On 07/12/2018 22:35, Steven Sistare wrote: [...] >>> @@ -4468,8 +4495,12 @@ static void throttle_cfs_rq(struct cfs_rq *cfs_rq) >>> dequeue = 0; >>> } >>> >>> - if (!se) >>> + if (!se) { >>> sub_nr_running(rq, task_delta); >>> + if (prev_nr >= 2 && prev_nr - task_delta < 2) >>> + overload_clear(rq); >>> + >>> + } >> >> Eventually it'd be nice to squash those into {add, sub}_nr_running(), but >> you already mentioned wanting to stick to CFS for now, so I don't think >> it's *too* much of a big deal. > > Maybe. It depends on a design decision to be made if/when we add bitmap > based stealing to other scheduling classes. Do we maintain one bitmap > for overloaded CPUs where the overload may be caused by any mix of different > task classes? If yes, then the bitmap search for one class such as RT > will inspect and reject overloaded CPUs that only have CFS tasks, which > making the search less efficient. I am leaning towards a separate bitmap > per class to avoid that. >
Didn't ponder too much about it, but a bitmap per class sounds sane. My comment was mostly about saving ourselves the need to decorate every {add, sub}_nr_running() call with overload_{set, clear}() calls. > - Steve > >