Re: [PATCH 1/5] memcg: disable kmem charging in nmi for unsupported arch

[Shakeel Butt <shakeel.butt@linux.dev>]

On Fri, May 16, 2025 at 08:37:23AM -0700, Shakeel Butt wrote:
> On Fri, May 16, 2025 at 11:30:17AM +0200, Vlastimil Babka wrote:
> > On 5/16/25 08:49, Shakeel Butt wrote:
> > > The memcg accounting and stats uses this_cpu* and atomic* ops. There are
> > > archs which define CONFIG_HAVE_NMI but does not define
> > > CONFIG_ARCH_HAS_NMI_SAFE_THIS_CPU_OPS and ARCH_HAVE_NMI_SAFE_CMPXCHG, so
> > > memcg accounting for such archs in nmi context is not possible to
> > > support. Let's just disable memcg accounting in nmi context for such
> > > archs.
> > > 
> > > Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
> > > ---
> > >  include/linux/memcontrol.h |  5 +++++
> > >  mm/memcontrol.c            | 15 +++++++++++++++
> > >  2 files changed, 20 insertions(+)
> > > 
> > > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > > index f7848f73f41c..53920528821f 100644
> > > --- a/include/linux/memcontrol.h
> > > +++ b/include/linux/memcontrol.h
> > > @@ -62,6 +62,11 @@ struct mem_cgroup_reclaim_cookie {
> > >  
> > >  #ifdef CONFIG_MEMCG
> > >  
> > > +#if defined(CONFIG_ARCH_HAS_NMI_SAFE_THIS_CPU_OPS) || \
> > > +	!defined(CONFIG_HAVE_NMI) || defined(ARCH_HAVE_NMI_SAFE_CMPXCHG)
> > > +#define MEMCG_SUPPORTS_NMI_CHARGING
> > > +#endif
> > > +
> > >  #define MEM_CGROUP_ID_SHIFT	16
> > >  
> > >  struct mem_cgroup_id {
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index e17b698f6243..dface07f69bb 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -2647,11 +2647,26 @@ static struct obj_cgroup *current_objcg_update(void)
> > >  	return objcg;
> > >  }
> > >  
> > > +#ifdef MEMCG_SUPPORTS_NMI_CHARGING
> > > +static inline bool nmi_charging_allowed(void)
> > > +{
> > > +	return true;
> > > +}
> > > +#else
> > > +static inline bool nmi_charging_allowed(void)
> > > +{
> > > +	return false;
> > > +}
> > > +#endif
> > > +
> > >  __always_inline struct obj_cgroup *current_obj_cgroup(void)
> > >  {
> > >  	struct mem_cgroup *memcg;
> > >  	struct obj_cgroup *objcg;
> > >  
> > > +	if (in_nmi() && !nmi_charging_allowed())
> > 
> > Exchange the two as the latter is compile-time constant, so it can shortcut
> > the in_nmi() check away in all the good cases?
> > 
> 
> Oh I thought compiler would figure that out but now that I think about
> it, it can only do so if the first condition does not have any
> side-effects and though in_nmi() does not, I am not sure if compiler can
> extract that information.
> 
> I will fix this and make sure that compiler is doing the right thing.

So, gcc 11.5 generates the same code irrespective of checking in_nmi()
first or second i.e. avoid in_nmi() check altogether on x86_64. I will
still rearrange the checks to not leave this optimization to compilers.