sched: Optimize finish_lock_switch()
The kernel test robot measured a -1.6% performance regression on
will-it-scale/sched_yield due to commit:
2558aacff858 ("sched/hotplug: Ensure only per-cpu kthreads run during hotplug")
Even though we were careful to replace a single load with another
single load from the same cacheline.
Restore finish_lock_switch() to the exact state before the offending
patch and solve the problem differently.
Fixes: 2558aacff858 ("sched/hotplug: Ensure only per-cpu kthreads run during hotplug")
Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20201210161408.GX3021@hirez.programming.kicks-ass.net
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index f5acb6c..12ada79 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -975,7 +975,6 @@ struct rq {
unsigned long cpu_capacity_orig;
struct callback_head *balance_callback;
- unsigned char balance_flags;
unsigned char nohz_idle_balance;
unsigned char idle_balance;
@@ -1226,6 +1225,8 @@ struct rq_flags {
#endif
};
+extern struct callback_head balance_push_callback;
+
/*
* Lockdep annotation that avoids accidental unlocks; it's like a
* sticky/continuous lockdep_assert_held().
@@ -1243,9 +1244,9 @@ static inline void rq_pin_lock(struct rq *rq, struct rq_flags *rf)
#ifdef CONFIG_SCHED_DEBUG
rq->clock_update_flags &= (RQCF_REQ_SKIP|RQCF_ACT_SKIP);
rf->clock_update_flags = 0;
-#endif
#ifdef CONFIG_SMP
- SCHED_WARN_ON(rq->balance_callback);
+ SCHED_WARN_ON(rq->balance_callback && rq->balance_callback != &balance_push_callback);
+#endif
#endif
}
@@ -1408,9 +1409,6 @@ init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
#ifdef CONFIG_SMP
-#define BALANCE_WORK 0x01
-#define BALANCE_PUSH 0x02
-
static inline void
queue_balance_callback(struct rq *rq,
struct callback_head *head,
@@ -1418,13 +1416,12 @@ queue_balance_callback(struct rq *rq,
{
lockdep_assert_held(&rq->lock);
- if (unlikely(head->next || (rq->balance_flags & BALANCE_PUSH)))
+ if (unlikely(head->next || rq->balance_callback == &balance_push_callback))
return;
head->func = (void (*)(struct callback_head *))func;
head->next = rq->balance_callback;
rq->balance_callback = head;
- rq->balance_flags |= BALANCE_WORK;
}
#define rcu_dereference_check_sched_domain(p) \
