On Thu, 24 Mar 2011, Eric Dumazet wrote:
> Le jeudi 24 mars 2011 Ã 13:47 -0500, Christoph Lameter a Ãcrit :
>
> > Hmmm.. Could be. KVM would not really disable interrupts so this may
> > explain that the test case works here.
> >
> > Simple fix would be to do a load before the cli I guess.
> >
>
> Hmm...
>
> If we have a preemption and migration right after this load...
Cannot be the issue here since init_kmem_cache_cpus already
touches the per cpu data. At least if CONFIG_PREEMPT is on.
Is it on?
void init_kmem_cache_cpus(struct kmem_cache *s)
{
#if defined(CONFIG_CMPXCHG_LOCAL) && defined(CONFIG_PREEMPT)
int cpu;
for_each_possible_cpu(cpu)
per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
#endif
}