Re: [PATCH] cgroup namespaces: add a 'nsroot=' mountinfo field

[Serge E. Hallyn]

Quoting Eric W. Biederman (ebiederm@xxxxxxxxxxxx):
> "Serge E. Hallyn" <serge@xxxxxxxxxx> writes:
> 
> > Quoting Eric W. Biederman (ebiederm@xxxxxxxxxxxx):
> >> "Serge E. Hallyn" <serge@xxxxxxxxxx> writes:
> >> 
> >> > This is so that userspace can distinguish a mount made in a cgroup
> >> > namespace from a bind mount from a cgroup subdirectory.
> >> 
> >> To do that do you need to print the path, or is an extra option that
> >> reveals nothing except that it was a cgroup mount sufficient?
> >> 
> >> Is there any practical difference between a mount in a namespace and a
> >> bind mount?
> >> 
> >> Given the way the conversation has been going I think it would be good
> >> to see the answers to these questions.  Perhaps I missed it but I
> >> haven't seen the answers to those questions.
> >
> > Yup, I tried to answer those in my last email, let me try again.
> >
> > Let's say I start a container using cgroup namespaces, /lxc/x1.  It mounts
> > freezer at /sys/fs/cgroup so it has field three of mountinfo as /lxc/x1,
> > and /sys/fs/cgroup/ is the path to the container's cgroup (/lxc/x1).  In
> > that container, I start another container x1, not using cgroup namespaces.
> > It also wants a cgroup mount, and a common way to handle that (to prevent
> > container rewriting its limits) is to mount a tmpfs at /sys/fs/cgroup,
> > create /sysfs/cgroup/lxc/x1, and bind mount /sys/fs/cgroup/lxc/x1 from
> > the parent container onto /sys/fs/cgroup/lxc/x1 in the child container.
> > Now for that bind mount, the mountinfo field 3 will show /lxc/x1/lxc/x1,
> > with mount target /sys/fs/cgroup/lxc/x1, while /proc/self/cgroup for a task
> > in that container will show '/lxc/x1'.  Unless it has been moved into
> > /lxc/x1/lxc/x1 in the container (/lxc/x1/lxc/x1/lxc/x1 on the host)...
> > Every time I've thought "maybe we can just..." I've found a case where it
> > wouldn't work.
> >
> > At first in lxc we simply said if /proc/self/ns/cgroup exists assume that
> > the cgroupfs mounts are not bind mounts.  However, old userspace (and
> > container drivers) on new kernels is certainly possible, especially an
> > older distro in a container on a newer distro on the host.  That completely
> > breaks with this approach.
> >
> > I also personally think there *is* value in letting a task know its
> > place on the system, so hiding the full cgroup path is imo not only not
> > a valid goal, it's counter-productive.  Part of making for better
> > virtualization is to give userspace all the info it needs about its
> > current limits.  Consider that with the unified hierarchy, you cannot
> > have tasks in a cgroup that also has child cgroups - except for the
> > root.  Cgroup namespaces do not make an exception for this, so knowing
> > that you are not in the absolute cgroup root actually can prevent you
> > from trying something that cannot work.  Or, I suppose, at least
> > understanding why you're unable to do what you're trying to do (namely
> > your container manager messed up).  I point this out because finding
> > a way to only show the namespaced root in field 3 of mountinfo would
> > fix the base problem, but at the cost of hiding useful information
> > from a container.
> 
> It is just the superblock show_path method.  And regardless of the rest
> of the usefullness of your mount option implementing show_path appears

Ugh.  Yeah as I've said implementing that would be the other way to go.
I'm somewhat loath to give up the extra information, but I can work
on that patch later this week.

> to be fundamentally the right thing in this context.  As that field
> appears to have the same issue as /proc/self/cgroup.

Well, /proc/self/cgroup could also have been fixed by adding a
':<nsroot>" field to each line, but it's used differently...

thanks,
-serge