eventdev: support Rx adapter event vector
authorPavan Nikhilesh <pbhagavatula@marvell.com>
Wed, 31 Mar 2021 09:29:58 +0000 (14:59 +0530)
committerJerin Jacob <jerinj@marvell.com>
Mon, 12 Apr 2021 07:23:34 +0000 (09:23 +0200)
Add event vector support for event eth Rx adapter, the implementation
creates vector flows based on port and queue identifier of the received
mbufs.
The flow id for SW Rx event vectorization will use 12-bits of queue
identifier and 8-bits port identifier when custom flow id is not set
for simplicity.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Acked-by: Jay Jayatheerthan <jay.jayatheerthan@intel.com>
doc/guides/prog_guide/event_ethernet_rx_adapter.rst
lib/librte_eventdev/eventdev_pmd.h
lib/librte_eventdev/rte_event_eth_rx_adapter.c
lib/librte_eventdev/rte_eventdev.c

index 5eefef3..c01e5a9 100644 (file)
@@ -224,3 +224,22 @@ A loop processing ``rte_event_vector`` containing mbufs is shown below.
         case ...
         ...
         }
+
+Rx event vectorization for SW Rx adapter
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+For SW based event vectorization, i.e., when the
+``RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT`` is not set in the adapter's
+capabilities flags for a particular ethernet device, the service function
+creates a single event vector flow for all the mbufs arriving on the given
+Rx queue.
+The 20-bit event flow identifier is set to 12-bits of Rx queue identifier
+and 8-bits of ethernet device identifier.
+Flow identifier is formatted as follows:
+
+.. code-block:: console
+
+    19      12,11            0
+    +---------+--------------+
+    | port_id |   queue_id   |
+    +---------+--------------+
index 9297f14..0f724ac 100644 (file)
@@ -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) | \
-                       (RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ))
+#define RTE_EVENT_ETH_RX_ADAPTER_SW_CAP                                        \
+       ((RTE_EVENT_ETH_RX_ADAPTER_CAP_OVERRIDE_FLOW_ID) |                     \
+        (RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ) |                         \
+        (RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR))
 
 #define RTE_EVENT_CRYPTO_ADAPTER_SW_CAP \
                RTE_EVENT_CRYPTO_ADAPTER_CAP_SESSION_PRIVATE_DATA
index ac8ba5b..e273b3a 100644 (file)
 #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 */
        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)
            &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 +773,77 @@ rxa_flush_event_buffer(struct rte_event_eth_rx_adapter *rx_adapter)
        return n;
 }
 
+static inline void
+rxa_init_vector(struct rte_event_eth_rx_adapter *rx_adapter,
+               struct eth_rx_vector_data *vec)
+{
+       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);
+}
+
+static inline uint16_t
+rxa_create_event_vector(struct rte_event_eth_rx_adapter *rx_adapter,
+                       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;
+
+       if (vec->vector_ev == NULL) {
+               if (rte_mempool_get(vec->vector_pool,
+                                   (void **)&vec->vector_ev) < 0) {
+                       rte_pktmbuf_free_bulk(mbufs, num);
+                       return 0;
+               }
+               rxa_init_vector(rx_adapter, vec);
+       }
+       while (num) {
+               if (vec->vector_ev->nb_elem == vec->max_vector_count) {
+                       /* 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);
+                       if (rte_mempool_get(vec->vector_pool,
+                                           (void **)&vec->vector_ev) < 0) {
+                               rte_pktmbuf_free_bulk(mbufs, num);
+                               return 0;
+                       }
+                       rxa_init_vector(rx_adapter, vec);
+               }
+
+               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) {
+               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;
+}
+
 static inline void
 rxa_buffer_mbufs(struct rte_event_eth_rx_adapter *rx_adapter,
                uint16_t eth_dev_id,
@@ -766,29 +868,33 @@ rxa_buffer_mbufs(struct rte_event_eth_rx_adapter *rx_adapter,
        uint16_t nb_cb;
        uint16_t dropped;
 
-       /* 0xffff ffff if PKT_RX_RSS_HASH is set, otherwise 0 */
-       rss_mask = ~(((m->ol_flags & PKT_RX_RSS_HASH) != 0) - 1);
-       do_rss = !rss_mask && !eth_rx_queue_info->flow_id_mask;
-
-       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) {
+               /* 0xffff ffff if PKT_RX_RSS_HASH is set, otherwise 0 */
+               rss_mask = ~(((m->ol_flags & PKT_RX_RSS_HASH) != 0) - 1);
+               do_rss = !rss_mask && !eth_rx_queue_info->flow_id_mask;
+               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,
+                                             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);
+                                       ETH_EVENT_BUFFER_SIZE, buf->count,
+                                       &buf->events[buf->count], num,
+                                       dev_info->cb_arg, &dropped);
                if (unlikely(nb_cb > num))
                        RTE_EDEV_LOG_ERR("Rx CB returned %d (> %d) events",
                                nb_cb, num);
@@ -1124,6 +1230,30 @@ rxa_poll(struct rte_event_eth_rx_adapter *rx_adapter)
        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 +1267,24 @@ 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) {
+                               uint64_t elapsed_time = rte_rdtsc() - vec->ts;
+
+                               if (elapsed_time >= vec->vector_timeout_ticks) {
+                                       rxa_vector_expire(vec, rx_adapter);
+                                       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);
@@ -1640,11 +1788,35 @@ rxa_update_queue(struct rte_event_eth_rx_adapter *rx_adapter,
        }
 }
 
+static void
+rxa_set_vector_data(struct eth_rx_queue_info *queue_info, uint16_t vector_count,
+                   uint64_t vector_ns, struct rte_mempool *mp, int32_t qid,
+                   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 & 0xFFF) | (port_id & 0xFF) << 12 : flow_id;
+       vector_data->event = (queue_info->event & ~0xFFFFF) | flow_id;
+}
+
 static void
 rxa_sw_del(struct rte_event_eth_rx_adapter *rx_adapter,
        struct eth_device_info *dev_info,
        int32_t rx_queue_id)
 {
+       struct eth_rx_vector_data *vec;
        int pollq;
        int intrq;
        int sintrq;
@@ -1663,6 +1835,14 @@ rxa_sw_del(struct rte_event_eth_rx_adapter *rx_adapter,
                return;
        }
 
+       /* Push all the partial event vectors to event device. */
+       TAILQ_FOREACH(vec, &rx_adapter->vector_list, next) {
+               if (vec->queue != rx_queue_id)
+                       continue;
+               rxa_vector_expire(vec, rx_adapter);
+               TAILQ_REMOVE(&rx_adapter->vector_list, vec, next);
+       }
+
        pollq = rxa_polled_queue(dev_info, rx_queue_id);
        intrq = rxa_intr_queue(dev_info, rx_queue_id);
        sintrq = rxa_shared_intr(dev_info, rx_queue_id);
@@ -1741,6 +1921,42 @@ rxa_add_queue(struct rte_event_eth_rx_adapter *rx_adapter,
        }
 }
 
+static void
+rxa_sw_event_vector_configure(
+       struct rte_event_eth_rx_adapter *rx_adapter, uint16_t eth_dev_id,
+       int rx_queue_id,
+       const struct rte_event_eth_rx_adapter_event_vector_config *config)
+{
+       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,
+                                                     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;
+       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;
+}
+
 static int rxa_sw_add(struct rte_event_eth_rx_adapter *rx_adapter,
                uint16_t eth_dev_id,
                int rx_queue_id,
@@ -1967,6 +2183,7 @@ rte_event_eth_rx_adapter_create_ext(uint8_t id, uint8_t dev_id,
        rx_adapter->conf_cb = conf_cb;
        rx_adapter->conf_arg = conf_arg;
        rx_adapter->id = id;
+       TAILQ_INIT(&rx_adapter->vector_list);
        strcpy(rx_adapter->mem_name, mem_name);
        rx_adapter->eth_devices = rte_zmalloc_socket(rx_adapter->mem_name,
                                        RTE_MAX_ETHPORTS *
@@ -2081,6 +2298,15 @@ rte_event_eth_rx_adapter_queue_add(uint8_t id,
                return -EINVAL;
        }
 
+       if ((cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR) == 0 &&
+           (queue_conf->rx_queue_flags &
+            RTE_EVENT_ETH_RX_ADAPTER_QUEUE_EVENT_VECTOR)) {
+               RTE_EDEV_LOG_ERR("Event vectorization is not supported,"
+                                " eth port: %" PRIu16 " adapter id: %" PRIu8,
+                                eth_dev_id, id);
+               return -EINVAL;
+       }
+
        if ((cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ) == 0 &&
                (rx_queue_id != -1)) {
                RTE_EDEV_LOG_ERR("Rx queues can only be connected to single "
@@ -2143,6 +2369,17 @@ rte_event_eth_rx_adapter_queue_add(uint8_t id,
        return 0;
 }
 
+static int
+rxa_sw_vector_limits(struct rte_event_eth_rx_adapter_vector_limits *limits)
+{
+       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,
                                int32_t rx_queue_id)
@@ -2333,7 +2570,8 @@ rte_event_eth_rx_adapter_queue_event_vector_config(
                ret = dev->dev_ops->eth_rx_adapter_event_vector_config(
                        dev, &rte_eth_devices[eth_dev_id], rx_queue_id, config);
        } else {
-               ret = -ENOTSUP;
+               rxa_sw_event_vector_configure(rx_adapter, eth_dev_id,
+                                             rx_queue_id, config);
        }
 
        return ret;
@@ -2371,7 +2609,7 @@ rte_event_eth_rx_adapter_vector_limits_get(
                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;
index be0499c..6282465 100644 (file)
@@ -122,7 +122,11 @@ rte_event_eth_rx_adapter_caps_get(uint8_t dev_id, uint16_t eth_port_id,
 
        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;
 
        return dev->dev_ops->eth_rx_adapter_caps_get ?
                                (*dev->dev_ops->eth_rx_adapter_caps_get)(dev,