writel(value, &io_apic->data);
 }
 
+static int io_apic_level_ack_pending(unsigned int irq)
+{
+       struct irq_pin_list *entry;
+       unsigned long flags;
+       int pending = 0;
+
+       spin_lock_irqsave(&ioapic_lock, flags);
+       entry = irq_2_pin + irq;
+       for (;;) {
+               unsigned int reg;
+               int pin;
+
+               pin = entry->pin;
+               if (pin == -1)
+                       break;
+               reg = io_apic_read(entry->apic, 0x10 + pin*2);
+               /* Is the remote IRR bit set? */
+               pending |= (reg >> 14) & 1;
+               if (!entry->next)
+                       break;
+               entry = irq_2_pin + entry->next;
+       }
+       spin_unlock_irqrestore(&ioapic_lock, flags);
+       return pending;
+}
+
 /*
  * Synchronize the IO-APIC and the CPU by doing
  * a dummy read from the IO-APIC
        ack_APIC_irq();
 
        /* Now we can move and renable the irq */
-       move_masked_irq(irq);
-       if (unlikely(do_unmask_irq))
+       if (unlikely(do_unmask_irq)) {
+               /* Only migrate the irq if the ack has been received.
+                *
+                * On rare occasions the broadcast level triggered ack gets
+                * delayed going to ioapics, and if we reprogram the
+                * vector while Remote IRR is still set the irq will never
+                * fire again.
+                *
+                * To prevent this scenario we read the Remote IRR bit
+                * of the ioapic.  This has two effects.
+                * - On any sane system the read of the ioapic will
+                *   flush writes (and acks) going to the ioapic from
+                *   this cpu.
+                * - We get to see if the ACK has actually been delivered.
+                *
+                * Based on failed experiments of reprogramming the
+                * ioapic entry from outside of irq context starting
+                * with masking the ioapic entry and then polling until
+                * Remote IRR was clear before reprogramming the
+                * ioapic I don't trust the Remote IRR bit to be
+                * completey accurate.
+                *
+                * However there appears to be no other way to plug
+                * this race, so if the Remote IRR bit is not
+                * accurate and is causing problems then it is a hardware bug
+                * and you can go talk to the chipset vendor about it.
+                */
+               if (!io_apic_level_ack_pending(irq))
+                       move_masked_irq(irq);
                unmask_IO_APIC_irq(irq);
+       }
 }
 
 static struct irq_chip ioapic_chip __read_mostly = {