Re: [patch] big IRQ lock removal, 2.5.27-G0

[Ingo Molnar]

On Wed, 24 Jul 2002, Oleg Nesterov wrote:

> > okay - i've fixed irq_exit() once more in -G4
> 
> found G5, your forgot to add preempt_disable() in irq_exit()
> 
>  #define irq_exit()
>  do {
> +               preempt_disable();
>                 preempt_count() -= IRQ_EXIT_OFFSET;
>                 if (!in_interrupt() && softirq_pending(smp_processor_id()))
> 
> Oleg.

nope, it's tricky, check out the define of IRQ_EXIT_OFFSET, it has the
plus 1 count added already.

using preempt_disable() has the problem of putting a barrier() between the
++ and the -IRQ_OFFSET, causing one more instruction to be generated.

but there was another (not too critical) bug here - irq_enter() needs to
have a barrier() after manipulating the preemption count, irq_exit() needs
to have a barrier() before. Fixed in my tree.

	Ingo