> The behavior in the regular case is unchanged, spinning for up to 1ms,
> but the open-coded poll loop is replaced with read_poll_timeout_atomic()
> as suggested by Andrew Lunn.

Hmm, spinning 1ms is still a lot. Can't we just use read_poll_timeout()
for the whole timeout? I can't see that it will cause a regression. But
please correct me if I am wrong.

The series looks good for me apart from the above.