examples/ioat: use always same lcore for enqueue/dequeue
[dpdk.git] / examples / ioat / ioatfwd.c
index b1aa276..fd690ce 100644 (file)
@@ -331,43 +331,36 @@ update_mac_addrs(struct rte_mbuf *m, uint32_t dest_portid)
 
 /* Perform packet copy there is a user-defined function. 8< */
 static inline void
-pktmbuf_sw_copy(struct rte_mbuf *src, struct rte_mbuf *dst)
+pktmbuf_metadata_copy(const struct rte_mbuf *src, struct rte_mbuf *dst)
 {
-       /* Copy packet metadata */
-       rte_memcpy(&dst->rearm_data,
-               &src->rearm_data,
-               offsetof(struct rte_mbuf, cacheline1)
-               - offsetof(struct rte_mbuf, rearm_data));
+       dst->data_off = src->data_off;
+       memcpy(&dst->rx_descriptor_fields1, &src->rx_descriptor_fields1,
+               offsetof(struct rte_mbuf, buf_len) -
+               offsetof(struct rte_mbuf, rx_descriptor_fields1));
+}
 
-       /* Copy packet data */
+/* Copy packet data */
+static inline void
+pktmbuf_sw_copy(struct rte_mbuf *src, struct rte_mbuf *dst)
+{
        rte_memcpy(rte_pktmbuf_mtod(dst, char *),
                rte_pktmbuf_mtod(src, char *), src->data_len);
 }
 /* >8 End of perform packet copy there is a user-defined function. */
 
 static uint32_t
-ioat_enqueue_packets(struct rte_mbuf **pkts,
+ioat_enqueue_packets(struct rte_mbuf *pkts[], struct rte_mbuf *pkts_copy[],
        uint32_t nb_rx, uint16_t dev_id)
 {
        int ret;
        uint32_t i;
-       struct rte_mbuf *pkts_copy[MAX_PKT_BURST];
-
-       const uint64_t addr_offset = RTE_PTR_DIFF(pkts[0]->buf_addr,
-               &pkts[0]->rearm_data);
-
-       ret = rte_mempool_get_bulk(ioat_pktmbuf_pool,
-               (void *)pkts_copy, nb_rx);
-
-       if (unlikely(ret < 0))
-               rte_exit(EXIT_FAILURE, "Unable to allocate memory.\n");
 
        for (i = 0; i < nb_rx; i++) {
                /* Perform data copy */
                ret = rte_ioat_enqueue_copy(dev_id,
-                       pkts[i]->buf_iova - addr_offset,
-                       pkts_copy[i]->buf_iova - addr_offset,
-                       rte_pktmbuf_data_len(pkts[i]) + addr_offset,
+                       rte_pktmbuf_iova(pkts[i]),
+                       rte_pktmbuf_iova(pkts_copy[i]),
+                       rte_pktmbuf_data_len(pkts[i]),
                        (uintptr_t)pkts[i],
                        (uintptr_t)pkts_copy[i]);
 
@@ -376,20 +369,50 @@ ioat_enqueue_packets(struct rte_mbuf **pkts,
        }
 
        ret = i;
-       /* Free any not enqueued packets. */
-       rte_mempool_put_bulk(ioat_pktmbuf_pool, (void *)&pkts[i], nb_rx - i);
-       rte_mempool_put_bulk(ioat_pktmbuf_pool, (void *)&pkts_copy[i],
-               nb_rx - i);
-
        return ret;
 }
 
+static inline uint32_t
+ioat_enqueue(struct rte_mbuf *pkts[], struct rte_mbuf *pkts_copy[],
+       uint32_t num, uint16_t dev_id)
+{
+       uint32_t n;
+
+       n = ioat_enqueue_packets(pkts, pkts_copy, num, dev_id);
+       if (n > 0)
+               rte_ioat_perform_ops(dev_id);
+
+       return n;
+}
+
+static inline uint32_t
+ioat_dequeue(struct rte_mbuf *src[], struct rte_mbuf *dst[], uint32_t num,
+       uint16_t dev_id)
+{
+       int32_t rc;
+       /* Dequeue the mbufs from IOAT device. Since all memory
+        * is DPDK pinned memory and therefore all addresses should
+        * be valid, we don't check for copy errors
+        */
+       rc = rte_ioat_completed_ops(dev_id, num, NULL, NULL,
+               (void *)src, (void *)dst);
+       if (rc < 0) {
+               RTE_LOG(CRIT, IOAT,
+                       "rte_ioat_completed_ops(%hu) failedi, error: %d\n",
+                       dev_id, rte_errno);
+               rc = 0;
+       }
+       return rc;
+}
+
 /* Receive packets on one port and enqueue to IOAT rawdev or rte_ring. 8< */
 static void
 ioat_rx_port(struct rxtx_port_config *rx_config)
 {
+       int32_t ret;
        uint32_t nb_rx, nb_enq, i, j;
        struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
+       struct rte_mbuf *pkts_burst_copy[MAX_PKT_BURST];
 
        for (i = 0; i < rx_config->nb_queues; i++) {
 
@@ -401,40 +424,54 @@ ioat_rx_port(struct rxtx_port_config *rx_config)
 
                port_statistics.rx[rx_config->rxtx_port] += nb_rx;
 
+               ret = rte_mempool_get_bulk(ioat_pktmbuf_pool,
+                       (void *)pkts_burst_copy, nb_rx);
+
+               if (unlikely(ret < 0))
+                       rte_exit(EXIT_FAILURE,
+                               "Unable to allocate memory.\n");
+
+               for (j = 0; j < nb_rx; j++)
+                       pktmbuf_metadata_copy(pkts_burst[j],
+                               pkts_burst_copy[j]);
+
                if (copy_mode == COPY_MODE_IOAT_NUM) {
-                       /* Perform packet hardware copy */
-                       nb_enq = ioat_enqueue_packets(pkts_burst,
+
+                       /* enqueue packets for  hardware copy */
+                       nb_enq = ioat_enqueue(pkts_burst, pkts_burst_copy,
                                nb_rx, rx_config->ioat_ids[i]);
-                       if (nb_enq > 0)
-                               rte_ioat_perform_ops(rx_config->ioat_ids[i]);
-               } else {
-                       /* Perform packet software copy, free source packets */
-                       int ret;
-                       struct rte_mbuf *pkts_burst_copy[MAX_PKT_BURST];
 
-                       ret = rte_mempool_get_bulk(ioat_pktmbuf_pool,
-                               (void *)pkts_burst_copy, nb_rx);
+                       /* free any not enqueued packets. */
+                       rte_mempool_put_bulk(ioat_pktmbuf_pool,
+                               (void *)&pkts_burst[nb_enq],
+                               nb_rx - nb_enq);
+                       rte_mempool_put_bulk(ioat_pktmbuf_pool,
+                               (void *)&pkts_burst_copy[nb_enq],
+                               nb_rx - nb_enq);
 
-                       if (unlikely(ret < 0))
-                               rte_exit(EXIT_FAILURE,
-                                       "Unable to allocate memory.\n");
+                       port_statistics.copy_dropped[rx_config->rxtx_port] +=
+                               (nb_rx - nb_enq);
 
+                       /* get completed copies */
+                       nb_rx = ioat_dequeue(pkts_burst, pkts_burst_copy,
+                               MAX_PKT_BURST, rx_config->ioat_ids[i]);
+               } else {
+                       /* Perform packet software copy, free source packets */
                        for (j = 0; j < nb_rx; j++)
                                pktmbuf_sw_copy(pkts_burst[j],
                                        pkts_burst_copy[j]);
+               }
 
-                       rte_mempool_put_bulk(ioat_pktmbuf_pool,
-                               (void *)pkts_burst, nb_rx);
+               rte_mempool_put_bulk(ioat_pktmbuf_pool,
+                       (void *)pkts_burst, nb_rx);
 
-                       nb_enq = rte_ring_enqueue_burst(
-                               rx_config->rx_to_tx_ring,
-                               (void *)pkts_burst_copy, nb_rx, NULL);
+               nb_enq = rte_ring_enqueue_burst(rx_config->rx_to_tx_ring,
+                       (void *)pkts_burst_copy, nb_rx, NULL);
 
-                       /* Free any not enqueued packets. */
-                       rte_mempool_put_bulk(ioat_pktmbuf_pool,
-                               (void *)&pkts_burst_copy[nb_enq],
-                               nb_rx - nb_enq);
-               }
+               /* Free any not enqueued packets. */
+               rte_mempool_put_bulk(ioat_pktmbuf_pool,
+                       (void *)&pkts_burst_copy[nb_enq],
+                       nb_rx - nb_enq);
 
                port_statistics.copy_dropped[rx_config->rxtx_port] +=
                        (nb_rx - nb_enq);
@@ -446,51 +483,33 @@ ioat_rx_port(struct rxtx_port_config *rx_config)
 static void
 ioat_tx_port(struct rxtx_port_config *tx_config)
 {
-       uint32_t i, j, nb_dq = 0;
-       struct rte_mbuf *mbufs_src[MAX_PKT_BURST];
-       struct rte_mbuf *mbufs_dst[MAX_PKT_BURST];
+       uint32_t i, j, nb_dq, nb_tx;
+       struct rte_mbuf *mbufs[MAX_PKT_BURST];
 
        for (i = 0; i < tx_config->nb_queues; i++) {
-               if (copy_mode == COPY_MODE_IOAT_NUM) {
-                       /* Dequeue the mbufs from IOAT device. Since all memory
-                        * is DPDK pinned memory and therefore all addresses should
-                        * be valid, we don't check for copy errors
-                        */
-                       nb_dq = rte_ioat_completed_ops(
-                               tx_config->ioat_ids[i], MAX_PKT_BURST, NULL, NULL,
-                               (void *)mbufs_src, (void *)mbufs_dst);
-               } else {
-                       /* Dequeue the mbufs from rx_to_tx_ring. */
-                       nb_dq = rte_ring_dequeue_burst(
-                               tx_config->rx_to_tx_ring, (void *)mbufs_dst,
-                               MAX_PKT_BURST, NULL);
-               }
-
-               if ((int32_t) nb_dq <= 0)
-                       return;
 
-               if (copy_mode == COPY_MODE_IOAT_NUM)
-                       rte_mempool_put_bulk(ioat_pktmbuf_pool,
-                               (void *)mbufs_src, nb_dq);
+               /* Dequeue the mbufs from rx_to_tx_ring. */
+               nb_dq = rte_ring_dequeue_burst(tx_config->rx_to_tx_ring,
+                               (void *)mbufs, MAX_PKT_BURST, NULL);
+               if (nb_dq == 0)
+                       continue;
 
                /* Update macs if enabled */
                if (mac_updating) {
                        for (j = 0; j < nb_dq; j++)
-                               update_mac_addrs(mbufs_dst[j],
+                               update_mac_addrs(mbufs[j],
                                        tx_config->rxtx_port);
                }
 
-               const uint16_t nb_tx = rte_eth_tx_burst(
-                       tx_config->rxtx_port, 0,
-                       (void *)mbufs_dst, nb_dq);
+               nb_tx = rte_eth_tx_burst(tx_config->rxtx_port, 0,
+                               (void *)mbufs, nb_dq);
 
                port_statistics.tx[tx_config->rxtx_port] += nb_tx;
 
                /* Free any unsent packets. */
                if (unlikely(nb_tx < nb_dq))
                        rte_mempool_put_bulk(ioat_pktmbuf_pool,
-                       (void *)&mbufs_dst[nb_tx],
-                               nb_dq - nb_tx);
+                       (void *)&mbufs[nb_tx], nb_dq - nb_tx);
        }
 }
 /* >8 End of transmitting packets from IOAT. */
@@ -852,9 +871,6 @@ port_init(uint16_t portid, struct rte_mempool *mbuf_pool, uint16_t nb_queues)
 
        local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
                dev_info.flow_type_rss_offloads;
-       if (dev_info.tx_offload_capa & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE)
-               local_port_conf.txmode.offloads |=
-                       RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
        ret = rte_eth_dev_configure(portid, nb_queues, 1, &local_port_conf);
        if (ret < 0)
                rte_exit(EXIT_FAILURE, "Cannot configure device:"
@@ -973,7 +989,8 @@ main(int argc, char **argv)
 
        /* Allocates mempool to hold the mbufs. 8< */
        nb_mbufs = RTE_MAX(nb_ports * (nb_queues * (nb_rxd + nb_txd +
-               4 * MAX_PKT_BURST) + rte_lcore_count() * MEMPOOL_CACHE_SIZE),
+               4 * MAX_PKT_BURST + ring_size) + ring_size +
+               rte_lcore_count() * MEMPOOL_CACHE_SIZE),
                MIN_POOL_SIZE);
 
        /* Create the mbuf pool */
@@ -1005,8 +1022,8 @@ main(int argc, char **argv)
 
        if (copy_mode == COPY_MODE_IOAT_NUM)
                assign_rawdevs();
-       else /* copy_mode == COPY_MODE_SW_NUM */
-               assign_rings();
+
+       assign_rings();
        /* >8 End of assigning each port resources. */
 
        start_forwarding_cores();