Re: [ANNOUNCE] ktimers subsystem

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, 2005-09-19 at 15:03 -0700, Christoph Lameter wrote:
> On Mon, 19 Sep 2005 tglx@linutronix.de wrote:
> 
> > sources. Another astonishing implementation detail of the current time 
> > keeping is the fact that we get the monotonic clock (defined by POSIX as 
> > a continous clock source which can not be set) by subtracting a variable 
> > offset from the real time clock, which can be set by the user and 
> > corrected by NTP or other mechanisms.
> 
> The benefit or drawback of that implementation depends which time is more 
> important: realtime or monotonic time. I think the most used time value is 
> realtime and not monotonic time. Having the real time value in xtime 
> saves one addition when retrieving realtime. 

Thats only partially true. 

Granted, the most used time in user space is clock_realtime
(gettimeofday() / clock_gettime(CLOCK_REALTIME). 

But do we really want to discuss one add instruction ?

The most used time in kernel space is clock_monotonic. 
Thats partially a result of the rather odd POSIX specs regarding
relative CLOCK_REALTIME timers. 

Also the basic prerequisite for for high resolution timers is a fast and
simple access to clock_monotonic rather than to a backward corrected
clock_realtime representation. 

Kernel code speed in hot pathes must have precedence over code executed
on behalf of userspace if its not completely out of bounds. One add/sub
is definitely not.

We should rather ask glibc people why gettimeofday() / clock_getttime()
is called inside the library code all over the place for non obvious
reasons.

tglx