Sean Hefty wrote: >> Does the the code below is what needs to be done by a CMA >> consumer to listen on a SID based on <port-space,port> over >> all the active IB ports of this node?
> Your code should be sufficient to listen across all RDMA devices. OK, thanks for confirming that. >> What's most important for me now is to understand that the code >> below does what i expect, however looking in ib_cm.h and the cma >> code while coding the example, one point that i run into was the >> granularity the IB CM allows consumers "bind" objects to. > The <port-space, port> binding is unique system wide if an IP address of 0 is > specified, so the generated SID will also be unique system wide. OK >> I see that CM IDs are bounded to **device** and not to >> <device,port> couple? if indeed, what was the rational? > Basically, the implementation was easier, and I didn't see a need to track > listens per port. OK, given the below info it makes sense >> It means that we don't support a config of a node connected to >> two IP subnets each over a different port of the same device, >> and two processes, each attempting to listen over the same port, >> but with different ip. > As long as the processes bind to different IP addresses, this is supported. > Rdma_bind() allows binding to the same port if the IP addresses are different. > Rdma_listen() will map both requests to the same SID, but distinguishes them > using the private data carried in the REQ. wow! i was not aware to that, thanks Or. _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
