without checking if there were any real blocks already in the
 speculative allocation area. This could result in an allocation
 that overlaps pre-allocated space. This would result in an ASSERT
 failure in debug kernels, or invalid output from xfs_bmap.
 The code will now only do speculative allocation if we are writing
 beyond the current allocation eof.

SGI Modid: xfs-linux:xfs-kern:171163a