[PATCH] drm/amdgpu: Forbid kfd using cpu to update pt if vm is shared with gfx

[Christian König]

I have no problem dropping the AMDGPU_GEM_CREATE_NO_CPU_ACCESS flag from 
page tables no matter if we have large BAR or not.

That should only be necessary when BOs are pinned and since we never pin 
page tables it shouldn't matter.

Just set AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED after switching to CPU 
based updates for all already allocated page tables.

Regards,
Christian.

Am 21.06.23 um 17:16 schrieb Felix Kuehling:
> Can we change the flags if needed. E.g. see what 
> amdgpu_bo_pin_restricted does:
>
>         if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
>                 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
>         amdgpu_bo_placement_from_domain(bo, domain);
>
> This shouldn't really change anything about the BO placement because 
> we only enable CPU page table updates on large-BAR GPUs by default. 
> Alternatively, we could create VM BOs with 
> AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED on large-BAR GPUs to make it 
> possible to switch to CPU page table updates for compute VMs.
>
> Regards,
>   Felix
>
>
> Am 2023-06-21 um 05:46 schrieb YuBiao Wang:
>> If a same GPU VM is shared by kfd and graphic operations, we must align
>> the vm update mode to sdma, or cpu kmap will fail and cause null pointer
>> issue.
>>
>> Signed-off-by: YuBiao Wang <YuBiao.Wang at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 5 +++++
>>   1 file changed, 5 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> index 291977b93b1d..e105ff9e8041 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>> @@ -2239,6 +2239,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, 
>> struct amdgpu_vm *vm)
>>   int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct 
>> amdgpu_vm *vm)
>>   {
>>       bool pte_support_ats = (adev->asic_type == CHIP_RAVEN);
>> +    struct amdgpu_bo *bo = vm->root.bo;
>>       int r;
>>         r = amdgpu_bo_reserve(vm->root.bo, true);
>> @@ -2265,6 +2266,10 @@ int amdgpu_vm_make_compute(struct 
>> amdgpu_device *adev, struct amdgpu_vm *vm)
>>       /* Update VM state */
>>       vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
>>                       AMDGPU_VM_USE_CPU_FOR_COMPUTE);
>> +
>> +    if (bo && !(bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED))
>> +        vm->use_cpu_for_update = false;
>> +
>>       DRM_DEBUG_DRIVER("VM update mode is %s\n",
>>                vm->use_cpu_for_update ? "CPU" : "SDMA");
>>       WARN_ONCE((vm->use_cpu_for_update &&