On Fri, 13 Jul 2018, Ming Lei wrote:
> This patch introduces blk_mq_pm_add_request() which is called after
> allocating one request. Also blk_mq_pm_put_request() is introduced
> and called after one request is freed.
>
> For blk-mq, it can be quite expensive to accounting in-flight IOs,
> so this patch calls pm_runtime_mark_last_busy() simply after each IO
> is done, instead of doing that only after the last in-flight IO is done.
> This way is still workable, since the active non-PM IO will be checked
> in blk_pre_runtime_suspend(), and runtime suspend will be prevented
> if there is any active non-PM IO.
>
> Turns out that sync between runtime PM and IO path has to be done
> for avoiding race, this patch applies one seqlock for this purpose.
> So the cost introduced in fast IO path can be minimized given seqlock
> is often used in fast path, such as reading jiffies &tick, or d_walk(),
> ...
>
> Cc: "Rafael J. Wysocki" <rjw@xxxxxxxxxxxxx>
> Cc: Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>
> Cc: linux-pm@xxxxxxxxxxxxxxx
> Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> Cc: Christoph Hellwig <hch@xxxxxx>
> Cc: Bart Van Assche <bart.vanassche@xxxxxxx>
> Cc: Hannes Reinecke <hare@xxxxxxx>
> Cc: Johannes Thumshirn <jthumshirn@xxxxxxx>
> Cc: Adrian Hunter <adrian.hunter@xxxxxxxxx>
> Cc: "James E.J. Bottomley" <jejb@xxxxxxxxxxxxxxxxxx>
> Cc: "Martin K. Petersen" <martin.petersen@xxxxxxxxxx>
> Cc: linux-scsi@xxxxxxxxxxxxxxx
> Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
> ---
> +static void blk_mq_post_runtime_suspend(struct request_queue *q, int err)
> +{
> + if (!blk_mq_support_runtime_pm(q))
> + return;
> +
> + write_seqlock_irq(&q->rpm_lock);
> + __blk_post_runtime_suspend(q, err);
> + write_sequnlock_irq(&q->rpm_lock);
> +}
> +
> +static void blk_mq_pre_runtime_resume(struct request_queue *q)
> +{
> + if (!blk_mq_support_runtime_pm(q))
> + return;
> +
> + write_seqlock_irq(&q->rpm_lock);
> + q->rpm_status = RPM_RESUMING;
> + write_sequnlock_irq(&q->rpm_lock);
> +}
> +
> +static void blk_mq_post_runtime_resume(struct request_queue *q, int err)
> +{
> + if (!blk_mq_support_runtime_pm(q))
> + return;
> +
> + write_seqlock_irq(&q->rpm_lock);
> + __blk_post_runtime_resume(q, err);
> + write_sequnlock_irq(&q->rpm_lock);
> +}
> +
> +static void blk_mq_set_runtime_active(struct request_queue *q)
> +{
> + if (!blk_mq_support_runtime_pm(q))
> + return;
> +
> + write_seqlock_irq(&q->rpm_lock);
> + __blk_set_runtime_active(q);
> + write_sequnlock_irq(&q->rpm_lock);
> +}
Would the code be cleaner if these routines were written inline,
like their non-mq counterparts?
Alan Stern
