Closes the v4.0 plan P3 'acpid pci_fd is not registered' item.
Two related defects:
1. scheme.rs::on_sendfd had a one-shot EINVAL when self.pci_fd was
already Some. Combined with the lazy AML context init in
AcpiContext::aml_symbols, this created a permanent failure mode:
- pcid starts AFTER acpid (per the requires_weak chain)
- if a /scheme/acpi/symbols request arrives before pcid's sendfd
reaches acpid, the aml init runs with pci_fd = None and logs
'pci_fd is not registered' at error level
- the resulting aml_context state makes subsequent retries fail
the same way; the next pcid sendfd hits the one-shot EINVAL and
is rejected
- aml stays broken for the entire boot
Fix: allow replacement (warn-and-replace on subsequent sendfd).
Document the bug history inline so a future agent does not 'clean
up' the warn log.
2. aml_physmem.rs::AmlPhysMemHandler::new logged the missing fd at
log::error. Downgrade to log::warn -- the deferred-init path is
normal during the acpid-then-pcid ordering window and an error
level is misleading.
3. scheme.rs::on_sendfd now eagerly calls self.ctx.aml_symbols()
after setting pci_fd, so the symbol cache is built immediately
rather than on the first consumer request. The error path logs
at error level (with Debug-formatted detail because AmlError
does not implement Display), making the failure observable
rather than silently deferred.
Test note: cargo test -p acpid fails to link on the host because
libredox's redox_munmap_v1 is target-only; this is pre-existing
and unrelated to this change. cargo check -p acpid is clean.
Red Bear OS
b906ad688a