Re: [PATCH] drm/amdgpu: skip huge page for PRT mapping

["Zhang, Jerry (Junwei)" <Jerry.Zhang-5C7GfCeVMHo@public.gmane.org>]

On 06/04/2018 07:01 PM, Christian König wrote:
> I've figured out what's going wrong here.
>
> Your analysis of the problem is correct, but the proposed fix just doesn't
> handle everything.
>
> The issue is that I assumed in the replace operation that the existing lose end
> of the mapping can stay as they are, but that assumption is actually not correct
> because of huge page handling.
>
> This not only affects PRTs, but all mappings in general, so that's a rather
> severe bug we need to fix.
>
> Just send a only compile tested patch with subject "drm/amdgpu: fix clear_all
> and replace handling in the VM" to the list, please give that one a try and with
> your change.

Sorry, I didn't find the patch in amdgfx or internal group.
If I missed any, please send the patch to me again.

I may get your meaning to handle the huge page in amdgpu_vm_bo_clear_mappings().
If the replaced mapping is huge page, then it has to do mapping for else part.
rather than map the target one only.
(in this case, the "after" part has to do mapping, inheriting the flags from the 
replaced one who has PRT flag)

Try to achieve that locally, found it's a bit tricky to do mapping for else 
part, since the else part(reserved PRT range) bo_va is different from the target 
one(tiled bo). so has to do bo update for else bo_va.
(under debugging...)

>
> Thanks,
> Christian.
>
> Am 04.06.2018 um 11:51 schrieb Christian König:
>> Am 04.06.2018 um 10:19 schrieb Zhang, Jerry (Junwei):
>>> On 06/04/2018 03:48 PM, Christian König wrote:
>>>> Am 04.06.2018 um 09:02 schrieb Zhang, Jerry (Junwei):
>>>>> On 06/04/2018 02:43 PM, Christian König wrote:
>>>>>> Actually that is not correct. According to the documentation the PRT flag
>>>>>> should
>>>>>> work for huge pages as well.
>>>>>
>>>>> Mmm, I checked the doc earlier, didn't find the PRT flag for PDE.
>>>>
>>>> The PDE indeed doesn't have the PRT flag, but it also doesn't have the fragment
>>>> and MTYPE fields and those still work normally when you enable huge page
>>>> handling.
>>>>
>>>> The trick is that when you set the huge page flag then the PDE is handled
>>>> like a
>>>> PTE and so all the extra fields (PRT, fragment size, MTYPE etc...) should
>>>> now be
>>>> handled correctly.
>>>>
>>>> Could be that there is a hardware bug related to PRT handling and the huge page
>>>> flag, but at least in theory it should work fine.
>>>
>>> The doc just skips these fields in PDE, if those fields really works
>>> expectedly, even if it doesn't describe the details(but we could refer to PTE
>>> fields), we may have to confirm that with HW guys.
>>
>> I've just confirmed that huge pages and PRTs should work together. We even
>> have an unit test for that.
>>
>> I mean we could ping Wade as well, but I'm pretty sure that he will tell you
>> just the same.
>>
>> I think by disabling huge pages for PRTs we actually hide the real issue
>> somehow, probably some problem with updating PRTs after they are split up.
>>
>> Can you check if the problem also vanishes when you disable the following
>> optimization in amdgpu_vm_update_ptes?
>>>                 /* We don't need to update PTEs for huge pages */
>>>                 if (entry->huge)
>>>                         continue;
>> Just comment this out for a test.
>>
>>>
>>> In my view in current stage, PDE doesn't support PRT and it may not make much
>>> sense to support PRT either.
>>> The huge page always happens when reserving PRT range, but later UMD is
>>> likely to bind a/some tiled bo(s) inside this range, that will break the huge
>>> page mapping and split into several pieces mappings, representing in PTE
>>> instead of huge page.
>>> In this case, just skipping the huge page for PRT may be an acceptable way.
>>
>> Well disabling huge pages for PRTs can definitely cause a huge performance
>> problem because it increases the TLB misses for PRTs by a factor of 512.
>>
>> And since PRTs are usually used for large only sparse allocated textures that
>> is really really important here.
>>
>>>
>>>>
>>>>>
>>>>> In CTS PRT test, the reserved PRT mapping introduces huge page mapping, so
>>>>> later tiled bo mapping cannot make sure the corresponding PTE is set as PRT.
>>>>> Then following access triggers VM fault.
>>>>
>>>> Interesting, but that rather sounds like a bug in the handling instead of a
>>>> hardware problem.
>>>
>>> On the 2nd thinking, we may handle that for huge page.
>>> e.g.when binding tiled bo in PRT range, we could split the huge page into
>>> pieces PTE mappings.
>>> However, it may just make the code more complex and get the same results as
>>> current fix.
>>
>> That should already be the case when everything works as expected.
>>
>>>
>>>>
>>>> Can you narrow down the CTS test further into a libdrm unit test? Or in other
>>>> words what exactly does the CTS test do?
>>>
>>> The issue could be reproduced by below command:
>>> {{{
>>> deqp-vk -n
>>> dEQP-VK.sparse_resources.buffer.ssbo.sparse_residency.buffer_size_2_24
>>> }}}
>>>
>>> In KMD view, the main process(with amdgpu mainline driver) is like below:
>>> 1) reserve PRT range [0x300400000, 0x300400000 + 0x1000000)
>>> 2) bind tiled bos each other page
>>>    0x300400000 ~ 0x300401000,
>>>    0x300402000 ~ 0x300403000,
>>>    0x300404000 ~ 0x300405000,
>>>    ...
>>> 3) access them all, VM fault at 0x300401000
>>
>> That sounds like something I should be able to reproduce when your quick test
>> doesn't bring the desired result.
>>
>> Christian.
>>
>>>
>>> Jerry
>>>
>>>>
>>>> Christian.
>>>>
>>>>>
>>>>> Jerry
>>>>>
>>>>>>
>>>>>> Christian.
>>>>>>
>>>>>> Am 04.06.2018 um 07:59 schrieb Zhou, David(ChunMing):
>>>>>>> Good catch, Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
>>>>>>>
>>>>>>> -----Original Message-----
>>>>>>> From: amd-gfx [mailto:amd-gfx-bounces@lists.freedesktop.org] On Behalf Of
>>>>>>> Junwei Zhang
>>>>>>> Sent: Monday, June 04, 2018 10:04 AM
>>>>>>> To: amd-gfx@lists.freedesktop.org
>>>>>>> Cc: Zhang, Jerry <Jerry.Zhang@amd.com>
>>>>>>> Subject: [PATCH] drm/amdgpu: skip huge page for PRT mapping
>>>>>>>
>>>>>>> PRT mapping doesn't support huge page, since it's per PTE basis.
>>>>>>>
>>>>>>> Signed-off-by: Junwei Zhang <Jerry.Zhang@amd.com>
>>>>>>> ---
>>>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 3 ++-
>>>>>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> index 850cd66..4ce8bb0 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> @@ -1111,7 +1111,8 @@ static void amdgpu_vm_handle_huge_pages(struct
>>>>>>> amdgpu_pte_update_params *p,
>>>>>>>       /* In the case of a mixed PT the PDE must point to it*/
>>>>>>>       if (p->adev->asic_type >= CHIP_VEGA10 && !p->src &&
>>>>>>> -        nptes == AMDGPU_VM_PTE_COUNT(p->adev)) {
>>>>>>> +        nptes == AMDGPU_VM_PTE_COUNT(p->adev) &&
>>>>>>> +        !(flags & AMDGPU_PTE_PRT)) {
>>>>>>>           /* Set the huge page flag to stop scanning at this PDE */
>>>>>>>           flags |= AMDGPU_PDE_PTE;
>>>>>>>       }
>>>>>>
>>>>
>>
>
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx