[openib-general] CMA: port 2 loopback problems

[Or Gerlitz]

On 5/9/06, Roland Dreier <rdreier at cisco.com> wrote:

> The DLID must be the LID of the port.
> If the LID changes when the port becomes active then packets sent to
> the old LID will start going into the fabric and either disappear or
> go to some other port.

OK, thanks this is what i was thinking.

> There was some discussion on openib-general a year or two ago about
> picking a pseudo-random LID for local ports when initializing HCAs, to
> allow loopback connections to stay alive in most cases (collisions are
> relatively unlikely and the SM should respect existing LIDs...).

Sure, i remember that. I think such an approach should be instantly
**rejected** on the spot by the openib community. Among other reasons
since basically, IB is not ment to that. The SM should set the LIDs of
the fabric and its easy to come up with senarios which would not be
supported under this approach, eg LMC > 0 would cause the SM to change
the random LID, systematic derivation of the LID eg from the IP is
problematic as only 48K of the 64K LIDs are legal for unicast,etc,etc
you named it.

Or.