Interrupt synchronisation questions for SMP experts

[Guillaume Knispel <gknispel@proformatique.com>]

Hi,

1. Imagine an interrupt storm from an evil device on an evil SMP
machine, is there any guarantee that if the ISR call
disable_irq_nosync() on its own irq it won't be called again just after
it returns?  I feel this could be a reasonable expectation while i fail
to see anything that could provide it in the for (;;) loop of
__do_IRQ().

2. Does using synchronizing functions like disable_irq(),
tasklet_kill(), or del_timer_sync() provide sequentiality between the
effects of the end of the interrupt servicing code and the effects of
the code immediately after the call to the synchronizing function?

3. Consider this:

(a initially set to 0)
  CPU 1.          CPU 2.

  z = 3;          while (!a) barrier();
  smp_wmb();      smp_mb();
  a = 1;          z = 42;

Does the semantic of smp_mb() and smp_wmb() guarantee that the final
value of z will be 42?  (This has something to do with 2.)

Cheers,
Guillaume KNISPEL