Re: [PATCH v2 17/17] drm/xe/pf: Add documentation for sriov_admin attributes

[Michal Wajdeczko <michal.wajdeczko@intel.com>]

On 10/30/2025 6:25 PM, Lucas De Marchi wrote:
> On Tue, Oct 28, 2025 at 06:58:31PM +0100, Michal Wajdeczko wrote:
>> Add initial documentation for all recently added Xe driver
>> specific SR-IOV sysfs files located under device/sriov_admin.
>>
>> Signed-off-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
>> Cc: Lucas De Marchi <lucas.demarchi@intel.com>
>> Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
>> ---
>> v2: update version (Rodrigo)
>> ---
>> .../ABI/testing/sysfs-driver-intel-xe-sriov   | 160 ++++++++++++++++++

...

>> +            Actual quantum value might be aligned per HW/FW requirements.
> 
> Is the aligned value returned on read so user can know the actual value?

yes, while GuC does not tell us about any specific low level alignment,
the PF driver clamps requested value based on the GuC ABI prior to sending
this new EQ settings to the GuC, see [1]

[1] https://elixir.bootlin.com/linux/v6.18-rc3/source/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c#L197

and you may also try it:

$ echo 150000 | /sys/bus/pci/drivers/xe/0000:00:02.0/sriov_admin/vf1/profile/exec_quantum_ms

[ ] xe 0000:00:02.0: [drm] PF: Tile0: GT0: VF1 provisioned with 100000ms execution quantum
[ ] xe 0000:00:02.0: [drm] PF: Tile0: GT1: VF1 provisioned with 100000ms execution quantum

$ cat /sys/bus/pci/drivers/xe/0000:00:02.0/sriov_admin/vf1/profile/exec_quantum_ms
100000

> 
> Does this need to be expanded to mention what is a GT execution quantum?
> I don't think we have it documented elsewhere.

we have small kernel-doc for the EQ KLV [2], but it looks it could be
a good candidate for a section in the xe_sriov.rst ;)

[2] https://elixir.bootlin.com/linux/v6.18-rc3/source/drivers/gpu/drm/xe/abi/guc_klvs_abi.h#L257

> How does a 0 in one of
> the functions interact with the others not being 0?

that's definitely a not recommended configuration, but still may work

please note that currently there is no synchronization between GuCs,
so engines on those GTs are utilized for the VF WLs independently

btw, there is one problematic configuration when EQ/PT is infinity/infinity
together with VF priority set to "normal" - I'm wondering if we should
somehow try to reject such setting over sysfs ...

> 
> In the scenario below, I was expecting the functions to get 1/3, 1/3, 1/6, and 1/6
> of the GPU, but I'm not entirely sure:
> 
>     cat sriov_admin/*/profile/exec_quantum_ms
>     100
>     100
>     50
>     50

with VFs priority set to "normal" this should work

but if VFs priority is still at default "low" then if other VFs are
idle, the one VF with active WL will get 100% GPU (it will be just
sliced at its EQ)