[PATCH, v6 3/3] cgroups: introduce timer slack controller

Matt Helsley matthltc at us.ibm.com
Mon Feb 14 15:39:36 PST 2011


On Tue, Feb 15, 2011 at 12:39:39AM +0200, Kirill A. Shutemov wrote:
> On Mon, Feb 14, 2011 at 06:01:06PM +0100, Thomas Gleixner wrote:
> > B1;2401;0cOn Mon, 14 Feb 2011, Kirill A. Shutemov wrote:
> > 
> > > On Mon, Feb 14, 2011 at 03:00:03PM +0100, Thomas Gleixner wrote:
> > > > On Mon, 14 Feb 2011, Kirill A. Shutsemov wrote:
> > > > > From: Kirill A. Shutemov <kirill at shutemov.name>

<snip>

> > > > > +	list_for_each_entry(cur, &cgroup->children, sibling) {
> > > > > +		child = cgroup_to_tslack_cgroup(cur);
> > > > > +		if (type == TIMER_SLACK_MIN && val > child->min_slack_ns)
> > > > > +			return -EBUSY;
> > > > 
> > > >   I thought the whole point is to propagate values through the group.
> > > 
> > > I think silent change here is wrong. cpuset returns -EBUSY in similar
> > > case.
> > 
> > And how is cpuset relevant for this ? Not at all. This is about

I agree with Thomas here -- cpusets aren't relevant.

> > timer_slack and we better have a well defined scheme for all of this
> > and not some cobbled together thing with tons of exceptions and corner
> > cases. Of course undocumented as far the code goes.
> 
> I don't like silent cascade changes. Userspace can implement it if

It need not be totally silent. memcg has a "use_hierarchy" flag file.
Alternately, you could punt for now and disable hierarchy somewhat like
blkio does.

> needed. -EBUSY is appropriate.

Hmm, I haven't thought about that method of cascading enough. The important
question to consider is how will the parent cgroup be constrained if the
owner/group of the children is different and thus disallows userspace from
implementing this cascade. I suppose it's consistent with the owner/group
ids but it hardly seems consistent with the "spirit" of using cgroups to
enable things like containers.

Cheers,
	-Matt Helsley


More information about the Containers mailing list