* Mel Gorman <mgor...@techsingularity.net> wrote:
> The following two patches optimise try_to_wake_up() when the wakee is
> descheduling. In a vanilla kernel, there can be excessive time spent
> spinning on p->on_rq. This is fine if it's a strictly synchronous wakeup
> and the waker is going to sleep but in other cases, the waker spins until
> it can do work that can be deferred to the wakee.
>
> The first patch frontloads work that can be done before p->on_rq is
> checked. If it's a wakeup on a CPU that does not share cache then the
> wakelist is used instead of spinning. The second patch goes a little
> further and uses the wakelist if the wakee is descheduling and is the
> only task running on the target CPU.
>
> The performance impact is documented in the changelog of the second patch.
>
> kernel/sched/core.c | 81 ++++++++++++++++++++++++++++++++------------
> kernel/sched/sched.h | 3 +-
> 2 files changed, 61 insertions(+), 23 deletions(-)
Thanks, these are really nice improvements, those latency spikes in netperf
runs were always annoying, and they've been there forever I think.
I've applied these two patches to tip:sched/core:
c6e7bd7afaeb: ("sched/core: Optimize ttwu() spinning on p->on_cpu")
2ebb17717550: ("sched/core: Offload wakee task activation if it the wakee is
descheduling")
In 2ebb17717550 I resolved the conflict with a recent cleanup by PeterZ:
90b5363acd47: ("sched: Clean up scheduler_ipi()")
... which had a drive-by cleanup of ttwu_queue_remote() as well.
Thanks,
Ingo
