Re: [PATCH 0/6] sched: TTWU, IPI, and assorted stuff
From: Paul E. McKenney
Date: Fri Jun 19 2020 - 13:20:50 EST
On Fri, Jun 19, 2020 at 03:44:23PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 16, 2020 at 07:17:21PM +0200, Peter Zijlstra wrote:
> > On Tue, Jun 16, 2020 at 07:04:10PM +0200, Peter Zijlstra wrote:
> > > [19324.795303] ------------[ cut here ]------------
> > > [19324.795304] WARNING: CPU: 10 PID: 76 at kernel/smp.c:138 __smp_call_single_queue+0x40/0x50
> > > [19324.795305] Modules linked in:
> > > [19324.795306] CPU: 10 PID: 76 Comm: ksoftirqd/10 Not tainted 5.8.0-rc1+ #8
> > > [19324.795307] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.12.0-1 04/01/2014
> > > [19324.795307] RIP: 0010:__smp_call_single_queue+0x40/0x50
> > > [19324.795308] Code: c2 40 91 02 00 4c 89 e6 4c 89 e7 48 03 14 c5 e0 56 2d b4 e8 b2 3a 2f 00 84 c0 75 04 5d 41 5c c3 89 ef 5d 41 5c e9 40 af f9 ff <0f> 0b eb cd 66 66 2e 0f 1f 84 00 00 00 00 00 90 41 54 49 89 f4 55
> > > [19324.795309] RSP: 0000:ffffb3cb4030bd18 EFLAGS: 00010046
> > > [19324.795310] RAX: 000000000000000a RBX: 0000000000000000 RCX: 00000000ffffffff
> > > [19324.795310] RDX: 00000000000090aa RSI: ffffffffb420bc3f RDI: ffffffffb4232e3e
> > > [19324.795311] RBP: 000000000000000a R08: 00001193646cd91c R09: ffff93c1df49c008
> > > [19324.795312] R10: ffffb3cb4030bdf8 R11: 000000000000032e R12: ffff93c1dbed5b30
> > > [19324.795312] R13: ffff93c1df4a8340 R14: 000000000000000a R15: ffff93c1df2e8340
> > > [19324.795313] FS: 0000000000000000(0000) GS:ffff93c1df480000(0000) knlGS:0000000000000000
> > > [19324.795313] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > [19324.795314] CR2: 00000000ffffffff CR3: 000000001e40a000 CR4: 00000000000006e0
> > > [19324.795315] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > [19324.795315] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > [19324.795316] Call Trace:
> > > [19324.795316] ttwu_queue_wakelist+0xa4/0xc0
> > > [19324.795316] try_to_wake_up+0x432/0x530
> >
> > This is indeed WF_ON_CPU... it had to be, but how ?!
>
> So my latest theory is that we have a memory ordering problem. It would
> fully explain the thing, but it would also render my patch #1
> insufficient.
>
> If we suppose the: task_cpu(p) load at the beginning of try_to_wake_up()
> returns an old value, and this old value happens to be this_cpu. Further
> assume that the p->on_cpu load accurately returns 1, it really is still
> running, just not here.
>
> Then, when we issue a local wakeup, we can crash in exactly the observed
> manner because p->se.cfs_rq != rq->cfs_rq, because p's cfs_rq is from
> the wrong CPU, therefore we'll iterate into the non-existant parents and
> NULL deref.
>
> The scenario is somewhat elaborate:
>
>
> X->cpu = 1
> rq(1)->curr = X
>
>
> CPU0 CPU1 CPU2
>
> // switch away from X
> LOCK rq(1)->lock
> smp_mb__after_spinlock
> dequeue_task(X)
> X->on_rq = 9
> switch_to(Z)
> X->on_cpu = 0
> UNLOCK rq(1)->lock
>
>
> // migrate X to cpu 0
> LOCK rq(1)->lock
> dequeue_task(X)
> set_task_cpu(X, 0)
> X->cpu = 0
> UNLOCK rq(1)->lock
>
> LOCK rq(0)->lock
> enqueue_task(X)
> X->on_rq = 1
> UNLOCK rq(0)->lock
>
> // switch to X
> LOCK rq(0)->lock
> smp_mb__after_spinlock
> switch_to(X)
> X->on_cpu = 1
> UNLOCK rq(0)->lock
>
> // X goes sleep
> X->state = TASK_UNINTERRUPTIBLE
> smp_mb(); // wake X
> ttwu()
> LOCK X->pi_lock
> smp_mb__after_spinlock
>
> if (p->state)
>
> cpu = X->cpu; // =? 1
>
> smp_rmb()
>
> // X calls schedule()
> LOCK rq(0)->lock
> smp_mb__after_spinlock
> dequeue_task(X)
> X->on_rq = 0
>
> if (p->on_rq)
>
> smp_rmb();
>
> if (p->on_cpu && ttwu_queue_wakelist(..)) [*]
>
> smp_cond_load_acquire(&p->on_cpu, !VAL)
>
> cpu = select_task_rq(X, X->wake_cpu, ...)
> if (X->cpu != cpu)
> switch_to(Y)
> X->on_cpu = 0
> UNLOCK rq(0)->lock
>
>
> Furthermore, without the fancy new path [*] we would have hit
> smp_cond_load_acquire(), and if we _really_ would have had ->on_cpu==1
> and cpu==this_cpu there, that'd have been a deadlock, but no such
> deadlocks have ever been observed.
>
> Also, note how the rest of the code never actually uses the @cpu value
> loaded earlier, all that is re-loaded after the load_aquire of
> X->on_cpu.
>
> I'm having trouble convincing myself that's actually possible on
> x86_64 -- after all, every LOCK implies an smp_mb there, so if ttwu
> observes ->state != RUNNING, it must also observe ->cpu != 1.
>
> Most of the previous ttwu() races were found on very large PowerPC
> machines which are far more 'interesting'. I suppose I should go write
> me litmus tests...
You should be able to get access to a modest PowerPC system here:
https://osuosl.org/services/powerdev/request_hosting/
> Anyway, IFF any of this holds true; then I suppose a patch like the below
> ought to cure things.
I have started an initial 140-hour test which should complete by about
4PM PDT.
Last night's all-scenarios run on current -rcu (without any scheduler
patch) saw two failures in three 10-hour runs of TREE03. And I might as
well use the failure times to get tighter bounds on the probability of
occurrence, though with only two failures it is pretty dicey -- getting
decent bounds would require something like 30 failures. But we must
work with what we have.
The failures occurred at 18680 and 20018 seconds, respectively, and
the failure-free run was 36000 seconds long. Using maxima to help:
load(distrib);
bfloat(18680/3600+20018/3600+10);
2.074944444444444b1
1-bfloat(cdf_poisson(0,140/(2.074944444444444b1/2)));
9.999986212554204b-1
So a 140-hour failure-free run gives a 99.9999% confidence of some degree
of improvement. But "some degree of improvement" might be a very small
improvement. So let's look at an order of magnitude:
1-bfloat(cdf_poisson(0,0.1*140/(2.074944444444444b1/2)));
7.406128951482777b-1
So a 140-hour failure-free run gives a 74% confidence that the
patch gave an order-of-magnitude improvement. So if that passes, an
additional 420-hour failure-free run would get to 99.5% confidence of
an order-of-magnitude improvement.
Statistics is rather unforgiving. :-/
> If not, I'm, once again, defeated by this...
Here is hoping that this patch cures things!
Thanx, Paul
> ---
> kernel/sched/core.c | 9 +++++++--
> 1 file changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 8298b2c240ce..5534eb1ab79a 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2378,6 +2378,9 @@ static inline bool ttwu_queue_cond(int cpu, int wake_flags)
> static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
> {
> if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(cpu, wake_flags)) {
> + if (WARN_ON(cpu == smp_processor_id()))
> + return false;
> +
> sched_clock_cpu(cpu); /* Sync clocks across CPUs */
> __ttwu_queue_wakelist(p, cpu, wake_flags);
> return true;
> @@ -2550,7 +2553,6 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
>
> /* We're going to change ->state: */
> success = 1;
> - cpu = task_cpu(p);
>
> /*
> * Ensure we load p->on_rq _after_ p->state, otherwise it would
> @@ -2615,7 +2617,8 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
> * let the waker make forward progress. This is safe because IRQs are
> * disabled and the IPI will deliver after on_cpu is cleared.
> */
> - if (READ_ONCE(p->on_cpu) && ttwu_queue_wakelist(p, cpu, wake_flags | WF_ON_RQ))
> + if (smp_load_acquire(&p->on_cpu) &&
> + ttwu_queue_wakelist(p, task_cpu(p), wake_flags | WF_ON_RQ))
> goto unlock;
>
> /*
> @@ -2635,6 +2638,8 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
> psi_ttwu_dequeue(p);
> set_task_cpu(p, cpu);
> }
> +#else
> + cpu = task_cpu(p);
> #endif /* CONFIG_SMP */
>
> ttwu_queue(p, cpu, wake_flags);
>