atom feed1 message in org.kernel.vger.linux-manCLOCK_MONOTONIC_RAW
FromSent OnAttachments
Erik ThieleAug 14, 2009 8:44 am 
From:Erik Thiele (
Date:Aug 14, 2009 8:44:43 am


clock_gettime talks about CLOCK_MONOTONIC_RAW clock_getres talks about CLOCK_MONOTONIC_RAW clock_nanosleep does NOT talk about CLOCK_MONOTONIC_RAW timerfd_create does NOT talk about CLOCK_MONOTONIC_RAW there seem to be other functions relevant to CLOCK_* selection that don't specify all

the ones that do not talk about the CLOCK_MONOTONIC_RAW should say why it is not possible to use that parameter or should specify that it is indeed possible to use that parameter.

maybe there should instead be a manpage "linux_clocks" that explains all clocks instead of replicating the information on the individual manpages.

manpage of clock_nanosleep should say that using CLOCK_MONOTONIC_RAW is the only possibility on linux to really wait a specific amount of time without being in danger of ntp fiddling around because CLOCK_MONOTONIC still leaves us with the problem of time slew generated by ntp.

the manpage could also say that CLOCK_MONOTONIC_RAW is not always "better" than CLOCK_MONOTONIC since it is on the one hand free of slewing, but on the other hand it has the problem that it might run too fast or too slow depending on hardware properties. at least it should run at constant speed, which CLOCK_MONOTONIC with ntp does certainly not.

maybe you should make a table to explain that:

CLOCK_REALTIME - can jump - can slew - if ntp is running this clock is always kept close to GMT. even if hardware is not 100% correct, ntp will correct everything over time.

CLOCK_MONOTONIC - cannot jump - can slew !!! (because of ntp) - it is not kept in sync with GMT. but the "speed" of seconds is kept in sync with GMT by varying it constantly by ntp.

CLOCK_MONOTONIC_RAW - cannot jump - cannot slew ! - the speed of seconds is not the same as the speed of GMT seconds since the hardware timer is never 100% exact and ntp daemon does NOT have influence here

in my program i have a main scheduler event loop which needs to do absolute timer events with nanosecond precision and independent of ntp time slewing and jumping.

- to get nanosecond precision i need to use a dual threaded approach where one thread has the only duty to wait with clock_nanosleep and report timer events over a pipe() to the mother thread since epoll_wait is only with millisecond resolution

- to get the independence of ntp time jumping i use CLOCK_MONOTONIC

- to get the independence of ntp time slewing i'd like to use CLOCK_MONOTONIC_RAW but it's at least not documented?

thanks! cya