[lxc-devel] cgroup V2 and LXC

[Serge Hallyn]

Quoting Christian Brauner (christianvanbrauner at gmail.com):
> On Mon, Feb 15, 2016 at 07:48:05PM +0000, Serge Hallyn wrote:
> > Quoting Christian Brauner (christian.brauner at mailbox.org):
> > > On Wed, Feb 10, 2016 at 05:45:48PM +0000, Serge Hallyn wrote:
> > > > Quoting Christian Brauner (christian.brauner at mailbox.org):
> > > > > On Mon, Feb 01, 2016 at 04:56:08AM +0000, Serge Hallyn wrote:
> > > > > > Quoting Kevin Wilson (wkevils at gmail.com):
> > > > > > > Hi, LXC developers,
> > > > > > > 
> > > > > > > The latest kernel release (4.4) includes initial support to cgroup v2
> > > > > > > with 2 controllers (memory and io). Also it seems that the PIDs
> > > > > > > controller works in cgroup v2, but I do not know if it is officially
> > > > > > > supported in v2.
> > > > > > > 
> > > > > > > Is there any intention to replace the existing cgroup v1 usage in LXC
> > > > > > > by cgroup v2 ? or at least to enable working with both of them ?
> > > > > > > 
> > > > > > > Regards,
> > > > > > > Kevin
> > > > > > 
> > > > > > Replace, no, support, yes.  I've added support for it to cgmanager, and have
> > > > > > used lxc with the unified hierarchy through cgmanager.  Without cgmanager
> > > > > > it will currently definately not work.  It's worth discussing how we should
> > > > > > handle it - and how init wants us to handle it.   With cgmanager I actually
> > > > > > built in the support so that you could treat it as a legacy hierarchy, and
> > > > > > upstart was happy with that since it used cgmanager.  Systemd will not be
> > > > > > happy with that, and it will be a problem.  The only exception to the "no
> > > > > > tasks in a non-leaf node" rule is for the / cgroup.  So lxc would need to
> > > > > > place init in say /lxc/c1/.leaf, and systemd would have to accept that
> > > > > > /lxc/c1 is the container's cgroup.  A few possibilities:
> > > > > > 
> > > > > > 1. maybe if we place systemd in /lxc/c1/init.scope it will be happy
> > > > > Well, here is how I thought it could go (sticking to systemd specifics here):
> > > > >         - create a slice for all lxc "lxc.slice" (similar to "machine.slice" of
> > > > >           systemd-nspawn backed containers)
> > > > >         - "lxc.slice" contains a scope for each container (e.g. "c1.scope"
> > > > >         - "c1.scope" contains an "init.scope"
> > > > >         - "init.scope" only contains the PID of "/sbin/init" as seen from the
> > > > >           host (obviously)
> > > > 
> > > > So if we are creating container c1, are you talking about
> > > > 
> > > > /lxc/c1/lxc.slice/c1.scope/init.scope
> > > > 
> > > > or are you talking about a host-global
> > > > 
> > > > /lxc.slice
> > > Yes, you have lxc.slice then you have all your machines under this. This is what
> > > systemd-nspawn does if I'm not mistaken.
> > > > with container-specific
> > > > 
> > > > /lxc.slice/c1.scope
> > > > 
> > > > per container?
> > > > 
> > > > ?
> > > Yes.
> > 
> > This doesn't seem to address the problem.  Where we put these on the host doesn't
> > matter.  The question is, we create container c1, in which cgroup do we put the
> > init process?
> > 
> > Assume we create /lxc/c1 on the host as we do now.  This becomes / in the container's
> > cgroup namespace.  Where do we put init?  If we put it into (namespaced) /, then
> > systemd will not be able to create any cgroups.  So we should probably put it into
> > /init.scope.  This is fine with cgroup namespaces since it can see it is in '/init.scope'
> > (or '/' if an unprivileged container couldn't create a cgroup for some controllers).
> > But if we do not have cgroup namespaces, systemd sees it is running in perhaps
> > /user.slice/user-1000.slice/session-c6.scope/lxc/lxdvm1/lxc/c1/init.scope.  In that
> > case we want systemd to recognize init.scope and create services under
> > /user.slice/user-1000.slice/session-c6.scope/lxc/lxdvm1/lxc/c1.
> > 
> > > > >         - All other processes are put in another slice "c1-something.slice"
> > > > 
> > > > Which other processes?
> > > Well, all processes, systemd starts are either put in system.slice or
> > > user.slice. All other things we start in the container (let it be e.g. vim) is
> > > put in a session.slice (e.g. session-0.slice, session-1000.slice).
> > 
> > wc -l /sys/fs/cgroup/memory/tasks
> > 548
> This is output from a legacy cgroup. (The tasks file is removed in cgroup
> unified hierarchy, no?) I was talking about unified cgroups.

Oh, of course.

> A typical layout for a container BB running a unified cgroup system inside on a
> host running a unified cgroup system with systemd-nspawn:
> 
> /sys/fs/cgroup/machine.slice/:
>         - non-leaf node --> cgroup.procs empty
> 
> /sys/fs/cgroup/machine.slice/machine-BB\x2dtree.scope/:
>         - non-leaf node --> cgroup.procs empty
> 
> The following are on the same level: (/sys/fs/cgroup/machine.slice/machine-BB\x2dtree.scope/)
> 
> - /sys/fs/cgroup/machine.slice/machine-BB\x2dtree.scope/init.scope/:
>         - leaf node --> cgroup.procs contains PID of init
> 
> - /sys/fs/cgroup/machine.slice/machine-BB\x2dtree.scope/system.slice/:
>         - non-leaf node --> cgroup.procs empty
>         - contains leaf nodes for system setup stuff (journald, logind etc.)
> 
> - /sys/fs/cgroup/machine.slice/machine-BB\x2dtree.scope/user.slice/user-0.slice/session-c1.scope
> and
> - /sys/fs/cgroup/machine.slice/machine-BB\x2dtree.scope/user.slice/user-0.slice/user at 0.service:
>         - filled with leaf-nodes for e.g. processes started by the user

I'm really not happy to use systemd's namespace though.  We cannot assume
that either the host (or host container) or guest (container) is running
systemd.

The question still is, what does systemd look at to decide what it should do as
first step?  If we just place it into /init.scope then will it know to leave
itself put and create /system.slice and such hierarchies for its services?

That would make the most sense...  It also seems to me with cgroup namespaces
that it would not matter whether we use /lxc/container or /machine.slice/container.
The full name does leak to the container as the mount root (third field in
/proc/self/mountinfo), but systemd would have to go out of its way to
check that.

> > > > AFAIK all other processes will be created by systemd.  The q is what will it
> > > > do.  If we put systemd in /lxc.slice/c1.scope/init.scope, will it take that
> > > > as its cgroup root and try to create and move itself into
> > > > /lxc.slice/c1.scope/init.scope ?  If so it will fail since it cannot create a
> > > > cgroup while it is in it.
> > > I don't think so but I need to test that again. Time to boot unified.
> > > 
> > > > 
> > > > So I think I've convinced myself that we need to collaborate with systemd
> > > > on this.  Perhaps we can agree with it on a default cgroup in which it should
> > > > be started to tell it "this is the leaf cgroup for your init".  So if it sees
> > > > it is in /a/b/c/.cg_leaf, then it will know that /a/b/c is its root.
> > > I thought the same that's why I started to read some of the code.
> > > fwiw, systemd-nspawn already works with the unified cgroup hierarchy and I think
> > > nesting works as well. But I'm not completely sure how nspawn handles nesting.
> > 
> > Looks like it puts systemd into '/supervisor' and the container into '/payload'?
> > (nspawn-cgroup.c)
> I don't think so. This seems to be a special case when systemd-nspawn is run
> from a service unit. Otherwise the layout seems to be as I sketched above.
> _______________________________________________
> lxc-devel mailing list
> lxc-devel at lists.linuxcontainers.org
> http://lists.linuxcontainers.org/listinfo/lxc-devel