[PATCH] Document Linux's memory barriers [try #4]

[Paul Mackerras]

David Howells writes:

> Paul Mackerras <paulus at samba.org> wrote:
> 
> > No, that's not the problem.  The problem is that you can get q == &b
> > and d == 1, believe it or not.  That is, you can see the new value of
> > the pointer but the old value of the thing pointed to.
> 
> But that doesn't make any sense!

It certainly violates the principle of least surprise. :)

Apparently this can occur on some Alpha machines that have a
partitioned cache.  Although CPU 1 sends out the updates to b and p in
the right order because of the smp_wmb(), it's possible that b and p
are present in CPU 2's cache, one in each half of the cache.  If there
are a lot of updates coming in for the half containing b, but the half
containing p is quiet, it is possible for CPU 2 to see a new value of
p but an old value of b, unless you put an rmb instruction between the
two loads from memory.

I haven't heard of this being an issue on any other architecture.  On
PowerPC it can't happen because the architecture specifies that a data
dependency creates an implicit read barrier.

Paul.