Re: [RFC][PATCH] sched: Fix a deadlock of cpu-hotplug

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2012-10-24 at 17:25 +0800, Huacai Chen wrote:
> We found poweroff sometimes fails on our computers, so we have the
> lock debug options configured. Then, when we do poweroff or take a
> cpu down via cpu-hotplug, kernel complain as below. To resove this,
> we modify sched_ttwu_pending(), disable the local irq when acquire
> rq->lock.
> 
> [   83.066406] =================================
> [   83.066406] [ INFO: inconsistent lock state ]
> [   83.066406] 3.5.0-3.lemote #428 Not tainted
> [   83.066406] ---------------------------------
> [   83.066406] inconsistent {IN-HARDIRQ-W} -> {HARDIRQ-ON-W} usage.
> [   83.066406] migration/1/7 [HC0[0]:SC0[0]:HE1:SE1] takes:
> [   83.066406]  (&rq->lock){?.-.-.}, at: [<ffffffff802585ac>] sched_ttwu_pending+0x64/0x98
> [   83.066406] {IN-HARDIRQ-W} state was registered at:
> [   83.066406]   [<ffffffff8027c9ac>] __lock_acquire+0x80c/0x1cc0
> [   83.066406]   [<ffffffff8027e3d0>] lock_acquire+0x60/0x9c
> [   83.066406]   [<ffffffff8074ba04>] _raw_spin_lock+0x3c/0x50
> [   83.066406]   [<ffffffff8025a2fc>] scheduler_tick+0x48/0x178
> [   83.066406]   [<ffffffff8023b334>] update_process_times+0x54/0x70
> [   83.066406]   [<ffffffff80277568>] tick_handle_periodic+0x2c/0x9c
> [   83.066406]   [<ffffffff8020a818>] c0_compare_interrupt+0x8c/0x94
> [   83.066406]   [<ffffffff8029ec8c>] handle_irq_event_percpu+0x7c/0x248
> [   83.066406]   [<ffffffff802a2774>] handle_percpu_irq+0x8c/0xc0
> [   83.066406]   [<ffffffff8029e2c8>] generic_handle_irq+0x48/0x58
> [   83.066406]   [<ffffffff80205c04>] do_IRQ+0x18/0x24
> [   83.066406]   [<ffffffff802016e4>] mach_irq_dispatch+0xe4/0x124
> [   83.066406]   [<ffffffff80203ca0>] ret_from_irq+0x0/0x4
> [   83.066406]   [<ffffffff8022d114>] console_unlock+0x3e8/0x4c0
> [   83.066406]   [<ffffffff811ff0d0>] con_init+0x370/0x398
> [   83.066406]   [<ffffffff811fe3e0>] console_init+0x34/0x50
> [   83.066406]   [<ffffffff811e4844>] start_kernel+0x2f8/0x4e0
> [   83.066406] irq event stamp: 971
> [   83.066406] hardirqs last  enabled at (971): [<ffffffff8021c384>] local_flush_tlb_all+0x134/0x17c
> [   83.066406] hardirqs last disabled at (970): [<ffffffff8021c298>] local_flush_tlb_all+0x48/0x17c
> [   83.066406] softirqs last  enabled at (0): [<ffffffff802298a4>] copy_process+0x510/0x117c
> [   83.066406] softirqs last disabled at (0): [<          (null)>] (null)
> [   83.066406]
> [   83.066406] other info that might help us debug this:
> [   83.066406]  Possible unsafe locking scenario:
> [   83.066406]
> [   83.066406]        CPU0
> [   83.066406]        ----
> [   83.066406]   lock(&rq->lock);
> [   83.066406]   <Interrupt>
> [   83.066406]     lock(&rq->lock);
> [   83.066406]
> [   83.066406]  *** DEADLOCK ***
> [   83.066406]
> [   83.066406] no locks held by migration/1/7.
> [   83.066406]
> [   83.066406] stack backtrace:
> [   83.066406] Call Trace:
> [   83.066406] [<ffffffff80747544>] dump_stack+0x8/0x34
> [   83.066406] [<ffffffff8027ba04>] print_usage_bug+0x2ec/0x314
> [   83.066406] [<ffffffff8027be28>] mark_lock+0x3fc/0x774
> [   83.066406] [<ffffffff8027ca48>] __lock_acquire+0x8a8/0x1cc0
> [   83.066406] [<ffffffff8027e3d0>] lock_acquire+0x60/0x9c
> [   83.066406] [<ffffffff8074ba04>] _raw_spin_lock+0x3c/0x50
> [   83.066406] [<ffffffff802585ac>] sched_ttwu_pending+0x64/0x98
> [   83.066406] [<ffffffff80745ff4>] migration_call+0x10c/0x2e0
> [   83.066406] [<ffffffff80253110>] notifier_call_chain+0x44/0x94
> [   83.066406] [<ffffffff8022eae0>] __cpu_notify+0x30/0x5c
> [   83.066406] [<ffffffff8072b598>] take_cpu_down+0x5c/0x70
> [   83.066406] [<ffffffff80299ba4>] stop_machine_cpu_stop+0x104/0x1e8
> [   83.066406] [<ffffffff802997cc>] cpu_stopper_thread+0x110/0x1ac
> [   83.066406] [<ffffffff8024c940>] kthread+0x88/0x90
> [   83.066406] [<ffffffff80205ee4>] kernel_thread_helper+0x10/0x18

Weird, that's from a CPU_DYING call, I thought those were with IRQs
disabled. 

Look at how __stop_machine() calls the function with IRQs disabled for !
stop_machine_initialized or !SMP. Also stop_machine_cpu_stop() seems to
disabled interrupts, so how do we end up calling take_cpu_down() with
IRQs enabled?

That simply doesn't make any sense.

> Signed-off-by: Huacai Chen <chenhc@lemote.com>
> ---
>  kernel/sched/core.c |    5 +++--
>  1 files changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 36e2666..703754a 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1468,9 +1468,10 @@ static void sched_ttwu_pending(void)
>  {
>  	struct rq *rq = this_rq();
>  	struct llist_node *llist = llist_del_all(&rq->wake_list);
> +	unsigned long flags;
>  	struct task_struct *p;
>  
> -	raw_spin_lock(&rq->lock);
> +	raw_spin_lock_irqsave(&rq->lock, flags);
>  
>  	while (llist) {
>  		p = llist_entry(llist, struct task_struct, wake_entry);
> @@ -1478,7 +1479,7 @@ static void sched_ttwu_pending(void)
>  		ttwu_do_activate(rq, p, 0);
>  	}
>  
> -	raw_spin_unlock(&rq->lock);
> +	raw_spin_unlock_irqrestore(&rq->lock, flags);
>  }
>  
>  void scheduler_ipi(void)


That's wrong though, you add the cost to the common case instead of the
hardly ever ran hotplug case.