(Emacs: -*- indented-text -*-)
Things to be done ($Date: 1999/02/07 00:32:42 $):

        0) Reset some PPS variables when setting the time...
	   Re-think about the tolerance being independent from MAXFREQ...
	   Really think about the mess with the line discipline...

	1) Prove that the new code is correct and that it does not
           break anything.  Check if the Linux timekeeping code has
           any benefit from the new code.  Most likely the precision
           with the new code should be around few microseconds while
           the method using select() is limited to a few (10)
           milliseconds in kernel 2.0. [looks good, obsolete in 2.2]

        2) Make sure that do_gettimeoffset returns a consistent time
           offset (it performs internal calibrations occasionally that
           also take time).  The offset should be as precise as
           possible. [looks good, obsolete in 2.2]

	3) Extend the patch for architecture other than i386.  As I
	   only have a Intel PC, I can't test other platforms.
	   Feedback is welcome!  [done, but completely untested by me]

	4) Get the time of the pulse more exactly (must be done in the
	   serial driver's interrupt routine); that is earlier and
	   with a fixed time offset (currently possibly variable).  At
	   least the propagation delay should be determined (a tricky
	   job).  I'm no longer sure whether it is really needed...

	5) Discussions with several interested users have shown that I
	   should write a NTP/PPS MINI-HOWTO (still to be done, but
	   have a look at README)... [I preferred to make a
	   hyperlinked HTML version of the NTP-FAQ instead]

	6) Improve Linux support in xntp (NTP software for UNIX) even
	   more. [temporary hack]

	7) Implement the official new PPS API in kernel 2.1/2.2.
