Newer kernel will reward iov iter back to original
position for direct IO, see following codes:

    
            iov_iter_revert(from, write_len -iov_iter_count(from));--------->here
    out:
            return written;
    }
    EXPORT_SYMBOL(generic_file_direct_write);

This break assumptions from Lustre and caused problem
when Lustre need split one IO to several io loop, considering
4M block IO for 1 MiB stripe file, it will submit first 1MiB IO
4 times and caused data corruptions finally.

Since generic kernel varies from different kernel versions,
we'd better fix this problem by move iov iter forward by
Lustre itself.