Re: [PATCH] clockevent: don't remove broadcast device when cpu is dead

[Xiaotian Feng <dfeng@redhat.com>]

On 01/14/2010 06:08 AM, Thomas Gleixner wrote:
> On Wed, 13 Jan 2010, Xiaotian Feng wrote:
>> On 01/12/2010 09:20 PM, Thomas Gleixner wrote:
>>> On Thu, 7 Jan 2010, Xiaotian Feng wrote:
>>>
>>>> Marc reported BUG during shutdown, after debugging, kernel is trying
>>>> to remove a broadcast device which mode is CLOCK_EVT_MODE_ONESHOT.
>>>>
>>>> The root cause for this bug is that in clockevents_notify,
>>>> "cpumask_weight(dev->cpumask) == 1" is always true even if dev is a
>>>
>>> Why is cpumask_weight(dev->cpumask) == 1 always true when we shutdown
>>> a non boot cpu ?
>>>
>>> The broadcast device is not a per cpu device and the cpumask should
>>> not only contain the CPU which is shut down !
>>
>> At least for hpet broadcast dev, it's dev->cpumask is only contain the CPU
>> which it is initialized from.
>
> Which is fundamentaly wrong and the root cause of the problem. I'll
> have a look tomorrow morning when my brain is more awake than now.

hpet_legacy_clockevent_register is trying to register new CE, but 
replace failed,
then in tick_check_new_device -> tick_check_broadcast_device, the legacy 
hpet CE
was registered as multicast device, but its dev->cpumask is cpumask of 
smp_processor_id().

on my system its dev->cpumask is cpumask of 0, but in Marc's, 
dev->cpumask is cpumask of 4.
So when kernel is trying to offline cpu 4, the broadcast hpet is removed.

>
>> And for broadcast device, kernel is using tick_broadcast_mask not
>> dev->cpumask, right?
>
> No, tick_broadcast_mask is the bitmask which tells us which cpus get
> the broadcast IPI.
>
> Thanks,
>
> 	tglx
>