SchOS/docs/architecture/networking.md

Networking

ARP, IPv4, ICMP, UDP, TCP and DHCP, over a driver-agnostic interface class, reached through the same socket syscalls as Unix domain sockets.

Verified end to end rather than by inspection: bootcheck -Net starts a listener and a UDP responder on the host, tells the guest to talk to them, and asserts on both ends — one side alone cannot tell a working stack from one that completes a handshake and drops every byte.

---

The four layers

        socket() bind() connect() send() recv()        syscalls, one set
    ┌────────────────┴───────────────────┐
    │  AF_UNIX              AF_INET      │   socknew() dispatches on domain
    │  kernel/net/unix.c    kernel/net/inet.c
    └────────────────┬───────────────────┘
                     │  lwIP raw API
    ┌────────────────┴───────────────────┐
    │  lwIP 2.2.0, NO_SYS=1              │   ports/lwip, compiled into the kernel
    │  port layer: kernel/net/lwip/      │   ~400 lines: a lock, a clock, a netif
    └────────────────┬───────────────────┘
                     │  struct netdev_ops
    ┌────────────────┴───────────────────┐
    │  e1000, and any future NIC driver  │
    └────────────────────────────────────┘

Realtek RTL8169/8168/8111 — drivers/net/r8169/

The NIC in a very large number of laptops and desktop boards. Every RTL8111/8168 variant answers to 10EC:8168, which is why one match entry covers two decades of silicon — and why the driver prints the chip's XID on probe, since the PCI device id tells you almost nothing about which part arrived.

Why it is native, when the plan said LDLS

The first plan was to run Linux's r8169.ko under ldls.md. Measuring it is what changed the answer: the module has 183 undefined symbols, 130 of them not implemented, and 35 of those 130 are phylib — phy_connect_direct, phy_start, mdiobus_read, genphy_soft_reset.

phylib is a subsystem with a link state machine, not a set of leaf calls, and LDLS's usual answer for an unimplemented symbol — a stub that reports itself and returns zero — is precisely wrong for it. phy_start returning zero means the link never comes up and there is nothing on screen to say why. The 8139too that ldls.md uses as its model has no phylib dependency, which is why that one was tractable.

A native driver reaches the PHY through the chip's own MII window (PHYAR, one register that is a whole MII bus) and skips the subsystem entirely. It is less work than the port, it carries no GPL entanglement, it needs no firmware blob, and it is not coupled to one Linux version.

The register layout is a fact about the hardware, taken the way ldls.md §3 says facts may be taken. The expression is this tree's: a descriptor-ring driver in the same shape as the e1000 beside it.

It has not run on hardware

Stated plainly because it cannot be tested here. QEMU emulates rtl8139 and nothing else from Realtek, and the 8139 is a different programming model — a circular receive buffer rather than descriptors — so it is deliberately not in the match table. Claiming it would break a NIC that works.

What has been verified: the driver builds, its presence changes nothing on a machine without the device, and with an rtl8139 on the bus it correctly leaves it unclaimed (10ec:8139 network controller [-]).

What has not: every register write, on a real part.

So the probe says everything a first bug report would need, with cprintf rather than kinfo so it reaches the screen on a machine with no serial port:

r8169: xid 5c0, 1c:1b:b5:xx:xx:xx, irq 11, link up (bmsr 796d)

xid is the hardware version out of the top of TxConfig; bmsr is MII register 1, where bit 2 is link and bit 5 is autonegotiation complete. Those four facts separate "the chip is not there", "the chip is there and the PHY sees nothing", and "everything is up and the rings are wrong".

Two constants that were wrong first

Recorded because both would have worked, quietly:

• RxConfig had 0x7700 for unlimited burst and threshold. MXDMA is bits 8-10 and RXFTH is bits 13-15, so that is one bit short of a full burst and a half-full FIFO threshold rather than none — slower, and invisible.
• CPCR bit 3 was being set under a comment calling it "PCI multiple read/write". Bit 3 is PCIDAC, 64-bit DMA addressing; bit 4 is the one the comment described. Neither is wanted, since the rings come from the DMA arena below 4 GB, so the write is gone rather than corrected.

The netdev class

A driver publishes a struct netdev_ops and is then finished with the network stack forever: it never mentions lwIP, and the stack never mentions it.

struct netdev_ops {
  int (*transmit)(struct device *, const void *frame, int len);
  int (*receive) (struct device *, void *out, int cap);
  const unsigned char *(*mac)(struct device *);
  int (*link_up) (struct device *);        // optional
};

transmit returning -1 is congestion, not failure — the ring is full and the caller is expected to retry. receive never blocks; it returns 0 when nothing is waiting, and the caller decides how to wait.

This class was added late, and the reason it had to be is instructive: the stack had reached straight into the driver. net.c called e1000_present(), the lwIP glue was named netif_e1000.c, and the e1000 shipped with .ops = 0 and a comment saying so. A second NIC driver would have meant editing the protocol stack to add itself. lwIP now gets one netif per bound device, /proc/netif lists them all, and the e1000 drives up to four cards rather than refusing the second — verified with two attached, each with its own MAC and its own DHCP lease.

netdev_selftest()

echo n > /proc/sysrq sends a hand-built ARP request and waits for the reply.

It proves the whole driver path in one exchange — descriptor formats, the DD handshake, tail bookkeeping, DMA addresses, the receive ring hand-back, and the MAC filter, because an ARP reply is addressed to this station and does not arrive if the filter is wrong.

It was written before there was any protocol stack, deliberately: a stack on an unproven driver fails in ways indistinguishable from its own bugs. It belongs to the class rather than to the e1000, because it is the check every new NIC driver will want first.

lwIP

lwIP 2.2.0, fetched by ports/lwip and compiled into the kernel — it is the one port that installs nothing into the sysroot, because it is kernel code.

NO_SYS=1, so lwIP has no threads, mailboxes or semaphores of its own. The port layer is kernel/net/lwip/: an architecture header, a lock, a clock, an Ethernet netif, and a set of shim headers standing in for the libc lwIP expects.

Why lwIP and not a BSD stack

The licence was not the deciding factor — FreeBSD's sys/netinet is BSD-3 too. It is that FreeBSD's stack is not a library. tcp_input.c alone is 122 KB and reaches into mbufs, epoch(9), netisr, VNET, uma(9), sockbufs and inpcb hash tables. Taking it means porting FreeBSD's kernel infrastructure first and finding the TCP somewhere underneath — the same conclusion already reached about FreeBSD's USB stack.

NetBSD was ruled out on licence: its tcp_input.c and tcp_output.c carry the WIDE Project's 4-clause advertising clause, which Berkeley's rescission does not reach. See ../licensing.md §3.

lwIP was written for exactly the case SchOS is in: a small operating system with a raw Ethernet driver and no infrastructure to speak of.

The reason not to write one, kept because it is the reason to resist later too: it is not the packet formats that are hard. It is congestion control, retransmission, window management and twenty years of corner cases that only appear against a real peer.

The network thread

kernel/net/net.c. One kernel thread, started by netinit(), feeds lwIP received frames and runs its timers.

A thread rather than the interrupt handler, and the reason is the lock: a call into lwIP can reach the driver, the driver can sleep, and an interrupt handler cannot. So the frame is always pulled, by this thread, through .receive.

The interrupt's only job is to end the thread's wait early. netdev_rx_signal() is what a driver calls from its handler; it delivers no frame and reads no ring. Without it a frame that arrived just after a poll sat there until the next clock tick — up to 10 ms, and the largest fixed latency in the system.

The wait stays on the timer tick rather than moving to a channel of its own, and that is the design rather than an accident. lwIP's timers must still run when the link is silent, so the clock has to remain a floor under the wait, and sleep() cannot wait on two things at once. Joining the tick's existing wakeup costs nothing and leaves one wait with one predicate.

The flag lives under tickslock for the same reason it does not get a lock of its own: sleep() releases tickslock atomically, so a driver setting the flag and waking under that lock cannot lose the wakeup by arriving between the predicate test and the sleep. A private lock would not be held across the sleep, and that window would be back.

/proc/netif reports wake_rx and wake_tick so this is observable rather than assumed — if wake_rx stays at zero while frames arrive, the interrupt is not reaching the stack and the thread is running on the clock exactly as it used to. Measured across three pings: rx_frames and wake_rx both went 2 → 6.

netinit() runs after sinit() because kthread_create needs a usable process table, and it returns immediately: everything happens in that thread, which cannot run until the scheduler is entered.

DHCP obtains an address at boot. /proc/netif reports each interface's address, netmask, gateway, MAC and frame counts, taken at the lwIP seam.

The lease announcement goes to the serial line only, through klog() rather than cprintf(). A lease arrives seconds after the shell prompt does, so on the display it lands in the middle of whatever the user is doing — which is the same rule that makes cons_serial_only start set: the display is the product, the serial line is the diagnostics. Someone who wants to know their address reads /proc/netif.

It is /proc/netif and not /proc/net deliberately: on Linux /proc/net is a directory of many files, and a single file by that name would be a lie about a well-known path.

Sockets

One set of syscalls, at their real Linux numbers. socknew() dispatches on the domain, and no syscall changes shape between the two families.

AF_UNIX

kernel/net/unix.c. SOCK_STREAM, socketpair, shutdown.

A connected pair is two buffers, not one: writing appends to the peer's buffer and reading drains your own, which is what makes a socket bidirectional where a pipe is not, without needing two of anything.

One lock covers the whole table. A pipe gets away with a lock per pipe because every operation touches one object; a socket operation touches two, and two-object locking needs an ordering rule every path obeys or it deadlocks. Worse, the wakeup races are subtle — a reader testing peer_closed under its own lock while the closer sets it under the peer's has no interlock at all.

A bound socket is a file. bind() creates a T_SOCK node at the path, and connect() resolves the path and looks for the socket bound to that inode.

Matching on the inode rather than on the path string is the part that matters. The name used to live only in this table, so two spellings of one path missed each other, a relative path never matched an absolute one, and there was nothing to stat() or unlink(). An inode is the thing a path resolves to — so connect() can walk the path however it likes and still arrive at the same socket, which is precisely what a string comparison could never do.

# ls -l /tmp
srwxr-xr-x 1 root root 0 2026-09-22 08:04 socktest

The node is a marker, not the binding. It holds no data, and the live association is in this table. So a socket file left behind by a crash is what Linux leaves behind: a dead file that connect() refuses until something binds to it again. Conversely, unlink()ing a live socket's path does not disturb anyone already connected — the socket simply stops being findable.

EADDRINUSE comes from the directory, not from a scan of this table, and that is deliberate: a stale socket file has to refuse a rebind for the same reason a running server's does, or two servers both believe they own it. The consequence is the one every real server already handles — unlink() before bind(), which cmd/socktest now does too.

The creation lives in sys_bind() rather than here, because that is the layer with both: kernel/net/unix.c knows about sockets and has no business calling create().

flock() is still missing. libwayland locks wayland-0.lock beside the socket so that two compositors cannot claim one display. The lock file can be created and opened now that paths are real; the advisory lock itself is a separate mechanism with no other caller yet.

Descriptor passing

sendmsg/recvmsg with SCM_RIGHTS, which is the one thing a socket can carry that a pipe cannot: a reference to an object the kernel already knows, rather than a copy of its contents. Wayland's wl_shm_create_pool passes a descriptor and has no integer-shaped alternative, and hybrid.md's mechanism 1 is how a driver process is handed the device it is allowed to touch. Both wanted the same primitive.

A batch of descriptors is pinned to the stream position of the first byte of the message that carried it. That single rule is the whole delivery model, and it has two halves that are easy to get half-right:

• a reader that arrives at that position is handed the descriptors, and
• a reader that has not reached it yet is stopped short of it.

Without the second half, a program that asked for 4096 bytes would swallow three messages in one call and be handed one message's descriptors, with nothing in the result saying which message they belonged to. socktest checks exactly this: two messages, each one byte and one descriptor, read with an eight-byte buffer, and the first read must return one byte.

The kernel holds a reference on every descriptor in flight, so a sender may close its own copy the moment sendmsg returns. That reference is also the thing that has to be given back, and there are three ways a message ends:

received by recvmsg with room	installed in the receiver's table
received by recvmsg without room, or by plain read(2)	closed, and MSG_CTRUNC set for the first case
never received — the socket is closed	closed by sockclose

The second row is the one that looks like a choice and is not. Holding kernel references for a receiver that has already proved it is not looking is a leak with extra steps, and it is what Linux does for the same reason.

Descriptors need at least one byte to ride on; sendmsg refuses a message of pure ancillary data with EINVAL rather than accepting it and pinning the descriptors to a position that never arrives.

There is no garbage collector. An AF_UNIX socket descriptor, sent over that same socket and never received, holds a reference to the socket that holds the reference to it, and neither close can reach zero. Linux carries several hundred lines to break these cycles. Nothing on either plan passes a socket over a socket — Wayland passes memory, the driver work passes devices — so this is a known and bounded leak (thirty-two sockets) rather than a silent one.

AF_INET

kernel/net/inet.c, on lwIP's raw API rather than lwIP's own sockets layer, which would need NO_SYS=0 and a second socket abstraction alongside this kernel's own.

Everything is protected by the lwIP lock and nothing else, and that is not a shortcut: lwIP's callbacks run in the network thread with that lock held and are the only writers to a socket's receive buffer. A syscall reading it takes the same lock. There is no second lock because there is no second writer.

How a blocking socket waits

Worth reading before touching, because the obvious implementation does not work. A blocking call cannot sleep(&socket, &lwip_lock): sleep() wants a spinlock and the lwIP lock is a sleeplock — it has to be, because a call into lwIP can reach the driver. And a waiter cannot hold that lock while sleeping, or the network thread could never deliver the thing being waited for.

So the wait is on a counter with a spinlock of its own:

  reader:  lwip_lock; check condition; s = net_seq(); lwip_unlock; net_wait(s)
  waker:   (inside a callback, lwIP lock held) net_wake()

No wakeup can be lost. s is read while the reader still holds the lwIP lock, so anything changing the condition must do so afterwards and therefore bumps the counter after the read; net_wait re-checks under its own spinlock and does not sleep.

One counter and one channel for all inet sockets rather than one each. At this scale waking sixteen sockets costs a handful of instructions in the ones with nothing to do, and a per-socket channel would need careful teardown in close() against callbacks that can still arrive.

Before this, a blocked socket woke on every timer tick and looked — costing a wakeup per socket per 10 ms and answering up to 10 ms late even when the data was already there. tcp_sent had to be registered for it as well, or a write blocked on a full send buffer slept with nobody left to wake it.

ICMP, and ping sockets

socket(AF_INET, SOCK_DGRAM, IPPROTO_ICMP) — Linux's ping socket, which gives an unprivileged process the one thing raw sockets were used for. The caller reads and writes the ICMP message alone; the kernel owns the IP header.

Built on lwIP's raw API: raw_new(IP_PROTO_ICMP), with the receive callback stripping the variable-length IP header before queueing the message on the same datagram queue UDP uses.

Only echo replies are consumed, and that restriction is load-bearing. raw_input runs before lwIP's own icmp_input, so a callback returning "eaten" for every ICMP message would swallow the echo requests too and this machine would silently stop answering pings from outside. Returning 0 leaves the packet for lwIP, which is what should happen to every ICMP message nobody asked for.

There is no per-socket echo-id filtering. Linux rewrites the id and matches on it; here a reply goes to every open ICMP socket, which is invisible with one ping running and means a second one may see a sequence number it never sent.

AF_NETLINK, and why it exists

getifaddrs(3) is how every program asks what interfaces does this machine have, and what addresses do they carry — and on Linux there is exactly one answer: a NETLINK_ROUTE socket. musl implements the libc call over it, so what looks like a syscall is really a small protocol conversation, and without this side of it every caller fails identically.

kernel/net/netlink.c is a responder, not a general socket family. There is no kernel-to-userspace event stream, no multicast groups, nothing subscribes to link state. A request arrives, the whole answer is generated into the socket's buffer, and reads drain it.

That shape is deliberate. getifaddrs sends a dump request and reads until NLMSG_DONE with MSG_DONTWAIT, so generating the answer at send time means a read never blocks and the non-blocking flag is honoured for free.

Two message types are served — RTM_GETLINK returns one RTM_NEWLINK per interface (name, flags, MAC), RTM_GETADDR returns one RTM_NEWADDR per address (address, prefix length, label). Anything else gets a bare NLMSG_DONE.

The data is not netlink's. Interfaces come from the netdev class and addresses from lwIP's netifs, exactly as /proc/netif reads them. Nothing in the file knows what an e1000 is.

Interface indices matter more than they look: getifaddrs matches address records to link records by index, so the two dumps have to agree. Loopback is 1 and the netdevs follow from 2.

What it produces, read back through musl:

lo     flags=0x49    fam=-1
eth0   flags=0x1043  fam=17                    (AF_PACKET, the MAC)
lo     flags=0x49    fam=2  addr=127.0.0.1  mask=255.0.0.0
eth0   flags=0x1043  fam=2  addr=10.0.2.15  mask=255.255.255.0

Loopback

127.0.0.1 is a real destination: LWIP_HAVE_LOOPIF creates the interface and lwip_init() brings it up. LWIP_NETIF_LOOPBACK is the half that is easy to miss — without it, a packet addressed to an interface's own address goes out of the wire and never comes back.

With NO_SYS=1 there is no thread to drain the loopback queue, so the network thread calls netif_poll_all() alongside sys_check_timeouts().

Datagrams

UDP has a datagram queue, not a byte ring. It shared TCP's ring at first, and two datagrams arriving together came back as one — which is the bug that had to be fixed before recvfrom could report a source address at all.

---

Names

For a long time nothing on this system could turn a name into an address, and the shape of that failure is worth keeping: it did not look like one missing feature. wget did not work. telnet did not work. Anything taking a hostname did not work, each in its own way, and each looked like a separate broken program — which is a good argument for a userland with fewer applets, and a wrong one.

There is still no DNS client in SchOS, and there does not need to be. musl contains a complete resolver: getaddrinfo reads /etc/resolv.conf, sends a query over an ordinary UDP socket and parses the reply. Every one of those pieces already worked. The single missing fact was which nameserver — and that arrives in a DHCP reply the kernel handled and never told anyone about.

So the mechanism is one file:

# cat /proc/resolv.conf
nameserver 10.0.2.3

net_resolv() renders resolv.conf's exact syntax, so that putting it where the resolver looks is a copy and not a translation. It is empty until DHCP answers, which is the honest report: a file saying nothing is right, and one naming a server that does not exist is not.

The policy is cmd/resolvconf, run once at boot by a schinit job. It waits for a nameserver, writes /etc/resolv.conf and exits — and will not overwrite an existing one without -f, because a machine with a hand-written resolv.conf has been configured on purpose.

That division is the usual one here: the kernel knows a fact a process cannot find out for itself and publishes it; whether that fact should become /etc/resolv.conf is not the kernel's business.

# wget -O - http://example.com/ | head -1
<!doctype html><html lang="en"><head><title>Example Domain</title>

Serving

cmd/httpd is the other direction, and it existed for the same reason this section does: a machine that could fetch a page but not serve one is a strange shape for something with a TCP stack and a filesystem.

GET and HEAD, static files out of a directory, index.html for a directory and a generated listing when there is none. A child per connection, so one client that stops reading cannot stop the server. No CGI, no keep-alive, no TLS, no configuration file — it is a file server, and each of those is a decision somebody should make on purpose.

The path check is the only security-relevant code in it, and it refuses rather than normalises. The tempting version resolves .., collapses slashes and then checks the result — and it has to agree exactly with the kernel's resolver, and where it does not, the difference is a file outside the root. So no component may be .. at all. Percent-decoding happens first, because %2e%2e is .. and a check that runs before decoding does not see it.

$ curl -s -o /dev/null -w '%{http_code}\n' --path-as-is host/%2e%2e/etc/passwd
403

A listening socket is a different object

Worth its own note, because it cost a kernel panic and the shape of it is not obvious from the API.

tcp_listen() does not put a pcb into a listening state. It frees the connection pcb and returns a new, smaller one — a struct tcp_pcb_listen, which has no recv, sent, err or poll field at all. The two are different types behind one struct tcp_pcb *, and the only callback a listening one has is accept.

inet_close() did not know that, and cleared the callbacks the same way for both. On a listening socket that is a write past the end of the allocation; lwIP asserts rather than allowing it, and an assert here is a panic:

lwip assert: invalid socket state for recv callback at tcp.c:2024
  #1 tcp_recv   #2 inet_close   #3 sockclose
  #4 fileclose  #5 exit         #6 syscall

Which is a user process taking the machine down by pressing Ctrl-C — the close path runs from exit(), so any program that listened and then died did it. Nothing had, because until cmd/httpd nothing in the tree listened on TCP and exited; schdbgd would have done it too, a few days later.

The lesson is not about lwIP. It is that listening was already tracked one field away in struct inet_sock and the close path did not consult it — the information was there and the code did not ask.

What it does not do

• ping is a first-party program, cmd/ping, not toybox's. recvmsg now exists, so half of what stopped toybox's is gone; what remains is IP_RECVTTL, an IPPROTO_IP control message that would have to be generated on the receive path rather than merely carried. Ours has no -I, no flood mode and no TTL display.
• SCM_RIGHTS is the only control message. SCM_CREDENTIALS is refused with EINVAL rather than ignored: a caller told its credentials were passed would have the receiver read a uid out of uninitialised memory and act on it.
• A message may carry eight descriptors, and a socket eight such messages. Past that sendmsg answers ENOBUFS. The cap exists because the kernel holds a reference to each one until it is received; a sender eight unreceived messages ahead of its receiver is not rescued by a ninth slot.
• ping still takes a dotted quad, even though names now work. It is a first-party program that does not link musl's resolver, and localhost is special-cased in it rather than resolved. Every ported program resolves names properly; ours is the one that does not.
• Nothing re-runs resolvconf. The nameserver is written once at boot. A lease that changes the DNS server, or a /dev/pci handover that brings up a new interface, leaves a stale /etc/resolv.conf until somebody runs it again.
• lwIP's own DNS client is still unused. It is compiled in and only dns_getserver() is ever called; resolution happens entirely in the process. Nothing in the kernel can resolve a name, which matters the day something in the kernel wants to.
• httpd does not serve HTTP/1.1. Every response is HTTP/1.0 with Connection: close, because the directory listing is generated straight into the socket and so has no Content-Length to give. One request per connection is the cost.
• A queued-but-never-accepted connection is closed without clearing its callbacks. inet_close() clears the argument and closes, which is enough because the pcb is freed — but it is a narrower guarantee than the one the main path makes, and it relies on tcp_close() succeeding.
• httpd has no timeouts. A client that connects and says nothing ties up a child process until it goes away; SO_RCVTIMEO does not exist here (see below), so there is nothing to set.
• No routing table, and no /proc/net/route to report one. With two interfaces up, the first is the default route and the second is reachable only by an address on its own subnet. That is a rule rather than an accident, and it is not routing. The missing file is the more immediately visible half: it is how a program asks which interface carries the default route, and without it anything that filters on "the default one" sees nothing.
• Netlink serves two message types. No RTM_GETROUTE, no RTM_NEWROUTE, no event stream, no multicast groups. A program that expects to subscribe to link changes gets a socket that answers dumps and never speaks again.
• No getsockname/getpeername, and accept() does not report the peer's address for an AF_INET socket. A server that wants to log who connected cannot.
• setsockopt covers six options, and refuses the rest with ENOPROTOOPT rather than pretending: SO_REUSEADDR, SO_KEEPALIVE, SO_BROADCAST and TCP_NODELAY reach lwIP's pcb; SO_SNDBUF/SO_RCVBUF are accepted and report the real size on read. Notably absent: SO_RCVTIMEO/SO_SNDTIMEO.
• SOCK_DGRAM over AF_UNIX does not exist — stream only.
• AF_UNIX names are not filesystem paths, so nothing can stat() or unlink() a socket.
• One NIC driver exists. The class is proven with two e1000 cards, not with two different drivers.
• No IPv6, no DNS resolver in the kernel (lwIP's is compiled but unused by anything), and no firewall or packet filter of any kind.