On 07/11/2019 06:20, Jens Axboe wrote:
> Pavel,
> 
> I didn't read this one closely enough for the latest change, we
> can't just return -EAGAIN if we don't submit anything, we have to
> explicitly do it for just the failure case of not getting a request.
> I'm going to fold in this incremental:
> 
My bad, it should be a common case to have @to_submit >= nr_sqes_in_ring
(i.e. I don't care, just submit all). Thanks for fixing it.
Also, if it submitted @n (>0), and can't allocate for @n+1, I'd rather
prefer to get submitted count.

BTW, I found this "one return value for both submit() and wait_cqes()"
a bit confusing. In addition, there is reporting through cqe, and
sometimes they could be interchanged like in this case. It's getting
even more trickier with links.
Maybe, you have some write-up on what to expect from this perspective?
Don't remember it to be in the pdf.


> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 951fd09612ec..af8a673c618f 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -2978,8 +2978,11 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>  		unsigned int sqe_flags;
>  
>  		req = io_get_req(ctx, statep);
> -		if (unlikely(!req))
> +		if (unlikely(!req)) {
> +			if (!submitted)
> +				submitted = -EAGAIN;
>  			break;
> +		}
>  		if (!io_get_sqring(ctx, &req->submit)) {
>  			__io_free_req(req);
>  			break;
> @@ -4316,8 +4319,6 @@ SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
>  		cur_mm = ctx->sqo_mm;
>  		submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
>  					   &cur_mm, false);
> -		if (!submitted)
> -			submitted = -EAGAIN;
>  		mutex_unlock(&ctx->uring_lock);
>  	}
>  	if (flags & IORING_ENTER_GETEVENTS) {
> 

-- 
Pavel Begunkov