Re: [PATCH V2 1/2] blk-mq: init hctx sched after update ctx and hctx mapping

["jianchao.wang" <jianchao.w.wang@oracle.com>]

On 08/20/2018 12:18 PM, Ming Lei wrote:
> On Mon, Aug 20, 2018 at 11:57:52AM +0800, jianchao.wang wrote:
>> Hi Ming
>>
>> On 08/20/2018 10:24 AM, Ming Lei wrote:
>>>> The code path is:
>>>>
>>>> elevator_release
>>>>   -> elevator_put(e->type)
>>>>     -> module_put(e->elevator_owner)
>>>>
>>>> In normal elevator switch path, elevator_get will hold a reference counter of the
>>>> elevator_owner.
>>>> In this patch set, the elevator_type is saved directly. To prevent the io scheduler module
>>>> is removed, we need to hold a reference of the module.
>>> Yeah, I agree that the module reference need to be held, but it need to be
>>> released too.
>>>
>>> My concern is that this introduced getting module reference in your patch
>>> isn't released. The module reference is a counter too, so the get and
>>> put operation should be matched.
>>
>> elevator_switch_mq
>>   -> elevator_exit
>> it will put a reference count of the elevator_queue associated with the request_queue.
>> and the elevator_release will be invoked when the reference count of elevator_queue->kobj reaches zero.
>> elevator_release will put the reference count of the io scheduler module.
>>
>> the elevator_queue structure will be allocated and freed every time when we switch io scheduler.
>> so the elevator_release will always be invoked.
>>
>> This is the put ref corresponding to the get one in this patch.
> 
> OK, got it, the added __module_get() is just like what elevator_get()
> does from __elevator_change(). Sorry for the noise.
> 
> Reviewed-by: Ming Lei <ming.lei@redhat.com>
> 

Thanks very much for your reviewing.

I will add more comment to describe the added __module_get() in the V3.

Thanks
Jianchao
 

> Thanks,
> Ming
>