On 2026/11/13 9:29, Marc-André Lureau wrote:
vhost_user_gpu_reset() clears VirtIOGPUBase::enable, which makes the
virtio-vga wrapper fall back to the legacy VGA display. The console is
left in a GL scanout though, so qemu_console_surface() returns NULL and
vga_update_display() dereferences it when the display refreshes: the
host crashes on guest reboot or driver reload with a GL display.
Release the frontend dmabufs, disable the GL scanout and install a
placeholder surface for each output, as virtio_gpu_reset() does for the
non vhost-user device.
This is unsafe independently: GL teardown runs on the vCPU thread. Patch
4 fixes this; please fold them together to preserve bisectability.
Regards,
Akihiko Odaki
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/4846
Reported-by: Dmitry R <[email protected]>
Tested-by: Dmitry R <[email protected]>
Signed-off-by: Marc-André Lureau <[email protected]>
---
hw/display/vhost-user-gpu.c | 14 ++++++++++++++
1 file changed, 14 insertions(+)
diff --git a/hw/display/vhost-user-gpu.c b/hw/display/vhost-user-gpu.c
index c09aca041135..d3b720d08a73 100644
--- a/hw/display/vhost-user-gpu.c
+++ b/hw/display/vhost-user-gpu.c
@@ -638,6 +638,20 @@ static void
vhost_user_gpu_reset(VirtIODevice *vdev)
{
VhostUserGPU *g = VHOST_USER_GPU(vdev);
+ int i;
+
+ /* release the scanouts before virtio_gpu_base_reset() drops ds */
+ for (i = 0; i < g->parent_obj.conf.max_outputs; i++) {
+ struct virtio_gpu_scanout *scanout = &g->parent_obj.scanout[i];
+
+ /* the guest may have left a GL scanout behind */
+ if (g->dmabuf[i]) {
+ qemu_console_gl_release_dmabuf(scanout->con, g->dmabuf[i]);
+ g_clear_pointer(&g->dmabuf[i], qemu_dmabuf_free);
+ }
+ qemu_console_gl_scanout_disable(scanout->con);
+ qemu_console_set_surface(scanout->con, NULL);
+ }
virtio_gpu_base_reset(VIRTIO_GPU_BASE(vdev));