[Impact]
All VPN users meaning to use split-tunnelling in a situation where leaking DNS data to the internet about what might be behind their VPN is undesirable.
[Test case]
1) connect to VPN
2) Use dig to request a name that should be on the VPN
3) Verify (kill -USR1 the dnsmasq binary from NM) that the request has only gone through the VPN nameservers (only its request number should have increased by one)
4) Use dig to request a name off-VPN, such as google.com.
5) Verify (kill -USR1 again) that the request has caused the non-VPN nameserver request number to increase, and that the VPN number has not increased.
It is easier to verify this when there is as little traffic as possible on the system, to avoid spurious DNS requests which would make it harder to validate the counters.
[Regression potential]
This change modifies the way in which DNS nameservers and search domains are passed to dnsmasq, as such, if a VPN is configured in a non-standard way and intended to be used to resolve all network requests (which is typically not the case for split-tunnelling) or if the public network is intended to always resolve all requests while the VPN still provides search domains, one might observe incorrect behavior.
Possible failure cases would include failure to resolve names correctly (resulting in NXDOMAIN or REFUSED from dnsmasq) or resolving to the incorrect values (if the wrong nameserver is used).
---
Currently, NM will write all search domains to both any DNS-handling plugins running, and also to resolv.conf / resolvconf; in all cases.
The issue is that doing so means that in the split-DNS case on VPNs, you might get a negative response from all nameservers, then a new request by glibc with the search tacked on, to nameservers again, which might cause DNS requests for "private" resources (say, on the VPN) to be sent to external, untrusted resolvers, or for DNS queries not meant for VPN nameservers to be sent through the VPN anyway.
This is fixable in the case where we have a caching plugin running (such as dnsmasq). dnsmasq will already know about the search domains and use that to limit queries to the right nameservers when a VPN is running. Writing search domains to resolv.conf is unnecessary in this case.
We should still write search domains if no caching gets done, as we then need to expect glibc to send requests as it otherwise would.
This bug was fixed in the package network-manager - 1.2.2-0ubuntu4
---------------
network-manager (1.2.2-0ubuntu4) yakkety; urgency=medium
[ Mathieu Trudel-Lapierre ]
* debian/
refreshed to make gbp pq happy.
* debian/
/run, which is to override whatever might be shipped in /usr/lib; and be
overriden by /etc or command-line arguments. (LP: #1591898)
* debian/
explicitly unmanage anything that is not wifi or wwan: we definitely want
NM to manage wifi and mobile data; and probably don't want it to touch
wired in many cases.
* debian/
make sure we migrate from no global "unmanaged" policy to something
equivalent where we may have a global policy, but explicitly override it
to be disabled; so that on upgrade users do not suddenly see some of their
network devices no longer being handled by NM.
* debian/
not add split DNS search domains to resolv.conf; doing so would risk
leaking names to non-VPN DNS nameservers when attempting to resolve non-
FQDN names. (LP: #1592721)
[ Martin Pitt ]
* debian/
applications such as Chrome don't use NSS but reimplement DNS resolution,
for those we need a local DNS server. Until resolved gets one, we continue
to use the NM specific dnsmasq on the desktop. Correspondingly, revert
libnss-resolve recommends back to dnsmasq-base depends.
-- Mathieu Trudel-Lapierre <email address hidden> Thu, 16 Jun 2016 09:54:02 +0300