Re: [PATCH] jump_label: jump_label for boot options.

[Johannes Weiner <hannes@cmpxchg.org>]

On Fri, Dec 02, 2011 at 01:53:27PM +0100, Peter Zijlstra wrote:
> On Fri, 2011-12-02 at 13:45 +0100, Johannes Weiner wrote:
> > On Fri, Dec 02, 2011 at 10:24:10AM +0100, Peter Zijlstra wrote:
> > > On Fri, 2011-12-02 at 09:22 +0900, KAMEZAWA Hiroyuki wrote:
> > > > 
> > > > Yes, that's an idea. But we also have another stupid shadow swap accounting
> > > > table.  This can be disabled at boot, too. 
> > > 
> > > Bah I thought it was just the page frame thing, hnaz any plans to kill
> > > this swap array as well?
> > 
> > I haven't looked at the swap accounting at all yet, sorry.  But where
> > did this discussion go all of a sudden? :-)
> > 
> > That array is not allocated at all when the memory controller is
> > disabled at boot-time.
> > 
> > Rather, we have those mem_cgroup_disabled() conditionals everytime we
> > enter the memory controller from the VM and the idea is to patch them
> > out during boot, since you can not re-enable the thing anyway.
> 
> Yeah, but without those arrays you could.. this boot time switch really
> is a wart and if you don't have the shadow page frame and shadow swap
> accounting muck stuff you could runtime flip all this..

Ah, got you.

The shadow page frame is on its way out.  The extra
swp_entry_t->cgroup_id array is allocated during swap-on in vmap
space, so I don't think we have much of a runtime problem with that
one, either.

The biggest problem I see is that we modify per-page(_cgroup) state
when charging/uncharging against the root_mem_cgroup.  But this sucks
anyway (overhead for distro-kernel users that don't care about memcg),
so we need to ditch that.  And once that is gone, we are much closer
to runtime-toggling the controller.

> I've no objection to using jump_labels fwiw, we're looking to do the
> same with the cpu controller for the nr_cgroups == 0 case. But boot time
> stuff just doesn't make sense to me.

Agreed.