Re: [RFC PATCH -lkmm] docs/memory-barriers: Fix inconsistent name of 'data dependency barrier'

From: Alan Stern
Date: Tue Jun 07 2022 - 10:34:23 EST


On Tue, Jun 07, 2022 at 02:34:33PM +0100, Will Deacon wrote:
> On Sat, May 28, 2022 at 01:15:30PM +0900, Akira Yokosawa wrote:
> > The term "data dependency barrier", which has been in
> > memory-barriers.txt ever since it was first authored by David Howells,
> > has become confusing due to the fact that in LKMM's explanations.txt
> > and elsewhere, "data dependency" is used mostly for load-to-store data
> > dependency.
> >
> > To prevent further confusions, do the following changes:
> >
> > - substitute "address-dependency barrier" for "data dependency barrier";
> > - add note on the removal of kernel APIs for explicit address-
> > dependency barriers in kernel release v5.9;
> > - add note on the section title rename;
> > - use READ_ONCE_OLD() for READ_ONCE() of pre-4.15 (no address-
> > dependency implication) in code snippets;
> > - fix number of CPU memory barrier APIs;
> > - and a few more context adjustments.

> I suppose this isn't really a comment on your patch, as I much prefer the
> updated terminology, but the way this section is now worded really makes it
> sounds like address dependencies only order load -> load, whereas they
> equally order load -> store. Saying that "An address-dependency barrier...
> is not required to have any effect on stores" is really confusing to me: the
> barrier should only ever be used in conjunction with an address-dependency
> _anyway_ so whether or not it's the barrier or the dependency giving the
> order is an implementation detail.

It would be more accurate to say that address-dependency barriers are
not _needed_ for load->store ordering because the dependencies
themselves already provide this ordering (even on Alpha).

> Perhaps the barrier should be called a "Read-read-address-dependency
> barrier", an "Address-dependency read barrier" or even a "Consume barrier"
> (:p) instead? Dunno, Alan is normally much better at naming these things
> than I am.

Well, "load-load-address-dependency barrier" would be okay as a generic
name, albeit unwieldy. Note however that on Alpha -- the only
architecture to need these barriers -- they aren't anything special; the
actual instruction is the equivalent of an ordinary smp_rmb(). (Please
correct me if my memory about this is wrong.)

So in principle you could simply call them "read memory barriers" while
pointing out the need for special use on demented architectures where
address dependencies do not guarantee load->load ordering.

> Alternatively, maybe we should be removing the historical stuff from the
> document altogether if it's no longer needed. We don't have any occurrences
> of read_barrier_depends() anymore, so why confuse people with it?

How about relegating discussion of these barriers to a special
"historical" or "niche architecture" section of the document? In a
separate patch, of course.

Alan