Messages in this thread | | | From | Suren Baghdasaryan <> | Date | Mon, 27 Apr 2020 11:47:50 -0700 | Subject | Re: lockdep warning about possible circular dependency in PSI |
| |
On Sun, Apr 26, 2020 at 5:00 PM Suren Baghdasaryan <surenb@google.com> wrote: > > On Fri, Apr 24, 2020 at 1:54 PM Shakeel Butt <shakeelb@google.com> wrote: > > > > On Fri, Apr 24, 2020 at 10:52 AM Suren Baghdasaryan <surenb@google.com> wrote: > > > > > > On Fri, Apr 24, 2020 at 10:40 AM Peter Zijlstra <peterz@infradead.org> wrote: > > > > > > > > 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. > > > > > > > > > > Thanks for the suggestion, Peter. Let me digest this and will post a > > > patch with your Suggested-by. > > > Cheers! > > > > > > > I tested on my simple repro and the patch fixes the lockdep splat. > > > > You can add > > Tested-by: Shakeel Butt <shakeelb@google.com> > > > > Thanks Shakeel! Will do. > > > > > --- > > > > 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); > > > >
Patch is posted at https://lkml.org/lkml/2020/4/27/985 . Had to include linux/sched/wake_q.h and fix a long line but other than that it's unchanged. Thanks!
> > > > -- > > > > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@android.com. > > > >
| |