Re: Documenting the ioctl interfaces to discover relationships between namespaces

[ebiederm@xmission.com (Eric W. Biederman)]

"Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> writes:

> On 12/11/2016 11:30 PM, Eric W. Biederman wrote:
>> "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> writes:
>> 
>>> [was: [PATCH 0/4 v3] Add an interface to discover relationships
>>> between namespaces]
>> 
>> One small comment below.
>> 
>>>
>>>    Introspecting namespace relationships
>>>        Since Linux 4.9, two ioctl(2) operations  are  provided  to  allow
>>>        introspection  of  namespace relationships (see user_namespaces(7)
>>>        and pid_namespaces(7)).  The form of the calls is:
>>>
>>>            ioctl(fd, request);
>>>
>>>        In each case, fd refers to a /proc/[pid]/ns/* file.
>>>
>>>        NS_GET_USERNS
>>>               Returns a file descriptor that refers to  the  owning  user
>>>               namespace for the namespace referred to by fd.
>>>
>>>        NS_GET_PARENT
>>>               Returns  a file descriptor that refers to the parent names‐
>>>               pace of the namespace referred to by fd.  This operation is
>>>               valid  only for hierarchical namespaces (i.e., PID and user
>>>               namespaces).  For user namespaces, NS_GET_PARENT is synony‐
>>>               mous with NS_GET_USERNS.
>>>
>>>        In each case, the returned file descriptor is opened with O_RDONLY
>>>        and O_CLOEXEC (close-on-exec).
>>>
>>>        By applying fstat(2) to the returned file descriptor, one  obtains
>>>        a  stat structure whose st_ino (inode number) field identifies the
>>>        owning/parent namespace.  This inode number can  be  matched  with
>>>        the  inode  number  of  another  /proc/[pid]/ns/{pid,user} file to
>>>        determine whether that is the owning/parent namespace.
>> 
>> Like all fstat inode comparisons to be fully accurate you need to
>> compare both the st_ino and st_dev.  I reserve the right for st_dev to
>> be significant when comparing namespaces.  Otherwise I might have to
>> create a namespace of namespaces someday and that is ugly.
>> 
>>>        Either of these ioctl(2) operations can fail  with  the  following
>>>        error:
>>>
>>>        EPERM  The  requested  namespace is outside of the caller's names‐
>>>               pace scope.  This error can occur if, for example, the own‐
>>>               ing  user  namespace is an ancestor of the caller's current
>>>               user namespace.  It can also occur on  attempts  to  obtain
>>>               the parent of the initial user or PID namespace.
>>>
>>>        Additionally,  the  NS_GET_PARENT operation can fail with the fol‐
>>>        lowing error:
>>>
>>>        EINVAL fd refers to a nonhierarchical namespace.
>>>
>>>        See the EXAMPLE section for an example of the use of these  opera‐
>>>        tions.
>
> So, after playing with this a bit, I have a question. 
>
> I gather that in order to, for example, elaborate the tree of user
> namespaces on the system, one would use NS_GET_PARENT on each of
> the /proc/*/ns/user files and match up the results. Right?
> 	   
> What happens if one of the parent user namespaces contains no
> processes? That is, the parent namespace exists by virtue of being
> pinned because a proc/PID/ns/user file is open or bind mounted.
> (Chrome seems to do this sort of dance with user namespaces, for
> example.) How do we find the ancestor of *that* user namespace?

What is returned from NS_GET_USERNS and NS_GET_PARENT is a file
descriptor, that you can call NS_GET_PARENT on.

Eric