UdpLibrary Release Notes
=================================================================================================================

-----------------------------------------------------------------------------------------------------------------
6/11/2003
-----------------------------------------------------------------------------------------------------------------
- There are a lot of significant changes in this release.  CPU usage, throughput ability, and memory-usage
  have all been improved substantially, particularly for high-stress server-to-server type environments.
  The optimizations are such that old settings that used up a lot of memory can be changed to more reasonable
  values without adversely effecting throughput ability or CPU usage.  To make things simpler in this regard, 
  below is a list of recommended settings for different types of connections.  If you have connections of
  different types, ideally you would setup your architecture such that the different types were managed by
  different UdpManager's, such that optimal settings can be applied to both.  As always, you should not have
  more managers than is needed (perhaps one for all client connections and one for all interserver connections).
  Settings not specified below should be left at their default.  It is still important that you keep the GiveTime
  frame rate reasonably high.  For high-bandwidth server to server connections, I recommend at least 20, 8 minimum.
  If your frame rate drops below that and you are sending more than 500k/sec of data, then you better seriously
  consider putting the UdpManager into it's own thread.  In a future release (next release probably), I plan on
  including a threaded-wrapper version of the UdpLibrary to make this painless.  Until then, you may wish to
  talk to the SWG team, which has already done this, and may be able to provide you with some getting-started code.

  Server side settings for incoming player connections (ie. clients connecting are relatively low bandwidth)
  (total bandwidth per connection is intended to be low)
  -----------------------------------------------------------------------------------------------------------
	Params::maxConnections					set as appropriate
	Params::port							set as appropriate
	Params::outgoingBufferSize				set to 1mb-4mb
	Params::incomingBufferSize				set to 1mb-4mb
	Params::keepAliveDelay					set to 15000 (15 second keep-alives generally do the trick)
	Params::noDataTimeout					set to 46000 (46 seconds, typically just over 3 times client side keepAliveDelay)
	Params::reliableOverflowBytes           set to 2mb (however much you are willing to let them have backlogged)
	Params::hashTableSize					set to 10 times max expected connections
	Params::pooledPacketMax					set to roughly 5 times max expected connections, but rarely over 10000
	Params::crcBytes						set to 2
	Params::bindIpAddress[32]				set as appropriate (I like to leave it blank)

	Params::clockSyncDelay					leave at default of 0 (client side will do clock syncing)
	Params::portRange                       leave at default
	Params::packetHistoryMax				leave at default
	Params::portAliveDelay					leave at default of 0
	Params::replyUnreachableConnection		leave at default of true
	Params::allowPortRemapping				leave at default of true
	Params::allowAddressRemapping			leave at default of false
	Params::icmpErrorRetryPeriod			leave at default of 5000
	Params::oldestUnacknowledgedTimeout		leave at default of 90000 (or set more aggressively even, down to 30000)
	Params::maxDataHoldTime					leave at default of 50 (or set longer up to 100 possibly)
	Params::maxDataHoldSize					leave at default of -1 (meaning same as raw packet size)
	Params::maxRawPacketSize				leave at default of 512
	Params::avoidPriorityQueue				leave at default of false
	Params::pooledPacketInitial             leave at default of 0 (or set to pooledPacketMax for extra efficiency at cost of memory)
	Params::pooledPacketSize				leave at default of -1 (uses same size as raw packet size, this is critical)
	Params::wrappedPoolMax                  leave at default of 1000
	Params::processIcmpErrors               leave at default of true
	Params::connectAttemptDelay             leave at default (irrelevant for server side)
	Params::reliable[0]						leave all settings at defaults


  Server-to-server settings for LAN connections (ie. connections are local on high-bandwidth)
  (total bandwidth per connection is intended to be high)
  (same settings used regardless of who intiates connection)
  -----------------------------------------------------------------------------------------------------------
	Params::maxRawPacketSize				set to 1460
	Params::maxConnections					set as appropriate
	Params::port							set as appropriate
	Params::outgoingBufferSize				set to 4mb
	Params::incomingBufferSize				set to 4mb
	Params::keepAliveDelay					set to 15000 (15 second keep-alives generally do the trick)
	Params::noDataTimeout					set to 61000 (61 seconds, typically just over 4 times client side keepAliveDelay)
	Params::reliableOverflowBytes           set to 10mb (however much you are willing to let them have backlogged before disconnecting)
	Params::hashTableSize					set to 10 times max expected connections
	Params::pooledPacketMax					set to 5000
	Params::crcBytes						set to 2
	Params::bindIpAddress[32]				set as appropriate (I like to leave it blank)
	Params::icmpErrorRetryPeriod			set to 2000 (optional)
	Params::allowPortRemapping				set to false (optional)
	Params::oldestUnacknowledgedTimeout		set to 30000 (longer for debugging tolerance)
	Params::maxDataHoldTime					set to 0 (this is critical)
	Params::reliable[].maxOutstandingBytes		set to 2mb
	Params::reliable[].maxOutstandingPackets	set to 4000
	Params::reliable[].resendDelayAdjust		set to 1500
	Params::reliable[].congestionWindowMinimum  set to 50000

	Params::reliable[]						leave all other settings at defaults
	Params::clockSyncDelay					leave at default of 0 (client side will do clock syncing)
	Params::portRange                       leave at default
	Params::packetHistoryMax				leave at default
	Params::portAliveDelay					leave at default of 0
	Params::replyUnreachableConnection		leave at default of true
	Params::allowAddressRemapping			leave at default of false
	Params::maxDataHoldSize					leave at default of -1 (meaning same as raw packet size)
	Params::avoidPriorityQueue				leave at default of false
	Params::pooledPacketInitial             leave at default of 0 (or set to pooledPacketMax for extra efficiency at cost of memory)
	Params::pooledPacketSize				leave at default of -1 (uses same size as raw packet size, this is critical)
	Params::wrappedPoolMax                  leave at default of 1000
	Params::processIcmpErrors               leave at default of true
	Params::connectAttemptDelay             leave at default


  Client side settings for player connections (ie. client is connecting over the internet to the server)
  (total bandwidth per connection is intended to be low)
  -----------------------------------------------------------------------------------------------------------
	Params::maxConnections					set as appropropriate (probably something small like 5)
	Params::port							set to 0 (let client pick port)
	Params::outgoingBufferSize				set to 64k (default I believe)
	Params::incomingBufferSize				set to 64k (default I believe)
	Params::keepAliveDelay					set to 15000 (15 second keep-alives generally do the trick)
	Params::noDataTimeout					set to 46000 (46 seconds, typically just over 3 times client side keepAliveDelay)
	Params::reliableOverflowBytes           set to 512k (however much you are willing to let them have backlogged)
	Params::hashTableSize					set to 10 times max connections
	Params::clockSyncDelay					set to 60000 if you want packetloss stats or clock-sync features, otherwise 0
	Params::pooledPacketMax					set to 50
	Params::crcBytes						set to 2
	Params::maxDataHoldTime					set to 20 (more agressive for client to server data generally)
	Params::packetHistoryMax				set to 5 (to save memory on client)

	Params::bindIpAddress[32]				leave at default
	Params::portRange                       leave at default
	Params::portAliveDelay					leave at default of 0
	Params::replyUnreachableConnection		leave at default of true
	Params::allowPortRemapping				leave at default of true
	Params::allowAddressRemapping			leave at default of false
	Params::icmpErrorRetryPeriod			leave at default of 5000
	Params::oldestUnacknowledgedTimeout		leave at default of 90000 (or set more aggressively even, down to 30000)
	Params::maxDataHoldTime					leave at default of 60 (or set longer up to 100 possibly)
	Params::maxDataHoldSize					leave at default of -1 (meaning same as raw packet size)
	Params::maxRawPacketSize				leave at default of 512
	Params::avoidPriorityQueue				leave at default of false
	Params::pooledPacketInitial             leave at default of 0 (or set to pooledPacketMax for extra efficiency at cost of memory)
	Params::pooledPacketSize				leave at default of -1 (uses same size as raw packet size, this is critical)
	Params::wrappedPoolMax                  leave at default of 1000
	Params::processIcmpErrors               leave at default of true
	Params::connectAttemptDelay             leave at default
	Params::reliable[0]						leave all settings at defaults


- added reliable-channel level coalescing support.  This is a major feature enhancements to the UdpLibrary.
  Originally, the UdpLibrary only did coalescing at the low-level, such that it could combine together
  unreliable, reliable, and internal data into the same physical packet.  This is still a vital and key
  feature of the UdpLibrary.  Not coalescing up at the higher level of the reliable channel had its
  problems: 1) it required that ACK packets on a per-logical packet basis, meaning that if
  lots of tiny reliable packets were being sent, there was a bit more ack-overhead in the process.  This
  overhead is generally negligable except in the most extreme circumstances.  2) it required that the
  reliable channel object track all those outstanding packets, which proved to be a bit of a performance
  burden.  3) it required the reliable channel object allocated LogicalPacket objects to hold onto each
  individual application packet (in the usual case where the application wasn't directly sending
  LogicalPacket objects).  To make matters worse in this regard, it ended up getting these objects from
  the memory pool, which often effectively meant that tiny applications packets were stealing much
  larger pooled packets and wasting memory.  The new scheme has the advantage that many reliable sends
  will now be performed without an allocation and without using up a pooled packet.
  
  One nice effect of this change will be that a much smaller pool needs to be reserved.  Due to a
  limitation in the low-level coalescing, if your application sends lots of medium sized packets
  (256 to maxRawPacketSize bytes), then you may see a significant reduction in physical packets sent
  per second.  This change may sound like it would have a drastic effect on the protocol, but that is not
  the case, we remain protocol compatible with old version.  This is because the underlying protocol
  has always supported the concept of a group logical packet.

  The benefits of this feature will be most evident in high-bandwidth server-to-server connections.  You
  should make sure you pooledPacketSize = maxRawPacketSize, and you should be able to reduce the size of
  your pool substantially.

  Internally, the reliable channel will coalesce things up to the fragment size, which defaults to being
  the same as the maxRawPacketSize.  There is no hold-time for coalescing.  The next time the connection
  is given time (next time manager is given time), if there is flow-control-window space available, it
  will flush the coalesce buffer in an attempt to use it up.  So, it will effectively coalesce application
  sends until there is an opening in the flow-control-window and GiveTime is called.

  The only disadvantage I can see to leaving this feature enabled (it is configurable in
  UdpManager::Params::reliable[].coalesce) is for client-side low-bandwidth connections.  In those situations,
  it is important that the ACK traffic back to the server coalesce with application data.  This will occur
  even if this feature is enabled, but because the low-level coalescer is incapable of coalescing packets
  larger than 256 bytes, it is possible the reliable-sends from the client headed to the server might
  result in more physical packets per second.  Even then, I don't see it being a performance issue,
  so I would leave it enabled.  The configurability was largely added such that the library could be
  configured to operate as before, just in case there were some unforseen issues that cropped up with this.

  It's worth noting, that configuring the reliable channel to processOnSend will effectively circumvent
  any coalescing that you might have been getting.

- if reliable-channel coalescing is used (it defaults to ON), then the pooledPacketSize should be set
  to the maxRawPacketSize.  This is critical to performance.  The previous recommendation of setting
  it to the typical packet size won't work anymore, since it is now coalescing at the reliable channel
  level, effectively increaseing the size of packet it will grab from the pool typically.

- made substantial CPU-usage improvements to the reliable channel resend algorithm in situations where
  thousands of outstanding packets needs to be supported to achieve server-to-server throughput.
  This does not effect the protocol in any way, just uses slightly less memory to manage the process and
  potentially substantially less CPU.

- made a slight change to the reliable channel fast-recovery algorithm (based on the TCP Reno algorithm)
  When we notice that packetloss is selective (ie. some packet after the lost packet is acked and serves
  as a notice to do an accelerated resend on the other), we used to shrink the flow-control window
  by 50% (as opposed to non-selective loss which completely resets it).  We now only shrink the window
  by 33%, in an attempt to make for an even faster recovery and less hiccup.  I don't believe being this
  aggressive will be an issue, as we know more about the nature of the circumstances than Reno does, since
  we have selective acks.

- fixed a nasty design flaw that was causing the maximum effective reliable logical packet transfer rate
  to be limited to 256 times the number of GiveTime iterations per second.  When small packets are being
  sent, this can turn into a serious limitation, particularly on server-to-server communications where
  these packet rates potentially get very high.  The limit has been increased to 8192, which should be
  plenty large for even the highest bandwidth connections.  Note: this fix causes that an additional
  31k of stack space is used during the GiveTime call -- I am doubting this will cause anybody grief.

- improved the pooled packet scheme.  The old scheme used to fall back to standard SimpleLogicalPackets
  after it has created the maximum number it was allowed in the pool.  The new scheme always creates
  packets that can go into the pool when they are returned (if possible size wise).  It then releases
  packets that are returned to the pool if the pool is already at its max.  So as before when max
  represent the max anywhere, the max now represents the max it will hold onto in the available queue.
  This will make a big performance difference in situations where just one of many connections is
  backlogging.  In the old scheme, a backlogging connection could eat up the entire pool, leaving
  all other connections left doing allocs and frees of their packets.  With the new scheme, the
  bad connection can't monopolize the pool, as when the bad connection finally clears up and release
  all those packets, those packets will be destroyed even though they were gotten from the pool, because
  the pool available will have likely grown as appropriate.  Between this change and the reliable
  coalescing, the pooledPacketMax setting should be set MUCH smaller than it previously was.  I
  wouldn't set it to over 5000, even for server-to-server connections.  Additionally, I would set
  pooledPacketInitial to under 1000, since the vast majority of the time I don't see this ever
  being exceeded.

- changed the default setting of pooledPacketMax down to 1000 (from 2000), in order to reflect the
  recent optimizations noted above and the lack of need for a large pool most of the time.

- changed the default setting of pooledPacketSize from 256 to -1.  -1 means use whatever size the
  maxRawPacketSize is set to.  As a general rule, you should never change this setting from this.
  Now that the reliable channel is coalescing, all allocations from the pool are of maxRawPacketSize.
  If you set the pool size smaller than this, then odds are the pool will be completely worthless.
  There are some odd cases where you may wish to set it higher, depending on the nature of the
  packets your application sends.

- got rid of the deque object that was used internally to track outstanding LogicalPackets that were
  queued up on the reliable channel.  Since we put in coalescing at the top, all packets smaller than
  the maxRawPacketSize were getting coalesced into a larger logical packet anyways, meaning that the
  logical packet that we owned was extremely unlikely to be shared.  If we knew for a fact that all the
  logical packets we had queued were not shared by other connections, then we could get rid of the deque
  object and have a much faster and simpler linked-list scheme.  Since shared packets are now so rare,
  I decided it would be worthwhile to switch to a linked list and in situations where the logical packet
  is shared, I create a WrapperLogicalPacket object to effectively make it appear unique from a linked-list
  management point of view.  This wrappers are actually gotten from a pool owned by the UdpManager.  You
  can set the size to allow this pool to grow with Params::wrappedPoolMax.  I recommend leaving this
  value at the default of 1000, though if you are a server doing a lot of large shared packets, then it
  may be worth your while setting this to a larger value.  This optimizations gets rid of the nasty
  deque expansion problem that was eating up so much CPU when backlogging occurred.

  Note:  This also means the PointerDeque.hpp has been removed from the UdpLibrary as it is no longer
  used.  You can remove this object from your build.

- added ack-deduping support.  This can be disabled (Params::reliable[].ackDeduping), but defaults to
  enabled (true).  I really can't think of any reason you would ever want to turn it off.  What it does
  is cause ack-packets in the low-level coalesce buffer to be deduped.  There are no performance
  implications (in fact, it is faster when enabled).  How it works is if the library is sending an
  ack-all packet and there is a previous ack-packet of some sort still sitting in the low-level
  multi-buffer, then it will replace the existing ack instead of appending  one onto the multi-buffer.
  In high-bandwidth situations (or if reliable coalescing is disabled), this can result in an appreciable
  bandwidth savings.  The only situation where you might want to disable this is if for some reason you
  were counting on the additional acks to intentionally overflow the multi-buffer and force it to send
  rather than waiting for multi-buffer timeout.  But, that doesn't make a whole lot of sense.

- changed the default value of maxDataHoldSize to -1 (used to be 512).  Setting it to -1 now causes it
  to use the maxRawPacketSize for this.  Generally you should always be wanting to coalesce to as
  large of a packet as is allowed.

- added new configuration parameter, icmpErrorRetryPeriod.  This setting allows you to not terminate the
  connection at the first sign of trouble (ie. an ICMP error), but instead wait for a period of time in
  hopes that the situation will correct itself.  Normally, a single ICMP error packet is sufficient cause
  to terminate a connection.  However, there are situations where it may be desireable to tolerate ICMP
  errors for a few seconds as sometimes these situations resolve themselves.  In particular, the new
  port-remapping feature (see below) requires that some tolerance be set in this regard, as ICMP errors
  will likely occur for a int16_t period of time while remapping is occurring.  Although it changes
  previous behavior of the library, I have decided to make this new setting start out at 5000ms.  This
  means that ICMP errors for port unreachable and such will not occur for at least 5 seconds, whereas
  before they may have occurred nearly instantly.  If you want the previous behavior, setting this value
  to 0 will cause it to disconnect on even a single ICMP error.  Once the first ICMP error is received,
  the timer is started.  If a subsequent ICMP error comes in after the timeout period has expired, then
  it will be processed and the connection will be closed.  Once a successful packet is received on the
  connection, the ICMP timeout period is reset, such that a subsequent ICMP error will be treated as if
  it were the first.

- added new configuration parameter, allowPortRemapping.  This setting allows the UdpLibrary to effectively
  change the port-number associated with a particular connection dynamically on-the-fly.  Now, normally
  this would never happen, but sometimes a user may be going through a NAT router that feels the need to
  change the associated mapping mid-session.  Normally when this occurs, connections are simply lost.
  However, due to a bug in the #1 selling Linksys router whereby it does this remapping every 10 minutes,
  we have decided to add this feature to deal with it.  This feature is also generally useful in other
  situations.  For example, with the EQ channel-chat system, the user often goes for quite a while without
  sending or receiving any messages.  Many NAT routers expire the port-mappings fairly quickly if data
  is not being sent.  The usual solution to this problem is to configure it to send keep-alives, and
  generally that works.  In the case of the EQ channel-chat, we have keep-alives configured to go out
  every 45 seconds, which keeps 99% of NAT routers alive and happy.  The other 1% seem to expire after as
  little as 30 seconds.  We put in an override setting in EQ to have it send the keep alives every 15
  seconds instead, but it is a less than ideal solution as it requires the user changing things.  We 
  don't want to configure everybody to send keep-alives that frequently, as there are 100,000 connections
  to the server, and the keep-alives alone start to really add up.  This port remapping system will
  effectively allow the mapping to expire and the connection never miss a beat.  See allowAddressRemapping
  for a discussion of security concerns.  This parameter defaults to 'true' or enabled.  While this
  feature adds some new packetting, it remains protocol compatible with old versions.  You will need to
  upgrade both ends of the connection for this feature to be available though.

- added new configuration parameter, allowAddressRemapping.  This setting is almost identical to the
  allowPortRemapping setting, only it takes the concept one step further, and allows for the IP address
  itself to be changed on-the-fly as well.  Normally when only port remapping is enabled, if the IP address
  is changed, it will not be able to pick up on it.  By default, this parameter is 'false' or disabled.  It
  is highly unlikely that under any sort of normal operating conditions this will occur.  About the only
  reason I can think of for this is if you wanted to be so tolerant as to allow a modem user to disconnect
  and reconnect from their ISP and maintain their connection.  The biggest danger in enabling this feature
  is the potential for somebody to hijack somebody elses connection.  This is extremely unlikely however
  as the UdpLibrary effectively uses a 64-bit randomly generated key to authenticate any request for
  a remapping to occur.  To hijack a connection, somebody would have to guess what that 64-bit key value
  is for a particular IP address.

- added a new static UdpConnection::DisconnectReasonText function that will translate a disconnect reason
  into a displayable reason-string to make logging of disconnect reasons easier to read.

- put in safety checks to try and identify packets that are corrupt for some reason.  When one of these
  packets is spotted, the application is notified via the UdpConnectionHandler::OnPacketCorrupt callback.
  Statistics are also kept on the connection and on the manager regarding how many of these occur.  If you
  have CRC bytes enabled, then the odds of a corrupt packet making it through the check are extremely low.
  In this case, odds are the actual cause of the corruption is intentional packet tampering by the user;
  that is, they modify the packet and touchup the CRC so it passes.  This is why we callback the application
  as they may wish to log this event as a means of monitoring possible cheating.  This new callback provides
  the application with the packet data at the time the corruption is sensed.  This means that if the
  corruption is sensed after the decryption has occurred (say, by the multi-packet unpacker), then the
  data passed to the application will reflect the decrypted packet.  Additionally, the callback provides
  a reason why the packet was deemed corrupt.  If you are seeing a lot of corrupt packets for some reason,
  it may be worth logging the reason and trying to find a pattern (I also like to log the IP of where it
  came from, as often times bad packets tend to come from the same IP over and over, often due to
  tampering).

- added new configuration parameter, reliableOverflowBytes.  Normally, you can send an unlimited amount
  of data to a connection and it will never overflow, instead it will queue up the data until it runs
  out of memory and crashes.  In the past, it has been the applications responsibility to periodically
  check how many bytes are pending using UdpConnection::TotalPendingBytes and take appropriate action.
  This new parameter causes the UdpLibrary to do this check for the application, and if a connection
  ever overflows this amount (all reliable channels added together), then the connection is disconnected
  with a reason of cDisconnectReasonReliableOverflow.  The default setting for this parameter is zero, 
  which disables this check and makes it operate as it always has.  The check for an overflow condition
  does not occur at send time as this would result in a callback to the application at an illegal time;
  instead, the overflow condition is checked the next time the connection object is given time.

- added a new sample-program to the distribution.  This program is called 'udpstress' and simulates what
  one might see in a backend server-to-server cluster implementation.  Each udpstress client that starts
  up connects to one of the existing udpstress clients that is already running, and effectively adds itself
  to the cluster.  Each node in the cluster transfers data to every other node in the cluster at a compile
  time configurable rate.  Since everybody is connecting to everybody else, this stress program operates as
  both a client and a server.  Unlike the other sample program, it fully takes advantage of the new
  callbacks that were added and more accurately reflects the proper way of setting up an architecture.
  This program is also portable to other platforms.

- removed the old client/server test programs from the distribution.  The udpstress program covers both
  client and server examples, is far more flexible for testing, and is written to take advantage of the
  latest features, providing a better example to follow.

- moved the processing of the flagged port unreachable further up in the InternalGiveTime function.  This
  would have only effected the sparc version of the UdpLibrary and the way it handles ICMP errors.  There
  was a bug whereby it was going through the connected give-time logic when in fact it may have been
  disconnected by a port unreachable.  This likely didn't cause any problems, but I moved it just to be
  safe.

- added new UdpMisc::Sleep function.  This is not used by the UdpLibrary, but is used by the sample
  applications and in general provides a portable way for applications to sleep.

- made some internal optimizations that allows the UdpLibrary to avoid making an extra copy of the data
  (in most circumstances) in order to append CRC bytes.  When coalescing is likely to occur and encryption
  is not being used, this optimization will speed up the cost of CRC checking on the send side a fair amount.

- created a new variation of the HashTable called an ObjectHashTable.  The UdpLibrary now uses this variation
  which is optimized to assume that the contents of the hash table are pointers to objects that are derived 
  from HashTableMember.  This allows the hash table to avoid allocations/deallocations entirely when new 
  entries are added/removed.  UdpConnection objects make use of this optimization.  The new hash table object
  has restrictions the old one does not, namely, its members can only ever be in one hash table, and then
  only in it one time.  For most applications, this is not an issue.  The old version is no longer used by
  the UdpLibrary but has been kept in the header file because some applications have started making use of it,
  and we don't want to mess them up.

- recently, the UdpLibrary has been seeing much heavier load as a server-to-server interprocess communications
  protocol.  While it was designed to handle this sort of thing, it did have one int16_t-sighted limitation; it
  was only capable of transferring 4 billion logical reliable packets per session before it would fall apart.
  Even at a very high average trasfer rate of 1000 packets per second on a single connection, it would take over
  a month to exceed this rate.  It seemed at first that this might be a reasonable limitation, until Planetside
  and SWG came along.  These games have interprocess packet rates that are sometimes 10 times that high, meaning
  that critical failure could occur in only a few days in extreme circumstances.  Fixing this limitation did
  not require a protocol change, merely some internal changes to the way packets are tracked.  You need only
  upgrade both sides of a connection simultaneously to maintain compatibility if that particular connection
  will exceed 4 billion packets in a single session.  There is now no limit on the amount of data that a single
  session can send reliably.

- for the Win32 version of the Clock function, I had come up with a rather clever technique to avoid threading
  issues and clock-wrap.  Turns out the thread-local-storage scheme I was using doesn't work if the UdpLibrary
  is stuck in a DLL.  The odds of a clock wrap that occurs once every 47 days thread-slicing between two
  consecutive interger assignments and causing a problem are astronomically rare, so I have decided not to
  worry about the potential issue.  What I really need is a 64-bit interlocked increment, but Windows doesn't
  add support for that until the next version (after XP).

- added new configuration parameter, processIcmpErrorsDuringNegotiating.  This boolean parameter defaults to
  false, which was the default behavior of the UdpLibrary before this became configurable.  Setting this to 
  true will cause a client side connection attempt to fail immediately if the server process is not running
  (ie. doesn't have the port open).  Usually you will want to leave this at the default as often times two
  processes that talk to each other are started at the same time, and you don't want a race-condition to cause
  the client to fail because it happened to startup quicker than the server.

- put a hard cap on the maximum allowed maxOutstandingPackets configuration value.  If this value were ever
  set past 30000, things potentially could have fallen apart.  I can't imagine there ever being a need for this
  many outstanding packets.  Eventually when I get the reliable-channel coalescing put in, this cap will become
  even more meaningless.

- made the maximum allowed resend delay configurable.  This allows the application to put a cap on the longest
  it will ever wait to resend a lost packet.  Without a cap, lost-packet throttling can potentially cause the
  resend delay to get excessively long.  You should generally leave this value alone, but tweaking with it for
  certain internal-network connections may be desireable to optimize performance.  For example, if you know you
  are on an internal network, then it may be safe to set this value to the larger of double the typical ping-time
  and the longest GiveTime cycle time of either end of the connection.  Like many values, unless you really
  know what you are doing, you probably shouldn't mess with this.  I would never set this value to less than
  1000 for anything, but setting it to a value like 2000 might be desireable for some internal server-to-server
  connections.  The default is 5000.  This cap has always existed in the UdpLibrary, it has just been hard-coded
  in the past.

- got rid of the temporary encryption and decryption buffers in the UdpManager and replaced with stack-based
  buffers.  Doing this has two side effects.  1) more stack space is used, though it won't be very much.
  2) there is now a hard-limit on how large you can set the maxRawPacketSize.  Currently that hard limit
  is set to 16384 bytes.  I can't imagine anybody in their right mind would want to set it anywhere near
  that high.  On some crazy chance that you did want to set it higher, you would simply need to change the
  cHardMaxRawPacketSize constant in the header file.  In theory it could never go over 65535 anyways, since
  that is an IP protocol hard limit.  The only implication to increasing this number is more stack space usage.

- added debug signatures to the end of the encryption buffer and an assert to catch cases where user-supplied
  encryption routines overflow the end of the destination buffer.  Have yet to see this happen, but I am
  learning more and more that the more stuff the UdpLibrary can do to help the application find bugs in its
  usage of the library, the better.

- made sure UdpManager constructor initialized mPassThroughData to NULL.

- got rid of UdpMisc::FindAvailablePort function.  The reason for getting rid of this is that any use of
  it will almost surely be an unsafe/bad practice, as once it has found a port for you, it is entirely possible
  that some other process on the machine may grab that port before you create a UdpManager and bind to it.  We
  found this out the hard way in EQ.

- added a new configuration option, Params::portRange.  This new parameter accomplished what FindAvailablePort
  was supposed to do, only it does it in a safe manner.  If you wanted a to bind to a random port between
  40000 and 40999, you would specify 'port' to be 40000 and portRange to be 1000.  Upon construction, the
  manager would then randomly pick a port in that range and try binding to it.  It will try all ports in the
  range before it completely fails and gives up.  You cannot specify a portRange to use if you have 'port'
  set to zero as this makes no sense.  By default this parameter is 0, which causes it to act as it always has
  previously.

- against my better judgement, I added a GetRefCount function to the UdpConnection, UdpManager, 
  and LogicalPacket objects.  I actually needed it on the UdpConnection object internally such that the
  UdpManager could tell if the application had accepted a connection during an OnConnectRequest callback.

- added a new disconnect reason, cDisconnectReasonConnectionRefused.  You will never actually see this reason
  returned from GetDisconnectReason, since the object that would return it will already be destroyed so you can't
  ask it.  Where you will actually see it is from the GetOtherSideDisconnectReason function.  When a client
  attempts to connect to a server and that server ignores their connect request (by simply not AddRef'ing 
  it during the OnConnectRequest callback), the server responds back to the client with a terminate-packet.  The
  client will see the disconnect reason as cDisconnectOtherSideTerminated, but can further query to see that the
  server explicitly refused the connection.

- added a new disconnect reason, cDisconnectReasonMutualConnectError.  You will only ever see this from
  GetDisconnectReasonOther and it is a form of connection-refusal as well.  In this case, what it means is
  that the guy you are trying to connect to is simulatanously in the process of trying to connect to you.  The
  odds of this happening are very rare, but since the code was able to distinguish it from other situations, I
  figured it was worth clarifying if it ever did happen.

- added a new disconnect reason, cDisconnectReasonConnectingToSelf.  This is identical to the MutualConnectError
  and occurs whenever you try to establish a connection to yourself.  This is somewhat more likely, and in fact
  was added to the library because it actually happened.  For the sake of others who might run into this, I will
  give a brief description of how we managed to get bit by it.  We had enabled ICMP error processing during
  connection-negotiation, such that we could instantly tell if the server were down.  Our client application would
  then fail the connection, and immediately try again.  Each time it tried, it picked the next available port.
  The client in this case was another server process that happened to be running on the same box as the server
  it was trying to connect to.  The server it was trying to connect to was down. The server it was trying to
  connect to was supposed to be listening on port 44452.  The application-level retry loop was very tight with
  no delay, and eventually (30 seconds or so), it had managed to cycle through enough ports that the one it
  randomly picked (next available) was 44452.  This effectively made it connect to itself, which actually caused
  the library to eventually crash.  This was SWG btw.

- fixed bug in CRC calculation whereby it was applying the encrypt code in a byte-order dependent manner.  Nobody
  actually found this bug since nobody is running on big-endian based hardware.  There may be other endian issues
  that have yet to be tested, but in general the library is supposed to handle endian issues.  Internally, all
  protocols and such are done in network-byte-order (motorola big-endian order that is)

- added new api-function UdpConnection::GetOtherSideDisconnectReason().  When a connection is terminated, if the 
  GetDisconnectReason function returns to you the value cDisconnectReasonOtherSideTerminated, then you will be able to
  call this new function and find out what the other-sides reason was for disconnecting.  While this latest version of
  the library remains backwards protocol compatible, this new API feature will only be available if both sides of the
  connection are using the latest version of the library.  If the other side is not using the latest version, then
  this function will return cDisconnectReasonNone, meaning that it doesn't know.  Likewise, if this function is ever
  called when the primary disconnect reason is not OtherSideTerminated, then it will return None.  This should make
  tracking client-disconnects easier in situations where you don't have easy access to the client that is having the 
  problem.

- added new api-function UdpConnection::SetSilentDisconnect(bool silent).  This function allows the application to tell
  a connection that it should not notify the other end of the connection when it terminates (by sending the usual terminate
  packet to the other side).  This was actually added for internal reasons such that we didn't reply to a terminate-packet
  from the other side with a terminate packet of our own; however, it was thought that in some circumstances it may
  be desireable to let the other side languish and have to time-out, rather than us explicitly telling them that we
  have shutdown our side of the connection.  An example of wanting to do this might be if you were testing timeout
  functionality, or if you wanted to dump a cheater hard without telling them explicitly that they had been disconnected.

- added statistics to track how many socket-overflow errors occurred.  A socket overflow error can be caused if the
  socket-buffer size is too small for the overall datarate that the manager is trying to send (which could very well
  happen on a busy server if you gave it a small buffer.  The net result is packetloss, which will cause extra resends
  later on (for reliable data), etc.  If you have a fairly big outgoing socket buffer size, then this may be caused
  if the application is spamming too much unreliable data out the pipe.  Remember, unreliable data is not flow-controlled
  in any way, so if the server instantly dumps 500k of unreliable data to all connections combined and the server only
  has a 256k outgoing socket buffer, then you are going to lose a healthy chunk of that data.  This statistic can be
  used to determine if you have this problem occurring on a frequent basis.

- made UdpManagerHandler::OnConnectRequest non-pure-virtual (ie. provided a default implementation).  It used to be the
  case that the only reason you would want a UdpManagerHandler is if you were listening for connections, now it is quite
  likely that you will want one for user-supplied encyrption on the client side, so it was thought that we really should
  provide a default implementation that basically ignored incoming connection requests.


-----------------------------------------------------------------------------------------------------------------
3/7/2003
-----------------------------------------------------------------------------------------------------------------
- Added new configuration parameter to reliable channels UdpManager::Params::reliable[].congestionWindowMinimum.  It is very
  unlikely that you would want to change this parameter from its default value of zero.  What this parameter does is allow
  you to control how small (in bytes) the flow-control congestion window is able to shrink.  How flow control works is that
  as connections get congested, they slow down sending data by controlling how much outstanding unacknowledged data there
  is.  The congestion-window itself represents how much data this is, and it changes constantly as appropriate.  When packets
  are getting lost it shrinks.  This new parameter added here lets you control how small you will let it shrink in response
  to bad connection conditions.  The UdpLibrary will never let the congestion-window shrink below the size of a single
  physical packet (512 bytes by default), so all values for this setting below maxRawPacketSize are meaningless.  Instead,
  the purpose of this setting is to allow you to set this minimum higher than that.  Setting this number higher can result
  in the connection getting flooded beyond its capacity, as it may not be able to slow itself down enough to meet
  current line conditions.
  
  A situation where you may want to toy with this number is in a game (surprise), where you expect to be the only 
  process using bandwidth in the system, and you don't really care to be 'friendly' to other connections running 
  on the system.  Normally, if a modem-user is playing the game and has an TCP transfer going on in the background,
  then the TCP connection and the UdpLibrary connection will both effectively settle-in on using about half the available
  bandwidth.  That is, the flow-control algorithms used by the UdpLibrary and TCP are such that over a fairly int16_t period
  of time, they will both settle in on a happy split.  However, if half-the-bandwidth was equal to a 2k window-size, and
  the UdpLibrary was configured to have a minimum window size of 3k, then the UdpLibrary would not scale down.  Because
  the TCP connection is willing to scale down further, it would, and you would effectively end up with a 3k/1k split
  with TCP kindly taking a backseat to the UdpLibrary (of course, if you had put two 3k-configured UdpLibrary connections
  in this same environment, the net result would be massive/wasteful packetloss).

  The advantage of setting a higher minimum is just the scenario described above, you can effectively configure your
  application to be a bandwidth pig at the expense of other connections in the system; like I said, sometimes this is
  desireable.  Another advantage of having a larger minimum window size is that it allows the connection to more
  quickly scale-up to optimal speed in situations where packets are lost, or there are long stalls in the outgoing
  packet stream.  So, if the game is plugging away at 500 bytes per second and has a window size of a minimum of 3k, then
  when a sudden burst of data transfer is needed for something, it can burst a good chunk of it all at once and not
  wait for the flow control window to scale up via the slow-start algorithm (which really should be called the fast-start
  algorithm).

  Another case where you might want to set this higher, is if you know for a fact that the connection is always going
  to be very high quality (internal ethernet).  In those situation, you may want to set the minimum window size to 30k
  or so, again, such that transfers get up to full-speed quicker.

  Basically, you can look at this value as the guaranteed bandwidth that you want the connection to claim.  The effects of
  playing with this setting are going to be very difficult to quantify, but if anybody does decide to try playing with it
  and notices an appreciable difference (good or bad), I would appreciate hearing about it, just so I know when others ask
  me about it.

- Added new helper function UdpMisc::FindAvailablePort(int startPortScan, int endPortScan), that will find you a randomly
  available port in the specified range, or return 0 if no available port was found.  This can be used in situations where
  you want the server to be listening on a random port in a particular range.  Sometimes it is nice to limit the random
  port selection to a particular range as it allows easier configuration of firewalls.  It's worth noting that when it
  finds an available port, it does not reserve it, but rather releases it and returns which one it is.  It is entirely
  possible that the port could be claimed by somebody else before you effectively bind to it.  A better way of handling
  this situation would be to actually create UdpManager's giving them a random listenPort in the range and then checking
  the UdpManager::GetErrorCondition to see if it successfully bound and try again if not.  In this way, you don't have to
  worry about your port getting stolen.  This helper function was added merely to handle a situation that was unique to
  Everquest where we did not have control of the underlying communications library in order to add this functionality.

- API CHANGE: UdpManager::GetHostByName has been moved and made a static function of UdpMisc.  This allows you to do this
  function without creating a UdpManager.

- API CHANGE: UdpConnection::Disconnect no longer gives the application the option of not being notified via OnTerminated for
  the disconnect event.  The application can only specify a flushTimeout.  When the application does specify a flushTimeout
  the OnTerminated callback does not occur until the actual connection is fully disconnected (by either timing out on the flush
  time, or successfully sending all the pending data and terminating).  While the connection is in a cStatusDisconnectPending
  state, it is possible for some other event (like an ICMP error for example) to cause the connection to switch to a
  cStatusDisconnected state before all pending data is sent.

- API CHANGE: OnTerminated callback will now be called ANYTIME a UdpConnection object switches to a cStatusDisconnected state,
  regardless of the reason.  Once a UdpConnection is in a cStatusDisconnected state, it will always remain as such.  Since
  newly created UdpConnection objects start out in a non-disconnected state (they are either connected or negotiating to start
  with), all UdpConnection objects created are guaranteed to eventually receive one and only-one OnTerminated callback.  In 
  previous versions of the library, a connection that was still negotiating (trying to connect) would not call OnTerminated
  if it failed, now it will.  In this manner, it will be easier and more clear for an application to maintain a strict balance
  between connection creation and destruction.  Furthermore, connections that are explicitly terminated by an application
  call to Disconnect will now callback via OnTerminated (if necessary).

- API CHANGE: OnConnectComplete callback is now only called when a connection is successfully established (as initiated by an
  EstablishConnection call).  Previously, any negotiating connection would callback via this function when it left the
  negotiating state (either succeed or fail).

- API CHANGE: moved user-supplied encryption callbacks from the UdpConnectionHandler to the UdpManagerHandler.  If you are
  using the user-supplied encryption callbacks, then you will likely need to make some minor changes to your code, by creating
  a handler object for the manager and moving the callback routines to there.  This should in no way effect the type of
  encryption you can do as you will still have access to the UdpConnection object.  The purpose of this change was to deal
  with a bug whereby the encryption routines could have been called on before the application had an opportunity to set
  a connection-specific handler.

- made API and library more const safe.  There are likely a few other places where I could have used const that I did not, I
  am sure they will all get tracked down eventually.

- fixed bug where in theory an infinite loop situation could occur.  This could have happened in a situation where one
  connection was bumping another connection to the top of the priority queue and vice-versa during priority-queue processing.
  This is the second time this issue has come up, so this time I put in a more comprehensive fix.  Now, the priority queue
  will refuse to let any connection schedule itself for processing time sooner than 1 ms after the current processing cutoff
  line.  This will also help to deal with issues where the clock might go backwards (though in theory that won't happen now
  either).

- documented the UdpHandler.hpp file more thoroughly to describe what the various callback functions do.

- fixed a bug where the congestion-window used for flow control was able to shrink down to zero when certain rare packet-loss
  patterns occurred.  It turns out that if you are under heavy flow and the timing mechanism used by the library goes
  backwards it can cause this to happen.  This only happened on Linux which uses the real time clock for timing, and the system
  clock was changed while the program was running.  The fix prevents the congestion-window from dropping below the size of
  a single packet.  (Thanks Thomas Farthing for help in finding this one)

- changed the Linux version of the UdpMisc::Clock function to compensate for a backwards moving clock caused by the system
  time being changed mid-application.  There is still and issue with time jumping forward when the clock is changed, but
  that will tend to cause less problems.  If anybody knows a relatively portable way of getting millisecond accurate timing
  that isn't relative to the real-time clock under Linux (something equivalent to Windows GetTickCount), let me know.
  (copied Vince Harron's time-correction routine for this purpose).

- changed the Windows version of the Clock function to better handle clock-wrap in situations where there are multiple
  threads in the process calling the function at the same time.  (used thread-local-storage to solve the issue)

- enhanced user-supplied encryption/decryption callback functions such that if they return -1 it is treated as an
  error-condition and the packet is thrown away by the UdpLibrary.  This could occur if the application detected during 
  decryption that something wasn't right.  Perhaps somebody tampered with the packet, or it got corrupted on-route yet 
  still managed to make it through the UDP header checksum and any UdpLibrary CRC checks that might have been in place.

- fixed bug in explicit bind where it was getting the network byte-ordering wrong (thanks Justin Randall for finding this one).

- fixed a bug where the OnTerminated callback was not being called in situations where an existing connection was terminated
  due to a subsequent new connect-request coming from the same IP and port.  Basically, the InternalDisconnect function was
  being called and told not to bother notifying the application of the event...I have no idea why I would have told it not
  to notify the application of termination for this reason, but it is now fixed.

- added direct winsock2 support.  If you want the UdpLibrary to use version winsock v2.2, then you should define the 
  symbol UDPLIBRARY_WINSOCK2.  This will cause it to use the winsock2.h header files and request version 2.2 of winsock
  on WSAStartup.  Since the library itself doesn't make use of any winsock2-only features, it really doesn't buy you much
  to do this.  The exception is if other parts of your program are using winsock2 functionality for some reason, you may
  have to define this in order to successfully compile, since the winsock.h and winsock2.h headers conflict.

- modified the hashtable template class and gave it the ability to iterate through its contents with WalkFirst and WalkNext.
  This feature isn't actually used by the UdpLibrary, but some other stuff I am working on that shares that hash table object
  needed it.

- fixed some code issues that were causing level 4 warnings (all future releases will be checked against level 4 warnings now)

- fixed problem where NULL was being implicitly converted to a UdpIpAddress inside of GetHostByName


-----------------------------------------------------------------------------------------------------------------
11/26/2002
-----------------------------------------------------------------------------------------------------------------
- thanks to Ben Hulse for porting the library to Solaris/sparc.  His changes have been rolled into the master distribution.
  The ICMP error processing on this platform is done in yet another way from the other platforms.  When you attempt
  to send data to a destination that has previously returned and ICMP unreachable error, the sendto function fails.  Given
  that we use the same port for multiple destinations and that hundreds of outgoing packets to other connections could have
  occurred before we send to the one that returned the error, I suspect this feature may not be working under Solaris.  I
  have seem some references on the web that indicate that UDP sockets do not receive ICMP errors unless they are connected
  (which obviously we can't do).  I didn't immediately see a way that this could work; though, I think it would be worth
  trying to do it the basically the same way Windows does on the recvfrom call.

- we have discovered that ICMP errors are not being forwarded up to the application under Windows NT and Windows 98.  These
  operating systems echo back ICMP errors as appropriate, but applications running on these platforms will not receive
  them on the other end.  Windows 2000/XP do not appear to have this problem.  Our NT testing was using service-pack 6.

- fixed bug in replyUnreachableConnection processing.  The net result is that the library should more accurately
  detect when a connection has gone dead in situations where the server has shutdown and restarted again fairly quickly.

- optimized the code dealing with removing a connection from the manager.  In previous versions of the library this required
  a linear search through all connections as they were tracked with a single linked list.  They are now tracked with a
  double linked list such that they can be removed without the search.  This change gets rid of the last linear operation in
  the library relative to the number of the connections (in normal processing).

- removed the UDPLIBRARY_SAME_ENDIAN flag.  This turned out to be a source of confusion and in practice probably should
  never have been used anyways.  It was merely an optimization that allowed the integer packing routine to operate more
  quickly in situation where both ends of the connection were guaranteed to be the same byte-order platform.

- added function UdpManager::GetHostByName, which does a DNS lookup to translate a name into an IP address.  Even though
  functions such as EstablishConnection will automatically do the DNS lookup if necessary, it is recommended that the 
  application do this translation themselves on startup if the architecture is such that the connection might be established 
  multiple times.  The reason for this is the lookup is a blocking call and it saves having to do it multiple times potentially.

- added support for binding the socket used to a particular IP address (UdpManager::Params::bindIpAddress)
  Normally, and by default, the library will bind the socket to any address in the machine.  This setting should
  not be messed with unless you really know what you are doing.  In multi-homed machines it is generally NOT
  necessary to bind to a particular address, even if there are firewall issues involved, and even if you want
  to limit traffic to a particular network (firewalls do a better job of serving that purpose).  If you are having
  problems communicating with a server on a multi-homed machine and think this might solve the problem, think again.
  You most likely need to configure the OS to route data appropriately, or make sure that internal network clients
  are connecting to the machine via the machines internal IP address (or vice versa).  The only reason I have come up
  with to bind to a specific IP is for security reasons above and beyond those that would normally be provided by
  a firewall.

- added configuration parameter UdpManager::Params::processIcmpErrors to turn on/off ICMP error processing.  By default
  ICMP error processing will be enabled.  If you are having problems with false ICMP errors causing disconnections, then
  you may want to disable this option.  False ICMP errors can occur when a router along the way momentarily goes down
  but is traffic is quickly re-routed, such that termination of the connection wasn't really necessary (at least, that's
  my theory).

- added configuration parameter UdpManager::Params::connectAttemptDelay.  During connection negotiation, the client will
  periodically send the server connect-attempt packets until the server acknowledges the request for a connection.  This
  value represents how frequently the client sends the connect-attempt packets.  By default this is set to 1000 and as a 
  general rule, should not be changed.  The reason this configuration parameter was added was such that I could slow down
  the attempt-rate used by clients of the EQ channel-server.  At peak, the channel server will have upwards of 90,000
  simultaneous connections.  Additionally, there will be up to 2000 people who are attempting to connect but cannot (most
  likely due to a firewall issue on their end).  These 2000 people will continously try to connect for 30 seconds, sending
  the channel-server a packet once per second, resulting in an extra 2000 packets/second coming into the channel server from
  people who will never be able to form a connection.  Since packetloss is relatively rare and rapid connection establishment
  is not critical, I wanted to slow down the rate at which these connection-attempts flooded the server.

- included <stdio.h> to get the definition of NULL.

- added carriage return to end of some header files to prevent warnings from gcc compiler

- removed an assert that could happen in normal processing after all.  Turns out that some ICMP error packets will show
  as coming from port 0 (presumeably when the entire IP address is not reachable?).  We have no way of doing an IP-only
  base connection query without looping through every connection, which we really don't want to do, so for now we are
  ignoring ICMP error packets that have a port of 0.  The appropriate behavior would probably be to terminate all
  connections that came from that IP address.  This doesn't appear to happen very often though; we had the EQ Channel Server
  running for 3 hours with 80,000 connections before it happened for the first time.  ICMP processing to accellerate
  disconnection recognition is really a bit of a luxury, so I don't feel too bad ignoring this rare case.


-----------------------------------------------------------------------------------------------------------------
10/3/2002	
-----------------------------------------------------------------------------------------------------------------
- added new function UdpConnection::IncomingBytesLastSecond(), which returns the number of raw bytes the connection object
  has received in the last second (accurate to within 25ms).

- made sure ScheduleTimeNow was never executed during a GiveTime call.  If it was, then the net result would potentially be
  an immediate re-giving of time again, which I am thinking could possible create an infinite processing loop.

- added new function UdpManager::GetErrorCondition(), which returns an enum ErrorCondition value (see header file).  The
  only error conditions that it currently checks for is the ability to create and bind the socket.  If you are using a fixed
  port and are concerned that somebody else (like another copy of the program running) might already have it, then you should
  check the error condition after constructing the UdpManager object.


-----------------------------------------------------------------------------------------------------------------
9/18/2002
-----------------------------------------------------------------------------------------------------------------
- fixed problem where it was unable to establish a new connection because the server-side thought it already had a connection.
  This actually would not happen except under very unique circumstances.  In particular, you had to establish a connection
  to the server, then, using the same UdpManager object, establish another connection to the same server.  It is illegal
  to have multiple connections between the same two UdpManagers, as the resultant UdpConnection objects would effectively
  have the same ip/port combo's, effectively making it impossible for the UdpManager to know who incoming data was coming
  from.  The problem arose when the client had actually terminated one of the connections, but the server was unaware of it.
  Then, the next conenct-attempt would simply be ignored since the server would think it already had a connection to that
  ip/port.  You would think that the server would eventually timeout the old connection object, but the problem was that
  the new connect-attempts (that were being ignored) were actually refreshing the no-data-timeout on the existing connection.
  This effectively put it in a state where the client could never re-establish the connection.

  This would actually only occur if the second-connection attempt was using the same UdpManager as it had for the first
  connection, or if the UdpManager it was using had been assigned the same port-number again (either by explicit override,
  or by dumb-luck).  The fix for this is to have the existing connection terminate if it gets a new connect-attempt that
  is using a different encrypt-code.  When this set of circumstances happens and is the cause for a connection to be
  terminated, the disconnect-reason will be set to cDisconnectReasonNewConnectionAttempt.

  On a side note, since clients generally only have one outgoing connection for a given manager, it is generally a good
  idea to have the UdpManager created and destroyed along with the connection. Each time the UdpManager is re-created, it
  picks a new port to use, which effectively eliminates these issues.  This is also a good reason not to explicitly
  set the Params::port on client side connections, but instead leave it at 0, so the library will let the OS pick a random
  port (that presumeably hasn't been used recently).

  The situation where this bug first appeared was one where the client was actually using the same UdpManager to connect
  to dozens of servers simultaneously, so re-making the UdpManager wasn't really an option, since all the other connections
  would still be operating.

- got rid of C-style callback functions.  If you are currently using this style of callback you will need to convert
  your code over to using handlers.  I don't think anybody was using user-supplied encryption/compression routines
  yet, but those have been removed as C-style callbacks as well and added to the UdpConnectionHandler class.  See the
  UdpLibrary.doc file or the udpserver.cpp sample code for a description on how to use handler based callbacks.  A quick
  and dirty way of converting an existing C-style callback implementation into a handler based one would be to use the
  following class:

	class MyHandler : public UdpManagerHandler, public UdpConnectionHandler
	{
		public:
			inline virtual void OnConnectRequest(UdpConnection *con) { AppConnectRequest(NULL, con); }
			inline virtual void OnRoutePacket(UdpConnection *con, const uchar *data, int dataLen) { AppRoutePacket(NULL, con, data, dataLen); }
			inline virtual void OnConnectComplete(UdpConnection *con) { AppConnectComplete(NULL, con); }
			inline virtual void OnTerminated(UdpConnection *con) { AppTerminated(NULL, con); }
	};

  Where the AppXXXXXXXXX functions are functions you had previously passed in as callback functions in the UdpManager::Params.
  The only thing lost is there is not a UdpManager mainPassThrough pointer anymore (first parameter to the ConnectRequest 
  callback in the old scheme), but odds are it wasn't used anyways.  Finally, you need to create an instance of the above
  class and set it up as the UdpManager::Params::handler for the UdpManager, and call UdpConnection::SetHandler(...) for
  each newly created connection as well (just after either EstablishConnection or on the server-side in the OnConnectRequest
  function).

  The above dummy-handler is primarily intended to handle converting existing code that uses the old method. New code
  should generally have one handler object for the UdpManager and a seperate handler object for each UdpConnection created.
  Again, see the udpserver.cpp code for an example of this should be set up.

- when reliable data is sent, it remains in the queue until it is acknowledged by the other side.  When you query
  the reliable channel status, you can find out the age of the oldest piece of data in the queue that has been sent
  yet has not been acknowledged yet.  As a general rule, if things are operating correctly, it should be very rare
  for something that has been sent to not be acknowledged within a few seconds, even if resending had to occur.
  Eventually, the sender could use this statistic to determine that the other side is no longer talking and terminate
  the connection.  In past version of the UdpLibrary, the sending-application has checked this statistic itself.  A new
  parameter has been added UdpManager::Params::oldestUnacknowledgedTimeout. This parameter will cause the UdpLibrary 
  to monitor this for you and automatically change the connection to a disconnected state when the value goes over 
  this setting (in milliseconds) (0 means do not perform this check at all)  The default is set fairly liberally, 
  on client-side connections, you could safely set this to as low as 30 seconds (30000) allowing for quicker realization 
  of lost connections.  Often times the connection will realize it is dead much quicker for other reasons.
  When disconnected due to this, the disconnect reason is set to cDisconnectReasonUnacknowledgedTimeout

- added UdpManager::Params::replyUnreachableConnection.  When this parameter is set to true (the default), it causes the 
  UdpManager to reply back to destinations with unreachable-connection error packets when the server does not have a 
  UdpConnection object representing the destination.  The purpose of this is to allow a client who has had its connection
  terminated by the server to quickly realize that the server no longer considers the connection valid.  Normally the
  client would be told when the server terminated the connection, but that termination notice can get lost leaving a client
  thinking it is connected for longer than it should.  See the header file for this new parameter for more details.  
  (note: the client can sort of sense this is occurring now by monitoring the oldestUnacknowledgedAge in the reliable
  channel status, this just makes it occur quickly instead of having to time-out).


-----------------------------------------------------------------------------------------------------------------
8/26/2002
-----------------------------------------------------------------------------------------------------------------
- when I fixing the bug below, I also took the liberty of having the UdpConnection object hold a reference to the
  UdpManager during processing, such that it would handle the case where the UdpManager was deleted during a callback
  as well.  The problem is, I made the same mistake that caused me to mess with the code in the first place, that is,
  referencing the mUdpManager pointer after a callback.  But the release still needs to take place, even if we lose
  our link to the UdpManager due to a disconnect, so to fix it, I have the UdpConnection object hold its reference to
  the UdpManager in a stack variable (effectively).


-----------------------------------------------------------------------------------------------------------------
8/23/2002
-----------------------------------------------------------------------------------------------------------------
- fixed bug where during a UdpConnection::GiveTime call, the mUdpManager member could become NULL (if the application
  did a hard-disconnect during a callback function), and the GiveTime function later on used the mUdpManager pointer
  without checking it.  I think in practice this would have only happened if the application did a hard Disconnect call
  during the processing of an incoming reliable application packet.  Whatever the case, I have went through the code
  and verified that mUdpManager is checked for NULL everywhere it could possibly be used after a callback.

- added the ability to modify the keep-alive delay on a per-connection basis (see UdpConnection::SetKeepAliveDelay)

- fixed some compiler warnings/errors for the Linux compile (I don't actually compile the Linux version, so I count on
  those who do to tell me when I break things :).


-----------------------------------------------------------------------------------------------------------------
8/19/2002
-----------------------------------------------------------------------------------------------------------------
- added a UdpConnection::GetDisconnectReason function that returns the reason that a connection was terminated.

- fixed bug where it wasn't tracking the duplicate-packet-received statistic correctly

- added two new parameters for configuring reliable-channels.  I now allow you to set the resendDelayAdjust and the 
  resendDelayPercent.  These two values allow you to fine-tune the resending algorithm to make the channel either more or
  less aggressive about correcting lost packets.  On some channels, you may be willing to unnecessarily send the packet twice
  to ensure that it gets there as soon as possible; whereas on other channels that are less critical, you may be willing
  to tolerate a slower correction time to avoid sending something you might not need to.  See header-file for details on these
  parameter values.

- changed several of the UdpMisc functions to take void* instead of uchar*.  This was to make it easier for applications to
  use these helper functions to hand-pack packets in order to be byte-ordering and structure-packing safe.

- added string-hashing functions to the hashtable.hpp file, since I have found myself using the hash table object in a lot
  of places for strings and it seemed generally useful.

- added the ability to force the UdpManager to pre-allocate a certain portion of the packet pool.  You might want to do this
  if you are very concerned about memory fragmentation and you know for sure that you are going to be using a certain minimum
  number of packets from the pool anyways.  The new parameter is UdpManager::Params::pooledPacketInitial.

- modified the ReliableChannel object to get temporary packets from the pool instead of allocating them.  It uses temporary
  LogicalPackets to hold onto incoming data that arrives out of order.  Again, if you are not configuring the UdpManager to
  have a pool and having your application use it as well, you are doing yourself a huge disservice.

- changed the logical-packet queue to expand itself 64 packet-pointers at a time instead of 1024 at a time.  This effectively
  reduces the minimum connection footprint by almost 4k at the cost of having to expand this buffer on the fly if the number
  of queued logical packets ever exceeds 64.  Most of the time this expansion will not have to occur, if I find otherwise in
  some application, I will make the expansion rate of this queue a setting instead of hardcoded.

- added a slightly more efficient, though not quite as effective non-buffer based Xor encryption method.  I ended up adding
  this as I am trying to avoid the 512 bytes/connection overhead of the XorBuffer method.  This should be good enough for
  most situations, we are just trying to make the packet-stream look like work to hack to the casual hacker.  This small amount
  of memory saved actually turns out to be critical for the application I am doing as I am trying to get 200,000 players
  connected to a single server.


-----------------------------------------------------------------------------------------------------------------
7/1/2002
-----------------------------------------------------------------------------------------------------------------
- added a new 'interations' statistic to the UdpManagerStatistics.  This represents how many times UdpManager::GiveTime has
  been called.  Divide this by elapsedTime to get an average number of times per second GiveTime is called.  This should
  generally be at least 10 or else you are likely creating too much processing-induced-lag (as I call it).  A similar statistic
  has been added to UdpConnectionStatistics

- changed code to ignore ICMP port unreachable errors while in cStatusNegotiating mode. This was necessary in order to prevent
  a client who happened to startup moments before the server did from aborting it's connection attempt because the port wasn't
  open yet when the first packet went out.

- added a new UdpManager::Params::portAliveDelay parameter which is intended to be used on the client side to keep the NAT or
  firewall port alive without having to actually send the server unnecessary keep alive packets.  This is very similar to the 
  keep alive delay, but serves a completely different purpose and may not have the desired effect on all platforms.  The purpose
  of the keepAliveDelay is generally to keep data flowing to the server so it knows that the client is still there.  In this 
  manner if the server doesn't get data within a certain period of time, it can know that the connection is probably dead.
  Sometimes it is the case that the server does not need to be kept alive, or at least kept alive very often (like for a chat 
  server perhaps where nobody is talking much).  For some people, it may be necessary to send data more frequently in order to 
  keep their NAT mapping fresh, or their firewall software happy.  However, we don't want to be in a situation where our server 
  is receiving a lot more data than it needs to just so these people can keep their port open.  I have seen NAT's that lose 
  mappings in as int16_t as 20 seconds.  What this feature does is a bit tricky.  It changes the time-to-live (TTL) for a special 
  keep-alive packet to some small value (5) which is enough for the packet to get past local firewalls and NAT's, but not make 
  it all the way to our server.  In this manner, the port gets kept alive, but we don't waste bandwidth with these packets.
  These special packets are not counted statistically in any way and they do not reset any timers of any kind.  Their sole 
  purpose is to keep the port alive on the client side.  Any other data (including standard keepAlive packets will reset the 
  timer for this packet, so obviously the portAliveDelay must be smaller than the keepAliveDelay in order to be meaningful.
  By default this feature is off.  It's worth noting that the client-side will receive-back an ICMP TTL-expired packet; however,
  the client will simply ignore it.  It is unknown as to whether all NAT's will ignore the ICMP return packet, it is assumed
  that they will.  You should not set this value too aggressively as the returning ICMP packets will be taking up valuable
  incoming bandwidth on the client side, even though they have no effect on our end at all.

  (update: we put this live on EQ and it fixed the problem for a lot of people; however, it also caused a different small
  set of players to start having problems...probably firewalls or NATs that interpret the ICMP error packet as a reason to
  shut down the port/mapping.  Anyways, I am not sure it is a viable option because of this.  You could make it an advanced
  user option, but clearly it isn't the fool-proof solution we had hoped it would be.)

- added an optional parameter to UdpConnection::Disconnect as to whether it should callback the application via OnTerminated
  to notify of it of this Disconnect call.  In either case, the callback would only occur if connection had previously
  been in a cStatusConnected state.  This api enhancement was primarily added for internal purposes, though it will
  allow the application to receive OnTerminated callbacks when it explicitly calls Disconnect itself (if it so desires)

- added the ability for the library to auto-terminate connections that are not receiving data.  This feature is supported
  via the UdpManager::Params::noDataTimeout setting.  The timeout can also be overridden/controlled on a per-connection
  basis via the UdpConnection::SetNoDataTimeout and UdpConnection::GetNoDataTimeout functions.  Setting the timeout to
  zero (the default) causes the connections to never auto-disconnect due to not receiving data.

- added UdpManager::GetHandler and UdpConnection::GetHandler calls (added for completeness, not currently used by anybody)

- some minor changes to the non-windows version so it compiled cleaner.

- made it so keep-alive packets do not cause UdpConnection objects to be rescheduled in the priority queue.  They really
  don't need to reschedule things since the worst thing that can happen is it might get processing time sooner than it
  might otherwise need to.  But scheduling it for immediate processing time just makes that worse if you think about it.


-----------------------------------------------------------------------------------------------------------------
5/15/2002
-----------------------------------------------------------------------------------------------------------------
- added the ability to have a Disconnect'ed connection continue processing until there is no more data outstanding
  on any of the reliable channels.  The UdpConnection::Disconnect function now takes a 'flushTimeout' value, which is
  how long (in milliseconds) the connection will attempt to finish sending queued reliable data before giving up.  If
  the application specifies a flushTimeout of zero (the default), then the connection is immediately terminated as before.

  When the application specifies a timeout, the connection status is changed to cStatusDisconnectPending.  From the
  applications point of view the connection should be thought of as disconnected the moment the call is made.  The
  connection will never callback the application again to deliver data and it will not allow the application to Send
  any more data on the connection.  About all the application can with the connection is query its various status', or
  force an immediate disconnect by calling Disconnect again giving it no flush timeout.

  It is generally thought that the application will call Disconnect with a timeout and then immediately Release the
  connection and have nothing else to do with it again.  This is how it is intended to work.  The Release by the application
  will not actually free the connection if there is still data pending to be sent.
  
  Internally how this works is when the application calls Disconnect with a timeout, the status of the connection
  is changed to cStatusDisconnectPending.  The connection object registers itself with the UdpManager telling it 
  to keep a reference to the connection until the connection status changes to cStatusDisconnected.  The connection will 
  be given scheduled time by the UdpManager as before and when the connection has no more reliable data to send, or when
  its 'flushTime' expires, the connection will change itself to a cStatusDisconnected state, at which point the UdpManager
  will 'Release' its reference to the connection and the connection will actually be deleted.  So, take the following
  code for example:

	void ApplicationFunction()
	{
		myConnection->Send(cUdpChannelReliable, data, 100000);	// send 100,000 byte packet
		myConnection->Disconnect(30000);		// disconnect us, but give it up to 30 seconds to get any pending reliable data sent
		myConnection->Release();				// we are finished with this connection
		myConnection = NULL;
	}

  The above code will not block and will deliver the 100k packet if it can manage to do it in the 30 second timeframe.
  Releasing the UdpManager while there is pending disconnects will result in everybody being disconnected immediately
  and the data will not be delivered.

- added a new function UdpConnection::TotalPendingBytes(), which returns the total number of reliable-bytes that
  are outstanding.  When this returns zero, you can be assured that all data on all reliable channels has been successfully
  delivered.  (note: this is same as adding up ChannelStatus::totalPendingBytes on all reliable channels)

- added a GetStatistics function to the internally used hash-table object (in case anybody is using it externally for their
  own purposes.  The UdpLibrary itself does not make use of the statistics.


-----------------------------------------------------------------------------------------------------------------
5/9/2002
-----------------------------------------------------------------------------------------------------------------
- seeded the CRC algorithm with the negotiated encryption key, such that the CRC values calculated for packets cannot
  be easily calculated by off the shelf CRC algorithms (to prevent packet tampering).  This gives us some level of
  protection without the huge disadvantage of encryption which is 1) potentially slow, and 2) encrypting of a typical
  data stream reduces how effective modems hardware compress.  In Infantry, we found turning encryption on (where the
  data did not change size) significantly lagged modem users.  For this reason, encryption should likely only be used
  in conjunction with compression.  General byte-level compression itself is probably sufficient encryption for out needs.

- added OnCrcReject handler function.  This function is only available if a handler object is used (there is no callback
  equivalent right now...if somebody wants one, let me know; but, I would encourage everybody to switch over to handler based
  callbacks instead).  Whenever a packet is rejected due to a CRC error, this function is called to report the incident to
  the application.  Corrupt packets are fairly rare as they must first get through the 16 bit checksum that the UDP packet
  inherently puts on it.  If it makes it through that, then our CRC check will probably catch it.  The reason why an
  application would want to know this, is because there is a very good chance that the corruption was caused by somebody
  intentionally trying to tamper with the packet data, but not touching up the CRC on the end.  The application may want
  to take note of the IP address and if too many corrupt packets come from the same IP, consider banning the player for
  cheating.

- made manager check the protocol version for compatibility before creating an associated UdpConnection object.  This prevents
  the application from seeing connections that may later be rejected because of protocol version incompatibility.

- tweaked the flow control algorithm a bit.  The two main changes were 1) selective acks now adjust the congestion window
  differently similar to the TCP Reno algorithm, and 2) The slow start threshhold was reset when the pipe went dead (before
  we only reset the congestion window, but this resulted in a circumstance whereby the slow-start threshhold once low would
  never be able to go back up.

- modified the resend-timer to not take into account the amount of outstanding data in the window.  It turns out that
  this adjustments is redundant in that the slow-start flow control mechanisms will ramp things up to speed in a manner
  such that the average ack time will grow longer as the window gets bigger as appropriate.

- Added a new parameter to the reliable channel configuration.  ReliableConfig::processOnSend if 'true' will configure that 
  particular reliable channel to do processing on sent packets at the time they are sent instead of waiting until the next
  UdpManager::GiveTime call.  Normally the application will want to leave this at its default value of 'false' because it
  is more efficient on CPU usage, since it can avoid doing all the ReliableChannel::GiveTime processing on a per-sent packet
  basis by delaying it until the next global GiveTime.  The downside of delaying is that it is often the case that the next
  GiveTime simply causes the reliable packet to get moved into the multi-buffer inside the UdpConnection object.  Then another
  GiveTime a bit later finally sends it.
  
  Setting this value to 'true' is similar to issuing a FlushChannels command after each reliable Send (with the 
  exception that rather than giving every channel and the UdpConnection object time, it only has to give the channel
  doing the send time).

  This feature is most useful when you are on the client side and the overhead of processing the packet on-send 
  will not add up to much, compared to what the server-side has to do managing thousands of connections.  Most clients
  should probably set this to 'true', since timeliness of sent packets generally exceeds any need for code efficiency on
  the client side in this regard.

  It's worth noting that the reliable channel object ends up scheduling time for the connection the next GiveTime, 
  regardless of how this value is set.

  If you have the multi-buffer set to 0 (no multibuffering) and you have this parameter set to 'true', then the actual
  send of a physical packet will end up occuring during your reliable-send call (assuming there is room in the reliable
  flow control window for the packet to go out at that time).  As a reminder, I don't recommend setting the multibuffer
  to zero, even on the client side, because it effectively circumvents the mechanism by which ACK packets get combined
  and piggy-backed on other packets (ie. every ACK would end up taking its own physical packet).  A better approach I feel
  is to have a multibuffer and then do explicit FlushMultiBuffer calls after sends.  In this manner, ACK's still have some
  chance of piggy-backing as they will wait around until the next send call (or the multibuffer timeout).

- fixed a tiny problem whereby packets that were rejected due to CRC mismatches were not causing the connection receiving
  them to schedule itself for processing time as it probably should.  In practice this didn't really effect anything.

- added a (UdpConnection *) parameter to the UdpConnectionHandler callback functions.  This was done such that an application
  could setup one object as the handler for multiple connections and be able to distinguish between which connection the
  callback was coming from.  You will have to change your handler functions appropriately to make them compile, and odds
  are you will be able to simply ignore the incoming UdpConnection parameter.

- added UdpManager::GetPassThroughData and UdpManager::SetPassThroughData calls to allow application to change the pass through
  data after construction (makes it more consistent with the way UdpConnection objects work as well)

- made it negotiate the maxRawPacketSize between the client and server side during connection establishment.  It will actually
  end up using the smaller of the two values specified on the client or server side.  In this manner, the server can be
  configured to support 1k packets (for example) and some client could be configured to only support 512 byte packets and they
  would still be able to talk to each other (they would end up doing all talking with 512 byte packets in this case).  This 
  will prevent the client and server from accidentally getting configured differently and then refusing to talk to each other.  
  It will also allow clients who are behind networks that can only handle tiny packets to be configured to talk in the smaller
  chunks to support that.  You will have to update both ends of the connection with this code in order for them to talk, since
  this change effected the initial negotiation process.


-----------------------------------------------------------------------------------------------------------------
4/12/2002
-----------------------------------------------------------------------------------------------------------------
- Removed the Handler base class declarations from being nested and put them in their own header file such that we could avoid
  having to expose the entire UdpLibrary header everywhere (along with all it's baggage).  You can now only include UdpHandler.hpp
  to get these declarations, which have now been named UdpConnectionHandler and UdpManagerHandler.

- Removed the nesting of ConnectionStatistics and ManagerStatistics, they are now UdpConnectionStatistics and UdpManagerStatistics.

- Optimized the UdpLibrary.hpp header file to avoid including stuff we really didn't need to.  In particular, with a few trivial
  changes, we managed to be able to avoid including any other header files (including winsock.h).  The only file we could not
  get rid of including is stddef.h, which has the offsetof() macro that we need in the hashtable.hpp template implementation.

- merged latest Linux implementation into main distribution.  Hopefully I didn't change anything in the Linux implementation in
  doing so as I have not tested it.


-----------------------------------------------------------------------------------------------------------------
4/9/2002
-----------------------------------------------------------------------------------------------------------------
- switched UdpManager over to an AddRef/Release scheme as well.  This is to allow the application to release/delete the UdpManager
  during a callback as well.  Internally the UdpManager keeps a reference to itself when it calls back the application, such
  that the actual delete doesn't occur until the UdpManager is safely off the stack.

- added UdpManager::SetHandler function.  Normally the handler for the UdpManager is set in the UdpManager::Params and passed to the
  constructor; however, I wanted to let the application change or get rid of the handler if needed.  In particular, if the application
  has set the handler to a particular object, and that object gets deleted before the manager, then that object should clear itself
  from being the handler for the manager by setting the handler to NULL (presumeably).  In particular, it is probably a good idea
  to get in the habit of having handler objects set and clear themselves on construction and destruction as appropriate, for both
  UdpConnection and UdpManager handlers.  For example, this is the best way to set things up (similar things should be done
  for UdpManagers):

		class Player : public UdpConnection::Handler;

		Player::Player(UdpConnection *con)
		{
			mConnection = con;
			mConnection->AddRef();
			mConnection->SetHandler(this);
		}

		Player::~Player()
		{
			mConnection->SetHandler(NULL);
			mConnection->Disconnect();
			mConnection->Release();
		}

  Clearing the Handler on destruction is probably not necessary, but on the odd chance that somebody else has a reference to
  your object and it lives on, it would seem to be a safe practice.  In particular, during callbacks the connection will actually keep
  a reference to itself momentarily.  The Disconnect call should keep it from every trying to call you back, but clearing the
  handler is a good safety step to be sure.

- fixed bugged array delete in PriorityQueue.

- fixed another bug in the ICMP handling that prevented the application from deleting the connection during that callback.

- changed ICMP processing such that it would not quit trying to poll packets when it got an ICMP error return.  I am not sure if
  Windows avoids sending ICMP error results via recvfrom if there are legit packets in the queue or not, but just in case, this
  change will make it handle the situation slightly better.


-----------------------------------------------------------------------------------------------------------------
4/8/2002	
-----------------------------------------------------------------------------------------------------------------
- added ICMP port-unreachable support.  This will allow the sending end of a connection to quickly determine that the other side
  has terminated servicing its port for some reason.  Again, props to Justin for this idea.  I tested it and it appears to work.

- changed UdpConnection objects to be AddRef/Release based.  This effectively meant we could get rid of the UDPGIFT ownership
  passing paradigm because ownership is now a function of having a reference to it.  This means that when you are called back
  and given a UdpConnection object from a ConnectRequest callback, you must AddRef it if you intend to keep a hold of it; otherwise
  it will be destroyed by the UdpManager when the callback returns.  If you have an existing application you must remember to put
  the AddRef in the callback, otherwise your connection will be destroyed immediately after it is created.  The compiler will give you
  errors for the destructors which need to be changed as noted below.
  
  Similarly, the application should Disconnect() and Release() the UdpConnection object when it is done using it and wants 
  it to go away.  This change was needed such that the application would be free to delete the connection object 
  during packet processing inside a callback.  The UdpManager keeps a reference to the object during packet processing 
  and releases it when done.  The problem was that the callback to the application to route the packet came from deep 
  inside the UdpConnection object which was being destoryed.  Because the UdpConnection object may be breaking apart
  a multi-packet at the time the application releases the UdpConnection object, it will continue to route the remaining parts of the
  multi-packet, as it has no idea that the application no longer wants to receive them.  The application may have released its reference
  to the UdpConnection, but the UdpManager maintains a reference until the raw packet is finished processing.  What this means is that
  if the application wants to ensure that when it releases the UdpConnection object that it also prevents all future callbacks from that
  connection, it must also call UdpConnection::Disconnect before calling Release().  The bottom line is the proper way for the application
  to destroy a connection during packet processing is to call Disconnect() and Release().  A Disconnected UdpConnection object will
  never callback the application to route a packet.

- fixed UdpMisc::Clock such that if different threads were calling it at the same time and got time sliced away, it wouldn't accidentally 
  double-increment or miscalculate in some way the next stamp.

- Justin Randall modified library to compile properly under Linux and his changes have been included into official distribution.
  The sample client/server programs are still WIN32 specific.


-----------------------------------------------------------------------------------------------------------------
4/2/2002
-----------------------------------------------------------------------------------------------------------------
- made API const safe for most things.

- added support for using object-based handler objects instead of callback functions for notifications.  See the header file
  under Params::handler and UdpConnection::SetHandler for more details on how these work.  The UdpLibrary.doc file has been updated 
  to talk briefly about them as well.  The UdpServer sample application included in the distribution has been modified to use 
  handlers instead of callbacks by default as well.

- changed hashtable template interface slightly.  It now has a simpler FindFirst/FindNext interface, avoiding the need to expose
  the linked-list implementation details.


-----------------------------------------------------------------------------------------------------------------
3/24/2002
-----------------------------------------------------------------------------------------------------------------
- added a pooled memory manager for logical packets to the UdpManager object.  This should virtually eliminate
  allocations for the vast majority of usage if configured properly.  Using the UdpManager::CreatePacket is now the recommended
  method for allocating logical packets to send since it will use the pool automatically if possible.

- implemented function UdpConnection::OutgoingBytesLastSecond (replaces the unimplemented BytesInLastSecond)

- added new FlushChannels function which allows the application to force all channels (particularly reliable ones) to send
  out any queued data that they have time for immediately, instead of waiting for the next time they get processing time.

- added new totalBytesPending statistic to the ChannelStatus, so you could tell for certainty that everything had made it

- tweaked resend-time calculation...I will probably toy with this for a while til I am perfectly happy with it.

- moved clock-sync packet sending ahead of reliable-queue packet processing to better the odds of getting a fast ping.


-----------------------------------------------------------------------------------------------------------------
3/22/2002
-----------------------------------------------------------------------------------------------------------------
- optimized the flow-control window algorithm for our unique situation.  Basically, I changed it so that the window-size does not
  increase when the application is pacing the data rate itself such that the window is not getting filled.  The problem was that
  a slow stream from the application would give the flow-control mechanism the impression that packets were never being lost and it
  would keep increasing it's window size forever.  Then if the application suddenly dumped a huge chunk of data into the pipe, the large
  window size would cause the client to get seriously over-flooded.  This is one area where TCP falls flat on its face as well.
  For more information on this flow-optimization, search for the mMaxxedOutCurrentWindow variable in the implementation and read the 
  associated comments.

- fixed nasty bug where UdpReliableChannel::mLastTimeStampAcknowledged was not being initialized.  This potentially caused the
  first reliable packet sent to think that it had been ack-accelerated by a later packet and resent a second time.  What made the bug
  really nasty is that would resend this packet over and over again very very rapidly until it had been acknowledged.  But because
  the packet had been sent more than once, it would not use that packets last send stamp to reseed the accelleration stamp.  This meant
  that the next reliable packet sent experienced the same problem.  The net effect was a ton of traffic and 80% packetloss to the destination.
  I'm surprised this did not show up more in testing, but it should be fixed now.

- fixed bug where it would crash if the timeout expired on EstablishConnection call (optional feature)
- fixed bug in trickle channels where they were sending faster than the specified trickle rate
- fixed bug in UdpConnection::LastReceive(...) (the version added last rev, accidentally compared the wrong stamp)
- added more ChannelStatus statistics (not real useful, but they have some gee-wiz value)


-----------------------------------------------------------------------------------------------------------------
3/12/2002
-----------------------------------------------------------------------------------------------------------------
- enhanced UdpManager::GiveTime to allow it to be told to process only 1 physical packet out of the socket per call.  See the docs
  for GiveTime for an explanation of when this might be useful.

- added support for reliable-channel specific fragment sizes (as opposed to using maxRawPacketSize for all channels).  See the docs
  for ReliableConfig::fragmentSize for details on when you might want to do this.

- fixed bug whereby a client-side connection could possibly try and process a packet before it received the connection confirmation
  This was a serious problem, but would only happen if the very first communication a client/server did originated from the server
  side and the connection confirmation packet had been lost, and encryption was being used.

- changed UdpMisc::CreateQuickLogicalPacket to allow it to take a NULL primary pointer (ie. just allocate space)

- added a UdpConnection::LastReceive function that takes the current time as a parameter, this saves the connection having
  to poll the clock to accomplish this, which can add up if you are looping through a ton of connections just to check this info

- moved several functions into the header as inline implementations for performance


-----------------------------------------------------------------------------------------------------------------
3/8/2002
-----------------------------------------------------------------------------------------------------------------
- Fixed bug in UdpConnection::GetLocalPort, it was incorrectly returning the port number in network-byte-order
- Added a UdpConnection::GetUdpManager call to get the UdpManager associated with a connection
- initialized UdpConnection's pass-through pointer to NULL
- got rid of UdpManager::Initialize and UdpManager::Terminate.  These functions weren't needed after all since WSAStartup/WSACleanup do ref-counting.
  You can now simply create the UdpManager immediately and it will take care of it.


-----------------------------------------------------------------------------------------------------------------
3/6/2002
-----------------------------------------------------------------------------------------------------------------
- UdpManager::ManagerStatistics::elapsedTime changed to milliseconds from seconds.
- Added bytesSent tracking to UdpConnection::ConnectionStatistics
- changed all byte and packet count tracking to 64-bit values to prevent rollover. (cleaned up stat tracking code internally while at it)
- fixed an off-by-one bug in statistics where it was improperly counting the clock-reflect packet it had just sent when calculating packet-loss statistics
- added a UdpManager::GetLocalPort call to complement the GetLocalIp call (oversight, it should have been there all along).
- documentation continues to be fleshed out (still a lot of sections missing though, UdpManager::GiveTime section is worth reading)


-----------------------------------------------------------------------------------------------------------------
3/5/2002	
-----------------------------------------------------------------------------------------------------------------
- Added two new callback functions, one for when an EstablishConnection
  call either succeeds or fails and one for when the connection is
  is terminated by the other end.  EstablishConnection function used for
  establishing client-side connections now takes a timeout value for how
  long it should attempt to establish the connection.  Both are completely optional
  if the application wishes to have callback notification of these events.

- Fixed compiler incompatibility with MSVC 7.0 (template friend function related)
- Fixed compiler level 4 warnings for unused parameters
- Modified the resend protocol to throttle back resend times when packetloss
  occurs to prevent unnecessary spamming on truely link-dead connections.


-----------------------------------------------------------------------------------------------------------------
2/26/2002
-----------------------------------------------------------------------------------------------------------------
- Initial release of library to EQ2 team (only)
  Documentation not yet finished, see header file comments for full documentation.


-----------------------------------------------------------------------------------------------------------------
Distribution Files
-----------------------------------------------------------
UdpLibrary.cpp			- main source file
UdpLibrary.hpp			- main include file
UdpHandler.hpp			- include needed to be a handler-only
hashtable.hpp			- include file needed internally
priority.hpp			- include file needed internally
udpstress.cpp			- sample application (acts as both client and server)
UdpLibrary.doc			- documentation file
UdpLibraryRelease.txt	- this file
