Meanwhile several people have tested my PPS extension to the Linux
time routines with success. As Linus indicated there should be some
documentation in addition to the code changes.
I put some usr/src/linux/Documentation/kernel-time.txt (8kB) on the tiny
FTP server mentioned below. The document just describes what all is
about (at least I tried).
Also I have done some tests of my own with a GNUplotted graph
(PostScript) that can be inspected from file PPSoffset.ps.gz (45kB).
I used a comination of synchronizing the kernel clock to the CMOS
clock's periodic interrupts, and then synchronize to PPS pulses
delivered by my DCF77 receiver. Interestingly the receiver failed
about in the middle of the graph, and you can see the consequences.
(Note: The precision of my DCF77 receiver is not good enough for PPS,
but it works a little bit!)
My dumb test program, rtctest3.c, is also on the server.
Before I'll continue work I'd like to get the current patch
integrated into the current kernels and architectures, because up to
tnow it does not break anything. (I'm planning to rename struct timex
that the kernel uses...)
P.S. in libc 5.2.18 (I think) the __ntp_gettime function calls
adjtimex with uninitialized data. If you use that function, be
prepared for surprise! (libc/sysdeps/linux/__ntpgttm.c).
P.P.S. For my Pentium-100 I had between 157 and 1748 (sorry!) clock
cycles for the PPS interrupt routine. That shouldn't be a problem for
a 1-pulse-per-second-interrupt, but if you put some white noise on
that pin...
Ulrich Windl