[dpdk-dev] [PATCH 1/2] event/cnxk: add Rx adapter vector support

[dpdk-dev] [PATCH 1/2] event/cnxk: add Rx adapter vector support

[pbhagavatula]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Add event vector support for cnxk event Rx adapter, add control path
APIs to get vector limits and ability to configure event vectorization
on a given Rx queue.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/event/cnxk/cn10k_eventdev.c      | 106 ++++++++++++++++++++++-
 drivers/event/cnxk/cnxk_eventdev.h       |   2 +
 drivers/event/cnxk/cnxk_eventdev_adptr.c |  25 ++++++
 3 files changed, 132 insertions(+), 1 deletion(-)

diff --git a/drivers/event/cnxk/cn10k_eventdev.c b/drivers/event/cnxk/cn10k_eventdev.c
index 817cb08480..c9e4ddf1bb 100644
--- a/drivers/event/cnxk/cn10k_eventdev.c
+++ b/drivers/event/cnxk/cn10k_eventdev.c
@@ -595,7 +595,8 @@ cn10k_sso_rx_adapter_caps_get(const struct rte_eventdev *event_dev,
 	else
 		*caps = RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT |
 			RTE_EVENT_ETH_RX_ADAPTER_CAP_MULTI_EVENTQ |
-			RTE_EVENT_ETH_RX_ADAPTER_CAP_OVERRIDE_FLOW_ID;
+			RTE_EVENT_ETH_RX_ADAPTER_CAP_OVERRIDE_FLOW_ID |
+			RTE_EVENT_ETH_RX_ADAPTER_CAP_EVENT_VECTOR;

 	return 0;
 }
@@ -652,6 +653,105 @@ cn10k_sso_rx_adapter_queue_del(const struct rte_eventdev *event_dev,
 	return cnxk_sso_rx_adapter_queue_del(event_dev, eth_dev, rx_queue_id);
 }

+static int
+cn10k_sso_rx_adapter_vector_limits(
+	const struct rte_eventdev *dev, const struct rte_eth_dev *eth_dev,
+	struct rte_event_eth_rx_adapter_vector_limits *limits)
+{
+	struct cnxk_eth_dev *cnxk_eth_dev;
+	int ret;
+
+	RTE_SET_USED(dev);
+	ret = strncmp(eth_dev->device->driver->name, "net_cn10k", 8);
+	if (ret)
+		return -ENOTSUP;
+
+	cnxk_eth_dev = cnxk_eth_pmd_priv(eth_dev);
+	limits->log2_sz = true;
+	limits->min_sz = 1 << ROC_NIX_VWQE_MIN_SIZE_LOG2;
+	limits->max_sz = 1 << ROC_NIX_VWQE_MAX_SIZE_LOG2;
+	limits->min_timeout_ns =
+		(roc_nix_get_vwqe_interval(&cnxk_eth_dev->nix) + 1) * 100;
+	limits->max_timeout_ns = BITMASK_ULL(8, 0) * limits->min_timeout_ns;
+
+	return 0;
+}
+
+static int
+cnxk_sso_rx_adapter_vwqe_enable(struct cnxk_eth_dev *cnxk_eth_dev,
+				uint16_t port_id, uint16_t rq_id, uint16_t sz,
+				uint64_t tmo_ns, struct rte_mempool *vmp)
+{
+	struct roc_nix_rq *rq;
+
+	rq = &cnxk_eth_dev->rqs[rq_id];
+
+	if (!rq->sso_ena)
+		return -EINVAL;
+	if (rq->flow_tag_width == 0)
+		return -EINVAL;
+
+	rq->vwqe_ena = 1;
+	rq->vwqe_first_skip = 0;
+	rq->vwqe_aura_handle = roc_npa_aura_handle_to_aura(vmp->pool_id);
+	rq->vwqe_max_sz_exp = rte_log2_u32(sz);
+	rq->vwqe_wait_tmo =
+		tmo_ns /
+		((roc_nix_get_vwqe_interval(&cnxk_eth_dev->nix) + 1) * 100);
+	rq->tag_mask = (port_id & 0xF) << 20;
+	rq->tag_mask |=
+		(((port_id >> 4) & 0xF) | (RTE_EVENT_TYPE_ETHDEV_VECTOR << 4))
+		<< 24;
+
+	return roc_nix_rq_modify(&cnxk_eth_dev->nix, rq, 0);
+}
+
+static int
+cn10k_sso_rx_adapter_vector_config(
+	const struct rte_eventdev *event_dev, const struct rte_eth_dev *eth_dev,
+	int32_t rx_queue_id,
+	const struct rte_event_eth_rx_adapter_event_vector_config *config)
+{
+	struct cnxk_eth_dev *cnxk_eth_dev;
+	struct cnxk_sso_evdev *dev;
+	int i, rc;
+
+	rc = strncmp(eth_dev->device->driver->name, "net_cn10k", 8);
+	if (rc)
+		return -EINVAL;
+
+	dev = cnxk_sso_pmd_priv(event_dev);
+	cnxk_eth_dev = cnxk_eth_pmd_priv(eth_dev);
+	if (rx_queue_id < 0) {
+		for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
+			cnxk_sso_updt_xae_cnt(dev, config->vector_mp,
+					      RTE_EVENT_TYPE_ETHDEV_VECTOR);
+			rc = cnxk_sso_xae_reconfigure(
+				(struct rte_eventdev *)(uintptr_t)event_dev);
+			rc = cnxk_sso_rx_adapter_vwqe_enable(
+				cnxk_eth_dev, eth_dev->data->port_id, i,
+				config->vector_sz, config->vector_timeout_ns,
+				config->vector_mp);
+			if (rc)
+				return -EINVAL;
+		}
+	} else {
+
+		cnxk_sso_updt_xae_cnt(dev, config->vector_mp,
+				      RTE_EVENT_TYPE_ETHDEV_VECTOR);
+		rc = cnxk_sso_xae_reconfigure(
+			(struct rte_eventdev *)(uintptr_t)event_dev);
+		rc = cnxk_sso_rx_adapter_vwqe_enable(
+			cnxk_eth_dev, eth_dev->data->port_id, rx_queue_id,
+			config->vector_sz, config->vector_timeout_ns,
+			config->vector_mp);
+		if (rc)
+			return -EINVAL;
+	}
+
+	return 0;
+}
+
 static int
 cn10k_sso_tx_adapter_caps_get(const struct rte_eventdev *dev,
 			      const struct rte_eth_dev *eth_dev, uint32_t *caps)
@@ -720,6 +820,10 @@ static struct rte_eventdev_ops cn10k_sso_dev_ops = {
 	.eth_rx_adapter_start = cnxk_sso_rx_adapter_start,
 	.eth_rx_adapter_stop = cnxk_sso_rx_adapter_stop,

+	.eth_rx_adapter_vector_limits_get = cn10k_sso_rx_adapter_vector_limits,
+	.eth_rx_adapter_event_vector_config =
+		cn10k_sso_rx_adapter_vector_config,
+
 	.eth_tx_adapter_caps_get = cn10k_sso_tx_adapter_caps_get,
 	.eth_tx_adapter_queue_add = cn10k_sso_tx_adapter_queue_add,
 	.eth_tx_adapter_queue_del = cn10k_sso_tx_adapter_queue_del,
diff --git a/drivers/event/cnxk/cnxk_eventdev.h b/drivers/event/cnxk/cnxk_eventdev.h
index 57c3327aa0..bca5d27912 100644
--- a/drivers/event/cnxk/cnxk_eventdev.h
+++ b/drivers/event/cnxk/cnxk_eventdev.h
@@ -94,6 +94,8 @@ struct cnxk_sso_evdev {
 	uint16_t tim_adptr_ring_cnt;
 	uint16_t *timer_adptr_rings;
 	uint64_t *timer_adptr_sz;
+	uint16_t vec_pool_cnt;
+	uint64_t *vec_pools;
 	/* Dev args */
 	uint32_t xae_cnt;
 	uint8_t qos_queue_cnt;
diff --git a/drivers/event/cnxk/cnxk_eventdev_adptr.c b/drivers/event/cnxk/cnxk_eventdev_adptr.c
index d4d07c793f..95483cb6ee 100644
--- a/drivers/event/cnxk/cnxk_eventdev_adptr.c
+++ b/drivers/event/cnxk/cnxk_eventdev_adptr.c
@@ -40,6 +40,31 @@ cnxk_sso_updt_xae_cnt(struct cnxk_sso_evdev *dev, void *data,
 		dev->adptr_xae_cnt += rxq->qconf.mp->size;
 		break;
 	}
+	case RTE_EVENT_TYPE_ETHDEV_VECTOR: {
+		struct rte_mempool *mp = data;
+		uint64_t *old_ptr;
+
+		for (i = 0; i < dev->vec_pool_cnt; i++) {
+			if ((uint64_t)mp == dev->vec_pools[i])
+				return;
+		}
+
+		dev->vec_pool_cnt++;
+		old_ptr = dev->vec_pools;
+		dev->vec_pools =
+			rte_realloc(dev->vec_pools,
+				    sizeof(uint64_t) * dev->vec_pool_cnt, 0);
+		if (dev->vec_pools == NULL) {
+			dev->adptr_xae_cnt += mp->size;
+			dev->vec_pools = old_ptr;
+			dev->vec_pool_cnt--;
+			return;
+		}
+		dev->vec_pools[dev->vec_pool_cnt - 1] = (uint64_t)mp;
+
+		dev->adptr_xae_cnt += mp->size;
+		break;
+	}
 	case RTE_EVENT_TYPE_TIMER: {
 		struct cnxk_tim_ring *timr = data;
 		uint16_t *old_ring_ptr;
--
2.17.1

[dpdk-dev] [PATCH 2/2] event/cnxk: add Rx event vector fastpath

[pbhagavatula]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Add Rx event vector fastpath to convert HW defined metadata into
rte_mbuf and rte_event_vector.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/event/cnxk/cn10k_worker.h | 50 +++++++++++++++++++++++++++++++
 1 file changed, 50 insertions(+)

diff --git a/drivers/event/cnxk/cn10k_worker.h b/drivers/event/cnxk/cn10k_worker.h
index b3f71202ad..8c2cd72873 100644
--- a/drivers/event/cnxk/cn10k_worker.h
+++ b/drivers/event/cnxk/cn10k_worker.h
@@ -5,6 +5,8 @@
 #ifndef __CN10K_WORKER_H__
 #define __CN10K_WORKER_H__

+#include <rte_vect.h>
+
 #include "cnxk_ethdev.h"
 #include "cnxk_eventdev.h"
 #include "cnxk_worker.h"
@@ -101,6 +103,44 @@ cn10k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
 			      mbuf_init.value, flags);
 }

+static __rte_always_inline void
+cn10k_process_vwqe(uintptr_t vwqe, uint16_t port_id, const uint32_t flags,
+		   void *lookup_mem)
+{
+	union mbuf_initializer mbuf_init = {
+		.fields = {.data_off = RTE_PKTMBUF_HEADROOM,
+			   .refcnt = 1,
+			   .nb_segs = 1,
+			   .port = port_id},
+	};
+	struct rte_event_vector *vec;
+	uint16_t nb_mbufs, non_vec;
+	uint64_t **wqe;
+
+	vec = (struct rte_event_vector *)vwqe;
+	wqe = vec->u64s;
+
+	nb_mbufs = RTE_ALIGN_FLOOR(vec->nb_elem, NIX_DESCS_PER_LOOP);
+	nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init.value, vec->mbufs,
+					      nb_mbufs, flags | NIX_RX_VWQE_F,
+					      lookup_mem);
+	wqe += nb_mbufs;
+	non_vec = vec->nb_elem - nb_mbufs;
+
+	while (non_vec) {
+		struct nix_cqe_hdr_s *cqe = (struct nix_cqe_hdr_s *)wqe[0];
+		struct rte_mbuf *mbuf;
+
+		mbuf = (struct rte_mbuf *)((char *)cqe -
+					   sizeof(struct rte_mbuf));
+		cn10k_nix_cqe_to_mbuf(cqe, cqe->tag, mbuf, lookup_mem,
+				      mbuf_init.value, flags);
+		wqe[0] = (uint64_t *)mbuf;
+		non_vec--;
+		wqe++;
+	}
+}
+
 static __rte_always_inline uint16_t
 cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
 		       const uint32_t flags, void *lookup_mem)
@@ -141,6 +181,16 @@ cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
 					  gw.u64[0] & 0xFFFFF, flags,
 					  lookup_mem);
 			gw.u64[1] = mbuf;
+		} else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
+			   RTE_EVENT_TYPE_ETHDEV_VECTOR) {
+			uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
+			__uint128_t vwqe_hdr = *(__uint128_t *)gw.u64[1];
+
+			vwqe_hdr = ((vwqe_hdr >> 64) & 0xFFF) | BIT_ULL(31) |
+				   ((vwqe_hdr & 0xFFFF) << 48) |
+				   ((uint64_t)port << 32);
+			*(uint64_t *)gw.u64[1] = (uint64_t)vwqe_hdr;
+			cn10k_process_vwqe(gw.u64[1], port, flags, lookup_mem);
 		}
 	}

--
2.17.1

Re: [dpdk-dev] [PATCH 2/2] event/cnxk: add Rx event vector fastpath

[Jerin Jacob]

On Mon, May 24, 2021 at 6:36 PM <pbhagavatula@marvell.com> wrote:
>
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
> Add Rx event vector fastpath to convert HW defined metadata into
> rte_mbuf and rte_event_vector.
>
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> ---
>  drivers/event/cnxk/cn10k_worker.h | 50 +++++++++++++++++++++++++++++++
>  1 file changed, 50 insertions(+)
>
> diff --git a/drivers/event/cnxk/cn10k_worker.h b/drivers/event/cnxk/cn10k_worker.h
> index b3f71202ad..8c2cd72873 100644
> --- a/drivers/event/cnxk/cn10k_worker.h
> +++ b/drivers/event/cnxk/cn10k_worker.h
> @@ -5,6 +5,8 @@
>  #ifndef __CN10K_WORKER_H__
>  #define __CN10K_WORKER_H__
>
> +#include <rte_vect.h>
> +
>  #include "cnxk_ethdev.h"
>  #include "cnxk_eventdev.h"
>  #include "cnxk_worker.h"
> @@ -101,6 +103,44 @@ cn10k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
>                               mbuf_init.value, flags);
>  }
>
> +static __rte_always_inline void
> +cn10k_process_vwqe(uintptr_t vwqe, uint16_t port_id, const uint32_t flags,
> +                  void *lookup_mem)
> +{
> +       union mbuf_initializer mbuf_init = {
> +               .fields = {.data_off = RTE_PKTMBUF_HEADROOM,
> +                          .refcnt = 1,
> +                          .nb_segs = 1,
> +                          .port = port_id},
> +       };


If there is performance improvement move mbuf_init population to slow
path and store it
in HWS software structure or so.

> +       struct rte_event_vector *vec;
> +       uint16_t nb_mbufs, non_vec;
> +       uint64_t **wqe;
> +
> +       vec = (struct rte_event_vector *)vwqe;
> +       wqe = vec->u64s;
> +
> +       nb_mbufs = RTE_ALIGN_FLOOR(vec->nb_elem, NIX_DESCS_PER_LOOP);
> +       nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init.value, vec->mbufs,
> +                                             nb_mbufs, flags | NIX_RX_VWQE_F,
> +                                             lookup_mem);
> +       wqe += nb_mbufs;
> +       non_vec = vec->nb_elem - nb_mbufs;
> +
> +       while (non_vec) {
> +               struct nix_cqe_hdr_s *cqe = (struct nix_cqe_hdr_s *)wqe[0];
> +               struct rte_mbuf *mbuf;
> +
> +               mbuf = (struct rte_mbuf *)((char *)cqe -
> +                                          sizeof(struct rte_mbuf));
> +               cn10k_nix_cqe_to_mbuf(cqe, cqe->tag, mbuf, lookup_mem,
> +                                     mbuf_init.value, flags);
> +               wqe[0] = (uint64_t *)mbuf;
> +               non_vec--;
> +               wqe++;
> +       }
> +}
> +
>  static __rte_always_inline uint16_t
>  cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
>                        const uint32_t flags, void *lookup_mem)
> @@ -141,6 +181,16 @@ cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
>                                           gw.u64[0] & 0xFFFFF, flags,
>                                           lookup_mem);
>                         gw.u64[1] = mbuf;
> +               } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
> +                          RTE_EVENT_TYPE_ETHDEV_VECTOR) {
> +                       uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
> +                       __uint128_t vwqe_hdr = *(__uint128_t *)gw.u64[1];
> +
> +                       vwqe_hdr = ((vwqe_hdr >> 64) & 0xFFF) | BIT_ULL(31) |
> +                                  ((vwqe_hdr & 0xFFFF) << 48) |
> +                                  ((uint64_t)port << 32);
> +                       *(uint64_t *)gw.u64[1] = (uint64_t)vwqe_hdr;
> +                       cn10k_process_vwqe(gw.u64[1], port, flags, lookup_mem);
>                 }
>         }
>
> --
> 2.17.1
>

Re: [dpdk-dev] [PATCH 1/2] event/cnxk: add Rx adapter vector support

[Jerin Jacob]

On Mon, May 24, 2021 at 6:36 PM <pbhagavatula@marvell.com> wrote:
>
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
> Add event vector support for cnxk event Rx adapter, add control path
> APIs to get vector limits and ability to configure event vectorization
> on a given Rx queue.
>
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>

> +
> +static int
> +cn10k_sso_rx_adapter_vector_config(
> +       const struct rte_eventdev *event_dev, const struct rte_eth_dev *eth_dev,
> +       int32_t rx_queue_id,
> +       const struct rte_event_eth_rx_adapter_event_vector_config *config)
> +{
> +       struct cnxk_eth_dev *cnxk_eth_dev;
> +       struct cnxk_sso_evdev *dev;
> +       int i, rc;
> +
> +       rc = strncmp(eth_dev->device->driver->name, "net_cn10k", 8);
> +       if (rc)
> +               return -EINVAL;
> +
> +       dev = cnxk_sso_pmd_priv(event_dev);
> +       cnxk_eth_dev = cnxk_eth_pmd_priv(eth_dev);
> +       if (rx_queue_id < 0) {
> +               for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
> +                       cnxk_sso_updt_xae_cnt(dev, config->vector_mp,
> +                                             RTE_EVENT_TYPE_ETHDEV_VECTOR);
> +                       rc = cnxk_sso_xae_reconfigure(
> +                               (struct rte_eventdev *)(uintptr_t)event_dev);
> +                       rc = cnxk_sso_rx_adapter_vwqe_enable(
> +                               cnxk_eth_dev, eth_dev->data->port_id, i,
> +                               config->vector_sz, config->vector_timeout_ns,
> +                               config->vector_mp);
> +                       if (rc)
> +                               return -EINVAL;
> +               }

See below

> +       } else {
> +
> +               cnxk_sso_updt_xae_cnt(dev, config->vector_mp,
> +                                     RTE_EVENT_TYPE_ETHDEV_VECTOR);
> +               rc = cnxk_sso_xae_reconfigure(
> +                       (struct rte_eventdev *)(uintptr_t)event_dev);
> +               rc = cnxk_sso_rx_adapter_vwqe_enable(
> +                       cnxk_eth_dev, eth_dev->data->port_id, rx_queue_id,
> +                       config->vector_sz, config->vector_timeout_ns,
> +                       config->vector_mp);
> +               if (rc)
> +                       return -EINVAL;
> +       }

Parameterize and move this separate function to avoid code duplication.

> +
> +       return 0;
> +}
> +


# Add update this feature in doc/guides/eventdevs/cnxk.rst.
# Please add a new section in the release not for the enhancement of
event cnxk driver and add new feature there
as you update the feature in code.