[openib-general] Re: fixes to the udapl ucm/uat patch you sent

[Arlin Davis]

James,

>
> Hi Arlin,
>
> Can I break this patch into 3 parts: the changes to dapl_evd_wait, the 
> changes to dapl_evd_resize, and the ib changes? I think it will be 
> easier to discuss each set of changes seperately (with so many 
> seperate issues, I'm afraid I've missed your reply to some of these 
> questions)

Yes please break them out. I need to go back and take a closer look at 
my error cases for the wait and resize. In the future I will keep any 
common code patches seperate from the IB code.

>
> dapl_evd_wait:
>
> I looked over the original implementation of dapl_evd_wait() with an 
> eye towards the situation you described (the caller polling and 
> finding fewer events than requested, the caller going to turn on 
> notification, an event occuring, the caller turning on notification, 
> the caller blocking unaware of the last event). I don't believe that 
> this would happen in the original implementation. Here's why: after 
> the caller turns on notification, the code loops, via the continue 
> statement on line 213, back to the begining of the for loop on line 
> 173 and repolls. Do you agree?

Hmmm. Yes it appears to be correct now that I take another look at it. 
Ok, ignore the patch for now and I will take another look
at the senario that was missing the completions. On a side note, does 
this call need to adjust the time if coming out of a wait but still not 
reaching threshold and going back into a subsequent wait?

>
> dapl_evd_resize:
>
> I'm still unsure of why you removed the call to 
> dapls_evd_event_realloc() and moved the work that was being performed 
> in that routine up into dapl_evd_resize(). If we don't call 
> dapls_evd_event_realloc() anymore, the code should be removed.
>
sorry. I had some issues with resize and just grabbed a gen1 IBAL 
version that rolled up everything into one call. Replaced the 
reallocates with the allocs. The event_realloc and the rbuf_reallocs are 
no longer used in this version. Let me go back to my error condition and 
see if I can isolate to one of the realloc calls instead of a wholesale 
change to resize.

Cancel this patch for now.

> ib changes:
>
> These look ok to me. I've checked them into revision 2955.

thanks,

-arlin

>
>
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
>
> To unsubscribe, please visit 
> http://openib.org/mailman/listinfo/openib-general
>