On Tue, Oct 20, 2020 at 02:54:20PM +0800, Jeffle Xu wrote:
> Design of cookie is initially constrained as a per-bio concept. It
> dosn't work well when bio-split needed, and it is really an issue when
> adding support of iopoll for dm devices.
>
> The current algorithm implementation is simple. The returned cookie of
> dm device is actually not used since it is just the cookie of one of
> the cloned bios. Polling of dm device is actually polling on all
> hardware queues (in poll mode) of all underlying target devices.
>
> Signed-off-by: Jeffle Xu <jefflexu@xxxxxxxxxxxxxxxxx>
> ---
> drivers/md/dm-core.h | 1 +
> drivers/md/dm-table.c | 30 ++++++++++++++++++++++++++++++
> drivers/md/dm.c | 39 +++++++++++++++++++++++++++++++++++++++
> 3 files changed, 70 insertions(+)
>
> diff --git a/drivers/md/dm-core.h b/drivers/md/dm-core.h
> index d522093cb39d..f18e066beffe 100644
> --- a/drivers/md/dm-core.h
> +++ b/drivers/md/dm-core.h
> @@ -187,4 +187,5 @@ extern atomic_t dm_global_event_nr;
> extern wait_queue_head_t dm_global_eventq;
> void dm_issue_global_event(void);
>
> +int dm_io_poll(struct request_queue *q, blk_qc_t cookie);
> #endif
> diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
> index ce543b761be7..634b79842519 100644
> --- a/drivers/md/dm-table.c
> +++ b/drivers/md/dm-table.c
> @@ -1809,6 +1809,31 @@ static bool dm_table_requires_stable_pages(struct dm_table *t)
> return false;
> }
>
> +static int device_not_support_poll(struct dm_target *ti, struct dm_dev *dev,
> + sector_t start, sector_t len, void *data)
> +{
> + struct request_queue *q = bdev_get_queue(dev->bdev);
> +
> + return q && !(q->queue_flags & QUEUE_FLAG_POLL);
> +}
> +
> +bool dm_table_supports_poll(struct dm_table *t)
> +{
> + struct dm_target *ti;
> + unsigned int i;
> +
> + /* Ensure that all targets support DAX. */
> + for (i = 0; i < dm_table_get_num_targets(t); i++) {
> + ti = dm_table_get_target(t, i);
> +
> + if (!ti->type->iterate_devices ||
> + ti->type->iterate_devices(ti, device_not_support_poll, NULL))
> + return false;
> + }
> +
> + return true;
> +}
> +
> void dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
> struct queue_limits *limits)
> {
> @@ -1901,6 +1926,11 @@ void dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
> #endif
>
> blk_queue_update_readahead(q);
> +
> + if (dm_table_supports_poll(t)) {
> + q->poll_fn = dm_io_poll;
> + blk_queue_flag_set(QUEUE_FLAG_POLL, q);
> + }
> }
>
> unsigned int dm_table_get_num_targets(struct dm_table *t)
> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> index c18fc2548518..4eceaf87ffd4 100644
> --- a/drivers/md/dm.c
> +++ b/drivers/md/dm.c
> @@ -1666,6 +1666,45 @@ static blk_qc_t dm_submit_bio(struct bio *bio)
> return ret;
> }
>
> +static int dm_poll_one_dev(struct request_queue *q, blk_qc_t cookie)
> +{
> + /* Iterate polling on all polling queues for mq device */
> + if (queue_is_mq(q)) {
> + struct blk_mq_hw_ctx *hctx;
> + int i, ret = 0;
> +
> + if (!percpu_ref_tryget(&q->q_usage_counter))
> + return 0;
> +
> + queue_for_each_poll_hw_ctx(q, hctx, i) {
> + ret += q->mq_ops->poll(hctx);
> + }
IMO, this way may not be accepted from performance viewpoint, .poll()
often requires per-hw-queue lock. So in case of > 1 io thread,
contention/cache ping-pong on hw queue resource can be very serious.
I guess you may have to find one way to pass correct cookie to ->poll().
Thanks,
Ming
