Re: [PATCH 3/6] bitops: define gen_test_bit() the same way as the rest of functions

[Marco Elver]

On Tue, 7 Jun 2022 at 18:05, Alexander Lobakin
<alexandr.lobakin@xxxxxxxxx> wrote:
>
> From: Marco Elver <elver@xxxxxxxxxx>
> Date: Tue, 7 Jun 2022 15:43:49 +0200
>
> > On Mon, Jun 06, 2022 at 01:49PM +0200, Alexander Lobakin wrote:
> > > Currently, the generic test_bit() function is defined as a one-liner
> > > and in case with constant bitmaps the compiler is unable to optimize
> > > it to a constant. At the same time, gen_test_and_*_bit() are being
> > > optimized pretty good.
> > > Define gen_test_bit() the same way as they are defined.
> > >
> > > Signed-off-by: Alexander Lobakin <alexandr.lobakin@xxxxxxxxx>
> > > ---
> > >  include/asm-generic/bitops/generic-non-atomic.h | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/asm-generic/bitops/generic-non-atomic.h b/include/asm-generic/bitops/generic-non-atomic.h
> > > index 7a60adfa6e7d..202d8a3b40e1 100644
> > > --- a/include/asm-generic/bitops/generic-non-atomic.h
> > > +++ b/include/asm-generic/bitops/generic-non-atomic.h
> > > @@ -118,7 +118,11 @@ gen___test_and_change_bit(unsigned int nr, volatile unsigned long *addr)
> > >  static __always_inline int
> > >  gen_test_bit(unsigned int nr, const volatile unsigned long *addr)
> > >  {
> > > -   return 1UL & (addr[BIT_WORD(nr)] >> (nr & (BITS_PER_LONG-1)));
> > > +   const unsigned long *p = (const unsigned long *)addr + BIT_WORD(nr);
> > > +   unsigned long mask = BIT_MASK(nr);
> > > +   unsigned long val = *p;
> > > +
> > > +   return !!(val & mask);
> >
> > Unfortunately this makes the dereference of 'addr' non-volatile, and
> > effectively weakens test_bit() to the point where I'd no longer consider
> > it atomic. Per atomic_bitops.txt, test_bit() is atomic.
> >
> > The generic version has been using a volatile access to make it atomic
> > (akin to generic READ_ONCE() casting to volatile). The volatile is also
> > the reason the compiler can't optimize much, because volatile forces a
> > real memory access.
>
> Ah-ha, I see now. Thanks for catching and explaining this!
>
> >
> > Yes, confusingly, test_bit() lives in non-atomic.h, and this had caused
> > confusion before, but the decision was made that moving it will cause
> > headaches for ppc so it was left alone:
> > https://lore.kernel.org/all/87a78xgu8o.fsf@xxxxxxxxxxxxxxxxxxxxxxx/T/#u
> >
> > As for how to make test_bit() more compiler-optimization friendly, I'm
> > guessing that test_bit() needs some special casing where even the
> > generic arch_test_bit() is different from the gen_test_bit().
> > gen_test_bit() should probably assert that whatever it is called with
> > can actually be evaluated at compile-time so it is never accidentally
> > used otherwise.
>
> I like the idea! Will do in v2.
> I can move the generics and after, right below them, define
> 'const_*' helpers which will mostly redirect to 'generic_*', but
> for test_bit() it will be a separate function with no `volatile`
> and with an assertion that the input args are constants.

Be aware that there's already a "constant_test_bit()" in
arch/x86/include/asm/bitops.h, which uses 2 versions of test_bit() if
'nr' is constant or not. I guess you can steer clear of that if you
use "const_", but they do sound similar.

> >
> > I would also propose adding a comment close to the deref that test_bit()
> > is atomic and the deref needs to remain volatile, so future people will
> > not try to do the same optimization.
>
> I think that's also the reason why it's not underscored, right?

Yes, the naming convention is that double-underscored ones are
non-atomic so that's one clue indeed. Documentation/atomic_bitops.txt
another, and unlike the other non-atomic bitops, its kernel-doc
comment also does not mention "This operation is non-atomic...".