Re: [PATCH] io_uring: unlock sqd->lock before sq thread release CPU

[Jens Axboe <axboe@kernel.dk>]

On 5/23/23 11:28?PM, Wenwen Chen wrote:
> The sq thread actively releases CPU resources by calling the 
> cond_resched() and schedule() interfaces when it is idle. Therefore,
> more resources are available for other threads to run.
> 
> There exists a problem in sq thread: it does not unlock sqd->lock before
> releasing CPU resources every time. This makes other threads pending on
> sqd->lock for a long time. For example, the following interfaces all 
> require sqd->lock: io_sq_offload_create(), io_register_iowq_max_workers()
> and io_ring_exit_work().
>        
> Before the sq thread releases CPU resources, unlocking sqd->lock will 
> provide the user a better experience because it can respond quickly to
> user requests.
> 
> Signed-off-by: Kanchan Joshi<joshi.k@samsung.com>
> Signed-off-by: Wenwen Chen<wenwen.chen@samsung.com>
> ---
>  io_uring/sqpoll.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/io_uring/sqpoll.c b/io_uring/sqpoll.c
> index 9db4bc1f521a..759c80fb4afa 100644
> --- a/io_uring/sqpoll.c
> +++ b/io_uring/sqpoll.c
> @@ -255,7 +255,9 @@ static int io_sq_thread(void *data)
>  			sqt_spin = true;
>  
>  		if (sqt_spin || !time_after(jiffies, timeout)) {
> +			mutex_unlock(&sqd->lock);
>  			cond_resched();
> +			mutex_lock(&sqd->lock);
>  			if (sqt_spin)
>  				timeout = jiffies + sqd->sq_thread_idle;
>  			continue;

Since this is the spin case, and we expect (by far) most of these
to NOT need a reschedule, I think we should do:

	if (need_resched()) {
		mutex_unlock(&sqd->lock);
		cond_resched();
		mutex_lock(&sqd->lock);
	}

to make that lock shuffle dependent on the need to reschedule. And
since we're marking the timeout at that point, timeout should be
assigned first as far as I can tell. So in total:

	if (sqt_spin || !time_after(jiffies, timeout)) {
		if (sqt_spin)
 			timeout = jiffies + sqd->sq_thread_idle;
 		if (unlikely(need_resched())) {
			mutex_unlock(&sqd->lock);
			cond_resched();
			mutex_lock(&sqd->lock);
		}
		continue;
	}

would probably be the better fix.

-- 
Jens Axboe