Re: [PATCH mmotm,next] mm: fix memcg-less page reclaim

[Michal Hocko <mhocko@suse.cz>]

[Let's CC Johannes, Kamezewa and Kosaki]

On Mon 19-08-13 09:44:07, Michal Hocko wrote:
> On Sun 18-08-13 23:05:25, Hugh Dickins wrote:
[...]
> > Adding mem_cgroup_disabled() and once++ test there is ugly.  Ideally,
> > even a !CONFIG_MEMCG build might in future have a stub root_mem_cgroup,
> > which would get around this: but that's not so at present.
> > 
> > However, it appears that nothing actually dereferences the memcg pointer
> > in the mem_cgroup_disabled() case, here or anywhere else that case can
> > reach mem_cgroup_iter() (mem_cgroup_iter_break() is not called in
> > global reclaim).
> > 
> > So, simply pass back an ordinarily-oopsing non-NULL address the first
> > time, and we shall hear about it if I'm wrong.
> 
> This is a bit tricky but it seems like the easiest way for now. I will
> look at the fake root cgroup for !CONFIG_MEMCG.

OK, the following builds for both CONFIG_MEMCG enabled and disabled and
should work with cgroup_disable=memory as well as we are allocating
root_mem_cgroup for disabled case as well AFAICS.

It looks less scary than I expected. I haven't tested it yet but if you
think that it looks promising I will send a full patch with changelog.
---