Otherwise we will have 10s-100s of thread wakeups in feedback profiles, create
RTCP packets, etc. just to suppress them in 99% of the cases (i.e. if no
feedback is actually pending and no regular RTCP has to be sent).
This improves CPU usage and battery life quite a lot.
https://bugzilla.gnome.org/show_bug.cgi?id=746543
If we may suppress the packet due to the rules of RFC4585 (i.e. when
below the t-rr-int), we can send a smaller RTCP packet without RRs
and full SDES. In theory we could even send a minimal RTCP packet
according to RFC5506, but we don't support that yet.
https://bugzilla.gnome.org/show_bug.cgi?id=746543
Otherwise we can't properly schedule RTCP in feedback profiles as we need to
distinguish the time when we last checked for sending RTCP (tp) but might have
suppressed it, and the time when we last actually sent a non-early RTCP
packet.
This together with the other changes should now properly implement RTCP
scheduling according to RFC4585, and especially allow us to send feedback
packets a lot if needed but only send regular RTCP packets every once in a
while.
https://bugzilla.gnome.org/show_bug.cgi?id=746543
And modify our RTCP scheduling algorithm accordingly. We now can send more
RTCP packets if needed for feedback, but will throttle full RTCP packets by
rtcp-min-interval (t-rr-int from RFC4585).
In non-feedback mode, rtcp-min-interval is Tmin from RFC3550, which is
statically set to 1s or 0s by RFC4585. Tmin defines how often we should
send RTCP packets at most.
https://bugzilla.gnome.org/show_bug.cgi?id=746543
It might just be a late retransmission or spurious packet from elsewhere, but
resetting everything would mean that we will cause a noticeable hickup. Let's
get some confidence first that the sequence numbers changed for whatever
reason.
https://bugzilla.gnome.org/show_bug.cgi?id=747922
This reverts commit d22ec49632.
Application code might expect that it only gets external sources on those
signals, and get confused by this. If anything we would need to add new
signals.
Without this it seems impossible for an application to easily get notified
about the internal ssrcs that are created, e.g. sender sources, and also
to know when they are active and produce RTCP packets.
https://bugzilla.gnome.org/show_bug.cgi?id=746747
We now take the maximum of 2*jitter and 0.5*packet_spacing for the extra
delay. If jitter is very low, this should prevent unnecessary retransmission
requests to some degree.
https://bugzilla.gnome.org/show_bug.cgi?id=748041
This also happens in the very beginning when we receive the first packet, a
warning would be very confusing here. In all places where we should warn about
this, we would've printed a warning already before.
Right above we consider lost_packet packets, each of them having duration,
as lost and triggered their timers immediately. Below we use expected_dts
to schedule retransmission or schedule lost timers for the packets that
come after expected_dts.
As we just triggered lost_packets packets as lost, there's no point in
scheduling new timers for them and we can just skip over all lost packets.
https://bugzilla.gnome.org/show_bug.cgi?id=739868
Resetting the jitterbuffer drops all packets and other things, and will cause
a discontinuity in the packets received by the depayloaders. They should now
also flush anything they had pending as the new data will start at a different
position.
https://bugzilla.gnome.org/show_bug.cgi?id=739868
Make sure that the sync_src pad has caps before the segment event.
Otherwise we might get a segment event before caps from the receive
RTCP pad, and then later when receiving RTCP packets will set caps.
This will results in a sticky event misordering warning
This fixes warnings in the rtpaux unit test but also in the
rtpaux and rtx examples in tests/examples/rtp
https://bugzilla.gnome.org/show_bug.cgi?id=746445
Before we only started it when either:
- there is no send RTP stream
or
- we received an RTP packet for sending
This could mean that if the send RTP pads are connected but never receive any
RTP data, and the same session is also used for receiving RTP/RTCP, we would
never start the RTCP thread and would never send RTCP for the receiving part
of the session.
This can be reproduced with a pipeline like:
gst-launch-1.0 rtpbin name=rtpbin \
udpsrc port=5000 ! "application/x-rtp, media=video, clock-rate=90000, encoding-name=H264" ! rtpbin.recv_rtp_sink_0 \
udpsrc port=5001 ! rtpbin.recv_rtcp_sink_0 \
rtpbin.send_rtcp_src_0 ! fakesink name=rtcp_fakesink silent=false async=false sync=false \
rtpbin.recv_rtp_src_0_2553225531_96 ! decodebin ! xvimagesink \
fakesrc ! valve drop=true ! rtpbin.send_rtp_sink_0 \
rtpbin.send_rtp_src_0 ! fakesink name=rtp_fakesink silent=false async=false sync=false -v
Before this change the rtcp_fakesink would never send RTCP for the receiving
part of the session (i.e. no receiver reports!), after the change it does.
And before and after this change it would send RTCP for the receiving part of
the session if the sender part was omitted (the last two lines).
These are outside the expected range of sequence numbers and should be
clipped, especially for RTSP they might belong to packets from before a seek
or a previous stream in general.
This reverts commit 1591adf4cd.
https://bugzilla.gnome.org/show_bug.cgi?id=745586#c1:
It's the beginning of an implementation of RFC 2762, which is needed for
large multicast groups. The implementation is not yet complete but why
not leave what is there and implement RFC 2762 instead?
rtpsession declares an array of maps to store srrcs but only the
the key 0 is being used. This patch replaces the array of maps
for just one map and remove useless parameters in rtpsession
https://bugzilla.gnome.org/show_bug.cgi?id=745586
According to RFC 4585 section 3.5.3 step 1 we are not allowed to send
an early RTCP packet for the very first one. It must be a regular one.
Also make sure to not use last_rtcp_send_time in any calculations until
we actually sent an RTCP packet already. In specific this means that we
must not use it for forward reconsideration of the current RTCP send time.
Instead we don't do any forward reconsideration for the first RTCP packet.
This will be emitted whenever an RTCP packet is received. Different to
on-feedback-rtcp, this signal gets every complete RTCP packet and not
just the individual feedback packets.
It had no effect since quite some time and also is not needed in general,
especially not to switch between immediate feedback mode and early feedback
mode. The latest understanding of the RFC is that from the endpoint point of
view, both modes are exactly the same. RTCP is only allowed to use the
bandwidth as given by the RFC constraints, as such it is only ever possible
to schedule a RTCP packet early but it's against the RFC to schedule more RTCP
packets.
The difference between immediate feedback mode and early feedback mode is that
the former guarantees that an RTCP packet can be sent for every event
"immediately", which means that the bandwidth calculations from the RFC have
resulted in an RTCP scheduling interval that is small enough. Early feedback
mode on the other hand means that we can schedule some packets early to make
that happen, but it's not guaranteed at all that it's possible to schedule
an RTCP packet per event (i.e. they need to be accumulated or dropped).
This indicates with a boolean return value if scheduling a new RTCP packet
within the requested delay was possible. Otherwise it behaves exactly like
send-rtcp. The only reason for adding a new signal is ABI compatibility.
We (currently?) can't really handle gaps between RTP packets if they're not
properly timestamped. The current code would go into calculations with
GST_CLOCK_TIME_NONE and then cause assertions everywhere. It's probably
better to error out cleanly instead.
Apparently linphone sends an invalid RTP packet as very
first packet. We want to ignore that instead of erroring
out (same for any other invalid packets really).
https://bugzilla.gnome.org/show_bug.cgi?id=741398
We used to setup an iterator with 1 GValue set with a NULL object
pointer which is not the normal way to do that. Instead we should make
sure that the first call to gst_iterator_next returns GST_ITERATOR_DONE.
When the RTT and jitter are very low (such as on a local network), the
calculated retransmission timeout is very small. Set some sensible lower
boundary to the timeout by adding a new property. We use the packet
spacing as a lower boundary by default.
In early retransmission we are allowed to schedule 1 regular RTCP packet
at an earlier time. When we do that, we need to set allow_early to FALSE
and ignore/drop (or merge) all future requests for early transmission.
We now first check if we can schedule an early RTCP and if we can,
actually prepare the data for the next RTCP interval.
After we send the next regular RTCP after the early RTCP, we set
allow_early to TRUE again to allow more early requests.
Remove the condition for the immediate feedback for now.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=738319
Add a need-resync state, this is when we need to try to lock on to a
time/RTPtime pair.
Always check the RTP timestamps and if they go backwards, mark ourselves
as need-resync.
Only resync when need-resync is TRUE and we have a valid time. Otherwise
we keep the old values. This avoids locking on to an invalid time and
causing us to timestamp everything with -1.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=730417
rtpmux behaves like a funnel in that it forwards whatever upstream is
sending buffers. So setting proxy caps doesn't make sense as the
upstream don't have to have compatible caps, thus resulting in an empty
caps set as a result of a caps query. Instead set fixed caps just
as funnel does.
https://bugzilla.gnome.org/show_bug.cgi?id=738722
We never initialize clock_rate explicitly, therefore it is 0 by default. The
parameter is a uint32 and the only caller ensure that it is >0, therefore it
won't become -1 ever.
The jitterbuffer shouldn't force clock-rate on its sink pad, this will cause a negotiation issue since rtpssrcdemux doesn't have the clock-rate and doesn't add it to the caps. The documentation states that the clock-rate can either be specified through the caps or through the request-pt-map signal, so we must remove clock-rate from the pad templates and we must accept the GST_EVENT_CAPS if the caps don't have the clock-rate.
https://bugzilla.gnome.org/show_bug.cgi?id=734322
Implement 3 different cases for handling the SR:
1) we don't have enough timing information to handle the SR packet and
we need to wait a little for more RTP packets. In that case we keep
the SR packet around and retry when we get an RTP packet in the
chain function.
2) the SR packet has a too old timestamp and should be discarded. It is
labeled invalid and the last_sr is cleared.
3) the SR packet is ok and there is enough timing information, proceed
with processing the SR packet.
Before this patch, case 2) and 1) were handled in the same way,
resulting that SR packets with too old timestamps were checked over and
over again for each RTP packet.
1) sources that have sent BYE in the past cannot be senders, since
they would have timed out to being receivers in the meantime...
2) sources that have sent BYE are now being removed earlier inside
this function
If we are inserting a packet into the jitter queue we need to keep
looping through the items until the right position is found. Currently,
the code stops as soon as an event is found in the queue.
Regarding events, we should only move packets before an event if there
is another packet before the event that has a larger seqnum.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=730078
If two streams request a retranmission for the same SSRC, ignore the second
one if the first oen is less than one second old, otherwise time out the first
one and ignore the second.
As we now replace the local RTPSource on a conflict, it's no longer possible
to keep local conflicts in the RTPSource, so they instead need to be kept
in the RTPSession.
Also fix the rtpcollision test to generate multiple collisions instead of
one by change the address, as otherwise we detected that it was a single one.
Rework the packet queue so that the most common action (insert a packet
at the tail of the queue) goes very fast.
Report if a packet was inserted at the head instead of the tail so that
we can know when to retry _pop or _peek.
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.
Make a new method to disable the jitterbuffer buffering.
Rework the update_estimated_eos() method. Calculate how much time
there is left to play. If we have less than the delay of the
jitterbuffer, we disabled buffering because we might never be able to
fill the complete jitterbuffer again.
If we receive an EOS event, disable buffering. We will drain the
buffer and eventually push the EOS event out.
When we reach the estimated NPT timeout and we didn't receive an EOS
event, make one and queue it so that it can be pushed.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=728017
When the internal-ssrc property changes, we want to send a reconfigure
upstream to make payloaders use the new suggested ssrc.
Using the internal-ssrc property to change the SSRC of a stream is not a
good idea and doesn't work when there are multiple senders, we want to
set the SSRC directly on the payloaders. Therefore, deprecate this
property.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=725361
Rework the logic to make buffering messages a little, make sure we
don't make the same message multiple times.
Consider the buffer full when EOS was received.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=728017
When we are buffering, we can't block and wait for the serialized query
to complete because the jitterbuffer will not try to forward the query
while buffering. Instead, just refuse the query.
It is possible that the DTS is invalid (when we receive RTP packets from
TCP, for example). As a fallback, use the reconstructed PTS value to
calculate the buffer level.
gstrtpjitterbuffer.c: In function 'gst_rtp_jitter_buffer_loop':
gstrtpjitterbuffer.c:2978:3: error: 'result' may be used uninitialized in this function
while (result == GST_FLOW_OK);
^
Several conditional statements perform comparison on RTP sequence
numbers without taking the sequence number rollover into account.
Instead, use the gst_rtp_buffer_compare_seqnum function to perform the
comparison.
https://bugzilla.gnome.org/show_bug.cgi?id=725159
recv_rtp_sink: allow proxying of the allocation query.
send_rtp_sink: allow proxying of caps and allocation. This allows us to
query caps downstream as well as get an allocator from downstream.
send_rtp_src: allow proxy of caps, this makes the caps query do
upstream.
See https://bugzilla.gnome.org/show_bug.cgi?id=723850
If the expected packet (do_next_seqnum is TRUE) is the one we requested
for retranmission earlier, do the logic to update the retransmission
statistics as well before setting up the timers for the next expected
packet.
Also reset the retransmission counter if the timer is reused for another
seqnum.
Add an accumulator that stops the signal emission as soon as a caps has
been retrieved. Otherwise the default handler would continue emitting
the signal and possibly overwrite the result with NULL again.
To do that, enqueue the EOS event to be sent from the srcpad task
thread and flush the queue right afterwards, so that no more rtx
buffers can be sent, even if there are more requests coming in.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=722370
The reason behind this is to minimize the retransmission delay.
Previously, when a NACK was received, rtprtxsend would put a
retransmission packet in a queue and it would send it from chain(),
i.e. only after a new buffer would arrive.
This unfortunately was causing big delays, in the order of 60-100 ms,
which can be critical for the receiver side.
By having a separate GstTask for pushing buffers out of rtxsend,
we can push buffers out right after receiving the event, without
waiting for chain() to get called.
1) pt can be lower than 96
2) there is no point in checking that because rtprtxsend will not
even store buffers for payload types that it doesn't know about,
so this case will never be reached
This patch moves the creation of rtx packets to be done early,
in the src_event() function, when they are requested. The purpose
is to run gst_rtp_rtx_buffer_new() with the object locked to
protect internal data, because if it is done at the pushing stage,
we would have to lock and unlock multiple times in a row while we
are pushing the rtx buffers.
Previously there was no locking at all, which was terribly wrong.
This property allows you to specify the amount of buffers
to keep in the retransmission queue expressed as time (ms)
instead of buffer count (which is the max_size_buffers property).
The purpose of the sender RTX object is to keep a history
of RTP packets up to a configurable limit (in time). It will
listen for custom retransmission events from downstream. When
it receives a request for retransmission, it will look up the
requested seqnum in its list of stored packets. If the packet
is available, it will create a RTX packet according to RFC 4588
and send this as an auxiliary stream.
The receiver will listen to the custom retransmission events
from the downstream jitterbuffer and will remember the SSRC1
of the stream and seqnum that was requested. When it sees a
packet with one of the stored seqnum, it associates the SSRC2
of the stream with the SSRC1 of the master stream. From then
on it knows that SSRC2 is the retransmission stream of SSRC1.
This algorithm is stated in RFC 4588. For this algorithm to
work, RFC4588 also states that no two pending retransmission
requests can exist for the same seqnum and different SSRCs or
else it would be impossible to associate the retransmission with
the original requester SSRC.
When the RTX receiver has associated the retransmission packets,
it can depayload and forward them to the source pad of the element.
RTX is SSRC-multiplexed
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=711084
AUX elements are elements that can be inserted into the rtpbin
pipeline right before or after 1 or more session elements.
The AUX elements are essential for implementing functionality such
as error correction (FEC) and retransmission (RTX).
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=711087
Keep track of elements that are added to multiple sessions and make sure
we only add them to the rtpbin once and that we clean them when no
session refers to them anymore.
When a collision is found on the internal ssrc, we have to change it.
Ideally, we want also the payloader upstream to follow this change and use
the new internal ssrc. Ideally we want this condition to be always met:
if there is one payloader sending on this session, its ssrc should match the
internal ssrc.
* gst/rtpmanager/gstrtpbin.[ch]: four new action signals have been
added (request-rtp-encoder, request-rtp-decoder, request-rtcp-encoder
and request-rtcp-decoder). The user will be able to provide encoders
or decoders dynamically. The encoders must follow the srtpenc API and
the decoders the srtpdec API. Having separate signals for RTP and RTCP
allows the user to use different encoders/decoders or provide the same
one (e.g. that would be the case for srtpenc).
Also, rtpbin now allows application/x-srtp in its pads.
https://bugzilla.gnome.org/show_bug.cgi?id=719938
Use the round-trip-time and average jitter to dynamically calculate the
retransmission interval and expected packet arrival time.
Based on patches from Torrie Fischer <torrie.fischer@collabora.co.uk>
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=711412
Don't use the current time calculated from the tmieout loop for when we
last scheduled the NACK because it might be unscheduled because of a max
packet misorder and then we don't accurately calculate the current time.
Instead, take the current element running time using the clock.
According to RFC 5104 section 4.3.1.2, RTCP PSFB FIR message SHALL
have a media_ssrc field set to 0. The actual media ssrc is in the FCI.
So in that case, we ignore the retained feedback and just let it through
to the rtp_session_process_fir() function which will check for the actual
SSRC inside the FCI.
Fixes a regression introduced by commit 57c27ec3
Previously, when the session had multiple internal sender SSRCs, it would
issue SR reports with RB blocks only on the first RTCP timeout and afterwards
SR reports would be sent empty. This was because the "generation" number
in RTPSource would increase more than once during the same cycle and afterwards
it would always be greater than the session's generation, which would cause
it to be skipped from being included in RBs.
This commit fixes this problem by:
1) Increasing the RTPSource generation only at the end of each cycle,
which essentially fixes the problem but only when the internal senders
are less than GST_RTCP_MAX_RB_COUNT.
2) Keeping for each RTPSource a set of SSRCs which stores which SSRC's
SR the given RTPSource has been reported in, which also fixes the problem
when the internal senders are more than GST_RTCP_MAX_RB_COUNT. This is
necessary because of the fact that any RTPSource is marked as reported
in itself's SR and makes it impossible to know if it has been reported
in other SRs too or not, and which.
Keep an extra stats structure for scheduling the BYE packets. When we
decide to schedule BYE, make a copy of the current stats into the
bye_stats. Then while we schedule the BYE, update and use only the
bye_stats. When we finished scheduling the BYE packet, we use the
regular stats again.
When we are scheduling BYE packets, ignore all RTCP for the sources that
are scheduling a BYE packet. Other sources that are not scheduling BYE
should continue receiving RTCP packets as usual.
Don't reset the expected output seqnum when clearing the pt map because this
could stall the jitterbuffer forever.
Add a unit test for this.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=709800
Remove bogus reconfigure event on collision, we don't want to send the event on
the receiving RTP pad and the collision event is now handling this
case.
See https://bugzilla.gnome.org/show_bug.cgi?id=711560
The problem here was that the jitterbuffer lock was unlocked to push
the event, but that caused another thread to remove the timer currently
being processed, probably because the amount of rtx events
(and therefore timers) was getting too high. The solution is to
unlock and push the event only after timer processing has finished.
fixes https://bugzilla.gnome.org/show_bug.cgi?id=711131
An internal sender in a session is also a receiver of its own packets so update
the receiver stats. Other senders in the session will use this info to generate
correct RB blocks in their SR reports.
Add a new timestamp mode that assumes the local and remote clock are
synchronized. It takes the first timestamp as a base time and then uses the RTP
timestamps for the output PTS.
When flush-stop arrives before we process the result of the _push() in the
loop function, we might pause even though we are not flushing anymore. Fix this
race by waiting for the srcpad loop function to completely pause after doing the
flush-start.
If we were not waiting for the missing seqnum when we insert the lost packet
event in the jitterbuffer, we end up not updating the next_seqnum and wait
forever for the lost packets to arrive. Instead, keep track of the amount of
packets contained by the jitterbuffer item and update the next expected
seqnum only after pushing the buffer/event. This makes sure we correctly handle
GAPS in the sequence numbers.
Always prepare a lost event in the jitterbuffer, it is to wake up and make the
pushing thread continue. We drop the event when we are not supposed to push lost
events downstream.
Schedule the lost event by placing it inside the jitterbuffer with the seqnum
that was lost so that the pushing thread can interleave and push it properly.
Make the jitterbuffer operate on a structure containing all the packet
information. This avoids mapping the buffer multiple times just to get the RTP
information. It will also make it possible to store other miniobjects such as
events later.
Improve the order of the timeout events, if there are timers with the same
timeout, we want to trigger the lowest seqnum first. For this we need to loop
over the complete array of timers to find the best one before triggering the
timeout.
First send the lost event, then update the next_seqnum counter and then
send the signal to the pushing thread that it can retry to push a buffer. This
avoids pushing out buffers before the lost event is pushed.
There is no need to unschedule the timer in flush-start, flush-stop will remove
the timers and unschedule.
Unschedule the current timer before attempting to join the timer thread.