[openib-general] cm_id state information

Sean Hefty sean.hefty at intel.com
Fri May 6 20:19:20 PDT 2005


>On Fri, May 06, 2005 at 11:28:01AM -0700, Sean Hefty wrote:
>> James Lentini wrote:
>> > Is there documentation or a piece of sample code that shows how to
>> > interact with the CM correctly?
>>
>> The only sample code is cmpost, under the util directory.  You could
>> probably use SDP as well.  Also, I'm currently working with Hal to make sure
>> that kDAPL works properly with the CM.  It will just take a little while
>> before a patch is ready.
>
>  I thought James' question was in the context of not using cm_id->state
>to drive the CM state machine, which SDP is currently doing, so it would
>make a bad example.

I've been looking at a way to make the cm_id->state usable to consumers, and I
think that it might be possible, but would require some rework in
the CM.  But if it would avoid consumers needing to maintain their own
connection states, it could be worthwhile.

Currently, the CM serializes all callbacks to a single client after it has
processed received MADs.  By moving the serialization higher up in the codepath,
before the MADs are processed, I might be able serialize the cm_id
state transitions with user callbacks.  So that the cm_id->state would not
change while in a user callback, unless the user invokes a routine that would
naturally change it.

I'm just not sure yet if such a change significantly simplifies CM clients
any, particularly clients that are themselves midlayers.

- Sean





More information about the general mailing list