On Tue, Jun 26, 2018 at 10:30 PM, Peter Zijlstra <pet...@infradead.org> wrote: > On Tue, Jun 26, 2018 at 09:53:18PM +0300, Andy Shevchenko wrote: >> On Mon, Jun 25, 2018 at 1:59 PM, Mark Rutland <mark.rutl...@arm.com> wrote: >> > From: Peter Zijlstra <pet...@infradead.org> >> > >> > Mark found ldsem_cmpxchg() needed an (atomic_long_t *) cast to keep >> > working after making the atomic_long interface type safe. >> > >> > Needing casts is bad form, which made me look at the code. There are no >> > ld_semaphore::count users outside of these functions so there is no >> > reason why it can not be an atomic_long_t in the first place, obviating >> > the need for this cast. >> > >> > That also ensures the loads use atomic_long_read(), which implies (at >> > least) READ_ONCE() in order to guarantee single-copy-atomic loads. >> > >> > When using atomic_long_try_cmpxchg() the ldsem_cmpxchg() wrapper gets >> > very thin (the only difference is not changing *old on success, which >> > most callers don't seem to care about). >> > >> > So rework the whole thing to use atomic_long_t and its accessors >> > directly. >> > >> > While there, fixup all the horrible comment styles. >> >> >> > - ldsem_atomic_update(-LDSEM_WAIT_BIAS, sem); >> > + atomic_long_add_return(-LDSEM_WAIT_BIAS, &sem->count); >> >> I suppose it's simple atomic_long_add() here? > > Different ordering rules for those two. I didn't look hard enough to see > if that mattered here.
Indeed. So, to follow semantics it would be something like smp_mb__before_atomic(); atomic_long_add_relaxed(); smp_mb__after_atomic(); though I do not dare to convert that way (as I understood the simple atomic_long_add_return() variant might be implemented better). -- With Best Regards, Andy Shevchenko