I like this idea, but there are a couple issues here.

First the flag per command really doesn't work - we need a creation
time flag.  Unfortunately the existing io_setup system call doesn't
take flags, so we'll need to add a new one.

Second we need a check that the polling mode is actually supported
for a given file descriptor.

Third this hardcodes block device knowlege into the block layer. We
really need move it into a method.   See my third attached patch
for an (untested) idea how to make that work.  The first two are
just cleanups.

Fourth we already have aio ->poll support.  So this needs a different
naming scheme, e.g. *iopoll*.