[PATCH V4 3/8] namespaces: expose ns instance serial numbers in proc

Andy Lutomirski luto at amacapital.net
Mon Aug 25 16:50:11 UTC 2014


On Mon, Aug 25, 2014 at 9:41 AM, Nicolas Dichtel
<nicolas.dichtel at 6wind.com> wrote:
> Le 25/08/2014 18:13, Andy Lutomirski a écrit :
>
>> On Mon, Aug 25, 2014 at 8:43 AM, Nicolas Dichtel
>> <nicolas.dichtel at 6wind.com> wrote:
>>>
>>> Le 25/08/2014 16:04, Andy Lutomirski a écrit :
>>>
>>>> On Aug 25, 2014 6:30 AM, "Nicolas Dichtel" <nicolas.dichtel at 6wind.com>
>>>> wrote:
>>>>>>
>>>>>>
>>>>>> CRIU wants to save the complete state of a namespace and then restore
>>>>>> it.  For that to work, any information exposed to things in the
>>>>>> namespace *cannot* be globally unique or unique per boot, since CRIU
>>>>>> needs to arrange for that information to match whatever it was when
>>>>>> CRIU saved it.
>>>>>
>>>>>
>>>>>
>>>>> How are ifindex of network devices managed? These ifindexes are unique
>>>>> per boot,
>>>>> thus can change depending on the order in which netdev are created.
>>>>> These ifindexes are unique per boot and exposed to userspace ...
>>>>>
>>>>
>>>> This does not appear to be true.
>>>>
>>>> $ sudo unshare --net
>>>> # ip link add veth0 type veth peer name veth1
>>>> # ip link
>>>> 1: lo: <LOOPBACK> mtu 65536 qdisc noop state DOWN mode DEFAULT group
>>>> default
>>>>       link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
>>>> 2: veth1: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode
>>>> DEFAULT group default qlen 1000
>>>>       link/ether 06:0d:59:c7:a6:a8 brd ff:ff:ff:ff:ff:ff
>>>> 3: veth0: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode
>>>> DEFAULT group default qlen 1000
>>>>       link/ether b2:5c:8b:f2:12:28 brd ff:ff:ff:ff:ff:ff
>>>> # logout
>>>> $ ip link
>>>> 1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN
>>>>       link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
>>>> 3: em1: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc pfifo_fast
>>>> state DOWN qlen 1000
>>>>
>>> I've probably misunderstood what you're trying to say. ifindexes are
>>> unique
>>> per
>>> boot and per netns.
>>
>>
>> I think we both misunderstood each other.  The ifindexes are unique
>> *per netns*, which means that, if you're unprivileged in a netns,
>> global information doesn't leak to you.  I think this is good.
>
> Ok, I agree. I think audit daemons are always running under privileged
> users.
>
>
>>
>>>>
>>>> Let me try again, with emphasis in the right place.
>>>>
>>>> I think that *code running in a namespace* has no business even
>>>> knowing a unique identity of *that namespace* from the perspective of
>>>> the host.
>>>>
>>>> In your example, if there's a veth device between netns A and netns B,
>>>> then code *in netns A* has no business knowing the identity of its
>>>> veth peer if its peer (B) is a sibling or ancestor.  It also IMO has
>>>> no business knowing the identity of its own netns (A) other than as
>>>> "my netns".
>>>
>>>
>>> I do not agree (see the example below).
>>>
>>>
>>>>
>>>> If A and B are siblings, then their parent needs to know where that
>>>> veth device goes, but I think this is already the case to a sufficient
>>>> extent today.
>>>
>>>
>>> I'm not aware of a hierarchy between netns. A daemon should be able to
>>> got the full network configuration, even if it's started when this
>>> configuration
>>> is already applied, ie even if it doesn't know what happen before it
>>> starts.
>>>
>>
>> I don't know exactly which namespaces have an explicit hierarchy, but
>> there is certainly a hierarchy of *user* namespaces, and network
>> namespaces live in user namespaces, so they at least have somewhat of
>> a hierarchy.
>>
>>>
>>>>
>>>> I feel like this discussion is falling into a common trap of new API
>>>> discussions.  Can one of you who wants this API please articulate,
>>>> with a reasonably precise example, what it is that you want to do, why
>>>> you can't easily do it already, and how this API helps?  I currently
>>>> understand how the API creates problems, but I don't understand how it
>>>> solves any problems, and I will NAK it (and I suspect that Eric will,
>>>> too, which is pretty much fatal) unless that changes.
>>>
>>>
>>> What I'm trying to solve is to have full info in netlink messages sent by
>>> the
>>> kernel, thus beeing able to identify a peer netns (and this is close from
>>> what
>>> audit guys are trying to have). Theorically, messages sent by the kernel
>>> can
>>> be
>>> reused as is to have the same configuration. This is not the case with
>>> x-netns
>>> devices. Here is an example, with ip tunnels:
>>>
>>> $ ip netns add 1
>>> $ ip link add ipip1 type ipip remote 10.16.0.121 local 10.16.0.249 dev
>>> eth0
>>> $ ip -d link ls ipip1
>>> 8: ipip1 at eth0: <POINTOPOINT,NOARP> mtu 1480 qdisc noop state DOWN mode
>>> DEFAULT group default
>>>      link/ipip 10.16.0.249 peer 10.16.0.121 promiscuity 0
>>>      ipip remote 10.16.0.121 local 10.16.0.249 dev eth0 ttl inherit
>>> pmtudisc
>>> $ ip link set ipip1 netns 1
>>> $ ip netns exec 1 ip -d link ls ipip1
>>> 8: ipip1 at tunl0: <POINTOPOINT,NOARP,M-DOWN> mtu 1480 qdisc noop state DOWN
>>> mode DEFAULT group default
>>>      link/ipip 10.16.0.249 peer 10.16.0.121 promiscuity 0
>>>      ipip remote 10.16.0.121 local 10.16.0.249 dev tunl0 ttl inherit
>>> pmtudisc
>>>
>>> Now informations got with 'ip link' are wrong and incomplete:
>>>   - the link dev is now tunl0 instead of eth0, because we only got an
>>> ifindex
>>>     from the kernel without any netns informations.
>>>   - the encapsulation addresses are not part of this netns but the user
>>> doesn't
>>>     known that (still because netns info is missing). These IPv4
>>> addresses
>>> may
>>>     exist into this netns.
>>>   - it's not possible to create the same netdevice with these infos.
>>>
>>
>> Aha.  That's a genuine problem.
>>
>> Perhaps we need a concept of which netnses should be able to see each
>> other.
>
> Yes, I agree. This is not required for all netns, only a subset of netns
> should
>
> be able to see each other.
>
>>
>> I think I would be okay with a somewhat different outcome from your
>> example:
>>
>> $ ip netns exec 1 ip -d link ls ipip1
>> 8: ipip1@[unknown device in another namespace]:
>> <POINTOPOINT,NOARP,M-DOWN> mtu 1480 qdisc noop state DOWN
>>
>> I think this outcome is mandatory if netns 1 lives in a subsidiary
>> user namespace.
>
> Yes.
>
>
>>
>> Certainly, if you do the 'ip link' in the original namespace, I agree
>> that this should work.
>
> And yes :)
>
> I will update my previous proposal
> (http://thread.gmane.org/gmane.linux.network/315933/focus=321753)
> to allow to get an id for a peer netns only when the user namespace is the
> same.
>

I think it should work if the peer userns is the same or a descendent.
I also wonder whether the peer's ifindex should be suppressed if peer
userns is not the same or a descendent.

Now you just have to get Eric to be happy with the id allocation. :)
This may be nontrivial.

>
>>
>> For most namespace types, this all works transparently, since
>> everything has an real identity all the way up the hierarchy.  Network
>> namespaces are different.
>>
>> I don't think that exposing serial numbers in /proc is a good
>> solution, both for the reasons already described and because I don't
>> think that iproute2 should need to muck around with /proc to function
>
> A netlink API is probably enough. But it will help only for the network
> problem, not for audit. I was hoping to find a common solution.

I still don't understand why audit needs anything beyond the audit
part of this patch set.  I have no problem with audit seeing that
migrated/restored namespaces are really brand-new namespaces, as long
as the code in those namespaces isn't exposed to it.

>
>
>> correctly.  Eric, any clever ideas here?  Do we need fancier netlink
>> messages for this?
>>
>> --Andy
>>
>



-- 
Andy Lutomirski
AMA Capital Management, LLC


More information about the Containers mailing list