[PATCH 0/2] cpufreq/amd-pstate register write fixes

[PATCH 0/2] cpufreq/amd-pstate register write fixes

[Christian Loehle]

Two issues that popped up when trying to integrate an AMD machine into
my testing for acpi-cppc, in particular for this (otherwise
unrelated) patch:
https://lore.kernel.org/lkml/20261020121846.691218-2-sumitg@nvidia.com/

The first is rather straightforward, I think the case was just not
considered during the refactor of the Fixes commit.
The second is part of the resume handling and was introduced when
that restored the register values (DESIRED derived from policy->cur).
My platform seems to handle it well, but it's still strange
behaviour, so why not fix it.
(Also of course the resume value will generally be overwritten and
therefore 'fixed' anyway.)

Christian Loehle (2):
  cpufreq/amd-pstate: Keep minimum requests within policy limits
  cpufreq/amd-pstate: Clamp desired perf on resume

 drivers/cpufreq/amd-pstate.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)


base-commit: c339f4010dea8b0d754f84ace70cd1b089878ef8
-- 
2.34.1

[PATCH 1/2] cpufreq/amd-pstate: Keep minimum requests within policy limits

[Christian Loehle]

The minimum performance hint can exceed the policy maximum, e.g.
for schedutil when an RT task requests full CPU capacity. After
converting that hint, amd_pstate_adjust_perf() raises max_perf to
min_perf. This replaces the policy cap with the scheduler's minimum
request, and the subsequent Desired clamp uses the expanded interval.

Clamp min_perf to max_limit_perf instead of raising max_perf, to
enforce the policy's upper-bound again.

Fixes: 3f7b835fa4d0 ("cpufreq/amd-pstate: Move limit updating code")
Signed-off-by: Christian Loehle <christian.loehle@arm.com>
---
 drivers/cpufreq/amd-pstate.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 4bedfbfc18b6..28df8507d93f 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -866,8 +866,7 @@ static void amd_pstate_adjust_perf(struct cpufreq_policy *policy,
 		min_perf = perf.min_limit_perf;
 
 	max_perf = perf.max_limit_perf;
-	if (max_perf < min_perf)
-		max_perf = min_perf;
+	min_perf = min(min_perf, max_perf);
 
 	amd_pstate_update(policy, min_perf, des_perf, max_perf, true,
 			policy->governor->flags);
-- 
2.34.1

[PATCH 2/2] cpufreq/amd-pstate: Clamp desired perf on resume

[Christian Loehle]

amd_pstate_resume() converts policy->cur to a performance level and
writes it directly alongside the cached policy limits. Both
frequency/performance conversions round up, so the converted value
can exceed max_limit_perf even when the request before suspend was
within the policy limits.

On a Ryzen 5 7430U with a fixed 1.2 GHz policy, a request with
Desired=46 becomes policy->cur=1216621 kHz and converts back to
Desired=47. Resume then writes MIN=46, Desired=47, MAX=46.

Clamp the converted Desired to the cached minimum and maximum before
restoring the request.

Fixes: 608a76b65288 ("cpufreq/amd-pstate: Add support for the "Requested CPU Min frequency" BIOS option")
Signed-off-by: Christian Loehle <christian.loehle@arm.com>
---
 drivers/cpufreq/amd-pstate.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 28df8507d93f..0b1b0e263577 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -2211,6 +2211,7 @@ static int amd_pstate_resume(struct cpufreq_policy *policy)
 	int ret;
 
 	/* Set CPPC_REQ to last sane value until the governor updates it */
+	cur_perf = clamp(cur_perf, perf.min_limit_perf, perf.max_limit_perf);
 	ret = amd_pstate_update_perf(policy, perf.min_limit_perf, cur_perf, perf.max_limit_perf,
 				     0U, false);
 	if (ret)
-- 
2.34.1

Re: [PATCH 0/2] cpufreq/amd-pstate register write fixes

[Mario Limonciello]

On 10/29/26 07:19, Christian Loehle wrote:
> Two issues that popped up when trying to integrate an AMD machine into
> my testing for acpi-cppc, in particular for this (otherwise
> unrelated) patch:
> https://lore.kernel.org/lkml/20261020121846.691218-2-sumitg@nvidia.com/
> 
> The first is rather straightforward, I think the case was just not
> considered during the refactor of the Fixes commit.
> The second is part of the resume handling and was introduced when
> that restored the register values (DESIRED derived from policy->cur).
> My platform seems to handle it well, but it's still strange
> behaviour, so why not fix it.
> (Also of course the resume value will generally be overwritten and
> therefore 'fixed' anyway.)
> 
> Christian Loehle (2):
>    cpufreq/amd-pstate: Keep minimum requests within policy limits
>    cpufreq/amd-pstate: Clamp desired perf on resume
> 
>   drivers/cpufreq/amd-pstate.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
> 
> 
> base-commit: c339f4010dea8b0d754f84ace70cd1b089878ef8

Thanks for the fixes.  I'll pick them up for 7.4-rc2.

Acked-by: Mario Limociello (AMD) <superm1@kernel.org>

Re: [PATCH 0/2] cpufreq/amd-pstate register write fixes

[K Prateek Nayak]

Hello Chris,

On 10/29/2026 5:49 PM, Christian Loehle wrote:
> Two issues that popped up when trying to integrate an AMD machine into
> my testing for acpi-cppc, in particular for this (otherwise
> unrelated) patch:
> https://lore.kernel.org/lkml/20261020121846.691218-2-sumitg@nvidia.com/
> 
> The first is rather straightforward, I think the case was just not
> considered during the refactor of the Fixes commit.
> The second is part of the resume handling and was introduced when
> that restored the register values (DESIRED derived from policy->cur).
> My platform seems to handle it well, but it's still strange
> behaviour, so why not fix it.
> (Also of course the resume value will generally be overwritten and
> therefore 'fixed' anyway.)

Thank you for catching them. For the whole series, feel free to include:

Reviewed-by: K Prateek Nayak <kprateek.nayak@amd.com>
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>

-- 
Thanks and Regards,
Prateek