Re: [PATCH 06/51] memcg: add mem_cgroup_root_css

[Michal Hocko]

On Wed 17-06-15 14:25:00, Tejun Heo wrote:
> Hey, Michal.
> 
> On Wed, Jun 17, 2015 at 04:56:42PM +0200, Michal Hocko wrote:
> > On Fri 22-05-15 17:13:20, Tejun Heo wrote:
> > > Add global mem_cgroup_root_css which points to the root memcg css.
> > 
> > Is there any reason to using css rather than mem_cgroup other than the
> > structure is not visible outside of memcontrol.c? Because I have a
> > patchset which exports it. It is not merged yet so a move to mem_cgroup
> > could be done later. I am just interested whether there is a stronger
> > reason.
> 
> It doesn't really matter either way but I think it makes a bit more
> sense to use css as the common type when external code interacts with
> cgroup controllers.  e.g. cgroup writeback interacts with both memcg
> and blkcg and in most cases it doesn't know or care about their
> internal states.  Most of what it wants is tracking them and doing
> some common css operations (refcnting, printing and so on) on them.

I see and yes, it makes some sense. I just think we can get rid of the
accessor functions when the struct mem_cgroup is visible and the code
can simply do &{page->}mem_cgroup->css.

> > > This will be used by cgroup writeback support.  If memcg is disabled,
> > > it's defined as ERR_PTR(-EINVAL).
> > 
> > Hmm. Why EINVAL? I can see only mm/backing-dev.c (in
> > review-cgroup-writeback-switch-20150528 branch) which uses it and that
> > shouldn't even try to compile if !CONFIG_MEMCG no? Otherwise we would
> > simply blow up.
> 
> Hmm... the code maybe has changed inbetween but there was something
> which depended on the root css being defined when
> !CONFIG_CGROUP_WRITEBACK or maybe it was on blkcg_root_css and memcg
> side was added for consistency.

I have tried to compile with !CONFIG_MEMCG and !CONFIG_CGROUP_WRITEBACK
without mem_cgroup_root_css defined for this configuration and
mm/backing-dev.c compiles just fine. So maybe we should get rid of it
rather than have a potentially tricky code?

> An ERR_PTR value is non-zero, which
> is an invariant which is often depended upon, while guaranteeing oops
> when deref'd.

Yeah, but css_{get,put} and others consumers of the pointer are not
checking for ERR_PTR. So I think this is really misleading.

-- 
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/