Establishing a connection: The WebRTC perfect negotiation pattern - Web APIs 编辑

Because WebRTC doesn't mandate a specific transport mechanism for signaling during the negotiation of a new peer connection, it's highly flexible. However, despite that flexibility in transport and communication of signaling messages, there's still a recommended design pattern you should follow when possible, known as perfect negotiation. This article introduces WebRTC perfect negotiation, describing how it works and why it's the recommended way to negotiate a WebRTC connection between peers, and provides sample code to demonstrate the technique.

After the first deployments of WebRTC-capable browsers, it was realized that parts of the negotiation process were more complicated than they needed to be for typical use cases. This was due to a small number of issues with the API and some potential race conditions that needed to be prevented. These issues have since been addressed, letting us simplify our WebRTC negotiation significantly. The perfect negotiation pattern is an example of the ways in which negotiation have improved since the early days of WebRTC.

Perfect negotiation concepts

Perfect negotiation makes it possible to seamlessly and completely separate the negotiation process from the rest of your application's logic. Negotiation is an inherently asymmetric operation: one side needs to serve as the "caller" while the other peer is the "callee." The perfect negotiation pattern smooths this difference away by separating that difference out into independent negotiation logic, so that your application doesn't need to care which end of the connection it is. As far as your application is concerned, it makes no difference whether you're calling out or receiving a call.

The best thing about perfect negotiation is that the same code is used for both the caller and the callee, so there's no repetition or otherwise added levels of negotiation code to write.

Perfect negotiation works by assigning each of the two peers a role to play in the negotiation process that's entirely separate from the WebRTC connection state:

• A polite peer, which uses ICE rollback to prevent collisions with incoming offers. A polite peer, essentially, is one which may send out offers, but then responds if an offer arrives from the other peer with "Okay, never mind, drop my offer and I'll consider yours instead."
• An impolite peer, which always ignores incoming offers that collide with its own offers. It never apologizes or gives up anything to the polite peer. Any time a collision occurs, the impolite peer wins.

This way, both peers know exactly what should happen if there are collisions between offers that have been sent. Responses to error conditions become far more predictable.

How you determine which peer is polite and which is impolite is generally up to you. It could be as simple as assigning the polite role to the first peer to connect to the signaling server, or you could do something more elaborate like having the peers exchange random numbers and assigning the polite role to the winner. However you make the determination, once these roles are assigned to the two peers, they can then work together to manage signaling in a way that doesn't deadlock and doesn't require a lot of extra code to manage.

An important thing to keep in mind is this: the roles of caller and callee can switch during perfect negotiation. If the polite peer is the caller and it sends an offer but there's a collision with the impolite peer, the polite peer drops its offer and instead replies to the offer it has received from the impolite peer. By doing so, the polite peer has switched from being the caller to the callee!

Implementing perfect negotiation

Let's take a look at an example that implements the perfect negotiation pattern. The code assumes that there's a SignalingChannel class defined that is used to communicate with the signaling server. Your own code, of course, can use any signaling technique you like.

Note that this code is identical for both peers involved in the connection.

Create the signaling and peer connections

First, the signaling channel needs to be opened and the RTCPeerConnection needs to be created. The STUN server listed here is obviously not a real one; you'll need to replace stun.myserver.tld with the address of a real STUN server.

const constraints = { audio: true, video: true };
const config = {
  iceServers: [{ urls: "stun:stun.mystunserver.tld" }]
};

const selfVideo = document.querySelector("video.selfview");
const remoteVideo = document.querySelector("video.remoteview");

const signaler = new SignalingChannel();
const pc = new RTCPeerConnection(config);

This code also gets the <video> elements using the classes "selfview" and "remoteview"; these will contain, respectively, the local user's self-view and the view of the incoming stream from the remote peer.

Connecting to a remote peer

The start() function shown here can be called by either of the two end-points that want to talk to one another. It doesn't matter who does it first; the negotiation will just work.

async function start() {
  try {
    const stream = await navigator.mediaDevices.getUserMedia(constraints);

    for (const track of stream.getTracks()) {
      pc.addTrack(track, stream);
    }
    selfVideo.srcObject = stream;
  } catch(err) {
    console.error(err);
  }
}

This isn't appreciably different from older WebRTC connection establishment code. The user's camera and microphone are obtained by calling getUserMedia(). The resulting media tracks are then added to the RTCPeerConnection by passing them into addTrack(). Then, finally, the media source for the self-view <video> element indicated by the selfVideo constant is set to the camera and microphone stream, allowing the local user to see what the other peer sees.

Handling incoming tracks

We next need to set up a handler for track events to handle inbound video and audio tracks that have been negotiatied to be received by this peer connection. To do this, we implement the RTCPeerConnection's ontrack event handler.

pc.ontrack = ({track, streams}) => {
  track.onunmute = () => {
    if (remoteVideo.srcObject) {
      return;
    }
    remoteVideo.srcObject = streams[0];
  };
};

When the track event occurs, this handler executes. Using destructuring, the RTCTrackEvent's track and streams properties are extracted. The former is either the video track or the audio track being received. The latter is an array of MediaStream objects, each representing a stream containing this track (a track may in rare cases belong to multiple streams at once). In our case, this will always contain one stream, at index 0, because we passed one stream into addTrack() earlier.

We add an unmute event handler to the track, because the track will become unmuted once it starts receiving packets. We put the remainder of our reception code in there.

If we already have video coming in from the remote peer (which we can see if the remote view's <video> element's srcObject property already has a value), we do nothing. Otherwise, we set srcObject to the stream at index 0 in the streams array.

The perfect negotiation logic

Now we get into the true perfect negotiation logic, which functions entirely independently from the rest of the application.

Handling the negotiationneeded event

First, we implement the RTCPeerConnection event handler onnegotiationneeded to get a local description and send it using the signaling channel to the remote peer.

let makingOffer = false;

pc.onnegotiationneeded = async () => {
  try {
    makingOffer = true;
    await pc.setLocalDescription();
    signaler.send({ description: pc.localDescription });
  } catch(err) {
    console.error(err);
  } finally {
    makingOffer = false;
  }
};

Note that setLocalDescription() without arguments automatically creates and sets the appropriate description based on the current signalingState. The set description is either an answer to the most recent offer from the remote peer or a freshly-created offer if there's no negotiation underway. Here, it will be always be an offer, because the negotiationneeded event is only fired in stable state.

We set a Boolean variable, makingOffer to true to mark that we're preparing an offer. To avoid races, we'll use this value later instead of the signaling state to determine whether or not an offer is being processed because the value of signalingState changes asynchronously, introducing a glare opportunity.

Once the offer has been created, set and sent (or an error occurs), makingOffer gets set back to false.

Handling incoming ICE candidates

Next, we need to handle the RTCPeerConnection event icecandidate, which is how the local ICE layer passes candidates to us for delivery to the remote peer over the signaling channel.

pc.onicecandidate = ({candidate}) => signaler.send({candidate});

This takes the candidate member of this ICE event and passes it through to the signaling channel's send() method to be sent over the signaling server to the remote peer.

Handling incoming messages on the signaling channel

The last piece of the puzzle is code to handle incoming messages from the signaling server. That's implemented here as an onmessage event handler on the signaling channel object. This method is invoked each time a message arrives from the signaling server.

let ignoreOffer = false;

signaler.onmessage = async ({ data: { description, candidate } }) => {
  try {
    if (description) {
      const offerCollision = (description.type == "offer") &&
                             (makingOffer || pc.signalingState != "stable");

      ignoreOffer = !polite && offerCollision;
      if (ignoreOffer) {
        return;
      }

      await pc.setRemoteDescription(description);
      if (description.type == "offer") {
        await pc.setLocalDescription();
        signaler.send({ description: pc.localDescription })
      }
    } else if (candidate) {
      try {
        await pc.addIceCandidate(candidate);
      } catch(err) {
        if (!ignoreOffer) {
          throw err;
        }
      }
    }
  } catch(err) {
    console.error(err);
  }
}

Upon receiving an incoming message from the SignalingChannel through its onmessage event handler, the received JSON object is destructured to obtain the description or candidate found within. If the incoming message has a description, it's either an offer or an answer sent by the other peer.

If, on the other hand, the message has a candidate, it's an ICE candidate received from the remote peer as part of trickle ICE. The candidate is destined to be delivered to the local ICE layer by passing it into addIceCandidate().

On receiving a description

If we received a description, we prepare to respond to the incoming offer or answer. First, we check to make sure we're in a state in which we can accept an offer. If the connection's signaling state isn't stable or if our end of the connection has started the process of making its own offer, then we need to look out for offer collision.

If we're the impolite peer, and we're receiving a colliding offer, we return without setting the description, and instead set ignoreOffer to true to ensure we also ignore all candidates the other side may be sending us on the signaling channel belonging to this offer. Doing so avoids error noise since we never informed our side about this offer.

If we're the polite peer, and we're receiving a colliding offer, we don't need to do anything special, because our existing offer will automatically be rolled back in the next step.

Having ensured that we want to accept the offer, we set the remote description to the incoming offer by calling setRemoteDescription(). This lets WebRTC know what the proposed configuration of the other peer is. If we're the polite peer, we will drop our offer and accept the new one.

If the newly-set remote description is an offer, we ask WebRTC to select an appropriate local configuration by calling the RTCPeerConnection method setLocalDescription() without parameters. This causes setLocalDescription() to automatically generate an appropriate answer in response to the received offer. Then we send the answer through the signaling channel back to the first peer.

On receiving an ICE candidate

On the other hand, if the received message contains an ICE candidate, we deliver it to the local ICE layer by calling the RTCPeerConnection method addIceCandidate(). If an error occurs and we've ignored the most recent offer, we also ignore any error that may occur when trying to add the candidate.

Making negotiation perfect

If you're curious what makes perfect negotiation so... perfect... this section is for you. Here, we'll look at each change made to the WebRTC API and to best practice recommendations to make perfect negotiation possible.

Glare-free setLocalDescription()

In the past, the negotiationneeded event was easily handled in a way that was susceptible to glare—that is, it was prone to collisions, where both peers could wind up attempting to make an offer at the same time, leading to one or the other peers getting an error and aborting the connection attempt.

The old way

Consider this onnegotiationneeded event handler:

pc.onnegotiationneeded = async () => {
  try {
    await pc.setLocalDescription(await pc.createOffer());
    signaler.send({description: pc.localDescription});
  } catch(err) {
    console.error(err);
  }
};

Because the createOffer() method is asynchronous and takes some time to complete, there's time in which the remote peer might attempt to send an offer of its own, causing us to leave the stable state and enter the have-remote-offer state, which means we are now waiting for a response to the offer. But once it receives the offer we just sent, so is the remote peer. This leaves both peers in a state in which the connection attempt cannot be completed.

Perfect negotiation with the updated API

As shown in the section Implementing perfect negotiation, we can eliminate this problem by introducing a variable (here called makingOffer) which we use to indicate that we are in the process of sending an offer, and making use of the updated setLocalDescription() method:

let makingOffer = false;

pc.onnegotiationneeded = async () => {
  try {
    makingOffer = true;
    await pc.setLocalDescription();
    signaler.send({ description: pc.localDescription });
  } catch(err) {
    console.error(err);
  } finally {
    makingOffer = false;
  }
};

We set makingOffer immediately before calling setLocalDescription() in order to lock against interfering with sending this offer, and we don't clear it back to false until the offer has been sent to the signaling server (or an error has occurred, preventing the offer from being made). This way, we avoid the risk of offers colliding.

Automatic rollback in setRemoteDescription()

A key component to perfect negotiation is the concept of the polite peer, which always rolls itself back if it receives an offer while itself waiting for an answer to an offer. Previously, triggering rollback involved manually checking for rollback conditions and triggering the rollback manually, by setting the local description to one with the type rollback, like this:

await pc.setLocalDescription({ type: "rollback" });

Doing so returns the local peer to the stable signalingState from whichever state it had previously been in. Since a peer can only accept offers when in the stable state, the peer has thus rescinded its offer and is ready to receive the offer from the remote (impolite) peer. As we'll see in a moment, there are problems with this approach, however.

Perfect negotiation with the old API

Using the previous API to implement incoming negotiation messages during perfect negotiation would look something like this:

signaler.onmessage = async({data: { description, candidate }}) => {
  try {
    if (description) {
      if (description.type == "offer" && pc.signalingState != "stable") {
        if (!polite) {
          return;
        }

        await Promise.all([
          pc.setLocalDescription({type: "rollback"}),
          pc.setRemoteDescription(description);
        ]);
      } else {
        await pc.setRemoteDescription(description);
      }

      if (description.type == "offer") {
        await pc.setLocalDescription(await pc.createAnswer());
        signaler.send({ description: pc.localDescription });
      }
    } else if (candidate) {
      try {
        await pc.addIceCandidate(candidate);
      } catch(err) {
        if (!ignoreOffer) {
          throw err;
        }
      }
    }
  } catch(err) {
    console.error(err);
  }
};

Since rollback works by postponing changes until the next negotiation (which will begin immediately after the current one is finished), the polite peer needs to know when it needs to throw away a received offer if it's currently waiting for a reply to an offer it's already sent.

The code checks to see if the message is an offer, and if so, if the local signaling state isn't stable. If it's not stable, and the local peer is the polite one, we need to trigger rollback so we can replace the outgoing offer with the new incoming one. and these must both be completed before we can proceed with handling the received offer.

Since there isn't a single "roll back and use this offer instead", performing this change on the polite peer requires two steps, executed in the context of Promise.all(), which is used to ensure that both statements execute completely before continuing to handle the received offer. The first statement triggers rollback and the second sets the remote description to the received one, thus completing the process of replacing the previously sent offer with the newly received offer. The impolite peer has now become the callee instead of the caller.

Alll other descriptions received from the impolite peer are processed as normal, by passing them into setRemoteDescription().

Finally, we process a received offer by calling setLocalDescription() to set our local description to the one returned by createAnswer(). Then that gets sent to the polite peer using the signaling channel.

If the incoming message is an ICE candidate rather than an SDP description, it's delivered to the ICE layer by passing it into the RTCPeerConnection method addIceCandidate(). If an error occurs here and we didn't just discard an offer due to being the impolite peer during a collision, we throw the error so the caller can handle it. Otherwise, we drop the error, ignoring it, since it doesn't matter in this context.

Perfect negotiation with the updated API

The updated code takes advantage of the fact that you can now call setLocalDescription() with no parameters so it just does the right thing for you, as well as the fact that setRemoteDescription() automatically rolls back if necessary. This lets us get rid of the need to use a Promise to keep the timing in order, since the rollback becomes an essentially atomic part of the setRemoteDescription() call.

let ignoreOffer = false;

signaler.onmessage = async ({ data: { description, candidate } }) => {
  try {
    if (description) {
      const offerCollision = (description.type == "offer") &&
                             (makingOffer || pc.signalingState != "stable");

      ignoreOffer = !polite && offerCollision;
      if (ignoreOffer) {
        return;
      }

      await pc.setRemoteDescription(description);
      if (description.type == "offer") {
        await pc.setLocalDescription();
        signaler.send({ description: pc.localDescription });
      }
    } else if (candidate) {
      try {
        await pc.addIceCandidate(candidate);
      } catch(err) {
        if (!ignoreOffer) {
          throw err;
        }
      }
    }
  } catch(err) {
    console.error(err);
  }
}

While the difference in code size is minor, and the complexity isn't reduced much either, the code is much, much more reliable. Let's take a dive into the code to see how it works now.

On receiving a description

In the revised code, if the received message is an SDP description, we check to see if it arrived while we're attempting to transmit an offer. If the received message is an offer and the local peer is the impolite peer, and a collision is occurring, we ignore the offer because we want to continue to try to use the offer that's already in the process of being sent. That's the impolite peer in action.

In any other case, we'll try instead to handle the incoming message. This begins by setting the remote description to the received description by passing it into setRemoteDescription(). This works regardless of whether we're handling an offer or an answer since rollback will be performed automaticaly as needed.

At that point, if the received message is an offer, we use setLocalDescription() to create and set an appropriate local description, then we send it to the remote peer over the signaling server.

On receiving an ICE candidate

On the other hand, if the received message is an ICE candidate—indicated by the JSON object containing a candidate member—we deliver it to the local ICE layer by calling the RTCPeerConnection method addIceCandidate(). Errors are, as before, ignored if we have just discarded an offer.

Explicit restartIce() method added

The techniques previously used to trigger an ICE restart while handling the eventnegotiationneeded have significant flaws. These flaws have made it difficult to safely and reliably trigger a restart during negotiation. The perfect negotiation improvements have fixed this by adding a new restartIce() method to RTCPeerConnection.

The old way

In the past, if you encountered an ICE error and needed to restart negotiation, you might have done something like this:

pc.onnegotiationneeded = async options => {
  await pc.setLocalDescription(await pc.createOffer(options));
  signaler.send({ description: pc.localDescription });
};
pc.oniceconnectionstatechange = () => {
  if (pc.iceConnectionState === "failed") {
    pc.onnegotiationneeded({ iceRestart: true });
  }
};

This has a number of reliability issues and outright bugs (such as failing if the iceconnectionstatechange event fires when the signaling state isn't stable), but there was no way you could actually request an ICE restart other than by creating and sending an offer with the iceRestart option set to true. Sending the restart request thus required directly invoking the negotiationneeded event's handler. Getting it right was tricky at best, and was so easy to get wrong that bugs are common.

Using restartIce()

Now, you can use restartIce() to do this much more cleanly:

pc.onnegotiationneeded = async options => {
  await pc.setLocalDescription(await pc.createOffer(options));
  signaler.send({ description: pc.localDescription });
};
pc.oniceconnectionstatechange = () => {
  if (pc.iceConnectionState === "failed") {
    pc.restartIce();
  }
};

With this improved technique, instead of directly calling onnegotiationneeded with options to trigger ICE restart, the failed ICE connection state calls restartIce(). restartIce() tells the ICE layer to automatically add the iceRestart flag to the next ICE message sent. Problem solved!

Rollback no longer supported in the pranswer state

The last of the API changes that stand out is that you can no longer roll back when in either of the have-remote-pranswer or the have-local-pranswer states. Fortunately, when using perfect negotiation there's no need to do this anyway, since the situations that cause this are caught and prevented before rolling these back ever becomes necessary.

Thus, attempting to trigger rollback while in one of the two pranswer states will now throw an InvalidStateError.

See also

• WebRTC API
• Lifetime of a WebRTC session