dhcpcd/BUILDING.md
Roy Marples 12cc2d9946 Write syslog entries in logfile:
date tag[pid]: message
This does add a few more bytes, but it allows for the same log
analysis tools to be used.
When building SMALL, this code is removed and you just get the
message in the logfile.
2017-04-11 18:04:03 +01:00

7.7 KiB

Building dhcpcd

This attempts to document various ways of building dhcpcd for your platform.

Building for distribution (ie making a dhcpcd source tarball) now requires gmake-4 or any BSD make.

Size is an issue

To compile small dhcpcd, maybe to be used for installation media where size is a concern, you can use the --small configure option to enable a reduced feature set within dhcpcd. Currently this just removes non important options out of dhcpcd-definitions.conf, the logfile option and support for DHCPv6 Prefix Delegation. Other features maybe dropped as and when required. dhcpcd can also be made smaller by removing the IPv4 or IPv6 stack:

  • --disable-inet
  • --disable-inet6

Or by removing the following features:

  • --disable-auth
  • --disable-arp
  • --disable-arping
  • --disable-ipv4ll
  • --disable-dhcp6

You can also move the embedded extended configuration from the dhcpcd binary to an external file (LIBEXECDIR/dhcpcd-definitions.conf)

  • --disable-embedded If dhcpcd cannot load this file at runtime, dhcpcd will work but will not be able to decode any DHCP/DHCPv6 options that are not defined by the user in /etc/dhcpcd.conf. This does not really change the total on disk size.

Cross compiling

If you're cross compiling you may need set the platform if OS is different from the host.
--target=sparc-sun-netbsd5.0

If you're building for an MMU-less system where fork() does not work, you should ./configure --disable-fork. This also puts the --no-background flag on and stops the --background flag from working.

Default directories

You can change the default dirs with these knobs. For example, to satisfy FHS compliance you would do this: ./configure --libexecdir=/lib/dhcpcd dbdir=/var/lib/dhcpcd

Compile Issues

We now default to using -std=c99. For 64-bit linux, this always works, but for 32-bit linux it requires either gnu99 or a patch to asm/types.h. Most distros patch linux headers so this should work fine. linux-2.6.24 finally ships with a working 32-bit header. If your linux headers are older, or your distro hasn't patched them you can set CSTD=gnu99 to work around this.

ArchLinux presently sanitises all kernel headers to the latest version regardless of the version for your CPU. As such, Arch presently ships a 3.12 kernel with 3.17 headers which claim that it supports temporary address management and no automatic prefix route generation, both of which are obviously false. You will have to patch support either in the kernel or out of the headers (or dhcpcd itself) to have correct operation.

OS specific issues

Some BSD systems do not allow the manipulation of automatically added subnet routes. You can find discussion here: http://mail-index.netbsd.org/tech-net/2008/12/03/msg000896.html BSD systems where this has been fixed or is known to work are: NetBSD-5.0 FreeBSD-10.0

Some BSD systems protect against IPv6 NS/NA messages by ensuring that the source address matches a prefix on the recieved by a RA message. This is an error as the correct check is for on-link prefixes as the kernel may not be handling RA itself. BSD systems where this has been fixed or is known to work are: NetBSD-7.0 OpenBSD-5.0 patch submitted against FreeBSD-10.0

Some BSD systems do not announce IPv6 address flag changes, such as IN6_IFF_TENTATIVE, IN6_IFF_DUPLICATED, etc. On these systems, dhcpcd will poll a freshly added address until either IN6_IFF_TENTATIVE is cleared or IN6_IFF_DUPLICATED is set and take action accordingly. BSD systems where this has been fixed or is known to work are: NetBSD-7.0

OpenBSD will always add it's own link-local address if no link-local address exists, because it doesn't check if the address we are adding is a link-local address or not.

Some BSD systems do not announce cached neighbour route changes based on reachability to userland. For such systems, IPv6 routers will always be assumed to be reachable until they either stop being a router or expire. BSD systems where this has been fixed or is known to work are: NetBSD-7.99.3

Linux prior to 3.17 won't allow userland to manage IPv6 temporary addresses. Either upgrade or don't allow dhcpcd to manage the RA, so don't set either ipv6ra_own or slaac private in dhcpcd.conf if you want to have working IPv6 temporary addresses. SLAAC private addresses are just as private, just stable.

Init systems

We try and detect how dhcpcd should interact with system services at runtime. If we cannot auto-detect how do to this, or it is wrong then you can change this by passing shell commands to --serviceexists, --servicecmd and optionally --servicestatus to ./configure or overriding the service variables in a hook.

/dev management

Some systems have /dev management systems and some of these like to rename interfaces. As this system would listen in the same way as dhcpcd to new interface arrivals, dhcpcd needs to listen to the /dev management sytem instead of the kernel. However, if the /dev management system breaks, stops working, or changes to a new one, dhcpcd should still try and continue to work. To facilitate this, dhcpcd allows a plugin to load to instruct dhcpcd when it can use an interface. As of the time of writing only udev support is included. You can disable this with --without-dev, or without-udev. NOTE: in Gentoo at least, sys-fs/udev as provided by systemd leaks memory sys-fs/eudev, the fork of udev does not and as such is recommended.

select

dhcpcd uses eloop.c, which is a portable main event loop with timeouts and signal handling. Unlike libevent and similar, it can be transplanted directly within the application - the only caveat outside of POSIX calls is that you provide queue.h based on a recent BSD (glibc sys/queue.h is not enough). eloop supports the following polling mechanisms, listed in order of preference: kqueue, epoll, pollts, ppoll and pselect. If signal handling is disabled (ie in RTEMS or other single process OS's) then eloop can use poll. You can decide which polling mechanism dhcpcd will select in eloop like so ./configure --with-poll=[kqueue|epoll|pselect|pollts|ppoll]

Importing into another source control system

To prepare dhcpcd for import into a platform source tree (like NetBSD) you can use the make import target to create /tmp/dhcpcd-$version and populate it with all the source files and hooks needed. In this instance, you may wish to disable some configured tests when the binary has to run on older versions which lack support, such as getline. ./configure --without-getline

Hooks

Not all the hooks in dhcpcd-hooks are installed by default. By default we install 01-test, 02-dump, 10-mtu, 20-resolv.conf and 30-hostname. The other hooks, 10-wpa_supplicant, 15-timezone and 29-lookup-hostname are installed to $(datadir)/dhcpcd/hooks by default and need to be copied to $(libexecdir)/dhcpcd-hooks for use. The configure program attempts to find hooks for systems you have installed. To add more simply ./configure -with-hook=ntp.conf

Some system services expose the name of the service we are in, by default dhcpcd will pick RC_SVCNAME from the environment. You can override this in CPPFLAGS+= -DRC_SVCNAME="YOUR_SVCNAME". This is important because dhcpcd will scrub the environment aside from $PATH before running hooks. This variable could be used to facilitate service re-entry so this chain could happen in a custom OS hook: dhcpcd service marked inactive && dhcpcd service starts dependant services are not started because dhcpcd is inactive (not stopped) dhcpcd hook tests if $if_up = true and $af_waiting is empty or unset. if true, mark the dhcpcd service as started and then start dependencies if false and the dhcpcd service was previously started, mark as inactive and stop any dependant services.