Re: [PATCH net] net: try harder to not reuse ifindex when moving interfaces

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

Le 21/10/2015 19:12, Hannes Frederic Sowa a écrit :
> Hello,
>
> On Wed, Oct 21, 2015, at 17:56, David Miller wrote:
>> From: Jiri Benc <jbenc@redhat.com>
>> Date: Wed, 21 Oct 2015 17:25:02 +0200
>>
>>> On Wed, 21 Oct 2015 08:32:14 -0700 (PDT), David Miller wrote:
>>>> As you say the apps are broken, so file a bug and have them fixed.
>>>>
>>>> The assumption is clearly invalid, so apps cannot make such an
>>>> assumption.
>>>
>>> Does it mean you would be okay with a patch that always allocates and
>>> assigns a new ifindex in the target netns when interface is moved
>>> between name spaces?
>>
>> I think you're misunderstanding me if you're still recommending
>> kernel changes.
>>
>> I'm plainly saying to remove the assumption in the apps.
>>
>> If you don't show me exactly how some kernel change can lead to
>> the apps implementing things properly, without the invalid
>> assumptions, then I can only assume you didn't hear what I
>> said.
>
> I think the reason why ifindexes exists as ints is that we want to have
> lightweight way to refer to interfaces without taking references or
> timestamps or generation ids which completely remove the possibility for
> races. But the racy nature in ifindexes is something we actually want,
> otherwise a user space program acquiring an ifindex would need to get a
> reference on the device and either during socket close or program
> termination release it, that would be very costly.
>
> This patch minimizes the race quite a lot, from something we could
> actually see in everydays container creation to probably something only
> some users will expire with depleting the ifindex pool or playing around
> with CRIU.
>
> We could come up with more heavy machinery to close the race further for
> CRIU by keeping track of "poisoned" ifindexes, which would need a
> hashmap which could become pretty big and we could recycle when ifindex
> wraps around, but this seems too heavy weight to me.
>
> I am in favor of a solution to minimize this race in the kernel even
> though we cannot ever close it completely.
I probably miss something, but if the app listens netlink, I don't see how such
app may have a race window.

With the proposed scenario:
1. create netns 'new_netns'
2. in root netns, move the interface with ifindex 2 to new_netns
3. in new_netns, delete the interface with ifindex 2
4. in new_netns, create an interface - it will get ifindex 2

Operation 2 and 4 are done by dev_change_net_namespace() under rtnl_lock().
RTM_DELLINK(root netns) and RTM_NEWLINK(new_netns) are sent by this function.
It means that operation 3 has been done before and that RTM_DELLINK(new_netns)
has been sent before.

Regards,
Nicolas