NetBSD/dist/ntp/sntp/README

SNTP (Simple Network Time Protocol Utility) - Version 1.6
----------------------------------------------------------

Please read the file Copyright first.  Also note that the file RFC2030.TXT is
David Mills's copyright and not the author's - it is just a copy of the RFC
that is available from so many Internet archives.

RFC 1305 (Network Time Protocol - NTP) is an attempt to provide globally
consistent timestamps in an extremely hostile environment; it is fiendishly
complicated and an impressive piece of virtuosity.  RFC 2030 (Simple Network
Time Protocol - SNTP) which supersedes RFC 1769 describes a subset of this that
will give excellent accuracy in most environments encountered in practice; it
uses only the obvious algorithms that have been used since time immemorial.

WARNING: the text version of RFC 1305 is incomplete, and omits the tables that
are in the Postscript version.  Unfortunately, these contain the only copy of
some critical information.

draft-mills-sntp-v4-00.txt is the next proposed revision of RFC 2030,
and the current goal is to have this code implement that specification.

SNTP Servers - Some Little-Known Facts
--------------------------------------

RFC 2030 states that SNTP clients should be used only at the lowest level,
which is good practice.  It then states that SNTP servers should be used only
at stratum 1 (i.e. top level), which is bizarre!  A far saner use of them would
be for the very lowest level of server, exporting solely to local clients that
do not themselves act as servers to ANY system (e.g. on a Netware server,
exporting only to the PCs that it manages).

[There is missing language in the previous paragraph.  SNTP is designed
to be used in 2 cases: as a client at the lowest levels of the timing
hierarchy, or as a server of last resort at stratum 1 when connected to
a modem or radio clock.]

[This is as far as I have updated this file as part of the upgrade.]

If the NTP network were being run as a directed acyclic graph (i.e. using SNTP
rather than full NTP), with a diameter of D links and a maximum error per link
of E, the maximum synchronisation error would be D*E.  Reasonable figures for D
and E are 5 and 0.1 seconds, so this would be adequate for most uses.  Note
that the fact that the graph is acyclic is critical, which is one reason why
SNTP client/servers must NEVER be embedded WITHIN an NTP network.

The other reason is that inserting SNTP client/servers at a low stratum (but
not the root) of an NTP network could easily break NTP!  See RFC 1305 for why,
but don't expect the answer to stand out at you.  It would be easy to extend
SNTP to a full-function client/server application, thus making it into a true
alternative to ntp, but this incompatibility is why it MUST NOT be done.

The above does not mean that the SNTP approach is unsatisfactory, but only that
it is incompatible with full NTP.  The author would favour a complete SNTP
network using the SNTP approach, and the statistical error reduction used in
SNTP, but it actually addresses a slightly different problem from that
addressed by NTP.  TANSTAAFL.

FINAL WARNING: do NOT use this program to serve NTP requests from outside the
systems that you manage.  If you do this, and manage to break the time
synchronisation on other people's systems, you will be regarded very
unfavourably.  Actually, this should be possible only if their NTP client is
completely broken, because SNTP does its damnedest to declare its packets as
the lowest form of NTP timestamp.



SNTP and its Assumptions
-------------------------

SNTP is intended to be a straightforward SNTP daemon/utility that is easy to
build on any reasonable Unix platform (and most near-Unix ones), whether or not
it has ever been ported to them before.  It is intended to answer the following
requirements, either by challenge and response or the less reliable broadcast
method:

    A simple command to run on Unix systems that will check the time
    and optionally drift compared with a known, local and reliable NTP
    time server.  No privilege is required just to read the time and
    estimate the drift.

    A client for Unix systems that will synchronise the time from a known,
    local and reliable NTP time server.  This is probably the most common
    one, and the need that caused the program to be written.

    A server for Unix systems that are synchronised other than by NTP
    methods and that need to synchronise other systems by NTP.  This is
    the classroom of PCs with a central server scenario.  It is NOT
    intended to work as a peer with true NTP servers, and won't.

    A simple method by which two or more Unix systems can keep themselves
    synchronised using what is becoming a standard protocol.  Yes, I know
    that there are half-a-dozen other such methods.

    A base for building non-Unix SNTP clients.  Some 3/4 of the code
    (including all of the complicated algorithms and NTP packet handling)
    should work, unchanged, on any system with an ANSI/ISO C compiler.

There are full tracing facilities and a lot of paranoia in the code to check
for bad packets (more than in ntp) which may need relaxing in the light of
experience.  Unfortunately, RFC 1305 does not include a precise description of
the data protocol, despite its length, and there are some internal
inconsistencies and differences between it and RFC 2030 and ntp's behaviour.

WARNING: SNTP has not been tested in conjunction with ntp broadcasts or ntp
clients, as the ability to do so was not available to the author.  It is very
unlikely that it won't work, but you should check.  Much of the paranoid code
is only partially tested, too, because it is dealing with cases that are very
hard to provoke.

It assumes that the local network is tolerably secure and that any accessible
NTP or SNTP servers are trustworthy.  It also makes no attempt to check that
it has been installed and is being used correctly (e.g. at an appropriate
priority) or that the changes it makes have the desired effect.  When you first 
use it, you should both run it in display mode and use the date command as a
cross-check.

Furthermore, it does not attempt to solve all of the problems addressed by the
NTP protocol and you should NOT use it if any of those problems are likely to
cause you serious trouble.  If they are, bite the bullet and implement ntp, or
buy a fancy time-server.


Building SNTP
-------------

The contents of the distribution are:

README        -    this file
Copyright     -    the copyright notice and conditions of use
Makefile      -    the makefile, with comments for several systems
header.h      -    the main header (almost entirely portable)
kludges.h     -    dirty kludges for difficult systems
internet.h    -    a very small header for internet.c and socket.c
main.c        -    most of the source (almost entirely portable)
unix.c        -    just for isatty, sleep and locking
internet.c    -    Internet host and service name lookup
socket.c      -    the Berkeley socket code
sntp.1        -    the man page
RFC2030.TXT   -    the SNTPv4 specification

All you SHOULD need to do is to uncomment the settings in file Makefile for
your system or to add new ones.  But real life is not always so simple.  As
POSIX does not yet define sub-second timers, Internet addressing facilities,
sockets etc., the code has to rely on the facilities described in the
ill-defined and non-standard 'X/Open' documents and the almost totally
unspecified 'BSD' extensions.

Most hacks should be limited to the compiler options (e.g. setting flags like
_XOPEN_SOURCE), but perverse systems may need additions to kludges.h - please
report them to the author.  See Makefile and kludges.h for documentation on
the standard hacks - there only 6, and most are only for obsolete systems.
But, generally, using the generic set of C options usually works with no
further ado.


Sick, Bizarre or non-Unix Systems
---------------------------------

A very few Unix systems and almost all non-Unix systems may need changes to the
code, such as:

    If the system doesn't have Berkeley sockets, you will need to replace
    socket.c and possibly modify internet.h and internet.c.  All of the
    systems for which the author needs this have Berkeley sockets.

    NTP is supposedly an Internet protocol, but is not Internet specific.
    For other types of network, you will need to replace internet.c and
    probably modify internet.h.

    If the system doesn't have gettimeofday or settimeofday, you will
    need to modify timing.c.  If it doesn't have adjtime (e.g. HP-UX
    on PA-RISC before 10.0), you can set -DADJTIME_MISSING and the code
    will compile but the -a option will always give an error.

    If the system has totally broken signal handling, the program will
    hang or crash if it can't reach its name server or responses time
    out.  You may be able to improve matters by hacking internet.c and
    socket.c, but don't bet on it.

    If the the program won't be able to create files in /etc when
    updating the clock, you can use another lock file or even set
    -DLOCKFILE=NULL, which will disable the locking code entirely.  On
    systems that have it, using /var/run would be better than /etc.

    If the the program hangs when flushing outstanding packets (which
    you can tell by setting -W), it may help to set -DNONBLOCK_BROKEN.
    This seems needed only for obsolete systems, like Ultrix.

    If the system isn't Unix, even vaguely, you will probably need to
    modify all of the above, and unix.c as well.

    Note that adjtime is commonly sick, but you don't need to change the
    code - just use the -r option whan making large corrections (see below
    for more details).

Any changes needed to header.h or main.c are bugs.  They may be bugs in the
code or in the compiler or libraries, but they are bugs.  Please prod the
people responsible and tell the author, who may be able to bypass them cleanly
even if they aren't bugs in his code.  The code also makes the following
assumptions, which would be quite hard to remove:

    8-bit bytes.  Strictly, neither ANSI/ISO C nor POSIX require these,
    and there were some very early versions of Unix on systems with other
    byte sizes.  But, without a defined sub-byte facility in C, ....

    At least 32-bit ints.  Well, actually, this wouldn't be too hard to
    remove.  But most Unix programs make this assumption, and I have very
    little interest in the more rudimentary versions of MS-DOS etc.

    An ANSI/ISO C compiler.  It didn't seem worth writing dual-language
    code in 1996.  Tough luck if you haven't got one.

    Tolerably efficient floating-point arithmetic, with at least 13 digits
    (decimal), preferably 15, in the mantissa of doubles.  Ditto.  If you
    want to port this to a toaster, please accept my insincerest sympathies
    and don't bother me.

    A trustworthy local network.  It does not check for DNS, Ethernet,
    packet or other spoofing, and assumes that any accessible NTP or SNTP
    servers are properly synchronised.


Warnings about Installation and Use
-----------------------------------

Anyone attempting to fiddle with the clock on their system should already know
how to write system administration scripts, install daemons and so on.  There
are a few warnings:

    Don't use the broadcast modes unless you really have to, as the
    client-server modes are far more reliable.  The broadcast modes were
    implemented more for virtuosity (a.k.a. SNTP conformance) than use.
    In particular, the error estimates are mere guesses, and may be low
    or even very low.  And even reading broadcasts needs privilege.

    The program is not intended to be installed setuid or setgid, and
    doing so is asking for trouble.  Its ownerships and access modes are
    not important.  It need not be run by root for merely displaying the
    time (even in daemon mode).

    The program does not need to run at a high priority (low in Unix
    terms!) even when being used to set the clock or as a server, except
    when the '-r' option is  used.  However, doing so may improve its
    accuracy.

    Unlike NTP, the SNTP protocol contains no protection against
    client-server loops.  If you set one up, your systems will spin
    themselves off into a disconnected vortex of unreality!

    It will get very confused if another process changes the local time
    while it is running.  There is some locking code in unix.c to prevent
    this program doing this to itself, but it will protect only against
    some errors.  However, the remaining failures should be harmless.

    Don't run it as a server unless you REALLY know what you are doing.
    It should be used as a server only on a system that is properly
    synchronised, by fair means or foul.  If it isn't, you will simply
    perpetrate misinformation.  And remember that broadcasts are most
    unpopular with overloaded administrators of overloaded networks.

    Watch out for multi-server broadcasts and systems with multiple ports
    onto the same Ethernet; there is some code to protect against this,
    but it is still easy to get confused.

    Don't put the lock file onto an automounted partition or delete it by
    hand, unless you really want to start two daemons at the same time.
    Both will probably fail horribly if you do this.

    The daemon save file is checked fairly carefully, but should be in a
    reasonably safe directory, unless you want hackers to cause trouble.
    /tmp is safe enough on most systems, but not all - /etc is better.


Installing and Using the Program
--------------------------------

Start by copying the executable and man page to where you want them.  If you
want only to display the time and as a replacement for the rdate or date
commands, the installation is finished!

You can use it as a simple unprivileged command to check the time, quite
independently of whether it is running as a time-updating daemon or server, or
whether you are running ntp.  You can run it in daemon mode without updating
the clock, to check for drift, but it may fail if the clock is changed under
its feet.  Unfortunately, you cannot listen to broadcasts without privilege.

If it is used with the -a option to keep the time synchronised, it is best to
run it as one of root's cron jobs - for many systems, running it once a day
should be adequate, but it will depend on the reliability of the local clock.
The author runs it this way with -a and -x - see below.

If it is used with the -r option to set the time (instead of the rdate or date
commands), it should be used interactively and either on a lightly loaded
system or at a high priority.  You should then check the result by running it
in display mode.

You are advised NOT to run it with the -r option in a cron job, though this is
not locked out.  If you have to (for example under HP-UX before 10.0), be sure
to run it as the highest priority that will not cause other system problems and
set the maximum automatic change to as low a value as you can get away with.

WARNING: adjtime is more than a bit sick on many systems, and will ignore large
corrections, usually without any form of hint that it has done so.  It is often
(even usually) necessary to reset the clock to approximately the right time
using the -r option before using the -a and -x options to keep it correct.

It can be started as a time-updating daemon with the -a and -x options (or -r
and -x if you must), and will perform some limited drift correction.  In this
case, start it from any suitable system initialisation script and leave it
running.  Note that it will stop if it thinks that the time difference or drift
has got out of control, and you will need to reset the time and restart it by
hand.

In daemon mode, it will survive its time server or network disappearing for a
while, but will eventually fail, and will fail immediately if the network call
returns an unexpected error.  If this is a problem, you can start it (say,
hourly or nightly) from cron, and it will fail if it is already running
(provided that you haven't disabled or deleted the lock file).

If it is used as a server, it should be started from any suitable system
initialisation script, just like any other daemon.  It must be started after
the networking, of course.  To run it in both server modes, start one copy with 
the -B option and one with the -S option.


Simple Examples of Use
----------------------

Many people use it solely to check the time of their system, especially as a
cross-check on ntpd.  You do not need privilege and it will not cause trouble
to the local network, so you can use it on someone else's system!  You can
specify one server or several.  For example:

    msntp ntp.server.local ntp.server.neighbour

You can use it to check how your system is drifting, but it isn't very good at
this if the system is drifting very badly (in which case use the previous
technique and dc) or if you are running ntp.  You do not need privilege and it
will not cause trouble to the local network.  For example:

    sntp -x 120 -f /tmp/msntp.state ntp.server.local

More generally, it is used to synchronise the clock, in which case you DO need
root privilege.  It can be used in many ways, but the author favours running it
in daemon mode, started from a cron job, which will restart after power cuts
with no attention, and send a mail message (if cron is configured to do that)
when it fails badly.  For example, the author uses a root crontab entry on one
system of:

    15 0 * * * /bin/nice --10 /usr/local/bin/sntp -a -x 480 ntp.server.local

If you have a home computer, it can be set up to resynchronise each time you
dial up.  For example, the author uses a /etc/ppp/ip-up.d/sntp file on his
home Linux system of:

    #!/bin/sh
    sleep 60
    /bin/nice --10 /usr/local/sbin/sntp -r -P 60 ntp.server.local

-a would be better, but adjtime is broken in Linux.


Debugging or Hacking the Program
--------------------------------

Almost everybody who does this is likely to need to modify only the system
interfaces.  While they are messy, they are pretty simple and have a simple
specification.  This is documented in comments in the source.  This is
described above.

The main program SHOULD need no attention, though it may need the odd tweak to
bypass compiler problems - please report these, if you encounter any.  If
something looks odd while it is running, start by setting the -v option (lower
case), as for investigating network problems, and checking any diagnostics that
appear.  Note that most of it can be checked in display mode without harming
your system.

The client will sometimes give up, complaining about inconsistent timestamps or
similar.  This can be caused by the server being rebooted and similar glitches
to the time - unfortunately, there is no reliable way to tell an ignorable
fluctuation from a server up the spout.  If this happens annoyingly often,
the -V option may help tie down the problem.  In actual use, it is simplest
just to restart the client in a cron job!

If it needs more than this, then you will need to debug the source seriously.
Start by putting an icepack on your head and pouring yourself a large whisky!
While it is commented, it is not well commented, and much of the code interacts
in complex and horrible ways.  This isn't so much because it lacks 'structure'
as because one part needs to make assumptions about the numerical properties of
another.

The -W option (upper case) will print out a complete trace of everything it
does, and this should be enough to tie down the problem.  It does distort the
timing a bit, but not usually too badly.  However, wading through that amount
of gibberish (let alone looking at the source) is not a pleasant task.  If you
are pretty sure that you have a bug, you may tell the author, and he may ask
for a copy of the output - but he will reply rudely if you send thousands of
lines of tracing to him by Email!

Note that there are a fair number of circumstances where its error recovery
could be better, but is left as it is to keep the code simple.  Most of these
should be pretty rare.


Changes in Version 1.2
----------------------

The main change was the addition of the daemon mode for drift correction (i.e.
the -x option).  The daemon code is complex and has a lot of special-casing for
strange circumstances, not all of which are testable in practice.

A lot of the code was reordered while doing this.  The output was slightly
different - considerably different with -V.

The error estimation for broadcasts was modified, and should bear more relation
to reality.  It remains a guess, as there is no way to get decent error error
estimates under such circumstances.

The -B option is now in minutes, and has a different permissible range and
default value.

The argument consistency checking for broadcasts was tightened up a bit, and a
few other internal checks added.  These should not affect any reasonable
requirement.

A couple of new functions were added to the portability base, but they don't
use any non-standard new facilities.  However, the specification of the
functions has changed slightly.


Changes in Version 1.3
----------------------

The main change was the addition of the restarting facility for daemon mode
(i.e. the -f option), which is pretty straightforward.

There were also a lot of minor changes to the paranoia code in daemon mode, to
try to separate out the case of a demented server from network and other
'ignorable' problems.  These are not entirely successful.


Changes in Version 1.4 and 1.5
------------------------------

There turned out to be a couple of places where the author misunderstood the
specification of NTP, which affect only its use in server mode.  The main
change is to use stratum 15 instead of stratum 0.

And there were some more relaxations of the paranoia code, to allow for more
erratic servers, plus a kludge to improve restarting in daemon mode after a
period of down time has unsynchronised the clock.  There is also an
incompatible change to the debugging options to add a new level - the old -V
option is now -W, and -V is an intermediate one for debugging daemon mode - but
they are both hacker's facilities, and not for normal use.

Version 1.5 adds some very minor fixes.


Changes in Version 1.6
----------------------

The first change is support for multiple server addresses - it uses these in a
round-robin fashion.  This may be useful when you have access to several
servers, all of which are a bit iffy.  This means that the restart file format
is incompatible with msntp 1.5.

It has also been modified to reset itself automatically after detecting an
inconsistency in its server's timestamps, because the author got sick of the
failures.  It writes a comment to syslog (uniquely) in such cases.

The ability to query a daemon save file was added.

Related to the above, the -E argument has been redefined to mean an error bound
on various internal times (which is what it had become, anyway) and a -P option
introduced to be what the -E argument was documented to be.

The lock and save file handling have been changed to allow defaults to be set
at installation time, and to be overridable at run-time.  To disable these
at either stage, simply set the file names to the null string.

And there have been the usual changes for portability, as standards have been
modified and/or introduced.


Future Versions
---------------

There are unlikely to be any, except probably one to fix bugs in version 1.6.

I attempted to put support for intermittent connexions (e.g. dial-up) into the
daemon mode, but doing so needs so much code reorganisation that it isn't worth
it.  What needs doing for that is to separate the socket handling from the
timekeeping, so that they can be run asynchronously (either in separate
processes or threads), and to look up a network name and open a socket only
when prodded (and to close it immediately thereafter).  So just running it
with the -r option is the current best solution.

I also attempted to put support for the "Unix 2000" interfaces into the code.
Ha, ha.  Not merely do very few systems define socklen_t (needed for IPv6
support), but "Unix 2000" neither addresses the leap second problem nor even
provides an adjtime replacement!  Some function like the latter is critical,
not so much because of the gradual change, but because of its atomicity;
without it, msntp really needs to be made non-interruptible, and that brings in
a ghastly number of system-dependencies.

Realistically, it needs a complete rewrite before adding any more function.
And, worse, the Unix 'standards' need fixing, too.



Miscellaneous
-------------

Thanks are due to Douglas M. Wells of Connection Technologies for helping the
author with several IP-related conventions, to Sam Nelson of Stirling
University for testing it on some very strange systems, and to David Mills for
clarifying what the NTP specification really is.

Thanks are also due to several other people with locating bugs, finding
appropriate options for the Makefile and passing on extension code and
suggestions.  As I am sure to leave someone out, I shall not name anyone else.

Version 1.0 - October 1996.
Version 1.1 - November 1996 - mainly portability improvements.
Version 1.2 - January 1997 - mainly drift handling, but much reorganisation.
Version 1.3 - February 1997 - daemon save file, and some robustness changes.
Version 1.4 - May 1997 - relatively minor fixes, more diagnostic levels etc.
Version 1.5 - December 1997 - some very minor fixes
Version 1.6 - October 2000 - quite a few miscellaneous changes


Nick Maclaren,
University of Cambridge Computer Laboratory,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email:  nmm1@cam.ac.uk
Tel.:  +44 1223 334761    Fax:  +44 1223 334679