drm/amdgpu: disable ring_muxer if mcbp is off

Using the ring_muxer without preemption adds overhead for no
reason since mcbp cannot be triggered.

Moving back to a single queue in this case also helps when
high priority app are used: in this case the gpu_scheduler
priority handling will work as expected - much better than
ring_muxer with its 2 independant schedulers competing for
the same hardware queue.

This change requires moving amdgpu_device_set_mcbp above
amdgpu_device_ip_early_init because we use adev->gfx.mcbp.
Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jiadong Zhu <Jiadong.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -4056,13 +4056,13 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 		goto unmap_memory;
 	}
 
+	amdgpu_device_set_mcbp(adev);
+
 	/* early init functions */
 	r = amdgpu_device_ip_early_init(adev);
 	if (r)
 		goto unmap_memory;
 
-	amdgpu_device_set_mcbp(adev);
-
 	/* Get rid of things like offb */
 	r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver);
 	if (r)

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c

@@ -2080,7 +2080,7 @@ static int gfx_v9_0_sw_init(void *handle)
 		ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1;
 
 		/* disable scheduler on the real ring */
-		ring->no_scheduler = true;
+		ring->no_scheduler = adev->gfx.mcbp;
 		ring->vm_hub = AMDGPU_GFXHUB(0);
 		r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq,
 				     AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP,
@@ -2090,7 +2090,7 @@ static int gfx_v9_0_sw_init(void *handle)
 	}
 
 	/* set up the software rings */
-	if (adev->gfx.num_gfx_rings) {
+	if (adev->gfx.mcbp && adev->gfx.num_gfx_rings) {
 		for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++) {
 			ring = &adev->gfx.sw_gfx_ring[i];
 			ring->ring_obj = NULL;
@@ -2180,7 +2180,7 @@ static int gfx_v9_0_sw_fini(void *handle)
 	int i;
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
-	if (adev->gfx.num_gfx_rings) {
+	if (adev->gfx.mcbp && adev->gfx.num_gfx_rings) {
 		for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++)
 			amdgpu_ring_fini(&adev->gfx.sw_gfx_ring[i]);
 		amdgpu_ring_mux_fini(&adev->gfx.muxer);
@@ -5899,12 +5899,15 @@ static int gfx_v9_0_eop_irq(struct amdgpu_device *adev,
 
 	switch (me_id) {
 	case 0:
-		if (adev->gfx.num_gfx_rings &&
-		    !amdgpu_mcbp_handle_trailing_fence_irq(&adev->gfx.muxer)) {
+		if (adev->gfx.num_gfx_rings) {
+			if (!adev->gfx.mcbp) {
+				amdgpu_fence_process(&adev->gfx.gfx_ring[0]);
+			} else if (!amdgpu_mcbp_handle_trailing_fence_irq(&adev->gfx.muxer)) {
 				/* Fence signals are handled on the software rings*/
 				for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++)
 					amdgpu_fence_process(&adev->gfx.sw_gfx_ring[i]);
 			}
+		}
 		break;
 	case 1:
 	case 2:
@@ -7038,7 +7041,7 @@ static void gfx_v9_0_set_ring_funcs(struct amdgpu_device *adev)
 	for (i = 0; i < adev->gfx.num_gfx_rings; i++)
 		adev->gfx.gfx_ring[i].funcs = &gfx_v9_0_ring_funcs_gfx;
 
-	if (adev->gfx.num_gfx_rings) {
+	if (adev->gfx.mcbp && adev->gfx.num_gfx_rings) {
 		for (i = 0; i < GFX9_NUM_SW_GFX_RINGS; i++)
 			adev->gfx.sw_gfx_ring[i].funcs = &gfx_v9_0_sw_ring_funcs_gfx;
 	}