Re: [PATCH 3/6] locking/atomic/x86: Use READ_ONCE before atomic{,64}_try_cmpxchg loops

[Uros Bizjak]

On Tue, Apr 9, 2024 at 1:07 PM Mark Rutland <mark.rutland@xxxxxxx> wrote:
>
> On Tue, Apr 09, 2024 at 12:03:54PM +0200, Uros Bizjak wrote:
> > The value preload before the cmpxchg loop does not need to be atomic,
> > but should use READ_ONCE to prevent compiler from merging, refetching
> > or reordering the read.
> >
>
> Yes, and that's what arch_atomic_read() and arch_atomic64_read() do...
>
> > This patch unifies arch_atomic{,64}_{,fetch}_{and,or,xor}() macros
> > between x86_32 and x86_64 targets.
> >
> > No functional changes intended.
> >
> > Signed-off-by: Uros Bizjak <ubizjak@xxxxxxxxx>
> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > Cc: Borislav Petkov <bp@xxxxxxxxx>
> > Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> > Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > ---
> >  arch/x86/include/asm/atomic.h      |  8 ++++----
> >  arch/x86/include/asm/atomic64_64.h | 20 ++++++++++----------
> >  2 files changed, 14 insertions(+), 14 deletions(-)
> >
> > diff --git a/arch/x86/include/asm/atomic.h b/arch/x86/include/asm/atomic.h
> > index 55a55ec04350..b166da21ee98 100644
> > --- a/arch/x86/include/asm/atomic.h
> > +++ b/arch/x86/include/asm/atomic.h
> > @@ -20,7 +20,7 @@ static __always_inline int arch_atomic_read(const atomic_t *v)
> >        * Note for KASAN: we deliberately don't use READ_ONCE_NOCHECK() here,
> >        * it's non-inlined function that increases binary size and stack usage.
> >        */
> > -     return __READ_ONCE((v)->counter);
> > +     return __READ_ONCE(v->counter);
>
> Removing the unncessary brackets is fine, but the commit message didn't mention this.

It was just a change in passing. I didn't think it even needed a comment.

> [...]
>
> >  static __always_inline int arch_atomic_fetch_and(int i, atomic_t *v)
> >  {
> > -     int val = arch_atomic_read(v);
> > +     int val = __READ_ONCE(v->counter);
>
> This is the wrong thing to do; arch_atomic_read() already has the required
> semantic, and it more clearly aligns with the use of arch_atomic_try_cmpxchg()
> below. It contains the documentation regarding why we use __READ_ONCE()
> specifically (which we should probably note in arch_atomic64_read()).
>
> Please leave this as-is, and likewise for the other cases below. Similarly, the
> prior patch should use arch_atomic{,_64}_read() rather than using
> __READ_ONCE().

Please note that arch_atomic64_read implements true 64-bit atomic read
on x86_32. I tried to bypass this using __READ_ONCE(), but your
suggestion to use arch_atomic64_read_tearable() is indeed a much
better approach.

> [...]
>
> >  static __always_inline s64 arch_atomic64_fetch_and(s64 i, atomic64_t *v)
> >  {
> > -     s64 val = arch_atomic64_read(v);
> > +     s64 val = __READ_ONCE(v->counter);
> > +
> > +     do { } while (!arch_atomic64_try_cmpxchg(v, &val, val & i));
> >
> > -     do {
> > -     } while (!arch_atomic64_try_cmpxchg(v, &val, val & i));
> >       return val;
>
> I think this reformatting is what you meant in the commit message when you said:
>
> | This patch unifies arch_atomic{,64}_{,fetch}_{and,or,xor}() macros
> | between x86_32 and x86_64 targets.

Actually, it was the change to use __READ_ONCE(). The reformatting
follows the functions from atomic.h, e.g. arch_atomic_fetch_and() and
was another case of change in passing, not worth mentioning in the
commit message. I will rewrite this in v2 of the patch, so these
functions will uniformly use arch_atomic64_read_tearable().

> Assuming so, can you please jsut do that, and say:
>
>   This patch reformats the x86_64 arch_atomic{,64}_{,fetch}_{and,or,xor}()
>   functions to match the x86_32 versions.

Thanks,
Uros.