On Wed, Apr 11, 2018 at 02:16:23PM -0500, Eric W. Biederman wrote: > Christian Brauner <christian.brau...@canonical.com> writes: > > > On Wed, Apr 11, 2018 at 01:37:18PM -0500, Eric W. Biederman wrote: > >> Christian Brauner <christian.brau...@canonical.com> writes: > >> > >> > On Wed, Apr 11, 2018 at 11:40:14AM -0500, Eric W. Biederman wrote: > >> >> Christian Brauner <christian.brau...@canonical.com> writes: > >> >> > Yeah, agreed. > >> >> > But I think the patch is not complete. To guarantee that no > >> >> > non-initial > >> >> > user namespace actually receives uevents we need to: > >> >> > 1. only sent uevents to uevent sockets that are located in network > >> >> > namespaces that are owned by init_user_ns > >> >> > 2. filter uevents that are sent to sockets in mc_list that have > >> >> > opened a > >> >> > uevent socket that is owned by init_user_ns *from* a > >> >> > non-init_user_ns > >> >> > > >> >> > We account for 1. by only recording uevent sockets in the global > >> >> > uevent > >> >> > socket list who are owned by init_user_ns. > >> >> > But to account for 2. we need to filter by the user namespace who owns > >> >> > the socket in mc_list. So in addition to that we also need to slightly > >> >> > change the filter logic in kobj_bcast_filter() I think: > >> >> > > >> >> > diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c > >> >> > index 22a2c1a98b8f..064d7d29ace5 100644 > >> >> > --- a/lib/kobject_uevent.c > >> >> > +++ b/lib/kobject_uevent.c > >> >> > @@ -251,7 +251,8 @@ static int kobj_bcast_filter(struct sock *dsk, > >> >> > struct sk_buff *skb, void *data) > >> >> > return sock_ns != ns; > >> >> > } > >> >> > > >> >> > - return 0; > >> >> > + /* Check if socket was opened from non-initial user namespace. > >> >> > */ > >> >> > + return sk_user_ns(dsk) != &init_user_ns; > >> >> > } > >> >> > #endif > >> >> > > >> >> > > >> >> > But correct me if I'm wrong. > >> >> > >> >> You are worrying about NETLINK_LISTEN_ALL_NSID sockets. That has > >> >> permissions and an explicit opt-in to receiving packets from multiple > >> >> network namespaces. > >> > > >> > I don't think that's what I'm talking about unless that is somehow the > >> > default for NETLINK_KOBJECT_UEVENT sockets. What I'm worried about is > >> > doing > >> > > >> > unshare -U --map-root > >> > > >> > then opening a NETLINK_KOBJECT_UEVENT socket and starting to listen to > >> > uevents. Imho, this should not be possible because I'm in a > >> > non-init_user_ns. But currently I'm able to - even with the patch to > >> > come - since the uevent socket in the kernel was created when init_net > >> > was created and hence is *owned* by the init_user_ns which means it is > >> > in the list of uevent sockets. Here's a demo of what I mean: > >> > > >> > https://asciinema.org/a/175632 > >> > >> Why do you care about this case? > > > > It's not so much that I care about this case since any workload that > > wants to run a separate udevd will have to unshare the network namespace > > and the user namespace for it to make complete sense. > > What I do care about is that the two of us are absolutely in the clear > > about what semantics we are going to expose to userspace and it seems > > that we were talking past each other wrt to this "corner case". > > For userspace, it needs to be very clear that the intention is to filter > > by *owning user namespace of the network namespace a given task resides > > in* and not by user namespace of the task per se. This is what this > > corner case basically shows, I think. > > If this is just a clarification of semantics then yes this is a > productive question. I almost agree with your definition above. > > I would make the definition very simple. Uevents will not be broadcast > via netlink in a network namespace where net->user_ns != &init_user_ns, > with the exception of uevents for network devices in that network > namespace.
Well, for the sake of posterity :) I should add that I'd prefer we'd add what I suggested above: - return 0; + /* Check if socket was opened from non-initial user namespace. */ + return sk_user_ns(dsk) != &init_user_ns; } to slam the door shut once and for all for all non-init_user_ns namespaces because it *seems* like the cleanest solution: uevents are owned by init_user_ns; period. Because it is the only user namespace that can do anything interesting with them *by default*. But what we have now right now with my upcoming patch is at least sufficient and safe. Christian > > The existing filtering by the sending uid and verifying that it is uid 0 > gives a little more room to filter if we want (as udev & friends will > ignore the uevent), but I don't see the point. > > Eric