Roland Dreier wrote:
> The problem:
> A multicast sender joins the MGID as full member, but does not leave
> (as long as the interface is up). This causes an MGID leakage in the
> SM.
I think ipoib has used send-only joins for a while now. But the problem
of never cleaning up these memberships stands.
The problem worse because ipoib is not really using send-only joins, it
joins as full member. See ipoib_multicast.c/ipoib_mcast_sendonly_join():
struct ib_sa_mcmember_rec rec = {
#if 0 /* Some SMs don't support send-only yet */
.join_state = 4
#else
.join_state = 1
#endif
};
Thanks for looking at this -- it would be nice to deal with this.
> An issue we found with this solution, is that it takes a while until the
> neighbour is cleaned, and it requires that the host will send some traffic
> (I supposed that unless we put more entries in the routing table, out
neighbour
> will not be garbage collected).
Would it work to have a more naive approach, and just keep track of the
last time each send-only membership was used, and set up a timer that
runs, say every minute, and garbage collect send-only memberships that
are more than 2 minutes old, or something like that?
I guess it would, and it will sure be simpler.
- R.
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
