Description of problem:
When chronyd operating as an NTP client with enabled hardware timestamping receives a response from a fast and close server (e.g. an appliance with a hardware-based NTP implementation) before the hardware transmit timestamp of the request, the timestamp is not included in the calculation of the offset. If the server is so fast and close that a large number of measurements is missing the hardware timestamp, the accuracy of the system clock will degrade significantly.
The fix is to rework the NTP I/O code to wait a bit for late HW transmit timestamps before processing the fast response.
A related issue is that enabling kernel receive timestamping has a delay and it may not be enabled in time to timestamp the fast response. If the hardware receive timestamp was missing for any reason, chronyd would have to use the daemon receive timestamp, which may be very inaccurate due to the fix for the previous issue. The fix is to permanently enable the kernel receive timestamping.
Version-Release number of selected component (if applicable):
Requires special hardware
Steps to Reproduce:
1. configure chronyd with:
server $HW_NTP_SERVER minpoll 0 maxpoll 0
hwtimestamp * rxfilter none
where $HW_NTP_SERVER is a close HW-based NTP server (peer delay + response
time should be smaller than few tens of microseconds)
2. observe the last two columns in the measurements log
Some measurements (other than the first few) have kernel-kernel (K K), kernel-daemon (K D), or hardware-daemon (H D) timestamps.
All measurements except the first few have hardware-kernel (H K) timestamps.
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.