On Wed, May 25, 2016 at 07:52:49AM +0800, Yuyang Du wrote: > On Mon, May 23, 2016 at 11:58:49AM +0100, Morten Rasmussen wrote: > > For systems with the SD_ASYM_CPUCAPACITY flag set on higher level in the > > sched_domain hierarchy we need a way to enable wake-up balancing for the > > lower levels as well as we may want to balance tasks that don't fit the > > capacity of the previous cpu. > > > > We have the option of introducing a new topology flag to express this > > requirement, or let the existing SD_BALANCE_WAKE flag be set by the > > architecture as a topology flag. The former means introducing yet > > another flag, the latter breaks the current meaning of topology flags. > > None of the options are really desirable. > > I'd propose to replace SD_WAKE_AFFINE with SD_BALANCE_WAKE. And the > SD_WAKE_AFFINE semantic is simply "waker allowed": > > waker_allowed = cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); > > This can be implemented without current functionality change. > > From there, the choice between waker and wakee, and fast path > select_idle_sibling() and the rest slow path should be reworked, which > I am thinking about.
I don't really understand how that would work. If you change the semantics of the flags you don't preserve current behaviour. To me it sounds like at total rewrite of everything. SD_BALANCE_WAKE controls whether we go slow path or not in case want_affine is false. SD_WAKE_AFFINE controls whether we should consider waking up near the waker instead of always waking up near the previous cpu.