Re: [QUESTION] problems report: rcu_read_unlock_special() called in irq_exit() causes dead loop

[Joel Fernandes <joelagnelf@nvidia.com>]

On Fri, May 30, 2025 at 09:55:45AM +0800, Xiongfeng Wang wrote:
> Hi Joel,
> 
> On 2025/5/29 0:30, Joel Fernandes wrote:
> > On Wed, May 21, 2025 at 5:43 AM Xiongfeng Wang
> > <wangxiongfeng2@huawei.com> wrote:
> >>
> >> Hi RCU experts,
> >>
> >> When I ran syskaller in Linux 6.6 with CONFIG_PREEMPT_RCU enabled, I got
> >> the following soft lockup. The Calltrace is too long. I put it in the end.
> >> The issue can also be reproduced in the latest kernel.
> >>
> >> The issue is as follows. CPU3 is waiting for a spin_lock, which is got by CPU1.
> >> But CPU1 stuck in the following dead loop.
> >>
> >> irq_exit()
> >>   __irq_exit_rcu()
> >>     /* in_hardirq() returns false after this */
> >>     preempt_count_sub(HARDIRQ_OFFSET)
> >>     tick_irq_exit()
> >>       tick_nohz_irq_exit()
> >>             tick_nohz_stop_sched_tick()
> >>               trace_tick_stop()  /* a bpf prog is hooked on this trace point */
> >>                    __bpf_trace_tick_stop()
> >>                       bpf_trace_run2()
> >>                             rcu_read_unlock_special()
> >>                               /* will send a IPI to itself */
> >>                               irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
> >>
> >> /* after interrupt is enabled again, the irq_work is called */
> >> asm_sysvec_irq_work()
> >>   sysvec_irq_work()
> >> irq_exit() /* after handled the irq_work, we again enter into irq_exit() */
> >>   __irq_exit_rcu()
> >>     ...skip...
> >>            /* we queue a irq_work again, and enter a dead loop */
> >>            irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
> > 
> > This seems legitimate, Boqun and I were just talking about it. He may
> > share more thoughts but here are a few:
> > 
> > Maybe we can delay subsequent clearing of the flag in
> > rcu_preempt_deferred_qs_handler() using a timer and an exponential
> > back-off? That way we are not sending too many self-IPIs.
> > 
> > And reset the process at the end of a grace period.
> > 
> > Or just don't send subsequent self-IPIs if we just sent one for the
> > rdp. Chances are, if we did not get the scheduler's attention during
> > the first one, we may not in subsequent ones I think. Plus we do send
> > other IPIs already if the grace period was over extended (from the FQS
> > loop), maybe we can tweak that?
> 
> Thanks a lot for your reply. I think it's hard for me to fix this issue as
> above without introducing new bugs. I barely understand the RCU code. But I'm
> very glad to help test if you have any code modifiction need to. I have
> the VM and the syskaller benchmark which can reproduce the problem.

Sure, I understand. This is already incredibly valuable so thank you again.
Will request for your testing help soon. I also have a test module now which
can sort-off reproduce this. Keep you posted!

thanks,

 - Joel