Re: [RFC PATCH] vhost, mm: make sure that oom_reaper doesn't reap memory read by vhost

[Michal Hocko]

On Sun 19-06-16 23:35:43, Michal Hocko wrote:
> On Sat 18-06-16 03:09:02, Michael S. Tsirkin wrote:
> > On Fri, Jun 17, 2016 at 11:00:17AM +0200, Michal Hocko wrote:
[...]
> > > diff --git a/include/linux/uaccess.h b/include/linux/uaccess.h
> > > index 349557825428..b1f314fca3c8 100644
> > > --- a/include/linux/uaccess.h
> > > +++ b/include/linux/uaccess.h
> > > @@ -76,6 +76,28 @@ static inline unsigned long __copy_from_user_nocache(void *to,
> > >  #endif		/* ARCH_HAS_NOCACHE_UACCESS */
> > >  
> > >  /*
> > > + * A safe variant of __get_user for for use_mm() users to have a
> > > + * gurantee that the address space wasn't reaped in the background
> > > + */
> > > +#define __get_user_mm(mm, x, ptr)				\
> > > +({								\
> > > +	int ___gu_err = __get_user(x, ptr);			\
> > > +	if (!___gu_err && test_bit(MMF_UNSTABLE, &mm->flags))	\
> > 
> > test_bit is somewhat expensive. See my old mail
> > 	x86/bitops: implement __test_bit
> 
> Do you have a msg_id?
> 
> > I dropped it as virtio just switched to simple &/| for features,
> > but we might need something like this now.
> 
> Is this such a hot path that something like this would make a visible
> difference? 

OK, so I've tried to apply your patch [1] and updated both __get_user_mm
and __copy_from_user_mm and the result is a code size reduction:
   text    data     bss     dec     hex filename
  12835       2      32   12869    3245 drivers/vhost/vhost.o
  12882       2      32   12916    3274 drivers/vhost/vhost.o.before

This is really tiny and I cannot tell anything about the performance. Should
I resurrect your patch and push it together with this change or this can happen
later?

[1] http://lkml.kernel.org/r/1440776707-22016-1-git-send-email-mst@xxxxxxxxxx
-- 
Michal Hocko
SUSE Labs