[PATCH v3 07/10] perf: arm_spe: Add support for filtering on data source

James Clark james.clark at linaro.org
Thu Jul 17 09:42:37 PDT 2025



On 17/07/2025 4:27 pm, Will Deacon wrote:
> On Thu, Jul 17, 2025 at 04:16:32PM +0100, James Clark wrote:
>>
>>
>> On 17/07/2025 3:29 pm, Will Deacon wrote:
>>> On Tue, Jul 15, 2025 at 02:04:18PM +0100, James Clark wrote:
>>>> On 14/07/2025 3:04 pm, Will Deacon wrote:
>>>>> On Thu, Jun 05, 2025 at 11:49:05AM +0100, James Clark wrote:
>>>>>> @@ -406,6 +416,9 @@ static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
>>>>>>     	if (ATTR_CFG_GET_FLD(attr, inv_event_filter))
>>>>>>     		reg |= PMSFCR_EL1_FnE;
>>>>>> +	if (ATTR_CFG_GET_FLD(attr, data_src_filter))
>>>>>> +		reg |= PMSFCR_EL1_FDS;
>>>>>
>>>>> Is the polarity correct here? The description of PMSDSFR_EL1.S<m> suggests
>>>>> that setting bits to 1 _excludes_ the FDS filtering.
>>>>>
>>>>
>>>> Setting filter bits to 1 means that samples matching are included. Setting
>>>> bits to 0 means that they are excluded. And PMSFCR_EL1.FDS enables filtering
>>>> as a whole, so if the user sets any filter bit to 1 we want to enable
>>>> filtering:
>>>>
>>>>     PMSDSFR_EL1.S<m>
>>>>
>>>>     0b0  If PMSFCR_EL1.FDS is 1, do not record load operations that have
>>>>          bits [5:0] of the Data Source packet set to <m>.
>>>>
>>>>     0b1  Load operations with Data Source <m> are unaffected by
>>>>          PMSFCR_EL1.FDS.
>>>>
>>>> I think it's all the right way around and it ends up being the same as the
>>>> other filters in SPE. Because we're using any bit being set to enable the
>>>> filtering, the only thing you can't do is enable filtering with a 0 filter,
>>>> but I didn't think that was useful. See the previous discussion on this
>>>> here:
>>>> https://lore.kernel.org/all/5752f039-51c1-4452-b5df-03ff06da7be3@linaro.org/
>>>>
>>>> Reading the "Data source filtering" section in the docs change at the end
>>>> might help too.
>>>
>>> Sorry, but I still don't get it :/
>>>
>>> afaict, if any of the bits in 'data_src_filter' are _zero_ then we
>>> should set PMSFCR_EL1.FDS. That also means that a mask of zero means all
>>> loads are filtered, which is what the architecture says and is what we
>>> should provide to userspace.
>>>
>>> Will
>>
>> We'd have to add another format flag to enable data source filtering then,
>> because otherwise the default would be zero and people's samples would
>> disappear.
>>
>> But the only use cases I could think of were more like "I want to see
>> samples from data source 1":
>>
>>    -e arm_spe/data_src_filter=0x1/
>>
>> Or "I want to see all data sources except 1":
>>
>>    -e arm_spe/data_src_filter=0xfffffffe/
>>
>> Filtering out all samples with any data source didn't seem to make sense to
>> me, and I think you can already do that with the other filters (remove loads
>> etc).
>>
>> It would be a shame to be inconsistent and to add an enable flag just for
>> that one case because the other filters in SPE are auto enabled for non-zero
>> values. Although to be fair for PMSFCR.FT and others, zero filters are
>> explicitly not allowed:
>>
>>    If this field is set to 1 and the PMSFCR_EL1.{ST, LD, B} bits are all
>>    set to zero, it is CONSTRAINED UNPREDICTABLE whether no samples are
>>    recorded or the PE behaves as if PMSFCR_EL1.FT is set to 0
>>
>> Seems like FDS doesn't end up as neat as the others, but IMO I can't see
>> anyone needing a zero filter. I did discuss it with Leo and we decided that
>> we could always add the enable flag at a later date if a use case turned up
>> and it wouldn't be a breaking change.
>>
>> But if you think it's there so it should be exposed I can add it.
> 
> What about if we expose the inverse of PMSDSFR_EL1 to userspace instead?
> 
> Will

That's exactly what I was thinking. It does seem a bit weird, but I can 
call it inv_data_src_filter and document it appropriately so it should 
be fine.




More information about the linux-arm-kernel mailing list