Kevin Veen-Birkenbach 499e61346d fix(net): detect interfaces without following /sys symlinks ...

/sys/class/net/<if> is always a symlink into /sys/devices; in nested
containers (sysbox runtime) that target is not visible, so the link is
dead. iface_exists() used Path.exists(), which follows the link and
therefore tested the visibility of the target instead of the existence of
the interface. detect_egress_iface() discarded every interface it had
correctly read from the routing table, so automtu aborted with "Could not
detect egress interface" and rc=2 even though `ip -4 route show default`
and `ip link show dev eth0` both worked. The documented escape hatch
--egress-if was equally dead, because core.py validates it through the
same call.

Chosen fix: os.path.lexists() rather than probing netlink for existence.
It asks the right question -- "is there an entry named <if>" -- and costs
no subprocess on a healthy host. Netlink (`ip link show`) is only the
fallback for when /sys/class/net itself is unavailable, so sysfs stays the
preferred path everywhere.

read_iface_mtu() gains the same cascade: sysfs first, MTU parsed from
`ip link show dev <if>` when the sysfs path is unreadable, RuntimeError
naming both sources when neither answers - no silent default. core.py
turns that into an error line plus rc=3 instead of a traceback.

list_ifaces() had the same defect (is_dir() on a dead symlink) and would
have left Docker bridge detection blind in the same environments.

Regression tests build a /sys replacement whose class/net/eth0 points at a
missing target and mock the ip command: they fail if lexists becomes
exists again, if the MTU fallback is removed, or if the symlink filter is
dropped.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

2026-09-08 18:35:25 +02:00

8.4 KiB

Raw Blame History

View Raw