On Tue, Apr 22, 2008 at 09:03:04AM +0800, Herbert Xu wrote: > On Mon, Apr 21, 2008 at 08:49:58AM -0700, Linus Torvalds wrote: > > > > That is *not* the main problem. > > > > If you use "rcu_dereference()" on the wrong access, it not only loses the > > "smp_read_barrier_depends()" (which is a no-op on all sane architectures > > anyway), but it loses the ACCESS_ONCE() thing *entirely*. > > Actually rcu_dereference didn't have ACCESS_ONCE when I did this. > That only appearaed later with the preemptible RCU work. Yep, ACCESS_ONCE() is quite recent -- within the last year. So I should have modified the list_for_each.*rcu() macros when I made that change. > The original purpose of rcu_dereference was exactly to replace the > explicit barriers that people were using for RCU, nothing more, > nothing less. > > Oh and I totally agree that the compiler is going to generate insane > code whenever ACCESS_ONCE is used. In this case we may have avoided > it by rearranging the code, but in general the introduction of ACCESS_ONCE > in rcu_dereference is likely to have a negative impact on the code > generated. > > Remember that "volatile" discussion? I think this is where it all came > from. And I still have the bug in to gcc: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33102 Interesting, currently in status "unconfirmed"... I guess I should supply a test case. Thanx, Paul -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
