Resuming a Session while it's waiting to auto-resume from a non-fatal error can crash

[sync-by-unito]

If you resume a session that has received a non-fatal error (i.e. InitialSyncNotCompleted) before the timer to auto-resume the session has fired and then you receive another non-fatal error you can crash when the sync client tries to create a new auto-resume timer.

[sync-by-unito]

➤ Jonathan Reams commented:

What are the cases where this could happen? When you call sync_manager->reconnect() it will eventually call sync::Session::cancel_reconnect_delay(), but only if the SyncSession is in the active state, which if you've received a fatal error the SyncSession should be in an Inactive state - so I don't see how calling reconnect would re-use an abandoned SessionWrapper.

[sync-by-unito]

➤ danieltabacaru commented:

Can't there be a race between the SyncSession becoming inactive (as a result of SessionImpl receiving a fatal error) and reconnect being called? There may be other cases when this can happen though.

[sync-by-unito]

➤ Jonathan Reams commented:

Got it, I've reproduced this in a test. I think I know what to do now.

[sync-by-unito]

➤ Jonathan Reams commented:

I was able to reproduce the original description of this ticket but couldn't actually get it to crash. So I went back to the logs and was able to reproduce the exact crash in RNET-1051 by calling SyncSession::handle_reconnect() after two successive retryable errors without having the try again timer complete between the two of them. I've added a test that reliably reproduces this and the fix is to just cancel the try again timer if we're explicitly resuming the session.