if (ret)
                return ret;
 
-       intel_runtime_pm_get(i915);
+       if (i915_gem_object_has_struct_page(obj)) {
+               /*
+                * Avoid waking the device up if we can fallback, as
+                * waking/resuming is very slow (worst-case 10-100 ms
+                * depending on PCI sleeps and our own resume time).
+                * This easily dwarfs any performance advantage from
+                * using the cache bypass of indirect GGTT access.
+                */
+               if (!intel_runtime_pm_get_if_in_use(i915)) {
+                       ret = -EFAULT;
+                       goto out_unlock;
+               }
+       } else {
+               /* No backing pages, no fallback, we must force GGTT access */
+               intel_runtime_pm_get(i915);
+       }
+
        vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
                                       PIN_MAPPABLE |
                                       PIN_NONFAULT |
        if (IS_ERR(vma)) {
                ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
                if (ret)
-                       goto out_unlock;
+                       goto out_rpm;
                GEM_BUG_ON(!node.allocated);
        }
 
        } else {
                i915_vma_unpin(vma);
        }
-out_unlock:
+out_rpm:
        intel_runtime_pm_put(i915);
+out_unlock:
        mutex_unlock(&i915->drm.struct_mutex);
        return ret;
 }