Testing GTT ready might be more correct but cp.ready
works fine and has been tested on irc by 2-3 ppl.
fixes bug k.org 15035 and fd.o 25733
Signed-off-by: Dave Airlie <airlied@redhat.com>
        rbo = container_of(bo, struct radeon_bo, tbo);
        switch (bo->mem.mem_type) {
        case TTM_PL_VRAM:
-               radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
+               if (rbo->rdev->cp.ready == false)
+                       radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
+               else
+                       radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
                break;
        case TTM_PL_TT:
        default: