Watch the network over netlink instead of polling operstate #122

New issue

Open

opened 2026-08-22 23:01:12 -04:00 by mysticalsoap · 0 comments

mysticalsoap commented

2026-08-22 23:01:12 -04:00

Owner

NetObserver (#111) is a straight port of NetMon's mechanism: poll /sys/class/net/*/operstate every 2 seconds and confirm a default gateway before reporting up. It works, but it is polling -- up to 2s of latency on a transition, plus a wakeup every 2s forever.

The kernel already broadcasts these events: an RTMGRP_LINK/RTMGRP_IPV4_ROUTE/RTMGRP_IPV6_ROUTE netlink socket would make the observer event-driven -- react on the route change itself, no sleep loop, no missed 2s window. Route messages would also replace the ip -j route subprocess calls in observer.default_routes.

Split out of #111 deliberately (poll-first port keeps that diff reviewable). Design points for whoever picks it up:

Dependency choice: raw socket.AF_NETLINK (stdlib, but hand-parsing rtattr) vs pyroute2 (clean, but a new hard dependency across distros -- check what target distros package).
Keep the "up means usable" rule: a link-up event without a default route must stay "down" until a route event backs it (today that is the gateway check + retry-next-poll).
The kernel does not drop subscribers the way a perf ring buffer drops events, but the socket buffer can still overflow -- on ENOBUFS, re-derive state from a fresh default_routes() read rather than trusting the event stream (an unbounded watcher that silently loses events is the upstream bug this fork exists to avoid).

NetObserver (#111) is a straight port of NetMon's mechanism: poll `/sys/class/net/*/operstate` every 2 seconds and confirm a default gateway before reporting up. It works, but it is polling -- up to 2s of latency on a transition, plus a wakeup every 2s forever. The kernel already broadcasts these events: an `RTMGRP_LINK`/`RTMGRP_IPV4_ROUTE`/`RTMGRP_IPV6_ROUTE` netlink socket would make the observer event-driven -- react on the route change itself, no sleep loop, no missed 2s window. Route messages would also replace the `ip -j route` subprocess calls in `observer.default_routes`. Split out of #111 deliberately (poll-first port keeps that diff reviewable). Design points for whoever picks it up: - Dependency choice: raw `socket.AF_NETLINK` (stdlib, but hand-parsing rtattr) vs `pyroute2` (clean, but a new hard dependency across distros -- check what target distros package). - Keep the "up means usable" rule: a link-up event without a default route must stay "down" until a route event backs it (today that is the gateway check + retry-next-poll). - The kernel does not drop subscribers the way a perf ring buffer drops events, but the socket buffer can still overflow -- on `ENOBUFS`, re-derive state from a fresh `default_routes()` read rather than trusting the event stream (an unbounded watcher that silently loses events is the upstream bug this fork exists to avoid).

mysticalsoap added the

architecture

label

2026-08-22 23:01:12 -04:00