>ok, given there is no single protocol and you define different >absolute max values (56 on Windows, 256 on Linux), how >do rdma_cm consumer's authenticate private data capabilities?
The private data differs based the underlying transport, connection message, implementation, and, possibly someday, the underlying link. The guaranteed *minimum* for any message/transport/implementation is 0. :) There is no guarantee on the maximum. For standard IB, if the RDMA IP CM service is employeed, 56 bytes are currently available in the REQ (rdma_connect() call). Other messages and functions, rdma_accept() and rdma_reject(), are higher. If the UDP port space is used instead of the TCP port space, then 180 bytes are available when calling rdma_connect(). If SDP is employeed, 28 bytes are available in the REQ. Future extensions to the RDMA CM, such as IB ACM support may adjust these values -- in this case, it could be more. So, even with IB there is not one correct value. The rdma_cm could expose a new call that would return the guaranteed maximum size of private data for a specific rdma_cm_id based on its state, but this doesn't really help DAPL, which is wanting to report one value for the device that is guaranteed to work for any call. - Sean _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
