Ring.onReadEnd and Ring.ReadDiscard call Reset() when the last readable
byte is consumed, which rewinds Off and End to 0. That is a valid empty
state, but it also moves the write position, and bytes staged past that
position with PeekWrite are addressed relative to it. After the rewind a
Commit hands back whatever bytes happen to live at the start of the
buffer instead of the staged ones.
tcp stages out-of-order segments exactly this way (see reassembly.store,
"the ring write pointer advances in lockstep with rcv.NXT, so the staged
bytes are always where seq implies"). Reading is driven by the
application, so a receiver that drains its stream while a gap is open
loses the staged tail silently: the stream arrives with the correct
length and the wrong contents.
Observed on a Sophgo SG2002 (100Mbit DWMAC, receive ring dropping frames
under load): a 16 MiB download arrived complete with a different
SHA-256, and TLS over the same path failed with "bad record MAC", while
the transmit path was bit-perfect.
Mark the ring empty without rewinding instead: End=0 is what empty
means, and the write position is then Off, which Ring.Write already
supports explicitly.
Two tests, both failing before the change:
- internal: staged bytes survive the ring being read empty.
- tcp: a reassembled stream is byte-identical when segments arrive
reordered (segment K arrived as the contents of an earlier segment).
Co-authored-by: Derek den Haas <i.pestano@easyflor.nl>
* feat(tcp): add out-of-order segment reassembly
Add an opt-in, bounded out-of-order reassembly buffer so a single lost
segment can be recovered by retransmitting the gap while later segments
are held and delivered once the gap fills.
The receiver also subtracts buffered out-of-order bytes from the
advertised receive window and avoids challenge-ACK aborts for in-window
future data. Reassembly is disabled by default.
Generated with LLM assistance.
Signed-off-by: Marvin Drees <marvin.drees@9elements.com>
* implement review feedback around rx buffer reuse
Signed-off-by: Marvin Drees <marvin.drees@9elements.com>
---------
Signed-off-by: Marvin Drees <marvin.drees@9elements.com>
- http/httpraw: cap header slice growth to pre-allocated capacity, fix
benchmark to reuse Header across iterations (0 allocs/op)
- internal/ring: replace TODO panic comments with invariant explanations
- tcp/conn: document truncated-frame offset check
- tcp/handler: replace TODO with net.ErrClosed on RST-closed connection
- internet/stack-udpport: filter incoming packets by remote IP address
when configured via SetStackNode; skip filter for IPv4 multicast
destinations (class D) so mDNS and similar protocols work correctly
Generated with LLM assistance.
Signed-off-by: Marvin Drees <marvin.drees@9elements.com>
* pcap: reuse Frame memory
* slog: reduce heap allocations of addresses; also prevent heap alloc of dhcp options in pcap
* dns: heapless improvement; add StackAsync buffer for more heapless operation; start thinking of errors
* errors: begin standardise errors in lneto
* errors: finish standardization of errors
* fix merge issues
* add more lneto errors to rest of package
* format errors.go
* do not rely on unspecified Go behaviour; fix TCP ring buffer bug
* even more stricter error handling on tcp connections
* stricter lock copying in tcp.Conn, even though likely not a problem
* add listener and tcp pool logging
* forgot reqAddr unused
* add logging to TCPConn and friends
Derek den Haas
Marvin Drees
Pat Whittingslow
ddirect