DF-4991

SMB_BREAD passes uninitialized stack byte as the driver's requested count; user's clamped rcount never communicated

Summary

bcount is declared uninit and passed as the in/out count to smbus_bread - the hardware transfer length is leaked stack garbage (ichsmb gates on it AND writes it to ICH_D0) while the user's rcount influences only the copyout. Nondeterministic lengths + a stack-content oracle. Seed-bcount fix.

Discussion (0)

No comments yet.

Severity: Low
Status: new
Bucket: Kernel memory leaks / uninitialized disclosure
Confidence: certain
Pass: 2 · GLM 5.3 second pass
CVSS 3.1: CVSS:3.1/AV:L/AC:L/PR:H/UI:N/S:U/C:N/I:L/A:N
CWE: CWE-457
File: sys/dev/smbus/smb/smb.c
Lines: 217-329
Area: dev
Discovered: 2026-09-25
Updated: 2026-09-25

Team triage

untriaged

priority: — · owner: —

PoC verification

untested

Status: untested
Reproduced: no
Impact: corruption
Confidence: certain
Tested: 2026-09-25 17:15:59
Attempts: 0
Runtime: —
Code hash: —
Guest: —
Build cmd: —
Run cmd: —

jump to evidence pack ↓

Fix verification

not_testable

Fix status: not_testable
Baseline reproduced: —
Patched reproduced: —
Confidence: certain
Tested: 2026-09-25 17:15:59
Runtime: —
Kernel: per-fix-DF-4991
fix.diff: findings/poc/DF-4991/fix.diff
Runner: df-bsd-reviewer

jump to fix verification ↓

PoC verification

Evidence pack

findings/poc/DF-4991 · 1 files

File	Type	Description	Size
verdict.json	file		991 B	view raw

Fix verification

not_testable

per-fix-DF-4991

Confirmed kernel references

Detail

Exploit chain

uninit bcount -> ichsmb EINVAL-or-transfer branch on stack garbage; length observable

Evidence (decisive lines)

findings/poc/DF-4991/ (this verdict)

Verdict

Uninit-input, root-gated (Low): SMB_BREAD declares u_char bcount UNINITIALIZED and passes &bcount as the smbus_bread count (an in/out param) - the hardware transfer length is a leaked stack byte (ichsmb gates the whole op on it and writes it to ICH_D0; ig4 reads it as rcount) while the user's clamped rcount influences only the copyout bound, never the transaction. Nondeterministic lengths + a mini oracle on prior stack contents. No OOB on in-tree backends. Seed-bcount-from-rcount fix.