[openib-general] Re: [PATCH 3/3] iWARP CM - iWARPConnectionManager
swise at opengridcomputing.com
Thu Mar 23 08:57:36 PST 2006
On Wed, 2006-03-22 at 18:10 -0800, Sean Hefty wrote:
> This is fairly abstract, so please let me know if my understanding or logic is
> off anywhere.
> Do we really have two conceptual contexts here? One that tracks the user's
> information and another that interacts with the provider?
> The user context is created / destroyed by calls to iw_create_cm_id() /
> The provider context is "created" before calling connect() or accept(). It's
> "destroyed" if connect() or accept() fail, CONNECT_REPLY returns a failure, or a
> CLOSE event occurs. For the events, the context can be considered "destroyed"
> after the callback that reported these events has returned to the provider.
> The issue is that we cannot touch either context from the other after it has
> been destroyed. In turn, this implies that we need to acquire a reference on
> each context before accessing it.
> The good news is that acquiring and releasing references is simple and doable.
> More good news is that destruction of each context will only be initiated once.
> The bad news is that we cannot block in the thread that reports events, which
> potentially complicates destruction of the provider context. So, while we are
> safe having the provider context reference/dereference the user context. The
> user context may never reference the provider context, because it can be
> destroyed at anytime after calling connect() or accept().
> So, what if we viewed the cm_id as having two pieces, or gave it multiple
> reference counts, or split it into two structures? Could any of that be used to
> simplify / validate the design?
IMO, doing this will result in another IWCM implementation that is about
the same complexity...
More information about the general