Re: [PATCH] driver core: fix shutdown races with probe/remove

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Jun 06, 2012 at 12:05:08PM -0400, Alan Stern wrote:
> On Wed, 6 Jun 2012, Paul E. McKenney wrote:
> 
> > On Wed, Jun 06, 2012 at 11:21:52AM -0400, Alan Stern wrote:
> > > On Wed, 6 Jun 2012, Paul E. McKenney wrote:
> > > 
> > > > No sane compiler would change it to a byte-at-a-time store, but the
> > > > compiler would nevertheless be within its rights to do so.  And a quick
> > > > review of certain LKML threads could easily cause anyone to question gcc's
> > > > sanity.  Furthermore, the compiler is permitted to make transformations
> > > > like the following, which it might well do to save a branch:
> > > > 
> > > > 	if (b)				a = 0;
> > > > 		a = 1;			if (b)
> > > > 	else					a = 1;
> > > > 		a = 0;
> > > 
> > > The compiler would be forbidden if the original code were
> > > 
> > > 	if (b)
> > > 		ACCESS_ONCE(a) = 1;
> > > 	else
> > > 		ACCESS_ONCE(a) = 0;
> > > 
> > > But if I remember correctly, the code snippet we were talking was more 
> > > like:
> > > 
> > > 	if (ACCESS_ONCE(b))
> > > 		a = 1;
> > > 
> > > Isn't this use of ACCESS_ONCE unnecessary?
> > 
> > That would depend on what else is nearby.
> 
> Here's the relevant part of the original patch:
> 
> @@ -467,6 +473,12 @@ EXPORT_SYMBOL_GPL(driver_attach);
>  static void __device_release_driver(struct device *dev)
>  {
>  	struct device_driver *drv;
> +	int idx;
> +
> +	idx = srcu_read_lock(&driver_srcu);
> +
> +	if (ACCESS_ONCE(device_shutdown_started))
> +		goto exit;
> 
>  	drv = dev->driver;
>  	if (drv) {
> @@ -494,6 +506,8 @@ static void __device_release_driver(struct device *dev)
>  						     dev);
> 
>  	}
> +exit:
> +	srcu_read_unlock(&driver_srcu, idx);
>  }

In this case, the ACCESS_ONCE() prevents the compiler from speculatively
executing the stuff following the "goto exit", which I freely admit is
insane even for compiler writers.  But the documentation benefits still
stand.

> > There are some limitations because volatile accesses are not allowed to
> > move past "sequence points", but it would be possible to come up with
> > similar examples.  This sort of thing is why C1x has a memory model and
> > why it allows variables to be designated as needing to be SMP-safe.
> 
> Almost certainly the kernel won't use this facility.  Or else it will 
> just require that _all_ global variables be SMP-safe.

I will reserve judgment until after I see what effect requiring all
globals to be SMP-safe has on code generation.  ;-)

							Thanx, Paul