Re: possible deadlock in generic_file_write_iter

[Byungchul Park <byungchul.park@lge.com>]

11/7/2017 5:31 PM에 Dmitry Vyukov 이(가) 쓴 글:
> On Tue, Nov 7, 2017 at 9:30 AM, Byungchul Park <byungchul.park@lge.com> wrote:
>> 11/7/2017 5:11 PM에 Peter Zijlstra 이(가) 쓴 글:
>>
>>> On Tue, Nov 07, 2017 at 09:54:42AM +0900, Byungchul Park wrote:
>>>>>
>>>>> The best I could come up with is something like the below; its not
>>>>> at all pretty and I could see people objecting; least of all myself for
>>>>> the __complete() thing, but I ran out of creative naming juice.
>>>>
>>>>
>>>> Patches assigning a lock_class per gendisk were already applied in tip.
>>>> I believe that solves this.
>>>>
>>>>      e319e1fbd9d42420ab6eec0bfd75eb9ad7ca63b1
>>>>      block, locking/lockdep: Assign a lock_class per gendisk used for
>>>>      wait_for_completion()
>>>>
>>>> I think the following proposal makes kernel too hacky.
>>>
>>>
>>> Ah, I tough this was with those included...
>>
>>
>> Please CC me for issues wrt. crossrelease.
> 
> Hi Byungchul,
> 
> Whom are you asking? And what is crossrelease?

Hi Dmitry,

Actually, I asked Peter since he know I am the author of
crossrelease which is a feature enhancing lockdep detection.

That makes deadlocks caused by wait_for_completion() or
lock_page() detectable.

The feature was merged and disabled by default for now in
vanilla because of its performance regression, though that
was fixed and enabled in tip latest.

--
Thanks,
Byungchul