Re: [PATCH] tick/rcu: fix false positive "softirq work is pending" messages on RT

[Wen Yang]

On 2023/8/19 04:07, paul.gortmaker@xxxxxxxxxxxxx wrote:
> From: Paul Gortmaker <paul.gortmaker@xxxxxxxxxxxxx>
>
> In commit 0345691b24c0 ("tick/rcu: Stop allowing RCU_SOFTIRQ in idle")
> the new function report_idle_softirq() was created by breaking code out
> of the existing can_stop_idle_tick() for kernels v5.18 and newer.
>
> In doing so, the code essentially went from a one conditional:
>
> 	if (a && b && c)
> 		warn();
>
> to a three conditional:
>
> 	if (!a)
> 		return;
> 	if (!b)
> 		return;
> 	if (!c)
> 		return;
> 	warn();
>
> However, it seems one of the conditionals didn't get a "!" removed.
> Compare the instance of local_bh_blocked() in the old code:
>
> -               if (ratelimit < 10 && !local_bh_blocked() &&
> -                   (local_softirq_pending() & SOFTIRQ_STOP_IDLE_MASK)) {
> -                       pr_warn("NOHZ tick-stop error: Non-RCU local softirq work is pending, handler #%02x!!!\n",
> -                               (unsigned int) local_softirq_pending());
> -                       ratelimit++;
> -               }
>
> ...to the usage in the new (5.18+) code:
>
> +       /* On RT, softirqs handling may be waiting on some lock */
> +       if (!local_bh_blocked())
> +               return false;
>
> It seems apparent that the "!" should be removed from the new code.
>
> This issue lay dormant until another fixup for the same commit was added
> in commit a7e282c77785 ("tick/rcu: Fix bogus ratelimit condition").
> This commit realized the ratelimit was essentially set to zero instead
> of ten, and hence *no* softirq pending messages would ever be issued.
>
> Once this commit was backported via linux-stable, both the v6.1 and v6.4
> preempt-rt kernels started printing out 10 instances of this at boot:
>
>    NOHZ tick-stop error: local softirq work is pending, handler #80!!!
>
> Just to double check my understanding of things, I confirmed that the
> v5.18-rt did print the pending-80 messages with a cherry pick of the
> ratelimit fix, and then confirmed no pending softirq messages were
> printed with a revert of mainline's 034569 on a v5.18-rt baseline.
>
> Finally I confirmed it fixed the issue on v6.1-rt and v6.4-rt, and
> also didn't break anything on a defconfig of mainline master of today.
>
> Fixes: 0345691b24c0 ("tick/rcu: Stop allowing RCU_SOFTIRQ in idle")
> Cc: Wen Yang <wenyang.linux@xxxxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Paul E. McKenney <paulmck@xxxxxxxxxx>
> Cc: Frederic Weisbecker <frederic@xxxxxxxxxx>
> Signed-off-by: Paul Gortmaker <paul.gortmaker@xxxxxxxxxxxxx>
>
> diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> index 2b865cb77feb..b52e1861b913 100644
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -1050,7 +1050,7 @@ static bool report_idle_softirq(void)
>   		return false;
>   
>   	/* On RT, softirqs handling may be waiting on some lock */
> -	if (!local_bh_blocked())
> +	if (local_bh_blocked())
>   		return false;
>   
>   	pr_warn("NOHZ tick-stop error: local softirq work is pending, handler #%02x!!!\n",


Good catch!

Reviewed-by: Wen Yang <wenyang.linux@xxxxxxxxxxx>

--
Thanks,
Wen