It seems that our kernels (and likely stock ones as well), both 2.4.x and
2.6.x based, violate POSIX for msync.
http://www.opengroup.org/onlinepubs/009695399/functions/msync.html
If msync() causes any write to a file, the file's st_ctime and st_mtime
fields shall be marked for update.

See the attached testcase.  If test_msync is changed to false,
the tests passes, so certainly write marks the st_mtime for update.

This is distilled from a LSB 1.3 testcase, but that test is flawed
and thus usually succeeds although msync really doesn't update
st_mtime.  The flaw is that the test doesn't sleep between writing
the file and calling time (or it could wait longer and increase start by 2
seconds), so usually st_mtime will be equal to start time.
With the added sleep in between writing/mmaping the file and time (&start)
it fails reliably on all arches.

I think msync.c needs to propagate up from filemap_sync_pte whether there were
any dirty pages and call inode_update_time before exiting if there were some.