Re: [PATCH 03/13] drm/xe/pf: Force new VFs prorities only once

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

On 4/2/2026 12:22 PM, Piotr Piórkowski wrote:
> Michal Wajdeczko <michal.wajdeczko@intel.com> wrote on pon [2026-mar-30 22:40:34 +0200]:
>> We should force change of VFs scheduling priorities only after
>> initial change of the SCHED_IF_IDLE policy. Doing that also during
>> any later policy reprovisioning will overwrite changes done on the
>> individual per-VF scheduling priorities (currently only for PF).
>>
>> While around also move priority change code to the _config component
>> to maintain a proper isolation. Also document our expectation about
>> the GuC version where the new meaning of the policy KLV is present.
>>
>> Signed-off-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
>> ---
>>  drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c | 28 ++++++++++++++
>>  drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h |  1 +
>>  drivers/gpu/drm/xe/xe_gt_sriov_pf_policy.c | 44 ++++++++++------------
>>  3 files changed, 48 insertions(+), 25 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
>> index 658e9b048751..e1c028176eb8 100644
>> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
>> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
>> @@ -2550,6 +2550,34 @@ u32 xe_gt_sriov_pf_config_get_sched_priority(struct xe_gt *gt, unsigned int vfid
>>  	return priority;
>>  }
>>  
>> +/**
>> + * xe_gt_sriov_pf_config_force_sched_priority_locked() - Force update scheduling priority.
>> + * @gt: the &xe_gt
>> + * @priority: new scheduling priority to set
>> + *
>> + * This function allows to update cached values of the scheduling priorities of all
>> + * VFs (and PF) as result of applying the `GUC_KLV_VGT_POLICY_SCHED_IF_IDLE`_ policy.
>> + *
>> + * This function can only be called on PF.
>> + */
>> +void xe_gt_sriov_pf_config_force_sched_priority_locked(struct xe_gt *gt, u32 priority)
>> +{
>> +	unsigned int total_vfs = 1 + xe_gt_sriov_pf_get_totalvfs(gt);
>> +	struct xe_gt_sriov_config *config;
>> +	unsigned int n;
>> +
>> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> I will add:
> lockdep_assert_held(xe_gt_sriov_pf_master_mutex(gt));

this is already part of the pf_pick_vf_config() called below

but yeah, I can add it here as kind of documenting function expectations
(will try to do it during merge)

> 
>> +
>> +	for (n = 0; n < total_vfs; n++) {
>> +		config = pf_pick_vf_config(gt, VFID(n));
>> +		config->sched_priority = priority;
>> +	}
>> +
>> +	pf_config_bulk_set_u32_done(gt, PFID, 1 + total_vfs, priority,
>> +				    pf_get_sched_priority, "scheduling priority",
>> +				    sched_priority_unit, n, 0);
>> +}
>> +

...

>>  
> 
> LGTM:
> Reviewed-by: Piotr Piórkowski <piotr.piorkowski@intel.com>

Thanks!

> 
> 
>>  /**
>> -- 
>> 2.47.1
>>
>