In preparation of using the balance_push state in ttwu() we need it to provide a reliable and consistent state.
The immediate problem is that rq->balance_callback gets cleared every schedule() and then re-set in the balance_push_callback() itself. This is not a reliable signal, so add a variable that stays set during the entire time. Also move setting it before the synchronize_rcu() in sched_cpu_deactivate(), such that we get guaranteed visibility to ttwu(), which is a preempt-disable region. Signed-off-by: Peter Zijlstra (Intel) <[email protected]> --- kernel/sched/core.c | 18 +++++++++++++----- kernel/sched/sched.h | 1 + 2 files changed, 14 insertions(+), 5 deletions(-) --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -7320,6 +7327,7 @@ static void balance_push_set(int cpu, bo struct rq_flags rf; rq_lock_irqsave(rq, &rf); + rq->balance_push = on; if (on) { WARN_ON_ONCE(rq->balance_callback); rq->balance_callback = &balance_push_callback; @@ -7489,17 +7497,17 @@ int sched_cpu_deactivate(unsigned int cp int ret; set_cpu_active(cpu, false); + balance_push_set(cpu, true); + /* - * We've cleared cpu_active_mask, wait for all preempt-disabled and RCU - * users of this state to go away such that all new such users will - * observe it. + * We've cleared cpu_active_mask / set balance_push, wait for all + * preempt-disabled and RCU users of this state to go away such that + * all new such users will observe it. * * Do sync before park smpboot threads to take care the rcu boost case. */ synchronize_rcu(); - balance_push_set(cpu, true); - rq_lock_irqsave(rq, &rf); if (rq->rd) { update_rq_clock(rq); --- a/kernel/sched/sched.h +++ b/kernel/sched/sched.h @@ -975,6 +975,7 @@ struct rq { unsigned long cpu_capacity_orig; struct callback_head *balance_callback; + unsigned char balance_push; unsigned char nohz_idle_balance; unsigned char idle_balance;
