| Commit message (Collapse) | Author |
|---|
|
A server could only ever hold ONE connection, which makes an ircd pointless:
listen() turned the listener into the connection, and net_find_tcp demuxed on
local port alone, so two clients on 6667 were indistinguishable.
- net_find_tcp now matches an established socket on the full four-tuple
(local port + peer IP + peer port) and only falls back to a LISTENing
socket when none matches.
- A SYN at a listener no longer consumes it: tcp_spawn_conn clones a new
socket (local port + owning pid, so exit/kill still reclaims it) and the
listener keeps listening. accept() returns that socket; closing it leaves
the listener alone. A full table drops the SYN, and the client's retransmit
is taken once a slot frees.
- tcp_peer_is_current preserves BC: net_find_tcp scans with the socket index
there, and the compare needs BC for the rx buffer base. Without this the
scan died at the first mismatching socket, so a second client's handshake
ACK never reached its socket and it hung in SYNRCVD - tcpdump showed our
SYN+ACK going out and the client's NICK retransmitted five times into
silence.
- listen() only conflicts with another LISTENER on the port; connections
share it by design now.
- MAX_SOCKS 4 -> 6 (listener + 3 clients + 2 spare; 537 bytes of WRAM0 still
free), and SK_ACC marks a spawned connection as not yet accepted.
tcp_close's settle is also time-boxed to ~250 ms of real time. It counted
PUMPS (8 x 8000), which is ~30s of wall clock when the peer never answers -
invisible with well-behaved peers, but a daemon refusing a client that stayed
connected froze itself, and every other client with it, for half a minute.
httpd keeps its listener open across requests now, so the next client's SYN
is accepted immediately instead of being dropped between close and re-listen:
three parallel clients go 0.1/1.1/2.1s -> 0.4/1.0/0.1s.
|
|
The stack could only ever dial out. Two new ops give it an ear:
- listen(sock, port): sockets demux on LOCAL PORT alone, so the listener IS
the connection - the SYN fills in the peer and the socket becomes the
connection. One at a time, re-armed by listening again after close, which
is all a 4-socket table can honestly promise. A second client's SYN is
dropped in silence; its retransmit is taken once we are free (~1s), so
three parallel clients complete in 0.1/1.1/2.1s.
- accept(sock): one RX pump per call, never blocks (like recv_nb), so a
server keeps its own main loop and stays killable.
- TCP_LISTEN/TCP_SYNRCVD in tcp_in: answer a SYN with SYN+ACK, become
established on the handshake's ACK, and process any data that ACK carries.
tcp_peer_is_current tells a peer's retransmitted SYN from a newcomer's.
Without it, "any SYN in SYNRCVD is a retransmission" let a second client
*hijack* a connection mid-request: the first client got its SYN+ACK, sent
its GET twice, and was never answered (found in a tcpdump of two parallel
curls).
listen() also refuses a port another socket already holds - two sockets on
one port silently black-hole the second listener, since demux takes the
first match.
Sockets now record the pid that opened them (SK_OWNER) and exit *and* kill
release them: with 4 slots, a few killed irc/nc/httpd runs used to leave the
box unable to open a socket at all until reboot.
tcp_close stops settling the moment the peer's FIN/ACK arrives instead of
burning its whole pump budget. It otherwise sits in FINWAIT for about a
second, and a server that closes and re-listens drops the next client's SYN
in that window: every request after the first paid a SYN retransmit.
1.1s -> 0.1s per request.
|
|
Found by writing nc(1): a 96-byte receive buffer against a host that
answers in 200-byte segments hung the process every time.
- recv/recv_nb copied RXLEN bytes into the caller's buffer and never
looked at the caller's maxlen, so any short buffer got overrun and its
stack smashed. Deliver min(RXLEN, maxlen) and keep the tail in RXBUF
for the next call. wget only survived because its buffer (220) happens
to exceed our MSS.
- tcp_buffer_data overwrote RXBUF from offset 0 on every segment, so a
segment arriving before the app drained the previous one destroyed
unread bytes - silent, undetectable data loss. Append at RXBUF+RXLEN
instead, and drop (without advancing rcv_nxt) when it doesn't fit, so
the peer retransmits once there's room. RXLEN is one byte, so 255 is
the capacity that matters, not RXBUF's 288.
- A draining recv now clears RXLEN as well as HASRX: RXLEN is the append
offset, and a stale one made every later segment look too big to fit,
stalling the connection until the peer gave up.
- A FIN past a gap is not end-of-stream. tcp_in accepted any FIN whose
segment carried no data, without a sequence check, marking the socket
DONE and making recv report EOF at the hole - truncating a transfer
right before its last segments. Decide in-order-ness once, before
rcv_nxt moves, and re-ACK an out-of-order FIN instead.
Verified against a 6300-byte reply received while sending 831 bytes:
byte-exact (6339 = payload + banner, 303 lines, first/mid/last intact),
plus no regression in ping / nslookup / wget / irc.
|
|
Retransmitted segments (GB ACKs lag behind slow LCD rendering, so real
servers do retransmit) were accepted as fresh data: the same line rendered
again on every retransmit, and rcv_nxt over-advanced so every later outgoing
segment carried an ACK beyond the peer's snd_nxt - which real stacks drop,
silently wedging the session ('/join does nothing' until reconnect).
|
|
Programs that own their main loop (the IRC client) need to poll for
typed input without blocking. pollin() only sees the on-screen keyboard;
bytes injected over the link (gbhub 'type', for scripted/hub-driven
sessions) land in the kernel console ring, previously only drained by the
blocking KGetc path. Add SYS_POLLCON: a non-blocking con_pop for
userland.
Also enlarge that console ring 16 -> 64. One net_pump drains an entire
serial burst into the ring at once, so a whole injected command line has
to fit or bytes are dropped and lines merge (a 20-char command came out
truncated and glued to the next). 64 covers a full line; mask stays a
power of two.
|
|
SK_STATE/SK_SND/SK_RCV sat at struct offsets 305/306/310, past the
288-byte SK_RXBUF. But the field accessors add the offset as an 8-bit
immediate (add SK_x; ld l,a), so rgbasm truncated 305->49, 306->50,
310->54 (the -Wtruncation warnings we'd been ignoring) - putting STATE
and the sequence numbers *inside* RXBUF at offsets 49/50/54.
Any TCP segment with >=33 payload bytes therefore overwrote the
connection's own STATE and rcv_nxt/snd_nxt with message text. The first
segment of a stream landed, its data clobbered STATE to a garbage value,
and every subsequent segment was dropped because tcp_in no longer saw
ESTABLISHED - so multi-segment TCP receives (an IRC MOTD, any HTTP body
past one segment) silently stalled. wget appeared to 'work' only because
a single-segment reply plus a never-honored FIN still printed once.
Fix: reorder the struct so STATE/SND/RCV precede the big RXBUF, keeping
every field offset < 256 (SK_SIZE unchanged at 314, all accessors are
symbolic). Zero -Wtruncation warnings remain. Verified an 11-line IRC
registration burst now arrives intact.
|
|
Pinging sl0p.foo through gbhub looked hung: DNS resolved, then nothing.
Packet-tracing showed the echo request leaving the hub's TUN and eth0
correctly NATed every time - but 80.78.19.56 blackholes ICMP for all
ids in windows of tens of seconds (provider rate limiting). One lost
reply wedged ping for minutes because net_recv's pump-counted timeout
is effectively unbounded at native emulation speed.
Two fixes:
- NET_RECVNB (op 8): non-blocking recv - one RX pump, $FE if nothing
buffered, same delivery/EOF semantics as NET_RECV otherwise. ping now
waits <=~2s per seq (net_recv_nb + msleep loop), prints 'seq=N
timeout' and moves on, like real ping.
- ICMP echo id was hardcoded $1234 for every GB, every boot, so all
sessions produced byte-identical flows - hostile to NAT conntrack
(keyed on icmp id). wNetEchoId is now our host octet + rDIV timing
noise sampled at DHCP lease, distinct per GB and per boot.
Verified: 8 back-to-back native-speed gbhub runs, zero hangs; a run
that hit a blackhole window printed seq=1 timeout then recovered to
3/4 received. GB<->GB ping and DNS/DHCP unaffected.
|
|
A Game Boy sitting at the shell prompt now services the network instead of
being deaf until you run netd: KGetc (the console input wait) pumps net_pump
each poll, so inbound pings are auto-answered while idle. net_pump gained an
in-frame model and routes any non-framed link bytes to a small console-input
ring (con_push/con_pop), so a headless-injected command still reaches the shell
while SLIP frames go to the stack. Removes the old SLIP-skip-in-KGetc hack.
Verified: two Game Boys on the hub, GB0 idle at the prompt (no netd), GB1
`ping 10.0.0.2` -> 4/4 replies, routed GB1->hub->GB0->hub->GB1.
(Separate, pre-existing: gbhub *spawn* mode and windowed gbjoin can drop an idle
link socket - under investigation; daemon mode + headless is solid.)
|
|
The address is no longer baked in. net_init starts at 0.0.0.0; a DHCP client
runs as the first thing on boot (init/shell forks it and waits), and only once
it has a lease (or gives up) does the prompt appear.
Kernel:
- IP is configurable: 0.0.0.0 until leased; NET_SETIP op stores it.
- 16-bit frame length. DHCP/BOOTP packets are ~272 bytes, over the old 255-byte
frame cap, so net_slip_send takes a 16-bit length, net_pump assembles into a
320-byte buffer with a 16-bit wNetRxLen, and udp_send writes a 16-bit IP total.
Payloads stay <=255 (kept small on purpose) so the per-protocol datalen math
is unchanged. wNetTx/wNetRxBuf 256->320, SK_RXBUF 208->288.
Userland:
- c/dhcp.c: DISCOVER->OFFER->REQUEST->ACK over a UDP socket, then net_setip();
times out gracefully (shell still boots) if there's no server. sh.c runs it
before the prompt.
Bridge (self-contained DHCP server, no dnsmasq):
- tunbridge.py + netboot intercept UDP->:67 and answer OFFER/ACK leasing
10.0.0.2 (gateway 10.0.0.1); everything else is bridged/NATed as before.
Regression-tested ICMP/UDP/TCP after the 16-bit change. Verified end to end:
dhcp: discovering
dhcp: leased 10.0.0.2
/# ping 10.0.0.1 -> 4/4 (traffic from the leased address)
|
|
Kernel TCP client on the socket layer: net_connect(SOCK_TCP) runs the 3-way
handshake in-kernel, send()/recv() drive the byte stream, close() does the FIN.
- Socket gains state + 32-bit snd_nxt/rcv_nxt (big-endian, add-with-carry-fold).
- tcp_send_seg builds IP+TCP with a pseudo-header checksum; the SYN carries an
MSS option (200) so the peer never sends a segment larger than our 256-byte
frame buffer (we don't do IP reassembly).
- tcp_in state machine: SYN_SENT->ESTABLISHED on SYN-ACK, buffers in-order data
and ACKs it, handles FIN -> recv() returns 0 (EOF).
- No retransmission: the GB<->host link is lossless and the host's real TCP
owns the internet side - which removes TCP's hardest part.
- net_pump now processes one frame per call so recv drains each segment before
the next arrives (single rx slot, no overwrite).
wget.c is now a thin socket client: connect -> send "GET / HTTP/1.0" -> recv to
EOF -> print. Verified end to end against a host HTTP server:
/# wget 10.0.0.1
HTTP/1.0 200 OK
Hello from a real HTTP server, fetched by a Game Boy!
with a clean SYN/SYN-ACK/ACK ... PSH ... FIN/ACK trace on the wire.
|
|
Adds UDP to the kernel socket layer on top of the ICMP core:
- net_sum() (raw folded sum) split out of net_cksum() so a UDP pseudo-header
(src/dst IP + proto + length) can seed the segment checksum.
- udp_send: builds IP+UDP with the pseudo-header checksum; NET_BIND sets the
local/source port; net_connect sets the peer.
- udp_in: demuxes inbound UDP by destination port to the bound socket
(net_find_udp), delivers the payload + source addr.
Also fixes a real recv bug: net_pump clobbers BC/DE/HL, so the old recv timeout
counted in registers and was effectively random. recv now counts in WRAM.
New `nslookup HOST` (PROG_NSLOOKUP=28, bank 30): builds a DNS A query and parses
the answer (with 0xC0 name-compression) entirely in userland over a UDP socket -
the kernel never sees DNS, just UDP. Verified through the bridge NAT:
/# nslookup example.com -> example.com -> 172.66.147.243
This gives us name resolution for the TCP/HTTP demo next.
|
|
The network stack moves into the kernel. src/socket.asm owns SLIP framing,
IPv4, RFC1071 checksums, and ICMP; programs now speak a socket API through one
syscall (SYS_NET, DE=&netreq dispatched by op): net_socket/connect/send/recv/
close/poll (c/sock.h). No program touches SLIP, IP headers, or checksums.
- Socket table (4 sockets) + tx/rx buffers in WRAM0; our IP = 10.0.0.2.
- net_pump: drains the link, reassembles SLIP frames, demuxes IPv4. Inbound
ICMP echo requests are auto-answered in-kernel, so the GB replies to pings
whenever any process pumps RX.
- ICMP sockets: send() emits an echo request to the connected peer; recv()
returns the matching reply (with a spin/yield timeout).
ping.c is now a ~15-line socket client; netd.c is just `for(;;){net_poll();
yield();}`. Verified over tunbridge:
/# ping 1.1.1.1 -> replies from the real internet (kernel builds it all)
host# ping 10.0.0.2 -> 4/4, 0% loss (kernel auto-answers)
Gotchas recorded: gbos.inc isn't a make dep (touch asm after editing); this
crt0 doesn't copy initializers (fill arrays at runtime); and the arg string at
0xA000 overlaps _DATA, so parse targets must sit past it (big buffer first).
UDP and TCP sockets build on this same core next.
|
