"W. Trevor King" [off-list ref] writes:

On Sat, Jul 23, 2016 at 04:56:44PM -0500, Eric W. Biederman wrote:

quoted

"W. Trevor King" [off-list ref] writes:

quoted

On Sat, Jul 23, 2016 at 02:38:56PM -0700, James Bottomley wrote:

quoted

On Sat, 2016-07-23 at 14:14 -0700, W. Trevor King wrote:

quoted

namespaces(7) and clone(2) both have:

  When a network namespace is freed (i.e., when the last
  process in the namespace terminates), its physical network
  devices are moved back to the initial network namespace (not
  to the parent of the process).

So the initial network namespace (the head of
net_namespace_list?)  is special [1].  To understand how
physical network devices will be handled, it seems like we want
to treat network devices as a depth-1 tree, with all
non-initial net namespaces as children of the initial net
namespace.  Can we extend this series' NS_GET_PARENT to return:

* EPERM for an unprivileged caller (like this series currently
  does for PID namespaces),
* ENOENT when called on net_namespace_list, and
* net_namespace_list when called on any other net namespace.

What's the practical application of this?  independent net
namespaces are managed by the ip netns command.  It pins them by
a bind mount in a flat fashion; if we make them hierarchical the
tool would probably need updating to reflect this, so we're going
to need a reason to give the network people.  Just having the
interfaces not go back to root when you do an ip netns delete
doesn't seem very compelling.

I'm not suggesting we add support for deeper nesting, I'm suggesting
we use NS_GET_PARENT to allow sufficiently privileged users to
determine if a given net namespace is the initial net namespace.  You
could do this already with something like:

1. Create a new net namespace.
2. Add a physical network device to that namespace.
3. Delete that namespace.
4. See if the physical network device shows up in your
   initial-net-namespace candidate.
5. Delete the physical network device (hopefully it ended up
   somewhere you can find it ;).

But using an NS_GET_PARENT call seems much safer and easier.

Have you had the problem in practice where you can't tell which
network namespace is the initial network namespace.  This all seems
like a theoretical problem rather than a real one.

I haven't had any practical problems here, I'm just trying to wrap my
head around namespace-relationship discovery.  The special physical
network device handling seems a lot like init re-parenting (with no
PR_SET_CHILD_SUBREAPER analog in a 1-deep namespace tree), so calling
the initial network namespace a parent (and all the other namespaces
its direct children) seems natural enough.  If that doesn't sound
convincing, I'm happy to punt this idea until someone runs into a
practical problem ;).

Then let's punt this until someone runs into a practical problem.

For scaling and for sanity it is desirable to keep the connections
between namespaces to a minimum.  Further the initial instances of a
namespace always tend to be a little bit special.

Eric