[openib-general] IBV_WC_REM_INV_REQ_ERR and IBV_WC_WR_FLUSH_ERR on rdma write

[Roland Dreier]

    Galen> After posting an RDMA write I then poll the completion
    Galen> queue, at which time I get an IBV_WC_REM_INV_REQ_ERR error
    Galen> followed by a bunch of IBV_WC_WR_FLUSH_ERR errors.

    Galen> The peer is seeing a bunch of IBV_WC_WR_FLUSH_ERR errors on
    Galen> the polling of the completion queue.

According to the IBA spec, a remote invalid request error is generated
when the responder detects an invalid message.  Possible causes are
"operation is not supported by this receive queue, insufficient
buffering to receive a new RDMA or Atomic Operation request...."

Have you set IBV_ACCESS_REMOTE_WRITE and/or IBV_ACCESS_REMOTE_READ in
your qp_attr.qp_access_flags?  How about max_dest_rd_atomic (number of
RDMAs outstanding as a responder) and max_rd_atomic (number of RDMAs
outstanding as an initiator)?

Flush errors will be generated for all remaining work requests after a
queue pair transitions to the error state.

 - R.