DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH] examples/l3fwd: add changes to use event vector
@ 2021-08-30 19:15 Shijith Thotton
  2021-09-04  7:46 ` [dpdk-dev] [PATCH v2] " Shijith Thotton
  0 siblings, 1 reply; 6+ messages in thread
From: Shijith Thotton @ 2021-08-30 19:15 UTC (permalink / raw)
  To: dev; +Cc: Shijith Thotton, jerinj, ndabilpuram, pbhagavatula, schalla

Added changes to receive packets as event vector. By default this is
disabled and can be enabled using the option --enable-vector. Vector
size and timeout to form the vector can be configured using options
--vector-size and --vector-tmo-ns.

Example:
	dpdk-l3fwd -l 0-3 -n 4 -- -p 0x03 --mode=eventdev \
	--eventq-sched=ordered --enable-vector --vector-size 16

Signed-off-by: Shijith Thotton <sthotton@marvell.com>
---
Depends-on: series=18322 (eventdev: simplify Rx adapter event vector config)

 doc/guides/sample_app_ug/l3_forward.rst    |   7 +
 examples/l3fwd/l3fwd.h                     |  26 ++++
 examples/l3fwd/l3fwd_em.c                  | 104 ++++++++++++++
 examples/l3fwd/l3fwd_em.h                  |  29 ++++
 examples/l3fwd/l3fwd_em_hlm.h              |  66 +++++++++
 examples/l3fwd/l3fwd_em_sequential.h       |  23 +++
 examples/l3fwd/l3fwd_event.c               |  57 +++++---
 examples/l3fwd/l3fwd_event.h               |  32 +++++
 examples/l3fwd/l3fwd_event_internal_port.c |  27 +++-
 examples/l3fwd/l3fwd_fib.c                 | 157 +++++++++++++++++++++
 examples/l3fwd/l3fwd_lpm.c                 | 113 +++++++++++++++
 examples/l3fwd/main.c                      |  58 ++++++++
 12 files changed, 676 insertions(+), 23 deletions(-)

diff --git a/doc/guides/sample_app_ug/l3_forward.rst b/doc/guides/sample_app_ug/l3_forward.rst
index 2d5cd5f1c0..e20f4a558d 100644
--- a/doc/guides/sample_app_ug/l3_forward.rst
+++ b/doc/guides/sample_app_ug/l3_forward.rst
@@ -74,6 +74,7 @@ The application has a number of command line options::
                              [--mode]
                              [--eventq-sched]
                              [--event-eth-rxqs]
+                             [--enable-vector [--vector-size SIZE] [--vector-tmo-ns NS]]"
                              [-E]
                              [-L]
 
@@ -115,6 +116,12 @@ Where,
 
 * ``--event-eth-rxqs:`` Optional, Number of ethernet RX queues per device. Only valid if --mode=eventdev.
 
+* ``--enable-vector:`` Optional, Enable event vectorization. Only valid if --mode=eventdev.
+
+* ``--vector-size:`` Optional, Max vector size if event vectorization is enabled.
+
+* ``--vector-tmo-ns:`` Optional, Max timeout to form vector in nanoseconds if event vectorization is enabled.
+
 * ``-E:`` Optional, enable exact match,
   legacy flag, please use ``--lookup=em`` instead.
 
diff --git a/examples/l3fwd/l3fwd.h b/examples/l3fwd/l3fwd.h
index a808d60247..9607ee0fbb 100644
--- a/examples/l3fwd/l3fwd.h
+++ b/examples/l3fwd/l3fwd.h
@@ -28,6 +28,8 @@
 #define MEMPOOL_CACHE_SIZE 256
 #define MAX_RX_QUEUE_PER_LCORE 16
 
+#define VECTOR_SIZE_DEFAULT   MAX_PKT_BURST
+#define VECTOR_TMO_NS_DEFAULT 1E6 /* 1ms */
 /*
  * Try to avoid TX buffering if we have at least MAX_TX_BURST packets to send.
  */
@@ -221,6 +223,14 @@ int
 lpm_event_main_loop_tx_q(__rte_unused void *dummy);
 int
 lpm_event_main_loop_tx_q_burst(__rte_unused void *dummy);
+int
+lpm_event_main_loop_tx_d_vector(__rte_unused void *dummy);
+int
+lpm_event_main_loop_tx_d_burst_vector(__rte_unused void *dummy);
+int
+lpm_event_main_loop_tx_q_vector(__rte_unused void *dummy);
+int
+lpm_event_main_loop_tx_q_burst_vector(__rte_unused void *dummy);
 
 int
 em_event_main_loop_tx_d(__rte_unused void *dummy);
@@ -230,6 +240,14 @@ int
 em_event_main_loop_tx_q(__rte_unused void *dummy);
 int
 em_event_main_loop_tx_q_burst(__rte_unused void *dummy);
+int
+em_event_main_loop_tx_d_vector(__rte_unused void *dummy);
+int
+em_event_main_loop_tx_d_burst_vector(__rte_unused void *dummy);
+int
+em_event_main_loop_tx_q_vector(__rte_unused void *dummy);
+int
+em_event_main_loop_tx_q_burst_vector(__rte_unused void *dummy);
 
 int
 fib_event_main_loop_tx_d(__rte_unused void *dummy);
@@ -239,6 +257,14 @@ int
 fib_event_main_loop_tx_q(__rte_unused void *dummy);
 int
 fib_event_main_loop_tx_q_burst(__rte_unused void *dummy);
+int
+fib_event_main_loop_tx_d_vector(__rte_unused void *dummy);
+int
+fib_event_main_loop_tx_d_burst_vector(__rte_unused void *dummy);
+int
+fib_event_main_loop_tx_q_vector(__rte_unused void *dummy);
+int
+fib_event_main_loop_tx_q_burst_vector(__rte_unused void *dummy);
 
 
 /* Return ipv4/ipv6 fwd lookup struct for LPM, EM or FIB. */
diff --git a/examples/l3fwd/l3fwd_em.c b/examples/l3fwd/l3fwd_em.c
index 2a8ab6aab5..4efa523cff 100644
--- a/examples/l3fwd/l3fwd_em.c
+++ b/examples/l3fwd/l3fwd_em.c
@@ -878,6 +878,110 @@ em_event_main_loop_tx_q_burst(__rte_unused void *dummy)
 	return 0;
 }
 
+/* Same eventdev loop for single and burst of vector */
+static __rte_always_inline void
+em_event_loop_vector(struct l3fwd_event_resources *evt_rsrc,
+		     const uint8_t flags)
+{
+	const int event_p_id = l3fwd_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 uint8_t event_d_id = evt_rsrc->event_d_id;
+	const uint16_t deq_len = evt_rsrc->deq_depth;
+	struct rte_event events[MAX_PKT_BURST];
+	struct lcore_conf *lconf;
+	unsigned int lcore_id;
+	int i, nb_enq, nb_deq;
+
+	if (event_p_id < 0)
+		return;
+
+	lcore_id = rte_lcore_id();
+	lconf = &lcore_conf[lcore_id];
+
+	RTE_LOG(INFO, L3FWD, "entering %s on lcore %u\n", __func__, lcore_id);
+
+	while (!force_quit) {
+		/* Read events from RX queues */
+		nb_deq = rte_event_dequeue_burst(event_d_id, event_p_id, events,
+						 deq_len, 0);
+		if (nb_deq == 0) {
+			rte_pause();
+			continue;
+		}
+
+		for (i = 0; i < nb_deq; i++) {
+			if (flags & L3FWD_EVENT_TX_ENQ) {
+				events[i].queue_id = tx_q_id;
+				events[i].op = RTE_EVENT_OP_FORWARD;
+			}
+
+#if defined RTE_ARCH_X86 || defined __ARM_NEON
+			l3fwd_em_process_event_vector(events[i].vec, lconf);
+#else
+			l3fwd_em_no_opt_process_event_vector(events[i].vec,
+							     lconf);
+#endif
+			if (flags & L3FWD_EVENT_TX_DIRECT)
+				event_vector_txq_set(events[i].vec, 0);
+		}
+
+		if (flags & L3FWD_EVENT_TX_ENQ) {
+			nb_enq = rte_event_enqueue_burst(event_d_id, event_p_id,
+							 events, nb_deq);
+			while (nb_enq < nb_deq && !force_quit)
+				nb_enq += rte_event_enqueue_burst(event_d_id,
+						event_p_id, events + nb_enq,
+						nb_deq - nb_enq);
+		}
+
+		if (flags & L3FWD_EVENT_TX_DIRECT) {
+			nb_enq = rte_event_eth_tx_adapter_enqueue(event_d_id,
+						event_p_id, events, nb_deq, 0);
+			while (nb_enq < nb_deq && !force_quit)
+				nb_enq += rte_event_eth_tx_adapter_enqueue(
+					event_d_id, event_p_id, events + nb_enq,
+					nb_deq - nb_enq, 0);
+		}
+	}
+}
+
+int __rte_noinline
+em_event_main_loop_tx_d_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	em_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_DIRECT);
+	return 0;
+}
+
+int __rte_noinline
+em_event_main_loop_tx_d_burst_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	em_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_DIRECT);
+	return 0;
+}
+
+int __rte_noinline
+em_event_main_loop_tx_q_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	em_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_ENQ);
+	return 0;
+}
+
+int __rte_noinline
+em_event_main_loop_tx_q_burst_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	em_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_ENQ);
+	return 0;
+}
+
 /* Initialize exact match (hash) parameters. 8< */
 void
 setup_hash(const int socketid)
diff --git a/examples/l3fwd/l3fwd_em.h b/examples/l3fwd/l3fwd_em.h
index b992a21da4..6c13b09bca 100644
--- a/examples/l3fwd/l3fwd_em.h
+++ b/examples/l3fwd/l3fwd_em.h
@@ -175,4 +175,33 @@ l3fwd_em_no_opt_process_events(int nb_rx, struct rte_event **events,
 		l3fwd_em_simple_process(events[j]->mbuf, qconf);
 }
 
+static inline void
+l3fwd_em_no_opt_process_event_vector(struct rte_event_vector *vec,
+				     struct lcore_conf *qconf)
+{
+	int32_t i;
+
+	/* Prefetch first packets */
+	for (i = 0; i < PREFETCH_OFFSET && i < vec->nb_elem; i++)
+		rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[i], void *));
+
+	event_vector_attr_init(vec, vec->mbufs[0]);
+
+	/*
+	 * Prefetch and forward already prefetched packets.
+	 */
+	for (i = 0; i < (vec->nb_elem - PREFETCH_OFFSET); i++) {
+		rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[i + PREFETCH_OFFSET],
+					       void *));
+		l3fwd_em_simple_process(vec->mbufs[i], qconf);
+		event_vector_attr_update(vec, vec->mbufs[i]);
+	}
+
+	/* Forward remaining prefetched packets */
+	for (; i < vec->nb_elem; i++) {
+		l3fwd_em_simple_process(vec->mbufs[i], qconf);
+		event_vector_attr_update(vec, vec->mbufs[i]);
+	}
+}
+
 #endif /* __L3FWD_EM_H__ */
diff --git a/examples/l3fwd/l3fwd_em_hlm.h b/examples/l3fwd/l3fwd_em_hlm.h
index 278707c18c..537fa54871 100644
--- a/examples/l3fwd/l3fwd_em_hlm.h
+++ b/examples/l3fwd/l3fwd_em_hlm.h
@@ -318,4 +318,70 @@ l3fwd_em_process_events(int nb_rx, struct rte_event **ev,
 		process_packet(pkts_burst[j], &pkts_burst[j]->port);
 	}
 }
+
+static inline void
+l3fwd_em_process_event_vector(struct rte_event_vector *vec,
+			      struct lcore_conf *qconf)
+{
+	int32_t i, j, n, pos;
+	uint16_t dst_port[MAX_PKT_BURST];
+
+	for (j = 0; j < EM_HASH_LOOKUP_COUNT && j < vec->nb_elem; j++)
+		rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[j],
+					       struct rte_ether_hdr *) + 1);
+
+	event_vector_attr_init(vec, vec->mbufs[0]);
+	n = RTE_ALIGN_FLOOR(vec->nb_elem, EM_HASH_LOOKUP_COUNT);
+	for (j = 0; j < n; j += EM_HASH_LOOKUP_COUNT) {
+		uint32_t pkt_type = RTE_PTYPE_L3_MASK |
+				    RTE_PTYPE_L4_TCP | RTE_PTYPE_L4_UDP;
+		uint32_t l3_type, tcp_or_udp;
+
+		for (i = 0; i < EM_HASH_LOOKUP_COUNT; i++)
+			pkt_type &= vec->mbufs[j + i]->packet_type;
+
+		l3_type = pkt_type & RTE_PTYPE_L3_MASK;
+		tcp_or_udp = pkt_type & (RTE_PTYPE_L4_TCP | RTE_PTYPE_L4_UDP);
+
+		for (i = 0, pos = j + EM_HASH_LOOKUP_COUNT;
+		     i < EM_HASH_LOOKUP_COUNT && pos < vec->nb_elem;
+		     i++, pos++) {
+			rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[pos],
+					       struct rte_ether_hdr *) + 1);
+		}
+
+		if (tcp_or_udp && (l3_type == RTE_PTYPE_L3_IPV4)) {
+			em_get_dst_port_ipv4xN_events(qconf, &vec->mbufs[j],
+						      &dst_port[j]);
+		} else if (tcp_or_udp && (l3_type == RTE_PTYPE_L3_IPV6)) {
+			em_get_dst_port_ipv6xN_events(qconf, &vec->mbufs[j],
+						      &dst_port[j]);
+		} else {
+			for (i = 0; i < EM_HASH_LOOKUP_COUNT; i++) {
+				vec->mbufs[j + i]->port = em_get_dst_port(
+					qconf, vec->mbufs[j + i],
+					vec->mbufs[j + i]->port);
+				process_packet(vec->mbufs[j + i],
+					       &vec->mbufs[j + i]->port);
+				event_vector_attr_update(vec,
+							 vec->mbufs[j + i]);
+			}
+			continue;
+		}
+		processx4_step3(&vec->mbufs[j], &dst_port[j]);
+
+		for (i = 0; i < EM_HASH_LOOKUP_COUNT; i++) {
+			vec->mbufs[j + i]->port = dst_port[j + i];
+			event_vector_attr_update(vec, vec->mbufs[j + i]);
+		}
+	}
+
+	for (; j < vec->nb_elem; j++) {
+		vec->mbufs[j]->port = em_get_dst_port(qconf, vec->mbufs[j],
+						      vec->mbufs[j]->port);
+		process_packet(vec->mbufs[j], &vec->mbufs[j]->port);
+		event_vector_attr_update(vec, vec->mbufs[j]);
+	}
+}
+
 #endif /* __L3FWD_EM_HLM_H__ */
diff --git a/examples/l3fwd/l3fwd_em_sequential.h b/examples/l3fwd/l3fwd_em_sequential.h
index 6170052cf8..3784058d3f 100644
--- a/examples/l3fwd/l3fwd_em_sequential.h
+++ b/examples/l3fwd/l3fwd_em_sequential.h
@@ -121,4 +121,27 @@ l3fwd_em_process_events(int nb_rx, struct rte_event **events,
 		process_packet(mbuf, &mbuf->port);
 	}
 }
+
+static inline void
+l3fwd_em_process_event_vector(struct rte_event_vector *vec,
+			      struct lcore_conf *qconf)
+{
+	int32_t i, j;
+
+	rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[0],
+				       struct rte_ether_hdr *) + 1);
+	event_vector_attr_init(vec, vec->mbufs[0]);
+
+	for (i = 0, j = 1; i < vec->nb_elem; i++, j++) {
+		struct rte_mbuf *mbuf = vec->mbufs[i];
+
+		if (j < vec->nb_elem)
+			rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[j],
+					       struct rte_ether_hdr *) + 1);
+		mbuf->port = em_get_dst_port(qconf, mbuf, mbuf->port);
+		process_packet(mbuf, &mbuf->port);
+		event_vector_attr_update(vec, mbuf);
+	}
+}
+
 #endif /* __L3FWD_EM_SEQUENTIAL_H__ */
diff --git a/examples/l3fwd/l3fwd_event.c b/examples/l3fwd/l3fwd_event.c
index 961860ea18..29172e590b 100644
--- a/examples/l3fwd/l3fwd_event.c
+++ b/examples/l3fwd/l3fwd_event.c
@@ -215,23 +215,35 @@ void
 l3fwd_event_resource_setup(struct rte_eth_conf *port_conf)
 {
 	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
-	const event_loop_cb lpm_event_loop[2][2] = {
-		[0][0] = lpm_event_main_loop_tx_d,
-		[0][1] = lpm_event_main_loop_tx_d_burst,
-		[1][0] = lpm_event_main_loop_tx_q,
-		[1][1] = lpm_event_main_loop_tx_q_burst,
+	const event_loop_cb lpm_event_loop[2][2][2] = {
+		[0][0][0] = lpm_event_main_loop_tx_d,
+		[0][0][1] = lpm_event_main_loop_tx_d_burst,
+		[0][1][0] = lpm_event_main_loop_tx_q,
+		[0][1][1] = lpm_event_main_loop_tx_q_burst,
+		[1][0][0] = lpm_event_main_loop_tx_d_vector,
+		[1][0][1] = lpm_event_main_loop_tx_d_burst_vector,
+		[1][1][0] = lpm_event_main_loop_tx_q_vector,
+		[1][1][1] = lpm_event_main_loop_tx_q_burst_vector,
 	};
-	const event_loop_cb em_event_loop[2][2] = {
-		[0][0] = em_event_main_loop_tx_d,
-		[0][1] = em_event_main_loop_tx_d_burst,
-		[1][0] = em_event_main_loop_tx_q,
-		[1][1] = em_event_main_loop_tx_q_burst,
+	const event_loop_cb em_event_loop[2][2][2] = {
+		[0][0][0] = em_event_main_loop_tx_d,
+		[0][0][1] = em_event_main_loop_tx_d_burst,
+		[0][1][0] = em_event_main_loop_tx_q,
+		[0][1][1] = em_event_main_loop_tx_q_burst,
+		[1][0][0] = em_event_main_loop_tx_d_vector,
+		[1][0][1] = em_event_main_loop_tx_d_burst_vector,
+		[1][1][0] = em_event_main_loop_tx_q_vector,
+		[1][1][1] = em_event_main_loop_tx_q_burst_vector,
 	};
-	const event_loop_cb fib_event_loop[2][2] = {
-		[0][0] = fib_event_main_loop_tx_d,
-		[0][1] = fib_event_main_loop_tx_d_burst,
-		[1][0] = fib_event_main_loop_tx_q,
-		[1][1] = fib_event_main_loop_tx_q_burst,
+	const event_loop_cb fib_event_loop[2][2][2] = {
+		[0][0][0] = fib_event_main_loop_tx_d,
+		[0][0][1] = fib_event_main_loop_tx_d_burst,
+		[0][1][0] = fib_event_main_loop_tx_q,
+		[0][1][1] = fib_event_main_loop_tx_q_burst,
+		[1][0][0] = fib_event_main_loop_tx_d_vector,
+		[1][0][1] = fib_event_main_loop_tx_d_burst_vector,
+		[1][1][0] = fib_event_main_loop_tx_q_vector,
+		[1][1][1] = fib_event_main_loop_tx_q_burst_vector,
 	};
 	uint32_t event_queue_cfg;
 	int ret;
@@ -265,12 +277,15 @@ l3fwd_event_resource_setup(struct rte_eth_conf *port_conf)
 	if (ret < 0)
 		rte_exit(EXIT_FAILURE, "Error in starting eventdev");
 
-	evt_rsrc->ops.lpm_event_loop = lpm_event_loop[evt_rsrc->tx_mode_q]
-						       [evt_rsrc->has_burst];
+	evt_rsrc->ops.lpm_event_loop =
+		lpm_event_loop[evt_rsrc->vector_enabled][evt_rsrc->tx_mode_q]
+			      [evt_rsrc->has_burst];
 
-	evt_rsrc->ops.em_event_loop = em_event_loop[evt_rsrc->tx_mode_q]
-						       [evt_rsrc->has_burst];
+	evt_rsrc->ops.em_event_loop =
+		em_event_loop[evt_rsrc->vector_enabled][evt_rsrc->tx_mode_q]
+			     [evt_rsrc->has_burst];
 
-	evt_rsrc->ops.fib_event_loop = fib_event_loop[evt_rsrc->tx_mode_q]
-						       [evt_rsrc->has_burst];
+	evt_rsrc->ops.fib_event_loop =
+		fib_event_loop[evt_rsrc->vector_enabled][evt_rsrc->tx_mode_q]
+			      [evt_rsrc->has_burst];
 }
diff --git a/examples/l3fwd/l3fwd_event.h b/examples/l3fwd/l3fwd_event.h
index 3ad1902ab5..a15a369943 100644
--- a/examples/l3fwd/l3fwd_event.h
+++ b/examples/l3fwd/l3fwd_event.h
@@ -65,6 +65,7 @@ struct l3fwd_event_resources {
 	uint8_t disable_implicit_release;
 	struct l3fwd_event_setup_ops ops;
 	struct rte_mempool * (*pkt_pool)[NB_SOCKETS];
+	struct rte_mempool **vec_pool;
 	struct l3fwd_event_queues evq;
 	struct l3fwd_event_ports evp;
 	uint32_t port_mask;
@@ -76,8 +77,39 @@ struct l3fwd_event_resources {
 	uint8_t has_burst;
 	uint8_t enabled;
 	uint8_t eth_rx_queues;
+	uint8_t vector_enabled;
+	uint16_t vector_size;
+	uint64_t vector_tmo_ns;
 };
 
+static inline void
+event_vector_attr_init(struct rte_event_vector *vec, struct rte_mbuf *mbuf)
+{
+	vec->attr_valid = 1;
+	vec->port = mbuf->port;
+	vec->queue = rte_event_eth_tx_adapter_txq_get(mbuf);
+}
+
+static inline void
+event_vector_attr_update(struct rte_event_vector *vec, struct rte_mbuf *mbuf)
+{
+	if (vec->attr_valid && (vec->port != mbuf->port))
+		vec->attr_valid = 0;
+}
+
+static inline void
+event_vector_txq_set(struct rte_event_vector *vec, uint16_t txq)
+{
+	int i;
+
+	if (vec->attr_valid)
+		vec->queue = txq;
+
+	for (i = 0; i < vec->nb_elem; i++)
+		rte_event_eth_tx_adapter_txq_set(vec->mbufs[i], txq);
+
+}
+
 struct l3fwd_event_resources *l3fwd_get_eventdev_rsrc(void);
 void l3fwd_event_resource_setup(struct rte_eth_conf *port_conf);
 int l3fwd_get_free_event_port(struct l3fwd_event_resources *eventdev_rsrc);
diff --git a/examples/l3fwd/l3fwd_event_internal_port.c b/examples/l3fwd/l3fwd_event_internal_port.c
index 9916a7f556..bb681e0d5a 100644
--- a/examples/l3fwd/l3fwd_event_internal_port.c
+++ b/examples/l3fwd/l3fwd_event_internal_port.c
@@ -215,12 +215,35 @@ l3fwd_rx_tx_adapter_setup_internal_port(void)
 		rte_panic("Failed to allocate memory for Rx adapter\n");
 	}
 
-
 	RTE_ETH_FOREACH_DEV(port_id) {
 		if ((evt_rsrc->port_mask & (1 << port_id)) == 0)
 			continue;
+
+		if (evt_rsrc->vector_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 = evt_rsrc->vector_size;
+				eth_q_conf.vector_timeout_ns =
+							evt_rsrc->vector_tmo_ns;
+				eth_q_conf.vector_mp = evt_rsrc->per_port_pool ?
+						evt_rsrc->vec_pool[port_id] :
+						evt_rsrc->vec_pool[0];
+				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);
+						      &evt_rsrc->def_p_conf);
 		if (ret)
 			rte_panic("Failed to create rx adapter[%d]\n",
 				  adapter_id);
diff --git a/examples/l3fwd/l3fwd_fib.c b/examples/l3fwd/l3fwd_fib.c
index f8d6a3ac39..e57f0c108f 100644
--- a/examples/l3fwd/l3fwd_fib.c
+++ b/examples/l3fwd/l3fwd_fib.c
@@ -412,6 +412,163 @@ fib_event_main_loop_tx_q_burst(__rte_unused void *dummy)
 	return 0;
 }
 
+static __rte_always_inline void
+fib_process_event_vector(struct rte_event_vector *vec)
+{
+	uint8_t ipv6_arr[MAX_PKT_BURST][RTE_FIB6_IPV6_ADDR_SIZE];
+	uint64_t hopsv4[MAX_PKT_BURST], hopsv6[MAX_PKT_BURST];
+	uint32_t ipv4_arr_assem, ipv6_arr_assem;
+	uint32_t ipv4_arr[MAX_PKT_BURST];
+	uint8_t type_arr[MAX_PKT_BURST];
+	uint32_t ipv4_cnt, ipv6_cnt;
+	struct lcore_conf *lconf;
+	uint16_t nh;
+	int i;
+
+	lconf = &lcore_conf[rte_lcore_id()];
+
+	/* Reset counters. */
+	ipv4_cnt = 0;
+	ipv6_cnt = 0;
+	ipv4_arr_assem = 0;
+	ipv6_arr_assem = 0;
+
+	/* Prefetch first packets. */
+	for (i = 0; i < FIB_PREFETCH_OFFSET && i < vec->nb_elem; i++)
+		rte_prefetch0(rte_pktmbuf_mtod(vec->mbufs[i], void *));
+
+	event_vector_attr_init(vec, vec->mbufs[0]);
+
+	/* Parse packet info and prefetch. */
+	for (i = 0; i < (vec->nb_elem - FIB_PREFETCH_OFFSET); i++) {
+		rte_prefetch0(rte_pktmbuf_mtod(
+			vec->mbufs[i + FIB_PREFETCH_OFFSET], void *));
+		fib_parse_packet(vec->mbufs[i], &ipv4_arr[ipv4_cnt], &ipv4_cnt,
+				 ipv6_arr[ipv6_cnt], &ipv6_cnt, &type_arr[i]);
+	}
+
+	/* Parse remaining packet info. */
+	for (; i < vec->nb_elem; i++)
+		fib_parse_packet(vec->mbufs[i], &ipv4_arr[ipv4_cnt], &ipv4_cnt,
+				 ipv6_arr[ipv6_cnt], &ipv6_cnt, &type_arr[i]);
+
+	/* Lookup IPv4 hops if IPv4 packets are present. */
+	if (likely(ipv4_cnt > 0))
+		rte_fib_lookup_bulk(lconf->ipv4_lookup_struct, ipv4_arr, hopsv4,
+				    ipv4_cnt);
+
+	/* Lookup IPv6 hops if IPv6 packets are present. */
+	if (ipv6_cnt > 0)
+		rte_fib6_lookup_bulk(lconf->ipv6_lookup_struct, ipv6_arr,
+				     hopsv6, ipv6_cnt);
+
+	/* Assign ports looked up in fib depending on IPv4 or IPv6 */
+	for (i = 0; i < vec->nb_elem; i++) {
+		if (type_arr[i])
+			nh = (uint16_t)hopsv4[ipv4_arr_assem++];
+		else
+			nh = (uint16_t)hopsv6[ipv6_arr_assem++];
+		if (nh != FIB_DEFAULT_HOP)
+			vec->mbufs[i]->port = nh;
+		event_vector_attr_update(vec, vec->mbufs[i]);
+	}
+}
+
+static __rte_always_inline void
+fib_event_loop_vector(struct l3fwd_event_resources *evt_rsrc,
+		      const uint8_t flags)
+{
+	const int event_p_id = l3fwd_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 uint8_t event_d_id = evt_rsrc->event_d_id;
+	const uint16_t deq_len = evt_rsrc->deq_depth;
+	struct rte_event events[MAX_PKT_BURST];
+	int nb_enq, nb_deq, i;
+
+	if (event_p_id < 0)
+		return;
+
+	RTE_LOG(INFO, L3FWD, "entering %s on lcore %u\n", __func__,
+		rte_lcore_id());
+
+	while (!force_quit) {
+		/* Read events from RX queues. */
+		nb_deq = rte_event_dequeue_burst(event_d_id, event_p_id, events,
+						 deq_len, 0);
+		if (nb_deq == 0) {
+			rte_pause();
+			continue;
+		}
+
+		for (i = 0; i < nb_deq; i++) {
+			if (flags & L3FWD_EVENT_TX_ENQ) {
+				events[i].queue_id = tx_q_id;
+				events[i].op = RTE_EVENT_OP_FORWARD;
+			}
+
+			fib_process_event_vector(events[i].vec);
+
+			if (flags & L3FWD_EVENT_TX_DIRECT)
+				event_vector_txq_set(events[i].vec, 0);
+		}
+
+		if (flags & L3FWD_EVENT_TX_ENQ) {
+			nb_enq = rte_event_enqueue_burst(event_d_id, event_p_id,
+							 events, nb_deq);
+			while (nb_enq < nb_deq && !force_quit)
+				nb_enq += rte_event_enqueue_burst(event_d_id,
+						event_p_id, events + nb_enq,
+						nb_deq - nb_enq);
+		}
+
+		if (flags & L3FWD_EVENT_TX_DIRECT) {
+			nb_enq = rte_event_eth_tx_adapter_enqueue(event_d_id,
+						event_p_id, events, nb_deq, 0);
+			while (nb_enq < nb_deq && !force_quit)
+				nb_enq += rte_event_eth_tx_adapter_enqueue(
+					event_d_id, event_p_id, events + nb_enq,
+					nb_deq - nb_enq, 0);
+		}
+	}
+}
+
+int __rte_noinline
+fib_event_main_loop_tx_d_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	fib_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_DIRECT);
+	return 0;
+}
+
+int __rte_noinline
+fib_event_main_loop_tx_d_burst_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	fib_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_DIRECT);
+	return 0;
+}
+
+int __rte_noinline
+fib_event_main_loop_tx_q_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	fib_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_ENQ);
+	return 0;
+}
+
+int __rte_noinline
+fib_event_main_loop_tx_q_burst_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	fib_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_ENQ);
+	return 0;
+}
+
 /* Function to setup fib. 8< */
 void
 setup_fib(const int socketid)
diff --git a/examples/l3fwd/l3fwd_lpm.c b/examples/l3fwd/l3fwd_lpm.c
index 7200160164..60e7a4df72 100644
--- a/examples/l3fwd/l3fwd_lpm.c
+++ b/examples/l3fwd/l3fwd_lpm.c
@@ -427,6 +427,119 @@ lpm_event_main_loop_tx_q_burst(__rte_unused void *dummy)
 	return 0;
 }
 
+static __rte_always_inline void
+lpm_process_event_vector(struct rte_event_vector *vec, struct lcore_conf *lconf)
+{
+	int i;
+
+	event_vector_attr_init(vec, vec->mbufs[0]);
+
+	for (i = 0; i < vec->nb_elem; i++) {
+		lpm_process_event_pkt(lconf, vec->mbufs[i]);
+		event_vector_attr_update(vec, vec->mbufs[i]);
+	}
+}
+
+/* Same eventdev loop for single and burst of vector */
+static __rte_always_inline void
+lpm_event_loop_vector(struct l3fwd_event_resources *evt_rsrc,
+		      const uint8_t flags)
+{
+	const int event_p_id = l3fwd_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 uint8_t event_d_id = evt_rsrc->event_d_id;
+	const uint16_t deq_len = evt_rsrc->deq_depth;
+	struct rte_event events[MAX_PKT_BURST];
+	struct lcore_conf *lconf;
+	unsigned int lcore_id;
+	int i, nb_enq, nb_deq;
+
+	if (event_p_id < 0)
+		return;
+
+	lcore_id = rte_lcore_id();
+	lconf = &lcore_conf[lcore_id];
+
+	RTE_LOG(INFO, L3FWD, "entering %s on lcore %u\n", __func__, lcore_id);
+
+	while (!force_quit) {
+		/* Read events from RX queues */
+		nb_deq = rte_event_dequeue_burst(event_d_id, event_p_id, events,
+						 deq_len, 0);
+		if (nb_deq == 0) {
+			rte_pause();
+			continue;
+		}
+
+		for (i = 0; i < nb_deq; i++) {
+			if (flags & L3FWD_EVENT_TX_ENQ) {
+				events[i].queue_id = tx_q_id;
+				events[i].op = RTE_EVENT_OP_FORWARD;
+			}
+
+			lpm_process_event_vector(events[i].vec, lconf);
+
+			if (flags & L3FWD_EVENT_TX_DIRECT)
+				event_vector_txq_set(events[i].vec, 0);
+		}
+
+		if (flags & L3FWD_EVENT_TX_ENQ) {
+			nb_enq = rte_event_enqueue_burst(event_d_id, event_p_id,
+							 events, nb_deq);
+			while (nb_enq < nb_deq && !force_quit)
+				nb_enq += rte_event_enqueue_burst(event_d_id,
+						event_p_id, events + nb_enq,
+						nb_deq - nb_enq);
+		}
+
+		if (flags & L3FWD_EVENT_TX_DIRECT) {
+			nb_enq = rte_event_eth_tx_adapter_enqueue(event_d_id,
+						event_p_id, events, nb_deq, 0);
+			while (nb_enq < nb_deq && !force_quit)
+				nb_enq += rte_event_eth_tx_adapter_enqueue(
+					event_d_id, event_p_id, events + nb_enq,
+					nb_deq - nb_enq, 0);
+		}
+	}
+}
+
+int __rte_noinline
+lpm_event_main_loop_tx_d_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	lpm_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_DIRECT);
+	return 0;
+}
+
+int __rte_noinline
+lpm_event_main_loop_tx_d_burst_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	lpm_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_DIRECT);
+	return 0;
+}
+
+int __rte_noinline
+lpm_event_main_loop_tx_q_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	lpm_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_ENQ);
+	return 0;
+}
+
+int __rte_noinline
+lpm_event_main_loop_tx_q_burst_vector(__rte_unused void *dummy)
+{
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
+
+	lpm_event_loop_vector(evt_rsrc, L3FWD_EVENT_TX_ENQ);
+	return 0;
+}
+
 void
 setup_lpm(const int socketid)
 {
diff --git a/examples/l3fwd/main.c b/examples/l3fwd/main.c
index 00ac267af1..7382a557aa 100644
--- a/examples/l3fwd/main.c
+++ b/examples/l3fwd/main.c
@@ -137,6 +137,7 @@ static struct rte_eth_conf port_conf = {
 };
 
 static struct rte_mempool *pktmbuf_pool[RTE_MAX_ETHPORTS][NB_SOCKETS];
+static struct rte_mempool *vector_pool[RTE_MAX_ETHPORTS];
 static uint8_t lkp_per_socket[NB_SOCKETS];
 
 struct l3fwd_lkp_mode {
@@ -334,6 +335,7 @@ print_usage(const char *prgname)
 		" [--per-port-pool]"
 		" [--mode]"
 		" [--eventq-sched]"
+		" [--enable-vector [--vector-size SIZE] [--vector-tmo-ns NS]]"
 		" [-E]"
 		" [-L]\n\n"
 
@@ -361,6 +363,9 @@ print_usage(const char *prgname)
 		"  --event-eth-rxqs: Number of ethernet RX queues per device.\n"
 		"                    Default: 1\n"
 		"                    Valid only if --mode=eventdev\n"
+		"  --enable-vector:  Enable event vectorization.\n"
+		"  --vector-size: Max vector size if event vectorization is enabled.\n"
+		"  --vector-tmo-ns: Max timeout to form vector in nanoseconds if event vectorization is enabled\n"
 		"  -E : Enable exact match, legacy flag please use --lookup=em instead\n"
 		"  -L : Enable longest prefix match, legacy flag please use --lookup=lpm instead\n\n",
 		prgname);
@@ -574,6 +579,10 @@ static const char short_options[] =
 #define CMD_LINE_OPT_EVENTQ_SYNC "eventq-sched"
 #define CMD_LINE_OPT_EVENT_ETH_RX_QUEUES "event-eth-rxqs"
 #define CMD_LINE_OPT_LOOKUP "lookup"
+#define CMD_LINE_OPT_ENABLE_VECTOR "enable-vector"
+#define CMD_LINE_OPT_VECTOR_SIZE "vector-size"
+#define CMD_LINE_OPT_VECTOR_TMO_NS "vector-tmo-ns"
+
 enum {
 	/* long options mapped to a short option */
 
@@ -592,6 +601,9 @@ enum {
 	CMD_LINE_OPT_EVENTQ_SYNC_NUM,
 	CMD_LINE_OPT_EVENT_ETH_RX_QUEUES_NUM,
 	CMD_LINE_OPT_LOOKUP_NUM,
+	CMD_LINE_OPT_ENABLE_VECTOR_NUM,
+	CMD_LINE_OPT_VECTOR_SIZE_NUM,
+	CMD_LINE_OPT_VECTOR_TMO_NS_NUM
 };
 
 static const struct option lgopts[] = {
@@ -608,6 +620,9 @@ static const struct option lgopts[] = {
 	{CMD_LINE_OPT_EVENT_ETH_RX_QUEUES, 1, 0,
 					CMD_LINE_OPT_EVENT_ETH_RX_QUEUES_NUM},
 	{CMD_LINE_OPT_LOOKUP, 1, 0, CMD_LINE_OPT_LOOKUP_NUM},
+	{CMD_LINE_OPT_ENABLE_VECTOR, 0, 0, CMD_LINE_OPT_ENABLE_VECTOR_NUM},
+	{CMD_LINE_OPT_VECTOR_SIZE, 1, 0, CMD_LINE_OPT_VECTOR_SIZE_NUM},
+	{CMD_LINE_OPT_VECTOR_TMO_NS, 1, 0, CMD_LINE_OPT_VECTOR_TMO_NS_NUM},
 	{NULL, 0, 0, 0}
 };
 
@@ -774,6 +789,16 @@ parse_args(int argc, char **argv)
 				return -1;
 			break;
 
+		case CMD_LINE_OPT_ENABLE_VECTOR_NUM:
+			printf("event vectorization is enabled\n");
+			evt_rsrc->vector_enabled = 1;
+			break;
+		case CMD_LINE_OPT_VECTOR_SIZE_NUM:
+			evt_rsrc->vector_size = strtol(optarg, NULL, 10);
+			break;
+		case CMD_LINE_OPT_VECTOR_TMO_NS_NUM:
+			evt_rsrc->vector_tmo_ns = strtoull(optarg, NULL, 10);
+			break;
 		default:
 			print_usage(prgname);
 			return -1;
@@ -795,6 +820,19 @@ parse_args(int argc, char **argv)
 		return -1;
 	}
 
+	if (evt_rsrc->vector_enabled && !evt_rsrc->vector_size) {
+		evt_rsrc->vector_size = VECTOR_SIZE_DEFAULT;
+		fprintf(stderr, "vector size set to default (%" PRIu16 ")\n",
+			evt_rsrc->vector_size);
+	}
+
+	if (evt_rsrc->vector_enabled && !evt_rsrc->vector_tmo_ns) {
+		evt_rsrc->vector_tmo_ns = VECTOR_TMO_NS_DEFAULT;
+		fprintf(stderr,
+			"vector timeout set to default (%" PRIu64 " ns)\n",
+			evt_rsrc->vector_tmo_ns);
+	}
+
 	/*
 	 * Nothing is selected, pick longest-prefix match
 	 * as default match.
@@ -833,6 +871,7 @@ print_ethaddr(const char *name, const struct rte_ether_addr *eth_addr)
 int
 init_mem(uint16_t portid, unsigned int nb_mbuf)
 {
+	struct l3fwd_event_resources *evt_rsrc = l3fwd_get_eventdev_rsrc();
 	struct lcore_conf *qconf;
 	int socketid;
 	unsigned lcore_id;
@@ -876,6 +915,24 @@ init_mem(uint16_t portid, unsigned int nb_mbuf)
 				lkp_per_socket[socketid] = 1;
 			}
 		}
+
+		if (evt_rsrc->vector_enabled && vector_pool[portid] == NULL) {
+			unsigned int nb_vec;
+
+			nb_vec = (nb_mbuf + evt_rsrc->vector_size - 1) /
+				 evt_rsrc->vector_size;
+			snprintf(s, sizeof(s), "vector_pool_%d", portid);
+			vector_pool[portid] = rte_event_vector_pool_create(
+				s, nb_vec, 0, evt_rsrc->vector_size, socketid);
+			if (vector_pool[portid] == NULL)
+				rte_exit(EXIT_FAILURE,
+					 "Failed to create vector pool for port %d\n",
+					 portid);
+			else
+				printf("Allocated vector pool for port %d\n",
+				       portid);
+		}
+
 		qconf = &lcore_conf[lcore_id];
 		qconf->ipv4_lookup_struct =
 			l3fwd_lkp.get_ipv4_lookup_struct(socketid);
@@ -1307,6 +1364,7 @@ main(int argc, char **argv)
 
 	evt_rsrc->per_port_pool = per_port_pool;
 	evt_rsrc->pkt_pool = pktmbuf_pool;
+	evt_rsrc->vec_pool = vector_pool;
 	evt_rsrc->port_mask = enabled_port_mask;
 	/* Configure eventdev parameters if user has requested */
 	if (evt_rsrc->enabled) {
-- 
2.25.1


^ permalink raw reply	[flat|nested] 6+ messages in thread
* Re: [dpdk-dev] [PATCH v2] examples/l3fwd: add changes to use event vector
@ 2021-09-22 15:36 Shijith Thotton
  0 siblings, 0 replies; 6+ messages in thread
From: Shijith Thotton @ 2021-09-22 15:36 UTC (permalink / raw)
  To: Thomas Monjalon
  Cc: dev, Jerin Jacob Kollanukkaran, Nithin Kumar Dabilpuram,
	Pavan Nikhilesh Bhagavatula, Srujana Challa, konstantin.ananyev,
	bruce.richardson, hemant.agrawal

>04/09/2021 09:46, Shijith Thotton:
>> +* ``--enable-vector:`` Optional, Enable event vectorization. Only valid if --
>mode=eventdev.
>
>As for the IPsec example, the option name --event-vector
>would be more appropriate because "vector" can refer to something else.
>

Will change option name to --event-vector.

>> +
>> +* ``--vector-size:`` Optional, Max vector size if event vectorization is enabled.
>> +
>> +* ``--vector-tmo-ns:`` Optional, Max timeout to form vector in nanoseconds if
>event vectorization is enabled.
>
>Better to have a common prefix "event" for all three options.
>

Makes sense, will change in next version.


^ permalink raw reply	[flat|nested] 6+ messages in thread
end of thread, other threads:[~2021-10-20 16:14 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-08-30 19:15 [dpdk-dev] [PATCH] examples/l3fwd: add changes to use event vector Shijith Thotton
2021-09-04  7:46 ` [dpdk-dev] [PATCH v2] " Shijith Thotton
2021-09-21 21:42   ` Thomas Monjalon
2021-09-27  4:57   ` [dpdk-dev] [PATCH v3] " Shijith Thotton
2021-10-20 16:13     ` Jerin Jacob
2021-09-22 15:36 [dpdk-dev] [PATCH v2] " Shijith Thotton

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).