Re: [PATCH 4/4] io_uring: implement multishot mode for accept

[Jens Axboe <axboe@kernel.dk>]

On 5/8/22 9:32 AM, Hao Xu wrote:
> From: Hao Xu <howeyxu@tencent.com>
> 
> Refactor io_accept() to support multishot mode.
> 
> theoretical analysis:
>   1) when connections come in fast
>     - singleshot:
>               add accept sqe(userpsace) --> accept inline

userspace here too, like in the cover letter.

> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index e0d12af04cd1..f21172913336 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -1146,6 +1146,7 @@ static const struct io_op_def io_op_defs[] = {
>  		.unbound_nonreg_file	= 1,
>  		.pollin			= 1,
>  		.poll_exclusive		= 1,
> +		.ioprio			= 1,	/* used for flags */
>  	},
>  	[IORING_OP_ASYNC_CANCEL] = {
>  		.audit_skip		= 1,
> @@ -5706,6 +5707,7 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
>  static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>  {
>  	struct io_accept *accept = &req->accept;
> +	unsigned flags;
>  
>  	if (sqe->len || sqe->buf_index)
>  		return -EINVAL;
> @@ -5714,19 +5716,26 @@ static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>  	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
>  	accept->flags = READ_ONCE(sqe->accept_flags);
>  	accept->nofile = rlimit(RLIMIT_NOFILE);
> +	flags = READ_ONCE(sqe->ioprio);
> +	if (flags & ~IORING_ACCEPT_MULTISHOT)
> +		return -EINVAL;
>  
>  	accept->file_slot = READ_ONCE(sqe->file_index);
> -	if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))
> +	if (accept->file_slot && ((accept->flags & SOCK_CLOEXEC) ||
> +	   flags & IORING_ACCEPT_MULTISHOT))
>  		return -EINVAL;

So something ala:

	if (accept_>file_slot) {
		if (accept->flags & SOCK_CLOEXEC)
			return -EINVAL;
		if (flags & IORING_ACCEPT_MULTISHOT &&
		    accept->file_slot != IORING_FILE_INDEX_ALLOC)
			return -EINVAL;
	}

when rebased as mentioned in the cover reply.


	if (accept->file_slot && (accept->flags & SOCK_CLOEXEC))

>  static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
>  {
> +	struct io_ring_ctx *ctx = req->ctx;
>  	struct io_accept *accept = &req->accept;
>  	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
>  	unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
> @@ -5734,6 +5743,7 @@ static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
>  	struct file *file;
>  	int ret, fd;
>  
> +retry:
>  	if (!fixed) {
>  		fd = __get_unused_fd_flags(accept->flags, accept->nofile);
>  		if (unlikely(fd < 0))
> @@ -5745,8 +5755,12 @@ static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
>  		if (!fixed)
>  			put_unused_fd(fd);
>  		ret = PTR_ERR(file);
> -		if (ret == -EAGAIN && force_nonblock)
> -			return -EAGAIN;
> +		if (ret == -EAGAIN && force_nonblock) {
> +			if ((req->flags & IO_APOLL_MULTI_POLLED) ==
> +			    IO_APOLL_MULTI_POLLED)
> +				ret = 0;
> +			return ret;
> +		}

Could probably do with a comment here for this check, it's not
immediately obvious unless you've already been deep in this part.

> @@ -5757,8 +5771,26 @@ static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
>  		ret = io_install_fixed_file(req, file, issue_flags,
>  					    accept->file_slot - 1);
>  	}
> -	__io_req_complete(req, issue_flags, ret, 0);
> -	return 0;
> +
> +	if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
> +		__io_req_complete(req, issue_flags, ret, 0);
> +		return 0;
> +	}
> +	if (ret >= 0) {
> +		bool filled;
> +
> +		spin_lock(&ctx->completion_lock);
> +		filled = io_fill_cqe_aux(ctx, req->cqe.user_data, ret,
> +					 IORING_CQE_F_MORE);
> +		io_commit_cqring(ctx);
> +		spin_unlock(&ctx->completion_lock);
> +		if (!filled)
> +			return -ECANCELED;
> +		io_cqring_ev_posted(ctx);
> +		goto retry;
> +	}
> +
> +	return ret;
>  }

As mentioned, I still think this should be:

		...
		if (filled) {
			io_cqring_ev_posted(ctx);
			goto retry;
		}
		ret = -ECANCELED;
	}

	return ret;

-- 
Jens Axboe