Because this happens at high thread counts on high IOPS devices
doing mixed read/write AIO-DIO to a single file at about a million
iops:

   64.09%     0.21%  [kernel]            [k] io_submit_one
   - 63.87% io_submit_one
      - 44.33% aio_write
         - 42.70% xfs_file_write_iter
            - 41.32% xfs_file_dio_write_aligned
               - 25.51% xfs_file_write_checks
                  - 21.60% _raw_spin_lock
                     - 21.59% do_raw_spin_lock
                        - 19.70% __pv_queued_spin_lock_slowpath

This also happens of the IO completion IO path:

   22.89%     0.69%  [kernel]            [k] xfs_dio_write_end_io
   - 22.49% xfs_dio_write_end_io
      - 21.79% _raw_spin_lock
         - 20.97% do_raw_spin_lock
            - 20.10% __pv_queued_spin_lock_slowpath

IOWs, fio is burning ~14 whole CPUs on this spin lock.

So, do an unlocked check against inode size first, then if we are
at/beyond EOF, take the spinlock and recheck. This makes the
spinlock disappear from the overwrite fastpath.

I'd like to report that fixing this makes things go faster. It
doesn't - it just exposes the the XFS_ILOCK as the next severe
contention point doing extent mapping lookups, and that now burns
all the 14 CPUs this spinlock was burning.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>