Reconnect errors

[pblazej]

Resolves #844

• Handles offerId (symptoms)
• Fixes double offer on ICE restart (root cause)
• Fixes deadlock, probably introduced by strict ordering in Fix serial runner cancellation #804 (the fix is still valid though)
• Handles leaveAction param vs legacy canReconnect

[github-actions]

⚠️ This PR does not contain any files in the .changes directory.

[pblazej]

JS around https://github.com/livekit/client-sdk-js/blob/65e0bfe38ba594bbfe73de60ce72dbc7b96be3a2/src/room/PCTransport.ts#L271

if (this._pc && this._pc.signalingState === 'have-local-offer') {
    const currentSD = this._pc.remoteDescription;
    if (options?.iceRestart && currentSD) {
        // 1. Rollback: Sets remote description
        await this._pc.setRemoteDescription(currentSD);
        // 2. DOES NOT set this.renegotiate = true
        // 3. Falls through to create offer (line 291)
    } else {
        // 1. Defer: Sets renegotiate = true
        this.renegotiate = true;
        // 2. Returns immediately (skips offer creation)
        return;
    }
}

vs Swift

if signalingState == .haveLocalOffer {
    if !(iceRestart && remoteDescription != nil) {
        // 1. Defer: Sets _reNegotiate = true
        _reNegotiate = true
        // 2. Returns immediately
        return
    }
    
    // Else: ICE Restart path falls through...
}

// ... (offer ID increment) ...

if signalingState == .haveLocalOffer, iceRestart, let sd = remoteDescription {
    // 1. Force clear _reNegotiate (Matches JS not setting it)
    _reNegotiate = false  
    // 2. Rollback: Sets remote description
    try await set(remoteDescription: sd)
    // 3. Creates new offer immediately
    return try await _negotiateSequence()
}

[pblazej]

SerialRunnerActor (inside SignalClient._delegate)
│
├─> [Task 1] didUpdateConnectionState
│   └─> await startReconnect()  ← blocking the actor
│       └─> waiting for offer...
│
└─> [Task 2] didReceiveOffer
    └─> Can't enter because actor is busy with Task 1

[pblazej]

@hiroshihorie this ain't trivial, I'd appreciate if you give it a spin in ✈️ mode

[pblazej]

This is more or less equivalent to JS:

        case LeaveRequest_Action.RECONNECT:
          this.fullReconnectOnNext = true;
          // reconnect immediately instead of waiting for next attempt
          this.handleDisconnect(leaveReconnect);

[lukasIO]

suggestion (if-minor): maybe this is a good moment to adopt the .action of the leave request as well here? (backwards compatible of course in case it's unset/0).

We don't actually need a full reconnect on every leave request. But if it makes more sense as a follow up that also sounds good.

[pblazej]

Good point, moved to the action param here 👍

Re: backwards compatibility, I discarded the legacy canReconnect param, which mimics JS and no-op if unknown.

[lukasIO]

looks good to me in general!

[lukasIO]

suggestion (if-minor): maybe this is a good moment to adopt the .action of the leave request as well here? (backwards compatible of course in case it's unset/0).

We don't actually need a full reconnect on every leave request. But if it makes more sense as a follow up that also sounds good.