net/ena: destroy queues if start failed
[dpdk.git] / drivers / net / ena / ena_ethdev.c
index 89080a9..deb9715 100644 (file)
@@ -55,7 +55,7 @@
 
 #define DRV_MODULE_VER_MAJOR   1
 #define DRV_MODULE_VER_MINOR   1
-#define DRV_MODULE_VER_SUBMINOR        0
+#define DRV_MODULE_VER_SUBMINOR        1
 
 #define ENA_IO_TXQ_IDX(q)      (2 * (q))
 #define ENA_IO_RXQ_IDX(q)      (2 * (q) + 1)
@@ -85,7 +85,6 @@
 
 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
 
-#define ENA_MAX_RING_DESC      ENA_DEFAULT_RING_SIZE
 #define ENA_MIN_RING_DESC      128
 
 enum ethtool_stringset {
@@ -117,6 +116,9 @@ struct ena_stats {
 #define ENA_STAT_GLOBAL_ENTRY(stat) \
        ENA_STAT_ENTRY(stat, dev)
 
+#define ENA_MAX_RING_SIZE_RX 8192
+#define ENA_MAX_RING_SIZE_TX 1024
+
 /*
  * Each rte_memzone should have unique name.
  * To satisfy it, count number of allocation and add it to name.
@@ -189,6 +191,8 @@ static const struct ena_stats ena_stats_ena_com_strings[] = {
 
 #define        ENA_TX_OFFLOAD_MASK     (\
        PKT_TX_L4_MASK |         \
+       PKT_TX_IPV6 |            \
+       PKT_TX_IPV4 |            \
        PKT_TX_IP_CKSUM |        \
        PKT_TX_TCP_SEG)
 
@@ -239,9 +243,13 @@ static void ena_rx_queue_release_bufs(struct ena_ring *ring);
 static void ena_tx_queue_release_bufs(struct ena_ring *ring);
 static int ena_link_update(struct rte_eth_dev *dev,
                           int wait_to_complete);
-static int ena_queue_restart(struct ena_ring *ring);
-static int ena_queue_restart_all(struct rte_eth_dev *dev,
-                                enum ena_ring_type ring_type);
+static int ena_create_io_queue(struct ena_ring *ring);
+static void ena_queue_stop(struct ena_ring *ring);
+static void ena_queue_stop_all(struct rte_eth_dev *dev,
+                             enum ena_ring_type ring_type);
+static int ena_queue_start(struct ena_ring *ring);
+static int ena_queue_start_all(struct rte_eth_dev *dev,
+                              enum ena_ring_type ring_type);
 static void ena_stats_restart(struct rte_eth_dev *dev);
 static void ena_infos_get(struct rte_eth_dev *dev,
                          struct rte_eth_dev_info *dev_info);
@@ -433,6 +441,7 @@ static void ena_config_host_info(struct ena_com_dev *ena_dev)
                (DRV_MODULE_VER_MINOR << ENA_ADMIN_HOST_INFO_MINOR_SHIFT) |
                (DRV_MODULE_VER_SUBMINOR <<
                        ENA_ADMIN_HOST_INFO_SUB_MINOR_SHIFT);
+       host_info->num_cpus = rte_lcore_count();
 
        rc = ena_com_set_host_attributes(ena_dev);
        if (rc) {
@@ -507,14 +516,31 @@ err:
 
 static void ena_close(struct rte_eth_dev *dev)
 {
+       struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
+       struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
        struct ena_adapter *adapter =
                (struct ena_adapter *)(dev->data->dev_private);
 
-       ena_stop(dev);
+       if (adapter->state == ENA_ADAPTER_STATE_RUNNING)
+               ena_stop(dev);
        adapter->state = ENA_ADAPTER_STATE_CLOSED;
 
        ena_rx_queue_release_all(dev);
        ena_tx_queue_release_all(dev);
+
+       rte_free(adapter->drv_stats);
+       adapter->drv_stats = NULL;
+
+       rte_intr_disable(intr_handle);
+       rte_intr_callback_unregister(intr_handle,
+                                    ena_interrupt_handler_rte,
+                                    adapter);
+
+       /*
+        * MAC is not allocated dynamically. Setting NULL should prevent from
+        * release of the resource in the rte_eth_dev_release_port().
+        */
+       dev->data->mac_addrs = NULL;
 }
 
 static int
@@ -569,11 +595,12 @@ ena_dev_reset(struct rte_eth_dev *dev)
        ena_com_admin_aenq_enable(ena_dev);
 
        for (i = 0; i < nb_queues; ++i)
-               ena_rx_queue_setup(eth_dev, i, adapter->rx_ring_size, 0, NULL,
-                       mb_pool_rx[i]);
+               ena_rx_queue_setup(eth_dev, i, adapter->rx_ring[i].ring_size, 0,
+                       NULL, mb_pool_rx[i]);
 
        for (i = 0; i < nb_queues; ++i)
-               ena_tx_queue_setup(eth_dev, i, adapter->tx_ring_size, 0, NULL);
+               ena_tx_queue_setup(eth_dev, i, adapter->tx_ring[i].ring_size, 0,
+                       NULL);
 
        adapter->trigger_reset = false;
 
@@ -746,26 +773,21 @@ static void ena_tx_queue_release_all(struct rte_eth_dev *dev)
 static void ena_rx_queue_release(void *queue)
 {
        struct ena_ring *ring = (struct ena_ring *)queue;
-       struct ena_adapter *adapter = ring->adapter;
-       int ena_qid;
 
        ena_assert_msg(ring->configured,
                       "API violation - releasing not configured queue");
        ena_assert_msg(ring->adapter->state != ENA_ADAPTER_STATE_RUNNING,
                       "API violation");
 
-       /* Destroy HW queue */
-       ena_qid = ENA_IO_RXQ_IDX(ring->id);
-       ena_com_destroy_io_queue(&adapter->ena_dev, ena_qid);
-
-       /* Free all bufs */
-       ena_rx_queue_release_bufs(ring);
-
        /* Free ring resources */
        if (ring->rx_buffer_info)
                rte_free(ring->rx_buffer_info);
        ring->rx_buffer_info = NULL;
 
+       if (ring->rx_refill_buffer)
+               rte_free(ring->rx_refill_buffer);
+       ring->rx_refill_buffer = NULL;
+
        if (ring->empty_rx_reqs)
                rte_free(ring->empty_rx_reqs);
        ring->empty_rx_reqs = NULL;
@@ -779,22 +801,16 @@ static void ena_rx_queue_release(void *queue)
 static void ena_tx_queue_release(void *queue)
 {
        struct ena_ring *ring = (struct ena_ring *)queue;
-       struct ena_adapter *adapter = ring->adapter;
-       int ena_qid;
 
        ena_assert_msg(ring->configured,
                       "API violation. Releasing not configured queue");
        ena_assert_msg(ring->adapter->state != ENA_ADAPTER_STATE_RUNNING,
                       "API violation");
 
-       /* Destroy HW queue */
-       ena_qid = ENA_IO_TXQ_IDX(ring->id);
-       ena_com_destroy_io_queue(&adapter->ena_dev, ena_qid);
-
-       /* Free all bufs */
-       ena_tx_queue_release_bufs(ring);
-
        /* Free ring resources */
+       if (ring->push_buf_intermediate_buf)
+               rte_free(ring->push_buf_intermediate_buf);
+
        if (ring->tx_buffer_info)
                rte_free(ring->tx_buffer_info);
 
@@ -803,6 +819,7 @@ static void ena_tx_queue_release(void *queue)
 
        ring->empty_tx_reqs = NULL;
        ring->tx_buffer_info = NULL;
+       ring->push_buf_intermediate_buf = NULL;
 
        ring->configured = 0;
 
@@ -854,8 +871,8 @@ static int ena_link_update(struct rte_eth_dev *dev,
        return 0;
 }
 
-static int ena_queue_restart_all(struct rte_eth_dev *dev,
-                                enum ena_ring_type ring_type)
+static int ena_queue_start_all(struct rte_eth_dev *dev,
+                              enum ena_ring_type ring_type)
 {
        struct ena_adapter *adapter =
                (struct ena_adapter *)(dev->data->dev_private);
@@ -883,18 +900,25 @@ static int ena_queue_restart_all(struct rte_eth_dev *dev,
                                        "Inconsistent state of tx queues\n");
                        }
 
-                       rc = ena_queue_restart(&queues[i]);
+                       rc = ena_queue_start(&queues[i]);
 
                        if (rc) {
                                PMD_INIT_LOG(ERR,
-                                            "failed to restart queue %d type(%d)",
+                                            "failed to start queue %d type(%d)",
                                             i, ring_type);
-                               return rc;
+                               goto err;
                        }
                }
        }
 
        return 0;
+
+err:
+       while (i--)
+               if (queues[i].configured)
+                       ena_queue_stop(&queues[i]);
+
+       return rc;
 }
 
 static uint32_t ena_get_mtu_conf(struct ena_adapter *adapter)
@@ -924,34 +948,74 @@ static int ena_check_valid_conf(struct ena_adapter *adapter)
 }
 
 static int
-ena_calc_queue_size(struct ena_com_dev *ena_dev,
-                   u16 *max_tx_sgl_size,
-                   struct ena_com_dev_get_features_ctx *get_feat_ctx)
+ena_calc_queue_size(struct ena_calc_queue_size_ctx *ctx)
 {
-       uint32_t queue_size = ENA_DEFAULT_RING_SIZE;
+       struct ena_admin_feature_llq_desc *llq = &ctx->get_feat_ctx->llq;
+       struct ena_com_dev *ena_dev = ctx->ena_dev;
+       uint32_t tx_queue_size = ENA_MAX_RING_SIZE_TX;
+       uint32_t rx_queue_size = ENA_MAX_RING_SIZE_RX;
+
+       if (ena_dev->supported_features & BIT(ENA_ADMIN_MAX_QUEUES_EXT)) {
+               struct ena_admin_queue_ext_feature_fields *max_queue_ext =
+                       &ctx->get_feat_ctx->max_queue_ext.max_queue_ext;
+               rx_queue_size = RTE_MIN(rx_queue_size,
+                       max_queue_ext->max_rx_cq_depth);
+               rx_queue_size = RTE_MIN(rx_queue_size,
+                       max_queue_ext->max_rx_sq_depth);
+               tx_queue_size = RTE_MIN(tx_queue_size,
+                       max_queue_ext->max_tx_cq_depth);
+
+               if (ena_dev->tx_mem_queue_type ==
+                   ENA_ADMIN_PLACEMENT_POLICY_DEV) {
+                       tx_queue_size = RTE_MIN(tx_queue_size,
+                               llq->max_llq_depth);
+               } else {
+                       tx_queue_size = RTE_MIN(tx_queue_size,
+                               max_queue_ext->max_tx_sq_depth);
+               }
 
-       queue_size = RTE_MIN(queue_size,
-                            get_feat_ctx->max_queues.max_cq_depth);
-       queue_size = RTE_MIN(queue_size,
-                            get_feat_ctx->max_queues.max_sq_depth);
+               ctx->max_rx_sgl_size = RTE_MIN(ENA_PKT_MAX_BUFS,
+                       max_queue_ext->max_per_packet_rx_descs);
+               ctx->max_tx_sgl_size = RTE_MIN(ENA_PKT_MAX_BUFS,
+                       max_queue_ext->max_per_packet_tx_descs);
+       } else {
+               struct ena_admin_queue_feature_desc *max_queues =
+                       &ctx->get_feat_ctx->max_queues;
+               rx_queue_size = RTE_MIN(rx_queue_size,
+                       max_queues->max_cq_depth);
+               rx_queue_size = RTE_MIN(rx_queue_size,
+                       max_queues->max_sq_depth);
+               tx_queue_size = RTE_MIN(tx_queue_size,
+                       max_queues->max_cq_depth);
+
+               if (ena_dev->tx_mem_queue_type ==
+                   ENA_ADMIN_PLACEMENT_POLICY_DEV) {
+                       tx_queue_size = RTE_MIN(tx_queue_size,
+                               llq->max_llq_depth);
+               } else {
+                       tx_queue_size = RTE_MIN(tx_queue_size,
+                               max_queues->max_sq_depth);
+               }
 
-       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV)
-               queue_size = RTE_MIN(queue_size,
-                                    get_feat_ctx->max_queues.max_llq_depth);
+               ctx->max_rx_sgl_size = RTE_MIN(ENA_PKT_MAX_BUFS,
+                       max_queues->max_packet_tx_descs);
+               ctx->max_tx_sgl_size = RTE_MIN(ENA_PKT_MAX_BUFS,
+                       max_queues->max_packet_rx_descs);
+       }
 
-       /* Round down to power of 2 */
-       if (!rte_is_power_of_2(queue_size))
-               queue_size = rte_align32pow2(queue_size >> 1);
+       /* Round down to the nearest power of 2 */
+       rx_queue_size = rte_align32prevpow2(rx_queue_size);
+       tx_queue_size = rte_align32prevpow2(tx_queue_size);
 
-       if (unlikely(queue_size == 0)) {
+       if (unlikely(rx_queue_size == 0 || tx_queue_size == 0)) {
                PMD_INIT_LOG(ERR, "Invalid queue size");
                return -EFAULT;
        }
 
-       *max_tx_sgl_size = RTE_MIN(ENA_PKT_MAX_BUFS,
-               get_feat_ctx->max_queues.max_packet_tx_descs);
+       ctx->rx_queue_size = rx_queue_size;
+       ctx->tx_queue_size = tx_queue_size;
 
-       return queue_size;
+       return 0;
 }
 
 static void ena_stats_restart(struct rte_eth_dev *dev)
@@ -1043,19 +1107,19 @@ static int ena_start(struct rte_eth_dev *dev)
        if (rc)
                return rc;
 
-       rc = ena_queue_restart_all(dev, ENA_RING_TYPE_RX);
+       rc = ena_queue_start_all(dev, ENA_RING_TYPE_RX);
        if (rc)
                return rc;
 
-       rc = ena_queue_restart_all(dev, ENA_RING_TYPE_TX);
+       rc = ena_queue_start_all(dev, ENA_RING_TYPE_TX);
        if (rc)
-               return rc;
+               goto err_start_tx;
 
        if (adapter->rte_dev->data->dev_conf.rxmode.mq_mode &
            ETH_MQ_RX_RSS_FLAG && adapter->rte_dev->data->nb_rx_queues > 0) {
                rc = ena_rss_init_default(adapter);
                if (rc)
-                       return rc;
+                       goto err_rss_init;
        }
 
        ena_stats_restart(dev);
@@ -1070,6 +1134,12 @@ static int ena_start(struct rte_eth_dev *dev)
        adapter->state = ENA_ADAPTER_STATE_RUNNING;
 
        return 0;
+
+err_rss_init:
+       ena_queue_stop_all(dev, ENA_RING_TYPE_TX);
+err_start_tx:
+       ena_queue_stop_all(dev, ENA_RING_TYPE_RX);
+       return rc;
 }
 
 static void ena_stop(struct rte_eth_dev *dev)
@@ -1078,16 +1148,116 @@ static void ena_stop(struct rte_eth_dev *dev)
                (struct ena_adapter *)(dev->data->dev_private);
 
        rte_timer_stop_sync(&adapter->timer_wd);
+       ena_queue_stop_all(dev, ENA_RING_TYPE_TX);
+       ena_queue_stop_all(dev, ENA_RING_TYPE_RX);
 
        adapter->state = ENA_ADAPTER_STATE_STOPPED;
 }
 
-static int ena_queue_restart(struct ena_ring *ring)
+static int ena_create_io_queue(struct ena_ring *ring)
+{
+       struct ena_adapter *adapter;
+       struct ena_com_dev *ena_dev;
+       struct ena_com_create_io_ctx ctx =
+               /* policy set to _HOST just to satisfy icc compiler */
+               { ENA_ADMIN_PLACEMENT_POLICY_HOST,
+                 0, 0, 0, 0, 0 };
+       uint16_t ena_qid;
+       unsigned int i;
+       int rc;
+
+       adapter = ring->adapter;
+       ena_dev = &adapter->ena_dev;
+
+       if (ring->type == ENA_RING_TYPE_TX) {
+               ena_qid = ENA_IO_TXQ_IDX(ring->id);
+               ctx.direction = ENA_COM_IO_QUEUE_DIRECTION_TX;
+               ctx.mem_queue_type = ena_dev->tx_mem_queue_type;
+               ctx.queue_size = adapter->tx_ring_size;
+               for (i = 0; i < ring->ring_size; i++)
+                       ring->empty_tx_reqs[i] = i;
+       } else {
+               ena_qid = ENA_IO_RXQ_IDX(ring->id);
+               ctx.direction = ENA_COM_IO_QUEUE_DIRECTION_RX;
+               ctx.queue_size = adapter->rx_ring_size;
+               for (i = 0; i < ring->ring_size; i++)
+                       ring->empty_rx_reqs[i] = i;
+       }
+       ctx.qid = ena_qid;
+       ctx.msix_vector = -1; /* interrupts not used */
+       ctx.numa_node = ena_cpu_to_node(ring->id);
+
+       rc = ena_com_create_io_queue(ena_dev, &ctx);
+       if (rc) {
+               RTE_LOG(ERR, PMD,
+                       "failed to create io queue #%d (qid:%d) rc: %d\n",
+                       ring->id, ena_qid, rc);
+               return rc;
+       }
+
+       rc = ena_com_get_io_handlers(ena_dev, ena_qid,
+                                    &ring->ena_com_io_sq,
+                                    &ring->ena_com_io_cq);
+       if (rc) {
+               RTE_LOG(ERR, PMD,
+                       "Failed to get io queue handlers. queue num %d rc: %d\n",
+                       ring->id, rc);
+               ena_com_destroy_io_queue(ena_dev, ena_qid);
+               return rc;
+       }
+
+       if (ring->type == ENA_RING_TYPE_TX)
+               ena_com_update_numa_node(ring->ena_com_io_cq, ctx.numa_node);
+
+       return 0;
+}
+
+static void ena_queue_stop(struct ena_ring *ring)
+{
+       struct ena_com_dev *ena_dev = &ring->adapter->ena_dev;
+
+       if (ring->type == ENA_RING_TYPE_RX) {
+               ena_com_destroy_io_queue(ena_dev, ENA_IO_RXQ_IDX(ring->id));
+               ena_rx_queue_release_bufs(ring);
+       } else {
+               ena_com_destroy_io_queue(ena_dev, ENA_IO_TXQ_IDX(ring->id));
+               ena_tx_queue_release_bufs(ring);
+       }
+}
+
+static void ena_queue_stop_all(struct rte_eth_dev *dev,
+                             enum ena_ring_type ring_type)
+{
+       struct ena_adapter *adapter =
+               (struct ena_adapter *)(dev->data->dev_private);
+       struct ena_ring *queues = NULL;
+       uint16_t nb_queues, i;
+
+       if (ring_type == ENA_RING_TYPE_RX) {
+               queues = adapter->rx_ring;
+               nb_queues = dev->data->nb_rx_queues;
+       } else {
+               queues = adapter->tx_ring;
+               nb_queues = dev->data->nb_tx_queues;
+       }
+
+       for (i = 0; i < nb_queues; ++i)
+               if (queues[i].configured)
+                       ena_queue_stop(&queues[i]);
+}
+
+static int ena_queue_start(struct ena_ring *ring)
 {
        int rc, bufs_num;
 
        ena_assert_msg(ring->configured == 1,
-                      "Trying to restart unconfigured queue\n");
+                      "Trying to start unconfigured queue\n");
+
+       rc = ena_create_io_queue(ring);
+       if (rc) {
+               PMD_INIT_LOG(ERR, "Failed to create IO queue!\n");
+               return rc;
+       }
 
        ring->next_to_clean = 0;
        ring->next_to_use = 0;
@@ -1098,6 +1268,8 @@ static int ena_queue_restart(struct ena_ring *ring)
        bufs_num = ring->ring_size - 1;
        rc = ena_populate_rx_queue(ring, bufs_num);
        if (rc != bufs_num) {
+               ena_com_destroy_io_queue(&ring->adapter->ena_dev,
+                                        ENA_IO_RXQ_IDX(ring->id));
                PMD_INIT_LOG(ERR, "Failed to populate rx ring !");
                return ENA_COM_FAULT;
        }
@@ -1111,17 +1283,10 @@ static int ena_tx_queue_setup(struct rte_eth_dev *dev,
                              __rte_unused unsigned int socket_id,
                              const struct rte_eth_txconf *tx_conf)
 {
-       struct ena_com_create_io_ctx ctx =
-               /* policy set to _HOST just to satisfy icc compiler */
-               { ENA_ADMIN_PLACEMENT_POLICY_HOST,
-                 ENA_COM_IO_QUEUE_DIRECTION_TX, 0, 0, 0, 0 };
        struct ena_ring *txq = NULL;
        struct ena_adapter *adapter =
                (struct ena_adapter *)(dev->data->dev_private);
        unsigned int i;
-       int ena_qid;
-       int rc;
-       struct ena_com_dev *ena_dev = &adapter->ena_dev;
 
        txq = &adapter->tx_ring[queue_idx];
 
@@ -1146,36 +1311,8 @@ static int ena_tx_queue_setup(struct rte_eth_dev *dev,
                return -EINVAL;
        }
 
-       ena_qid = ENA_IO_TXQ_IDX(queue_idx);
-
-       ctx.direction = ENA_COM_IO_QUEUE_DIRECTION_TX;
-       ctx.qid = ena_qid;
-       ctx.msix_vector = -1; /* admin interrupts not used */
-       ctx.mem_queue_type = ena_dev->tx_mem_queue_type;
-       ctx.queue_size = adapter->tx_ring_size;
-       ctx.numa_node = ena_cpu_to_node(queue_idx);
-
-       rc = ena_com_create_io_queue(ena_dev, &ctx);
-       if (rc) {
-               RTE_LOG(ERR, PMD,
-                       "failed to create io TX queue #%d (qid:%d) rc: %d\n",
-                       queue_idx, ena_qid, rc);
-               return rc;
-       }
-       txq->ena_com_io_cq = &ena_dev->io_cq_queues[ena_qid];
-       txq->ena_com_io_sq = &ena_dev->io_sq_queues[ena_qid];
-
-       rc = ena_com_get_io_handlers(ena_dev, ena_qid,
-                                    &txq->ena_com_io_sq,
-                                    &txq->ena_com_io_cq);
-       if (rc) {
-               RTE_LOG(ERR, PMD,
-                       "Failed to get TX queue handlers. TX queue num %d rc: %d\n",
-                       queue_idx, rc);
-               goto err_destroy_io_queue;
-       }
-
-       ena_com_update_numa_node(txq->ena_com_io_cq, ctx.numa_node);
+       if (nb_desc == RTE_ETH_DEV_FALLBACK_TX_RINGSIZE)
+               nb_desc = adapter->tx_ring_size;
 
        txq->port_id = dev->data->port_id;
        txq->next_to_clean = 0;
@@ -1188,8 +1325,7 @@ static int ena_tx_queue_setup(struct rte_eth_dev *dev,
                                          RTE_CACHE_LINE_SIZE);
        if (!txq->tx_buffer_info) {
                RTE_LOG(ERR, PMD, "failed to alloc mem for tx buffer info\n");
-               rc = -ENOMEM;
-               goto err_destroy_io_queue;
+               return -ENOMEM;
        }
 
        txq->empty_tx_reqs = rte_zmalloc("txq->empty_tx_reqs",
@@ -1197,8 +1333,19 @@ static int ena_tx_queue_setup(struct rte_eth_dev *dev,
                                         RTE_CACHE_LINE_SIZE);
        if (!txq->empty_tx_reqs) {
                RTE_LOG(ERR, PMD, "failed to alloc mem for tx reqs\n");
-               rc = -ENOMEM;
-               goto err_free;
+               rte_free(txq->tx_buffer_info);
+               return -ENOMEM;
+       }
+
+       txq->push_buf_intermediate_buf =
+               rte_zmalloc("txq->push_buf_intermediate_buf",
+                           txq->tx_max_header_size,
+                           RTE_CACHE_LINE_SIZE);
+       if (!txq->push_buf_intermediate_buf) {
+               RTE_LOG(ERR, PMD, "failed to alloc push buff for LLQ\n");
+               rte_free(txq->tx_buffer_info);
+               rte_free(txq->empty_tx_reqs);
+               return -ENOMEM;
        }
 
        for (i = 0; i < txq->ring_size; i++)
@@ -1214,13 +1361,6 @@ static int ena_tx_queue_setup(struct rte_eth_dev *dev,
        dev->data->tx_queues[queue_idx] = txq;
 
        return 0;
-
-err_free:
-       rte_free(txq->tx_buffer_info);
-
-err_destroy_io_queue:
-       ena_com_destroy_io_queue(ena_dev, ena_qid);
-       return rc;
 }
 
 static int ena_rx_queue_setup(struct rte_eth_dev *dev,
@@ -1230,16 +1370,10 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
                              __rte_unused const struct rte_eth_rxconf *rx_conf,
                              struct rte_mempool *mp)
 {
-       struct ena_com_create_io_ctx ctx =
-               /* policy set to _HOST just to satisfy icc compiler */
-               { ENA_ADMIN_PLACEMENT_POLICY_HOST,
-                 ENA_COM_IO_QUEUE_DIRECTION_RX, 0, 0, 0, 0 };
        struct ena_adapter *adapter =
                (struct ena_adapter *)(dev->data->dev_private);
        struct ena_ring *rxq = NULL;
-       uint16_t ena_qid = 0;
-       int i, rc = 0;
-       struct ena_com_dev *ena_dev = &adapter->ena_dev;
+       int i;
 
        rxq = &adapter->rx_ring[queue_idx];
        if (rxq->configured) {
@@ -1249,6 +1383,9 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
                return ENA_COM_FAULT;
        }
 
+       if (nb_desc == RTE_ETH_DEV_FALLBACK_RX_RINGSIZE)
+               nb_desc = adapter->rx_ring_size;
+
        if (!rte_is_power_of_2(nb_desc)) {
                RTE_LOG(ERR, PMD,
                        "Unsupported size of RX queue: %d is not a power of 2.",
@@ -1263,36 +1400,6 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
                return -EINVAL;
        }
 
-       ena_qid = ENA_IO_RXQ_IDX(queue_idx);
-
-       ctx.qid = ena_qid;
-       ctx.direction = ENA_COM_IO_QUEUE_DIRECTION_RX;
-       ctx.mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
-       ctx.msix_vector = -1; /* admin interrupts not used */
-       ctx.queue_size = adapter->rx_ring_size;
-       ctx.numa_node = ena_cpu_to_node(queue_idx);
-
-       rc = ena_com_create_io_queue(ena_dev, &ctx);
-       if (rc) {
-               RTE_LOG(ERR, PMD, "failed to create io RX queue #%d rc: %d\n",
-                       queue_idx, rc);
-               return rc;
-       }
-
-       rxq->ena_com_io_cq = &ena_dev->io_cq_queues[ena_qid];
-       rxq->ena_com_io_sq = &ena_dev->io_sq_queues[ena_qid];
-
-       rc = ena_com_get_io_handlers(ena_dev, ena_qid,
-                                    &rxq->ena_com_io_sq,
-                                    &rxq->ena_com_io_cq);
-       if (rc) {
-               RTE_LOG(ERR, PMD,
-                       "Failed to get RX queue handlers. RX queue num %d rc: %d\n",
-                       queue_idx, rc);
-               ena_com_destroy_io_queue(ena_dev, ena_qid);
-               return rc;
-       }
-
        rxq->port_id = dev->data->port_id;
        rxq->next_to_clean = 0;
        rxq->next_to_use = 0;
@@ -1304,7 +1411,17 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
                                          RTE_CACHE_LINE_SIZE);
        if (!rxq->rx_buffer_info) {
                RTE_LOG(ERR, PMD, "failed to alloc mem for rx buffer info\n");
-               ena_com_destroy_io_queue(ena_dev, ena_qid);
+               return -ENOMEM;
+       }
+
+       rxq->rx_refill_buffer = rte_zmalloc("rxq->rx_refill_buffer",
+                                           sizeof(struct rte_mbuf *) * nb_desc,
+                                           RTE_CACHE_LINE_SIZE);
+
+       if (!rxq->rx_refill_buffer) {
+               RTE_LOG(ERR, PMD, "failed to alloc mem for rx refill buffer\n");
+               rte_free(rxq->rx_buffer_info);
+               rxq->rx_buffer_info = NULL;
                return -ENOMEM;
        }
 
@@ -1315,7 +1432,8 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
                RTE_LOG(ERR, PMD, "failed to alloc mem for empty rx reqs\n");
                rte_free(rxq->rx_buffer_info);
                rxq->rx_buffer_info = NULL;
-               ena_com_destroy_io_queue(ena_dev, ena_qid);
+               rte_free(rxq->rx_refill_buffer);
+               rxq->rx_refill_buffer = NULL;
                return -ENOMEM;
        }
 
@@ -1326,7 +1444,7 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
        rxq->configured = 1;
        dev->data->rx_queues[queue_idx] = rxq;
 
-       return rc;
+       return 0;
 }
 
 static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
@@ -1337,7 +1455,7 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
        uint16_t ring_mask = ring_size - 1;
        uint16_t next_to_use = rxq->next_to_use;
        uint16_t in_use, req_id;
-       struct rte_mbuf **mbufs = &rxq->rx_buffer_info[0];
+       struct rte_mbuf **mbufs = rxq->rx_refill_buffer;
 
        if (unlikely(!count))
                return 0;
@@ -1345,13 +1463,8 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
        in_use = rxq->next_to_use - rxq->next_to_clean;
        ena_assert_msg(((in_use + count) < ring_size), "bad ring state");
 
-       count = RTE_MIN(count,
-                       (uint16_t)(ring_size - (next_to_use & ring_mask)));
-
        /* get resources for incoming packets */
-       rc = rte_mempool_get_bulk(rxq->mb_pool,
-                                 (void **)(&mbufs[next_to_use & ring_mask]),
-                                 count);
+       rc = rte_mempool_get_bulk(rxq->mb_pool, (void **)mbufs, count);
        if (unlikely(rc < 0)) {
                rte_atomic64_inc(&rxq->adapter->drv_stats->rx_nombuf);
                PMD_RX_LOG(DEBUG, "there are no enough free buffers");
@@ -1360,15 +1473,17 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 
        for (i = 0; i < count; i++) {
                uint16_t next_to_use_masked = next_to_use & ring_mask;
-               struct rte_mbuf *mbuf = mbufs[next_to_use_masked];
+               struct rte_mbuf *mbuf = mbufs[i];
                struct ena_com_buf ebuf;
 
-               rte_prefetch0(mbufs[((next_to_use + 4) & ring_mask)]);
+               if (likely((i + 4) < count))
+                       rte_prefetch0(mbufs[i + 4]);
 
                req_id = rxq->empty_rx_reqs[next_to_use_masked];
                rc = validate_rx_req_id(rxq, req_id);
                if (unlikely(rc < 0))
                        break;
+               rxq->rx_buffer_info[req_id] = mbuf;
 
                /* prepare physical address for DMA transaction */
                ebuf.paddr = mbuf->buf_iova + RTE_PKTMBUF_HEADROOM;
@@ -1377,17 +1492,19 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
                rc = ena_com_add_single_rx_desc(rxq->ena_com_io_sq,
                                                &ebuf, req_id);
                if (unlikely(rc)) {
-                       rte_mempool_put_bulk(rxq->mb_pool, (void **)(&mbuf),
-                                            count - i);
                        RTE_LOG(WARNING, PMD, "failed adding rx desc\n");
+                       rxq->rx_buffer_info[req_id] = NULL;
                        break;
                }
                next_to_use++;
        }
 
-       if (unlikely(i < count))
+       if (unlikely(i < count)) {
                RTE_LOG(WARNING, PMD, "refilled rx qid %d with only %d "
                        "buffers (from %d)\n", rxq->id, i, count);
+               rte_mempool_put_bulk(rxq->mb_pool, (void **)(&mbufs[i]),
+                                    count - i);
+       }
 
        /* When we submitted free recources to device... */
        if (likely(i > 0)) {
@@ -1445,7 +1562,7 @@ static int ena_device_init(struct ena_com_dev *ena_dev,
        ena_dev->dma_addr_bits = ena_com_get_dma_width(ena_dev);
 
        /* ENA device administration layer init */
-       rc = ena_com_admin_init(ena_dev, &aenq_handlers, true);
+       rc = ena_com_admin_init(ena_dev, &aenq_handlers);
        if (rc) {
                RTE_LOG(ERR, PMD,
                        "cannot initialize ena admin queue with device\n");
@@ -1546,15 +1663,87 @@ static void ena_timer_wd_callback(__rte_unused struct rte_timer *timer,
        }
 }
 
-static int ena_calc_io_queue_num(__rte_unused struct ena_com_dev *ena_dev,
+static inline void
+set_default_llq_configurations(struct ena_llq_configurations *llq_config)
+{
+       llq_config->llq_header_location = ENA_ADMIN_INLINE_HEADER;
+       llq_config->llq_ring_entry_size = ENA_ADMIN_LIST_ENTRY_SIZE_128B;
+       llq_config->llq_stride_ctrl = ENA_ADMIN_MULTIPLE_DESCS_PER_ENTRY;
+       llq_config->llq_num_decs_before_header =
+               ENA_ADMIN_LLQ_NUM_DESCS_BEFORE_HEADER_2;
+       llq_config->llq_ring_entry_size_value = 128;
+}
+
+static int
+ena_set_queues_placement_policy(struct ena_adapter *adapter,
+                               struct ena_com_dev *ena_dev,
+                               struct ena_admin_feature_llq_desc *llq,
+                               struct ena_llq_configurations *llq_default_configurations)
+{
+       int rc;
+       u32 llq_feature_mask;
+
+       llq_feature_mask = 1 << ENA_ADMIN_LLQ;
+       if (!(ena_dev->supported_features & llq_feature_mask)) {
+               RTE_LOG(INFO, PMD,
+                       "LLQ is not supported. Fallback to host mode policy.\n");
+               ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+               return 0;
+       }
+
+       rc = ena_com_config_dev_mode(ena_dev, llq, llq_default_configurations);
+       if (unlikely(rc)) {
+               PMD_INIT_LOG(WARNING, "Failed to config dev mode. "
+                       "Fallback to host mode policy.\n");
+               ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+               return 0;
+       }
+
+       /* Nothing to config, exit */
+       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_HOST)
+               return 0;
+
+       if (!adapter->dev_mem_base) {
+               RTE_LOG(ERR, PMD, "Unable to access LLQ bar resource. "
+                       "Fallback to host mode policy.\n.");
+               ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+               return 0;
+       }
+
+       ena_dev->mem_bar = adapter->dev_mem_base;
+
+       return 0;
+}
+
+static int ena_calc_io_queue_num(struct ena_com_dev *ena_dev,
                                 struct ena_com_dev_get_features_ctx *get_feat_ctx)
 {
-       int io_sq_num, io_cq_num, io_queue_num;
+       uint32_t io_tx_sq_num, io_tx_cq_num, io_rx_num, io_queue_num;
+
+       /* Regular queues capabilities */
+       if (ena_dev->supported_features & BIT(ENA_ADMIN_MAX_QUEUES_EXT)) {
+               struct ena_admin_queue_ext_feature_fields *max_queue_ext =
+                       &get_feat_ctx->max_queue_ext.max_queue_ext;
+               io_rx_num = RTE_MIN(max_queue_ext->max_rx_sq_num,
+                                   max_queue_ext->max_rx_cq_num);
+               io_tx_sq_num = max_queue_ext->max_tx_sq_num;
+               io_tx_cq_num = max_queue_ext->max_tx_cq_num;
+       } else {
+               struct ena_admin_queue_feature_desc *max_queues =
+                       &get_feat_ctx->max_queues;
+               io_tx_sq_num = max_queues->max_sq_num;
+               io_tx_cq_num = max_queues->max_cq_num;
+               io_rx_num = RTE_MIN(io_tx_sq_num, io_tx_cq_num);
+       }
 
-       io_sq_num = get_feat_ctx->max_queues.max_sq_num;
-       io_cq_num = get_feat_ctx->max_queues.max_cq_num;
+       /* In case of LLQ use the llq number in the get feature cmd */
+       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV)
+               io_tx_sq_num = get_feat_ctx->llq.max_llq_num;
 
-       io_queue_num = RTE_MIN(io_sq_num, io_cq_num);
+       io_queue_num = RTE_MIN(rte_lcore_count(), ENA_MAX_NUM_IO_QUEUES);
+       io_queue_num = RTE_MIN(io_queue_num, io_rx_num);
+       io_queue_num = RTE_MIN(io_queue_num, io_tx_sq_num);
+       io_queue_num = RTE_MIN(io_queue_num, io_tx_cq_num);
 
        if (unlikely(io_queue_num == 0)) {
                RTE_LOG(ERR, PMD, "Number of IO queues should not be 0\n");
@@ -1566,14 +1755,16 @@ static int ena_calc_io_queue_num(__rte_unused struct ena_com_dev *ena_dev,
 
 static int eth_ena_dev_init(struct rte_eth_dev *eth_dev)
 {
+       struct ena_calc_queue_size_ctx calc_queue_ctx = { 0 };
        struct rte_pci_device *pci_dev;
        struct rte_intr_handle *intr_handle;
        struct ena_adapter *adapter =
                (struct ena_adapter *)(eth_dev->data->dev_private);
        struct ena_com_dev *ena_dev = &adapter->ena_dev;
        struct ena_com_dev_get_features_ctx get_feat_ctx;
-       int queue_size, rc;
-       u16 tx_sgl_size = 0;
+       struct ena_llq_configurations llq_config;
+       const char *queue_type_str;
+       int rc;
 
        static int adapters_found;
        bool wd_state;
@@ -1627,20 +1818,36 @@ static int eth_ena_dev_init(struct rte_eth_dev *eth_dev)
        }
        adapter->wd_state = wd_state;
 
-       ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+       set_default_llq_configurations(&llq_config);
+       rc = ena_set_queues_placement_policy(adapter, ena_dev,
+                                            &get_feat_ctx.llq, &llq_config);
+       if (unlikely(rc)) {
+               PMD_INIT_LOG(CRIT, "Failed to set placement policy");
+               return rc;
+       }
+
+       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_HOST)
+               queue_type_str = "Regular";
+       else
+               queue_type_str = "Low latency";
+       RTE_LOG(INFO, PMD, "Placement policy: %s\n", queue_type_str);
+
+       calc_queue_ctx.ena_dev = ena_dev;
+       calc_queue_ctx.get_feat_ctx = &get_feat_ctx;
        adapter->num_queues = ena_calc_io_queue_num(ena_dev,
                                                    &get_feat_ctx);
 
-       queue_size = ena_calc_queue_size(ena_dev, &tx_sgl_size, &get_feat_ctx);
-       if (queue_size <= 0 || adapter->num_queues <= 0) {
+       rc = ena_calc_queue_size(&calc_queue_ctx);
+       if (unlikely((rc != 0) || (adapter->num_queues <= 0))) {
                rc = -EFAULT;
                goto err_device_destroy;
        }
 
-       adapter->tx_ring_size = queue_size;
-       adapter->rx_ring_size = queue_size;
+       adapter->tx_ring_size = calc_queue_ctx.tx_queue_size;
+       adapter->rx_ring_size = calc_queue_ctx.rx_queue_size;
 
-       adapter->max_tx_sgl_size = tx_sgl_size;
+       adapter->max_tx_sgl_size = calc_queue_ctx.max_tx_sgl_size;
+       adapter->max_rx_sgl_size = calc_queue_ctx.max_rx_sgl_size;
 
        /* prepare ring structures */
        ena_init_rings(adapter);
@@ -1659,6 +1866,12 @@ static int eth_ena_dev_init(struct rte_eth_dev *eth_dev)
        ether_addr_copy((struct ether_addr *)get_feat_ctx.dev_attr.mac_addr,
                        (struct ether_addr *)adapter->mac_addr);
 
+       /*
+        * Pass the information to the rte_eth_dev_close() that it should also
+        * release the private port resources.
+        */
+       eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
+
        adapter->drv_stats = rte_zmalloc("adapter stats",
                                         sizeof(*adapter->drv_stats),
                                         RTE_CACHE_LINE_SIZE);
@@ -1697,13 +1910,11 @@ err:
 
 static int eth_ena_dev_uninit(struct rte_eth_dev *eth_dev)
 {
-       struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
-       struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
        struct ena_adapter *adapter =
                (struct ena_adapter *)(eth_dev->data->dev_private);
 
        if (rte_eal_process_type() != RTE_PROC_PRIMARY)
-               return -EPERM;
+               return 0;
 
        if (adapter->state != ENA_ADAPTER_STATE_CLOSED)
                ena_close(eth_dev);
@@ -1713,14 +1924,6 @@ static int eth_ena_dev_uninit(struct rte_eth_dev *eth_dev)
        eth_dev->tx_pkt_burst = NULL;
        eth_dev->tx_pkt_prepare = NULL;
 
-       rte_free(adapter->drv_stats);
-       adapter->drv_stats = NULL;
-
-       rte_intr_disable(intr_handle);
-       rte_intr_callback_unregister(intr_handle,
-                                    ena_interrupt_handler_rte,
-                                    adapter);
-
        adapter->state = ENA_ADAPTER_STATE_FREE;
 
        return 0;
@@ -1761,6 +1964,7 @@ static void ena_init_rings(struct ena_adapter *adapter)
                ring->type = ENA_RING_TYPE_RX;
                ring->adapter = adapter;
                ring->id = i;
+               ring->sgl_size = adapter->max_rx_sgl_size;
        }
 }
 
@@ -1833,15 +2037,19 @@ static void ena_infos_get(struct rte_eth_dev *dev,
        adapter->tx_supported_offloads = tx_feat;
        adapter->rx_supported_offloads = rx_feat;
 
-       dev_info->rx_desc_lim.nb_max = ENA_MAX_RING_DESC;
+       dev_info->rx_desc_lim.nb_max = adapter->rx_ring_size;
        dev_info->rx_desc_lim.nb_min = ENA_MIN_RING_DESC;
+       dev_info->rx_desc_lim.nb_seg_max = RTE_MIN(ENA_PKT_MAX_BUFS,
+                                       adapter->max_rx_sgl_size);
+       dev_info->rx_desc_lim.nb_mtu_seg_max = RTE_MIN(ENA_PKT_MAX_BUFS,
+                                       adapter->max_rx_sgl_size);
 
-       dev_info->tx_desc_lim.nb_max = ENA_MAX_RING_DESC;
+       dev_info->tx_desc_lim.nb_max = adapter->tx_ring_size;
        dev_info->tx_desc_lim.nb_min = ENA_MIN_RING_DESC;
        dev_info->tx_desc_lim.nb_seg_max = RTE_MIN(ENA_PKT_MAX_BUFS,
-                                       feat.max_queues.max_packet_tx_descs);
+                                       adapter->max_tx_sgl_size);
        dev_info->tx_desc_lim.nb_mtu_seg_max = RTE_MIN(ENA_PKT_MAX_BUFS,
-                                       feat.max_queues.max_packet_tx_descs);
+                                       adapter->max_tx_sgl_size);
 }
 
 static uint16_t eth_ena_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
@@ -1877,7 +2085,7 @@ static uint16_t eth_ena_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
        for (completed = 0; completed < nb_pkts; completed++) {
                int segments = 0;
 
-               ena_rx_ctx.max_bufs = rx_ring->ring_size;
+               ena_rx_ctx.max_bufs = rx_ring->sgl_size;
                ena_rx_ctx.ena_bufs = rx_ring->ena_bufs;
                ena_rx_ctx.descs = 0;
                /* receive packet context */
@@ -1886,6 +2094,8 @@ static uint16_t eth_ena_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                                    &ena_rx_ctx);
                if (unlikely(rc)) {
                        RTE_LOG(ERR, PMD, "ena_com_rx_pkt error %d\n", rc);
+                       rx_ring->adapter->reset_reason =
+                               ENA_REGS_RESET_TOO_MANY_RX_DESCS;
                        rx_ring->adapter->trigger_reset = true;
                        return 0;
                }
@@ -1921,10 +2131,12 @@ static uint16_t eth_ena_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                        segments++;
                        next_to_clean++;
                }
+               if (unlikely(rc))
+                       break;
 
                /* fill mbuf attributes if any */
                ena_rx_mbuf_prepare(mbuf_head, &ena_rx_ctx);
-               mbuf_head->hash.rss = (uint32_t)rx_ring->id;
+               mbuf_head->hash.rss = ena_rx_ctx.hash;
 
                /* pass to DPDK application head mbuf */
                rx_pkts[recv_idx] = mbuf_head;
@@ -2037,13 +2249,21 @@ static void ena_update_hints(struct ena_adapter *adapter,
 static int ena_check_and_linearize_mbuf(struct ena_ring *tx_ring,
                                        struct rte_mbuf *mbuf)
 {
-       int num_segments, rc;
+       struct ena_com_dev *ena_dev;
+       int num_segments, header_len, rc;
 
+       ena_dev = &tx_ring->adapter->ena_dev;
        num_segments = mbuf->nb_segs;
+       header_len = mbuf->data_len;
 
        if (likely(num_segments < tx_ring->sgl_size))
                return 0;
 
+       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV &&
+           (num_segments == tx_ring->sgl_size) &&
+           (header_len < tx_ring->tx_max_header_size))
+               return 0;
+
        rc = rte_pktmbuf_linearize(mbuf);
        if (unlikely(rc))
                RTE_LOG(WARNING, PMD, "Mbuf linearize failed\n");
@@ -2058,6 +2278,7 @@ static uint16_t eth_ena_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
        uint16_t next_to_use = tx_ring->next_to_use;
        uint16_t next_to_clean = tx_ring->next_to_clean;
        struct rte_mbuf *mbuf;
+       uint16_t seg_len;
        unsigned int ring_size = tx_ring->ring_size;
        unsigned int ring_mask = ring_size - 1;
        struct ena_com_tx_ctx ena_tx_ctx;
@@ -2065,6 +2286,8 @@ static uint16_t eth_ena_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
        struct ena_com_buf *ebuf;
        uint16_t rc, req_id, total_tx_descs = 0;
        uint16_t sent_idx = 0, empty_tx_reqs;
+       uint16_t push_len = 0;
+       uint16_t delta = 0;
        int nb_hw_desc;
 
        /* Check adapter state */
@@ -2097,17 +2320,32 @@ static uint16_t eth_ena_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
                       sizeof(struct ena_com_tx_meta));
                ena_tx_ctx.ena_bufs = ebuf;
                ena_tx_ctx.req_id = req_id;
+
+               delta = 0;
+               seg_len = mbuf->data_len;
+
                if (tx_ring->tx_mem_queue_type ==
                                ENA_ADMIN_PLACEMENT_POLICY_DEV) {
-                       /* prepare the push buffer with
-                        * virtual address of the data
-                        */
-                       ena_tx_ctx.header_len =
-                               RTE_MIN(mbuf->data_len,
-                                       tx_ring->tx_max_header_size);
-                       ena_tx_ctx.push_header =
-                               (void *)((char *)mbuf->buf_addr +
-                                        mbuf->data_off);
+                       push_len = RTE_MIN(mbuf->pkt_len,
+                                          tx_ring->tx_max_header_size);
+                       ena_tx_ctx.header_len = push_len;
+
+                       if (likely(push_len <= seg_len)) {
+                               /* If the push header is in the single segment,
+                                * then just point it to the 1st mbuf data.
+                                */
+                               ena_tx_ctx.push_header =
+                                       rte_pktmbuf_mtod(mbuf, uint8_t *);
+                       } else {
+                               /* If the push header lays in the several
+                                * segments, copy it to the intermediate buffer.
+                                */
+                               rte_pktmbuf_read(mbuf, 0, push_len,
+                                       tx_ring->push_buf_intermediate_buf);
+                               ena_tx_ctx.push_header =
+                                       tx_ring->push_buf_intermediate_buf;
+                               delta = push_len - seg_len;
+                       }
                } /* there's no else as we take advantage of memset zeroing */
 
                /* Set TX offloads flags, if applicable */
@@ -2122,25 +2360,44 @@ static uint16_t eth_ena_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
                /* Process first segment taking into
                 * consideration pushed header
                 */
-               if (mbuf->data_len > ena_tx_ctx.header_len) {
+               if (seg_len > push_len) {
                        ebuf->paddr = mbuf->buf_iova +
                                      mbuf->data_off +
-                                     ena_tx_ctx.header_len;
-                       ebuf->len = mbuf->data_len - ena_tx_ctx.header_len;
+                                     push_len;
+                       ebuf->len = seg_len - push_len;
                        ebuf++;
                        tx_info->num_of_bufs++;
                }
 
                while ((mbuf = mbuf->next) != NULL) {
-                       ebuf->paddr = mbuf->buf_iova + mbuf->data_off;
-                       ebuf->len = mbuf->data_len;
+                       seg_len = mbuf->data_len;
+
+                       /* Skip mbufs if whole data is pushed as a header */
+                       if (unlikely(delta > seg_len)) {
+                               delta -= seg_len;
+                               continue;
+                       }
+
+                       ebuf->paddr = mbuf->buf_iova + mbuf->data_off + delta;
+                       ebuf->len = seg_len - delta;
                        ebuf++;
                        tx_info->num_of_bufs++;
+
+                       delta = 0;
                }
 
                ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
 
-               /* Write data to device */
+               if (ena_com_is_doorbell_needed(tx_ring->ena_com_io_sq,
+                                              &ena_tx_ctx)) {
+                       RTE_LOG(DEBUG, PMD, "llq tx max burst size of queue %d"
+                               " achieved, writing doorbell to send burst\n",
+                               tx_ring->id);
+                       rte_wmb();
+                       ena_com_write_sq_doorbell(tx_ring->ena_com_io_sq);
+               }
+
+               /* prepare the packet's descriptors to dma engine */
                rc = ena_com_prepare_tx(tx_ring->ena_com_io_sq,
                                        &ena_tx_ctx, &nb_hw_desc);
                if (unlikely(rc))
@@ -2220,9 +2477,7 @@ RTE_PMD_REGISTER_PCI(net_ena, rte_ena_pmd);
 RTE_PMD_REGISTER_PCI_TABLE(net_ena, pci_id_ena_map);
 RTE_PMD_REGISTER_KMOD_DEP(net_ena, "* igb_uio | uio_pci_generic | vfio-pci");
 
-RTE_INIT(ena_init_log);
-static void
-ena_init_log(void)
+RTE_INIT(ena_init_log)
 {
        ena_logtype_init = rte_log_register("pmd.net.ena.init");
        if (ena_logtype_init >= 0)