weave reset fails with "Cannot find device" on kernel 4.2+

[rade]

Extracted from #1356...

@MaximeHeckel has seen weave --local reset fail with Cannot find device "vethwepl14191@if322". This can happen when a container dies during reset, which is what #1356 is about, but it could also happen if somehow our interface name extraction is broken - that @if322 in the interface name looks weird.

@MaximeHeckel as @bboreham said, please get us the output of ip link show from the host. Also what OS and kernel is that host running? And what version of iptables (ip -V)?

[MaximeHeckel]

@rade The user might have reset his host, so not sure I can provide that to you. I'm going to check with him.

[rade]

@MaximeHeckel I am going to close this; if you ever encounter such strange interface names in the wild again, please re-open with some extra info.

[dpw]

It looks to me that an interface name like that will occur in the ip link output for anyone using kernel 4.1 or later:

torvalds/linux@a45253b

See also:

iproute2/iproute2@ccdcbf35

(That iproute2 change appeared in v4.0.0, or a snapshot after 20150217)

[rade]

iproute2/iproute2@ccdcbf35

Hmm. According to that, it was outputing @ before too. What am I missing?

In the weave script we extract the name and then call ip link del on it. Will s.t. like ip link del vethwepl14191@if322 break?

[dpw]

Hmm. According to that, it was outputing @ before too. What am I missing?

The kernel change is the crucial one, because it causes the IFLA_LINK attribute to be reported for veth netdevs. ip link could output '@' before, going back to when it migrated to git, but until that kernel change it wouldn't do so for veths.

That iproute2 commit is interesting because it introduces @ifNNN (to denote a unique name for a netdev in a different namespace). So it ties down the relevant version of iproute2 a bit.

(It also shows that there are moves to make network namespaces and associated plumbing a bit more transparent. Which is nice - it might have made debugging the dangling netns problems a bit easier. It's just a shame the improvements are not yet widely deployed..)

In the weave script we extract the name and then call ip link del on it. Will s.t. like ip link del vethwepl14191@if322 break?

Yes, it will break. The part after '@' is informational. It's not part of the netdev name in any sense, and no ip subcommands accept it. We'll need to strip it off.

[rade]

@dpw thanks for tracking down the cause.