tls.CipherSuiteName calls tls.CipherSuites, which in turn allocates a
large slice of values. We only need the name when we're logging, and
doing so, reduces the amount of allocations (and therefore garbage
created) during a QUIC handshake by roughly 5%.
Especially the sentPacketHistory linked list shows up in allocation
profiles, since a new list element is allocated for every single packet
we send.
Using a pool for the receiving path, as well as for the frame sorter, is
less critical, since we're tracking ranges there instead of individual
packets / frames, but it doesn't hurt either.
The other occurrences where we use a linked list (connection ID tracking
and the token store) are used so rarely (a few times over the lifetime
of the connection) that using a pool doesn't make any sense there.
There's no point in having the timer fire multiple times for the same
timestamp. By using a monotonuos timer we avoid busy-looping in cases where the
timer fires, but we can't actually send a packet.
The speedup comes from multiple sources:
1. We now preallocate a byte slice, instead of appending multiple times.
2. Marshaling into a byte slice is faster than using a bytes.Buffer.
3. quicvarint.Write allocates, while quicvarint.Append doesn't.
The new parsing function is vastly faster than the combination of header
and extended header parsing:
BenchmarkShortHeaderParsing-10 44192314 26.79 ns/op 48 B/op 1 allocs/op
BenchmarkShortHeaderParsingOld-10 12627363 99.99 ns/op 228 B/op 3 allocs/op
* Add support for providing a custom ConnectionID generator via Config
This work makes it possible for servers or clients to control how
ConnectionIDs are generated, which in turn will force peers in the
connection to use those ConnectionIDs as destination connection IDs when sending packets.
This is useful for scenarios where we want to perform some kind
selection on the QUIC packets at the L4 level.
* add more doc
* refactor populate config to not use provided config
* add an integration test for custom connection ID generators
* fix linter warnings
Co-authored-by: Marten Seemann <martenseemann@gmail.com>
* optimize FirstOutstanding
* fix variable naming
* bug fix
* minor code improvements
* add a test to make sure that `Iterate` iterates in the right order
* add comment
