{
  "finding_id": "DF-4370",
  "status": "not_reproduced",
  "reproduced": 0,
  "impact": "none",
  "confidence": "certain",
  "verdict": "Certain by static proof: amdgpu_bo_placement_from_domain() (amdgpu_object.c:139-209) appends one ttm_place per set domain bit into placements[AMDGPU_BO_MAX_PLACEMENTS=3] and checks BUG_ON(c >= 3) only AFTER storing. Any unprivileged render-node client can pass domains=VRAM|GTT|CPU (0x7) through AMDGPU_GEM_CREATE (mask 0x3f accepted at amdgpu_gem.c:233, DRM_RENDER_ALLOW at amdgpu_kms.c:1197): c==3 is in-bounds but BUG_ON(3>=3) panics the kernel. domains=0x3f additionally writes places[3..5] (36 bytes of constants) over bo->placement/bo->tbo before the panic; a second path sets preferred_domains=0x7 via AMDGPU_GEM_OP SET_PLACEMENT (amdgpu_gem.c:727) and panics on the next validate. Not reproduced at runtime because the audit VM has no GPU (no /dev/dri; only QEMU virtio VGA): PoC builds (BUILD_OK) and runs, printing NO_DRM_DEVICES, exit 2. Real-world impact ceiling: reliable local kernel panic (DoS) on amdgpu systems; the OOB write is intra-allocation with constant data and BUG_ON is an unconditional panic (asm/bug.h:33), so no escalation path.",
  "exploit_chain": "",
  "evidence": "VERDICT.md (full line-by-line proof incl. struct layout math); poc_df4370.c (self-contained trigger, both variants); build.log (BUILD_OK); run.log (NO_DRM_DEVICES EXIT=2); env.txt (uname 6.5-DEVELOPMENT, no /dev/dri, virtio-vga only); fix.diff verified APPLIES_CLEAN via patch -p1 --dry-run on a scratch copy",
  "kernel_refs": [
    "sys/dev/drm/amd/amdgpu/amdgpu_object.c:209",
    "sys/dev/drm/amd/amdgpu/amdgpu_object.c:139",
    "sys/dev/drm/amd/amdgpu/amdgpu_object.c:493",
    "sys/dev/drm/amd/amdgpu/amdgpu_object.h:35",
    "sys/dev/drm/amd/amdgpu/amdgpu_gem.c:233",
    "sys/dev/drm/amd/amdgpu/amdgpu_gem.c:727",
    "sys/dev/drm/amd/amdgpu/amdgpu_kms.c:1197",
    "sys/dev/drm/include/asm/bug.h:33"
  ],
  "poc_changes": "No seed present; PoC written from scratch. Self-contained ioctl definitions (DRM_IOCTL_VERSION and AMDGPU GEM_CREATE as _IOWR(0x40,0x00,uint64_t[4])) to avoid userspace header dependency; scans renderD128..130/card0..2, verifies driver==amdgpu via VERSION before firing. One iteration: fixed device-scan bookkeeping that mis-reported NO_AMDGPU when no device opened.",
  "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": 240,
  "guest_dirty": 0,
  "build_cmd": "cc -O2 -o /tmp/poc_df4370 /tmp/poc_df4370.c",
  "run_cmd": "/tmp/poc_df4370  (as uid=1001 maxx)",
  "code_hash": "bc7d84f654953aa2b10c8eb91eeed0ff583c4c489f5cab4be16230086a344230",
  "notes": "Guest phase executed honestly: device absent, hence not_reproduced/none on this VM. Do NOT read impact=none as a refutation: the panic is certain on amdgpu hardware; the sibling finding DF-1717 (different file) shares the requires AMD hw testability limit. BUG_ON in this port maps to panic() unconditionally (sys/dev/drm/include/asm/bug.h:33-39), so the corruption never survives to be exploited.",
  "recommended_fix": "Guard each placement store with if (c < AMDGPU_BO_MAX_PLACEMENTS) and relax BUG_ON(c >= AMDGPU_BO_MAX_PLACEMENTS) to c > AMDGPU_BO_MAX_PLACEMENTS (see fix.diff)",
  "fix_status": "not_testable",
  "fix_baseline_reproduced": 0,
  "fix_patched_reproduced": 0,
  "fix_verdict": "fix.diff authored post-verification; applies cleanly (patch -p1 --dry-run) to a scratch copy of amdgpu_object.c; runtime before/after comparison impossible on this VM (no AMD GPU, no /dev/dri) \u2014 flagged not_testable rather than guessed",
  "fix_evidence": "findings/poc/DF-4370/fix.diff; dry-run output checking file sys/dev/drm/amd/amdgpu/amdgpu_object.c / APPLIES_CLEAN"
}