[openib-general] create QP failure

Kyle Schochenmaier kschoche at scl.ameslab.gov
Fri Feb 24 08:31:06 PST 2006


Fabian Tillier wrote:

>On 2/22/06, Roland Dreier <rdreier at cisco.com> wrote:
>  
>
>>   Kyle> The second ibv_create_qp() call fails for some reason, even
>>   Kyle> though it's being initialized with the same attributes as
>>   Kyle> the first call/qp.  I didnt see anything in particular that
>>   Kyle> said this should break on updating to the latest rc's.  Have
>>   Kyle> I missed a major change which would cause this to break now?
>>
>>It should still work.  There was a change to the interface, so that
>>the kernel now returns the real QP capacities so libibverbs can give
>>you the true QP attributes that your QP was created with.  This
>>probably introduced a bug somewhere.
>>    
>>
>
>Roland, is the attribute structure now being used as an output
>parameter when it wasn't before?  If so, what happens if the output
>from one create_qp call gets passed as input into another?
>
>And Kyle, did the application reset the attributes before the second
>call, or just pass the output of the first call as input to the
>second?
>
>- Fab
>
>
>
>  
>
   In my case, where I was pretty much literally calling:
    ibv_create_qp(c->qp,&attr);
    check_for_errors();
    ibv_create_qp(c->qp_ack,&attr);
    check_for_errors();

    The second create_qp call failed over.
    So maybe the answer to your question about parameter passing can be 
discovered from that ;)

   Apparently it is (now?) necessary to use a seperate - or 'reset' the 
old one - ibv_qp_init_attr struct to pass to subsequent ibv_create_qp() 
calls.
   Previously I had been unaware of this, and was able to get by with 
using the same init_attr's passed to both create_qp(*) calls.
   I guess I just needed to zero out some of the values that got set 
inside the create_qp() call, it makes sense that I would need to do this..
   
thanks,

- Kyle

-- 
Kyle Schochenmaier
kschoche at scl.ameslab.gov
Research Assistant, Dr. Brett Bode
AmesLab - US Dept.Energy
Scalable Computing Laboratory 




More information about the general mailing list