3.7.0 - 2021-06-08
    * TCP threads now shut down faster on daemon stop/replace with active
      connections; previously they took up to 10s in a two-phased shutdown
      approach, now they take up to 5s in a single phase.
    * TCP threads' timeout management fudge factor for coalescing timer events
      from multiple connections was increased from 100ms to 250ms.
    * The UDP send-side code is now non-blocking; it will immediately fail for
      lack of SO_SNDBUF space for the response instead of blocking up to ~257ms
      in an attempt to wait for for buffer availability.  This has less chance
      of negatively impacting other requests waiting in the receive buffer
      during short spikes, at the cost of burning a little more CPU (to create
      answers which are quickly dropped on the output side) when buffers are
      overwhelmed with traffic.
    * Control sock proto now documented (imperfectly!) at docs/ControlSock.md
    * Control sock clients now send their version in REQ_INFO
    * Clearer messaging about certain corner-case configuration parser failures
      related to bad DNS escapes and $include files.
    * Bugfix for parsing zonefiles > 4GiB in size
    * Some possible race conditions in the testsuite which could cause false
      failures were fixed
    * Various internal refactorings and code QA improvements

3.6.0 - 2021-02-24
    * UDP DNS threads now idle much more efficiently when there's little to no
      network traffic arriving for them.  Previously all such threads woke up
      and re-executed recvmsg() at least once every ~3.1 seconds.  Now they
      block indefinitely in a ppoll() call once they've been idle for ~257ms or
      longer.  This reduces thread wakeup churn and thus power/cpu efficiency
      waste in the case of unused threads for extra listen addresses.  It also
      reduces the worst corner case for daemon shutdown delays caused by UDP
      threads from ~3.1s to ~257ms.
    * TCP DNS threads now invoke rcu_quiescent_state() more often, which
      can help data reloads complete faster when there are many active TCP
      clients per thread.  Previously they quiesced or went offline from RCU
      once per eventloop iteration, where one iteration could handle up to 1
      pending request per client connection for many connections.  Now they
      also quiesce once for each response sent.
    * SO_REUSEPORT_LB is now used for better UDP and TCP socket loadbalancing
      on BSDs which have this socket option available.
    * Zone and config files are now loaded more efficiently when there are many
      small ones, by switching to a malloc()+read() pattern instead of mmap()
      for files that are smaller than 1MB.  mmap modifies process-global memory
      mappings once per file, which is expensive for a multi-threaded daemon
      loading many small files (especially in parallel!), so it's just not
      worth it unless the files are large-ish.

3.5.2 - 2021-02-23
    * Fixed a very rare potential inter-thread race on daemon shutdown due to a
      'stop' or 'replace'.  If a still-in-progress zone reload operation had
      started just before the stop/replace, and happened to finish its job at
      nearly the exact same moment that the main thread was deallocating the
      zone data, both the reloader thread and the main thread could attempt to
      deallocate it at the same time.  This happens after DNS listeners are
      stopped, but it's still an unclean crash instead of a clean termination.
      In the 'replace' case it would also cause the new daemon to get no stats
      data handoff thanks to the crash (but the new daemon would still
      successfully start).  This issue hasn't ever been observed or reported in
      the wild, but I'm pretty sure it was possible with some extremely low
      probability.
    * The timeout for some control socket network operations between two
      daemons during a 'replace' takeover was extended from a fixed 13s to a
      fixed 37s.  There is no perfect value here that balances all concerns
      without adding a ton of new code complexity, but this new value should be
      more robust in edge cases in practice.  Making it configurable for those
      with known edge cases would be an easy next step!
    * Fixed an snprintf format bug that effectively limited stats outputs to
      32-bit counters on x86_64 x32 ABI builds, even though they were being
      tracked as 64-bit internally.
    * Fixed a minor snprintf format bug in a log message about very early
      failures in zonefile processing, which probably wasn't causing any real
      bug on most platforms
    * Hopefully improved the reliability of the 011random.t test on hosts
      that are slower and/or have smaller default UDP receive buffers.
    * Added a docs and log warning about not relying on the default 'any'
      listen option without explicitly requesting it, for compat with future
      versions.
    * Relaxed a constraint in the socket handoff process between daemons
      during 'replace' in the interest of compat with future versions.
    * Added some additional pedantic safety measures to how TCP thread
      shutdowns are sequenced.
    * Improved the utility and/or accuracy of various error log messages
    * A few small code QA nits of no real consequence

3.5.1 - 2021-02-10
    * [Fixes marked with [BP] are also backported to 3.[234].x branches for
      convenience, but new releases from those branches are not imminent]
    * Fixed issue #202, which was causing non-ANY UDPv4 listen sockets to fail
      to work correctly on FreeBSD 12.2 due to some new strictness in its
      parsing of sendmsg() msg_hdr contents since sometime after 12.0. [BP]
    * Fixed a bug which may have caused issues listening on global unicast
      UDPv6 addresses bound to loopback interfaces on some platforms.  This
      was previously fixed for just FreeBSD in 2.4.2 and 3.0.0.  The existing
      fix was just extended to all other platforms, as it probably should've
      been the first time around. [BP]
    * Fixed a bug that was causing escape sequences for binary bytes in
      domainnames to be emitted incorrectly in log message outputs. [BP]
    * A workaround was added for building on platforms which lack
      MSG_CMSG_CLOEXEC for SCM_RIGHTS socket handoff.
    * Some fatal error message outputs in the case of a failed "replace"
      operation were cleaned up for consistency.
    * A new configure argument "--with-buildinfo=FOO" was added so that
      packagers can build with a consistent version output encoded in the
      binary regardless of whether they're building directly from git.
    * Some docs, comments, and log message typos fixed by lintian
    * A few small code QA nits of no real consequence

3.5.0 - 2021-01-11
    * Multiple efficiency improvements for TCP conn handling, primarily aimed
      at reducing the per-conn memory footprint and malloc churn.
    * Address (A/AAAA) records are now properly randomly shuffled each time
      they're emitted in responses.  Previously we were doing a basic rotation
      of them in the supplied order.  Shuffling has better load distribution
      characteristics in some scenarios.
    * NS responses now include true glue, when warranted, for qtype=NS at the
      root of a zone.  This covers some important corner cases, including the
      root-of-dns delegated glue hints case.  Previously, we were only
      attempting to attach glue to delegation responses, but never zone root
      qtype=NS responses.

3.4.2 - 2021-01-11
    * Several bugfixes are included here which came to light while reworking
      major chunks of the code in a 4.x prototype branch.  All are also
      backported to the 3.[23].x branches for packagers.
    * Fixed a few memory leaks that can happen in certain uncommon zonefile
      parsing failures related to TXT data, CAA data, and/or the zone
      filename itself
    * Fixed some missed response packet compression opportunities when the
      query name from the client was not already normalized to lower case.
    * Fixed some log message output bugs when complaining about SOA records
    * Added missing min_ttl/max_ttl clamping to RFC 3597 records' TTLs.
    * Fixed a corner case that could cause mixed TTLs in a single rrset, in
      the case of ACME challenges defined for the same name both statically
      in zone data and via gdnsdctl.
    * Fixed a bug which could force an RR's TTL to the value "1" if it were
      outside the range declared by ttl_min and ttl_max (default 5 and
      3600000) and the non-default "zones_strict_data" (or -S flag) was
      enabled.
    * Configurations of plugins and zone data which allow DYNC to return
      address (A/AAAA) records are deprecated.  They still work, but they
      emit a log warning message every time a zone is loaded which creates
      such a situation.

3.4.1 - 2020-11-20
    * Bugfix for a rare TCP connection state management issue which can cause a
      TCP thread to consume excess CPU and prevent the clean exit of the daemon
      process on shutdown.

3.4.0 - 2020-11-19
    * The default value of 'experimental_no_chain' has changed from false to
      true.  This change affects on-the-wire DNS responses in the case of a
      CNAME record which points to another record in the same zone. This is
      kind of an edge case for a semver minor version bump: this is not
      strictly a feature-add, but the behavioral change here is deep in the
      weeds most users will never see and has been about as widely tested as it
      can be without becoming a default like it is now.  In case of issues, it
      can still be disabled by explicitly configuring it to false (which will
      also emit an error to syslog at startup pleading for bug reports about
      any such necessary case).  This path seemed better than putting all the
      risks of this change in a future 4.0 release for the first time, where
      there would already be a ton of other core work built on top of it,
      making reversion much more difficult.
    * A new statistic 'tcp_acceptfail' was created to track all non-trivial
      failure returns from accept4().
    * The daemon now makes a soft, non-fatal attempt to raise the soft limit on
      total file descriptors if internal calculations indicate it may be too
      small for the configuration (especially tcp listener config).  It will
      complain to syslog on startup if it sees a potentially-bad fd limit
      situation that it can't fix due to the hard limits.
    * The TCP code now at least attempts to handle running into such a file
      descriptor limit more-gracefully by shutting down the most-idle of its
      older connections if possible, which is the same mechanism used by our
      existing internal limiter tcp_clients_per_thread.
    * Several sites in the code which may emit syslog errors driven by network
      input (e.g. recv() errors, accept4() errors, etc) now use a different
      logging call which applies some sanity-level ratelimiting to avoid
      excessive log spam.
    * max_edns_response[_v6] default values both changed to 1232, from previous
      defaults of 1410 and 1212, respectively.
    * The server now prefers the Linux IP_PMTUDISC_OMIT socket option over
      IP_PMTUDISC_DONT, when OMIT exists at compile time and works at runtime,
      which increases resilience against certain kinds of fragmentation
      attacks.

3.3.0 - 2020-09-09
    * New Feature: 'skip_first' flag added for geoip resource definitions.
      This allows defining a backup geoip resource which shares the same map as
      a primary resource, but always resolves to the second-best choice for a
      given client, as if the (dynamic) first choice were down.

3.2.2 - 2020-02-19
    * Bugfix: In the case that a "replace" daemon takeover happened while
      unexpired ephemeral ACME challenge data was present (which is handed
      off to the new replacement daemon), the new daemon could pause for the
      entire value of "acme_challenge_ttl" (default: 10 minutes) during its
      startup sequence.  The takeover/replace process would eventually succeed
      between the two daemons, but the long wait would cause various timeouts
      for control/notification by default (e.g. gdnsdctl won't wait that long
      to report the success).

3.2.1 - 2019-07-19
    * Fix CVE-2019-13951: IPv4 addresses in local zone file data which are
      longer than the maximum legitimate IPv4 address cause a stack buffer
      overflow and crash.
    * Fix CVE-2019-13952: IPv6 addresses in local zone file data which are
      longer than the maximum legitimate IPv6 address cause a stack buffer
      overflow and crash.

3.2.0 - 2019-05-05
    * ACME challenge response TXT RRs are now emitted with a TTL of zero by
      default, which makes things work reliably in some tricky corner cases
      with ACME providers.  This can be overridden via the new config option
      "acme_challenge_dns_ttl".
    * The minimum configurable value for "min_ttl" was reduced from 1 to 0.
    * The "min_ttl" setting now correctly clamps small TTLs for SOA records.
    * All out of zone glue addresses (truly out of zone, not just within a
      local delegation) now generate a warning on zonefile loading (or an error
      if zones_strict_data is in effect).  Support for out of zone glue will
      likely be removed in a future major version bump, as it's basically
      useless and non-recommended in the DNS in general at this point in
      history, and supporting it is needlessly complicated.
    * Minor refactorings, nit-fixes, and docs updates.

3.1.0 - 2019-03-30
    * RFC 8490 "DNS Stateful Operations" (DSO) implemented and enabled by
      default for all TCP listeners.  Can be disabled in case of interop or
      other issues via the new global config option "disable_tcp_dso".
    * Zonefile parsing can now use multiple threads for speed.  There's a new
      config option "zones_rfc1035_threads" (default: 2) to specify how many
      threads to use when parsing zones on start or reload.  You'll have to do
      some testing to find an optimal value for your setup.
    * TCP control sockets: you can now define a limited-functionality TCP
      control socket listener which gdnsdctl can connect to from another
      machine.  See the documentation for the new config option "tcp_control".
    * experimental_no_chain config option: As the name implies, this is for
      experimentation only.  If set to true, gdnsd will never chain through
      zone-local CNAME entries while generating responses, instead emitting
      just a singular CNAME RR.  See the documentation for more details!
    * Some substantial internal refactoring was done for code cleanliness and
      for making runtime queries more CPU/cache-efficient.

3.0.1 - 2019-03-04
    * Bugfixes:
      * The code which validates that loaded zonefile RRs are not oversized for
        implementation limits (~16KB encoded response data per RR Set) was
        having its results thrown away, nullifying the validation.  Now the
        limit is checked properly, and zonefiles containing such data will be
        rejected at load time.
      * Minor fix for UDP error statistics, in a corner case on platforms that
        do not have the recvmmsg()/sendmmsg() syscalls, where sometimes what
        should've incremented the UDP "recvfail" stat instead incremented the
        "dropped" stat.
      * Daemon no longer exits with an error if asked for resource states info
        over the control socket ("gdnsdctl states") when no such states are
        configured at all.
      * Removed redundant and outdated help string for "-i" from gdnsdctl's
        usage output.

3.0.0 - 2019-02-15
    * Note that this is a major version number bump, and for some users and
      configurations this means major compatibility-breaking changes.  Do not
      expect a simple upgrade like a minor version release!  Please see the
      separate file VERSION3.md which offers an overview of all the major
      changes from 2.x; there's too much to succinctly summarize here.
    * Also, all previous NEWS entries were shifted to NEWS.old, so that we can
      start this file over and leave previous 2146 lines of updates behind.
    * For beta users: there were no functional changes since the final
      2.99.9950-beta release, only QA/refactoring work and a couple of
      FreeBSD-specific fixups to the TCP parts of the testsuite.
