[rds-devel] [ofa-general] Re: [PATCH] opensm: fix iser service-id used for SL assignment

Or Gerlitz or.gerlitz at gmail.com
Thu Nov 6 13:25:46 PST 2008


On Thu, Nov 6, 2008 at 8:58 PM, Andy Grover <andy.grover at oracle.com> wrote:
>> can you guys revert to 18634 to make things simpler wrt RDS/QoS configuration?

> It appears this is a fix for multiple rds transports each trying to bind to that port with
> INADDR_ANY, see commit f0af6566. I think the correct fix is to use a single port but have > transports listen on their specific interfaces only.

Andy,

commit f0af6566 came to handle the case where RDS iWARP and RDS TCP
listeners co-exist on the same node.

In the general case, there's would be no special interface nor IP
address for iWARP, the same IP is used for both TCP connections served
by the OS stack and iWARP connections serves by the NIC TOE stack.
This creates the "TOE port space problem" since when the NIC gets TCP
connection request for port X it has no clue if it need to be served
by the TOE stack or the OS stack, so RDS iWARP connection request to
port 18635 could be routed to iperf server that was spawned to listen
on that port. A possible solution that was suggested by the iWARP guys
was to have the TCP port space being shared between TCP and RDMA
listeners, currently the Linux kernel netdev maintainers are not
willing to accept such patch, and the current suggestion was applied
in ofed 1.4 see cma_0100_unified_tcp_ports.patch under
kernel_patches/fixes

> I think this is too big a fix for 1.4.0 so I will simply disable TCP transport there (leaving just > IB transport, thus no problem) and move the port back to 18634. For 1.4.1 we will have
> multiple transports again and will need to fix this by not using INADDR_ANY, as described > above.

Yes, lets have the IB transport use port 18634. As I explained above
the INADDR_ANY usage is not the problem.

Or.



More information about the rds-devel mailing list