examples/eventdev_pipeline: support Tx adapter
authorPavan Nikhilesh <pbhagavatula@caviumnetworks.com>
Mon, 24 Sep 2018 10:12:08 +0000 (15:42 +0530)
committerJerin Jacob <jerin.jacob@caviumnetworks.com>
Mon, 1 Oct 2018 14:52:47 +0000 (16:52 +0200)
Redo the worker pipelines and offload transmission to service cores
seamlessly through Tx adapter.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Tested-by: Nikhil Rao <nikhil.rao@intel.com>
examples/eventdev_pipeline/main.c
examples/eventdev_pipeline/pipeline_common.h
examples/eventdev_pipeline/pipeline_worker_generic.c
examples/eventdev_pipeline/pipeline_worker_tx.c

index 700bc69..92e08bc 100644 (file)
@@ -26,20 +26,6 @@ core_in_use(unsigned int lcore_id) {
                fdata->tx_core[lcore_id] || fdata->worker_core[lcore_id]);
 }
 
-static void
-eth_tx_buffer_retry(struct rte_mbuf **pkts, uint16_t unsent,
-                       void *userdata)
-{
-       int port_id = (uintptr_t) userdata;
-       unsigned int _sent = 0;
-
-       do {
-               /* Note: hard-coded TX queue */
-               _sent += rte_eth_tx_burst(port_id, 0, &pkts[_sent],
-                                         unsent - _sent);
-       } while (_sent != unsent);
-}
-
 /*
  * Parse the coremask given as argument (hexadecimal string) and fill
  * the global configuration (core role and core count) with the parsed
@@ -263,6 +249,7 @@ parse_app_args(int argc, char **argv)
 static inline int
 port_init(uint8_t port, struct rte_mempool *mbuf_pool)
 {
+       struct rte_eth_rxconf rx_conf;
        static const struct rte_eth_conf port_conf_default = {
                .rxmode = {
                        .mq_mode = ETH_MQ_RX_RSS,
@@ -291,6 +278,8 @@ port_init(uint8_t port, struct rte_mempool *mbuf_pool)
        if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
                port_conf.txmode.offloads |=
                        DEV_TX_OFFLOAD_MBUF_FAST_FREE;
+       rx_conf = dev_info.default_rxconf;
+       rx_conf.offloads = port_conf.rxmode.offloads;
 
        port_conf.rx_adv_conf.rss_conf.rss_hf &=
                dev_info.flow_type_rss_offloads;
@@ -311,7 +300,8 @@ port_init(uint8_t port, struct rte_mempool *mbuf_pool)
        /* Allocate and set up 1 RX queue per Ethernet port. */
        for (q = 0; q < rx_rings; q++) {
                retval = rte_eth_rx_queue_setup(port, q, rx_ring_size,
-                               rte_eth_dev_socket_id(port), NULL, mbuf_pool);
+                               rte_eth_dev_socket_id(port), &rx_conf,
+                               mbuf_pool);
                if (retval < 0)
                        return retval;
        }
@@ -350,7 +340,7 @@ port_init(uint8_t port, struct rte_mempool *mbuf_pool)
 static int
 init_ports(uint16_t num_ports)
 {
-       uint16_t portid, i;
+       uint16_t portid;
 
        if (!cdata.num_mbuf)
                cdata.num_mbuf = 16384 * num_ports;
@@ -367,36 +357,26 @@ init_ports(uint16_t num_ports)
                        rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu16 "\n",
                                        portid);
 
-       RTE_ETH_FOREACH_DEV(i) {
-               void *userdata = (void *)(uintptr_t) i;
-               fdata->tx_buf[i] =
-                       rte_malloc(NULL, RTE_ETH_TX_BUFFER_SIZE(32), 0);
-               if (fdata->tx_buf[i] == NULL)
-                       rte_panic("Out of memory\n");
-               rte_eth_tx_buffer_init(fdata->tx_buf[i], 32);
-               rte_eth_tx_buffer_set_err_callback(fdata->tx_buf[i],
-                                                  eth_tx_buffer_retry,
-                                                  userdata);
-       }
-
        return 0;
 }
 
 static void
 do_capability_setup(uint8_t eventdev_id)
 {
+       int ret;
        uint16_t i;
-       uint8_t mt_unsafe = 0;
+       uint8_t generic_pipeline = 0;
        uint8_t burst = 0;
 
        RTE_ETH_FOREACH_DEV(i) {
-               struct rte_eth_dev_info dev_info;
-               memset(&dev_info, 0, sizeof(struct rte_eth_dev_info));
-
-               rte_eth_dev_info_get(i, &dev_info);
-               /* Check if it is safe ask worker to tx. */
-               mt_unsafe |= !(dev_info.tx_offload_capa &
-                               DEV_TX_OFFLOAD_MT_LOCKFREE);
+               uint32_t caps = 0;
+
+               ret = rte_event_eth_tx_adapter_caps_get(eventdev_id, i, &caps);
+               if (ret)
+                       rte_exit(EXIT_FAILURE,
+                               "Invalid capability for Tx adptr port %d\n", i);
+               generic_pipeline |= !(caps &
+                               RTE_EVENT_ETH_TX_ADAPTER_CAP_INTERNAL_PORT);
        }
 
        struct rte_event_dev_info eventdev_info;
@@ -406,21 +386,42 @@ do_capability_setup(uint8_t eventdev_id)
        burst = eventdev_info.event_dev_cap & RTE_EVENT_DEV_CAP_BURST_MODE ? 1 :
                0;
 
-       if (mt_unsafe)
+       if (generic_pipeline)
                set_worker_generic_setup_data(&fdata->cap, burst);
        else
-               set_worker_tx_setup_data(&fdata->cap, burst);
+               set_worker_tx_enq_setup_data(&fdata->cap, burst);
 }
 
 static void
 signal_handler(int signum)
 {
+       static uint8_t once;
+       uint16_t portid;
+
        if (fdata->done)
                rte_exit(1, "Exiting on signal %d\n", signum);
-       if (signum == SIGINT || signum == SIGTERM) {
+       if ((signum == SIGINT || signum == SIGTERM) && !once) {
                printf("\n\nSignal %d received, preparing to exit...\n",
                                signum);
+               if (cdata.dump_dev)
+                       rte_event_dev_dump(0, stdout);
+               once = 1;
                fdata->done = 1;
+               rte_smp_wmb();
+
+               RTE_ETH_FOREACH_DEV(portid) {
+                       rte_event_eth_rx_adapter_stop(portid);
+                       rte_event_eth_tx_adapter_stop(portid);
+                       rte_eth_dev_stop(portid);
+               }
+
+               rte_eal_mp_wait_lcore();
+
+               RTE_ETH_FOREACH_DEV(portid) {
+                       rte_eth_dev_close(portid);
+               }
+
+               rte_event_dev_close(0);
        }
        if (signum == SIGTSTP)
                rte_event_dev_dump(0, stdout);
@@ -499,7 +500,7 @@ main(int argc, char **argv)
        if (worker_data == NULL)
                rte_panic("rte_calloc failed\n");
 
-       int dev_id = fdata->cap.evdev_setup(&cons_data, worker_data);
+       int dev_id = fdata->cap.evdev_setup(worker_data);
        if (dev_id < 0)
                rte_exit(EXIT_FAILURE, "Error setting up eventdev\n");
 
@@ -524,8 +525,8 @@ main(int argc, char **argv)
 
                if (fdata->tx_core[lcore_id])
                        printf(
-                               "[%s()] lcore %d executing NIC Tx, and using eventdev port %u\n",
-                               __func__, lcore_id, cons_data.port_id);
+                               "[%s()] lcore %d executing NIC Tx\n",
+                               __func__, lcore_id);
 
                if (fdata->sched_core[lcore_id])
                        printf("[%s()] lcore %d executing scheduler\n",
@@ -555,9 +556,6 @@ main(int argc, char **argv)
 
        rte_eal_mp_wait_lcore();
 
-       if (cdata.dump_dev)
-               rte_event_dev_dump(dev_id, stdout);
-
        if (!cdata.quiet && (port_stat(dev_id, worker_data[0].port_id) !=
                        (uint64_t)-ENOTSUP)) {
                printf("\nPort Workload distribution:\n");
index 9703396..a6cc912 100644 (file)
@@ -16,6 +16,7 @@
 #include <rte_ethdev.h>
 #include <rte_eventdev.h>
 #include <rte_event_eth_rx_adapter.h>
+#include <rte_event_eth_tx_adapter.h>
 #include <rte_service.h>
 #include <rte_service_component.h>
 
 #define BATCH_SIZE 16
 #define MAX_NUM_CORE 64
 
-struct cons_data {
-       uint8_t dev_id;
-       uint8_t port_id;
-       uint8_t release;
-} __rte_cache_aligned;
-
 struct worker_data {
        uint8_t dev_id;
        uint8_t port_id;
 } __rte_cache_aligned;
 
 typedef int (*worker_loop)(void *);
-typedef int (*consumer_loop)(void);
 typedef void (*schedule_loop)(unsigned int);
-typedef int (*eventdev_setup)(struct cons_data *, struct worker_data *);
-typedef void (*rx_adapter_setup)(uint16_t nb_ports);
+typedef int (*eventdev_setup)(struct worker_data *);
+typedef void (*adapter_setup)(uint16_t nb_ports);
 typedef void (*opt_check)(void);
 
 struct setup_data {
        worker_loop worker;
-       consumer_loop consumer;
        schedule_loop scheduler;
        eventdev_setup evdev_setup;
-       rx_adapter_setup adptr_setup;
+       adapter_setup adptr_setup;
        opt_check check_opt;
 };
 
 struct fastpath_data {
        volatile int done;
-       uint32_t tx_lock;
        uint32_t evdev_service_id;
        uint32_t rxadptr_service_id;
+       uint32_t txadptr_service_id;
        bool rx_single;
        bool tx_single;
        bool sched_single;
@@ -62,7 +55,6 @@ struct fastpath_data {
        unsigned int tx_core[MAX_NUM_CORE];
        unsigned int sched_core[MAX_NUM_CORE];
        unsigned int worker_core[MAX_NUM_CORE];
-       struct rte_eth_dev_tx_buffer *tx_buf[RTE_MAX_ETHPORTS];
        struct setup_data cap;
 } __rte_cache_aligned;
 
@@ -88,6 +80,8 @@ struct config_data {
        int16_t next_qid[MAX_NUM_STAGES+2];
        int16_t qid[MAX_NUM_STAGES];
        uint8_t rx_adapter_id;
+       uint8_t tx_adapter_id;
+       uint8_t tx_queue_id;
        uint64_t worker_lcore_mask;
        uint64_t rx_lcore_mask;
        uint64_t tx_lcore_mask;
@@ -99,8 +93,6 @@ struct port_link {
        uint8_t priority;
 };
 
-struct cons_data cons_data;
-
 struct fastpath_data *fdata;
 struct config_data cdata;
 
@@ -142,12 +134,11 @@ schedule_devices(unsigned int lcore_id)
                }
        }
 
-       if (fdata->tx_core[lcore_id] && (fdata->tx_single ||
-                        rte_atomic32_cmpset(&(fdata->tx_lock), 0, 1))) {
-               fdata->cap.consumer();
-               rte_atomic32_clear((rte_atomic32_t *)&(fdata->tx_lock));
+       if (fdata->tx_core[lcore_id]) {
+               rte_service_run_iter_on_app_lcore(fdata->txadptr_service_id,
+                               !fdata->tx_single);
        }
 }
 
 void set_worker_generic_setup_data(struct setup_data *caps, bool burst);
-void set_worker_tx_setup_data(struct setup_data *caps, bool burst);
+void set_worker_tx_enq_setup_data(struct setup_data *caps, bool burst);
index 2215e9e..1690649 100644 (file)
@@ -119,153 +119,13 @@ worker_generic_burst(void *arg)
        return 0;
 }
 
-static __rte_always_inline int
-consumer(void)
-{
-       const uint64_t freq_khz = rte_get_timer_hz() / 1000;
-       struct rte_event packet;
-
-       static uint64_t received;
-       static uint64_t last_pkts;
-       static uint64_t last_time;
-       static uint64_t start_time;
-       int i;
-       uint8_t dev_id = cons_data.dev_id;
-       uint8_t port_id = cons_data.port_id;
-
-       do {
-               uint16_t n = rte_event_dequeue_burst(dev_id, port_id,
-                               &packet, 1, 0);
-
-               if (n == 0) {
-                       RTE_ETH_FOREACH_DEV(i)
-                               rte_eth_tx_buffer_flush(i, 0, fdata->tx_buf[i]);
-                       return 0;
-               }
-               if (start_time == 0)
-                       last_time = start_time = rte_get_timer_cycles();
-
-               received++;
-               uint8_t outport = packet.mbuf->port;
-
-               exchange_mac(packet.mbuf);
-               rte_eth_tx_buffer(outport, 0, fdata->tx_buf[outport],
-                               packet.mbuf);
-
-               if (cons_data.release)
-                       rte_event_enqueue_burst(dev_id, port_id,
-                                                               &packet, n);
-
-               /* Print out mpps every 1<22 packets */
-               if (!cdata.quiet && received >= last_pkts + (1<<22)) {
-                       const uint64_t now = rte_get_timer_cycles();
-                       const uint64_t total_ms = (now - start_time) / freq_khz;
-                       const uint64_t delta_ms = (now - last_time) / freq_khz;
-                       uint64_t delta_pkts = received - last_pkts;
-
-                       printf("# %s RX=%"PRIu64", time %"PRIu64 "ms, "
-                                       "avg %.3f mpps [current %.3f mpps]\n",
-                                       __func__,
-                                       received,
-                                       total_ms,
-                                       received / (total_ms * 1000.0),
-                                       delta_pkts / (delta_ms * 1000.0));
-                       last_pkts = received;
-                       last_time = now;
-               }
-
-               cdata.num_packets--;
-               if (cdata.num_packets <= 0)
-                       fdata->done = 1;
-       /* Be stuck in this loop if single. */
-       } while (!fdata->done && fdata->tx_single);
-
-       return 0;
-}
-
-static __rte_always_inline int
-consumer_burst(void)
-{
-       const uint64_t freq_khz = rte_get_timer_hz() / 1000;
-       struct rte_event packets[BATCH_SIZE];
-
-       static uint64_t received;
-       static uint64_t last_pkts;
-       static uint64_t last_time;
-       static uint64_t start_time;
-       unsigned int i, j;
-       uint8_t dev_id = cons_data.dev_id;
-       uint8_t port_id = cons_data.port_id;
-
-       do {
-               uint16_t n = rte_event_dequeue_burst(dev_id, port_id,
-                               packets, RTE_DIM(packets), 0);
-
-               if (n == 0) {
-                       RTE_ETH_FOREACH_DEV(j)
-                               rte_eth_tx_buffer_flush(j, 0, fdata->tx_buf[j]);
-                       return 0;
-               }
-               if (start_time == 0)
-                       last_time = start_time = rte_get_timer_cycles();
-
-               received += n;
-               for (i = 0; i < n; i++) {
-                       uint8_t outport = packets[i].mbuf->port;
-
-                       exchange_mac(packets[i].mbuf);
-                       rte_eth_tx_buffer(outport, 0, fdata->tx_buf[outport],
-                                       packets[i].mbuf);
-
-                       packets[i].op = RTE_EVENT_OP_RELEASE;
-               }
-
-               if (cons_data.release) {
-                       uint16_t nb_tx;
-
-                       nb_tx = rte_event_enqueue_burst(dev_id, port_id,
-                                                               packets, n);
-                       while (nb_tx < n)
-                               nb_tx += rte_event_enqueue_burst(dev_id,
-                                               port_id, packets + nb_tx,
-                                               n - nb_tx);
-               }
-
-               /* Print out mpps every 1<22 packets */
-               if (!cdata.quiet && received >= last_pkts + (1<<22)) {
-                       const uint64_t now = rte_get_timer_cycles();
-                       const uint64_t total_ms = (now - start_time) / freq_khz;
-                       const uint64_t delta_ms = (now - last_time) / freq_khz;
-                       uint64_t delta_pkts = received - last_pkts;
-
-                       printf("# consumer RX=%"PRIu64", time %"PRIu64 "ms, "
-                                       "avg %.3f mpps [current %.3f mpps]\n",
-                                       received,
-                                       total_ms,
-                                       received / (total_ms * 1000.0),
-                                       delta_pkts / (delta_ms * 1000.0));
-                       last_pkts = received;
-                       last_time = now;
-               }
-
-               cdata.num_packets -= n;
-               if (cdata.num_packets <= 0)
-                       fdata->done = 1;
-       /* Be stuck in this loop if single. */
-       } while (!fdata->done && fdata->tx_single);
-
-       return 0;
-}
-
 static int
-setup_eventdev_generic(struct cons_data *cons_data,
-               struct worker_data *worker_data)
+setup_eventdev_generic(struct worker_data *worker_data)
 {
        const uint8_t dev_id = 0;
        /* +1 stages is for a SINGLE_LINK TX stage */
        const uint8_t nb_queues = cdata.num_stages + 1;
-       /* + 1 is one port for consumer */
-       const uint8_t nb_ports = cdata.num_workers + 1;
+       const uint8_t nb_ports = cdata.num_workers;
        struct rte_event_dev_config config = {
                        .nb_event_queues = nb_queues,
                        .nb_event_ports = nb_ports,
@@ -285,11 +145,6 @@ setup_eventdev_generic(struct cons_data *cons_data,
                        .nb_atomic_flows = 1024,
                .nb_atomic_order_sequences = 1024,
        };
-       struct rte_event_port_conf tx_p_conf = {
-                       .dequeue_depth = 128,
-                       .enqueue_depth = 128,
-                       .new_event_threshold = 4096,
-       };
        struct rte_event_queue_conf tx_q_conf = {
                        .priority = RTE_EVENT_DEV_PRIORITY_HIGHEST,
                        .event_queue_cfg = RTE_EVENT_QUEUE_CFG_SINGLE_LINK,
@@ -297,7 +152,6 @@ setup_eventdev_generic(struct cons_data *cons_data,
 
        struct port_link worker_queues[MAX_NUM_STAGES];
        uint8_t disable_implicit_release;
-       struct port_link tx_queue;
        unsigned int i;
 
        int ret, ndev = rte_event_dev_count();
@@ -314,7 +168,6 @@ setup_eventdev_generic(struct cons_data *cons_data,
                        RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE);
 
        wkr_p_conf.disable_implicit_release = disable_implicit_release;
-       tx_p_conf.disable_implicit_release = disable_implicit_release;
 
        if (dev_info.max_event_port_dequeue_depth <
                        config.nb_event_port_dequeue_depth)
@@ -372,8 +225,7 @@ setup_eventdev_generic(struct cons_data *cons_data,
                printf("%d: error creating qid %d\n", __LINE__, i);
                return -1;
        }
-       tx_queue.queue_id = i;
-       tx_queue.priority = RTE_EVENT_DEV_PRIORITY_HIGHEST;
+       cdata.tx_queue_id = i;
 
        if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
                wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
@@ -403,26 +255,6 @@ setup_eventdev_generic(struct cons_data *cons_data,
                w->port_id = i;
        }
 
-       if (tx_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
-               tx_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
-       if (tx_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
-               tx_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
-
-       /* port for consumer, linked to TX queue */
-       if (rte_event_port_setup(dev_id, i, &tx_p_conf) < 0) {
-               printf("Error setting up port %d\n", i);
-               return -1;
-       }
-       if (rte_event_port_link(dev_id, i, &tx_queue.queue_id,
-                               &tx_queue.priority, 1) != 1) {
-               printf("%d: error creating link for port %d\n",
-                               __LINE__, i);
-               return -1;
-       }
-       *cons_data = (struct cons_data){.dev_id = dev_id,
-                                       .port_id = i,
-                                       .release = disable_implicit_release };
-
        ret = rte_event_dev_service_id_get(dev_id,
                                &fdata->evdev_service_id);
        if (ret != -ESRCH && ret != 0) {
@@ -431,76 +263,107 @@ setup_eventdev_generic(struct cons_data *cons_data,
        }
        rte_service_runstate_set(fdata->evdev_service_id, 1);
        rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
-       if (rte_event_dev_start(dev_id) < 0) {
-               printf("Error starting eventdev\n");
-               return -1;
-       }
 
        return dev_id;
 }
 
 static void
-init_rx_adapter(uint16_t nb_ports)
+init_adapters(uint16_t nb_ports)
 {
        int i;
        int ret;
+       uint8_t tx_port_id = 0;
        uint8_t evdev_id = 0;
        struct rte_event_dev_info dev_info;
 
        ret = rte_event_dev_info_get(evdev_id, &dev_info);
 
-       struct rte_event_port_conf rx_p_conf = {
-               .dequeue_depth = 8,
-               .enqueue_depth = 8,
-               .new_event_threshold = 1200,
+       struct rte_event_port_conf adptr_p_conf = {
+               .dequeue_depth = cdata.worker_cq_depth,
+               .enqueue_depth = 64,
+               .new_event_threshold = 4096,
        };
 
-       if (rx_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
-               rx_p_conf.dequeue_depth = dev_info.max_event_port_dequeue_depth;
-       if (rx_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
-               rx_p_conf.enqueue_depth = dev_info.max_event_port_enqueue_depth;
+       if (adptr_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
+               adptr_p_conf.dequeue_depth =
+                       dev_info.max_event_port_dequeue_depth;
+       if (adptr_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
+               adptr_p_conf.enqueue_depth =
+                       dev_info.max_event_port_enqueue_depth;
 
        /* Create one adapter for all the ethernet ports. */
        ret = rte_event_eth_rx_adapter_create(cdata.rx_adapter_id, evdev_id,
-                       &rx_p_conf);
+                       &adptr_p_conf);
        if (ret)
                rte_exit(EXIT_FAILURE, "failed to create rx adapter[%d]",
                                cdata.rx_adapter_id);
 
+       ret = rte_event_eth_tx_adapter_create(cdata.tx_adapter_id, evdev_id,
+                       &adptr_p_conf);
+       if (ret)
+               rte_exit(EXIT_FAILURE, "failed to create tx adapter[%d]",
+                               cdata.tx_adapter_id);
+
        struct rte_event_eth_rx_adapter_queue_conf queue_conf;
        memset(&queue_conf, 0, sizeof(queue_conf));
        queue_conf.ev.sched_type = cdata.queue_type;
        queue_conf.ev.queue_id = cdata.qid[0];
 
        for (i = 0; i < nb_ports; i++) {
-               uint32_t cap;
-
-               ret = rte_event_eth_rx_adapter_caps_get(evdev_id, i, &cap);
-               if (ret)
-                       rte_exit(EXIT_FAILURE,
-                                       "failed to get event rx adapter "
-                                       "capabilities");
-
                ret = rte_event_eth_rx_adapter_queue_add(cdata.rx_adapter_id, i,
                                -1, &queue_conf);
                if (ret)
                        rte_exit(EXIT_FAILURE,
                                        "Failed to add queues to Rx adapter");
+
+               ret = rte_event_eth_tx_adapter_queue_add(cdata.tx_adapter_id, i,
+                               -1);
+               if (ret)
+                       rte_exit(EXIT_FAILURE,
+                                       "Failed to add queues to Tx adapter");
        }
 
+       ret = rte_event_eth_tx_adapter_event_port_get(cdata.tx_adapter_id,
+                       &tx_port_id);
+       if (ret)
+               rte_exit(EXIT_FAILURE,
+                               "Failed to get Tx adapter port id");
+       ret = rte_event_port_link(evdev_id, tx_port_id, &cdata.tx_queue_id,
+                       NULL, 1);
+       if (ret != 1)
+               rte_exit(EXIT_FAILURE,
+                               "Unable to link Tx adapter port to Tx queue");
+
        ret = rte_event_eth_rx_adapter_service_id_get(cdata.rx_adapter_id,
                                &fdata->rxadptr_service_id);
        if (ret != -ESRCH && ret != 0) {
                rte_exit(EXIT_FAILURE,
-                       "Error getting the service ID for sw eventdev\n");
+                       "Error getting the service ID for Rx adapter\n");
        }
        rte_service_runstate_set(fdata->rxadptr_service_id, 1);
        rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id, 0);
 
+       ret = rte_event_eth_tx_adapter_service_id_get(cdata.tx_adapter_id,
+                               &fdata->txadptr_service_id);
+       if (ret != -ESRCH && ret != 0) {
+               rte_exit(EXIT_FAILURE,
+                       "Error getting the service ID for Tx adapter\n");
+       }
+       rte_service_runstate_set(fdata->txadptr_service_id, 1);
+       rte_service_set_runstate_mapped_check(fdata->txadptr_service_id, 0);
+
        ret = rte_event_eth_rx_adapter_start(cdata.rx_adapter_id);
        if (ret)
                rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
                                cdata.rx_adapter_id);
+
+       ret = rte_event_eth_tx_adapter_start(cdata.tx_adapter_id);
+       if (ret)
+               rte_exit(EXIT_FAILURE, "Tx adapter[%d] start failed",
+                               cdata.tx_adapter_id);
+
+       if (rte_event_dev_start(evdev_id) < 0)
+               rte_exit(EXIT_FAILURE, "Error starting eventdev");
 }
 
 static void
@@ -510,6 +373,7 @@ generic_opt_check(void)
        int ret;
        uint32_t cap = 0;
        uint8_t rx_needed = 0;
+       uint8_t sched_needed = 0;
        struct rte_event_dev_info eventdev_info;
 
        memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
@@ -519,6 +383,8 @@ generic_opt_check(void)
                                RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES))
                rte_exit(EXIT_FAILURE,
                                "Event dev doesn't support all type queues\n");
+       sched_needed = !(eventdev_info.event_dev_cap &
+               RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED);
 
        RTE_ETH_FOREACH_DEV(i) {
                ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
@@ -531,9 +397,8 @@ generic_opt_check(void)
 
        if (cdata.worker_lcore_mask == 0 ||
                        (rx_needed && cdata.rx_lcore_mask == 0) ||
-                       cdata.tx_lcore_mask == 0 || (cdata.sched_lcore_mask == 0
-                               && !(eventdev_info.event_dev_cap &
-                                       RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))) {
+                       (cdata.tx_lcore_mask == 0) ||
+                       (sched_needed && cdata.sched_lcore_mask == 0)) {
                printf("Core part of pipeline was not assigned any cores. "
                        "This will stall the pipeline, please check core masks "
                        "(use -h for details on setting core masks):\n"
@@ -545,23 +410,24 @@ generic_opt_check(void)
                rte_exit(-1, "Fix core masks\n");
        }
 
-       if (eventdev_info.event_dev_cap & RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED)
+       if (!sched_needed)
                memset(fdata->sched_core, 0,
                                sizeof(unsigned int) * MAX_NUM_CORE);
+       if (!rx_needed)
+               memset(fdata->rx_core, 0,
+                               sizeof(unsigned int) * MAX_NUM_CORE);
 }
 
 void
 set_worker_generic_setup_data(struct setup_data *caps, bool burst)
 {
        if (burst) {
-               caps->consumer = consumer_burst;
                caps->worker = worker_generic_burst;
        } else {
-               caps->consumer = consumer;
                caps->worker = worker_generic;
        }
 
-       caps->adptr_setup = init_rx_adapter;
+       caps->adptr_setup = init_adapters;
        caps->scheduler = schedule_devices;
        caps->evdev_setup = setup_eventdev_generic;
        caps->check_opt = generic_opt_check;
index 3dbde92..85eb075 100644 (file)
@@ -36,10 +36,11 @@ worker_event_enqueue_burst(const uint8_t dev, const uint8_t port,
 }
 
 static __rte_always_inline void
-worker_tx_pkt(struct rte_mbuf *mbuf)
+worker_tx_pkt(const uint8_t dev, const uint8_t port, struct rte_event *ev)
 {
-       exchange_mac(mbuf);
-       while (rte_eth_tx_burst(mbuf->port, 0, &mbuf, 1) != 1)
+       exchange_mac(ev->mbuf);
+       rte_event_eth_tx_adapter_txq_set(ev->mbuf, 0);
+       while (!rte_event_eth_tx_adapter_enqueue(dev, port, ev, 1))
                rte_pause();
 }
 
@@ -64,15 +65,15 @@ worker_do_tx_single(void *arg)
                received++;
 
                if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
-                       worker_tx_pkt(ev.mbuf);
+                       worker_tx_pkt(dev, port, &ev);
                        tx++;
-                       continue;
+               } else {
+                       work();
+                       ev.queue_id++;
+                       worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
+                       worker_event_enqueue(dev, port, &ev);
+                       fwd++;
                }
-               work();
-               ev.queue_id++;
-               worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
-               worker_event_enqueue(dev, port, &ev);
-               fwd++;
        }
 
        if (!cdata.quiet)
@@ -100,14 +101,14 @@ worker_do_tx_single_atq(void *arg)
                received++;
 
                if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
-                       worker_tx_pkt(ev.mbuf);
+                       worker_tx_pkt(dev, port, &ev);
                        tx++;
-                       continue;
+               } else {
+                       work();
+                       worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
+                       worker_event_enqueue(dev, port, &ev);
+                       fwd++;
                }
-               work();
-               worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
-               worker_event_enqueue(dev, port, &ev);
-               fwd++;
        }
 
        if (!cdata.quiet)
@@ -141,7 +142,7 @@ worker_do_tx_single_burst(void *arg)
                        rte_prefetch0(ev[i + 1].mbuf);
                        if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 
-                               worker_tx_pkt(ev[i].mbuf);
+                               worker_tx_pkt(dev, port, &ev[i]);
                                ev[i].op = RTE_EVENT_OP_RELEASE;
                                tx++;
 
@@ -188,7 +189,7 @@ worker_do_tx_single_burst_atq(void *arg)
                        rte_prefetch0(ev[i + 1].mbuf);
                        if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
 
-                               worker_tx_pkt(ev[i].mbuf);
+                               worker_tx_pkt(dev, port, &ev[i]);
                                ev[i].op = RTE_EVENT_OP_RELEASE;
                                tx++;
                        } else
@@ -232,7 +233,7 @@ worker_do_tx(void *arg)
 
                if (cq_id >= lst_qid) {
                        if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
-                               worker_tx_pkt(ev.mbuf);
+                               worker_tx_pkt(dev, port, &ev);
                                tx++;
                                continue;
                        }
@@ -280,7 +281,7 @@ worker_do_tx_atq(void *arg)
 
                if (cq_id == lst_qid) {
                        if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
-                               worker_tx_pkt(ev.mbuf);
+                               worker_tx_pkt(dev, port, &ev);
                                tx++;
                                continue;
                        }
@@ -330,7 +331,7 @@ worker_do_tx_burst(void *arg)
 
                        if (cq_id >= lst_qid) {
                                if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
-                                       worker_tx_pkt(ev[i].mbuf);
+                                       worker_tx_pkt(dev, port, &ev[i]);
                                        tx++;
                                        ev[i].op = RTE_EVENT_OP_RELEASE;
                                        continue;
@@ -387,7 +388,7 @@ worker_do_tx_burst_atq(void *arg)
 
                        if (cq_id == lst_qid) {
                                if (ev[i].sched_type == RTE_SCHED_TYPE_ATOMIC) {
-                                       worker_tx_pkt(ev[i].mbuf);
+                                       worker_tx_pkt(dev, port, &ev[i]);
                                        tx++;
                                        ev[i].op = RTE_EVENT_OP_RELEASE;
                                        continue;
@@ -413,10 +414,8 @@ worker_do_tx_burst_atq(void *arg)
 }
 
 static int
-setup_eventdev_worker_tx(struct cons_data *cons_data,
-               struct worker_data *worker_data)
+setup_eventdev_worker_tx_enq(struct worker_data *worker_data)
 {
-       RTE_SET_USED(cons_data);
        uint8_t i;
        const uint8_t atq = cdata.all_type_queues ? 1 : 0;
        const uint8_t dev_id = 0;
@@ -575,10 +574,9 @@ setup_eventdev_worker_tx(struct cons_data *cons_data,
        }
        rte_service_runstate_set(fdata->evdev_service_id, 1);
        rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
-       if (rte_event_dev_start(dev_id) < 0) {
-               printf("Error starting eventdev\n");
-               return -1;
-       }
+
+       if (rte_event_dev_start(dev_id) < 0)
+               rte_exit(EXIT_FAILURE, "Error starting eventdev");
 
        return dev_id;
 }
@@ -602,7 +600,7 @@ service_rx_adapter(void *arg)
 }
 
 static void
-init_rx_adapter(uint16_t nb_ports)
+init_adapters(uint16_t nb_ports)
 {
        int i;
        int ret;
@@ -613,17 +611,18 @@ init_rx_adapter(uint16_t nb_ports)
        ret = rte_event_dev_info_get(evdev_id, &dev_info);
        adptr_services = rte_zmalloc(NULL, sizeof(struct rx_adptr_services), 0);
 
-       struct rte_event_port_conf rx_p_conf = {
-               .dequeue_depth = 8,
-               .enqueue_depth = 8,
-               .new_event_threshold = 1200,
+       struct rte_event_port_conf adptr_p_conf = {
+               .dequeue_depth = cdata.worker_cq_depth,
+               .enqueue_depth = 64,
+               .new_event_threshold = 4096,
        };
 
-       if (rx_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
-               rx_p_conf.dequeue_depth = dev_info.max_event_port_dequeue_depth;
-       if (rx_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
-               rx_p_conf.enqueue_depth = dev_info.max_event_port_enqueue_depth;
-
+       if (adptr_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
+               adptr_p_conf.dequeue_depth =
+                       dev_info.max_event_port_dequeue_depth;
+       if (adptr_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
+               adptr_p_conf.enqueue_depth =
+                       dev_info.max_event_port_enqueue_depth;
 
        struct rte_event_eth_rx_adapter_queue_conf queue_conf;
        memset(&queue_conf, 0, sizeof(queue_conf));
@@ -633,11 +632,11 @@ init_rx_adapter(uint16_t nb_ports)
                uint32_t cap;
                uint32_t service_id;
 
-               ret = rte_event_eth_rx_adapter_create(i, evdev_id, &rx_p_conf);
+               ret = rte_event_eth_rx_adapter_create(i, evdev_id,
+                               &adptr_p_conf);
                if (ret)
                        rte_exit(EXIT_FAILURE,
-                                       "failed to create rx adapter[%d]",
-                                       cdata.rx_adapter_id);
+                                       "failed to create rx adapter[%d]", i);
 
                ret = rte_event_eth_rx_adapter_caps_get(evdev_id, i, &cap);
                if (ret)
@@ -654,7 +653,6 @@ init_rx_adapter(uint16_t nb_ports)
                        rte_exit(EXIT_FAILURE,
                                        "Failed to add queues to Rx adapter");
 
-
                /* Producer needs to be scheduled. */
                if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
                        ret = rte_event_eth_rx_adapter_service_id_get(i,
@@ -680,9 +678,29 @@ init_rx_adapter(uint16_t nb_ports)
                ret = rte_event_eth_rx_adapter_start(i);
                if (ret)
                        rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
-                                       cdata.rx_adapter_id);
+                                       i);
        }
 
+       /* We already know that Tx adapter has INTERNAL port cap*/
+       ret = rte_event_eth_tx_adapter_create(cdata.tx_adapter_id, evdev_id,
+                       &adptr_p_conf);
+       if (ret)
+               rte_exit(EXIT_FAILURE, "failed to create tx adapter[%d]",
+                               cdata.tx_adapter_id);
+
+       for (i = 0; i < nb_ports; i++) {
+               ret = rte_event_eth_tx_adapter_queue_add(cdata.tx_adapter_id, i,
+                               -1);
+               if (ret)
+                       rte_exit(EXIT_FAILURE,
+                                       "Failed to add queues to Tx adapter");
+       }
+
+       ret = rte_event_eth_tx_adapter_start(cdata.tx_adapter_id);
+       if (ret)
+               rte_exit(EXIT_FAILURE, "Tx adapter[%d] start failed",
+                               cdata.tx_adapter_id);
+
        if (adptr_services->nb_rx_adptrs) {
                struct rte_service_spec service;
 
@@ -695,8 +713,7 @@ init_rx_adapter(uint16_t nb_ports)
                                &fdata->rxadptr_service_id);
                if (ret)
                        rte_exit(EXIT_FAILURE,
-                               "Rx adapter[%d] service register failed",
-                               cdata.rx_adapter_id);
+                               "Rx adapter service register failed");
 
                rte_service_runstate_set(fdata->rxadptr_service_id, 1);
                rte_service_component_runstate_set(fdata->rxadptr_service_id,
@@ -708,23 +725,19 @@ init_rx_adapter(uint16_t nb_ports)
                rte_free(adptr_services);
        }
 
-       if (!adptr_services->nb_rx_adptrs && fdata->cap.consumer == NULL &&
-                       (dev_info.event_dev_cap &
+       if (!adptr_services->nb_rx_adptrs && (dev_info.event_dev_cap &
                         RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))
                fdata->cap.scheduler = NULL;
-
-       if (dev_info.event_dev_cap & RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED)
-               memset(fdata->sched_core, 0,
-                               sizeof(unsigned int) * MAX_NUM_CORE);
 }
 
 static void
-worker_tx_opt_check(void)
+worker_tx_enq_opt_check(void)
 {
        int i;
        int ret;
        uint32_t cap = 0;
        uint8_t rx_needed = 0;
+       uint8_t sched_needed = 0;
        struct rte_event_dev_info eventdev_info;
 
        memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
@@ -734,32 +747,38 @@ worker_tx_opt_check(void)
                                RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES))
                rte_exit(EXIT_FAILURE,
                                "Event dev doesn't support all type queues\n");
+       sched_needed = !(eventdev_info.event_dev_cap &
+               RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED);
 
        RTE_ETH_FOREACH_DEV(i) {
                ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
                if (ret)
                        rte_exit(EXIT_FAILURE,
-                                       "failed to get event rx adapter "
-                                       "capabilities");
+                               "failed to get event rx adapter capabilities");
                rx_needed |=
                        !(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
        }
 
        if (cdata.worker_lcore_mask == 0 ||
                        (rx_needed && cdata.rx_lcore_mask == 0) ||
-                       (cdata.sched_lcore_mask == 0 &&
-                        !(eventdev_info.event_dev_cap &
-                                RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED))) {
+                       (sched_needed && cdata.sched_lcore_mask == 0)) {
                printf("Core part of pipeline was not assigned any cores. "
                        "This will stall the pipeline, please check core masks "
                        "(use -h for details on setting core masks):\n"
-                       "\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
-                       "\n\tworkers: %"PRIu64"\n",
-                       cdata.rx_lcore_mask, cdata.tx_lcore_mask,
-                       cdata.sched_lcore_mask,
-                       cdata.worker_lcore_mask);
+                       "\trx: %"PRIu64"\n\tsched: %"PRIu64
+                       "\n\tworkers: %"PRIu64"\n", cdata.rx_lcore_mask,
+                       cdata.sched_lcore_mask, cdata.worker_lcore_mask);
                rte_exit(-1, "Fix core masks\n");
        }
+
+       if (!sched_needed)
+               memset(fdata->sched_core, 0,
+                               sizeof(unsigned int) * MAX_NUM_CORE);
+       if (!rx_needed)
+               memset(fdata->rx_core, 0,
+                               sizeof(unsigned int) * MAX_NUM_CORE);
+
+       memset(fdata->tx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
 }
 
 static worker_loop
@@ -821,18 +840,15 @@ get_worker_multi_stage(bool burst)
 }
 
 void
-set_worker_tx_setup_data(struct setup_data *caps, bool burst)
+set_worker_tx_enq_setup_data(struct setup_data *caps, bool burst)
 {
        if (cdata.num_stages == 1)
                caps->worker = get_worker_single_stage(burst);
        else
                caps->worker = get_worker_multi_stage(burst);
 
-       memset(fdata->tx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
-
-       caps->check_opt = worker_tx_opt_check;
-       caps->consumer = NULL;
+       caps->check_opt = worker_tx_enq_opt_check;
        caps->scheduler = schedule_devices;
-       caps->evdev_setup = setup_eventdev_worker_tx;
-       caps->adptr_setup = init_rx_adapter;
+       caps->evdev_setup = setup_eventdev_worker_tx_enq;
+       caps->adptr_setup = init_adapters;
 }