net/qede: fix performance bottleneck in Rx path
authorShahed Shaikh <shshaikh@marvell.com>
Fri, 18 Jan 2019 10:29:29 +0000 (02:29 -0800)
committerThomas Monjalon <thomas@monjalon.net>
Fri, 18 Jan 2019 23:36:17 +0000 (00:36 +0100)
Allocating replacement buffer per received packet is expensive.
Instead, process received packets first and allocate
replacement buffers in bulk later.

This improves performance by ~25% in terms of PPS on AMD
platforms.

Fixes: 2ea6f76aff40 ("qede: add core driver")
Cc: stable@dpdk.org
Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
Acked-by: Rasesh Mody <rmody@marvell.com>
drivers/net/qede/qede_rxtx.c
drivers/net/qede/qede_rxtx.h

index 0e33be1..684c4ae 100644 (file)
@@ -35,6 +35,52 @@ static inline int qede_alloc_rx_buffer(struct qede_rx_queue *rxq)
        return 0;
 }
 
+#define QEDE_MAX_BULK_ALLOC_COUNT 512
+
+static inline int qede_alloc_rx_bulk_mbufs(struct qede_rx_queue *rxq, int count)
+{
+       void *obj_p[QEDE_MAX_BULK_ALLOC_COUNT] __rte_cache_aligned;
+       struct rte_mbuf *mbuf = NULL;
+       struct eth_rx_bd *rx_bd;
+       dma_addr_t mapping;
+       int i, ret = 0;
+       uint16_t idx;
+
+       idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
+
+       if (count > QEDE_MAX_BULK_ALLOC_COUNT)
+               count = QEDE_MAX_BULK_ALLOC_COUNT;
+
+       ret = rte_mempool_get_bulk(rxq->mb_pool, obj_p, count);
+       if (unlikely(ret)) {
+               PMD_RX_LOG(ERR, rxq,
+                          "Failed to allocate %d rx buffers "
+                           "sw_rx_prod %u sw_rx_cons %u mp entries %u free %u",
+                           count, idx, rxq->sw_rx_cons & NUM_RX_BDS(rxq),
+                           rte_mempool_avail_count(rxq->mb_pool),
+                           rte_mempool_in_use_count(rxq->mb_pool));
+               return -ENOMEM;
+       }
+
+       for (i = 0; i < count; i++) {
+               mbuf = obj_p[i];
+               if (likely(i < count - 1))
+                       rte_prefetch0(obj_p[i + 1]);
+
+               idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
+               rxq->sw_rx_ring[idx].mbuf = mbuf;
+               rxq->sw_rx_ring[idx].page_offset = 0;
+               mapping = rte_mbuf_data_iova_default(mbuf);
+               rx_bd = (struct eth_rx_bd *)
+                       ecore_chain_produce(&rxq->rx_bd_ring);
+               rx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
+               rx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
+               rxq->sw_rx_prod++;
+       }
+
+       return 0;
+}
+
 /* Criterias for calculating Rx buffer size -
  * 1) rx_buf_size should not exceed the size of mbuf
  * 2) In scattered_rx mode - minimum rx_buf_size should be
@@ -1131,7 +1177,7 @@ qede_reuse_page(__rte_unused struct qede_dev *qdev,
                struct qede_rx_queue *rxq, struct qede_rx_entry *curr_cons)
 {
        struct eth_rx_bd *rx_bd_prod = ecore_chain_produce(&rxq->rx_bd_ring);
-       uint16_t idx = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
+       uint16_t idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
        struct qede_rx_entry *curr_prod;
        dma_addr_t new_mapping;
 
@@ -1364,7 +1410,6 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
        uint8_t bitfield_val;
 #endif
        uint8_t tunn_parse_flag;
-       uint8_t j;
        struct eth_fast_path_rx_tpa_start_cqe *cqe_start_tpa;
        uint64_t ol_flags;
        uint32_t packet_type;
@@ -1373,6 +1418,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
        uint8_t offset, tpa_agg_idx, flags;
        struct qede_agg_info *tpa_info = NULL;
        uint32_t rss_hash;
+       int rx_alloc_count = 0;
 
        hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
        sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
@@ -1382,6 +1428,25 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
        if (hw_comp_cons == sw_comp_cons)
                return 0;
 
+       /* Allocate buffers that we used in previous loop */
+       if (rxq->rx_alloc_count) {
+               if (unlikely(qede_alloc_rx_bulk_mbufs(rxq,
+                            rxq->rx_alloc_count))) {
+                       struct rte_eth_dev *dev;
+
+                       PMD_RX_LOG(ERR, rxq,
+                                  "New buffer allocation failed,"
+                                  "dropping incoming packetn");
+                       dev = &rte_eth_devices[rxq->port_id];
+                       dev->data->rx_mbuf_alloc_failed +=
+                                                       rxq->rx_alloc_count;
+                       rxq->rx_alloc_errors += rxq->rx_alloc_count;
+                       return 0;
+               }
+               qede_update_rx_prod(qdev, rxq);
+               rxq->rx_alloc_count = 0;
+       }
+
        while (sw_comp_cons != hw_comp_cons) {
                ol_flags = 0;
                packet_type = RTE_PTYPE_UNKNOWN;
@@ -1553,16 +1618,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                        rx_mb->hash.rss = rss_hash;
                }
 
-               if (unlikely(qede_alloc_rx_buffer(rxq) != 0)) {
-                       PMD_RX_LOG(ERR, rxq,
-                                  "New buffer allocation failed,"
-                                  "dropping incoming packet\n");
-                       qede_recycle_rx_bd_ring(rxq, qdev, fp_cqe->bd_num);
-                       rte_eth_devices[rxq->port_id].
-                           data->rx_mbuf_alloc_failed++;
-                       rxq->rx_alloc_errors++;
-                       break;
-               }
+               rx_alloc_count++;
                qede_rx_bd_ring_consume(rxq);
 
                if (!tpa_start_flg && fp_cqe->bd_num > 1) {
@@ -1574,17 +1630,9 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
                        if (qede_process_sg_pkts(p_rxq, seg1, num_segs,
                                                 pkt_len - len))
                                goto next_cqe;
-                       for (j = 0; j < num_segs; j++) {
-                               if (qede_alloc_rx_buffer(rxq)) {
-                                       PMD_RX_LOG(ERR, rxq,
-                                               "Buffer allocation failed");
-                                       rte_eth_devices[rxq->port_id].
-                                               data->rx_mbuf_alloc_failed++;
-                                       rxq->rx_alloc_errors++;
-                                       break;
-                               }
-                               rxq->rx_segs++;
-                       }
+
+                       rx_alloc_count += num_segs;
+                       rxq->rx_segs += num_segs;
                }
                rxq->rx_segs++; /* for the first segment */
 
@@ -1626,7 +1674,8 @@ next_cqe:
                }
        }
 
-       qede_update_rx_prod(qdev, rxq);
+       /* Request number of bufferes to be allocated in next loop */
+       rxq->rx_alloc_count = rx_alloc_count;
 
        rxq->rcv_pkts += rx_pkt;
 
index 454daa0..5b249cb 100644 (file)
@@ -192,6 +192,8 @@ struct qede_rx_queue {
        uint16_t queue_id;
        uint16_t port_id;
        uint16_t rx_buf_size;
+       uint16_t rx_alloc_count;
+       uint16_t unused;
        uint64_t rcv_pkts;
        uint64_t rx_segs;
        uint64_t rx_hw_errors;