/**
* @param rx_headroom Headroom requested by the user.
* @param rx_priv_len Length of transport private data to reserve (0 if unused)
* @param rx_hdr_len Length of transport network header.
* @param mss Maximal segment size (transport limit).
*/
UCS_CLASS_INIT_FUNC(uct_ib_iface_t, uct_ib_iface_ops_t *ops, uct_md_h md,
uct_worker_h worker, const uct_iface_params_t *params,
unsigned rx_priv_len, unsigned rx_hdr_len, unsigned tx_cq_len,
size_t mss, const uct_ib_iface_config_t *config)
{
uct_ib_device_t *dev = &ucs_derived_of(md, uct_ib_md_t)->dev;
ucs_status_t status;
uint8_t port_num;
UCS_CLASS_CALL_SUPER_INIT(uct_base_iface_t, &ops->super, md, worker,
&config->super UCS_STATS_ARG(dev->stats));
status = uct_ib_device_find_port(dev, params->dev_name, &port_num);
if (status != UCS_OK) {
goto err;
}
self->ops = ops;
self->config.rx_payload_offset = sizeof(uct_ib_iface_recv_desc_t) +
ucs_max(sizeof(uct_am_recv_desc_t) +
params->rx_headroom,
rx_priv_len + rx_hdr_len);
self->config.rx_hdr_offset = self->config.rx_payload_offset - rx_hdr_len;
self->config.rx_headroom_offset= self->config.rx_payload_offset -
params->rx_headroom;
self->config.seg_size = ucs_min(mss, config->super.max_bcopy);
self->config.tx_max_poll = config->tx.max_poll;
self->config.rx_max_poll = config->rx.max_poll;
self->config.rx_max_batch = ucs_min(config->rx.max_batch,
config->rx.queue_len / 4);
self->config.port_num = port_num;
self->config.sl = config->sl;
self->config.gid_index = config->gid_index;
status = uct_ib_iface_init_pkey(self, config);
if (status != UCS_OK) {
goto err;
}
status = uct_ib_device_query_gid(dev, self->config.port_num,
self->config.gid_index, &self->gid);
if (status != UCS_OK) {
goto err;
}
status = uct_ib_iface_init_lmc(self, config);
if (status != UCS_OK) {
goto err;
}
self->comp_channel = ibv_create_comp_channel(dev->ibv_context);
if (self->comp_channel == NULL) {
ucs_error("ibv_create_comp_channel() failed: %m");
status = UCS_ERR_IO_ERROR;
goto err_free_path_bits;
}
status = ucs_sys_fcntl_modfl(self->comp_channel->fd, O_NONBLOCK, 0);
if (status != UCS_OK) {
goto err_destroy_comp_channel;
}
status = uct_ib_iface_create_cq(self, tx_cq_len, 0, &self->send_cq);
if (status != UCS_OK) {
goto err_destroy_comp_channel;
}
status = uct_ib_iface_create_cq(self, config->rx.queue_len, config->rx.inl,
&self->recv_cq);
if (status != UCS_OK) {
goto err_destroy_send_cq;
}
/* Address scope and size */
if (config->addr_type == UCT_IB_IFACE_ADDRESS_TYPE_AUTO) {
if (IBV_PORT_IS_LINK_LAYER_ETHERNET(uct_ib_iface_port_attr(self))) {
self->addr_type = UCT_IB_ADDRESS_TYPE_ETH;
} else {
self->addr_type = uct_ib_address_scope(self->gid.global.subnet_prefix);
}
} else {
ucs_assert(config->addr_type < UCT_IB_ADDRESS_TYPE_LAST);
self->addr_type = config->addr_type;
}
self->addr_size = uct_ib_address_size(self->addr_type);
ucs_debug("created uct_ib_iface_t headroom_ofs %d payload_ofs %d hdr_ofs %d data_sz %d",
self->config.rx_headroom_offset, self->config.rx_payload_offset,
self->config.rx_hdr_offset, self->config.seg_size);
return UCS_OK;
err_destroy_send_cq:
ibv_destroy_cq(self->send_cq);
err_destroy_comp_channel:
ibv_destroy_comp_channel(self->comp_channel);
err_free_path_bits:
ucs_free(self->path_bits);
err:
return status;
}
static ucs_status_t uct_ib_md_umr_qp_create(uct_ib_md_t *md)
{
#if HAVE_EXP_UMR
struct ibv_exp_qp_init_attr qp_init_attr;
struct ibv_qp_attr qp_attr;
uint8_t port_num;
int ret;
uct_ib_device_t *ibdev;
struct ibv_exp_port_attr *port_attr;
ibdev = &md->dev;
if (!(ibdev->dev_attr.exp_device_cap_flags & IBV_EXP_DEVICE_UMR)) {
return UCS_ERR_UNSUPPORTED;
}
/* TODO: fix port selection. It looks like active port should be used */
port_num = ibdev->first_port;
port_attr = uct_ib_device_port_attr(ibdev, port_num);
memset(&qp_init_attr, 0, sizeof(qp_init_attr));
md->umr_cq = ibv_create_cq(ibdev->ibv_context, 1, NULL, NULL, 0);
if (md->umr_cq == NULL) {
ucs_error("failed to create UMR CQ: %m");
goto err;
}
qp_init_attr.qp_type = IBV_QPT_RC;
qp_init_attr.send_cq = md->umr_cq;
qp_init_attr.recv_cq = md->umr_cq;
qp_init_attr.cap.max_inline_data = 0;
qp_init_attr.cap.max_recv_sge = 1;
qp_init_attr.cap.max_send_sge = 1;
qp_init_attr.srq = NULL;
qp_init_attr.cap.max_recv_wr = 16;
qp_init_attr.cap.max_send_wr = 16;
qp_init_attr.pd = md->pd;
qp_init_attr.comp_mask = IBV_EXP_QP_INIT_ATTR_PD|IBV_EXP_QP_INIT_ATTR_MAX_INL_KLMS;
qp_init_attr.max_inl_recv = 0;
#if (HAVE_IBV_EXP_QP_CREATE_UMR_CAPS || HAVE_EXP_UMR_NEW_API)
qp_init_attr.max_inl_send_klms = ibdev->dev_attr.umr_caps.max_send_wqe_inline_klms;
#else
qp_init_attr.max_inl_send_klms = ibdev->dev_attr.max_send_wqe_inline_klms;
#endif
#if HAVE_IBV_EXP_QP_CREATE_UMR
qp_init_attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_CREATE_FLAGS;
qp_init_attr.exp_create_flags = IBV_EXP_QP_CREATE_UMR;
#endif
md->umr_qp = ibv_exp_create_qp(ibdev->ibv_context, &qp_init_attr);
if (md->umr_qp == NULL) {
ucs_error("failed to create UMR QP: %m");
goto err_destroy_cq;
}
memset(&qp_attr, 0, sizeof(qp_attr));
/* Modify QP to INIT state */
qp_attr.qp_state = IBV_QPS_INIT;
qp_attr.pkey_index = 0;
qp_attr.port_num = port_num;
qp_attr.qp_access_flags = UCT_IB_MEM_ACCESS_FLAGS;
ret = ibv_modify_qp(md->umr_qp, &qp_attr,
IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT | IBV_QP_ACCESS_FLAGS);
if (ret) {
ucs_error("Failed to modify UMR QP to INIT: %m");
goto err_destroy_qp;
}
/* Modify to RTR */
qp_attr.qp_state = IBV_QPS_RTR;
qp_attr.dest_qp_num = md->umr_qp->qp_num;
memset(&qp_attr.ah_attr, 0, sizeof(qp_attr.ah_attr));
qp_attr.ah_attr.port_num = port_num;
qp_attr.ah_attr.dlid = port_attr->lid;
qp_attr.ah_attr.is_global = 1;
if (uct_ib_device_query_gid(ibdev, port_num, 0, &qp_attr.ah_attr.grh.dgid) != UCS_OK) {
goto err_destroy_qp;
}
qp_attr.rq_psn = 0;
qp_attr.path_mtu = IBV_MTU_512;
qp_attr.min_rnr_timer = 7;
qp_attr.max_dest_rd_atomic = 1;
ret = ibv_modify_qp(md->umr_qp, &qp_attr,
IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU | IBV_QP_DEST_QPN |
IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC | IBV_QP_MIN_RNR_TIMER);
if (ret) {
ucs_error("Failed to modify UMR QP to RTR: %m");
goto err_destroy_qp;
}
/* Modify to RTS */
qp_attr.qp_state = IBV_QPS_RTS;
qp_attr.sq_psn = 0;
qp_attr.timeout = 7;
qp_attr.rnr_retry = 7;
qp_attr.retry_cnt = 7;
qp_attr.max_rd_atomic = 1;
ret = ibv_modify_qp(md->umr_qp, &qp_attr,
IBV_QP_STATE | IBV_QP_TIMEOUT | IBV_QP_RETRY_CNT |
IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY | IBV_QP_SQ_PSN |
IBV_QP_MAX_QP_RD_ATOMIC);
if (ret) {
ucs_error("Failed to modify UMR QP to RTS: %m");
goto err_destroy_qp;
}
return UCS_OK;
err_destroy_qp:
ibv_destroy_qp(md->umr_qp);
err_destroy_cq:
ibv_destroy_cq(md->umr_cq);
err:
return UCS_ERR_IO_ERROR;
#else
return UCS_ERR_UNSUPPORTED;
#endif
}
