[ewg] Re: [PATCH] IPOIB/CM Increase retry counts for OFED-1.3

[Pradeep Satyanarayana]

Tziporet Koren wrote:
> Pradeep Satyanarayana wrote:
>> This patch change retry counts to small values. This helps
>> interoperability
>> between ehca and mthca. Without this patch I had seen "send completion
>> errors".
>>
>> Or Gerlitz has started a thread on the general mailing list and the
>> complete
>> discussion will be available there. This is the second part of the patch
>> submitted yesterday and is split up as per Eli's request.
>>
>> Signed-off-by: Pradeep Satyanarayana <pradeeps at linux.vnet.ibm.com>
>> ---
>>
>> --- ofa_kernel-1.3_a/drivers/infiniband/ulp/ipoib/ipoib_cm.c   
>> 2008-02-12 17:46:03.000000000 -0500
>> +++ ofa_kernel-1.3_b/drivers/infiniband/ulp/ipoib/ipoib_cm.c   
>> 2008-02-12 17:46:58.000000000 -0500
>> @@ -1016,8 +1016,8 @@ static int ipoib_cm_send_req(struct net_
>>      req.responder_resources          = 4;
>>      req.remote_cm_response_timeout = 20;
>>      req.local_cm_response_timeout  = 20;
>> -    req.retry_count           = 0; /* RFC draft warns against retries */
>> -    req.rnr_retry_count           = 0; /* RFC draft warns against
>> retries */
>> +    req.retry_count           = 3;
>> +    req.rnr_retry_count           = 3;
>>      req.max_cm_retries           = 15;
>>      req.srq                   = ipoib_cm_has_srq(dev);
>>      return ib_send_cm_req(id, &req);
>>
>>
>>   
> 
> I wish to see Roland's respond to this patch.
> I think its enough to enlarge only the rnr_retry_count since the
> retry_count is for cases were packages are dropped by a very busy subnet.
> The case you want to solve should be covered by the rnr_retry_count.
> 

The case that I saw was the other way round. The sender saw "send completion
errors" and that was solved by changing the retry_count. The rnr_retry_count
was added to cover any other corner cases I had not seen (Table 78 in the spec).

Pradeep