On Fri, Apr 23, 2021 at 03:05:58PM -0700, Keith Busch wrote:

The block layer knows how to deal with polled requests. Let the NVMe
driver use the previously reserved user "flags" fields to define an
option to allocate the request from the polled hardware contexts. If
polling is not enabled, then the block layer will automatically fallback
to a non-polled request.

So this only support synchronous polling for a single command.  What
use case do we have for that?  I think io_uring based polling would
be much more useful once we support NVMe passthrough through that.