[ofa-general] SubnAdmGet (6777)

[Eli Dorfman (Voltaire)]

Eli Dorfman (Voltaire) wrote:
> Hal Rosenstock wrote:
>> On Mon, Jun 1, 2009 at 5:36 PM, Hal Rosenstock <hal.rosenstock at gmail.com> wrote:
>>> On Mon, Jun 1, 2009 at 4:27 PM, Sean Hefty <sean.hefty at intel.com> wrote:
>>>>> Yes, RMPP is an overhead when the response is a single MAD but is this
>>>>> significant ? Anyhow, how can the spec be changed in a way that
>>>>> doesn't break existing implementations ?
>>>> But the implementations are assuming different things about SubnAdmGet.  The SA
>>>> is assuming that the query should fail if multiple records match.  The client
>>>> side software (ipoib and rdma_cm) assume that it will obtain a single record
>>>> even if multiple paths are present.  So, something needs to change.
>>> Seems so.
>>>
>>>> The spec indicates that value in the request is ignored and NumbPath is 1, not
>>>> that NumbPath is completely ignored.
>>> For Get, it doesn't say that the matches are paired down to this
>>> number as it does for GetTable.
>>>
>>>>  Also see page 1242 in the SDP annex which
>>>> reads: 'NumbPath could be 1 (in which case the SA query may use SubnAdmGet
>>>> rather than SubnAdmGetTable)'.
>>> SDP annex is not the primary source for this (chapter 15 is) and is
>>> inconsistent and no one caught this.
>>>
>>>> To me, this implies that SubnAdmGet should be
>>>> treated equivalent as SubnAdmGetTable with NumbPath = 1.
>>>> It just seems really odd to treat NumbPath differently for PR SubnAdmGet versus
>>>> PR SubnAdmGetTable and MPR SubAdmGetMulti.  Basically, this makes PR SubnAdmGet
>>>> useless.
>>> when there's a subnet with multiple paths and the requests are not
>>> specific enough to use get.
>>>
>>> Seems like either the queries need to use RMPP, or the spec modified
>>> (if that's possible) and the SAs updated.
>> I sit corrected :-) Your interpretation of the spec is correct. Also,
>> in looking at OpenSM, the intent is as you indicate: it does try to
>> only return 1 attibute for get PR. If when returning the response,
>> there is more than 1 attribute in the list, it returns the too many
>> records error. There must be some code path I don't see right now
>> which is doing this. It would be useful to know the details of the
>> query (get request) causing this.
>>
> 
> This may happen when pr_rcv_get_port_pair_paths() is called several times.
> The only case i see is pr_rcv_process_world() that means the request is without or wrong 
> src and dest port or component mask for SGID and DGID is 0.

correction - this may happen only when component mask for SGID and DGID is 0.

Eli