On Mon, 2008-01-07 at 14:49 +0200, Dotan Barak wrote: > Dotan Barak wrote: > > Yevgeny Kliteynik wrote: > >> Dotan Barak wrote: > >>> Or Gerlitz wrote: > >>>> Dotan Barak wrote: > >>>>> Fix the value of the pkey_index in the completion to get a valid > >>>>> value for GSI QPs. > >>>> > >>>> Is libmthca fine in that respect? > >>> As much as i know, everything is fine with mthca/libmthca. > >>> > >>> We saw several problems only in ConnectX (because of the new low > >>> level driver). > >>> > >>> Right now, we are doing some more checks to check the mlx4_0 low > >>> level driver as well as the IB core. > >>> After that we'll check the mthca low level driver too. > >> > >> Currently OpenSM doesn't support any non-default pkey > >> (or any pkey at index other than 0) in sa queries. > >> When a request is received, opensm doesn't extract the > >> right pkey from the mad header - it replaces it with a > >> default pkey, and when a response is sent, OpenSM always > >> uses pkey at index 0. > >> > >> -- Yevgeny > > FYI: after several testings it seems that mthca low level driver don't > > have this problem. > > > > Dotan > > > > Just to make sure that everything is clear: I checked that the mthca low > level driver can extract > the right pkey_index in the completion of GSI QP. > > The problem that Yevgeny mentioned exists in the openSM and i opened a > bug on this issue.
Some comments on the issues raised above: 1. There has already been discussion on this list of other OpenFabrics components assuming the default PKey at index 0 (and yes, this is not mandated by IBA). 2. As to the impact of using a non default PKey (in the BTH) for querying SA, is this really used in any implementations ? It makes deployment of SM difficult (needing much more configuration). That's not to say this isn't a bug but more speaks to the severity of it. IMO it should be documented as a current limitation. [Also, note that user MAD API only supports pkey index at recent kernel and library versions.] -- Hal > Dotan _______________________________________________ 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
