dhcpcd not passing NTP servers to systemd-timesyncd

[colinhogben]

I have a Compute Module 3 running Stretch (2018-04-18 version with latest update/upgrade applied) and trying to get time synchronised with our local NTP servers.

By default, systemd-timesyncd is trying various servers *.debian.pool.ntp.org, but these all time out due to our site firewall configuration.

The local NTP server addresses are supplied by dhcp, and I can see these written to the file /run/ntp.conf.dhcp . However, since Raspbian is using systemd-timesyncd rather than ntpd, that doesn't appear to have any effect. It looks like some glue is missing.

Googling and manpage-reading found a temporary workaround - I added a file /etc/systemd/timesyncd.conf.d/site.conf containing approximately

Code: Select all

[Time]
NTP=ntp-0.localdomain ntp-1.localdomain

However, this is only good until system administrators decide to move things or we deploy the module on a different network - which is why we need to use the server addresses from DHCP.

Does anyone know how this should be done properly? Or have I missed something?

[donbrew]

Is there some reason it can't get the time from the network, that is the default?

[colinhogben]

Is there some reason it can't get the time from the network, that is the default?

Yes. The Pi is on a network which is isolated from the outside world by a firewall which does not allow NTP traffic through. However, there are local NTP servers which are dual-hosted i.e. one interface on the internal network (to act as server to internal hosts) and another on the external network (to allow synchronisation with the rest of the world).

[DougieLawson]

Have you enabled ntp with sudo timedatectl set-ntp 1?
Have you installed ntpd with a /etc/ntp.conf update to point to your LAN server?
Change this stuff

Code: Select all

server 0.ubuntu.pool.ntp.org burst iburst maxpoll 11
server 1.ubuntu.pool.ntp.org burst iburst maxpoll 11
server 2.ubuntu.pool.ntp.org burst iburst maxpoll 11
server 3.ubuntu.pool.ntp.org burst iburst maxpoll 11

to

Code: Select all

server 192.168.xxx.yyy

where that's the address of the server on your LAN.

[ejolson]

Does anyone know how this should be done properly?

I think the idea behind including simple versions of ntpd, dhcpcd, udev and others in systemd is to speedup how fast the system reboots in the desktop usage case. For servers and situations not related to desktop use the trade-off between reliability, flexibility and reboot speed favours a different init philosophy using the original daemons to provide full featured services. If you install ntpd using

# apt-get install ntpd ntpdate

systemd should be graceful enough to allow ntpd handle the time synchronization. You may also need to remove the

timesyncd.conf.d/site.conf

file you created or manually disable the systemd time synchronization daemon. After this ntpd should work normally.

Do you usually use an includefile directive in ntp.conf to source the file with the names of the time servers provided by dhcp?

I have a similar setup with a cluster of Pi Zero computers that can't reach the Internet. In that case, however, there is no need to automatically configure ntpd using dhcp.

[colinhogben]

Have you enabled ntp with sudo timedatectl set-ntp 1?

Yes, "timedatectl status" shows that ntp is enabled. The issue is that it's not using the servers obtained from DHCP.

I think the idea behind including simple versions of ntpd, dhcpcd, udev and others in systemd is to speedup how fast the system reboots in the desktop usage case. For servers and situations not related to desktop use the trade-off between reliability, flexibility and reboot speed favours a different init philosophy using the original daemons to provide full featured services. If you install ntpd ...

Well, yes, I could go the route of switching from systemd-timesyncd to traditional ntpd and then the dhcp hook machinery would work as it used to.

Or, perhaps, go in the other direction and use systemd-networkd instead of dhcpcd. But, as you allude to, that may not give me the flexibility I need (e.g. to deal with some other DHCP options).

I was hoping/expecting that the default combination of packages and services chosen by the distro (Raspbian Lite Stretch) would be fully integrated together and work out-of-the-box - hence my original comment about some missing "glue". Maybe I should raise an issue against Raspbian.

[colinhogben]

I've created /lib/dhcpcd/dhcpcd-hooks/50-timesyncd.conf containing the following:

Code: Select all

# Set NTP servers for systemd-timesyncd

confd=/run/systemd/timesyncd.conf.d

set_servers() {
    mkdir -p "$confd"
    (
        echo "# Created by dhcpcd hook"
        echo "[Time]"
        echo "NTP=$new_ntp_servers"
    ) > "$confd/dhcp-ntp.conf"

    # Tell timesyncd it has an updated configuration
    systemctl try-reload-or-restart systemd-timesyncd
}

if $if_up; then
    set_servers
fi

which seems to work.

[ejolson]

I've created /lib/dhcpcd/dhcpcd-hooks/50-timesyncd.conf

That script looks reasonable. Thanks for posting. Do you need to check whether $new_ntp_servers is non-null before updating?