[ofa-general] Re: NOSRQ misc patch [PATCH V1]
Pradeep Satyanarayana
pradeeps at linux.vnet.ibm.com
Sun Jul 22 11:39:00 PDT 2007
Michael S. Tsirkin wrote:
>> Quoting Pradeep Satyanarayana <pradeeps at linux.vnet.ibm.com>:
>> Subject: Re: NOSRQ misc patch [PATCH V1]
>>
>> Michael S. Tsirkin wrote:
>>>> Quoting Pradeep Satyanarayana <pradeeps at linux.vnet.ibm.com>:
>>>> Subject: Re: NOSRQ misc patch [PATCH V1]
>>>>
>>>> Michael S. Tsirkin wrote:
>>>>>> @@ -1168,9 +1170,9 @@ static struct ib_qp *ipoib_cm_create_tx_
>>>>>> attr.recv_cq = priv->cq;
>>>>>> attr.srq = priv->cm.srq;
>>>>>> attr.cap.max_send_wr = ipoib_sendq_size;
>>>>>> - attr.cap.max_recv_wr = 1;
>>>>>> + attr.cap.max_recv_wr = 0;
>>>>>> attr.cap.max_send_sge = 1;
>>>>>> - attr.cap.max_recv_sge = 1;
>>>>>> + attr.cap.max_recv_sge = 0;
>>>>>> attr.sq_sig_type = IB_SIGNAL_ALL_WR;
>>>>>> attr.qp_type = IB_QPT_RC;
>>>>>> attr.send_cq = cq;
>>>>> I don't see how does this fix things.
>>>>> This line
>>>>>> attr.srq = priv->cm.srq;
>>>>> connected the TX QP to SRQ, making it possible to get packets on this QP.
>>>>> But if cm.srq is NULL, and a remote sends a packet on this connection,
>>>>> the connection will get closed. Which is a quality of implementation issue.
>>>>>
>>>> When the QP numbers are exchanged correctly, then it should not receive
>>>> a packet on this QP in the first place.
>>> Re-read the RFC. It is perfectly legal to reuse a passive QP for transmitting
>>> packets. We don't do this currently but we might in the future.
>> I presume you mean passive side for receiving.
>
> A passive side is the one that gets a REQ (look in IB spec section 12.9.6).
> Under IPoIB passive side can perform post send on the QP created.
> To make this work, I connect the QP to the SRQ on the active side:
>> attr.srq = priv->cm.srq;
>
> However, with your patch, priv->cm.srq might be NULL, which
> means that the QP won't be attached to SRQ. This is
> a quality of implementation issue that your patch is introducing.
>
I do not understand -for one you mention transmitting packets and, on the
other hand you mention SRQ. Are you hinting at Shared Send Queues (which may
be in the future as you state)?
I have already tested the series of NOSRQ patches for interoperability between
IBM and Mellanox adapters and it works. I do not see the quality of implementation
issues that you keep referring to.
I believe this should not pose issues for merging this immediately into 2.6.23.
If you have ideas that can be implemented in the future, we can discuss that
but outside the context of this patch.
Pradeep
More information about the general
mailing list