Re: [PATCH 10/11] cpufreq: amd-pstate: fix the highest frequency issue which limit performance

From: Mario Limonciello
Date: Tue May 07 2024 - 11:23:50 EST


On 5/7/2024 02:15, Perry Yuan wrote:
To address the performance drop issue, an optimization has been implemented.
The incorrect highest performance value previously set by the low-level power
firmware for AMD CPUs with Family ID 0x19 and Model ID ranging from 0x70 to 0x7F
series has been identified as the cause.

To resolve this, a check has been implemented to accurately determine the CPU family
and model ID. The correct highest performance value is now set and the performance
drop caused by the incorrect highest performance value are eliminated.

Before the fix, the highest frequency was set to 4200MHz, now it is set
to 4971MHz which is correct.

CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE MAXMHZ MINMHZ MHZ
0 0 0 0 0:0:0:0 yes 4971.0000 400.0000 400.0000
1 0 0 0 0:0:0:0 yes 4971.0000 400.0000 400.0000
2 0 0 1 1:1:1:0 yes 4971.0000 400.0000 4865.8140
3 0 0 1 1:1:1:0 yes 4971.0000 400.0000 400.0000

Closes: https://bugzilla.kernel.org/show_bug.cgi?id=218759
Signed-off-by: Perry Yuan <perry.yuan@xxxxxxx>
---
drivers/cpufreq/amd-pstate.c | 8 ++++++++
1 file changed, 8 insertions(+)

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 7fe8a8fc6227..3ff381c4edf7 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -348,6 +348,7 @@ static u32 amd_pstate_highest_perf_set(struct amd_cpudata *cpudata)
{
u32 highest_perf;
int core_type;
+ struct cpuinfo_x86 *c = &cpu_data(0);
core_type = amd_pstate_get_cpu_type(cpudata->cpu);
pr_debug("core_type %d found\n", core_type);
@@ -355,6 +356,13 @@ static u32 amd_pstate_highest_perf_set(struct amd_cpudata *cpudata)
switch (core_type) {
case CPU_CORE_TYPE_NO_HETERO_SUP:
highest_perf = CPPC_HIGHEST_PERF_DEFAULT;
+ /*
+ * For AMD CPUs with Family ID 19H and Model ID range 0x70 to 0x7F,
+ * the highest performance level is set to 196.
+ * https://bugzilla.kernel.org/show_bug.cgi?id=218759
+ */
+ if (c->x86 == 0x19 && (c->x86_model >= 0x70 && c->x86_model <= 0x7F))
+ highest_perf = CPPC_HIGHEST_PERF_PERFORMANCE;

I agree this is the right type of change to make for the reported issue, but since it's actually a performance regression from 6.9, can you move this to the start of the series and add a Fixes tag and stable tag so we can get the regression fixed for 6.10 and 6.9.y?

This will of course mean you need to adjust patch 9 as well for such a change, but I think it's better this specific patch goes into 6.10 as a fix and the rest of the series can aim for 6.11.

break;
case CPU_CORE_TYPE_PERFORMANCE:
highest_perf = CPPC_HIGHEST_PERF_PERFORMANCE;