⬢ DragonFlyBSD Kernel Audit
DF-4752 / verdict.json
← back to finding ↓ download raw
{
  "finding_id": "DF-4752",
  "status": "inconclusive",
  "reproduced": 0,
  "impact": "leak",
  "confidence": "certain",
  "verdict": "Source-certain defect: gpio_unregister (gpio.c:587-617) never returns the master devices DEVFS_CLONE_BITMAP(gpio) unit (allocated at gpio.c:573 for the gpio/<drv>N/master cdev) and never frees the shared struct gpio_driver + kstrdupd name, unlinking it at gpio.c:609-611 even while sibling controllers of the same driver_name still reference it. Permanent privileged-trigger kernel heap leak per controller detach cycle; bitmap auto-extends (limit=0) so it is unbounded creep, not exhaustion. Distinct from known DF-1845 (pin-loop get-vs-put) and DF-1846. Runtime reproduction is hardware/kernel-config-gated: the gpio(4) framework is optional gpio (sys/conf/files:2119), absent from the guests X86_64_GENERIC (no /dev/gpio*, no module), so the trigger could not be executed on the audit guest; fix.diff compile-validated as a KLD module with -Werror on the guest (fix_build.log rc=0).",
  "exploit_chain": "",
  "evidence": "VERDICT.md (source trace with path:line); run.log (guest feature-absence markers); build.log vs fix_build.log (baseline and patched both rc=0, -Werror); fix.diff (refs + master-unit put + kfree on last ref)",
  "kernel_refs": [
    "sys/dev/misc/gpio/gpio.c:573",
    "sys/dev/misc/gpio/gpio.c:602",
    "sys/dev/misc/gpio/gpio.c:609-611",
    "sys/vfs/devfs/devfs_helper.c:devfs_clone_bitmap_get"
  ],
  "poc_changes": "No seed PoC existed; trigger.sh authored fresh and honestly exits SKIP=3 on guests without the gpio(4) framework.",
  "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": "cd /root/gpiobuild/fix && make   (KMOD= gpiotest SRCS= gpio.c, bsd.kmod.mk)",
  "run_cmd": "sh run.sh / trigger.sh",
  "code_hash": "",
  "notes": "Phase V effort concentrated on (1) proving the leak accounting across register/unregister by whole-function source trace incl. devfs_clone_bitmap_get limit semantics, and (2) -Werror compile validation of the fix on the guest kernel source. Guest /usr/src restored to pristine (sha256 verified) afterwards.",
  "recommended_fix": "Put the master clone unit back after destroy_dev; refcount struct gpio_driver and kfree gpd+name when the last same-name controller unregisters.",
  "fix_status": "inconclusive",
  "fix_baseline_reproduced": 0,
  "fix_patched_reproduced": 0,
  "fix_verdict": "fix.diff compiles clean as a KLD module (-Werror, rc=0) against guest 6.5-DEVELOPMENT source. Behavioral validation (unit returned, gpd freed, siblings safe) requires gpio(4) hardware absent from this guest.",
  "fix_evidence": "fix_build.log (patched module link line + rc=0); build.log (baseline rc=0); fix.diff"
}