> On 15 May 2018, at 02:38, Hal Rosenstock <hal@xxxxxxxxxxxxxxxxxx> wrote: > > On 5/14/2018 5:02 PM, Jason Gunthorpe wrote: >> On Thu, May 10, 2018 at 05:16:28PM +0200, Håkon Bugge wrote: >> >>> We are talking about two things here. The PKey in the BTH and the >>> PKey in the CM REQ payload. They differ. >>> >>> I am out of office, but if my memory serves me correct, the PKey in >>> the BTH in the MAD packet will be the default PKey. Further, we have >>> per IBTA: >> >> This sounds like a Linux bug. >> >> Linux does not do a PR to get a reversible path dedicated to the GMP> so it always uses the data flow path, thus the GMP path paramenters >> and those in the REQ should always exactly match. >> >> Where is Linux setting the BTH.PKey and how did it choose to use the >> default pkey? Lets fix that at least for sure. Linux isn’t. The BTH.PKey is inserted by the HCA (hw or fw) coming from the P_Key table (10.9.2 in IBTA), selected by a pkey_index associated with the QP. As per C10-133: Packets sent from the Send Queue of a GSI QP shall attach a P_Key associated with that QP, just as a P_Key is associated with nonmanagement QPs >> Once that is fixed the rest of the series makes no sense since a REQ >> with invalid PKey will never arrive. >> >> However... >> >> This series seems inconsistent with the spec. >> >> IIRC the spec doesn't say if a full or limited pkey should be placed >> in the REQ (Hal?). > > CM spec for REQ just says partition key without indicating whether this > means P_Key or just the partition (15 bits) so my read is that either > full or limited pkey is allowed in REQ. > >> It is designed so that the requestor can get a >> single reversible path and put that results into the REQ without >> additional processing, however the PR returns only one PKey and again, >> it is not really specified if it should be the full or limited pkey >> (Hal?). > > Correct; it's not specified. > >> Basically this means that any pkey in the REQ could randomly be the >> full or limited value, and that in-of-itself has not bearing on the >> connection. >> >> So it is quite wrong to insist that the pkey be limited or full when >> processing the REQ. The end port is expected to match against the >> local table. > > Note that there is thorny issue with shared (physical) port > virtualization. In shared port virtualization, the VF pkey assignment is > a local matter. Only thing SM knows is the physical port pkeys where > both full and limited membership of same partition is possible. It is > conceivable that CM REQ contains limited partition key between 2 limited > VFs and for that a new REJ reason code appears to be needed. +1 Håkon > > -- Hal > >> The real answer to your trap problem is to fix the SM to not create >> paths that are non-functional, that is just flat out broken SM >> behavior. >> >> Jason >> -- 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
