[Lguest] [PATCH 4/8] lguest: update commentry

[Rusty Russell]

On Sat, 25 Jul 2009 01:20:05 am Paul E. McKenney wrote:
> On Fri, Jul 24, 2009 at 08:12:14PM +0930, Rusty Russell wrote:
> > 	 * rcu_dereference is the counter-side of rcu_assign_pointer(); it
> > 	 * makes sure we don't access the memory pointed to by
> > 	 * cpu->lg->eventfds before cpu->lg->eventfds is set.  As you might
> > 	 * expect, that's impossible on almost every architecture anyway.
> 
> Perhaps add a sentence about how aggressive optimizing compilers can
> get this effect?  For whatever it is worth, one such optimization is as
> follows:
> 
> 1.	Compiler guesses the value of the pointer, and also issues a
> 	load of the pointer value.  Note that if the pointer almost
> 	never changes, the guess will almost always be correct.
> 	But to keep things interesting, let's assume that the guess
> 	is -not- correct.
> 
> 2.	The compiler dereferences its guess, picking up garbage.
> 
> 3.	Meanwhile, the updater allocates new memory, just happening
> 	to get memory whose address matches the compiler's incorrect
> 	guess.  The updater initializes this new memory and updates
> 	the pointer.

Ah, if the "guess" is a previously cached copy of the pointer (thus an
indefinite number of allocs behind us) this is indeed possible.

Here's the new version:

	/*
	 * rcu_dereference is the counter-side of rcu_assign_pointer(); it
	 * makes sure we don't access the memory pointed to by
	 * cpu->lg->eventfds before cpu->lg->eventfds is set.  Sounds crazy,
	 * but Alpha allows this!  Paul McKenney points out that a really
	 * aggressive compiler could have the same effect:
	 *   http://lists.ozlabs.org/pipermail/lguest/2009-July/001560.html
	 *
	 * So play safe, use rcu_dereference to get the rcu-protected pointer:
	 */
	map = rcu_dereference(cpu->lg->eventfds);

Thanks!
Rusty.