[rds-devel] how can RDMA ops kill a QP?

[Zach Brown]

So, rds-stress -I 2 reliably kills a QP.  -I is the super secret undocumented flag which builds rdma ops from vectors of memory instead of from one buffer.

The rds-stress -r side always sees a send complete with an error:

[ 6764.919515] RDS/IB: send completion on 10.3.7.17 had status 10, disconnecting and reconnecting

The RDMA additions were after my time.  Before I go debugging this from first principles, can someone give a quick rundown of what we might be messing up in an RDMA op to kill a QP?  Are there places in particular that I should check for the details of what went wrong?

- z