[BUG] 2.6.37-rc3 massive interactivity regression on ARM

[Eric Dumazet]

Le vendredi 10 décembre 2010 à 18:49 +0100, Peter Zijlstra a écrit :
> On Fri, 2010-12-10 at 18:18 +0100, Eric Dumazet wrote:
> > Le vendredi 10 décembre 2010 à 14:47 +0100, Peter Zijlstra a écrit :
> 
> > Also irq_time_write_begin() and irq_time_write_end() could be faster
> > (called for current cpu)
> > 
> > static inline void irq_time_write_begin(void)
> > {
> > 	__this_cpu_inc(irq_time_seq.sequence);
> > 	smp_wmb();
> > }
> >  
> > static inline void irq_time_write_end(void)
> > {
> > 	smp_wmb();
> > 	__this_cpu_inc(irq_time_seq.sequence);
> > }
> 
> Yeah, but that kinda defeats the purpose of having it implemented in
> seqlock.h. Ideally we'd teach gcc about these long pointers and have
> something like: 
> 
>   write_seqcount_begin(&this_cpu_read(irq_time_seq));
> 
> do the right thing.

gcc wont be able to do this yet (%fs/%gs selectors)

But we can provide this_cpu_write_seqcount_{begin|end}()

static inline void this_cpu_write_seqcount_begin(seqcount_t *s)
{
	__this_cpu_inc(s->sequence);
	smp_wmb();
}