Re: [PATCH v1 2/9] drm/amdgpu: hand out the dedicated SPM VMID from the reserved path

["Khatri, Sunil" <sukhatri@amd.com>]

On 29-09-2026 05:18 pm, Christian König wrote:
> On 9/28/26 20:57, Sunil Khatri wrote:
>> The SPM reservation borrowed a VMID from the GFXHUB round-robin pool
>> shared with normal graphics contexts (VMIDs 1-7). SPM streams its output
>> through the selected VMID's page tables, so sharing that slot with
>> dynamically scheduled contexts is a poor isolation boundary and makes
>> the profiling VMID non-deterministic.
>>
>> Reservation now uses the VMID a hub dedicates to it. When
>> vm_manager.spm_vmid is set, amdgpu_vmid_alloc_reserved() hands out that
>> fixed VMID directly; when it is unset the reservation is rejected with
>> -ENOENT rather than borrowing a pool VMID, so a reserved submission can
>> never end up on VMID 0 (the kernel VM) or on a slot shared with graphics
>> contexts. The dedicated VMID is not part of the round-robin LRU, so it
>> is given an initialized active-fence sync on every GFXHUB instance in
>> amdgpu_vmid_mgr_init() and amdgpu_vmid_free_reserved() has nothing to
>> return to the LRU.
>>
>> The behaviour is gated on spm_vmid, so hubs that do not dedicate a VMID
>> simply do not support reservation.
> That won't work like this. The SPM VMID would still be used round robin with that.
>
> You need to take the VMID completely out of the mask. Take a look at when amdgpu_vmid_mgr_set_vmid_mask() is called.
Noted
Regards
Sunil Khatri
>
> Regards,
> Christian.
>
>> Signed-off-by: Sunil Khatri <sunil.khatri@amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 15 ++++++++++-----
>>   1 file changed, 10 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
>> index 5e424e3b6e71f..c8cbd0851d478 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
>> @@ -551,9 +551,12 @@ int amdgpu_vmid_alloc_reserved(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>   		r = -ENOENT;
>>   		goto unlock;
>>   	}
>> -	/* Remove from normal round robin handling */
>> -	id = list_first_entry(&id_mgr->ids_lru, struct amdgpu_vmid, list);
>> -	list_del_init(&id->list);
>> +	if (!adev->vm_manager.spm_vmid) {
>> +		r = -ENOENT;
>> +		goto unlock;
>> +	}
>> +
>> +	id = &id_mgr->ids[adev->vm_manager.spm_vmid];
>>   	vm->reserved_vmid[vmhub] = id;
>>   	id_mgr->reserved_vmid = true;
>>   	mutex_unlock(&id_mgr->lock);
>> @@ -577,8 +580,6 @@ void amdgpu_vmid_free_reserved(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>>   
>>   	mutex_lock(&id_mgr->lock);
>>   	if (vm->reserved_vmid[vmhub]) {
>> -		list_add(&vm->reserved_vmid[vmhub]->list,
>> -			&id_mgr->ids_lru);
>>   		vm->reserved_vmid[vmhub] = NULL;
>>   		id_mgr->reserved_vmid = false;
>>   	}
>> @@ -670,6 +671,10 @@ void amdgpu_vmid_mgr_init(struct amdgpu_device *adev)
>>   			amdgpu_sync_create(&id_mgr->ids[j].active);
>>   			list_add_tail(&id_mgr->ids[j].list, &id_mgr->ids_lru);
>>   		}
>> +
>> +		/* out-of-pool VMID; create its sync here */
>> +		if (adev->vm_manager.spm_vmid && AMDGPU_IS_GFXHUB(i))
>> +			amdgpu_sync_create(&id_mgr->ids[adev->vm_manager.spm_vmid].active);
>>   	}
>>   }
>>