Hi,

With the support in 5.16-rc1 for allocating and completing batches of
IO, the one missing piece is passing down a list of requests for issue.
Drivers can take advantage of this by defining an mq_ops->queue_rqs()
hook.

This implements it for NVMe, allowing copy of multiple commands in one
swoop.

This is good for around a 500K IOPS/core improvement in my testing,
which is around a 5-6% improvement in efficiency.

Changes since v1:

- Addressed review comments
- Rebase on top of Ming's hctx lock change
- Clean ups
- Bypass for shared tags

-- 
Jens Axboe