On Wednesday, 01/28/2004 at 01:06 CET, Franco Mignogna
<[EMAIL PROTECTED]> wrote:

> I suspect the problem here is  that Gbit OSA uses QDIO mode; according
> http://www.redbooks.ibm.com/redbooks/pdfs/sg245948.pdf in QDIO mode an
OSA
> card  answers  to ARP request by itself, having IP addresses registered
by
> the IP stack. In the above redbook the process of ARP takeover is
described
> (is this what you mean for IP takeover ?).
> In the above radbook a reference to spantree exists. I don't know if it
> applies to Lucius's problem, and I don't whanto to add confusion, but it

ARP takeover and IP takeover are the same thing:  Moving the association
of an IP address from one adapter to another.  It really doesn't matter if
it is a VIPA or the base IP address.  z/OS can move the base IP address
*and* the VIPA between adapters.  To the best of my knowledge, Linux
doesn't, as it requires cooperation between the IP layer and the device
driver that doesn't exist on Linux.

The *intent* of VIPA is to insulate the host from IP address changes AND
to provide failover.  If you put the VIPA in the same subnet as the
adapters, then you lose half of the insulating value (hey! it's cold
outside!) of the VIPA.  When in a different subnet, it is not necessary
for the adapters to respond to ARPs for the VIPAs - they just have to add
it to their IP filters.

But as you suggest, the core of the problem is the gratuitous ARP
performed by OSA when it is in QDIO mode.  If Linux does not tell the
adapter "Hey, don't do that!", then OSA will not register the IP address
and will not respond to ARPs for it.   Fixing that is relatively easy.  To
have both adapters running and moving traffic, with transparent failover
for all traffic to a single adapter is much harder (requiring that
cooperation thing).  When Linux can do that, then you can failover
same-subnet VIPA the same way.

Alan Altmark
Sr. Software Engineer
IBM z/VM Development

Reply via email to