[PATCH 1/3] ARM: Introduce atomic MMIO clear/set

[ezequiel.garcia@free-electrons.com (Ezequiel Garcia)]

On Tue, Aug 20, 2013 at 09:32:13AM -0500, Matt Sealey wrote:
> On Mon, Aug 19, 2013 at 11:59 AM, Ezequiel Garcia
> <ezequiel.garcia@free-electrons.com> wrote:
> > On Mon, Aug 12, 2013 at 07:29:42PM +0100, Will Deacon wrote:
> >
> >> This means that you don't have ordering guarantees between the two accesses
> >> outside of the CPU, potentially giving you:
> >>
> >>       spin_lock(&__io_lock);
> >>       spin_unlock(&__io_lock);
> >>       writel((readl(reg) & ~clear) | set, reg);
> >>
> >> which is probably not what you want.
> >>
> >> I suggest adding an iowmb after the writel if you really need this ordering
> >> to be enforced (but this may have a significant performance impact,
> >> depending on your SoC).
> >
> > I don't want to argue with you, given I have zero knowledge about this
> > ordering issue. However let me ask you a question.
> >
> > In arch/arm/include/asm/spinlock.h I'm seeing this comment:
> >
> > ""ARMv6 ticket-based spin-locking.
> > A memory barrier is required after we get a lock, and before we
> > release it, because V6 CPUs are assumed to have weakly ordered
> > memory.""
> >
> > and also:
> >
> > static inline void arch_spin_unlock(arch_spinlock_t *lock)
> > {
> >         smp_mb();
> >         lock->tickets.owner++;
> >         dsb_sev();
> > }
> >
> > So, knowing this atomic API should work for every ARMv{N}, and not being very
> > sure what the call to dsb_sev() does. Would you care to explain how the above
> > is *not* enough to guarantee a memory barrier before the spin unlocking?
> 
> arch_spin_[un]lock as an API is not guaranteed to use a barrier before
> or after doing anything, even if this particular implementation does.
> 
> dsb_sev() is an SMP helper which does a synchronization barrier and
> then sends events to other CPUs which informs them of the unlock. If
> the other CPUs were in WFE state waiting on that spinlock, they can
> now thunder in and attempt to lock it themselves. It's not really
> relevant to the discussion as arch_spin_unlock is not guaranteed to
> perform a barrier before returning.
> 
> On some other architecture there may be ISA additions which make
> locking barrier-less, or on a specific implementation of an ARM
> architecture SoC whereby there may be a bank of "hardware spinlocks"
> available.
> 
> So, in this sense, you shouldn't rely on implementation-specific
> behaviors of a function. If you need to be sure C follows B follows A,
> insert a barrier yourself. Don't expect A to barrier for you just
> because you saw some source code that does it today, as tomorrow it
> may be different. It's not an optimization, just a potential source of
> new bugs if the implementation changes underneath you later.
> 

Of course. I agree completely.

Thanks a lot,
-- 
Ezequiel Garc?a, Free Electrons
Embedded Linux, Kernel and Android Engineering
http://free-electrons.com