Skip to contentSkip to conversation

ntp.c : possible inconsistency?

PATCH Bernhard Schiffner 2 participants 4 messages Jan 10, 2007 – Jan 15, 2007 Timekeeping, clocksource core, NTP, alarmtimer

lore
Share & tools
lore
#0 lore raw

Hi,

trying to find reasons for some crazy ntpd-behavior I read
http://lkml.org/lkml/2006/10/27/67

This thread doesn't result in a pulished patch, so I (hopefully) did what was said there. The patch doesn't break my system, but it doesn't change ntpd's crazyness too.
Nevertheless it should be discussed again in the sense of preventing an inconsistency.

Bernhard

PS:
Can someone point me to the reason for doing txc->constant + 4, please?

References:
lkml-thread (start)
http://lkml.org/lkml/2006/10/26/47

ntpd-crazyness (comparing 2.6.15 with 2.6.19)
http://ml.enneenne.com/pipermail/linuxpps/2006-December/000482.html

Patch:
index 3afeaa3..36d7ecc 100644

kernel/time/ntp.c +1 -1
@@ -263,7 +263,7 @@ int do_adjtimex(struct timex *txc)
                    result = -EINVAL;
                    goto leave;
                }
-               time_constant = min(txc->constant + 4, (long)MAXTC);
+               time_constant = min(txc->constant + 4, (long)MAXTC + 4);
            }
 
            if (txc->modes & ADJ_OFFSET) {      /* values checked earlier */
@@ -329,7 +329,7 @@ leave:      if ((time_status & (STA_UNSYNC|

STA_CLOCKERR)) != 0)

        txc->maxerror      = time_maxerror;
        txc->esterror      = time_esterror;
        txc->status        = time_status;
-       txc->constant      = time_constant;
+       txc->constant      = time_constant - 4;
        txc->precision     = 1;
        txc->tolerance     = MAXFREQ;
        txc->tick          = tick_usec;
#1 lore raw

Hi,

Quoted 8 lines · Bernhard Schiffner

trying to find reasons for some crazy ntpd-behavior I read
http://lkml.org/lkml/2006/10/27/67

This thread doesn't result in a pulished patch, so I (hopefully) did what was said there. The patch doesn't break my system, but it doesn't change ntpd's crazyness too.
Nevertheless it should be discussed again in the sense of preventing an inconsistency.

Without a further explanation of this craziness, it's a little hard to discuss...

Quoted 2 lines

PS:
Can someone point me to the reason for doing txc->constant + 4, please?

The main reason is this part in ntpd/ntp_loopfilter.c of the ntp package:

        if (pll_nano) {
                ntv.offset = (int32)(clock_offset *
                    1e9 + dtemp);
                ntv.constant = sys_poll;
        } else {
                ntv.offset = (int32)(clock_offset *
                    1e6 + dtemp);
                ntv.constant = sys_poll - 4;
        }

It uses the pll_nano switch to distinguish between the old and new ntp model. The kernel now implements the new ntp model, but the actual userspace interface hasn't changed yet, so the kernel undoes the compensation.

Here is bit more information about where this adjustments comes from:
http://www.ntp.org/ntpfaq/NTP-s-compat.htm#Q-COMPAT

bye, Roman

#2 lore raw
Quoted 2 lines

Without a further explanation of this craziness, it's a little hard to discuss...

Let's try it:
time_constant is created for internal use of ntp.c and added by 4
-               time_constant = min(txc->constant + 4, (long)MAXTC);
+               time_constant = min(txc->constant + 4, (long)MAXTC + 4);

But sometimes it is written back to data referenced from outside. So let's do the + 4 backwards ...
-       txc->constant      = time_constant;
+       txc->constant      = time_constant - 4;

otherwise I'd like to speak of an inconsistency. It doesn't depend of actual use and other topics.

Bernhard

#3 lore raw

Hi,

Quoted 6 lines · Bernhard Schiffner
Quoted 2 lines

Without a further explanation of this craziness, it's a little hard to discuss...

Let's try it:
time_constant is created for internal use of ntp.c and added by 4
-               time_constant = min(txc->constant + 4, (long)MAXTC);
+               time_constant = min(txc->constant + 4, (long)MAXTC + 4);

MAXTC is already adjusted.

Quoted 4 lines

But sometimes it is written back to data referenced from outside. So let's do the + 4 backwards ...
-       txc->constant      = time_constant;
+       txc->constant      = time_constant - 4;

ntpd doesn't read it back for it's own purposes, it only prints it, when the kernel info is queried, it doesn't adjust the constant there, so I didn't do it in the kernel either.

bye, Roman