Re: [PATCH HACK RFC] cpu: Prevent late-arriving interrupts from disrupting offline

[Paul E. McKenney]

On Mon, Jun 03, 2019 at 03:39:18PM +0200, Dietmar Eggemann wrote:
> On 6/3/19 1:44 PM, Mark Rutland wrote:
> >On Mon, Jun 03, 2019 at 10:38:48AM +0200, Peter Zijlstra wrote:
> >>On Sat, Jun 01, 2019 at 06:12:53PM -0700, Paul E. McKenney wrote:
> >>>Scheduling-clock interrupts can arrive late in the CPU-offline process,
> >>>after idle entry and the subsequent call to cpuhp_report_idle_dead().
> >>>Once execution passes the call to rcu_report_dead(), RCU is ignoring
> >>>the CPU, which results in lockdep complaints when the interrupt handler
> >>>uses RCU:
> >>
> >>>diff --git a/kernel/cpu.c b/kernel/cpu.c
> >>>index 448efc06bb2d..3b33d83b793d 100644
> >>>--- a/kernel/cpu.c
> >>>+++ b/kernel/cpu.c
> >>>@@ -930,6 +930,7 @@ void cpuhp_report_idle_dead(void)
> >>>  	struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
> >>>  	BUG_ON(st->state != CPUHP_AP_OFFLINE);
> >>>+	local_irq_disable();
> >>>  	rcu_report_dead(smp_processor_id());
> >>>  	st->state = CPUHP_AP_IDLE_DEAD;
> >>>  	udelay(1000);
> >>
> >>Urgh... I'd almost suggest we do something like the below.
> >>
> >>
> >>But then I started looking at the various arch_cpu_idle_dead()
> >>implementations and ran into arm's implementation, which is calling
> >>complete() where generic code already established this isn't possible
> >>(see for example cpuhp_report_idle_dead()).
> >
> >IIRC, that should have been migrated over to cpu_report_death(), as
> >arm64 was in commit:
> >
> >   05981277a4de1ad6 ("arm64: Use common outgoing-CPU-notification code")
> >
> >... but it looks like Paul's patch to do so [1] fell through the cracks;
> >I'm not aware of any reason that shouldn't have been taken.
> >[1] https://lore.kernel.org/lkml/1431467407-1223-8-git-send-email-paulmck@xxxxxxxxxxxxxxxxxx/
> >
> >Paul, do you want to resend that?
> 
> Please do. We're carrying this patch out-of-tree for while now in
> our EAS integration to get cpu hotplug tests passing on TC2 (arm).

Huh.  It still applies.  But I have no means of testing it.

And it looks like the reason I dropped it was that I didn't get any
response from the maintainer.  I sent a message to this effect to
linux-arm-kernel@xxxxxxxxxxxxxxxxxxx and linux@xxxxxxxxxxxxxxxx on May
21, 2015.

So here it is again.  ;-)

I have queued this locally.  Left to myself, I add the two of you on its
Cc: list and run it through my normal process.  But given the history,
I would still want either an ack from the maintainer or, better, for
the maintainer to take the patch.

Or is there a better way for us to proceed on this?

							Thanx, Paul

------------------------------------------------------------------------

arm: Use common outgoing-CPU-notification code

This commit removes the open-coded CPU-offline notification with new
common code.  In particular, this change avoids calling scheduler code
using RCU from an offline CPU that RCU is ignoring.  This is a minimal
change.  A more intrusive change might invoke the cpu_check_up_prepare()
and cpu_set_state_online() functions at CPU-online time, which would
allow onlining throw an error if the CPU did not go offline properly.

Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
Cc: Russell King <linux@xxxxxxxxxxxxxxxx>

diff --git a/arch/arm/kernel/smp.c b/arch/arm/kernel/smp.c
index ebc53804d57b..8687d619260f 100644
--- a/arch/arm/kernel/smp.c
+++ b/arch/arm/kernel/smp.c
@@ -267,15 +267,13 @@ int __cpu_disable(void)
 	return 0;
 }
 
-static DECLARE_COMPLETION(cpu_died);
-
 /*
  * called on the thread which is asking for a CPU to be shutdown -
  * waits until shutdown has completed, or it is timed out.
  */
 void __cpu_die(unsigned int cpu)
 {
-	if (!wait_for_completion_timeout(&cpu_died, msecs_to_jiffies(5000))) {
+	if (!cpu_wait_death(cpu, 5)) {
 		pr_err("CPU%u: cpu didn't die\n", cpu);
 		return;
 	}
@@ -322,7 +320,7 @@ void arch_cpu_idle_dead(void)
 	 * this returns, power and/or clocks can be removed at any point
 	 * from this CPU and its cache by platform_cpu_kill().
 	 */
-	complete(&cpu_died);
+	(void)cpu_report_death();
 
 	/*
 	 * Ensure that the cache lines associated with that completion are