diff mbox series

[3/3] ACPI: cpufreq: Add ITMT support when CPPC enabled for Zhaoxin CPUs

Message ID 20231228075705.26652-4-TonyWWang-oc@zhaoxin.com
State New
Headers show
Series [1/3] ACPI: CPPC: Add get the highest perf register value support | expand

Commit Message

Tony W Wang-oc Dec. 28, 2023, 7:57 a.m. UTC
For Zhaoxin CPUs, the cores' highest frequencies may be different, which
means that cores may run at different max frequencies,

According to ACPI-spec6 chapter 8.4.7, the per-core highest frequency
value can be obtained via cppc.

The core with the higher frequency have better performance, which can be
called as preferred core. And better performance can be achieved by
making the scheduler to run tasks on these preferred cores.

The cpufreq driver can use the highest frequency value as the prioriy of
core to make the scheduler try to get better performace. More specifically,
in the acpi-cpufreq driver use cppc_get_highest_perf() to get highest
frequency value of each core, use sched_set_itmt_core_prio() to set
highest frequency value as core priority, and use sched_set_itmt_support()
provided by ITMT to tell the scheduler to favor on the preferred cores.

Signed-off-by: Tony W Wang-oc <TonyWWang-oc@zhaoxin.com>
---
 drivers/cpufreq/acpi-cpufreq.c | 56 +++++++++++++++++++++++++++++++++-
 1 file changed, 55 insertions(+), 1 deletion(-)

Comments

Tony W Wang-oc Jan. 4, 2024, 6:26 a.m. UTC | #1
On 2024/1/3 23:37, Rafael J. Wysocki wrote:
> On Thu, Dec 28, 2023 at 8:57 AM Tony W Wang-oc <TonyWWang-oc@zhaoxin.com> wrote:
>> For Zhaoxin CPUs, the cores' highest frequencies may be different, which
>> means that cores may run at different max frequencies,
>>
>> According to ACPI-spec6 chapter 8.4.7, the per-core highest frequency
>> value can be obtained via cppc.
>>
>> The core with the higher frequency have better performance, which can be
>> called as preferred core. And better performance can be achieved by
>> making the scheduler to run tasks on these preferred cores.
>>
>> The cpufreq driver can use the highest frequency value as the prioriy of
>> core to make the scheduler try to get better performace. More specifically,
>> in the acpi-cpufreq driver use cppc_get_highest_perf() to get highest
>> frequency value of each core, use sched_set_itmt_core_prio() to set
>> highest frequency value as core priority, and use sched_set_itmt_support()
>> provided by ITMT to tell the scheduler to favor on the preferred cores.
>>
>> Signed-off-by: Tony W Wang-oc <TonyWWang-oc@zhaoxin.com>
>> ---
>>   drivers/cpufreq/acpi-cpufreq.c | 56 +++++++++++++++++++++++++++++++++-
>>   1 file changed, 55 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/cpufreq/acpi-cpufreq.c b/drivers/cpufreq/acpi-cpufreq.c
>> index 37f1cdf46d29..f4c1ff9e4bb0 100644
>> --- a/drivers/cpufreq/acpi-cpufreq.c
>> +++ b/drivers/cpufreq/acpi-cpufreq.c
>> @@ -663,8 +663,56 @@ static u64 get_max_boost_ratio(unsigned int cpu)
>>
>>          return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
>>   }
>> +
>> +/* The work item is needed to avoid CPU hotplug locking issues */
>> +static void sched_itmt_work_fn(struct work_struct *work)
>> +{
>> +       sched_set_itmt_support();
>> +}
>> +
>> +static DECLARE_WORK(sched_itmt_work, sched_itmt_work_fn);
>> +
>> +static void set_itmt_prio(int cpu)
>> +{
>> +       static bool cppc_highest_perf_diff;
>> +       static struct cpumask core_prior_mask;
>> +       u64 highest_perf;
>> +       static u64 max_highest_perf = 0, min_highest_perf = U64_MAX;
>> +       int ret;
>> +
>> +       ret = cppc_get_highest_perf(cpu, &highest_perf);
>> +       if (ret)
>> +               return;
>> +
>> +       sched_set_itmt_core_prio(highest_perf, cpu);
>> +       cpumask_set_cpu(cpu, &core_prior_mask);
>> +
>> +       if (max_highest_perf <= min_highest_perf) {
>> +               if (highest_perf > max_highest_perf)
>> +                       max_highest_perf = highest_perf;
>> +
>> +               if (highest_perf < min_highest_perf)
>> +                       min_highest_perf = highest_perf;
>> +
>> +               if (max_highest_perf > min_highest_perf) {
>> +                       /*
>> +                        * This code can be run during CPU online under the
>> +                        * CPU hotplug locks, so sched_set_itmt_support()
>> +                        * cannot be called from here.  Queue up a work item
>> +                        * to invoke it.
>> +                        */
>> +                       cppc_highest_perf_diff = true;
>> +               }
>> +       }
>> +
>> +       if (cppc_highest_perf_diff && cpumask_equal(&core_prior_mask, cpu_online_mask)) {
>> +               pr_debug("queue a work to set itmt enabled\n");
>> +               schedule_work(&sched_itmt_work);
>> +       }
>> +}
>>   #else
>>   static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
>> +static void set_itmt_prio(int cpu) { }
>>   #endif
>>
>>   static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
>> @@ -677,7 +725,7 @@ static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
>>          unsigned int valid_states = 0;
>>          unsigned int result = 0;
>>          u64 max_boost_ratio;
>> -       unsigned int i;
>> +       unsigned int i, j;
>>   #ifdef CONFIG_SMP
>>          static int blacklisted;
>>   #endif
>> @@ -742,6 +790,12 @@ static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
>>          }
>>   #endif
>>
>> +       if (c->x86_vendor == X86_VENDOR_CENTAUR || c->x86_vendor == X86_VENDOR_ZHAOXIN) {
>> +               for_each_cpu(j, policy->cpus) {
>> +                       set_itmt_prio(j);
>> +               }
>> +       }
>> +
>>          /* capability check */
>>          if (perf->state_count <= 1) {
>>                  pr_debug("No P-States\n");
>> --
> Have you considered using the CPPC cpufreq driver on those platforms?


Thanks for your reply.
The ACPI cpufreq driver is used by default on Zhaoxin platforms. We added
Zhaoxin preferred core support and did related tests based on the ACPI
cpufreq driver.
The CPPC cpufreq driver is currently used on the ARM platforms. We have
not yet considered using the CPPC cpufreq driver to support the Zhaoxin
preferred core feature, and we also unclear how well the CPPC cpufreq
driver works for the X86 platform.
At the moment, it seems that it is more appropriate to add Zhaoxin preferred
core support to the ACPI cpufreq Driver.

Sincerely
TonyWWangoc
Ricardo Neri Jan. 17, 2024, 10:51 p.m. UTC | #2
On Thu, Dec 28, 2023 at 03:57:05PM +0800, Tony W Wang-oc wrote:
> For Zhaoxin CPUs, the cores' highest frequencies may be different, which
> means that cores may run at different max frequencies,
> 
> According to ACPI-spec6 chapter 8.4.7, the per-core highest frequency
> value can be obtained via cppc.
> 
> The core with the higher frequency have better performance, which can be
> called as preferred core. And better performance can be achieved by
> making the scheduler to run tasks on these preferred cores.
> 
> The cpufreq driver can use the highest frequency value as the prioriy of
> core to make the scheduler try to get better performace. More specifically,
> in the acpi-cpufreq driver use cppc_get_highest_perf() to get highest
> frequency value of each core, use sched_set_itmt_core_prio() to set
> highest frequency value as core priority, and use sched_set_itmt_support()
> provided by ITMT to tell the scheduler to favor on the preferred cores.
> 
> Signed-off-by: Tony W Wang-oc <TonyWWang-oc@zhaoxin.com>
> ---
>  drivers/cpufreq/acpi-cpufreq.c | 56 +++++++++++++++++++++++++++++++++-
>  1 file changed, 55 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/cpufreq/acpi-cpufreq.c b/drivers/cpufreq/acpi-cpufreq.c
> index 37f1cdf46d29..f4c1ff9e4bb0 100644
> --- a/drivers/cpufreq/acpi-cpufreq.c
> +++ b/drivers/cpufreq/acpi-cpufreq.c
> @@ -663,8 +663,56 @@ static u64 get_max_boost_ratio(unsigned int cpu)
>  
>  	return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
>  }
> +
> +/* The work item is needed to avoid CPU hotplug locking issues */
> +static void sched_itmt_work_fn(struct work_struct *work)
> +{
> +	sched_set_itmt_support();
> +}
> +
> +static DECLARE_WORK(sched_itmt_work, sched_itmt_work_fn);
> +
> +static void set_itmt_prio(int cpu)
> +{
> +	static bool cppc_highest_perf_diff;
> +	static struct cpumask core_prior_mask;
> +	u64 highest_perf;
> +	static u64 max_highest_perf = 0, min_highest_perf = U64_MAX;
> +	int ret;
> +
> +	ret = cppc_get_highest_perf(cpu, &highest_perf);
> +	if (ret)
> +		return;
> +
> +	sched_set_itmt_core_prio(highest_perf, cpu);
> +	cpumask_set_cpu(cpu, &core_prior_mask);
> +
> +	if (max_highest_perf <= min_highest_perf) {
> +		if (highest_perf > max_highest_perf)
> +			max_highest_perf = highest_perf;
> +
> +		if (highest_perf < min_highest_perf)
> +			min_highest_perf = highest_perf;
> +
> +		if (max_highest_perf > min_highest_perf) {
> +			/*
> +			 * This code can be run during CPU online under the
> +			 * CPU hotplug locks, so sched_set_itmt_support()
> +			 * cannot be called from here.  Queue up a work item
> +			 * to invoke it.
> +			 */
> +			cppc_highest_perf_diff = true;
> +		}
> +	}
> +
> +	if (cppc_highest_perf_diff && cpumask_equal(&core_prior_mask, cpu_online_mask)) {
> +		pr_debug("queue a work to set itmt enabled\n");
> +		schedule_work(&sched_itmt_work);
> +	}
> +}

sched_itmt_work and this function is a duplicate of what the intel_pstate
driver already does. It might be good if consolidate in a single place
if you are going to pursue this approach.
diff mbox series

Patch

diff --git a/drivers/cpufreq/acpi-cpufreq.c b/drivers/cpufreq/acpi-cpufreq.c
index 37f1cdf46d29..f4c1ff9e4bb0 100644
--- a/drivers/cpufreq/acpi-cpufreq.c
+++ b/drivers/cpufreq/acpi-cpufreq.c
@@ -663,8 +663,56 @@  static u64 get_max_boost_ratio(unsigned int cpu)
 
 	return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
 }
+
+/* The work item is needed to avoid CPU hotplug locking issues */
+static void sched_itmt_work_fn(struct work_struct *work)
+{
+	sched_set_itmt_support();
+}
+
+static DECLARE_WORK(sched_itmt_work, sched_itmt_work_fn);
+
+static void set_itmt_prio(int cpu)
+{
+	static bool cppc_highest_perf_diff;
+	static struct cpumask core_prior_mask;
+	u64 highest_perf;
+	static u64 max_highest_perf = 0, min_highest_perf = U64_MAX;
+	int ret;
+
+	ret = cppc_get_highest_perf(cpu, &highest_perf);
+	if (ret)
+		return;
+
+	sched_set_itmt_core_prio(highest_perf, cpu);
+	cpumask_set_cpu(cpu, &core_prior_mask);
+
+	if (max_highest_perf <= min_highest_perf) {
+		if (highest_perf > max_highest_perf)
+			max_highest_perf = highest_perf;
+
+		if (highest_perf < min_highest_perf)
+			min_highest_perf = highest_perf;
+
+		if (max_highest_perf > min_highest_perf) {
+			/*
+			 * This code can be run during CPU online under the
+			 * CPU hotplug locks, so sched_set_itmt_support()
+			 * cannot be called from here.  Queue up a work item
+			 * to invoke it.
+			 */
+			cppc_highest_perf_diff = true;
+		}
+	}
+
+	if (cppc_highest_perf_diff && cpumask_equal(&core_prior_mask, cpu_online_mask)) {
+		pr_debug("queue a work to set itmt enabled\n");
+		schedule_work(&sched_itmt_work);
+	}
+}
 #else
 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
+static void set_itmt_prio(int cpu) { }
 #endif
 
 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
@@ -677,7 +725,7 @@  static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 	unsigned int valid_states = 0;
 	unsigned int result = 0;
 	u64 max_boost_ratio;
-	unsigned int i;
+	unsigned int i, j;
 #ifdef CONFIG_SMP
 	static int blacklisted;
 #endif
@@ -742,6 +790,12 @@  static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 	}
 #endif
 
+	if (c->x86_vendor == X86_VENDOR_CENTAUR || c->x86_vendor == X86_VENDOR_ZHAOXIN) {
+		for_each_cpu(j, policy->cpus) {
+			set_itmt_prio(j);
+		}
+	}
+
 	/* capability check */
 	if (perf->state_count <= 1) {
 		pr_debug("No P-States\n");