[PATCH 1/2] block: add queue flag to always poll

[Jeff Moyer]

Jon Derrick <jonathan.derrick at intel.com> writes:

> This patch adds poll-by-default functionality back for 4.6 by adding a
> queue flag which specifies that it should always try polling, rather
> than only if the io specifies it.

I'm not against the functionality, but it somehow feels like you've
implemented this at the wrong layer.  I'd much rather polling be forced
on for the file or the mountpoint, and then all requests would come down
with RWF_HIPRI and there would be no special casing in the direct I/O
code.

In case others disagree, I've provided a couple of comments below.
Really, though, I think this is implemented upside-down.

-Jeff

> Signed-off-by: Jon Derrick <jonathan.derrick at intel.com>
> ---
>  block/blk-core.c       | 8 ++++++++
>  fs/direct-io.c         | 7 ++++++-
>  include/linux/blkdev.h | 2 ++
>  3 files changed, 16 insertions(+), 1 deletion(-)
>
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 827f8ba..d85f913 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -3335,6 +3335,14 @@ void blk_finish_plug(struct blk_plug *plug)
>  }
>  EXPORT_SYMBOL(blk_finish_plug);
>  
> +bool blk_force_poll(struct request_queue *q)
> +{
> +	if (!q->mq_ops || !q->mq_ops->poll ||
> +	    !test_bit(QUEUE_FLAG_POLL_FORCE, &q->queue_flags))
> +		return false;
> +	return true;
> +}

The flag shouldn't be set if it doesn't make sense, and these checks are
re-done inside blk_poll, anyway.  Just do the test bit:

	return test_bit(QUEUE_FLAG_POLL_FORCE, &q->queue_flags);

>  bool blk_poll(struct request_queue *q, blk_qc_t cookie)
>  {
>  	struct blk_plug *plug;
> diff --git a/fs/direct-io.c b/fs/direct-io.c
> index 476f1ec..2775552 100644
> --- a/fs/direct-io.c
> +++ b/fs/direct-io.c
> @@ -450,7 +450,12 @@ static struct bio *dio_await_one(struct dio *dio)
>  		__set_current_state(TASK_UNINTERRUPTIBLE);
>  		dio->waiter = current;
>  		spin_unlock_irqrestore(&dio->bio_lock, flags);
> -		if (!(dio->iocb->ki_flags & IOCB_HIPRI) ||
> +		/*
> +		 * Polling must be enabled explicitly on a per-IO basis,
> +		 * or through the queue's sysfs io_poll_force control
> +		 */
> +		if (!((dio->iocb->ki_flags & IOCB_HIPRI) ||
> +		      (blk_force_poll(bdev_get_queue(dio->bio_bdev)))) ||
>  		    !blk_poll(bdev_get_queue(dio->bio_bdev), dio->bio_cookie))

Make a local variable for the queue, please.

-Jeff