drm/amdgpu: fix using the reserved VMID with gang submit

We need to ensure that even when using a reserved VMID that the gang
members can still run in parallel.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -1441,6 +1441,7 @@ u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
 				u32 reg);
 void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
 				u32 reg, u32 v);
+struct dma_fence *amdgpu_device_get_gang(struct amdgpu_device *adev);
 struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
 					    struct dma_fence *gang);
 bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev);

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

@@ -6522,6 +6522,22 @@ void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
 	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 }
 
+/**
+ * amdgpu_device_get_gang - return a reference to the current gang
+ * @adev: amdgpu_device pointer
+ *
+ * Returns: A new reference to the current gang leader.
+ */
+struct dma_fence *amdgpu_device_get_gang(struct amdgpu_device *adev)
+{
+	struct dma_fence *fence;
+
+	rcu_read_lock();
+	fence = dma_fence_get_rcu_safe(&adev->gang_submit);
+	rcu_read_unlock();
+	return fence;
+}
+
 /**
  * amdgpu_device_switch_gang - switch to a new gang
  * @adev: amdgpu_device pointer
@@ -6538,10 +6554,7 @@ struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
 
 	do {
 		dma_fence_put(old);
-		rcu_read_lock();
-		old = dma_fence_get_rcu_safe(&adev->gang_submit);
-		rcu_read_unlock();
-
+		old = amdgpu_device_get_gang(adev);
 		if (old == gang)
 			break;
 

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

@@ -290,7 +290,24 @@ static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,
 	     !dma_fence_is_signaled((*id)->last_flush))) {
 		struct dma_fence *tmp;
 
-		/* Don't use per engine and per process VMID at the same time */
+		/* Wait for the gang to be assembled before using a
+		 * reserved VMID or otherwise the gang could deadlock.
+		 */
+		tmp = amdgpu_device_get_gang(adev);
+		if (!dma_fence_is_signaled(tmp) && tmp != job->gang_submit) {
+			*id = NULL;
+			*fence = tmp;
+			return 0;
+		}
+		dma_fence_put(tmp);
+
+		/* Make sure the id is owned by the gang before proceeding */
+		if (!job->gang_submit ||
+		    (*id)->owner != vm->immediate.fence_context) {
+
+			/* Don't use per engine and per process VMID at the
+			 * same time
+			 */
 			if (adev->vm_manager.concurrent_flush)
 				ring = NULL;
 
@@ -302,6 +319,7 @@ static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,
 				*fence = dma_fence_get(tmp);
 				return 0;
 			}
+		}
 		needs_flush = true;
 	}