Unvalidated controller-supplied subq_id and cmd_id in nvme_poll_completions enable kernel memory corruption
Summary
nvme_poll_completions reads DMA-written completion entry and uses its 16-bit subq_id and cmd_id fields directly as indices into sc->subqueues[1024] and subq->reqary[nqe<=256] with no bounds checks. Only protection is KKASSERT (nvme.c:722-723) compiled out in production/release kernels. Malicious/faulty NVMe controller supplies out-of-range indices driving three subsequent writes (subq->subq_head at :714 req->res=*res at :724 req->state=COMPLETED at :727) to arbitrary kernel addresses yielding write-what-where primitive or instant NULL-deref panic during early boot device enumeration. subq_id range 0..65535 OOB for >=1024. cmd_id OOB for >=nqe. Attacker: Thunderbolt/USB4 NVMe enclosure custom FPGA PCIe PCI-passthrough VM buggy SSD firmware. No userspace privilege or interaction required kernel issues Identify Controller during attach malicious device answers with crafted completion. Impact: kernel panic DoS or full memory-corruption write-what-where privilege escalation.
Discussion (0)
PoC verification
Evidence pack
findings/poc/DF-2511 · 7 files| File | Type | Description | Size | |
|---|---|---|---|---|
| VERDICT.md | verdict | HW-gate analysis + source trace + fix summary | 1.8 KB | ↓ raw |
| fix.diff | suggested-fix | git-apply-able defense-in-depth fix (applies + compiles) | 1.1 KB | view raw |
| fix_build.log | build-log | combined build proof: all 18 fix.diffs compile in X86_64_GENERIC (-Werror, rc=0); each diff also passes git apply --check | 404 B | view raw |
| manifest.json | manifest | this catalog | 1.5 KB | view raw |
| env.txt | environment | guest uname + PCI/kldstat/dev inventory proving the gate | 3.6 KB | view raw |
| build.sh | build-script | no-op (HW-gated, nothing to build) | 356 B | view raw |
| run.sh | run-script | no-op (HW-gated, nothing to run) | 268 B | view raw |
DF-2511 — Unvalidated controller-supplied subq_id and cmd_id index fixed arrays (write-what-where)
Verdict
NOT REPRODUCED (hardware / trigger gated). The cited vulnerable code path exists
in sys/dev/disk/nvme/nvme.c and is compiled into the X86_64_GENERIC kernel, but it is not
exercisable at runtime on this QEMU guest because the required HBA / device /
trigger is absent. Confidence in the source bug itself: certain.
Why it cannot be reproduced here (the gate)
Hard-gated: needs an NVMe controller. The guest has NO NVMe controller; nvme_poll_completions never runs.
See env.txt for the full guest PCI/kldstat/dev-node inventory that proves the
gate.
The bug is real in source (traced line-by-line)
Cited path:
- sys/dev/disk/nvme/nvme.c:713-727
- sys/dev/disk/nvme/nvme.h:212
- The vulnerable construct is present verbatim in the current master source
(confirmed by direct read of the cited lines during verification).
- A defense-in-depth fix.diff that closes the path is included and was
validated to apply (git apply --check) and to compile cleanly in a full
X86_64_GENERIC kernel build with -Werror (build rc=0).
Fix
Validate res->tail.subq_id < NVME_MAX_QUEUES and res->tail.cmd_id < subq->nqe before indexing subqueues[]/reqary[] (the KKASSERT is compiled out of production kernels).
The standalone git-apply-able diff is fix.diff.
Reproduce
./build.sh && ./run.sh — both are no-ops on this guest by design (the gate
holds). Exercising the path requires the corresponding HBA/device/trigger
(ATA disk / ATAPI floppy or tape / NVMe controller / AdvanSys HBA / AHCI
controller / a malicious ATAPI device / a CD burner / hot-unplug).
Status
status: not_reproduced | reproduced: 0 | impact: none (HW-gated) fix_status: not_testable (path cannot run on this guest; diff applies + compiles)
Fix verification
not_testablenot_testable: the vulnerable path is unreachable on this guest (no target HW/malicious device). Fix validated structurally: git apply --check OK + full kernel build rc=0 (-Werror).
git apply --check findings/poc/DF-2511/fix.diff -> OK. Combined kernel build: 'NK_DONE rc=0'. No runtime before/after possible (HW-gated).
Confirmed kernel references
Detail
Exploit chain
none — valid hard blocker (driver/device path dead at runtime on this guest: no target HBA / no malicious device / no removable media). No unprivileged->root path.
Evidence (decisive lines)
kldstat -> kernel/ehci/xhci only (no target driver); pciconf -l -> PIIX3 IDE + virtio only (no AHCI/NVMe/AdvanSys HBA); camcontrol devlist -> only <QEMU QEMU DVD-ROM>; ls /dev/<target> -> No such file. Source confirmed at cited lines.
PoC changes
Created findings/poc/DF-2511/{VERDICT.md,fix.diff,env.txt,build.sh,run.sh,manifest.json,fix_build.log}. No PoC source (HW-gated).
Verified recommended fix
Defense-in-depth fix.diff adds the validation/bounds check closing the cited path (see findings/poc/DF-2511/fix.diff; git apply --check OK).
Verdict
NOT REPRODUCED (HW/trigger-gated on this guest). The bug is REAL in source (traced line-by-line): nvme unvalidated controller-supplied subq_id/cmd_id (no NVMe controller). Gate confirmed via kldstat (target driver not loaded; only kernel+ehci+xhci), pciconf -l (no target HBA — only PIIX3 IDE atapci0 + virtio), camcontrol devlist (only QEMU DVD-ROM), and ls /dev (no target disk/tape/fd/nvme nodes). The benign QEMU devices cannot produce the malicious device responses the bugs require.
No comments yet.