Hello!

As discussed on IRC, my initial advice about MS_PRIVATE was misguided.
The real issue is the “rm_rf (new_root);” call, which removes the root
directory and thus leaves child processes (the daemon) with nothing.

The attached patch adds a test loosely based on yours and a fix for
that.  The fix (for the “userns” engine) is to make NEW_ROOT a tmpfs,
such that upon completion, all we need to do is to unmount it and remove
it; it lives on as the root file system of child processes.

In the “fakechroot” case, we have to leave NEW_ROOT behind, which is not
great but acceptable (it’s user-owned, #o700, and it’s under /tmp).  The
test only checks the “userns” engine.

If you confirm that it works for you and looks reasonable, we can apply
it.

Thanks,
Ludo’.