⬢ DragonFlyBSD Kernel Audit
DF-4601 / verdict.json
← back to finding ↓ download raw 
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
{
  "finding_id": "DF-4601",
  "status": "not_reproduced",
  "reproduced": 0,
  "impact": "none",
  "confidence": "certain",
  "verdict": "Info-severity init-order defects, both provable statically and confirmed by a host-side logic mirror (construct_order.c): (1) dce112_compressor_construct() evaluates the single DRAM channel => no LPT guard (memory_bus_width == 64, :802) while the field is still 0 (kzallocd object; assigned from asic_id.vram_width only at :810) \u2014 the guard is dead and LPT_SUPPORT is never cleared on 64-bit-bus ASICs; (2) lpt_channels_num=1 set at :797 is unconditionally wiped to 0 at :818, making the only valid case of the program_lpt_control() switch (:675-691) unreachable \u2014 LOW_POWER_TILING_MODE is never programmed. Knock-on: banks_num/raw_size/dram_channels_num are never set anywhere in the tree, so lpt_size_alignment() is always 0 and enable_lpt() (:594-651) can arm LPT (LOW_POWER_TILING_ENABLE=1, :645-650) with ROWS_PER_CHAN=0 and a stale MODE whenever FBC+LPT come up on eDP. Runtime kernel-path reproduction is impossible on the audit guest (no AMD GPU; QEMU virtio only), hence not_reproduced/impact none; on-HW consequence (display underflow / GPU hang on affected Polaris eDP systems) is speculative and requires modeset privilege (/dev/dri).",
  "exploit_chain": "",
  "evidence": "run.log (PROOF-1 OK / PROOF-2 OK / KNOCK-ON alignment=0 / RESULT: DEFECT CONFIRMED, exit 0); README.md path:line walkthrough; fix.diff",
  "kernel_refs": [
    "sys/dev/drm/amd/display/dc/dce112/dce112_compressor.c:797",
    "sys/dev/drm/amd/display/dc/dce112/dce112_compressor.c:802",
    "sys/dev/drm/amd/display/dc/dce112/dce112_compressor.c:810",
    "sys/dev/drm/amd/display/dc/dce112/dce112_compressor.c:818",
    "sys/dev/drm/amd/display/dc/dce112/dce112_compressor.c:675",
    "sys/dev/drm/amd/display/dc/dce112/dce112_compressor.c:645"
  ],
  "poc_changes": "n/a (demonstrator authored fresh; no seed PoC existed)",
  "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": 30,
  "guest_dirty": 0,
  "build_cmd": "cd /tmp/df4601 && sh build.sh",
  "run_cmd": "/tmp/df4601/run.sh",
  "code_hash": "a9a43abd13da32bd153311fc5cc53a0ea259bc2b21a922381a8acbef5dc6dd35",
  "notes": "attempt 1 failed on the audit host (no cc installed) \u2014 built and run on the guest instead (pure userspace logic mirror, no GPU needed). Enum note: the guest lacks the target device table row maps to not_testable, which the orchestrator mapping disallows; recorded as not_reproduced with impact none. The finding itself is filed at Info severity: no kernel memory-safety consequence, register-programming logic only.",
  "recommended_fix": "In dce112_compressor_construct(), assign memory_bus_width from ctx->asic_id.vram_width BEFORE the ==64 check, and set lpt_channels_num=1 after the zeroing block instead of wiping it (fix.diff).",
  "fix_status": "not_testable",
  "fix_baseline_reproduced": 0,
  "fix_patched_reproduced": 0,
  "fix_verdict": "Single-fix kernel validation impossible: the amd drm driver cannot attach to the QEMU virtio guest, so neither baseline nor patched kernels can exercise the path. fix.diff is a minimal reorder authored against the pristine sys/ tree (applies by construction).",
  "fix_evidence": "fix.diff (assign memory_bus_width before the 64-bit-bus guard; stop wiping lpt_channels_num \u2014 set it to 1 after the zeroing block)"
}