Re: do_posix_clock_monotonic_gettime() returns negative nsec

[George Anzinger <george@mvista.com>]

Herbert Poetzl wrote:
> On Thu, Dec 02, 2004 at 07:08:23PM -0800, Andrew Morton wrote:
> 
>>Herbert Poetzl <herbert@13thfloor.at> wrote:
>>
>>>
>>>Hi Folks!
>>>
>>>recent kernels (tested 2.6.10-rc2 and 2.6.10-rc2-bk15)
>>>produce funny output in /proc/uptime like this:
>>>
>>>	# cat /proc/uptime
>>>	  12.4294967218 9.05
>>>	# cat /proc/uptime
>>>	  13.4294967251 10.33
>>>	# cat /proc/uptime
>>>	  14.4294967295 11.73
>>>
>>>a short investigation of the issue, ended at
>>>do_posix_clock_monotonic_gettime() which can (and 
>>>often does) return negative nsec values (within
>>>one second), so while the actual 'time' returned
>>>is correct, some parts of the kernel assume that
>>>those part is within the range (0 - NSEC_PER_SEC)
>>>
>>>        len = sprintf(page,"%lu.%02lu %lu.%02lu\n",
>>>                        (unsigned long) uptime.tv_sec,
>>>                        (uptime.tv_nsec / (NSEC_PER_SEC / 100)),
>>>
>>>as the function itself corrects overflows, it would
>>>make sense to me to correct underflows too, for 
>>>example with the following patch:
>>>
>>>--- ./kernel/posix-timers.c.orig	2004-11-19 21:11:05.000000000 +0100
>>>+++ ./kernel/posix-timers.c	2004-12-03 02:23:56.000000000 +0100
>>>@@ -1208,7 +1208,10 @@ int do_posix_clock_monotonic_gettime(str
>>> 	tp->tv_sec += wall_to_mono.tv_sec;
>>> 	tp->tv_nsec += wall_to_mono.tv_nsec;
>>> 
>>>-	if ((tp->tv_nsec - NSEC_PER_SEC) > 0) {
>>>+	if (tp->tv_nsec < 0) {
>>>+		tp->tv_nsec += NSEC_PER_SEC;
>>>+		tp->tv_sec--;
>>>+	} else if ((tp->tv_nsec - NSEC_PER_SEC) > 0) {
>>> 		tp->tv_nsec -= NSEC_PER_SEC;
>>> 		tp->tv_sec++;
>>> 	}
>>
>>Doesn't this imply that do_posix_clock_monotonic_gettime_parts() is
>>returning a negative tv_nsec?
> 
> 
> nope, not necessarily, because after that ...
> 
>         tp->tv_sec += wall_to_mono.tv_sec;
>         tp->tv_nsec += wall_to_mono.tv_nsec;
> 
> might add a negative value, which explains the
> underflow ...
> 
> and if you look closer:
> 
> 	xtime.tv_sec = get_cmos_time();
>         wall_to_monotonic.tv_sec = -xtime.tv_sec;
>         xtime.tv_nsec = (INITIAL_JIFFIES % HZ) * (NSEC_PER_SEC / HZ);
>         wall_to_monotonic.tv_nsec = -xtime.tv_nsec;

Yep, that IS the problem.  It should be normalized here,  I.e.
	set_normalized_timespec(wall_to_monotonic,
		wall_to_monotonic.tv_sec - xtime.tv_sec,
		wall_to_monotonic.tv_nsec - xtime.tv_nsec);
with the obvious delets :)

Still, this should be corrected by the first settimeofday, which most systems do 
on the way up, or is that just those who use NTP?


-- 
George Anzinger   george@mvista.com
High-res-timers:  http://sourceforge.net/projects/high-res-timers/