examples/l2fwd-event: support event vector
authorShijith Thotton <sthotton@marvell.com>
Mon, 27 Sep 2021 06:21:35 +0000 (11:51 +0530)
committerJerin Jacob <jerinj@marvell.com>
Thu, 21 Oct 2021 08:16:00 +0000 (10:16 +0200)
Added changes to receive packets as event vector. By default this is
disabled and can be enabled using the option --event-vector. Vector
size and timeout to form the vector can be configured using options
--event-vector-size and --event-vector-tmo.

Example:
dpdk-l2fwd-event -l 0-3 -n 4 -- -p 0x03 --mode=eventdev \
--eventq-sched=ordered --event-vector --event-vector-size 16

Signed-off-by: Shijith Thotton <sthotton@marvell.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
doc/guides/sample_app_ug/l2_forward_event.rst
examples/l2fwd-event/l2fwd_common.h
examples/l2fwd-event/l2fwd_event.c
examples/l2fwd-event/l2fwd_event_generic.c
examples/l2fwd-event/l2fwd_event_internal_port.c
examples/l2fwd-event/main.c

index 540ca01..904f6f1 100644 (file)
@@ -52,7 +52,12 @@ The application requires a number of command line options:
 
 .. code-block:: console
 
-    ./<build_dir>/examples/dpdk-l2fwd-event [EAL options] -- -p PORTMASK [-q NQ] --[no-]mac-updating --mode=MODE --eventq-sched=SCHED_MODE
+    ./<build_dir>/examples/dpdk-l2fwd-event [EAL options] -- -p PORTMASK
+                                                        [-q NQ]
+                                                        [--[no-]mac-updating]
+                                                        [--mode=MODE]
+                                                        [--eventq-sched=SCHED_MODE]
+                                                        [--event-vector [--event-vector-size SIZE] [--event-vector-tmo NS]]
 
 where,
 
@@ -68,6 +73,12 @@ where,
 
 *   --config: Configure forwarding port pair mapping. Alternate port pairs by default.
 
+*   --event-vector: Enable event vectorization. Only valid if --mode=eventdev.
+
+*   --event-vector-size: Max vector size if event vectorization is enabled.
+
+*   --event-vector-tmo: Max timeout to form vector in nanoseconds if event vectorization is enabled.
+
 Sample usage commands are given below to run the application into different mode:
 
 Poll mode with 4 lcores, 16 ports and 8 RX queues per lcore and MAC address updating enabled,
index ad453b4..396e238 100644 (file)
@@ -55,6 +55,9 @@
 #define DEFAULT_TIMER_PERIOD   10 /* default period is 10 seconds */
 #define MAX_TIMER_PERIOD       86400 /* 1 day max */
 
+#define VECTOR_SIZE_DEFAULT   MAX_PKT_BURST
+#define VECTOR_TMO_NS_DEFAULT 1E6 /* 1ms */
+
 /* Per-port statistics struct */
 struct l2fwd_port_statistics {
        uint64_t dropped;
@@ -62,6 +65,13 @@ struct l2fwd_port_statistics {
        uint64_t rx;
 } __rte_cache_aligned;
 
+/* Event vector attributes */
+struct l2fwd_event_vector_params {
+       uint8_t enabled;
+       uint16_t size;
+       uint64_t timeout_ns;
+};
+
 struct l2fwd_resources {
        volatile uint8_t force_quit;
        uint8_t event_mode;
@@ -74,9 +84,11 @@ struct l2fwd_resources {
        uint32_t enabled_port_mask;
        uint64_t timer_period;
        struct rte_mempool *pktmbuf_pool;
+       struct rte_mempool *evt_vec_pool;
        uint32_t dst_ports[RTE_MAX_ETHPORTS];
        struct rte_ether_addr eth_addr[RTE_MAX_ETHPORTS];
        struct l2fwd_port_statistics port_stats[RTE_MAX_ETHPORTS];
+       struct l2fwd_event_vector_params evt_vec;
        void *evt_rsrc;
        void *poll_rsrc;
 } __rte_cache_aligned;
index acfacfa..6df3cdf 100644 (file)
@@ -345,19 +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;
@@ -393,8 +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];
 }
index c19b6f1..f31569a 100644 (file)
@@ -245,6 +245,27 @@ l2fwd_rx_tx_adapter_setup_generic(struct l2fwd_resources *rsrc)
                if ((rsrc->enabled_port_mask & (1 << port_id)) == 0)
                        continue;
                eth_q_conf.ev.queue_id = evt_rsrc->evq.event_q_id[i];
+               if (rsrc->evt_vec.enabled) {
+                       uint32_t cap;
+
+                       if (rte_event_eth_rx_adapter_caps_get(event_d_id,
+                                                             port_id, &cap))
+                               rte_panic(
+                                       "Failed to get event rx adapter capability");
+
+                       if (cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR) {
+                               eth_q_conf.vector_sz = rsrc->evt_vec.size;
+                               eth_q_conf.vector_timeout_ns =
+                                       rsrc->evt_vec.timeout_ns;
+                               eth_q_conf.vector_mp = rsrc->evt_vec_pool;
+                               eth_q_conf.rx_queue_flags |=
+                               RTE_EVENT_ETH_RX_ADAPTER_QUEUE_EVENT_VECTOR;
+                       } else {
+                               rte_panic(
+                                       "Rx adapter doesn't support event vector");
+                       }
+               }
+
                ret = rte_event_eth_rx_adapter_queue_add(rx_adptr_id, port_id,
                                                         -1, &eth_q_conf);
                if (ret)
index 16a52d6..86d772d 100644 (file)
@@ -230,6 +230,28 @@ l2fwd_rx_tx_adapter_setup_internal_port(struct l2fwd_resources *rsrc)
        RTE_ETH_FOREACH_DEV(port_id) {
                if ((rsrc->enabled_port_mask & (1 << port_id)) == 0)
                        continue;
+
+               if (rsrc->evt_vec.enabled) {
+                       uint32_t cap;
+
+                       if (rte_event_eth_rx_adapter_caps_get(event_d_id,
+                                                             port_id, &cap))
+                               rte_panic(
+                                       "Failed to get event rx adapter capability");
+
+                       if (cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR) {
+                               eth_q_conf.vector_sz = rsrc->evt_vec.size;
+                               eth_q_conf.vector_timeout_ns =
+                                       rsrc->evt_vec.timeout_ns;
+                               eth_q_conf.vector_mp = rsrc->evt_vec_pool;
+                               eth_q_conf.rx_queue_flags |=
+                               RTE_EVENT_ETH_RX_ADAPTER_QUEUE_EVENT_VECTOR;
+                       } else {
+                               rte_panic(
+                                       "Rx adapter doesn't support event vector");
+                       }
+               }
+
                ret = rte_event_eth_rx_adapter_create(adapter_id, event_d_id,
                                                &evt_rsrc->def_p_conf);
                if (ret)
index 1db89f2..f2adc66 100644 (file)
@@ -25,6 +25,9 @@ l2fwd_event_usage(const char *prgname)
               "  --eventq-sched: Event queue schedule type, ordered, atomic or parallel.\n"
               "                  Default: atomic\n"
               "                  Valid only if --mode=eventdev\n"
+              "  --event-vector:  Enable event vectorization.\n"
+              "  --event-vector-size: Max vector size if event vectorization is enabled.\n"
+              "  --event-vector-tmo: Max timeout to form vector in nanoseconds if event vectorization is enabled\n"
               "  --config: Configure forwarding port pair mapping\n"
               "            Default: alternate port pairs\n\n",
               prgname);
@@ -175,6 +178,9 @@ static const char short_options[] =
 #define CMD_LINE_OPT_MODE "mode"
 #define CMD_LINE_OPT_EVENTQ_SCHED "eventq-sched"
 #define CMD_LINE_OPT_PORT_PAIR_CONF "config"
+#define CMD_LINE_OPT_ENABLE_VECTOR "event-vector"
+#define CMD_LINE_OPT_VECTOR_SIZE "event-vector-size"
+#define CMD_LINE_OPT_VECTOR_TMO_NS "event-vector-tmo"
 
 enum {
        /* long options mapped to a short option */
@@ -186,6 +192,9 @@ enum {
        CMD_LINE_OPT_MODE_NUM,
        CMD_LINE_OPT_EVENTQ_SCHED_NUM,
        CMD_LINE_OPT_PORT_PAIR_CONF_NUM,
+       CMD_LINE_OPT_ENABLE_VECTOR_NUM,
+       CMD_LINE_OPT_VECTOR_SIZE_NUM,
+       CMD_LINE_OPT_VECTOR_TMO_NS_NUM
 };
 
 /* Parse the argument given in the command line of the application */
@@ -202,6 +211,12 @@ l2fwd_event_parse_args(int argc, char **argv, struct l2fwd_resources *rsrc)
                                                CMD_LINE_OPT_EVENTQ_SCHED_NUM},
                { CMD_LINE_OPT_PORT_PAIR_CONF, required_argument, NULL,
                                        CMD_LINE_OPT_PORT_PAIR_CONF_NUM},
+               {CMD_LINE_OPT_ENABLE_VECTOR, no_argument, NULL,
+                                       CMD_LINE_OPT_ENABLE_VECTOR_NUM},
+               {CMD_LINE_OPT_VECTOR_SIZE, required_argument, NULL,
+                                       CMD_LINE_OPT_VECTOR_SIZE_NUM},
+               {CMD_LINE_OPT_VECTOR_TMO_NS, required_argument, NULL,
+                                       CMD_LINE_OPT_VECTOR_TMO_NS_NUM},
                {NULL, 0, 0, 0}
        };
        int opt, ret, timer_secs;
@@ -270,6 +285,16 @@ l2fwd_event_parse_args(int argc, char **argv, struct l2fwd_resources *rsrc)
                                return -1;
                        }
                        break;
+               case CMD_LINE_OPT_ENABLE_VECTOR_NUM:
+                       printf("event vectorization is enabled\n");
+                       rsrc->evt_vec.enabled = 1;
+                       break;
+               case CMD_LINE_OPT_VECTOR_SIZE_NUM:
+                       rsrc->evt_vec.size = strtol(optarg, NULL, 10);
+                       break;
+               case CMD_LINE_OPT_VECTOR_TMO_NS_NUM:
+                       rsrc->evt_vec.timeout_ns = strtoull(optarg, NULL, 10);
+                       break;
 
                /* long options */
                case 0:
@@ -283,6 +308,18 @@ l2fwd_event_parse_args(int argc, char **argv, struct l2fwd_resources *rsrc)
 
        rsrc->mac_updating = mac_updating;
 
+       if (rsrc->evt_vec.enabled && !rsrc->evt_vec.size) {
+               rsrc->evt_vec.size = VECTOR_SIZE_DEFAULT;
+               printf("vector size set to default (%" PRIu16 ")\n",
+                      rsrc->evt_vec.size);
+       }
+
+       if (rsrc->evt_vec.enabled && !rsrc->evt_vec.timeout_ns) {
+               rsrc->evt_vec.timeout_ns = VECTOR_TMO_NS_DEFAULT;
+               printf("vector timeout set to default (%" PRIu64 " ns)\n",
+                      rsrc->evt_vec.timeout_ns);
+       }
+
        if (optind >= 0)
                argv[optind-1] = prgname;
 
@@ -636,6 +673,17 @@ main(int argc, char **argv)
                rte_panic("Cannot init mbuf pool\n");
        /* >8 End of creation of mbuf pool. */
 
+       if (rsrc->evt_vec.enabled) {
+               unsigned int nb_vec, vec_size;
+
+               vec_size = rsrc->evt_vec.size;
+               nb_vec = (nb_mbufs + vec_size - 1) / vec_size;
+               rsrc->evt_vec_pool = rte_event_vector_pool_create(
+                       "vector_pool", nb_vec, 0, vec_size, rte_socket_id());
+               if (rsrc->evt_vec_pool == NULL)
+                       rte_panic("Cannot init event vector pool\n");
+       }
+
        nb_ports_available = l2fwd_event_init_ports(rsrc);
        if (!nb_ports_available)
                rte_panic("All available ports are disabled. Please set portmask.\n");