Thanks - agreed, a lock shape reads better here, and v4 will use one. To be clear about why it isn't a plain mutex: admission runs in hard IRQ context (mana_gd_intr() -> mana_gd_process_eqe() -> mana_schedule_serv_work()), and mana_gd_remove() can be entered on the service work's own stack (its rescan exit calls pci_stop_and_remove_bus_device()), so cancel_work_sync() is out and remove() can only wait until the work has stopped touching gc, not for it to finish. v4 expresses this with a small gc-embedded spinlock (hardirq-safe, never held across sleeps) plus a waitqueue for the drain, replacing the flag protocol.
v4 will also fix the two real issues the AI review flagged, which the lock alone does not address: - GC_REMOVING currently stays latched if a rescan exit bails out early (no parent bus) without removing the device; v4 checks for that before closing admission, so service is never permanently disabled on a bound device. - mana_gd_shutdown() now drains as well, since it tears down the same hardware paths remove() does.
