Re: [PATCH v3 0/7] Support L3 Smart Data Cache Injection Allocation Enforcement (SDCIAE)

[Reinette Chatre]

Hi Babu,

On 4/9/25 5:58 PM, Moger, Babu wrote:
> Hi Reinette,
> 
> On 4/8/2025 8:41 PM, Reinette Chatre wrote:
>> Hi Babu,
>>
>> On 4/8/25 5:41 PM, Moger, Babu wrote:
>>> Hi Reinette,
>>>
>>> On 4/8/2025 4:44 PM, Reinette Chatre wrote:
>>>> Hi Babu,
>>>>
>>>> On 4/7/25 1:12 PM, Moger, Babu wrote:
>>>>> On 3/21/25 17:50, Reinette Chatre wrote:
>>>>>> On 1/30/25 1:20 PM, Babu Moger wrote:
>>>>
>>>>
>>>>>>>
>>>>>>
>>>>>> AMD also supports what is exposed to user space as "shareable_bits". According
>>>>>> to APM:
>>>>>>      Depending on the implementation, some portions of the L3 Cache may be
>>>>>>      shared by other system functions or used for some other purpose not
>>>>>>      under the control of the PQOS feature set. The L3 Cache Allocation
>>>>>>      Sharing Mask returned by CPUID Fn0000_0010_EBX_x1[L3ShareAllocMask] is a
>>>>>>      bitmask that represents portions of the L3 that may be shared by those
>>>>>>      functions.
>>>>>
>>>>> Here is the complete text.
>>>>>
>>>>> The L3 Cache allocation sharing mask (L3ShareAllocMask) returned in EBX by
>>>>> CPUID Fn0000_0010 with ECX=1 is a bit vector which represents portions of
>>>>> the cache which may be shared with other system entities or used for some
>>>>> other purpose not under the control of the QOS feature set. When software
>>>>> sets a bit in one of the L3_MASK_n registers at the same bit positions a
>>>>> bit in the L3ShareAllocMask, processors executing with the corresponding
>>>>> COS will competitively share that portion of the cache with the other
>>>>> function. If this mask is all 0’s, then there is no other entity in the
>>>>> system competing with the processors for use of the L3 cache.
>>>>>
>>>>> The "L3ShareAllocMask" is always reported as 0 on AMD systems.
>>>>>
>>>>>> Could you please include what (if any) the relationship is between the CBM
>>>>>> discoverable via Fn0000_0010_EBX_x1[L3ShareAllocMask] and the CBM of
>>>>>> "highest-supported L3_MASK_n register" when SDCIAE is enabled?
>>>>>
>>>>> No. There is no relationship in here.
>>>>>
>>>>>>
>>>>>> On the resctrl interface side the documentation currently states:
>>>>>>
>>>>>>      "shareable_bits":
>>>>>>          Bitmask of shareable resource with other executing
>>>>>>          entities (e.g. I/O). User can use this when
>>>>>>          setting up exclusive cache partitions. Note that
>>>>>>          some platforms support devices that have their
>>>>>>          own settings for cache use which can over-ride
>>>>>>          these bits.
>>>>>>
>>>>>> Even though this was originally used to expose the content of
>>>>>> Fn0000_0010_EBX_x1[L3ShareAllocMask] the intent of the content does
>>>>>> seem to also apply to the "io_alloc" CBM also.
>>>>>
>>>>> It says "shared by other system functions or used for some other purpose
>>>>> not under the control of the PQOS feature set".
>>>>
>>>> This is a quote from the AMD spec, not the resctrl user interface documentation.
>>>>
>>>> Please consider this from resctrl user interface perspective.
>>>>
>>>>>
>>>>> "io_alloc" is PQOS feature set. I feel it should not affect "shareable_bits".
>>>>
>>>> When I read the resctrl user interface documentation for "shareable_bits" it
>>>> sounds relevant to io_alloc. The "shareable_bits" contains a bitmask "of
>>>> shareable resource with other executing entities (e.g. I/O)" ... is this
>>>> not exactly io_alloc?
>>>
>>> I agree the text is pretty generic. Actually, the whole bit mask (0xfffF) is shareable with io_alloc.
>>
>> I think the value of "shareable_bits" presented to user space could be the
>> actual io_alloc_cbm value. Thus, not the "possible IO bitmask" but the actual
> 
> Confused little bit here. The shareable_bits is resource property.
> io_alloc_cbm is domain specific value. Not sure how to display it.

ah, right. We should still aim to not let the new "io_alloc" interfaces cause
confusion with the existing "shareable_bits" that is already part of interface and
used to communicate IO allocation to user space. Perhaps we are just left with
needing a modification to the existing documentation surrounding "shareable_bits"?

> 
>> value. This seems to be the best match of what "shareable_bits" represents, which
>> is the region currently used by IO devices. This partners well with the "bit_usage"
>> output, for example, "X" can be used to show which portions of cache are currently
>> used by both software (via schemata of resource groups) and hardware (via io_alloc_cbm).
> 
> Haven't looked at this code much. Will look into it.

Thank you. This is per domain so should be a good fit ... but again the documentation
creates strong connection with "shareable_bits" that should be reworked.

Reinette