stop using ioends for direct write completions

[Christoph Hellwig]

On Tue, Feb 02, 2016 at 10:31:18AM -0500, Brian Foster wrote:
> FWIW, I don't see any such review comments against the three versions of
> the "DIO needs an ioend for writes" patch I have in my mailbox, but I
> easily could have missed something..? But if there wasn't time, then
> fair enough.

I'll have to look at the mailboxes, but I remember Dave sending this
out and complaining.

> I'm just looking for context. I don't have much of an opinion on which
> approach is used here. If it simplifies COW, then that seems good enough
> reason to me to take this approach. I'm pointing this out more because
> this code seems to have been rewritten the last couple of times we
> needed to fix something, which makes backports particularly annoying.
> The two patches above were associated with a broader enhancement and a
> bug fix (respectively) as a sort of justification, whereas this post had
> a much more vague purpose from what I could tell, and therefore why I at
> least hadn't taken the time to review it.
> 
> If COW is the primary motivator, perhaps we can bundle it with that
> work?

The prime motivator is to:

 (1) avoid a pointless memory allocation
 (2) avoid a pointless context switch
 (3) avoid pointless code complexity

COW is just another case where these show up.