[PATCH 08/15] io_uring: don't sleep when polling for I/O

[Sagi Grimberg]

On 5/12/21 6:15 AM, Christoph Hellwig wrote:
> There is no point in sleeping for the expected I/O completion timeout
> in the io_uring async polling model as we never poll for a specific
> I/O.  Split the boolean spin argument to blk_poll into a set of flags
> to control sleeping and the oneshot behavior separately.
> 
> Signed-off-by: Christoph Hellwig <hch at lst.de>
> ---
>   block/blk-mq.c           | 18 ++++++++----------
>   drivers/nvme/host/core.c |  2 +-
>   fs/block_dev.c           |  8 ++++----
>   fs/io_uring.c            | 14 +++++++-------
>   fs/iomap/direct-io.c     |  6 +++---
>   include/linux/blkdev.h   |  6 +++++-
>   include/linux/fs.h       |  2 +-
>   include/linux/iomap.h    |  2 +-
>   mm/page_io.c             |  2 +-
>   9 files changed, 31 insertions(+), 29 deletions(-)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index ac0b517c5503..164e39d34bf6 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -3873,7 +3873,7 @@ static bool blk_mq_poll_hybrid(struct request_queue *q, blk_qc_t qc)
>   }
>   
>   static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
> -		bool spin)
> +		unsigned int flags)
>   {
>   	struct blk_mq_hw_ctx *hctx = blk_qc_to_hctx(q, cookie);
>   	long state = current->state;
> @@ -3896,7 +3896,7 @@ static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
>   		if (current->state == TASK_RUNNING)
>   			return 1;
>   
> -		if (ret < 0 || !spin)
> +		if (ret < 0 || (flags & BLK_POLL_ONESHOT))
>   			break;
>   		cpu_relax();
>   	} while (!need_resched());
> @@ -3909,15 +3909,13 @@ static int blk_mq_poll_classic(struct request_queue *q, blk_qc_t cookie,
>    * blk_poll - poll for IO completions
>    * @q:  the queue
>    * @cookie: cookie passed back at IO submission time
> - * @spin: whether to spin for completions
> + * @flags: BLK_POLL_* flags that control the behavior
>    *
>    * Description:
>    *    Poll for completions on the passed in queue. Returns number of
> - *    completed entries found. If @spin is true, then blk_poll will continue
> - *    looping until at least one completion is found, unless the task is
> - *    otherwise marked running (or we need to reschedule).
> + *    completed entries found.
>    */
> -int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
> +int blk_poll(struct request_queue *q, blk_qc_t cookie, unsigned int flags)
>   {
>   	if (cookie == BLK_QC_T_NONE ||
>   	    !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
> @@ -3926,12 +3924,12 @@ int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
>   	if (current->plug)
>   		blk_flush_plug_list(current->plug, false);
>   
> -	/* If specified not to spin, we also should not sleep. */
> -	if (spin && q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
> +	if (!(flags & BLK_POLL_NOSLEEP) &&
> +	    q->poll_nsec != BLK_MQ_POLL_CLASSIC) {
>   		if (blk_mq_poll_hybrid(q, cookie))
>   			return 1;
>   	}
> -	return blk_mq_poll_classic(q, cookie, spin);
> +	return blk_mq_poll_classic(q, cookie, flags);

I think that the combination of oneshot and nosleep flags to replace
a boolen spin is a little hard to follow (especially that spin doesn't
mean spinning without sleeping).

Maybe we should break it to:
1. replace spin to flags with ONESHOT passed from io_uring (direct
    replacement)
2. add NOSLEEP passed from io_uring as there is no need for it.

Just a suggestion though that would help (me at least) to follow
this more easily.