Re: [PATCH 4/5] io_uring/msg_ring: improve handling of target CQE posting

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

On 6/18/24 19:48, Jens Axboe wrote:
> Use the exported helper for queueing task_work for message passing,
> rather than rolling our own.
> 
> Note that this is only done for strict data messages for now, file
> descriptor passing messages still rely on the kernel task_work. It could
> get converted at some point if it's performance critical.
> 
> This improves peak performance of message passing by about 5x in some
> basic testing, with 2 threads just sending messages to each other.
> Before this change, it was capped at around 700K/sec, with the change
> it's at over 4M/sec.
> 
> Signed-off-by: Jens Axboe <axboe@kernel.dk>
> ---
>   io_uring/msg_ring.c | 90 +++++++++++++++++++++++----------------------
>   1 file changed, 47 insertions(+), 43 deletions(-)
> 
> diff --git a/io_uring/msg_ring.c b/io_uring/msg_ring.c
> index 9fdb0cc19bfd..ad7d67d44461 100644
> --- a/io_uring/msg_ring.c
> +++ b/io_uring/msg_ring.c
> @@ -13,7 +13,6 @@
>   #include "filetable.h"
>   #include "msg_ring.h"
>   
> -
>   /* All valid masks for MSG_RING */
>   #define IORING_MSG_RING_MASK		(IORING_MSG_RING_CQE_SKIP | \
>   					IORING_MSG_RING_FLAGS_PASS)
> @@ -71,54 +70,43 @@ static inline bool io_msg_need_remote(struct io_ring_ctx *target_ctx)
>   	return target_ctx->task_complete;
>   }
>   
> -static int io_msg_exec_remote(struct io_kiocb *req, task_work_func_t func)
> +static void io_msg_tw_complete(struct io_kiocb *req, struct io_tw_state *ts)
>   {
> -	struct io_ring_ctx *ctx = req->file->private_data;
> -	struct io_msg *msg = io_kiocb_to_cmd(req, struct io_msg);
> -	struct task_struct *task = READ_ONCE(ctx->submitter_task);

Not about this series particularly, but sounds like msg requests
should be REQ_F_INFLIGHT, but there is a chance lazy file assignment
is enough.

> -
> -	if (unlikely(!task))
> -		return -EOWNERDEAD;
> +	struct io_ring_ctx *ctx = req->ctx;
>   
> -	init_task_work(&msg->tw, func);
> -	if (task_work_add(task, &msg->tw, TWA_SIGNAL))
> -		return -EOWNERDEAD;
> +	io_add_aux_cqe(ctx, req->cqe.user_data, req->cqe.res, req->cqe.flags);
> +	kmem_cache_free(req_cachep, req);
> +	percpu_ref_put(&ctx->refs);
> +}
>   
> -	return IOU_ISSUE_SKIP_COMPLETE;
> +static void io_msg_remote_post(struct io_ring_ctx *ctx, struct io_kiocb *req,
> +			       int res, u32 cflags, u64 user_data)
> +{
> +	req->cqe.user_data = user_data;
> +	io_req_set_res(req, res, cflags);
> +	percpu_ref_get(&ctx->refs);
> +	req->ctx = ctx;
> +	req->task = READ_ONCE(ctx->submitter_task);

Missing a null check, apart from that the patchset looks right

> +	req->io_task_work.func = io_msg_tw_complete;
> +	io_req_task_work_add_remote(req, ctx, IOU_F_TWQ_LAZY_WAKE);
>   }

-- 
Pavel Begunkov