[PATCH 2/2] drm/amdgpu: skip vm entries checking while in sriov vf

[Koenig, Christian]

Am 12.10.2018 um 06:05 schrieb Huang, Ray:
>> -----Original Message-----
>> From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On Behalf
>> Of Felix Kuehling
>> Sent: Friday, October 12, 2018 3:46 AM
>> To: Koenig, Christian <Christian.Koenig at amd.com>; Min, Frank
>> <Frank.Min at amd.com>; amd-gfx at lists.freedesktop.org
>> Subject: Re: [PATCH 2/2] drm/amdgpu: skip vm entries checking while in
>> sriov vf
>>
>> On 2018-10-11 08:32 AM, Christian König wrote:
>>> Am 11.10.2018 um 14:09 schrieb Frank.Min:
>>>> vm page table would be allocated while in amdgpu_vm_init by
>>>> amdgpu_allocate_static_csa for sriov, so the checking here would be
>>>> skipped.
>>> NAK, that checking is actually correct and points out that this
>>> doesn't work correctly.
>>>
>>> Fix the function to be able to handle the case when there are already
>>> mappings in the VM.
>> KFD and the Thunk assume that they own the entire address space. If
>> something is already mapped, KFD and the Thunk would not be aware of it.
>>
>> amdgpu_allocate_static_csa runs during driver initialization, before there are
>> any VMs. Looks like amdgpu_map_static_csa is the more important function
>> here, which currently runs in amdgpu_driver_open_kms.
>>
>> Could amdgpu_map_static_csa be delayed to the first time user mode maps
>> memory into the VM through GEM APIs? That way it would allow turning a
>> VM into a compute VM that hasn't been used for any GEM memory
>> mappings.
>>
> Yes. Frank and I found, while we do acquire_process_vm(), the vm->va RB trees are always not empty under SRIOV, because amdgpu_map_static_csa() will be called while vm inits under SRIOV, that will actually map the CSA BO to the VM and insert the node into vm->va before it acquires process VM in KFD.

Yeah and exactly that will clash with the KFD.

Could be that it still works initially because the CSA is not mapped at 
the beginning of the address space IIRC, but we need to clean that up in 
the long term.

We should either delay adding the CSA BO to the VM (probably best done 
during context creation).

Or we start to make nails with heads and move the CSA BO into the 
context altogether. We are going to need this anyway for MCBP, so that 
would be a good first step.

Ray/Frank can anybody of you take care of this?

Thanks,
Christian.

>   
> Thanks,
> Ray
>
>> Regards,
>>    Felix
>>
>>> Christian.
>>>
>>>> Change-Id: Id30b86ad15ae509aeed9ed8ab60c259c88af3df5
>>>> Signed-off-by: Frank.Min <Frank.Min at amd.com>
>>>> ---
>>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 9 +++++----
>>>>    1 file changed, 5 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> index 6904d79..6066f87 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>> @@ -3100,11 +3100,12 @@ int amdgpu_vm_make_compute(struct
>>>> amdgpu_device *adev, struct amdgpu_vm *vm, uns
>>>>            return r;
>>>>          /* Sanity checks */
>>>> -    if (!RB_EMPTY_ROOT(&vm->va.rb_root) || vm->root.entries) {
>>>> -        r = -EINVAL;
>>>> -        goto unreserve_bo;
>>>> +    if (!amdgpu_sriov_vf(adev)) {
>>>> +        if (!RB_EMPTY_ROOT(&vm->va.rb_root) || vm->root.entries) {
>>>> +            r = -EINVAL;
>>>> +            goto unreserve_bo;
>>>> +        }
>>>>        }
>>>> -
>>>>        if (pasid) {
>>>>            unsigned long flags;
>>>>
>>> _______________________________________________
>>> amd-gfx mailing list
>>> amd-gfx at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx