When synapse receives an event for a room its not in over federation, it
double checks with the remote server to see if it is in fact in the
room. This is done so that if the server has forgotten about the room
(usually as a result of the database being dropped) it can recover from
it.
However, in the presence of state resets in large rooms, this can cause
a lot of work for servers that have legitimately left. As a hacky
solution that supports both cases we drop incoming events for rooms that
we have explicitly left.
This means that we no longer support the case of servers having
forgotten that they've rejoined a room, but that is sufficiently rare
that we're not going to support it for now.
preserve_context_over_fn is essentially broken, because (a) it pointlessly
drops the current logcontext before calling its wrapped function, which means
we don't get any useful logcontexts for _handle_key_deferred; (b) it wraps the
resulting deferred in a _PreservingContextDeferred, which is very dangerous
because you then can't yield on it without leaking context back into the
reactor.
Instead, let's specify that the resultant deferreds call their callbacks with
no logcontext.
... which means that logcontexts can be correctly preserved for the stuff it
does.
get_server_verify_keys is now called with the logcontext, so needs to
preserve_fn when it fires off its nested inlineCallbacks function.
Also renames get_server_verify_keys to reflect the fact it's meant to be
private.
If the verify_request.deferred has already completed, then `remove_deferreds`
will be called immediately. It therefore might resolve the server_to_deferred
deferred while there are still other requests for that server in flight.
To avoid that, we should build the complete list of requests, and *then* add the
callbacks.
Define that it is run with no log context, and make sure that happens.
If we aren't careful to reset the logcontext, we can't bung the deferreds into
defer.gatherResults etc. We don't actually do that directly, but we *do*
resolve other deferreds from affected callbacks (notably the server_to_deferred
map in _start_key_lookups), and those *do* get passed into
defer.gatherResults. It turns out that this way ends up being least confusing.
