[openib-general][PATCH 1 of 3] repost: Client Reregister support for kernel space

[Roland Dreier]

 >  the most urgent and critical case is
 >  the SM failure/restart when the SM is not connected to the host directly.
 >  In this cases neither PortError no PortActive events will be
 > generated on the host.
 >  The SM will lose the multicast group configuration for the host and
 > the host will need to rejoin its multicast groups in this case.
 >  IPoIB shall handle the problem by catching the ClientReregister event.

It seems your patch doesn't help in this situation at all.  Right now
mthca will generate a LID_CHANGE event for any set of PortInfo; IPoIB
will catch that event and rejoin all multicast groups.  Your patch
changes some of those events to CLIENT_REREGISTER events and has IPoIB
treat them exactly the same way as LID_CHANGE events.  So the behavior
won't change at all.

 >   There are additional cases. Any client which registers itself on the
 > SA, will need to handle this event in order to work properly after the
 > SM failure/restart. We'll need it very soon for a user mode
 > application.

OK, but you could use LID_CHANGE events the same way as IPoIB does
now.  Since ClientReregister support is optional, and in fact you
didn't fix ipath to generate these events, your app can't count on
CLIENT_REREGISTER events being generated anyway.

I'm not really opposed to these changes, but it is adding additional
code for what looks like very minimal improvement.  So I'm trying to
understand how this really helps you.

 - R.