OK, we have basic UDP-only caching (with the ability to read and write
the cache to disk).  Let's improve that:

* If the server sends us "truncated", just give the user this reply
  without caching it. (DONE; caching disabled for tuncated replies)
  Make sure we fully validate a truncated packet.  (No, we just tell them
  it's truncated and they can use DNS-over-TCP if enabled)

* Bugfix: The Google problem.  Google.com's DNS reply is in the form
  "Cache for one day: Google.com is a CNAME for www.l.google.com;
   Cache for one hour: www.l.google.com has the ip 192.168.1.55"
  Deadwood 2 only looks at the first TTL for a DNS packet and use that
  to determine how long to cache a reply; Deadwood 2 should look at the
  first 2 (or first N) TTLs and cache the entire reply for the minimum 
  TTL Deadwood sees (right now Deadwood 2 caches Google.com for one week
  when it really should only cache it for one hour) (DONE)

* RFC compliance: Records with a TTL of 0 should not be cached. (DONE
  by Neeo; undone by myself because recursive Deadwood needs to put
  items in the cache before giving them to the user)

* Safe cache writing: Write to a temporary file and rename the file to 
  the name of the cache file when the write is done. (Never done)

* Timed cache writing: Allow Deadwood to write to the cache every N minutes;
  this will allow Windows users to have cachefiles. (The Windows port just
  writes the cache when the service is stopped, so Never done)

* Windows support: (Windows is important because MaraDNS is the only 
  open source DNS server I know of, besides BIND, to run on Windows)
  How does Windows do signals, so writing the cache to a file works
  in Windows? (DONE; Deadwood now Windows service)

* OK, once all of that is done, another wishlist item: Handle DNS-over-TCP
  correctly.  Instead of having TCP connections handled by a separate
  non-caching daemon, have the main daemon get TCP connections, 
  convert them in to UDP connections.  Also, should we get a truncated 
  UDP packet, try again using TCP. (DONE)

* OK, another wishlist item: Synthetic replies.   If they ask for an A
  record, and we can't talk to upstream -or- upstream gives us a 
  not there/NXDOMAIN, we give them a synthetic A record.  Think "You're
  not logged on to our network" pages. (I've done the opposite: If you 
  have an ISP with an annoying page with ads when you mistype a domain 
  name, Deadwood can make the DNS reply "not there" instead).

* Wishlist item: TTL aging.  Not really sure how I'll pull this off
  (give each hash entry an "atime" entry?; make it so each entry has the
   same TTL for all DNS entries?) (DONE)

Some thoughts about the cache design.

* The code that retries may end up trying to contact the same dead upstream
  server again.  We should change this; when making a connection upstream,
  we should note which server we tried to contact, and have it so we don't
  contact that server the next time we try to make a connection.  Maybe even
  have an array of the number of times we have failed to connect to a given
  upstream server, and have the code use a heuristic to try less connections
  with an unreliable server. (DONE; we round-robin rotate if we can't connect
  to an upstream NS server)