Re: [Linuxptp-users] Polling ptp clock not allowed?

[Keller, J. E <jac...@in...>]

On Fri, 2016-02-12 at 15:03 +0100, Richard Cochran wrote:
> On Fri, Feb 12, 2016 at 12:09:25PM +0100, frank wrote:
> > During printf debugging I have seen large jumps in the timespecs
> > returned by clock_gettime().
> > 
> > Using a Intel i5 with a  "Intel Corporation Ethernet Connection
> > I217-V
> > (rev 05)"
> > network card around 100-200 calls to clock_gettime() in a loop are
> > enough.
> 
> This uses the e1000e driver, right?  That driver has a work around
> for
> a HW bug when reading the time, but it is restricted to 82574/82583
> devices:
> 
> drivers/net/ethernet/intel/e1000e/netdev.c:e1000e_cyclecounter_read()
> :
> 		/* errata for 82574/82583 possible bad bits read from
> SYSTIMH/L
> 		 * check to see that the time is incrementing at a
> reasonable
> 		 * rate and is a multiple of incvalue
> 		 */
> 
> Maybe the i217 is also affected?  You could try enabling this work
> around for i217 as well.
> 
> >     struct timespec oldts;
> >     struct timespec newts;
> >     clock_gettime(clkid, &oldts);
> > 
> >     for(;;)
> >     {
> >        clock_gettime(clkid, &newts);
> >        time_t delta = newts.tv_sec - oldts.tv_sec;
> >        oldts = newts;
> >        assert(delta >= 0);
> >        assert(delta <= 1);
> 
> Here it would help to see the values of oldts and newts when the
> assertion fails.
> 
> >     }
> > }
> 
> This sounds like the same old hw bug.  I would ask intel and/or on
> the
> e1000 list.
> 
> Thanks,
> Richard

It is possible this is the same bug. Let me know your results, I can
forward it to the team responsible.

Regards,
Jake