It could have been resumed without going into hibernation first, i.e.
when the client notices the disconnect before the server, or if it
switches networks etc.
Still having the connection on the session may cause unintentional
behavior, such as the session being treated as if connected, even tho
the connection has been closed.
To ensure that if a session is replaced after it has gone into
hibernation, it does not come back and cause trouble for the new session
(see previous commit).
The resumption_token is removed when the session is closed via the
pre-session-close event, signaling that it cannot be resumed, and
therefore no hibernation timeout logic should be invoked.
Fixes that if a session somehow is replaced by a new one using the same
resource (which is the common behavior), the old session would still be
around until it times out at which point it sends `<presence
type="unavailable"/>` which would look as if it came from the new
session, ie appearing offline to everyone including MUCs.
This overloads that flag a bit, but it has the intended effect of
stopping outgoing_stanza_filter() from queueing stanzas.
Fixes a traceback because of the queue having been removed somewhere
around here, since it is no longer needed.
Thanks Martin for reporting
This brings back the queue size limit that was once added, then removed
because destroying the session when reaching the limit was not great.
Instead, the queue wraps and overwrites the oldest unacked stanza on the
assumption that it will probably be acked anyway and thus does not need
to be delivered. If those discarded stanzas turn out to be needed on
resumption then the resumption fails.
All that was a complicated way to limit the number of resumable
sessions. Let's control resource usage some other way. This leaves the
essence of mapping resumption tokens to live sessions.
This keeps resumption state across reloads.
This allows clients that try to resume a session after a server restart
to at least know which of their pending outgoing stanzas were received
and which need to be re-sent.
This removes the limit on how many of those counters are kept, which
should be fixed eventually.
I have no idea what is going on in this code, which session is which?
Something has one of the sessions as an upvalue which is where the
filter checks for it.
As per a86ae74da96c the 'session' object here is the wrong session, so
the attempt to block stanzas from being added to the queue twice did not
work causing something of a leak.
Instead we have a leak of the previous session.
a86ae74da96c caused the stanza queue to double on resumption because
session.send() keeps a reference to the session which is what gets
passed to filters, so the added flag was not seen in the filter.
Taking advantage of the new callbacks added in dcf38ac6a38c and
9c450185bac1 avoids extra timers, extra syscalls and sending the `<r>`
in its own TCP segment, improving efficiency.
