Re: [PATCH] Fix the race between smp_call_function and CPU booting

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 03/20/2012 07:11 PM, Peter Zijlstra wrote:

> On Tue, 2012-03-20 at 18:44 +0530, Srivatsa S. Bhat wrote:
>>
>>
>> I don't think this patch would change anything, atleast it wouldn't get
>> rid of the warning that Liu reported. Because, he is running his stress
>> tests on a machine which has only 2 CPUs. So effectively, we are hotplugging
>> only CPU1 (since CPU0 can't be taken offline, on Intel boxes).
>>
>> Also, CPU1 is removed from cpu_active_mask during CPU_DOWN_PREPARE time itself,
>> and migrate_tasks() comes much later (during CPU_DYING). And in any case,
>> dest_cpu will never be CPU1, because it is the CPU that is going down. So it
>> *has* to be CPU0 anyway.
>>
>> So, I don't think changes to select_fallback_rq() to make it more careful is
>> going to make any difference in the particular scenario that Liu is testing.
>>
>> That said, even I don't know what the root cause of the warning is.. :-(
> 
> Its a race in cpu-up, we set active before online, when we do a wakeup
> select_task_rq() will see !cpu_online(), we then call
> select_fallback_rq() to compute a new cpu, select_fallback_rq() computes
> a new cpu against cpu_active (which is set) and can thus return cpu 1,
> even though it is still offline.
> 
> So we queue the task on cpu 1 and send a reschedule ipi, at which point
> we'll get the reported warning.
> 
> My change modifies select_fallback_rq() to require online && active.
> 


Ok, that makes sense.. Thanks a lot for the explanation!

Regards,
Srivatsa S. Bhat