net/bnxt: improve CPR handling in vector path

Reduce overhead of CPR descriptor validity checking in vector
receive and transmit functions.

Preserve raw cpr consumer index in vector transmit completion
function.

Remove an unneeded prefetch (per benchmarking) from vector
transmit completion function.

Fixes: bc4a000f2f53 ("net/bnxt: implement SSE vector mode")
Cc: stable@dpdk.org
Signed-off-by: Lance Richardson <lance.richardson@broadcom.com>
Reviewed-by: Ajit Khaparde <ajit.khaparde@broadcom.com>

diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
index 029053e..22d9f9e 100644 (file)
--- a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
@@ -245,10 +245,6 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
                if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
                        break;
 
-               cpr->valid = FLIP_VALID(cons,
-                                       cpr->cp_ring_struct->ring_mask,
-                                       cpr->valid);
-
                if (likely(CMP_TYPE(rxcmp) == RX_PKT_CMPL_TYPE_RX_L2)) {
                        struct rx_pkt_cmpl_hi *rxcmp1;
                        uint32_t tmp_raw_cons;
@@ -272,10 +268,6 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
                        rte_prefetch0(mbuf);
                        rxr->rx_buf_ring[cons].mbuf = NULL;
 
-                       cpr->valid = FLIP_VALID(cp_cons,
-                                               cpr->cp_ring_struct->ring_mask,
-                                               cpr->valid);
-
                        /* Set constant fields from mbuf initializer. */
                        _mm_store_si128((__m128i *)&mbuf->rearm_data,
                                        mbuf_init);
@@ -318,22 +310,13 @@ bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 
        rxq->rxrearm_nb += nb_rx_pkts;
        cpr->cp_raw_cons = raw_cons;
+       cpr->valid = !!(cpr->cp_raw_cons & cpr->cp_ring_struct->ring_size);
        if (nb_rx_pkts || evt)
                bnxt_db_cq(cpr);
 
        return nb_rx_pkts;
 }
 
-static inline void bnxt_next_cmpl(struct bnxt_cp_ring_info *cpr, uint32_t *idx,
-                                 bool *v, uint32_t inc)
-{
-       *idx += inc;
-       if (unlikely(*idx == cpr->cp_ring_struct->ring_size)) {
-               *v = !*v;
-               *idx = 0;
-       }
-}
-
 static void
 bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
 {
@@ -379,10 +362,8 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
                cons = RING_CMPL(ring_mask, raw_cons);
                txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
 
-               if (!CMPL_VALID(txcmp, cpr->valid))
+               if (!CMP_VALID(txcmp, raw_cons, cp_ring_struct))
                        break;
-               bnxt_next_cmpl(cpr, &cons, &cpr->valid, 1);
-               rte_prefetch0(&cp_desc_ring[cons]);
 
                if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
                        nb_tx_pkts += txcmp->opaque;
@@ -390,9 +371,10 @@ bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
                        RTE_LOG_DP(ERR, PMD,
                                   "Unhandled CMP type %02x\n",
                                   CMP_TYPE(txcmp));
-               raw_cons = cons;
+               raw_cons = NEXT_RAW_CMP(raw_cons);
        } while (nb_tx_pkts < ring_mask);
 
+       cpr->valid = !!(raw_cons & cp_ring_struct->ring_size);
        if (nb_tx_pkts) {
                bnxt_tx_cmp_vec(txq, nb_tx_pkts);
                cpr->cp_raw_cons = raw_cons;