[ofw] RE: patch: Support APM on IBAL code

[Fab Tillier]

Hi Tzachi,

Tzachi Dar wrote on Thu, 24 Sep 2009 at 07:33:35:

> It seems that one change in my code is causing build break, and I have
> there for reverted this part on 2467.
> 
> It seems that the function al_cep_pre_req is now replaced with
> kal_cep_pre_req witch doesn't have the IB QP, as part of it.
> 
> In order to allow me to return this code, it seems that the best
> solution is to add a new member to the struct iba_cm_req:
> iq_qp_hqndle_t h_qp.

The CID is stored in the QP in the __cep_conn_req function (al_cm_qp.c), and if you want to have the CID stored in the kernel QP object for a user-mode QP, you should do that in the CM proxy.  The problem you have if you do that, is that you must be careful that the CID, if destroyed by the kernel, is no longer referenced anywhere in user-mode (otherwise the CID could get reused for that same process and later erroneously destroyed by user-mode when cleaning up an old reference.)

You could store the CID in the QP in the proxy_cep_pre_req and proxy_cep_pre_rep handlers.

You need to be very careful about managing the CID's lifetime.  I had to rewrite a ton of the CM code to fix a bug where the CID was being recycled before all user-mode references had been released.  You likely need to add code to destroying_qp to only destroy the CEP if it's a kernel QP object.  You'll also need to make sure you don't end up with potential deadlocks during cleanup.  The code looks fine here though (destroying_al will cleanup CEPs before destroying children objects).

 
> Are there any objections to this? (In other words, am I breaking some
> kind of abstraction here?)

Yes, the we're trying to make the CEP manager more independent of AL, so that WinVerbs can use it more naturally.

-Fab