On Fri, 2005-09-30 at 14:28, Pradeep Satyanarayana wrote: > I have been following this and the other thread on CMA. There appears > to be some opinions to removing the ib_at module and introduce CMA. > Is that correct? > > True, CMA will need some form of address translation. Can we not use > some incarnation of ib_at for that? I realize that ib_at has a > net_device > refcnt problem. Is this refcnt problem a usage issue rather than just > a bug in the implementation?
It's an implementation issue which can be fixed. The address translation portion of CMA will need to do something similar (as SDP does or a fixed AT would). -- Hal > How would CMA solve the refcnt issue? > What am I missing? > > Pradeep > [EMAIL PROTECTED] > > > [EMAIL PROTECTED] wrote on 09/30/2005 07:27:32 AM: > > > On Tue, 2005-09-27 at 00:08, Michael S. Tsirkin wrote: > > > Why does AT need to keep netdev reference for longer? > > > > I don't think it really does and could be changed. I think (but am > not > > sure) it was a convenience of implementation to try to make the > netdev > > reference counting simpler. > > > > It only needs to hold the netdev for sending the ARP (like SDP). > > > > It needs the underlying ib_device and port for ATS and path queries > as > > well as reregistration if the interface address changes and > > deregistration if the IPoIB interface is removed. (SDP doesn't need > to > > worry about these aspects (only path queries).) > > > > -- Hal > > > > _______________________________________________ > > openib-general mailing list > > [email protected] > > http://openib.org/mailman/listinfo/openib-general > > > > To unsubscribe, please visit > http://openib.org/mailman/listinfo/openib-general _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
