From: Jun Yang <[email protected]> Add a public API to transmit an array of packets whose destination port and Tx queue are given per packet, and clean up the ordered Tx helpers it builds on.
- New API rte_dpaa2_dev_tx_multi_ports(): takes per-packet port_id[] and txq_id[] arrays (txq_id NULL means queue 0 of each port), resolves each dpaa2_queue, and forwards to dpaa2_dev_tx_multi_txq_ordered(). This allows a single burst to span multiple ports. - dpaa2_dev_tx_multi_txq_ordered() takes struct dpaa2_queue ** instead of void ** and no longer loops with tx_again or accumulates a running "sent" count; it enqueues one batch and returns the number sent. Tx confirmation is now done after enqueue only for queues that have a tx_conf_queue, and the buffer-pool lookup is deferred to the slow path. - The dy_conf out-parameter is dropped from the mbuf-to-fd helpers (dpaa2_dev_tx_mbuf_to_sg_fd, dpaa2_dev_tx_mbuf_to_simple_fd, dpaa2_dev_tx_fast_mbuf_to_fd); it was written but never consumed. dpaa2_dev_tx_fast_mbuf_to_fd() now takes the private structure directly and reads tx_conf_type from it. Signed-off-by: Jun Yang <[email protected]> --- drivers/event/dpaa2/dpaa2_eventdev.c | 2 +- drivers/net/dpaa2/dpaa2_ethdev.h | 4 +- drivers/net/dpaa2/dpaa2_rxtx.c | 127 ++++++++++++--------------- drivers/net/dpaa2/rte_pmd_dpaa2.h | 5 ++ 4 files changed, 66 insertions(+), 72 deletions(-) diff --git a/drivers/event/dpaa2/dpaa2_eventdev.c b/drivers/event/dpaa2/dpaa2_eventdev.c index 476decbbdc..be6d8bc3da 100644 --- a/drivers/event/dpaa2/dpaa2_eventdev.c +++ b/drivers/event/dpaa2/dpaa2_eventdev.c @@ -1095,7 +1095,7 @@ static uint16_t dpaa2_eventdev_txa_enqueue(void *port, struct rte_event ev[], uint16_t nb_events) { - void *txq[MAX_TX_RING_SLOTS]; + struct dpaa2_queue *txq[MAX_TX_RING_SLOTS]; struct rte_mbuf *m[MAX_TX_RING_SLOTS]; uint16_t sent = 0, burst, i, port_id, qid; diff --git a/drivers/net/dpaa2/dpaa2_ethdev.h b/drivers/net/dpaa2/dpaa2_ethdev.h index 8c0be81c95..e4dc7b7e4f 100644 --- a/drivers/net/dpaa2/dpaa2_ethdev.h +++ b/drivers/net/dpaa2/dpaa2_ethdev.h @@ -975,8 +975,8 @@ uint16_t dpaa2_dev_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts); uint16_t dpaa2_dev_tx_ordered(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts); __rte_internal -uint16_t dpaa2_dev_tx_multi_txq_ordered(void **queue, - struct rte_mbuf **bufs, uint16_t nb_pkts); +uint16_t dpaa2_dev_tx_multi_txq_ordered(struct dpaa2_queue **dpaa2_q, + struct rte_mbuf **bufs, uint16_t nb_pkts); void dpaa2_dev_free_eqresp_buf(uint16_t eqresp_ci, struct dpaa2_queue *dpaa2_q); void dpaa2_flow_clean(struct rte_eth_dev *dev, uint8_t tc_id, diff --git a/drivers/net/dpaa2/dpaa2_rxtx.c b/drivers/net/dpaa2/dpaa2_rxtx.c index 0d4fc18455..a44b991d1b 100644 --- a/drivers/net/dpaa2/dpaa2_rxtx.c +++ b/drivers/net/dpaa2/dpaa2_rxtx.c @@ -738,8 +738,7 @@ dpaa2_dev_tx_conf_mbuf_to_sge(struct qbman_sge *sgt, static int __rte_noinline __rte_hot dpaa2_dev_tx_mbuf_to_sg_fd(struct rte_mempool *hw_mp, struct rte_mbuf *mbuf, struct qbman_fd *fd, - struct dpaa2_queue *txq, enum dpaa2_tx_conf_type conf, - uint8_t *dy_conf, int tstamp) + struct dpaa2_queue *txq, enum dpaa2_tx_conf_type conf, int tstamp) { struct rte_mbuf *cur_seg = mbuf, *sg_mbuf; struct qbman_sge *sgt; @@ -841,8 +840,6 @@ dpaa2_dev_tx_mbuf_to_sg_fd(struct rte_mempool *hw_mp, if (ret) return ret; } - if (dy_conf && need_dyconf) - *dy_conf = true; return 0; } @@ -867,8 +864,7 @@ dpaa2_dev_prefetch_next_psr(const struct qbman_result *dq) static int __rte_noinline __rte_hot dpaa2_dev_tx_mbuf_to_simple_fd(struct rte_mempool *hw_mp, struct rte_mbuf *mbuf, struct qbman_fd *fd, - struct dpaa2_queue *txq, enum dpaa2_tx_conf_type conf, - uint8_t *dy_conf, int tstamp) + struct dpaa2_queue *txq, enum dpaa2_tx_conf_type conf, int tstamp) { int need_dyconf = false, need_conf = false; struct rte_mbuf *mi; @@ -905,13 +901,10 @@ dpaa2_dev_tx_mbuf_to_simple_fd(struct rte_mempool *hw_mp, rte_pktmbuf_free(mbuf); } - if (need_dyconf) { + if (need_dyconf) dpaa2_dev_tx_config_dynamic_confirm(fd, txq, true, tstamp); - if (dy_conf) - *dy_conf = true; - } else if (tstamp) { + else if (tstamp) dpaa2_dev_tx_enable_tstamp(fd); - } return ret; } @@ -1613,12 +1606,10 @@ uint16_t dpaa2_dev_tx_conf(void *txq, int drain) } static inline int -dpaa2_dev_tx_fast_mbuf_to_fd(struct rte_eth_dev_data *dev, - struct rte_mbuf *buf, struct qbman_fd *fd, - enum dpaa2_tx_conf_type conf) +dpaa2_dev_tx_fast_mbuf_to_fd(struct dpaa2_dev_priv *priv, + struct rte_mbuf *buf, struct qbman_fd *fd) { struct rte_mempool *mp = buf->pool; - struct dpaa2_dev_priv *priv = dev->dev_private; if (likely(RTE_MBUF_DIRECT(buf) && mp && mp->ops_index == @@ -1629,7 +1620,7 @@ dpaa2_dev_tx_fast_mbuf_to_fd(struct rte_eth_dev_data *dev, DPAA2_PMD_WARN("Single mbuf has next segment(%p)", buf->next); } - if (unlikely(conf == DPAA2_TX_ABSOLUTE_CONF)) + if (unlikely(priv->tx_conf_type == DPAA2_TX_ABSOLUTE_CONF)) dpaa2_mbuf_to_conf_contig_fd(buf, fd); else DPAA2_MBUF_TO_CONTIG_FD(buf, fd, mempool_to_bpid(mp)); @@ -1662,7 +1653,6 @@ dpaa2_dev_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts) struct dpaa2_dev_priv *priv = eth_data->dev_private; struct rte_mempool *hw_mp; uint32_t flags[MAX_TX_RING_SLOTS] = {0}; - uint8_t dy_conf[MAX_TX_RING_SLOTS]; if (unlikely(!DPAA2_PER_LCORE_DPIO)) { ret = dpaa2_affine_qbman_swp(); @@ -1709,7 +1699,6 @@ dpaa2_dev_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts) dpaa2_eqcr_size : nb_pkts; for (loop = 0; loop < frames_to_send; loop++) { - dy_conf[loop] = false; if (*dpaa2_seqn(*bufs)) { uint8_t dqrr_index = *dpaa2_seqn(*bufs) - 1; @@ -1739,8 +1728,7 @@ dpaa2_dev_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts) goto skip_fast_mbuf2fd; } - ret = dpaa2_dev_tx_fast_mbuf_to_fd(eth_data, - *bufs, &fd_arr[loop], priv->tx_conf_type); + ret = dpaa2_dev_tx_fast_mbuf_to_fd(priv, *bufs, &fd_arr[loop]); if (likely(!ret)) { bufs++; continue; @@ -1749,10 +1737,10 @@ dpaa2_dev_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts) skip_fast_mbuf2fd: if (unlikely((*bufs)->nb_segs > 1)) { ret = dpaa2_dev_tx_mbuf_to_sg_fd(hw_mp, *bufs, &fd_arr[loop], - dpaa2_q, priv->tx_conf_type, &dy_conf[loop], tstamp[loop]); + dpaa2_q, priv->tx_conf_type, tstamp[loop]); } else { ret = dpaa2_dev_tx_mbuf_to_simple_fd(hw_mp, *bufs, &fd_arr[loop], - dpaa2_q, priv->tx_conf_type, &dy_conf[loop], tstamp[loop]); + dpaa2_q, priv->tx_conf_type, tstamp[loop]); } if (ret) goto send_n_return; @@ -1849,11 +1837,9 @@ dpaa2_dev_free_eqresp_buf(uint16_t eqresp_ci, static void dpaa2_set_enqueue_descriptor(struct dpaa2_queue *dpaa2_q, - struct rte_mbuf *m, - struct qbman_eq_desc *eqdesc) + struct rte_mbuf *m, struct qbman_eq_desc *eqdesc) { - struct rte_eth_dev_data *eth_data = dpaa2_q->eth_data; - struct dpaa2_dev_priv *priv = eth_data->dev_private; + struct dpaa2_dev_priv *priv = dpaa2_q->eth_data->dev_private; struct dpaa2_dpio_dev *dpio_dev = DPAA2_PER_LCORE_DPIO; struct eqresp_metadata *eqresp_meta; uint16_t orpid, seqnum; @@ -1896,22 +1882,20 @@ dpaa2_set_enqueue_descriptor(struct dpaa2_queue *dpaa2_q, RTE_EXPORT_INTERNAL_SYMBOL(dpaa2_dev_tx_multi_txq_ordered) uint16_t -dpaa2_dev_tx_multi_txq_ordered(void **queue, +dpaa2_dev_tx_multi_txq_ordered(struct dpaa2_queue **dpaa2_q, struct rte_mbuf **bufs, uint16_t nb_pkts) { /* Function to transmit the frames to multiple queues respectively.*/ - uint32_t loop, retry_count, sent = 0, i; + uint32_t loop, retry_count, i; int32_t ret = 0, tstamp[MAX_TX_RING_SLOTS], ptp_set_count, ptp_set; struct qbman_fd fd_arr[MAX_TX_RING_SLOTS]; uint32_t frames_to_send, num_free_eq_desc = 0; struct rte_mempool *hw_mp; struct qbman_eq_desc eqdesc[MAX_TX_RING_SLOTS]; - struct dpaa2_queue *dpaa2_q[MAX_TX_RING_SLOTS]; struct qbman_swp *swp; - struct rte_eth_dev_data *eth_data; struct dpaa2_dev_priv *priv; struct dpaa2_queue *order_sendq; - uint8_t dy_conf[MAX_TX_RING_SLOTS]; + struct dpaa2_queue *cnf_q; if (unlikely(!DPAA2_PER_LCORE_DPIO)) { ret = dpaa2_affine_qbman_swp(); @@ -1923,25 +1907,12 @@ dpaa2_dev_tx_multi_txq_ordered(void **queue, } swp = DPAA2_PER_LCORE_PORTAL; -tx_again: ptp_set_count = 0; frames_to_send = (nb_pkts > dpaa2_eqcr_size) ? dpaa2_eqcr_size : nb_pkts; for (loop = 0; loop < frames_to_send; loop++) { - dpaa2_q[loop] = queue[loop]; - if (dpaa2_q[loop]->tx_conf_queue) - dpaa2_dev_tx_conf(dpaa2_q[loop], false); - eth_data = dpaa2_q[loop]->eth_data; - priv = eth_data->dev_private; - dy_conf[loop] = false; - if (unlikely(!priv->bp_list)) { - DPAA2_PMD_ERR("%s's buffer pool not initialized!", - eth_data->name); - ret = -ENOMEM; - goto send_frames; - } - hw_mp = priv->bp_list->mp; + priv = dpaa2_q[loop]->eth_data->dev_private; if (!priv->en_loose_ordered && (*dpaa2_seqn(*bufs) & DPAA2_ENQUEUE_FLAG_ORP)) { if (!num_free_eq_desc) { @@ -1954,9 +1925,6 @@ dpaa2_dev_tx_multi_txq_ordered(void **queue, num_free_eq_desc--; } - DPAA2_PMD_DP_DEBUG("===> eth_data =%p, fqid =%d", - eth_data, dpaa2_q[loop]->fqid); - /* Check if the queue is congested */ retry_count = 0; while (qbman_result_SCN_state(dpaa2_q[loop]->cscn)) { @@ -1973,8 +1941,7 @@ dpaa2_dev_tx_multi_txq_ordered(void **queue, if (*dpaa2_seqn(*bufs) && priv->en_ordered) { order_sendq = priv->tx_vq[0]; - dpaa2_set_enqueue_descriptor(order_sendq, *bufs, - &eqdesc[loop]); + dpaa2_set_enqueue_descriptor(order_sendq, *bufs, &eqdesc[loop]); } else { qbman_eq_desc_set_no_orp(&eqdesc[loop], DPAA2_EQ_RESP_ERR_FQ); qbman_eq_desc_set_fq(&eqdesc[loop], dpaa2_q[loop]->fqid); @@ -1993,27 +1960,27 @@ dpaa2_dev_tx_multi_txq_ordered(void **queue, goto skip_fast_mbuf2fd; } - ret = dpaa2_dev_tx_fast_mbuf_to_fd(eth_data, - *bufs, &fd_arr[loop], priv->tx_conf_type); + ret = dpaa2_dev_tx_fast_mbuf_to_fd(priv, *bufs, &fd_arr[loop]); if (likely(!ret)) { - if (priv->tx_conf_type == DPAA2_TX_ABSOLUTE_CONF) - dy_conf[loop] = true; bufs++; continue; } skip_fast_mbuf2fd: + if (unlikely(!priv->bp_list)) { + ret = -ENOMEM; + goto send_frames; + } + hw_mp = priv->bp_list->mp; if (unlikely((*bufs)->nb_segs > 1)) { ret = dpaa2_dev_tx_mbuf_to_sg_fd(hw_mp, *bufs, &fd_arr[loop], - dpaa2_q[loop], priv->tx_conf_type, &dy_conf[loop], tstamp[loop]); + dpaa2_q[loop], priv->tx_conf_type, tstamp[loop]); } else { ret = dpaa2_dev_tx_mbuf_to_simple_fd(hw_mp, *bufs, &fd_arr[loop], - dpaa2_q[loop], priv->tx_conf_type, &dy_conf[loop], tstamp[loop]); + dpaa2_q[loop], priv->tx_conf_type, tstamp[loop]); } if (ret) goto send_frames; - if (priv->tx_conf_type == DPAA2_TX_ABSOLUTE_CONF) - dy_conf[loop] = true; bufs++; } @@ -2023,8 +1990,7 @@ dpaa2_dev_tx_multi_txq_ordered(void **queue, retry_count = 0; while (loop < frames_to_send) { ret = qbman_swp_enqueue_multiple_desc(swp, &eqdesc[loop], - &fd_arr[loop], - frames_to_send - loop); + &fd_arr[loop], frames_to_send - loop); if (likely(ret > 0)) { loop += ret; retry_count = 0; @@ -2034,18 +2000,41 @@ dpaa2_dev_tx_multi_txq_ordered(void **queue, break; } } - nb_pkts -= loop; - sent += loop; for (i = 0; i < loop; i++) { - if (tstamp[i] && dpaa2_q[i]->tx_conf_queue) { - rte_atomic_fetch_add_explicit(&dpaa2_q[i]->tx_conf_queue->ts_to_cnfd, - 1, rte_memory_order_relaxed); + if (unlikely(dpaa2_q[i]->tx_conf_queue)) { + dpaa2_dev_tx_conf(dpaa2_q[i], false); + if (tstamp[i]) { + cnf_q = dpaa2_q[i]->tx_conf_queue; + rte_atomic_fetch_add_explicit(&cnf_q->ts_to_cnfd, + 1, rte_memory_order_relaxed); + } } } - if (nb_pkts > 0 && !ret) - goto tx_again; - return sent; + return loop; +} + +RTE_EXPORT_EXPERIMENTAL_SYMBOL(rte_dpaa2_dev_tx_multi_ports, 26.11) +uint16_t +rte_dpaa2_dev_tx_multi_ports(uint16_t port_id[], + uint16_t txq_id[], struct rte_mbuf **bufs, uint16_t nb_pkts) +{ + uint16_t i; + struct dpaa2_queue *txq[nb_pkts]; + struct rte_eth_dev_data *data; + + if (txq_id) { + for (i = 0; i < nb_pkts; i++) { + data = rte_eth_devices[port_id[i]].data; + txq[i] = data->tx_queues[txq_id[i]]; + } + } else { + for (i = 0; i < nb_pkts; i++) { + data = rte_eth_devices[port_id[i]].data; + txq[i] = data->tx_queues[0]; + } + } + return dpaa2_dev_tx_multi_txq_ordered(txq, bufs, nb_pkts); } /* Callback to handle sending ordered packets through WRIOP based interface */ @@ -2054,7 +2043,7 @@ dpaa2_dev_tx_ordered(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts) { uint16_t i; - void *mq[nb_pkts]; + struct dpaa2_queue *mq[nb_pkts]; for (i = 0; i < nb_pkts; i++) mq[i] = queue; diff --git a/drivers/net/dpaa2/rte_pmd_dpaa2.h b/drivers/net/dpaa2/rte_pmd_dpaa2.h index 5125ec5bf7..74b7e63c54 100644 --- a/drivers/net/dpaa2/rte_pmd_dpaa2.h +++ b/drivers/net/dpaa2/rte_pmd_dpaa2.h @@ -162,6 +162,11 @@ uint16_t rte_dpaa2_scheduler_rx(void *scheduler_handle, struct rte_mbuf **mbuf, uint16_t nb_pkts); +__rte_experimental +uint16_t +rte_dpaa2_dev_tx_multi_ports(uint16_t port_id[], + uint16_t txq_id[], struct rte_mbuf **bufs, uint16_t nb_pkts); + #define RTE_DPAA2_EVENT_PORT_CFG_ATOMIC RTE_BIT32(31) /** RTE_DPAA2_MIX_GROUP_FLOW has same behavio as RTE_DPAA2_FLAT_GROUP_FLOW*/ -- 2.43.0
