Skip to contentSkip to conversation

kthread: Fix the race condition when kthread is parked

PATCH Daniel J Blueman 2 participants 2 messages Nov 2, 2014 – Nov 3, 2014

lore
Share & tools
lore

Daniel J Blueman

author Tested-by
#0 lore raw
Quoted 195 lines · Thomas Gleixner
Quoted 21 lines · Subbaraman Narayanamurthy

While stressing the CPU hotplug path, sometimes we hit a problem as shown below.

[57056.416774] ------------[ cut here ]------------
[57056.489232] ksoftirqd/1 (14): undefined instruction: pc=c01931e8
[57056.489245] Code: e594a000 eb085236 e15a0000 0a000000 (e7f001f2)
[57056.489259] ------------[ cut here ]------------
[57056.492840] kernel BUG at kernel/kernel/smpboot.c:134!
[57056.513236] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP ARM
[57056.519055] Modules linked in: wlan(O) mhi(O)
[57056.523394] CPU: 0 PID: 14 Comm: ksoftirqd/1 Tainted: G        W O
3.10.0-g3677c61-00008-g180c060 #1
[57056.532595] task: f0c8b000 ti: f0e78000 task.ti: f0e78000
[57056.537991] PC is at smpboot_thread_fn+0x124/0x218
[57056.542750] LR is at smpboot_thread_fn+0x11c/0x218
[57056.547528] pc : [<c01931e8>]    lr : [<c01931e0>] psr: 200f0013
[57056.547528] sp : f0e79f30  ip : 00000000  fp : 00000000
[57056.558983] r10: 00000001  r9 : 00000000  r8 : f0e78000
[57056.564192] r7 : 00000001  r6 : c1195758  r5 : f0e78000  r4 : f0e5fd00
[57056.570701] r3 : 00000001  r2 : f0e79f20  r1 : 00000000  r0 : 00000000
Quoted 4 lines

This issue was always seen in the context of "ksoftirqd". It seems to be happening because of a potential race condition in __kthread_parkme where just after completing the parked completion, before the ksoftirqd task has been scheduled again, it can go into running state.

This explanation does not make any sense. You completely fail to explain the details of the race. And your patch does not make any sense either, because the real issue is this:

Task CPU 0 CPU 1

T1 unplug cpu1

       kthread_park(T2)
       set_bit(KTHREAD_SHOULD_PARK);
	  wait_for_completion()
T2					parkme(X)
				   	  __set_current_state(TASK_PARKED);
				   	  while (test_bit(KTHREAD_SHOULD_PARK)) {
				     	    if (!test_and_set_bit(KTHREAD_IS_PARKED))
				              complete();
			             	    schedule();

T1 plug cpu1

--> premature wakeup of T2, i.e. before unpark, so T2 gets scheduled on
CPU 0

T2 __set_current_state(TASK_PARKED);

--> Preemption by the plug thread

T1 thread_unpark(T2)

         clear_bit(KTHREAD_SHOULD_PARK);

--> Preemption by the softirq thread which breaks out of the
while(test_bit(KTHREAD_SHOULD_PARK)) loop because KTHREAD_SHOULD_PARK is not longer set.

T2 }

    clear_bit(KTHREAD_IS_PARKED);

--> Now T2 happily continues to run on CPU0 which rightfully casues
the BUG to trigger.

T1

	__kthread_bind(T2)

--> Too late ....

So the real issue is that the park/unpark code is not able to handle the premature wakeup of T2 and that needs to be fixed.

Your changelog says:

Quoted 1 line

It seems to be happening because of a potential race condition in

Potential is wrong to begin with. A race condition is either real and explainable or it does not exist.

Quoted 3 lines

__kthread_parkme where just after completing the parked completion, before the ksoftirqd task has been scheduled again, it can go into running state.

What exactly has this to do with state RUNNING or PARKED?

Nothing, the task state is completely irrelevant as the real issue is the task->*PARK flags state.

So what is your patch solving here ?

You put a wait for task->state == TASK_PARKED after the wait_for_completion. What does it solve? Actually nothing. It just changes the propability of that issue. Go and apply it between any step of the above and figure out what it solves. Nothing, really.

Now just as an extra thought experiment assume that you have only two cpus and T1 is a SCHED_FIFO task and T2 is SCHED_OTHER ....

Please do not misunderstand me, but "fixing" races without proper understanding them is plain wrong. Providing a vague changelog which does neither explain what the issue is and why the fix works is even more wrong.

The next time you hit something like this, please take the time and sit down, get out the old fashioned piece of paper and a pencil and draw the picture so you can actually understand what the root cause of the observed issue is before sending out halfarsed duct tape fixes which just paper over the root cause. If you cannot figure it out, send a proper bug report.

Thanks,

tglx
------------------>

Subject: kthread: Plug park/ unplug race
From: Thomas Gleixner <[email protected]>
Date: Thu, 26 Jun 2014 01:24:36 +0200

The kthread park/unpark logic has the following issue:

Task CPU 0 CPU 1

T1 unplug cpu1

       kthread_park(T2)
       set_bit(KTHREAD_SHOULD_PARK);
	  wait_for_completion()
T2					parkme(X)
				   	  __set_current_state(TASK_PARKED);
				   	  while (test_bit(KTHREAD_SHOULD_PARK)) {
				     	    if (!test_and_set_bit(KTHREAD_IS_PARKED))
				              complete();
			             	    schedule();

T1 plug cpu1

--> premature wakeup of T2, i.e. before unpark, so T2 gets scheduled on
CPU 0

T2 __set_current_state(TASK_PARKED);

--> Preemption by the plug thread

T1 thread_unpark(T2)

         clear_bit(KTHREAD_SHOULD_PARK);

--> Preemption by the softirq thread which breaks out of the
while(test_bit(KTHREAD_SHOULD_PARK)) loop because KTHREAD_SHOULD_PARK is not longer set.

T2 }

    clear_bit(KTHREAD_IS_PARKED);

--> Now T2 happily continues to run on CPU0 which rightfully causes

    the BUG_ON(T2->cpu != smp_processor_id()) to trigger.

T1

	__kthread_bind(T2)

--> Too late ....

Reorder the logic so that the unplug code binds the thread to the target cpu before clearing the KTHREAD_SHOULD_PARK bit.

---
 kernel/kthread.c |   14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

Index: linux/kernel/kthread.c
===================================================================

kernel/kthread.c +9 -0
@@ -382,6 +382,15 @@ struct task_struct *kthread_create_on_cp
 
 static void __kthread_unpark(struct task_struct *k, struct kthread *kthread)
 {
+	/*
+	 * Rebind the thread to the target cpu first if it is a per
+	 * cpu thread unconditionally because it must be bound to the
+	 * target cpu before it can observe the KTHREAD_SHOULD_PARK
+	 * bit cleared.
+	 */
+	if (test_bit(KTHREAD_IS_PER_CPU, &kthread->flags))
+		__kthread_bind(k, kthread->cpu, TASK_PARKED);
+
 	clear_bit(KTHREAD_SHOULD_PARK, &kthread->flags);
 	/*
 	 * We clear the IS_PARKED bit here as we don't wait
@@ -389,11 +398,8 @@ static void __kthread_unpark(struct task
 	 * park before that happens we'd see the IS_PARKED bit
 	 * which might be about to be cleared.

kernel/kthread.c static void __kthread_unpark(struct task · line 405

Quoted 10 lines
kernel/kthread.c +1 -4
 	 */
-	if (test_and_clear_bit(KTHREAD_IS_PARKED, &kthread->flags)) {
-		if (test_bit(KTHREAD_IS_PER_CPU, &kthread->flags))
-			__kthread_bind(k, kthread->cpu, TASK_PARKED);
+	if (test_and_clear_bit(KTHREAD_IS_PARKED, &kthread->flags))
 		wake_up_state(k, TASK_PARKED);
-	}
 }
 
 /**

I just got a window to test this, and it reliably addresses the boot-time core onlining race that we've seen occasionally on a 2000-core customer system. Splendid work, Thomas.

Many thanks,
Daniel

Signature · 2 lines
Daniel J Blueman
Principal Software Engineer, Numascale