Re: [PATCH 3/4] hv_util: use do_adjtimex() to update system time

[Stephen Hemminger]

On Mon, 09 Jan 2017 19:14:30 +0100
Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote:

> Stephen Hemminger <stephen@xxxxxxxxxxxxxxxxxx> writes:
> 
> > On Mon, 09 Jan 2017 18:40:15 +0100
> > Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> wrote:
> >  
> >> Stephen Hemminger <stephen@xxxxxxxxxxxxxxxxxx> writes:
> >>   
> >> > An alternative would be for hyper-v util to provide a clocksource device and
> >> > let NTP manage the adjustment. The advantage of this would be HV util not fighting
> >> > with NTP, and using standard API's. The downside would be the complexity of configuring
> >> > NTP, and difficulty of writing a clock source pseudo device.    
> >> 
> >> Yes, I see this option. But as I wrote to John I'm afraid we'll have to
> >> come up with a custom interface from hv_util to userspace which no NTP
> >> server will want to support (because, first of all, it's not about
> >> 'network time' any more). We can write our own daemon which will read
> >> from this interface and do adjtimex but in this case I don't see much
> >> value in this data traveling from kernel to userspace and back...
> >>   
> >
> > Master NTP servers are connected to authoritative clock sources. I have no idea how
> > that is configured, but it should be possible.  
> 
> As far as I understand these servers are connected to GPS receivers or
> something like that and we can probably pretend being one but I'm not
> sure that the TimeSync v4 protocol fits there, we'll probably lose the
> precision - currently we calculate the delta in a small kernel function
> with interrupts disabled, in my tests I see it floating around several
> hundred - few thousand nanoseconds from host's time.

My understanding is that NTP doesn't work very well at small time intervals.
Probably the ideal solution is something where kernel corrects time but
there is also way to communicate to NTP server that the kernel time is being maintained by
other entitity.