Re: [PATCH 2/2] io_uring: don't hold uring_lock when calling io_run_task_work*

[Hao Xu <haoxu@linux.alibaba.com>]

在 2021/2/5 下午6:18, Pavel Begunkov 写道:
> On 05/02/2021 09:57, Hao Xu wrote:
>> 在 2021/2/4 下午11:26, Pavel Begunkov 写道:
>>> On 04/02/2021 11:17, Pavel Begunkov wrote:
>>>> On 04/02/2021 03:25, Hao Xu wrote:
>>>>> 在 2021/2/4 上午12:45, Pavel Begunkov 写道:
>>>>>> On 03/02/2021 16:35, Pavel Begunkov wrote:
>>>>>>> On 03/02/2021 14:57, Hao Xu wrote:
>>>>>>>> This is caused by calling io_run_task_work_sig() to do work under
>>>>>>>> uring_lock while the caller io_sqe_files_unregister() already held
>>>>>>>> uring_lock.
>>>>>>>> we need to check if uring_lock is held by us when doing unlock around
>>>>>>>> io_run_task_work_sig() since there are code paths down to that place
>>>>>>>> without uring_lock held.
>>>>>>>
>>>>>>> 1. we don't want to allow parallel io_sqe_files_unregister()s
>>>>>>> happening, it's synchronised by uring_lock atm. Otherwise it's
>>>>>>> buggy.
>>>>> Here "since there are code paths down to that place without uring_lock held" I mean code path of io_ring_ctx_free().
>>>>
>>>> I guess it's to the 1/2, but let me outline the problem again:
>>>> if you have two tasks userspace threads sharing a ring, then they
>>>> can both and in parallel call syscall:files_unregeister. That's
>>>> a potential double percpu_ref_kill(&data->refs), or even worse.
>>>>
>>>> Same for 2, but racing for the table and refs.
>>>
>>> There is a couple of thoughts for this:
>>>
>>> 1. I don't like waiting without holding the lock in general, because
>>> someone can submit more reqs in-between and so indefinitely postponing
>>> the files_unregister.
>> Thanks, Pavel.
>> I thought this issue before, until I saw this in __io_uring_register:
>>
>>    if (io_register_op_must_quiesce(opcode)) {
>>            percpu_ref_kill(&ctx->refs);
> 
> It is different because of this kill, it will prevent submissions.
> 
I saw percpu_ref_is_dying(&ctx->refs) check in sq thread but not
in io_uring_enter(), so I guess there could be another thread doing
io_uring_enter() and submiting sqes.
>>
>>            /*
>>            ¦* Drop uring mutex before waiting for references to exit. If
>>            ¦* another thread is currently inside io_uring_enter() it might
>>            ¦* need to grab the uring_lock to make progress. If we hold it
>>            ¦* here across the drain wait, then we can deadlock. It's safe
>>            ¦* to drop the mutex here, since no new references will come in
>>            ¦* after we've killed the percpu ref.
>>            ¦*/
>>            mutex_unlock(&ctx->uring_lock);
>>            do {
>>                    ret = wait_for_completion_interruptible(&ctx->ref_comp);
>>                    if (!ret)
>>                            break;
>>                    ret = io_run_task_work_sig();
>>                    if (ret < 0)
>>                            break;
>>            } while (1);
>>
>>            mutex_lock(&ctx->uring_lock);
>>
>>            if (ret) {
>>                    percpu_ref_resurrect(&ctx->refs);
>>                    goto out_quiesce;
>>            }
>>    }
>>
>> So now I guess the postponement issue also exits in the above code since
>> there could be another thread submiting reqs to the shared ctx(or we can say uring fd).
>>
>>> 2. I wouldn't want to add checks for that in submission path.
>>>
>>> So, a solution I think about is to wait under the lock, If we need to
>>> run task_works -- briefly drop the lock, run task_works and then do
>>> all unregister all over again. Keep an eye on refs, e.g. probably
>>> need to resurrect it.
>>>
>>> Because we current task is busy nobody submits new requests on
>>> its behalf, and so there can't be infinite number of in-task_work
>>> reqs, and eventually it will just go wait/sleep forever (if not
>>> signalled) under the mutex, so we can a kind of upper bound on
>>> time.
Sorry Pavel, I don't quiet understand "so we can a kind of upper bound 
on time". :(
>>>
>> Do you mean sleeping with timeout rather than just sleeping? I think this works, I'll work on this and think about the detail.
> 
> Without timeout -- it will be awaken when new task_works are coming in,
> but Jens knows better.
So we can just put unlock and lock around io_run_task_work_sig()
to address the issue 2?
> 
>> But before addressing this issue, Should I first send a patch to just fix the deadlock issue?
> 
> Do you mean the deadlock 2/2 was trying to fix? Or some else? The thread
> is all about fixing it, but doing it right. Not sure there is a need for
> faster but incomplete solution, if that's what you meant.
>