Re: [RFC][PATCHv2 1/8] printk: move printk_pending out of per-cpu

[Petr Mladek <pmladek@suse.com>]

On Mon 2017-04-03 20:23:01, Sergey Senozhatsky wrote:
> On (03/31/17 15:33), Peter Zijlstra wrote:
> > On Fri, Mar 31, 2017 at 03:09:50PM +0200, Petr Mladek wrote:
> > > On Wed 2017-03-29 18:25:04, Sergey Senozhatsky wrote:
> > 
> > > >  	if (waitqueue_active(&log_wait)) {
> > > > -		this_cpu_or(printk_pending, PRINTK_PENDING_WAKEUP);
> > > > +		set_bit(PRINTK_PENDING_WAKEUP, &printk_pending);
> > > 
> > > We should add here a write barrier:
> > > 
> > > 	/*
> > > 	 * irq_work_queue() uses cmpxchg() and implies the memory
> > > 	 * barrier only when the work is queued. An explicit barrier
> > > 	 * is needed here to make sure that wake_up_klogd_work_func()
> > > 	 * sees printk_pending set even when the work was already queued
> > > 	 * because of an other pending event.
> > > 	 */
> > > 	 smp_wmb();
> > > 
> > > >  		irq_work_queue(this_cpu_ptr(&wake_up_klogd_work));
> > > >  	}
> > > >  	preempt_enable();
> > 
> > smp_mb__after_atomic() is probably better, because if you're not
> > ordering with the cmpxchg, you're ordering against a load done by
> > cmpxchg to see it doesn't need to do anything.
> 
> Petr and Peter, thanks for the review.
> 
> can you educate me, what exactly is broken there?

Good point!

> when called from console_unlock(), we have something as follows
> 
> 	console_unlock()
> 	{
> 		for (;;) {
> 			spin_lock_irqsave();
> 			...
> 			spin_unlock_irqrestore();
> 			...
> 		}
> 
> 		spin_unlock_irqrestore();
> 
> <<IRQs enabled>>
> 
> 		if (wake_klogd)
> 			wake_up_klogd()
> 			{
> 				set_bit(PRINTK_PENDING_WAKEUP, &printk_pending);
> 				irq_work_queue(this_cpu_ptr(&wake_up_klogd_work));
> 			}
> 	}
> 
> 
> we queue a per-CPU irq_work.

Ah, I forgot that irq_work is still per-CPU. In this case, everything
seems to be safe even without the barrier. The important thing is that
there always will be queued an irq_work that will see and handle the
bit. I believe that the barrier would be needed if the irq_work was
global.

I am sorry for the noise.

Best Regards,
Petr