Re: [Lse-tech] Re: RFC: patch to allow lock-free traversal of lists with insertion

[torvalds@transmeta.com (Linus Torvalds)]

In article <20011010185848.D726@athlon.random>,
Andrea Arcangeli  <andrea@suse.de> wrote:
>
>However the more I think about it the more I suspect we'd better use
>rmb() in all readers in the common code

Absolutely.  It's not that expensive an operation on sane hardware.  And
it's definitely conceptually the only right thing to do - we're saying
that we're doing a read that depends on a previous read having seen
previous memory.  Ergo, "rmb()". 

Of course, right now Linux only exports a subset of the potential memory
barriers, and maybe we should export a fuller set - allowing CPU's that
have stricter ordering to possibly make it a no-op.  But thinking about
even something like x86, I don't see where Intel would guarantee that
two reads (data-dependent or not) would have some implicit memory
ordering. 

Re-ordering reads with data dependencies is hard, but it happens quite
naturally in a CPU that does address speculation. I don't know of
anybody who does that, but I bet _somebody_ will. Maybe even the P4?

			Linus