[LinuxPPS] Recommendation / Re: Stats & architectures

[James Boddington]

Bernhard Schiffner wrote:
> Am Dienstag, 18. März 2008 00:06 schrieb clemens at dwf.com:
> ...
>>> Linuxpps uses getnstimeofday() so a nanosecond time stamp is
>>> available.
>> I hate to break it to you, but if you look at the kernel code,
>> getnstimeofday is a dmy routine,- it just multiplies the us value by
>> 1000 and returns it.  It does NO interpolation.
>>
>> That DOES mean that you dont have to change code when a REAL ns clock
>> becomes available, but thats about it...
> 
> Once I used  getnstimeofday() for some debugging. The readout showed the 
> full 10^-9 resolution and no visible *1000 steps. NTPD was not active this 
> time.
> From this time on I was a beliver in increased numerical resolution 
> provided by this call.
> 
> Can you prove this on your machine please?
> (Which clocksoure is used?)
> 
> Do you see an easy way to determine getnstimeofday() "grain"?

I have not seen the email where I was quoted so I'll reply here.

I tried the following module

#include <linux/module.h>

int init_module(void)

{
	struct timespec ts;
	int loop;
	
	for (loop = 0; loop < 10; loop++) {
		getnstimeofday(&ts);
		printk("%lu %09lu\n", ts.tv_sec, ts.tv_nsec);
	}

	return 0;
}

void cleanup_module(void)

{
}

MODULE_AUTHOR("James");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("play with getnstimeofday()");

Checking dmesg afterwards I get 10 time stamps that are not us * 1000

1205832048 169178484
1205832048 169493048
1205832048 169586077
1205832048 169674356
1205832048 169760679
1205832048 169846165
1205832048 169931092
1205832048 170015460
1205832048 170100107
1205832048 170185313

Replacing the loop with

	getnstimeofday(&ts1);
	getnstimeofday(&ts2);
	getnstimeofday(&ts3);
	printk("%lu %09lu\n", ts1.tv_sec, ts1.tv_nsec);
	printk("%lu %09lu\n", ts2.tv_sec, ts2.tv_nsec);
	printk("%lu %09lu\n", ts3.tv_sec, ts3.tv_nsec);

gives me 3 time stamps that increase in less than 1us increments

1205832931 266274358
1205832931 266274637
1205832931 266275196


Looking at kernel/time/timekeeping.c

getnstimeofday() calls __get_realtime_clock_ts()
__get_realtime_clock_ts() calls __get_nsec_offset()

The comment for __get_nsec_offset() is
/**
  * __get_nsec_offset - Returns nanoseconds since last call to periodic_hook
  *
  * private function, must hold xtime_lock lock when being
  * called. Returns the number of nanoseconds since the
  * last call to update_wall_time() (adjusted by NTP scaling)
  */

  __get_nsec_offset() calculates nanoseconds

I get the same results for tsc, acpi_pm and pit but not jiffies. For jiffies I 
got the same time stamp 10 times.

Digging a little deeper I found kernel/time.c also has a getnstimeofday(). It 
depends on whether CONFIG_GENERIC_TIME is defined.

#ifdef CONFIG_GENERIC_TIME you get __get_nsec_offset() returning nanoseconds 
and getnstimeofday() returning nanoseconds.

#ifndef CONFIG_GENERIC_TIME you get __get_nsec_offset() returning 0 and 
getnstimeofday() returning us * 1000.

I have CONFIG_GENERIC_TIME=y in .config

-- 
    James