[openib-general] [RFC] CM reject handling and redirection

[Sean Hefty]

Hal Rosenstock wrote:

>>Code 25 (port redirection only) raises a separate issue that I think
>>is orthogonal to REJ handling.  The CM spec always allows the CM to
> 
> How much complexity is added to the CM to support this ? Do we really need to
> (and if so what is the application which requires this) ? 

I think it can be as simple as adding a new field, cm_path, that can be 
used.  (The user gets back a ClassPortInfo, which would need to be 
converted to a PathRecord...)

Is it needed?  I doubt it.  I'm more concerned that if it ever does 
become needed that we don't have an architecture that prevents adding it.

> I'm not sure I get the "connection" between the CM path and data path.
> Doesn't failover and APM apply only to the data path ?

It does.  However, if the CM established the connection along the 
failed path, it will not be able to send future LAP, DREQ, etc. 
messages unless it sends those along a new path.

> Is it purely the
> detected of a failed path between CMs (which seems orthogonal to this to
> me) ? That's not to say that this is not an issue.

These are somewhat separate issues.  My intention was to have the CM 
send to the destination port that matched the currently active/primary 
path, with the hope that this would result in CM "failover" as well. 
The CM has some of the framework in place to do this, but is lacking 
most of the actual implementation.

- Sean