On Fri, Sep 18, 2020 at 12:48:24PM +0200, Oleg Nesterov wrote: > Of course, this assumes that atomic_t->counter underflows "correctly", just > like "unsigned int".
We're documented that we do. Lots of code relies on that. See Documentation/atomic_t.txt TYPES > But again, do we really want this? I like the two counters better, avoids atomics entirely, some archs hare horridly expensive atomics (*cough* power *cough*). I just tried to be clever and use a single u64 load (where possible) instead of two 32bit loads and got the sum vs split order wrong.