> The only trick here is that you may get, in the driver, EARLY WARNING
> but still have satisfied the I/O request completely. In this case, the
> driver has to maintain state so that the *next* write gets a (synthesized
> and immediate) 'short' write (with a value returned such that the
> application believes (correctly) that no data was moved for this I/O
> operation).

Okay, but unless I'm missing something, that means a synthesized return of
exactly zero. This seems (to me) to conflict with your earlier statement
that the driver would guarantee a "non-zero residual" since I have been
reading that as implying a "short write" returning something between zero
and the request count.

Admittedly I came into this thread late, but I thought I smelled an API
leak in your proposal -- an ambiguity when the correct value to return is
a "short write" of exactly zero.

Todd Whitesel
toddpw @ best.com