net/bnxt: support fast mbuf free
PATCH v2 1/1 net/bnxt: support fast mbuf free
Add support for DEV_TX_OFFLOAD_MBUF_FAST_FREE to bnxt vector mode transmit. This offload may be enabled only when multi-segment transmit is not needed, all transmitted mbufs for a given queue will be allocated from the same pool, and all transmitted mbufs will have a reference count of 1.
Since this offload is only supported in vector mode at this time, support in the features list is indicated as "partial".
---
- v2:
Fixed mangling of advertised tx offload capabilities.
doc/guides/nics/features/bnxt.ini | 1 + drivers/net/bnxt/bnxt_ethdev.c | 11 +++++++---- drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 28 ++++++++++++++++++++++++++- drivers/net/bnxt/bnxt_rxtx_vec_sse.c | 28 ++++++++++++++++++++++++++- drivers/net/bnxt/bnxt_txq.c | 2 ++ drivers/net/bnxt/bnxt_txq.h | 1 + 6 files changed, 65 insertions(+), 6 deletions(-)
doc/guides/nics/features/bnxt.ini +1 -0
@@ -8,6 +8,7 @@ Speed capabilities = Y Link status = Y Link status event = Y Rx interrupt = Y +Fast mbuf free = P Queue start/stop = Y Burst mode info = Y MTU update = Y
drivers/net/bnxt/bnxt_ethdev.c +7 -4
@@ -920,7 +920,9 @@ static int bnxt_dev_info_get_op(struct rte_eth_dev *eth_dev, dev_info->rx_offload_capa =BNXT_DEV_RX_OFFLOAD_SUPPORT ;if (bp->flags &BNXT_FLAG_PTP_SUPPORTED ) dev_info->rx_offload_capa |=DEV_RX_OFFLOAD_TIMESTAMP ; - dev_info->tx_offload_capa =BNXT_DEV_TX_OFFLOAD_SUPPORT ; + dev_info->tx_queue_offload_capa =DEV_TX_OFFLOAD_MBUF_FAST_FREE ; + dev_info->tx_offload_capa =BNXT_DEV_TX_OFFLOAD_SUPPORT | + dev_info->tx_queue_offload_capa; dev_info->flow_type_rss_offloads =BNXT_ETH_RSS_SUPPORT ; dev_info->speed_capa =bnxt_get_speed_capabilities (bp);
@@ -1191,6 +1193,7 @@ bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev) {#if defined(RTE_ARCH_X86 ) || defined(RTE_ARCH_ARM64 )#ifndef RTE_LIBRTE_IEEE1588 +uint64_t offloads = eth_dev->data->dev_conf.txmode.offloads;struct bnxt *bp = eth_dev->data->dev_private;/*
@@ -1198,7 +1201,7 @@ bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev)* or tx offloads. */ if (!eth_dev->data->scattered_rx && - !eth_dev->data->dev_conf.txmode.offloads && + !(offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE ) && !BNXT_TRUFLOW_EN (bp)) {PMD_DRV_LOG (INFO ,"Using vector mode transmit for port %d\n" , eth_dev->data->port_id);
@@ -1210,7 +1213,7 @@ bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev)"Port %d scatter: %d tx offload: %" PRIX64 "\n" , eth_dev->data->port_id, eth_dev->data->scattered_rx, - eth_dev->data->dev_conf.txmode.offloads); + offloads);#endif #endif return bnxt_xmit_pkts;
@@ -2685,7 +2688,7 @@ bnxt_txq_info_get_op(struct rte_eth_dev *dev, uint16_t queue_id, qinfo->conf.tx_free_thresh = txq->tx_free_thresh; qinfo->conf.tx_rs_thresh =0 ; qinfo->conf.tx_deferred_start = txq->tx_deferred_start; - qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads; + qinfo->conf.offloads = txq->offloads; } static const struct {
drivers/net/bnxt/bnxt_rxtx_vec_neon.c +27 -1
@@ -338,6 +338,29 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,return nb_rx_pkts; } +static void +bnxt_tx_cmp_vec_fast (struct bnxt_tx_queue *txq,int nr_pkts) +{ +struct bnxt_tx_ring_info *txr = txq->tx_ring; +struct rte_mbuf **free = txq->free; +uint16_t cons = txr->tx_cons; +unsigned int blk =0 ; +uint32_t ring_mask = txr->tx_ring_struct->ring_mask; + +while (nr_pkts--) { +struct bnxt_sw_tx_bd *tx_buf; + + tx_buf = &txr->tx_buf_ring[cons]; + cons = (cons +1 ) & ring_mask; + free[blk++] = tx_buf->mbuf; + tx_buf->mbuf =NULL ; + } +if (blk) +rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); + + txr->tx_cons = cons; +} +static void bnxt_tx_cmp_vec (struct bnxt_tx_queue *txq,int nr_pkts) {
@@ -399,7 +422,10 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq) cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);if (nb_tx_pkts) { -bnxt_tx_cmp_vec (txq, nb_tx_pkts); +if (txq->offloads &DEV_TX_OFFLOAD_MBUF_FAST_FREE ) +bnxt_tx_cmp_vec_fast (txq, nb_tx_pkts); +else +bnxt_tx_cmp_vec (txq, nb_tx_pkts); cpr->cp_raw_cons = raw_cons;bnxt_db_cq (cpr); }
drivers/net/bnxt/bnxt_rxtx_vec_sse.c +27 -1
@@ -309,6 +309,29 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,return nb_rx_pkts; } +static void +bnxt_tx_cmp_vec_fast (struct bnxt_tx_queue *txq,int nr_pkts) +{ +struct bnxt_tx_ring_info *txr = txq->tx_ring; +struct rte_mbuf **free = txq->free; +uint16_t cons = txr->tx_cons; +unsigned int blk =0 ; +uint32_t ring_mask = txr->tx_ring_struct->ring_mask; + +while (nr_pkts--) { +struct bnxt_sw_tx_bd *tx_buf; + + tx_buf = &txr->tx_buf_ring[cons]; + cons = (cons +1 ) & ring_mask; + free[blk++] = tx_buf->mbuf; + tx_buf->mbuf =NULL ; + } +if (blk) +rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); + + txr->tx_cons = cons; +} +static void bnxt_tx_cmp_vec (struct bnxt_tx_queue *txq,int nr_pkts) {
@@ -371,7 +394,10 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq) cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);if (nb_tx_pkts) { -bnxt_tx_cmp_vec (txq, nb_tx_pkts); +if (txq->offloads &DEV_TX_OFFLOAD_MBUF_FAST_FREE ) +bnxt_tx_cmp_vec_fast (txq, nb_tx_pkts); +else +bnxt_tx_cmp_vec (txq, nb_tx_pkts); cpr->cp_raw_cons = raw_cons;bnxt_db_cq (cpr); }
drivers/net/bnxt/bnxt_txq.c +2 -0
@@ -131,6 +131,8 @@ int bnxt_tx_queue_setup_op(struct rte_eth_dev *eth_dev, txq->nb_tx_desc = nb_desc; txq->tx_free_thresh =RTE_MIN (rte_align32pow2 (nb_desc) /4 ,RTE_BNXT_MAX_TX_BURST ); + txq->offloads = eth_dev->data->dev_conf.txmode.offloads | + tx_conf->offloads; txq->tx_deferred_start = tx_conf->tx_deferred_start;
drivers/net/bnxt/bnxt_txq.h +1 -0
@@ -39,6 +39,7 @@ struct bnxt_tx_queue {struct bnxt_cp_ring_info *cp_ring;const struct rte_memzone *mz;struct rte_mbuf **free; +uint64_t offloads; };void bnxt_free_txq_stats (struct bnxt_tx_queue *txq);
Signature · 1 line
2.25.1
PATCH v3 1/1 net/bnxt: support fast mbuf free
Add support for DEV_TX_OFFLOAD_MBUF_FAST_FREE to bnxt vector mode transmit. This offload may be enabled only when multi-segment transmit is not needed, all transmitted mbufs for a given queue will be allocated from the same pool, and all transmitted mbufs will have a reference count of 1.
---
- v2:
Fixed mangling of advertised tx offload capabilities.
- v3:
Reduced duplicated code.
Enable for non-vector transmit path as well.
doc/guides/nics/features/bnxt.ini | 1 + drivers/net/bnxt/bnxt_ethdev.c | 11 ++++--- drivers/net/bnxt/bnxt_rxtx_vec_common.h | 32 ++++++++++++++++++++ drivers/net/bnxt/bnxt_rxtx_vec_neon.c | 38 ++++-------------------- drivers/net/bnxt/bnxt_rxtx_vec_sse.c | 39 ++++--------------------- drivers/net/bnxt/bnxt_txq.c | 2 ++ drivers/net/bnxt/bnxt_txq.h | 1 + drivers/net/bnxt/bnxt_txr.c | 5 +++- drivers/net/bnxt/bnxt_txr.h | 27 +++++++++++++++++ 9 files changed, 84 insertions(+), 72 deletions(-)
doc/guides/nics/features/bnxt.ini +1 -0
@@ -8,6 +8,7 @@ Speed capabilities = Y Link status = Y Link status event = Y Rx interrupt = Y +Fast mbuf free = Y Queue start/stop = Y Burst mode info = Y MTU update = Y
drivers/net/bnxt/bnxt_ethdev.c +7 -4
@@ -920,7 +920,9 @@ static int bnxt_dev_info_get_op(struct rte_eth_dev *eth_dev, dev_info->rx_offload_capa =BNXT_DEV_RX_OFFLOAD_SUPPORT ;if (bp->flags &BNXT_FLAG_PTP_SUPPORTED ) dev_info->rx_offload_capa |=DEV_RX_OFFLOAD_TIMESTAMP ; - dev_info->tx_offload_capa =BNXT_DEV_TX_OFFLOAD_SUPPORT ; + dev_info->tx_queue_offload_capa =DEV_TX_OFFLOAD_MBUF_FAST_FREE ; + dev_info->tx_offload_capa =BNXT_DEV_TX_OFFLOAD_SUPPORT | + dev_info->tx_queue_offload_capa; dev_info->flow_type_rss_offloads =BNXT_ETH_RSS_SUPPORT ; dev_info->speed_capa =bnxt_get_speed_capabilities (bp);
@@ -1191,6 +1193,7 @@ bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev) {#if defined(RTE_ARCH_X86 ) || defined(RTE_ARCH_ARM64 )#ifndef RTE_LIBRTE_IEEE1588 +uint64_t offloads = eth_dev->data->dev_conf.txmode.offloads;struct bnxt *bp = eth_dev->data->dev_private;/*
@@ -1198,7 +1201,7 @@ bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev)* or tx offloads. */ if (!eth_dev->data->scattered_rx && - !eth_dev->data->dev_conf.txmode.offloads && + !(offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE ) && !BNXT_TRUFLOW_EN (bp)) {PMD_DRV_LOG (INFO ,"Using vector mode transmit for port %d\n" , eth_dev->data->port_id);
@@ -1210,7 +1213,7 @@ bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev)"Port %d scatter: %d tx offload: %" PRIX64 "\n" , eth_dev->data->port_id, eth_dev->data->scattered_rx, - eth_dev->data->dev_conf.txmode.offloads); + offloads);#endif #endif return bnxt_xmit_pkts;
@@ -2685,7 +2688,7 @@ bnxt_txq_info_get_op(struct rte_eth_dev *dev, uint16_t queue_id, qinfo->conf.tx_free_thresh = txq->tx_free_thresh; qinfo->conf.tx_rs_thresh =0 ; qinfo->conf.tx_deferred_start = txq->tx_deferred_start; - qinfo->conf.offloads = dev->data->dev_conf.txmode.offloads; + qinfo->conf.offloads = txq->offloads; } static const struct {
drivers/net/bnxt/bnxt_rxtx_vec_common.h +32 -0
@@ -94,4 +94,36 @@ bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr) rxq->rxrearm_nb -= nb; } + +static inline void +bnxt_tx_cmp_vec (struct bnxt_tx_queue *txq,int nr_pkts) +{ +struct bnxt_tx_ring_info *txr = txq->tx_ring; +struct rte_mbuf **free = txq->free; +uint16_t cons = txr->tx_cons; +unsigned int blk =0 ; +uint32_t ring_mask = txr->tx_ring_struct->ring_mask; + +while (nr_pkts--) { +struct bnxt_sw_tx_bd *tx_buf; +struct rte_mbuf *mbuf; + + tx_buf = &txr->tx_buf_ring[cons]; + cons = (cons +1 ) & ring_mask; + mbuf =rte_pktmbuf_prefree_seg (tx_buf->mbuf); +if (unlikely (mbuf ==NULL )) +continue ; + tx_buf->mbuf =NULL ; + +if (blk && mbuf->pool != free[0 ]->pool) { +rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); + blk =0 ; + } + free[blk++] = mbuf; + } +if (blk) +rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); + + txr->tx_cons = cons; +}#endif /* _BNXT_RXTX_VEC_COMMON_H_ */
drivers/net/bnxt/bnxt_rxtx_vec_neon.c +5 -33
@@ -13,10 +13,10 @@#include "bnxt.h" #include "bnxt_cpr.h" #include "bnxt_ring.h" -#include "bnxt_rxtx_vec_common.h" #include "bnxt_txq.h" #include "bnxt_txr.h" +#include "bnxt_rxtx_vec_common.h" /* * RX Ring handling
@@ -338,37 +338,6 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,return nb_rx_pkts; } -static void -bnxt_tx_cmp_vec (struct bnxt_tx_queue *txq,int nr_pkts) -{ -struct bnxt_tx_ring_info *txr = txq->tx_ring; -struct rte_mbuf **free = txq->free; -uint16_t cons = txr->tx_cons; -unsigned int blk =0 ; - -while (nr_pkts--) { -struct bnxt_sw_tx_bd *tx_buf; -struct rte_mbuf *mbuf; - - tx_buf = &txr->tx_buf_ring[cons]; - cons =RING_NEXT (txr->tx_ring_struct, cons); - mbuf =rte_pktmbuf_prefree_seg (tx_buf->mbuf); -if (unlikely (mbuf ==NULL )) -continue ; - tx_buf->mbuf =NULL ; - -if (blk && mbuf->pool != free[0 ]->pool) { -rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); - blk =0 ; - } - free[blk++] = mbuf; - } -if (blk) -rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); - - txr->tx_cons = cons; -} -static void bnxt_handle_tx_cp_vec (struct bnxt_tx_queue *txq) {
@@ -399,7 +368,10 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq) cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);if (nb_tx_pkts) { -bnxt_tx_cmp_vec (txq, nb_tx_pkts); +if (txq->offloads &DEV_TX_OFFLOAD_MBUF_FAST_FREE ) +bnxt_tx_cmp_fast (txq, nb_tx_pkts); +else +bnxt_tx_cmp_vec (txq, nb_tx_pkts); cpr->cp_raw_cons = raw_cons;bnxt_db_cq (cpr); }
drivers/net/bnxt/bnxt_rxtx_vec_sse.c +5 -34
@@ -13,10 +13,10 @@#include "bnxt.h" #include "bnxt_cpr.h" #include "bnxt_ring.h" -#include "bnxt_rxtx_vec_common.h" #include "bnxt_txq.h" #include "bnxt_txr.h" +#include "bnxt_rxtx_vec_common.h" /* * RX Ring handling
@@ -309,38 +309,6 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,return nb_rx_pkts; } -static void -bnxt_tx_cmp_vec (struct bnxt_tx_queue *txq,int nr_pkts) -{ -struct bnxt_tx_ring_info *txr = txq->tx_ring; -struct rte_mbuf **free = txq->free; -uint16_t cons = txr->tx_cons; -unsigned int blk =0 ; -uint32_t ring_mask = txr->tx_ring_struct->ring_mask; - -while (nr_pkts--) { -struct bnxt_sw_tx_bd *tx_buf; -struct rte_mbuf *mbuf; - - tx_buf = &txr->tx_buf_ring[cons]; - cons = (cons +1 ) & ring_mask; - mbuf =rte_pktmbuf_prefree_seg (tx_buf->mbuf); -if (unlikely (mbuf ==NULL )) -continue ; - tx_buf->mbuf =NULL ; - -if (blk && mbuf->pool != free[0 ]->pool) { -rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); - blk =0 ; - } - free[blk++] = mbuf; - } -if (blk) -rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); - - txr->tx_cons = cons; -} -static void bnxt_handle_tx_cp_vec (struct bnxt_tx_queue *txq) {
@@ -371,7 +339,10 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq) cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);if (nb_tx_pkts) { -bnxt_tx_cmp_vec (txq, nb_tx_pkts); +if (txq->offloads &DEV_TX_OFFLOAD_MBUF_FAST_FREE ) +bnxt_tx_cmp_fast (txq, nb_tx_pkts); +else +bnxt_tx_cmp_vec (txq, nb_tx_pkts); cpr->cp_raw_cons = raw_cons;bnxt_db_cq (cpr); }
drivers/net/bnxt/bnxt_txq.c +2 -0
@@ -131,6 +131,8 @@ int bnxt_tx_queue_setup_op(struct rte_eth_dev *eth_dev, txq->nb_tx_desc = nb_desc; txq->tx_free_thresh =RTE_MIN (rte_align32pow2 (nb_desc) /4 ,RTE_BNXT_MAX_TX_BURST ); + txq->offloads = eth_dev->data->dev_conf.txmode.offloads | + tx_conf->offloads; txq->tx_deferred_start = tx_conf->tx_deferred_start;
drivers/net/bnxt/bnxt_txq.h +1 -0
@@ -39,6 +39,7 @@ struct bnxt_tx_queue {struct bnxt_cp_ring_info *cp_ring;const struct rte_memzone *mz;struct rte_mbuf **free; +uint64_t offloads; };void bnxt_free_txq_stats (struct bnxt_tx_queue *txq);
drivers/net/bnxt/bnxt_txr.c +4 -1
@@ -428,7 +428,10 @@ static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq) }while (nb_tx_pkts < ring_mask);if (nb_tx_pkts) { -bnxt_tx_cmp (txq, nb_tx_pkts); +if (txq->offloads &DEV_TX_OFFLOAD_MBUF_FAST_FREE ) +bnxt_tx_cmp_fast (txq, nb_tx_pkts); +else +bnxt_tx_cmp (txq, nb_tx_pkts); cpr->cp_raw_cons = raw_cons;bnxt_db_cq (cpr); }
drivers/net/bnxt/bnxt_txr.h +27 -0
@@ -52,6 +52,33 @@ static inline uint32_t bnxt_tx_avail(struct bnxt_tx_queue *txq) bnxt_tx_bds_in_hw(txq)) - 1); } +/* +* Transmit completion function for use when DEV_TX_OFFLOAD_MBUF_FAST_FREE +* is enabled. +*/ +static inline void +bnxt_tx_cmp_fast (struct bnxt_tx_queue *txq,int nr_pkts) +{ +struct bnxt_tx_ring_info *txr = txq->tx_ring; +uint32_t ring_mask = txr->tx_ring_struct->ring_mask; +struct rte_mbuf **free = txq->free; +uint16_t cons = txr->tx_cons; +unsigned int blk =0 ; + +while (nr_pkts--) { +struct bnxt_sw_tx_bd *tx_buf; + + tx_buf = &txr->tx_buf_ring[cons]; + cons = (cons +1 ) & ring_mask; + free[blk++] = tx_buf->mbuf; + tx_buf->mbuf =NULL ; + } +if (blk) +rte_mempool_put_bulk (free[0 ]->pool, (void **)free, blk); + + txr->tx_cons = cons; +} +void bnxt_free_tx_rings (struct bnxt *bp);int bnxt_init_one_tx_ring (struct bnxt_tx_queue *txq);int bnxt_init_tx_ring_struct (struct bnxt_tx_queue *txq,unsigned int socket_id);
Signature · 1 line
2.25.1
Quoted 9 lines · Lance Richardson
Add support for
DEV_TX_OFFLOAD_MBUF_FAST_FREEto bnxt vector mode transmit. This offload may be enabled only when multi-segment transmit is not needed, all transmitted mbufs for a given queue will be allocated from the same pool, and all transmitted mbufs will have a reference count of 1.Signed-off-by Lance Richardson <[email protected]>
Patch applied to dpdk-next-net-brcm. Thanks