GitLab

Prevents Meson error:

ERROR:  requires argument not a string, library with pkgconfig-generated
file or pkgconfig-dependency object, got <InternalDependency null: True>

Requires Meson 0.49.

When building libnice with glib fallback, its 'libgio_dep' internal
dependency doesn't provide all needed include and library search paths.

To avoid unresolved header files and linker errors in such case,
'libglib_dep', 'libmodule_dep' and 'libgobject_dep' from glib submodule
must be pulled in as well.

The problem should be fixed in GLib 2.60.

nice_stream_close() calls agent_find_stream(), which iterates over
agent->streams list.

Therefore, when we're freeing agent->streams in a loop, we must also
delete the respective GSList item from the list at the end of each
iteration. Otherwise agent_find_stream() would stumble over already
deleted NiceStream instance the next time it gets called.

Fixes random crashes when running test-different-number-streams on
Windows.

Fixes failing test cases on Windows.

The underlying GSocket implementation on Windows returns an error when
the user-provided buffer isn't large enough to fit the whole datagram
received on a message-oriented socket.

When this occurs, in order to preserve identical behavior of udp-bsd
NiceSocket across platforms, we have to mute the error and set the
received message length to the size of the provided NiceInputMessage.
Any excess portion of the message gets discarded.

Fixed udp-bsd test on Windows.

GLib 2.48 is required in order to use G_IO_ERROR_MESSAGE_TOO_LARGE.

https://gitlab.freedesktop.org/gstreamer/gst-build/issues/14

test-fullmode-with-stun needs to run in non-parallel mode to work.

test-pseudotcp-fuzzy sometimes just times out, which needs investigating.

meson test -C builddir --repeat=20

stdint.h is available since vs2010 and stdbool.h since vs2013,
so we should be able to just use that these days.

Tested inside Cerbero with VS 2017

It's disabled by default, so saying it wasn't found is confusing.

Just implement argument parsing ourselves here.

It's not really more lines of code than the getopt.h variant
and makes things build on Windows with MSVC without having to
ship a separate getopt implementation.

No tests are being built or run here.

And drop gstreamer1-plugins-base-devel dep
which is not needed.

It's not a requirement, there's a fallback.

The streams are matched by the order in which they appear in the SDP,
not by name, so insisting on local streams to have one makes no sense.

This likely should have been removed in 5972db04.

If this isn't done, then there may be invalid points left inside the
passive socket which could be used and cause a crash.

Fixes #33

Avoids race between socket_send_more() and socket_close() leading either
to crash on use of freed NiceSocket, or abort upon attempting to
g_mutex_clear() a locked mutex.

Akin to what d1611c0c did to udp-turn.

To be able to call it from a context that is already locked.

nice_socket_recv_messages() may return a NiceInputMessage of length = 0,
so before attempting to read the RFC4571 header check the message really
has at least sizeof (guint16) bytes of data.

The bug's always been there, the previous commit only made it more
apparent.

There might be multiple RFC4571-framed messages (or fragments thereof)
within a single TCP-TURN message. Make sure each NiceInputMessage
passed by the user into socket_recv_messages() gets exactly one RFC4571
frame, or remains empty if there aren't any messages to receive.

We should keep any data that doesn't fit into the user buffers for
the next time socket_recv_messages() gets called with the socket.

After being parsed, a TURN control message turns into a NiceInputMessage
with zero length. Such message doesn't increment the iteration counter i
and so is re-processed in the next iteration, which detects right away
that message->length == 0 and continues to the next element in
recv_messages.

Thus, n_valid_messages variable serves no real purpose and to achieve
the same result we can simply increment the iteration counter after each
message.

The socket abstraction not being reference counted, we need a global
lock for them in the short term.

It's not locked in any way.

This should make it safer.

Remove all pointers that don't include have a reference except to agents

Work on libnice's bug #1 in Gitlab. This work is composed of multiple
merged parts:

- "Global lock contention removed"
Phabricator D1900: https://phabricator.freedesktop.org/D1900
By @nifigase
Opened in GitLab as Merge Request !12

- "agent: properly handle NiceAgent ref in callbacks from timeout
sources"
Phabricator D1898: https://phabricator.freedesktop.org/D1898
By @mparis
This patch was itself based upon a previous version of the work done in
D1900. After the switch of hosting, it got lost.

On top of these, additions to follow some review comments from @ocrete:
- https://phabricator.freedesktop.org/D1900#40412
- https://phabricator.freedesktop.org/D1898#39332