Linus Torvalds (torvalds.delete@this.linux-foundation.org) on August 21, 2018 6:02 pm wrote:
> anon (anon.delete@this.anon.com) on August 21, 2018 4:27 pm wrote:
> >
> > "Remote atomics"? "Lovely model"? Who is this crazy person impersonating Linus?
>
> I think they conceptually make sense.
>
> Whether they actually work in practice or not I will leave to others. Most of the users are likely
> user space, although I think the kernel could find use for them too. There are many possible gotchas,
> like the fact that the remote atomics have to participate in the TLB shootdown etc.
>
> But yes, it may be an implementation nightmare. Still, I'd rather see that kind of architecture
> work than yet another vector execution thing that slows down everything else.
>
> Linus

Maybe the assembly instruction "inc [memory]" which is probably mostly used for statistics (not transfering the content to a register, not reusing the result) could on cacheline-in-invalid-state state, register interrest of this cacheline to be seen by local processor in exclusive mode in "fair" share.
Maybe cachelines marked with this flag "to be shared fairly" should be held in exclusive mode by the local processor only until all the instructions currently in flight referencing it are retired, and then only given to the processor with an higher index: ((processor_ID + 1) % nb_processor).
Maybe it would be an implementation nightmare.