[openib-general] Re: OpenIB documentation

[Christopher C. Aycock]

Creating a thread is no problem. I notice that the queue data  
structures have pthread mutex locks and condition variables  
predefined; what is this for? Could I simply use pthread's  
synchronization primitives against these constructs? If so, are they  
already initialized during queue creation, or will I have to do that  
manually?

Getting back to ibv_get_cq_event(), what is it then that triggers a  
completion event?

And finally, with regard to ibv_req_notify_cq(), just what does the  
second parameter mean? The example programme passes a 0 value for  
solicited_only; this is in a section that is relying on events. I'm  
totally confused by what is meant regarding "only solicited for  
notification."

Thanks again,
Chris

On Mar 23, 2006, at 4:44 AM, Michael S. Tsirkin wrote:

> Quoting r. Christopher C. Aycock <christopher.aycock at chch.ox.ac.uk>:
>> For example, is there any way to have IBverbs invoke a callback
>> function when an item arrives on the completion queue? The IBA spec
>> mentions a "Completion Event Handler"; is this what I'm looking for?
>> If so, how do I use it in IBverbs?
>
> No. This would require creating a thread behind the scenes and Roland
> didn't want to do that. You can create a thread yourself and
> call ibv_get_cq_event from there.
>
>> One function I'm kind of confused with is ibv_req_notify_cq(). Is the
>> second parameter simply a true / false value? This isn't very obvious
>> since the rc_pingpong.c example file passes a 0 value when it
>> requests CQ notification.
>
> Yes.
>
>> Another confusing one is ibv_get_cq_event(). Am I correct in assuming
>> that this is a blocking function? That's the impression I get from
>> the example codes and from the fact that struct ibv_comp_channel
>> appears to merely consist of a file descriptor. My impression is that
>> it "wakes up" when an element appears on the CQ,
>
> No, when a completion event is triggered. A single event might
> signal multiple completions.
>
>> but that it requires
>> the use of ibv_poll_cq() to actually obtain that element. Is this
>> correct?
>
> Yes.
>
> -- 
> Michael S. Tsirkin
> Staff Engineer, Mellanox Technologies
>