Re: [PATCH] sched: unify the check on atomic sleeping in __might_sleep() and schedule_bug()

[Michael Wang <wangyun@linux.vnet.ibm.com>]

On 09/14/2012 11:02 AM, Michael Wang wrote:
> On 09/13/2012 06:04 PM, Peter Zijlstra wrote:
>> On Wed, 2012-08-22 at 10:40 +0800, Michael Wang wrote:
>>> From: Michael Wang <wangyun@linux.vnet.ibm.com>
>>>
>>> Fengguang Wu <wfg@linux.intel.com> has reported the bug:
>>>
>>> [    0.043953] BUG: scheduling while atomic: swapper/0/1/0x10000002
>>> [    0.044017] no locks held by swapper/0/1.
>>> [    0.044692] Pid: 1, comm: swapper/0 Not tainted 3.6.0-rc1-00420-gb7aebb9 #34
>>> [    0.045861] Call Trace:
>>> [    0.048071]  [<c106361e>] __schedule_bug+0x5e/0x70
>>> [    0.048890]  [<c1b28701>] __schedule+0x91/0xb10
>>> [    0.049660]  [<c14472ea>] ? vsnprintf+0x33a/0x450
>>> [    0.050444]  [<c1060006>] ? lg_local_lock+0x6/0x70
>>> [    0.051256]  [<c14fb5b1>] ? wait_for_xmitr+0x31/0x90
>>> [    0.052019]  [<c144fd55>] ? do_raw_spin_unlock+0xa5/0xf0
>>> [    0.052903]  [<c1b2a532>] ? _raw_spin_unlock+0x22/0x30
>>> [    0.053759]  [<c105cdbb>] ? up+0x1b/0x70
>>> [    0.054421]  [<c1065d6b>] __cond_resched+0x1b/0x30
>>> [    0.055228]  [<c1b292d5>] _cond_resched+0x45/0x50
>>> [    0.056020]  [<c1b26c58>] mutex_lock_nested+0x28/0x370
>>> [    0.056884]  [<c1034222>] ? console_unlock+0x3a2/0x4e0
>>> [    0.057741]  [<c1ac8559>] __irq_alloc_descs+0x39/0x1c0
>>> [    0.058589]  [<c10223bc>] io_apic_setup_irq_pin+0x2c/0x310
>>> [    0.060042]  [<c20638df>] setup_IO_APIC+0x101/0x744
>>> [    0.060878]  [<c1021d51>] ? clear_IO_APIC+0x31/0x50
>>> [    0.061695]  [<c20600f4>] native_smp_prepare_cpus+0x538/0x680
>>> [    0.062644]  [<c2056a91>] ? do_one_initcall+0x12c/0x12c
>>> [    0.063517]  [<c2056a91>] ? do_one_initcall+0x12c/0x12c
>>> [    0.064016]  [<c2056adc>] kernel_init+0x4b/0x17f
>>> [    0.064790]  [<c2056a91>] ? do_one_initcall+0x12c/0x12c
>>> [    0.065660]  [<c1b2bbd6>] kernel_thread_helper+0x6/0x10
>>>
>>> It was caused by that:
>>>
>>> 	native_smp_prepare_cpus()
>>> 	preempt_disable()		//preempt_count++
>>> 	mutex_lock()			//in __irq_alloc_descs
>>> 	__might_sleep()			//system is booting, avoid check
>>> 	might_resched()
>>> 	__schedule()
>>> 	preempt_disable()		//preempt_count++
>>> 	schedule_bug()			//preempt_count > 1, report bug
>>>
>>> The __might_sleep() avoid check on atomic sleeping until the system booted
>>> while the schedule_bug() doesn't, it's the reason for the bug.
>>>
>>> This patch will add one additional check in schedule_bug() to avoid check
>>> until the system booted, so the check on atomic sleeping will be unified.
>>>
>>> Signed-off-by: Michael Wang <wangyun@linux.vnet.ibm.com>
>>> Tested-by: Fengguang Wu <wfg@linux.intel.com>
>>> ---
>>>  kernel/sched/core.c |    3 ++-
>>>  1 files changed, 2 insertions(+), 1 deletions(-)
>>>
>>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>>> index 4376c9f..3396c33 100644
>>> --- a/kernel/sched/core.c
>>> +++ b/kernel/sched/core.c
>>> @@ -3321,7 +3321,8 @@ static inline void schedule_debug(struct task_struct *prev)
>>>  	 * schedule() atomically, we ignore that path for now.
>>>  	 * Otherwise, whine if we are scheduling when we should not be.
>>>  	 */
>>> -	if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
>>> +	if (unlikely(in_atomic_preempt_off() && !prev->exit_state
>>> +					&& system_state == SYSTEM_RUNNING))
>>>  		__schedule_bug(prev);
>>>  	rcu_sleep_check();
>>>  
>>
>>
>> No this is very very wrong.. we avoid the might_sleep bug on !
>> SYSTEM_RUNNING because while we _might_ sleep, we should _never_
>> actually sleep under those conditions.
>>
>> So hitting a schedule() here is an actual bug.
> 
> I see, so the rule is that we never allowed invoke schedule() with
> preempt disabled.
> 
> The actual reason trigger this bug is that:
> 	we invoke irq_alloc_descs() which will use mutex_lock() while
> 	!SYSTEM_RUNNING.
> And mutex_lock() invoke the might_sleep(), which do the schedule()
> without any warning.
> 
> So if we want to follow the rule, should_resched() should never return
> true if preempt disabled.
> 
> I think we could do changes like:
> 
> 
> 
> index c46a011..36fe510 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -4209,7 +4209,7 @@ SYSCALL_DEFINE0(sched_yield)
>  
>  static inline int should_resched(void)
>  {
> -       return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
> +       return need_resched() && !preempt_count();
>  }
>  
>  static void __cond_resched(void)
> 
> 
> 
> Then the should_resched() will return false when the preempt disabled or
> PREEMPT_ACTIVE bit is on.
> 
> Could we use this solution?

Let me send out the patch so we could have a thread to discuss, but
please warn me if it's a totally foolish one...

Regards,
Michael Wang

> 
> Regards,
> Michael Wang
> 
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at  http://www.tux.org/lkml/
>>
>