With client_mss_bulk the whole authenticated FakeTLS response (ServerHello,
CCS, emulated certificate records, tickets - typically ~4 KiB) was written in
client_mss-sized chunks: about 50 segments at `tspu`. With ~2% loss at least
one segment is lost in ~64% of handshakes, and the burst takes several RTTs
to pass the initial congestion window.
- New server.client_mss_prefix (Linux, requires client_mss_bulk): `full`
(default, current behavior), `server-hello` (through the end of the first
TLS record, capped at 2048 bytes, falls back to `full` on unexpected input)
or a byte count. Only that prefix is written in MSG_EOR chunks; the rest
goes out in one write at the bulk MSS.
- Chunk payload accounts for TCP timestamps and never exceeds tcpi_snd_mss.
- Removed TcpMaxSegmentGuard: TCP_MAXSEG on an established socket does not
change the MSS of outgoing segments on Linux, and its getsockopt/setsockopt
failures no longer abort the handshake. TCP_NODELAY is set once.
- Docs (en/ru/de), API discovery metadata, strict keys, reload diff.