Re: [RFC][PATCH] Catch xtime_nsec underflows and fix them

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Wed, 2008-12-03 at 04:48 +0100, Roman Zippel wrote:
> Hi,
> 
> On Wed, 3 Dec 2008, Zhang, Yanmin wrote:
> 
> > > This doesn't explain the problem entirely, I considered a negative 
> > > xtime_nsec before, but xtime_nsec+offset should still be positive
> > xtime_nsec underflows after clocksource_adjust. Before clocksource_adjust,
> > xtime_nsec is a small positive.
> > 
> > When xtime_nsec underflows at the first time, xtime.tv_nsec becomes -1.
> > Later on when the second tick arrives, below statement in the while loop
> >  clock->xtime_nsec += clock->xtime_interval;
> > will cause clock->xtime_nsec becomes positive again. So the second tick
> > appears a going-backward time.
> 
> Yes, but only by 1nsec, so normally it wouldn't be noticable.
Not 1nsec. At the second tick, go back about 4294967296 nsec (2^32), about
4 seconds. That matches the output of testing process.

> 
> > >  and 
> > > produce the correct result, at least I can't find anything in 
> > > getnstimeofday().
> > The testing uses vsyscall to get call gettimeofday. vsyscall_gtod_data.wall_time_nsec
> > is a u32 while timespec->tv_nsec is a signed long.
> 
> Ok, I was missing this part, I looked at the 32bit version of 
> getnstimeofday() and there xtime.tv_nsec was correctly sign extended.
> To be safe for the future wall_time_nsec should also be a s32.