Re: possible deadlock in __wake_up_common_lock

[Peter Zijlstra]

On Wed, Jan 02, 2019 at 01:51:01PM +0100, Vlastimil Babka wrote:
> > -> #3 (&base->lock){-.-.}:
> >         __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline]
> >         _raw_spin_lock_irqsave+0x99/0xd0 kernel/locking/spinlock.c:152
> >         lock_timer_base+0xbb/0x2b0 kernel/time/timer.c:937
> >         __mod_timer kernel/time/timer.c:1009 [inline]
> >         mod_timer kernel/time/timer.c:1101 [inline]
> >         add_timer+0x895/0x1490 kernel/time/timer.c:1137
> >         __queue_delayed_work+0x249/0x380 kernel/workqueue.c:1533
> >         queue_delayed_work_on+0x1a2/0x1f0 kernel/workqueue.c:1558
> >         queue_delayed_work include/linux/workqueue.h:527 [inline]
> >         schedule_delayed_work include/linux/workqueue.h:628 [inline]
> >         psi_group_change kernel/sched/psi.c:485 [inline]
> >         psi_task_change+0x3f1/0x5f0 kernel/sched/psi.c:534
> >         psi_enqueue kernel/sched/stats.h:82 [inline]
> >         enqueue_task kernel/sched/core.c:727 [inline]
> >         activate_task+0x21a/0x430 kernel/sched/core.c:751
> >         wake_up_new_task+0x527/0xd20 kernel/sched/core.c:2423
> >         _do_fork+0x33b/0x11d0 kernel/fork.c:2247
> >         kernel_thread+0x34/0x40 kernel/fork.c:2281
> >         rest_init+0x28/0x372 init/main.c:409
> >         arch_call_rest_init+0xe/0x1b
> >         start_kernel+0x873/0x8ae init/main.c:741
> >         x86_64_start_reservations+0x29/0x2b arch/x86/kernel/head64.c:470
> >         x86_64_start_kernel+0x76/0x79 arch/x86/kernel/head64.c:451
> >         secondary_startup_64+0xa4/0xb0 arch/x86/kernel/head_64.S:243

That thing is fairly new; I don't think we used to have this dependency
prior to PSI.

Johannes, can we move that mod_timer out from under rq->lock? At worst
we can use an irq_work to self-ipi.