On Wed, Apr 12, 2017 at 12:51:25PM -0400, Hal Rosenstock wrote: > On 4/12/2017 10:49 AM, Honggang LI wrote: > > On Wed, Apr 12, 2017 at 10:34:47AM -0400, Hal Rosenstock wrote: > >> On 4/12/2017 10:18 AM, Honggang LI wrote: > >>> On Wed, Apr 12, 2017 at 09:34:20AM -0400, Hal Rosenstock wrote: > >>>> On 4/12/2017 9:29 AM, Honggang LI wrote: > >>>>> On Tue, Apr 11, 2017 at 09:49:44AM -0600, Jason Gunthorpe wrote: > >>>>>> On Tue, Apr 11, 2017 at 11:28:53PM +0800, Honggang LI wrote: > >>>>>>> From: Honggang Li <honli@xxxxxxxxxx> > >>>>>>> > >>>>>>> According to InfiniBand Architecture Release 1.2.1, Table 208 > >>>>>>> Example PathRecord Request MAD Header Fields, MADHeader:Method > >>>>>>> should setup to 0x12 (SubnAdmGetTable()). > >>>>>> > >>>>>> That is just an example, Table 192 shows that Get and GetTable are > >>>>>> both valid query types. > >>>>> > >>>>> I don't understand this as "Table 192 PortInfoRecord" seems nothing to > >>>>> do with PathRecord request. > >>>> > >>>> In IBA 1.3 volume 1, it's Table 209 Subnet Administration > >>>> Attribute/Method Map shows PathRecord methods are Get and GetTable. > >>>> > >>> > >>> http://www.intel.com/content/dam/www/public/us/en/documents/product-briefs/truescale-infiniband-12300-switch-brief.pdf > >>> Interoperability > >>> • Compliant with IBTA* specifications 1.0a, 1.1, 1.2, and 1.2. > >>> > >>> It seems the embedded subnet manager does not support IBA 1.3 > >> > >> It's same for IBA 1.2.1. This part hasn't changed. > >> > >>>>>> > >>>>>> I however expect that the SM will fail when using GetTable if the > >>>>>> query returns more than one result, could that be what is happening? > >>>>>> When you convert it to GET_TABLE does it return more than one result? > >>>>> > >>>>> strace output shows srptools always received MAD packet in 320 bytes. So > >>>>> the embedded subnet manager only returns one result. > >>>> > >>>> What is MAD status ? > >>> > >>> ib_user_mad->status = 0x0 > >>> > >>> I attached the MAD packet dump out with strace. Please see the > >>> attachments p4.in amd p4.decode. > >> > >> p4.decode has: > >> rp_dm_rmpp_sa_mad->method = 0x81 //IB_MAD_METHOD_GET_RESP > >> opensm/include/iba/ib_types.h // GetResp() Method (13.4.5) > >> > >> Proper response to GetTable (0x12) method is GetTableResponse (0x92) > >> method but this is GetResp (0x81) response. > > > > Yes, you are right. p4.decode was captured with unpatched upstream srptools-1.0.3. > > against the embedded SM ? Yes. > If so, this response looks (mostly) good to me and should be accepted by srp_daemon. I think I see the issue in srp_daemon: it's due to the check of the attribute offset which is not being set in the response: > > size = ib_get_attr_size(in_sa_mad->attr_offset); > if (!size) { > if (config->verbose) > printf("PathRec Query did not find any targets " > "over P_Key %x\n", pkey); > continue; > } > > p4.decode has: > srp_dm_rmpp_sa_mad->attr_offset = 0x0 > > Since a get rather than get table is being used, a good MAD status means that a valid PR was returned and there is no need for this check. > > Would you try using the original get method without this code block and verify that it works with the embedded SM ? > I sent the scratch RPM with the patch you suggested to third party for testing. Please wait. -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
