On Wed, Oct 18, 2017 at 03:15:03PM +0200, Thomas Gleixner wrote:
> On Wed, 18 Oct 2017, Linus Torvalds wrote:
> > On Tue, Oct 17, 2017 at 3:33 AM, Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> wrote:
> > >
> > > It looks like a compiler bug. The code of slob_units() try to read two
> > > bytes at ffff88001c4afffe. It's valid. But the compiler generates
> > > wrong code that try to read four bytes.
> > >
> > > static slobidx_t slob_units(slob_t *s)
> > > {
> > > if (s->units > 0)
> > > return s->units;
> > > return 1;
> > > }
> > >
> > > s->units is defined as two bytes in this setup.
> > >
> > > Wrongly generated code for this part.
> > >
> > > 'mov 0x0(%rbp), %ebp'
> > >
> > > %ebp is four bytes.
> > >
> > > I guess that this wrong four bytes read cross over the valid memory
> > > boundary and this issue happend.
> >
> > Hmm. I can see why the compiler would do that (16-bit accesses are
> > slow), but it's definitely wrong.
> >
> > Does it work ok if that slob_units() code is written as
> >
> > static slobidx_t slob_units(slob_t *s)
> > {
> > int units = READ_ONCE(s->units);
> >
> > if (units > 0)
> > return units;
> > return 1;
> > }
> >
> > which might be an acceptable workaround for now?
>
> Discussed exactly that with Peter Zijlstra yesterday, but we came to the
> conclusion that this is a whack a mole game. It might fix this slob issue,
> but what guarantees that we don't have the same problem in some other
> place? Just duct taping this particular instance makes me nervous.
I have checked that above patch works fine but I agree with Thomas.
> Joonsoo says:
>
> > gcc 4.8 and 4.9 fails to generate proper code. gcc 5.1 and
> > the latest version works fine.
>
> > I guess that this problem is related to the corner case of some
> > optimization feature since minor code change makes the result
> > different. And, with -O2, proper code is generated even if gcc 4.8 is
> > used.
>
> So it would be useful to figure out which optimization bit is causing that
> and blacklist it for the affected compiler versions.
I have tried it but cannot find any clue. What I did is that compiling
with -O2 and disabling some options to make option list as same as
-Os. Some guide line is roughly mentioned in gcc man page. However, I
cannot reproduce the issue by this way.
Thanks.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
