Re: [RFC Patch 1/10] Introducing generic hardware breakpointhandler interfaces

From: Paul E. McKenney
Date: Sun Feb 01 2009 - 08:55:21 EST


On Fri, Jan 30, 2009 at 10:55:39AM -0500, Alan Stern wrote:
> On Fri, 30 Jan 2009, K.Prasad wrote:
>
> > > A few RCU-related questions below.
> > >
> > > Thanx, Paul
>
> Paul, you've got to learn to trim your replies! It's not nice to have
> to skim over hundreds and hundreds lines of quoted text while searching
> for your interpolated comments. In fact, the phrase "needle in a
> haystack" springs to mind...

I should have said "search for empty lines", but yes, I should have
trimmed a bit. My apologies!!!

> > > > + thr_kbpdata = chbi->cur_kbpdata;
> > > > + barrier();
> > >
> > > Couldn't the above two lines instead be:
> > >
> > > thr_kbpdata = ACCESS_ONCE(chbi->cur_kbpdata);
> > >
> > > This would prevent the pointer aliasing, but would make it very clear
> > > exactly how the compiler was to be restricted.
> >
> > Ok. Using a barrier() could be an overkill. I will change it.
>
> IIRC, the original code above was written before ACCESS_ONCE came into
> being. But I could be wrong about that...

Could well be, ACCESS_ONCE() showed up in 2.6.24, and moved out of
rcupdate.h a couple of releases later.

> > > > +/*
> > > > + * Tell all CPUs to update their debug registers.
> > > > + *
> > > > + * The caller must hold hw_breakpoint_mutex.
> > > > + */
> > > > +static void update_all_cpus(void)
> > > > +{
> > > > + /* We don't need to use any sort of memory barrier. The IPI
> > > > + * carried out by on_each_cpu() includes its own barriers.
> > > > + */
> > > > + on_each_cpu(update_this_cpu, NULL, 0);
> > > > + synchronize_rcu();
> > >
> > > Don't we need the rcu_read_lock() / rcu_read_unlock() pair from
> > > load_debug_registers() to move down into update_this_cpu() in order
> > > for this to be guaranteed to work? As the code reads now, the
> > > update_this_cpu() calls running on other CPUs are not running under
> > > RCU protection, right?
>
> Maybe I'm misunderstanding the question. update_this_cpu() is called
> from only two places: on_each_cpu() as shown above, and
> load_debug_registers(). It seems clear that contexts resulting from
> on_each_cpu() don't need RCU protection, because on_each_cpu() won't
> return until those routines have completed.
>
> This leaves only contexts resulting from load_debug_registers(). But
> the first thing load_debug_registers() does is disable local
> interrupts, thus blocking IPI delivery. Hence any simultaneous
> on_each_cpu() won't complete until after load_debug_registers() is
> done.
>
> So there doesn't seem to be any need for RCU protection in
> update_this_cpu().
>
> > Yes, indeed. With the current implementation, there's a possibility of
> > two instances of update_this_cpu() function executing - one with an
> > rcu_read_lock() taken (when called from load_debug_registers) while the
> > other without (when invoked through update_all_cpus()).
>
> No, this isn't possible unless I have misunderstood the nature of
> IPIs. Isn't is true that calling local_irq_save() will block delivery
> of IPIs?

Touche! ;-)

But in that case, why do you need the synchronize_rcu() following the
on_each_cpu() above? Is this needed to make sure that any concurrent
load_debug_registers() call has completed?

Thanx, Paul
--
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/