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

Le 25/08/2014 18:13, Andy Lutomirski a écrit : > On Mon, Aug 25, 2014 at 8:43 AM, Nicolas Dichtel > <nicolas.dichtel(a)6wind.com&gt; wrote: >> >> Le 25/08/2014 16:04, Andy Lutomirski a écrit : >> >>> On Aug 25, 2014 6:30 AM, "Nicolas Dichtel" <nicolas.dichtel(a)6wind.com&gt; >>> 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 >