Re: [PATCH] io_uring: reduce one unnecessary io_sqring_entries() call

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

On 08/01/2021 08:36, Xiaoguang Wang wrote:
> In __io_sq_thread(), we have gotten the number of sqes to submit,
> then in io_submit_sqes(), we can use this number directly, no need
> to call io_sqring_entries() again in io_submit_sqes().
> 
> Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
> ---
>  fs/io_uring.c | 5 ++---
>  1 file changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index ca46f314640b..200a9eb72788 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -6830,9 +6830,6 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
>  			return -EBUSY;
>  	}
>  
> -	/* make sure SQ entry isn't read before tail */
> -	nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));

the userspace can modify rings->sq.tail, and so io_sqring_entries()
can turn out to be very large. sqpoll still needs min(ctx->sq_entries, ...)

> -
>  	if (!percpu_ref_tryget_many(&ctx->refs, nr))
>  		return -EAGAIN;
>  
> @@ -9211,6 +9208,8 @@ SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
>  		if (unlikely(ret))
>  			goto out;
>  		mutex_lock(&ctx->uring_lock);
> +		/* make sure SQ entry isn't read before tail */
> +		to_submit = min3(to_submit, ctx->sq_entries, io_sqring_entries(ctx));
>  		submitted = io_submit_sqes(ctx, to_submit);
>  		mutex_unlock(&ctx->uring_lock);
>  
> 

-- 
Pavel Begunkov