Re: [PATCH] workqueue: Ensure that cpumask set for pools created after boot

[Michael Bringmann <mwb@linux.vnet.ibm.com>]

On 05/16/2017 10:55 AM, Tejun Heo wrote:
> Hello, Michael.
> 
> On Mon, May 15, 2017 at 10:48:04AM -0500, Michael Bringmann wrote:
>>>> --- a/kernel/workqueue.c
>>>> +++ b/kernel/workqueue.c
>>>> @@ -3366,6 +3366,8 @@ static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
>>>>  	copy_workqueue_attrs(pool->attrs, attrs);
>>>>  	pool->node = target_node;
>>>>  
>>>> +	cpumask_copy(pool->attrs->cpumask, cpumask_of(smp_processor_id()));
>>>
>>> What prevents a cpu getting added right here tho?
>>
>> PowerPC has only one control path to add/remove CPUs via DLPAR operations.
>> Even so, the underlying code is protected through multiple locks.
> 
> The more I look at the patch, the less sense it seems to make.  So,
> whenever we create a new pool, we ignore the requested cpumask and
> override it with the cpumask of the current thread?

No.  As I mentioned previously, the operation/problem occurs within a DLPAR
hotplug add/remove operation.  This is happening to a node which previously
did not have any CPUs associated to it -- we are trying to add more resources
to an LPAR / partition.  At this point, the cpumask for the node is empty / zero.
Sorry for not being more clear on this point earlier.
 
>>> Maybe the right thing to do is protecting the whole thing with hotplug
>>> readlock?
>>
>> The operation is already within a hotplug readlock when performing DLPAR
>> add/remove.  Adding a CPU to the system, requires it to be brought online.
>> Removing a CPU from the system, requires it to be taken offline.  These
>> involve calls to cpu_up / cpu_down, which go through _cpu_up / _cpu_down,
>> which acquire the hotplug locks, among others along the path of execution.
>>
>> The locks are acquired before getting to the workqueue code, the pool
>> creation/attachment code (which is where the cpu mask needs to be set),
>> or trying to wakeup the initial created task in 'sched.c'.
> 
> A new unbound workqueue and thus unbound pool can also be created from
> paths outside cpu hotplug, so get_unbound_pool() can race against
> hotplug.  Can you please explain the failures that you see in more
> detail?  I'm sure your patch works around the issue somehow but it
> doesn't look like the right fix.

We fill in an empty cpumask field with a guaranteed non-empty value.
I verified that the incoming cpumask in the attrs was zero at this point
preceding the failure.  If we proceed without putting in a useful value,
we go to 'wake_up_process()' (kernel/sched/core.c) next to wakeup the new
worker for the new unbound pool.  While there, the code runs through
'select_task_rq()' and invokes cpumask_any() on a copy of the cpumask.
Unfortunately, running that function over an empty/non-initialized cpumask
returns an index beyond the end of the list, resulting shortly thereafter
in an instruction/data fetch exception.

If you have a suggestion for an alternate non-empty value to use, I would
be happy to try it.

> 
> Thanks.
> 

-- 
Michael W. Bringmann
Linux Technology Center
IBM Corporation
Tie-Line  363-5196
External: (512) 286-5196
Cell:       (512) 466-0650
mwb@linux.vnet.ibm.com