Re: time / gtod seconds value out of sync?

[Nishanth Aravamudan <nacc@linux.vnet.ibm.com>]

Hi John!

On 19.02.2015 [11:03:26 -0800], John Stultz wrote:
> Hey Nish! Long time!

yep :)

> On Thu, Feb 19, 2015 at 10:35 AM, Nishanth Aravamudan
> <nacc@linux.vnet.ibm.com> wrote:
> > Hi John,
> >
> > We're seeing an interesting issue with the openposix testcase
> > difftime/1-1, which basically calls gtod/time, sleeps, calls time/gtod,
> 
> I'm not familiar with the test... Do you have a link?

Sorry about that:

http://sourceforge.net/projects/posixtest/files/posixtest/posixtestsuite-1.5.2/posixtestsuite-1.5.2.tar.gz/download

conformance/interfaces/difftime/1-1.c

It takes quite a few iterations in a loop, but it eventually fails with
mainline on both x86-64 and ppc64le.

> > then difftime and sees if they disagree. The issue occurs with either
> > vDSO implementations or direct syscalls.
> >
> > We are seeing failures on ppc64le and x86_64 (probably other places too,
> > just not tested yet), because (I'm pretty sure), the time() syscalls
> > granularity is not accounting for the nsecs value at all. Instead, it
> > just returns get_seconds().
> 
> 
> Right, so there is always a problem mixing calls of different
> granularity (similar issues crop up with gettimeofday() and filesystem
> timestamps), so the basis of the test worries me a little bit from the
> description, but I'd have to look at it to really get a sense.

Yep, that makes sense to me too -- the only concern I had was that the
point where time() is returning X seconds, a simultaneous (theoretical)
call to gtod() would return the correct X+1 seconds (presuming nsecs had
exceeded 1000000000).

> > In one case, I see, in sys_time():
> >
> > [  313.001823] NACC: timekeeping_get_ns = 1000121642
> > [  314.001889] NACC: timekeeping_get_ns = 188401
> >
> > gtod correctly accumulates those nsecs into the secs value:
> >
> >                 ts->tv_sec = tk->xtime_sec;
> >                 nsecs = timekeeping_get_ns(&tk->tkr);
> >                 ts->tv_nsec = 0;
> >                 timespec64_add_ns(ts, nsecs);
> >
> > but time() does:
> >
> >                 return tk->xtime_sec;
> >
> > It seems like overkill to do the full timekeeping_get_ns() in time(),
> 
> Right, so looking into the git history,
> f20bf6125605acbbc7eb8c9420d7221c91aa83eb (time: introduce
> xtime_seconds) changed this specifically for performance reasons
> (cc'ed Ingo here, in case he remembers more context).

Ah I see. I can see the performance impact of calling into gtod being
high.

> The idea that time() would be ok as being HZ granular, and its been
> this way since 2.6.23.  Thus you have a < HZ sized window where
> gettimeofday() will return the next second before time() gets updated
> by the tick.

Right.

> > but maybe it's also necessary to account for leap seconds? That is, we
> > need to ensure that accumulate_nsecs_to_secs() has been called before
> > return tk->xtime_sec?
> 
> So leapseconds are also applied at tick time, so I don't think you'd
> see any different behavior with them.

Yep, ok.

> There was a thread on this quite awhile back and I if I recall I think
> the general consensus was to keep time() tick granular  (so it aligns
> with filesystem timestamps) and gettimeofday() hardware granular.
> Though we also introduced the CLOCK_REALTIME_COARSE to match
> sub-second filesystem timestamps as well.
> 
> So yea... I don't think we want to make a change here, but maybe I'm
> not understanding the underlying issue... so feel free to push back
> here. :)

Oh that's fine. I mostly wanted the subsystem experts to chime in on if
the the testcase was valid, etc.

Jan, do you have any other concerns?

Thanks,
Nish