examples/distributor: add dedicated core
authorDavid Hunt <david.hunt@intel.com>
Mon, 20 Mar 2017 10:08:37 +0000 (10:08 +0000)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Wed, 29 Mar 2017 14:46:57 +0000 (16:46 +0200)
Give the distribution functionality it's own core for performance,
otherwise it's limited by the Rx core.

Signed-off-by: David Hunt <david.hunt@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
examples/distributor/main.c

index 75c001d..96d6454 100644 (file)
@@ -49,6 +49,8 @@
 #define NUM_MBUFS ((64*1024)-1)
 #define MBUF_CACHE_SIZE 250
 #define BURST_SIZE 32
+#define SCHED_RX_RING_SZ 8192
+#define SCHED_TX_RING_SZ 65536
 #define RTE_RING_SZ 1024
 
 #define RTE_LOGTYPE_DISTRAPP RTE_LOGTYPE_USER1
@@ -193,37 +195,14 @@ port_init(uint8_t port, struct rte_mempool *mbuf_pool)
 struct lcore_params {
        unsigned worker_id;
        struct rte_distributor *d;
-       struct rte_ring *r;
+       struct rte_ring *rx_dist_ring;
+       struct rte_ring *dist_tx_ring;
        struct rte_mempool *mem_pool;
 };
 
-static int
-quit_workers(struct rte_distributor *d, struct rte_mempool *p)
-{
-       const unsigned num_workers = rte_lcore_count() - 2;
-       unsigned i;
-       struct rte_mbuf *bufs[num_workers];
-
-       if (rte_mempool_get_bulk(p, (void *)bufs, num_workers) != 0) {
-               printf("line %d: Error getting mbufs from pool\n", __LINE__);
-               return -1;
-       }
-
-       for (i = 0; i < num_workers; i++)
-               bufs[i]->hash.rss = i << 1;
-
-       rte_distributor_process(d, bufs, num_workers);
-       rte_mempool_put_bulk(p, (void *)bufs, num_workers);
-
-       return 0;
-}
-
 static int
 lcore_rx(struct lcore_params *p)
 {
-       struct rte_distributor *d = p->d;
-       struct rte_mempool *mem_pool = p->mem_pool;
-       struct rte_ring *r = p->r;
        const uint8_t nb_ports = rte_eth_dev_count();
        const int socket_id = rte_socket_id();
        uint8_t port;
@@ -260,9 +239,15 @@ lcore_rx(struct lcore_params *p)
                }
                app_stats.rx.rx_pkts += nb_rx;
 
-               rte_distributor_process(d, bufs, nb_rx);
-               const uint16_t nb_ret = rte_distributor_returned_pkts(d,
-                               bufs, BURST_SIZE*2);
+/*
+ * You can run the distributor on the rx core with this code. Returned
+ * packets are then send straight to the tx core.
+ */
+#if 0
+       rte_distributor_process(d, bufs, nb_rx);
+       const uint16_t nb_ret = rte_distributor_returned_pktsd,
+                       bufs, BURST_SIZE*2);
+
                app_stats.rx.returned_pkts += nb_ret;
                if (unlikely(nb_ret == 0)) {
                        if (++port == nb_ports)
@@ -270,7 +255,22 @@ lcore_rx(struct lcore_params *p)
                        continue;
                }
 
-               uint16_t sent = rte_ring_enqueue_burst(r, (void *)bufs, nb_ret);
+               struct rte_ring *tx_ring = p->dist_tx_ring;
+               uint16_t sent = rte_ring_enqueue_burst(tx_ring,
+                               (void *)bufs, nb_ret);
+#else
+               uint16_t nb_ret = nb_rx;
+               /*
+                * Swap the following two lines if you want the rx traffic
+                * to go directly to tx, no distribution.
+                */
+               struct rte_ring *out_ring = p->rx_dist_ring;
+               /* struct rte_ring *out_ring = p->dist_tx_ring; */
+
+               uint16_t sent = rte_ring_enqueue_burst(out_ring,
+                               (void *)bufs, nb_ret);
+#endif
+
                app_stats.rx.enqueued_pkts += sent;
                if (unlikely(sent < nb_ret)) {
                        RTE_LOG_DP(DEBUG, DISTRAPP,
@@ -281,20 +281,9 @@ lcore_rx(struct lcore_params *p)
                if (++port == nb_ports)
                        port = 0;
        }
-       rte_distributor_process(d, NULL, 0);
-       /* flush distributor to bring to known state */
-       rte_distributor_flush(d);
        /* set worker & tx threads quit flag */
+       printf("\nCore %u exiting rx task.\n", rte_lcore_id());
        quit_signal = 1;
-       /*
-        * worker threads may hang in get packet as
-        * distributor process is not running, just make sure workers
-        * get packets till quit_signal is actually been
-        * received and they gracefully shutdown
-        */
-       if (quit_workers(d, mem_pool) != 0)
-               return -1;
-       /* rx thread should quit at last */
        return 0;
 }
 
@@ -331,6 +320,58 @@ flush_all_ports(struct output_buffer *tx_buffers, uint8_t nb_ports)
        }
 }
 
+
+
+static int
+lcore_distributor(struct lcore_params *p)
+{
+       struct rte_ring *in_r = p->rx_dist_ring;
+       struct rte_ring *out_r = p->dist_tx_ring;
+       struct rte_mbuf *bufs[BURST_SIZE * 4];
+       struct rte_distributor *d = p->d;
+
+       printf("\nCore %u acting as distributor core.\n", rte_lcore_id());
+       while (!quit_signal_dist) {
+               const uint16_t nb_rx = rte_ring_dequeue_burst(in_r,
+                               (void *)bufs, BURST_SIZE*1);
+               if (nb_rx) {
+                       app_stats.dist.in_pkts += nb_rx;
+
+                       /* Distribute the packets */
+                       rte_distributor_process(d, bufs, nb_rx);
+                       /* Handle Returns */
+                       const uint16_t nb_ret =
+                               rte_distributor_returned_pkts(d,
+                                       bufs, BURST_SIZE*2);
+
+                       if (unlikely(nb_ret == 0))
+                               continue;
+                       app_stats.dist.ret_pkts += nb_ret;
+
+                       uint16_t sent = rte_ring_enqueue_burst(out_r,
+                                       (void *)bufs, nb_ret);
+                       app_stats.dist.sent_pkts += sent;
+                       if (unlikely(sent < nb_ret)) {
+                               app_stats.dist.enqdrop_pkts += nb_ret - sent;
+                               RTE_LOG(DEBUG, DISTRAPP,
+                                       "%s:Packet loss due to full out ring\n",
+                                       __func__);
+                               while (sent < nb_ret)
+                                       rte_pktmbuf_free(bufs[sent++]);
+                       }
+               }
+       }
+       printf("\nCore %u exiting distributor task.\n", rte_lcore_id());
+       quit_signal_work = 1;
+
+       rte_distributor_flush(d);
+       /* Unblock any returns so workers can exit */
+       rte_distributor_clear_returns(d);
+       quit_signal_rx = 1;
+       return 0;
+}
+
+
 static int
 lcore_tx(struct rte_ring *in_r)
 {
@@ -403,7 +444,7 @@ int_handler(int sig_num)
 {
        printf("Exiting on signal %d\n", sig_num);
        /* set quit flag for rx thread to exit */
-       quit_signal_rx = 1;
+       quit_signal_dist = 1;
 }
 
 static void
@@ -517,7 +558,7 @@ lcore_worker(struct lcore_params *p)
                buf[i] = NULL;
 
        printf("\nCore %u acting as worker core.\n", rte_lcore_id());
-       while (!quit_signal) {
+       while (!quit_signal_work) {
                num = rte_distributor_get_pkt(d, id, buf, buf, num);
                /* Do a little bit of work for each packet */
                for (i = 0; i < num; i++) {
@@ -608,7 +649,8 @@ main(int argc, char *argv[])
 {
        struct rte_mempool *mbuf_pool;
        struct rte_distributor *d;
-       struct rte_ring *output_ring;
+       struct rte_ring *dist_tx_ring;
+       struct rte_ring *rx_dist_ring;
        unsigned lcore_id, worker_id = 0;
        unsigned nb_ports;
        uint8_t portid;
@@ -630,10 +672,11 @@ main(int argc, char *argv[])
        if (ret < 0)
                rte_exit(EXIT_FAILURE, "Invalid distributor parameters\n");
 
-       if (rte_lcore_count() < 3)
+       if (rte_lcore_count() < 4)
                rte_exit(EXIT_FAILURE, "Error, This application needs at "
-                               "least 3 logical cores to run:\n"
-                               "1 lcore for packet RX and distribution\n"
+                               "least 4 logical cores to run:\n"
+                               "1 lcore for packet RX\n"
+                               "1 lcore for distribution\n"
                                "1 lcore for packet TX\n"
                                "and at least 1 lcore for worker threads\n");
 
@@ -673,30 +716,52 @@ main(int argc, char *argv[])
        }
 
        d = rte_distributor_create("PKT_DIST", rte_socket_id(),
-                       rte_lcore_count() - 2,
+                       rte_lcore_count() - 3,
                        RTE_DIST_ALG_BURST);
        if (d == NULL)
                rte_exit(EXIT_FAILURE, "Cannot create distributor\n");
 
        /*
-        * scheduler ring is read only by the transmitter core, but written to
-        * by multiple threads
+        * scheduler ring is read by the transmitter core, and written to
+        * by scheduler core
         */
-       output_ring = rte_ring_create("Output_ring", RTE_RING_SZ,
-                       rte_socket_id(), RING_F_SC_DEQ);
-       if (output_ring == NULL)
+       dist_tx_ring = rte_ring_create("Output_ring", SCHED_TX_RING_SZ,
+                       rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
+       if (dist_tx_ring == NULL)
+               rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
+
+       rx_dist_ring = rte_ring_create("Input_ring", SCHED_RX_RING_SZ,
+                       rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
+       if (rx_dist_ring == NULL)
                rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
 
        RTE_LCORE_FOREACH_SLAVE(lcore_id) {
-               if (worker_id == rte_lcore_count() - 2)
+               if (worker_id == rte_lcore_count() - 3) {
+                       printf("Starting distributor on lcore_id %d\n",
+                                       lcore_id);
+                       /* distributor core */
+                       struct lcore_params *p =
+                                       rte_malloc(NULL, sizeof(*p), 0);
+                       if (!p)
+                               rte_panic("malloc failure\n");
+                       *p = (struct lcore_params){worker_id, d,
+                               rx_dist_ring, dist_tx_ring, mbuf_pool};
+                       rte_eal_remote_launch(
+                               (lcore_function_t *)lcore_distributor,
+                               p, lcore_id);
+               } else if (worker_id == rte_lcore_count() - 4) {
+                       printf("Starting tx  on worker_id %d, lcore_id %d\n",
+                                       worker_id, lcore_id);
+                       /* tx core */
                        rte_eal_remote_launch((lcore_function_t *)lcore_tx,
-                                       output_ring, lcore_id);
-               else {
+                                       dist_tx_ring, lcore_id);
+               else {
                        struct lcore_params *p =
                                        rte_malloc(NULL, sizeof(*p), 0);
                        if (!p)
                                rte_panic("malloc failure\n");
-                       *p = (struct lcore_params){worker_id, d, output_ring, mbuf_pool};
+                       *p = (struct lcore_params){worker_id, d, rx_dist_ring,
+                                       dist_tx_ring, mbuf_pool};
 
                        rte_eal_remote_launch((lcore_function_t *)lcore_worker,
                                        p, lcore_id);
@@ -704,7 +769,7 @@ main(int argc, char *argv[])
                worker_id++;
        }
        /* call lcore_main on master core only */
-       struct lcore_params p = { 0, d, output_ring, mbuf_pool};
+       struct lcore_params p = { 0, d, rx_dist_ring, dist_tx_ring, mbuf_pool};
 
        if (lcore_rx(&p) != 0)
                return -1;