On 12/03/2021 19:40, Jens Axboe wrote:
> On 3/12/21 12:35 PM, Pavel Begunkov wrote:
>> On 11/03/2021 23:29, Pavel Begunkov wrote:
>>> 1) The first problem is io_uring_cancel_sqpoll() ->
>>> io_uring_cancel_task_requests() basically doing park(); park(); and so
>>> hanging.
>>>
>>> 2) Another one is more subtle, when the master task is doing cancellations,
>>> but SQPOLL task submits in-between the end of the cancellation but
>>> before finish() requests taking a ref to the ctx, and so eternally
>>> locking it up.
>>>
>>> 3) Yet another is a dying SQPOLL task doing io_uring_cancel_sqpoll() and
>>> same io_uring_cancel_sqpoll() from the owner task, they race for
>>> tctx->wait events. And there probably more of them.
>>>
>>> Instead do SQPOLL cancellations from within SQPOLL task context via
>>> task_work, see io_sqpoll_cancel_sync(). With that we don't need temporal
>>> park()/unpark() during cancellation, which is ugly, subtle and anyway
>>> doesn't allow to do io_run_task_work() properly.>
>>> io_uring_cancel_sqpoll() is called only from SQPOLL task context and
>>> under sqd locking, so all parking is removed from there. And so,
>>> io_sq_thread_[un]park() and io_sq_thread_stop() are not used now by
>>> SQPOLL task, and that spare us from some headache.
>>>
>>> Also remove ctx->sqd_list early to avoid 2). And kill tctx->sqpoll,
>>> which is not used anymore.
>>
>>
>> Looks, the chunk below somehow slipped from the patch. Not important
>> for 5.12, but can can be folded anyway
>>
>> diff --git a/include/linux/io_uring.h b/include/linux/io_uring.h
>> index 9761a0ec9f95..c24c62b47745 100644
>> --- a/include/linux/io_uring.h
>> +++ b/include/linux/io_uring.h
>> @@ -22,7 +22,6 @@ struct io_uring_task {
>> void *io_wq;
>> struct percpu_counter inflight;
>> atomic_t in_idle;
>> - bool sqpoll;
>>
>> spinlock_t task_lock;
>> struct io_wq_work_list task_list;
>
> Let's do it as a separate patch instead.
Ok, I'll send it for-5.13 when it's appropriate.
--
Pavel Begunkov
