[dpdk-dev] [PATCH v5 5/8] net/ice: support Tx SSE vector

[Maxime Coquelin]

On 3/22/19 3:58 AM, Wenzhuo Lu wrote:
> Signed-off-by: Wenzhuo Lu <wenzhuo.lu at intel.com>
> ---
>   doc/guides/nics/features/ice_vec.ini  |   2 +
>   drivers/net/ice/ice_rxtx.c            |  17 +++++
>   drivers/net/ice/ice_rxtx.h            |   4 +
>   drivers/net/ice/ice_rxtx_vec_common.h | 133 +++++++++++++++++++++++++++++++++
>   drivers/net/ice/ice_rxtx_vec_sse.c    | 135 ++++++++++++++++++++++++++++++++++
>   5 files changed, 291 insertions(+)
> 
> diff --git a/doc/guides/nics/features/ice_vec.ini b/doc/guides/nics/features/ice_vec.ini
> index 1a19788..173c8f2 100644
> --- a/doc/guides/nics/features/ice_vec.ini
> +++ b/doc/guides/nics/features/ice_vec.ini
> @@ -12,6 +12,7 @@ Queue start/stop     = Y
>   MTU update           = Y
>   Jumbo frame          = Y
>   Scattered Rx         = Y
> +TSO                  = Y
>   Promiscuous mode     = Y
>   Allmulticast mode    = Y
>   Unicast MAC filter   = Y
> @@ -22,6 +23,7 @@ RSS reta update      = Y
>   VLAN filter          = Y
>   Packet type parsing  = Y
>   Rx descriptor status = Y
> +Tx descriptor status = Y
>   Basic stats          = Y
>   Extended stats       = Y
>   FW version           = Y
> diff --git a/drivers/net/ice/ice_rxtx.c b/drivers/net/ice/ice_rxtx.c
> index 5409dd0..f9ecffa 100644
> --- a/drivers/net/ice/ice_rxtx.c
> +++ b/drivers/net/ice/ice_rxtx.c
> @@ -2332,6 +2332,23 @@ void __attribute__((cold))
>   {
>   	struct ice_adapter *ad =
>   		ICE_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
> +#ifdef RTE_ARCH_X86
> +	struct ice_tx_queue *txq;
> +	int i;
> +
> +	if (!ice_tx_vec_dev_check(dev)) {
> +		for (i = 0; i < dev->data->nb_tx_queues; i++) {
> +			txq = dev->data->tx_queues[i];
> +			(void)ice_txq_vec_setup(txq);
> +		}
> +		PMD_DRV_LOG(DEBUG, "Using Vector Tx (port %d).",
> +			    dev->data->port_id);
> +		dev->tx_pkt_burst = ice_xmit_pkts_vec;
> +		dev->tx_pkt_prepare = NULL;
> +
> +		return;
> +	}
> +#endif
>   
>   	if (ad->tx_simple_allowed) {
>   		PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
> diff --git a/drivers/net/ice/ice_rxtx.h b/drivers/net/ice/ice_rxtx.h
> index 6ef0a84..1dde4e7 100644
> --- a/drivers/net/ice/ice_rxtx.h
> +++ b/drivers/net/ice/ice_rxtx.h
> @@ -170,9 +170,13 @@ void ice_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
>   const uint32_t *ice_dev_supported_ptypes_get(struct rte_eth_dev *dev);
>   
>   int ice_rx_vec_dev_check(struct rte_eth_dev *dev);
> +int ice_tx_vec_dev_check(struct rte_eth_dev *dev);
>   int ice_rxq_vec_setup(struct ice_rx_queue *rxq);
> +int ice_txq_vec_setup(struct ice_tx_queue *txq);
>   uint16_t ice_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
>   			   uint16_t nb_pkts);
>   uint16_t ice_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
>   				     uint16_t nb_pkts);
> +uint16_t ice_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
> +			   uint16_t nb_pkts);
>   #endif /* _ICE_RXTX_H_ */
> diff --git a/drivers/net/ice/ice_rxtx_vec_common.h b/drivers/net/ice/ice_rxtx_vec_common.h
> index cfef91b..be079a3 100644
> --- a/drivers/net/ice/ice_rxtx_vec_common.h
> +++ b/drivers/net/ice/ice_rxtx_vec_common.h
> @@ -71,6 +71,73 @@
>   	return pkt_idx;
>   }
>   
> +static __rte_always_inline int
> +ice_tx_free_bufs(struct ice_tx_queue *txq)
> +{
> +	struct ice_tx_entry *txep;
> +	uint32_t n;
> +	uint32_t i;
> +	int nb_free = 0;
> +	struct rte_mbuf *m, *free[ICE_TX_MAX_FREE_BUF_SZ];
> +
> +	/* check DD bits on threshold descriptor */
> +	if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
> +			rte_cpu_to_le_64(ICE_TXD_QW1_DTYPE_M)) !=
> +			rte_cpu_to_le_64(ICE_TX_DESC_DTYPE_DESC_DONE))
> +		return 0;
> +
> +	n = txq->tx_rs_thresh;
> +
> +	 /* first buffer to free from S/W ring is at index
> +	  * tx_next_dd - (tx_rs_thresh-1)
> +	  */
> +	txep = &txq->sw_ring[txq->tx_next_dd - (n - 1)];
> +	m = rte_pktmbuf_prefree_seg(txep[0].mbuf);
> +	if (likely(m)) {
> +		free[0] = m;
> +		nb_free = 1;
> +		for (i = 1; i < n; i++) {
> +			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> +			if (likely(m)) {
> +				if (likely(m->pool == free[0]->pool)) {
> +					free[nb_free++] = m;
> +				} else {
> +					rte_mempool_put_bulk(free[0]->pool,
> +							     (void *)free,
> +							     nb_free);
> +					free[0] = m;
> +					nb_free = 1;
> +				}
> +			}
> +		}
> +		rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
> +	} else {
> +		for (i = 1; i < n; i++) {
> +			m = rte_pktmbuf_prefree_seg(txep[i].mbuf);
> +			if (m)
> +				rte_mempool_put(m->pool, m);
> +		}
> +	}
> +
> +	/* buffers were freed, update counters */
> +	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
> +	txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
> +	if (txq->tx_next_dd >= txq->nb_tx_desc)
> +		txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
> +
> +	return txq->tx_rs_thresh;
> +}
> +
> +static __rte_always_inline void
> +tx_backlog_entry(struct ice_tx_entry *txep,
Consider prefixing it with 'ice_tx_'.

> +		 struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
> +{
> +	int i;
> +
> +	for (i = 0; i < (int)nb_pkts; ++i)
> +		txep[i].mbuf = tx_pkts[i];
> +}
> +
>   static inline void
>   _ice_rx_queue_release_mbufs_vec(struct ice_rx_queue *rxq)
>   {
> @@ -101,6 +168,34 @@
>   	memset(rxq->sw_ring, 0, sizeof(rxq->sw_ring[0]) * rxq->nb_rx_desc);
>   }
>   
> +static inline void
> +_ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
> +{
> +	uint16_t i;
> +
> +	if (!txq || !txq->sw_ring) {

if (unlikely(...)) {

> +		PMD_DRV_LOG(DEBUG, "Pointer to rxq or sw_ring is NULL");

s/rxq/txq/

> +		return;
> +	}
> +
> +	/**
> +	 *  vPMD tx will not set sw_ring's mbuf to NULL after free,
> +	 *  so need to free remains more carefully.
> +	 */
> +	i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
> +	if (txq->tx_tail < i) {
> +		for (; i < txq->nb_tx_desc; i++) {
> +			rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
> +			txq->sw_ring[i].mbuf = NULL;
> +		}
> +		i = 0;
> +	}
> +	for (; i < txq->tx_tail; i++) {
> +		rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
> +		txq->sw_ring[i].mbuf = NULL;
> +	}
> +}
> +
>   static inline int
>   ice_rxq_vec_setup_default(struct ice_rx_queue *rxq)
>   {
> @@ -137,6 +232,29 @@
>   	return 0;
>   }
>   
> +#define ICE_NO_VECTOR_FLAGS (				 \
> +		DEV_TX_OFFLOAD_MULTI_SEGS |		 \
> +		DEV_TX_OFFLOAD_VLAN_INSERT |		 \
> +		DEV_TX_OFFLOAD_SCTP_CKSUM |		 \
> +		DEV_TX_OFFLOAD_UDP_CKSUM |		 \
> +		DEV_TX_OFFLOAD_TCP_CKSUM)
> +
> +static inline int
> +ice_tx_vec_queue_default(struct ice_tx_queue *txq)
> +{
> +	if (!txq)
> +		return -1;
> +
> +	if (txq->offloads & ICE_NO_VECTOR_FLAGS)
> +		return -1;
> +
> +	if (txq->tx_rs_thresh < ICE_VPMD_TX_BURST ||
> +	    txq->tx_rs_thresh > ICE_TX_MAX_FREE_BUF_SZ)
> +		return -1;
> +
> +	return 0;
> +}
> +
>   static inline int
>   ice_rx_vec_dev_check_default(struct rte_eth_dev *dev)
>   {
> @@ -152,4 +270,19 @@
>   	return 0;
>   }
>   
> +static inline int
> +ice_tx_vec_dev_check_default(struct rte_eth_dev *dev)
> +{
> +	int i;
> +	struct ice_tx_queue *txq;
> +
> +	for (i = 0; i < dev->data->nb_tx_queues; i++) {
> +		txq = dev->data->tx_queues[i];
> +		if (ice_tx_vec_queue_default(txq))
> +			return -1;

return ice_tx_vec_queue_default(txq);

Applies also to rx path.

> +	}
> +
> +	return 0;
> +}
> +
>   #endif
> diff --git a/drivers/net/ice/ice_rxtx_vec_sse.c b/drivers/net/ice/ice_rxtx_vec_sse.c
> index e1f057a..4e148d6 100644
> --- a/drivers/net/ice/ice_rxtx_vec_sse.c
> +++ b/drivers/net/ice/ice_rxtx_vec_sse.c
> @@ -514,12 +514,131 @@
>   				      &split_flags[i]);
>   }
>   
> +static inline void
> +ice_vtx1(volatile struct ice_tx_desc *txdp, struct rte_mbuf *pkt,
> +	 uint64_t flags)
> +{
> +	uint64_t high_qw =
> +		(ICE_TX_DESC_DTYPE_DATA |
> +		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
> +		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
> +
> +	__m128i descriptor = _mm_set_epi64x(high_qw,
> +					    pkt->buf_iova + pkt->data_off);
> +	_mm_store_si128((__m128i *)txdp, descriptor);
> +}
> +
> +static inline void
> +ice_vtx(volatile struct ice_tx_desc *txdp, struct rte_mbuf **pkt,
> +	uint16_t nb_pkts, uint64_t flags)
> +{
> +	int i;
> +
> +	for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
> +		ice_vtx1(txdp, *pkt, flags);
> +}
> +
> +static uint16_t
> +ice_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
> +			 uint16_t nb_pkts)
> +{
> +	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
> +	volatile struct ice_tx_desc *txdp;
> +	struct ice_tx_entry *txep;
> +	uint16_t n, nb_commit, tx_id;
> +	uint64_t flags = ICE_TD_CMD;
> +	uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
> +	int i;
> +
> +	/* cross rx_thresh boundary is not allowed */
> +	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
> +
> +	if (txq->nb_tx_free < txq->tx_free_thresh)
> +		ice_tx_free_bufs(txq);
> +
> +	nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
> +	nb_commit = nb_pkts;
> +	if (unlikely(nb_pkts == 0))
> +		return 0;
> +
> +	tx_id = txq->tx_tail;
> +	txdp = &txq->tx_ring[tx_id];
> +	txep = &txq->sw_ring[tx_id];
> +
> +	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
> +
> +	n = (uint16_t)(txq->nb_tx_desc - tx_id);
> +	if (nb_commit >= n) {
> +		tx_backlog_entry(txep, tx_pkts, n);
> +
> +		for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
> +			ice_vtx1(txdp, *tx_pkts, flags);
> +
> +		ice_vtx1(txdp, *tx_pkts++, rs);
> +
> +		nb_commit = (uint16_t)(nb_commit - n);
> +
> +		tx_id = 0;
> +		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
> +
> +		/* avoid reach the end of ring */
> +		txdp = &txq->tx_ring[tx_id];
> +		txep = &txq->sw_ring[tx_id];
> +	}
> +
> +	tx_backlog_entry(txep, tx_pkts, nb_commit);
> +
> +	ice_vtx(txdp, tx_pkts, nb_commit, flags);
> +
> +	tx_id = (uint16_t)(tx_id + nb_commit);
> +	if (tx_id > txq->tx_next_rs) {
> +		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
> +			rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
> +					 ICE_TXD_QW1_CMD_S);
> +		txq->tx_next_rs =
> +			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
> +	}
> +
> +	txq->tx_tail = tx_id;
> +
> +	ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
> +
> +	return nb_pkts;
> +}
> +
> +uint16_t
> +ice_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
> +		  uint16_t nb_pkts)
> +{
> +	uint16_t nb_tx = 0;
> +	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
> +
> +	while (nb_pkts) {
> +		uint16_t ret, num;
> +
> +		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
> +		ret = ice_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx], num);
> +		nb_tx += ret;
> +		nb_pkts -= ret;
> +		if (ret < num)
> +			break;
> +	}
> +
> +	return nb_tx;
> +}
> +
>   static void __attribute__((cold))
>   ice_rx_queue_release_mbufs_vec(struct ice_rx_queue *rxq)
>   {
>   	_ice_rx_queue_release_mbufs_vec(rxq);
>   }
>   
> +static void __attribute__((cold))
> +ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
> +{
> +	_ice_tx_queue_release_mbufs_vec(txq);

As for Rx, consider putting the code firectly here as
_ice_tx_queue_release_mbufs_vec() is called only once here.

> +}
> +
>   int __attribute__((cold))
>   ice_rxq_vec_setup(struct ice_rx_queue *rxq)
>   {
> @@ -531,7 +650,23 @@ int __attribute__((cold))
>   }
>   
>   int __attribute__((cold))
> +ice_txq_vec_setup(struct ice_tx_queue __rte_unused *txq)
> +{
> +	if (!txq)
> +		return -1;
> +
> +	txq->tx_rel_mbufs = ice_tx_queue_release_mbufs_vec;
> +	return 0;
> +}
> +
> +int __attribute__((cold))
>   ice_rx_vec_dev_check(struct rte_eth_dev *dev)
>   {
>   	return ice_rx_vec_dev_check_default(dev);
>   }
> +
> +int __attribute__((cold))
> +ice_tx_vec_dev_check(struct rte_eth_dev *dev)
> +{
> +	return ice_tx_vec_dev_check_default(dev);
> +}
>