Re: cm_process_routed_req() does not resonate well with RoCE systems

[Jason Gunthorpe <jgg@nvidia.com>]

On Fri, Mar 19, 2021 at 03:07:07PM +0000, Haakon Bugge wrote:
> 
> 
> > On 18 Mar 2021, at 20:21, Haakon Bugge <haakon.bugge@oracle.com> wrote:
> > 
> > With the introduction of RoCE systems, a CM REQ message will contain (pasted from Wireshark):
> > 
> > Primary Hop Limit: 0x40
> > .... 0... = Primary Subnet Local: 0x0
> > 
> > This because cma_resolve_iboe_route() has:
> > 
> >        if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
> >                /* TODO: get the hoplimit from the inet/inet6 device */
> >                route->path_rec->hop_limit = addr->dev_addr.hoplimit;
> >        else
> >                route->path_rec->hop_limit = 1;
> > 
> > The addr->dev_addr.hoplimit is coming from addr4_resolve(), which has:
> > 
> > 	addr->hoplimit = ip4_dst_hoplimit(&rt->dst);
> > 
> > ip4_dst_hoplimit() returns the value of the sysctl net.ipv4.ip_default_ttl in this case (64).
> > 
> > For the purpose of this case, consider the CM REQ to have the Primary SL != 0.
> > 
> > When this REQ gets processed by cm_req_handler(), the cm_process_routed_req() function is called.
> > 
> > Since the Primary Subnet Local value is zero in the request, and since this is RoCE (Primary Local LID is permissive), the following statement will be executed:
> > 
> > 	IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl);
> > 
> > At least on the system I ran on, which was equipped with a
> > Mellanox CX-5 HCA, the wc->sl is zero. Hence, the request to setup
> > a connection using an SL != zero, will not be honoured, and a
> > connection using SL zero will be created instead.
> > 
> > As a side note, in cm_process_routed_req(), we have:
> > 
> > 	IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, wc->dlid_path_bits);
> > 
> > which is strange, since a LID is 16 bits, whereas dlid_path_bits is only eight.
> > 
> > I am uncertain about the correct fix here. Any input appreciated.
> 
> I intend to send a patch doing:
> 
> +++ b/drivers/infiniband/core/cm.c
> @@ -2138,7 +2138,8 @@ static int cm_req_handler(struct cm_work *work)
>                 goto destroy;
>         }
>  
> -       cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
> +       if (cm_id_priv->av.ah_attr.type != RDMA_AH_ATTR_TYPE_ROCE)
> +               cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
>  
>         memset(&work->path[0], 0, sizeof(work->path[0]));
>         if (cm_req_has_alt_path(req_msg))
> > 
> if I do not get a push back.

This does seem reasonable, but I don't understand the underlying
issue, why is anything in roce land touching the SL? Are you trying to
use the SL as a proxy for the TOS bits?

Jason