[openib-general] [RFC] libibverbs completion event handling

[Caitlin Bestler]

On 9/21/05, Roland Dreier <rolandd at cisco.com> wrote:
>
> Caitlin> I'm not sure I follow what a "completion channel" is. My
> Caitlin> understanding is that work completions are stored in
> Caitlin> user-accessible memory (typically a ring buffer). This
> Caitlin> enables fast-path reaping of work completions. The OS has
> Caitlin> no involvement unless notifications are enabled.
>
> Right. Notifications ("events" in the terminology I used) are the
> only thing I was talking about.
>
> Caitlin> The "completion vector" is used to report completion
> Caitlin> notifications. So is the completion vector a *single*
> Caitlin> resource used by the driver/verbs to report completions,
> Caitlin> where said notifications are then split into user context
> Caitlin> dependent "completion channels"?
>
> Yes.



And how does a completion vector relate to callbacks? Would this
be typically identifying the number of "processing resources" (such
as distinct execution contexts) that will be allocated?

Or more formally, if two CQs use the same completion vector, are
we guaranteed that they will not have concurrent callbacks?


Caitlin> The RDMAC verbs did not define callbacks to userspace at
> Caitlin> all. Instead it is assumed that the proxy for user mode
> Caitlin> services will receive the callbacks, and how it relays
> Caitlin> those notifications to userspace is outside the scope of
> Caitlin> the verbs.
>
> That "outside the scope" part is exactly what I'm talking about
> implementing here.



How would this enable kernel mode processing of notifications
for a user-mode CQ by the proxy?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20050922/d99c84af/attachment.html>