Linus Torvalds (torvalds.delete@this.linux-foundation.org) on August 21, 2018 3:31 pm wrote:
> So you can solve it other ways. One way is to have the remote atomics, and just do the statistics update
> "somewhere else" and asynchronously. That's a lovely model in many ways, but it's not portable.

The only implementation of that I know of are in Freescale's POWER network processors. They added remote atomics which they advertised specifically for use in statistics. IIRC the actual operations were executed in the L3 cache. They worked by setting up a cache-inhibited range of memory and then used what they called "decorated operations" (i.e. loads or stores plus an operation) to access it; you couldn't access the same area as cached memory, it just didn't work. Here's some more info:

Decorated Operations for QorIQ P3/P4/P5 Processors

One big downside they had was that they could not be executed speculatively, they were effectively serializing instructions. The issue was that a speculatively executed instruction would trigger an atomic operation even if the instruction was later squashed if the speculation turned out to be wrong. The remote atomic machinery had no access to the core and thus no way to know that the instruction had been squashed, and would merrily execute it anyway.