DF-5467 VERDICT — vga_pci_unmap_bios() shadow-path teardown of a ROM-BAR mapping

STATUS: reproduced (kernel panic on the unpatched guest) / impact: kernel-memory-corruption primitive (PTE teardown + kernel_map VA free over a 128 KiB range owning a 64 KiB mapping, plus a leaked live resource). FIX VALIDATED (fixed kernel: identical PoC sequence completes cleanly, guest stays up).

Root cause (line-accurate)

1. vga_pci_map_bios() chooses the mapping kind by evaluating vga_pci_is_boot_display(dev) at entry — sys/bus/pci/vga_pci.c:175. TRUE ⇒ return a fresh pmap_mapbios(0xC0000, 128K) shadow mapping (:185-186). FALSE ⇒ ROM-BAR resource path (:190-274) returning rman_get_virtual(res) of size rman_get_size(res) (here 64 KiB).

2. The predicate is mutable and latching: - on the ROM path vga_pci_map_bios() itself sets PCIB_BCR_VGA_ENABLE on the parent PCI-PCI bridge (:197-201) — one of the exact inputs vga_pci_is_boot_display() consults (:120-124); - vga_pci_default_unit is sticky: the first evaluation that passes all checks latches it (:94-101, :139).

3. vga_pci_unmap_bios() re-evaluates the predicate (:288) instead of remembering how the BIOS was mapped. When it returns TRUE for a mapping that was created via the ROM path, unmap executes the shadow teardown

   pmap_unmapdev((vm_offset_t)bios, VGA_PCI_BIOS_SHADOW_SIZE) / 128 KiB / (:290)

on the ROM-BAR kva. pmap_unmapdev() clears PTEs via pmap_qremove() (pmap.c:6230, atomic_readandclear_long per page — see pmap.c:2177-2192) and frees kernel_map VA via kmem_free(kernel_map, base, size) (pmap.c:6231) for the full rounded 128 KiB — over a mapping that owns only 64 KiB. The 16 foreign pages of neighboring kernel mappings lose their PTEs, and their vm_map entries are destroyed by the over-range free (kernel_map VA-space corruption: the freed range is re-handable to future allocations while the true owner still holds it). The KASSERTs that would have caught the mismatch (:314-316) sit after the early return (:292) and never execute on this path.

4. Secondary effect: the shadow path returns without touching sc->vga_bios — the ROM resource stays cached (vr_res != NULL, vr_refs == 1), so the next vga_pci_map_bios()/vga_pci_alloc_resource() hands back the stale kva whose PTEs are gone → kernel page fault inside the kernel's own poll loop (:228).

Reachability / threat model

Attacker = firmware/hardware state (or a malicious PCIe device presenting GPU-class + bridge topology), same trust boundary as the DF-0990/DF-2324 family; no userland privilege is involved:

• Secondary GPU behind a PCI-PCI bridge whose BCR.VGA_ENABLE is clear (typical for non-boot GPUs on UEFI systems without CSM), with the GPU's own MEM/IO decode enabled, on a system where no display device passed the boot-display checks during probe (vga_pci_default_unit == -1: headless boxes, or firmware that left all VGA decoding unselected).
• drm attach then runs radeon_read_bios() / amdgpu's analogue (sys/dev/drm/radeon/radeon_bios.c:86-104): vga_pci_map_bios() sees predicate=0 (bridge bit clear at :121), takes the ROM path and sets the bridge bit (:200); the immediately following vga_pci_unmap_bios() sees predicate=1 (first passing evaluation latches :139) and runs the 128 KiB teardown on the 64 KiB ROM mapping.
• The same mismatch arises from any concurrent latch of vga_pci_default_unit between map and unmap (dual-GPU attach interleave).
• Note DF-5468: on devices whose ROM BAR readback preserves the enable bit, DF's lazy pci_alloc_map() fails the ROM allocation first, masking this chain; it is live when the readback masks bit0 or the ROM BAR is pre-reserved (hints / FreeBSD-style enumeration). QEMU's std VGA is in the first category — the PoC therefore seeds the resource-list entry exactly as FreeBSD's enumeration does, and the CONTROL phases prove the seeded path is balanced whenever the predicate does not flip.

Exploitation ceiling (primitive characterization)

The demonstrated fault is the simplest consequence (the ROM mapping's own PTEs). The primitive is stronger: kmem_free(kernel_map, base, 128K) frees a 64 KiB span the kernel still believes allocated (VA double-assignment → two owners for one kernel VA range: a write through the surviving owner and a fresh allocation alias), and pmap_qremove strips PTEs from up to 64 KiB of arbitrary neighboring kernel mappings chosen by allocator adjacency — device MMIO windows, kernel heap, KLD text. A weaponized version (choose neighbor, free VA, re-allocate over a live victim object) plausibly reaches uid=0; this requires controlling kernel_map adjacency, which we did not weaponize in this run (trigger is firmware-state-gated, not unpriv-user- gated). Panic + leak + VA-space corruption are proven; privesc is characterized but not demonstrated.

Reproduction record (this pack)

• serial-full.log — full serial console of the baseline run.

• panic.txt — decisive excerpt:

  DF5467: BUG map -> bios=0xfffff8011887c000 size=0x10000 (ROM-BAR path) DF5467: flip: is_boot_display=1 default_unit=0 DF5467: BUG unmap returned — shadow-path pmap_unmapdev(ROM kva, 128K) ran DF5467: touching bios[0]=0xfffff8011887c000 — expect kernel page fault... Fatal trap 12: page fault while in kernel mode fault virtual address = 0xfffff8011887c000 fault code = supervisor read data, page not present Stopped at df5467_modevent+0x3ff: movzbl (%rbx),%esi

CONTROL phases (predicate consistent) map/unmap/remap/unmap cleanly on the same kernel — the only variable is the predicate flip.

• Baseline kernel: stock snapshot #0: Thu Jul 2 06:02:54 UTC 2026 (env.txt).

Fix validation

• fix.diff — record in the softc how map_bios mapped the BIOS (vga_bios_shadow); unmap tears down that kind. git-apply clean against the repo tree.
• Applied to the guest /usr/src, kernel rebuilt (make -j6 nativekernel KERNCONF=X86_64_GENERIC, rc=0, -Werror; fix_build.log), installed, rebooted (#1: Tue Oct 6 00:18:55 UTC 2026).
• Identical PoC on the fixed kernel (fix_run.log): the BUG-phase unmap now performs a proper ROM-path release; the post-bug remap allocates a fresh mapping and reads the ROM signature 0x55/0xaa; unmap clean; kldload rc=0; guest alive. Baseline panic ⇒ baseline_reproduced=1; patched bad-behavior ⇒ patched_reproduced=0.
• Note: the module's "shadow-path ... ran" kprintf fires unconditionally after vga_pci_unmap_bios() returns; on the fixed kernel the shadow path did NOT run (the remap discriminator proves it). Kept as-is to keep the two runs byte-identical except for the kernel.

PoC changes vs the seeded sketch (honesty log)

• run_root does not forward stdin — initial cat > push produced an empty .c and a vacuous "successful" build; switched to scp -F config.
• KLD Makefile needed generated headers (device_if.h bus_if.h pci_if.h).
• DF kernel env accessor is kgetenv, not getenv.
• Seeded rle size math fixed (32-bit ~mask + 1); stale rle now deleted before re-adding.
• ROM rle seeding added after the live discovery that DF's lazy pci_alloc_map() cannot allocate ROM BARs at all on this guest (DF-5468) — the seed emulates FreeBSD-style pre-reservation.
• Added CONTROL phases and the remap discriminator (a correct ROM-path release also unmaps the ROM, so "touch bios[0]" alone cannot distinguish fixed from unfixed).

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.