static UCS_F_ALWAYS_INLINE void
uct_rc_mlx5_iface_poll_tx(uct_rc_mlx5_iface_t *iface)
{
uct_rc_iface_send_op_t *op;
struct mlx5_cqe64 *cqe;
uct_rc_mlx5_ep_t *ep;
unsigned qp_num;
uint16_t hw_ci;
cqe = uct_ib_mlx5_get_cqe(&iface->tx.cq, UCT_IB_MLX5_CQE64_SIZE_LOG);
if (cqe == NULL) {
return;
}
UCS_STATS_UPDATE_COUNTER(iface->super.stats, UCT_RC_IFACE_STAT_TX_COMPLETION, 1);
ucs_memory_cpu_load_fence();
qp_num = ntohl(cqe->sop_drop_qpn) & UCS_MASK(UCT_IB_QPN_ORDER);
ep = ucs_derived_of(uct_rc_iface_lookup_ep(&iface->super, qp_num), uct_rc_mlx5_ep_t);
ucs_assert(ep != NULL);
hw_ci = ntohs(cqe->wqe_counter);
ep->super.available = uct_ib_mlx5_txwq_update_bb(&ep->tx.wq, hw_ci);
++iface->super.tx.cq_available;
/* Process completions */
ucs_queue_for_each_extract(op, &ep->super.outstanding, queue,
UCS_CIRCULAR_COMPARE16(op->sn, <=, hw_ci)) {
op->handler(op);
}
}
static UCS_F_ALWAYS_INLINE void
uct_ud_mlx5_iface_poll_tx(uct_ud_mlx5_iface_t *iface)
{
struct mlx5_cqe64 *cqe;
cqe = uct_ib_mlx5_get_cqe(&iface->tx.cq, UCT_IB_MLX5_CQE64_SIZE_LOG);
if (cqe == NULL) {
return;
}
uct_ib_mlx5_log_cqe(cqe);
iface->super.tx.available = uct_ib_mlx5_txwq_update_bb(&iface->tx.wq, ntohs(cqe->wqe_counter));
}
static UCS_F_ALWAYS_INLINE void
uct_dc_mlx5_poll_tx(uct_dc_mlx5_iface_t *iface)
{
uint8_t dci;
struct mlx5_cqe64 *cqe;
uint32_t qp_num;
uint16_t hw_ci;
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
cqe = uct_ib_mlx5_poll_cq(&iface->super.super.super, &iface->mlx5_common.tx.cq);
if (cqe == NULL) {
return;
}
UCS_STATS_UPDATE_COUNTER(iface->super.super.stats, UCT_RC_IFACE_STAT_TX_COMPLETION, 1);
ucs_memory_cpu_load_fence();
qp_num = ntohl(cqe->sop_drop_qpn) & UCS_MASK(UCT_IB_QPN_ORDER);
dci = uct_dc_iface_dci_find(&iface->super, qp_num);
txqp = &iface->super.tx.dcis[dci].txqp;
txwq = &iface->dci_wqs[dci];
hw_ci = ntohs(cqe->wqe_counter);
ucs_trace_poll("dc_mlx5 iface %p tx_cqe: dci[%d] qpn 0x%x txqp %p hw_ci %d",
iface, dci, qp_num, txqp, hw_ci);
uct_rc_txqp_available_set(txqp, uct_ib_mlx5_txwq_update_bb(txwq, hw_ci));
uct_dc_iface_dci_put(&iface->super, dci);
uct_rc_mlx5_txqp_process_tx_cqe(txqp, cqe, hw_ci);
iface->super.super.tx.cq_available++;
if (uct_dc_iface_dci_can_alloc(&iface->super)) {
ucs_arbiter_dispatch(uct_dc_iface_dci_waitq(&iface->super), 1,
uct_dc_iface_dci_do_pending_wait, NULL);
}
ucs_arbiter_dispatch(uct_dc_iface_tx_waitq(&iface->super), 1,
uct_dc_iface_dci_do_pending_tx, NULL);
}
static UCS_F_ALWAYS_INLINE void
uct_dc_mlx5_poll_tx(uct_dc_mlx5_iface_t *iface)
{
uint8_t dci;
struct mlx5_cqe64 *cqe;
uint32_t qp_num;
uint16_t hw_ci;
UCT_DC_MLX5_TXQP_DECL(txqp, txwq);
cqe = uct_ib_mlx5_get_cqe(&iface->super.super.super, &iface->mlx5_common.tx.cq,
iface->mlx5_common.tx.cq.cqe_size_log);
if (cqe == NULL) {
return;
}
UCS_STATS_UPDATE_COUNTER(iface->super.super.stats, UCT_RC_IFACE_STAT_TX_COMPLETION, 1);
ucs_memory_cpu_load_fence();
ucs_assertv(!(cqe->op_own & (MLX5_INLINE_SCATTER_32|MLX5_INLINE_SCATTER_64)),
"tx inline scatter not supported");
qp_num = ntohl(cqe->sop_drop_qpn) & UCS_MASK(UCT_IB_QPN_ORDER);
dci = uct_dc_iface_dci_find(&iface->super, qp_num);
txqp = &iface->super.tx.dcis[dci].txqp;
txwq = &iface->dci_wqs[dci];
hw_ci = ntohs(cqe->wqe_counter);
uct_rc_txqp_available_set(txqp, uct_ib_mlx5_txwq_update_bb(txwq, hw_ci));
uct_rc_txqp_completion(txqp, hw_ci);
iface->super.super.tx.cq_available++;
uct_dc_iface_dci_put(&iface->super, dci);
if (uct_dc_iface_dci_can_alloc(&iface->super)) {
ucs_arbiter_dispatch(&iface->super.super.tx.arbiter, 1, uct_dc_iface_dci_do_pending_wait, NULL);
}
ucs_arbiter_dispatch(&iface->super.tx.dci_arbiter, 1, uct_dc_iface_dci_do_pending_tx, NULL);
}
