{
  "finding_id": "DF-3180",
  "status": "reproduced",
  "reproduced": 1,
  "impact": "panic",
  "confidence": "likely",
  "verdict": "REPRODUCED (window + panic): a 4-thread fhopen(2) race against creat/unlink/close on a tmpfs file panicked the stock INVARIANTS guest within ~60s with assertion (vp)!=NULL && (vp)->v_data!=NULL failed at tmpfs.h:493 and backtrace tmpfs_fhtovp -> tmpfs_alloc_vp -> vrele -> vop_inactive -> tmpfs_inactive \u2014 the vnode reached inactive from inside tmpfs_alloc_vp (called by tmpfs_fhtovp with NO node pin) after reclaim had already torn down the node (v_data NULL). This is strictly PAST DF-3072s DOOMED-interval assert (tmpfs_subr.c:402 did not fire): the node had been reclaimed. On stock (no-INVARIANTS) kernels the same window re-reads the freed node at tmpfs_subr.c:382 (vget ENOENT -> vdrop -> goto loop -> node->tn_vnode on kfree_objd memory) and can install it as vp->v_data at :468 \u2014 heap UAF traced but not executed on a NOINVARIANTS build in this run. Remote reachability requires a tmpfs exported via mountctl(MOUNTCTL_SET_EXPORT) (supported: tmpfs_vnops.c:1879) \u2014 NFS client stale-FH replay; local fhopen analog is root-only, so no local unpriv->root chain claimed.",
  "exploit_chain": "NFS-exported tmpfs: client replays stale FH while last ref drops -> server fhtovp finds node -> alloc_vp blocks -> reclaim frees node -> vget ENOENT -> loop re-reads freed node -> stale node attached to live vnode as v_data -> chunk recycled as another files node -> cross-node type confusion (not executed; requires export + NOINVARIANTS kernel)",
  "evidence": [
    "panic.txt (raw serial)",
    "console_full.txt",
    "VERDICT.md (reconstruction + trace)",
    "fix.diff (tn_links pin in tmpfs_fhtovp)"
  ],
  "kernel_refs": [
    "sys/vfs/tmpfs/tmpfs_vfsops.c:453",
    "sys/vfs/tmpfs/tmpfs_vfsops.c:463",
    "sys/vfs/tmpfs/tmpfs_subr.c:382",
    "sys/vfs/tmpfs/tmpfs_subr.c:432",
    "sys/vfs/tmpfs/tmpfs_subr.c:468",
    "sys/vfs/tmpfs/tmpfs_vnops.c:1852",
    "sys/kern/vfs_lock.c:595"
  ],
  "poc_changes": "seeded concept adjusted: statfs include removed; single shared fhandle refreshed per round with 200us post-getfh pause; 4 racer threads; 600s budget (panicked at ~60s)",
  "attempts": 1,
  "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": 60,
  "guest_dirty": 0,
  "build_cmd": "cc -O2 -pthread -o fh_race fh_race.c",
  "run_cmd": "./fh_race /tmp/fhrace 600   (root; fhopen is SYSCAP_RESTRICTEDROOT)",
  "code_hash": "6d849b3c61111093cb59e74bbf43bb464d814b410aad07fe83f313b3fbf2866d",
  "notes": "Panic string garbled on serial console by concurrent CPU output; reconstructed in VERDICT.md, raw bytes preserved in panic.txt. Distinct from DF-3072 (earlier DOOMED window, assert at tmpfs_subr.c:402): this race completed reclaim (v_data NULL / node freed) and the missing pin is owned by tmpfs_fhtovp. Fix not kernel-validated (time budget spent on DF-3179 fix build); fix.diff pins tn_links around tmpfs_alloc_vp exactly like tmpfs_unmount does.",
  "recommended_fix": "pin node->tn_links across tmpfs_alloc_vp in tmpfs_fhtovp (mirror tmpfs_unmount), or re-validate the node under the token after blocking",
  "fix_status": "not_testable",
  "fix_kernel_uname": "",
  "fix_baseline_reproduced": 1,
  "fix_patched_reproduced": 0,
  "fix_verdict": "fix.diff authored (tn_links pin) but not kernel-build-validated in this run; guest kernel-build budget was spent validating DF-3179",
  "fix_evidence": [
    "fix.diff"
  ]
}