Re: [PATCH 2/6] io_uring: move poll_refs up a cacheline to fill a hole

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

On 4/19/23 17:25, Jens Axboe wrote:
> After bumping the flags to 64-bits, we now have two holes in io_kiocb.
> The best candidate for moving is poll_refs, as not to split the task_work
> related items.
> 
> Signed-off-by: Jens Axboe <axboe@kernel.dk>
> ---
>   include/linux/io_uring_types.h | 6 ++++--
>   1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
> index 84f436cc6509..4dd54d2173e1 100644
> --- a/include/linux/io_uring_types.h
> +++ b/include/linux/io_uring_types.h
> @@ -535,6 +535,9 @@ struct io_kiocb {
>   	 * and after selection it points to the buffer ID itself.
>   	 */
>   	u16				buf_index;
> +
> +	atomic_t			poll_refs;

poll wake is often done by a random CPU, would be great to limit
cache sharing, i.e. the 2nd cache line is never touched by
io_poll_wake() and it'd only need the poll entry and
the 3rd line with io_task_work().

There is one place that doesn't care about it, i.e.
clearing REQ_F_[SINGLE,DOUBLE]_POLL in poll wake, but
it's a place to improve.


> +
>   	u64				flags;
>   
>   	struct io_cqe			cqe;
> @@ -565,9 +568,8 @@ struct io_kiocb {
>   		__poll_t apoll_events;
>   	};
>   	atomic_t			refs;
> -	atomic_t			poll_refs;
> -	struct io_task_work		io_task_work;
>   	unsigned			nr_tw;
> +	struct io_task_work		io_task_work;
>   	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
>   	union {
>   		struct hlist_node	hash_node;

-- 
Pavel Begunkov