Re: Fix 80d20d35af1e ("nohz: Fix local_timer_softirq_pending()") may have revealed another problem

[Frederic Weisbecker <frederic@kernel.org>]

On Fri, Dec 28, 2018 at 12:11:12AM +0100, Heiner Kallweit wrote:
> 
> OK, did as you advised and here comes the trace. That's the related dmesg part:
> 
> [ 1479.025092] x86: Booting SMP configuration:
> [ 1479.025129] smpboot: Booting Node 0 Processor 1 APIC 0x2
> [ 1479.094715] NOHZ: local_softirq_pending 202
> [ 1479.096557] smpboot: CPU 1 is now offline
> 
> Hope it helps.
> Heiner
> 
> 
> # tracer: nop
> #
> #                              _-----=> irqs-off
> #                             / _----=> need-resched
> #                            | / _---=> hardirq/softirq
> #                            || / _--=> preempt-depth
> #                            ||| /     delay
> #           TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION
> #              | |       |   ||||       |         |
[...]
>           <idle>-0     [001] d.h2  1479.111017: softirq_raise: vec=9 [action=RCU]
>           <idle>-0     [001] d.h2  1479.111026: softirq_raise: vec=7 [action=SCHED]
>           <idle>-0     [001] ..s2  1479.111035: softirq_entry: vec=1 [action=TIMER]
>           <idle>-0     [001] ..s2  1479.111040: softirq_exit: vec=1 [action=TIMER]
>           <idle>-0     [001] ..s2  1479.111040: softirq_entry: vec=7 [action=SCHED]
>           <idle>-0     [001] ..s2  1479.111052: softirq_exit: vec=7 [action=SCHED]
>           <idle>-0     [001] ..s2  1479.111052: softirq_entry: vec=9 [action=RCU]
>           <idle>-0     [001] .Ns2  1479.111079: softirq_exit: vec=9 [action=RCU]
>          cpuhp/1-13    [001] dNh2  1479.112930: softirq_raise: vec=1 [action=TIMER]
>          cpuhp/1-13    [001] dNh2  1479.112935: softirq_raise: vec=9 [action=RCU]

Interesting, the softirq is raised from hardirq but it's not handled in the end of
the IRQ. Are you running threaded IRQS by any chance? If so I would expect ksoftirqd
to handle the pending work before we go idle. However I can imagine a small window
where such an expectation may not be met: if the softirq is raised after the ksoftirqd
thread is parked (CPUHP_AP_SMPBOOT_THREADS), which is right before we disable the CPU
(CPUHP_TEARDOWN_CPU).

I don't know if we can afford to ignore a softirq even at this late stage. We should
probably avoid leaking any. So here is a possible fix, if you don't mind trying:

diff --git a/kernel/softirq.c b/kernel/softirq.c
index d288133..716096b 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -56,6 +56,7 @@ EXPORT_PER_CPU_SYMBOL(irq_stat);
 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
 
 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
+DEFINE_PER_CPU(int, ksoftirqd_parked);
 
 const char * const softirq_to_name[NR_SOFTIRQS] = {
 	"HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
@@ -363,7 +364,7 @@ static inline void invoke_softirq(void)
 	if (ksoftirqd_running(local_softirq_pending()))
 		return;
 
-	if (!force_irqthreads) {
+	if (!force_irqthreads || __this_cpu_read(ksoftirqd_parked)) {
 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
 		/*
 		 * We can safely execute softirq on the current stack if
@@ -659,6 +660,22 @@ static void run_ksoftirqd(unsigned int cpu)
 	local_irq_enable();
 }
 
+static void ksoftirqd_park(unsigned int cpu)
+{
+	local_irq_disable();
+	__this_cpu_write(ksoftirqd_parked, 1);
+
+	if (local_softirq_pending())
+		run_ksoftirqd(cpu);
+
+	local_irq_enable();
+}
+
+static void ksoftirqd_unpark(unsigned int cpu)
+{
+	__this_cpu_write(ksoftirqd_parked, 0);
+}
+
 #ifdef CONFIG_HOTPLUG_CPU
 /*
  * tasklet_kill_immediate is called to remove a tasklet which can already be
@@ -724,6 +741,8 @@ static int takeover_tasklets(unsigned int cpu)
 static struct smp_hotplug_thread softirq_threads = {
 	.store			= &ksoftirqd,
 	.thread_should_run	= ksoftirqd_should_run,
+	.park			= ksoftirqd_park,
+	.unpark			= ksoftirqd_unpark,
 	.thread_fn		= run_ksoftirqd,
 	.thread_comm		= "ksoftirqd/%u",
 };