[dpdk-dev] [PATCH v7 4/5] eventdev/rx_adapter: implement per queue event buffer
Jayatheerthan, Jay
jay.jayatheerthan at intel.com
Wed Oct 6 08:42:52 CEST 2021
> -----Original Message-----
> From: Naga Harish K, S V <s.v.naga.harish.k at intel.com>
> Sent: Wednesday, October 6, 2021 9:33 AM
> To: jerinj at marvell.com; Jayatheerthan, Jay <jay.jayatheerthan at intel.com>
> Cc: dev at dpdk.org
> Subject: [PATCH v7 4/5] eventdev/rx_adapter: implement per queue event buffer
>
> this patch implement the per queue event buffer with
> required validations.
>
> Signed-off-by: Naga Harish K S V <s.v.naga.harish.k at intel.com>
> ---
> lib/eventdev/rte_event_eth_rx_adapter.c | 206 ++++++++++++++++++------
> 1 file changed, 153 insertions(+), 53 deletions(-)
>
> diff --git a/lib/eventdev/rte_event_eth_rx_adapter.c b/lib/eventdev/rte_event_eth_rx_adapter.c
> index 5ccea168ea..1a2aa23475 100644
> --- a/lib/eventdev/rte_event_eth_rx_adapter.c
> +++ b/lib/eventdev/rte_event_eth_rx_adapter.c
> @@ -102,10 +102,12 @@ struct rte_event_eth_rx_adapter {
> uint8_t rss_key_be[RSS_KEY_SIZE];
> /* Event device identifier */
> uint8_t eventdev_id;
> - /* Per ethernet device structure */
> - struct eth_device_info *eth_devices;
> /* Event port identifier */
> uint8_t event_port_id;
> + /* Flag indicating per rxq event buffer */
> + bool use_queue_event_buf;
> + /* Per ethernet device structure */
> + struct eth_device_info *eth_devices;
> /* Lock to serialize config updates with service function */
> rte_spinlock_t rx_lock;
> /* Max mbufs processed in any service function invocation */
> @@ -241,6 +243,7 @@ struct eth_rx_queue_info {
> 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;
> + struct rte_eth_event_enqueue_buffer *event_buf;
> };
>
> static struct rte_event_eth_rx_adapter **event_eth_rx_adapter;
> @@ -262,6 +265,22 @@ rxa_validate_id(uint8_t id)
> return id < RTE_EVENT_ETH_RX_ADAPTER_MAX_INSTANCE;
> }
>
> +static inline struct rte_eth_event_enqueue_buffer *
> +rxa_event_buf_get(struct rte_event_eth_rx_adapter *rx_adapter,
> + uint16_t eth_dev_id, uint16_t rx_queue_id)
> +{
> + struct rte_eth_event_enqueue_buffer *buf = NULL;
> +
> + if (rx_adapter->use_queue_event_buf) {
> + struct eth_device_info *dev_info =
> + &rx_adapter->eth_devices[eth_dev_id];
> + buf = dev_info->rx_queue[rx_queue_id].event_buf;
We can return here. It may save an instr or two.
> + } else
> + buf = &rx_adapter->event_enqueue_buffer;
Same here.
> +
> + return buf;
> +}
> +
> #define RTE_EVENT_ETH_RX_ADAPTER_ID_VALID_OR_ERR_RET(id, retval) do { \
> if (!rxa_validate_id(id)) { \
> RTE_EDEV_LOG_ERR("Invalid eth Rx adapter id = %d\n", id); \
> @@ -767,10 +786,9 @@ rxa_enq_block_end_ts(struct rte_event_eth_rx_adapter *rx_adapter,
>
> /* Enqueue buffered events to event device */
> static inline uint16_t
> -rxa_flush_event_buffer(struct rte_event_eth_rx_adapter *rx_adapter)
> +rxa_flush_event_buffer(struct rte_event_eth_rx_adapter *rx_adapter,
> + struct rte_eth_event_enqueue_buffer *buf)
> {
> - struct rte_eth_event_enqueue_buffer *buf =
> - &rx_adapter->event_enqueue_buffer;
> struct rte_event_eth_rx_adapter_stats *stats = &rx_adapter->stats;
> uint16_t count = buf->last ? buf->last - buf->head : buf->count;
>
> @@ -888,15 +906,14 @@ rxa_buffer_mbufs(struct rte_event_eth_rx_adapter *rx_adapter,
> uint16_t eth_dev_id,
> uint16_t rx_queue_id,
> struct rte_mbuf **mbufs,
> - uint16_t num)
> + uint16_t num,
> + struct rte_eth_event_enqueue_buffer *buf)
> {
> uint32_t i;
> struct eth_device_info *dev_info =
> &rx_adapter->eth_devices[eth_dev_id];
> struct eth_rx_queue_info *eth_rx_queue_info =
> &dev_info->rx_queue[rx_queue_id];
> - struct rte_eth_event_enqueue_buffer *buf =
> - &rx_adapter->event_enqueue_buffer;
> uint16_t new_tail = buf->tail;
> uint64_t event = eth_rx_queue_info->event;
> uint32_t flow_id_mask = eth_rx_queue_info->flow_id_mask;
> @@ -995,11 +1012,10 @@ rxa_eth_rx(struct rte_event_eth_rx_adapter *rx_adapter,
> uint16_t queue_id,
> uint32_t rx_count,
> uint32_t max_rx,
> - int *rxq_empty)
> + int *rxq_empty,
> + struct rte_eth_event_enqueue_buffer *buf)
> {
> struct rte_mbuf *mbufs[BATCH_SIZE];
> - struct rte_eth_event_enqueue_buffer *buf =
> - &rx_adapter->event_enqueue_buffer;
> struct rte_event_eth_rx_adapter_stats *stats =
> &rx_adapter->stats;
> uint16_t n;
> @@ -1012,7 +1028,7 @@ rxa_eth_rx(struct rte_event_eth_rx_adapter *rx_adapter,
> */
> while (rxa_pkt_buf_available(buf)) {
> if (buf->count >= BATCH_SIZE)
> - rxa_flush_event_buffer(rx_adapter);
> + rxa_flush_event_buffer(rx_adapter, buf);
>
> stats->rx_poll_count++;
> n = rte_eth_rx_burst(port_id, queue_id, mbufs, BATCH_SIZE);
> @@ -1021,14 +1037,14 @@ rxa_eth_rx(struct rte_event_eth_rx_adapter *rx_adapter,
> *rxq_empty = 1;
> break;
> }
> - rxa_buffer_mbufs(rx_adapter, port_id, queue_id, mbufs, n);
> + rxa_buffer_mbufs(rx_adapter, port_id, queue_id, mbufs, n, buf);
> nb_rx += n;
> if (rx_count + nb_rx > max_rx)
> break;
> }
>
> if (buf->count > 0)
> - rxa_flush_event_buffer(rx_adapter);
> + rxa_flush_event_buffer(rx_adapter, buf);
>
> return nb_rx;
> }
> @@ -1169,7 +1185,7 @@ rxa_intr_ring_dequeue(struct rte_event_eth_rx_adapter *rx_adapter)
> ring_lock = &rx_adapter->intr_ring_lock;
>
> if (buf->count >= BATCH_SIZE)
> - rxa_flush_event_buffer(rx_adapter);
> + rxa_flush_event_buffer(rx_adapter, buf);
>
> while (rxa_pkt_buf_available(buf)) {
> struct eth_device_info *dev_info;
> @@ -1221,7 +1237,7 @@ rxa_intr_ring_dequeue(struct rte_event_eth_rx_adapter *rx_adapter)
> continue;
> n = rxa_eth_rx(rx_adapter, port, i, nb_rx,
> rx_adapter->max_nb_rx,
> - &rxq_empty);
> + &rxq_empty, buf);
> nb_rx += n;
>
> enq_buffer_full = !rxq_empty && n == 0;
> @@ -1242,7 +1258,7 @@ rxa_intr_ring_dequeue(struct rte_event_eth_rx_adapter *rx_adapter)
> } else {
> n = rxa_eth_rx(rx_adapter, port, queue, nb_rx,
> rx_adapter->max_nb_rx,
> - &rxq_empty);
> + &rxq_empty, buf);
> rx_adapter->qd_valid = !rxq_empty;
> nb_rx += n;
> if (nb_rx > rx_adapter->max_nb_rx)
> @@ -1273,13 +1289,12 @@ rxa_poll(struct rte_event_eth_rx_adapter *rx_adapter)
> {
> uint32_t num_queue;
> uint32_t nb_rx = 0;
> - struct rte_eth_event_enqueue_buffer *buf;
> + struct rte_eth_event_enqueue_buffer *buf = NULL;
> uint32_t wrr_pos;
> uint32_t max_nb_rx;
>
> wrr_pos = rx_adapter->wrr_pos;
> max_nb_rx = rx_adapter->max_nb_rx;
> - buf = &rx_adapter->event_enqueue_buffer;
>
> /* Iterate through a WRR sequence */
> for (num_queue = 0; num_queue < rx_adapter->wrr_len; num_queue++) {
> @@ -1287,24 +1302,31 @@ rxa_poll(struct rte_event_eth_rx_adapter *rx_adapter)
> uint16_t qid = rx_adapter->eth_rx_poll[poll_idx].eth_rx_qid;
> uint16_t d = rx_adapter->eth_rx_poll[poll_idx].eth_dev_id;
>
> + buf = rxa_event_buf_get(rx_adapter, d, qid);
> +
> /* Don't do a batch dequeue from the rx queue if there isn't
> * enough space in the enqueue buffer.
> */
> if (buf->count >= BATCH_SIZE)
> - rxa_flush_event_buffer(rx_adapter);
> + rxa_flush_event_buffer(rx_adapter, buf);
> if (!rxa_pkt_buf_available(buf)) {
> - rx_adapter->wrr_pos = wrr_pos;
> - return nb_rx;
> + if (rx_adapter->use_queue_event_buf)
> + goto poll_next_entry;
> + else {
> + rx_adapter->wrr_pos = wrr_pos;
> + return nb_rx;
> + }
> }
>
> nb_rx += rxa_eth_rx(rx_adapter, d, qid, nb_rx, max_nb_rx,
> - NULL);
> + NULL, buf);
> if (nb_rx > max_nb_rx) {
> rx_adapter->wrr_pos =
> (wrr_pos + 1) % rx_adapter->wrr_len;
> break;
> }
>
> +poll_next_entry:
> if (++wrr_pos == rx_adapter->wrr_len)
> wrr_pos = 0;
> }
> @@ -1315,12 +1337,13 @@ 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_eth_event_enqueue_buffer *buf = NULL;
> struct rte_event *ev;
>
> + buf = rxa_event_buf_get(rx_adapter, vec->port, vec->queue);
> +
> if (buf->count)
> - rxa_flush_event_buffer(rx_adapter);
> + rxa_flush_event_buffer(rx_adapter, buf);
>
> if (vec->vector_ev->nb_elem == 0)
> return;
> @@ -1947,9 +1970,16 @@ rxa_sw_del(struct rte_event_eth_rx_adapter *rx_adapter,
> rx_adapter->num_rx_intr -= intrq;
> dev_info->nb_rx_intr -= intrq;
> dev_info->nb_shared_intr -= intrq && sintrq;
> + if (rx_adapter->use_queue_event_buf) {
> + struct rte_eth_event_enqueue_buffer *event_buf =
> + dev_info->rx_queue[rx_queue_id].event_buf;
> + rte_free(event_buf->events);
> + rte_free(event_buf);
> + dev_info->rx_queue[rx_queue_id].event_buf = NULL;
> + }
> }
>
> -static void
> +static int
> rxa_add_queue(struct rte_event_eth_rx_adapter *rx_adapter,
> struct eth_device_info *dev_info,
> int32_t rx_queue_id,
> @@ -1961,15 +1991,21 @@ rxa_add_queue(struct rte_event_eth_rx_adapter *rx_adapter,
> int intrq;
> int sintrq;
> struct rte_event *qi_ev;
> + struct rte_eth_event_enqueue_buffer *new_rx_buf = NULL;
> + uint16_t eth_dev_id = dev_info->dev->data->port_id;
> + int ret;
>
> 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_add_queue(rx_adapter, dev_info, i, conf);
> - return;
> + for (i = 0; i < nb_rx_queues; i++) {
> + ret = rxa_add_queue(rx_adapter, dev_info, i, conf);
> + if (ret)
> + return ret;
> + }
> + return 0;
> }
>
> pollq = rxa_polled_queue(dev_info, rx_queue_id);
> @@ -2032,6 +2068,37 @@ rxa_add_queue(struct rte_event_eth_rx_adapter *rx_adapter,
> dev_info->next_q_idx = 0;
> }
> }
> +
> + if (!rx_adapter->use_queue_event_buf)
> + return 0;
> +
> + new_rx_buf = rte_zmalloc_socket("rx_buffer_meta",
> + sizeof(*new_rx_buf), 0,
> + rte_eth_dev_socket_id(eth_dev_id));
> + if (new_rx_buf == NULL) {
> + RTE_EDEV_LOG_ERR("Failed to allocate event buffer meta for "
> + "dev_id: %d queue_id: %d",
> + eth_dev_id, rx_queue_id);
> + return -ENOMEM;
> + }
> +
> + new_rx_buf->events_size = RTE_ALIGN(conf->event_buf_size, BATCH_SIZE);
> + new_rx_buf->events_size += (2 * BATCH_SIZE);
> + new_rx_buf->events = rte_zmalloc_socket("rx_buffer",
> + sizeof(struct rte_event) *
> + new_rx_buf->events_size, 0,
> + rte_eth_dev_socket_id(eth_dev_id));
> + if (new_rx_buf->events == NULL) {
> + rte_free(new_rx_buf);
> + RTE_EDEV_LOG_ERR("Failed to allocate event buffer for "
> + "dev_id: %d queue_id: %d",
> + eth_dev_id, rx_queue_id);
> + return -ENOMEM;
> + }
> +
> + queue_info->event_buf = new_rx_buf;
> +
> + return 0;
> }
>
> static int rxa_sw_add(struct rte_event_eth_rx_adapter *rx_adapter,
> @@ -2060,6 +2127,16 @@ static int rxa_sw_add(struct rte_event_eth_rx_adapter *rx_adapter,
> temp_conf.servicing_weight = 1;
> }
> queue_conf = &temp_conf;
> +
> + if (queue_conf->servicing_weight == 0 &&
> + rx_adapter->use_queue_event_buf) {
> +
> + RTE_EDEV_LOG_ERR("Use of queue level event buffer "
> + "not supported for interrupt queues "
> + "dev_id: %d queue_id: %d",
> + eth_dev_id, rx_queue_id);
> + return -EINVAL;
> + }
> }
>
> nb_rx_queues = dev_info->dev->data->nb_rx_queues;
> @@ -2139,7 +2216,9 @@ static int rxa_sw_add(struct rte_event_eth_rx_adapter *rx_adapter,
>
>
>
> - rxa_add_queue(rx_adapter, dev_info, rx_queue_id, queue_conf);
> + ret = rxa_add_queue(rx_adapter, dev_info, rx_queue_id, queue_conf);
> + if (ret)
> + goto err_free_rxqueue;
> rxa_calc_wrr_sequence(rx_adapter, rx_poll, rx_wrr);
>
> rte_free(rx_adapter->eth_rx_poll);
> @@ -2160,7 +2239,7 @@ static int rxa_sw_add(struct rte_event_eth_rx_adapter *rx_adapter,
> rte_free(rx_poll);
> rte_free(rx_wrr);
>
> - return 0;
> + return ret;
> }
>
> static int
> @@ -2286,20 +2365,25 @@ rxa_create(uint8_t id, uint8_t dev_id,
> rx_adapter->eth_devices[i].dev = &rte_eth_devices[i];
>
> /* Rx adapter event buffer allocation */
> - buf = &rx_adapter->event_enqueue_buffer;
> - buf->events_size = rxa_params->event_buf_size;
> -
> - events = rte_zmalloc_socket(rx_adapter->mem_name,
> - buf->events_size * sizeof(*events),
> - 0, socket_id);
> - if (events == NULL) {
> - RTE_EDEV_LOG_ERR("Failed to allocate mem for event buffer\n");
> - rte_free(rx_adapter->eth_devices);
> - rte_free(rx_adapter);
> - return -ENOMEM;
> - }
> + rx_adapter->use_queue_event_buf = rxa_params->use_queue_event_buf;
> +
> + if (!rx_adapter->use_queue_event_buf) {
> + buf = &rx_adapter->event_enqueue_buffer;
> + buf->events_size = rxa_params->event_buf_size;
> +
> + events = rte_zmalloc_socket(rx_adapter->mem_name,
> + buf->events_size * sizeof(*events),
> + 0, socket_id);
> + if (events == NULL) {
> + RTE_EDEV_LOG_ERR("Failed to allocate memory "
> + "for adapter event buffer");
> + rte_free(rx_adapter->eth_devices);
> + rte_free(rx_adapter);
> + return -ENOMEM;
> + }
>
> - rx_adapter->event_enqueue_buffer.events = events;
> + rx_adapter->event_enqueue_buffer.events = events;
> + }
>
> event_eth_rx_adapter[id] = rx_adapter;
>
> @@ -2327,6 +2411,7 @@ rte_event_eth_rx_adapter_create_ext(uint8_t id, uint8_t dev_id,
>
> /* use default values for adapter params */
> rxa_params.event_buf_size = ETH_EVENT_BUFFER_SIZE;
> + rxa_params.use_queue_event_buf = false;
>
> return rxa_create(id, dev_id, &rxa_params, conf_cb, conf_arg);
> }
> @@ -2346,11 +2431,15 @@ rte_event_eth_rx_adapter_create_with_params(uint8_t id, uint8_t dev_id,
> /* use default values if rxa_params is NULL */
> if (rxa_params == NULL) {
> rxa_params = &temp_params;
> - rxa_params->event_buf_size = ETH_EVENT_BUFFER_SIZE;
> - }
> -
> - if (rxa_params->event_buf_size == 0)
> + rxa_params->event_buf_size = 4 * BATCH_SIZE;
This assumes ETH_EVENT_BUFFER_SIZE to be set to 6 * BATCH_SIZE so we can have 4x here and 2x later. It may break if ETH_EVENT_BUFFER_SIZE is changed later.
Can we change the code to just use ETH_EVENT_BUFFER_SIZE here. See below.
> + rxa_params->use_queue_event_buf = false;
> + } else if ((!rxa_params->use_queue_event_buf &&
> + rxa_params->event_buf_size == 0) ||
> + (rxa_params->use_queue_event_buf &&
> + rxa_params->event_buf_size != 0)) {
> + RTE_EDEV_LOG_ERR("Invalid adapter params\n");
> return -EINVAL;
> + }
>
> pc = rte_malloc(NULL, sizeof(*pc), 0);
> if (pc == NULL)
> @@ -2362,9 +2451,11 @@ rte_event_eth_rx_adapter_create_with_params(uint8_t id, uint8_t dev_id,
> * from NIC rx queues to get full buffer utilization and prevent
> * unnecessary rollovers.
> */
> - rxa_params->event_buf_size = RTE_ALIGN(rxa_params->event_buf_size,
> - BATCH_SIZE);
> - rxa_params->event_buf_size += (BATCH_SIZE + BATCH_SIZE);
> + if (!rxa_params->use_queue_event_buf) {
> + rxa_params->event_buf_size =
> + RTE_ALIGN(rxa_params->event_buf_size, BATCH_SIZE);
> + rxa_params->event_buf_size += (BATCH_SIZE + BATCH_SIZE);
> + }
Above if condition can be added as an else part of rxa_params == NULL. Something like:
if (rxa_params == NULL) {
rxa_params = &temp_params;
rxa_params->event_buf_size = ETH_EVENT_BUFFER_SIZE;
rxa_params->use_queue_event_buf = false;
} else if ((!rxa_params->use_queue_event_buf &&
rxa_params->event_buf_size == 0) ||
(rxa_params->use_queue_event_buf &&
rxa_params->event_buf_size != 0)) {
RTE_EDEV_LOG_ERR("Invalid adapter params\n");
return -EINVAL;
} else if (!rxa_params->use_queue_event_buf) {
rxa_params->event_buf_size =
RTE_ALIGN(rxa_params->event_buf_size, BATCH_SIZE);
rxa_params->event_buf_size += (BATCH_SIZE + BATCH_SIZE);
}
>
> ret = rxa_create(id, dev_id, rxa_params, rxa_default_conf_cb, pc);
> if (ret)
> @@ -2418,7 +2509,8 @@ rte_event_eth_rx_adapter_free(uint8_t id)
> if (rx_adapter->default_cb_arg)
> rte_free(rx_adapter->conf_arg);
> rte_free(rx_adapter->eth_devices);
> - rte_free(rx_adapter->event_enqueue_buffer.events);
> + if (!rx_adapter->use_queue_event_buf)
> + rte_free(rx_adapter->event_enqueue_buffer.events);
> rte_free(rx_adapter);
> event_eth_rx_adapter[id] = NULL;
>
> @@ -2522,6 +2614,14 @@ rte_event_eth_rx_adapter_queue_add(uint8_t id,
> return -EINVAL;
> }
>
> + if ((rx_adapter->use_queue_event_buf &&
> + queue_conf->event_buf_size == 0) ||
> + (!rx_adapter->use_queue_event_buf &&
> + queue_conf->event_buf_size != 0)) {
> + RTE_EDEV_LOG_ERR("Invalid Event buffer size for the queue");
> + return -EINVAL;
> + }
> +
> dev_info = &rx_adapter->eth_devices[eth_dev_id];
>
> if (cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT) {
> --
> 2.25.1
More information about the dev
mailing list