drm/i915: Flush ggtt writes through the old fenced vma before changing fences

This is a precautionary measure as I have no evidence to suggest we've
hit a bug here (I was hoping this might explain gdg's odd behaviour, but
alas), but given that we have a function to flush the ggtt writes it
seems prudent to use it prior to changing the fence register. Due to the
intrinsic nature of the GTT often operating as an independent mmio path,
we should not just rely on the write to the fence acting as a full flush
for GTT writes.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180130164457.14037-1-chris@chris-wilson.co.uk

diff --git a/drivers/gpu/drm/i915/i915_gem_fence_reg.c b/drivers/gpu/drm/i915/i915_gem_fence_reg.c
index 012250f..b8338d7 100644 (file)
--- a/drivers/gpu/drm/i915/i915_gem_fence_reg.c
+++ b/drivers/gpu/drm/i915/i915_gem_fence_reg.c
@@ -230,10 +230,14 @@ static int fence_update(struct drm_i915_fence_reg *fence,
        }
 
        if (fence->vma) {
-               ret = i915_gem_active_retire(&fence->vma->last_fence,
-                                     &fence->vma->obj->base.dev->struct_mutex);
+               struct i915_vma *old = fence->vma;
+
+               ret = i915_gem_active_retire(&old->last_fence,
+                                            &old->obj->base.dev->struct_mutex);
                if (ret)
                        return ret;
+
+               i915_vma_flush_writes(old);
        }
 
        if (fence->vma && fence->vma != vma) {