sched: Fix balance_callback()

The intent of balance_callback() has always been to delay executing
balancing operations until the end of the current rq->lock section.  This
is because balance operations must often drop rq->lock, and that isn't safe
in general.

However, as noted by Scott, there were a few holes in that scheme;
balance_callback() was called after rq->lock was dropped, which means
another CPU can interleave and touch the callback list.

Rework code to call the balance callbacks before dropping rq->lock where
possible, and otherwise splice the balance list onto a local stack.

This guarantees that the balance list must be empty when rq->lock is
acquired. IOW, this will only ever run balance callbacks which are queued
while holding rq->lock.

Reported-by: Scott Wood <swood@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Link: https://lore.kernel.org/r/20200911082536.470013100@infradead.org

kernel/sched/core.c

@@ -3489,6 +3489,69 @@ static inline void finish_task(struct task_struct *prev)
 #endif
 }
 
+#ifdef CONFIG_SMP
+
+static void do_balance_callbacks(struct rq *rq, struct callback_head *head)
+{
+	void (*func)(struct rq *rq);
+	struct callback_head *next;
+
+	lockdep_assert_held(&rq->lock);
+
+	while (head) {
+		func = (void (*)(struct rq *))head->func;
+		next = head->next;
+		head->next = NULL;
+		head = next;
+
+		func(rq);
+	}
+}
+
+static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
+{
+	struct callback_head *head = rq->balance_callback;
+
+	lockdep_assert_held(&rq->lock);
+	if (head)
+		rq->balance_callback = NULL;
+
+	return head;
+}
+
+static void __balance_callbacks(struct rq *rq)
+{
+	do_balance_callbacks(rq, splice_balance_callbacks(rq));
+}
+
+static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
+{
+	unsigned long flags;
+
+	if (unlikely(head)) {
+		raw_spin_lock_irqsave(&rq->lock, flags);
+		do_balance_callbacks(rq, head);
+		raw_spin_unlock_irqrestore(&rq->lock, flags);
+	}
+}
+
+#else
+
+static inline void __balance_callbacks(struct rq *rq)
+{
+}
+
+static inline struct callback_head *splice_balance_callbacks(struct rq *rq)
+{
+	return NULL;
+}
+
+static inline void balance_callbacks(struct rq *rq, struct callback_head *head)
+{
+}
+
+#endif
+
 static inline void
 prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
 {
@@ -3514,6 +3577,7 @@ static inline void finish_lock_switch(struct rq *rq)
 	 * prev into current:
 	 */
 	spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
+	__balance_callbacks(rq);
 	raw_spin_unlock_irq(&rq->lock);
 }
 
@@ -3655,43 +3719,6 @@ static struct rq *finish_task_switch(struct task_struct *prev)
 	return rq;
 }
 
-#ifdef CONFIG_SMP
-
-/* rq->lock is NOT held, but preemption is disabled */
-static void __balance_callback(struct rq *rq)
-{
-	struct callback_head *head, *next;
-	void (*func)(struct rq *rq);
-	unsigned long flags;
-
-	raw_spin_lock_irqsave(&rq->lock, flags);
-	head = rq->balance_callback;
-	rq->balance_callback = NULL;
-	while (head) {
-		func = (void (*)(struct rq *))head->func;
-		next = head->next;
-		head->next = NULL;
-		head = next;
-
-		func(rq);
-	}
-	raw_spin_unlock_irqrestore(&rq->lock, flags);
-}
-
-static inline void balance_callback(struct rq *rq)
-{
-	if (unlikely(rq->balance_callback))
-		__balance_callback(rq);
-}
-
-#else
-
-static inline void balance_callback(struct rq *rq)
-{
-}
-
-#endif
-
 /**
  * schedule_tail - first thing a freshly forked thread must call.
  * @prev: the thread we just switched away from.
@@ -3711,7 +3738,6 @@ asmlinkage __visible void schedule_tail(struct task_struct *prev)
 	 */
 
 	rq = finish_task_switch(prev);
-	balance_callback(rq);
 	preempt_enable();
 
 	if (current->set_child_tid)
@@ -4527,10 +4553,11 @@ static void __sched notrace __schedule(bool preempt)
 		rq = context_switch(rq, prev, next, &rf);
 	} else {
 		rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
-		rq_unlock_irq(rq, &rf);
-	}
 
-	balance_callback(rq);
+		rq_unpin_lock(rq, &rf);
+		__balance_callbacks(rq);
+		raw_spin_unlock_irq(&rq->lock);
+	}
 }
 
 void __noreturn do_task_dead(void)
@@ -4941,9 +4968,11 @@ void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
 out_unlock:
 	/* Avoid rq from going away on us: */
 	preempt_disable();
-	__task_rq_unlock(rq, &rf);
 
-	balance_callback(rq);
+	rq_unpin_lock(rq, &rf);
+	__balance_callbacks(rq);
+	raw_spin_unlock(&rq->lock);
+
 	preempt_enable();
 }
 #else
@@ -5217,6 +5246,7 @@ static int __sched_setscheduler(struct task_struct *p,
 	int retval, oldprio, oldpolicy = -1, queued, running;
 	int new_effective_prio, policy = attr->sched_policy;
 	const struct sched_class *prev_class;
+	struct callback_head *head;
 	struct rq_flags rf;
 	int reset_on_fork;
 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
@@ -5455,6 +5485,7 @@ static int __sched_setscheduler(struct task_struct *p,
 
 	/* Avoid rq from going away on us: */
 	preempt_disable();
+	head = splice_balance_callbacks(rq);
 	task_rq_unlock(rq, p, &rf);
 
 	if (pi) {
@@ -5463,7 +5494,7 @@ static int __sched_setscheduler(struct task_struct *p,
 	}
 
 	/* Run balance callbacks after we've adjusted the PI chain: */
-	balance_callback(rq);
+	balance_callbacks(rq, head);
 	preempt_enable();
 
 	return 0;

kernel/sched/sched.h

@@ -1220,6 +1220,8 @@ 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;
+
+	SCHED_WARN_ON(rq->balance_callback);
 #endif
 }