PyTCP

Pure-Python, zero-dependency TCP/IP stack — Ethernet through RFC 9293 TCP — running in user space on a TAP/TUN interface, embeddable in-process or run as a daemon, with a Berkeley-sockets API.

from pmd_pytcp import socket, stack

What it is

A full, RFC-grounded TCP/IP stack implemented entirely in Python: Ethernet II / 802.3 (LLC/SNAP), ARP, IPv4 / IPv6 (with Hop-by-Hop / Destination-Options / Routing / Fragment extension headers), ICMPv4 / ICMPv6 (incl. Neighbor Discovery, MLDv2 with MLDv1 fallback, and IGMP IPv4 multicast group membership), DHCPv4 and DHCPv6 clients, UDP, and a RFC 9293 TCP with a real FSM, congestion control (Reno / NewReno / CUBIC), SACK / timestamps / window-scaling, and a BSD-sockets facade. It runs on a TAP/TUN interface in user space — no kernel module, no privileged data path. It can be embedded in-process as a library, or run as a daemon that out-of-process clients drive over an AF_UNIX control boundary — the kernel/userspace split described below.

The project's north star is feature-equivalence with the Linux host network stack: where an RFC is unambiguous PyTCP follows it, and where it is silent or offers a menu PyTCP picks the Linux choice. Per-RFC adherence is audited under docs/rfc/.

The three distributions

PyTCP is strictly layered into three independently-published dists (one invariant: project folder == import name):

Installing PyTCP pulls the other two automatically (lockstep version pin).

Runtime architecture

TAP/TUN fd ─> RxRing ─> PacketHandler (per protocol, RX) ─> Socket queues / ARP+ND caches / fragment store
           <─ TxRing <─ PacketHandler (per protocol, TX) <─ Socket send / ND / DHCP / ACD

• Subsystem base — every background service (RX/TX rings, neighbor caches, timer, DHCPv4 / DHCPv6 clients, link-local / ACD) extends Subsystem and runs its own thread with an event-driven loop.
• Packet handlers — RX and TX paths are composed from per-protocol sub-handlers (packet_handler__<proto>__<rx|tx>.py). Every branch bumps a per-protocol stat counter for observability.
• Event-driven timer — a heap-based deadline scheduler (no polling tick); subsystems register deadlines and are woken on the nearest one.
• Per-interface model — a PacketHandler is an interface. A multi-homed host runs one handler per interface; global tables (routing FIB, socket table, neighbor caches) are shared and lock-guarded.

Free-threaded (no-GIL) safety

Per-interface state is partitioned (single-writer TX ring hand-off); the shared global tables (RouteTable, SocketTable, InterfaceTable) guard their compound (check-then-act) operations with a small threading.Lock and hand readers consistent snapshots. Single built-in dict/list ops are left lock-free (individually atomic).

Control-plane APIs (the Phase-3 kernel/userspace boundary)

Consumers talk to the stack only through sanctioned surfaces — never by reaching into runtime internals — mirroring how a Linux process talks to its kernel:

Lifecycle

stack.init(...) builds the singletons, stack.add_interface(...) / stack.remove_interface(...) attach / detach interfaces at runtime (RTNETLINK RTM_NEWLINK / RTM_DELLINK semantics, including the address / route / neighbor / session teardown cascade), stack.start() spawns the subsystem threads, and stack.stop() winds them down. A stack can init() with zero interfaces and gain them later — the daemon / multi-homed shape.

Sockets

pmd_pytcp.socket mirrors the stdlib socket module: a socket(...) factory returns TcpSocket / UdpSocket / RawSocket / PacketSocket, with bind / listen / accept / connect / send / recv / close, fileno() + eventfd for selectors integration, blocking & non-blocking modes, errno-mapped OSError, getaddrinfo, common setsockopt options, IPv4/IPv6 multicast group membership and source-filter options (IP_ADD_MEMBERSHIP, IP_ADD_SOURCE_MEMBERSHIP, IPV6_JOIN_GROUP, …), and an IP_RECVERR / MSG_ERRQUEUE error queue. Stdlib-parity constants (AF_INET, SOCK_STREAM, IP_*, SO_*, MSG_*) are exposed as bare module names backed by IntEnums.

Daemon mode — out-of-process clients

The stack runs as a daemon: a normal in-process stack that also listens on an AF_UNIX control socket, so a separate process opens sockets and drives the control-plane APIs through pmd_pytcp.client — exactly the way a Linux process talks to the kernel. The client never boots the stack; it holds real, selectors-pollable socket fds handed to it across the boundary via SCM_RIGHTS.

Start the daemon (it owns the TAP interface); the first-class entry point ships in the package as python -m pmd_pytcp.daemon (or the pytcpd console script after install), defaulting the control socket to $XDG_RUNTIME_DIR/pmd_pytcp.sock:

python -m pmd_pytcp.daemon --ipc-socket /tmp/pmd_pytcp.sock

Then, from any other process — note it imports pmd_pytcp.client, boots no stack, and calls no stack.init():

from pmd_pytcp.client import connect
from pmd_pytcp.socket import AddressFamily, SocketType

with connect(socket_path="/tmp/pmd_pytcp.sock") as client:
    sock = client.socket(AddressFamily.INET4, SocketType.STREAM)
    sock.connect(("10.0.1.1", 7))   # a real, selectable fd backs this socket
    sock.send(b"hello")
    print(sock.recv(5))

The same client.socket(...) factory returns UDP / raw / AF_PACKET sockets, and client.sysctl / .route / .link / .address / .neighbor / .membership mirror the in-process control APIs across the boundary. See examples/client__tcp_echo_ipc.py.

Install

pip install PyTCP

Brings in PyTCP-pmd_net_proto and PyTCP-pmd_net_addr — no other runtime dependencies (the whole stack is stdlib-only). Fully typed (ships py.typed, PEP 561); strict-mypy clean.

Running the stack

Running needs one or more TAP/TUN interfaces (root for interface / bridge setup). Bridged TAP interfaces (Ethernet) are created on the br0 bridge, so the bridge comes first:

make bridge      # create the br0 bridge (sudo)
make tap7        # create tap7, add it to br0 (sudo)
make tap9        # create a second tap, tap9, on br0 (sudo)
make run         # run the stack on tap7
make run_multi   # multi-interface demo (runs on tap7 + tap9)

Point-to-point TUN interfaces (IP), each created pre-addressed and needing no bridge, are also available — make tun3 (172.16.1.1/24, 2001:db8:1::1/64) and make tun5 (172.16.2.1/24, 2001:db8:2::1/64) set up the host side, and make run_tun / make run_tun5 run the stack on the matching device (taking the .2 host in each subnet). A stack can init() with zero interfaces and add / remove them at runtime, so any mix of taps and tuns can be attached to one running stack.

PyTCP is consumed as a library through the stack lifecycle API and the pmd_pytcp.socket BSD-sockets API. See examples/ — examples/stack.py is the complete runnable reference (TAP/TUN open, stack.init(...), multi-interface bind, runtime interface removal on SIGUSR1).

Requirements

Python 3.14+, Linux (TAP/TUN), POSIX.

Current state (3.0.7)

• ~210 source modules; 3.0.7 adds the kernel/userspace IPC layer (an ipc AF_UNIX RPC + SCM_RIGHTS fd-passing core, a client out-of-process mirror, and a first-class daemon entry point). The pmd_pytcp suite runs ~4,000 unit + integration tests (the full repo suite, across all three packages + examples, is ~12,500). Lint clean (codespell + isort + black + flake8 + mypy strict + pylint).
• Host-stack feature-complete (North Star Phase 1), now reachable both in-process and over an out-of-process daemon boundary (AF_UNIX control plane + SCM_RIGHTS socket-fd passing for TCP / UDP / raw / AF_PACKET). Phase-2 router/forwarding sits behind the forward_or_deliver seam as a stub. Authoring contracts in .claude/rules/pmd_pytcp.md; per-RFC adherence in docs/rfc/.

License

GPL-3.0-or-later. PyTCP by Sebastian Majewski.