Cause: When a Ceph rbd image is exported through multiple iSCSI targets, the rbd kernel module will grab the exclusive lock when it receives a write request. Consequence: If a write request is hung in the rbd module, and the iSCSI initiator and multipath layer decide to fail the request to the application, read requests sent after the failure may not reflect the state of the write. Fix: The rbd module will grab the exclusive lock for both reads and writes. This will cause hung writes to be flushed and or failed before executing reads. Result: Read requests sent after failed writes will always return the same data.