On Tue, 2 Aug 2022, Will Deacon wrote:
> On Tue, Aug 02, 2022 at 07:38:17AM -0400, Mikulas Patocka wrote:
> >
> >
> > On Tue, 2 Aug 2022, Will Deacon wrote:
> >
> > > On Mon, Aug 01, 2022 at 12:12:47PM -0400, Mikulas Patocka wrote:
> > > > On Mon, 1 Aug 2022, Will Deacon wrote:
> > > > > On Mon, Aug 01, 2022 at 06:42:15AM -0400, Mikulas Patocka wrote:
> > > > >
> > > > > > Index: linux-2.6/arch/x86/include/asm/bitops.h
> > > > > > ===================================================================
> > > > > > --- linux-2.6.orig/arch/x86/include/asm/bitops.h 2022-08-01 12:27:43.000000000 +0200
> > > > > > +++ linux-2.6/arch/x86/include/asm/bitops.h 2022-08-01 12:27:43.000000000 +0200
> > > > > > @@ -203,8 +203,10 @@ arch_test_and_change_bit(long nr, volati
> > > > > >
> > > > > > static __always_inline bool constant_test_bit(long nr, const volatile unsigned long *addr)
> > > > > > {
> > > > > > - return ((1UL << (nr & (BITS_PER_LONG-1))) &
> > > > > > + bool r = ((1UL << (nr & (BITS_PER_LONG-1))) &
> > > > > > (addr[nr >> _BITOPS_LONG_SHIFT])) != 0;
> > > > > > + barrier();
> > > > > > + return r;
> > > > >
> > > > > Hmm, I find it a bit weird to have a barrier() here given that 'addr' is
> > > > > volatile and we don't need a barrier() like this in the definition of
> > > > > READ_ONCE(), for example.
> > > >
> > > > gcc doesn't reorder two volatile accesses, but it can reorder non-volatile
> > > > accesses around volatile accesses.
> > > >
> > > > The purpose of the compiler barrier is to make sure that the non-volatile
> > > > accesses that follow test_bit are not reordered by the compiler before the
> > > > volatile access to addr.
> > >
> > > If we need these accesses to be ordered reliably, then we need a CPU barrier
> > > and that will additionally prevent the compiler reordering. So I still don't
> > > think we need the barrier() here.
> >
> > This is x86-specific code. x86 has strong memory ordering, so we only care
> > about compiler reordering.
>
> Indeed, but what I'm trying to say is that the _caller_ would have a memory
> barrier in this case, and so there's no need for one in here. test_bit() does
> not have ordering semantics.
>
> Will
But constant_test_bit() is also used for test_bit_acquire(), and for
test_bit_acquire(), the barrier is needed. Without the barrier, it doesn't
have acquire semantics, because the compiler (not CPU) can move the
following non-volatile accesses before the volatile access to "addr[nr >>
_BITOPS_LONG_SHIFT]".
See this piece of code in arch/x86/include/asm/bitops.h in the patch:
+#define arch_test_bit_acquire(nr, addr) \
+ arch_test_bit(nr, addr)
We could split constant_test_bit() to two functions: constant_test_bit()
and constant_test_bit_acquire() and put the barrier only in
constant_test_bit_acquire(). But I chose not to do it because code
duplication is bad and because the overhead of the compiler barrier is
none.
Mikulas
