Message-ID: <53FB59A3.5030804@6wind.com>
Date:	Mon, 25 Aug 2014 17:43:31 +0200
From:	Nicolas Dichtel <nicolas.dichtel@...nd.com>
To:	Andy Lutomirski <luto@...capital.net>
CC:	Linux Containers <containers@...ts.linux-foundation.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"Serge E. Hallyn" <serge@...lyn.com>,
	"Eric W. Biederman" <ebiederm@...ssion.com>,
	linux-audit@...hat.com, Linux API <linux-api@...r.kernel.org>,
	Richard Guy Briggs <rgb@...hat.com>,
	netdev <netdev@...r.kernel.org>
Subject: Re: [PATCH V4 3/8] namespaces: expose ns instance serial numbers
 in proc

Le 25/08/2014 16:04, Andy Lutomirski a écrit :
> On Aug 25, 2014 6:30 AM, "Nicolas Dichtel" <nicolas.dichtel@...nd.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. These ifindexes depend on the interface creation order:

$ ip netns add 1
$ ip link set eth1 netns 1
$ ip netns exec 1 ip link add veth0 type veth peer name veth1
$ ip netns exec 1 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 9a:a0:89:99:a0:3c brd ff:ff:ff:ff:ff:ff
3: eth1: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT group 
default qlen 1000
     link/ether 52:54:00:12:34:57 brd ff:ff:ff:ff:ff:ff
4: veth0: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT 
group default qlen 1000
     link/ether 96:86:44:49:ce:a8 brd ff:ff:ff:ff:ff:ff
$ ip netns del 1
$ ip netns add 1
$ ip netns exec 1 ip link add veth0 type veth peer name veth1
$ ip link set eth1 netns 1
$ ip netns exec 1 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 86:92:90:01:32:6b 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 ae:8b:d2:71:48:a2 brd ff:ff:ff:ff:ff:ff
4: eth1: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT group 
default qlen 1000
     link/ether 52:54:00:12:34:57 brd ff:ff:ff:ff:ff:ff

Note: when an interface is moved to another netns, the ifindex is kept if
possible, else another ifindex is chosen.
I will dig a bit to understand how CRIU save these netns informations.

>>
>>>
>>> Also, 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.
>>
>> Another scenario is when you have virtual network devices across two netns. You
>> need to identify the peer netns to have a netlink message which is fully interpretable by the userspace.
>
> 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 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@...0: <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@...l0: <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.


Hope it's more clear now.


Regards,
Nicolas
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists