[PATCH 0/3] Add OMAP hardware spinlock misc driver

[Catalin Marinas]

Peter Zijlstra <peterz at infradead.org> wrote:
> On Mon, 2010-10-18 at 16:27 +0100, Catalin Marinas wrote:
>> Peter Zijlstra <peterz at infradead.org> wrote:
>> > On Mon, 2010-10-18 at 14:35 +0100, Russell King - ARM Linux wrote:
>> >> In any case, Linux's spinlock API (or more accurately, the ARM exclusive
>> >> access instructions) relies upon hardware coherency support (a piece of
>> >> hardware called an exclusive monitor) which isn't present on the M3 nor
>> >> DSP processors.  So there's no way to ensure that updates from the M3
>> >> and DSP are atomic wrt the A9 updates.
>> >
>> > Right, so the problem is that there simply is no way to do atomic memory
>> > access from these auxiliary processing units wrt the main CPU? Seeing as
>> > they operate on the same memory space, wouldn't it make sense to have
>> > them cache-coherent and thus provide atomicy guarantees through that?
>> 
>> With cache coherency you may get atomicity of writes or reads but
>> usually not atomic modifications.
>
> Sure, but you can 'easily' extend your coherency protocols with support
> for things like ll/sc (or larger transactions).
>
> Have ll bring the cacheline into exclusive state and tag it, then
> anything that demotes the cacheline will clear the tag and make sc fail.

For the ll/sc operations on ARM (exclusive load/store) there is a
per-CPU local exclusive monitor and a (virtual) global one. The global
one may either be a separate piece of hardware or emulated via cache
lines as you said. But if you need synchronisation with a CPU (or DSP)
like Cortex-M3 which doesn't have any built-in caches, you can only get
atomic operations on the main processor (A9) but not on the M3 (as you
can't have a cache line in exclusive state on the M3).

The M3 may have a local exclusive monitor (like the main CPU) but it
isn't cleared by memory accesses from the main CPU.

-- 
Catalin