831ae4ef65
This is a reland of d4089cae47334a4228b69d6bb23f2e49ebb7496e with the following fix: Invoke MaybeStartGathering as the last step of DoSetLocalDescription. This ensures that candidates and onicegatheringstatechange does not happen before SLD is resolved. This is important for passing external/wpt/webrtc/RTCPeerConnection-iceGatheringState.html. Original change's description: > [Perfect Negotiation] Implement non-racy version of SetLocalDescription. > > BACKGROUND > > When SLD is invoked with SetSessionDescriptionObserver, the observer is > called by posting a message back to the execution thread, delaying the > call. This delay is "artificial" - it's not necessary; the operation is > already complete. It's a post from the signaling thread to the signaling > thread. The rationale for the post was to avoid the observer making > recursive calls back into the PeerConnection. The problem with this is > that by the time the observer is called, the PeerConnection could > already have executed other operations and modified its states. > > This causes the referenced bug: one can have a race where SLD is > resolved "too late" (after a pending SRD is executed) and the signaling > state observed when SLD resolves doesn't make sense. > > When implementing Unified Plan, we fixed similar issues for SRD by > adding a version that takes SetRemoteDescriptionObserverInterface as > argument instead of SetSessionDescriptionObserver. The new version did > not have the delay. The old version had to be kept around not to break > downstream projects that had dependencies both on he delay and on > allowing the PC to be destroyed midst-operation without informing its > observers. > > THIS CL > > This does the old SRD fix for SLD as well: A new observer interface is > added, SetLocalDescriptionObserverInterface, and > PeerConnection::SetLocalDescription() is overloaded. If you call it with > the old observer, you get the delay, but if you call it with the new > observer, you don't get a delay. > > - SetLocalDescriptionObserverInterface is added. > - SetLocalDescription is overloaded. > - The adapter for SetSessionDescriptionObserver that causes the delay > previously only used for SRD is updated to handle both SLD and SRD. > - FakeSetLocalDescriptionObserver is added and > MockSetRemoteDescriptionObserver is renamed "Fake...". > > Bug: chromium:1071733 > Change-Id: I920368e648bede481058ac22f5b8794752a220b3 > Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/179100 > Commit-Queue: Henrik Boström <hbos@webrtc.org> > Reviewed-by: Harald Alvestrand <hta@webrtc.org> > Cr-Commit-Position: refs/heads/master@{#31798} TBR=hta@webrtc.org Bug: chromium:1071733 Change-Id: Ic6e8d96afa1c19604762f373716c08dbfa9d178c Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/180481 Commit-Queue: Henrik Boström <hbos@webrtc.org> Reviewed-by: Henrik Boström <hbos@webrtc.org> Cr-Commit-Position: refs/heads/master@{#31804}
How to write code in the api/ directory
Mostly, just follow the regular style guide, but:
- Note that
api/code is not exempt from the “.hand.ccfiles come in pairs” rule, so if you declare something inapi/path/to/foo.h, it should be defined inapi/path/to/foo.cc. - Headers in
api/should, if possible, not#includeheaders outsideapi/. It’s not always possible to avoid this, but be aware that it adds to a small mountain of technical debt that we’re trying to shrink. .ccfiles inapi/, on the other hand, are free to#includeheaders outsideapi/.
That is, the preferred way for api/ code to access non-api/ code is to call
it from a .cc file, so that users of our API headers won’t transitively
#include non-public headers.
For headers in api/ that need to refer to non-public types, forward
declarations are often a lesser evil than including non-public header files. The
usual rules still apply, though.
.cc files in api/ should preferably be kept reasonably small. If a
substantial implementation is needed, consider putting it with our non-public
code, and just call it from the api/ .cc file.