[RFC PATCH v2 11/31] kvx: Add atomic/locking headers

[Jules Maselbas]

Hi Mark,

On Fri, Jan 20, 2023 at 03:18:48PM +0000, Mark Rutland wrote:
> On Fri, Jan 20, 2023 at 03:09:42PM +0100, Yann Sionneau wrote:
> > Add common headers (atomic, bitops, barrier and locking) for basic
> > kvx support.
> > 
> > Co-developed-by: Clement Leger <clement at clement-leger.fr>
> > Signed-off-by: Clement Leger <clement at clement-leger.fr>
> > Co-developed-by: Jules Maselbas <jmaselbas at kalray.eu>
> > Signed-off-by: Jules Maselbas <jmaselbas at kalray.eu>
> > Co-developed-by: Julian Vetter <jvetter at kalray.eu>
> > Signed-off-by: Julian Vetter <jvetter at kalray.eu>
> > Co-developed-by: Julien Villette <jvillette at kalray.eu>
> > Signed-off-by: Julien Villette <jvillette at kalray.eu>
> > Co-developed-by: Yann Sionneau <ysionneau at kalray.eu>
> > Signed-off-by: Yann Sionneau <ysionneau at kalray.eu>
> > ---
> > 
> > Notes:
> >     V1 -> V2:
> >      - use {READ,WRITE}_ONCE for arch_atomic64_{read,set}
> >      - use asm-generic/bitops/atomic.h instead of __test_and_*_bit
> >      - removed duplicated includes
> >      - rewrite xchg and cmpxchg in C using builtins for acswap insn
> 
> Thanks for those changes. I see one issue below (instantiated a few times), but
> other than that this looks good to me.
> 
> [...]
> 
> > +#define ATOMIC64_RETURN_OP(op, c_op)					\
> > +static inline long arch_atomic64_##op##_return(long i, atomic64_t *v)	\
> > +{									\
> > +	long new, old, ret;						\
> > +									\
> > +	do {								\
> > +		old = v->counter;					\
> 
> This should be arch_atomic64_read(v), in order to avoid the potential for the
> compiler to replay the access and introduce ABA races and other such problems.
Thanks for the suggestion, this will be into v3.

> For details, see:
> 
>   https://lore.kernel.org/lkml/Y70SWXHDmOc3RhMd@osiris/
>   https://lore.kernel.org/lkml/Y71LoCIl+IFdy9D8@FVFF77S0Q05N/
> 
> I see that the generic 32-bit atomic code suffers from that issue, and we
> should fix it.
I took a look at the generic 32-bit atomic, but I am unsure if this
needs to be done for both the SMP and non-SMP implementations. But I
can send a first patch and we can discuss from there.

> > +		new = old c_op i;					\
> > +		ret = arch_cmpxchg(&v->counter, old, new);		\
> > +	} while (ret != old);						\
> > +									\
> > +	return new;							\
> > +}
> > +
> > +#define ATOMIC64_OP(op, c_op)						\
> > +static inline void arch_atomic64_##op(long i, atomic64_t *v)		\
> > +{									\
> > +	long new, old, ret;						\
> > +									\
> > +	do {								\
> > +		old = v->counter;					\
> 
> Likewise, arch_atomic64_read(v) here.
ack

> > +		new = old c_op i;					\
> > +		ret = arch_cmpxchg(&v->counter, old, new);		\
> > +	} while (ret != old);						\
> > +}
> > +
> > +#define ATOMIC64_FETCH_OP(op, c_op)					\
> > +static inline long arch_atomic64_fetch_##op(long i, atomic64_t *v)	\
> > +{									\
> > +	long new, old, ret;						\
> > +									\
> > +	do {								\
> > +		old = v->counter;					\
> 
> Likewise, arch_atomic64_read(v) here.
ack

> > +		new = old c_op i;					\
> > +		ret = arch_cmpxchg(&v->counter, old, new);		\
> > +	} while (ret != old);						\
> > +									\
> > +	return old;							\
> > +}
> > +
> > +#define ATOMIC64_OPS(op, c_op)						\
> > +	ATOMIC64_OP(op, c_op)						\
> > +	ATOMIC64_RETURN_OP(op, c_op)					\
> > +	ATOMIC64_FETCH_OP(op, c_op)
> > +
> > +ATOMIC64_OPS(and, &)
> > +ATOMIC64_OPS(or, |)
> > +ATOMIC64_OPS(xor, ^)
> > +ATOMIC64_OPS(add, +)
> > +ATOMIC64_OPS(sub, -)
> > +
> > +#undef ATOMIC64_OPS
> > +#undef ATOMIC64_FETCH_OP
> > +#undef ATOMIC64_OP
> > +
> > +static inline int arch_atomic_add_return(int i, atomic_t *v)
> > +{
> > +	int new, old, ret;
> > +
> > +	do {
> > +		old = v->counter;
> 
> Likewise, arch_atomic64_read(v) here.
ack, this will bt arch_atomic_read(v) here since this is not atomic64_t
here.


Thanks
-- Jules