Re: [PATCH v2 1/5] cpufreq/amd-pstate: Add dynamic energy performance preference

From: Mario Limonciello
Date: Wed Mar 19 2025 - 13:17:14 EST


On 3/19/2025 11:50, Mario Limonciello wrote:
On 3/18/2025 22:43, Dhananjay Ugwekar wrote:
On 3/19/2025 1:06 AM, Mario Limonciello wrote:
On 3/12/2025 07:16, Dhananjay Ugwekar wrote:
On 3/4/2025 8:53 PM, Mario Limonciello wrote:
From: Mario Limonciello <mario.limonciello@xxxxxxx>

Dynamic energy performance preference will change the EPP profile
based on whether the machine is running on AC or DC power.

A notification chain from the power supply core is used to adjust
EPP values on plug in or plug out events.

For non-server systems:
      * the default EPP for AC mode is `performance`.
      * the default EPP for DC mode is `balance_performance`.

Signed-off-by: Mario Limonciello <mario.limonciello@xxxxxxx>
---
v1->v2:
   * Change defaults to performance (AC) and balance_performance (DC)
   * Default Kconfig to disabled for now
   * Rebase on latest branch
---
   Documentation/admin-guide/pm/amd-pstate.rst |  18 ++-
   drivers/cpufreq/Kconfig.x86                 |  12 ++
   drivers/cpufreq/amd-pstate.c                | 129 ++++++++++++++ ++++--
   drivers/cpufreq/amd-pstate.h                |   5 +-
   4 files changed, 155 insertions(+), 9 deletions(-)

[Snip]
@@ -1556,6 +1667,10 @@ static int amd_pstate_epp_set_policy(struct cpufreq_policy *policy)
       if (!policy->cpuinfo.max_freq)
           return -ENODEV;
   +    /* policy can't be changed to performance policy while dynamic epp is enabled */
+    if (policy->policy == CPUFREQ_POLICY_PERFORMANCE && cpudata- >dynamic_epp)
+        return -EBUSY;

We might need to tweak this condition, because if we enable "CONFIG_X86_AMD_PSTATE_DYNAMIC_EPP" in config
and boot with "amd_pstate=active" it lands here (cpufreq_online()- >amd_pstate_epp_set_policy()) driver init fails
as the default governor is performance.


The check is important to make sure that you can't go to performance mode after init.

I think this is the way I would want to solve it.
Set policy to powersave before enabling dynamic epp for amd_pstate_epp_cpu_init().

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 824756ac0010e..4a0f561d0e2d1 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -1729,8 +1729,10 @@ static int amd_pstate_epp_cpu_init(struct cpufreq_policy *policy)
                 WRITE_ONCE(cpudata->cppc_req_cached, value);
         }

-       if (dynamic_epp)
+       if (dynamic_epp) {
+               policy->policy = CPUFREQ_POLICY_POWERSAVE;
                 ret = amd_pstate_set_dynamic_epp(policy);
+       }
         else
                 ret = amd_pstate_set_epp(policy, amd_pstate_get_balanced_epp(policy));
            ^^^^^^^^^^^^ (mentioned below)
         if (ret)

Thoughts?

Yes, this looks good, because anyway there is no point in having performance governor and dynamic
epp set at the same time.

I found one related quirk though, we are setting performance governor for server platforms in
amd_pstate_epp_cpu_init() and then setting epp at the line highlighted above. We dont have a
check in *_set_epp() functions for performance governor. This could alter the performance governor
behavior if we set a "balanced" epp for it. I haven't tested this part yet.

In that case we probably want the "default" ACPI platform profile to be "performance" when on a server instead of balanced.

Actually this is already done:

cpudata->current_profile = PLATFORM_PROFILE_PERFORMANCE;

I'll post an updated patch series for you to test.



Thanks,
Dhananjay


+
       cpudata->policy = policy->policy;
         ret = amd_pstate_epp_update_limit(policy);
diff --git a/drivers/cpufreq/amd-pstate.h b/drivers/cpufreq/amd- pstate.h
index fbe1c08d3f061..6882876f895de 100644
--- a/drivers/cpufreq/amd-pstate.h
+++ b/drivers/cpufreq/amd-pstate.h
@@ -104,7 +104,10 @@ struct amd_cpudata {
       /* EPP feature related attributes*/
       u32    policy;
       bool    suspended;
-    u8    epp_default;
+    u8    epp_default_ac;
+    u8    epp_default_dc;
+    bool    dynamic_epp;
+    struct notifier_block power_nb;
   };
     /*