Skip to contentSkip to conversation

x86: Do not override performance EPB MSR value on servers

PATCH In review Thomas Renninger 2 participants 3 messages Apr 8, 2019 – Apr 16, 2019 x86

lore
Share & tools
lore

Thomas Renninger

author
#0 lore raw

PATCH 1/1 x86: Do not override performance EPB MSR value on servers

It is about this kernel message showing up on quite a lot servers:

[    0.072652] ENERGY_PERF_BIAS: Set to 'normal', was 'performance'
[    0.076003] ENERGY_PERF_BIAS: View and update with x86_energy_perf_policy(8)

With this patch the EPB value is not overridden anymore on systems which expose themselves as appliance or enterprise server via ACPI perferred profile variable.

Signed-off-by [email protected]
arch/x86/kernel/cpu/intel_epb.c +6 -1
Index: perf_bias_acpi_server_only/arch/x86/kernel/cpu/intel_epb.c
===================================================================
@@ -19,6 +19,8 @@
 #include <asm/cpufeature.h>
 #include <asm/msr.h>
 
+#include <linux/acpi.h>
+
 /**
  * DOC: overview
  *

@@ -89,7 +91,10 @@ static void intel_epb_restore(void)
 		 * ('normal').
 		 */
 		val = epb & EPB_MASK;
-		if (val == ENERGY_PERF_BIAS_PERFORMANCE) {
+		if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
+		    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER) {
+			/* Keep performance setting on servers */
+		} else	if (val == ENERGY_PERF_BIAS_PERFORMANCE) {
 			val = ENERGY_PERF_BIAS_NORMAL;
 			pr_warn_once("ENERGY_PERF_BIAS: Set to 'normal', was 'performance'\n");
 		}
author
#1 lore raw
Quoted 10 lines · CEST Thomas Renninger

It is about this kernel message showing up on quite a lot servers:

[    0.072652] ENERGY_PERF_BIAS: Set to 'normal', was 'performance'
[    0.076003] ENERGY_PERF_BIAS: View and update with

x86_energy_perf_policy(8)

With this patch the EPB value is not overridden anymore on systems which expose themselves as appliance or enterprise server via ACPI perferred profile variable.

Signed-off-by [email protected]

The patch has successfully been tested on top of Rafeals latest EPB fixes:
[PATCH 0/2] PM / arch: x86: MSR_IA32_ENERGY_PERF_BIAS handling fixes and sysfs i/f

EPB value is now successfully and correctly restored after CPU off-/onlining It can be set back to performance value if kernel set it to normal and this value also survives CPU off-/onlining.

Still the kernel should not override the performance value on servers, so it would be nice to see the patch applied.
Once applied, I may send a patch based on code before EPB changes similar to this for stable kernels. There it is even more urgent, due to the wrong unconditional setting to "normal" value after CPU off-/onlining.

I tested the patch to make sure the preferred value is considered and it works.

Thomas

#2 lore raw
In reply to Thomas Renninger
Quoted 26 lines · Thomas Renninger

It is about this kernel message showing up on quite a lot servers:

[    0.072652] ENERGY_PERF_BIAS: Set to 'normal', was 'performance'
[    0.076003] ENERGY_PERF_BIAS: View and update with x86_energy_perf_policy(8)

With this patch the EPB value is not overridden anymore on systems which expose themselves as appliance or enterprise server via ACPI perferred profile variable.

Signed-off-by [email protected]

Index: perf_bias_acpi_server_only/arch/x86/kernel/cpu/intel_epb.c
===================================================================

arch/x86/kernel/cpu/intel_epb.c 2019-04-08 10:01:26.029278425 +0200 renamed from arch/x86/kernel/cpu/intel_epb.c 2019-04-05 14:43:04.280222577 +0200 +2 -0
@@ -19,6 +19,8 @@
 #include <asm/cpufeature.h>
 #include <asm/msr.h>
 
+#include <linux/acpi.h>
+
 /**
  * DOC: overview
  *

arch/x86/kernel/cpu/intel_epb.c 2019-04-08 10:01:26.029278425 +0200 static void intel_epb_restore(void) · line 97

Quoted 8 lines
arch/x86/kernel/cpu/intel_epb.c 2019-04-08 10:01:26.029278425 +0200 renamed from arch/x86/kernel/cpu/intel_epb.c 2019-04-05 14:43:04.280222577 +0200 +4 -1
@@ -89,7 +91,10 @@ static void intel_epb_restore(void)
                 * ('normal').
                 */
                val = epb & EPB_MASK;
-               if (val == ENERGY_PERF_BIAS_PERFORMANCE) {
+               if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
+                   acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER) {
+                       /* Keep performance setting on servers */
+               } else  if (val == ENERGY_PERF_BIAS_PERFORMANCE) {

I would write this as

if (val == ENERGY_PERF_BIAS_PERFORMANCE &&

    acpi_gbl_FADT.preferred_profile != PM_ENTERPRISE_SERVER &&
    acpi_gbl_FADT.preferred_profile != PM_PERFORMANCE_SERVER) {

but overall the approach is not particularly objectionable.

Quoted 4 lines
arch/x86/kernel/cpu/intel_epb.c 2019-04-08 10:01:26.029278425 +0200 renamed from arch/x86/kernel/cpu/intel_epb.c 2019-04-05 14:43:04.280222577 +0200 +0 -0
                        val = ENERGY_PERF_BIAS_NORMAL;
                        pr_warn_once("ENERGY_PERF_BIAS: Set to 'normal', was 'performance'\n");
                }