DNS over DTLS (DNSoD)Cisco Systems, Inc.Cessna Business Park, Varthur HobliSarjapur Marathalli Outer Ring RoadBangaloreKarnataka560103Indiatireddy@cisco.comCisco Systems, Inc.170 West Tasman DriveSan JoseCalifornia95134USAdwing@cisco.comCisco Systems, Inc.BangaloreIndiapraspati@cisco.comDPRIVEDNS queries and responses are visible to network elements on the path
between the DNS client and its server. These queries and responses can
contain privacy-sensitive information which is valuable to protect. An
active attacker can send bogus responses causing misdirection of the
subsequent connection.To counter passive listening and active attacks, this document
proposes the use of Datagram Transport Layer Security (DTLS) for DNS, to
protect against passive listeners and certain active attacks. As DNS
needs to remain fast, this proposal also discusses mechanisms to reduce
DTLS round trips and reduce DTLS handshake size. The proposed mechanism
runs over the default DNS port and can also run over an alternate
port.The Domain Name System is specified in
and . DNS queries and responses are
normally exchanged unencrypted and are thus vulnerable to eavesdropping.
Such eavesdropping can result in an undesired entity learning domains
that a host wishes to access, thus resulting in privacy leakage. DNS
privacy problem is further discussed in .Active attackers have long been successful at injecting bogus
responses, causing cache poisoning and causing misdirection of the
subsequent connection (if attacking A or AAAA records). A popular
mitigation against that attack is to use ephemeral and random source
ports for DNS queries.This document defines DNS over DTLS (DNSoD, pronounced "dee-enn-sod")
which provides confidential DNS communication for stub resolvers,
recursive resolvers, iterative resolvers and authoritative servers.The motivations for proposing DNSoD are thatTCP suffers from network head-of-line blocking, where the loss of
a packet causes all other TCP segments to not be delivered to the
application until the lost packet is re-transmitted. DNSoD, because
it uses UDP, does not suffer from network head-of-line blocking.DTLS session resumption consumes 1 round trip whereas TLS session
resumption can start only after TCP handshake is complete. Although
TCP Fast Open can reduce that
handshake, TCP Fast Open is not yet available in
commercially-popular operating systems.DNS queries can be sent over UDP or TCP. The scope of this document,
however, is only UDP. DNS over TCP could be protected with TLS, as
described in .
Alternatively, a shim protocol could be defined between DTLS and DNS,
allowing large responses to be sent over DTLS itself, see .DNS Security Extensions (DNSSEC)
provides object integrity of DNS resource records, allowing end-users
(or their resolver) to verify legitimacy of responses. However, DNSSEC
does not protect privacy of DNS requests or responses. DNSoD works in
conjunction with DNSSEC, but DNSoD does not replace the need or value of
DNSSEC.This section describes problems common to any DNS privacy solution.
To achieve DNS privacy an encrypted and integrity-protected channel is
needed between the client and server. This channel can be blocked, and
the client needs to react to such blockages.When sending DNS over an encrypted channel, there are two choices:
send the encrypted traffic over the DNS ports (UDP 53, TCP 53) or send
the encrypted traffic over a different port. The encrypted traffic is
not normal DNS traffic, but rather is a cryptographic handshake
followed by encrypted payloads. There can be firewalls, other security
devices, or intercepting DNS proxies which block the non-DNS traffic
or otherwise react negatively (e.g., quarantining the host for
suspicious behavior). Alternatively, if a different port is used for
the encrypted traffic, a firewall or other security device might block
that port or otherwise react negatively.There is no panacea, and only experiments on the Internet will
uncover which technique or combination of techniques will work best.
The authors believe a combination of techniques will be necessary, as
that has proven necessary with other protocols that desire to work on
existing networks.DNS privacy requires encrypting the query (and response) from
passive attacks. Such encryption typically provides integrity
protection as a side-effect, which means on-path attackers cannot
simply inject bogus DNS responses. However, to provide stronger
protection from active attackers pretending to be the server, the
server itself needs to be authenticated.To authenticate the server providing DNS privacy, the DNS client
needs to be configured with the names of those DNS privacy servers.
When connecting a DNS privacy server, the server's IP address can be
converted to its hostname by doing a DNS PTR lookup, verifying that
the name matches the pre-configured list of DNS privacy servers, and
finally validating its certificate trust chain or a local list of
certificates. For DNS privacy servers that don't have a certificate
trust chain (e.g.,, because they are on a home network or a corporate
network), the configured list of DNS privacy servers can contain the
certificate fingerprint of the DNS privacy server (i.e., a simple
whitelist of name and certificate fingerprint).Using DNS privacy with an authenticated server is most preferred,
DNS privacy with an unauthenticated server is next preferred, and
plain DNS is least preferred. An implementation will attempt to obtain
DNS privacy by contacting DNS servers on the local network (provided
by DHCP) and on the Internet, and will make those attempts in parallel
to reduce user impact. If DNS privacy cannot be successfully
negotiated for whatever reason, client can do three things: refuse to send DNS queries on this network, which means the
client can not make effective use of this network, as modern
networks require DNS; or,use DNS privacy with an un-authorized server, which means an
attacker could be spoofing the handshake with the DNS privacy
server; or,send plain DNS queries on this network, which means no DNS
privacy is provided.Heuristics can improve this situation, but only to a degree (e.g.,
previous success of DNS privacy on this network may be reason to alert
the user about failure to establish DNS privacy on this network now).
Still, the client (in cooperation with the end user) has to decide to
use the network without the protection of DNS privacy.The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in .DNSoD can be deployed incrementally by the Internet Service Provider
or as an Internet service.If the ISP's DNS resolver supports DNSoD, then DNS queries are
protected from passive listening and from many active attacks along that
path.DNSoD can be offered as an Internet service, and a stub resolver or
DNS resolver can be configured to point to that DNSoD server (rather
than to the ISP-provided DNS server).[Note - This section requires further discussion]Many modern operating systems already detect if a web proxy is
interfering with Internet communications, using proprietary mechanisms
that are out of scope of this document. After that mechanism has run
(and detected Internet connectivity is working), the DNSoD procedure
described in this document should commence. This timing avoids delays in
joining the network (and displaying an icon indicating successful
Internet connection), at the risk that those initial DNS queries will be
sent without protection afforded by DNSoD.DNSoD can run over standard UDP port 53 as defined in . A DNS client or server that does not implement
this specification will not respond to the incoming DTLS packets because
they don't parse as DNS packets (the DNS Opcode would be 15, which is
undefined). A DNS client or server that does implement this
specification can demultiplex DNS and DTLS packets by examining the
third octet. For TLS 1.2, which is what is defined by this
specification, a DTLS packet will contain 253 in the third octet,
whereas a DNS packet will never contain 253 in the third octet.There has been some concern with sending DNSoD traffic over the same
port as normal, un-encrypted DNS traffic. The intent of this section is
to show that DNSoD could successfully be sent over port 53. Further
analysis and testing on the Internet may be valuable to determine if
multiplexing on port 53, using a separate port, or some fallback between
a separate port and port 53 brings the most success.After performing the above steps, the host should determine if the
DNS server supports DNSoD by sending a DTLS ClientHello message. A DNS
server that does not support DNSoD will not respond to ClientHello
messages sent by the client, because they are not valid DNS requests
(specifically, the DNS Opcode is invalid). The client MUST use timer
values defined in Section 4.2.4.1 of for
retransmission of ClientHello message and if no response is received
from the DNS server. After 15 seconds, it MUST cease attempts to
re-transmit its ClientHello. Thereafter, the client MAY repeat that
procedure in the event the DNS server has been upgraded to support
DNSoD, but such probing SHOULD NOT be done more frequently than every 24
hours and MUST NOT be done more frequently than every 15 minutes. This
mechanism requires no additional signaling between the client and
server.To reduce number of octets of the DTLS handshake, especially the size
of the certificate in the ServerHello (which can be several kilobytes),
we should consider using plain public keys . Considering that to authorize
a certain DNS server the client already needs explicit configuration of
the DNS servers it trusts, maybe the public key configuration problem is
really no worse than the configuration problem of those whitelisted
certificates?Multiple DNS queries can be sent over a single DNSoD security
association. The existing QueryID allows multiple requests and responses
to be interleaved in whatever order they can be fulfilled by the DNS
server. This means DNSoD reduces the consumption of UDP port numbers,
and because DTLS protects the communication between the DNS client and
its server, the resolver SHOULD NOT use random ephemeral source ports
(Section 9.2 of ) because such source port
use would incur additional, unnecessary DTLS load on the DNSoD
server.It is highly advantageous to avoid server-side DTLS state and reduce
the number of new DTLS security associations on the server which can be
done with . This also eliminates a
round-trip for subsequent DNSoD queries, because with the DTLS security association does not need to
be re-established. Note: with the shim (described below) perhaps we
could send the query and the restore server-side state in the
ClientHello packet.Compared to normal DNS, DTLS adds at least 13 octets of header, plus
cipher and authentication overhead to every query and every response.
This reduces the size of the DNS payload that can be carried. Certain
DNS responses are large (e.g., many AAAA records, TXT, SRV) and don't
fit into a single UDP packet, causing a partial response with the
truncation (TC) bit set. The client is then expected to repeat the query
over TCP, which causes additional name resolution delay. We have
considered two ideas, one that reduces the need to switch to TCP and
another that eliminates the need to switch to TCP: Path MTU can be determined using Packetization Layer Path MTU Discovery using
DTLS heartbeat. does not rely on ICMP
or ICMPv6, and would not affect DNS state or responsiveness on the
client or server. However, it would be additional chattiness.To avoid IP fragmentation, DTLS handshake messages incorporate
their own fragment offset and fragment length. We might utilize a
similar mechanism in a shim layer between DTLS and DNS, so that
large DNS messages could be carried without causing IP
fragmentation.DNSoD puts an additional computational load on servers. The largest
gain for privacy is to protect the communication between the DNS client
(the end user's machine) and its caching resolver. Because of the load
imposed, and because of the infrequency of queries to root servers means
the DTLS overhead is unlikely to be amoritized over the DNS queries sent
over that DTLS connection, implementing DNSoD on root servers is NOT
RECOMMENDED.In DTLS, all data is protected using the same record encoding and
mechanisms. When the mechanism described in this document is in effect,
DNS messages are encrypted using the standard DTLS record encoding. When
a user of DTLS wishes to send an DNS message, it delivers it to the DTLS
implementation as an ordinary application data write (e.g.,
SSL_write()). A single DTLS session can be used to receive multiple DNS
requests and generate DNS multiple responses.To improve interoperability, the set of DTLS features and cipher
suites is restricted. The DTLS implementation MUST disable compression.
DTLS compression can lead to the exposure of information that would not
otherwise be revealed . Generic
compression is unnecessary since DNS provides compression features
itself. DNS over DTLS MUST only be used with cipher suites that have
ephemeral key exchange, such as the ephemeral Diffie-Hellman (DHE) or the elliptic curve variant (ECDHE) . Ephemeral key exchange MUST have a minimum
size of 2048 bits for DHE or security level of 128 bits for ECDHE.
Authenticated Encryption with Additional Data (AEAD) modes, such as the
Galois Counter Model (GCM) mode for AES
are acceptable.DNS servers are often configured with anycast addresses. While the
network is stable, packets transmitted from a particular source to an
anycast address will reach the same server that has the cryptographic
context from the DNS over DTLS handshake. But when the network
configuration changes,a DNS over DTLS packet can be received by a server
that does not have the necessary cryptographic context. To encourage the
client to initiate a new DTLS handshake, DNS servers SHOULD generate a
DTLS Alert message in response to receiving a DTLS packet for which the
server does not have any cryptographic context.If demultiplexing DTLS and DNS (using the third octet, ) is useful, we should reserve DNS Opcode 15 to
ensure DNS always has a 0 bit where DTLS always has a 1 bit.Once a DNSoD client has established a security association with a
particular DNS server, and outstanding normal DNS queries with that
server (if any) have been received, the DNSoD client MUST ignore any
subsequent normal DNS responses from that server, as all subsequent
responses should be inside DNSoD. This behavior mitigates all (?)
attacks described in Measures for Making DNS More
Resilient against Forged Answers.Security considerations discussed in DTLS also apply to this document.Thanks to Phil Hedrick for his review comments on TCP and to Josh
Littlefield for pointing out DNSoD load on busy servers (most notably
root servers).