/**
* @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 char *dev_name, unsigned rx_headroom,
unsigned rx_priv_len, unsigned rx_hdr_len, unsigned tx_cq_len,
size_t mss, 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, dev_name, &port_num);
if (status != UCS_OK) {
goto err;
}
self->port_num = port_num;
self->sl = config->sl;
self->config.rx_payload_offset = sizeof(uct_ib_iface_recv_desc_t) +
ucs_max(sizeof(uct_am_recv_desc_t) + 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 - 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->ops = ops;
status = uct_ib_iface_init_pkey(self, config);
if (status != UCS_OK) {
goto err;
}
status = uct_ib_iface_init_gid(self, config);
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("Failed to create completion channel: %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;
}
/* TODO inline scatter for send SQ */
self->send_cq = ibv_create_cq(dev->ibv_context, tx_cq_len,
NULL, self->comp_channel, 0);
if (self->send_cq == NULL) {
ucs_error("Failed to create send cq: %m");
status = UCS_ERR_IO_ERROR;
goto err_destroy_comp_channel;
}
if (config->rx.inl > 32 /*UCT_IB_MLX5_CQE64_MAX_INL*/) {
ibv_exp_setenv(dev->ibv_context, "MLX5_CQE_SIZE", "128", 1);
}
self->recv_cq = ibv_create_cq(dev->ibv_context, config->rx.queue_len,
NULL, self->comp_channel, 0);
ibv_exp_setenv(dev->ibv_context, "MLX5_CQE_SIZE", "64", 1);
if (self->recv_cq == NULL) {
ucs_error("Failed to create recv cq: %m");
status = UCS_ERR_IO_ERROR;
goto err_destroy_send_cq;
}
if (!uct_ib_device_is_port_ib(dev, self->port_num)) {
ucs_error("Unsupported link layer");
status = UCS_ERR_UNSUPPORTED;
goto err_destroy_recv_cq;
}
/* Address scope and size */
self->addr_scope = uct_ib_address_scope(self->gid.global.subnet_prefix);
self->addr_size = uct_ib_address_size(self->addr_scope);
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_recv_cq:
ibv_destroy_cq(self->recv_cq);
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;
}
ucs_status_t uct_ib_iface_query(uct_ib_iface_t *iface, size_t xport_hdr_len,
uct_iface_attr_t *iface_attr)
{
uct_ib_device_t *dev = uct_ib_iface_device(iface);
static const unsigned ib_port_widths[] = {
[0] = 1,
[1] = 4,
[2] = 8,
[3] = 12
};
uint8_t active_width, active_speed, active_mtu;
double encoding, signal_rate, wire_speed;
size_t mtu, width, extra_pkt_len;
if (!uct_ib_device_is_port_ib(dev, iface->port_num)) {
return UCS_ERR_UNSUPPORTED;
}
active_width = uct_ib_iface_port_attr(iface)->active_width;
active_speed = uct_ib_iface_port_attr(iface)->active_speed;
active_mtu = uct_ib_iface_port_attr(iface)->active_mtu;
/* Get active width */
if (!ucs_is_pow2(active_width) ||
(active_width < 1) || (ucs_ilog2(active_width) > 3))
{
ucs_error("Invalid active_width on %s:%d: %d",
uct_ib_device_name(dev), iface->port_num, active_width);
return UCS_ERR_IO_ERROR;
}
memset(iface_attr, 0, sizeof(*iface_attr));
iface_attr->device_addr_len = iface->addr_size;
switch (active_speed) {
case 1: /* SDR */
iface_attr->latency = 5000e-9;
signal_rate = 2.5e9;
encoding = 8.0/10.0;
break;
case 2: /* DDR */
iface_attr->latency = 2500e-9;
signal_rate = 5.0e9;
encoding = 8.0/10.0;
break;
case 4: /* QDR */
iface_attr->latency = 1300e-9;
signal_rate = 10.0e9;
encoding = 8.0/10.0;
break;
case 8: /* FDR10 */
iface_attr->latency = 700e-9;
signal_rate = 10.3125e9;
encoding = 64.0/66.0;
break;
case 16: /* FDR */
iface_attr->latency = 700e-9;
signal_rate = 14.0625e9;
encoding = 64.0/66.0;
break;
case 32: /* EDR */
iface_attr->latency = 600e-9;
signal_rate = 25.78125e9;
encoding = 64.0/66.0;
break;
default:
ucs_error("Invalid active_speed on %s:%d: %d",
uct_ib_device_name(dev), iface->port_num, active_speed);
return UCS_ERR_IO_ERROR;
}
/* Wire speed calculation: Width * SignalRate * Encoding */
width = ib_port_widths[ucs_ilog2(active_width)];
wire_speed = (width * signal_rate * encoding) / 8.0;
/* Calculate packet overhead */
mtu = ucs_min(uct_ib_mtu_value(active_mtu),
iface->config.seg_size);
extra_pkt_len = UCT_IB_LRH_LEN + UCT_IB_BTH_LEN + xport_hdr_len +
UCT_IB_ICRC_LEN + UCT_IB_VCRC_LEN + UCT_IB_DELIM_LEN;
iface_attr->bandwidth = (wire_speed * mtu) / (mtu + extra_pkt_len);
return UCS_OK;
}
/**
* @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_iface_ops_t *ops, uct_pd_h pd,
uct_worker_h worker, const char *dev_name, unsigned rx_headroom,
unsigned rx_priv_len, unsigned rx_hdr_len, unsigned tx_cq_len,
size_t mss, uct_ib_iface_config_t *config)
{
uct_ib_device_t *dev = &ucs_derived_of(pd, uct_ib_pd_t)->dev;
ucs_status_t status;
uint8_t port_num;
UCS_CLASS_CALL_SUPER_INIT(uct_base_iface_t, ops, pd, worker,
&config->super UCS_STATS_ARG(dev->stats));
status = uct_ib_iface_find_port(dev, dev_name, &port_num);
if (status != UCS_OK) {
ucs_error("Failed to find port %s: %s", dev_name, ucs_status_string(status));
goto err;
}
self->port_num = port_num;
self->sl = config->sl;
self->config.rx_payload_offset = sizeof(uct_ib_iface_recv_desc_t) +
ucs_max(sizeof(uct_am_recv_desc_t) + 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 - rx_headroom;
self->config.seg_size = config->super.max_bcopy;
status = uct_ib_iface_init_pkey(self, config);
if (status != UCS_OK) {
goto err;
}
status = uct_ib_iface_init_gid(self, config);
if (status != UCS_OK) {
goto err;
}
status = uct_ib_iface_init_lmc(self, config);
if (status != UCS_OK) {
goto err;
}
/* TODO comp_channel */
/* TODO inline scatter for send SQ */
self->send_cq = ibv_create_cq(dev->ibv_context, tx_cq_len, NULL, NULL, 0);
if (self->send_cq == NULL) {
ucs_error("Failed to create send cq: %m");
status = UCS_ERR_IO_ERROR;
goto err_free_path_bits;
}
if (config->rx.inl > 32 /*UCT_IB_MLX5_CQE64_MAX_INL*/) {
ibv_exp_setenv(dev->ibv_context, "MLX5_CQE_SIZE", "128", 1);
}
self->recv_cq = ibv_create_cq(dev->ibv_context, config->rx.queue_len,
NULL, NULL, 0);
ibv_exp_setenv(dev->ibv_context, "MLX5_CQE_SIZE", "64", 1);
if (self->recv_cq == NULL) {
ucs_error("Failed to create recv cq: %m");
status = UCS_ERR_IO_ERROR;
goto err_destroy_send_cq;
}
if (!uct_ib_device_is_port_ib(dev, self->port_num)) {
ucs_error("Unsupported link layer");
status = UCS_ERR_UNSUPPORTED;
goto err_destroy_recv_cq;
}
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_recv_cq:
ibv_destroy_cq(self->recv_cq);
err_destroy_send_cq:
ibv_destroy_cq(self->send_cq);
err_free_path_bits:
ucs_free(self->path_bits);
err:
return status;
}
