-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
The Saturday 2007-12-08 at 01:05 +0100, Carlos E. R. wrote:
> Not yet. I think I will compile the kernel tomorrow: it takes about
> three hours in this machine.
So, what happened Carlos. Any change?
Not yet, I couldn't... however, I have been running tests. I wrote a
cronjob to compare the system clock to the cmos clock every 5 minutes, and
the difference varies between 0 and 1 seconds for several hours. Then I
disconnected the router, to force ntpd to be unable to connect to its
peers... and not 1" error! No drift! I'm astonished. I have an email
pending with some logs of this, I'll send it promptly.
The only thing I did was to force the kernel to use the 'tsc' clock
instead of the 'acpi_pm' clocksource it was using:
I left the computer on all night, with no network, and... no drift! The
error this morning was, surprise:
8 Dec 11:31:06 ntpd[26893]: offset 0.011621 sec freq -98.861 ppm error
0.002463 poll 7
8 Dec 11:51:44 ntpd[26893]: synchronized to 193.62..., stratum 1
8 Dec 11:51:49 ntpd[26893]: time reset +0.252153 s
Less than half a second of error, with no network! This is indeed
surprising.
It seems that my computer works better with the 'tsc' clocksource than
with the 'acpi_pm' one. Now the problem is to know why 'tsc' was rejected
on boot or later.
Googling for 'acpi_pm' I find references to problems similar to mine on
other distros, so there is definitely a problem somewhere in the kernel.
The question on what are each clock type is also asked with no answer.
The 'acpi_pm.c' source says:
* Driver to use the Power Management Timer (PMTMR) available in some
* southbridges as primary timing source for the Linux kernel.
So that's not the typical cpu, IRQ based clock!
The 'tsc.c' source says nothing:
* This code largely moved from arch/i386/kernel/timer/timer_tsc.c
* which was originally moved from arch/i386/kernel/time.c.
* See comments there for proper credits.
And I have no 'arch/i386/kernel/timer/' dir at all, not
'arch/i386/kernel/time.c' file. Good grief. Free code may be good and all
that, but things are difficult to find out, unless you are an insider!
[...]
For instance, in <http://marc.info/?l=linux-kernel&m=117812987231998&w=2>
(march 2007) they say the clock is lossing time when the computer is
iddling - which matches my findings, too, even though mine is a desktop,
i686, 32 bits.
- --
Cheers,
Carlos E. R.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.4-svn0 (GNU/Linux)
iD8DBQFHWoL1tTMYHG2NR9URArBaAJ4kcaM7fCb7LpE8OlUvCE0hv/lmagCfS+r0
JkCxmhH8V6rKcF19wvIFndo=
=OFzo
-----END PGP SIGNATURE-----
--
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]