[ofa-general] [PATCH] opensm: set hop limit when creating ipoib multicast groups

[Hal Rosenstock]

On 8/30/07, Rolf Manderscheid <rvm at obsidianresearch.com> wrote:
> On Thu, Aug 30, 2007 at 01:25:43PM -0400, Hal Rosenstock wrote:
> > On 8/30/07, Rolf Manderscheid <rvm at obsidianresearch.com> wrote:
> > > On Thu, Aug 30, 2007 at 12:01:57PM -0400, Hal Rosenstock wrote:
> > > > On 8/30/07, Rolf Manderscheid <rvm at obsidianresearch.com> wrote:
> > > > > Hi Sasha,
> > > > >
> > > > > This patch sets the hop limit for the IPv4 broadcast groups so that broadcasts work
> > > > > through IB routers.
> > > > >
> > > > > Signed-off-by: Rolf Manderscheid <rvm at obsidianresearch.com>
> > > > >
> > > > > ---
> > > > >
> > > > > diff --git a/opensm/opensm/osm_prtn.c b/opensm/opensm/osm_prtn.c
> > > > > index 46ee429..fc000b1 100644
> > > > > --- a/opensm/opensm/osm_prtn.c
> > > > > +++ b/opensm/opensm/osm_prtn.c
> > > > > @@ -213,7 +213,7 @@ ib_api_status_t osm_prtn_add_mcgroup(osm_log_t * p_log,
> > > > >        mc_rec.pkey = pkey;
> > > > >        mc_rec.rate = (rate ? rate : OSM_DEFAULT_MGRP_RATE) | (2 << 6); /* 10Gb/sec */
> > > > >        mc_rec.pkt_life = OSM_DEFAULT_SUBNET_TIMEOUT;
> > > > > -       mc_rec.sl_flow_hop = ib_member_set_sl_flow_hop(p->sl, 0, 0);
> > > > > +       mc_rec.sl_flow_hop = ib_member_set_sl_flow_hop(p->sl, 0, 0xff);
> > > >
> > > > Shouldn't this be based on whether the group is local scope or not ?
> > >
> > > If the group has local scope then the hop limit is irrelevant, packets will get
> > > dropped by the router anyway.  I certainly don't object to making the hop limit
> > > conditional on scope if that's the preference.
> >
> > Looking at the PathRecord code, it only checks whether a DGID was
> > requested or not and then sets the hop limit only being dependent on
> > whether ROUTER_EXP was defined or not. Should we follow continue to
> > follow this or make PathRecord responses consistent with this ?
>
>
> Isn't the plan to incorporate the ROUTER_EXP code eventually?

Eventually is relatively a long time IMO as it when there is an
official router spec.
Certainly well beyond the OFED 1.3 timeframe.

>I would
> rather not add any more #ifdefs if we can avoid it.  I think your suggestion
> of setting the hop limit according to the scope of the multicast group seems
> like the better option to me ... I'll post a new patch.  Are you OK with the
> other patch to fix the scope of the second MGID?  If so, I'll base the new
> patch on that.

Yes, I think that is OK as this would work for the no routers
supported case as does the PathRecord code.

In the longer term for both of these, I wonder whether hop limit will
continue to be returned as the max or whether some actual limit would
be returned.

-- Hal

>    Rolf
>