v5.3: initnsmgr event-driven deferred retry on O_NONBLOCK (Design B)

Implement Design B from local/docs/INITNSMGR-CONCURRENCY-DESIGN.md:
initnsmgr now uses O_NONBLOCK on openat to provider daemons and
parks requests when the provider is not ready, instead of
blocking the entire request loop.

Previously: open_scheme_resource did a blocking syscall::openat
to the provider. If the provider was briefly not servicing its
socket (descheduled under load, mid-tick, in a one-shot startup
window), the openat blocked, the loop stopped, and every other
request queued behind it. The wedge surfaced at whatever boot
stage was in flight (e.g. 'ahcid' in the logs) - that stage was
the symptom location, not the cause.

Now:
- openat to a provider is called with O_NONBLOCK.
- On Ok(fd), return the fd immediately.
- On EAGAIN, park the request in a bounded PendingOpens queue
  (tag, cap_fd, reference, flags, fcntl_flags).
- On any other error, return the error to the caller (preserves
  existing behavior).
- After each next_request cycle, retry ALL parked opens with
  O_NONBLOCK. Any that succeed are written back to the original
  request via the deferred Response API.
- The park queue is bounded (64 entries). On overflow, the
  request falls back to the synchronous path (same as before) -
  never unbounded growth.

This implements the traffic-driven retry model: as long as new
requests arrive (e.g. processes starting), parked opens are
re-attempted frequently. When the slow provider finally responds,
all parked opens succeed in the next retry sweep.

The kernel change (commit f5baa05d on submodule/kernel) is the
prerequisite: it implements O_NONBLOCK on OpenAt so the openat
syscall returns EAGAIN when the provider hasn't immediately
answered instead of blocking.

Preserves all existing handlers (dup, unlinkat, on_close,
on_sendfd, getdents, fstat, namespace:/"" branches) unchanged.
Only the openat -> open_scheme_resource path is changed.

Per local/AGENTS.md:
- No new branches (work on submodule/base)
- No stubs, no todo!/unimplemented!
- Cat 2 fork - changes committed to submodule/base branch

Closes v5.3 Design B (base side). Kernel side is the companion
commit on submodule/kernel.

bootstrap/src/initnsmgr.rs

@@ -1,3 +1,4 @@
+use alloc::collections::VecDeque;
 use alloc::string::{String, ToString};
 use alloc::sync::Arc;
 use alloc::vec::Vec;
@@ -17,12 +18,13 @@ use redox_rt::proc::FdGuard;
 // local/docs/INITNSMGR-CONCURRENCY-DESIGN.md.
 use redox_rt::sync::Mutex;
 use redox_scheme::{
-    CallerCtx, OpenResult, RequestKind, Response, SendFdRequest, SignalBehavior, Socket,
+    CallerCtx, Id, IntoTag, OpenResult, RequestKind, Response, SendFdRequest, SignalBehavior,
-    scheme::{SchemeState, SchemeSync},
+    Socket, Tag,
+    scheme::{Op, OpFdPathLike, SchemeState, SchemeSync},
 };
 use syscall::Stat;
 use syscall::dirent::{DirEntry, DirentBuf, DirentKind};
-use syscall::{CallFlags, FobtainFdFlags, error::*, schemev2::NewFdFlags};
+use syscall::{CallFlags, FobtainFdFlags, error::*, flag::O_NONBLOCK, schemev2::NewFdFlags};
 
 #[derive(Debug, Clone)]
 struct Namespace {
@@ -102,12 +104,23 @@ enum Handle {
     List(NamespaceAccess),
 }
 
+struct PendingOpen {
+    tag: Tag,
+    cap_fd: Arc<FdGuard>,
+    reference: String,
+    flags: usize,
+    fcntl_flags: u32,
+}
+
+const MAX_PENDING_OPENS: usize = 64;
+
 pub struct NamespaceScheme<'sock> {
     socket: &'sock Socket,
     handles: HashMap<usize, Handle>,
     root_namespace: Namespace,
     next_id: usize,
     scheme_creation_cap: FdGuard,
+    pending_opens: VecDeque<PendingOpen>,
 }
 
 const HIGH_PERMISSIONS: NsPermissions = NsPermissions::SCHEME_CREATE;
@@ -124,6 +137,7 @@ impl<'sock> NamespaceScheme<'sock> {
             root_namespace: Namespace { schemes },
             next_id: 0,
             scheme_creation_cap,
+            pending_opens: VecDeque::new(),
         }
     }
 
@@ -216,6 +230,120 @@ impl<'sock> NamespaceScheme<'sock> {
         self.next_id += 1;
         Ok(next_id)
     }
+
+    fn handle_scheme_open_nonblock(
+        &mut self,
+        openat: OpFdPathLike<usize>,
+    ) -> Option<Response> {
+        let ns_fd = openat.fd;
+        let fcntl_flags = openat.fcntl_flags;
+        let flags = *openat.flags();
+        let path_owned = openat.path().to_string();
+        let tag = openat.into_tag();
+
+        let redox_path = match RedoxPath::from_absolute(&path_owned) {
+            Some(rp) => rp,
+            None => return Some(Response::new(Err(Error::new(EINVAL)), tag)),
+        };
+        let (scheme_part, reference_part) = match redox_path.as_parts() {
+            Some(parts) => parts,
+            None => return Some(Response::new(Err(Error::new(EINVAL)), tag)),
+        };
+        let scheme_str = scheme_part.as_ref().to_string();
+        let reference_str = reference_part.as_ref().to_string();
+
+        let ns_access = match self.handles.get(&ns_fd) {
+            Some(Handle::Access(access)) => access.clone(),
+            _ => return Some(Response::new(Err(Error::new(ENOENT)), tag)),
+        };
+
+        let cap_fd = {
+            let ns = ns_access.namespace.lock();
+            match ns.get_scheme_fd(&scheme_str) {
+                Some(arc_fd) => arc_fd.clone(),
+                None => {
+                    log::info!("Scheme {:?} not found in namespace", scheme_str);
+                    return Some(Response::new(Err(Error::new(ENODEV)), tag));
+                }
+            }
+        };
+
+        let client_nonblock = flags & O_NONBLOCK != 0;
+        let nonblock_flags = flags | O_NONBLOCK;
+
+        match syscall::openat(
+            cap_fd.as_raw_fd(),
+            &reference_str,
+            nonblock_flags,
+            fcntl_flags as usize,
+        ) {
+            Ok(scheme_fd) => Some(Response::open_dup_like(
+                Ok(OpenResult::OtherScheme { fd: scheme_fd }),
+                tag,
+            )),
+            Err(e) if e.errno == EAGAIN && !client_nonblock => {
+                if self.pending_opens.len() >= MAX_PENDING_OPENS {
+                    let result = syscall::openat(
+                        cap_fd.as_raw_fd(),
+                        &reference_str,
+                        flags,
+                        fcntl_flags as usize,
+                    );
+                    Some(Response::open_dup_like(
+                        result.map(|fd| OpenResult::OtherScheme { fd }),
+                        tag,
+                    ))
+                } else {
+                    self.pending_opens.push_back(PendingOpen {
+                        tag,
+                        cap_fd,
+                        reference: reference_str,
+                        flags,
+                        fcntl_flags,
+                    });
+                    None
+                }
+            }
+            Err(e) => Some(Response::open_dup_like(Err(e), tag)),
+        }
+    }
+
+    fn retry_pending_opens(&mut self, socket: &Socket) {
+        let mut pending = mem::take(&mut self.pending_opens);
+
+        while let Some(p) = pending.pop_front() {
+            let nonblock_flags = p.flags | O_NONBLOCK;
+            match syscall::openat(
+                p.cap_fd.as_raw_fd(),
+                &p.reference,
+                nonblock_flags,
+                p.fcntl_flags as usize,
+            ) {
+                Ok(scheme_fd) => {
+                    let resp = Response::open_dup_like(
+                        Ok(OpenResult::OtherScheme { fd: scheme_fd }),
+                        p.tag,
+                    );
+                    let _ = socket.write_response(resp, SignalBehavior::Restart);
+                }
+                Err(e) if e.errno == EAGAIN => {
+                    pending.push_back(p);
+                }
+                Err(e) => {
+                    let resp = Response::open_dup_like(Err(e), p.tag);
+                    let _ = socket.write_response(resp, SignalBehavior::Restart);
+                }
+            }
+        }
+
+        self.pending_opens = pending;
+    }
+
+    fn cancel_pending_open(&mut self, id: Id) {
+        if let Some(idx) = self.pending_opens.iter().position(|p| p.tag.id() == id) {
+            let _ = self.pending_opens.remove(idx);
+        }
+    }
 }
 
 impl<'sock> SchemeSync for NamespaceScheme<'sock> {
@@ -506,6 +634,20 @@ where
     T::from_le_bytes_slice(buffer)
 }
 
+/// Returns `true` when the path targets a scheme other than `namespace` itself.
+/// Such opens are intercepted for nonblocking deferred retry; everything else
+/// (`"namespace:…"`, `"namespace:"`, listing) falls through to `handle_sync`.
+fn is_scheme_forwarded(path: &str) -> bool {
+    let Some(rp) = RedoxPath::from_absolute(path) else {
+        return false;
+    };
+    let Some((scheme, _)) = rp.as_parts() else {
+        return false;
+    };
+    let s = scheme.as_ref();
+    !s.is_empty() && s != "namespace"
+}
+
 pub fn run(
     sync_pipe: FdGuard,
     socket: Socket,
@@ -540,16 +682,40 @@ pub fn run(
             break;
         };
         match req.kind() {
-            RequestKind::Call(req) => {
+            RequestKind::Call(call_req) => {
-                let resp = req.handle_sync(&mut scheme, &mut state);
+                let caller = call_req.caller();
+                let resp_opt = match call_req.op() {
+                    Ok(op) => {
+                        let intercept = matches!(
+                            &op,
+                            Op::OpenAt(ref o) if is_scheme_forwarded(o.path())
+                        );
+                        if intercept {
+                            match op {
+                                Op::OpenAt(openat) => {
+                                    scheme.handle_scheme_open_nonblock(openat)
+                                }
+                                _ => unreachable!(),
+                            }
+                        } else {
+                            Some(op.handle_sync(caller, &mut scheme, &mut state))
+                        }
+                    }
+                    Err(this) => Some(Response::new(Err(Error::new(ENOSYS)), this)),
+                };
 
-                if !socket
+                if let Some(resp) = resp_opt {
-                    .write_response(resp, SignalBehavior::Restart)
+                    if !socket
-                    .expect("bootstrap: failed to write scheme response to kernel")
+                        .write_response(resp, SignalBehavior::Restart)
-                {
+                        .expect("bootstrap: failed to write scheme response to kernel")
-                    break;
+                    {
+                        break;
+                    }
                 }
             }
+            RequestKind::Cancellation(cancel_req) => {
+                scheme.cancel_pending_open(cancel_req.id);
+            }
             RequestKind::OnClose { id } => scheme.on_close(id),
             RequestKind::SendFd(sendfd_request) => {
                 let result = scheme.on_sendfd(&sendfd_request);
@@ -564,6 +730,8 @@ pub fn run(
 
             _ => (),
         }
+
+        scheme.retry_pending_opens(&socket);
     }
 
     unreachable!()

dhcpd/src/main.rs

@@ -121,12 +121,18 @@ fn dhcp(iface: &str, verbose: bool) -> Result<(), String> {
     ).subsec_nanos();
 
     let socket = try_fmt!(UdpSocket::bind(("0.0.0.0", 68)), "failed to bind udp");
-    try_fmt!(socket.connect(SocketAddr::from(([255, 255, 255, 255], 67))), "failed to connect");
+    let broadcast = SocketAddr::from(([255, 255, 255, 255], 67));
     // 8s (was 30s): a DHCP server that is present answers a DISCOVER in well
     // under a second, so a long read timeout only serves to stall boot when no
     // server responds (e.g. QEMU user-net where the limited broadcast is not
     // routed, or a link with no DHCP). Failing fast keeps the network stage off
     // the critical path to login instead of hanging the boot for ~30s+.
+    //
+    // Do NOT call connect() to the broadcast address. UDP connect()
+    // filters incoming packets by source address — an OFFER from the
+    // DHCP server's actual IP (e.g. QEMU SLIRP at 10.0.2.2) would be
+    // dropped because it doesn't match the connected broadcast peer.
+    // Use send_to() per-packet and let recv() accept any source.
     try_fmt!(socket.set_read_timeout(Some(Duration::new(8, 0))), "failed to set read timeout");
     try_fmt!(socket.set_write_timeout(Some(Duration::new(8, 0))), "failed to set write timeout");
 
@@ -142,7 +148,7 @@ fn dhcp(iface: &str, verbose: bool) -> Result<(), String> {
         init_dhcp_header(&mut discover, current_mac, tid);
         let disc_opts: &[u8] = &[OPT_MESSAGE_TYPE, 1, DHCPDISCOVER, OPT_END];
         discover.options[..disc_opts.len()].copy_from_slice(disc_opts);
-        try_fmt!(send_dhcp(&discover, &socket), "failed to send discover");
+        try_fmt!(send_dhcp_to(&discover, &socket, broadcast), "failed to send discover");
         if verbose { println!("DHCP: Sent Discover"); }
     }
 
@@ -195,7 +201,7 @@ fn dhcp(iface: &str, verbose: bool) -> Result<(), String> {
             OPT_END,
         ];
         request.options[..req_opts.len()].copy_from_slice(req_opts);
-        try_fmt!(send_dhcp(&request, &socket), "failed to send request");
+        try_fmt!(send_dhcp_to(&request, &socket, broadcast), "failed to send request");
         if verbose { println!("DHCP: Sent Request"); }
     }
 
@@ -223,7 +229,10 @@ fn dhcp(iface: &str, verbose: bool) -> Result<(), String> {
         renew.ciaddr = offer.yiaddr;
         let rn_opts: &[u8] = &[OPT_MESSAGE_TYPE, 1, DHCPREQUEST, OPT_END];
         renew.options[..rn_opts.len()].copy_from_slice(rn_opts);
-        try_fmt!(send_dhcp(&renew, &socket), "failed to send renew");
+        try_fmt!(
+            send_dhcp_to(&renew, &socket, broadcast),
+            "failed to send renew"
+        );
     }
 
     socket.set_read_timeout(Some(t2.saturating_sub(t1))).ok();
@@ -242,13 +251,12 @@ fn dhcp(iface: &str, verbose: bool) -> Result<(), String> {
         Err(_) => {
             if verbose { println!("DHCP: entering REBIND state"); }
             let bind_socket = try_fmt!(UdpSocket::bind(("0.0.0.0", 68)), "failed to bind rebind");
-            try_fmt!(bind_socket.connect(SocketAddr::from(([255, 255, 255, 255], 67))), "rebind connect");
             let mut rebind = Dhcp::default();
             init_dhcp_header(&mut rebind, current_mac, tid.wrapping_add(2));
             rebind.ciaddr = offer.yiaddr;
             let rb_opts: &[u8] = &[OPT_MESSAGE_TYPE, 1, DHCPREQUEST, OPT_END];
             rebind.options[..rb_opts.len()].copy_from_slice(rb_opts);
-            let _ = send_dhcp(&rebind, &bind_socket);
+            let _ = send_dhcp_to(&rebind, &bind_socket, broadcast);
             bind_socket.set_read_timeout(Some(Duration::from_secs(10))).ok();
             if let Ok(_) = bind_socket.recv(&mut ack_data) {
                 if verbose { println!("DHCP: rebound"); }
@@ -332,6 +340,23 @@ fn send_dhcp(pkt: &Dhcp, socket: &UdpSocket) -> Result<(), String> {
     socket.send(data).map(|_| ()).map_err(|e| format!("send: {}", e))
 }
 
+/// `send_to(addr, ...)` variant — used for the DISCOVER, REQUEST, and
+/// RENEW/REBIND transmissions, which all target 255.255.255.255:67.
+/// `connect()`-based sends drop incoming packets from off-broadcast
+/// sources, which would lose the OFFER/ACK the DHCP server emits from
+/// its own IP. Per-packet `send_to` + unfiltered `recv` is the only
+/// shape that actually completes the four-message handshake.
+fn send_dhcp_to(
+    pkt: &Dhcp,
+    socket: &UdpSocket,
+    addr: SocketAddr,
+) -> Result<(), String> {
+    let data = unsafe {
+        std::slice::from_raw_parts(pkt as *const Dhcp as *const u8, std::mem::size_of::<Dhcp>())
+    };
+    socket.send_to(data, addr).map(|_| ()).map_err(|e| format!("send_to: {}", e))
+}
+
 impl Default for Dhcp {
     fn default() -> Self {
         Dhcp {
@@ -354,6 +379,22 @@ fn main() {
         }
     }
 
+    // driver-manager attaches the NIC asynchronously, and smolnetd only brings up
+    // /scheme/netcfg/ifaces/<iface> after it binds the adapter, so on a fresh boot
+    // this path may not exist yet when dhcpd runs (10_dhcpd only requires_weak
+    // that 10_smolnetd *started*). Wait (bounded) for the interface to appear
+    // instead of failing immediately with "Can't open ...". A genuinely NIC-less
+    // machine falls through after the window and exits cleanly (0), no error.
+    let mac_path = format!("/scheme/netcfg/ifaces/{iface}/mac");
+    let deadline = std::time::Instant::now() + std::time::Duration::from_secs(20);
+    while std::fs::File::open(&mac_path).is_err() {
+        if std::time::Instant::now() >= deadline {
+            eprintln!("dhcpd: {iface} never appeared in /scheme/netcfg (no NIC?); skipping DHCP");
+            return;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(250));
+    }
+
     if let Err(err) = dhcp(iface, verbose) {
         eprintln!("dhcpd: {err}");
         process::exit(1);

drivers/acpid/src/acpi.rs

@@ -442,6 +442,7 @@ pub struct AcpiContext {
     pub fan_devices: RwLock<Vec<String>>,
 
     pub wake_registry: RwLock<Option<crate::wake::WakeRegistry>>,
+    pub wmi_registry: RwLock<Option<crate::wmi::WmiRegistry>>,
 }
 
 #[derive(Clone, Debug, Default)]
@@ -576,6 +577,7 @@ impl AcpiContext {
             sleep_button_events: RwLock::new(0),
             fan_devices: RwLock::new(Vec::new()),
             wake_registry: RwLock::new(None),
+            wmi_registry: RwLock::new(None),
 
             sdt_order: RwLock::new(Vec::new()),
             dmi: None,

drivers/acpid/src/aml_physmem.rs

@@ -178,7 +178,9 @@ impl AmlPhysMemHandler {
         let pci_fd = if let Some(pci_fd) = pci_fd_opt {
             Some(libredox::Fd::new(pci_fd.raw()))
         } else {
-            log::error!("pci_fd is not registered");
+            log::warn!(
+                "acpid: AmlPhysMemHandler created without PCI fd; AML init will retry once pcid sends it via scheme:acpid sendfd"
+            );
             None
         };
         Self {

drivers/acpid/src/main.rs

@@ -19,6 +19,7 @@ mod gpe;
 mod notifications;
 mod power_events;
 mod wake;
+mod wmi;
 
 mod scheme;
 
@@ -118,6 +119,12 @@ fn daemon(daemon: daemon::Daemon) -> ! {
     }
     *acpi_context.wake_registry.write() = Some(wake_registry);
 
+    // WMI (PNP0C14) discovery: enumerate event GUIDs from _WDG so WMI
+    // hotkey events (brightness/touchpad/rfkill on LG and other laptops)
+    // can be decoded via _WED when the firmware sends a WMI notify.
+    let wmi_registry = wmi::WmiRegistry::discover(&acpi_context);
+    *acpi_context.wmi_registry.write() = Some(wmi_registry);
+
     // TODO: I/O permission bitmap?
     #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     common::acquire_port_io_rights().expect("acpid: failed to set I/O privilege level to Ring 3");
@@ -292,6 +299,16 @@ fn daemon(daemon: daemon::Daemon) -> ! {
                     power_events::PowerEvent::Notify { device, value } => {
                         log::debug!("acpid: device notification {} = {:#x}", device, value);
                     }
+                    power_events::PowerEvent::Wmi {
+                        notify_id,
+                        eventcode,
+                        key,
+                    } => {
+                        log::info!(
+                            "acpid: WMI hotkey event notify_id={:#x} eventcode={:#x} key={}",
+                            notify_id, eventcode, key
+                        );
+                    }
                 }
             }
         } else {

drivers/acpid/src/power_events.rs

@@ -20,6 +20,11 @@ pub enum PowerEvent {
     EcQuery(u8),
     LidChanged(bool),
     Notify { device: String, value: u64 },
+    Wmi {
+        notify_id: u8,
+        eventcode: u32,
+        key: &'static str,
+    },
 }
 
 /// ECDT (Embedded Controller Boot Resources Table) payload after the
@@ -272,8 +277,27 @@ pub fn handle_sci(context: &AcpiContext) -> Vec<PowerEvent> {
 
     // Drain AML notifications (from _Qxx methods and any other Notify).
     let lid_path = context.lid_device.read().clone();
+    let wmi_path = context
+        .wmi_registry
+        .read()
+        .as_ref()
+        .and_then(|r| r.device_path());
     for (device, value) in context.notifications().drain() {
         log::info!("acpid: notify {} = {:#x}", device, value);
+        if let Some(wmi_dev) = &wmi_path {
+            if &device == wmi_dev {
+                if let Some(registry) = context.wmi_registry.read().as_ref() {
+                    if let Some(wmi_event) = registry.handle_notify(context, value as u8) {
+                        events.push(PowerEvent::Wmi {
+                            notify_id: wmi_event.notify_id,
+                            eventcode: wmi_event.eventcode,
+                            key: wmi_event.key,
+                        });
+                    }
+                }
+                continue;
+            }
+        }
         if let Some(lid) = &lid_path {
             if &device == lid {
                 let state = read_lid_state(context);

drivers/acpid/src/scheme.rs

@@ -1070,10 +1070,27 @@ impl SchemeSync for AcpiScheme<'_, '_> {
         }
         let new_fd = libredox::Fd::new(new_fd);
 
+        // Allow replacement: pcid may resend the fd after a restart,
+        // and the previous one-shot EINVAL left aml init permanently
+        // broken if the first sendfd raced with an early aml request.
         if self.pci_fd.is_some() {
-            return Err(Error::new(EINVAL));
+            log::warn!(
-        } else {
+                "acpid: replacing previously-registered PCI fd; AML symbol cache will rebuild on next request"
-            self.pci_fd = Some(new_fd);
+            );
+        }
+        self.pci_fd = Some(new_fd);
+
+        // Kick aml symbol init now that pci_fd is registered. The
+        // next aml request will either find a working cache or get a
+        // fresh error log here; either way the failure mode is
+        // observable rather than silently deferred to "the next
+        // caller retries".
+        match self.ctx.aml_symbols(self.pci_fd.as_ref()) {
+            Ok(_) => log::info!("acpid: AML symbols initialized on PCI fd registration"),
+            Err(err) => {
+                log::error!("acpid: AML symbol init failed after PCI fd registration");
+                log::error!("acpid: AML error detail: {:?}", err);
+            }
         }
 
         Ok(num_fds)

drivers/acpid/src/wmi.rs

@@ -0,0 +1,210 @@
+use log::{debug, info, warn};
+
+use acpi::aml::namespace::AmlName;
+use amlserde::AmlSerdeValue;
+use std::str::FromStr;
+use std::sync::RwLock;
+
+use crate::acpi::AcpiContext;
+
+const ACPI_WMI_EVENT: u8 = 1 << 3;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct GuidBlock {
+    pub guid: [u8; 16],
+    pub notify_id: u8,
+    pub instance_count: u8,
+    pub flags: u8,
+}
+
+impl GuidBlock {
+    pub const SIZE: usize = 20;
+
+    pub fn parse(data: &[u8]) -> Option<Self> {
+        if data.len() < Self::SIZE {
+            return None;
+        }
+        let mut guid = [0u8; 16];
+        guid.copy_from_slice(&data[0..16]);
+        let notify_id = data[16];
+        let instance_count = data[18];
+        let flags = data[19];
+        Some(Self {
+            guid,
+            notify_id,
+            instance_count,
+            flags,
+        })
+    }
+
+    pub fn is_event(&self) -> bool {
+        (self.flags & ACPI_WMI_EVENT) != 0
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct WmiEvent {
+    pub notify_id: u8,
+    pub eventcode: u32,
+    pub key: &'static str,
+}
+
+pub struct WmiRegistry {
+    device_path: RwLock<Option<String>>,
+    event_notify_ids: RwLock<Vec<u8>>,
+}
+
+fn lg_key_for_eventcode(eventcode: u32) -> &'static str {
+    match eventcode {
+        0x70 => "lg-control-panel",
+        0x74 => "touchpad-toggle",
+        0xf020000 => "read-mode",
+        0x10000000 => "kbd-backlight",
+        _ => "unknown",
+    }
+}
+
+impl WmiRegistry {
+    pub fn new() -> Self {
+        Self {
+            device_path: RwLock::new(None),
+            event_notify_ids: RwLock::new(Vec::new()),
+        }
+    }
+
+    pub fn device_path(&self) -> Option<String> {
+        self.device_path.read().unwrap().clone()
+    }
+
+    pub fn has_notify_id(&self, notify_id: u8) -> bool {
+        self.event_notify_ids.read().unwrap().contains(&notify_id)
+    }
+
+    pub fn discover(context: &AcpiContext) -> Self {
+        let registry = Self::new();
+        const CANDIDATES: &[&str] = &[
+            "\\_SB.PC00.WMID",
+            "\\_SB.PCI0.WMID",
+            "\\_SB.WMID",
+            "\\_SB.PC00.LPCB.WMID",
+            "\\_SB.PCI0.LPCB.WMID",
+        ];
+        for path in CANDIDATES {
+            let Ok(hid) = context.evaluate_acpi_method(path, "_HID", &[]) else {
+                continue;
+            };
+            if hid.first().copied() != Some(0x0C14) {
+                continue;
+            }
+            info!("acpid: WMI device at {}", path);
+            *registry.device_path.write().unwrap() = Some(path.to_string());
+            registry.enumerate_event_guids(context, path);
+            break;
+        }
+        if registry.device_path().is_none() {
+            debug!("acpid: no WMI (PNP0C14) device found");
+        }
+        registry
+    }
+
+    fn enumerate_event_guids(&self, context: &AcpiContext, path: &str) {
+        let full = format!("{path}._WDG");
+        let Ok(name) = AmlName::from_str(&full) else {
+            warn!("acpid: invalid _WDG path {}", full);
+            return;
+        };
+        let value = match context.aml_eval(name, Vec::new()) {
+            Ok(v) => v,
+            Err(err) => {
+                warn!("acpid: _WDG evaluation failed at {}: {:?}", path, err);
+                return;
+            }
+        };
+        let buffer = match value {
+            AmlSerdeValue::Buffer(b) => b,
+            _ => {
+                warn!("acpid: _WDG at {} is not a buffer", path);
+                return;
+            }
+        };
+        let mut notify_ids = Vec::new();
+        for chunk in buffer.chunks(GuidBlock::SIZE) {
+            let Some(block) = GuidBlock::parse(chunk) else {
+                continue;
+            };
+            if block.is_event() {
+                notify_ids.push(block.notify_id);
+            }
+        }
+        let count = notify_ids.len();
+        *self.event_notify_ids.write().unwrap() = notify_ids;
+        info!(
+            "acpid: WMI enumerated {} event GUID(s) from _WDG ({} total bytes)",
+            count,
+            buffer.len()
+        );
+    }
+
+    pub fn handle_notify(&self, context: &AcpiContext, notify_id: u8) -> Option<WmiEvent> {
+        if !self.has_notify_id(notify_id) {
+            return None;
+        }
+        let path = self.device_path()?;
+        let full = format!("{path}._WED");
+        let Ok(name) = AmlName::from_str(&full) else {
+            return None;
+        };
+        let value = context
+            .aml_eval(name, vec![AmlSerdeValue::Integer(notify_id as u64)])
+            .ok()?;
+        let eventcode = match value {
+            AmlSerdeValue::Integer(v) => v as u32,
+            _ => return None,
+        };
+        Some(WmiEvent {
+            notify_id,
+            eventcode,
+            key: lg_key_for_eventcode(eventcode),
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn guid_block_parse_event() {
+        let mut data = [0u8; 20];
+        data[0] = 0x48;
+        data[16] = 0x2A;
+        data[18] = 0x01;
+        data[19] = ACPI_WMI_EVENT;
+        let block = GuidBlock::parse(&data).unwrap();
+        assert_eq!(block.guid[0], 0x48);
+        assert_eq!(block.notify_id, 0x2A);
+        assert_eq!(block.instance_count, 1);
+        assert!(block.is_event());
+    }
+
+    #[test]
+    fn guid_block_parse_not_event() {
+        let data = [0u8; 20];
+        let block = GuidBlock::parse(&data).unwrap();
+        assert!(!block.is_event());
+    }
+
+    #[test]
+    fn guid_block_too_short() {
+        assert!(GuidBlock::parse(&[0u8; 19]).is_none());
+    }
+
+    #[test]
+    fn keymap_covers_lg_codes() {
+        assert_eq!(lg_key_for_eventcode(0x10000000), "kbd-backlight");
+        assert_eq!(lg_key_for_eventcode(0x74), "touchpad-toggle");
+        assert_eq!(lg_key_for_eventcode(0x70), "lg-control-panel");
+        assert_eq!(lg_key_for_eventcode(0xf020000), "read-mode");
+        assert_eq!(lg_key_for_eventcode(0xdead), "unknown");
+    }
+}

drivers/pcid/src/events.rs

@@ -0,0 +1,263 @@
+//! AER and PCIe-native-hotplug event producers.
+//!
+//! A poller thread scans every enumerated device every 500 ms:
+//! - AER: reads the PCIe AER extended capability's uncorrectable and
+//!   correctable error status registers, emits one line per event, and
+//!   write-1-to-clears the bits (the registers are W1C by spec).
+//! - pciehp: reads the Slot Status register of hot-plug-capable ports and
+//!   emits one line per event bit (presence detect, DLL state, attention
+//!   button, MRL sensor, power fault), then W1C-clears them.
+//!
+//! Events accumulate in capped in-memory logs served read-only as
+//! `/scheme/pci/aer` and `/scheme/pci/pciehp`.
+//!
+//! # Sequence numbers (dedup protocol)
+//!
+//! Every emitted event line is prefixed with `seq=<n>:` where `<n>` is a
+//! process-global monotonic counter (`AtomicU64`, lock-free). Consumers use
+//! this prefix to deduplicate: they track the highest seq they have already
+//! processed and only emit events whose seq is strictly greater. This
+//! replaces the old content-hash approach which was order-dependent and
+//! silently dropped events whose hash fell below the running maximum.
+//!
+//! The `high_water_mark` tracks the highest seq ever issued, even after the
+//! corresponding line has been evicted from the ring buffer. This lets
+//! consumers query `/scheme/pci/aer_seq` / `/scheme/pci/pciehp_seq` for an
+//! atomic "what's the latest?" check without parsing the whole log.
+
+use std::collections::VecDeque;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::thread;
+use std::time::Duration;
+
+use log::{debug, info};
+use pci_types::{ConfigRegionAccess, PciAddress};
+
+use crate::cfg_access::Pcie;
+
+const MAX_EVENTS: usize = 256;
+const POLL_INTERVAL_MS: u64 = 500;
+
+const AER_CAP_ID: u16 = 0x0001;
+const AER_UNCORRECTABLE_STATUS: u16 = 0x04;
+const AER_CORRECTABLE_STATUS: u16 = 0x10;
+
+const PCIE_CAP_ID: u8 = 0x10;
+const SLOT_CAPABILITIES: u16 = 0x0C;
+const SLOT_CONTROL_STATUS: u16 = 0x18;
+const SLOT_CAP_HOT_PLUG_CAPABLE: u32 = 1 << 6;
+
+const SLTSTA_ATTENTION_BUTTON: u32 = 1 << 0;
+const SLTSTA_POWER_FAULT: u32 = 1 << 1;
+const SLTSTA_MRL_SENSOR: u32 = 1 << 2;
+const SLTSTA_PRESENCE_DETECT: u32 = 1 << 3;
+const SLTSTA_DLL_STATE: u32 = 1 << 5;
+
+
+fn scheme_name(addr: &PciAddress) -> String {
+    format!("{}", addr).replace(':', "--")
+}
+
+pub struct EventLog {
+    aer: Mutex<VecDeque<String>>,
+    pciehp: Mutex<VecDeque<String>>,
+    seq_counter: AtomicU64,
+    high_water_mark: AtomicU64,
+}
+
+impl EventLog {
+    pub fn new() -> Self {
+        Self {
+            aer: Mutex::new(VecDeque::new()),
+            pciehp: Mutex::new(VecDeque::new()),
+            seq_counter: AtomicU64::new(0),
+            high_water_mark: AtomicU64::new(0),
+        }
+    }
+
+    /// Allocate the next monotonic sequence number and advance
+    /// `high_water_mark` if necessary. The counter starts at 1 so seq 0
+    /// is never a valid event (consumers can safely treat seq 0 as
+    /// "nothing seen yet").
+    fn next_seq(&self) -> u64 {
+        let seq = self.seq_counter.fetch_add(1, Ordering::Relaxed) + 1;
+        let mut current = self.high_water_mark.load(Ordering::Relaxed);
+        while seq > current {
+            match self.high_water_mark.compare_exchange_weak(
+                current,
+                seq,
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(actual) => current = actual,
+            }
+        }
+        seq
+    }
+
+    /// Returns the highest sequence number ever issued by this `EventLog`,
+    /// even if the corresponding event line has been evicted from the ring
+    /// buffer. Returns 0 when no events have been emitted yet.
+    pub fn latest_seq(&self) -> u64 {
+        self.high_water_mark.load(Ordering::Relaxed)
+    }
+
+    fn push(&self, log: &Mutex<VecDeque<String>>, line: String) {
+        let seq = self.next_seq();
+        let prefixed = format!("seq={seq}:{line}");
+        let mut queue = log.lock().unwrap_or_else(|e| e.into_inner());
+        if queue.len() >= MAX_EVENTS {
+            queue.pop_front();
+        }
+        queue.push_back(prefixed);
+    }
+
+    fn snapshot(log: &Mutex<VecDeque<String>>) -> String {
+        let queue = log.lock().unwrap_or_else(|e| e.into_inner());
+        queue.iter().cloned().collect()
+    }
+
+    pub fn push_aer(&self, line: String) {
+        self.push(&self.aer, line);
+    }
+
+    pub fn push_pciehp(&self, line: String) {
+        self.push(&self.pciehp, line);
+    }
+
+    pub fn aer_snapshot(&self) -> String {
+        Self::snapshot(&self.aer)
+    }
+
+    pub fn pciehp_snapshot(&self) -> String {
+        Self::snapshot(&self.pciehp)
+    }
+}
+
+pub fn spawn_event_poller(
+    pcie: Arc<Pcie>,
+    devices: Vec<PciAddress>,
+    log: Arc<EventLog>,
+) -> thread::JoinHandle<()> {
+    thread::Builder::new()
+        .name("pcid-events".to_string())
+        .spawn(move || {
+            info!(
+                "pcid-events: AER/pciehp poller started ({} devices, {} ms poll)",
+                devices.len(),
+                POLL_INTERVAL_MS
+            );
+            loop {
+                poll_aer(&pcie, &devices, &log);
+                poll_pciehp(&pcie, &devices, &log);
+                thread::sleep(Duration::from_millis(POLL_INTERVAL_MS));
+            }
+        })
+        .expect("spawn pcid event poller")
+}
+
+fn poll_aer(pcie: &Pcie, devices: &[PciAddress], log: &EventLog) {
+    for addr in devices {
+        let Some(cap_off) = find_extended_capability(pcie, *addr, AER_CAP_ID) else {
+            continue;
+        };
+
+        let uncorrectable = unsafe { pcie.read(*addr, cap_off + AER_UNCORRECTABLE_STATUS) };
+        if uncorrectable != 0 && uncorrectable != u32::MAX {
+            debug!("AER: uncorrectable error on {addr}: 0x{uncorrectable:08x}");
+            log.push_aer(format!(
+                "severity=NonFatal device={} raw=uncorrectable_status=0x{uncorrectable:08x}\n", scheme_name(addr)
+            ));
+            unsafe { pcie.write(*addr, cap_off + AER_UNCORRECTABLE_STATUS, uncorrectable) };
+        }
+
+        let correctable = unsafe { pcie.read(*addr, cap_off + AER_CORRECTABLE_STATUS) };
+        if correctable != 0 && correctable != u32::MAX {
+            debug!("AER: correctable error on {addr}: 0x{correctable:08x}");
+            log.push_aer(format!(
+                "severity=Correctable device={} raw=correctable_status=0x{correctable:08x}\n", scheme_name(addr)
+            ));
+            unsafe { pcie.write(*addr, cap_off + AER_CORRECTABLE_STATUS, correctable) };
+        }
+    }
+}
+
+fn poll_pciehp(pcie: &Pcie, devices: &[PciAddress], log: &EventLog) {
+    for addr in devices {
+        let Some(cap_off) = find_standard_capability(pcie, *addr, PCIE_CAP_ID) else {
+            continue;
+        };
+
+        let slot_capabilities = unsafe { pcie.read(*addr, cap_off + SLOT_CAPABILITIES) };
+        if slot_capabilities == u32::MAX || slot_capabilities & SLOT_CAP_HOT_PLUG_CAPABLE == 0 {
+            continue;
+        }
+
+        let status = unsafe { pcie.read(*addr, cap_off + SLOT_CONTROL_STATUS) } >> 16;
+        if status == 0 || status == 0xFFFF {
+            continue;
+        }
+
+        if status & SLTSTA_ATTENTION_BUTTON != 0 {
+            log.push_pciehp(format!("kind=attention_button device={}\n", scheme_name(addr)));
+        }
+        if status & SLTSTA_POWER_FAULT != 0 {
+            log.push_pciehp(format!("kind=power_fault device={}\n", scheme_name(addr)));
+        }
+        if status & SLTSTA_MRL_SENSOR != 0 {
+            log.push_pciehp(format!("kind=mrl_sensor device={}\n", scheme_name(addr)));
+        }
+        if status & SLTSTA_PRESENCE_DETECT != 0 {
+            log.push_pciehp(format!("kind=presence_detect device={}\n", scheme_name(addr)));
+        }
+        if status & SLTSTA_DLL_STATE != 0 {
+            log.push_pciehp(format!("kind=dll_state device={}\n", scheme_name(addr)));
+        }
+
+        // Write-1-to-clear the event bits in Slot Status (upper word of the
+        // Slot Control/Status dword), preserving the current Slot Control
+        // value in the lower word.
+        let control_status = unsafe { pcie.read(*addr, cap_off + SLOT_CONTROL_STATUS) };
+        let clear_value = (control_status & 0x0000_FFFF) | (status << 16);
+        unsafe { pcie.write(*addr, cap_off + SLOT_CONTROL_STATUS, clear_value) };
+    }
+}
+
+fn find_extended_capability(pcie: &Pcie, addr: PciAddress, cap_id: u16) -> Option<u16> {
+    let mut offset = 0x100u16;
+    for _ in 0..48 {
+        if offset == 0 {
+            return None;
+        }
+        let header = unsafe { pcie.read(addr, offset) };
+        if header == u32::MAX || header == 0 {
+            return None;
+        }
+        if (header & 0xFFFF) as u16 == cap_id {
+            return Some(offset);
+        }
+        offset = ((header >> 20) & 0xFFF) as u16;
+    }
+    None
+}
+
+fn find_standard_capability(pcie: &Pcie, addr: PciAddress, cap_id: u8) -> Option<u16> {
+    let command_status = unsafe { pcie.read(addr, 0x04) };
+    if (command_status >> 16) & 0x10 == 0 {
+        return None;
+    }
+    let mut ptr = (unsafe { pcie.read(addr, 0x34) } & 0xFC) as u16;
+    for _ in 0..48 {
+        if ptr == 0 {
+            return None;
+        }
+        let entry = unsafe { pcie.read(addr, ptr) };
+        if (entry & 0xFF) as u8 == cap_id {
+            return Some(ptr);
+        }
+        ptr = ((entry >> 8) & 0xFC) as u16;
+    }
+    None
+}

drivers/pcid/src/main.rs

@@ -3,6 +3,7 @@
 #![feature(non_exhaustive_omitted_patterns_lint)]
 
 use std::collections::BTreeMap;
+use std::sync::Arc;
 
 use log::{debug, info, trace, warn};
 use pci_types::capability::PciCapability;
@@ -18,6 +19,7 @@ use pcid_interface::{FullDeviceId, LegacyInterruptLine, PciBar, PciFunction, Pci
 
 mod cfg_access;
 mod driver_handler;
+mod events;
 mod scheme;
 
 pub struct Func {
@@ -284,11 +286,11 @@ fn daemon(daemon: daemon::Daemon) -> ! {
         common::file_level(),
     );
 
-    let pcie = Pcie::new();
+    let pcie = Arc::new(Pcie::new());
 
     info!("PCI SG-BS:DV.F VEND:DEVI CL.SC.IN.RV");
 
-    let mut scheme = scheme::PciScheme::new(pcie);
+    let mut scheme = scheme::PciScheme::new(Arc::clone(&pcie));
     let socket = redox_scheme::Socket::create().expect("failed to open pci scheme socket");
     let handler = Blocking::new(&socket, 16);
 
@@ -342,6 +344,11 @@ fn daemon(daemon: daemon::Daemon) -> ! {
     }
     debug!("Enumeration complete, now starting pci scheme");
 
+    let event_log = Arc::new(events::EventLog::new());
+    let devices: Vec<PciAddress> = scheme.tree.keys().copied().collect();
+    scheme.set_event_log(Arc::clone(&event_log));
+    let _events_thread = events::spawn_event_poller(Arc::clone(&pcie), devices, event_log);
+
     register_sync_scheme(&socket, "pci", &mut scheme)
         .expect("failed to register pci scheme to namespace");
 

drivers/pcid/src/scheme.rs

@@ -6,7 +6,7 @@ use redox_scheme::{CallerCtx, OpenResult};
 use scheme_utils::HandleMap;
 use syscall::dirent::{DirEntry, DirentBuf, DirentKind};
 use syscall::error::{Error, Result, EACCES, EBADF, EINVAL, EIO, EISDIR, ENOENT, ENOTDIR};
-use syscall::flag::{MODE_CHR, MODE_DIR, O_DIRECTORY, O_STAT};
+use syscall::flag::{MODE_CHR, MODE_DIR, MODE_FILE, O_DIRECTORY, O_STAT};
 use syscall::schemev2::NewFdFlags;
 use syscall::ENOLCK;
 
@@ -14,8 +14,9 @@ use crate::cfg_access::Pcie;
 
 pub struct PciScheme {
     handles: HandleMap<HandleWrapper>,
-    pub pcie: Pcie,
+    pub pcie: std::sync::Arc<Pcie>,
     pub tree: BTreeMap<PciAddress, crate::Func>,
+    events: std::sync::Arc<crate::events::EventLog>,
 }
 enum Handle {
     TopLevel { entries: Vec<String> },
@@ -23,6 +24,10 @@ enum Handle {
     Device,
     Config { addr: PciAddress },
     Channel { addr: PciAddress, st: ChannelState },
+    Aer,
+    Pciehp,
+    AerSeq,
+    PciehpSeq,
     SchemeRoot,
 }
 struct HandleWrapper {
@@ -33,7 +38,13 @@ impl Handle {
     fn is_file(&self) -> bool {
         matches!(
             self,
-            Self::Access | Self::Config { .. } | Self::Channel { .. }
+            Self::Access
+                | Self::Config { .. }
+                | Self::Channel { .. }
+                | Self::Aer
+                | Self::Pciehp
+                | Self::AerSeq
+                | Self::PciehpSeq
         )
     }
     fn is_dir(&self) -> bool {
@@ -95,16 +106,27 @@ impl SchemeSync for PciScheme {
         let path = path.trim_matches('/');
 
         let handle = if path.is_empty() {
-            Handle::TopLevel {
+            let mut entries: Vec<String> = self
-                entries: self
+                .tree
-                    .tree
+                .iter()
-                    .iter()
+                // FIXME remove replacement of : once the old scheme format is no longer supported.
-                    // FIXME remove replacement of : once the old scheme format is no longer supported.
+                .map(|(addr, _)| format!("{}", addr).replace(':', "--"))
-                    .map(|(addr, _)| format!("{}", addr).replace(':', "--"))
+                .collect::<Vec<_>>();
-                    .collect::<Vec<_>>(),
+            entries.push("aer".to_string());
-            }
+            entries.push("aer_seq".to_string());
+            entries.push("pciehp".to_string());
+            entries.push("pciehp_seq".to_string());
+            Handle::TopLevel { entries }
         } else if path == "access" {
             Handle::Access
+        } else if path == "aer" {
+            Handle::Aer
+        } else if path == "aer_seq" {
+            Handle::AerSeq
+        } else if path == "pciehp" {
+            Handle::Pciehp
+        } else if path == "pciehp_seq" {
+            Handle::PciehpSeq
         } else {
             let idx = path.find('/').unwrap_or(path.len());
             let (addr_str, after) = path.split_at(idx);
@@ -142,6 +164,10 @@ impl SchemeSync for PciScheme {
             Handle::Device => (DEVICE_CONTENTS.len(), MODE_DIR | 0o755),
             Handle::Config { .. } => (config_size, MODE_CHR | 0o600),
             Handle::Access | Handle::Channel { .. } => (0, MODE_CHR | 0o600),
+            Handle::Aer => (self.events.aer_snapshot().len(), MODE_FILE | 0o444),
+            Handle::Pciehp => (self.events.pciehp_snapshot().len(), MODE_FILE | 0o444),
+            Handle::AerSeq => (self.events.latest_seq().to_string().len(), MODE_FILE | 0o444),
+            Handle::PciehpSeq => (self.events.latest_seq().to_string().len(), MODE_FILE | 0o444),
             Handle::SchemeRoot => return Err(Error::new(EBADF)),
         };
         stat.st_size = len as u64;
@@ -186,6 +212,39 @@ impl SchemeSync for PciScheme {
                 addr: _,
                 ref mut st,
             } => Self::read_channel(st, buf),
+            Handle::Aer => {
+                let data = self.events.aer_snapshot();
+                let bytes = data.as_bytes();
+                let offset = usize::try_from(_offset).map_err(|_| Error::new(EINVAL))?;
+                if offset >= bytes.len() {
+                    return Ok(0);
+                }
+                let count = (bytes.len() - offset).min(buf.len());
+                buf[..count].copy_from_slice(&bytes[offset..offset + count]);
+                Ok(count)
+            }
+            Handle::Pciehp => {
+                let data = self.events.pciehp_snapshot();
+                let bytes = data.as_bytes();
+                let offset = usize::try_from(_offset).map_err(|_| Error::new(EINVAL))?;
+                if offset >= bytes.len() {
+                    return Ok(0);
+                }
+                let count = (bytes.len() - offset).min(buf.len());
+                buf[..count].copy_from_slice(&bytes[offset..offset + count]);
+                Ok(count)
+            }
+            Handle::AerSeq | Handle::PciehpSeq => {
+                let data = self.events.latest_seq().to_string();
+                let bytes = data.as_bytes();
+                let offset = usize::try_from(_offset).map_err(|_| Error::new(EINVAL))?;
+                if offset >= bytes.len() {
+                    return Ok(0);
+                }
+                let count = (bytes.len() - offset).min(buf.len());
+                buf[..count].copy_from_slice(&bytes[offset..offset + count]);
+                Ok(count)
+            }
             Handle::SchemeRoot => Err(Error::new(EBADF)),
             _ => Err(Error::new(EBADF)),
         }
@@ -219,7 +278,13 @@ impl SchemeSync for PciScheme {
                 return Ok(buf);
             }
             Handle::Device => DEVICE_CONTENTS,
-            Handle::Access | Handle::Config { .. } | Handle::Channel { .. } => {
+            Handle::Access
+            | Handle::Config { .. }
+            | Handle::Channel { .. }
+            | Handle::Aer
+            | Handle::Pciehp
+            | Handle::AerSeq
+            | Handle::PciehpSeq => {
                 return Err(Error::new(ENOTDIR));
             }
             Handle::SchemeRoot => return Err(Error::new(EBADF)),
@@ -383,11 +448,16 @@ impl PciScheme {
         if self.pcie.has_ecam() { 4096 } else { 256 }
     }
 
-    pub fn new(pcie: Pcie) -> Self {
+    pub fn set_event_log(&mut self, events: std::sync::Arc<crate::events::EventLog>) {
+        self.events = events;
+    }
+
+    pub fn new(pcie: std::sync::Arc<Pcie>) -> Self {
         Self {
             handles: HandleMap::new(),
             pcie,
             tree: BTreeMap::new(),
+            events: std::sync::Arc::new(crate::events::EventLog::new()),
         }
     }
     fn parse_after_pci_addr(&mut self, addr: PciAddress, after: &str) -> Result<Handle> {

init/src/main.rs

@@ -130,16 +130,6 @@ fn main() {
     scheduler
         .schedule_start_and_report_errors(&mut unit_store, UnitId("00_logd.service".to_owned()));
     scheduler.step(&mut unit_store, &mut init_config);
-    // NOTE: init/daemon stdio is deliberately kept on the kernel console
-    // rather than being redirected into /scheme/log. Redirecting made the boot
-    // "go dark" after logd on text/recovery ISOs (bare/mini): the graphical
-    // bootlog (fbbootlogd, VT1) does not reliably surface /scheme/log to the
-    // framebuffer, and logd's console mirror is racy, so the console showed
-    // nothing between logd and the login VT — indistinguishable from a hang.
-    // Keeping stdio on the console makes the full boot observable on
-    // `-serial` and the framebuffer. (A quiet/splash mode for the graphical
-    // full ISO can re-enable the redirect once fbbootlogd rendering is fixed.)
-    let _ = switch_stdio;
 
     let runtime_target = UnitId("00_runtime.target".to_owned());
     scheduler.schedule_start_and_report_errors(&mut unit_store, runtime_target.clone());
@@ -155,6 +145,29 @@ fn main() {
         Path::new("/usr"),
         Path::new("/etc"),
     );
+
+    // Pin init/daemon stdio to the dedicated boot-log VT (fbcon/1) before the
+    // /usr units run. Rationale: stdio otherwise stays on the kernel console
+    // (/scheme/debug), which fbcond renders onto the *active* VT. Once the
+    // `29_activate_console` unit switches the active VT to 2 for the getty
+    // login, any daemon that logs afterwards — late async driver attach
+    // (i2c-hidd, evdevd, dw-acpi-i2cd) or smolnetd's bounded NIC wait — paints
+    // its output on top of the login banner/prompt, so the prompt is never the
+    // last thing on screen. Pinning daemon output to VT1 (a fixed VT, not the
+    // active one) keeps the boot log fully visible — VT1 is the active console
+    // during boot, and fbcond still mirrors it to the serial line — while
+    // leaving VT2 clean for a login prompt that lands last.
+    //
+    // This is deliberately NOT the old `/scheme/log` redirect that made boot
+    // "go dark": that depended on fbbootlogd surfacing /scheme/log to the
+    // framebuffer (racy/unreliable). A real VT renders directly, so full-boot
+    // observability is preserved. Best-effort: if fbcon/1 cannot be opened
+    // (e.g. fbcond absent on a headless/serial-only boot), stdio stays on the
+    // kernel console and the boot remains observable exactly as before.
+    if let Err(err) = switch_stdio("/scheme/fbcon/1") {
+        eprintln!("INIT: keeping stdio on kernel console (fbcon/1 unavailable: {err})");
+    }
+
     {
         // FIXME introduce multi-user.target unit and replace the config dir iteration
         // scheduler.schedule_start_and_report_errors(&mut unit_store, UnitId("multi-user.target".to_owned()));

netstack/src/main.rs

@@ -35,22 +35,58 @@ fn get_all_network_adapters() -> Result<Vec<String>> {
         adapters.push(scheme);
     }
 
-    if adapters.is_empty() {
-        // No NIC present is a normal state on a diskless/headless/bare machine,
-        // or simply before an unsupported NIC's driver attaches. Report it as
-        // info and let the caller exit cleanly instead of emitting a spurious
-        // ERROR + non-zero exit on every boot of a NIC-less system.
-        info!("no network adapter found; network stack idle");
-        return Ok(adapters);
-    }
-    info!("Found {} network adapter(s): {:?}", adapters.len(), adapters);
     Ok(adapters)
 }
 
+/// Wait for at least one NIC driver to register its `network.*` scheme.
+///
+/// driver-manager attaches drivers ASYNCHRONOUSLY, so on a real boot the NIC's
+/// `network.*` scheme does not exist yet when smolnetd starts (10_smolnetd only
+/// requires_weak that driver-manager has *started*). The previous code scanned
+/// `/scheme` exactly once and, finding nothing, exited — losing the race with
+/// virtio-netd and leaving `/scheme/netcfg` absent, so dhcpd/netctl failed with
+/// "Can't open /scheme/netcfg/ifaces/eth0/mac". Poll for a bounded window
+/// instead. smolnetd is oneshot_async, so this wait never blocks init/boot; a
+/// genuinely NIC-less machine simply gets an empty list after the window and
+/// exits cleanly.
+fn wait_for_network_adapters() -> Vec<String> {
+    use std::time::{Duration, Instant};
+    let deadline = Instant::now() + Duration::from_secs(20);
+    let mut logged_wait = false;
+    loop {
+        match get_all_network_adapters() {
+            Ok(adapters) if !adapters.is_empty() => {
+                info!("Found {} network adapter(s): {:?}", adapters.len(), adapters);
+                return adapters;
+            }
+            _ => {}
+        }
+        if Instant::now() >= deadline {
+            return Vec::new();
+        }
+        if !logged_wait {
+            info!("no network adapter yet; waiting for a NIC driver to attach...");
+            logged_wait = true;
+        }
+        std::thread::sleep(Duration::from_millis(250));
+    }
+}
+
 fn run(daemon: daemon::Daemon) -> Result<()> {
-    let adapters = get_all_network_adapters()?;
+    let adapters = wait_for_network_adapters();
     if adapters.is_empty() {
-        // Nothing to service; exit cleanly (see get_all_network_adapters).
+        // No `network.*` scheme registered within the wait window. Two cases:
+        //   (a) genuinely NIC-less machine  -> expected, harmless;
+        //   (b) a NIC IS present but its driver never registered `network.*`
+        //       -> a driver bug or a STALE/broken NIC driver binary that no
+        //          longer speaks the current scheme protocol.
+        // If this box has a NIC, this is case (b): rebuild the net driver
+        // (virtio-netd / e1000d / rtl8168d). Exit cleanly either way (smolnetd
+        // is oneshot_async, so waiting never blocked boot/login).
+        warn!(
+            "no network.* scheme after waiting; network stack idle. If a NIC is \
+             present, its driver failed to register it (driver bug or stale driver binary)."
+        );
         return Ok(());
     }
     trace!("found {} network adapter(s)", adapters.len());