Re: [RFC PATCH 0/8] Rework READ_ONCE() to improve codegen

[Arnd Bergmann]

On Fri, Jan 10, 2020 at 9:15 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Fri, Jan 10, 2020 at 11:47 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> >
> > Isn't the read_barrier_depends() the only reason for actually needing
> > the temporary local variable that must not be volatile?
> >
> > If you make alpha provide its own READ_ONCE() as the first
> > step, it would seem that the rest of the series gets much easier
> > as the others can go back to the simple statement from your
>
> Hmm.. The union still would cause that "take the address of a volatile
> thing on the stack" problem, wouldn't it? And that was what caused
> most of the issues.

Ah, I was missing that there is still the union in smp_load_acquire(),
I only saw that the one in READ_ONCE() is needed only on alpha.

The number of files using smp_load_acquire() is fairly small though,
so we could consider changing it to pass both source and destination
as macro arguments and use typeof(dest) instad of typeof(source)
to avoid the volatile pointer access.

> I think the _real_ issue is how KASAN forces that odd pair of inline
> functions in order to have the annotations on the accesses.

But the inline functions (I assume you mean __write_once_size
and __read_once_size_nocheck?) are completely removed after
Will's series, so those no longer cause harm, right?

      Arnd