Watch
0
0
Fork
You've already forked hyperhive
0

egress-pin: a sweep says why a port stays unbound, which veths are off the bridge and which machines it could not read

A port nobody owns, a vb- veth that never joined the bridge and a machine whose
state file or namespace can't be read all left the bridge sweep silent, so a
container stayed unbound for 11 hours with nothing in the journal.

A timer run now prints one line for each, after the finding has stood for 60 s
and once until it clears, kept in <state file>.diag. Nothing it binds changes.
This commit is contained in:
maia 2026-11-19 19:54:26 +01:00 • committed by mara
commit 94dd24fa99

View file

@ -911,6 +911,16 @@ source check lives in that table and the host-side filtering stays in
is what the table is for: without it the egress log can't tell agents apart.
A binding that failed shows as `hive-egress-pin` errors in the journal and
`IN=` lines for that port.
- **A port that stays unbound.** The timer's run of `hive-egress-pin` says why
a port is still unbound after a minute (a port is unbound for a moment at
every container start, so it waits). `journalctl -u hive-egress-pin` then
shows `<port> (ifindex N, peer M) has no owner; machines seen: <name>=<index>/<peer>, …`
when no container's link pairs with the port, or `… stays unbound: <reason>`
when one does. A `vb-` veth that isn't on the bridge gets a line the same
way (`<veth> is not on hive-br0`), and so does a registered machine the run
couldn't read (`machine <name> is skipped: <reason>`). Each line comes once,
and again only after it has cleared (the run keeps
`/run/hive-egress-pins.diag`). None of them changes which ports get a binding.
- **What it doesn't cover.** A container can still send from any IPv6
link-local address and answer DHCP as a server on its own leased address. A
DHCP client message to the broadcast or bridge address carries any IPv4 source