Re: [PATCH] ACPI: CPPC: Scale EPP values to the register width
From: Sumit Gupta
Date: Fri Nov 13 2026 - 11:11:56 EST
On 13/11/26 18:02, Christian Loehle wrote:
External email: Use caution opening links or attachments
On 11/13/26 12:28, Rafael J. Wysocki (Intel) wrote:
On Fri, Nov 13, 2026 at 1:22 PM Christian Loehle
<christian.loehle@xxxxxxx> wrote:
On 11/12/26 17:19, Christian Loehle wrote:
On 11/12/26 17:11, Mario Limonciello wrote:
On 11/12/26 08:05, Christian Loehle wrote:
On 11/12/26 14:01, Rafael J. Wysocki (Intel) wrote:
On Thu, Nov 12, 2026 at 2:58 PM Christian Loehle
<christian.loehle@xxxxxxx> wrote:
On 11/12/26 13:13, Christian Loehle wrote:
The EPP sysfs interface uses 0-255, but ACPI permits 4-8 bit fields.
Writing 128 to a 4- or 6-bit SystemMemory field currently succeeds but
stores zero, selecting maximum performance.
Scale writes to supported narrow EPP registers and normalize reads to
0-255. Cover both EPP setters and cppc_get_perf() so saved firmware values
round-trip correctly.
Fixes: 922607a2b462 ("cpufreq: CPPC: Add support for autonomous selection")
Signed-off-by: Christian Loehle <christian.loehle@xxxxxxx>
---
.../ABI/testing/sysfs-devices-system-cpu | 5 ++
drivers/acpi/cppc_acpi.c | 46 +++++++++++++++++--
2 files changed, 47 insertions(+), 4 deletions(-)
diff --git a/Documentation/ABI/testing/sysfs-devices-system-cpu b/Documentation/ABI/testing/sysfs-devices-system-cpu
index d219fc27a0c9..c63b5cfd993c 100644
--- a/Documentation/ABI/testing/sysfs-devices-system-cpu
+++ b/Documentation/ABI/testing/sysfs-devices-system-cpu
@@ -322,6 +322,11 @@ Description: Energy performance preference
performance increase/decrease and the result of the hardware's
energy efficiency and performance optimization policies.
+ For supported EPP registers narrower than eight bits, values are
+ scaled and rounded to the nearest hardware preference.
+ Reads use the same 0-255 scale and may differ from the requested
+ value due to the reduced precision.
+
Writing to this file only has meaning when Autonomous Selection is
enabled.
diff --git a/drivers/acpi/cppc_acpi.c b/drivers/acpi/cppc_acpi.c
index 11a9f9eb6b4a..10b52163eb56 100644
--- a/drivers/acpi/cppc_acpi.c
+++ b/drivers/acpi/cppc_acpi.c
@@ -38,6 +38,7 @@
#include <linux/iopoll.h>
#include <linux/ktime.h>
#include <linux/list.h>
+#include <linux/math64.h>
#include <linux/mutex.h>
#include <linux/rbtree.h>
#include <linux/rwsem.h>
@@ -2841,6 +2842,25 @@ int cppc_get_highest_perf(int cpunum, u64 *highest_perf)
}
EXPORT_SYMBOL_GPL(cppc_get_highest_perf);
+/* Keep the EPP interface on an 8-bit scale, including firmware readback. */
+static u64 cppc_scale_epp(const struct cpc_register_resource *reg,
+ u64 epp, bool to_reg)
+{
+ unsigned int width = reg->cpc_entry.reg.bit_width;
+ u32 max;
+
+ if (reg->type != ACPI_TYPE_BUFFER || width < 4 || width >= 8)
+ return epp;
+
+ max = BIT(width) - 1;
+ if (to_reg)
+ return DIV_ROUND_CLOSEST_ULL(epp * max,
+ CPPC_EPP_ENERGY_EFFICIENCY_PREF);
+
+ return DIV_ROUND_CLOSEST_ULL(epp * CPPC_EPP_ENERGY_EFFICIENCY_PREF,
+ max);
+}
+
/**
* cppc_get_epp_perf - Get the epp register value.
* @cpunum: CPU from which to get epp preference value.
@@ -2850,7 +2870,15 @@ EXPORT_SYMBOL_GPL(cppc_get_highest_perf);
*/
int cppc_get_epp_perf(int cpunum, u64 *epp_perf)
{
- return cppc_get_reg_val(cpunum, ENERGY_PERF, epp_perf);
+ struct cpc_desc *cpc_desc = per_cpu(cpc_desc_ptr, cpunum);
+ int ret;
+
+ ret = cppc_get_reg_val(cpunum, ENERGY_PERF, epp_perf);
+ if (!ret)
+ *epp_perf = cppc_scale_epp(&cpc_desc->cpc_regs[ENERGY_PERF],
+ *epp_perf, false);
+
+ return ret;
}
EXPORT_SYMBOL_GPL(cppc_get_epp_perf);
@@ -3150,6 +3178,7 @@ int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
bool auto_sel_non_pcc;
bool epp_pcc;
bool epp_non_pcc;
+ u64 epp;
int ret;
if (!cpc_desc) {
@@ -3159,6 +3188,7 @@ int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
auto_sel_reg = &cpc_desc->cpc_regs[AUTO_SEL_ENABLE];
epp_set_reg = &cpc_desc->cpc_regs[ENERGY_PERF];
+ epp = cppc_scale_epp(epp_set_reg, perf_ctrls->energy_perf, true);
if (!enable && auto_sel_reg->type == ACPI_TYPE_INTEGER &&
auto_sel_reg->cpc_entry.int_value == 1)
return -EOPNOTSUPP;
@@ -3177,7 +3207,7 @@ int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
return ret;
}
if (epp_non_pcc) {
- ret = cpc_write(cpu, epp_set_reg, perf_ctrls->energy_perf);
+ ret = cpc_write(cpu, epp_set_reg, epp);
if (ret)
return ret;
}
@@ -3205,7 +3235,7 @@ int cppc_set_epp_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls, bool enable)
}
if (epp_pcc) {
- ret = cpc_write(cpu, epp_set_reg, perf_ctrls->energy_perf);
+ ret = cpc_write(cpu, epp_set_reg, epp);
if (ret)
goto out_unlock;
}
@@ -3235,9 +3265,17 @@ EXPORT_SYMBOL_GPL(cppc_set_epp_perf);
*/
int cppc_set_epp(int cpu, u64 epp_val)
{
+ struct cpc_desc *cpc_desc = per_cpu(cpc_desc_ptr, cpu);
+
if (epp_val > CPPC_EPP_ENERGY_EFFICIENCY_PREF)
return -EINVAL;
+ if (!cpc_desc)
+ return -ENODEV;
+
+ epp_val = cppc_scale_epp(&cpc_desc->cpc_regs[ENERGY_PERF],
+ epp_val, true);
+
return cppc_set_reg_val(cpu, ENERGY_PERF, epp_val);
}
EXPORT_SYMBOL_GPL(cppc_set_epp);
@@ -3496,7 +3534,7 @@ int cppc_get_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls)
if (ret)
goto out_err;
}
- perf_ctrls->energy_perf = energy_perf;
+ perf_ctrls->energy_perf = cppc_scale_epp(energy_perf_reg, energy_perf, false);
if (CPC_SUPPORTED(auto_sel_reg)) {
ret = cpc_read(cpu, auto_sel_reg, &auto_sel);
+CC Sumit's actual email, sorry about that.
So I'm assuming that this one is for Sumit to pick up and send again
along with a new version of
https://lore.kernel.org/linux-acpi/20261112080442.4007723-1-sumitg@xxxxxxxxxx/
Or please let me know if that is not the case.
Yup, while it fixes an issue by itself technically, I don't think there
are any affected platforms.
I'm hoping that Sumit this (and the limit-write-at-init) patch as a series
for all our convenience (and hopefully to get Sashiko to review them too).
(I guess I should've mentioned that in the notes.)
And sorry about ACPI/CPPC having such high traffic lately, I'm expecting once
we've got the issues ironed out and the new features supported it'll be a
lot more boring :)
So the basic idea with this patch that the userspace interface is always 8 bit so higher level software will not need to know the lower level implementation detail at the expense of some granularity.
Makes sense to me, but I'd like to know how realistic this hardware is going to be. If it's purely hypothetical hardware; how about proposing to ASWG that this field must always be 8 bit?
Good point, IMO it's hypothetical hardware that hopefully doesn't and won't exist.
I'll try to raise this through that channel as well.
Quick update on this, after proposing it, Souvik came back and made me aware that
we've already proposed this:
https://github.com/tianocore/edk2/issues/10838
I guess forget about $SUBJECT after all and I'll probe-fail <8bit and get the amd-pstate
handling correct.
So should I wait for an update of
https://lore.kernel.org/linux-acpi/20261112080442.4007723-1-sumitg@xxxxxxxxxx/
or is it applicable as-is? In which case, care to give it a tag?
I guess the issue is arguably preexisting and the fix doesn't touch the same code,
so feel free to apply,
Yes, [1] can be applied as is. The issue Sashiko reported on it exists
without [1] too, and [2] fixes it.
[2] replaces the firmware MIN/MAX at init(). v5 of [1] kept them in
autonomous mode. I'm not aware of any platform that needs the OS to
keep them, but if one does, I will look at that separately.
I can also resend [1] and [2] as a series if you prefer.
[1] https://lore.kernel.org/linux-acpi/20261112080442.4007723-1-sumitg@xxxxxxxxxx/
[2] https://lore.kernel.org/all/20261112113204.778904-1-christian.loehle@xxxxxxx/
Thanks,
Sumit