[PATCH 3/4] drm/amdgpu: Support snooped PTE flag

[Kuehling, Felix]

On 2019-08-24 7:13, Christian König wrote:
> Am 23.08.19 um 23:33 schrieb Kuehling, Felix:
>> From: Oak Zeng <Oak.Zeng at amd.com>
>>
>> Set snooped PTE flag according to mapping flag. Write request to a
>> page with snooped bit set, will send out invalidate probe request
>> to TCC of the remote GPU where the vram page resides.
>>
>> Change-Id: I799f68ec7a5a1abf32075f5ef31051641a0b3736
>> Signed-off-by: Oak Zeng <Oak.Zeng at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c 
>> b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> index 9aafcda6c488..8a7c4ec69ae8 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> @@ -604,6 +604,9 @@ static uint64_t gmc_v9_0_get_vm_pte_flags(struct 
>> amdgpu_device *adev,
>>       if (flags & AMDGPU_VM_PAGE_PRT)
>>           pte_flag |= AMDGPU_PTE_PRT;
>>   +    if (flags & AMDGPU_VM_PAGE_INVALIDATE_PROBE)
>> +        pte_flag |= AMDGPU_PTE_SNOOPED;
>> +
>
> That is still a NAK without further checks. We need to make absolutely 
> sure that we don't set this when PCIe routing is in use.

The only place where AMDGPU_VM_... flags are accepted from user mode 
seems to be amdgpu_gem_va_ioctl. It has an explicit set of valid_flags 
it accepts. The new INVALIDATE_PROBE flag is not part of it. That means 
user mode will  not be able to set it directly. If we added it to the 
set of valid_flags in amdgpu_gem_va_ioctl, we'd need to add appropriate 
checks at the same time.

KFD does not expose AMDGPU_VM_... flags directly to user mode. It only 
sets the INVALIDATE_PROBE flag in kernel mode for mappings in the same 
XGMI hive on Arturus (in patch 4).

If there is something I'm missing, please point it out. But AFAICT the 
checking that is currently done should satisfy your requirements.

Regards,
   Felix

>
> Christian.
>
>>       return pte_flag;
>>   }
>