systemd-networkd [DHCPServer] should send requested options also in DHCPOFFER

[mwetterw]

systemd version the issue has been seen with

253

Used distribution

Arch Linux

Linux kernel version used

6.2.10-arch1-1

CPU architectures issue was seen on

x86_64

Component

systemd-networkd

Expected behaviour you didn't see

According to RFC2131, when a DHCP server is answering to a DHCPDISCOVER with a DHCPOFFER, it should:

[...] The
server MUST return to the client:

o The client's network address, as determined by the rules given
earlier in this section,

o The expiration time for the client's lease, as determined by the
rules given earlier in this section,

o Parameters requested by the client, according to the following
rules:

    -- IF the server has been explicitly configured with a default
       value for the parameter, the server MUST include that value
       in an appropriate option in the 'option' field, ELSE

So if a client requested DNS option in the DHCPDISCOVER for example, we should see this in the DHCPOFFER if systemd-networkd has a specific value to give for that parameter.

Unexpected behaviour you saw

There is a discrepancy between options provided in the DHCPOFFER and DHCPACK.
The DHCPACK lists much more options than its DHCPOFFER counterpart:
https://github.com/systemd/systemd/blob/main/src/libsystemd-network/sd-dhcp-server.c#L622-L630

So for example, Linux kernel PNP DHCP (cmdline with ip=dhcp) fails to get the correct DNS server.
This is because in Linux implementation, they only consider the DNS Server of the DHCPOFFER.
If the DHCPOFFER contains no DNS option, Linux will fall back to choosing the same IP as the DHCP Server.
https://github.com/torvalds/linux/blob/v6.2/net/ipv4/ipconfig.c#L1161-L1162

Steps to reproduce the problem

Configure systemd-networkd to be a DHCPServer and to distribute a DNS server different from the server itself.

[Match]
Name=eth0

[Link]
RequiredForOnline=no
ActivationPolicy=manual

[Network]
Address=10.10.10.1/24
DHCPServer=yes

[DHCPServer]
EmitRouter=yes
EmitNTP=yes
NTP=192.168.0.27
EmitDNS=yes
DNS=192.168.0.27

[DHCPServerStaticLease]
MACAddress=aa:aa:aa:aa:aa:aa
Address=10.10.10.251

Use a DHCP client which for at least one option, only cares if it's in the DHCPOFFER.
For example, the Linux kernel built-in DHCP client during boot:

CONFIG_IP_PNP=y
CONFIG_IP_PNP_DHCP=y

And with ip=dhcp in the command line.

For DNS option, it will only care if it's present in the DHCPOFFER. If it was not present, it will set it to be equal to the IP of the DHCP Server, even if the value comes later in the DHCPACK.

Additional program output to the terminal or log subsystem illustrating the issue

# tcpdump -vvvnei eth0 port 68
tcpdump: listening on eth0, link-type EN10MB (Ethernet), snapshot length 262144 bytes
23:55:26.948929 aa:aa:aa:aa:aa:aa > ff:ff:ff:ff:ff:ff, ethertype IPv4 (0x0800), length 590: (tos 0x0, ttl 64, id 0, offset 0, flags [DF], proto UDP (17), length 576)
    0.0.0.0.68 > 255.255.255.255.67: [no cksum] BOOTP/DHCP, Request from aa:aa:aa:aa:aa:aa, length 548, xid 0xaaaaaaaa, secs 2, Flags [none] (0x0000)
	  Client-Ethernet-Address aa:aa:aa:aa:aa:aa
	  Vendor-rfc1048 Extensions
	    Magic Cookie 0x22222222
	    DHCP-Message (53), length 1: Discover
	    Parameter-Request (55), length 9: 
	      Subnet-Mask (1), Default-Gateway (3), Domain-Name-Server (6), Hostname (12)
	      Domain-Name (15), RP (17), MTU (26), YD (40)
	      NTP (42)
	    END (255), length 0
	    PAD (0), length 0, occurs 293
23:55:26.949070 bb:bb:bb:bb:bb:bb > aa:aa:aa:aa:aa:aa, ethertype IPv4 (0x0800), length 310: (tos 0xc0, ttl 64, id 0, offset 0, flags [none], proto UDP (17), length 296)
    10.10.10.1.67 > 10.10.10.251.68: [udp sum ok] BOOTP/DHCP, Reply, length 268, xid 0xaaaaaaaa, Flags [none] (0x0000)
	  Your-IP 10.10.10.251
	  Client-Ethernet-Address aa:aa:aa:aa:aa:aa
	  Vendor-rfc1048 Extensions
	    Magic Cookie 0x22222222
	    DHCP-Message (53), length 1: Offer
	    Lease-Time (51), length 4: 3600
	    Subnet-Mask (1), length 4: 255.255.255.0
	    Default-Gateway (3), length 4: 10.10.10.1
	    Server-ID (54), length 4: 10.10.10.1
	    END (255), length 0
23:55:26.968926 aa:aa:aa:aa:aa:aa > ff:ff:ff:ff:ff:ff, ethertype IPv4 (0x0800), length 590: (tos 0x0, ttl 64, id 0, offset 0, flags [DF], proto UDP (17), length 576)
    0.0.0.0.68 > 255.255.255.255.67: [no cksum] BOOTP/DHCP, Request from aa:aa:aa:aa:aa:aa, length 548, xid 0xaaaaaaaa, secs 2, Flags [none] (0x0000)
	  Client-Ethernet-Address aa:aa:aa:aa:aa:aa
	  Vendor-rfc1048 Extensions
	    Magic Cookie 0x22222222
	    DHCP-Message (53), length 1: Request
	    Server-ID (54), length 4: 10.10.10.1
	    Requested-IP (50), length 4: 10.10.10.251
	    Parameter-Request (55), length 9: 
	      Subnet-Mask (1), Default-Gateway (3), Domain-Name-Server (6), Hostname (12)
	      Domain-Name (15), RP (17), MTU (26), YD (40)
	      NTP (42)
	    END (255), length 0
	    PAD (0), length 0, occurs 281
23:55:26.969054 bb:bb:bb:bb:bb:bb > aa:aa:aa:aa:aa:aa, ethertype IPv4 (0x0800), length 336: (tos 0xc0, ttl 64, id 0, offset 0, flags [none], proto UDP (17), length 322)
    10.10.10.1.67 > 10.10.10.251.68: [udp sum ok] BOOTP/DHCP, Reply, length 294, xid 0xaaaaaaaa, Flags [none] (0x0000)
	  Your-IP 10.10.10.251
	  Client-Ethernet-Address aa:aa:aa:aa:aa:aa
	  Vendor-rfc1048 Extensions
	    Magic Cookie 0x22222222
	    DHCP-Message (53), length 1: ACK
	    Lease-Time (51), length 4: 3600
	    Subnet-Mask (1), length 4: 255.255.255.0
	    Default-Gateway (3), length 4: 10.10.10.1
	    Domain-Name-Server (6), length 4: 192.168.0.27
	    NTP (42), length 4: 192.168.0.27
	    TZ-Name (101), length 12: "Europe/Paris"
	    Server-ID (54), length 4: 10.10.10.1
	    END (255), length 0


Linux kernel gets the wrong DNS (falls back to 10.10.10.1 instead of using 192.168.0.27).
But it gets the correct NTP server though.

[    4.808940] IP-Config: Got DHCP answer from 10.10.10.1, my address is 10.10.10.251
[    4.816578] IP-Config: Complete:
[    4.819830]      device=eth0, hwaddr=aa:aa:aa:aa:aa:aa, ipaddr=10.10.10.251, mask=255.255.255.0, gw=10.10.10.1
[    4.829887]      host=10.10.10.251, domain=, nis-domain=(none)
[    4.835781]      bootserver=0.0.0.0, rootserver=0.0.0.0, rootpath=
[    4.835802]      nameserver0=10.10.10.1
[    4.846152]      ntpserver0=192.168.0.27
[    4.922446] uart-pl011 20201000.serial: no DMA platform data
[    5.017119] Freeing unused kernel image (initmem) memory: 29696K
[    5.023568] Run /init as init process

[mwetterw]

(It's possible Linux has an issue by only looking at the DHCPOFFER for the DNS, but I think systemd-networkd also has an issue here, and then the 2 issues get intertwined together and create a "protocol misunderstanding".)

[mwetterw]

I've changed the behavior on Linux side in net-next.
https://git.kernel.org/netdev/net-next/c/81ac2722fa19

The patch should probably land on Linux 6.5-rc1.

[yuwata]

Fix is waiting in #27607.