But keep the turn connection that's currently being used if it is
the selected candidate. Also clear the TURN candidate refresh.

Instead of just blindling assuming that all of the pointers are valid
until the end.

Change the SocketSource handling to make it possible
to remove a source, not only to add one.

It can now take it from the agent.

Skips one iterator over g_slist_foreach() + g_slist_free(). It is also easier
to read.

It is never used outside of this file.

Youness Alaoui authored Apr 19, 2014 and Olivier Crête committed Apr 23, 2014

If doing a recv on UDP-TURN, the 'from' address was kept set to
the turn server which causes connchecks to fail with "mismatched address"
error. This fixes it

Nice to make debug logs more friendly to read.

Philip Withnall authored Apr 16, 2014 and Olivier Crête committed Apr 17, 2014

If an incoming ACK acknowledges part of a pending SSegment, correctly
increment the base offset of that SSegment before attempting to transmit
it. Otherwise, the wrong data will be transmitted and the offset passed
to packet() will be incorrect.

Philip Withnall authored Apr 16, 2014 and Olivier Crête committed Apr 17, 2014

If a packet has an invalid tsecr field, invalid (e.g. negative) RTTs may
be calculated from it. If so, discard the packet as invalid.

Philip Withnall authored Apr 16, 2014 and Olivier Crête committed Apr 17, 2014

As documented at the top of the file, this creates two PseudoTcpSockets
in a loopback configuration, and mutates the raw packets which are sent
between them, hoping to cause assertion failures and illegal memory
accesses in the pseudotcp code.

Best run under Valgrind or GDB.

If the error handling block in the middle of the recv() loop is hit,
various things were previously leaked (context, cancellable_source) and
the return values were potentially incorrect.

Fix the break out of the error handling block to clean up correctly.

At this point in the code, neither the username nor password are
nul-terminated, so printing them in the debug output as nul-terminated
strings can easily result in rubbish being printed after them.

Fix this by specifying the string length in the printf() format string.
Finally my terminal will no longer ask me to install N’Ko fonts when
running the libnice tests.

If @cancellable is NULL in a call to nice_output_stream_write(), the
WriteData struct is created with a reference count of 4, but only two
operations are scheduled which will result in its reference count being
decremented. The third operation is only scheduled if @cancellable is
non-NULL (and the final reference is dropped unconditionally at the end
of the function).

Fix this by properly implementing reference counting for WriteData,
rather than hard-coding the expected number of references in a fragile
and unmaintainable way.

The GOutputStream based systems can sometimes call this, decide the stream is
not writable and then poll for it. We have to make sure that poll is woken up.