I have a Cisco 2960-X with IGMP snooping enabled (all IGMP options are default). There are two VLANs on the 2960-X. A single PC is connected to each VLAN on the Cisco 2960-X. A trunk leads from the Cisco to a non-Cisco switch on which SVIs are configured. Both SVIs have PIM/IGMP enabled and multicast routing is enabled. IGMPv3 is running on the SVIs. The two client PCs support both IGMP v2 and v3. Multicast between the two PCs works, but then temporarily breaks for 125 seconds and then starts working again "forever". (Note that 125 seconds is about the duration of one IGMP query interval.)
Packet captures helped deduce the problem. Initially, the client sends a IGMP v2 join message. Later, the router sends IGMP v3 Membership queries. The client then changes to sending IGMP v3 join messages. The Cisco switch no longer sees IGMP v2 join messages for the group. After 3 x the IGMP query interval, the Cisco switch times out the IGMP v2 snooping group entry and sends a IGMP leave message to the mrouter. (The client does not originate the IGMP leave message.) The Cisco switch removes the IGMP v2 group entry but leaves the IGMP v3 group entry. On the non-Cisco switch, the leave message causes the non-Cisco switch to remove the group from the port which breaks multicast to the client.
The question is: should the non-Cisco switch pay attention to the IGMP version of the leave request, or should the Cisco switch refrain from sending a leave message that wasn't originated by the client?
I don't see where where RFC 4541 (IGMP Snooping) mentions that a switch can create a leave message. It does say that a switch can proxy a leave message received from a client if the client is the last receiver of the group. But the client didn't send a leave message.
Shane
