On Fri, Feb 12, 2016 at 12:32:15PM -0500, Waiman Long wrote:
> When a task prepares to sleep and then aborts it somehow, there is
> a small chance that a waker may be spinning on the on_cpu flag of
> that task waiting for the flag to turn off before doing the wakeup
> operation. It may keep on spinning for a long time until that task
> actually sleeps leading to spurious wakeup.
>
> This patch adds code to detect the change in task state and abort
> the wakeup operation, when appropriate, to free up the waker's cpu
> to do other useful works.
>
> Signed-off-by: Waiman Long <waiman.l...@hp.com>
> ---
> kernel/sched/core.c | 9 ++++++++-
> 1 files changed, 8 insertions(+), 1 deletions(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 7e548bd..e4b6e84 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2075,8 +2075,15 @@ try_to_wake_up(struct task_struct *p, unsigned int
> state, int wake_flags)
> *
> * This ensures that tasks getting woken will be fully ordered against
> * their previous state and preserve Program Order.
> + *
> + * If the owning cpu decides not to sleep after all by changing back
> + * its task state, we can return immediately.
> */
> - smp_cond_acquire(!p->on_cpu);
> + smp_cond_acquire(!p->on_cpu || !(p->state & state));
> + if (!(p->state & state)) {
> + success = 0;
> + goto out;
> + }
This doesn't make sense, if we managed to get here, p->on_rq must be
false, which means the other side is already in the middle of
schedule().
