In message: <Pine.LNX.4.64.0601011839140.6056_at_digraph.polyomino.org.uk>
"Joseph S. Myers" <jsm_at_POLYOMINO.ORG.UK> writes:
: The Linux kernel (with NTP synchronisation) duly syslogged
:
: Dec 31 23:59:59 digraph kernel: Clock: inserting leap second 23:59:60 UTC
:
: and Markus's program showed a transition from 1136073600.005623 to
: 1136073599.013130, i.e. 23:59:59 was repeated (whereas a POSIX clock in
: perfect synchronisation with UTC should appear to repeat 00:00:00, and a
: kernel source code comment
:
: /* The timer interpolator will make time change gradually instead
: * of an immediate jump by one second.
: */
:
: describes what would practically be a safer approach, if the code followed
: the comment).
ntp synced computers repeat the last second of the day. POSIX is just
flat wrong here. Actually, POSIX is completely silent on what is to
happen at the leap second to system time. Once can infer what might
be the right behavior by working backwards from what mktime does for
23:59:60, but that's a weak inference for what the 'right' posix
behavior is for the system time accross a leap second. POSIX simply
does not have leap seconds in any meaningful way. It is as if they do
not exist at all. There's no clarifications to the POSIX standard
that say that 0:00:00 should be repeated at a leap second, at least
that I've been able to find. If there is such an explicit
clarification, I'd like to know about it.
FreeBSD does the same thing (as do all kernels that use the ntp kernel
stuff supplied with udel ntpd and successors). FreeBSD doesn't
actually log the leap second via syslog, but maybe it should.
Ideally, we'd switch to running computers with TAI and doing all the
conversion on input/output of time. However, *THAT*'s definitely not
POSIX.
Warner
Received on Sun Jan 01 2006 - 11:33:47 PST