Re: [PATCH 2/5] atomic: introduce atomic_inc_not_zero

[Andrew Morton <akpm@osdl.org>]

Roman Zippel <zippel@linux-m68k.org> wrote:
>
> Hi,
> 
> On Thu, 15 Sep 2005, Nick Piggin wrote:
> 
> > Roman: any ideas about what you would prefer? You'll notice
> > atomic_inc_not_zero replaces rcuref_inc_lf, which is used several times
> > in the VFS.
> 
> In the larger picture I'm not completely happy with these scalibilty 
> patches, as they add extra overhead at the lower end. On a UP system in 
> general nothing beats:
> 
> 	spin_lock();
> 	if (*ptr)
> 		ptr += 1;
> 	spin_unlock();
> 
> The main problem is here that the atomic functions are used in two basic 
> situation:
> 
> 1) interrupt synchronization
> 2) multiprocessor synchronization
> 
> The atomic functions have to assume both, but on UP systems it often is 
> a lot cheaper if they don't have to synchronize with interrupts. So 
> replacing a spinlock with a few atomic operations can hurt UP performance.
> 

Nope.  On uniprocessor systems, atomic_foo() doesn't actually do the
buslocked atomic thing.

#ifdef CONFIG_SMP
#define LOCK "lock ; "
#else
#define LOCK ""
#endif

On x86, at least.  Other architectures can do the same thing if they have
an atomic-wrt-IRQs add and sub.