Re: lockdep warning about possible circular dependency in PSI

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Apr 24, 2020 at 09:34:42AM -0700, Suren Baghdasaryan wrote:
> Sorry to bother you again folks. Any suggestions on how to silence
> this lockdep warning which I believe to be a false positive?
> 
> On Wed, Apr 15, 2020 at 4:01 PM Suren Baghdasaryan <surenb@google.com> wrote:
> >
> > I received a report about possible circular locking dependency warning
> > generated from PSI polling code. I think we are protected from this
> > scenario by poll_scheduled atomic but wanted to double-check and I’m
> > looking for an advice on how to annotate this case to fix the lockdep
> > warning. I copied the detailed information at the end of this email
> > but the short story is this:
> >
> > "WARNING: possible circular locking dependency detected" is generated
> > with CONFIG_PSI and CONFIG_LOCKDEP enabled. The dependency chain it
> > describes is:
> >
> > #0
> > kthread_delayed_work_timer_fn()
> >  |
> > worker->lock
> >  |
> > try_to_wake_up()
> >  |
> > p->pi_lock
> >
> > #1
> > sched_fork()
> >  |
> > p->pi_lock
> >  |
> > task_fork_fair()
> >  |
> > rq->lock
> >
> > #2
> > psi_memstall_enter
> >  |
> > rq->lock
> >  |
> > kthread_queue_delayed_work()
> >  |
> > worker->lock

Irrespective of it actually being a deadlock or not, it is fairly
fragile. Ideally we'd fix #2, we really should minimize the number of
locks nested under rq->lock.

That said, here's the easy fix, which breaks #0.

---
diff --git a/kernel/kthread.c b/kernel/kthread.c
index bfbfa481be3a..b443bba7dd21 100644
--- a/kernel/kthread.c
+++ b/kernel/kthread.c
@@ -806,14 +806,15 @@ static void kthread_insert_work_sanity_check(struct kthread_worker *worker,
 /* insert @work before @pos in @worker */
 static void kthread_insert_work(struct kthread_worker *worker,
 				struct kthread_work *work,
-				struct list_head *pos)
+				struct list_head *pos,
+				struct wake_q_head *wake_q)
 {
 	kthread_insert_work_sanity_check(worker, work);
 
 	list_add_tail(&work->node, pos);
 	work->worker = worker;
 	if (!worker->current_work && likely(worker->task))
-		wake_up_process(worker->task);
+		wake_q_add(wake_q, worker->task);
 }
 
 /**
@@ -831,15 +832,19 @@ static void kthread_insert_work(struct kthread_worker *worker,
 bool kthread_queue_work(struct kthread_worker *worker,
 			struct kthread_work *work)
 {
-	bool ret = false;
+	DEFINE_WAKE_Q(wake_q);
 	unsigned long flags;
+	bool ret = false;
 
 	raw_spin_lock_irqsave(&worker->lock, flags);
 	if (!queuing_blocked(worker, work)) {
-		kthread_insert_work(worker, work, &worker->work_list);
+		kthread_insert_work(worker, work, &worker->work_list, &wake_q);
 		ret = true;
 	}
 	raw_spin_unlock_irqrestore(&worker->lock, flags);
+
+	wake_up_q(&wake_q);
+
 	return ret;
 }
 EXPORT_SYMBOL_GPL(kthread_queue_work);
@@ -857,6 +862,7 @@ void kthread_delayed_work_timer_fn(struct timer_list *t)
 	struct kthread_delayed_work *dwork = from_timer(dwork, t, timer);
 	struct kthread_work *work = &dwork->work;
 	struct kthread_worker *worker = work->worker;
+	DEFINE_WAKE_Q(wake_q);
 	unsigned long flags;
 
 	/*
@@ -873,15 +879,18 @@ void kthread_delayed_work_timer_fn(struct timer_list *t)
 	/* Move the work from worker->delayed_work_list. */
 	WARN_ON_ONCE(list_empty(&work->node));
 	list_del_init(&work->node);
-	kthread_insert_work(worker, work, &worker->work_list);
+	kthread_insert_work(worker, work, &worker->work_list, &wake_q);
 
 	raw_spin_unlock_irqrestore(&worker->lock, flags);
+
+	wake_up_q(&wake_q);
 }
 EXPORT_SYMBOL(kthread_delayed_work_timer_fn);
 
 static void __kthread_queue_delayed_work(struct kthread_worker *worker,
 					 struct kthread_delayed_work *dwork,
-					 unsigned long delay)
+					 unsigned long delay,
+					 struct wake_q_head *wake_q)
 {
 	struct timer_list *timer = &dwork->timer;
 	struct kthread_work *work = &dwork->work;
@@ -895,7 +904,7 @@ static void __kthread_queue_delayed_work(struct kthread_worker *worker,
 	 * on that there's no such delay when @delay is 0.
 	 */
 	if (!delay) {
-		kthread_insert_work(worker, work, &worker->work_list);
+		kthread_insert_work(worker, work, &worker->work_list, wake_q);
 		return;
 	}
 
@@ -928,17 +937,21 @@ bool kthread_queue_delayed_work(struct kthread_worker *worker,
 				unsigned long delay)
 {
 	struct kthread_work *work = &dwork->work;
+	DEFINE_WAKE_Q(wake_q);
 	unsigned long flags;
 	bool ret = false;
 
 	raw_spin_lock_irqsave(&worker->lock, flags);
 
 	if (!queuing_blocked(worker, work)) {
-		__kthread_queue_delayed_work(worker, dwork, delay);
+		__kthread_queue_delayed_work(worker, dwork, delay, &wake_q);
 		ret = true;
 	}
 
 	raw_spin_unlock_irqrestore(&worker->lock, flags);
+
+	wake_up_q(&wake_q);
+
 	return ret;
 }
 EXPORT_SYMBOL_GPL(kthread_queue_delayed_work);
@@ -967,6 +980,7 @@ void kthread_flush_work(struct kthread_work *work)
 		KTHREAD_WORK_INIT(fwork.work, kthread_flush_work_fn),
 		COMPLETION_INITIALIZER_ONSTACK(fwork.done),
 	};
+	DEFINE_WAKE_Q(wake_q);
 	struct kthread_worker *worker;
 	bool noop = false;
 
@@ -979,15 +993,17 @@ void kthread_flush_work(struct kthread_work *work)
 	WARN_ON_ONCE(work->worker != worker);
 
 	if (!list_empty(&work->node))
-		kthread_insert_work(worker, &fwork.work, work->node.next);
+		kthread_insert_work(worker, &fwork.work, work->node.next, &wake_q);
 	else if (worker->current_work == work)
 		kthread_insert_work(worker, &fwork.work,
-				    worker->work_list.next);
+				    worker->work_list.next, &wake_q);
 	else
 		noop = true;
 
 	raw_spin_unlock_irq(&worker->lock);
 
+	wake_up_q(&wake_q);
+
 	if (!noop)
 		wait_for_completion(&fwork.done);
 }
@@ -1065,6 +1081,7 @@ bool kthread_mod_delayed_work(struct kthread_worker *worker,
 			      unsigned long delay)
 {
 	struct kthread_work *work = &dwork->work;
+	DEFINE_WAKE_Q(wake_q);
 	unsigned long flags;
 	int ret = false;
 
@@ -1083,9 +1100,12 @@ bool kthread_mod_delayed_work(struct kthread_worker *worker,
 
 	ret = __kthread_cancel_work(work, true, &flags);
 fast_queue:
-	__kthread_queue_delayed_work(worker, dwork, delay);
+	__kthread_queue_delayed_work(worker, dwork, delay, &wake_q);
 out:
 	raw_spin_unlock_irqrestore(&worker->lock, flags);
+
+	wake_up_q(&wake_q);
+
 	return ret;
 }
 EXPORT_SYMBOL_GPL(kthread_mod_delayed_work);