Re: [RFC PATCH 03/21] list: Annotate lockless list primitives with data_race()

[Will Deacon]

On Wed, Apr 01, 2020 at 08:34:36AM +0200, Marco Elver wrote:
> On Tue, 31 Mar 2020 at 15:10, Will Deacon <will@xxxxxxxxxx> wrote:
> > On Tue, Mar 24, 2020 at 05:23:30PM +0100, Marco Elver wrote:
> > > Then, my suggestion would be to mark the write with data_race() and
> > > just leave this as a READ_ONCE(). Having a data_race() somewhere only
> > > makes KCSAN stop reporting the race if the paired access is also
> > > marked (be it with data_race() or _ONCE, etc.).
> >
> > The problem with taking that approach is that it ends up much of the
> > list implementation annotated with either WRITE_ONCE() or data_race(),
> > meaning that concurrent, racy list operations will no longer be reported
> > by KCSAN. I think that's a pretty big deal and I'm strongly against
> > annotating the internals of library code such as this because it means
> > that buggy callers will largely go undetected.
> >
> > The situation we have here is that some calls, e.g. hlist_empty() are
> > safe even in the presence of a racy write and I'd like to suppress KCSAN
> > reports without annotating the writes at all.
> >
> > > Alternatively, if marking the write is impossible, you can surround
> > > the access with kcsan_disable_current()/kcsan_enable_current(). Or, as
> > > a last resort, just leaving as-is is fine too, because KCSAN's default
> > > config (still) has KCSAN_ASSUME_PLAIN_WRITES_ATOMIC selected.
> >
> > Hmm, I suppose some bright spark will want to change the default at the some
> > point though, no? ;) I'll look at using
> > kcsan_disable_current()/kcsan_enable_current(), thanks.
> 
> I think this will come up again (it did already come up in some other
> patch I reviewed, and Paul also mentioned it), so it seems best to
> change data_race() to match the intuitive semantics of just completely
> ignoring the access marked with it. I.e. marking accesses racing with
> accesses marked with data_race() is now optional:
>   https://lkml.kernel.org/r/20200331193233.15180-1-elver@xxxxxxxxxx

/me goes look. Thanks!

> In which case, the original patch you had here works just fine.

Ah yes, so now data_race(READ_ONCE(...)) does make sense as a combination.
It's tempting to wrap that up as an accessor, but actually forcing people to
spell it out might not be a bad thing after all.

Will