On Wed, May 19, 2021 at 5:19 AM Mark Kettenis <mark.kette...@xs4all.nl>
wrote:
> > Date: Tue, 18 May 2021 13:24:42 +0200
> > From: Martin Pieuchot <m...@openbsd.org>
>
...
> > There's only a couple of 'volatile' usages in sys/sys. These annotations
> > do not explicitly indicate which piece of code requires it. Maybe it
> would
> > be clearer to use a cast or a macro where necessary. This might help us
> > understand why and where "volatile" is needed.
>
> There are the READ_ONCE() and WRITE_ONCE() macros. I'm not a big fan
> of those (since they add clutter) but they do take care of dependency
> ordering issues that exist in the alpha memory model. Must admit that
> I only vaguely understand that issue, but I think it involves ordered
> access to two atomic variables which doesn't seem to be the case.
>
> On non-alpha systems, READ_ONCE() and WRITE_ONCE() just do a volatile
> pointer cast.
>
READ/WRITE_ONCE() are 99% about keeping the compiler from deciding to
rearrange code such that the indicated variable is read/written more than
once. To point to Linus posts from 2008/2009, when it was still
ACCESS_ONCE():
https://yarchive.net/comp/linux/ACCESS_ONCE.html
ISTR a paper by Ousterhout describing this same problem to the C standard
committee(?) in the early 90's, which kinda opened the "memory model"
rathole.
If the variable is actually being protected by a lock then these are indeed
noise/pessimization; it's the lock-less accesses where the compiler can
pull a rabbit from its hat and stab you with it.
Philip Guenther
