ndhc/README

ndhc, Copyright (C) 2004-2015 Nicholas J. Kain.
See LICENSE for licensing information.  In short: Two-clause / New BSD.

Requirements:

Linux kernel
GNU Make (tested: 3.82) or CMake (tested: 2.8)
Ragel (tested: 6.7)

INTRODUCTION
------------

ndhc is a multi-process, privilege-separated dhcp client.  Each subprocess runs
with the minimal necessary privileges in order to perform its task.  Currently,
ndhc consists of three subprocesses: the ndhc-master, ndhc-ifch, and
ndhc-sockd.

ndhc-master communicates with dhcp servers and handles the vagaries of the dhcp
client protocol.  It runs as a non-root user inside a chroot.  ndhc runs as a
normal user with no special privileges and is restricted to a chroot that
contains nothing more than a domain socket filesystem object (if using syslog),
a urandom device node, and a null device node.

ndhc-ifch handles interface change requests.  It listens on a unix socket for
such requests.  ndhc-ifch runs as a non-root user inside a chroot, and retains
only the power to configure network interfaces.  ndhc-ifch automatically forks
from ndhc-master to perform its job.

ndhc-sockd plays a similar role to ndhc-ifch, but it instead has the ability to
bind to a low port, the ability to open a raw socket, and the ability to
communicate on broadcast channels.  ndhc communicates with ndhc-sockd
over a unix socket, and the file descriptors that ndhc-sockd creates are
passed back to ndhc over the unix socket.

ndhc fully implements RFC5227's address conflict detection and defense.  Great
care is taken to ensure that address conflicts will be detected, and ndhc also
has extensive support for address defense.  Care is taken to prevent
unintentional ARP flooding under any circumstance.

ndhc also monitors hardware link status via netlink events and reacts
appropriately when interface carrier status changes or an interface is
explicitly deconfigured.  This functionality can be useful on wired networks
when transient carrier downtimes occur (or cables are changed), but it is
particularly useful on wireless networks.

RFC3927's IPv4 Link Local Addressing is not supported.  I have found v4 LLAs
to be more of an annoyance than a help.  v6 LLAs work much better in practice.

FEATURES
--------

Privilege-separated.  ndhc does not run as root after initial startup, and
capabilities are divided between the subprocesses.  All processes run in a
chroot.

Robust.  ndhc performs no runtime heap allocations -- malloc() (more
specifically, brk(), mmap(), etc) is never called after initialization (libc
behavior during initialization time will vary), and ndhc never performs
recursive calls and only stack-allocates fixed-length types, so stack depth is
bounded, too.

Active defense of IP address and IP collision avoidance.  ndhc fully implements
RFC5227.  It is capable of both a normal level of tenacity in defense, where
it will eventually back off and request a new lease if a peer won't relent
in the case of a conflict, and of relentlessly defending a lease forever.  In
either mode, it rate-limits defense messages, so it can't be tricked into
flooding by a hostile peer or DHCP server, either.

Small.  Both ndhc avoids unnecessary outside dependencies and is written in
plain C.

Fast.  ndhc filters input using the BPF/LPF mechanism so that uninteresting
packets are dropped by the operating system before ndhc even sees the data.
ndhc also only listens to DHCP traffic when it's necessary.

Flexible.  ndhc can request particular IPs, send user-specified client IDs,
write a file that contains the current lease IP, write PID files, etc.

Self-contained.  ndhc does not exec other processes, or rely on the shell.
Further, ndhc relies on no external libraries aside from the system libc.

Aware of the hardware link status.  If you disconnect an interface on which
ndhc is providing dhcp service, it will be aware.  When the link status
returns, ndhc will fingerprint the reconnected network and make sure that it
corresponds to the one on which it has a lease.  If the new network is
different, it will forget about the old lease and request a new one.

USAGE
-----

1) Compile and install ndhc.
    a) make
    b) Install the build/ndhc executable in a normal place.  I would suggest
       /usr/sbin or /usr/local/sbin.

1alt) Compile and install ndhc.
    a) Create a build directory:
       mkdir build && cd build
    b) Create the makefiles:
       cmake ..
    c) Build ndhc:
       make
    d) Install the ndhc/ndhc executable in a normal place.  I would suggest
       /usr/sbin or /usr/local/sbin.

2) Time to create the jail in which ndhc will run.
    a) Become root and create new group "ndhc".

        $ su -
        # umask 077
        # groupadd ndhc

    b) Create new users "dhcpifch" and "dhcp".  The primary group of these
       users should be "ndhc".

        # useradd -d /var/lib/ndhc -s /sbin/nologin -g ndhc dhcpsockd
        # useradd -d /var/lib/ndhc -s /sbin/nologin -g ndhc dhcpifch
        # useradd -d /var/lib/ndhc -s /sbin/nologin -g ndhc dhcp

    c) Create the state directory where DUIDs and IAIDs will be stored.

        # mkdir /etc/ndhc
        # chown root.root /etc/ndhc
        # chmod 0755 /etc/ndhc

    d) Create the jail directory and set its ownership properly.

        # mkdir /var/lib/ndhc
        # chown root.root /var/lib/ndhc
        # chmod a+rx /var/lib/ndhc
        # cd /var/lib/ndhc
        # mkdir var
        # mkdir var/state
        # mkdir var/run
        # chown -R dhcp.ndhc var
        # chmod -R a+rx var
        # chmod g+w var/run

    e) Create a urandom device for ndhc to use within the jail.

        # mkdir dev
        # mknod dev/urandom c 1 9
        # mknod dev/null c 1 3
        # chown -R root.root dev
        # chmod a+rx dev
        # chmod a+r dev/urandom
        # chmod a+rw dev/null

        f) (optional) If you wish for logging to properly work, you
           will need to properly configure your logging daemon so that it
           opens a domain socket in the proper location within the jail.
           Since this varies per-daemon, I cannot provide a general
           configuration.

3) At this point the jail is usable; ndhc is ready to be used.  An example
   of invoking ndhc:

    # ndhc -i wan0 -u dhcp -U dhcpifch -D dhcpsockd -C /var/lib/ndhc

4o) If you encounter problems, I suggest running ndhc in the foreground and
    examining the printed output.

5) I suggest running ndhc under some sort of process supervision.  This will
   allow for reliable functioning in the case of unforseen or
   unrecoverable errors.


BEHAVIOR NOTES
--------------

ndhc does not enable updates of the local hostname and resolv.conf by default.
If you wish to enable these functions, use the --resolve (-R) and --hostname
(-H) flags.  See ndhc --help.

STATE STORAGE NOTES
-------------------

ndhc requires a read/writable directory to store the DUID/IAID states.  By
default this directory is /etc/ndhc.  It exists outside the chroot.  The DUID
will be stored in a single file, DUID.  The IAIDs exist per-interface and are
stored in files with names similar to IAID-xx:xx:xx:xx:xx:xx, where the xx
values are replaced by the Ethernet hardware address of the interface.

If it is impossible to read or store the DUIDs or IAIDs, ndhc will
fail at start time before it performs any network activity or forks any
subprocesses.

If the host system lacks volatile storage, then a clientid should manually
be specified using the -I or --clientid command arguments.

RANDOMNESS NOTES
----------------

Each ndhc subprocess maintains a PCG PRNG that is uniquely seeded from the
kernel random device at startup.  Each PRNG consumes 128 bits of entropy for
its initial state.

DHCP does not require cryptographic randomness, so this arrangement should
be more than sufficient to ensure proper UUIDs, assuming only that the
kernel random device is even minimally seeded with real entropy.

PORTING NOTES
-------------

DHCP clients aren't naturally very portable.  It's necessary to perform a lot
of tasks that are platform-specific.  ndhc is rather platform-dependent, and it
extensively uses Linux-specific features.  Some of these features are also
available on the BSDs.

1) ndhc takes advantage of Linux capabilities so that it does not need full
root privileges.  Capabilities were a proposed POSIX feature that was not made
part of the official standard, so any implemention that may exist will be
system-dependent.

2) ndhc configures network interfaces and routes.  Interface and route
configuration is entirely non-portable.

3) ndhc uses netlink sockets extensively for fetching data, setting data,
and hardware link state change notification events.

4) ndhc uses the Berkeley Packet Filter / Linux Packet Filter interfaces to
drop unwanted packets in kernelspace.  This functionality is available on
most modern unix systems, but it is not standard.

5) ndhc uses epoll() and signalfd().  These are Linux-specific.

6) Numerous socket options are used, and the AF_PACKET socket family is used
for raw sockets and ARP.  These are largely Linux-specific, too.

7) ndhc can optionally use seccomp-filter to allow only a set of whitelisted
syscalls.  This functionality is Linux-specific.

HISTORY
-------

I started writing ndhc back in 2004.  My ISP at the time required a dhcp
client for connection authentication, and I was not comfortable with any
of the existing clients, which all ran as root and had colorful security
histories.  DHCP is generally not a routed protocol, and lacks real
authentication mechanisms in real world deployments (some largely
abandoned RFCs for such behavior do exist), so no program existed to
fill the niche of a truly secure DHCP client.

My router/server at the time ran a custom Linux distro that was designed
for extreme security.  A root privileged DHCP client would be nearly the
only root-owned process running on the machine, so I was highly motivated
to develop an alternative.

A separate ifchd was first written entirely from scratch.  It did not take long
to write, since it was by design rather simple, and I was already familiar with
the quirks of Linux capabilities.  That left me with the choice of adapting an
existing DHCP client or writing my own from scratch.

At the time, I just wanted something that would work, so my choice was to
adapt udhcpc to work with ifchd.  udhcpc was chosen since it was intended to
be used with resource-constrained or embedded systems, and was thus very
small.  ISC dhclient was another alternative, but it is an extremely large
program, and it would have been very hard to audit it for correctness.

udhcpc was not did not really fit my requirements well, since it was designed
to be small at all costs, sacrificing correctness when necessary.  The code was
hard to follow, and had many quirks.  Bounds-checking was rare, type aliasing
common, and state transitions were convoluted.  Not all of the client was
asynchronous, and no precautions were taken against conflicting peers.  ARP was
not used at all.

However, it was small.  With a lot of work, I ripped out the script-calling
mechanisms and replaced them with ifchd requests.  Bounds-checking was
aggressively (and somewhat hamfistedly) retrofitted into the code.  It was
cleaned to a degree, and importantly it worked for connecting to my ISP.

Then I changed ISPs.  My new ISP used PPPoE, not dhcp.  Around the same time, I
also switched to using Gentoo rather than a hand-built distribution.  I didn't
have time to maintain the old custom setup, and it was very hard keeping up
with library vulnerabilties in eg, zlib or openssl, and ensuring that all
installed binaries, dynamic and static, were updated.  ndhc was abandoned for
many years.  It wasn't needed on my server, and it was "too much effort" to
deviate from the stock distro dhcp clients on other machines.

Then, around 2008, I changed ISPs again.  This time my new ISP used dhcp and
not PPPoE.  So, after a few months, I decided to dust off the old ndhc/ifchd
project and adapt it to my modern standards and machines.

ifchd was in good shape and required little work.  I ended up rewriting
ndhc.  The only parts that remained from the original were the parts that
I had already rewritten before, and some of those were rewritten, too.

Eventually ifchd was rewritten to extensively use a Ragel-generated DFA-based
parser to make it easier to verify correct behavior for all possible inputs.

Quite a while later, I eventually merged ifchd into the same binary as ndhc
and instead rely on forking subprocesses and using socketpairs for IPC.  This
brought a lot of simplifications, particularly for user configuration.

Afterwards, privilege separation was applied to the remaining capabilities,
creating the ndhc-sockd subprocess.  After this change, the main ndhc
process runs completely unprivileged.

The end result is a modern DHCP client is largely RFC-compliant, except where
the RFCs dictate behavior that would be problematic, overly complex, useless,
or exploitable.  DHCP is poorly specified, and real-world servers and clients
vary a lot from the RFCs, so these conditions are necessary for a useful
program.

Although ndhc's implementation and behavior are different, I have to credit
the idea of using netlink events to discover hardware link status transitions
to Stefan Rompf and his 'dhcpclient' program.  The Linux netlink events that
are used are otherwise rather obscure and poorly documented, and I wouldn't
have known about them otherwise.

GRSECURITY NOTES
----------------

Make sure that CONFIG_GRKERNSEC_CHROOT_CAPS is disabled.  Otherwise, ndhc will
lose its capabilities (in particular, the ability to reconfigure interfaces)
when it chroots.

DHCP PROTOCOL QUIRKS
--------------------

Send a packet that has an options field set to:
'DHCP-OPTION-OVERLOAD:3'
Then in the file and sname fields:
'DHCP-OPTION-OVERLOAD:3'
I suspect some bad dhcp programs will hang given this input.

DHCP explicitly specifies that there is no minimum lease time and also
specifies that the minimum default rebinding time is leasetime*0.875 and
the minimum default renewing time is leasetime*0.500.  All times are relative
to the instant when the lease is bound and are specified in seconds.  Taken
together, this means that a client strictly implementing the RFC should
accept a lease that either is perpetually rebinding (lease == 1s) or instantly
expires (lease == 0s).  ndhc ignores the RFC and specifies a minimum lease
time of one minute.

Renew and rebind times are optionally specified and may take on any value.
This means that a malicious server could demand a rebind time before a renew
time, or make these times ridiculously short, or specify both times past
that of the lease duration.  ndhc avoids all of this nonsense by simply
ignoring these options and using the default values specified by the RFC.

There are other quirks, but these are just several interesting ones that
immediately occur to me while I'm writing this document.