vms-7bd: virtio-net NIC on the OVMX QEMU runtime (user-mode default, opt-in tap/bridge)

[baron-3dl]

What

Adds a virtio-net NIC to the OVMX QEMU runtime — the Linux-layer NIC infra prerequisite for TCP/IP Services (parent vms-67f; unblocks vms-3be, vms-9d2). The VMS device face (EWAn:) is a separate downstream item and is out of scope here.

Both runtime launch paths — distro/boot/run-qemu.sh and the distro/Dockerfile.bootable runtime CMD — previously passed -nic none (no guest network device). They now attach exactly one virtio-net-pci device, keeping -nodefaults and adding only -netdev/-device.

Modes (OVMX_NET_MODE)

Mode	Behavior
user (default)	user-mode NAT (SLIRP) — zero host config, unprivileged, CI-safe. No inbound by default (opt-in via OVMX_NET_HOSTFWD).
tap	pre-existing host tap (OVMX_NET_TAP, default tap0), script=no — operator provisions/bridges it.
bridge	existing host bridge (OVMX_NET_BRIDGE, default br0) via qemu-bridge-helper.
none	legacy no-NIC escape hatch.

OVMX_NET_MAC sets an explicit MAC. The PXE option ROM is disabled (romfile=) so the guest never pauses to network-boot.

tap/bridge need privileged, operator-provisioned host state and never work in unprivileged CI — hence user mode is the safe default. The launcher never creates a tap or bridge as a side effect (shared-host rule: no host-networking mutation).

Kernel support

The guest kernel (ubuntu generic 6.8) has virtio_net built in (CONFIG_VIRTIO_NET=y) — verified from the shipped vmlinux .modinfo (the virtio_net.* module params appear in vmlinux only for builtin drivers). So the interface enumerates with no in-guest module load and no kernel-config change.

Test — tests/qemu/test_virtio_nic.sh

Wired into the persistent-boot CI job (reuses the already-built ovmx-boot image) and the boot path filter, mirroring test_distrib_boot.sh.

• Part 1/2 (deterministic, no boot): the launcher + Dockerfile CMD emit the right -netdev/-device for every mode (via a new OVMX_QEMU_DRYRUN hook) — user default, opt-in tap/bridge/hostfwd, -nic none gone, -nodefaults kept, bad mode rejected.
• Part 3 (real boot): with the default user-mode NIC the guest kernel enumerates the virtio-net Ethernet controller (PCI 1af4:1000/1041, class 0x020000), the PXE ROM is suppressed, and the executive still attaches (NIC did not regress boot).

Verified empirically in-session: 18/18 checks pass (15/15 deterministic on the host + 3/3 real-boot against the extracted kernel + slim initramfs). virtio_net is silent on a successful probe (unlike virtio_blk), so the PCI enumeration line — made observable by raising loglevel in the test's own invocation — is the honest guest-side ground-source available without a guest shell.

Documented gaps (not faked)

• Asserting the netdev interface is up (ip link) and that user-mode NAT reaches the host needs an in-guest introspection channel the bootstrap initramfs lacks (init is STARTUP.EXE; no shell/dhclient). That arrives with the EWAn: device face + TCP/IP stack downstream.
• tap/bridge modes need a privileged host and are covered only at the arg-construction level.

🤖 Generated with Claude Code