For example, as of January 2001, one of the w3.org DNS servers is w3csun1.cis.rl.ac.uk. This server has the power to define the address of www.w3.org. It can flood the other servers to prevent them from providing contradictory information.
When the cache wants the address of w3csun1.cis.rl.ac.uk, it may contact the rl.ac.uk DNS servers, the ac.uk DNS servers, the uk DNS servers, or the root DNS servers. For example, ns.eu.net, one of the ac.uk DNS servers, has the power to define the address of w3csun1.cis.rl.ac.uk. Consequently it also has the power to define the address of www.w3.org.
Similarly, all names under eu.net, hence ac.uk and w3.org, are controlled by sunic.sunet.se; all names under sunet.se, hence eu.net and ac.uk and w3.org, are controlled by beer.pilsnet.sunet.se; and beer.pilsnet.sunet.se is running an ancient version of BIND, known to allow anyone on the Internet to take over the machine.
Are the www.w3.org administrators aware that their DNS service relies on beer.pilsnet.sunet.se and 200 other obscure computers around the world?
In contrast, if w3.org had used in-bailiwick names for its servers, such as a.ns.w3.org and b.ns.w3.org and c.ns.w3.org and d.ns.w3.org, then it would not be relying on the servers for ac.uk and eu.net and sunet.se.
I pointed out this type of problem in January 2000. At that time, these same 200 computers had control over practically all names on the Internet, including *.com. The .com server names were subsequently fixed to avoid the problem. Most country-code TLDs have not been fixed.
www.nasty.dom NS ns1.yahoo.com www.nasty.dom NS ns2.dca.yahoo.com www.nasty.dom NS ns3.europe.yahoo.com www.nasty.dom NS ns5.dcx.yahoo.com ns1.yahoo.com A 220.127.116.11 ns2.dca.yahoo.com A 18.104.22.168 ns3.europe.yahoo.com A 22.214.171.124 ns5.dcx.yahoo.com A 126.96.36.199The nasty.dom server can now wait for (or encourage) the cache to ask about www.nasty.dom. When the cache receives the answer, it will, according to RFC 1034, save the forged yahoo.com addresses for future reference. Subsequent queries for yahoo.com will be misdirected.
Cache poisoning was widely known in 1990. But it was viewed as merely a reliability issue, a result of sloppy administration. Someone who listed munnari.oz.au as a backup server with an out-of-date IP address would accidentally poison caches and destroy legitimate connections to munnari.oz.au.
Vixie's first BIND release, version 4.9 in 1992, featured a notion of ``credibility'' that managed to prevent the most severe cases of accidental poisoning. From a security point of view, Vixie's ``credibility'' is garbage; it doesn't even stop the yahoo.com attack described above.
It's obvious how to eliminate all poisoning. Caches must discard yahoo.com information except from the yahoo.com servers, the com servers, and the root servers. This stops malicious poisoning, so of course it stops accidental poisoning too. End of problem.
BIND finally adopted this poison-elimination rule in 1997, after cache poisoning became a popular attack tool. Did Vixie scrap his obsolete ``credibility'' rules? No! As of January 2000, they were still in BIND 8.2.2-P5, more incoherent than ever. For example, if records had ``additional section credibility,'' and if someone sent a query asking for those records, BIND would reduce the TTL of the records by 5%. Some of the other rules appear in RFC 2181.
I pointed out on bugtraq in January 2000 that, when a domain changed all its DNS server names (e.g., to switch ISPs), an attacker could trivially exploit BIND's ``credibility'' rules to break access to that domain. I also tried to point this out on namedroppers, but my message was censored by Randy Bush.
dnscache doesn't discriminate against additional records. Valid records are accepted whether they're additional records in one packet or answer records in the next; timing doesn't affect the semantics.
In practice, however, some parents limit the number of NS records that they will list; some parents have painful update procedures; and, for many years, the largest .com registrar pointlessly refused NS records listing host names with IP addresses that had already been registered under different host names.
So a child server often lists more NS records than its parent. It includes the NS records along with its answers, so that caches will replace the NS records from the parent with the NS records from the child. If the NS records (and associated addresses) expire after the answers do, the caches will use the complete NS list to find the new answers, and will obtain a fresh NS list at that point. The load is spread among all the servers, though not as evenly as it would be if the parent listed more servers.
Unfortunately, BIND 8.2 won't cache the fresh NS list. After the old list expires, BIND contacts the parent servers and again obtains the incomplete NS list.
Beware that, because of the ``credibility'' rules described above, the NS records from the child servers must include the NS records from the parent. Otherwise an attacker can break BIND's access to the child servers.
So you contact ns-1.disney.corp. But what's the address of ns-1.disney.corp? You have to put the original question on hold while you search for the address of ns-1.disney.corp. You happen to know an address of a .corp DNS server, so you ask it for the address of ns-1.disney.corp. ``I don't know, but I know that .disney.corp has two DNS servers, zone.espn.tv and night.espn.tv,'' it says. ``Try asking them.''
Bottom line: You can't reach espn.tv, and you can't reach disney.corp.
If zone.espn.tv had been a DNS server for .espn.tv, the .tv server would have provided glue for zone.espn.tv, i.e., the IP address of zone.espn.tv. So you would have been able to contact zone.espn.tv. RFC 1034 specifically requires glue for referrals to in-bailiwick DNS servers. (Some people use the word ``glue'' only in this case.)
For referrals to out-of-bailiwick DNS servers, however, RFC 1034 says that glue is unnecessary. RFC 1537 says the same thing. RFC 1912 says the same thing. The comp.protocols.tcp-ip.domains FAQ says that ``you do not need a glue record, and, in fact, adding one is a very bad idea.'' (This is an obsolete reference to accidental poisoning; see above.) Some DNS server implementations ignore out-of-bailiwick glue by default. So the glueless domains espn.tv and disney.corp are following the rules---yet neither of them is reachable.
There can be trouble even when there are no loops. Suppose a BIND cache is looking up www.espn.tv in the following situation:
espn.tv NS ns-1.disney.corp espn.tv NS ns-2.disney.corp disney.corp NS ns-1.disney.corp disney.corp NS ns-2.disney.corpWhen BIND sees the glueless delegation to ns-1.disney.corp, it drops the www.espn.tv query and begins a ``sysquery'' for ns-1.disney.corp, hoping to have the ns-1.disney.corp address cached by the time the www.espn.tv query is retried. (The BIND developers refer to this bug as ``no query restart.'') Clients generally don't retry more than four times, so an initial query for a domain with four levels of gluelessness will fail; an initial query for a domain with three levels of gluelessness will be very likely to fail, and very slow if it succeeds.
``As far as I know, the Internet has not yet lost any domains to gluelessness,'' I wrote in 2000. ``But there are an increasing number of glueless domains, and I've spotted a glueless domain with glueless DNS servers. How much gluelessness must a cache tolerate? Currently dnscache allows three levels of gluelessness. This seems to be enough for now, but will it be enough in the future?''
I subsequently learned about www.monty.de, which had so many levels of gluelessness that BIND caches were completely unable to reach it:
monty.de NS ns.norplex.net monty.de NS ns2.norplex.net norplex.net NS vserver.neptun11.de norplex.net NS ns1.mars11.de neptun11.de NS ns.germany.net neptun11.de NS ns2.germany.net mars11.de NS ns1.neptun11.de mars11.de NS www.gilching.de gilching.de NS ecrc.de gilching.de NS name.muenchen.roses.dednscache was able to find the address of www.monty.de, but it needed fourteen queries to various servers.
I recommend that all DNS servers be in-bailiwick servers with glue. External DNS servers should be given internal names, with address records copied automatically (preferably by some secure mechanism) from the external names to the internal names.
DNS should have been designed with addresses, not names, in NS records and MX records. The ``additional section'' of DNS responses should have been eliminated. RFC 1035 observes correctly that NS indirection and MX indirection ``insure [sic] consistency'' of addresses; however, this indirection should have been handled by the server, not the client. (On a related note: Microsoft Exchange Server 2000 reportedly fails to deliver to MX records that point to CNAME records; fixed in the first Service Pack.)
I have a separate page discussing A6 and DNAME from this perspective.
aol.com NS dns-01.ns.aol.com aol.com NS dns-02.ns.aol.comUsually this means that the A records accompanied the NS records but with lower TTLs, and the cache didn't contact the servers soon enough to refresh the A records as described above. (If the cache is BIND 8.2, then the A records won't be refreshed anyway, and an attacker can force the TTLs down even if they originally matched.)
In this situation, the RFC 1034 resolution algorithm fails. According to RFC 1034, if the cache wants the address of yb.mx.aol.com, it looks for the ``best servers'' among ``locally-available name server RRs,'' obtaining the names dns-01.ns.aol.com and dns-02.ns.aol.com; it then starts ``parallel resolver processes looking for the addresses'' of dns-01.ns.aol.com and dns-02.ns.aol.com; those resolver processes look for the ``best servers,'' and so on. The cache loops until it runs out of patience and gives up.
Fortunately, real caches use a different algorithm. dnscache starts from the roots, ignoring cached NS records, when it reaches gluelessness level 2. BIND reportedly starts all its glue requests from the roots.
RFC 1034 says that an alias ``should'' not point to another alias. In reality, however, if an administrator decides to set up www.espn.go.com as an alias for espn.go.com, he probably won't remember to change www.espn.tv---but users will kick and scream if www.espn.tv breaks. ``CNAME chains should be followed,'' RFC 1034 says.
Aliases, like gluelessness, force DNS clients to chew up time and memory. How many layers of aliases must a cache tolerate? Currently dnscache allows four levels of aliases. This seems to be enough for now, but will it be enough in the future?
I recommend that all CNAME records be eliminated. DNS should have been designed without aliases.
In practice, however, the ISP might instead use CNAME records. It makes 188.8.131.52.in-addr.arpa an alias for 100.cust184.108.40.206.in-addr.arpa, and similarly for 101 and 102; and then it delegates cust220.127.116.11.in-addr.arpa to the customer's DNS server. This is a valid configuration, although RFC 2317 says that some old versions of BIND can't handle it.
Why would an ISP want to add this extra layer of complication? Answer: With the simple approach, if the customer is running BIND, he'll have to put the 100 and 101 and 102 records in three separate files. With the complicated approach, the customer can put the records into a single file.
I recommend that, in this situation, the CNAME records be eliminated, and the customer upgrade to a better DNS server.
dnscache simply contacts a random server, to balance the load as effectively as possible. BIND keeps track of the round-trip times for its queries to each server, with various bonuses and penalties, and then sends all its queries to the ``best'' server.
Simulations show that the increasingly frequent .com server overloads (as of March 2000) could be caused by BIND's transmission strategy.
This procedure requires an incredible amount of bug-prone parsing for a very small amount of information. The underlying problem is that DNS was designed to declare information in a human-oriented format, rather than to support crucial operations in the simplest possible way.
Warning about NXDOMAIN: It is clear from RFC 1034 and RFC 1035 that an NXDOMAIN guarantees the nonexistence of every subdomain of the query domain. For example, if a cache sees an NXDOMAIN for ns.heaven.af.mil, it can conclude that a.ns.heaven.af.mil and b.ns.heaven.af.mil don't exist. If a server has records for a.ns.heaven.af.mil and b.ns.heaven.af.mil, but no records for ns.heaven.af.mil. it sends a zero-records (#5) response, not an NXDOMAIN. However, RFC 2308 allows NXDOMAIN even when the domain exists, to indicate that there are no records of any type under the query name. So it is essential for interoperability that caches not draw the above conclusion.
DNS clients and DNS caches begin by transmitting a query (which always fits into 512 bytes) through UDP. The response is sent back through UDP. If the response does not fit into a UDP packet, it is truncated, and the TC bit is set at the beginning of the UDP packet. Clients and caches that support TCP see the TC bit and retry their query through TCP.
RFC 1035 does not make clear exactly what ``truncated'' means. The obvious interpretation is to end the packet at exactly 512 bytes. However, this causes interoperability problems: in particular, the Squid cache dies if a packet is truncated between records. BIND ends the packet before the first record that went past 512 bytes. dnscache ends the packet before all records.
One problem with DNS compression is the amount of code required to parse it. Reliably locating all these names takes quite a bit of work that would otherwise have been unnecessary for a DNS cache. LZ77 compression would have been much easier to implement.
Another problem with DNS compression is the amount of code required to correctly generate it. (RFC 1035 allowed servers to not bother compressing their responses; however, caches have to implement compression, so that address lists from some well-known sites don't burst the seams of a DNS UDP packet.) Not only does the compressor need to figure out which names can be compressed, but it also needs to keep track of compression targets earlier in the packet. RFC 1035 doesn't make clear exactly what targets are allowed. (Most versions of BIND do not use pointers except to compressible names; suffixes of the query name are excluded. dnscache uses pointers to suffixes of the query name.)
Another problem with DNS compression is that it's not particularly effective. LZ77 would have done a noticeably better job on current data, and a much better job on new record types that might become popular in the future. (BIND versions 4.9.* through 8.1.2 compress names in new record types, such as RP and SRV, in blatant violation of RFC 1035. The names are not decompressed by caches that do not know about the new types. This is an interoperability disaster.)
Experienced programmers stored hostnames in lowercase, and converted uppercase to lowercase as part of the user interface. Hostname comparisons were simple binary comparisons.
DNS, however, was not designed by experienced programmers. DNS clients send hostnames exactly as typed by the user, without converting uppercase to lowercase. DNS servers send some hostnames as typed by the system administrator, without converting uppercase to lowercase. All implementors are forced to waste time worrying about case.
The DNS protocol allows arbitrary bytes in hostnames. This flexibility would have been convenient for several applications, notably in-addr.arpa, if the designers hadn't screwed up their case handling. As is, binary names in DNS are practically useless.
Unfortunately, in DNS packets, the list of mail exchangers is divided into separate MX records. The MX records can even be (and, in responses to * queries, often are) interleaved with other records. A cache has to sort the list of records, preferably using a method that isn't painfully slow for large packets, and partition the result into complete record sets.
Queries ask for records in a particular class. RFC 1034 allows queries to ask for records in all classes, but this makes no sense: if multiple classes were actually used then they would almost never be on the same server. The client knows what class it's looking for, so it can specify a class; RFC 1123 section 18.104.22.168 requires this for the Internet class and recommends it in all cases.
RFC 1034 says that classes ``allow parallel use of different formats for data of type address.'' This doesn't make sense. If DNS is used in a network with multiple address formats, then one DNS server will want to provide addresses in more than one format; but that DNS server is only in charge of one class. Address format extensibility should have been provided in the address data itself.
dnscache discards queries for non-Internet classes.
The Ultrix version of BIND sends queries with AD+CD set.
A client will receive a bogus response from a BIND cache if the client asks about X, the cache already knows X CNAME Y, and the cache has to ask a server about Y. The cache will forward the server's Y response, with Y as the query, to the client. This bug was fixed in BIND 8.2.3.
There is at least one server that incorrectly produces NXDOMAIN for all non-A queries, even for domains that exist:
% date Sat Nov 2 15:45:22 CST 2002 % dnsq any www.css.vtext.com njbdcss.vtext.com 255 www.css.vtext.com: 35 bytes, 1+0+0+0 records, response, nxdomain query: 255 www.css.vtext.com % dnsq aaaa www.css.vtext.com njbdcss.vtext.com 28 www.css.vtext.com: 35 bytes, 1+0+0+0 records, response, nxdomain query: 28 www.css.vtext.com % dnsq a www.css.vtext.com njbdcss.vtext.com 1 www.css.vtext.com: 51 bytes, 1+1+0+0 records, response, authoritative, noerror query: 1 www.css.vtext.com answer: www.css.vtext.com 0 A 22.214.171.124 %If a client looks up AAAA before A, the NXDOMAIN will be cached, so the A query will fail.