read_acpi_id() attempts to evaluate _CST using a stack buffer of
sizeof(union acpi_object) (48 bytes), but _CST returns a nested Package
of sub-Packages (one per C-state, each containing a register descriptor,
type, latency, and power) requiring hundreds of bytes. The evaluation
always fails with AE_BUFFER_OVERFLOW.

On modern systems using FFH/MWAIT entry (where pblk is zero), this
causes the function to return before setting the acpi_id_cst_present
bit. In check_acpi_ids(), flags.power is then zero for all Phase 2 CPUs
(physical CPUs beyond dom0's vCPU count), so push_cxx_to_hypervisor() is
never called for them.

On a system with dom0_max_vcpus=2 and 8 physical CPUs, only PCPUs 0-1
receive C-state data. PCPUs 2-7 are stuck in C0/C1 idle, unable to
enter C2/C3. This costs measurable wall power (4W observed on an Intel
Core Ultra 7 265K with Xen 4.20).

The function never uses the _CST return value -- it only needs to know
whether _CST exists. Replace the broken acpi_evaluate_object() call with
acpi_has_method(), which correctly detects _CST presence using
acpi_get_handle() without any buffer allocation. This brings C-state
detection to parity with the P-state path, which already works correctly
for Phase 2 CPUs.

Fixes: 59a568029181 ("xen/acpi-processor: C and P-state driver that uploads said data to hypervisor.")
Signed-off-by: David Thomson <dt@linux-mail.net>
---
Changes in v2:
- Check pblk first to avoid unnecessary acpi_has_method() call when
  pblk is zero (Jan Beulich)

 drivers/xen/xen-acpi-processor.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/drivers/xen/xen-acpi-processor.c b/drivers/xen/xen-acpi-processor.c
index 2967039..67a4afc 100644
--- a/drivers/xen/xen-acpi-processor.c
+++ b/drivers/xen/xen-acpi-processor.c
@@ -379,11 +379,8 @@ read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv)
 			 acpi_psd[acpi_id].domain);
 	}

-	status = acpi_evaluate_object(handle, "_CST", NULL, &buffer);
-	if (ACPI_FAILURE(status)) {
-		if (!pblk)
-			return AE_OK;
-	}
+	if (!pblk && !acpi_has_method(handle, "_CST"))
+		return AE_OK;
 	/* .. and it has a C-state */
 	__set_bit(acpi_id, acpi_id_cst_present);

--
2.34.1

On Tue, Feb 24, 2026 at 09:37:11AM +0000, David Thomson wrote:
> 
> On a system with dom0_max_vcpus=2 and 8 physical CPUs, only PCPUs 0-1
> receive C-state data. PCPUs 2-7 are stuck in C0/C1 idle, unable to
> enter C2/C3. This costs measurable wall power (4W observed on an Intel
> Core Ultra 7 265K with Xen 4.20).
> 
> The function never uses the _CST return value -- it only needs to know
> whether _CST exists. Replace the broken acpi_evaluate_object() call with
> acpi_has_method(), which correctly detects _CST presence using
> acpi_get_handle() without any buffer allocation. This brings C-state
> detection to parity with the P-state path, which already works correctly
> for Phase 2 CPUs.
> 
> Fixes: 59a568029181 ("xen/acpi-processor: C and P-state driver that uploads said data to hypervisor.")
> Signed-off-by: David Thomson <dt@linux-mail.net>

Tested-by: Elliott Mitchell <ehem+xen@m5p.com>

The window of opportunity happened and I confirmed it fixed the issue for
me.  I don't know whether this already got in though.

> -	status = acpi_evaluate_object(handle, "_CST", NULL, &buffer);
> -	if (ACPI_FAILURE(status)) {
> -		if (!pblk)
> -			return AE_OK;
> -	}
> +	if (!pblk && !acpi_has_method(handle, "_CST"))
> +		return AE_OK;
>  	/* .. and it has a C-state */
>  	__set_bit(acpi_id, acpi_id_cst_present);

I had traced the problem to this spot a while back, but I didn't know
what an appropriate change was.  Since I had a workaround this wasn't
urgent.  I do know this effects many others, I don't know how urgent this
bug is for everyone else.


-- 
(\___(\___(\______          --=> 8-) EHM <=--          ______/)___/)___/)
 \BS (    |         ehem+sigmsg@m5p.com  PGP 87145445         |    )   /
  \_CS\   |  _____  -O #include <stddisclaimer.h> O-   _____  |   /  _/
8A19\___\_|_/58D2 7E3D DDF4 7BA6 <-PGP-> 41D1 B375 37D0 8714\_|_/___/5445

On 24.02.2026 10:37, David Thomson wrote:
> read_acpi_id() attempts to evaluate _CST using a stack buffer of
> sizeof(union acpi_object) (48 bytes), but _CST returns a nested Package
> of sub-Packages (one per C-state, each containing a register descriptor,
> type, latency, and power) requiring hundreds of bytes. The evaluation
> always fails with AE_BUFFER_OVERFLOW.
> 
> On modern systems using FFH/MWAIT entry (where pblk is zero), this
> causes the function to return before setting the acpi_id_cst_present
> bit. In check_acpi_ids(), flags.power is then zero for all Phase 2 CPUs
> (physical CPUs beyond dom0's vCPU count), so push_cxx_to_hypervisor() is
> never called for them.
> 
> On a system with dom0_max_vcpus=2 and 8 physical CPUs, only PCPUs 0-1
> receive C-state data. PCPUs 2-7 are stuck in C0/C1 idle, unable to
> enter C2/C3. This costs measurable wall power (4W observed on an Intel
> Core Ultra 7 265K with Xen 4.20).
> 
> The function never uses the _CST return value -- it only needs to know
> whether _CST exists. Replace the broken acpi_evaluate_object() call with
> acpi_has_method(), which correctly detects _CST presence using
> acpi_get_handle() without any buffer allocation. This brings C-state
> detection to parity with the P-state path, which already works correctly
> for Phase 2 CPUs.
> 
> Fixes: 59a568029181 ("xen/acpi-processor: C and P-state driver that uploads said data to hypervisor.")
> Signed-off-by: David Thomson <dt@linux-mail.net>

Reviewed-by: Jan Beulich <jbeulich@suse.com>