[CRIU] [PATCH 1/2] Re-create cgroups if necessary

[Serge Hallyn]

Quoting Saied Kazemi (saied at google.com):
> Pavel,
> 
> Just a quick follow up:
> 
> > 1. cgroups sub-hierarchy
> > 2. configuration of cgroups
> > 3. populating cgroups with tasks
> > 4. creating cgroup mountpoints (if any)
> 
> Pertaining to the list above, one of our interns is working on #2 and will

Note that what Tycho is working on (1) should be coordinated with (2).
The cgroups should be checkpointed as a hierarchical set of structs
mirrorring the cgroup fs trees (and imo the cg_ctls should then point
into these trees), so something like

struct cgroup {
	char *name;
	struct list_head children;
	int n_contollers;  // # of co-mounted controllers
	char *controllers[];
	struct mem_limits *mem_limits;
	struct devcg_limits *devcg_limits;
	...
};

struct list_head init_cgroup_heads;

struct cg_ctl {
	struct list_head list;  // connect all (not-comounted) controllers
	struct cgroup *cgroup;
}

> probably have a patch in a month.  Please let me know if someone is already
> working on this to avoid duplication of effort.
> 
> > With a CLI option we tell CRIU:
> >
> > 1. Expect the cgroup to already exist, just put the process back in it.
>  If cgroup doesn't exist, fail.
> > 2. Expect the cgroup not to exist, create it and put the process in it.
>  If cgroup exists, fail.
> 
> I was wondering if you had a chance to think about this.  As explained
> before, there are legitimate cases where the cgroups do already exist prior
> to criu restore, so their absence is a definite error.  Also, there are
> legitimate cases where the cgroups do not already exist prior to criu
> restore, so CRIU should create them.  CRIU on its own cannot determine
> which scenario it's dealing with.  We can make the default action to always
> create cgroups, but we still need a mechanism to tell it otherwise.

And there are many legitimate cases where we care about the parent
cgroup under which the restarted tasks should be grouped, but the
actual cgroup name doesn't matter.  At the very least criu should have
an option to rename the head cgroup, so that the caller can create a
new cgroup with mkstemp, and pass that name on to criu, which can then
populate all child cgroups under there.

My preference is still for a template option, to say that if we
checkpointed u1, then we will restarted some template beginning with
u1, but just allowing the caller to specify a name would be fine.

To be more precise, if the init task of the set being checkpointed
was in cgroups /lxc/u1, and criu restore is called with "--cgroup-into x1",
then criu would restore into /lxc/x1.

-serge