In the scalar datapath, there is a loop to handle multi-segment, and multi-descriptor packets on Tx. After that loop, the end-of-packet bit was written to the descriptor separately, meaning that for each single-descriptor packet there were two writes to the second quad-word - basically 3 x 64-bit writes rather than just 2. Adjusting the code to compute the EOP bit inside the loop saves that extra write per packet and so improves performance.
Signed-off-by: Bruce Richardson <[email protected]> --- drivers/net/intel/common/tx_scalar_fns.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/drivers/net/intel/common/tx_scalar_fns.h b/drivers/net/intel/common/tx_scalar_fns.h index 6079a558e4..7b643fcf44 100644 --- a/drivers/net/intel/common/tx_scalar_fns.h +++ b/drivers/net/intel/common/tx_scalar_fns.h @@ -378,6 +378,10 @@ ci_xmit_pkts(struct ci_tx_queue *txq, txn = &sw_ring[txe->next_id]; } + /* fill the last descriptor with End of Packet (EOP) bit */ + if (m_seg->next == NULL) + td_cmd |= CI_TX_DESC_CMD_EOP; + txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr); txd->cmd_type_offset_bsz = rte_cpu_to_le_64(CI_TX_DESC_DTYPE_DATA | ((uint64_t)td_cmd << CI_TXD_QW1_CMD_S) | @@ -390,21 +394,17 @@ ci_xmit_pkts(struct ci_tx_queue *txq, txe = txn; m_seg = m_seg->next; } while (m_seg); - - /* fill the last descriptor with End of Packet (EOP) bit */ - td_cmd |= CI_TX_DESC_CMD_EOP; txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used); txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used); /* set RS bit on the last descriptor of one packet */ if (txq->nb_tx_used >= txq->tx_rs_thresh) { - td_cmd |= CI_TX_DESC_CMD_RS; + txd->cmd_type_offset_bsz |= + rte_cpu_to_le_64(CI_TX_DESC_CMD_RS << CI_TXD_QW1_CMD_S); /* Update txq RS bit counters */ txq->nb_tx_used = 0; } - txd->cmd_type_offset_bsz |= - rte_cpu_to_le_64(((uint64_t)td_cmd) << CI_TXD_QW1_CMD_S); if (ts_fns != NULL) ts_id = ts_fns->write_ts_desc(txq, tx_pkt, tx_id, ts_id); -- 2.51.0
