[ofw] RE: IPoIB partitioning support potential bug?

[Fab Tillier]

The driver should be fixed, and IPoIB too.

Note that all places where you take a pkey in the driver needs to expect it in network order.  I believe this applies to UD send work requests.

-Fab

> -----Original Message-----
> From: Leonid Keller [mailto:leonid at mellanox.co.il]
> Sent: Tuesday, June 03, 2008 3:27 AM
> To: Sean Hefty; Fab Tillier; Slava Strebkov; ofw at lists.openfabrics.org
> Subject: RE: [ofw] RE: IPoIB partitioning support potential bug?
>
> Yes, the drivers return pkeys in the host order.
> Is it safe to fix that ?
> What should be changed to adjust to that fix ?
>
>
> Index: mlx4/kernel/bus/ib/main.c
> ===================================================================
> --- mlx4/kernel/bus/ib/main.c   (revision 1222)
> +++ mlx4/kernel/bus/ib/main.c   (working copy)
> @@ -230,7 +230,7 @@
>         if (err)
>                 goto out;
>
> -       *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
> +       *pkey = ((__be16 *) out_mad->data)[index % 32];
>
>  out:
>         kfree(in_mad);
> Index: mthca/kernel/mthca_provider.c
> ===================================================================
> --- mthca/kernel/mthca_provider.c       (revision 1222)
> +++ mthca/kernel/mthca_provider.c       (working copy)
> @@ -264,7 +264,7 @@
>                 int i;
>                 __be16 *pkey_chunk = (__be16 *)out_mad->data;
>                 for (i=0; i<32; ++i)
> -                       pkey[i] = cl_ntoh16(pkey_chunk[i]);
> +                       pkey[i] = pkey_chunk[i];
>         }
>
>   out:
>
> > -----Original Message-----
> > From: ofw-bounces at lists.openfabrics.org
> > [mailto:ofw-bounces at lists.openfabrics.org] On Behalf Of Sean Hefty
> > Sent: Monday, June 02, 2008 8:57 PM
> > To: 'Fab Tillier'; Slava Strebkov; ofw at lists.openfabrics.org
> > Subject: RE: [ofw] RE: IPoIB partitioning support potential bug?
> >
> > >Ok, so does this mean that the pkey value was returned by the HCA
> > >driver in host-order?  If so, this is a bug in the HCA driver.
> >
> > I agree with Fab here.  Looking at the code, it looks wrong.
> >
> > >From __port_create_bcast() and __port_get_bcast():
> >
> >       mcast_req.member_rec.mgid.raw[4] = (uint8_t)
> >               p_port->p_adapter->guids.port_guid.pkey;
> >       mcast_req.member_rec.mgid.raw[5] = (uint8_t)
> >               (p_port->p_adapter->guids.port_guid.pkey >> 8);
> >       mcast_req.member_rec.pkey =
> >               p_port->p_adapter->guids.port_guid.pkey;
> >
> > The pkay is swapped in one case, but not the other.
> >
> > The pkey in the mcmember record is correct, but the created
> > mgid is wrong.  The pkey for the group itself is correct, so
> > I don't think the switches or SM would detect the problem.
> > But the mgid won't be the same.  Has anyone tried to use this
> > support mixing windows and linux?  I'm guessing that's when
> > we would see the problem.
> >
> > - Sean
> >
> > _______________________________________________
> > ofw mailing list
> > ofw at lists.openfabrics.org
> > http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
> >