drm/amdgpu: write PTEs directly into the IB.

Write the PTEs at the end of the IB instead of directly into the SDMA commands.
This can save quite some CPU cycles building the entries.

This doesn't change the DW estimation because PTEs where embedded into the IB
before as well. It just moves them to the end of the IB.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -911,15 +911,15 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	/* padding, etc. */
 	ndw = 64;
 
-	if (params.src) {
+	if (src) {
 		/* only copy commands needed */
 		ndw += ncmds * 7;
 
-	} else if (params.pages_addr) {
-		/* header for write data commands */
-		ndw += ncmds * 4;
+	} else if (pages_addr) {
+		/* copy commands needed */
+		ndw += ncmds * 7;
 
-		/* body of write data command */
+		/* and also PTEs */
 		ndw += nptes * 2;
 
 	} else {
@@ -936,6 +936,22 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 	params.ib = &job->ibs[0];
 
+	if (!src && pages_addr) {
+		uint64_t *pte;
+		unsigned i;
+
+		/* Put the PTEs at the end of the IB. */
+		i = ndw - nptes * 2;
+		pte= (uint64_t *)&(job->ibs->ptr[i]);
+		params.src = job->ibs->gpu_addr + i * 4;
+
+		for (i = 0; i < nptes; ++i) {
+			pte[i] = amdgpu_vm_map_gart(pages_addr, addr + i *
+						    AMDGPU_GPU_PAGE_SIZE);
+			pte[i] |= flags;
+		}
+	}
+
 	r = amdgpu_sync_fence(adev, &job->sync, exclusive);
 	if (r)
 		goto error_free;