Re: KCSAN: data-race in __alloc_file / __alloc_file

[Eric Dumazet]

On Fri, Nov 8, 2019 at 12:30 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Fri, Nov 8, 2019 at 9:56 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> >
> > BTW, I would love an efficient ADD_ONCE(variable, value)
> >
> > Using WRITE_ONCE(variable, variable + value) is not good, since it can
> > not use the optimized instructions operating directly on memory.
>
> So I'm having a hard time seeing how this could possibly ever be valid.
>
> Is this a "writer is locked, readers are unlocked" case or something?

per cpu SNMP counters mostly, with no IRQ safety requirements.

Note that this could be implemented using local{64}_add() on arches like x86_64,
while others might have to fallback to WRITE_ONCE(variable, variable + add)

>
> Because we don't really have any sane way to do that any more
> efficiently, unless we'd have to add new architecture-specific
> functions for it (like we do have fo the percpu ops).
>
> Anyway, if you have a really hot case you care about, maybe you could
> convince the gcc people to just add it as a peephole optimization?
> Right now, gcc ends up doing some strange things with volatiles, and
> basically disables a lot of stuff over them. But with a test-case,
> maybe you can convince somebody that certain optimizations are still
> fine. A "read+add+write" really does the exact same accesses as an
> add-to-memory instruction, but gcc has some logic to disable that
> instruction fusion.
>
>           Linus