On Thu, 26 Oct 2017 13:58:38 +1100
"Tobin C. Harding" <[email protected]> wrote:
> > +static bool have_filled_random_ptr_key;
> > +static siphash_key_t ptr_key __read_mostly;
> > +
> > +static void fill_random_ptr_key(struct random_ready_callback *unused)
> > +{
> > + get_random_bytes(&ptr_key, sizeof(ptr_key));
> > + WRITE_ONCE(have_filled_random_ptr_key, true);
>
> This usage of WRITE_ONCE was suggested by Jason A. Donenfeld. I read
> include/linux/compiler.h but was not able to grok it. Is this enough to
> stop the compiler re-ordering these two statements?
>
> Or do I need to read Documentation/memory-barriers.txt [again]?
No, the WRITE_ONCE does not stop the compiler from reordering those
statements. If you need that, then you need to do:
get_random_bytes(&ptr_key, sizeof(ptr_key));
barrier();
WRITE_ONCE(have_filled_random_ptr_key, true);
and that only works against interrupts. If you need synchronization
across CPUs, then you need smp_mb().
-- Steve
