A Common Operational Problem in DNS Servers

The DNS , is a query / response protocol. Failure to respond to queries causes both immediate operational problems and long term problems with protocol development. Failure to respond to a query is indistinguishable from a packet loss without doing a analysis of query response patterns and results in unnecessary additional queries being made by DNS clients and unnecessary delays being introduced to the resolution process. Due to the inability to distingish between packet loss and nameservers dropping EDNS queries, packet loss is sometimes misclassified as lack of EDNS support which can lead to DNSSEC validation failures. Allowing servers which fail to respond to queries to remain results in developers being afraid to deploy implementations of recent standards. Such servers need to be identified and corrected / replaced. The DNS has response codes that cover almost any conceivable query response. A nameserver should be able to respond to any conceivable query using them. Unless a nameserver is under attack, it should respond to all queries directed to it as a result of following delegations. Additionally code should not assume that there isn't a delegation to the server even if it is not configured to serve the zone. Broken delegation are a common occurrence in the DNS and receiving queries for zones that you are not configured for is not a necessarily a indication that you are under attack. Parent zone operators are supposed to regularly check that the delegating NS records are consistent with those of the delegated zone and to correct them when they are not . If this was being done regularly the instances of broke delegations would be much lower. When a nameserver is under attack it may wish to drop packets. A common attack is to use a nameserver as a amplifier by sending spoofed packets. This is done because response packets are bigger than the queries and big amplification factors are available especially if EDNS is supported. Limiting the rate of responses is reasonable when this is occuring and the client should retry. This however only works if legitimate clients are not being forced to guess whether EDNS queries are accept or not. While there is still a pool of servers that don't repsond to EDNS requests, clients have no way to know if the lack of response is due to packet loss, EDNS packets not being supported or rate limiting due to the server being under attack. Mis-classifications of server characteristics are unavoidable when rate limiting is done.

There are three common query classes that result in non responses today. These are EDNS queries, queries for unknown (unallocated) or unsupported types and filtering of TCP queries.

Identifying servers that fail to respond to EDNS queries can be done by first identifying that the server responds to regular DNS queries then making a series of otherwise identical responses using EDNS, then making the original query again. A series of EDNS queries is needed as at least one DNS implementation responds to the first EDNS query with FORMERR but fails to respond to subsequent queries from the same address for a period until a regular DNS query is made. The EDNS query should specify a UDP buffer size of 512 bytes to avoid false classification of not supporting EDNS due to response packet size. If the server responds to the first and last queries but fails to respond to most or all of the EDNS queries it is probably faulty. The test should be repeated a number of times to eliminate the likelihood of a false positive due to packet loss. Firewalls may also block larger EDNS responses but there is no easy way to check authoritative servers to see if the firewall is misconfigured.

Some servers respond correctly to EDNS version 0 queries but fail to respond to EDNS queries with version numbers that are higher than zero. Servers should respond with BADVERS to EDNS queries with version numbers that they do not support.

Identifying servers that fail to respond to unknown or unsupported types can be done by making an initial DNS query for an A record, making a number of queries for an unallocated type, them making a query for an A record again. IANA maintains a registry of allocated types. If the server responds to the first and last queries but fails to respond to the queries for the unallocated type it is probably faulty. The test should be repeated a number of times to eliminate the likely hood of a false positive due to packet loss.

All DNS servers are supposed to respond to queries over TCP . Firewalls that drop TCP connection attempts rather that resetting the connect attempt or send a ICMP/ICMPv6 administratively prohibited message introduce excessive delays to the resolution process. Whether a server accepts TCP connections can be tested by first checking that it responds to UDP queries to confirm that it is up and operating then attempting the same query over TCP. An additional query should be made over UDP if the TCP connection attempt fails to confirm that the server under test is still operating.

While the first step in remediating this problem is to get the offending nameserver code corrected, there is a very long tail problem with DNS servers in that it can often take over a decade between the code being corrected and a nameserver being upgraded with corrected code. With that in mind it is requested that TLD, and other similar zone operators, take steps to identify and inform their customers, directly or indirectly through registrars, that they are running such servers and that the customers need to correct the problem. TLD operators should construct a list of servers child zones are delegated to along with a delegated zone name. This name shall be the query name used to test the server as it is supposed to exist. For each server the TLD operator shall make an SOA query the delegated zone name. This should result in the SOA record being returned in the answer section. If the SOA record is not return but some other response is returned this is a indication of a bad delegation and the TLD operator should take whatever steps it normally takes to rectify a bad delegation. If more that one zone is delegated to the server it should choose another zone until it finds a zone which responds correctly or it exhausts the list of zones delegated to the server. If the server fails to get a response to a SOA query the TLD operator should make a A query as some nameservers fail to respond to SOA queries but respond to A queries. If it gets no response to the A query another delegated zone should be queried for as some nameservers fail to respond to zones they are not configured for. If subsequent queries find a responding zone all delegation to this server need to be checked and rectified using the TLD's normal procedures. Having identified a working <server, query name> tuple the TLD operator should now check that the server responds to EDNS, Unknown Query Type and TCP tests as described above. If the TLD operator finds that server fails any of the tests, the TLD operator shall take steps to inform the operator of the server that they are running a faulty nameserver and that they need to take steps to correct the matter. The TLD operator shall also record the <server, query name> for followup testing. If repeated attempts to inform and get the customer to correct / replace the faulty server are unsuccessful the TLD operator shall remove all delegations to said server from the zone. It will also be necessary for TLD operators to repeat the scans periodically. It is recommended that this be performed monthly backing off to bi-annually once the numbers of faulty servers found drops off to less than 1 in 100000 servers tested. Follow up tests for faulty servers still need to be performed monthly. Some operators claim that they can't perform checks at registration time. If a check is not performed at registration time it needs to be performed within a week of registration in order to detect faulty servers swiftly. Checking of delegations by TLD operators should be nothing new as they have been required from the very beginings of DNS to do this . Checking for compliance of nameserver operations should just be a extension of such testing. It is recommended that TLD operators setup a test web page which performs the tests the TLD operator performs as part of their regular audits to allow nameserver operators to test that they have correctly fixed their servers. Such tests should be rate limited to avoid these pages being a denial of service vector.

Firewalls and load balancers can affect the externally visible behaviour of a nameserver. Tests for conformance need to be done from outside of any firewall so that the system as a whole is tested. Firewalls and load balancers should not drop DNS packets that they don't understand. They should either pass through the packets or generate a appropriate error response. Requests for unknown query types are not attacks and should not be treated as such.

Choosing the correct response code when fixing a nameserver is important. Just because a type is not implemented does not mean that NOTIMP is the correct response code to return. Response codes need to be choosen considering how clients will handle them. For unimplemented opcodes NOTIMP is the expected response code. In general, for unimplemented type codes Name Error (NXDOMAIN) and NOERROR (no data) are the expected response codes. A server is not supposed to serve a zone which contains unsupported types () so the only thing left is return if the QNAME exists or not. NOTIMP and REFUSED are not useful responses as they force the clients to try all the authoritative servers for a zone looking for a server which will answer the query. Meta queries type may be the exception but these need to be thought about on a case by case basis. If you support EDNS and get a query with a unsupported EDNS version the correct response is BADVERS . If you do not support EDNS at all FORMERR and NOTIMP are the expected error codes. That said a mimimal EDNS server implementation just requires parsing the OPT records and responding with a empty OPT record. There is no need to interpret any EDNS options present in the request as unsupported options are expected to be ignored .