[dpdk-dev] [PATCH v2 08/15] examples/eventdev: add thread safe Tx worker pipeline

[Pavan Nikhilesh]

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 at caviumnetworks.com>
---

 v2 Changes:
 - Redo function names(Harry)

 examples/eventdev_pipeline_sw_pmd/Makefile         |   1 +
 examples/eventdev_pipeline_sw_pmd/main.c           |  18 +-
 .../eventdev_pipeline_sw_pmd/pipeline_common.h     |   5 +
 .../eventdev_pipeline_sw_pmd/pipeline_worker_tx.c  | 425 +++++++++++++++++++++
 4 files changed, 447 insertions(+), 2 deletions(-)
 create mode 100644 examples/eventdev_pipeline_sw_pmd/pipeline_worker_tx.c

diff --git a/examples/eventdev_pipeline_sw_pmd/Makefile b/examples/eventdev_pipeline_sw_pmd/Makefile
index 5e30556fb..59ee9840a 100644
--- a/examples/eventdev_pipeline_sw_pmd/Makefile
+++ b/examples/eventdev_pipeline_sw_pmd/Makefile
@@ -43,6 +43,7 @@ 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)
diff --git a/examples/eventdev_pipeline_sw_pmd/main.c b/examples/eventdev_pipeline_sw_pmd/main.c
index 947c5f786..f877e695b 100644
--- a/examples/eventdev_pipeline_sw_pmd/main.c
+++ b/examples/eventdev_pipeline_sw_pmd/main.c
@@ -381,9 +381,20 @@ 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));

@@ -391,7 +402,10 @@ 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
diff --git a/examples/eventdev_pipeline_sw_pmd/pipeline_common.h b/examples/eventdev_pipeline_sw_pmd/pipeline_common.h
index d58059b78..e06320050 100644
--- a/examples/eventdev_pipeline_sw_pmd/pipeline_common.h
+++ b/examples/eventdev_pipeline_sw_pmd/pipeline_common.h
@@ -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
index 000000000..397b1013f
--- /dev/null
+++ b/examples/eventdev_pipeline_sw_pmd/pipeline_worker_tx.c
@@ -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;
+}
--
2.15.1