gpu/cuda: map GPU memory with GDRCopy
[dpdk.git] / examples / l2fwd-event / l2fwd_event.c
index 7ba5311..6df3cdf 100644 (file)
@@ -5,7 +5,6 @@
 #include <stdbool.h>
 #include <getopt.h>
 
-#include <rte_atomic.h>
 #include <rte_cycles.h>
 #include <rte_ethdev.h>
 #include <rte_eventdev.h>
@@ -346,19 +345,198 @@ l2fwd_event_main_loop_tx_q_brst_mac(struct l2fwd_resources *rsrc)
                        L2FWD_EVENT_TX_ENQ | L2FWD_EVENT_BURST);
 }
 
+static __rte_always_inline void
+l2fwd_event_vector_fwd(struct l2fwd_resources *rsrc,
+                      struct rte_event_vector *vec,
+                      const uint64_t timer_period, const uint32_t flags)
+{
+       struct rte_mbuf **mbufs = vec->mbufs;
+       uint16_t i, j;
+
+       rte_prefetch0(rte_pktmbuf_mtod(mbufs[0], void *));
+
+       /* If vector attribute is valid, mbufs will be from same port/queue */
+       if (vec->attr_valid) {
+               vec->port = rsrc->dst_ports[mbufs[0]->port];
+               if (flags & L2FWD_EVENT_TX_DIRECT)
+                       vec->queue = 0;
+
+               if (timer_period > 0)
+                       __atomic_fetch_add(&rsrc->port_stats[mbufs[0]->port].rx,
+                                          vec->nb_elem, __ATOMIC_RELAXED);
+
+               for (i = 0, j = 1; i < vec->nb_elem; i++, j++) {
+                       if (j < vec->nb_elem)
+                               rte_prefetch0(
+                                       rte_pktmbuf_mtod(mbufs[j], void *));
+
+                       if (flags & L2FWD_EVENT_UPDT_MAC)
+                               l2fwd_mac_updating(
+                                       mbufs[i], vec->port,
+                                       &rsrc->eth_addr[vec->port]);
+               }
+
+               if (timer_period > 0)
+                       __atomic_fetch_add(&rsrc->port_stats[vec->port].tx,
+                                          vec->nb_elem, __ATOMIC_RELAXED);
+       } else {
+               for (i = 0, j = 1; i < vec->nb_elem; i++, j++) {
+                       if (timer_period > 0)
+                               __atomic_fetch_add(
+                                       &rsrc->port_stats[mbufs[i]->port].rx, 1,
+                                       __ATOMIC_RELAXED);
+
+                       if (j < vec->nb_elem)
+                               rte_prefetch0(
+                                       rte_pktmbuf_mtod(mbufs[j], void *));
+
+                       mbufs[i]->port = rsrc->dst_ports[mbufs[i]->port];
+
+                       if (flags & L2FWD_EVENT_UPDT_MAC)
+                               l2fwd_mac_updating(
+                                       mbufs[i], mbufs[i]->port,
+                                       &rsrc->eth_addr[mbufs[i]->port]);
+
+                       if (flags & L2FWD_EVENT_TX_DIRECT)
+                               rte_event_eth_tx_adapter_txq_set(mbufs[i], 0);
+
+                       if (timer_period > 0)
+                               __atomic_fetch_add(
+                                       &rsrc->port_stats[mbufs[i]->port].tx, 1,
+                                       __ATOMIC_RELAXED);
+               }
+       }
+}
+
+static __rte_always_inline void
+l2fwd_event_loop_vector(struct l2fwd_resources *rsrc, const uint32_t flags)
+{
+       struct l2fwd_event_resources *evt_rsrc = rsrc->evt_rsrc;
+       const int port_id = l2fwd_get_free_event_port(evt_rsrc);
+       const uint8_t tx_q_id =
+               evt_rsrc->evq.event_q_id[evt_rsrc->evq.nb_queues - 1];
+       const uint64_t timer_period = rsrc->timer_period;
+       const uint8_t event_d_id = evt_rsrc->event_d_id;
+       const uint8_t deq_len = evt_rsrc->deq_depth;
+       struct rte_event ev[MAX_PKT_BURST];
+       uint16_t nb_rx, nb_tx;
+       uint8_t i;
+
+       if (port_id < 0)
+               return;
+
+       printf("%s(): entering eventdev main loop on lcore %u\n", __func__,
+              rte_lcore_id());
+
+       while (!rsrc->force_quit) {
+               nb_rx = rte_event_dequeue_burst(event_d_id, port_id, ev,
+                                               deq_len, 0);
+               if (nb_rx == 0)
+                       continue;
+
+               for (i = 0; i < nb_rx; i++) {
+                       if (flags & L2FWD_EVENT_TX_ENQ) {
+                               ev[i].queue_id = tx_q_id;
+                               ev[i].op = RTE_EVENT_OP_FORWARD;
+                       }
+
+                       l2fwd_event_vector_fwd(rsrc, ev[i].vec, timer_period,
+                                              flags);
+               }
+
+               if (flags & L2FWD_EVENT_TX_ENQ) {
+                       nb_tx = rte_event_enqueue_burst(event_d_id, port_id, ev,
+                                                       nb_rx);
+                       while (nb_tx < nb_rx && !rsrc->force_quit)
+                               nb_tx += rte_event_enqueue_burst(
+                                       event_d_id, port_id, ev + nb_tx,
+                                       nb_rx - nb_tx);
+               }
+
+               if (flags & L2FWD_EVENT_TX_DIRECT) {
+                       nb_tx = rte_event_eth_tx_adapter_enqueue(
+                               event_d_id, port_id, ev, nb_rx, 0);
+                       while (nb_tx < nb_rx && !rsrc->force_quit)
+                               nb_tx += rte_event_eth_tx_adapter_enqueue(
+                                       event_d_id, port_id, ev + nb_tx,
+                                       nb_rx - nb_tx, 0);
+               }
+       }
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_d_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc, L2FWD_EVENT_TX_DIRECT);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_d_brst_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc, L2FWD_EVENT_TX_DIRECT);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_q_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc, L2FWD_EVENT_TX_ENQ);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_q_brst_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc, L2FWD_EVENT_TX_ENQ);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_d_mac_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc,
+                               L2FWD_EVENT_UPDT_MAC | L2FWD_EVENT_TX_DIRECT);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_d_brst_mac_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc,
+                               L2FWD_EVENT_UPDT_MAC | L2FWD_EVENT_TX_DIRECT);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_q_mac_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc,
+                               L2FWD_EVENT_UPDT_MAC | L2FWD_EVENT_TX_ENQ);
+}
+
+static void __rte_noinline
+l2fwd_event_main_loop_tx_q_brst_mac_vec(struct l2fwd_resources *rsrc)
+{
+       l2fwd_event_loop_vector(rsrc,
+                               L2FWD_EVENT_UPDT_MAC | L2FWD_EVENT_TX_ENQ);
+}
+
 void
 l2fwd_event_resource_setup(struct l2fwd_resources *rsrc)
 {
        /* [MAC_UPDT][TX_MODE][BURST] */
-       const event_loop_cb event_loop[2][2][2] = {
-               [0][0][0] = l2fwd_event_main_loop_tx_d,
-               [0][0][1] = l2fwd_event_main_loop_tx_d_brst,
-               [0][1][0] = l2fwd_event_main_loop_tx_q,
-               [0][1][1] = l2fwd_event_main_loop_tx_q_brst,
-               [1][0][0] = l2fwd_event_main_loop_tx_d_mac,
-               [1][0][1] = l2fwd_event_main_loop_tx_d_brst_mac,
-               [1][1][0] = l2fwd_event_main_loop_tx_q_mac,
-               [1][1][1] = l2fwd_event_main_loop_tx_q_brst_mac,
+       const event_loop_cb event_loop[2][2][2][2] = {
+               [0][0][0][0] = l2fwd_event_main_loop_tx_d,
+               [0][0][0][1] = l2fwd_event_main_loop_tx_d_brst,
+               [0][0][1][0] = l2fwd_event_main_loop_tx_q,
+               [0][0][1][1] = l2fwd_event_main_loop_tx_q_brst,
+               [0][1][0][0] = l2fwd_event_main_loop_tx_d_mac,
+               [0][1][0][1] = l2fwd_event_main_loop_tx_d_brst_mac,
+               [0][1][1][0] = l2fwd_event_main_loop_tx_q_mac,
+               [0][1][1][1] = l2fwd_event_main_loop_tx_q_brst_mac,
+               [1][0][0][0] = l2fwd_event_main_loop_tx_d_vec,
+               [1][0][0][1] = l2fwd_event_main_loop_tx_d_brst_vec,
+               [1][0][1][0] = l2fwd_event_main_loop_tx_q_vec,
+               [1][0][1][1] = l2fwd_event_main_loop_tx_q_brst_vec,
+               [1][1][0][0] = l2fwd_event_main_loop_tx_d_mac_vec,
+               [1][1][0][1] = l2fwd_event_main_loop_tx_d_brst_mac_vec,
+               [1][1][1][0] = l2fwd_event_main_loop_tx_q_mac_vec,
+               [1][1][1][1] = l2fwd_event_main_loop_tx_q_brst_mac_vec,
        };
        struct l2fwd_event_resources *evt_rsrc;
        uint32_t event_queue_cfg;
@@ -394,8 +572,7 @@ l2fwd_event_resource_setup(struct l2fwd_resources *rsrc)
        if (ret < 0)
                rte_panic("Error in starting eventdev\n");
 
-       evt_rsrc->ops.l2fwd_event_loop = event_loop
-                                       [rsrc->mac_updating]
-                                       [evt_rsrc->tx_mode_q]
-                                       [evt_rsrc->has_burst];
+       evt_rsrc->ops.l2fwd_event_loop =
+               event_loop[rsrc->evt_vec.enabled][rsrc->mac_updating]
+                         [evt_rsrc->tx_mode_q][evt_rsrc->has_burst];
 }