Re: [RFC PATCH net-next v2 0/5] netns: allow to identify peer netns

[Andy Lutomirski <luto@amacapital.net>]

On Thu, Oct 2, 2014 at 12:20 PM, Eric W. Biederman
<ebiederm@xmission.com> wrote:
> Nicolas Dichtel <nicolas.dichtel@6wind.com> writes:
>
>> Le 29/09/2014 20:43, Eric W. Biederman a écrit :
>>> Nicolas Dichtel <nicolas.dichtel@6wind.com> writes:
>>>
>>>> Le 26/09/2014 20:57, Eric W. Biederman a écrit :
>>>>> Andy Lutomirski <luto@amacapital.net> writes:
>>>>>
>>>>>> On Fri, Sep 26, 2014 at 11:10 AM, Eric W. Biederman
>>>>>> <ebiederm@xmission.com> wrote:
>>>>>>> I see two ways to go with this.
>>>>>>>
>>>>>>> - A per network namespace table to that you can store ids for ``peer''
>>>>>>>     network namespaces.  The table would need to be populated manually by
>>>>>>>     the likes of ip netns add.
>>>>>>>
>>>>>>>     That flips the order of assignment and makes this idea solid.
>>>> I have a preference for this solution, because it allows to have a full
>>>> broadcast messages. When you have a lot of network interfaces (> 10k),
>>>> it saves a lot of time to avoid another request to get all informations.
>>>
>>> My practical question is how often does it happen that we care?
>> In fact, I don't think that scenarii with a lot of netns have a full mesh of
>> x-netns interfaces. It will be more one "link" netns with the physical
>> interface and all other with one interface with the link part in this "link"
>> netns. Hence, only one nsid is needing in each netns.
>
> I will buy that a full mesh is unlikely.
>
> For people doing simulations anything physical has a limited number of
> links.
>
> For people wanting all to all connectivity setting up an internal
> macvlan (or the equivalent) is likely much simpler and more efficient
> that a full mesh.
>
> So the question in my mind is how do we create these identifiers at need
> (when we create the cross network namespace links) instead of at network
> namespace creation time.  I don't see an answer to that in your patches,
> and perhaps it obvious.
>

I wonder whether part of the problem is that we're thinking about
scoping wrong.  What if we made the hierarchy more explicit?

For example, we could give each netns an admin-assigned identifier
(e.g. a 64-bit number, maybe required to be unique, maybe not)
relative to its containing userns.  Then we could come up with a way
to identify user namespaces (i.e. inode number relative to containing
user ns, if that's well-defined).

>From user code's perspective, netnses that are in the requester's
userns or its descendents are identified by a path through a (possibly
zero-length) sequence of userns ids followed by a netns id.  Netnses
outside the requester's userns hierarchy cannot be named at all.

Would this make sense?  It should keep the asymptotic complexity of
everything under control and, for users of very large numbers of
network namespaces with complex routing, it doesn't require a
correspondingly large number of fds. It would have the added benefit
of allowing the same scheme to be used for all the other namespace
types, although it could be a bit odd for pid namespaces, which really
do have their own hierarchy.

--Andy