[openib-general] ibv_cmd_create_qp() question

[Caitlin Bestler]

openib-general-bounces at openib.org wrote:
>     Steve> Just curious, why don't all the verbs have this support?
> 
>     Steve> Maybe we should align all the verbs to support commands and
>     Steve> responses that allow provider-specific extensions?
> 
> It seemed like over-engineering to me.  I don't think we need
> arbitrary data in every single verb.  I could easily be wrong
> but I think we'll very quickly converge on a set of verbs
> that covers all HW.
> 

Allowing provider-specific data to be established when an object
is *created* would definitely be adequate. That allows sharing
info about provider-specific data structures. Once that information
is shared, however, there should be no need to have private
exchanges on each and every verb. The modifications being
requested in a qp_modify are *not* provider-specific, merely
the implementation data structures that the modifications 
will be made upon. But since the user-provider-verbs and
kernel-provider-verbs already share that information there
is nothing provider-specific that has to be communicated
with a modify, query or delete verb.

So the relevant set is create verbs for objects that are
accessed on the fast path (and hence might have user-space
created data structures): QP, SRQ and CQ.