Re: [PATCH tip/core/rcu 1/5] rcu: Fix soft lockup for rcu_nocb_kthread

[Nikolay Borisov]

On 22.08.2016 19:44, Paul E. McKenney wrote:
> On Mon, Aug 22, 2016 at 07:19:53PM +0300, Nikolay Borisov wrote:
>>
>> [SNIP]
>>>
>>> Signed-off-by: Ding Tianhong <dingtianhong@xxxxxxxxxx>
>>> [ paulmck: Substituted cond_resched_rcu_qs for cond_resched. ]
>>
>> This contradicts...
>>
>>> Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
>>> ---
>>>  kernel/rcu/tree_plugin.h | 1 +
>>>  1 file changed, 1 insertion(+)
>>>
>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
>>> index 0082fce402a0..85c5a883c6e3 100644
>>> --- a/kernel/rcu/tree_plugin.h
>>> +++ b/kernel/rcu/tree_plugin.h
>>> @@ -2173,6 +2173,7 @@ static int rcu_nocb_kthread(void *arg)
>>>  				cl++;
>>>  			c++;
>>>  			local_bh_enable();
>>> +			cond_resched_rcu_qs();
>>
>> with what's here?
> 
> Ding Tianhong's original patch:
> 
> http://www.mail-archive.com/linux-kernel@xxxxxxxxxxxxxxx/msg1167918.html
> 
> had cond_resched() here, which works for his workload, but can result
> in stall warnings in other cases.  I therfore changed his cond_resched()
> to the cond_resched_rcu_qs() that you see above, and documented this
> change in the "paulmck" note after Ding Tianhong's Signed-off-by.

I think my english escaped me since I took your paulmck note as "we had
cond_resched_rcu_qs initially and I replaced it with cond_resched". But
apparently it was the opposite.

Cheers,
Nik

> 
> So all is as it should be.
> 
> 							Thanx, Paul
> 
>>>  			list = next;
>>>  		}
>>>  		trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
>>>
>>
>