Re: [PATCH] synchronize_irq needs a barrier

[Nick Piggin <nickpiggin@yahoo.com.au>]

On Friday 19 October 2007 13:28, Herbert Xu wrote:
> Nick Piggin <nickpiggin@yahoo.com.au> wrote:
> >> First of all let's agree on some basic assumptions:
> >>
> >> * A pair of spin lock/unlock subsumes the effect of a full mb.
> >
> > Not unless you mean a pair of spin lock/unlock as in
> > 2 spin lock/unlock pairs (4 operations).
> >
> > *X = 10;
> > spin_lock(&lock);
> > /* *Y speculatively loaded here */
> > /* store to *X leaves CPU store queue here */
> > spin_unlock(&lock);
> > y = *Y;
>
> Good point.
>
> Although in this case we're still safe because in the worst
> cases:
>
> CPU0				CPU1
> irq_sync = 1
> synchronize_irq
> 	spin lock
> 	load IRQ_INPROGRESS
> irq_sync sync is visible
> 	spin unlock
> 				spin lock
> 					load irq_sync
> 	while (IRQ_INPROGRESS)
> 		wait
> 	return
> 				set IRQ_INPROGRESS
> 				spin unlock
> 				tg3_msi
> 					ack IRQ
> 					if (irq_sync)
> 						return
> 				spin lock
> 				clear IRQ_INPROGRESS
> 				spin unlock
>
> ------------------------------------------------------------
>
> CPU0				CPU1
> 				spin lock
> 					load irq_sync
> irq_sync = 1
> synchronize_irq
> 				set IRQ_INPROGRESS
> 				spin unlock
> 	spin lock
> 	load IRQ_INPROGRESS
> irq_sync sync is visible
> 	spin unlock
> 	while (IRQ_INPROGRESS)
> 		wait
> 				tg3_msi
> 					ack IRQ
> 					if (irq_sync)
> 						return
> 					do work
> 				spin lock
> 				clear IRQ_INPROGRESS
> 				spin unlock
> 	return
>
> So because we're using the same lock on both sides, it does
> do the right thing even without the memory barrier.


Yeah, if you've got the lock on both sides there, then I
agree it will be correct.