We document here that users can kill an operation running on a shard by sending a killOp command to the mongos with an opId of the format "shardName:opIdOnShard". But, this format doesn't include any information about what node in the shard it's running on and the mongos unconditionally forwards it to the primary.

We had a report of a related issue in SERVER-8878 and now include information in currentOp about the host in a shard a command is running on, but didn't appear to hook it up to KillOp in any meaningful way or provide a way for a user to kill a job running on a secondary through mongos.