{
  "finding_id": "DF-4939",
  "status": "reproduced",
  "reproduced": 1,
  "impact": "leak",
  "confidence": "certain",
  "verdict": "REPRODUCED: every SIOCSIFFLAGS down->up cycle on if_ic leaks 2 x (if_mtu+4) bytes of M_DEVBUF because the down branch (if_ic.c:179-187) never frees ic_obuf/ic_ifbuf and the up branch (if_ic.c:193-197) unconditionally re-kmallocs both. Demonstrated on the QEMU guest with a virtual-I2C-bridge harness KLD: 300/300 cycles leaked, M_DEVBUF grew 219->619->819 allocs (214K->814K->1.09M) for 200+100 cycles, exactly +2 allocs/+3008B per cycle. Deterministic, unbounded, privileged trigger (root loop can exhaust kernel heap). Fix (allocate-only-if-NULL in the up branch) validated by rebuilding if_ic.ko on the guest: same 200-cycle workload is byte-flat (219/214K -> 219/214K).",
  "exploit_chain": "",
  "evidence": "run.log: baseline devbuf 219/214K -> 619/814K after 200 cycles -> 819/1.09M after +100 (exactly 2 allocs per cycle); FIX section of run.log: patched module byte-flat 219/214K -> 219/214K across the same 200 cycles. Harness lpbbfake.c + Makefile.harness attach ic0 with no i2c hardware.",
  "kernel_refs": [
    "sys/dev/netif/ic/if_ic.c:179-187",
    "sys/dev/netif/ic/if_ic.c:193-197",
    "sys/net/if.c:2134-2175"
  ],
  "poc_changes": "Pass-1-style PoC sketch did not exist for this finding (new pass-2 finding). Had to invent the reproduction environment: guest has no I2C controller, so a virtual lpbb bit-bang bridge KLD (lpbbfake.c) was written to instantiate lpbb->iicbb->iicbus->ic0. Also: ifconfig requires a destination address for the point-to-point ic0 (ifconfig ic0 A B netmask ... up).",
  "attempts": 2,
  "guest_uname": "DragonFly dfbsd 6.5-DEVELOPMENT DragonFly 6.5-DEVELOPMENT #0: Thu Jul  2 06:02:54 UTC 2026  root@dfbsd:/usr/obj/usr/src/sys/X86_64_GENERIC  x86_64",
  "runtime_sec": 420,
  "guest_dirty": 0,
  "build_cmd": "see build.sh (make lpbbfake.ko; fix build: cd /usr/src/sys/dev/netif/ic && make after applying fix.diff)",
  "run_cmd": "see run.sh (kldload stack; ifconfig ic0 up; 200x down/up; vmstat -m)",
  "code_hash": "0879aeea3cdef90f59261e285450c38a94c646738086108e20bac23bfedb3084",
  "notes": "Privileged trigger (SIOCSIFFLAGS needs SYSCAP_RESTRICTEDROOT, if.c:2135); leak rate up to ~128KB/cycle at mtu 65535. Non-overlapping with DF-2000 (MTU-change UAF race) and DF-2001 (NULL deref): this is the steady-state down/up lifecycle leak. Guest reset to clean-source snapshot after validation. One uncharacterized observation: a vm.sh clean-shutdown hung (DDB suspected) after the leak demo with modules loaded; console log was overwritten by the next boot, so it is recorded as an observation only, not part of this finding.",
  "recommended_fix": "In icioctl SIOCSIFFLAGS up-branch, allocate ic_obuf/ic_ifbuf only if the pointer is NULL (reuse the pair retained across down), per fix.diff.",
  "fix_status": "fixed",
  "fix_baseline_reproduced": 1,
  "fix_patched_reproduced": 0,
  "fix_verdict": "Patched if_ic.ko (fix.diff) rebuilt and installed on the guest; the exact baseline workload (200 down/up cycles) no longer leaks: M_DEVBUF InUse stays 219 allocs / 214K before and after (baseline unfixed: +400 allocs / +600K).",
  "fix_evidence": "run.log FIX VALIDATION section (FIX-CYCLES-DONE); build.log.fix note in build.log"
}