[ofa-general] Re: [IPoIB][RFC] remove redundant gid query

[Roland Dreier]

 > > It does look like we're doing some work we don't need to do.  However
 > > ipoib_add_port() could run before an SM has brought up the local port,
 > 
 > The same could be true for ipoib_mcast_join_task()
 > 
 > These are both instances of the general problem that if the GID at 
 > index 0 changes, the IPoIB code is not automatically notified. Agree?

Yes, although what is there now should be semi-OK: a multicast join
can't succeed until the port is up, so ipoib should eventually get the
right GID.  And I would argue that an SM that changes a port's GID
prefix without at least generating a client reregister event is broken.

 > > so the GID prefix might change later.
 > > 
 > > I'm not sure what the best way to clean this up is.
 > 
 > As an aside: Why does ipoib_add_port() treat an error return from 
 > ib_query_gid() as fatal while ipoib_mcast_join_task() only emits a 
 > warning?

I guess because it's much easier to bail out of ipoib_add_port() than
it is to do something intelligent in ipoib_mcast_join_task().