[Devel] Re: [PATCH v2 08/11] memcg: disable kmem code when not in use.

[Michal Hocko]

On Fri 17-08-12 11:01:06, Glauber Costa wrote:
> On 08/17/2012 11:02 AM, Michal Hocko wrote:
> > On Thu 09-08-12 17:01:16, Glauber Costa wrote:
> >> We can use jump labels to patch the code in or out when not used.
> >>
> >> Because the assignment: memcg->kmem_accounted = true is done after the
> >> jump labels increment, we guarantee that the root memcg will always be
> >> selected until all call sites are patched (see memcg_kmem_enabled).
> > 
> > Not that it would be really important because kmem_accounted goes away

And just found out it doesn't go away completely, it just transforms
from bool to unsigned log (with flags). The rest still holds...

> > in a subsequent patch but I think the wording is a bit misleading here.
> > First of all there is no guanratee that kmem_accounted=true is seen
> > before atomic_inc(&key->enabled) because there is no memory barrier and
> > the lock serves just a leave barrier. But I do not think this is
> > important at all because key->enabled is what matters here. Even if
> > memcg_kmem_enabled is true we do not consider it if the key is disabled,
> > right?
> > 
> 
> Right.

-- 
Michal Hocko
SUSE Labs