static ucs_status_t UCS_F_ALWAYS_INLINE
uct_dc_mlx5_ep_put_short_inline(uct_ep_h tl_ep, const void *buffer,
unsigned length, uint64_t remote_addr,
uct_rkey_t rkey)
{
uct_dc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_dc_mlx5_iface_t);
uct_dc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_dc_mlx5_ep_t);
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
UCT_RC_MLX5_CHECK_PUT_SHORT(length, UCT_IB_MLX5_AV_FULL_SIZE);
UCT_DC_MLX5_CHECK_RES(iface, ep);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq);
uct_rc_mlx5_txqp_inline_post(&iface->super, UCT_IB_QPT_DCI,
txqp, txwq,
MLX5_OPCODE_RDMA_WRITE,
buffer, length, 0, 0, 0,
remote_addr, uct_ib_md_direct_rkey(rkey),
&ep->av, uct_dc_mlx5_ep_get_grh(ep),
uct_ib_mlx5_wqe_av_size(&ep->av), 0, INT_MAX);
UCT_TL_EP_STAT_OP(&ep->super, PUT, SHORT, length);
return UCS_OK;
}
ucs_status_t uct_rc_mlx5_ep_flush(uct_ep_h tl_ep, unsigned flags,
uct_completion_t *comp)
{
uct_rc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_rc_mlx5_ep_t);
uct_rc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_rc_mlx5_iface_t);
uint16_t sn;
if (!uct_rc_iface_has_tx_resources(&iface->super)) {
return UCS_ERR_NO_RESOURCE;
}
if (uct_rc_txqp_available(&ep->super.txqp) == ep->tx.wq.bb_max) {
UCT_TL_EP_STAT_FLUSH(&ep->super.super);
return UCS_OK;
}
if (uct_rc_txqp_unsignaled(&ep->super.txqp) != 0) {
sn = ep->tx.wq.sw_pi;
UCT_RC_CHECK_RES(&iface->super, &ep->super);
uct_rc_mlx5_txqp_inline_post(&iface->super, IBV_QPT_RC,
&ep->super.txqp, &ep->tx.wq,
MLX5_OPCODE_NOP, NULL, 0,
0, 0,
0, 0,
NULL, 0);
} else if (!uct_rc_ep_has_tx_resources(&ep->super)) {
return UCS_ERR_NO_RESOURCE;
} else {
sn = ep->tx.wq.sig_pi;
}
uct_rc_txqp_add_send_comp(&iface->super, &ep->super.txqp, comp, sn);
UCT_TL_EP_STAT_FLUSH_WAIT(&ep->super.super);
return UCS_INPROGRESS;
}
static ucs_status_t UCS_F_ALWAYS_INLINE
uct_dc_mlx5_ep_am_short_inline(uct_ep_h tl_ep, uint8_t id, uint64_t hdr,
const void *buffer, unsigned length)
{
uct_dc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_dc_mlx5_iface_t);
uct_dc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_dc_mlx5_ep_t);
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
UCT_RC_MLX5_CHECK_AM_SHORT(id, length, UCT_IB_MLX5_AV_FULL_SIZE);
UCT_DC_CHECK_RES_AND_FC(iface, ep);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq);
uct_rc_mlx5_txqp_inline_post(&iface->super, UCT_IB_QPT_DCI,
txqp, txwq,
MLX5_OPCODE_SEND,
buffer, length, id, hdr, 0,
0, 0,
&ep->av, uct_dc_mlx5_ep_get_grh(ep),
uct_ib_mlx5_wqe_av_size(&ep->av),
MLX5_WQE_CTRL_SOLICITED, INT_MAX);
UCT_RC_UPDATE_FC_WND(&iface->super.super, &ep->fc);
UCT_TL_EP_STAT_OP(&ep->super, AM, SHORT, sizeof(hdr) + length);
return UCS_OK;
}
ucs_status_t uct_rc_mlx5_ep_fc_ctrl(uct_rc_ep_t *rc_ep)
{
uct_rc_mlx5_ep_t *ep = ucs_derived_of(rc_ep, uct_rc_mlx5_ep_t);
uct_rc_iface_t *iface = ucs_derived_of(rc_ep->super.super.iface, uct_rc_iface_t);
UCT_RC_CHECK_RES(iface, &ep->super);
uct_rc_mlx5_txqp_inline_post(iface, IBV_QPT_RC,
&ep->super.txqp, &ep->tx.wq,
MLX5_OPCODE_SEND|UCT_RC_MLX5_OPCODE_FLAG_RAW,
NULL, 0,
UCT_RC_EP_FC_PURE_GRANT, 0 ,
0, 0,
NULL, 0);
UCT_TL_EP_STAT_OP(&ep->super.super, AM, SHORT, 0);
return UCS_OK;
}
ucs_status_t uct_rc_mlx5_ep_put_short(uct_ep_h tl_ep, const void *buffer, unsigned length,
uint64_t remote_addr, uct_rkey_t rkey)
{
uct_rc_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_rc_iface_t);
uct_rc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_rc_mlx5_ep_t);
UCT_RC_MLX5_CHECK_PUT_SHORT(length, 0);
UCT_RC_CHECK_RES(iface, &ep->super);
uct_rc_mlx5_txqp_inline_post(iface, IBV_QPT_RC,
&ep->super.txqp, &ep->tx.wq,
MLX5_OPCODE_RDMA_WRITE,
buffer, length, 0, 0,
remote_addr, uct_ib_md_direct_rkey(rkey),
NULL, 0);
UCT_TL_EP_STAT_OP(&ep->super.super, PUT, SHORT, length);
return UCS_OK;
}
ucs_status_t uct_dc_mlx5_ep_am_short(uct_ep_h tl_ep, uint8_t id, uint64_t hdr,
const void *buffer, unsigned length)
{
uct_dc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_dc_mlx5_iface_t);
uct_dc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_dc_mlx5_ep_t);
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
UCT_RC_MLX5_CHECK_AM_SHORT(id, length, UCT_IB_MLX5_AV_FULL_SIZE);
UCT_DC_CHECK_RES(&iface->super, &ep->super);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq);
uct_rc_mlx5_txqp_inline_post(&iface->super.super, IBV_EXP_QPT_DC_INI,
txqp, txwq,
MLX5_OPCODE_SEND,
buffer, length, id, hdr,
0, 0,
&ep->av, uct_ib_mlx5_wqe_av_size(&ep->av));
UCT_TL_EP_STAT_OP(&ep->super.super, AM, SHORT, sizeof(hdr) + length);
return UCS_OK;
}
ucs_status_t uct_rc_mlx5_ep_am_short(uct_ep_h tl_ep, uint8_t id, uint64_t hdr,
const void *payload, unsigned length)
{
uct_rc_iface_t *iface = ucs_derived_of(tl_ep->iface, uct_rc_iface_t);
uct_rc_mlx5_ep_t *ep = ucs_derived_of(tl_ep, uct_rc_mlx5_ep_t);
UCT_RC_MLX5_CHECK_AM_SHORT(id, length, 0);
UCT_RC_CHECK_RES(iface, &ep->super);
UCT_RC_CHECK_FC_WND(iface, &ep->super, id);
uct_rc_mlx5_txqp_inline_post(iface, IBV_QPT_RC,
&ep->super.txqp, &ep->tx.wq,
MLX5_OPCODE_SEND,
payload, length,
id, hdr,
0, 0,
NULL, 0);
UCT_TL_EP_STAT_OP(&ep->super.super, AM, SHORT, sizeof(hdr) + length);
UCT_RC_UPDATE_FC_WND(iface, &ep->super, 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);
}
ucs_status_t uct_dc_mlx5_ep_fc_ctrl(uct_ep_t *tl_ep, unsigned op,
uct_rc_fc_request_t *req)
{
uintptr_t sender_ep;
uct_ib_iface_t *ib_iface;
uct_ib_mlx5_base_av_t av;
uct_dc_fc_request_t *dc_req;
uct_dc_mlx5_ep_t *dc_mlx5_ep;
uct_dc_ep_t *dc_ep = ucs_derived_of(tl_ep, uct_dc_ep_t);
uct_dc_mlx5_iface_t *iface = ucs_derived_of(tl_ep->iface,
uct_dc_mlx5_iface_t);
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
ucs_assert((sizeof(uint8_t) + sizeof(sender_ep)) <=
UCT_IB_MLX5_AV_FULL_SIZE);
UCT_DC_CHECK_RES(&iface->super, dc_ep);
UCT_DC_MLX5_IFACE_TXQP_GET(iface, dc_ep, txqp, txwq);
if (op == UCT_RC_EP_FC_PURE_GRANT) {
ucs_assert(req != NULL);
dc_req = ucs_derived_of(req, uct_dc_fc_request_t);
sender_ep = (uintptr_t)dc_req->sender_ep;
ib_iface = &iface->super.super.super;
/* 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 = dc_req->lid | htons(ib_iface->path_bits[0]);
av.dqp_dct = htonl(dc_req->dct_num);
if (!iface->ud_common.config.compact_av) {
av.dqp_dct |= UCT_IB_MLX5_EXTENDED_UD_AV;
}
uct_rc_mlx5_txqp_inline_post(&iface->super.super, IBV_EXP_QPT_DC_INI,
txqp, txwq, MLX5_OPCODE_SEND,
NULL, 0, op, sender_ep, 0,
0, 0,
&av, uct_ib_mlx5_wqe_av_size(&av));
} else {
ucs_assert(op == UCT_RC_EP_FC_FLAG_HARD_REQ);
sender_ep = (uintptr_t)dc_ep;
dc_mlx5_ep = ucs_derived_of(tl_ep, uct_dc_mlx5_ep_t);
UCS_STATS_UPDATE_COUNTER(dc_ep->fc.stats,
UCT_RC_FC_STAT_TX_HARD_REQ, 1);
uct_rc_mlx5_txqp_inline_post(&iface->super.super, IBV_EXP_QPT_DC_INI,
txqp, txwq, MLX5_OPCODE_SEND_IMM,
NULL, 0, op, sender_ep,
iface->super.rx.dct->dct_num,
0, 0,
&dc_mlx5_ep->av,
uct_ib_mlx5_wqe_av_size(&dc_mlx5_ep->av));
}
return UCS_OK;
}
