From: Marc-André Lureau <marcandre.lureau@redhat.com>

virtio_gpu_cleanup_mapping() tears down the udmabuf mapping (or, for the
small-iov case, the guest memory mapping) via virtio_gpu_fini_udmabuf(),
but leaves res->blob pointing at it afterwards. Once a blob resource
goes through RESOURCE_DETACH_BACKING, res->blob is left dangling while
res->blob_size is still non-zero.

virtio_gpu_update_cursor_data() only checks res->blob_size before
copying from res->blob, so UPDATE_CURSOR on a detached blob resource
makes QEMU memcpy from freed memory.

virtio_gpu_do_set_scanout() also creates scanout->ds over
res->blob + fb->offset. RESOURCE_DETACH_BACKING leaves that surface
installed, so a subsequent display refresh can still read the backing
after virtio_gpu_fini_udmabuf() has unmapped it.

Disable any active scanouts referencing the resource before tearing down
its backing, clear res->blob afterwards, and check it before
dereferencing in virtio_gpu_update_cursor_data().

Fixes: CVE-2026-66020
Fixes: bdd53f739273 ("virtio-gpu: Update cursor data using blob")
Fixes: 32db3c63ae11 ("virtio-gpu: Add virtio_gpu_set_scanout_blob")
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3881
Reported-by: 章鱼哥@aipy (www.aipyaipy.com)
Reported-by: swing <swing@mail.exp.sh>
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
---
v3:
 - only disable blob-based scanouts on detach_backing
---
 hw/display/virtio-gpu.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/hw/display/virtio-gpu.c b/hw/display/virtio-gpu.c
index 574e68992925..9bacd38b4521 100644
--- a/hw/display/virtio-gpu.c
+++ b/hw/display/virtio-gpu.c
@@ -63,8 +63,16 @@ void virtio_gpu_update_cursor_data(VirtIOGPU *g,
         }
         data = pixman_image_get_data(res->image);
     } else {
+        if (!res->blob) {
+            qemu_log_mask(LOG_GUEST_ERROR, "%s: resource %d has no blob\n",
+                          __func__, resource_id);
+            return;
+        }
         if (res->blob_size < (s->current_cursor->width *
                               s->current_cursor->height * 4)) {
+            qemu_log_mask(LOG_GUEST_ERROR,
+                          "%s: blob size too small for resource %d\n",
+                          __func__, resource_id);
             return;
         }
         data = res->blob;
@@ -974,6 +982,7 @@ void virtio_gpu_cleanup_mapping(VirtIOGPU *g,
 
     if (res->blob) {
         virtio_gpu_fini_udmabuf(g, res);
+        res->blob = NULL;
     }
 }
 
@@ -1039,6 +1048,15 @@ virtio_gpu_resource_detach_backing(VirtIOGPU *g,
     if (!res) {
         return;
     }
+
+    if (res->blob) {
+        for (int i = 0; i < g->parent_obj.conf.max_outputs; i++) {
+            if (res->scanout_bitmask & (1 << i)) {
+                virtio_gpu_disable_scanout(g, i);
+            }
+        }
+    }
+
     virtio_gpu_cleanup_mapping(g, res);
 }
 
-- 
2.55.0

On 2026/07/28 19:58, marcandre.lureau@redhat.com wrote:
> From: Marc-André Lureau <marcandre.lureau@redhat.com>
> 
> virtio_gpu_cleanup_mapping() tears down the udmabuf mapping (or, for the
> small-iov case, the guest memory mapping) via virtio_gpu_fini_udmabuf(),
> but leaves res->blob pointing at it afterwards. Once a blob resource
> goes through RESOURCE_DETACH_BACKING, res->blob is left dangling while
> res->blob_size is still non-zero.
> 
> virtio_gpu_update_cursor_data() only checks res->blob_size before
> copying from res->blob, so UPDATE_CURSOR on a detached blob resource
> makes QEMU memcpy from freed memory.
> 
> virtio_gpu_do_set_scanout() also creates scanout->ds over
> res->blob + fb->offset. RESOURCE_DETACH_BACKING leaves that surface
> installed, so a subsequent display refresh can still read the backing
> after virtio_gpu_fini_udmabuf() has unmapped it.
> 
> Disable any active scanouts referencing the resource before tearing down
> its backing, clear res->blob afterwards, and check it before
> dereferencing in virtio_gpu_update_cursor_data().
> 
> Fixes: CVE-2026-66020
> Fixes: bdd53f739273 ("virtio-gpu: Update cursor data using blob")
> Fixes: 32db3c63ae11 ("virtio-gpu: Add virtio_gpu_set_scanout_blob")
> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3881
> Reported-by: 章鱼哥@aipy (www.aipyaipy.com)
> Reported-by: swing <swing@mail.exp.sh>
> Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
> ---
> v3:
>   - only disable blob-based scanouts on detach_backing
> ---
>   hw/display/virtio-gpu.c | 18 ++++++++++++++++++
>   1 file changed, 18 insertions(+)
> 
> diff --git a/hw/display/virtio-gpu.c b/hw/display/virtio-gpu.c
> index 574e68992925..9bacd38b4521 100644
> --- a/hw/display/virtio-gpu.c
> +++ b/hw/display/virtio-gpu.c
> @@ -63,8 +63,16 @@ void virtio_gpu_update_cursor_data(VirtIOGPU *g,
>           }
>           data = pixman_image_get_data(res->image);
>       } else {
> +        if (!res->blob) {
> +            qemu_log_mask(LOG_GUEST_ERROR, "%s: resource %d has no blob\n",
> +                          __func__, resource_id);
> +            return;
> +        }
>           if (res->blob_size < (s->current_cursor->width *
>                                 s->current_cursor->height * 4)) {
> +            qemu_log_mask(LOG_GUEST_ERROR,
> +                          "%s: blob size too small for resource %d\n",
> +                          __func__, resource_id);
>               return;
>           }
>           data = res->blob;
> @@ -974,6 +982,7 @@ void virtio_gpu_cleanup_mapping(VirtIOGPU *g,
>   
>       if (res->blob) {
>           virtio_gpu_fini_udmabuf(g, res);
> +        res->blob = NULL;
>       }
>   }
>   
> @@ -1039,6 +1048,15 @@ virtio_gpu_resource_detach_backing(VirtIOGPU *g,
>       if (!res) {
>           return;
>       }
> +
> +    if (res->blob) {
> +        for (int i = 0; i < g->parent_obj.conf.max_outputs; i++) {
> +            if (res->scanout_bitmask & (1 << i)) {
> +                virtio_gpu_disable_scanout(g, i);

virtio_gpu_disable_scanout() switches the software surface but does not 
call qemu_console_gl_scanout_disable() and 
qemu_console_gl_release_dmabuf(). Consequently, GTK/SDL can remain in 
dmabuf scanout mode and continue displaying/pinning the imported guest 
backing after QEMU considers it detached. virtio_gpu_fini_udmabuf() only 
closes the fd.

Regards,
Akihiko Odaki

> +            }
> +        }
> +    }
> +
>       virtio_gpu_cleanup_mapping(g, res);
>   }
>