On Sun, Mar 22, 2020 at 11:30:30AM +0200, Leon Romanovsky wrote: > From: Mark Zhang <markz@xxxxxxxxxxxx> > > In the request handler of the response side, Set flow label of the > recv_wc if it is not net. It will be used for all messages sent > by the responder. > > Signed-off-by: Mark Zhang <markz@xxxxxxxxxxxx> > Reviewed-by: Maor Gottlieb <maorg@xxxxxxxxxxxx> > Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx> > drivers/infiniband/core/cm.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/drivers/infiniband/core/cm.c b/drivers/infiniband/core/cm.c > index bbbfa77dbce7..4ab2f71da522 100644 > +++ b/drivers/infiniband/core/cm.c > @@ -2039,6 +2039,7 @@ static int cm_req_handler(struct cm_work *work) > struct cm_req_msg *req_msg; > const struct ib_global_route *grh; > const struct ib_gid_attr *gid_attr; > + struct ib_grh *ibgrh; > int ret; > > req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; > @@ -2048,6 +2049,12 @@ static int cm_req_handler(struct cm_work *work) > if (IS_ERR(cm_id_priv)) > return PTR_ERR(cm_id_priv); > > + ibgrh = work->mad_recv_wc->recv_buf.grh; > + if (!(be32_to_cpu(ibgrh->version_tclass_flow) & IB_GRH_FLOWLABEL_MASK)) > + ibgrh->version_tclass_flow |= > + cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, > + req_msg)); This doesn't seem right. Up until the path is established the response should follow the reversible GMP rules and the flow_label should come out of the request's GRH. Once we established the return data path and the GMP's switch to using the datapath, the flowlabel should be set in something like cm_format_paths_from_req() If you want to switch to using the return data path for REP replies earlier then it should be done completely and not only the flow label. But somehow I suspect we cannot as this could fail too. Jason
