Re: [PATCH 0/5] fixes for REQ_F_COMP_LOCKED

[Pavel Begunkov <asml.silence@gmail.com>]

On 13/10/2020 15:57, Jens Axboe wrote:
> On 10/13/20 3:46 AM, Pavel Begunkov wrote:
>> On 13/10/2020 09:43, Pavel Begunkov wrote:
>>> This removes REQ_F_COMP_LOCKED to fix a couple of problems with it.
>>>
>>> [5/5] is harsh and some work should be done to ease the aftermath,
>>> i.e. io_submit_flush_completions() and maybe fail_links().
>>>
>>> Another way around would be to replace the flag with an comp_locked
>>> argument in put_req(), free_req() and so on, but IMHO in a long run
>>> removing it should be better.
>>>
>>> note: there is a new io_req_task_work_add() call in [5/5]. Jens,
>>> could you please verify whether passed @twa_signal_ok=true is ok,
>>> because I don't really understand the difference.
> 
> It should be fine, the only case that can't use 'true' is when it's
> called from within the waitqueue handler as we can recurse on that
> lock.

Got it. And thanks for fixing descriptions!

> 
> Luckily that'll all go away once the TWA_SIGNAL improvement patches
> are ready.
> 
>> btw, when I copied task_work_add(TWA_RESUME) from __io_free_req(),
>> tasks were hanging sleeping uninterruptibly, and fail_links()
>> wasn't waking them. It looks like the deferring branch of
>> __io_free_req() is buggy as well and should use
>> io_req_task_work_add().
> 
> Probably related to exit conditions.

Yep, kind of
    close() -> ->flush() -> io_uring_cancel_files() -> schedule()

-- 
Pavel Begunkov