[lxc-devel] CLONE_PARENT after setns(CLONE_NEWPID)

[Serge Hallyn]

Quoting Eric W. Biederman (ebiederm at xmission.com):
> Oleg Nesterov <oleg at redhat.com> writes:
> 
> > Hi Serge,
> >
> > On 11/06, Serge Hallyn wrote:
> >>
> >> Hi Oleg,
> >>
> >> commit 40a0d32d1eaffe6aac7324ca92604b6b3977eb0e :
> >> "fork: unify and tighten up CLONE_NEWUSER/CLONE_NEWPID checks"
> >> breaks lxc-attach in 3.12.  That code forks a child which does
> >> setns() and then does a clone(CLONE_PARENT).  That way the
> >> grandchild can be in the right namespaces (which the child was
> >> not) and be a child of the original task, which is the monitor.
> 
> Serge that is a clever trick to get around the limitation that we can
> not change the pid namespace of our current process.  Given the
> challenging relaying of signals etc I can see why you would use this.
> 
> At the same time it makes me a little sad to see new users of
> CLONE_PARENT.  With CLONE_THREAD in existence the original reasons for
> CLONE_PARENT are gone now.
> 
> Having used bash as an init process I know it can handle unexpeted
> children.  However using CLONE_PARENT in this way still seems a little
> dodgy.  Or am I misunderstanding why you are using CLONE_PARENT?

FWIW Christian (cc:d from the start) was the author of that code, so he
can correct me if i mis-speak, but IIUC the design is:

1. pid X is the first process running lxc-attach.  It will be a monitor
for the process which is entered into the container

2. pid X forks pid Y, which does setns().  Now if it is setns()ing into
a pidns, it won't itself be in the new pidns, which is not satisfactory.
So

3. pid Y clones pid Z with CLONE_PARENT.  Y exists.  Z continues, as a
full member of the container, and a child of the monitor process.

So yes, as you said it's exactly to work around the fact that pid Y
can't change its own pidns.

-serge