* Disable audio loss proxying.
Added a config which is off by default.
With audio NACKs, that is the preferred repair mechanism.
With RED, repair is built in via packet redundancy to recover from
isolated losses.
So, proxying is not required. But, leaving it in there with a config
that is disabled by default.
* fix test
* Move caching of publisher sender report to subscriber side.
Please see inline for descriptive comments on why. Basically,
pause/unpause using replaceTrack(null)/replaceTrack(actualTrack) can
cause time stamp in sender report sent to subscribers jump ahead.
This prevents that.
With the caching on subscriber side, cleaning up the caching on
publisher side.
* fix compile, test still failing, need to debug
* skip reference TS for testing
With the change in https://github.com/livekit/livekit/pull/2611,
the dummy receiver was replaced with real receiver. But, the close check
was using the dummy receiver.
Doing two things
- Use the dummy receiver post upgrade also
(NOTE: this is not needed, but just keeping old behaviour)
- Fix the close check to count number of open receivers.
When migrating out, it is possible that a published layer is not
notified to the migrating in node. Reduce chances of that layer not
getting published to the new node by curbing RTCP during migration.
It could still happen if stars line up, but this should reduce the
window to a much smaller one.
The receiver should not change, but code wise, the option of replacing
receiver object makes more sense, i.e. otherwise, it could look like we
are leaving the stale object in there without replacing with new
receiver of same type.
Another case of duplciate tracks in SDP.
During migration (if both publisher and subscriber migrate), subscriber
could attach the remote track of the publisher. But, while that is
happening, publisher could migrate into the node and close the remote
media track. This was causing subscriber to switch from attaching to
remote media track -> attaching to local media track.
But, as remote media track was closed while add subscription was
happening, the subscriber is removed without subscription manager being
aware of it.
So, the subscription manager's reconcile and the remove subscriber is
racing and when subscription manager re-subscribes, caching has not run
yet and that creates a duplicate.
Delay removing subscribed track till after caching is done. That means,
even if the reconciler runs, it will get an `errAlreadySubscribed` error
and it will force it to reconcile again. By the time the subscribed
track is deleted from the subscriptions map, caching is done.
* Notify initial permissions
NOTE: This does add an initial subscription permission notification
which should be fine, but something to watch for.
A stress test combining
- mute/unmute on publisher side.
- allowing/revoking permission for subscriber from publisher side.
- subscribing/unsubscribing from subscriber side.
results in a scenario where a subscription permission update of
`not_allowed` being sent and on a re-subscribe, an `allowed` update does
not happen.
It happens like so
- Subscription revoke cloes the down track of subscriber.
- The subscription is still desired.
- So, a subscription reconcile runs and sees `permission: false`. This
sends subscription permission of `not_allowed`.
- Unsubscribe request comes in and sets `desired: false`.
- Reconsiler runs again and sees `desired: false` and `subscribedTrack:
nil`. This cleans up the subscription.
- Publisher grants permission for the subscriber.
- Subscriber subscribes to the track again. A new subscription is
created.
- Reconciler runs and sees `permission: true`, but there is no
permission change as it is a new subscription object. So, `allowed`
subscription permission update is not sent and the client is stuck at
`not_allowed`.
Fix, maintain if permission has been initialized. Has the effect of
sending an initial update which should be fine.
* clean up comment
* no default
On migration, when subscription moved from remote -> local,
transceiver caching was racing. Although a very small possibility,
it could happen like so
1. down track close
2. down track close callback fires go routine to close subscribed track
3. subscribed track close handler in subscription manager tries to
reconcile
4. reconcile adds subscribed track again
5. cannot find cached transceiver as caching happens after down track
close finishes in stap 1 above. Although there are a couple of
gortouine jumps (step 2 fires a goroutine to close subscribed track
and step 4 will reconcile in a goroutine too), it is theoretically
possible that the step 1 has not finished and hence transceiver is
not cached.
Fix is to move caching to before closing subscribed track.
* Forward publisher sender report.
Publisher side RTCP sernfer report is rebased to SFU time base
and used to send sender rerport to subscriber.
Will wait to merge till previous versions are out as this will require a
bunch of testing.
* - Add rebased report drift
- update protocol dep
- fix path change check, it has to check against delta of propagation
delay and not propagation delay as the two side clocks could be way
off.
It is possible that migration could trigger without migrating out node
knowing about it. So, when a migration started notification comes in,
set up migration timer if not already set.
* Support XR request/response for rtt calculation
* Update pkg/sfu/downtrack.go
Co-authored-by: David Zhao <dz@livekit.io>
---------
Co-authored-by: David Zhao <dz@livekit.io>
* Remove subscriber if track closed while adding subscriber.
It is possible that the track is closed when subscriber add is
processed. That subscriber would have been dangling off a closed track.
Check again after adding subscriber if track is closed.
If it is, remove the subscriber and return error so that subscription
manager re-resolves.
* oops, wrong unlock
When a fallback is not applied, it is due to signal interruption.
ICE connection failing happens. And every time there is error, it is due
to "no selected pair".
Move all of it to `Debugw`. `setting ICE config` is the definitive log
which says if a different ICE config was applied.
* Maintain subscription count.
Does not affect function as it is not decremented only if limits are
configured. But, good to maintain proper count anyway.
* wire
Simplify a bit. Pending migratiion tracks need not be maintained as when
a migrated track is added, it is added to up track manager and treated
as a published track. When up track manager closes, published tracks
will close. So, no need to maintain a separate list.
* Buffer size config for video and audio.
There was only one buffer size in config.
In upstream, config value was used for video.
Audio used a hard coded value of 200 packets.
But, in the down stream sequencer, the config value was used for both
video and audio. So, if video was set up for high bit rate (deep
buffers), audio sequencer ended up using a lot of memory too in
sequencer.
Split config to be able to control that and also not hard code audio.
Another optimisation here would be to not instantiate sequencer unkess
NACK is negotiated.
* deprecate packet_buffer_size
It is possible that participant state and migration state updates are
racing. And a participant update could end up with no tracks when
migration is being processed.
Moving handling of migrated tracks to when the migration state moves to
completed. Pending data channel were already happening only on complete.
Move tracks also to that point.
Handling it earlier meant that track published callback happened and
ownership of track moved to new node before the new node could finish
peer connection. So, in cases where migration did not go through, this
caused confusion of track ownership.
