As it is some callers of synchronize_irq rely on memory barriers
to provide synchronisation against the IRQ handlers.  For example,
the tg3 driver does
	tp->irq_sync = 1;
	smp_mb();
	synchronize_irq();
and then in the IRQ handler:
	if (!tp->irq_sync)
		netif_rx_schedule(dev, &tp->napi);
Unfortunately memory barriers only work well when they come in
pairs.  Because we don't actually have memory barriers on the
IRQ path, the memory barrier before the synchronize_irq() doesn't
actually protect us.
In particular, synchronize_irq() may return followed by the
result of netif_rx_schedule being made visible.
This patch (mostly written by Linus) fixes this by using spin
locks instead of memory barries on the synchronize_irq() path.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
 void synchronize_irq(unsigned int irq)
 {
        struct irq_desc *desc = irq_desc + irq;
+       unsigned int status;
 
        if (irq >= NR_IRQS)
                return;
 
-       while (desc->status & IRQ_INPROGRESS)
-               cpu_relax();
+       do {
+               unsigned long flags;
+
+               /*
+                * Wait until we're out of the critical section.  This might
+                * give the wrong answer due to the lack of memory barriers.
+                */
+               while (desc->status & IRQ_INPROGRESS)
+                       cpu_relax();
+
+               /* Ok, that indicated we're done: double-check carefully. */
+               spin_lock_irqsave(&desc->lock, flags);
+               status = desc->status;
+               spin_unlock_irqrestore(&desc->lock, flags);
+
+               /* Oops, that failed? */
+       } while (status & IRQ_INPROGRESS);
 }
 EXPORT_SYMBOL(synchronize_irq);