driver-manager: v3.0 — major assessment + plan realignment to Linux-driver reuse
Assessment (local/docs/evidence/driver-manager/ASSESSMENT-2026-07-22.md)
from three parallel audits (pcid layer, linux-kpi binding model, Linux
7.1 PCI core) + full first-hand review:
Blockers found:
- B1: /scheme/pci/<addr>/bind never existed in pcid — the P3
bind/aer/uevent patches are orphaned (path-fork recipes don't apply
patches). Every probe would ENOENT -> defer forever. Resolution: claim
collapses into pcid's existing channel ENOLCK exclusivity (the
pcid-spawner model) — driver-manager-only change, no fork surgery.
- B2: linux-kpi pci_register_driver and driver-manager are two competing
claim systems with zero mutual awareness; amdgpu (redox-drm FFI) and
redbear-iwlwifi (manual scan CLI) bypass both. linux-kpi covers ~15%
of the PCI API surface; MSI is synthetic.
- B3: /scheme/pci/pciehp and /scheme/acpi/aer have no producers; both
listeners were inert fail-soft loops.
Plan v3.0:
- Removes stale claims (bind endpoint 'done', pciehp/AER listeners
'done', exclusive_with as 'the CachyOS pattern' — CachyOS actually
uses probe-time -ENODEV handoff (ahci->intel-nvme-remap in 6.17.9
patch) + userspace modprobe.d blacklists; modern_tech 'wired' was
advisory theater).
- New work program: P0 correctness blockers (claim-via-channel,
orphan-patch resolution, modern_tech/exec.rs cleanup, QEMU runtime
gate) -> LDR Linux-Driver-Reuse (unified claim model, spawned-mode
pci_register_driver honoring PCID_CLIENT_CHANNEL, iwlwifi/amdgpu
onboarding, linux_loader TOML pipeline, linux-kpi API completion:
real MSI via pcid_interface, request_regions, pcie_capability, PM)
-> P2 operator/event surface (pcid AER/pciehp producers, scheme
bind/unbind/new_id/remove_id/driver_override/rescan endpoints,
trigger-based deferred retry) -> P3 policy/hygiene (quirk pass
phases, AER recovery actions, per-vendor firmware packaging).
- D5 audit gains a Runtime column: no capability counts without a QEMU
boot proving it.
vasilito
7e8d63ecc1