(I also have some vague memory that some bonding events mapped to this
cma event although the address was still available but that was a long
time ago so I'm not sure, and the change log says nothing about it...)

The test case where I reproduced the problem with processing RDMA_CM_EVENT_ADDR_CHANGE 
is the disconnection of the active link from bonding under load (through pull out the cable).

With my patch, the bug is not reproduced, the load is not interrupted.

Best regards,
Chesnokov Gleb