Otherwise we constantly create/close event file descriptors,
every single time we call g_socket_condition_timed_wait() or
g_socket_receive_message(), i.e. twice per packet received!
This was not particularly good for performance.
Also only create GCancellable on start-up.
From the API documentation: "Note that it is generally not
a good idea to reuse an existing cancellable for more
operations after it has been cancelled once, as this
function might tempt you to do. The recommended practice
is to drop the reference to a cancellable after cancelling
it, and let it die with the outstanding async operations.
You should create a fresh cancellable for further async
operations."
https://bugzilla.gnome.org/show_bug.cgi?id=739132
Drop use of g_socket_get_available_bytes() which is
not useful on all systems (where it returns the size
of the entire buffer not that of the next pending
packet), and is yet another syscall and apparently
very inefficient on Windows in the UDP case.
Instead, when reading UDP packets, use the more featureful
g_socket_receive_message() call that allows to read into
scattered memory, and allocate one memory chunk which is
likely to be large enough for a packet, while also providing
a larger allocated memory chunk just in case the packet
is larger than expected. If the received data fits into the
first chunk, we'll just add that to the buffer we return
and re-use the fallback buffer for next time, otherwise we
add both chunks to the buffer.
This reduces memory waste more reliably on systems where
get_available_bytes() doesn't work properly.
In a multimedia streaming scenario, incoming UDP packets
are almost never fragmented and thus almost always smaller
than the MTU size, which is also why we don't try to do
something smarter with more fallback memory chunks of
different sizes. The fallback scenario is just for when
someone built a broken sender pipeline (not using a
payloader or somesuch)
https://bugzilla.gnome.org/show_bug.cgi?id=610364
udpsrc gtk-doc documentation refers to sockfd and closefd properties which has
been removed. This patch replaces those references to socket and close-socket
respectively.
https://bugzilla.gnome.org/show_bug.cgi?id=734987
They are very confusing for people, and more often than not
also just not very accurate. Seeing 'last reviewed: 2005' in
your docs is not very confidence-inspiring. Let's just remove
those comments.
Protect caps with the lock.
Don't push the caps event from the set_property function but mark the
pad for reconfiguration so that it will renegotiate and push the new
caps event in the streaming thread.
We should open the socket when going to NULL<->READY and not in the
start/stop vemthod, which is called in READY<->PAUSED. This makes it
possible to allocate a socket without going to PAUSED (and starting the
negotiation).
On Windows it's not possible to bind to a multicast address
but the OS will make sure to filter out all packets that
arrive not for the multicast address the socket joined.
On Linux and others it is necessary to bind to a multicast
address to let the OS filter out all packets that are received
on the same port but for different addresses than the multicast
address
And deprecate the multicast-group property and replace it with the
address property.
https://bugzilla.gnome.org/show_bug.cgi?id=707042
This is equivalent to multicast-group currently for backwards compatibility.
In 2.0 this should be handled separately, the former only being the multicast
group and the latter always being the address the socket is bound to, even if
a multicast group is given.
On Windows and OS/X, _get_available_bytes() may not return the size
of the next pending packet, but the size of all pending packets in
the kernel-side buffer, which might be rather large depending on
configuration. Sanity-check the size returned by _get_available_bytes()
to make sure we never allocate more memory than the max. size for
a packet, if it's an IPv4 socket.
https://bugzilla.gnome.org/show_bug.cgi?id=610364
Make it possible to set the timeout after we went to the READY state by using
the timeout when checking the condition. This also makes it possible to set the
timeout with a higher granularity than seconds.
udpsrc/udpsink are almost always used with RTP, so let's use an
RTP port as the default port. It's unclear why 4951 was used, it
goes back to early commits in CVS.
It is allowed to send/receive UDP packets with no data. When such
a packet is available, select() will return with success but
ioctl(FIONREAD) will return 0. But a read() must still occur in
order to clear off the UDP packet from the queue.
This patch will read the dataless packet from the socket. If
select() was woken for other reasons (and FIONREAD returns 0),
this may result in a UDP packet getting accidentally dropped.
But since UDP is not reliable, this is acceptable.
NOTE: This patch fixes a nasty bug where sending a dataless
UDP packet to a udpsrc instance will cause an infinite
loop.
https://bugzilla.gnome.org/show_bug.cgi?id=666644
Signed-off-by: John Ogness <john.ogness@linutronix.de>
