On Wed, Nov 02, 2022 at 11:16:48PM +0100, Peter Zijlstra wrote:
> On Wed, Nov 02, 2022 at 06:57:51PM +0200, Ville Syrjälä wrote:
> > On Thu, Oct 27, 2022 at 06:53:23PM +0200, Peter Zijlstra wrote:
> > > On Thu, Oct 27, 2022 at 04:09:01PM +0300, Ville Syrjälä wrote:
> > > > On Wed, Oct 26, 2022 at 01:43:00PM +0200, Peter Zijlstra wrote:
> > >
> > > > > Could you please give the below a spin?
> > > >
> > > > Thanks. I've added this to our CI branch. I'll try to keep and eye
> > > > on it in the coming days and let you know if anything still trips.
> > > > And I'll report back maybe ~middle of next week if we haven't caught
> > > > anything by then.
> > >
> > > Thanks!
> >
> > Looks like we haven't caught anything since I put the patch in.
> > So the fix seems good.
>
> While writing up the Changelog, it occured to me it might be possible to
> fix another way, could I bother you to also run the below patch for a
> bit?
I swapped in the new patch to the CI branch. I'll check back
after a few days.
>
> ---
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index cb2aa2b54c7a..daff72f00385 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -4200,6 +4200,40 @@ try_to_wake_up(struct task_struct *p, unsigned int
> state, int wake_flags)
> return success;
> }
>
> +static bool __task_needs_rq_lock(struct task_struct *p)
> +{
> + unsigned int state = READ_ONCE(p->__state);
> +
> + /*
> + * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
> + * the task is blocked. Make sure to check @state since ttwu() can drop
> + * locks at the end, see ttwu_queue_wakelist().
> + */
> + if (state == TASK_RUNNING || state == TASK_WAKING)
> + return true;
> +
> + /*
> + * Ensure we load p->on_rq after p->__state, otherwise it would be
> + * possible to, falsely, observe p->on_rq == 0.
> + *
> + * See try_to_wake_up() for a longer comment.
> + */
> + smp_rmb();
> + if (p->on_rq)
> + return true;
> +
> +#ifdef CONFIG_SMP
> + /*
> + * Ensure the task has finished __schedule() and will not be referenced
> + * anymore. Again, see try_to_wake_up() for a longer comment.
> + */
> + smp_rmb();
> + smp_cond_load_acquire(&p->on_cpu, !VAL);
> +#endif
> +
> + return false;
> +}
> +
> /**
> * task_call_func - Invoke a function on task in fixed state
> * @p: Process for which the function is to be invoked, can be @current.
> @@ -4217,28 +4251,12 @@ try_to_wake_up(struct task_struct *p, unsigned int
> state, int wake_flags)
> int task_call_func(struct task_struct *p, task_call_f func, void *arg)
> {
> struct rq *rq = NULL;
> - unsigned int state;
> struct rq_flags rf;
> int ret;
>
> raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
>
> - state = READ_ONCE(p->__state);
> -
> - /*
> - * Ensure we load p->on_rq after p->__state, otherwise it would be
> - * possible to, falsely, observe p->on_rq == 0.
> - *
> - * See try_to_wake_up() for a longer comment.
> - */
> - smp_rmb();
> -
> - /*
> - * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
> - * the task is blocked. Make sure to check @state since ttwu() can drop
> - * locks at the end, see ttwu_queue_wakelist().
> - */
> - if (state == TASK_RUNNING || state == TASK_WAKING || p->on_rq)
> + if (__task_needs_rq_lock(p))
> rq = __task_rq_lock(p, &rf);
>
> /*
--
Ville Syrjälä
Intel
