[PATCH v2] drm/amdgpu: Sync KFD fence only for move/evict

[Harish.Kasiviswanathan@xxxxxxx (Kasiviswanathan, Harish)]

-----Original Message-----
From: Christian König [mailto:ckoenig.leichtzumerken@xxxxxxxxx] 
Sent: Friday, October 20, 2017 2:38 PM
To: Kasiviswanathan, Harish <Harish.Kasiviswanathan at amd.com>; amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH v2] drm/amdgpu: Sync KFD fence only for move/evict

Am 20.10.2017 um 19:18 schrieb Harish Kasiviswanathan:
> A single KFD eviction fence is attached to all the BOs of a process 
> including BOs imported. This fence ensures that all BOs belonging to 
> that process stays resident when the process queues are active.
>
> Don't add this eviction fence to any sync object unless it is a move 
> or evict job. These jobs are identified by the fence owner 
> AMDGPU_FENCE_OWNER_UNDEFINED
>
> v2: Always sync to exclusive fence
>
> Change-Id: I8752d1cf6b2a1c4f2a57292b7c2cd308d5b6f9b7
> Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c | 8 ++++----
>   1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
> index a4f0ecf..88e49f7 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
> @@ -199,10 +199,10 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
>   		return -EINVAL;
>   
>   	f = reservation_object_get_excl(resv);
> -	fence_owner = amdgpu_sync_get_owner(f);
> -	if (fence_owner != AMDGPU_FENCE_OWNER_KFD ||
> -			owner != AMDGPU_FENCE_OWNER_VM)
> +	if (f) {
>   		r = amdgpu_sync_fence(adev, sync, f);
> +		return r;
> +	}

That is not correct either.

[HK]: I thought resv. object can have either an exclusive fence or shared fences but not both at the same time. Is that correct? 

It must look like this:
> Â Â Â Â Â Â Â  /* always sync to the exclusive fence */
> Â Â Â Â Â Â Â  f = reservation_object_get_excl(resv);
> Â Â Â Â Â Â Â  r = amdgpu_sync_fence(adev, sync, f);
>
> Â Â Â Â Â Â Â  flist = reservation_object_get_list(resv);
> Â Â Â Â Â Â Â  if (!flist || r)
> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â  return r
Otherwise you ignore all shared fences and completely mess up the dependencies handling.

I suggest to just revert the patch which originally changed the code away from what it looks like on amd-staging-drm-next.

[HK]: This patch is going to 17.40 release branch. There is a serious performance issue in SSG benchmark and this is caused by unnecessary evictions.

Regards,
Christian.


>   
>   	if (explicit_sync)
>   		return r;
> @@ -216,7 +216,7 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
>   					      reservation_object_held(resv));
>   		fence_owner = amdgpu_sync_get_owner(f);
>   		if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
> -				owner == AMDGPU_FENCE_OWNER_VM)
> +		    owner != AMDGPU_FENCE_OWNER_UNDEFINED)
>   			continue;
>   
>   		if (amdgpu_sync_same_dev(adev, f)) {