RNSChannelOutlet.send() can return a packet that never reached the wire
(link not ACTIVE, no capable interface, etc). The old Channel.send()
queued the envelope in _tx_ring before calling outlet.send(), then
tried to rewind _next_sequence and remove the envelope if the outlet
returned a failed packet. Two problems:
- Between queueing and outlet.send() returning, _tx_ring held an
envelope with packet.raw=None. Any worker thread iterating the
ring (timeout fire, proof callback) crashed in get_packet_id's
packet.get_hash() with a TypeError on None.raw.
- The rewind was only safe for a single-threaded sender: it checked
"is _next_sequence one past mine?" and skipped the rewind otherwise.
Under concurrent senders, the rewind silently failed, leaving a
hole in the on-wire sequence stream. The receiver's contiguous
seqnum rule then stalled the channel permanently with no error.
This fix serializes the reservation-and-transmit pair with a per-channel
_send_lock so the rewind is always correct, and defers queueing until
outlet.send() returns a real packet so _tx_ring never contains a
packet-less envelope. _packet_tx_op() and get_packet_id() now also
defensively skip/return-None for packet-less envelopes.
Also handle the small race where a proof arrives between outlet.send()
registering the receipt and us installing the delivery callback: after
registration, re-read the receipt status and synthesize the
_packet_delivered() call if it's already DELIVERED.
- StreamDataMessage now packed by struct rather than umsgpack for a more predictable size
- Added protected variable on LocalInterface to allow tests to simulate a low bandwidth connection
- Retry timer now has exponential backoff and a more sane starting value
- Link proves packet _before_ sending contents to Channel; this should help prevent spurious retries especially on half-duplex links
- Prevent Transport packet filter from filtering out duplicate packets for Channel; handle duplicates in Channel to ensure the packet is reproven (in case the original proof packet was lost)
- Fix up other tests broken by these changes
- a message handler can return logical True to prevent subsequent message handlers from running
- Message types >= 0xff00 are reserved for system/framework messages
Jeremy O'Brien