Get this timeout in to the timer system as well to make things
simpler and more robust.

It is functionality specific to WinVNC, so move the code there
to make things more clear.

Removed the last parts of VNCSConnectionST's back door in to
VNCServerST and let the parent class fully handle coordination of
clients, and access to the desktop.

This makes it generally useful and other code doesn't have to
know of the specific sub-class.

More encapsulation of functions that require coordinate between
clients.

There is some client coordination needed which is better encapsulated
inside VNCServerST. This also helps hiding the desktop from the
individual clients.

Avoid having VNCSConnectionST poking around in VNCServerST's internals
and instead access things via safer methods.

Make things simpler by making this a part of the SDesktop interface
that always needs to be implemented.

Force queryConnection() to always call back to approveConnection()
rather than return special values. This makes the flow easier to
follow as it will be the same in all cases.

If we have plenty of bandwidth then CPU might be the limiting resource.
Unfortunately we don't have a good number for that limit, so add a
conservative hard coded value.

If we're out of time then we will definitely overshoot, so don't even
try.

If an area recently changed then we can guess that it will most likely
change again very soon. In such a case it is meaningless to send a
lossless refresh as it will directly be overwritten. Keep track of
such areas and avoid refreshing them until we no longer see any
changes to them.

There might be parts of the screen that haven't changed and can
therefore be refreshed. Figure which parts these are and send just
those.

We don't want to waste bandwidth on the lossless refresh if we might
need that bandwidth for a normal update. Try to estimate how much
data we can safely send without interfering.

Resend pixel perfect copies of areas that were previously sent
using a lossy encoder. This is done when there is no normal update
to send, and no congestion.

We already have an assignment operator, so no need for this method.

We cannot send updates for a cursor that is outside the requested
region, so make sure we track things properly. This also has the nice
side effect of just re-sending the bits needed when the cursor overlaps
a changed part of the framebuffer.

Allows us to compare our computed congestion window with the
underlying one used by the TCP layer.

Refine the previous method by interpolating the values we need.
This reduces the effect of the problem that we cannot send enough
ping packets.

It's a general function and it's better to have that particular
complexity in its own place.

This reduces the data sent, and avoids any problems with the
client corrupting it.

We shouldn't unconditionally remove the server side cursor just
because the client supports client side cursors. We need to also
check that the client fulfills the other criteria for taking over
the cursor.

In most cases we will send an update at the end of processing client
messages so these extra calls only slow things down. Server initiated
events still need an explicit call though.

Improve the tracking of what cursor we've sent to the client to make
sure the initial cursor is sent properly. We previously tried to infer
this information based on if we've rendered a server side cursor or
not. This logic broke down if things triggered before we've sent the
first update to the client.

This makes it possible to create a derived class from Socket which is
not TCP based, without having VNCSConnectionST.cxx trying to call
setsockopt() on a non-socket.

The logic got inverted resulting in the cursor being sent needlessly
whenever the cursor moved.

It was moved to the common code in f8e3b34c6, but it was unreliable
because the state could sometimes get out of sync. Push it back in
to Xvnc since it isn't necessarily something all servers will have.

Otherwise the client can end up with two visible cursors.

This makes it easier to provide more sane events to the backend code
even with crazy clients.

This adds the basic infrastructure and handshake for the QEMU
Extended Key Events extension. No viewer or server makes use of
the extra functionality yet though.

Based on QEMU's behaviour.

If client sent fence with some data, followed by fence with no data (length 0), the original fence data were freed, but the pointer kept pointing at them. Sending one more fence would attempt to free them again.

This redesigns the old "deferred updates" mechanism in to a frame
clock that governs how often updates are sent out. The goal is still
the same, to aggregate updates and avoid pointless updates, all in
the name of efficiency. This model should however be more robust
against delays that sometimes causes us to miss the desired rate.

This is not the code path commonly used in performance sensitive
scenarios, so favour correctness over performance.

Make it a bit more readable.

Make the control flow a bit clearer.

This will allow us to use better formats that preserve the entire
alpha channel.

There might be stuff lingering in the buffer simply because flush()
hasn't been called in a while, rather than because the transport is
congested.

We now get notifications when the output buffer empties, and we
already caught incoming RTT pongs, meaning we can now react at the
proper time to retry a congested update rather than use a timer.

What to do when a socket is writeable should be handled in the
RFB core code as there may be other events we want to fire off
when this happens.

Makes it easier to troubleshoot keyboard issues by printing exactly
what key events the server is getting from the client.

The bug fix in b64dbf2 didn't account for the proper request
region in the case of continuous updates. Make sure we use the
proper variable for which region we've sent updates for.