to feed that CPU frequency reduction into thermal_pressure [1]. The same applies for the DTPM which is missing currently the proper impact to the CPU reduced capacity in the scheduler. IMHO any limit set into FREQ_QOS_MAX should be visible in this new design of capacity reduction signaling. [1] https://lore.kernel.org/lkml/20220930094821.31665-2-lukasz.luba@arm.com/
On 12/12/23 14:27, Vincent Guittot wrote: > Provide to the scheduler a feedback about the temporary max available > capacity. Unlike arch_update_thermal_pressure, this doesn't need to be > filtered as the pressure will happen for dozens ms or more. > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> > --- > drivers/cpufreq/cpufreq.c | 48 +++++++++++++++++++++++++++++++++++++++ > include/linux/cpufreq.h | 10 ++++++++ > 2 files changed, 58 insertions(+) > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c > index 44db4f59c4cc..7d5f71be8d29 100644 > --- a/drivers/cpufreq/cpufreq.c > +++ b/drivers/cpufreq/cpufreq.c > @@ -2563,6 +2563,50 @@ int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu) > } > EXPORT_SYMBOL(cpufreq_get_policy); > > +DEFINE_PER_CPU(unsigned long, cpufreq_pressure); > +EXPORT_PER_CPU_SYMBOL_GPL(cpufreq_pressure); Why do we export this variable when we have get/update functions? Do we expect modules would manipulate those per-cpu variables independently and not like we do per-cpumask in the update func.?
diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c index 44db4f59c4cc..7d5f71be8d29 100644 --- a/drivers/cpufreq/cpufreq.c +++ b/drivers/cpufreq/cpufreq.c @@ -2563,6 +2563,50 @@ int cpufreq_get_policy(struct cpufreq_policy *policy, unsigned int cpu) } EXPORT_SYMBOL(cpufreq_get_policy); +DEFINE_PER_CPU(unsigned long, cpufreq_pressure); +EXPORT_PER_CPU_SYMBOL_GPL(cpufreq_pressure); + +/** + * cpufreq_update_pressure() - Update cpufreq pressure for CPUs + * @cpus : The related CPUs for which max capacity has been reduced + * @capped_freq : The maximum allowed frequency that CPUs can run at + * + * Update the value of cpufreq pressure for all @cpus in the mask. The + * cpumask should include all (online+offline) affected CPUs, to avoid + * operating on stale data when hot-plug is used for some CPUs. The + * @capped_freq reflects the currently allowed max CPUs frequency due to + * freq_qos capping. It might be also a boost frequency value, which is bigger + * than the internal 'capacity_freq_ref' max frequency. In such case the + * pressure value should simply be removed, since this is an indication that + * there is no capping. The @capped_freq must be provided in kHz. + */ +static void cpufreq_update_pressure(const struct cpumask *cpus, + unsigned long capped_freq) +{ + unsigned long max_capacity, capacity, pressure; + u32 max_freq; + int cpu; + + cpu = cpumask_first(cpus); + max_capacity = arch_scale_cpu_capacity(cpu); + max_freq = arch_scale_freq_ref(cpu); + + /* + * Handle properly the boost frequencies, which should simply clean + * the thermal pressure value. + */ + if (max_freq <= capped_freq) + capacity = max_capacity; + else + capacity = mult_frac(max_capacity, capped_freq, max_freq); + + pressure = max_capacity - capacity; + + + for_each_cpu(cpu, cpus) + WRITE_ONCE(per_cpu(cpufreq_pressure, cpu), pressure); +} + /** * cpufreq_set_policy - Modify cpufreq policy parameters. * @policy: Policy object to modify. @@ -2584,6 +2628,7 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, { struct cpufreq_policy_data new_data; struct cpufreq_governor *old_gov; + struct cpumask *cpus; int ret; memcpy(&new_data.cpuinfo, &policy->cpuinfo, sizeof(policy->cpuinfo)); @@ -2618,6 +2663,9 @@ static int cpufreq_set_policy(struct cpufreq_policy *policy, policy->max = __resolve_freq(policy, policy->max, CPUFREQ_RELATION_H); trace_cpu_frequency_limits(policy); + cpus = policy->related_cpus; + cpufreq_update_pressure(cpus, policy->max); + policy->cached_target_freq = UINT_MAX; pr_debug("new min and max freqs are %u - %u kHz\n", diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h index afda5f24d3dd..b1d97edd3253 100644 --- a/include/linux/cpufreq.h +++ b/include/linux/cpufreq.h @@ -241,6 +241,12 @@ struct kobject *get_governor_parent_kobj(struct cpufreq_policy *policy); void cpufreq_enable_fast_switch(struct cpufreq_policy *policy); void cpufreq_disable_fast_switch(struct cpufreq_policy *policy); bool has_target_index(void); + +DECLARE_PER_CPU(unsigned long, cpufreq_pressure); +static inline unsigned long cpufreq_get_pressure(int cpu) +{ + return per_cpu(cpufreq_pressure, cpu); +} #else static inline unsigned int cpufreq_get(unsigned int cpu) { @@ -263,6 +269,10 @@ static inline bool cpufreq_supports_freq_invariance(void) return false; } static inline void disable_cpufreq(void) { } +static inline unsigned long cpufreq_get_pressure(int cpu) +{ + return 0; +} #endif #ifdef CONFIG_CPU_FREQ_STAT
Provide to the scheduler a feedback about the temporary max available capacity. Unlike arch_update_thermal_pressure, this doesn't need to be filtered as the pressure will happen for dozens ms or more. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> --- drivers/cpufreq/cpufreq.c | 48 +++++++++++++++++++++++++++++++++++++++ include/linux/cpufreq.h | 10 ++++++++ 2 files changed, 58 insertions(+)