Re: [PATCH net-next 1/1] rtnetlink: request RTM_GETLINK by pid or fd

[Jiri Benc <jbenc@redhat.com>]

On Tue, 23 Jan 2018 11:26:58 +0100, Wolfgang Bumiller wrote:
> Even if you know the netnsid, do the mentioned watches work for
> nested/child namespaces if eg. a container creates new namespace before
> and/or after the watch was established and moves interfaces to these
> child namespaces, would you just see them disappear, or can you keep
> track of them later on as well?

What do you mean by "nested namespaces"? There's no such thing for net
name spaces.

As for missing API to get netnsid of the netns the interface is moved
to, see my previous emails in this thread. This needs to be added.

> Even if that works, from what the documentation tells me netlink is an
> unreliable protocol, so if my watcher's socket buffer is full, wouldn't
> I be losing important tracking information?

Sure. But that's fundamentally unfixable independently on netlink, the
kernel needs to take an action if a program is not reading its
messages. Either some messages get dropped or the program is killed or
infinite amount of memory is consumed. This has nothing to do with uAPI
design.

> I think one possible solution to tracking interfaces would be to have a
> unique identifier that never changes (even if it's just a simple
> uint64_t incremented whenever an interface is created). But since
> they're not local to the current namespace that may require a lot of
> extra permission checks (but I'm just speculating here...).

You'll get a hard NACK from CRIU folks if you try to propose this.

> In any case, IFLA_NET_NS_FD/PID are already there and I had been
> wondering previously why they couldn't be used with RTM_GETLINK, it
> would just make sense.

Those predate netnsids and we can't get rid of them now, since they're
part of uAPI. But we can (and should) make sure we don't add more of
those.

 Jiri