Re: [PATCH] perf: Fix lockdep warning in perf_event_throttle_group

[Ian Rogers]

On Thu, Nov 13, 2025 at 12:15 AM Mi, Dapeng <dapeng1.mi@xxxxxxxxxxxxxxx> wrote:
>
>
> On 11/12/2025 8:34 PM, Jianhui Zhou wrote:
> > When perf samples a tracepoint like trace_lock_acquire with high
> > frequency, it can trigger a WARN_ON in perf_event_throttle_group()
> > due to lockdep assertion failure in for_each_sibling_event().
> >
> > The issue is that for_each_sibling_event() requires either:
> > 1. Holding ctx->mutex, OR
> > 2. Having IRQs disabled (since ctx->lock is IRQ-safe)
> >
> > The call trace shows the problematic path:
> > futex_wake
> >   lock_acquire
> >     trace_lock_acquire
> >       perf_trace_lock_acquire
> >         perf_tp_event
> >           perf_swevent_event
> >             __perf_event_overflow
> >               __perf_event_account_interrupt
> >                 perf_event_throttle_group
> >                   for_each_sibling_event [WARN]
> >
> > In this tracepoint path, IRQs may be enabled and neither ctx->lock
> > nor ctx->mutex is held.
> >
> > The fix adds guard(irqsave)() protection around the sibling list
> > iteration in perf_event_throttle_group().
> >
> > Reported-by: syzbot+a945e9d15c8a49a7a7f0@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Closes: https://lore.kernel.org/all/69122cb3.a70a0220.22f260.00ff.GAE@xxxxxxxxxx/T/
> > Signed-off-by: Jianhui Zhou <jianhuizz@xxxxxx>
> > ---
> >  kernel/events/core.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> >
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 1fd347da9026..253c514e9452 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -2700,6 +2700,11 @@ static void perf_event_throttle_group(struct perf_event *event)
> >       struct perf_event *sibling, *leader = event->group_leader;
> >
> >       perf_event_throttle(leader);
> > +    /*
> > +     * Disable IRQs to protect sibling iteration; for_each_sibling_event()
> > +     * needs ctx->mutex or IRQs off.
> > +     */
> > +     guard(irqsave)();
>
> This seems the simplest way to fix the issue although it may bring a slight
> overhead if hard irq is already disabled. Thanks.
>
> Reviewed-by: Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx>

It seems this was never submitted, should it be?

Thanks,
Ian

> >       for_each_sibling_event(sibling, leader)
> >               perf_event_throttle(sibling);
> >  }