[RESEND PATCHSET] direct-io: unify asyn/sync completion paths and fix completion bugs

[Tejun Heo <htejun@gmail.com>]

[forgot to include lkml, resending.  sorry ppl.]

Hello, Zach Brown, all.

Chul-Woong Yang of Aratech reported kernel oops after failed aio
(partially mapped request) on 2.6.19-rc4 and I spent three days
chasing the bug and hacking direct-io only to find out that there are
pending patches in -mm.

Oh, well.. As my patches are almost ready now, I'm posting it, for
cross-check if nothing else.  The biggest difference between your fix
and mine is that instead of merging waiting code, I just unified whole
completion paths.  ie. Sync requests are handled exactly the same as
async requests.  They are pre-dirtied, completed from bio completion
callback and redirtied from bio_dirty_work if necessary.  That
simplified the code a lot and made it humanly readable. :-)

The only concern is possible performance impact for sync DIOs due to
the change in how dirtying is done.  As it's done per-bio, the chance
of being written back between pre-issue dirtying and post-completion
checking should be pretty slim, so I don't think it would have any
noticeable effect.  Async DIOs have been handled that way after all.

I reshuffled things a bit just before finding out your patch and
didn't test it thoroughly after that, but it should be enough to get
the idea.

Arghh... The lesson here is... check -mm before starting hacking
unfamiliar code.  I'm back to libata now.

Thanks.

--
tejun