Re: IORING_OP_READ and O_NONBLOCK behaviour

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

On 02/09/2020 18:35, Jens Axboe wrote:
> On 9/2/20 9:26 AM, Pavel Begunkov wrote:
>> On 02/09/2020 17:45, Jens Axboe wrote:
>>> On 9/2/20 4:09 AM, Norman Maurer wrote:
>>>> Hi there,
>>>>
>>>> We are currently working on integrating io_uring into netty and found
>>>> some “suprising” behaviour which seems like a bug to me.
>>>>
>>>> When a socket is marked as non blocking (accepted with O_NONBLOCK flag
>>>> set) and there is no data to be read IORING_OP_READ should complete
>>>> directly with EAGAIN or EWOULDBLOCK. This is not the case and it
>>>> basically blocks forever until there is some data to read. Is this
>>>> expected ?
>>>>
>>>> This seems to be somehow related to a bug that was fixed for
>>>> IO_URING_ACCEPT with non blocking sockets:
>>>>
>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.8&id=e697deed834de15d2322d0619d51893022c90ea2
>>>
>>> I agree with you that this is a bug, in general it's useful (and
>>> expected) that we'd return -EAGAIN for that case. I'll take a look.
>>>
>>
>> That's I mentioned that doing retries for nonblock requests in
>> io_wq_submit_work() doesn't look consistent. I think killing it
>> off may help.
> 
> Right, we should not retry those _in general_, the exception is regular
> files or block devices to handle IOPOLL retry where we do need it. The
> below is what I came up with for this one. Might not hurt to make this
> more explicit for 5.10.

Hmm, I didn't checked it, but if we 

> 
> 
> commit c78e0f02c3861b5b176b2f79552677b3604deb76
> Author: Jens Axboe <axboe@kernel.dk>
> Date:   Wed Sep 2 09:30:31 2020 -0600
> 
>     io_uring: no read-retry on -EAGAIN error and O_NONBLOCK marked file
>     
>     Actually two things that need fixing up here:
>     
>     - The io_rw_reissue() -EAGAIN retry is explicit to block devices and
>       regular files, so don't ever attempt to do that on other types of
>       files.
>     
>     - If we hit -EAGAIN on a nonblock marked file, don't arm poll handler for
>       it. It should just complete with -EAGAIN.
>     
>     Cc: stable@vger.kernel.org
>     Reported-by: Norman Maurer <norman.maurer@googlemail.com>
>     Signed-off-by: Jens Axboe <axboe@kernel.dk>
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index b1ccd7072d93..dc27cd5b8ad6 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -2300,8 +2300,11 @@ static bool io_resubmit_prep(struct io_kiocb *req, int error)
>  static bool io_rw_reissue(struct io_kiocb *req, long res)
>  {
>  #ifdef CONFIG_BLOCK
> +	umode_t mode = file_inode(req->file)->i_mode;
>  	int ret;
>  
> +	if (!S_ISBLK(mode) && !S_ISREG(mode))
> +		return false;
>  	if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
>  		return false;
>  
> @@ -3146,6 +3149,9 @@ static int io_read(struct io_kiocb *req, bool force_nonblock,
>  		/* IOPOLL retry should happen for io-wq threads */
>  		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
>  			goto done;
> +		/* no retry on NONBLOCK marked file */
> +		if (kiocb->ki_flags & IOCB_NOWAIT)
> +			goto done;

We clearing and setting IOCB_NOWAIT depending on @force_nonblock, so it may not
work. E.g. with IOSQE_IO_ASYNC io_read() will clear it at the beginning.
Maybe REQ_F_NOWAIT?

>  		/* some cases will consume bytes even on error returns */
>  		iov_iter_revert(iter, iov_count - iov_iter_count(iter));
>  		ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
> 

-- 
Pavel Begunkov