[openib-general] IPoIB CM for merge?
Michael S. Tsirkin
mst at mellanox.co.il
Sun Feb 4 05:06:06 PST 2007
> Quoting Pradeep Satyanarayana <pradeep at us.ibm.com>:
> Subject: Re: [openib-general] IPoIB CM for merge?
>
>
> Hello Michael,
>
> Here are a few more observations :
Pradeep, I think you are posting in the wrong thread: it seems you are not
talking about my code, but rather about the project you mentioned of
implementing IPoIB CM without SRQ.
IPoIB CM currently falls back on UD mode for HCAs that do not support SRQ,
so there should be no problem for the ehca - as new code won't be activated.
As I said already, I do not see a clean way to address this limitation,
so I would rather have current IPoIB CM code merged upstream first, and think
about enhancements later.
>
> 1. For the SRQ case, the skbs and recieve biffers are posted during init and even before the rx_qp is created. This causes a problem (atleast for non SRQs) for the ehca. We need to call the ipoib_cm_alloc_skb() and ipoib_cm_post_recieve() after the rx_qp
> is in the RTR state.
>
> 2. Also found that in ipoib_cm_create_rx_qp() one needs to initialize .cap.max_recv_wr and .cap.max_recv_sge. Otherwise this leads to some problems like rq overflows and causing communication failures.
Yes, I think these are some of the things that would need to be done to make IPoIB CM
work without SRQ. It is clearly not something we want to do for SRQ case however:
for example, posting WRs to SRQ during connection setup would race
against completion events for other connections. And assigning .cap.max_recv_wr > 0
for a QP not connected to SRQ does not make sense, and might thinkably confuse
low level drivers.
--
MST
More information about the general
mailing list