examples/eventdev: add thread safe Tx worker pipeline
authorPavan Nikhilesh <pbhagavatula@caviumnetworks.com>
Wed, 10 Jan 2018 11:10:06 +0000 (16:40 +0530)
committerJerin Jacob <jerin.jacob@caviumnetworks.com>
Fri, 19 Jan 2018 15:09:56 +0000 (16:09 +0100)
Add worker pipeline when Tx is multi thread safe.
Probe Ethernet dev capabilities and select it it is supported.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
examples/eventdev_pipeline_sw_pmd/Makefile
examples/eventdev_pipeline_sw_pmd/main.c
examples/eventdev_pipeline_sw_pmd/pipeline_common.h
examples/eventdev_pipeline_sw_pmd/pipeline_worker_tx.c [new file with mode: 0644]

index f2578da..c099adf 100644 (file)
@@ -16,6 +16,7 @@ APP = eventdev_pipeline_sw_pmd
 # all source are stored in SRCS-y
 SRCS-y := main.c
 SRCS-y += pipeline_worker_generic.c
+SRCS-y += pipeline_worker_tx.c
 
 CFLAGS += -O3
 CFLAGS += $(WERROR_FLAGS)
index 357584d..e07f68e 100644 (file)
@@ -364,9 +364,20 @@ init_ports(unsigned int num_ports)
 static void
 do_capability_setup(uint16_t nb_ethdev, uint8_t eventdev_id)
 {
-       RTE_SET_USED(nb_ethdev);
+       int i;
+       uint8_t mt_unsafe = 0;
        uint8_t burst = 0;
 
+       for (i = 0; i < nb_ethdev; 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);
+       }
+
        struct rte_event_dev_info eventdev_info;
        memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
 
@@ -374,7 +385,10 @@ do_capability_setup(uint16_t nb_ethdev, uint8_t eventdev_id)
        burst = eventdev_info.event_dev_cap & RTE_EVENT_DEV_CAP_BURST_MODE ? 1 :
                0;
 
-       set_worker_generic_setup_data(&fdata->cap, burst);
+       if (mt_unsafe)
+               set_worker_generic_setup_data(&fdata->cap, burst);
+       else
+               set_worker_tx_setup_data(&fdata->cap, burst);
 }
 
 static void
index d58059b..e063200 100644 (file)
@@ -79,6 +79,10 @@ struct config_data {
        int dump_dev_signal;
        unsigned int num_stages;
        unsigned int worker_cq_depth;
+       unsigned int rx_stride;
+       /* Use rx stride value to reduce congestion in entry queue when using
+        * multiple eth ports by forming multiple event queue pipelines.
+        */
        int16_t next_qid[MAX_NUM_STAGES+2];
        int16_t qid[MAX_NUM_STAGES];
        uint8_t rx_adapter_id;
@@ -144,3 +148,4 @@ schedule_devices(unsigned int lcore_id)
 }
 
 void set_worker_generic_setup_data(struct setup_data *caps, bool burst);
+void set_worker_tx_setup_data(struct setup_data *caps, bool burst);
diff --git a/examples/eventdev_pipeline_sw_pmd/pipeline_worker_tx.c b/examples/eventdev_pipeline_sw_pmd/pipeline_worker_tx.c
new file mode 100644 (file)
index 0000000..397b101
--- /dev/null
@@ -0,0 +1,425 @@
+/*
+ * SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2010-2014 Intel Corporation
+ * Copyright 2017 Cavium, Inc.
+ */
+
+#include "pipeline_common.h"
+
+static __rte_always_inline void
+worker_fwd_event(struct rte_event *ev, uint8_t sched)
+{
+       ev->event_type = RTE_EVENT_TYPE_CPU;
+       ev->op = RTE_EVENT_OP_FORWARD;
+       ev->sched_type = sched;
+}
+
+static __rte_always_inline void
+worker_event_enqueue(const uint8_t dev, const uint8_t port,
+               struct rte_event *ev)
+{
+       while (rte_event_enqueue_burst(dev, port, ev, 1) != 1)
+               rte_pause();
+}
+
+static __rte_always_inline void
+worker_tx_pkt(struct rte_mbuf *mbuf)
+{
+       exchange_mac(mbuf);
+       while (rte_eth_tx_burst(mbuf->port, 0, &mbuf, 1) != 1)
+               rte_pause();
+}
+
+/* Multi stage Pipeline Workers */
+
+static int
+worker_do_tx(void *arg)
+{
+       struct rte_event ev;
+
+       struct worker_data *data = (struct worker_data *)arg;
+       const uint8_t dev = data->dev_id;
+       const uint8_t port = data->port_id;
+       const uint8_t lst_qid = cdata.num_stages - 1;
+       size_t fwd = 0, received = 0, tx = 0;
+
+
+       while (!fdata->done) {
+
+               if (!rte_event_dequeue_burst(dev, port, &ev, 1, 0)) {
+                       rte_pause();
+                       continue;
+               }
+
+               received++;
+               const uint8_t cq_id = ev.queue_id % cdata.num_stages;
+
+               if (cq_id >= lst_qid) {
+                       if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
+                               worker_tx_pkt(ev.mbuf);
+                               tx++;
+                               continue;
+                       }
+
+                       worker_fwd_event(&ev, RTE_SCHED_TYPE_ATOMIC);
+                       ev.queue_id = (cq_id == lst_qid) ?
+                               cdata.next_qid[ev.queue_id] : ev.queue_id;
+               } else {
+                       ev.queue_id = cdata.next_qid[ev.queue_id];
+                       worker_fwd_event(&ev, cdata.queue_type);
+               }
+               work();
+
+               worker_event_enqueue(dev, port, &ev);
+               fwd++;
+       }
+
+       if (!cdata.quiet)
+               printf("  worker %u thread done. RX=%zu FWD=%zu TX=%zu\n",
+                               rte_lcore_id(), received, fwd, tx);
+
+       return 0;
+}
+
+static int
+setup_eventdev_worker_tx(struct cons_data *cons_data,
+               struct worker_data *worker_data)
+{
+       RTE_SET_USED(cons_data);
+       uint8_t i;
+       const uint8_t dev_id = 0;
+       const uint8_t nb_ports = cdata.num_workers;
+       uint8_t nb_slots = 0;
+       uint8_t nb_queues = rte_eth_dev_count() * cdata.num_stages;
+       nb_queues += rte_eth_dev_count();
+
+       struct rte_event_dev_config config = {
+                       .nb_event_queues = nb_queues,
+                       .nb_event_ports = nb_ports,
+                       .nb_events_limit  = 4096,
+                       .nb_event_queue_flows = 1024,
+                       .nb_event_port_dequeue_depth = 128,
+                       .nb_event_port_enqueue_depth = 128,
+       };
+       struct rte_event_port_conf wkr_p_conf = {
+                       .dequeue_depth = cdata.worker_cq_depth,
+                       .enqueue_depth = 64,
+                       .new_event_threshold = 4096,
+       };
+       struct rte_event_queue_conf wkr_q_conf = {
+                       .schedule_type = cdata.queue_type,
+                       .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
+                       .nb_atomic_flows = 1024,
+                       .nb_atomic_order_sequences = 1024,
+       };
+
+       int ret, ndev = rte_event_dev_count();
+
+       if (ndev < 1) {
+               printf("%d: No Eventdev Devices Found\n", __LINE__);
+               return -1;
+       }
+
+
+       struct rte_event_dev_info dev_info;
+       ret = rte_event_dev_info_get(dev_id, &dev_info);
+       printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
+
+       if (dev_info.max_event_port_dequeue_depth <
+                       config.nb_event_port_dequeue_depth)
+               config.nb_event_port_dequeue_depth =
+                               dev_info.max_event_port_dequeue_depth;
+       if (dev_info.max_event_port_enqueue_depth <
+                       config.nb_event_port_enqueue_depth)
+               config.nb_event_port_enqueue_depth =
+                               dev_info.max_event_port_enqueue_depth;
+
+       ret = rte_event_dev_configure(dev_id, &config);
+       if (ret < 0) {
+               printf("%d: Error configuring device\n", __LINE__);
+               return -1;
+       }
+
+       printf("  Stages:\n");
+       for (i = 0; i < nb_queues; i++) {
+
+               uint8_t slot;
+
+               nb_slots = cdata.num_stages + 1;
+               slot = i % nb_slots;
+               wkr_q_conf.schedule_type = slot == cdata.num_stages ?
+                       RTE_SCHED_TYPE_ATOMIC : cdata.queue_type;
+
+               if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
+                       printf("%d: error creating qid %d\n", __LINE__, i);
+                       return -1;
+               }
+               cdata.qid[i] = i;
+               cdata.next_qid[i] = i+1;
+               if (cdata.enable_queue_priorities) {
+                       const uint32_t prio_delta =
+                               (RTE_EVENT_DEV_PRIORITY_LOWEST) /
+                               nb_slots;
+
+                       /* higher priority for queues closer to tx */
+                       wkr_q_conf.priority =
+                               RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta *
+                               (i % nb_slots);
+               }
+
+               const char *type_str = "Atomic";
+               switch (wkr_q_conf.schedule_type) {
+               case RTE_SCHED_TYPE_ORDERED:
+                       type_str = "Ordered";
+                       break;
+               case RTE_SCHED_TYPE_PARALLEL:
+                       type_str = "Parallel";
+                       break;
+               }
+               printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
+                               wkr_q_conf.priority);
+       }
+
+       printf("\n");
+       if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
+               wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
+       if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
+               wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
+
+       /* set up one port per worker, linking to all stage queues */
+       for (i = 0; i < cdata.num_workers; i++) {
+               struct worker_data *w = &worker_data[i];
+               w->dev_id = dev_id;
+               if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
+                       printf("Error setting up port %d\n", i);
+                       return -1;
+               }
+
+               if (rte_event_port_link(dev_id, i, NULL, NULL, 0)
+                               != nb_queues) {
+                       printf("%d: error creating link for port %d\n",
+                                       __LINE__, i);
+                       return -1;
+               }
+               w->port_id = i;
+       }
+       /*
+        * Reduce the load on ingress event queue by splitting the traffic
+        * across multiple event queues.
+        * for example, nb_stages =  2 and nb_ethdev = 2 then
+        *
+        *      nb_queues = (2 * 2) + 2 = 6 (non atq)
+        *      rx_stride = 3
+        *
+        * So, traffic is split across queue 0 and queue 3 since queue id for
+        * rx adapter is chosen <ethport_id> * <rx_stride> i.e in the above
+        * case eth port 0, 1 will inject packets into event queue 0, 3
+        * respectively.
+        *
+        * This forms two set of queue pipelines 0->1->2->tx and 3->4->5->tx.
+        */
+       cdata.rx_stride = nb_slots;
+       ret = rte_event_dev_service_id_get(dev_id,
+                               &fdata->evdev_service_id);
+       if (ret != -ESRCH && ret != 0) {
+               printf("Error getting the service ID\n");
+               return -1;
+       }
+       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;
+}
+
+
+struct rx_adptr_services {
+       uint16_t nb_rx_adptrs;
+       uint32_t *rx_adpt_arr;
+};
+
+static int32_t
+service_rx_adapter(void *arg)
+{
+       int i;
+       struct rx_adptr_services *adptr_services = arg;
+
+       for (i = 0; i < adptr_services->nb_rx_adptrs; i++)
+               rte_service_run_iter_on_app_lcore(
+                               adptr_services->rx_adpt_arr[i], 1);
+       return 0;
+}
+
+static void
+init_rx_adapter(uint16_t nb_ports)
+{
+       int i;
+       int ret;
+       uint8_t evdev_id = 0;
+       struct rx_adptr_services *adptr_services = NULL;
+       struct rte_event_dev_info dev_info;
+
+       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,
+       };
+
+       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;
+
+
+       struct rte_event_eth_rx_adapter_queue_conf queue_conf = {
+               .ev.sched_type = cdata.queue_type,
+       };
+
+       for (i = 0; i < nb_ports; i++) {
+               uint32_t cap;
+               uint32_t service_id;
+
+               ret = rte_event_eth_rx_adapter_create(i, evdev_id, &rx_p_conf);
+               if (ret)
+                       rte_exit(EXIT_FAILURE,
+                                       "failed to create rx adapter[%d]",
+                                       cdata.rx_adapter_id);
+
+               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");
+
+               queue_conf.ev.queue_id = cdata.rx_stride ?
+                       (i * cdata.rx_stride)
+                       : (uint8_t)cdata.qid[0];
+
+               ret = rte_event_eth_rx_adapter_queue_add(i, i, -1, &queue_conf);
+               if (ret)
+                       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,
+                                       &service_id);
+                       if (ret != -ESRCH && ret != 0) {
+                               rte_exit(EXIT_FAILURE,
+                               "Error getting the service ID for rx adptr\n");
+                       }
+
+                       rte_service_runstate_set(service_id, 1);
+                       rte_service_set_runstate_mapped_check(service_id, 0);
+
+                       adptr_services->nb_rx_adptrs++;
+                       adptr_services->rx_adpt_arr = rte_realloc(
+                                       adptr_services->rx_adpt_arr,
+                                       adptr_services->nb_rx_adptrs *
+                                       sizeof(uint32_t), 0);
+                       adptr_services->rx_adpt_arr[
+                               adptr_services->nb_rx_adptrs - 1] =
+                               service_id;
+               }
+
+               ret = rte_event_eth_rx_adapter_start(i);
+               if (ret)
+                       rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
+                                       cdata.rx_adapter_id);
+       }
+
+       if (adptr_services->nb_rx_adptrs) {
+               struct rte_service_spec service;
+
+               memset(&service, 0, sizeof(struct rte_service_spec));
+               snprintf(service.name, sizeof(service.name), "rx_service");
+               service.callback = service_rx_adapter;
+               service.callback_userdata = (void *)adptr_services;
+
+               int32_t ret = rte_service_component_register(&service,
+                               &fdata->rxadptr_service_id);
+               if (ret)
+                       rte_exit(EXIT_FAILURE,
+                               "Rx adapter[%d] service register failed",
+                               cdata.rx_adapter_id);
+
+               rte_service_runstate_set(fdata->rxadptr_service_id, 1);
+               rte_service_component_runstate_set(fdata->rxadptr_service_id,
+                               1);
+               rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id,
+                               0);
+       } else {
+               memset(fdata->rx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
+               rte_free(adptr_services);
+       }
+
+       if (!adptr_services->nb_rx_adptrs && fdata->cap.consumer == NULL &&
+                       (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)
+{
+       int i;
+       int ret;
+       uint32_t cap = 0;
+       uint8_t rx_needed = 0;
+       struct rte_event_dev_info eventdev_info;
+
+       memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
+       rte_event_dev_info_get(0, &eventdev_info);
+
+       for (i = 0; i < rte_eth_dev_count(); 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");
+               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))) {
+               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);
+               rte_exit(-1, "Fix core masks\n");
+       }
+}
+
+void
+set_worker_tx_setup_data(struct setup_data *caps, bool burst)
+{
+       RTE_SET_USED(burst);
+       caps->worker = worker_do_tx;
+
+       memset(fdata->tx_core, 0, sizeof(unsigned int) * MAX_NUM_CORE);
+
+       caps->check_opt = worker_tx_opt_check;
+       caps->consumer = NULL;
+       caps->scheduler = schedule_devices;
+       caps->evdev_setup = setup_eventdev_worker_tx;
+       caps->adptr_setup = init_rx_adapter;
+}