Re: [PATCH] kprobes for s390 architecture

[Heiko Carstens]

On Thu, Jun 22, 2006 at 09:36:43AM -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.  

This won't solve anything. What Martin probably meant is something like a poor
man's stop_machine_run() implemented by using smp_call_function(). This way
you synchronize all cpus and when all cpus are in a known state, you change
the instruction in question and make sure that serialization happens before
cpus leave the handler again... Except for the cpu that called
smp_call_function() you get the serialization for free, since the last
instruction of the handler is always an lpsw/lpswe instruction.

Otherwise there is still the possibility that a different cpu is fetching the
instruction concurrently while you change it. This doesn't sound very good,
especially if you take this paragraph of the Principles of Operation into
account (p.5-89 of SA22-7832-04):

"It is possible, if another CPU or a channel program concurrently modifies
the instruction, for one CPU to recognize the changes to some but not all bit
positions of an instruction."
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/