Re: [BK PATCHES] 2.6.x net driver fixes

[Linus Torvalds <torvalds@xxxxxxxx>]

On Wed, 14 Jul 2004, Jeff Garzik wrote:
> > 
> > The weaker ordering constraint is called "read_barrier_depends()", and
> > should be a no-op on all but alpha. Would you want to use that to fix the
> > alpha case too?
> 
> What would the proper fix be?
> 
>       rmb();
>       read_barrier_depends();

No, just a

        index = *indexpointer;
        read_barrier_depends();
        data = dataptr[index];

which on most architectures is a no-op, but in case the hardware might do
data or address speculation, and needs a dependent load barrier, it will
do the right thing.

For stuff that isn't data dependent, you have to use the full rmb(), which 
obviously isn't a no-op on most setups.

                Linus