Re: [PATCH net-next 0/3 V1] rtnetlink: enable IFLA_IF_NETNSID for RTM_{DEL,SET}LINK

[Nicolas Dichtel <nicolas.dichtel@6wind.com>]

Le 26/01/2018 à 09:36, Jiri Benc a écrit :
> On Fri, 26 Jan 2018 00:34:51 +0100, Nicolas Dichtel wrote:
>> Why meaningful? The user knows that the answer is like if if was done in another
>> netns. It enables to have only one netlink socket instead of one per netns. But
>> the code using it will be the same.  
> 
> Because you can't use it to query the linked interface. You can't even
> use it as an opaque value to track interfaces (netnsid+ifindex) because
> netnsids are not unique across net name spaces. You can easily have two
> interfaces that have all the ifindex, ifname, netnsid (and basically
> everything else) identical but being completely different interfaces.
Yes, the user have to map those info correctly. And this complexifies the (user)
code a lot.

> That's really not helpful.
> 
>> I fear that with your approach, it will results to a lot of complexity in the
>> kernel.  
> 
> The complexity is (at least partly) already there. It's an inevitable
> result of the design decision to have relative identifiers.
Yes, you're right. My approach moves the complexity to the user, which make this
feature hard to use.

> 
> I agree that we should think about how to make this easy to implement.
> I like your idea of doing this somehow generically. Perhaps it's
> possible to do while keeping the netnsids valid in the caller's netns?
Yes. I agree that it will be a lot easier to use if the conversion is done in
the kernel. And having a generic mechanism will also help a lot to use it.

> 
>> What is really missing for me, is a way to get a fd from an nsid. The user
>> should be able to call RTM_GETNSID with an fd and a nsid and the kernel performs
>> the needed operations so that the fd points to the corresponding netns.  
> 
> That's what I was missing, too. I even looked into implementing it. But
> opening a fd on behalf of the process and returning it over netlink is a
> wrong thing to do. Netlink messages can get lost. Then you have a fd
> leak you can do nothing about.
Yes, I also looked at this ;-)

> 
> Given that we have netnsids used for so much stuff already (like
> NETLINK_LISTEN_ALL_NSID) you need to track them anyway. And if you need
> to track them, why bother with another identifier? It would be better
> if netnsid can be used universally for anything. Then there will be no
> need for the conversion.
I like this idea a lot. So the missing part is a setns() using the nsid ;-)


Regards,
Nicolas