Skip to contentSkip to conversation

intel_idle: Fix C6S and C10 latency values for Pantherlake

PATCH In review Kaushlendra Kumar 4 participants 7 messages Aug 11 – Oct 2 Intel idle driver

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 intel_idle: Fix C6S and C10 latency values for Pantherlake

Correct the values using BIOS ACPI _CST characterization data read from a Pantherlake device via /sys/devices/system/cpu/cpu0/cpuidle/stateN/{latency,residency}:

C6S: exit_latency 300->127, target_residency 300->381
C10: exit_latency 370->1048, target_residency 2500->3144

---
 drivers/idle/intel_idle.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/idle/intel_idle.c +4 -4
index 651408df9c24..0b5c94c7cd3a 100644
@@ -1010,16 +1010,16 @@ static struct cpuidle_state ptl_cstates[] __initdata = {
 		.name = "C6S",
 		.desc = "MWAIT 0x21",
 		.flags = MWAIT2flg(0x21) | CPUIDLE_FLAG_TLB_FLUSHED,
-		.exit_latency = 300,
-		.target_residency = 300,
+		.exit_latency = 127,
+		.target_residency = 381,
 		.enter = &intel_idle,
 		.enter_s2idle = intel_idle_s2idle, },
 	{
 		.name = "C10",
 		.desc = "MWAIT 0x60",
 		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
-		.exit_latency = 370,
-		.target_residency = 2500,
+		.exit_latency = 1048,
+		.target_residency = 3144,
 		.enter = &intel_idle,
 		.enter_s2idle = intel_idle_s2idle, },
 	{
Signature · 1 line
2.34.1
maintainer
#1 lore raw

The current value are based on what I measured using wult/pbe tool.

I think it will be better to explain the problem and how this patch fixes it, give some details about workload. And ideally, also provided turbostat output when it runs in parallel with the workload before and after this patch. Then we have can see CPU frequency and idle profile of your workload with and without this patch.

Thanks!

Quoted 37 lines · Kaushlendra Kumar

Correct the values using BIOS ACPI _CST characterization data read from a Pantherlake device via /sys/devices/system/cpu/cpu0/cpuidle/stateN/{latency,residency}:

C6S: exit_latency 300->127, target_residency 300->381
C10: exit_latency 370->1048, target_residency 2500->3144

---
 drivers/idle/intel_idle.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/idle/intel_idle.c +4 -4
index 651408df9c24..0b5c94c7cd3a 100644
@@ -1010,16 +1010,16 @@ static struct cpuidle_state ptl_cstates[] __initdata = {
 		.name = "C6S",
 		.desc = "MWAIT 0x21",
 		.flags = MWAIT2flg(0x21) | CPUIDLE_FLAG_TLB_FLUSHED,
-		.exit_latency = 300,
-		.target_residency = 300,
+		.exit_latency = 127,
+		.target_residency = 381,
 		.enter = &intel_idle,
 		.enter_s2idle = intel_idle_s2idle, },
 	{
 		.name = "C10",
 		.desc = "MWAIT 0x60",
 		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
-		.exit_latency = 370,
-		.target_residency = 2500,
+		.exit_latency = 1048,
+		.target_residency = 3144,
 		.enter = &intel_idle,
 		.enter_s2idle = intel_idle_s2idle, },
 	{
maintainer
#2 lore raw
In reply to Kaushlendra Kumar

Hi,

The current value are based on what I measured using wult/pbe tool.

I think it will be better to explain the problem and how this patch fixes it, give some details about workload. And ideally, also provided turbostat output when it runs in parallel with the workload before and after this patch. Then we have can see CPU frequency and idle profile of your workload with and without this patch.

Thanks!

Quoted 37 lines · Kaushlendra Kumar

Correct the values using BIOS ACPI _CST characterization data read from a Pantherlake device via /sys/devices/system/cpu/cpu0/cpuidle/stateN/{latency,residency}:

C6S: exit_latency 300->127, target_residency 300->381
C10: exit_latency 370->1048, target_residency 2500->3144

---
 drivers/idle/intel_idle.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/idle/intel_idle.c +4 -4
index 651408df9c24..0b5c94c7cd3a 100644
@@ -1010,16 +1010,16 @@ static struct cpuidle_state ptl_cstates[] __initdata = {
 		.name = "C6S",
 		.desc = "MWAIT 0x21",
 		.flags = MWAIT2flg(0x21) | CPUIDLE_FLAG_TLB_FLUSHED,
-		.exit_latency = 300,
-		.target_residency = 300,
+		.exit_latency = 127,
+		.target_residency = 381,
 		.enter = &intel_idle,
 		.enter_s2idle = intel_idle_s2idle, },
 	{
 		.name = "C10",
 		.desc = "MWAIT 0x60",
 		.flags = MWAIT2flg(0x60) | CPUIDLE_FLAG_TLB_FLUSHED,
-		.exit_latency = 370,
-		.target_residency = 2500,
+		.exit_latency = 1048,
+		.target_residency = 3144,
 		.enter = &intel_idle,
 		.enter_s2idle = intel_idle_s2idle, },
 	{
#3 lore raw
Quoted 6 lines · On Tue, 2026-08-11 at 13:47 +0530, <reviewer> wrote:

The current values are based on what I measured using wult/pbe tool.

I think it will be better to explain the problem and how this patch fixes it, give some details about workload. And ideally, also provided turbostat output when it runs in parallel with the workload before and after this patch.

Thank you for the feedback.

Looking at other platforms in intel_idle, C-state latency values are generally sourced from BIOS ACPI _CST characterization data. For example, MTL-L C6 (exit_latency=140, target_residency=420) and ADL-L C6 (exit_latency=170, target_residency=500) match their respective ACPI _CST tables.

For PTL, I observed a discrepancy. The ACPI _CST on a Pantherlake-L device reports:

C6S: latency=127us, residency=381us
C10: latency=1048us, residency=3144us

Could you clarify which reference was used for other platforms — ACPI _CST or wult — so we can apply a consistent methodology for PTL?

I will add the turbostat data and workload details as requested.

Thanks,
Kaushlendra

maintainer
#4 lore raw
Quoted 8 lines · Kaushlendra
Quoted 6 lines · On Tue, 2026-08-11 at 13:47 +0530, <reviewer> wrote:

The current values are based on what I measured using wult/pbe tool.

I think it will be better to explain the problem and how this patch fixes it, give some details about workload. And ideally, also provided turbostat output when it runs in parallel with the workload before and after this patch.

Quoted 7 lines

Thank you for the feedback.

Looking at other platforms in intel_idle, C-state latency values are generally sourced from BIOS ACPI _CST characterization data. For example, MTL-L C6 (exit_latency=140, target_residency=420) and ADL-L C6 (exit_latency=170, target_residency=500) match their respective ACPI _CST tables.

So there are platforms that have intel_idle table and that do not have it.

Quoted 5 lines

For PTL, I observed a discrepancy. The ACPI _CST on a Pantherlake-L device reports:

C6S: latency=127us, residency=381us
C10: latency=1048us, residency=3144us

Discrepancy -> difference.

Quoted 2 lines

Could you clarify which reference was used for other platforms — ACPI _CST or wult — so we can apply a consistent methodology for PTL?

Platforms that do not have intel_idle custom table, use latency from ACPI _CST table. It only provides latency. Target residency is calculated as 3x latency.

I do not know about _CST numbers. I only know that in many cases they are not what I can observe using wult tool.

Panther Lake is the only client platform that I measured with the tools I developed, and submitted the numbers based on wult and pbe data that I got.

Artem.

#5 lore raw
Quoted 10 lines · Artem Bityutskiy

Platforms that do not have intel_idle custom table, use latency from ACPI _CST table. It only provides latency. Target residency is calculated as 3x latency.

I do not know about _CST numbers. I only know that in many cases they are not what I can observe using wult tool.

Panther Lake is the only client platform that I measured with the tools I developed, and submitted the numbers based on wult and pbe data that I got.

Thank you for the clarification, Artem.

Below is turbostat data from a 10-minute idle run on Pantherlake-L with the current values and with the proposed values:

Metric Base with Patch
-------- -------- -------
C6S% 0.29 0.60
C10% 99.34 99.00
Pkg%pc6 87.84 89.48

C6S usage doubles and package idle improves with the proposed values. C10 remains dominant for long idle periods — no regression.

Would it be possible to re-run the wult/pbe measurements on PTL?

BR,
Kaushlendra

reviewer
#6 lore raw

Kaushlendra,
The intel_idle numbers are measured by Intel's upstream Linux team on real hardware --
both for max latency and average residency targets.

ACPI _CST numbers are not the result of any known hardware test and not validated by any known test on any operating system. Our experience with them is that often they are random,
which is why the native intel_idle driver was created in the first place.

ACPI _CST lacks any concept of the residency target required
by Linux cpuidle. To fill this void, Linux invents
a residency number in ACPI mode by multiplying the _CST max latency by 3. ISTR that at one time we multiplied it by 2. It is not science, it is a hack. (We are hoping to see this gap filled in the future by ACPI _LPI, BTW)

And thus the max latency and average residency numbers you see in ACPI _CST mode are
not something to aspire to. More often they are something to run away from.

Quoted 10 lines

Below is turbostat data from a 10-minute idle run on Pantherlake-L with the current values and with the proposed values:

Metric Base with Patch
-------- -------- -------
C6S% 0.29 0.60
C10% 99.34 99.00

C6S usage doubles and package idle improves with the proposed values. C10 remains dominant for long idle periods — no regression.

The rows above are software requested C-states, not the hardware C-states achieved.
It is actually worse, not better, that C6S is requested more in deep idle, and it is worse, not better, that C10% is requested less in deep idle. This can be due to a number of things -- most likely background OS noise. You'd need to include the Busy% to understand if these columns are even a valid comparison.

Quoted 1 line

Pkg%pc6 87.84 89.48

This is an actual hardware package state, and more is better.
More important, however, is Pkg%PC10. Achieving PC10
is worth more than a couple of % +/- on Pkg%pc6, and I'm guessing that on the s

Quoted 1 line

Would it be possible to re-run the wult/pbe measurements on PTL?

Go ahead and run them and let us know if you see different results on your system.

thanks,
Len Brown, Intel Open Source Technology Center