⬢ DragonFlyBSD Kernel Audit
← triage · dashboard
DF-2474

Stale base pointer in isp_pci_mbxdma error cleanup causes kernel-heap multi-free

Summary

isp_pci_mbxdma() uses single local caddr_t base for both control-area allocation and every per-channel FC scratch allocation. FC scratch loop overwrites base on each iteration. When any later step fails (most reachable: bus_dmamap_create ENOMEM in per-command map loop ~256 maps) control reaches bad: label. Cleanup loop frees every FC channel using SAME stale base and final control-area free also uses stale base. Result: contigfree() double/triple-free of one channel virtual address and leak of every other allocation. For isp_nchan=1 (default): fc[0]_base freed twice (once correctly once at :1673 with wrong/larger size) real control-area base leaked. For isp_nchan=N: last channel base freed N+1 times. Unprivileged user can contribute memory pressure while admin kldload/hotplug supplies attach. Classic kernel-heap double-free primitive.

Discussion (0)

No comments yet.

PoC verification

Evidence pack

findings/poc/DF-2474 · 6 files
FileTypeDescriptionSize
VERDICT.md verdict Source trace confirming DF-2474 bug is real but HW-gated 1.4 KB ↓ raw
fix.diff suggested-fix Save control-area base separately and store per-channel base in struct isp_fc 1.6 KB view raw
env.txt environment Guest gate proof: kldstat, pciconf, camcontrol (no target HBA) 1.1 KB view raw
build.sh build-script No buildable PoC (HW-gated) 384 B view raw
run.sh run-script Gate verification commands 411 B view raw
README.md readme Overview of HW-gated finding 791 B ↓ raw
README.md readme Overview of HW-gated finding
↓ download raw

HW-gated SCSI/FC HBA driver finding

This finding targets a SCSI/FC/SATA HBA driver that is not present in this QEMU/KVM guest (only PIIX3 IDE + virtio). The driver cannot attach and the cited code path is unreachable at runtime.

Reproduction status: NOT REPRODUCED (HW-gated)

  • The source bug is confirmed real by line-by-line source trace (see VERDICT.md).
  • It cannot be triggered because the HBA hardware/driver does not exist on this guest.
  • A defense-in-depth fix.diff has been authored and validated with git apply --check.

Evidence

  • VERDICT.md — full source trace and analysis
  • env.txt — guest environment (kldstat, pciconf, camcontrol gate proof)
  • fix.diff — git-apply-able defense-in-depth fix
  • manifest.json — machine-readable catalog
VERDICT.md verdict Source trace confirming DF-2474 bug is real but HW-gated
↓ download raw

DF-2474 — Stale base pointer in isp_pci_mbxdma error cleanup

Verdict: NOT REPRODUCED (HW-gated) — source bug CONFIRMED

Hardware gate

No QLogic ISP HBA in guest: kldstat shows only kernel/ehci/xhci; pciconf -l shows no QLogic device. The isp PCI attach path is never exercised.

Source trace (confirmed real bug)

File: sys/dev/disk/isp/isp_pci.c:1600-1673

isp_pci_mbxdma() uses a single local caddr_t base for both the control-area allocation (line 1600) and every per-channel FC scratch allocation (line 1629). The FC loop overwrites base on each iteration. When any later step fails (e.g., bus_dmamap_create ENOMEM in the per-command map loop at line 1646-1655), control reaches bad: at line 1667. The cleanup loop (1668-1672) frees every FC channel using the same stale base (last channel's address), and the final control-area free at line 1673 also uses the stale base. Result: double/triple-free of one channel's virtual address; leak of every other allocation.

Fix

  • Added caddr_t ctrl_base to save the control-area base before the FC loop
  • Added caddr_t rscratch field to struct isp_fc to store each channel's base
  • bad: cleanup uses fc->rscratch for per-channel frees and ctrl_base for the control-area free

Spans isp_freebsd.h and isp_pci.c. See fix.diff.

Impact (on HW that has the HBA)

Medium — kernel-heap double-free primitive on error cleanup paths.

Fix verification

not_testable
baseline no→ patch + rebuild →patched clean

not_testable: target driver cannot attach (no target HBA). fix.diff validated with git apply --check (EXIT=0); fix is correct by source trace.

git apply --check findings/poc/DF-2474/fix.diff => EXIT=0. No runtime test possible (no target HBA).
↓ fix.diffDragonFly 6.5-DEVELOPMENT #0 (unpatched with-src baseline — target driver not loadable on this guest)

Confirmed kernel references

Detail

Exploit chain

none — valid hard blocker (driver code path dead at runtime on this guest: no target HBA). No unprivileged->root path.

Evidence (decisive lines)

kldstat -> kernel/ehci/xhci only (no target driver); pciconf -l -> PIIX3 IDE + virtio only (no target HBA); camcontrol devlist -> only <QEMU QEMU DVD-ROM>. Source confirmed at cited lines.

PoC changes

Created findings/poc/DF-2474/{VERDICT.md,fix.diff,manifest.json,env.txt,build.sh,run.sh,README.md}. No PoC source (HW-gated).

Verified recommended fix

Defense-in-depth fix.diff closes the cited path (see findings/poc/DF-2474/fix.diff; git apply --check OK).

Verdict

NOT REPRODUCED (HW-gated). The bug is REAL in source (traced line-by-line): isp_pci stale base pointer in isp_pci_mbxdma error cleanup (double/triple-free; needs QLogic ISP + mem pressure). Gate confirmed: kldstat (only kernel+ehci+xhci; no isp/mpt/sili/trm/sbp module loaded), pciconf -l (no QLogic ISP/LSI MPT/SiliconImage/Tekram/FireWire HBA — only PIIX3 IDE atapci0 + virtio), camcontrol devlist (only QEMU DVD-ROM). The target driver cannot attach so the cited code path is dead at runtime on this guest.