On Wed, Aug 14, 2013 at 11:10:54AM +0200, Jan Kara wrote:

  Hello,

  this is second iteration of patches to fix handling of O_SYNC AIO DIO.
Since previous version I've addressed Dave's comments:
 - slightly expanded changelog of the first patch
 - workqueue is now created with parameters allowing paralelism
 - workqueue name contains sb->s_id
 - workqueue is created on demand (I decided to do this to reduce the overhead
   in unnecessary cases)

The patchset survives xfstests run for ext4 & xfs so it should be sane. Since
this touches several filesystems (although only ext4 & xfs are non-trivial),
the question is who should carry these patches. Maybe Al? But since xfs and
ext4 changes are non-trivial, I'd like to have a review from their
developers...

Looks sane, except that I'd probably put destroying the queue after
evict_inodes(), next to ->put_super() call.

Said that, there's another interesting problem in the code affected by that
sucker: generic_file_aio_write() might very well sync the wrong range.
Consider O_APPEND case; __generic_file_aio_write() will call
generic_write_checks(), which will update its copy of pos, and proceed to
write starting from there.  All right and proper, but then we return into
generic_file_aio_write() and sync the range of the right length, starting
at the *original* value of pos...