So Ilooked into the "unlocked" direct I/O case, and I think the current
code using dio_sem is really sketchy.  What btrfs really needs to do is
take i_rwsem shared by default for direct writes, and only upgrade to
the exclusive lock when needed, similar to xfs and the WIP ext4 code.

While looking for that I also noticed two other things:

 - check_direct_IO looks pretty bogus
 - btrfs_direct_IO really should be split and folded into the two
   callers

Untested patches attached.  The first should probably go into a prep
patch, and the second could be folded into this one.