Re: [PATCH 1/1] sched: CPU remove deadlock fix

[Brian King <brking@linux.vnet.ibm.com>]

Peter Zijlstra wrote:
> On Tue, 2008-12-09 at 15:52 +0100, Peter Zijlstra wrote:
>> On Tue, 2008-12-09 at 08:47 -0600, Brian King wrote:
>>> This patch fixes a possible deadlock scenario in the CPU remove path.
>>> migration_call grabs rq->lock, then wakes up everything on rq->migration_queue
>>> with the lock held. Then one of the tasks on the migration queue ends up
>>> calling tg_shares_up which then also tries to acquire the same rq->lock.
>> Looks ok, does lockdep agree?
> 
> On second thought, I'm not seeing it at all..
> 
> why doesn't every wakeup deadlock?

>From what I can tell, the only other place that does a complete(&req->done)
is in migration_thread, and there the lock is released before doing the wakeup.

-Brian

> 
>>> Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
>>> ---
>>>
>>>  kernel/sched.c |    2 ++
>>>  1 file changed, 2 insertions(+)
>>>
>>> diff -puN kernel/sched.c~sched_cpu_down_deadlock_fix kernel/sched.c
>>> --- linux-2.6/kernel/sched.c~sched_cpu_down_deadlock_fix	2008-12-09 08:42:09.000000000 -0600
>>> +++ linux-2.6-bjking1/kernel/sched.c	2008-12-09 08:42:09.000000000 -0600
>>> @@ -6587,7 +6587,9 @@ migration_call(struct notifier_block *nf
>>>  			req = list_entry(rq->migration_queue.next,
>>>  					 struct migration_req, list);
>>>  			list_del_init(&req->list);
>>> +			spin_unlock_irq(&rq->lock);
>>>  			complete(&req->done);
>>> +			spin_lock_irq(&rq->lock);
>>>  		}
>>>  		spin_unlock_irq(&rq->lock);
>>>  		break;
>>> _


-- 
Brian King
Linux on Power Virtualization
IBM Linux Technology Center