diff --git a/src/gallium/frontends/va/image.c b/src/gallium/frontends/va/image.c
index 39768a030b7..22713776d10 100644
--- a/src/gallium/frontends/va/image.c
+++ b/src/gallium/frontends/va/image.c
@@ -548,6 +548,7 @@ vlVaGetImage(VADriverContextP ctx, VASurfaceID surface, int x, int y,
       ret = vlVaPostProcCompositor(drv, &src_rect, &dst_rect,
                                    surf->buffer, tmp_surf.buffer,
                                    VL_COMPOSITOR_NONE, &proc);
+      drv->pipe->flush(drv->pipe, NULL, 0);
       if (ret != VA_STATUS_SUCCESS) {
          tmp_surf.buffer->destroy(tmp_surf.buffer);
          mtx_unlock(&drv->mutex);
@@ -715,14 +716,14 @@ vlVaPutImage(VADriverContextP ctx, VASurfaceID surface, VAImageID image,
       ret = vlVaPostProcCompositor(drv, &src_rect, &dst_rect,
                                    tmp_surf.buffer, surf->buffer,
                                    VL_COMPOSITOR_NONE, &proc);
+      vlVaSurfaceFlush(drv, surf);
       tmp_surf.buffer->destroy(tmp_surf.buffer);
       mtx_unlock(&drv->mutex);
       return ret;
    }
 
    vlVaUploadImage(drv, surf, img_buf, vaimage);
-
-   drv->pipe->flush(drv->pipe, NULL, 0);
+   vlVaSurfaceFlush(drv, surf);
    mtx_unlock(&drv->mutex);
 
    return VA_STATUS_SUCCESS;
diff --git a/src/gallium/frontends/va/postproc.c b/src/gallium/frontends/va/postproc.c
index b9802213fbb..9ff64fd4476 100644
--- a/src/gallium/frontends/va/postproc.c
+++ b/src/gallium/frontends/va/postproc.c
@@ -232,7 +232,6 @@ vlVaPostProcCompositor(vlVaDriver *drv,
 
    drv->cstate.chroma_location = VL_COMPOSITOR_LOCATION_NONE;
 
-   drv->pipe->flush(drv->pipe, NULL, 0);
    return VA_STATUS_SUCCESS;
 }
 
@@ -572,6 +571,8 @@ vlVaHandleVAProcPipelineParameterBufferType(vlVaDriver *drv, vlVaContext *contex
          return VA_STATUS_SUCCESS;
    }
 
-   return vlVaPostProcCompositor(drv, src_region, dst_region,
-                                 src, context->target, deinterlace, param);
+   VAStatus ret = vlVaPostProcCompositor(drv, src_region, dst_region,
+                                         src, context->target, deinterlace, param);
+   vlVaSurfaceFlush(drv, dst_surface);
+   return ret;
 }
diff --git a/src/gallium/frontends/va/surface.c b/src/gallium/frontends/va/surface.c
index 006421537f5..f93e4f0660c 100644
--- a/src/gallium/frontends/va/surface.c
+++ b/src/gallium/frontends/va/surface.c
@@ -119,6 +119,8 @@ vlVaDestroySurfaces(VADriverContextP ctx, VASurfaceID *surface_list, int num_sur
       }
       if (surf->buffer)
          surf->buffer->destroy(surf->buffer);
+      if (surf->pipe_fence)
+         drv->pipe->screen->fence_reference(drv->pipe->screen, &surf->pipe_fence, NULL);
       if (surf->ctx) {
          assert(_mesa_set_search(surf->ctx->surfaces, surf));
          _mesa_set_remove_key(surf->ctx->surfaces, surf);
@@ -176,13 +178,22 @@ _vlVaSyncSurface(VADriverContextP ctx, VASurfaceID render_target, uint64_t timeo
       fence = surf->fence;
    }
 
+   if (surf->pipe_fence) {
+      struct pipe_screen *pscreen = drv->pipe->screen;
+      if (!pscreen->fence_finish(pscreen, NULL, surf->pipe_fence, timeout_ns)) {
+         mtx_unlock(&drv->mutex);
+         return VA_STATUS_ERROR_TIMEDOUT;
+      }
+      pscreen->fence_reference(pscreen, &surf->pipe_fence, NULL);
+   }
+
    /* This is checked before getting the context below as
     * surf->ctx is only set in begin_frame
     * and not when the surface is created
     * Some apps try to sync/map the surface right after creation and
     * would get VA_STATUS_ERROR_INVALID_CONTEXT
     */
-   if (!surf->buffer || !fence) {
+   if (!surf->fence) {
       // No outstanding encode/decode operation: nothing to do.
       mtx_unlock(&drv->mutex);
       return VA_STATUS_SUCCESS;
@@ -441,7 +452,7 @@ vlVaPutSurface(VADriverContextP ctx, VASurfaceID surface_id, void* draw, short s
    /* flush before calling flush_frontbuffer so that rendering is flushed
     * to back buffer so the texture can be copied in flush_frontbuffer
     */
-   drv->pipe->flush(drv->pipe, NULL, 0);
+   vlVaSurfaceFlush(drv, surf);
 
    screen->flush_frontbuffer(screen, drv->pipe, tex, 0, 0,
                              vscreen->get_private(vscreen), 0, NULL);
@@ -996,7 +1007,7 @@ vlVaHandleSurfaceAllocate(vlVaDriver *drv, vlVaSurface *surface,
                   surfaces[i]->width, surfaces[i]->height,
                   false);
       }
-      drv->pipe->flush(drv->pipe, NULL, 0);
+      vlVaSurfaceFlush(drv, surface);
    }
 
    return VA_STATUS_SUCCESS;
@@ -1013,6 +1024,13 @@ vlVaGetSurfaceBuffer(vlVaDriver *drv, vlVaSurface *surface)
    return surface->buffer;
 }
 
+void
+vlVaSurfaceFlush(vlVaDriver *drv, vlVaSurface *surf)
+{
+   drv->pipe->flush(drv->pipe, &surf->pipe_fence,
+                    drv->has_external_handles ? 0 : PIPE_FLUSH_ASYNC);
+}
+
 static int
 rt_format_to_fourcc(uint32_t format)
 {
diff --git a/src/gallium/frontends/va/va_private.h b/src/gallium/frontends/va/va_private.h
index 6eb09b1248e..c163318dc2d 100644
--- a/src/gallium/frontends/va/va_private.h
+++ b/src/gallium/frontends/va/va_private.h
@@ -448,7 +448,8 @@ typedef struct vlVaSurface {
    vlVaContext *ctx;
    vlVaBuffer *coded_buf;
    bool full_range;
-   struct pipe_fence_handle *fence;
+   struct pipe_fence_handle *fence; /* pipe_video_codec fence */
+   struct pipe_fence_handle *pipe_fence; /* pipe_context fence */
    struct vlVaSurface *efc_surface; /* input surface for EFC */
    bool is_dpb;
 } vlVaSurface;
@@ -569,6 +570,7 @@ VAStatus vlVaHandleVAProcPipelineParameterBufferType(vlVaDriver *drv, vlVaContex
 VAStatus vlVaHandleSurfaceAllocate(vlVaDriver *drv, vlVaSurface *surface, struct pipe_video_buffer *templat,
                                    const uint64_t *modifiers, unsigned int modifiers_count);
 struct pipe_video_buffer *vlVaGetSurfaceBuffer(vlVaDriver *drv, vlVaSurface *surface);
+void vlVaSurfaceFlush(vlVaDriver *drv, vlVaSurface *surf);
 void vlVaAddRawHeader(struct util_dynarray *headers, uint8_t type, uint32_t size, uint8_t *buf,
                       bool is_slice, uint32_t emulation_bytes_start);
 void vlVaGetBufferFeedback(vlVaBuffer *buf);