CPU frequency cannot be boosted when using the amd_pstate driver in
passive or guided mode.
On a host that has an AMD EPYC 7662 processor, while running with
amd-pstate configured for passive mode on full CPU load, the processor
only reaches 2.0 GHz. On later kernels the CPU can reach 3.3GHz.
The CPU frequency is dependent on a setting called highest_perf which is
the multiplier used to compute it. The highest_perf value comes from
cppc_init_perf when the driver is built-in and from pstate_init_perf when
it is a loaded module. Both of these calls have the following condition:
highest_perf = amd_get_highest_perf();
if (highest_perf > __cppc_highest_perf_)
highest_perf = __cppc_highest_perf;
Where again __cppc_highest_perf is either the return from
cppc_get_perf_caps in the built-in case or AMD_CPPC_HIGHEST_PERF in the
module case. Both of these functions actually return the nominal value,
whereas the call to amd_get_highest_perf returns the correct boost value,
so the condition tests true and highest_perf always ends up being the
nominal value, therefore never having the ability to boost CPU frequency.
Since amd_get_highest_perf already returns the boost value, we have
eliminated this check.
The issue was introduced in v6.1 via commit bedadcfb011f ("cpufreq:
amd-pstate: Fix initial highest_perf value"), and exists in stable v6.1
kernels. This has been fixed in v6.6.y and newer but due to refactoring that
change isn't feasible to bring back to v6.1.y. Thus, v6.1 kernels are
affected by this significant performance issue, and cannot be easily
remediated.
Signed-off-by: Nabil S. Alramli <dev@xxxxxxxxxxxx>
Reviewed-by: Joe Damato <jdamato@xxxxxxxxxx>
Reviewed-by: Kyle Hubert <khubert@xxxxxxxxxx>
Fixes: bedadcfb011f ("cpufreq: amd-pstate: Fix initial highest_perf value")
See-also: 1ec40a175a48 ("cpufreq: amd-pstate: Enable amd-pstate preferred core support")
Cc: mario.limonciello@xxxxxxx
Cc: Perry.Yuan@xxxxxxx
Cc: li.meng@xxxxxxx
Cc: stable@xxxxxxxxxxxxxxx # v6.1
---
v2:
- Omit cover letter
- Converted from RFC to PATCH
- Expand commit message based on feedback from Mario Limonciello
- Added Reviewed-by tags
- No functional/code changes
rfc:
https://lore.kernel.org/lkml/20241025010527.491605-1-dev@xxxxxxxxxxxx/
---
drivers/cpufreq/amd-pstate.c | 8 ++------
1 file changed, 2 insertions(+), 6 deletions(-)
diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 90dcf26f0973..c66086ae624a 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -102,9 +102,7 @@ static int pstate_init_perf(struct amd_cpudata *cpudata)
*
* CPPC entry doesn't indicate the highest performance in some ASICs.
*/
- highest_perf = amd_get_highest_perf();
- if (highest_perf > AMD_CPPC_HIGHEST_PERF(cap1))
- highest_perf = AMD_CPPC_HIGHEST_PERF(cap1);
+ highest_perf = max(amd_get_highest_perf(), AMD_CPPC_HIGHEST_PERF(cap1));
WRITE_ONCE(cpudata->highest_perf, highest_perf);
@@ -124,9 +122,7 @@ static int cppc_init_perf(struct amd_cpudata *cpudata)
if (ret)
return ret;
- highest_perf = amd_get_highest_perf();
- if (highest_perf > cppc_perf.highest_perf)
- highest_perf = cppc_perf.highest_perf;
+ highest_perf = max(amd_get_highest_perf(), cppc_perf.highest_perf);
WRITE_ONCE(cpudata->highest_perf, highest_perf);
--
2.35.1
