Re: [patch 04/19] mutex subsystem, add include/asm-i386/mutex.h

[Ingo Molnar <mingo@elte.hu>]

* Nick Piggin <nickpiggin@yahoo.com.au> wrote:

> >+#ifdef __HAVE_ARCH_CMPXCHG
> >+	if (likely(atomic_cmpxchg(count, 1, 0)) == 1)
> >+		return 1;
> >+	return 0;
> >+#else
> >+	return fail_fn(count);
> >+#endif
> >+}
> 
> asm-i386 version I think really should just use atomic_cmpxchg 
> unconditionally, because otherwise an i386 compatible kernel will not 
> use cmpxchg even when running on 486+ (not sure how important that is 
> these days, but still...).

yeah. This code predates the generic-atomic-cmpxchg code. (Feel free to 
review all __HAVE_ARCH_CMPXCHG users after this goes in, and remove 
__HAVE_ARCH_CMPXCHG altogether with a CONFIG_ flag.)

but ... the spinlock based variant is quite likely faster (on i386) than 
generic-cmpxchg. Couldnt we introduce a new API, something along the 
lines of:

	atomic_cmpxchg_lock(&count, &lock->wait_lock);

and if the cmpxchg fails, it would hold the spinlock? The cmpxchg 
semantics could be guaranteed by the spinlock. (because it is 'global' 
for that particular critical section)

	Ingo