drm/amdgpu: Attach exclusive fence to prime exported bo's. (v5)

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.

v5: - Drop lret, instead use int ret -> long ret, as proposed
      by Christian.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=95472
Tested-by: Mike Lothian <mike@fireburn.co.uk> (v1)
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: Christian König <christian.koenig@amd.com>.
Cc: Christian König <christian.koenig@amd.com>
Cc: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

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
 	 */

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;

drivers/gpu/drm/amd/amdgpu/amdgpu_prime.c

@@ -74,20 +74,36 @@ 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;
 }
 
 int amdgpu_gem_prime_pin(struct drm_gem_object *obj)
 {
 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
-	int ret = 0;
+	long ret = 0;
 
 	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.
+	 */
+	ret = reservation_object_wait_timeout_rcu(bo->tbo.resv, true, false,
+						  MAX_SCHEDULE_TIMEOUT);
+	if (unlikely(ret < 0)) {
+		DRM_DEBUG_PRIME("Fence wait failed: %li\n", ret);
+		amdgpu_bo_unreserve(bo);
+		return ret;
+	}
+
 	/* 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 +118,8 @@ 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);
 }