Add SLIRP user-mode networking (outbound TCP/UDP)

[jserv]

Bridge LKL's virtio-net device with minislirp to provide outbound networking without privileges. Guest processes can resolve DNS and make HTTP connections through the host network via --net.

Core design:

• LKL virtio-net bridged to SLIRP via length-prefixed TX/RX pipes
• RX packet queue (SPSC ring buffer) between rx_reader thread and LKL net_rx callback; returns -1 when empty to prevent kworker refill spin that would starve lkl_cpu_get callers
• Shadow sockets for INET SOCK_STREAM/SOCK_DGRAM: AF_UNIX socketpair injected into tracee via ADDFD, event loop pumps data to LKL socket
• Non-INET, SOCK_RAW, and no-SLIRP sockets use virtual FD path
• Interface configured via ioctl (SIOCSIFADDR creates connected routes)
• net_poll gated on net_ready flag during boot/config
• Registration and deregistration both synchronous under shadow_lock

Socket dispatch additions:

• forward_socket, forward_bind, forward_connect, forward_sendto, forward_recvfrom, forward_recvmsg, forward_getsockopt, forward_setsockopt, forward_getsockname, forward_getpeername, forward_shutdown with shadow socket support via resolve_lkl_socket
• Shadow socket dup/dup2/dup3/fcntl(F_DUPFD) tracked via shadow_sp field with ADDFD injection and fd_table propagation
• forward_close ref-counts shared lkl_fd before teardown
• close_cloexec_entry ref-counts via lkl_fd_has_other_ref
• forward_execve calls kbox_fd_table_close_cloexec for cleanup
• fcntl(F_SETFD) syncs entry->cloexec for shadow sockets
• host_fd close guarded by shadow_sp >= 0 (tracee-namespace FD)

Change-Id: I48b05f90fac55f843330896f6b6bb84851ff1a65

---

Summary by cubic

Adds SLIRP user‑mode networking via minislirp to enable outbound TCP/UDP (DNS, HTTP) from guests without privileges. Enable it with --net, which bridges LKL virtio‑net to a host SLIRP instance.

• New Features

  • Outbound TCP/UDP via SLIRP; DNS verified by a new net-dns-test, HTTP via wget.
  • Register netdev before boot; configure IP/routes/DNS after boot; synchronous teardown and cleanup on failures.
  • Bridge LKL virtio‑net to minislirp using length‑prefixed pipes and an RX SPSC queue to avoid spin; event loop pumps frames.
  • Shadow INET sockets via AF_UNIX socketpair injected with ADDFD; dispatcher forwards bind/connect/sendto/recv* and socket options; dup/close tracked. sendmsg stays allow‑listed for pre‑exec FD transfer—use sendto for addressed datagrams; recvmsg is intercepted to return source addresses.
  • Test runner auto‑runs networking checks when --net is used.

• Dependencies

  • Replaced externals/minislirp submodule with scripts/fetch-minislirp.sh; Makefile auto‑fetches when KBOX_HAS_SLIRP=1 (and .gitignore updated).
  • Added socket syscall wrappers and seccomp updates, including kbox_notify_addfd_at for dup2/dup3 injection.

^{Written for commit ba08934. Summary will update on new commits.}

[cubic-dev-ai]

6 issues found across 21 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="Makefile">

<violation number="1" location="Makefile:142">
P2: Avoid running `fetch-minislirp.sh` via `$(shell ...)` at parse time; it causes side effects for unrelated targets like `clean` when SLIRP is enabled.</violation>
</file>

<file name="scripts/fetch-minislirp.sh">

<violation number="1" location="scripts/fetch-minislirp.sh:20">
P1: Validate `SLIRP_DIR` before `rm -rf` to avoid deleting unintended paths when the override is misconfigured.</violation>
</file>

<file name="src/seccomp-dispatch.c">

<violation number="1" location="src/seccomp-dispatch.c:868">
P2: FD leak when ADDFD result doesn't satisfy minfd. Once `kbox_notify_addfd` succeeds, the FD is already injected into the tracee. Setting `nh = -1` and falling through to CONTINUE causes the host kernel to create a second dup, leaving the ADDFD FD orphaned. Consider either not calling ADDFD when minfd > 0 and relying on CONTINUE, or closing the orphaned tracee FD before returning CONTINUE.</violation>
</file>

<file name="src/fd-table.c">

<violation number="1" location="src/fd-table.c:143">
P1: `kbox_fd_table_remove` may close unrelated supervisor FDs by treating tracee ADDFD numbers as local `host_fd`s.</violation>

<violation number="2" location="src/fd-table.c:216">
P1: CLOEXEC cleanup closes `host_fd` as if local, but it is a tracee ADDFD number and can corrupt supervisor FD state.</violation>
</file>

<file name="src/net-slirp.c">

<violation number="1" location="src/net-slirp.c:224">
P2: The `cmd_pipe` infrastructure is created and polled but never written to. Both `kbox_net_register_socket()` and `kbox_net_deregister_socket()` perform synchronous operations under `shadow_lock` instead of using the command pipe. This leaves `struct net_cmd`, `NET_CMD_REGISTER`, `NET_CMD_DEREGISTER`, and `process_cmd_pipe()` as dead code.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: Validate SLIRP_DIR before rm -rf to avoid deleting unintended paths when the override is misconfigured.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At scripts/fetch-minislirp.sh, line 20:

<comment>Validate `SLIRP_DIR` before `rm -rf` to avoid deleting unintended paths when the override is misconfigured.</comment>

<file context>
@@ -0,0 +1,26 @@
+fi
+
+echo "Fetching minislirp from ${REPO} ..."
+rm -rf "${SLIRP_DIR}"
+git clone --depth=1 "${REPO}" "${SLIRP_DIR}"
+
</file context>