diff mbox series

[v2,1/2] arm64: Provide an AMU-based version of arch_freq_get_on_cpu

Message ID 20231127160838.1403404-2-beata.michalska@arm.com (mailing list archive)
State Changes Requested, archived
Headers show
Series Add support for AArch64 AMUv1-based arch_freq_get_on_cpu | expand

Commit Message

Beata Michalska Nov. 27, 2023, 4:08 p.m. UTC 
With the Frequency Invariance Engine (FIE) being already wired up with
sched tick and making use of relevant (core counter and constant
counter) AMU counters, getting the current frequency for a given CPU
on supported platforms, can be achieved by utilizing the frequency scale
factor which reflects an average CPU frequency for the last tick period
length.

Suggested-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Beata Michalska <beata.michalska@arm.com>
Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>
---

Notes:
    Due to [1], if merged, there might be a need to modify the patch to
    accommodate changes [1] introduces:
    
    	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT
    	to
    	freq = per_cpu(capacity_freq_ref, cpu); >> SCHED_CAPACITY_SHIFT
    [1]
    https://lore.kernel.org/linux-arm-kernel/20231121154349.GA1938@willie-the-truck/T/#mcb018d076dbce6f60ed2779634a9b6ffe622641e

 arch/arm64/kernel/topology.c | 39 ++++++++++++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)

Comments

Ionela Voinescu Nov. 28, 2023, 3:13 p.m. UTC | #1 
Hi Beata,

On Monday 27 Nov 2023 at 16:08:37 (+0000), Beata Michalska wrote:
> With the Frequency Invariance Engine (FIE) being already wired up with
> sched tick and making use of relevant (core counter and constant
> counter) AMU counters, getting the current frequency for a given CPU
> on supported platforms, can be achieved by utilizing the frequency scale
> factor which reflects an average CPU frequency for the last tick period
> length.
> 
> Suggested-by: Ionela Voinescu <ionela.voinescu@arm.com>
> Signed-off-by: Beata Michalska <beata.michalska@arm.com>
> Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>
> ---
> 
> Notes:
>     Due to [1], if merged, there might be a need to modify the patch to
>     accommodate changes [1] introduces:
>     
>     	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT
>     	to
>     	freq = per_cpu(capacity_freq_ref, cpu); >> SCHED_CAPACITY_SHIFT
>     [1]
>     https://lore.kernel.org/linux-arm-kernel/20231121154349.GA1938@willie-the-truck/T/#mcb018d076dbce6f60ed2779634a9b6ffe622641e
> 
>  arch/arm64/kernel/topology.c | 39 ++++++++++++++++++++++++++++++++++++
>  1 file changed, 39 insertions(+)
> 
> diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> index 615c1a20129f..ae2445f6e7da 100644
> --- a/arch/arm64/kernel/topology.c
> +++ b/arch/arm64/kernel/topology.c
> @@ -17,6 +17,7 @@
>  #include <linux/cpufreq.h>
>  #include <linux/init.h>
>  #include <linux/percpu.h>
> +#include <linux/sched/isolation.h>
>  
>  #include <asm/cpu.h>
>  #include <asm/cputype.h>
> @@ -186,6 +187,44 @@ static void amu_scale_freq_tick(void)
>  	this_cpu_write(arch_freq_scale, (unsigned long)scale);
>  }
>  
> +unsigned int arch_freq_get_on_cpu(int cpu)
> +{
> +	unsigned int freq;
> +	u64 scale;
> +
> +	if (!cpumask_test_cpu(cpu, amu_fie_cpus))
> +		return 0;
> +
> +	/*
> +	 * For those CPUs that are in full dynticks mode, try an alternative
> +	 * source for the counters (and thus freq scale),
> +	 * if available for given policy
> +	 */
> +	if (!housekeeping_cpu(cpu, HK_TYPE_TICK)) {
> +		struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
> +		int ref_cpu = nr_cpu_ids;
> +
> +		if (cpumask_intersects(housekeeping_cpumask(HK_TYPE_TICK),
> +				       policy->cpus))
> +			ref_cpu = cpumask_nth_and(cpu, policy->cpus,
> +						  housekeeping_cpumask(HK_TYPE_TICK));
> +		cpufreq_cpu_put(policy);
> +		if (ref_cpu >= nr_cpu_ids)
> +			return 0;
> +		cpu = ref_cpu;
> +	}
> +
> +	/*
> +	 * Reversed computation to the one used to determine
> +	 * the arch_freq_scale value
> +	 * (see amu_scale_freq_tick for details)
> +	 */
> +	scale = per_cpu(arch_freq_scale, cpu);

Any reason for not using arch_scale_freq_capacity() here?

To me it seems a bit nicer to use the "official" function to return the
frequency scale factor.

> +	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT;

Given Vincent's patch at [1] I think here might be best to call
arch_scale_freq_ref() instead. That's because the frequency scale factor
will use that frequency as the maximum frequency in its calculations and
we'd not want to use a different one here.

The annoyance is coping with capacity_freq_ref not having been set
yet, and that would be easy if capacity_freq_ref was initialized to 0.
Luckily with Vincent's changes it can now be 0. I'll comments on his
patches and ask him to make this change.

So I think you can safely use arch_scale_freq_ref() here. If
arch_scale_freq_ref() returns 0, arch_freq_get_on_cpu() will just return
0 as well.

[1] https://lore.kernel.org/lkml/20231109101438.1139696-8-vincent.guittot@linaro.org/

> +	freq *= scale;

In some scenarios the frequencies visible to cpufreq might not look like
actual frequencies, but some scaled abstract performance values. One
example is cppc_cpufreq when one does not provide the optional frequency
information in the CPC objects but just the performance information.

Therefore the maximum frequency seen here can be quite a small value, so
it might be best to do the multiplication first and the shift after that.

Thanks,
Ionela.

> +	return freq;
> +}
> +
>  static struct scale_freq_data amu_sfd = {
>  	.source = SCALE_FREQ_SOURCE_ARCH,
>  	.set_freq_scale = amu_scale_freq_tick,
> -- 
> 2.25.1
>
Beata Michalska Feb. 2, 2024, 9:20 a.m. UTC | #2 
On Tue, Nov 28, 2023 at 03:13:51PM +0000, Ionela Voinescu wrote:
> Hi Beata,
> 
> On Monday 27 Nov 2023 at 16:08:37 (+0000), Beata Michalska wrote:
> > With the Frequency Invariance Engine (FIE) being already wired up with
> > sched tick and making use of relevant (core counter and constant
> > counter) AMU counters, getting the current frequency for a given CPU
> > on supported platforms, can be achieved by utilizing the frequency scale
> > factor which reflects an average CPU frequency for the last tick period
> > length.
> > 
> > Suggested-by: Ionela Voinescu <ionela.voinescu@arm.com>
> > Signed-off-by: Beata Michalska <beata.michalska@arm.com>
> > Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>
> > ---
> > 
> > Notes:
> >     Due to [1], if merged, there might be a need to modify the patch to
> >     accommodate changes [1] introduces:
> >     
> >     	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT
> >     	to
> >     	freq = per_cpu(capacity_freq_ref, cpu); >> SCHED_CAPACITY_SHIFT
> >     [1]
> >     https://lore.kernel.org/linux-arm-kernel/20231121154349.GA1938@willie-the-truck/T/#mcb018d076dbce6f60ed2779634a9b6ffe622641e
> > 
> >  arch/arm64/kernel/topology.c | 39 ++++++++++++++++++++++++++++++++++++
> >  1 file changed, 39 insertions(+)
> > 
> > diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
> > index 615c1a20129f..ae2445f6e7da 100644
> > --- a/arch/arm64/kernel/topology.c
> > +++ b/arch/arm64/kernel/topology.c
> > @@ -17,6 +17,7 @@
> >  #include <linux/cpufreq.h>
> >  #include <linux/init.h>
> >  #include <linux/percpu.h>
> > +#include <linux/sched/isolation.h>
> >  
> >  #include <asm/cpu.h>
> >  #include <asm/cputype.h>
> > @@ -186,6 +187,44 @@ static void amu_scale_freq_tick(void)
> >  	this_cpu_write(arch_freq_scale, (unsigned long)scale);
> >  }
> >  
> > +unsigned int arch_freq_get_on_cpu(int cpu)
> > +{
> > +	unsigned int freq;
> > +	u64 scale;
> > +
> > +	if (!cpumask_test_cpu(cpu, amu_fie_cpus))
> > +		return 0;
> > +
> > +	/*
> > +	 * For those CPUs that are in full dynticks mode, try an alternative
> > +	 * source for the counters (and thus freq scale),
> > +	 * if available for given policy
> > +	 */
> > +	if (!housekeeping_cpu(cpu, HK_TYPE_TICK)) {
> > +		struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
> > +		int ref_cpu = nr_cpu_ids;
> > +
> > +		if (cpumask_intersects(housekeeping_cpumask(HK_TYPE_TICK),
> > +				       policy->cpus))
> > +			ref_cpu = cpumask_nth_and(cpu, policy->cpus,
> > +						  housekeeping_cpumask(HK_TYPE_TICK));
> > +		cpufreq_cpu_put(policy);
> > +		if (ref_cpu >= nr_cpu_ids)
> > +			return 0;
> > +		cpu = ref_cpu;
> > +	}
> > +
> > +	/*
> > +	 * Reversed computation to the one used to determine
> > +	 * the arch_freq_scale value
> > +	 * (see amu_scale_freq_tick for details)
> > +	 */
> > +	scale = per_cpu(arch_freq_scale, cpu);
> 
> Any reason for not using arch_scale_freq_capacity() here?
> 
> To me it seems a bit nicer to use the "official" function to return the
> frequency scale factor.
> 
Noted.
> > +	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT;
> 
> Given Vincent's patch at [1] I think here might be best to call
> arch_scale_freq_ref() instead. That's because the frequency scale factor
> will use that frequency as the maximum frequency in its calculations and
> we'd not want to use a different one here.
> 
OK.
> The annoyance is coping with capacity_freq_ref not having been set
> yet, and that would be easy if capacity_freq_ref was initialized to 0.
> Luckily with Vincent's changes it can now be 0. I'll comments on his
> patches and ask him to make this change.
> 
> So I think you can safely use arch_scale_freq_ref() here. If
> arch_scale_freq_ref() returns 0, arch_freq_get_on_cpu() will just return
> 0 as well.
> 
Will do.
> [1] https://lore.kernel.org/lkml/20231109101438.1139696-8-vincent.guittot@linaro.org/
> 
> > +	freq *= scale;
> 
> In some scenarios the frequencies visible to cpufreq might not look like
> actual frequencies, but some scaled abstract performance values. One
> example is cppc_cpufreq when one does not provide the optional frequency
> information in the CPC objects but just the performance information.
> 
> Therefore the maximum frequency seen here can be quite a small value, so
> it might be best to do the multiplication first and the shift after that.
> 
Right, that was in v1! Must have mixed up things ending with stale data.
Will address that in the next version - if one is out.

---
BR
Beata
> Thanks,
> Ionela.
> 
> > +	return freq;
> > +}
> > +
> >  static struct scale_freq_data amu_sfd = {
> >  	.source = SCALE_FREQ_SOURCE_ARCH,
> >  	.set_freq_scale = amu_scale_freq_tick,
> > -- 
> > 2.25.1
> >
Vanshidhar Konda Feb. 22, 2024, 7:55 p.m. UTC | #3 
Hello Beata,

I tested this patch based on the discussion in [1] on an AmpereOne
system when the system was mostly idle. The results below are when only
I applied the first patch in this series to the kernel. I noticed that
the frequency reported through scaling_cur_freq vs cpuinfo_cur_freq is
quite different for some cores. When the cores are loaded using
stress-ng, the scaling_cur_freq and cpuinfo_cur_freq values are quite
similar.

Applying the second patch in this series causes the difference between
scaling_cur_freq and cpuinfo_cur_freq to disappear, but calculating the
core frequency based on the feedback_ctrs shows that the value is
incorrect for both.

The kernel I compiled for testing is based on the Fedora 39 config file.
These configs seem relevant to the discussion:

CONFIG_NO_HZ_COMMON=y
# CONFIG_HZ_PERIODIC is not set
# CONFIG_NO_HZ_IDLE is not set
CONFIG_NO_HZ_FULL=y
CONFIG_CONTEXT_TRACKING_USER=y
# CONFIG_CONTEXT_TRACKING_USER_FORCE is not set
CONFIG_NO_HZ=y

[1]: https://lore.kernel.org/linux-arm-kernel/20231212072617.14756-1-lihuisong@huawei.com/

Results:
cpu_num scaling_cur_freq  cpuinfo_cur_freq
[11]:	  1874560	          1000000
[12]:	  2056158	          1385000
[13]:	  1974146	          1000000
[21]:	  1587518	          1000000
[23]:	  1593376           1000000
...
... skipping similar results for ~50 other cores for brevity
...
nohz_full=113-118
[113]:	1874560	          1000000
[114]:	1968288	          1000000
[115]:	1962430	          1000000
[116]:	1871631	          1000000
[117]:	1877489	          1000000
[118]:	1877489	          1000000
isolcpus=119-127
[119]:	2999296	          1000000
[120]:	2999296	          1000000
[121]:	2999296	          1000000
[125]:	2999296	          1000000
[126]:	2999296	          1000000
[127]:	2999296	          1000000

Thanks,
Vanshi

On Mon, Nov 27, 2023 at 04:08:37PM +0000, Beata Michalska wrote:
>With the Frequency Invariance Engine (FIE) being already wired up with
>sched tick and making use of relevant (core counter and constant
>counter) AMU counters, getting the current frequency for a given CPU
>on supported platforms, can be achieved by utilizing the frequency scale
>factor which reflects an average CPU frequency for the last tick period
>length.
>
>Suggested-by: Ionela Voinescu <ionela.voinescu@arm.com>
>Signed-off-by: Beata Michalska <beata.michalska@arm.com>
>Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>
>---
>
>Notes:
>    Due to [1], if merged, there might be a need to modify the patch to
>    accommodate changes [1] introduces:
>
>    	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT
>    	to
>    	freq = per_cpu(capacity_freq_ref, cpu); >> SCHED_CAPACITY_SHIFT
>    [1]
>    https://lore.kernel.org/linux-arm-kernel/20231121154349.GA1938@willie-the-truck/T/#mcb018d076dbce6f60ed2779634a9b6ffe622641e
>
> arch/arm64/kernel/topology.c | 39 ++++++++++++++++++++++++++++++++++++
> 1 file changed, 39 insertions(+)
>
>diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
>index 615c1a20129f..ae2445f6e7da 100644
>--- a/arch/arm64/kernel/topology.c
>+++ b/arch/arm64/kernel/topology.c
>@@ -17,6 +17,7 @@
> #include <linux/cpufreq.h>
> #include <linux/init.h>
> #include <linux/percpu.h>
>+#include <linux/sched/isolation.h>
>
> #include <asm/cpu.h>
> #include <asm/cputype.h>
>@@ -186,6 +187,44 @@ static void amu_scale_freq_tick(void)
> 	this_cpu_write(arch_freq_scale, (unsigned long)scale);
> }
>
>+unsigned int arch_freq_get_on_cpu(int cpu)
>+{
>+	unsigned int freq;
>+	u64 scale;
>+
>+	if (!cpumask_test_cpu(cpu, amu_fie_cpus))
>+		return 0;
>+
>+	/*
>+	 * For those CPUs that are in full dynticks mode, try an alternative
>+	 * source for the counters (and thus freq scale),
>+	 * if available for given policy
>+	 */
>+	if (!housekeeping_cpu(cpu, HK_TYPE_TICK)) {
>+		struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
>+		int ref_cpu = nr_cpu_ids;
>+
>+		if (cpumask_intersects(housekeeping_cpumask(HK_TYPE_TICK),
>+				       policy->cpus))
>+			ref_cpu = cpumask_nth_and(cpu, policy->cpus,
>+						  housekeeping_cpumask(HK_TYPE_TICK));
>+		cpufreq_cpu_put(policy);
>+		if (ref_cpu >= nr_cpu_ids)
>+			return 0;
>+		cpu = ref_cpu;
>+	}
>+
>+	/*
>+	 * Reversed computation to the one used to determine
>+	 * the arch_freq_scale value
>+	 * (see amu_scale_freq_tick for details)
>+	 */
>+	scale = per_cpu(arch_freq_scale, cpu);
>+	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT;
>+	freq *= scale;
>+	return freq;
>+}
>+
> static struct scale_freq_data amu_sfd = {
> 	.source = SCALE_FREQ_SOURCE_ARCH,
> 	.set_freq_scale = amu_scale_freq_tick,
>-- 
>2.25.1
>
>
>_______________________________________________
>linux-arm-kernel mailing list
>linux-arm-kernel@lists.infradead.org
>http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
diff mbox series

Patch

diff --git a/arch/arm64/kernel/topology.c b/arch/arm64/kernel/topology.c
index 615c1a20129f..ae2445f6e7da 100644
--- a/arch/arm64/kernel/topology.c
+++ b/arch/arm64/kernel/topology.c
@@ -17,6 +17,7 @@ 
 #include <linux/cpufreq.h>
 #include <linux/init.h>
 #include <linux/percpu.h>
+#include <linux/sched/isolation.h>
 
 #include <asm/cpu.h>
 #include <asm/cputype.h>
@@ -186,6 +187,44 @@  static void amu_scale_freq_tick(void)
 	this_cpu_write(arch_freq_scale, (unsigned long)scale);
 }
 
+unsigned int arch_freq_get_on_cpu(int cpu)
+{
+	unsigned int freq;
+	u64 scale;
+
+	if (!cpumask_test_cpu(cpu, amu_fie_cpus))
+		return 0;
+
+	/*
+	 * For those CPUs that are in full dynticks mode, try an alternative
+	 * source for the counters (and thus freq scale),
+	 * if available for given policy
+	 */
+	if (!housekeeping_cpu(cpu, HK_TYPE_TICK)) {
+		struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
+		int ref_cpu = nr_cpu_ids;
+
+		if (cpumask_intersects(housekeeping_cpumask(HK_TYPE_TICK),
+				       policy->cpus))
+			ref_cpu = cpumask_nth_and(cpu, policy->cpus,
+						  housekeeping_cpumask(HK_TYPE_TICK));
+		cpufreq_cpu_put(policy);
+		if (ref_cpu >= nr_cpu_ids)
+			return 0;
+		cpu = ref_cpu;
+	}
+
+	/*
+	 * Reversed computation to the one used to determine
+	 * the arch_freq_scale value
+	 * (see amu_scale_freq_tick for details)
+	 */
+	scale = per_cpu(arch_freq_scale, cpu);
+	freq = cpufreq_get_hw_max_freq(cpu) >> SCHED_CAPACITY_SHIFT;
+	freq *= scale;
+	return freq;
+}
+
 static struct scale_freq_data amu_sfd = {
 	.source = SCALE_FREQ_SOURCE_ARCH,
 	.set_freq_scale = amu_scale_freq_tick,