gstreamer/docs/design/design-rtpretransmission.txt

RTP retransmission design


GstRTPRetransmissionRequest
---------------------------

Custom upstream event which mainly contains the ssrc and the seqnum of the
packet which is asked to be retransmisted.

On the pipeline receiver side this event is generated by the
gstrtpjitterbuffer element. Then it is translated to a NACK to be sent over
the network.

On the pipeline sender side, this event is generated by the gstrtpsession
element when it receives a NACK from the network.


rtprtxsend element
------------------

-- basic mechanism

rtprtxsend keeps a history of rtp packets that it has already sent.
When it receives the event GstRTPRetransmissionRequest from the downstream
gstrtpsession element, it loopkup the requested seqnum in its stored packets.
If the packet is present in its history, it will create a RTX packet according
to RFC 4588. Then this rtx packet is pushed to its src pad as other packets.

rtprtxsend works in SSRC-multiplexed mode, so it has one always sink and
src pad.

-- building retransmission packet fron original packet

A rtx packet is mostly the same as an orignal packet, except it has its own
ssrc and its own seqnum. That's why rtprtxsend works in SSRC-multiplexed mode.
It also means that the same session is used.
Another difference between rtx packet and its original is that it inserts the
original seqnum (OSN: 2 bytes) at the beginning of the payload.
Also rtprtxsend builds rtx packet without padding, to let other elements do that.
The last difference is the payload type. For now the user has to set it through
the rtx-payload-type property. Later it will be automatically retreive this
information from SDP. See fmtp field as specifies in the RPC4588
(a=fmtp:99 apt=98) fmtp is the payload type of the retransmission stream
and apt the payload type of its associated master stream.

-- restransmission ssrc and seqnum

To choose rtx_ssrc it randomly selects a number between 0 and 2^32-1 until
it is different than master_ssrc.
rtx_seqnum is randomly selected between 0 and 2^16-1

-- deeper in the stored buffer history

For the history it uses a GSequence with 2^15-1 as its maximum size.
Which is resonable as the default value is 100.
It contains the packets in reverse order they have been sent
(head:newest, tail:oldest)
GSequence allows to add and remove an element in constant time (like a queue).
Also GSequence allows to do a binary search when rtprtxsend lookup in its
history.
It's important if it receives a lot of requests or if the history is large.

-- pending rtx packets

When looking up in its history, if seqnum is found then it pushes the buffer
into a GQueue to its tail.
Before to send the current master stream packet, rtprtxsend sends all the
buffers which are in this GQueue. Taking care of converting them to rtx
packets.
This way, rtx packets are sent in the same order they have been requested.
(g_list_foreach traverse the queue from head to tail)
The GQueue is cleared between sending 2 master stream packets.
So for this GQueue to contain more than one element, it means that rtprtxsend
receives more than one rtx request between sending 2 master packets.

-- collision

When handling a GstRTPCollision event, if the ssrc is its rtx ssrc then
rtprtxsend clear its history and its pending retransmission queue.
Then it chooses a rtx_ssrc until it's different than master ssrc.
If the GstRTPCollision event does not contain its rtx ssrc, for example
its master ssrc or other, then it just forwards the event to upstream.
So that it can be handled by the rtppayloader.


rtprtxreceive element
------------------

-- basic mechanism

The same rtprtxreceive instance can receive several master streams and several
retransmission streams.
So it will try to dynamically associate a rtx ssrc with its master ssrc.
So that it can reconstruct the original from the proper rtx packet.

The algorithm is based on the fact that seqnums of different streams
(considering all master and all rtx streams) evolve at a different rate.
It means that the initial seqnum is random for each one and the offset could
also be different. So that they are statistically all different at a given
time. If bad luck then the association is delayed to the next rtx request.

The algorithm also needs to know if a given packet is a rtx packet or not.
To know this information there is the rtx-payload-types property. For now the
user as to configure it but later it will be automatically retreive this
information from SDP.
It needs to know if the current packet is rtx or not in order to know if
it can extract the OSN from the payload. Otherwise it would extract the OSN
even on master streams which means nothing and so it could do bad things.
In theory maybe it could work but we have this information in SDP so why not
using it to avoid bad associations.

Note that it also means that several master streams can have the same payload
type. And also several rtx streams can have the same payload type.
So the information from SDP which gives us which rtx payload type belong to
a give master payload type is not enough to do the association between rtx ssrc
and master ssrc.

rtprtxreceive works in SSRC-multiplexed mode, so it has one always sink and
src pad.

-- deeper in the association algorithm

When it receives a GstRTPRetransmissionRequest event it will remember the ssrc
and the seqnum from this request.

On incoming packets, if the packet has its ssrc already associated then it
knows if the ssrc is an rtx ssrc or a master stream ssrc.
If this is a rtx packet then it recontructs the original and pushs the result to
src pad as if it was a master packet.

If the ssrc is not yet associated rtprtxreceive checks the payload type.
if the packet has its payload type marked as rtx then it will extract the OSN
(original seqnum number) and lookup in its stored requests if a seqnum matchs.
If found, then it associates the current ssrc to the master ssrc marked in the
request. If not found it just drops the packet.
Then it removes the request from the stored requests.

If there are 2 requests with the same seqnum and different ssrc, then the
couple seqnum,ssrc is removed from the stored requests.
A stored request actually means that actually the couple seqnum,ssrc is stored.
If it's happens the request is droped but it avoids to do bad associations.
In this case the association is just delayed to the next request.

-- building original packet from rtx packet

Header, extensions, payload and padding are mostly the same. Except that the
OSN is removed from the payload. Then ssrc, seqnum, and original payload type
are correctly set. Original payload type is actually also stored when the
rtx request is handled.