UCS_CLASS_INIT_FUNC(uct_dc_ep_t, uct_dc_iface_t *iface, const uct_dc_iface_addr_t *if_addr) { UCS_CLASS_CALL_SUPER_INIT(uct_base_ep_t, &iface->super.super.super); self->atomic_mr_offset = uct_ib_md_atomic_offset(if_addr->atomic_mr_id); return uct_dc_ep_basic_init(iface, self); }
ucs_status_t uct_rc_verbs_ep_connect_to_ep(uct_ep_h tl_ep, const uct_device_addr_t *dev_addr, const uct_ep_addr_t *ep_addr) { uct_rc_verbs_ep_t *ep = ucs_derived_of(tl_ep, uct_rc_verbs_ep_t); uct_rc_verbs_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_rc_verbs_iface_t); const uct_ib_address_t *ib_addr = (const uct_ib_address_t *)dev_addr; const uct_rc_ep_address_t *rc_addr = (const uct_rc_ep_address_t*)ep_addr; struct ibv_ah_attr ah_attr; ucs_status_t status; uint32_t qp_num; uct_ib_iface_fill_ah_attr(&iface->super.super, ib_addr, ep->super.path_bits, &ah_attr); #if HAVE_IBV_EX_HW_TM if(UCT_RC_VERBS_TM_ENABLED(iface)) { /* For HW TM we need 2 QPs, one of which will be used by the device for * RNDV offload (for issuing RDMA reads and sending RNDV ACK). No WQEs * should be posted to the send side of the QP which is owned by device. */ uct_rc_verbs_ep_tm_address_t *tm_addr = ucs_derived_of(rc_addr, uct_rc_verbs_ep_tm_address_t); ucs_assert_always(tm_addr->super.type == UCT_RC_EP_ADDR_TYPE_TM); status = uct_rc_iface_qp_connect(&iface->super, ep->tm_qp, uct_ib_unpack_uint24(rc_addr->qp_num), &ah_attr); if (status != UCS_OK) { return status; } /* Need to connect local ep QP to the one owned by device * (and bound to XRQ) on the peer. */ qp_num = uct_ib_unpack_uint24(tm_addr->tm_qp_num); } else #endif { ucs_assert_always(rc_addr->type == UCT_RC_EP_ADDR_TYPE_BASIC); qp_num = uct_ib_unpack_uint24(rc_addr->qp_num); } status = uct_rc_iface_qp_connect(&iface->super, ep->super.txqp.qp, qp_num, &ah_attr); if (status != UCS_OK) { return status; } ep->super.atomic_mr_offset = uct_ib_md_atomic_offset(rc_addr->atomic_mr_id); return UCS_OK; }
ucs_status_t uct_dc_mlx5_ep_fc_ctrl(uct_ep_t *tl_ep, unsigned op, uct_rc_fc_request_t *req) { uct_dc_mlx5_ep_t *dc_ep = ucs_derived_of(tl_ep, uct_dc_mlx5_ep_t); uct_dc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_dc_mlx5_iface_t); uct_ib_iface_t *ib_iface = &iface->super.super.super; struct ibv_ah_attr ah_attr = {.is_global = 0}; uct_dc_fc_sender_data_t sender; uct_dc_fc_request_t *dc_req; struct mlx5_wqe_av mlx5_av; uct_ib_mlx5_base_av_t av; ucs_status_t status; uintptr_t sender_ep; struct ibv_ah *ah; UCT_DC_MLX5_TXQP_DECL(txqp, txwq); ucs_assert((sizeof(uint8_t) + sizeof(sender_ep)) <= UCT_IB_MLX5_AV_FULL_SIZE); UCT_DC_MLX5_CHECK_RES(iface, dc_ep); UCT_DC_MLX5_IFACE_TXQP_GET(iface, dc_ep, txqp, txwq); dc_req = ucs_derived_of(req, uct_dc_fc_request_t); if (op == UCT_RC_EP_FC_PURE_GRANT) { ucs_assert(req != NULL); sender_ep = (uintptr_t)dc_req->sender.ep; /* TODO: look at common code with uct_ud_mlx5_iface_get_av */ if (dc_req->sender.global.is_global) { uct_ib_iface_fill_ah_attr_from_gid_lid(ib_iface, dc_req->lid, ucs_unaligned_ptr(&dc_req->sender.global.gid), ib_iface->path_bits[0], &ah_attr); status = uct_ib_iface_create_ah(ib_iface, &ah_attr, &ah); if (status != UCS_OK) { return status; } uct_ib_mlx5_get_av(ah, &mlx5_av); } /* Note av initialization is copied from exp verbs */ av.stat_rate_sl = ib_iface->config.sl; /* (attr->static_rate << 4) | attr->sl */ av.fl_mlid = ib_iface->path_bits[0] & 0x7f; /* lid in dc_req is in BE already */ av.rlid = uct_ib_iface_is_roce(ib_iface) ? 0 : (dc_req->lid | htons(ib_iface->path_bits[0])); av.dqp_dct = htonl(dc_req->dct_num); uct_dc_mlx5_iface_set_av_sport(iface, &av, dc_req->dct_num); if (!iface->ud_common.config.compact_av || ah_attr.is_global) { av.dqp_dct |= UCT_IB_MLX5_EXTENDED_UD_AV; } uct_rc_mlx5_txqp_inline_post(&iface->super, UCT_IB_QPT_DCI, txqp, txwq, MLX5_OPCODE_SEND, &av /*dummy*/, 0, op, sender_ep, 0, 0, 0, &av, ah_attr.is_global ? mlx5_av_grh(&mlx5_av) : NULL, uct_ib_mlx5_wqe_av_size(&av), 0, INT_MAX); } else { ucs_assert(op == UCT_RC_EP_FC_FLAG_HARD_REQ); sender.ep = (uint64_t)dc_ep; sender.global.gid = ib_iface->gid; sender.global.is_global = dc_ep->flags & UCT_DC_MLX5_EP_FLAG_GRH; UCS_STATS_UPDATE_COUNTER(dc_ep->fc.stats, UCT_RC_FC_STAT_TX_HARD_REQ, 1); uct_rc_mlx5_txqp_inline_post(&iface->super, UCT_IB_QPT_DCI, txqp, txwq, MLX5_OPCODE_SEND_IMM, &sender.global, sizeof(sender.global), op, sender.ep, uct_dc_mlx5_get_dct_num(iface), 0, 0, &dc_ep->av, uct_dc_mlx5_ep_get_grh(dc_ep), uct_ib_mlx5_wqe_av_size(&dc_ep->av), MLX5_WQE_CTRL_SOLICITED, INT_MAX); } return UCS_OK; } UCS_CLASS_INIT_FUNC(uct_dc_mlx5_ep_t, uct_dc_mlx5_iface_t *iface, const uct_dc_mlx5_iface_addr_t *if_addr, uct_ib_mlx5_base_av_t *av) { uint32_t remote_dctn; ucs_trace_func(""); UCS_CLASS_CALL_SUPER_INIT(uct_base_ep_t, &iface->super.super.super.super); self->atomic_mr_offset = uct_ib_md_atomic_offset(if_addr->atomic_mr_id); remote_dctn = uct_ib_unpack_uint24(if_addr->qp_num); memcpy(&self->av, av, sizeof(*av)); self->av.dqp_dct |= htonl(remote_dctn); uct_dc_mlx5_iface_set_av_sport(iface, &self->av, remote_dctn); return uct_dc_mlx5_ep_basic_init(iface, self); }