RE: [PATCH] drm/amd/amdgpu: vm entities should have kernel priority

["Chen, JingWen" <JingWen.Chen2@amd.com>]

[AMD Official Use Only]

Hi Christian,

I agree that changing the priority is not the right way.
So to not consider paging job guilty, do you think it's OK to not increase_karma for it(if (job->vmid == 0))?

Best Regards,
JingWen Chen

-----Original Message-----
From: Christian König <ckoenig.leichtzumerken@gmail.com>
Sent: Tuesday, July 20, 2021 4:29 PM
To: Chen, JingWen <JingWen.Chen2@amd.com>; Liu, Monk <Monk.Liu@amd.com>; amd-gfx@lists.freedesktop.org
Cc: Chen, Horace <Horace.Chen@amd.com>
Subject: Re: [PATCH] drm/amd/amdgpu: vm entities should have kernel priority

Hi JingWen,

you can look at the job->vm pointer to distinct an userspace submission
from a kernel submission.

The priority is not really related to the submission type, we just
happen to treat paging jobs with the highest priority since they are
important for the system as a whole.

Regards,
Christian.

Am 20.07.21 um 09:49 schrieb Chen, JingWen:
> [AMD Official Use Only]
>
> Hi Christian,
>
> Even if this is a userspace mapping, it's still packaged by the kernel, so it's always assumed to be correct. In which case modifying the priority should have no side effects. May I know the detailed reason for your concern?
>
> And if we eventually decide not to change the priority, do you have any suggestions about how to make sure the paging jobs are not considered guilty?
>
> Best Regards,
> JingWen Chen
>
> -----Original Message-----
> From: Christian König <ckoenig.leichtzumerken@gmail.com>
> Sent: Monday, July 19, 2021 7:10 PM
> To: Liu, Monk <Monk.Liu@amd.com>; Chen, JingWen <JingWen.Chen2@amd.com>; amd-gfx@lists.freedesktop.org
> Cc: Chen, Horace <Horace.Chen@amd.com>
> Subject: Re: [PATCH] drm/amd/amdgpu: vm entities should have kernel priority
>
> Am 19.07.21 um 11:42 schrieb Liu, Monk:
>> [AMD Official Use Only]
>>
>> Besides, I think our current KMD have three types of kernel sdma jobs:
>> 1) adev->mman.entity, it is already a KERNEL priority entity
>> 2) vm->immediate
>> 3) vm->delay
>>
>> Do you mean now vm->immediate or delay are used as moving jobs instead of mman.entity ?
> No, exactly that's the point. vm->immediate and vm->delayed are not for kernel paging jobs.
>
> Those are used for userspace page table updates.
>
> I agree that those should probably not considered guilty, but modifying the priority of them is not the right way of doing that.
>
> Regards,
> Christian.
>
>> Thanks
>>
>> ------------------------------------------
>> Monk Liu | Cloud-GPU Core team
>> ------------------------------------------
>>
>> -----Original Message-----
>> From: Liu, Monk
>> Sent: Monday, July 19, 2021 5:40 PM
>> To: 'Christian König' <ckoenig.leichtzumerken@gmail.com>; Chen,
>> JingWen <JingWen.Chen2@amd.com>; amd-gfx@lists.freedesktop.org
>> Cc: Chen, Horace <Horace.Chen@amd.com>
>> Subject: RE: [PATCH] drm/amd/amdgpu: vm entities should have kernel
>> priority
>>
>> [AMD Official Use Only]
>>
>> If there is move jobs clashing there we probably need to fix the bugs
>> of those move jobs
>>
>> Previously I believe you also remember that we agreed to always trust
>> kernel jobs especially paging jobs,
>>
>> Without set paging jobs' priority to KERNEL level how can we keep that protocol ? do you have a better idea?
>>
>> Thanks
>>
>> ------------------------------------------
>> Monk Liu | Cloud-GPU Core team
>> ------------------------------------------
>>
>> -----Original Message-----
>> From: Christian König <ckoenig.leichtzumerken@gmail.com>
>> Sent: Monday, July 19, 2021 4:25 PM
>> To: Chen, JingWen <JingWen.Chen2@amd.com>;
>> amd-gfx@lists.freedesktop.org
>> Cc: Chen, Horace <Horace.Chen@amd.com>; Liu, Monk <Monk.Liu@amd.com>
>> Subject: Re: [PATCH] drm/amd/amdgpu: vm entities should have kernel
>> priority
>>
>> Am 19.07.21 um 07:57 schrieb Jingwen Chen:
>>> [Why]
>>> Current vm_pte entities have NORMAL priority, in SRIOV multi-vf use
>>> case, the vf flr happens first and then job time out is found.
>>> There can be several jobs timeout during a very small time slice.
>>> And if the innocent sdma job time out is found before the real bad
>>> job, then the innocent sdma job will be set to guilty as it only has
>>> NORMAL priority. This will lead to a page fault after resubmitting
>>> job.
>>>
>>> [How]
>>> sdma should always have KERNEL priority. The kernel job will always
>>> be resubmitted.
>> I'm not sure if that is a good idea. We intentionally didn't gave the page table updates kernel priority to avoid clashing with the move jobs.
>>
>> Christian.
>>
>>> Signed-off-by: Jingwen Chen <Jingwen.Chen2@amd.com>
>>> ---
>>>     drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 ++--
>>>     1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> index 358316d6a38c..f7526b67cc5d 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> @@ -2923,13 +2923,13 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
>>>       INIT_LIST_HEAD(&vm->done);
>>>
>>>       /* create scheduler entities for page table updates */
>>> -    r = drm_sched_entity_init(&vm->immediate, DRM_SCHED_PRIORITY_NORMAL,
>>> +    r = drm_sched_entity_init(&vm->immediate,
>>> +DRM_SCHED_PRIORITY_KERNEL,
>>>                                 adev->vm_manager.vm_pte_scheds,
>>>                                 adev->vm_manager.vm_pte_num_scheds, NULL);
>>>       if (r)
>>>               return r;
>>>
>>> -    r = drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_NORMAL,
>>> +    r = drm_sched_entity_init(&vm->delayed, DRM_SCHED_PRIORITY_KERNEL,
>>>                                 adev->vm_manager.vm_pte_scheds,
>>>                                 adev->vm_manager.vm_pte_num_scheds, NULL);
>>>       if (r)

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx