This test was incorrectly exercising a member of the endpoint record returned
by `Manager.get_endpoints()`. In the test environment, the `.listen` Endpoint
is actually a Mock, so calling e.g. `listen()` on `endpoints.listen` returns
another Mock instead of a Deferred. Twisted's `assertNoResult` used to
tolerate this silently, but as of Twisted-19.10 it throws an error, causing
the test to fail.
The fix is to assert that the record has attributes with the right names, but
not assume they behave like normal Endpoints, and not call `.listen()` on
them.
closes#366
The previous implementation would call the control/receiving Protocol
completely backwards: dataReceived first, then connectionLost, then finally
connectionMade. Which didn't work at all, of course.
This should fix the immediate issue of the remote side opening a
subchannel (and sending data on it) before the local side even sees the
Endpoints, so before it can register a listening factory to receive the OPEN.
We were already buffering early OPENs in the SubchannelListenerEndpoint, but
this makes sure that endpoint is available (for the manager's Inbound half to
deliver) them as soon as the dilation connection is established.
The downside to buffering OPENs (and all data written to inbound subchannels)
is that the application has no way to reject or pause them, until it
registers the listening factory. If the application never calls
`listen_ep.listen()`, we'll buffer this data forever (or until the wormhole
is closed). The upside is that we don't lose a roundtrip waiting for an ack
on the OPEN. See ticket #335 for more details.
refs #335
I think I just managed to forget that inbound_close requires we respond with
a close ourselves. Also outbound open means we must add the subchannel to the
inbound table, so we can receive any data on it at all.
I'm sure I had a good reason for avoiding integers, but it makes logging and
testing more difficult, and both sides are using integers to generate them
anyways (so one side can pick the odd ones, and the other can pick the even
ones).
Without this, the Follower would see data for subchannel 0 before it had a
chance to create the SubChannel object that could accept it. We already have
a mechanism for inbound data to be queued inside the SubChannel until the
endpoint has had a chance to create the Protocol object: we rely on that
mechanism here. We just need to create the SubChannel before telling the
Manager to start, even though we don't reveal the SubChannel to the
caller (via the control endpoint) until the connection is known to succeed.
This helps a manual test get data from one side to the other without throwing
exceptions.
When the follower's connection is accepted, they'll observe a single
dataReceived chunk containing both the leader's KCM and the leader's first
actual data record. The state machine considers the KCM for an eventual-turn
before selecting the connection, so the data record will arrive while the
connection isn't quite ready for it (if consider() were immediate, this
wouldn't be a problem, but Automat doesn't deal with reentrant calls very
well). So we queue any records that arrive before we're selected.
Dilator.stop() now shuts everything down, and returns a Deferred when it all
stops moving. This needed some Manager state machine changes (to notify
Dilator when it enters the STOPPED state). This also revealed problems in the
delivery of connector_connection_made() (which was misnamed) and
connector_connection_lost() (which wasn't being called at all).
This is only used by tests so far (and will simplify the integration test
that hasn't landed yet), but is not yet wired up to Boss, so there's no way
for applications to enable it yet.
I mistakenly believed that Noise handshakes are simultaneous. In fact, the
Responder waits until it sees the Initiator's handshake before sending its
own. I had to update the Connection state machines to work this way (the
Record machine now has set_role_leader and set_role_follower), and update the
tests to match.
For debugging I added a `_role` property to Record, but it should probably be
removed.
only install 'noiseprotocol' (which is necessary for dilation to work) if the
"dilate" feature is requested (e.g. `pip install magic-wormhole[dilate]`)
