Re: [PATCH] arm64: fix unreachable code issue with cmpxchg

[Andrew Murray <andrew.murray@arm.com>]

On Tue, Sep 10, 2019 at 11:38:37AM +0200, Arnd Bergmann wrote:
> On Tue, Sep 10, 2019 at 11:23 AM Andrew Murray <andrew.murray@arm.com> wrote:
> 
> >
> > >  arch/arm64/include/asm/cmpxchg.h | 15 ++++++++-------
> > >  1 file changed, 8 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/arch/arm64/include/asm/cmpxchg.h b/arch/arm64/include/asm/cmpxchg.h
> > > index a1398f2f9994..fd64dc8a235f 100644
> > > --- a/arch/arm64/include/asm/cmpxchg.h
> > > +++ b/arch/arm64/include/asm/cmpxchg.h
> > > @@ -19,7 +19,7 @@
> > >   * acquire+release for the latter.
> > >   */
> > >  #define __XCHG_CASE(w, sfx, name, sz, mb, nop_lse, acq, acq_lse, rel, cl)    \
> > > -static inline u##sz __xchg_case_##name##sz(u##sz x, volatile void *ptr)              \
> > > +static __always_inline u##sz __xchg_case_##name##sz(u##sz x, volatile void *ptr)\
> >
> > This hunk isn't needed, there is no BUILD_BUG here.
> 
> Right, I noticed this, but it seemed like a good idea regardless given the small
> size of the function compared with the overhead of a function call.  We clearly
> want these to be inlined all the time.
> 
> Same for the others.

I'm not so sure - isn't the point of something like OPTIMIZE_INLINING to give
more freedom to the tooling (and by virtue of the option - the user)?

Surely any decent optimising compiler will do the right thing by inlining small
trivial functions that are annotated with inline? And if not, the compiler
should be fixed not the kernel - unless of course it causes an issue - and then
we should fix those specific cases.

There must be dozens of trivial functions that are marked with __inline, I
don't think it would make sense to mark those as __always_inline. For example the
atomics in atomic_lse.h are trivial but only marked inline. We obviously want
them inline, though I don't think we should babysit the compiler to do the
right thing.

(Also the commit message implies that all the hunks are required to fix this
particular issue which they are not).

Thanks,

Andrew Murray

> 
> > Alternatively is it possible to replace the BUILD_BUG's with something else?
> >
> > I think because we use BUILD_BUG at the end of a switch statement, we make
> > the assumption that size is known at compile time, for this reason we should
> > ensure the function containing the BUILD_BUG is __always_inline.
> >
> > Looking across the kernel where BUILD_BUG is used as a default in a switch
> > statment ($ git grep -B 3 BUILD_BUG\( | grep default), most instances are
> > within macros, but many are found in an __always_inline function:
> >
> > arch/x86/kvm/cpuid.h
> > mm/kasan/generic.c
> >
> > Though some are not:
> >
> > include/linux/signal.h
> > arch/arm64/include/asm/arm_dsu/pmu.h
> >
> > I wonder if there may be a latent mole ready to whack with pmu.h?
> 
> Right, it can't hurt to annotate those as well. I actually have another
> fixup for linux/signal.h that I would have to revisit at some point.
> See https://bugs.llvm.org/show_bug.cgi?id=38789, I think this is
> fixed with clang-9 now, but maybe not with clang-8.
> 
>       Arnd