[openib-general] [PATCH] RDMA/iwcm: Bugs in cm_conn_req_handler()

[Tom Tucker]

On Mon, 2007-02-12 at 08:19 -0800, Sean Hefty wrote:
> > This design is based on the RDMA_CM and IB_CM behavior.  If the app
> > issues the destroy via rdma_destroy_cm_id, then we block that thread
> > until all references are gone.  If the app returns non-zero in a
> > callback for a given cm_id, then the CM owns destroying the cm_id and
> > the application is done with it. That's the short of it.  Here's the
> > long of it:
> 
> Note that the goal of this behavior is simply to ensure that no thread will 
> touch any code in their callback after destroying their cm_id.  That is all that 
> needs to be guaranteed, if this helps any.

It help a lot actually. We've discussed simplifying this code by not
blocking the destroy and guaranteeing that events received after the
destroy are never delivered, but we didn't want to do something this
drastic without some time to get it right.

> 
> - Sean