net/dpaa2: avoid multiple PtoV calls
PATCH 1/1 net/dpaa2: avoid multiple PtoV calls
Avoids recalculation of DPAA2_IOVA_TO_VADDR, since this is somewhat costly function on no-IOMMU setups.
Brings about 0.5 Gbps improvement at 19 Gbps speeds on LX2160 with 512b packets.
Signed-off-by Roman Kapl <[email protected]>
--- drivers/net/dpaa2/dpaa2_rxtx.c | 32 +++++++++++++++----------------- 1 file changed, 15 insertions(+), 17 deletions(-)
drivers/net/dpaa2/dpaa2_rxtx.c +15 -17
@@ -42,10 +42,12 @@ static void enable_tx_tstamp(struct qbman_fd *fd) __rte_unused; }while (0 ) static inline void __rte_hot -dpaa2_dev_rx_parse_new (struct rte_mbuf *m,const struct qbman_fd *fd) +dpaa2_dev_rx_parse_new (struct rte_mbuf *m,const struct qbman_fd *fd, +void *hw_annot_addr) { -struct dpaa2_annot_hdr *annotation;uint16_t frc =DPAA2_GET_FD_FRC_PARSE_SUM (fd); +struct dpaa2_annot_hdr *annotation = + (struct dpaa2_annot_hdr *)hw_annot_addr; m->packet_type =RTE_PTYPE_UNKNOWN ;switch (frc) {
@@ -101,17 +103,12 @@ dpaa2_dev_rx_parse_new(struct rte_mbuf *m, const struct qbman_fd *fd)RTE_PTYPE_L3_IPV6 |RTE_PTYPE_L4_ICMP ;break ;default : - m->packet_type =dpaa2_dev_rx_parse_slow (m, - (void *)((size_t )DPAA2_IOVA_TO_VADDR (DPAA2_GET_FD_ADDR (fd)) - +DPAA2_FD_PTA_SIZE )); + m->packet_type =dpaa2_dev_rx_parse_slow (m, annotation); } m->hash.rss = fd->simple.flc_hi; m->ol_flags |=PKT_RX_RSS_HASH ;if (dpaa2_enable_ts ==PMD_DPAA2_ENABLE_TS ) { - annotation = (struct dpaa2_annot_hdr *) - ((size_t )DPAA2_IOVA_TO_VADDR ( -DPAA2_GET_FD_ADDR (fd)) +DPAA2_FD_PTA_SIZE ); m->timestamp = annotation->word2; m->ol_flags |=PKT_RX_TIMESTAMP ;DPAA2_PMD_DP_DEBUG ("pkt timestamp:0x%" PRIx64"" , m->timestamp);
@@ -266,9 +263,11 @@ eth_sg_fd_to_mbuf(const struct qbman_fd *fd,struct qbman_sge *sgt, *sge;size_t sg_addr, fd_addr;int i =0 ; +void *hw_annot_addr;struct rte_mbuf *first_seg, *next_seg, *cur_seg, *temp; fd_addr = (size_t )DPAA2_IOVA_TO_VADDR (DPAA2_GET_FD_ADDR (fd)); + hw_annot_addr = (void *)(fd_addr +DPAA2_FD_PTA_SIZE );/* Get Scatter gather table address */ sgt = (struct qbman_sge *)(fd_addr +DPAA2_GET_FD_OFFSET (fd));
@@ -289,11 +288,10 @@ eth_sg_fd_to_mbuf(const struct qbman_fd *fd, first_seg->next =NULL ; first_seg->port = port_id;if (dpaa2_svr_family ==SVR_LX2160A ) -dpaa2_dev_rx_parse_new (first_seg, fd); +dpaa2_dev_rx_parse_new (first_seg, fd, hw_annot_addr);else - first_seg->packet_type =dpaa2_dev_rx_parse (first_seg, - (void *)((size_t )DPAA2_IOVA_TO_VADDR (DPAA2_GET_FD_ADDR (fd)) - +DPAA2_FD_PTA_SIZE )); + first_seg->packet_type = +dpaa2_dev_rx_parse (first_seg, hw_annot_addr);rte_mbuf_refcnt_set (first_seg,1 ); cur_seg = first_seg;
@@ -324,8 +322,9 @@ static inline struct rte_mbuf *__rte_hot eth_fd_to_mbuf(const struct qbman_fd *fd, int port_id) { - void *iova_addr = DPAA2_IOVA_TO_VADDR(DPAA2_GET_FD_ADDR(fd)); -struct rte_mbuf *mbuf =DPAA2_INLINE_MBUF_FROM_BUF (iova_addr, + void *v_addr = DPAA2_IOVA_TO_VADDR(DPAA2_GET_FD_ADDR(fd)); +void *hw_annot_addr = (void *)((size_t )v_addr +DPAA2_FD_PTA_SIZE ); +struct rte_mbuf *mbuf =DPAA2_INLINE_MBUF_FROM_BUF (v_addr, rte_dpaa2_bpid_info[DPAA2_GET_FD_BPID (fd)].meta_data_size);/* need to repopulated some of the fields,
@@ -347,10 +346,9 @@ eth_fd_to_mbuf(const struct qbman_fd *fd,*/ if (dpaa2_svr_family ==SVR_LX2160A ) -dpaa2_dev_rx_parse_new (mbuf, fd); +dpaa2_dev_rx_parse_new (mbuf, fd, hw_annot_addr);else - mbuf->packet_type =dpaa2_dev_rx_parse (mbuf, - (void *)((size_t )iova_addr +DPAA2_FD_PTA_SIZE )); + mbuf->packet_type =dpaa2_dev_rx_parse (mbuf, hw_annot_addr); DPAA2_PMD_DP_DEBUG("to mbuf - mbuf =%p, mbuf->buf_addr =%p, off = %d," "fd_off=%d fd =%" PRIx64 ", meta = %d bpid =%d, len=%d\n",
Signature · 1 line
2.20.1
Reviewed-by Hemant Agrawal <[email protected]>
Quoted 7 lines · Roman Kapl
Avoids recalculation of
DPAA2_IOVA_TO_VADDR, since this is somewhat costly function on no-IOMMU setups.Brings about 0.5 Gbps improvement at 19 Gbps speeds on LX2160 with 512b packets.
Signed-off-by Roman Kapl <[email protected]>>
Reviewed-by Hemant Agrawal <[email protected]>
Applied to dpdk-next-net/master, thanks.