Skip to content
Iñaki Baz Castillo edited this page Jul 12, 2019 · 40 revisions

Here the roadmap for adding DataChannel support to mediasoup. It includes a preliminary architecture design among other considerations (below) such as implementation details.

Architecture (new)

The DataChannel design and exposed API must follow the principles of the already exiting audio/video Producer and Consumer design in mediasoup v3. So, the DataChannel design must conform to the following requirements:

  • The DataChannel message unit is a SCTP message (in the same way, the message unit in audio/video is a RTP packet).

  • There must two new classes: DataProducer and DataConsumer (in both mediasoup server and mediasoup-client).

    • A DataProducer represents a data channel source (so a SCTP stream with a specific stream_id) being injected into a mediasoup router. It's created on top of a transport that defines how the SCTP packets are carried.
    • DataProducers are created using the already existing transport.produce() JavaScript API with new arguments.
    • A DataConsumer represents a data channel source (so a SCTP stream with a specific stream_id) being forwarded from a mediasoup router to an endpoint. It's created on top of a transport that defines how the SCTP packets are carried.
    • DataConsumers are created using the already existing transport.consume() JavaScript API with new arguments.
    • DataConsumers inherit the same SCTP settings as the DataProducer they consume from (ordered/unordered, reliable/unreliable, etc).
  • STCP packets can be injected into a mediasoup router on top of a WebRtcTransport (so STCP over DTLS over UDP/TCP) and also on top of a PlainRtpTransport (so SCTP over UDP).

    • NOTE: We may want to rename PlainRtpTransport to PlainTransport.
  • The WebRtcTransport and PlainRtpTransport will hold a new SctpAssociation class instance to manage the SCTP connection itself, STCP retransmission, etc.

  • When a SCTP endpoint sends a SCTP packet on a DataProducer:

    • The SCTP packet reaches the mediasoup WebRtcTransport or PlainRtpTransport which delivers the SCTP packet to its SctpAssociation class instance.
    • The SctpAssociation will pass it to the SCTP stack which will eventually call recv callbacks (notifications and messages). Let's assume it's a full message.
    • The SctpAssociation will notify the transport about "new SCTP message with this stream_id".
    • The transport looks for the corresponding DataProducer using its SctpListener (similar to the RtpListener) and passes the message to it.
    • The DataProducer may need to do something with the SCTP message (stats and so on) then notifies its transport back with the same message.
    • The transport notifies it to the router (Router C++ class).
    • The router iterates the DataConsumers associated to the sender DataProducer and calls dataConsumer->SendSctpMessage() on all them.
    • Each DataConsumer gets the message and does some stats stuff and so on, and the notifies its transport about it.
    • The transport of the DataConsumer receives the message and passes it to its SctpAssociation instance (that will manage retransmissions, etc) by also indicating the stream_id it must use.
    • The SctpAssociation passes it to the SCTP stack which will eventually invoke the send callback, in which the generated SCTP packets will be delivered to the transport again to deliver them to the remote endpoint:
      • If it's a WebRtcTransport, it will send the SCTP packet as DTLS data using the already existing DtlsTransport::SendApplicationData() method.
      • If it's a PlainRtpTransport it will send the SCTP packet using raw UDP.
    • The DataConsumer then restores the SCTP packet fields.
  • Similar to audio/video Producers and Consumers, it must be possible for the Node.js app running mediasoup to decide which endpoints consume (by means of a DataConsumer) from a specific DataProducer. Once a DataConsumer is created, the Router (at C++ level) will route to it SCTP messages from the associated DataProducer.

  • It must be possible for the Node.js app (the application that integrates mediasoup lib and uses its JS API) to inject SCTP messages into a mediasoup router so those messages are routed to endpoints via WebRTC DataChannel. And vice-versa: the Node.js app must be able to receive SCTP packets sent by endpoints to the mediasoup router. To do it, the Node.js app:

    • must create a PlainRtpTransport in the mediasoup router,
    • must call produceData() on it with appropriate arguments to create a DataProducer instance,
    • must create a UDP+SCTP client that sends (and/or receives) SCTP packets (with the announced stream_id) to the IP:port of the previously created mediasoup PlainRtpTransport.
      • The sctp_port must have a hardcoded value (5000 is good) to demux it with RTP, RTCP, STUN and DTLS.
  • Such a "UDP+SCTP client" could be a Node.js library/module which, for example, uses the Node.js UDP stack plus the node-sctp or @nodertc/sctp Node.js libraries.

    • Of course, we (the mediasoup team) will provide developers with such a client, probably as a separate library.
    • The node-sctp can, indeed, send SCTP packets over UDP: example.
    • Since Chrome does not allow SCTP messages bigger than 256KB (sad!!!) we may need to tell the application to use chunked-dc-js (credits to @@lgrahl).

Architecture (old)

UPDATE: Unfortunately this cool design is not possible due to usrsctp (no SCTP I-DATA support) and Chrome (which limits max SCTP message size to 256 KB). So mediasoup must route full SCTP messages from DataProducers to DataConsumers and not SCTP I-DATA chunks.

The DataChannel design and exposed API must follow the principles of the already exiting audio/video Producer and Consumer design in mediasoup v3. So, the DataChannel design must conform to the following requirements:

  • The DataChannel message unit is a SCTP I-DATA chunk (in the same way, the message unit in audio/video is a RTP packet).

  • There must two new classes: DataProducer and DataConsumer (in both mediasoup server and mediasoup-client).

    • A DataProducer represents a data channel source (so a SCTP stream with a specific stream_id) being injected into a mediasoup router. It's created on top of a transport that defines how the SCTP packets are carried.
    • DataProducers are created using the already existing transport.produce() JavaScript API with new arguments.
    • A DataConsumer represents a data channel source (so a SCTP stream with a specific stream_id) being forwarded from a mediasoup router to an endpoint. It's created on top of a transport that defines how the SCTP packets are carried.
    • DataConsumers are created using the already existing transport.consume() JavaScript API with new arguments.
    • DataConsumers inherit the same SCTP settings as the DataProducer they consume from (ordered/unordered, reliable/unreliable, etc).
  • STCP packets can be injected into a mediasoup router on top of a WebRtcTransport (so STCP over DTLS over UDP/TCP) and also on top of a PlainRtpTransport (so SCTP over UDP).

    • NOTE: We may want to rename PlainRtpTransport to PlainTransport.
  • The WebRtcTransport and PlainRtpTransport will hold a new SctpAssociation class instance to manage the SCTP connection itself, STCP retransmission, etc.

  • When a SCTP endpoint sends a SCTP packet on a DataProducer:

    • The SCTP packet reaches the mediasoup WebRtcTransport or PlainRtpTransport which delivers the SCTP packet to its SctpAssociation class instance.
    • The SctpAssociation will split the packet into chunks and determine whether each chunk is a SCTP control packet (init, reset, new stream created, alert, etc) or a SCTP I-DATA chunk. Let's assume it's a I-DATA chunk.
    • The SctpAssociation will notify the transport about "new SCTP I-DATA chunk".
    • The transport reads the SCTP stream_id of the chunk and looks for the corresponding DataProducer using its SctpListener (similar to the RtpListener) and passes the chunk to it.
    • The DataProducer may need to do something with the SCTP chunk (TSN, etc) and then notifies its transport back with the same SCTP I-DATA chunk.
    • The transport notifies it to the router (Router C++ class).
    • The router iterates the DataConsumers associated to the sender DataProducer and calls dataConsumer->SendSctpIDataChunk(chunk) on all them.
    • Each DataConsumer gets the chunk, mangles its SCTP streamId field and other fields in the chunk and notifies its transport to deliver the chunk.
    • The transport of the DataConsumer receives the chunk and passes it to its SctpAssociation instance (that will manage retransmissions, etc).
    • The SctpAssociation will then create a SCTP packet, put the chunk into it and notify back the SCTP chunk to the transport, which will finally deliver the packet to the remote endpoint:
      • If it's a WebRtcTransport, it will send the SCTP packet as DTLS data using the already existing DtlsTransport::SendApplicationData() method.
      • If it's a PlainRtpTransport it will send the SCTP packet using raw UDP.
    • The DataConsumer then restores the SCTP packet fields.
  • The WebRtcTransport and PlainRtpTransport may need a new SctpAssociation class instance to manage the SCTP connection itself.

  • Similar to audio/video Producers and Consumers, it must be possible for the Node.js app running mediasoup to decide which endpoints consume (by means of a DataConsumer) from a specific DataProducer. Once a DataConsumer is created, the Router (at C++ level) will route to it SCTP I-DATA chunks from the associated DataProducer.

  • It must be possible for the Node.js app (the application that integrates mediasoup lib and uses its JS API) to inject SCTP I-DATA chunks into a mediasoup router so those messages are routed to endpoints via WebRTC DataChannel. And vice-versa: the Node.js app must be able to receive SCTP packets sent by endpoints to the mediasoup router. To do it, the Node.js app:

    • must create a PlainRtpTransport in the mediasoup router,
    • must call produce() on it with appropriate arguments to create a DataProducer instance,
    • must create a UDP+SCTP client that sends (and/or receives) SCTP packets (with the announced stream_id) to the IP:port of the previously created mediasoup PlainRtpTransport.
      • The sctp_port must have a hardcoded value to demux it with RTP, RTCP, STUN and DTLS.
  • Such a "UDP+SCTP client" could be a Node.js library/module which, for example, uses the Node.js UDP stack plus the node-sctp or @nodertc/sctp Node.js libraries.

    • Of course, we (the mediasoup team) will provide developers with such a client, probably as a separate library.
    • The node-sctp can, indeed, send SCTP packets over UDP: example.
    • Since Chrome does not allow SCTP messages bigger than 256KB (sad!!!) we may need to tell the application to use chunked-dc-js (credits to @@lgrahl).

Issues

  • An issue in the design above is that a DataProducer with vey good uplink connection may send chunks super fast, while a DataConsumer may not send those chunks so fast, and that could break the Internet connection.
    • So mediasoup may need to "slow down" the data sending to the consuming endpoint, and that means buffering, congestion control, etc.
    • If we used usrsctp instead (SCTP message based), the lib itself would slow down the data transmission of the given message.

Implementation at C++ level

The WebRtcTransport of mediasoup already has an (unused yet) API to send and receive DTLS application data (so SCTP packets):

(we would add the same OnDtlsApplicationData callback into the PlainRtpTransport for SCTP over UDP).

So we need to plug a C++ SCTP stack to handle SCTP packets. It would be done within a new C++ class named SctpHandler or whatever.

As per now there are the following C or C++ SCTP stacks out there we can consider:

  • Create our own stack, which is the only one that can implement the design above (based on STCP I-DATA chunk routing instead of SCTP message routing).

  • sctplab/usrsctp, the "SCTP stack". Used by libwebrtc, Janus among many others. Problems:

    • It's multi-thread (question). We need a single-thread stack.
    • There is a Pull Request that makes it single thread, but was not merged.
    • Then @lgrahl said "No need to wait, you can use usrsctp-neat which will be merged back eventually", but that did never happen.
    • I also do not like the idea of integrating a library with more than 50 open issues, some of them definitively relevant as a memory leak.
    • It does not provide a GYP file, so we should do it (plus PR of course).
  • usrsctp-neat, which is a fork of usrsctp with same changes to make it single thread (which is far from perfect). Unfortunately it's not maintained at all and it does not merge new commits in the main stream project (sctplab/usrsctp).

  • rawrtc/rawrtc-data-channel, which is a layer on top of usrsctp-neat. Problems:

    • It depends on usrsctp-neat, so same concerns as above apply.
    • It also depends on re and rew C libraries (due the usage of some helpers and utils included in them). Not a super big problem, but I don't like the idea of having mediasoup depend on a C library that implements SIP, WebSocket, DNS, HTTP, etc.
    • It depends on meson and ninja for the building stage. This is a no-go for mediasoup. Users of mediasoup are supposed to just run npm install mediasoup and they are done. The "postinstall": "make -C worker" in package.json builds the mediasoup C++ code and its C/C++ deps (included in the source) by using gyp (which is also included in the package). All C/C++ dependencies of mediasoup include GYP files.
    • In other words, we should create GYP files for rawrtc-data-channel, usrsctp-neat, re and rew libraries...
  • medooze/libdatachannels, whose API matches the most what we need (single thread, and provides an API to send and receive packets by letting the external application manage the socket, timers, the DTLS, etc). Unfortunately it's nor finished neither active.

  • Another option is coding our own stack. Not desirable, but we have a clear design and whichever C/C++ SCTP stack we choose (or build) must conform to it.

Final Decision

We are gonna fork sctplab/usrsctp, make it single thread (as usrsctp-neat and a PR do) and add GYP support.

Specifications

  • peer-to-peer-data-api: Peer-to-Peer Data API
  • draft-ietf-rtcweb-transports: Transports for WebRTC
  • draft-ietf-rtcweb-data-channel: WebRTC Data Channels
  • draft-ietf-rtcweb-data-protocol: WebRTC Data Channel Establishment Protocol
    • This is for in-band DataChannel (pair of send & recv SCTP streams with same stream_id over the same STCP association). We (mediasoup) may prefer to use out-of-band negotiation (so DataChannel stream_id and other parameters are exchanged between client and server via signaling).
  • draft-ietf-mmusic-sctp-sdp: SDP Offer/Answer Procedures For SCTP over DTLS
    • Q: It's not clear to me which endpoint (DTLS client? DLTS server?) sends the SCTP INIT message. Section 9.3 says "When an SCTP association is established, both SCTP endpoints MUST initiate the SCTP association (i.e. both SCTP endpoints take the 'active' role)". I don't understand it at all. Isn't the SCTP association established with the INIT chunk?
    • R: The SCTP association is just determined by the SCTP ports negotiated in the SDP O/A (plus 'max-message-size', etc), so it "just" exists. Both endpoints must "initiate" it (via SCTP INIT) to setup the initial TSN and so on.
  • RFC 4960: Stream Control Transmission Protocol
  • RFC 3758: SCTP Partial Reliability Extension
  • RFC 5061: SCTP Dynamic Address Reconfiguration
    • Needed by RFC 6525.
  • RFC 6458: Sockets API Extensions for SCTP
  • RFC 6525: SCTP Stream Reconfiguration
  • RFC 7496: Additional Policies for the Partially Reliable SCTP Extension
    • Provides limited retransmission policy (limiting the number of retransmissions).
  • RFC 8260: Stream Schedulers and User Message Interleaving for SCTP
  • RFC 8261: DTLS Encapsulation of SCTP Packets