Re: [PATCH] kprobes for s390 architecture

[Jan Glauber <jan.glauber@de.ibm.com>]

On Thu, 2006-06-22 at 09:36 -0700, Mike Grundy wrote:
> On Thu, Jun 22, 2006 at 01:28:36PM +0200, Jan Glauber wrote:
> > On Wed, 2006-06-21 at 10:34 -0700, Mike Grundy wrote:
> > > On Wed, Jun 21, 2006 at 06:38:40PM +0200, Martin Schwidefsky wrote:
> > > > You misunderstood me here. I'm not talking about storing the same piece
> > > > of data to memory on each processor. I'm talking about isolating all
> > > > other cpus so that the initiating cpu can store the breakpoint to memory
> > > > without running into the danger that another cpu is trying to execute it
> > > > at the same time. But probably the store should be atomic in regard to
> > > > instruction fetching on the other cpus. It is only two bytes and it
> > > > should be aligned.
> > 
> > Preemption disabling is not necessary around smp_call_function(), since
> > smp_call_function() takes a spin lock. But smp_call_function() is wrong
> > here, it calls the code on all other CPUs but not on our own. Please use
> > on_each_cpu() instead.
> 
> But on_each_cpu() does:
> 
>         preempt_disable();
>         ret = smp_call_function(func, info, retry, wait);
>         local_irq_disable();
>         func(info);
>         local_irq_enable();
>         preempt_enable();
>  
> I'm confused. I really don't need to swap the instruction on each cpu. I really
> need to make sure each cpu is not fetching that instruction while I change it.
> s390 doesn't have a flush_icache_range() (which the other arches use after the 
> swap). I thought that the synchronization that smp_call_function() does was the
> primary reason for using it here, not repeatedly changing the same area of 
> memory.  If you'd prefer I use on_each_cpu() instead of smp_call_function(), 
> no problem.  

If I'm not completely off-track you _do_ swap the instruction on all
other CPUs with the smp_call_function(). But since we don't have a
flush_icache_range() interface on s390 we must understand how the
instruction cache works and then we will know whether we need the smp
call at all.