virgl: when reading back wait first, then do the transfer

Otherwise the transfer from the host might not be finished
when we try to read the data.

v2: always wait before readback, but also keep the wait after
  a possible readback, Itherwise TF fails in the CI.

Fixes: e87186fc67
    virgl: add virgl_resource_transfer_prepare

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18243>

src/gallium/drivers/virgl/ci/virpipe-gl-fails.txt

@@ -434,8 +434,6 @@ spec@arb_point_sprite@arb_point_sprite-mipmap,Fail
 spec@arb_program_interface_query@arb_program_interface_query-getprogramresourceindex,Fail
 spec@arb_program_interface_query@arb_program_interface_query-getprogramresourceindex@'vs_input2[1][0]' on GL_PROGRAM_INPUT,Fail
 
-spec@arb_shader_atomic_counters@fragment-discard,Fail
-
 spec@arb_sample_shading@builtin-gl-sample-position 2,Fail
 spec@arb_shader_image_load_store@early-z,Fail
 spec@arb_shader_image_load_store@early-z@occlusion query test/early-z pass,Fail

src/gallium/drivers/virgl/virgl_resource.c

@@ -253,11 +253,6 @@ virgl_resource_transfer_prepare(struct virgl_context *vctx,
          else
             return VIRGL_TRANSFER_MAP_WRITE_TO_STAGING_WITH_READBACK;
       }
-      /* Readback is yet another command and is transparent to the state
-       * trackers.  It should be waited for in all cases, including when
-       * PIPE_MAP_UNSYNCHRONIZED is set.
-       */
-      wait = true;
 
       /* When the transfer queue has pending writes to this transfer's region,
        * we have to flush before readback.
@@ -281,8 +276,16 @@ virgl_resource_transfer_prepare(struct virgl_context *vctx,
       return VIRGL_TRANSFER_MAP_ERROR;
 
    if (readback) {
+      /* Readback is yet another command and is transparent to the state
+       * trackers.  It should be waited for in all cases, including when
+       * PIPE_MAP_UNSYNCHRONIZED is set.
+       */
+      vws->resource_wait(vws, res->hw_res);
       vws->transfer_get(vws, res->hw_res, &xfer->base.box, xfer->base.stride,
                         xfer->l_stride, xfer->offset, xfer->base.level);
+      /* transfer_get puts the resource into a maybe_busy state, so we will have
+       * to wait another time if we want to use that resource. */
+      wait = true;
    }
 
    if (wait)