Comment 7 for bug 1726124

Paul Smith (psmith-gnu) wrote :

I'm not sure I understand what you mean. In a typical configuration you never send "foo" to the nameservice, there's always a search domain and those lookups are always tried first (because the default value for the ndots is 1). This is handled by the libc resolver linked into every program, it's not handled by a central service.

I don't have any idea how systemd-resolve works. But, my understanding of how it used to work with dnsmasq using split tunneling was that the nameservice mapped domains to DNS servers, and the resolver would only forward requests to the DNS server that matched the domain for the host being requested.

If you have a VPN interface that adds "mycorp.com" to the search domain that appears in /etc/resolv.conf search, so it contains "mycorp.com localdomain" for example. Then when someone tries to resolve "myhost", the libc resolver sees that there are no dots here and so it starts appending search paths to the hostname, in order, and sending them to the DNS service to look up. So first it will send "myhost.mycorp.com" to the resolver. The resolver sees that the hostname ends in "mycorp.com" and it knows that the VPN DNS servers "own" that domain, so it forwards that request to those servers for lookup.

If that doesn't match, then the libc resolver will try to look up "myhost.localdomain". That does NOT match the VPN domain, so it will not try to forward that to the DNS servers for the VPN connection and instead use a different resolver.

I don't think there's any information leakage here.