GET_VRING_BASE stops the queue, clearing the call and kick fds. However,
we don't clear vring.avail. That means that if vu_queue_notify() is called
it won't realise the queue isn't ready and will die with an EBADFD.
We get this during migration, because for some reason, qemu reconfigures
the vhost-user device when a migration is triggered. There's a window
between the GET_VRING_BASE and re-establishing the call fd where the
notify function can be called, causing a crash.
Signed-off-by: David Gibson <david(a)gibson.dropbear.id.au>
---
vhost_user.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/vhost_user.c b/vhost_user.c
index 7ab13774..be1aa942 100644
--- a/vhost_user.c
+++ b/vhost_user.c
@@ -732,6 +732,7 @@ static bool vu_get_vring_base_exec(struct vu_dev *vdev,
msg->hdr.size = sizeof(msg->payload.state);
vdev->vq[idx].started = false;
+ vdev->vq[idx].vring.avail = 0;
if (vdev->vq[idx].call_fd != -1) {
close(vdev->vq[idx].call_fd);
--
2.48.1