Linux Power Management development
 help / color / mirror / Atom feed
From: K Prateek Nayak <kprateek.nayak@amd.com>
To: Vishal Badole <Vishal.Badole@amd.com>, <ray.huang@amd.com>,
	<mario.limonciello@amd.com>, <rafael@kernel.org>,
	<viresh.kumar@linaro.org>, <perry.yuan@amd.com>,
	<hansg@kernel.org>, <ilpo.jarvinen@linux.intel.com>,
	<platform-driver-x86@vger.kernel.org>, <linux-pm@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 1/2] cpufreq: amd-pstate: Set asymmetric CPU capacity on hybrid platforms
Date: Tue, 17 Nov 2026 11:49:31 +0530	[thread overview]
Message-ID: <69aee432-502b-4013-83c6-dde70cb59629@amd.com> (raw)
In-Reply-To: <20261114150310.1377654-2-Vishal.Badole@amd.com>

Hello Vishal,

On 11/14/2026 8:33 PM, Vishal Badole wrote:
> diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
> index 4bedfbfc18b6..59eb2c4c7867 100644
> --- a/drivers/cpufreq/amd-pstate.c
> +++ b/drivers/cpufreq/amd-pstate.c
> @@ -29,6 +29,7 @@
>  #include <linux/smp.h>
>  #include <linux/sched.h>
>  #include <linux/cpufreq.h>
> +#include <linux/sched/smt.h>
>  #include <linux/compiler.h>
>  #include <linux/dmi.h>
>  #include <linux/slab.h>
> @@ -39,6 +40,8 @@
>  #include <linux/power_supply.h>
>  #include <linux/static_call.h>
>  #include <linux/topology.h>
> +#include <linux/cpuset.h>
> +#include <linux/amd-pstate.h>
>  
>  #include <acpi/processor.h>
>  #include <acpi/cppc_acpi.h>
> @@ -86,6 +89,10 @@ static struct cpufreq_driver amd_pstate_driver;
>  static struct cpufreq_driver amd_pstate_epp_driver;
>  static int cppc_state = AMD_PSTATE_UNDEFINED;
>  static bool amd_pstate_prefcore = true;
> +static bool amd_pstate_cas = true;
> +static bool amd_pstate_cas_initialized;
> +static u32 amd_pstate_ref_freq_khz;
> +static u32 amd_pstate_ref_highest_perf;
>  static struct quirk_entry *quirks;
>  
>  /*
> @@ -957,8 +964,8 @@ static void amd_pstate_init_prefcore(struct amd_cpudata *cpudata)
>  		return;
>  
>  	/* should use amd-hfi instead */
> -	if (cpu_feature_enabled(X86_FEATURE_AMD_WORKLOAD_CLASS) &&
> -	    IS_ENABLED(CONFIG_AMD_HFI)) {
> +	if ((cpu_feature_enabled(X86_FEATURE_AMD_WORKLOAD_CLASS) &&
> +	     IS_ENABLED(CONFIG_AMD_HFI)) || amd_pstate_cas) {

This keeps getting uglier. I think you should have a preperatory patch to
negotiate this and maybe a late initcall like Mario suggested to actually
toggle the itmt support.

I was thinking something like folows to negotiate the who / what gets
the owenership:

diff --git a/arch/x86/include/asm/amd/hfi.h b/arch/x86/include/asm/amd/hfi.h
new file mode 100644
index 000000000000..2a86d0049b8f
--- /dev/null
+++ b/arch/x86/include/asm/amd/hfi.h
@@ -0,0 +1,27 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+#ifndef __ASM_X86_AMD_HFI_H__
+#define __ASM_X86_AMD_HFI_H__
+
+enum amd_hfi_mode {
+	AMD_HFI_UNDEFINED,
+	AMD_HFI_PREFCORE,	/* amd-pstate defined preferred core. */
+	AMD_HFI_ACPI,		/* amd_hfi defined from ACPI PCCT. */
+	AMD_HFI_ASYM_CAPACITY,	/* Amd-pstate defined asymmetric capacity. */
+};
+
+static enum amd_hfi_mode amd_hfi_current_mode = AMD_HFI_UNDEFINED;
+
+static void amd_hfi_set_mode(enum amd_hfi_mode mode)
+{
+	if (mode <= amd_hfi_current_mode)
+		return;
+
+	amd_hfi_current_mode = mode;
+}
+
+static enum amd_hfi_mode amd_hfi_get_mode(enum amd_hfi_mode mode)
+{
+	return amd_hfi_current_mode;
+}
+
+#endif /* __ASM_X86_AMD_HFI_H__ */
---

>  		amd_pstate_prefcore = false;
>  		return;
>  	}
> @@ -969,6 +976,64 @@ static void amd_pstate_init_prefcore(struct amd_cpudata *cpudata)
>  	sched_set_itmt_core_prio((int)READ_ONCE(cpudata->prefcore_ranking), cpudata->cpu);
>  }
>  
> +bool amd_pstate_cas_active(void)
> +{
> +	return READ_ONCE(amd_pstate_cas_initialized);
> +}
> +EXPORT_SYMBOL_GPL(amd_pstate_cas_active);

You can then just use amd_hfi_get_mode() instead of this.

> +
> +static bool __init amd_pstate_init_cas(void)
> +{
> +	int highest_perf, max_freq;
> +
> +	highest_perf = amd_get_effective_highest_perf(0);

Didn't you mention an efficiency core / LP core could also be
the boot CPU? How are you considering it as your highest
reference then?

I think we can tak a page from how intel_pstate does a
hybrid_set_cpu_capacity() which iterates all online CPUs after driver
registration / if onlining changes the hybrid_max_perf_cpu.

> +	if (highest_perf <= 0) {
> +		pr_warn("Invalid reference highest performance (%d)\n", highest_perf);
> +		return false;
> +	}
> +
> +	max_freq = amd_get_max_frequency(0);
> +	if (max_freq < 0 || (u32)max_freq > U32_MAX / 1000) {
> +		pr_warn("Invalid reference maximum frequency (%d)\n", max_freq);
> +		return false;
> +	}
> +
> +	amd_pstate_ref_highest_perf = highest_perf;
> +	amd_pstate_ref_freq_khz = max_freq * 1000U;
> +
> +	return arch_enable_hybrid_capacity_scale();
> +}
> +
> +static void amd_pstate_set_cpu_capacity(struct amd_cpudata *cpudata,
> +					bool highest_freq_available)
> +{
> +	unsigned long capacity, max_capacity, capacity_freq, base_freq;
> +
> +	if (!IS_ENABLED(CONFIG_SMP))
> +		return;

What is the CONFIG_SMP dependnecy for?

> +
> +	if (highest_freq_available && amd_pstate_ref_freq_khz) {
> +		capacity = cpudata->max_freq;
> +		max_capacity = amd_pstate_ref_freq_khz;
> +		capacity_freq = capacity;
> +		base_freq = max_capacity;
> +	} else {
> +		union perf_cached perf = READ_ONCE(cpudata->perf);
> +
> +		capacity = perf.highest_perf;
> +		max_capacity = amd_pstate_ref_highest_perf;

You can just cache the freq corresponding to
amd_pstate_ref_highest_perf and not need two seperate global
variables. 

> +		capacity_freq = perf_to_freq(perf, cpudata->nominal_freq, perf.highest_perf);
> +		base_freq = perf_to_freq(perf, cpudata->nominal_freq, max_capacity);
> +	}
> +
> +	guard(mutex)(&sched_domains_mutex);

Why "sched_domains_mutex"? Probably means you are setting
x86_topology_update at the wrong time.

> +	arch_set_cpu_capacity(cpudata->cpu, capacity, max_capacity,
> +			      capacity_freq, base_freq);
> +	topology_set_cpu_scale(cpudata->cpu, arch_scale_cpu_capacity(cpudata->cpu));
> +	x86_topology_update = true;
> +	WRITE_ONCE(amd_pstate_cas_initialized, true);
> +}
> +
>  static void amd_pstate_update_limits(struct cpufreq_policy *policy)
>  {
>  	struct amd_cpudata *cpudata;
> @@ -1043,6 +1108,7 @@ static int amd_pstate_init_freq(struct amd_cpudata *cpudata)
>  	u32 min_freq, max_freq, nominal_freq, lowest_nonlinear_freq;
>  	struct cppc_perf_caps cppc_perf;
>  	union perf_cached perf;
> +	bool highest_freq_available;
>  	int ret;
>  
>  	ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf);
> @@ -1069,6 +1135,7 @@ static int amd_pstate_init_freq(struct amd_cpudata *cpudata)
>  
>  	/* try to look up from BIOS/quirk first, fall back to (nominal_freq * highest_perf)/nominal_perf */
>  	max_freq = amd_get_max_frequency(cpudata->cpu) * 1000;
> +	highest_freq_available = !!max_freq;
>  	if (!max_freq)
>  		max_freq = perf_to_freq(perf, nominal_freq, perf.highest_perf);
>  	WRITE_ONCE(cpudata->max_freq, max_freq);
> @@ -1100,6 +1167,9 @@ static int amd_pstate_init_freq(struct amd_cpudata *cpudata)
>  			 bios_min_freq, cpudata->cpu);
>  	}
>  
> +	if (amd_pstate_cas)
> +		amd_pstate_set_cpu_capacity(cpudata, highest_freq_available);
> +
>  	return 0;
>  }
>  
> @@ -1923,12 +1993,20 @@ static ssize_t prefcore_show(struct device *dev,
>  	return sysfs_emit(buf, "%s\n", str_enabled_disabled(amd_pstate_prefcore));
>  }
>  
> +static ssize_t cas_show(struct device *dev,
> +			struct device_attribute *attr, char *buf)
> +{
> +	return sysfs_emit(buf, "%s\n", str_enabled_disabled(amd_pstate_cas_active()));
> +}
> +
>  static DEVICE_ATTR_RW(status);
>  static DEVICE_ATTR_RO(prefcore);
> +static DEVICE_ATTR_RO(cas);
>  
>  static struct attribute *pstate_global_attributes[] = {
>  	&dev_attr_status.attr,
>  	&dev_attr_prefcore.attr,
> +	&dev_attr_cas.attr,
>  	NULL
>  };
>  
> @@ -2438,12 +2516,23 @@ static int __init amd_pstate_init(void)
>  			return ret;
>  	}
>  
> +	amd_pstate_cas = amd_pstate_cas && IS_ENABLED(CONFIG_SMP) &&
> +			cpu_feature_enabled(X86_FEATURE_AMD_HTR_CORES);
> +
> +	/* Enable CAS for processors with heterogeneous cores. */
> +	if (amd_pstate_cas)
> +		amd_pstate_cas = amd_pstate_init_cas();

That two step disable is pretty ugly. Why couldn't you push the
feature checks into amd_pstate_init_cas()?

> +
>  	ret = amd_pstate_register_driver(cppc_state);
>  	if (ret) {
>  		pr_err("failed to register with return %d\n", ret);
>  		return ret;
>  	}
>  
> +	/* Per-CPU capacities are set during registration; rebuild domains once. */
> +	if (amd_pstate_cas)
> +		rebuild_sched_domains();
> +

Isn't this super early into the boot? You don't need to rebuild sched
domains here since they will be done later (I'm having a deja vu moment
from a previous review)

>  	dev_root = bus_get_dev_root(&cpu_subsys);
>  	if (dev_root) {
>  		ret = sysfs_create_group(&dev_root->kobj, &amd_pstate_global_attr_group);
> @@ -2484,8 +2573,17 @@ static int __init amd_prefcore_param(char *str)
>  	return 0;
>  }
>  
> +static int __init amd_cas_param(char *str)
> +{
> +	if (!strcmp(str, "disable"))
> +		amd_pstate_cas = false;
> +
> +	return 0;
> +}
> +
>  early_param("amd_pstate", amd_pstate_param);
>  early_param("amd_prefcore", amd_prefcore_param);
> +early_param("amd_cas", amd_cas_param);
>  
>  MODULE_AUTHOR("Huang Rui <ray.huang@amd.com>");
>  MODULE_DESCRIPTION("AMD Processor P-state Frequency Driver");
> diff --git a/include/linux/amd-pstate.h b/include/linux/amd-pstate.h

Why not arch/x86/include/asm/amd/amd-pstate.h since all of this is
AMd related?

> new file mode 100644
> index 000000000000..eb750abd20f7
> --- /dev/null
> +++ b/include/linux/amd-pstate.h
> @@ -0,0 +1,14 @@
> +/* SPDX-License-Identifier: GPL-2.0-or-later */
> +/*
> + * Public interface for the AMD P-state driver.
> + */
> +#ifndef __LINUX_AMD_PSTATE_H__
> +#define __LINUX_AMD_PSTATE_H__
> +
> +#if IS_ENABLED(CONFIG_X86_AMD_PSTATE)
> +bool amd_pstate_cas_active(void);
> +#else
> +static inline bool amd_pstate_cas_active(void) { return false; }
> +#endif
> +
> +#endif /* __LINUX_AMD_PSTATE_H__ */

-- 
Thanks and Regards,
Prateek


  parent reply	other threads:[~2026-11-17  6:19 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-14 15:03 [PATCH 0/2] cpufreq: amd-pstate: Capacity Aware Scheduling on hybrid platforms Vishal Badole
2026-11-14 15:03 ` [PATCH 1/2] cpufreq: amd-pstate: Set asymmetric CPU capacity " Vishal Badole
2026-11-15 15:54   ` Mario Limonciello
2026-11-17  6:19   ` K Prateek Nayak [this message]
2026-11-14 15:03 ` [PATCH 2/2] platform/x86/amd: hfi: Stand down when amd-pstate CAS is active Vishal Badole
2026-11-15 14:16   ` Mario Limonciello
2026-11-15 14:22     ` Mario Limonciello

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=69aee432-502b-4013-83c6-dde70cb59629@amd.com \
    --to=kprateek.nayak@amd.com \
    --cc=Vishal.Badole@amd.com \
    --cc=hansg@kernel.org \
    --cc=ilpo.jarvinen@linux.intel.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-pm@vger.kernel.org \
    --cc=mario.limonciello@amd.com \
    --cc=perry.yuan@amd.com \
    --cc=platform-driver-x86@vger.kernel.org \
    --cc=rafael@kernel.org \
    --cc=ray.huang@amd.com \
    --cc=viresh.kumar@linaro.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox