Re: [PATCH v2 0/2] ACPI: CPPC: Add CPPC v4 support (ACPI 6.6)

[Sumit Gupta <sumitg@nvidia.com>]

Hi Rafael,


On 09/05/26 00:31, Rafael J. Wysocki wrote:
> External email: Use caution opening links or attachments
>
>
> On Thu, Apr 30, 2026 at 4:25 PM Sumit Gupta <sumitg@nvidia.com> wrote:
>> Add initial kernel support for CPPC v4 (ACPI 6.6, Section 8.4.6),
>> which extends the _CPC package from 23 to 25 entries with two
>> optional fields:
>>
>>    - OSPM Nominal Performance (8.4.6.1.2.6): register used by OSPM
>>      to tell the platform what it considers nominal. The platform
>>      classifies performance above this as boost and below as
>>      throttle for power/thermal decisions.
>>
>>    - Resource Priority (8.4.6.1.2.7): Package of Resource Priority
>>      Register Descriptor sub-packages. Full parsing is not yet
>>      implemented; such entries are marked as unsupported.
>>
>> Patch 1: Add v4 _CPC parsing - validate the 25-entry layout,
>> handle the Resource Priority package, and mark the two new
>> registers optional.
>>
>> Patch 2: Add acpi_cppc/ospm_nominal_perf as a read-write sysfs
>> attribute, and initialize it to the platform nominal value
>> during cppc_cpufreq policy init.
>>
>> ---
>> v1[1] -> v2:
>> - Patch 1: added Reviewed-by from Mario Limonciello.
>> - Patch 2:
>>    - Make ospm_nominal_perf sysfs read-write; cache last write in
>>      cpc_desc and skip redundant register writes.
>>    - Validate input in cppc_set_ospm_nominal_perf.
>>
>> Sumit Gupta (2):
>>    ACPI: CPPC: Add support for CPPC v4
>>    ACPI: CPPC: Add ospm_nominal_perf support
>>
>>   drivers/acpi/cppc_acpi.c       | 93 +++++++++++++++++++++++++++++++---
>>   drivers/cpufreq/cppc_cpufreq.c | 10 ++++
>>   include/acpi/cppc_acpi.h       | 14 ++++-
>>   3 files changed, 109 insertions(+), 8 deletions(-)
>>
>> [1] https://lore.kernel.org/lkml/20260427051823.280419-1-sumitg@nvidia.com/
>>
>> --
> Can you please see the sashiko.dev feedback on this set:
>
> https://sashiko.dev/#/patchset/20260430142430.755437-1-sumitg%40nvidia.com
>
> and let me know what you think?  Especially regarding the second patch?


Thank you for sharing this.

Patch 1:
- Comments #1 and #2 are pre-existing issues with rare occurrence.
   I will address them in a separate hardening patch.

- Comment #3: In v3, will limit the ACPI_TYPE_PACKAGE handling to the
   RESOURCE_PRIORITY entry. So a Package at any other slot will be
   treated as invalid and abort probe, as it did before this patch.

----------------

Patch 2:
Discussed the changes for v3 in some detail on this thread already
which address most of the points (Please see my reply to Pierre [1]).

Summary of how each point will be addressed below:


 > The commit message states the valid range is [Lowest Performance,
 > Nominal Performance]. Does this code allow writing arbitrary values
 > outside that range by only checking against U32_MAX, without fetching
 > the CPU's capabilities to validate the input?
Will fetch the bounds via cppc_get_perf_caps() and reject values
outside [lowest_perf, nominal_perf] in v3.


 > If the hardware loses state during a logical CPU hotplug or system
 > suspend, but the software cache is not invalidated, will this check
 > prevent the register from being correctly re-initialized when the CPU
 > comes back online?
The redundant write check will be removed in v3, so the stale cache
failure mode won't be possible.


 > Can concurrent sysfs writes permanently desynchronize the software
 > cache from the hardware register?
 > ...
 > Is a lock needed around the read-modify-write cycle?
This will not occur in v3 since concurrent calls for the same
policy are serialized by policy->rwsem at the cpufreq layer (see [1]).


 > Additionally, can a time-of-check to time-of-use race lead to a NULL
 > pointer dereference if cpc_desc_ptr is initialized concurrently?
 > ...
 > Would this cause the WRITE_ONCE() to dereference the locally fetched NULL
 > cpc_desc pointer? Should this explicitly return -ENODEV early if 
!cpc_desc?
Will add the early -ENODEV return at the top of the function in v3,
eliminating the NULL cpc_desc race.


 > For shared cpufreq policies where policy->cpus contains multiple
 > logical cores (such as CPUFREQ_SHARED_TYPE_ANY), does this skip
 > initializing the secondary CPUs in the domain?
 >
 > If they are uninitialized, will their local cache remain 0, causing
 > sysfs reads for those secondary CPUs to incorrectly return -ENODATA?
Will move the rw sysfs from the per-CPU acpi_cppc interface to a
per-policy cpufreq interface in v3, and write the register on every
CPU in policy->cpus/domain.
The -ENODATA on unwritten read path will go away with the per-CPU node,
and the per-policy show returns 0 until user-space writes a value. See [1].


 > Also, since the sysfs attribute is tied to the physical CPU device
 > lifetime and persists independently of cpufreq policy teardowns, will
 > unconditionally setting the nominal performance here silently clobber
 > any persistent userspace configurations when a CPU is taken offline
 > and online?
Will drop the unconditional cpu_init write in v3, so the user-set
value won't be overwritten on CPU hotplug.

[1] 
https://lore.kernel.org/all/9c32f75a-294f-4cea-810e-c011c4dd91ab@nvidia.com/

Thank you,
Sumit Gupta