This is a regression from ad0f061. If a VMware cursor rect was split up
over multiple read()s then the stream would become corrupted as we set
the restore point at the wrong place.

The generally recommended way is to include it from source files, not
headers. We had a mix of both. Let's try to be consistent and follow the
recommended way.

Can be helpful to see what encodings were actually used during a
connection, and how they performed.

Gives a bit more context where this error happened.

Single CPU machines are extremely rare now, so let's avoid the hassle of
multiple code paths.

The certificates might still be fine using the system trust store, or
the user can make an exception. So let's just log and continue on.

This got very common after 960c7d2 where we now always have a default
value for these settings.

This fixes regression introduced by the extended clipboard extension.
Previously it was possible for the server to hold on to the CLIPBOARD
selection even if another application took ownership of PRIMARY. This is
important to handle the common use case of selecting something in order
to paste over it.

The new request based model doesn't readily support this as we assume
the client has lost its data once we push the new PRIMARY selection to
it. So to handle this we have the maintain a cache of the client's data,
and make sure to fill that cache before we do anything that might cause
the client to lose the data.

The call of gnutls_set_default_priority_append() expects a normal priority
string, which means it must not start with ':'.

Let's avoid making this too complex and force every user to know about
magical functions.

Allows the default value to be dynamically generated in a buffer that is
reused.

Our current architecture doesn't support waiting for a response here, so
don't even try or we'll just get an error.

The session might have died, or failed to initialise properly, so be
prepared for gnutls_bye() to be unable to do its job.

We would leak the security module below the top client VeNCrypt module,
meaning that those modules would not get a chance to shut down
gracefully.

It's perfectly safe to delete NULL pointers, so simplify things by
removing these checks.

CharArray should always be null-terminated. There is a potential
scenario where this all might lead to crash. In Password we call
memset(), passing length of the array we get with strlen(), but
this won't return correct value when the array is not properly
null-terminated.

This change makes it possible for re-synchronizing the remote cursor on
the vncviewer when in fullscreen mode. This is done by locally moving
the cursor position to what the server thinks it should be.

Now SDL games should work!

This change adds support for the VMware Mouse Position
pseudo-encoding[1], which is used to notify VNC clients when X11 clients
call `XWarpPointer()`[2]. This function is called by SDL (and other
similar libraries)  when they detect that the server does not support
native relative motion, like some RFB clients.

With this, RFB clients can choose to adjust the local cursor position
under certain circumstances to match what the server has set. For
instance, if pointer lock has been enabled on the client's machine and
the cursor is not being drawn locally, the local position of the cursor
is irrelevant, so the RFB client can use what the server sends as the
canonical absolute position of the cursor. This ultimately enables the
possibility of games (especially FPS games) to behave how users expect
(if the clients implement the corresponding change).

Part of: #619

1: https://github.com/rfbproto/rfbproto/blob/master/rfbproto.rst#vmware-cursor-position-pseudo-encoding
2: https://tronche.com/gui/x/xlib/input/XWarpPointer.html
3: https://hg.libsdl.org/SDL/file/28e3b60e2131/src/events/SDL_mouse.c#l804

Some of these were incorrectly calculated so the server or client would
wait too long before proceeding with decoding.

Change all of these to be a more explicit calculation to avoid such
issues in the future.

We computed a safe area if a client gave us a bogus one, but we didn't
actually use it. Fix this properly and make sure we don't pass on bad
coordinates further.

So the current clipboard state is properly reflected in the desktop
session.

This was out of sync with the client handling for no good reason.

The peer expects a response, so we should also be able to respond that
there is no clipboard data currently available.

The extended clipboard protocol has the ability for the peer to request
things to be sent automatically, without a request message. Make sure we
honor such settings.

Each character is more than one byte, so adjust the clearing of the
buffer to reflect that.

Some code points are reserved for the UTF-16 coding itself and must not
appear as input data to the algorithm.

Signed bug prevented anything not ASCII from being coded correctly.

Everything outside of BMP was handled incorrectly and was coded as
completely different code points.

This would mess up most conversions from UTF-8 as the caller wouldn't
know how far to step to get to the next valid character, resulting in
markers for invalid data to be injected here and there.

Also add some unit tests to avoid this reoccurring.

It should be using the safe wrappers for everything so make sure it
cannot bypass those and call the SConnection methods directly.

We incorrectly called the underlying functions instead of the safe
wrappers for the new clipboard functions. This had the effect of a)
crashing the entire server if one of these functions failed, and b) not
respecting the settings disabling the clipboard.

64x64 changed block can be large for fine changes such as cursor
movement and typing in terminal windows, or an update to a clock.

If the block can be efficiently cropped, this will reduce latency
and bandwidth. Every pixel cropped is a pixel less to analyze, encode,
transmit, and decode.

The previous code already detected the top of the change in order
to determine if the block had changed. However, it did not use
this information to reduce the size of the change rectangle, nor
did it calculate any of the other edges.

The new code introduces detection of the other edges, and uses
the information to build a reduced area change rectangle. This
has the additional effect of reducing the number of discrete pixel
values in the change block which may allow a more efficient
encoding algorithm to be selected.

As this section of code is performance sensitive, the method
of detecting the edges has been optimized to quickly fall back
to pessimistic values as soon as a single comparison fails on
each edge. In the case that full 64x64 block are changing,
there will be three extra comparisons per block.

In cases where the change rectangle can be reduced from 64x64,
the reduced size of the change rectangle represents reduced
effort to encode, transfer, and decode the contained pixels.

In the case of images with high frequency changes, which
specifically includes text, the lossy JPEG encoding can be
highly distorted, especially with JPEG level 6 or less. The
quick flash from a distorted JPEG to a lossless JPEG can
appear as a flickering to some people. This effect was more
obvious when the surrounding area is not expected to change,
but is being distorted anyways due to being part of the 64x64
blocking algorithm.

In the case of a user typing in a terminal window, this change
may commonly reduce the number of pixels updated with every
character typed from 4096 pixels (64x64) to 640 pixels (32x20)
or less.

Since 8e09912 this wasn't triggered properly as we checked if all
clients were gone before we actually removed the last client from our
list.

The method it overloads got tweaked some time ago, so we need to make
sure this method follows suit.

Sends response for SetDesktopSize as per the community wiki
specification

We'll just crash later if we try to use such a large screen, so reject
the request from the client instead and keep the server running.

The previous method stored the certificates as authorities, meaning that
the owner of that certificate could impersonate any server it wanted
after a client had added an exception.

Handle this more properly by only storing exceptions for specific
hostname/certificate combinations, the same way browsers or SSH does
things.

Major restructuring of how streams work. Neither input nor output
streams are now blocking. This avoids stalling the rest of the client or
server when a peer is slow or unresponsive.

Note that this puts an extra burden on users of streams to make sure
they are allowed to do their work once the underlying transports are
ready (e.g. monitoring fds).

These are not universal in the protocol so having functions for them
only obfuscates things.

The specification only states a single result byte and not any reason
after a TLS authentication failure.

Provide some safety checks when directly accessing the underlying
pointer of streams.

Some systems (like TLS) need to send some final data before closing
a connection. Make sure this is properly handled by cleaning up the
security object before closing the underlying network socket.