If we are generating a LID CHANGE event, why do we need the CLIENT REREGISTER at all? If both are generated, you will do BOTH light and heavy sweeps -- isn't this a waste? Preferable to check for LID CHANGE first (and generate the event if needed). Only if there is no LID CHANGE event generated, check for CLIENT REREG.
Thoughts? --Original Message----- > From: Moni Shoua [mailto:[email protected]] > Sent: Wednesday, January 14, 2009 6:21 PM > To: Roland Dreier > Cc: Jack Morgenstein; Olga Stern; Yossi Etigin; OpenFabrics General > Subject: Re: [ofa-general] [PATCH] mlx4_ib: Fix dispatch of > IB_EVENT_LID_CHANGE > > > Roland Dreier wrote: > > > Yes. This one (as its twin patch for ib_mthca) fixes a real > > problem. > > > > What is the real problem? > > > > - 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 > > > Voltaire SM, for example, sends one MAD when it takes over > another SM. This one MAD tells the client to re-register and > if necessary, changes the client LID. Old code won't generate > LID change event and the IPoIB would only do a "light flush" > of paths. > This causes a communication breakdown for a very long time. > _______________________________________________ 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
