Confusing lockdep splat

Confusing lockdep splat

[Paul E. McKenney]

Hello!

I got the lockdep splat below from an SRCU-T rcutorture run, which uses
a !SMP !PREEMPT kernel.  This is a random event, and about half the time
it happens within an hour or two.  My reproducer (on current -rcu "dev"
branch for a 16-CPU system) is:

	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200

My points of confusion are as follows:

1.	The locks involved in this deadlock cycle are irq-disabled
	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
	There is only one CPU.	There is no possibility of preemption
	or interrupts.	So how can this deadlock actually happen?

2.	If there was more than one CPU, then yes, there would be
	a deadlock.  The PI lock is acquired by the wakeup code after
	acquiring the workqueue lock, and rcutorture tests the new ability
	of the scheduler to hold the PI lock across rcu_read_unlock(),
	and while it is at it, across the rest of the unlock primitives.

	But if there was more than one CPU, Tree SRCU would be used
	instead of Tiny SRCU, and there would be no wakeup invoked from
	srcu_read_unlock().

	Given only one CPU, there is no way to complete the deadlock
	cycle.

For now, I am working around this by preventing rcutorture from holding
the PI lock across Tiny srcu_read_unlock().

Am I missing something subtle here?

						Thanx, Paul

======================================================
WARNING: possible circular locking dependency detected
5.15.0-rc1+ #3766 Not tainted
------------------------------------------------------
rcu_torture_rea/53 is trying to acquire lock:
ffffffffa074e6a8 (srcu_ctl.srcu_wq.lock){....}-{2:2}, at: swake_up_one+0xa/0x30

but task is already holding lock:
ffffa03502479d80 (&p->pi_lock){-.-.}-{2:2}, at: rcutorture_one_extend+0x370/0x400

which lock already depends on the new lock.


the existing dependency chain (in reverse order) is:

-> #1 (&p->pi_lock){-.-.}-{2:2}:
       _raw_spin_lock_irqsave+0x2f/0x50
       try_to_wake_up+0x50/0x580
       swake_up_locked.part.7+0xe/0x30
       swake_up_one+0x22/0x30
       rcutorture_one_extend+0x1b6/0x400
       rcu_torture_one_read+0x290/0x5d0
       rcu_torture_reader+0xac/0x200
       kthread+0x12d/0x150
       ret_from_fork+0x22/0x30

-> #0 (srcu_ctl.srcu_wq.lock){....}-{2:2}:
       __lock_acquire+0x130c/0x2440
       lock_acquire+0xc2/0x270
       _raw_spin_lock_irqsave+0x2f/0x50
       swake_up_one+0xa/0x30
       rcutorture_one_extend+0x387/0x400
       rcu_torture_one_read+0x290/0x5d0
       rcu_torture_reader+0xac/0x200
       kthread+0x12d/0x150
       ret_from_fork+0x22/0x30

other info that might help us debug this:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&p->pi_lock);
                               lock(srcu_ctl.srcu_wq.lock);
                               lock(&p->pi_lock);
  lock(srcu_ctl.srcu_wq.lock);

 *** DEADLOCK ***

1 lock held by rcu_torture_rea/53:
 #0: ffffa03502479d80 (&p->pi_lock){-.-.}-{2:2}, at: rcutorture_one_extend+0x370/0x400

stack backtrace:
CPU: 0 PID: 53 Comm: rcu_torture_rea Not tainted 5.15.0-rc1+ #3766
Hardware name: Red Hat KVM/RHEL-AV, BIOS 1.13.0-2.module_el8.5.0+746+bbd5d70c 04/01/2014
Call Trace:
 check_noncircular+0xfe/0x110
 ? find_held_lock+0x2d/0x90
 __lock_acquire+0x130c/0x2440
 lock_acquire+0xc2/0x270
 ? swake_up_one+0xa/0x30
 ? find_held_lock+0x72/0x90
 _raw_spin_lock_irqsave+0x2f/0x50
 ? swake_up_one+0xa/0x30
 swake_up_one+0xa/0x30
 rcutorture_one_extend+0x387/0x400
 rcu_torture_one_read+0x290/0x5d0
 rcu_torture_reader+0xac/0x200
 ? rcutorture_oom_notify+0xf0/0xf0
 ? __kthread_parkme+0x61/0x90
 ? rcu_torture_one_read+0x5d0/0x5d0
 kthread+0x12d/0x150
 ? set_kthread_struct+0x40/0x40
 ret_from_fork+0x22/0x30

Re: Confusing lockdep splat

[Waiman Long]

On 9/24/21 5:02 PM, Paul E. McKenney wrote:
> Hello!
>
> I got the lockdep splat below from an SRCU-T rcutorture run, which uses
> a !SMP !PREEMPT kernel.  This is a random event, and about half the time
> it happens within an hour or two.  My reproducer (on current -rcu "dev"
> branch for a 16-CPU system) is:
>
> 	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200
>
> My points of confusion are as follows:
>
> 1.	The locks involved in this deadlock cycle are irq-disabled
> 	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
> 	There is only one CPU.	There is no possibility of preemption
> 	or interrupts.	So how can this deadlock actually happen?
>
> 2.	If there was more than one CPU, then yes, there would be
> 	a deadlock.  The PI lock is acquired by the wakeup code after
> 	acquiring the workqueue lock, and rcutorture tests the new ability
> 	of the scheduler to hold the PI lock across rcu_read_unlock(),
> 	and while it is at it, across the rest of the unlock primitives.
>
> 	But if there was more than one CPU, Tree SRCU would be used
> 	instead of Tiny SRCU, and there would be no wakeup invoked from
> 	srcu_read_unlock().
>
> 	Given only one CPU, there is no way to complete the deadlock
> 	cycle.
>
> For now, I am working around this by preventing rcutorture from holding
> the PI lock across Tiny srcu_read_unlock().
>
> Am I missing something subtle here?

I would say that the lockdep code just doesn't have enough intelligence 
to identify that deadlock is not possible in this special case. There 
are certainly false positives, and it can be hard to get rid of them.

Cheers,
Longman

Re: Confusing lockdep splat

[Paul E. McKenney]

On Fri, Sep 24, 2021 at 05:41:17PM -0400, Waiman Long wrote:
> On 9/24/21 5:02 PM, Paul E. McKenney wrote:
> > Hello!
> > 
> > I got the lockdep splat below from an SRCU-T rcutorture run, which uses
> > a !SMP !PREEMPT kernel.  This is a random event, and about half the time
> > it happens within an hour or two.  My reproducer (on current -rcu "dev"
> > branch for a 16-CPU system) is:
> > 
> > 	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200
> > 
> > My points of confusion are as follows:
> > 
> > 1.	The locks involved in this deadlock cycle are irq-disabled
> > 	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
> > 	There is only one CPU.	There is no possibility of preemption
> > 	or interrupts.	So how can this deadlock actually happen?
> > 
> > 2.	If there was more than one CPU, then yes, there would be
> > 	a deadlock.  The PI lock is acquired by the wakeup code after
> > 	acquiring the workqueue lock, and rcutorture tests the new ability
> > 	of the scheduler to hold the PI lock across rcu_read_unlock(),
> > 	and while it is at it, across the rest of the unlock primitives.
> > 
> > 	But if there was more than one CPU, Tree SRCU would be used
> > 	instead of Tiny SRCU, and there would be no wakeup invoked from
> > 	srcu_read_unlock().
> > 
> > 	Given only one CPU, there is no way to complete the deadlock
> > 	cycle.
> > 
> > For now, I am working around this by preventing rcutorture from holding
> > the PI lock across Tiny srcu_read_unlock().
> > 
> > Am I missing something subtle here?
> 
> I would say that the lockdep code just doesn't have enough intelligence to
> identify that deadlock is not possible in this special case. There are
> certainly false positives, and it can be hard to get rid of them.

Would it make sense for lockdep to filter out reports involving more
than one CPU unless there is at least one sleeplock in the cycle?

Of course, it gets more complicated when interrupts are involved...

							Thanx, Paul

Re: Confusing lockdep splat

[Waiman Long]

On 9/24/21 6:43 PM, Paul E. McKenney wrote:
> On Fri, Sep 24, 2021 at 05:41:17PM -0400, Waiman Long wrote:
>> On 9/24/21 5:02 PM, Paul E. McKenney wrote:
>>> Hello!
>>>
>>> I got the lockdep splat below from an SRCU-T rcutorture run, which uses
>>> a !SMP !PREEMPT kernel.  This is a random event, and about half the time
>>> it happens within an hour or two.  My reproducer (on current -rcu "dev"
>>> branch for a 16-CPU system) is:
>>>
>>> 	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200
>>>
>>> My points of confusion are as follows:
>>>
>>> 1.	The locks involved in this deadlock cycle are irq-disabled
>>> 	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
>>> 	There is only one CPU.	There is no possibility of preemption
>>> 	or interrupts.	So how can this deadlock actually happen?
>>>
>>> 2.	If there was more than one CPU, then yes, there would be
>>> 	a deadlock.  The PI lock is acquired by the wakeup code after
>>> 	acquiring the workqueue lock, and rcutorture tests the new ability
>>> 	of the scheduler to hold the PI lock across rcu_read_unlock(),
>>> 	and while it is at it, across the rest of the unlock primitives.
>>>
>>> 	But if there was more than one CPU, Tree SRCU would be used
>>> 	instead of Tiny SRCU, and there would be no wakeup invoked from
>>> 	srcu_read_unlock().
>>>
>>> 	Given only one CPU, there is no way to complete the deadlock
>>> 	cycle.
>>>
>>> For now, I am working around this by preventing rcutorture from holding
>>> the PI lock across Tiny srcu_read_unlock().
>>>
>>> Am I missing something subtle here?
>> I would say that the lockdep code just doesn't have enough intelligence to
>> identify that deadlock is not possible in this special case. There are
>> certainly false positives, and it can be hard to get rid of them.
> Would it make sense for lockdep to filter out reports involving more
> than one CPU unless there is at least one sleeplock in the cycle?
>
> Of course, it gets more complicated when interrupts are involved...

Actually, lockdep keeps track of all the possible lock orderings and put 
out a splat whenever these lock orderings suggest that a circular 
deadlock is possible. It doesn't keep track if a lock is sleepable or 
not. Also lockdep deals with lock classes each of which can have many 
instances. So all the pi_lock's in different task_struct's are all 
treated as the same lock from lockdep's perspective. We can't treat all 
different instances separately or we will run out of lockdep table space 
very quickly.

Cheers,
Longman

Re: Confusing lockdep splat

[Paul E. McKenney]

On Fri, Sep 24, 2021 at 08:30:20PM -0400, Waiman Long wrote:
> On 9/24/21 6:43 PM, Paul E. McKenney wrote:
> > On Fri, Sep 24, 2021 at 05:41:17PM -0400, Waiman Long wrote:
> > > On 9/24/21 5:02 PM, Paul E. McKenney wrote:
> > > > Hello!
> > > > 
> > > > I got the lockdep splat below from an SRCU-T rcutorture run, which uses
> > > > a !SMP !PREEMPT kernel.  This is a random event, and about half the time
> > > > it happens within an hour or two.  My reproducer (on current -rcu "dev"
> > > > branch for a 16-CPU system) is:
> > > > 
> > > > 	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200
> > > > 
> > > > My points of confusion are as follows:
> > > > 
> > > > 1.	The locks involved in this deadlock cycle are irq-disabled
> > > > 	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
> > > > 	There is only one CPU.	There is no possibility of preemption
> > > > 	or interrupts.	So how can this deadlock actually happen?
> > > > 
> > > > 2.	If there was more than one CPU, then yes, there would be
> > > > 	a deadlock.  The PI lock is acquired by the wakeup code after
> > > > 	acquiring the workqueue lock, and rcutorture tests the new ability
> > > > 	of the scheduler to hold the PI lock across rcu_read_unlock(),
> > > > 	and while it is at it, across the rest of the unlock primitives.
> > > > 
> > > > 	But if there was more than one CPU, Tree SRCU would be used
> > > > 	instead of Tiny SRCU, and there would be no wakeup invoked from
> > > > 	srcu_read_unlock().
> > > > 
> > > > 	Given only one CPU, there is no way to complete the deadlock
> > > > 	cycle.
> > > > 
> > > > For now, I am working around this by preventing rcutorture from holding
> > > > the PI lock across Tiny srcu_read_unlock().
> > > > 
> > > > Am I missing something subtle here?
> > > I would say that the lockdep code just doesn't have enough intelligence to
> > > identify that deadlock is not possible in this special case. There are
> > > certainly false positives, and it can be hard to get rid of them.
> > Would it make sense for lockdep to filter out reports involving more
> > than one CPU unless there is at least one sleeplock in the cycle?
> > 
> > Of course, it gets more complicated when interrupts are involved...
> 
> Actually, lockdep keeps track of all the possible lock orderings and put out
> a splat whenever these lock orderings suggest that a circular deadlock is
> possible. It doesn't keep track if a lock is sleepable or not. Also lockdep
> deals with lock classes each of which can have many instances. So all the
> pi_lock's in different task_struct's are all treated as the same lock from
> lockdep's perspective. We can't treat all different instances separately or
> we will run out of lockdep table space very quickly.

We shouldn't need additional classes, but only instead a marking of a
given lock class to tell whether or not it was a sleeplock.  Either way,
I now have a workaround within Tiny SRCU that appears to handle this case,
so it is not as urgent as it might be.

							Thanx, Paul

Re: Confusing lockdep splat

[Boqun Feng]

On Fri, Sep 24, 2021 at 03:43:37PM -0700, Paul E. McKenney wrote:
> On Fri, Sep 24, 2021 at 05:41:17PM -0400, Waiman Long wrote:
> > On 9/24/21 5:02 PM, Paul E. McKenney wrote:
> > > Hello!
> > > 
> > > I got the lockdep splat below from an SRCU-T rcutorture run, which uses
> > > a !SMP !PREEMPT kernel.  This is a random event, and about half the time
> > > it happens within an hour or two.  My reproducer (on current -rcu "dev"
> > > branch for a 16-CPU system) is:
> > > 
> > > 	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200
> > > 
> > > My points of confusion are as follows:
> > > 
> > > 1.	The locks involved in this deadlock cycle are irq-disabled
> > > 	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
> > > 	There is only one CPU.	There is no possibility of preemption
> > > 	or interrupts.	So how can this deadlock actually happen?
> > > 
> > > 2.	If there was more than one CPU, then yes, there would be
> > > 	a deadlock.  The PI lock is acquired by the wakeup code after
> > > 	acquiring the workqueue lock, and rcutorture tests the new ability
> > > 	of the scheduler to hold the PI lock across rcu_read_unlock(),
> > > 	and while it is at it, across the rest of the unlock primitives.
> > > 
> > > 	But if there was more than one CPU, Tree SRCU would be used
> > > 	instead of Tiny SRCU, and there would be no wakeup invoked from
> > > 	srcu_read_unlock().
> > > 
> > > 	Given only one CPU, there is no way to complete the deadlock
> > > 	cycle.
> > > 
> > > For now, I am working around this by preventing rcutorture from holding
> > > the PI lock across Tiny srcu_read_unlock().
> > > 
> > > Am I missing something subtle here?
> > 
> > I would say that the lockdep code just doesn't have enough intelligence to
> > identify that deadlock is not possible in this special case. There are
> > certainly false positives, and it can be hard to get rid of them.
> 
> Would it make sense for lockdep to filter out reports involving more
> than one CPU unless there is at least one sleeplock in the cycle?
> 

I think SRCU is special here, because it has different implementations
in SMP and UP. For other code, if the implemenation in SMP and UP is the
same, we want lockdep to detect the deadlock even if it's not in UP.

We can provide an annotation similar to data_race() for SRCU to mark
UP-only code

	#define LOCKDEP_UP_ONLY(expr) ({			\
		BUILD_BUG_ON(IS_ENABLED(CONFIG_SMP));		\

		<disable lockdep>
		<...> v = expr;
		<enable lockdep>
		v
	})

and in __srcu_read_unlock():

	LOCKDEP_UP_ONLY(swake_up_one(...));

Thoughts?

Regards,
Boqun

> Of course, it gets more complicated when interrupts are involved...
> 
> 							Thanx, Paul

Re: Confusing lockdep splat

[Paul E. McKenney]

On Sat, Sep 25, 2021 at 10:38:28AM +0800, Boqun Feng wrote:
> On Fri, Sep 24, 2021 at 03:43:37PM -0700, Paul E. McKenney wrote:
> > On Fri, Sep 24, 2021 at 05:41:17PM -0400, Waiman Long wrote:
> > > On 9/24/21 5:02 PM, Paul E. McKenney wrote:
> > > > Hello!
> > > > 
> > > > I got the lockdep splat below from an SRCU-T rcutorture run, which uses
> > > > a !SMP !PREEMPT kernel.  This is a random event, and about half the time
> > > > it happens within an hour or two.  My reproducer (on current -rcu "dev"
> > > > branch for a 16-CPU system) is:
> > > > 
> > > > 	tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 16 --configs "16*SRCU-T" --duration 7200
> > > > 
> > > > My points of confusion are as follows:
> > > > 
> > > > 1.	The locks involved in this deadlock cycle are irq-disabled
> > > > 	raw spinlocks.	The claimed deadlock cycle uses two CPUs.
> > > > 	There is only one CPU.	There is no possibility of preemption
> > > > 	or interrupts.	So how can this deadlock actually happen?
> > > > 
> > > > 2.	If there was more than one CPU, then yes, there would be
> > > > 	a deadlock.  The PI lock is acquired by the wakeup code after
> > > > 	acquiring the workqueue lock, and rcutorture tests the new ability
> > > > 	of the scheduler to hold the PI lock across rcu_read_unlock(),
> > > > 	and while it is at it, across the rest of the unlock primitives.
> > > > 
> > > > 	But if there was more than one CPU, Tree SRCU would be used
> > > > 	instead of Tiny SRCU, and there would be no wakeup invoked from
> > > > 	srcu_read_unlock().
> > > > 
> > > > 	Given only one CPU, there is no way to complete the deadlock
> > > > 	cycle.
> > > > 
> > > > For now, I am working around this by preventing rcutorture from holding
> > > > the PI lock across Tiny srcu_read_unlock().
> > > > 
> > > > Am I missing something subtle here?
> > > 
> > > I would say that the lockdep code just doesn't have enough intelligence to
> > > identify that deadlock is not possible in this special case. There are
> > > certainly false positives, and it can be hard to get rid of them.
> > 
> > Would it make sense for lockdep to filter out reports involving more
> > than one CPU unless there is at least one sleeplock in the cycle?
> 
> I think SRCU is special here, because it has different implementations
> in SMP and UP. For other code, if the implemenation in SMP and UP is the
> same, we want lockdep to detect the deadlock even if it's not in UP.

Ah, fair point!  There are a few others, for example, kernel/up.c, but
it seems to just disable interrupts as its "big UP kernel lock".

> We can provide an annotation similar to data_race() for SRCU to mark
> UP-only code
> 
> 	#define LOCKDEP_UP_ONLY(expr) ({			\
> 		BUILD_BUG_ON(IS_ENABLED(CONFIG_SMP));		\
> 
> 		<disable lockdep>
> 		<...> v = expr;
> 		<enable lockdep>
> 		v
> 	})
> 
> and in __srcu_read_unlock():
> 
> 	LOCKDEP_UP_ONLY(swake_up_one(...));
> 
> Thoughts?

With the workaround I have now, all is well unless someone needs to hold
a PI lock across an rcu_read_unlock(), which seems unlikely.  If such a
case does arise, lockdep will let us know.  In which case what you are
suggesting might be a good way to go.  Alternatively, I could use the
trick that RCU Tasks Trace uses, with the swake_up_one() deferred to an
irq_work_queue() handler.

It does appear that !SMP kernels are nowhere near as important to the
community as they were 20 years ago.  ;-)

							Thanx, Paul