Re: [syzbot] possible deadlock in io_sq_thread_finish

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

On 10/03/2021 04:10, Hillf Danton wrote:> 
> Fix 05ff6c4a0e07 ("io_uring: SQPOLL parking fixes") in the current tree
> by removing the extra set of IO_SQ_THREAD_SHOULD_STOP in response to
> the arrival of urgent signal because it misleads io_sq_thread_stop(),
> though a followup cleanup should go there.

That's actually reasonable, just like
8bff1bf8abeda ("io_uring: fix io_sq_offload_create error handling")

Are you going to send a patch?

> 
> --- x/fs/io_uring.c
> +++ y/fs/io_uring.c
> @@ -6689,10 +6689,8 @@ static int io_sq_thread(void *data)
>  			io_sqd_init_new(sqd);
>  			timeout = jiffies + sqd->sq_thread_idle;
>  		}
> -		if (fatal_signal_pending(current)) {
> -			set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
> +		if (fatal_signal_pending(current))
>  			break;
> -		}
>  		sqt_spin = false;
>  		cap_entries = !list_is_singular(&sqd->ctx_list);
>  		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
> 

-- 
Pavel Begunkov