[PATCH] drm/amdgpu: Attach exclusive fence to prime exported bo's. (v4)

[Christian König]

Am 08.11.2016 um 04:15 schrieb Mario Kleiner:
> External clients which import our bo's wait only
> for exclusive dmabuf-fences, not on shared ones,
> ditto for bo's which we import from external
> providers and write to.
>
> Therefore attach exclusive fences on prime shared buffers
> if our exported buffer gets imported by an external
> client, or if we import a buffer from an external
> exporter.
>
> See discussion in thread:
> https://lists.freedesktop.org/archives/dri-devel/2016-October/122370.html
>
> Prime export tested on Intel iGPU + AMD Tonga dGPU as
> DRI3/Present Prime render offload, and with the Tonga
> standalone as primary gpu.
>
> v2: Add a wait for all shared fences before prime export,
>      as suggested by Christian Koenig.
>
> v3: - Mark buffer prime_exported in amdgpu_gem_prime_pin,
>      so we only use the exclusive fence when exporting a
>      bo to external clients like a separate iGPU, but not
>      when exporting/importing from/to ourselves as part of
>      regular DRI3 fd passing.
>
>      - Propagate failure of reservation_object_wait_rcu back
>      to caller.
>
> v4: - Switch to a prime_shared_count counter instead of a
>        flag, which gets in/decremented on prime_pin/unpin, so
>        we can switch back to shared fences if all clients
>        detach from our exported bo.
>
>      - Also switch to exclusive fence for prime imported bo's.
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95472
> Tested-by: Mike Lothian <mike at fireburn.co.uk> (v1)
> Signed-off-by: Mario Kleiner <mario.kleiner.de at gmail.com>
> Cc: Christian König <christian.koenig at amd.com>
> Cc: Michel Dänzer <michel.daenzer at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu.h         |  1 +
>   drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c |  2 +-
>   drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c   | 19 +++++++++++++++++++
>   3 files changed, 21 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 039b57e..496f72b 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -459,6 +459,7 @@ struct amdgpu_bo {
>   	u64				metadata_flags;
>   	void				*metadata;
>   	u32				metadata_size;
> +	unsigned			prime_shared_count;
>   	/* list of all virtual address to which this bo
>   	 * is associated to
>   	 */
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c
> index 651115d..c02db01f6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_bo_list.c
> @@ -132,7 +132,7 @@ static int amdgpu_bo_list_set(struct amdgpu_device *adev,
>   		entry->priority = min(info[i].bo_priority,
>   				      AMDGPU_BO_LIST_MAX_PRIORITY);
>   		entry->tv.bo = &entry->robj->tbo;
> -		entry->tv.shared = true;
> +		entry->tv.shared = !entry->robj->prime_shared_count;
>   
>   		if (entry->robj->prefered_domains == AMDGPU_GEM_DOMAIN_GDS)
>   			gds_obj = entry->robj;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
> index 7700dc2..31aac7e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c
> @@ -74,6 +74,7 @@ amdgpu_gem_prime_import_sg_table(struct drm_device *dev,
>   	if (ret)
>   		return ERR_PTR(ret);
>   
> +	bo->prime_shared_count = 1;
>   	return &bo->gem_base;
>   }
>   
> @@ -81,13 +82,29 @@ int amdgpu_gem_prime_pin(struct drm_gem_object *obj)
>   {
>   	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
>   	int ret = 0;
> +	long lret;

No need for an extra long sized variable here, just change the type of 
the existing one.

The compiler can usually optimize away any size conversions perfectly fine.

With that fixed the patch is Reviewed-by: Christian König 
<christian.koenig at amd.com>.

Thanks a lot for the help,
Christian.

>   
>   	ret = amdgpu_bo_reserve(bo, false);
>   	if (unlikely(ret != 0))
>   		return ret;
>   
> +	/*
> +	 * Wait for all shared fences to complete before we switch to future
> +	 * use of exclusive fence on this prime shared bo.
> +	 */
> +	lret = reservation_object_wait_timeout_rcu(bo->tbo.resv, true, false,
> +						   MAX_SCHEDULE_TIMEOUT);
> +	if (unlikely(lret < 0)) {
> +		DRM_DEBUG_PRIME("Fence wait failed: %li\n", lret);
> +		amdgpu_bo_unreserve(bo);
> +		return lret;
> +	}
> +
>   	/* pin buffer into GTT */
>   	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT, NULL);
> +	if (likely(ret == 0))
> +		bo->prime_shared_count++;
> +
>   	amdgpu_bo_unreserve(bo);
>   	return ret;
>   }
> @@ -102,6 +119,8 @@ void amdgpu_gem_prime_unpin(struct drm_gem_object *obj)
>   		return;
>   
>   	amdgpu_bo_unpin(bo);
> +	if (bo->prime_shared_count)
> +		bo->prime_shared_count--;
>   	amdgpu_bo_unreserve(bo);
>   }
>