On 06/07/2012 12:32 PM, Steven Rostedt wrote:

On Thu, 2012-06-07 at 12:08 -0700, David Daney wrote:

quoted

quoted

Should it go to mainline stable?

I don't think it is necessary.  As far as I know, RT may be the only
thing that needs it.

Ah, you're right. As this is just an issue because it is called with
interrupts disabled (from stop_machine). Although it's interesting that
the mips code, re-enables interrupts from that function.

 From kernel/cpu.c:

_cpu_down() {
	 __stop_machine(take_cpu_down,&tcd_param, cpumask_of(cpu));

take_cpu_down() {
	err = __cpu_disable();

kernel/stop_machine.c:

__stop_machine(int (*fn)(void *) ...) {
	local_irq_save(flags);
	hard_irq_disable();
	ret = (*fn)(data);
	local_irq_restore(flags);


arch/mips/include/asm/smp.h:

static inline int __cpu_disable(void)
{
         extern struct plat_smp_ops *mp_ops;     /* private */

         return mp_ops->cpu_disable();
}

arch/mips/cavium-octeon/smp.c:

octeon_cpu_disable(void) {
         local_irq_disable();
         fixup_irqs();
         local_irq_enable();

struct plat_smp_ops octeon_smp_ops = {
         .cpu_disable            = octeon_cpu_disable,


Is this expected? It causes the cpu notifiers to be called with
interrupts enabled. Not sure if that's a problem or not.

I am inclined to go with your instinct here.  Probably we shouldn't 
unconditionally local_irq_enable() here.

Perhaps {,raw}_local_irq_save/{,raw}local_irq_restore would be better. 
Or even no local irq enable manipulation...

In any event, I may let Ralf sort it out.

David Daney