Multicast handling was susceptible to several race conditions, where different sources would try to start/stop multicast thread and create/destroy multicast groups in parallel.
The following two patches (sent separately) solve these races by 1. Forcing all entries into ipoib_multicast.c to go through the single-threaded ipoib workqueue. 2. Protecting each access to mcast->queue, priv->broadcast and priv->mcast_list by priv->lock. With these two patches applied (and with the mthca patch that I've sent previously), I've been running the up/down test for more than 24 hours without failures now. -- MST _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
