TSI: bind(port 0) + listen() lands the VMM proxy on a different ephemeral port than the guest — ephemeral listeners unreachable

Environment

• libkrun 1.19.4 (Arch package libkrun 1.19.4-1), libkrunfw 5.5.0-1, crun-based krun 1.29.1-1
• Host: Arch Linux x86_64, kernel 6.12.91, KVM; rootless podman 5.x with pasta
• Guest: python:3.12-slim (bookworm, glibc 2.36)
• Networking: default — no NIC, so implicit TSI with TsiFlags::HIJACK_INET (tsi_hijack on the guest kernel command line)
• Verified still present in current main (2.0 dev): TsiListenRsp { result: i32 } in src/devices/src/virtio/vsock/packet.rs is identical there — no address is communicated back in either version.

Summary

bind(port 0) followed by listen() under TSI leaves the listener unreachable at its own port, including for the process that created it:

1. Guest app calls bind(127.0.0.1:0); the guest kernel assigns port X.
2. listen() sends TSI_LISTEN to the VMM with the recorded request address (port 0) — net/tsi/af_tsi.c:tsi_listen() sends tsk->bound_addr, which still contains port 0.
3. The VMM (src/devices/src/virtio/vsock/tsi_stream.rs:try_listen()) binds that address in the VMM process's own netns, where bind(0) gets a different port Y from that netns's ephemeral allocator.
4. TsiListenRsp returns only a result code — Y is never communicated back, so the guest keeps reporting X in getsockname().

Result: connect(127.0.0.1:X) gets ECONNREFUSED. The listener is alive at port Y, unreachable and unknown.

Minimal repro (bare krun container; tested 3/3)

cat > t.py <<'EOF'
import socket, time
s = socket.socket(); s.bind(('127.0.0.1', 0))
port = s.getsockname()[1]; s.listen(1)
print('bound:', port)
time.sleep(0.5)
c = socket.socket(); c.settimeout(5)
c.connect(('127.0.0.1', port))   # -> ConnectionRefusedError
EOF

podman run --rm --runtime=krun \
    --annotation krun.ram_mib=1024 --annotation krun.cpus=2 \
    --mount type=bind,src=$PWD/t.py,target=/t.py \
    python:3.12-slim python3 /t.py

Output:

bound: 43859
ConnectionRefusedError: [Errno 111] Connection refused

Control: the same script with a fixed port (bind(('127.0.0.1', 45677)), connect to 45677) succeeds — fixed-port binds do not diverge, which confirms the mechanism.

Direct proof of the divergence

Run the same ephemeral listener inside a --network host krun container and inspect the host side while the guest sleeps on accept():

guest: bound 127.0.0.1:53603 (X)
host:  ss -ltn  →  LISTEN 127.0.0.1:45551 (Y — assigned by the VMM's netns)
connect to X → refused; connect to Y → accepted, guest accept() fires

Real-world impact

debugpy (python -m debugpy --listen 0.0.0.0:5678 …) always creates a 127.0.0.1:0 endpoints listener whose adapter child connects back, so every debugpy session dies after 30 s with RuntimeError: timed out waiting for adapter to connect. Reproduced in a bare krun container with debugpy alone. Anything else using ephemeral listeners is similarly affected.

Suggested fix directions

Return the actually-bound address/port in TsiListenRsp and have the guest update the socket's local address; or have the guest request a concrete port for the shadow bind (it already knows X). We currently carry a userspace workaround (LD_PRELOAD rewriting port-0 binds to concrete ports) and are happy to contribute a VMM/kernel fix if this direction sounds right.

Related: #819 — same investigation, TSI locking bug