[ofa-general] Non-Signaled RDMA Read (cxgb3)

Steve Wise swise at opengridcomputing.com
Tue Jul 29 12:45:27 PDT 2008


Caitlin Bestler wrote:
> On Tue, Jul 29, 2008 at 9:47 AM, Felix Marti <felix at chelsio.com> wrote:
>   
>
>   
>> Roland, AFAIK, even an unsignaled WR generates a CQE when an error
>> occurs (and then all subsequent WRs complete in error as well). Doesn't
>> that imply that the CQ must be sized assuming that _every_ WR can
>> generate a CQE?
>>
>>     
>
> Yes, that is exactly what application writers should assume.
>
> They should also assume that spurious completion interrupts are possible
> (although obviously they should be minimized) and therefore the application
> should always be ready be woken up early and find nothing to do.
>
> More complete descriptions that could be used as guidelines can be found
> in the DAPL manuals.
>
> About the only thing I think that can be done about this specific quirk is
> to document it so that application developers are aware of it, and know that
> it is a model-specific limitation and do not think it is a limitation of RDMA
> or iWARP. If their application does not need the notification it should
> keep requesting unsignaled completion.
>   
Well if applications should handle infrequent spurious notifcations 
where no cqe is available, then I can indeed add code in my provider 
poll to silently throw away unsignaled read completions.


Steve.




More information about the general mailing list