[time-nuts] Re: Network interface cards that support timestamping

[Hal Murray]

John Miller said:
> I imagine most of us here use traditional RS232 serial ports to get this
> signal into an x86 computer, usually using the DCD pin. I've found some
> implementations of RS232 on PCs don't implement all of the additional
> signaling pins, so sometimes DCD is flat out missing. In other cases -
> especially with smaller embedded-type boards only UARTs with Rx/Tx are
> available to the user. (I'm fiddling with using GPIO pins on a few such
> systems, but it's quite a struggle.) So, there is a need for getting PPS
> signals into systems. 

There is PPS support on GPIO pins on Raspberry Pi.
There are a couple of HOWTOs with details for use with NTP.

Quick start NTP on the Raspberry Pi
  https://www.satsignal.eu/ntp/Raspberry-Pi-quickstart.html

The Raspberry Pi as a Stratum-1 NTP Server
  https://www.satsignal.eu/ntp/Raspberry-Pi-NTP.html

Stratum-1-Microserver HOWTO
  https://www.ntpsec.org/white-papers/stratum-1-microserver-howto/

On linux, you can get at the PPS time stamps with
   $ cat /sys/devices/virtual/pps/pps0/assert
   1675144263.000009495#1771119
   $
The stuff before the # is the time stamp, seconds and nanoseconds.  After the 
# is a counter.
Use clear rather than assert for the falling edge.
ntpd turns off the edge it isn't using.
   
---------

I haven't worked with PTP.  Fiddling bits in packets while they are in flight 
breaks security.

One problem with doing timing in the NIC is that it gets you the time in the 
NIC.  If you are interested in NTP, you probably want the time on the main CPU.

I've daydreamed about an FPGA with a clock input from your best house clock.  
The idea is that you program it to keep POSIX time: seconds and nanoseconds 
and set things up so that  users can map it into memory read-only.  All you 
have to do to get the time is read a 64 bit register from a device.  Extra 
credit if you have PPS inputs that capture a time stamp etc.

Anybody know how long it takes to read a device register?

Modern read-the-clock on Linux doesn't do a system call.  The kernel keeps info in shared memory (man vdso) so user code can get the time from the TSC.  This PC is 3.6GHz.  Here is a histogram of reading the clock a million times in a loop, warm cache and such:
Histogram: CLOCK_REALTIME, 1 ns per bucket, 1000000 samples.
        ns      hits
        10      6012
        11    152279
        12    459189
        13    212678
        14    137852
        15     29083
        16       401
        17       538
        18       714
        19       441
813 samples were bigger than 19.
Slowest was 69632 ns



-- 
These are my opinions.  I hate spam.