Re: [PATCH]kprobes fix bug when probed on task and isr functions

[Andrew Morton <akpm@osdl.org>]

Keshavamurthy Anil S <anil.s.keshavamurthy@intel.com> wrote:
>
> > >  void __kprobes lock_kprobes(void)
> > >  {
> > > +	unsigned long flags = 0;
> > > +
> > > +	local_irq_save(flags);
> > >  	spin_lock(&kprobe_lock);
> > >  	kprobe_cpu = smp_processor_id();
> > > + 	local_irq_restore(flags);
> > >  }
> > 
> > what is this change trying to do?  If a lock is taken from both process and
> > irq contexts then local IRQs must be disabled for the entire period when the
> > lock is held, not just for a little blip like this.  If IRQ-context code is
> > running this function then the code is deadlockable.
> 
> In the kprobe exception handling we relay on kprobe_cpu = smp_processor_id() to determine
> whether we are inside the kprobe or not. It was so happeing that when we
> take the lock and before kprobe_cpu gets updated if an H/W interrupt happens
> and if kprobe is enabled on ISR routine, then in the kprobe execption handler
> for isr, we miss the indication that we are already in kprobes(since interrupt
> happened before we get to update kprobe_cpu) and we were trying to 
> take the lock again and there by causing the deadlock. This deadlock is avoided
> by disabling the ISR for a short period while we take the spin_lock() and update
> the kprobe_cpu.

OK.

Are you sure that other CPUs can safely read kprobe_cpu without taking the
lock?  I don't see any memory barriers in there...

> > 
> > Now, probably there's deep magic happening here and I'm wrong.  If so then
> > please explain the code's magic via a comment patch so the question doesn't
> > arise again, thanks.
> > 
> 
> This whole serialization will go away when we introduce the scalability patch.

Yes, it does look rather unscalable.