Being a dork and a lawyer, I actually looked up the DHCP protocol. This is the relevant part:

"3.7 When clients should use DHCP

   A client SHOULD use DHCP to reacquire or verify its IP address and
   network parameters whenever the local network parameters may have
   changed; e.g., at system boot time or after a disconnection from the
   local network, as the local network configuration may change without
   the client's or user's knowledge.

   If a client has knowledge of a previous network address and is unable
   to contact a local DHCP server, the client may continue to use the
   previous network address until the lease for that address expires.
   If the lease expires before the client can contact a DHCP server, the
   client must immediately discontinue use of the previous network
   address and may inform local users of the problem."

So no there is no indication in the protocol that the clients should assume that the DHCP server will keep its promises after a disconnection from the network. Furthermore, if you use common sense there is no reason to make this assumption in most real life situations. In a coffee house somebody can always trip over the router's cable, for example.

It says "SHOULD" not "MUST", so the client can assume the server will keep its promises, it's just not recommended.

    "SHOULD   This word, or the adjective "RECOMMENDED", mean that there
    may exist valid reasons in particular circumstances to ignore a
    particular item, but the full implications must be understood and
    carefully weighed before choosing a different course."

http://www.ietf.org/rfc/rfc2119.txt

Yes, I am aware the RFC used should and not must, but nevertheless their recommendation is clear.

And of course if that item is to be ignored "the full implications must be understood and carefully weighed before choosing a different course." Here the implications are that the device may cause other devices to drop their connections, which is a pretty severe implication, imo.

Note that it doesn't say that such verification has to be done prior to recommencing use of the leased address - just that it should be done. Arguably then the described implementation follows the letter of the recommendation. Your witness, counsel.

The small print in paragraph 2 explicitly allows reusing the previous address until the lease expires.

Paragraph 2 only allows this "If a client ... is unable to contact a local DHCP server." Note that in the present example, the client started using its old address before even trying to contact the DHCP server and eventually was able to contact the DHCP server.

This could alternately be described as "a client used its old address until it could contact the DHCP server".

So here's my understanding of the situation: when a DHCP server runs out of leases to assign, it will kill the oldest (in terms of use) lease. If the device that held the lease is not on, for relevant values of "on", at the time, it will not be notified about this.

The usual (perhaps standard, I'm not sure) process has the device confirming its DHCP lease when coming back on, and in this situation the device would be notified that it can't have the old lease as it's been repurposed, and the server will provide another lease (perhaps after killing some yet another lease).

If this is an Apple device and it behaves in the manner described, after coming back "on" it will not consult DHCP, but rather it will reuse the lease it held previously. If the server has given that lease to another device, the Apple device will butt in, causing an IP conflict. Perversely, the Apple device will shortly discover this, actually do a DHCP request, and switch over with no indication to the user, leaving the other device to wonder why it had an IP conflict and how to handle it.

Is this incorrect in any way?

As far as I can see from the RFC and from Googling, there is no mechanism for a DHCP server to take back an IP address before the lease has expired, unless the client explicitly initiates relinquishing the lease.

Once the address is given out, it is the client's for the term of the lease.

Given that most routers are managing a class C network, and even then have allocated the first 100 addresses for internal use (starting the DHCP pool at .100), this would mean that it would be trivial to DoS any DHCP server in seconds by simply cycling MAC addresses and leasing all of their addresses.

Obviously, this is not realistic, and no one would ever implement a DHCP server that allowed this to happen: it is going to keep track of some fixed number of leases (probably/hopefully as many IP addresses as it is managing) and, when it runs out, it will be forced to reclaim an address from another user.

If anything, the fact that such a mechanism is not specified is simply going to make this "worse", as the specific implementation of how and when the reclamation happens is going to end up as "whatever makes sense to the guy working on that specific DHCP server implementation", and may constitute seemingly random behavior.

You don't even need to cycle MAC addresses; just send DHCPDECLINE messages for any address the server gives. DHCP requires that any address that elicits that response from a client MUST be marked unavailable, and provides no provision for it ever becoming available again.

Ignoring the client is the only automatic mechanism DHCP provides for countering this.

You have to remember that DHCP was written back when it was assumed that a competent systems administrator managed the server and could manually respond to running out of available addresses (which was considered a likely sign of misconfiguration), and didn't attempt to specify what to do in the case where it happens.

Not giving out an address and notifying the systems administrator is what the DHCP spec recommends in this case. In fact, section 2.2 claims "The allocation mechanism (the collection of DHCP servers) guarantees not to reallocate that address within the requested time" so any reclamation of IP address on the part of the server is technically in violation of RFC 2131.

I'm confused... are you therefore suggesting that this is how people like Linksys should handle this problem? I fully realize that the spec was written in another time, but I don't think tzs did: he is actually citing the spec as a reason why these leases should stay around. When you go to a coffee shop, the router does not and probably should not play a chime when it has run out of DHCP leases, causing the barista to go around the room finding laptops that have been setup incorrectly... the router should, and almost certainly does, just "figure something out", likely by reclaiming addresses.

There are sensible options available other than violating the leases.

For example, for a highly-used access point, it would make sense to give out only very short leases - say, 10 minutes. That'd support a churn of 25 clients per minute if you're giving out addresses from a /24 (and there's no reason why you couldn't use a smaller netmask, down to /8). It'd also be sensible to ignore a single MAC address once it had more than a handful of outstanding leases.

It doesn't help with a malicious client - but then neither does breaking leases. The malicious client can cause enough address flapping to make the network completely unuseable. You really can't protect a wifi network against a malicious client trying to DoS it - there's a myriad ways to do that apart from DHCP.

It's trivial to DoS any layer 2 network. This is trivial netsec.

There is a fundamental difference between sitting on the network flooding it (an activity that requires you to actually be connected to the network, although this could of course be a small and difficult to detect device), and sending 256 packets and walking away, content in the knowledge that for the next 24 hours no one will be able to use the network because you own all of the DHCP leases. "This is trivial netsec."

I never said flood, I said DoS. I don't think this is on-topic so I'm not going to go in to detail but you should Google ARP spoofing. These tactics are quite trivial and I'm teaching undergrads to do it next year.

I was, and continue to be, under the impression that this kind of attack is not possible when an access point has node isolation (a very common feature that even my MiFi router supports, despite it often being turned off in home and even office environments), yet the aforementioned protocol-level DHCP DoS I described would work on any device that implemented the DHCP specification. I therefore fail to see how ARP spoofing is relevant. Remember: we are discussing whether it is a fair belief that any reasonable implementation of DHCP would have implemented the spec as stated.

That's where I'm at with this. What Jarek describes can happen, but that's a failure case, not the normal-operation case. And as portrayed in the article, the engineers at Apple have accounted for the failure case, so all's well. They simply have made the very reasonable assumption that the DHCP server will normally be operational, not in a crisis state, and that it will honor its contracts as specified in the RFC. How hard is this?

There is that principle of "conservative in what you send, liberal in what you accept" on which UNIX and half of the internet was built.

I think your description is correct.

The problem is the impact of the false "ip conflict" that is caused by this optimization. A device that detects an ip conflict should not have to guess whether it is an ip conflict caused by an optimization on another device. As pointed out in other comments an ip conflict can cause other systems to drop existing connections.

No one likes ip conflict notifications, or to have their connections dropped. Play nice. That's all.

The problem is that the DHCP server re-issued a valid lease to a second computer. Surely this is an issue with the DHCP server not playing nice, rather than Apple?

Do you feel there is a better, more tolerant, and less error prone course of action the DHCP server could take if it runs out of leases?

If we are able to modify the DHCP server to do something sensible, why not modify default router settings to issue /16 addresses instead of /24? Or even 10/8? There's plenty of RFC1918 space to not run out.

NAT state tables won't have to stretch that far as they can time out unused connections just like they do already.

Would there be any unintended side-effects with doing this?

If you don't like the NAT, go ahead and issue IPv6 :-)

Not really. The DHCP server may have many valid reasons for reissuing a technically valid lease, such as it being restarted and losing its table of valid leases. As I pointed out in a response above, the DHCP protocol makes it clear that devices should re-verify their addresses after waking up.

This optimization does not cause IP address conflicts. If a DHCP server is not correctly keeping track of lease times and ensuring that addresses already assigned are not reused while the client still has a valid lease, you can get IP conflicts regardless of whether or not their are any Apple devices on the network.

Please name one or more DHCP server which will knowingly violate a lease it has granted, via "kill the oldest" or any other means.