Re: [RFC][PATCH] new timeofday core subsystem (v.A0)

From: George Anzinger (george_at_mvista.com)
Date: 09/09/04

  • Next message: James Cleverdon: "[PATCH][2.6.8.1-mm4] phys_proc_id change for x86-64" 
    Date:	Wed, 08 Sep 2004 17:08:55 -0700
To: john stultz <johnstul@us.ibm.com>

    john stultz wrote:
    > On Fri, 2004-09-03 at 17:02, George Anzinger wrote:
    >
    >>>Again, monotonic_clock() and friends are NTP adjusted, so drift caused
    >>>by inaccurate calibration shouldn't be a problem the interval timer code
    >>>should need to worry about (outside of maybe adjusting its interval time
    >>>if its always arriving late/early). If possible the timesource
    >>>calibration code should be improved, but that's icing on the cake and
    >>>isn't critical.
    >>>
    >>
    >>Are you providing a way to predict what clock count provide a given time offset
    >>INCLUDING ntp? If so, cool. If not we need to get this conversion right. We
    >>will go into this more on your return.
    >
    >
    > Sorry, I'm not sure what you mean. Mind expanding on the idea while my
    > brain warms back up?

    The issue is this: A user wants a timer to fire at exactly time B which is
    several hours later than now (time A). We need to know how to measure this time
    with the timer resources (not the clock as you are talking about it). Currently
    we do something like delta_jiffies = timespec_to_jiffies(B - A) and set up a
    jiffies timer to expire in delta_jiffies. At this time we "assume" in
    timespec_to_jiffies() that we _know_ how long a jiffie is in terms of wall clock
    nanoseconds. We also assume (possibly incorrectly) that this number is "good
    enough" even with ntp messing things up. I think this means that we assume
    that, on the average, we have the right conversion and that any drift will be a)
    small and b) on the average 0 (or real close to it).

    To my mind this is a few too many assumptions, but this is what we do today. If
    you are proposing that ntp corrections take care of systemic constant drift
    (which by the way, linux use to do WRT HZ=1024 division errors), we will need a
    way of working this back into the timesepec_to_jiffies() code or we will not be
    able to do timers with any real accuracy.

    An alternative, and one that I prefer, but, I think, runs counter to your
    proposal, is to derive the timer interrupts from the clock after the ntp
    correction. Some hardware will have a problem with this (x86) but other
    hardware is rather constrained to do things this way. The PPC counter interrupt
    generator, for example, counts the same pulses as the clock counter.
    > 

    -- 
George Anzinger   george@mvista.com
High-res-timers:  http://sourceforge.net/projects/high-res-timers/
Preemption patch: http://www.kernel.org/pub/linux/kernel/people/rml
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
  • Next message: James Cleverdon: "[PATCH][2.6.8.1-mm4] phys_proc_id change for x86-64"

    Relevant Pages

    • Re: [PATCH/RFC 10/10] example of simple continuous gettimeofday
      ... Uses NTP to modify the clock frequency value rather then just ... > adjustment, which should be applied every f cycles. ... Your design seems to suggest keeping an NTP calculated reference time ...
      (Linux-Kernel)
    • Re: ntp discipline of local time?
      ... Principles and Precision Time Synchronization briefings on the NTP project page are old but applicable. ... The details you are asking for are carefully explained in the NTPv4 spec. ... like the clock state machine and poll-adjust algorithm continue in the daemon. ... factor from CPU cycles to time) and get rid of the offset. ...
      (comp.protocols.time.ntp)
    • Re: [PATCH/RFC 10/10] example of simple continuous gettimeofday
      ... >> The basic idea behind this is that ntp defines the frequency of the ... >> adjustment, which should be applied every f cycles. ... >> which allows for a very stable clock. ... >> resolution is still well within the practical limits of the clock. ...
      (Linux-Kernel)
    • Re: ntp discipline of local time?
      ... Principles and Precision Time Synchronization briefings on the NTP ... does two things-- set the system clock frequency (ie adjust the conversion ... factor from CPU cycles to time) and get rid of the offset. ... frequency and phase errors. ...
      (comp.protocols.time.ntp)
    • Re: Unix Time and Leap Seconds
      ... That is an obviously nonsensical reading of the passage. ... out based on a less accurate time source. ... The timesource cannot be "a locally attached reference clock ... or NTP" because NTP alone *cannot* serve as a timesource. ...
      (comp.unix.programmer)