[networkd] ConfigureWithoutCarrier makes the DHCP server fail (-EBUSY) when carrier gained

[hvenev]

I've tested this with systemd 238, but it should be the same in master.

This happens when the .network file specifies ConfigureWithoutCarrier=yes:

[Match]
Name=eth0

[Network]
Address=10.123.42.1/24
ConfigureWithoutCarrier=yes
DHCPServer=yes

When ConfigureWithoutCarrier, addresses are set (link_enter_set_addresses). This also starts the DHCP server.

Then, the carrier appears. link_carrier_gained sees that the link is in one of the following states:

• LINK_STATE_SETTING_ADDRESSES if the kernel is very slow to set them
• LINK_STATE_SETTING_ROUTES if link-local addresses still haven't been set
• LINK_STATE_CONFIGURED if everything has been set properly and a few seconds have passed

link_carrier_gained calls link_acquire_conf (again) and link_enter_set_addresses (again). The problem is that link_enter_set_addresses tries to start the already running DHCP server, which fails with -EBUSY:

Failed to read $container of PID 1, ignoring: Permission denied
Found container virtualization none.
Failed to open configuration file '/etc/systemd/networkd.conf': Success
timestamp of '/etc/systemd/network' changed
Ignoring /etc/systemd/network/10-eth0.link, because it's not a regular file with suffix .netdev.
Ignoring /etc/systemd/network/10-eth0.network, because it's not a regular file with suffix .netdev.
Ignoring /usr/lib/systemd/network/80-container-host0.network, because it's not a regular file with suffix .netdev.
Ignoring /usr/lib/systemd/network/99-default.link, because it's not a regular file with suffix .netdev.
Ignoring /usr/lib/systemd/network/80-container-ve.network, because it's not a regular file with suffix .netdev.
Ignoring /usr/lib/systemd/network/80-container-vz.network, because it's not a regular file with suffix .netdev.
Ignoring /etc/systemd/network/10-eth0.link, because it's not a regular file with suffix .network.
Ignoring /usr/lib/systemd/network/99-default.link, because it's not a regular file with suffix .network.
eth0: Flags change: +UP +MULTICAST +BROADCAST
eth0: Link 4 added
eth0: udev initialized link
eth0: Saved original MTU: 1500
eth0: Adding address: fe80::dead:beff:feef:0001/64 (valid forever)
eth0: Gained IPv6LL
eth0: Adding address: 10.123.42.1/24 (valid forever)
rtnl: received address with invalid family 129, ignoring
rtnl: received address with invalid family 128, ignoring
Enumeration completed
sd-netlink: ignored message with unknown type: 2
sd-netlink: ignored message with unknown type: 2
sd-netlink: ignored message with unknown type: 2
sd-netlink: ignored message with unknown type: 2
Successfully acquired requested service name.
eth0: Link state is up-to-date
eth0: found matching network '/etc/systemd/network/10-eth0.network'
LLDP: Started LLDP client
eth0: Started LLDP.
eth0: Discovering IPv6 routers
NDISC: Started IPv6 Router Solicitation client
eth0: Not emitting DNS server information on link, couldn't find suitable uplink.
eth0: Not emitting NTP server information on link, couldn't find suitable uplink.
DHCP SERVER: STARTED
eth0: Offering DHCPv4 leases
eth0: Setting addresses
NDISC: Sent Router Solicitation, next solicitation in 3s
eth0: Removing address: 10.123.42.1/24 (valid forever)
eth0: Adding address: 10.123.42.1/24 (valid forever)
eth0: Addresses set
NDISC: Sent Router Solicitation, next solicitation in 7s
NDISC: Sent Router Solicitation, next solicitation in 14s
NDISC: No RA received before link confirmation timeout
NDISC: Invoking callback for 't'.
eth0: Configured

Carrier appears

eth0: Flags change: +LOWER_UP +RUNNING
eth0: Gained carrier
eth0: Discovering IPv6 routers
eth0: Not emitting DNS server information on link, couldn't find suitable uplink.
eth0: Not emitting NTP server information on link, couldn't find suitable uplink.
Assertion '!server->receive_message' failed at ../src/libsystemd-network/sd-dhcp-server.c:1026, function sd_dhcp_server_start(). Ignoring.
eth0: Could not start DHCPv4 server instance: Device or resource busy
eth0: Failed
NDISC: Stopping IPv6 Router Solicitation client

I haven't tested the behaviour of the DHCP client, but I'm not sure if it will behave well because link_acquire_conf is also called twice.

[poettering]

cc @ssahani

[ssahani]

@poettering we can skip start of the DHCP server till the carrier gain. I guess that would be the right approach.