[PATCH v4 5/5] drm/xe: Refactor default device atomic settings

[Nirmoy Das]

Hi Brian,

On 4/29/2024 10:14 PM, Welty, Brian wrote:
>
>
> On 4/25/2024 3:23 PM, Nirmoy Das wrote:
>> The default behavior of device atomics depends on the
>> VM type and buffer allocation types. Device atomics are
>> expected to function with all types of allocations for
>> traditional applications/APIs. Additionally, in compute/SVM
>> API scenarios with fault mode or LR mode VMs, device atomics
>> must work with single-region allocations. 
>
> I think additional patch may be needed..... for the Compute mode
> handling.
> I'm not sure correct thing will happened for 'shared' 
> (multi-placement) allocations.
> HW will raise an atomic access page fault.
> Our page fault handler only checks this was atomic fault and proceeds to
> migrate the BO to VRAM. I think PTE_AE bit won't be set with your
> changes, so seems are in infinite loop of atomic access faults on same 
> address.
>
> I think we need to fail the page fault in this case, so HW raise CAT 
> error here? As atomic access not supposed to be allowed for above case
> until future uAPI is added?
>
> And need IGT test for this.


Thinking more into that, I just realized we need preferred location 
madvise hint. Currently I am not aware of

a good way to change backend of SMEM+LMEM bo. We don't migrate the BO to 
smem on CPU access, that would've helped too

but we need those to be implemented which is my next priority after 
getting the defaults fixed with this series.


Regards,

Nirmoy

>
> -Brian
>
>
>> In all other cases
>> device atomics should be disabled by default.
>>
>> Signed-off-by: Nirmoy Das <nirmoy.das at intel.com>
>> ---
>>   drivers/gpu/drm/xe/xe_pt.c | 24 ++++++++++++++++++++----
>>   drivers/gpu/drm/xe/xe_vm.c |  3 ++-
>>   2 files changed, 22 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
>> index 5b7930f46cf3..a8e9e8592c43 100644
>> --- a/drivers/gpu/drm/xe/xe_pt.c
>> +++ b/drivers/gpu/drm/xe/xe_pt.c
>> @@ -597,7 +597,6 @@ static int
>>   xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
>>            struct xe_vm_pgtable_update *entries, u32 *num_entries)
>>   {
>> -    struct xe_device *xe = tile_to_xe(tile);
>>       struct xe_bo *bo = xe_vma_bo(vma);
>>       bool is_devmem = !xe_vma_is_userptr(vma) && bo &&
>>           (xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo));
>> @@ -619,9 +618,26 @@ xe_pt_stage_bind(struct xe_tile *tile, struct 
>> xe_vma *vma,
>>       struct xe_pt *pt = xe_vma_vm(vma)->pt_root[tile->id];
>>       int ret;
>>   -    if ((vma->gpuva.flags & XE_VMA_ATOMIC_PTE_BIT) &&
>> -        (is_devmem || !IS_DGFX(xe)))
>> -        xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
>> +    /**
>> +     * Default atomic expectations for different allocation 
>> scenarios are as follows:
>> +     *
>> +     * 1. Traditional API: When the VM is not in fault mode or LR mode:
>> +     *    - Device atomics are expected to function with all 
>> allocations.
>> +     *
>> +     * 2. Compute/SVM API: When the VM is either in fault mode or LR 
>> mode:
>> +     *    - Device atomics are the default behavior when the bo is 
>> placed in a single region.
>> +     *    - In all other cases device atomics will be disabled with 
>> AE=0 until an application
>> +     *      request differently using a ioctl like madvise.
>> +     */
>> +    if (vma->gpuva.flags & XE_VMA_ATOMIC_PTE_BIT) {
>> +        if (xe_vm_in_fault_mode(xe_vma_vm(vma)) ||
>> +            xe_vm_in_lr_mode(xe_vma_vm(vma))) {
>> +            if (bo && xe_bo_has_single_placement(bo))
>> +                xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
>> +        } else {
>> +            xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
>> +        }
>> +    }
>>         if (is_devmem) {
>>           xe_walk.default_pte |= XE_PPGTT_PTE_DM;
>> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
>> index e41345c1627d..ac08b6fd537e 100644
>> --- a/drivers/gpu/drm/xe/xe_vm.c
>> +++ b/drivers/gpu/drm/xe/xe_vm.c
>> @@ -805,7 +805,8 @@ static struct xe_vma *xe_vma_create(struct xe_vm 
>> *vm,
>>       for_each_tile(tile, vm->xe, id)
>>           vma->tile_mask |= 0x1 << id;
>>   -    if (GRAPHICS_VER(vm->xe) >= 20 || vm->xe->info.platform == 
>> XE_PVC)
>> +    if (vm->xe->info.has_atomic_enable_pte_bit &&
>> +        vm->xe->info.has_device_atomics_on_smem)
>>           vma->gpuva.flags |= XE_VMA_ATOMIC_PTE_BIT;
>>         vma->pat_index = pat_index;