Simon McVittie reported in Debian bug #654876 that his laptop running Linux 3.0 or 3.1 would sometimes crash (panic) while idle. He initially suspected a driver bug, but the screen did not show any information about where the original fault occurred. However, using netconsole, he was able to capture a full log of the crash. This showed that a packet received through the wireless interface was being processed by IGMP, which then divided by zero.
IGMP is part of the IPv4 protocol suite, supporting multicast routing. Every multicast address corresponds to a dynamic set of hosts, called a multicast group. Multicast routers can send query messages asking which hosts belong to which groups, and hosts using multicast report back at intervals. Routers can then limit forwarding of multicast packets to the interfaces where the group has members. More sophisticated switches can also snoop IGMP and use it to limit their multicast forwarding. There are unfortunately three different versions with semi-compatible message formats. In version 1, the maximum reporting interval (Max Response Time) is fixed as 10 seconds, but from version 2 it is specified in query messages.
The Linux IGMP implementation supports all three versions. It distinguishes query messages as specified in RFC 3376 section 7.1: v3 messages are longer than v1 or v2; v2 messages have a non-zero Max Response Time whereas v1 messages always have zero. It is possible to force use of a specific protocol version, but normally if there are multiple multicast routers using different protocol versions it will respond according to the earliest protocol version in use so that all routers can understand its responses.
Linux 2.6.36 included two fixes to the version selection logic. Unfortunately, the second of these introduced the bug in question. While v2 query messages cannot possibly have zero Max Response Time (as that would make them v1), v3 query messages can. What this means is unspecified, but the Linux IGMP code previously treated it as the minimum valid value of 1/10 second. But in the case where a v3 query is received and a v2 query has also recently been received, this is no longer done. This results in a reporting interval of 0 seconds and a division by zero when deciding the initial random delay.
The fix for this is pretty obvious. It is included in Linux 3.0.17, 3.1.9, 3.2.1, and the Debian package version 3.1.8-2.
This is easily exploitable for denial of service within a local network. Linux does not check the destination address of IGMP queries, so it may also be possible to attack a target through its unicast address from several hops away. The attacker only needs to send a single IGMPv2 query followed by a single IGMPv3 query with zero Max Response Time. All systems running Linux 2.6.36 or later (up to the above fixed versions) with any active IPv4 multicast listeners (other than for the 'all hosts' address) are vulnerable.