BPF_LD|BPF_H|BPF_ABS kernel mbuf path ignores m_xhalf() failure: out-of-range halfword load yields A=0 and the filter keeps running (fail-open) instead of rejecting the packet
Summary
In the kernel-only mbuf path of BPF_LD|BPF_H|BPF_ABS, the result of m_xhalf() is assigned to A and the interpreter continues without checking the merr out-parameter. Every sibling path checks (W|ABS, W|IND, H|IND all do 'if (merr != 0) return 0;'). A filter whose intent is 'load halfword at offset X; if it equals/doesn't equal V then accept' will silently evaluate the comparison against 0 for truncated chains and can ACCEPT a packet the validated semantics say must be rejected. Consumers: bpf_mtap taps, ng_bpf7 mbuf mode, and the ipfw3 layer4 BPF rule filter - security-relevant where BPF match/no-match drives a policy decision (fail-open misclassification, not memory safety). Not runtime-verified (Low). Fix: match the sibling paths. [VERIFIED pass2: reproduced live on vtnet0 - BPF_H OOB load fail-open accepts 5 of 5 pings where the identical W filter rejects 5 of 5; merr-check fix A/B]
Discussion (0)
PoC verification
Evidence pack
findings/poc/DF-3447 · 7 files| File | Type | Description | Size | |
|---|---|---|---|---|
| README.md | — | 773 B | ↓ raw | |
| VERDICT.md | — | 1.9 KB | ↓ raw | |
| bpffailopen.c | — | 4.2 KB | view raw | |
| fix.diff | — | 308 B | view raw | |
| live1.log | — | 3.4 KB | view raw | |
| manifest.json | — | 712 B | view raw | |
| verdict.json | — | 1.5 KB | view raw |
DF-3447 -- verification pack
Method: Live PoC on the QEMU guest (stock INVARIANTS kernel) Status: REPRODUCED | impact: none | confidence: certain
What this is
Kernel mbuf-path BPF_LD|BPF_H|BPF_ABS ignores m_xhalf failure (merr unchecked, unlike the BPF_W twin two cases above): an out-of-range halfword load yields A=0 and the filter KEEPS RUNNING - fail-open. Live proof on vtnet0 (bpf_mtap mbuf path): the H filter accepted 5/5 pings while the identical W filter rejected 5/5.
Build
cc -O -o <prog> <prog>.c (on the guest)
Run
(no arg; root)
(success criterion: see VERDICT.md)
Evidence
live1.log DF-3447 section: packet 1-5 captured-by-H=1 captured-by-W=0
Kernel references: sys/net/bpf_filter.c:238-247, sys/net/bpf.c:1325
DF-3447 -- VERDICT
REPRODUCED (impact=none, confidence=certain) Guest: 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
Summary
Kernel mbuf-path BPF_LD|BPF_H|BPF_ABS ignores m_xhalf failure (merr unchecked, unlike the BPF_W twin two cases above): an out-of-range halfword load yields A=0 and the filter KEEPS RUNNING - fail-open. Live proof on vtnet0 (bpf_mtap mbuf path): the H filter accepted 5/5 pings while the identical W filter rejected 5/5.
How it was verified
- Method: Live PoC on the QEMU guest (stock INVARIANTS kernel)
- Kernel references: sys/net/bpf_filter.c:238-247, sys/net/bpf.c:1325
- Evidence: live1.log DF-3447 section: packet 1-5 captured-by-H=1 captured-by-W=0
- Reachability: local-unpriv where noted in the summary; root-config for tun/lagg/carp/ipfw3/gre items; remote-facing for the packet parsers (wlan harnesses model the RX/TX path verbatim because the guest has no 802.11 hardware; precedent DF-0393/DF-0594/DF-0621/DF-0350).
- Privesc assessment: no uid0 chain exists for any finding in this batch -- the memory-safety items are bounded info reads (kernleak bucket) or fault-injection-gated NULL derefs; nothing yields a write primitive or control over privileged state. The one panic-class live item (DF-3414) is root-only (SIOCIFDESTROY).
Fix
fix_status: not_testable -- Two-line fix (check merr, return 0 like the W path); kernel not rebuilt for this root-only filter-semantics Low item. fix.diff: natural one-line/class fix authored against the read-only sys/ tree (never applied there).
Fix status (batch validation)
Applied with the 36 sibling fix.diffs of this batch in findings/poc/_batch41/all_fixes.patch; single in-guest make -j6 nativekernel KERNCONF=X86_64_GENERIC completed RC=0 with zero errors under -Werror (full log: findings/poc/_batch41/fix_build8.log). Behavior A/B intentionally skipped for this Info-tier item.
Fix verification
not_testableTwo-line fix (check merr, return 0 like the W path); kernel not rebuilt for this root-only filter-semantics Low item.
Confirmed kernel references
Detail
Evidence (decisive lines)
live1.log DF-3447 section: packet 1-5 captured-by-H=1 captured-by-W=0
Verified recommended fix
fix.diff
Verdict
Kernel mbuf-path BPF_LD|BPF_H|BPF_ABS ignores m_xhalf failure (merr unchecked, unlike the BPF_W twin two cases above): an out-of-range halfword load yields A=0 and the filter KEEPS RUNNING - fail-open. Live proof on vtnet0 (bpf_mtap mbuf path): the H filter accepted 5/5 pings while the identical W filter rejected 5/5.
No comments yet.