Hi Peter,
On Fri, Jan 27, 2023 at 12:18:13PM +0100, Peter Zijlstra wrote:
> On Thu, Jan 26, 2023 at 06:33:54PM +0100, Jules Maselbas wrote:
>
> > @@ -58,9 +61,11 @@ static inline int generic_atomic_fetch_##op(int i, atomic_t *v) \
> > static inline void generic_atomic_##op(int i, atomic_t *v) \
> > { \
> > unsigned long flags; \
> > + int c; \
> > \
> > raw_local_irq_save(flags); \
> > - v->counter = v->counter c_op i; \
> > + c = arch_atomic_read(v); \
> > + arch_atomic_set(v, c c_op i); \
> > raw_local_irq_restore(flags); \
> > }
>
> This and the others like it are a bit sad, it explicitly dis-allows the
> compiler from using memops and forces a load-store.
Good point, I don't know much about atomic memops but this is indeed a
bit sad to prevent such instructions to be used.
> The alternative is writing it like:
>
> *(volatile int *)&v->counter c_op i;
I wonder if it could be possible to write something like:
*(volatile int *)&v->counter += i;
I also noticed that GCC has some builtin/extension to do such things,
__atomic_OP_fetch and __atomic_fetch_OP, but I do not know if this
can be used in the kernel.
Thanks,
-- Jules
