On Wed, 21 Nov 2012, Michal Hocko wrote:
> On Tue 20-11-12 13:49:32, Andrew Morton wrote:
> > On Mon, 19 Nov 2012 17:44:34 -0800 (PST)
> > David Rientjes <rientjes@xxxxxxxxxx> wrote:
> >
> > > While profiling numa/core v16 with cgroup_disable=memory on the command
> > > line, I noticed mem_cgroup_count_vm_event() still showed up as high as
> > > 0.60% in perftop.
> > >
> > > This occurs because the function is called extremely often even when memcg
> > > is disabled.
> > >
> > > To fix this, inline the check for mem_cgroup_disabled() so we avoid the
> > > unnecessary function call if memcg is disabled.
> > >
> > > ...
> > >
> > > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > > --- a/include/linux/memcontrol.h
> > > +++ b/include/linux/memcontrol.h
> > > @@ -181,7 +181,14 @@ unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
> > > gfp_t gfp_mask,
> > > unsigned long *total_scanned);
> > >
> > > -void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx);
> > > +void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx);
> > > +static inline void mem_cgroup_count_vm_event(struct mm_struct *mm,
> > > + enum vm_event_item idx)
> > > +{
> > > + if (mem_cgroup_disabled() || !mm)
> > > + return;
> > > + __mem_cgroup_count_vm_event(mm, idx);
> > > +}
> >
> > Does the !mm case occur frequently enough to justify inlining it, or
> > should that test remain out-of-line?
>
> Now that you've asked about it I started looking around and I cannot see
> how mm can ever be NULL. The condition is there since the very beginning
> (456f998e memcg: add the pagefault count into memcg stats) but all the
> callers are page fault handlers and those shouldn't have mm==NULL.
> Or is there anything obvious I am missing?
>
> Ying, the whole thread starts https://lkml.org/lkml/2012/11/19/545 but
> the primary question is why we need !mm test for mem_cgroup_count_vm_event
> at all.
Here's a guess: as Ying's 456f998e patch started out in akpm's tree,
shmem.c was calling mem_cgroup_count_vm_event(current->mm, PGMAJFAULT).
Then I insisted that was inconsistent with how we usually account when
one task touches another's address space, and rearranged it to work on
vma->vm_mm instead.
Done the original way, if the touching task were a kernel daemon (KSM's
ksmd comes to my mind), then the current->mm could well have been NULL.
I agree with you that it looks redundant now.
Hugh
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
