Re: [PATCH 07/10] drm/xe/sriov: Prep for multiple exec quantums and preemption timeouts

[Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>]

On 12/2/2025 8:42 AM, Michal Wajdeczko wrote:
>
> On 11/27/2025 2:45 AM, Daniele Ceraolo Spurio wrote:
>> Each scheduler group can be independently configured with its own exec
>> quantum and preemption timeouts. The existing KLVs to configure those
>> parameter will apply the value to all groups (even if they're not
> typo: parameters
>
>> enabled at the moment).
>>
>> When scheduler groups are disable the GuC used the values from Group 0.
> typo: ... disabled then GuC uses ... ?
>
>> Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
>> Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
>> ---
>>   drivers/gpu/drm/xe/abi/guc_klvs_abi.h         |  7 ++++--
>>   drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c    | 25 +++++++++++++------
>>   .../gpu/drm/xe/xe_gt_sriov_pf_config_types.h  |  5 ++--
>>   3 files changed, 25 insertions(+), 12 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/abi/guc_klvs_abi.h b/drivers/gpu/drm/xe/abi/guc_klvs_abi.h
>> index a6dce9da339f..48f47e26132d 100644
>> --- a/drivers/gpu/drm/xe/abi/guc_klvs_abi.h
>> +++ b/drivers/gpu/drm/xe/abi/guc_klvs_abi.h
>> @@ -290,7 +290,9 @@ enum  {
>>    *      infinitely long compute or shader kernel. In such a scenario, the
>>    *      PF would need to trigger a VM PAUSE and then change the KLV to force
>>    *      it to take effect. Such cases might typically happen on a 1PF+1VF
>> - *      Virtualization config enabled for heavier workloads like AI/ML.
>> + *      Virtualization config enabled for heavier workloads like AI/ML. If
> start with new empty line so it will rendered as separate paragraph
>
>> + *      scheduling groups are supported, the provided value is applied to all
>> + *      groups (even if they've not yet been enabled).
> should we document also here that "The scheduling groups are available starting from GuC 70.53" ?

ok

>
>>    *
>>    *      The max value for this KLV is 100 seconds, anything exceeding that
>>    *      will be clamped to the max.
>> @@ -312,7 +314,8 @@ enum  {
>>    *      In this case, the PF would need to trigger a VM PAUSE and then change
>>    *      the KLV to force it to take effect. Such cases might typically happen
>>    *      on a 1PF+1VF Virtualization config enabled for heavier workloads like
>> - *      AI/ML.
>> + *      AI/ML. If scheduling groups are supported, the provided value is applied
>> + *      to all groups (even if they've not yet been enabled).
> ditto
>
>>    *
>>    *      The max value for this KLV is 100 seconds, anything exceeding that
>>    *      will be clamped to the max.
>> 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 59c5c6b4d994..eb547fedb6da 100644
>> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
>> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
>> @@ -298,10 +298,10 @@ static u32 encode_config(u32 *cfg, const struct xe_gt_sriov_config *config, bool
>>   	}
>>   
>>   	cfg[n++] = PREP_GUC_KLV_TAG(VF_CFG_EXEC_QUANTUM);
>> -	cfg[n++] = config->exec_quantum;
>> +	cfg[n++] = config->exec_quantum[0];
>>   
>>   	cfg[n++] = PREP_GUC_KLV_TAG(VF_CFG_PREEMPT_TIMEOUT);
>> -	cfg[n++] = config->preempt_timeout;
>> +	cfg[n++] = config->preempt_timeout[0];
>>   
>>   #define encode_threshold_config(TAG, ...) ({					\
>>   	cfg[n++] = PREP_GUC_KLV_TAG(VF_CFG_THRESHOLD_##TAG);			\
>> @@ -1857,12 +1857,15 @@ static int pf_provision_exec_quantum(struct xe_gt *gt, unsigned int vfid,
>>   {
>>   	struct xe_gt_sriov_config *config = pf_pick_vf_config(gt, vfid);
>>   	int err;
>> +	int i;
>>   
>>   	err = pf_push_vf_cfg_exec_quantum(gt, vfid, &exec_quantum);
>>   	if (unlikely(err))
>>   		return err;
>>   
>> -	config->exec_quantum = exec_quantum;
>> +	for (i = 0; i < GUC_KLV_VGT_POLICY_ENGINE_GROUP_MAX_COUNT; i++)
> maybe:
>
> 	for (i = 0; i < ARRAY_SIZE(config->exec_quantum); i++)

ok

>
>> +		config->exec_quantum[i] = exec_quantum;
>> +
>>   	return 0;
>>   }
>>   
>> @@ -1870,7 +1873,7 @@ static u32 pf_get_exec_quantum(struct xe_gt *gt, unsigned int vfid)
>>   {
>>   	struct xe_gt_sriov_config *config = pf_pick_vf_config(gt, vfid);
>>   
>> -	return config->exec_quantum;
>> +	return config->exec_quantum[0];
>>   }
>>   
>>   /**
>> @@ -1987,12 +1990,14 @@ static int pf_provision_preempt_timeout(struct xe_gt *gt, unsigned int vfid,
>>   {
>>   	struct xe_gt_sriov_config *config = pf_pick_vf_config(gt, vfid);
>>   	int err;
>> +	int i;
>>   
>>   	err = pf_push_vf_cfg_preempt_timeout(gt, vfid, &preempt_timeout);
>>   	if (unlikely(err))
>>   		return err;
>>   
>> -	config->preempt_timeout = preempt_timeout;
>> +	for (i = 0; i < GUC_KLV_VGT_POLICY_ENGINE_GROUP_MAX_COUNT; i++)
> ditto
>
>> +		config->preempt_timeout[i] = preempt_timeout;
>>   
>>   	return 0;
>>   }
>> @@ -2001,7 +2006,7 @@ static u32 pf_get_preempt_timeout(struct xe_gt *gt, unsigned int vfid)
>>   {
>>   	struct xe_gt_sriov_config *config = pf_pick_vf_config(gt, vfid);
>>   
>> -	return config->preempt_timeout;
>> +	return config->preempt_timeout[0];
>>   }
>>   
>>   /**
>> @@ -2180,10 +2185,14 @@ u32 xe_gt_sriov_pf_config_get_sched_priority(struct xe_gt *gt, unsigned int vfid
>>   
>>   static void pf_reset_config_sched(struct xe_gt *gt, struct xe_gt_sriov_config *config)
>>   {
>> +	int i;
>> +
>>   	lockdep_assert_held(xe_gt_sriov_pf_master_mutex(gt));
>>   
>> -	config->exec_quantum = 0;
>> -	config->preempt_timeout = 0;
>> +	for (i = 0; i < GUC_KLV_VGT_POLICY_ENGINE_GROUP_MAX_COUNT; i++) {
>> +		config->exec_quantum[i] = 0;
>> +		config->preempt_timeout[i] = 0;
>> +	}
>>   }
>>   
>>   static int pf_provision_threshold(struct xe_gt *gt, unsigned int vfid,
>> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config_types.h
>> index 686c7b3b6d7a..abf003946242 100644
>> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config_types.h
>> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config_types.h
>> @@ -6,6 +6,7 @@
>>   #ifndef _XE_GT_SRIOV_PF_CONFIG_TYPES_H_
>>   #define _XE_GT_SRIOV_PF_CONFIG_TYPES_H_
>>   
>> +#include "abi/guc_klvs_abi.h"
>>   #include "xe_ggtt_types.h"
>>   #include "xe_guc_klv_thresholds_set_types.h"
>>   
>> @@ -30,9 +31,9 @@ struct xe_gt_sriov_config {
>>   	/** @begin_db: start index of GuC doorbell ID range. */
>>   	u16 begin_db;
>>   	/** @exec_quantum: execution-quantum in milliseconds. */
>> -	u32 exec_quantum;
>> +	u32 exec_quantum[GUC_KLV_VGT_POLICY_ENGINE_GROUP_MAX_COUNT];
>>   	/** @preempt_timeout: preemption timeout in microseconds. */
>> -	u32 preempt_timeout;
>> +	u32 preempt_timeout[GUC_KLV_VGT_POLICY_ENGINE_GROUP_MAX_COUNT];
> maybe we should rename
>
> 	GUC_KLV_VGT_POLICY_ENGINE_GROUP_MAX_COUNT
> to
> 	GUC_MAX_ENGINE_GROUPS
>
> and make it independent from KLVs and define like we have
>
> 	GUC_MAX_ENGINE_CLASSES

ok, I'll do it in an earlier patch.

Daniele

>
>
>>   	/** @sched_priority: scheduling priority. */
>>   	u32 sched_priority;
>>   	/** @thresholds: GuC thresholds for adverse events notifications. */