Self-review followups after v4.1, plus two correctness fixes the audit
uncovered.
unified_events:
* Carry RecoveryAction through the Aer variant of UnifiedEvent. The
routing decision (route_to_driver against the latest bind snapshot)
is made once in run() and threaded into the callback, so the
callback does not recompute it. Eliminates a duplicate
route_to_driver call per AER event.
* Update the unified_events test to the new Aer { event, action }
shape.
main.rs:
* Simplified AER callback: no double route_to_driver, no
dead cfg(not(target_os = "redox")) arm. Dispatch gated on
cfg(target_os = "redox") so host builds compile.
* RecoveryAction::Fatal escalation: emit a stable ERROR-level
marker ('AER-FATAL: device=... driver-already-dead escalation
marker ...') before the action_str match returns None. Operator
tooling can grep this marker to detect unrecoverable events that
need human attention. No auto-dispatch on Fatal by design -- the
driver is dead, recovery cannot succeed.
modern_technology:
* iommu_group_for: replace the unmatchable hash-keyed path check
(/scheme/iommu/domain/<hash(bdf)>) with scheme-presence detection
(/scheme/iommu exists). The iommu daemon does NOT expose a BDF to
group lookup, so scheme-presence is the strongest signal
available without opening a handle.
* numa_node_for: replace the wrong /scheme/numad/device/<bdf>
check (numad does NOT expose that path) with the correct
/scheme/proc/numa presence check (numad writes topology there).
* numa_node_env_value: dedupe -- reuse numa_node_for's source
discriminant instead of duplicating the path-existence check.
Tests:
- driver-manager: 63 passed (no change)
- redox-driver-core: 32 passed (no change)
- host build clean
vasilito
1ef6e6c893