-----Original Message-----
From: Jayatheerthan, Jay <redacted>
Sent: Thursday, March 25, 2021 4:07 PM
To: Pavan Nikhilesh Bhagavatula <redacted>; Jerin
Jacob Kollanukkaran [off-list ref]; Carrillo, Erik G
[off-list ref]; Gujjar, Abhinandan S
[off-list ref]; McDaniel, Timothy
[off-list ref]; hemant.agrawal@nxp.com; Van
Haaren, Harry [off-list ref]; mattias.ronnblom
[off-list ref]; Ma, Liang J
[off-list ref]
Cc: dev@dpdk.org
Subject: [EXT] RE: [dpdk-dev] [PATCH v5 4/8] eventdev: add Rx adapter
event vector support

External Email

----------------------------------------------------------------------

quoted

-----Original Message-----
From: pbhagavatula@marvell.com <redacted>
Sent: Wednesday, March 24, 2021 10:35 AM
To: jerinj@marvell.com; Jayatheerthan, Jay

[off-list ref]; Carrillo, Erik G
[off-list ref]; Gujjar,

quoted

Abhinandan S [off-list ref]; McDaniel, Timothy

[off-list ref]; hemant.agrawal@nxp.com; Van

quoted

Haaren, Harry [off-list ref]; mattias.ronnblom

[off-list ref]; Ma, Liang J

quoted

[off-list ref]
Cc: dev@dpdk.org; Pavan Nikhilesh <redacted>
Subject: [dpdk-dev] [PATCH v5 4/8] eventdev: add Rx adapter event

vector support

quoted

From: Pavan Nikhilesh <redacted>

Add event vector support for event eth Rx adapter, the

implementation

quoted

creates vector flows based on port and queue identifier of the

received

quoted

mbufs.

Signed-off-by: Pavan Nikhilesh <redacted>
---
 lib/librte_eventdev/eventdev_pmd.h            |   7 +-
 .../rte_event_eth_rx_adapter.c                | 257 ++++++++++++++++--
 lib/librte_eventdev/rte_eventdev.c            |   6 +-
 3 files changed, 250 insertions(+), 20 deletions(-)

diff --git a/lib/librte_eventdev/eventdev_pmd.h

b/lib/librte_eventdev/eventdev_pmd.h

quoted

index 9297f1433..0f724ac85 100644

--- a/lib/librte_eventdev/eventdev_pmd.h
+++ b/lib/librte_eventdev/eventdev_pmd.h

@@ -69,9 +69,10 @@ extern "C" {
 	} \
 } while (0)

-#define RTE_EVENT_ETH_RX_ADAPTER_SW_CAP \
-

((RTE_EVENT_ETH_RX_ADAPTER_CAP_OVERRIDE_FLOW_ID) |
\

quoted

-

(RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ))

quoted

+#define RTE_EVENT_ETH_RX_ADAPTER_SW_CAP

\

quoted

+	((RTE_EVENT_ETH_RX_ADAPTER_CAP_OVERRIDE_FLOW_ID) |

\

quoted

+	 (RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ) |

\

quoted

+	 (RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR))

 #define RTE_EVENT_CRYPTO_ADAPTER_SW_CAP \

RTE_EVENT_CRYPTO_ADAPTER_CAP_SESSION_PRIVATE_DATA

quoted

diff --git a/lib/librte_eventdev/rte_event_eth_rx_adapter.c

b/lib/librte_eventdev/rte_event_eth_rx_adapter.c

quoted

index ac8ba5bf0..c71990078 100644

--- a/lib/librte_eventdev/rte_event_eth_rx_adapter.c
+++ b/lib/librte_eventdev/rte_event_eth_rx_adapter.c

@@ -26,6 +26,10 @@
 #define BATCH_SIZE		32
 #define BLOCK_CNT_THRESHOLD	10
 #define ETH_EVENT_BUFFER_SIZE	(4*BATCH_SIZE)
+#define MAX_VECTOR_SIZE		1024
+#define MIN_VECTOR_SIZE		4
+#define MAX_VECTOR_NS		1E9
+#define MIN_VECTOR_NS		1E5

 #define ETH_RX_ADAPTER_SERVICE_NAME_LEN	32
 #define ETH_RX_ADAPTER_MEM_NAME_LEN	32

@@ -59,6 +63,20 @@ struct eth_rx_poll_entry {
 	uint16_t eth_rx_qid;
 };

+struct eth_rx_vector_data {
+	TAILQ_ENTRY(eth_rx_vector_data) next;
+	uint16_t port;
+	uint16_t queue;
+	uint16_t max_vector_count;
+	uint64_t event;
+	uint64_t ts;
+	uint64_t vector_timeout_ticks;
+	struct rte_mempool *vector_pool;
+	struct rte_event_vector *vector_ev;
+} __rte_cache_aligned;
+
+TAILQ_HEAD(eth_rx_vector_data_list, eth_rx_vector_data);
+
 /* Instance per adapter */
 struct rte_eth_event_enqueue_buffer {
 	/* Count of events in this buffer */

@@ -92,6 +110,14 @@ struct rte_event_eth_rx_adapter {
 	uint32_t wrr_pos;
 	/* Event burst buffer */
 	struct rte_eth_event_enqueue_buffer event_enqueue_buffer;
+	/* Vector enable flag */
+	uint8_t ena_vector;
+	/* Timestamp of previous vector expiry list traversal */
+	uint64_t prev_expiry_ts;
+	/* Minimum ticks to wait before traversing expiry list */
+	uint64_t vector_tmo_ticks;
+	/* vector list */
+	struct eth_rx_vector_data_list vector_list;
 	/* Per adapter stats */
 	struct rte_event_eth_rx_adapter_stats stats;
 	/* Block count, counts up to BLOCK_CNT_THRESHOLD */

@@ -198,9 +224,11 @@ struct eth_device_info {
 struct eth_rx_queue_info {
 	int queue_enabled;	/* True if added */
 	int intr_enabled;
+	uint8_t ena_vector;
 	uint16_t wt;		/* Polling weight */
 	uint32_t flow_id_mask;	/* Set to ~0 if app provides flow id else

0 */

quoted

 	uint64_t event;
+	struct eth_rx_vector_data vector_data;
 };

 static struct rte_event_eth_rx_adapter **event_eth_rx_adapter;

@@ -722,6 +750,9 @@ rxa_flush_event_buffer(struct

rte_event_eth_rx_adapter *rx_adapter)

quoted

 	    &rx_adapter->event_enqueue_buffer;
 	struct rte_event_eth_rx_adapter_stats *stats = &rx_adapter-
stats;

+	if (!buf->count)
+		return 0;
+
 	uint16_t n = rte_event_enqueue_new_burst(rx_adapter-
eventdev_id,
 					rx_adapter->event_port_id,
 					buf->events,

@@ -742,6 +773,72 @@ rxa_flush_event_buffer(struct

rte_event_eth_rx_adapter *rx_adapter)

quoted

 	return n;
 }

+static inline uint16_t
+rxa_create_event_vector(struct rte_event_eth_rx_adapter

*rx_adapter,

quoted

+			struct eth_rx_queue_info *queue_info,
+			struct rte_eth_event_enqueue_buffer *buf,
+			struct rte_mbuf **mbufs, uint16_t num)
+{
+	struct rte_event *ev = &buf->events[buf->count];
+	struct eth_rx_vector_data *vec;
+	uint16_t filled, space, sz;
+
+	filled = 0;
+	vec = &queue_info->vector_data;
+	while (num) {
+		if (vec->vector_ev == NULL) {
+			if (rte_mempool_get(vec->vector_pool,
+					    (void **)&vec->vector_ev) <

0) {

quoted

+				rte_pktmbuf_free_bulk(mbufs, num);
+				return 0;
+			}
+			vec->vector_ev->nb_elem = 0;
+			vec->vector_ev->port = vec->port;
+			vec->vector_ev->queue = vec->queue;
+			vec->vector_ev->attr_valid = true;
+			TAILQ_INSERT_TAIL(&rx_adapter->vector_list,

vec, next);

quoted

+		} else if (vec->vector_ev->nb_elem == vec-
max_vector_count) {

Is there a case where nb_elem > max_vector_count as we accumulate
sz to it ?

I don't think so, that would overflow the vector event.

quoted

+			/* Event ready. */
+			ev->event = vec->event;
+			ev->vec = vec->vector_ev;
+			ev++;
+			filled++;
+			vec->vector_ev = NULL;
+			TAILQ_REMOVE(&rx_adapter->vector_list, vec,

next);

quoted

+			if (rte_mempool_get(vec->vector_pool,
+					    (void **)&vec->vector_ev) <

0) {

quoted

+				rte_pktmbuf_free_bulk(mbufs, num);
+				return 0;
+			}
+			vec->vector_ev->nb_elem = 0;
+			vec->vector_ev->port = vec->port;
+			vec->vector_ev->queue = vec->queue;
+			vec->vector_ev->attr_valid = true;
+			TAILQ_INSERT_TAIL(&rx_adapter->vector_list,

vec, next);

quoted

+		}
+
+		space = vec->max_vector_count - vec->vector_ev-
nb_elem;
+		sz = num > space ? space : num;
+		memcpy(vec->vector_ev->mbufs + vec->vector_ev-
nb_elem, mbufs,
+		       sizeof(void *) * sz);
+		vec->vector_ev->nb_elem += sz;
+		num -= sz;
+		mbufs += sz;
+		vec->ts = rte_rdtsc();
+	}
+
+	if (vec->vector_ev->nb_elem == vec->max_vector_count) {

Same here.

quoted

+		ev->event = vec->event;
+		ev->vec = vec->vector_ev;
+		ev++;
+		filled++;
+		vec->vector_ev = NULL;
+		TAILQ_REMOVE(&rx_adapter->vector_list, vec, next);
+	}
+
+	return filled;
+}

I am seeing more than one repeating code chunks in this function.
Perhaps, you can give it a try to not repeat. We can drop if its
performance affecting.

I will try to move them to inline functions and test.

quoted

+
 static inline void
 rxa_buffer_mbufs(struct rte_event_eth_rx_adapter *rx_adapter,
 		uint16_t eth_dev_id,

@@ -770,25 +867,30 @@ rxa_buffer_mbufs(struct

rte_event_eth_rx_adapter *rx_adapter,

quoted

 	rss_mask = ~(((m->ol_flags & PKT_RX_RSS_HASH) != 0) - 1);
 	do_rss = !rss_mask && !eth_rx_queue_info->flow_id_mask;

The RSS related code is executed for vector case as well. Can this be
moved inside ena_vector if condition ?

RSS is used to generate the event flowid, in vector case the flow id 
Will be a combination of port and queue id. 
The idea is that flows having the same RSS LSB will end up in the same 
queue.

quoted

-	for (i = 0; i < num; i++) {
-		m = mbufs[i];
-
-		rss = do_rss ?
-			rxa_do_softrss(m, rx_adapter->rss_key_be) :
-			m->hash.rss;
-		ev->event = event;
-		ev->flow_id = (rss & ~flow_id_mask) |
-				(ev->flow_id & flow_id_mask);
-		ev->mbuf = m;
-		ev++;
+	if (!eth_rx_queue_info->ena_vector) {
+		for (i = 0; i < num; i++) {
+			m = mbufs[i];
+
+			rss = do_rss ? rxa_do_softrss(m, rx_adapter-
rss_key_be)
+				     : m->hash.rss;
+			ev->event = event;
+			ev->flow_id = (rss & ~flow_id_mask) |
+				      (ev->flow_id & flow_id_mask);
+			ev->mbuf = m;
+			ev++;
+		}
+	} else {
+		num = rxa_create_event_vector(rx_adapter,

eth_rx_queue_info,

quoted

+					      buf, mbufs, num);
 	}

-	if (dev_info->cb_fn) {
+	if (num && dev_info->cb_fn) {

 		dropped = 0;
 		nb_cb = dev_info->cb_fn(eth_dev_id, rx_queue_id,
-					ETH_EVENT_BUFFER_SIZE, buf-
count, ev,
-					num, dev_info->cb_arg,

&dropped);

quoted

+					ETH_EVENT_BUFFER_SIZE, buf-
count,
+					&buf->events[buf->count],

num,

quoted

+					dev_info->cb_arg, &dropped);

Before this patch, we pass ev which is &buf->events[buf->count] + num
as fifth param when calling cb_fn. Now, we are passing &buf-

quoted

events[buf->count] for non-vector case. Do you see this as an issue?

The callback function takes in the array newly formed events i.e. we need
to pass the start of array and the count.

the previous code had a bug where it passes the end of the event list.

Also, for vector case would it make sense to do pass &buf->events[buf-

quoted

count] + num ?

quoted

 		if (unlikely(nb_cb > num))
 			RTE_EDEV_LOG_ERR("Rx CB returned %d (> %d)

events",

quoted

 				nb_cb, num);

@@ -1124,6 +1226,30 @@ rxa_poll(struct rte_event_eth_rx_adapter

*rx_adapter)

quoted

 	return nb_rx;
 }

+static void
+rxa_vector_expire(struct eth_rx_vector_data *vec, void *arg)
+{
+	struct rte_event_eth_rx_adapter *rx_adapter = arg;
+	struct rte_eth_event_enqueue_buffer *buf =
+		&rx_adapter->event_enqueue_buffer;
+	struct rte_event *ev;
+
+	if (buf->count)
+		rxa_flush_event_buffer(rx_adapter);
+
+	if (vec->vector_ev->nb_elem == 0)
+		return;
+	ev = &buf->events[buf->count];
+
+	/* Event ready. */
+	ev->event = vec->event;
+	ev->vec = vec->vector_ev;
+	buf->count++;
+
+	vec->vector_ev = NULL;
+	vec->ts = 0;
+}
+
 static int
 rxa_service_func(void *args)
 {

@@ -1137,6 +1263,24 @@ rxa_service_func(void *args)
 		return 0;
 	}

+	if (rx_adapter->ena_vector) {
+		if ((rte_rdtsc() - rx_adapter->prev_expiry_ts) >=
+		    rx_adapter->vector_tmo_ticks) {
+			struct eth_rx_vector_data *vec;
+
+			TAILQ_FOREACH(vec, &rx_adapter->vector_list,

next) {

quoted

+				uint64_t elapsed_time = rte_rdtsc() -

vec->ts;

quoted

+
+				if (elapsed_time >= vec-
vector_timeout_ticks) {
+					rxa_vector_expire(vec,

rx_adapter);

quoted

+					TAILQ_REMOVE(&rx_adapter-
vector_list,
+						     vec, next);
+				}
+			}
+			rx_adapter->prev_expiry_ts = rte_rdtsc();
+		}
+	}
+
 	stats = &rx_adapter->stats;
 	stats->rx_packets += rxa_intr_ring_dequeue(rx_adapter);
 	stats->rx_packets += rxa_poll(rx_adapter);

@@ -1640,6 +1784,28 @@ rxa_update_queue(struct

rte_event_eth_rx_adapter *rx_adapter,

quoted

 	}
 }

+static void
+rxa_set_vector_data(struct eth_rx_queue_info *queue_info,

uint16_t vector_count,

quoted

+		    uint64_t vector_ns, struct rte_mempool *mp, int32_t

qid,

quoted

+		    uint16_t port_id)
+{
+#define NSEC2TICK(__ns, __freq) (((__ns) * (__freq)) / 1E9)
+	struct eth_rx_vector_data *vector_data;
+	uint32_t flow_id;
+
+	vector_data = &queue_info->vector_data;
+	vector_data->max_vector_count = vector_count;
+	vector_data->port = port_id;
+	vector_data->queue = qid;
+	vector_data->vector_pool = mp;
+	vector_data->vector_timeout_ticks =
+		NSEC2TICK(vector_ns, rte_get_timer_hz());
+	vector_data->ts = 0;
+	flow_id = queue_info->event & 0xFFFFF;
+	flow_id = flow_id == 0 ? (qid & 0xFF) | (port_id & 0xFFFF) :

flow_id;

Maybe I am missing something here. Looking at the code it looks like
qid and port_id may overlap. For e.g., if qid = 0x10 and port_id = 0x11,
flow_id would end up being 0x11. Is this the expectation? Also, it may
be useful to document flow_id format.

The flow_id is 20 bit, I guess we could do 12bit queue_id and 8bit port
as a flow.

Comparing this format with existing RSS hash based method, are we
saying that all mbufs received in a rx burst are part of same flow when
vectorization is used?

Yes, the hard way to do this is to use a hash table and treating each 
mbuf having an unique flow.

quoted

+	vector_data->event = (queue_info->event & ~0xFFFFF) |

flow_id;

quoted

+}
+
 static void
 rxa_sw_del(struct rte_event_eth_rx_adapter *rx_adapter,
 	struct eth_device_info *dev_info,

@@ -1741,6 +1907,44 @@ rxa_add_queue(struct

rte_event_eth_rx_adapter *rx_adapter,

quoted

 	}
 }

+static void
+rxa_sw_event_vector_configure(
+	struct rte_event_eth_rx_adapter *rx_adapter, uint16_t

eth_dev_id,

quoted

+	int rx_queue_id,
+	const struct rte_event_eth_rx_adapter_event_vector_config

*config)

quoted

+{
+	struct eth_device_info *dev_info = &rx_adapter-
eth_devices[eth_dev_id];
+	struct eth_rx_queue_info *queue_info;
+	struct rte_event *qi_ev;
+
+	if (rx_queue_id == -1) {
+		uint16_t nb_rx_queues;
+		uint16_t i;
+
+		nb_rx_queues = dev_info->dev->data->nb_rx_queues;
+		for (i = 0; i < nb_rx_queues; i++)
+			rxa_sw_event_vector_configure(rx_adapter,

eth_dev_id, i,

quoted

+						      config);
+		return;
+	}
+
+	queue_info = &dev_info->rx_queue[rx_queue_id];
+	qi_ev = (struct rte_event *)&queue_info->event;
+	queue_info->ena_vector = 1;
+	qi_ev->event_type =

RTE_EVENT_TYPE_ETH_RX_ADAPTER_VECTOR;

quoted

+	rxa_set_vector_data(queue_info, config->vector_sz,
+			    config->vector_timeout_ns, config-
vector_mp,
+			    rx_queue_id, dev_info->dev->data->port_id);
+	rx_adapter->ena_vector = 1;
+	rx_adapter->vector_tmo_ticks =
+		rx_adapter->vector_tmo_ticks ?
+			      RTE_MIN(config->vector_timeout_ns << 1,
+				      rx_adapter->vector_tmo_ticks) :
+			      config->vector_timeout_ns << 1;
+	rx_adapter->prev_expiry_ts = 0;
+	TAILQ_INIT(&rx_adapter->vector_list);
+}
+
 static int rxa_sw_add(struct rte_event_eth_rx_adapter *rx_adapter,
 		uint16_t eth_dev_id,
 		int rx_queue_id,

@@ -2081,6 +2285,15 @@

rte_event_eth_rx_adapter_queue_add(uint8_t id,

quoted

 		return -EINVAL;
 	}

+	if ((cap &

RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR) == 0 &&

quoted

+	    (queue_conf->rx_queue_flags &
+	     RTE_EVENT_ETH_RX_ADAPTER_QUEUE_EVENT_VECTOR)) {
+		RTE_EDEV_LOG_ERR("Event vectorization is not

supported,"

quoted

+				 " eth port: %" PRIu16 " adapter id: %"

PRIu8,

quoted

+				 eth_dev_id, id);
+		return -EINVAL;
+	}
+
 	if ((cap &

RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ) == 0 &&

quoted

 		(rx_queue_id != -1)) {
 		RTE_EDEV_LOG_ERR("Rx queues can only be connected

to single "

quoted

@@ -2143,6 +2356,17 @@

rte_event_eth_rx_adapter_queue_add(uint8_t id,

quoted

 	return 0;
 }

+static int
+rxa_sw_vector_limits(struct

rte_event_eth_rx_adapter_vector_limits *limits)

quoted

+{
+	limits->max_sz = MAX_VECTOR_SIZE;
+	limits->min_sz = MIN_VECTOR_SIZE;
+	limits->max_timeout_ns = MAX_VECTOR_NS;
+	limits->min_timeout_ns = MIN_VECTOR_NS;
+
+	return 0;
+}
+
 int
 rte_event_eth_rx_adapter_queue_del(uint8_t id, uint16_t

eth_dev_id,

quoted

 				int32_t rx_queue_id)

@@ -2333,7 +2557,8 @@

rte_event_eth_rx_adapter_queue_event_vector_config(

quoted

 		ret = dev->dev_ops-
eth_rx_adapter_event_vector_config(
 			dev, &rte_eth_devices[eth_dev_id],

rx_queue_id, config);

quoted

 	} else {
-		ret = -ENOTSUP;
+		rxa_sw_event_vector_configure(rx_adapter,

eth_dev_id,

quoted

+					      rx_queue_id, config);
 	}

 	return ret;

@@ -2371,7 +2596,7 @@

rte_event_eth_rx_adapter_vector_limits_get(

quoted

 		ret = dev->dev_ops-
eth_rx_adapter_vector_limits_get(
 			dev, &rte_eth_devices[eth_port_id], limits);
 	} else {
-		ret = -ENOTSUP;
+		ret = rxa_sw_vector_limits(limits);
 	}

 	return ret;

diff --git a/lib/librte_eventdev/rte_eventdev.c

b/lib/librte_eventdev/rte_eventdev.c

quoted

index f95edc075..254a31b1f 100644

--- a/lib/librte_eventdev/rte_eventdev.c
+++ b/lib/librte_eventdev/rte_eventdev.c

@@ -122,7 +122,11 @@ rte_event_eth_rx_adapter_caps_get(uint8_t

dev_id, uint16_t eth_port_id,

quoted

 	if (caps == NULL)
 		return -EINVAL;
-	*caps = 0;
+
+	if (dev->dev_ops->eth_rx_adapter_caps_get == NULL)
+		*caps = RTE_EVENT_ETH_RX_ADAPTER_SW_CAP;
+	else
+		*caps = 0;

Any reason why we had to set default caps value? I am thinking if sw
event device is used, it would set it anyways.

There are multiple sw event devices which don't implement caps_get 
function, this changes solves that.

quoted

 	return dev->dev_ops->eth_rx_adapter_caps_get ?
 				(*dev->dev_ops-
eth_rx_adapter_caps_get)(dev,
--
2.17.1