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

[Nicolas Dichtel <nicolas.dichtel-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>]

Le 25/08/2014 18:13, Andy Lutomirski a écrit :
> On Mon, Aug 25, 2014 at 8:43 AM, Nicolas Dichtel
> <nicolas.dichtel-pdR9zngts4EAvxtiuMwx3w@public.gmane.org> wrote:
>> Le 25/08/2014 16:04, Andy Lutomirski a écrit :
>>
>>> On Aug 25, 2014 6:30 AM, "Nicolas Dichtel" <nicolas.dichtel@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@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@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.

>
> 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.

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