Copyright (C) Internet Systems Consortium, Inc. ("ISC") See COPYRIGHT in the source root or http://isc.org/copyright.html for terms. BIND 8 to BIND 9 Migration Notes BIND 9 is designed to be mostly upwards compatible with BIND 8, but there is still a number of caveats you should be aware of when upgrading an existing BIND 8 installation to use BIND 9. 1. Configuration File Compatibility 1.1. Unimplemented Options and Changed Defaults BIND 9 supports most, but not all of the named.conf options of BIND 8. For a complete list of implemented options, see doc/misc/options. If your named.conf file uses an unimplemented option, named will log a warning message. A message is also logged about each option whose default has changed unless the option is set explicitly in named.conf. The default of the "transfer-format" option has changed from "one-answer" to "many-answers". If you have slave servers that do not understand the many-answers zone transfer format (e.g., BIND 4.9.5 or older) you need to explicitly specify "transfer-format one-answer;" in either the options block or a server statement. BIND 9.4 onwards implements "allow-query-cache". The "allow-query" option is no longer used to specify access to the cache. The "allow-query" option continues to specify which hosts are allowed to ask ordinary DNS questions. The new "allow-query-cache" option is used to specify which hosts are allowed to get answers from the cache. Since BIND 9.4.1, if "allow-query-cache" is not set then "allow-recursion" is used if it is set, otherwise "allow-query" is used if it is set, otherwise the default localnets and localhost is used. 1.2. Handling of Configuration File Errors In BIND 9, named refuses to start if it detects an error in named.conf. Earlier versions would start despite errors, causing the server to run with a partial configuration. Errors detected during subsequent reloads do not cause the server to exit. Errors in master files do not cause the server to exit, but they do cause the zone not to load. 1.3. Logging The set of logging categories in BIND 9 is different from that in BIND 8. If you have customised your logging on a per-category basis, you need to modify your logging statement to use the new categories. Another difference is that the "logging" statement only takes effect after the entire named.conf file has been read. This means that when the server starts up, any messages about errors in the configuration file are always logged to the default destination (syslog) when the server first starts up, regardless of the contents of the "logging" statement. In BIND 8, the new logging configuration took effect immediately after the "logging" statement was read. 1.4. Notify messages and Refresh queries The source address and port for these is now controlled by "notify-source" and "transfer-source", respectively, rather that query-source as in BIND 8. 1.5. Multiple Classes. Multiple classes have to be put into explicit views for each class. 2. Zone File Compatibility 2.1. Strict RFC1035 Interpretation of TTLs in Zone Files BIND 9 strictly complies with the RFC1035 and RFC2308 rules regarding omitted TTLs in zone files. Omitted TTLs are replaced by the value specified with the $TTL directive, or by the previous explicit TTL if there is no $TTL directive. If there is no $TTL directive and the first RR in the file does not have an explicit TTL field, the zone file is illegal according to RFC1035 since the TTL of the first RR is undefined. Unfortunately, BIND 4 and many versions of BIND 8 accept such files without warning and use the value of the SOA MINTTL field as a default for missing TTL values. BIND 9.0 and 9.1 completely refused to load such files. BIND 9.2 emulates the nonstandard BIND 4/8 SOA MINTTL behaviour and loads the files anyway (provided the SOA is the first record in the file), but will issue the warning message "no TTL specified; using SOA MINTTL instead". To avoid problems, we recommend that you use a $TTL directive in each zone file. 2.2. Periods in SOA Serial Numbers Deprecated Some versions of BIND allow SOA serial numbers with an embedded period, like "3.002", and convert them into integers in a rather unintuitive way. This feature is not supported by BIND 9; serial numbers must be integers. 2.3. Handling of Unbalanced Quotes TXT records with unbalanced quotes, like 'host TXT "foo', were not treated as errors in some versions of BIND. If your zone files contain such records, you will get potentially confusing error messages like "unexpected end of file" because BIND 9 will interpret everything up to the next quote character as a literal string. 2.4. Handling of Line Breaks Some versions of BIND accept RRs containing line breaks that are not properly quoted with parentheses, like the following SOA: @ IN SOA ns.example. hostmaster.example. ( 1 3600 1800 1814400 3600 ) This is not legal master file syntax and will be treated as an error by BIND 9. The fix is to move the opening parenthesis to the first line. 2.5. Unimplemented BIND 8 Extensions $GENERATE: The "$$" construct for getting a literal $ into a domain name is deprecated. Use \$ instead. 2.6. TXT records are no longer automatically split. Some versions of BIND accepted strings in TXT RDATA consisting of more than 255 characters and silently split them to be able to encode the strings in a protocol conformant way. You may now see errors like this dns_rdata_fromtext: local.db:119: ran out of space if you have TXT RRs with too longs strings. Make sure to split the string in the zone data file at or before a single one reaches 255 characters. 3. Interoperability Impact of New Protocol Features 3.1. EDNS0 BIND 9 uses EDNS0 (RFC2671) to advertise its receive buffer size. It also sets DO EDNS flag bit in queries to indicate that it wishes to receive DNSSEC responses. Most older servers that do not support EDNS0, including prior versions of BIND, will send a FORMERR or NOTIMP response to these queries. When this happens, BIND 9 will automatically retry the query without EDNS0. Unfortunately, there exists at least one non-BIND name server implementation that silently ignores these queries instead of sending an error response. Resolving names in zones where all or most authoritative servers use this server will be very slow or fail completely. We have contacted the manufacturer of the name server in case, and they are working on a solution. When BIND 9 communicates with a server that does support EDNS0, such as another BIND 9 server, responses of up to 4096 bytes may be transmitted as a single UDP datagram which is subject to fragmentation at the IP level. If a firewall incorrectly drops IP fragments, it can cause resolution to slow down dramatically or fail. 3.2. Zone Transfers Outgoing zone transfers now use the "many-answers" format by default. This format is not understood by certain old versions of BIND 4. You can work around this problem using the option "transfer-format one-answer;", but since these old versions all have known security problems, the correct fix is to upgrade the slave servers. Zone transfers to Windows 2000 DNS servers sometimes fail due to a bug in the Windows 2000 DNS server where DNS messages larger than 16K are not handled properly. Obtain the latest service pack for Windows 2000 from Microsoft to address this issue. In the meantime, the problem can be worked around by setting "transfer-format one-answer;". http://support.microsoft.com/default.aspx?scid=kb;en-us;297936 4. Unrestricted Character Set BIND 9.2 only BIND 9 does not restrict the character set of domain names - it is fully 8-bit clean in accordance with RFC2181 section 11. It is strongly recommended that hostnames published in the DNS follow the RFC952 rules, but BIND 9 will not enforce this restriction. Historically, some applications have suffered from security flaws where data originating from the network, such as names returned by gethostbyaddr(), are used with insufficient checking and may cause a breach of security when containing unexpected characters; see for details. Some earlier versions of BIND attempt to protect these flawed applications from attack by discarding data containing characters deemed inappropriate in host names or mail addresses, under the control of the "check-names" option in named.conf and/or "options no-check-names" in resolv.conf. BIND 9 provides no such protection; if applications with these flaws are still being used, they should be upgraded. BIND 9.3 onwards implements check-names. 5. Server Administration Tools 5.1 Ndc Replaced by Rndc The "ndc" program has been replaced by "rndc", which is capable of remote operation. Unlike ndc, rndc requires a configuration file. The easiest way to generate a configuration file is to run "rndc-confgen -a"; see the man pages for rndc(8), rndc-confgen(8), and rndc.conf(5) for details. 5.2. Nsupdate Differences The BIND 8 implementation of nsupdate had an undocumented feature where an update request would be broken down into multiple requests based upon the discovered zones that contained the records. This behaviour has not been implemented in BIND 9. Each update request must pertain to a single zone, but it is still possible to do multiple updates in a single invocation of nsupdate by terminating each update with an empty line or a "send" command. 6. No Information Leakage between Zones BIND 9 stores the authoritative data for each zone in a separate data structure, as recommended in RFC1035 and as required by DNSSEC and IXFR. When a BIND 9 server is authoritative for both a child zone and its parent, it will have two distinct sets of NS records at the delegation point: the authoritative NS records at the child's apex, and a set of glue NS records in the parent. BIND 8 was unable to properly distinguish between these two sets of NS records and would "leak" the child's NS records into the parent, effectively causing the parent zone to be silently modified: responses and zone transfers from the parent contained the child's NS records rather than the glue configured into the parent (if any). In the case of children of type "stub", this behaviour was documented as a feature, allowing the glue NS records to be omitted from the parent configuration. Sites that were relying on this BIND 8 behaviour need to add any omitted glue NS records, and any necessary glue A records, to the parent zone. Although stub zones can no longer be used as a mechanism for injecting NS records into their parent zones, they are still useful as a way of directing queries for a given domain to a particular set of name servers. 7. Umask not Modified The BIND 8 named unconditionally sets the umask to 022. BIND 9 does not; the umask inherited from the parent process remains in effect. This may cause files created by named, such as journal files, to be created with different file permissions than they did in BIND 8. If necessary, the umask should be set explicitly in the script used to start the named process.