qemu_namespace: Umount the original /dev before replacing it with tmpfs

Our code relies on mount events propagating into the namespace we
create for a domain. However, there's one caveat. In v8.8.0-rc1~8
I've tried to make us detect differences in mount tables between
the namespace in which libvirtd runs and the domain namespace.
This is crucial for any mounts that happen after the domain was
started (for instance new hugetlbfs can be mounted on say
/dev/hugepages1G).

Therefore, we take a look into /proc/$(pgrep qemu)/mounts to see
what filesystems are mounted under /dev. Now, since we don't
umount the original /dev, just mount a tmpfs over it, we get all
the events (e.g. aforementioned hugetlbfs mount on
/dev/hugepages1G), but we are not really able to access it
because of the tmpfs that's placed on top. This then confuses our
algorithm for detecting which filesystems are mounted (the
algorithm is implemented in qemuDomainGetPreservedMounts()).

To break the link between host's and guest's /dev we just need to
umount() the original /dev in the namespace. Just before our
artificially created tmpfs is moved into its place.

Fixes: 46b03819ae
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2151869#c6
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Martin Kletzander <mkletzan@redhat.com>

src/qemu/qemu_namespace.c

@@ -776,6 +776,11 @@ qemuDomainUnshareNamespace(virQEMUDriverConfig *cfg,
             goto cleanup;
     }
 
+    if (umount("/dev") < 0) {
+        virReportSystemError(errno, "%s", _("failed to umount devfs on /dev"));
+        return -1;
+    }
+
     if (virFileMoveMount(devPath, "/dev") < 0)
         goto cleanup;