net: ena: add functions for handling Low Latency Queues in ena_netdev
authorArthur Kiyanovski <akiyano@amazon.com>
Thu, 11 Oct 2018 08:26:20 +0000 (11:26 +0300)
committerDavid S. Miller <davem@davemloft.net>
Thu, 11 Oct 2018 17:13:51 +0000 (10:13 -0700)
This patch includes all code changes necessary in ena_netdev to enable
packet sending via the LLQ placemnt mode.

Signed-off-by: Arthur Kiyanovski <akiyano@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/amazon/ena/ena_ethtool.c
drivers/net/ethernet/amazon/ena/ena_netdev.c
drivers/net/ethernet/amazon/ena/ena_netdev.h

index 521607bc43937f0954fb20414c5b1002123c244d..fd28bd0d1c1ed7b773ae9b49360d39d11d1f3ab6 100644 (file)
@@ -81,6 +81,7 @@ static const struct ena_stats ena_stats_tx_strings[] = {
        ENA_STAT_TX_ENTRY(doorbells),
        ENA_STAT_TX_ENTRY(prepare_ctx_err),
        ENA_STAT_TX_ENTRY(bad_req_id),
+       ENA_STAT_TX_ENTRY(llq_buffer_copy),
        ENA_STAT_TX_ENTRY(missed_tx),
 };
 
index 79a4e87326b7e6fc13713d6665ac3da53fd54584..98314b28a000fcd1578eb7df3896c282bf1840f3 100644 (file)
@@ -237,6 +237,17 @@ static int ena_setup_tx_resources(struct ena_adapter *adapter, int qid)
                }
        }
 
+       size = tx_ring->tx_max_header_size;
+       tx_ring->push_buf_intermediate_buf = vzalloc_node(size, node);
+       if (!tx_ring->push_buf_intermediate_buf) {
+               tx_ring->push_buf_intermediate_buf = vzalloc(size);
+               if (!tx_ring->push_buf_intermediate_buf) {
+                       vfree(tx_ring->tx_buffer_info);
+                       vfree(tx_ring->free_tx_ids);
+                       return -ENOMEM;
+               }
+       }
+
        /* Req id ring for TX out of order completions */
        for (i = 0; i < tx_ring->ring_size; i++)
                tx_ring->free_tx_ids[i] = i;
@@ -265,6 +276,9 @@ static void ena_free_tx_resources(struct ena_adapter *adapter, int qid)
 
        vfree(tx_ring->free_tx_ids);
        tx_ring->free_tx_ids = NULL;
+
+       vfree(tx_ring->push_buf_intermediate_buf);
+       tx_ring->push_buf_intermediate_buf = NULL;
 }
 
 /* ena_setup_all_tx_resources - allocate I/O Tx queues resources for All queues
@@ -602,6 +616,36 @@ static void ena_free_all_rx_bufs(struct ena_adapter *adapter)
                ena_free_rx_bufs(adapter, i);
 }
 
+static inline void ena_unmap_tx_skb(struct ena_ring *tx_ring,
+                                   struct ena_tx_buffer *tx_info)
+{
+       struct ena_com_buf *ena_buf;
+       u32 cnt;
+       int i;
+
+       ena_buf = tx_info->bufs;
+       cnt = tx_info->num_of_bufs;
+
+       if (unlikely(!cnt))
+               return;
+
+       if (tx_info->map_linear_data) {
+               dma_unmap_single(tx_ring->dev,
+                                dma_unmap_addr(ena_buf, paddr),
+                                dma_unmap_len(ena_buf, len),
+                                DMA_TO_DEVICE);
+               ena_buf++;
+               cnt--;
+       }
+
+       /* unmap remaining mapped pages */
+       for (i = 0; i < cnt; i++) {
+               dma_unmap_page(tx_ring->dev, dma_unmap_addr(ena_buf, paddr),
+                              dma_unmap_len(ena_buf, len), DMA_TO_DEVICE);
+               ena_buf++;
+       }
+}
+
 /* ena_free_tx_bufs - Free Tx Buffers per Queue
  * @tx_ring: TX ring for which buffers be freed
  */
@@ -612,9 +656,6 @@ static void ena_free_tx_bufs(struct ena_ring *tx_ring)
 
        for (i = 0; i < tx_ring->ring_size; i++) {
                struct ena_tx_buffer *tx_info = &tx_ring->tx_buffer_info[i];
-               struct ena_com_buf *ena_buf;
-               int nr_frags;
-               int j;
 
                if (!tx_info->skb)
                        continue;
@@ -630,21 +671,7 @@ static void ena_free_tx_bufs(struct ena_ring *tx_ring)
                                   tx_ring->qid, i);
                }
 
-               ena_buf = tx_info->bufs;
-               dma_unmap_single(tx_ring->dev,
-                                ena_buf->paddr,
-                                ena_buf->len,
-                                DMA_TO_DEVICE);
-
-               /* unmap remaining mapped pages */
-               nr_frags = tx_info->num_of_bufs - 1;
-               for (j = 0; j < nr_frags; j++) {
-                       ena_buf++;
-                       dma_unmap_page(tx_ring->dev,
-                                      ena_buf->paddr,
-                                      ena_buf->len,
-                                      DMA_TO_DEVICE);
-               }
+               ena_unmap_tx_skb(tx_ring, tx_info);
 
                dev_kfree_skb_any(tx_info->skb);
        }
@@ -735,8 +762,6 @@ static int ena_clean_tx_irq(struct ena_ring *tx_ring, u32 budget)
        while (tx_pkts < budget) {
                struct ena_tx_buffer *tx_info;
                struct sk_buff *skb;
-               struct ena_com_buf *ena_buf;
-               int i, nr_frags;
 
                rc = ena_com_tx_comp_req_id_get(tx_ring->ena_com_io_cq,
                                                &req_id);
@@ -756,24 +781,7 @@ static int ena_clean_tx_irq(struct ena_ring *tx_ring, u32 budget)
                tx_info->skb = NULL;
                tx_info->last_jiffies = 0;
 
-               if (likely(tx_info->num_of_bufs != 0)) {
-                       ena_buf = tx_info->bufs;
-
-                       dma_unmap_single(tx_ring->dev,
-                                        dma_unmap_addr(ena_buf, paddr),
-                                        dma_unmap_len(ena_buf, len),
-                                        DMA_TO_DEVICE);
-
-                       /* unmap remaining mapped pages */
-                       nr_frags = tx_info->num_of_bufs - 1;
-                       for (i = 0; i < nr_frags; i++) {
-                               ena_buf++;
-                               dma_unmap_page(tx_ring->dev,
-                                              dma_unmap_addr(ena_buf, paddr),
-                                              dma_unmap_len(ena_buf, len),
-                                              DMA_TO_DEVICE);
-                       }
-               }
+               ena_unmap_tx_skb(tx_ring, tx_info);
 
                netif_dbg(tx_ring->adapter, tx_done, tx_ring->netdev,
                          "tx_poll: q %d skb %p completed\n", tx_ring->qid,
@@ -1300,7 +1308,6 @@ static int ena_enable_msix(struct ena_adapter *adapter, int num_queues)
 
        /* Reserved the max msix vectors we might need */
        msix_vecs = ENA_MAX_MSIX_VEC(num_queues);
-
        netif_dbg(adapter, probe, adapter->netdev,
                  "trying to enable MSI-X, vectors %d\n", msix_vecs);
 
@@ -1593,7 +1600,7 @@ static int ena_up_complete(struct ena_adapter *adapter)
 
 static int ena_create_io_tx_queue(struct ena_adapter *adapter, int qid)
 {
-       struct ena_com_create_io_ctx ctx = { 0 };
+       struct ena_com_create_io_ctx ctx;
        struct ena_com_dev *ena_dev;
        struct ena_ring *tx_ring;
        u32 msix_vector;
@@ -1606,6 +1613,8 @@ static int ena_create_io_tx_queue(struct ena_adapter *adapter, int qid)
        msix_vector = ENA_IO_IRQ_IDX(qid);
        ena_qid = ENA_IO_TXQ_IDX(qid);
 
+       memset(&ctx, 0x0, sizeof(ctx));
+
        ctx.direction = ENA_COM_IO_QUEUE_DIRECTION_TX;
        ctx.qid = ena_qid;
        ctx.mem_queue_type = ena_dev->tx_mem_queue_type;
@@ -1659,7 +1668,7 @@ create_err:
 static int ena_create_io_rx_queue(struct ena_adapter *adapter, int qid)
 {
        struct ena_com_dev *ena_dev;
-       struct ena_com_create_io_ctx ctx = { 0 };
+       struct ena_com_create_io_ctx ctx;
        struct ena_ring *rx_ring;
        u32 msix_vector;
        u16 ena_qid;
@@ -1671,6 +1680,8 @@ static int ena_create_io_rx_queue(struct ena_adapter *adapter, int qid)
        msix_vector = ENA_IO_IRQ_IDX(qid);
        ena_qid = ENA_IO_RXQ_IDX(qid);
 
+       memset(&ctx, 0x0, sizeof(ctx));
+
        ctx.qid = ena_qid;
        ctx.direction = ENA_COM_IO_QUEUE_DIRECTION_RX;
        ctx.mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
@@ -1981,73 +1992,70 @@ static int ena_check_and_linearize_skb(struct ena_ring *tx_ring,
        return rc;
 }
 
-/* Called with netif_tx_lock. */
-static netdev_tx_t ena_start_xmit(struct sk_buff *skb, struct net_device *dev)
+static int ena_tx_map_skb(struct ena_ring *tx_ring,
+                         struct ena_tx_buffer *tx_info,
+                         struct sk_buff *skb,
+                         void **push_hdr,
+                         u16 *header_len)
 {
-       struct ena_adapter *adapter = netdev_priv(dev);
-       struct ena_tx_buffer *tx_info;
-       struct ena_com_tx_ctx ena_tx_ctx;
-       struct ena_ring *tx_ring;
-       struct netdev_queue *txq;
+       struct ena_adapter *adapter = tx_ring->adapter;
        struct ena_com_buf *ena_buf;
-       void *push_hdr;
-       u32 len, last_frag;
-       u16 next_to_use;
-       u16 req_id;
-       u16 push_len;
-       u16 header_len;
        dma_addr_t dma;
-       int qid, rc, nb_hw_desc;
-       int i = -1;
-
-       netif_dbg(adapter, tx_queued, dev, "%s skb %p\n", __func__, skb);
-       /*  Determine which tx ring we will be placed on */
-       qid = skb_get_queue_mapping(skb);
-       tx_ring = &adapter->tx_ring[qid];
-       txq = netdev_get_tx_queue(dev, qid);
-
-       rc = ena_check_and_linearize_skb(tx_ring, skb);
-       if (unlikely(rc))
-               goto error_drop_packet;
-
-       skb_tx_timestamp(skb);
-       len = skb_headlen(skb);
+       u32 skb_head_len, frag_len, last_frag;
+       u16 push_len = 0;
+       u16 delta = 0;
+       int i = 0;
 
-       next_to_use = tx_ring->next_to_use;
-       req_id = tx_ring->free_tx_ids[next_to_use];
-       tx_info = &tx_ring->tx_buffer_info[req_id];
-       tx_info->num_of_bufs = 0;
-
-       WARN(tx_info->skb, "SKB isn't NULL req_id %d\n", req_id);
-       ena_buf = tx_info->bufs;
+       skb_head_len = skb_headlen(skb);
        tx_info->skb = skb;
+       ena_buf = tx_info->bufs;
 
        if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
-               /* prepared the push buffer */
-               push_len = min_t(u32, len, tx_ring->tx_max_header_size);
-               header_len = push_len;
-               push_hdr = skb->data;
+               /* When the device is LLQ mode, the driver will copy
+                * the header into the device memory space.
+                * the ena_com layer assume the header is in a linear
+                * memory space.
+                * This assumption might be wrong since part of the header
+                * can be in the fragmented buffers.
+                * Use skb_header_pointer to make sure the header is in a
+                * linear memory space.
+                */
+
+               push_len = min_t(u32, skb->len, tx_ring->tx_max_header_size);
+               *push_hdr = skb_header_pointer(skb, 0, push_len,
+                                              tx_ring->push_buf_intermediate_buf);
+               *header_len = push_len;
+               if (unlikely(skb->data != *push_hdr)) {
+                       u64_stats_update_begin(&tx_ring->syncp);
+                       tx_ring->tx_stats.llq_buffer_copy++;
+                       u64_stats_update_end(&tx_ring->syncp);
+
+                       delta = push_len - skb_head_len;
+               }
        } else {
-               push_len = 0;
-               header_len = min_t(u32, len, tx_ring->tx_max_header_size);
-               push_hdr = NULL;
+               *push_hdr = NULL;
+               *header_len = min_t(u32, skb_head_len,
+                                   tx_ring->tx_max_header_size);
        }
 
-       netif_dbg(adapter, tx_queued, dev,
+       netif_dbg(adapter, tx_queued, adapter->netdev,
                  "skb: %p header_buf->vaddr: %p push_len: %d\n", skb,
-                 push_hdr, push_len);
+                 *push_hdr, push_len);
 
-       if (len > push_len) {
+       if (skb_head_len > push_len) {
                dma = dma_map_single(tx_ring->dev, skb->data + push_len,
-                                    len - push_len, DMA_TO_DEVICE);
-               if (dma_mapping_error(tx_ring->dev, dma))
+                                    skb_head_len - push_len, DMA_TO_DEVICE);
+               if (unlikely(dma_mapping_error(tx_ring->dev, dma)))
                        goto error_report_dma_error;
 
                ena_buf->paddr = dma;
-               ena_buf->len = len - push_len;
+               ena_buf->len = skb_head_len - push_len;
 
                ena_buf++;
                tx_info->num_of_bufs++;
+               tx_info->map_linear_data = 1;
+       } else {
+               tx_info->map_linear_data = 0;
        }
 
        last_frag = skb_shinfo(skb)->nr_frags;
@@ -2055,18 +2063,75 @@ static netdev_tx_t ena_start_xmit(struct sk_buff *skb, struct net_device *dev)
        for (i = 0; i < last_frag; i++) {
                const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 
-               len = skb_frag_size(frag);
-               dma = skb_frag_dma_map(tx_ring->dev, frag, 0, len,
-                                      DMA_TO_DEVICE);
-               if (dma_mapping_error(tx_ring->dev, dma))
+               frag_len = skb_frag_size(frag);
+
+               if (unlikely(delta >= frag_len)) {
+                       delta -= frag_len;
+                       continue;
+               }
+
+               dma = skb_frag_dma_map(tx_ring->dev, frag, delta,
+                                      frag_len - delta, DMA_TO_DEVICE);
+               if (unlikely(dma_mapping_error(tx_ring->dev, dma)))
                        goto error_report_dma_error;
 
                ena_buf->paddr = dma;
-               ena_buf->len = len;
+               ena_buf->len = frag_len - delta;
                ena_buf++;
+               tx_info->num_of_bufs++;
+               delta = 0;
        }
 
-       tx_info->num_of_bufs += last_frag;
+       return 0;
+
+error_report_dma_error:
+       u64_stats_update_begin(&tx_ring->syncp);
+       tx_ring->tx_stats.dma_mapping_err++;
+       u64_stats_update_end(&tx_ring->syncp);
+       netdev_warn(adapter->netdev, "failed to map skb\n");
+
+       tx_info->skb = NULL;
+
+       tx_info->num_of_bufs += i;
+       ena_unmap_tx_skb(tx_ring, tx_info);
+
+       return -EINVAL;
+}
+
+/* Called with netif_tx_lock. */
+static netdev_tx_t ena_start_xmit(struct sk_buff *skb, struct net_device *dev)
+{
+       struct ena_adapter *adapter = netdev_priv(dev);
+       struct ena_tx_buffer *tx_info;
+       struct ena_com_tx_ctx ena_tx_ctx;
+       struct ena_ring *tx_ring;
+       struct netdev_queue *txq;
+       void *push_hdr;
+       u16 next_to_use, req_id, header_len;
+       int qid, rc, nb_hw_desc;
+
+       netif_dbg(adapter, tx_queued, dev, "%s skb %p\n", __func__, skb);
+       /*  Determine which tx ring we will be placed on */
+       qid = skb_get_queue_mapping(skb);
+       tx_ring = &adapter->tx_ring[qid];
+       txq = netdev_get_tx_queue(dev, qid);
+
+       rc = ena_check_and_linearize_skb(tx_ring, skb);
+       if (unlikely(rc))
+               goto error_drop_packet;
+
+       skb_tx_timestamp(skb);
+
+       next_to_use = tx_ring->next_to_use;
+       req_id = tx_ring->free_tx_ids[next_to_use];
+       tx_info = &tx_ring->tx_buffer_info[req_id];
+       tx_info->num_of_bufs = 0;
+
+       WARN(tx_info->skb, "SKB isn't NULL req_id %d\n", req_id);
+
+       rc = ena_tx_map_skb(tx_ring, tx_info, skb, &push_hdr, &header_len);
+       if (unlikely(rc))
+               goto error_drop_packet;
 
        memset(&ena_tx_ctx, 0x0, sizeof(struct ena_com_tx_ctx));
        ena_tx_ctx.ena_bufs = tx_info->bufs;
@@ -2082,14 +2147,22 @@ static netdev_tx_t ena_start_xmit(struct sk_buff *skb, struct net_device *dev)
        rc = ena_com_prepare_tx(tx_ring->ena_com_io_sq, &ena_tx_ctx,
                                &nb_hw_desc);
 
+       /* ena_com_prepare_tx() can't fail due to overflow of tx queue,
+        * since the number of free descriptors in the queue is checked
+        * after sending the previous packet. In case there isn't enough
+        * space in the queue for the next packet, it is stopped
+        * until there is again enough available space in the queue.
+        * All other failure reasons of ena_com_prepare_tx() are fatal
+        * and therefore require a device reset.
+        */
        if (unlikely(rc)) {
                netif_err(adapter, tx_queued, dev,
                          "failed to prepare tx bufs\n");
                u64_stats_update_begin(&tx_ring->syncp);
-               tx_ring->tx_stats.queue_stop++;
                tx_ring->tx_stats.prepare_ctx_err++;
                u64_stats_update_end(&tx_ring->syncp);
-               netif_tx_stop_queue(txq);
+               adapter->reset_reason = ENA_REGS_RESET_DRIVER_INVALID_STATE;
+               set_bit(ENA_FLAG_TRIGGER_RESET, &adapter->flags);
                goto error_unmap_dma;
        }
 
@@ -2152,35 +2225,11 @@ static netdev_tx_t ena_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
        return NETDEV_TX_OK;
 
-error_report_dma_error:
-       u64_stats_update_begin(&tx_ring->syncp);
-       tx_ring->tx_stats.dma_mapping_err++;
-       u64_stats_update_end(&tx_ring->syncp);
-       netdev_warn(adapter->netdev, "failed to map skb\n");
-
-       tx_info->skb = NULL;
-
 error_unmap_dma:
-       if (i >= 0) {
-               /* save value of frag that failed */
-               last_frag = i;
-
-               /* start back at beginning and unmap skb */
-               tx_info->skb = NULL;
-               ena_buf = tx_info->bufs;
-               dma_unmap_single(tx_ring->dev, dma_unmap_addr(ena_buf, paddr),
-                                dma_unmap_len(ena_buf, len), DMA_TO_DEVICE);
-
-               /* unmap remaining mapped pages */
-               for (i = 0; i < last_frag; i++) {
-                       ena_buf++;
-                       dma_unmap_page(tx_ring->dev, dma_unmap_addr(ena_buf, paddr),
-                                      dma_unmap_len(ena_buf, len), DMA_TO_DEVICE);
-               }
-       }
+       ena_unmap_tx_skb(tx_ring, tx_info);
+       tx_info->skb = NULL;
 
 error_drop_packet:
-
        dev_kfree_skb(skb);
        return NETDEV_TX_OK;
 }
@@ -2616,7 +2665,9 @@ static int ena_restore_device(struct ena_adapter *adapter)
 
        set_bit(ENA_FLAG_DEVICE_RUNNING, &adapter->flags);
        mod_timer(&adapter->timer_service, round_jiffies(jiffies + HZ));
-       dev_err(&pdev->dev, "Device reset completed successfully\n");
+       dev_err(&pdev->dev,
+               "Device reset completed successfully, Driver info: %s\n",
+               version);
 
        return rc;
 err_disable_msix:
@@ -2979,18 +3030,52 @@ static int ena_calc_io_queue_num(struct pci_dev *pdev,
        return io_queue_num;
 }
 
-static void ena_set_push_mode(struct pci_dev *pdev, struct ena_com_dev *ena_dev,
-                             struct ena_com_dev_get_features_ctx *get_feat_ctx)
+static int ena_set_queues_placement_policy(struct pci_dev *pdev,
+                                          struct ena_com_dev *ena_dev,
+                                          struct ena_admin_feature_llq_desc *llq,
+                                          struct ena_llq_configurations *llq_default_configurations)
 {
        bool has_mem_bar;
+       int rc;
+       u32 llq_feature_mask;
+
+       llq_feature_mask = 1 << ENA_ADMIN_LLQ;
+       if (!(ena_dev->supported_features & llq_feature_mask)) {
+               dev_err(&pdev->dev,
+                       "LLQ is not supported Fallback to host mode policy.\n");
+               ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+               return 0;
+       }
 
        has_mem_bar = pci_select_bars(pdev, IORESOURCE_MEM) & BIT(ENA_MEM_BAR);
 
-       /* Enable push mode if device supports LLQ */
-       if (has_mem_bar && get_feat_ctx->max_queues.max_legacy_llq_num > 0)
-               ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_DEV;
-       else
+       rc = ena_com_config_dev_mode(ena_dev, llq, llq_default_configurations);
+       if (unlikely(rc)) {
+               dev_err(&pdev->dev,
+                       "Failed to configure the device 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 (!has_mem_bar) {
+               dev_err(&pdev->dev,
+                       "ENA device does not expose LLQ bar. Fallback to host mode policy.\n");
                ena_dev->tx_mem_queue_type = ENA_ADMIN_PLACEMENT_POLICY_HOST;
+               return 0;
+       }
+
+       ena_dev->mem_bar = devm_ioremap_wc(&pdev->dev,
+                                          pci_resource_start(pdev, ENA_MEM_BAR),
+                                          pci_resource_len(pdev, ENA_MEM_BAR));
+
+       if (!ena_dev->mem_bar)
+               return -EFAULT;
+
+       return 0;
 }
 
 static void ena_set_dev_offloads(struct ena_com_dev_get_features_ctx *feat,
@@ -3115,6 +3200,15 @@ static void ena_release_bars(struct ena_com_dev *ena_dev, struct pci_dev *pdev)
        pci_release_selected_regions(pdev, release_bars);
 }
 
+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_calc_queue_size(struct pci_dev *pdev,
                               struct ena_com_dev *ena_dev,
                               u16 *max_tx_sgl_size,
@@ -3163,7 +3257,9 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
        static int version_printed;
        struct net_device *netdev;
        struct ena_adapter *adapter;
+       struct ena_llq_configurations llq_config;
        struct ena_com_dev *ena_dev = NULL;
+       char *queue_type_str;
        static int adapters_found;
        int io_queue_num, bars, rc;
        int queue_size;
@@ -3217,16 +3313,13 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
                goto err_free_region;
        }
 
-       ena_set_push_mode(pdev, ena_dev, &get_feat_ctx);
+       set_default_llq_configurations(&llq_config);
 
-       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
-               ena_dev->mem_bar = devm_ioremap_wc(&pdev->dev,
-                                                  pci_resource_start(pdev, ENA_MEM_BAR),
-                                                  pci_resource_len(pdev, ENA_MEM_BAR));
-               if (!ena_dev->mem_bar) {
-                       rc = -EFAULT;
-                       goto err_device_destroy;
-               }
+       rc = ena_set_queues_placement_policy(pdev, ena_dev, &get_feat_ctx.llq,
+                                            &llq_config);
+       if (rc) {
+               dev_err(&pdev->dev, "ena device init failed\n");
+               goto err_device_destroy;
        }
 
        /* initial Tx interrupt delay, Assumes 1 usec granularity.
@@ -3241,8 +3334,10 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
                goto err_device_destroy;
        }
 
-       dev_info(&pdev->dev, "creating %d io queues. queue size: %d\n",
-                io_queue_num, queue_size);
+       dev_info(&pdev->dev, "creating %d io queues. queue size: %d. LLQ is %s\n",
+                io_queue_num, queue_size,
+                (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) ?
+                "ENABLED" : "DISABLED");
 
        /* dev zeroed in init_etherdev */
        netdev = alloc_etherdev_mq(sizeof(struct ena_adapter), io_queue_num);
@@ -3332,9 +3427,15 @@ static int ena_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
        timer_setup(&adapter->timer_service, ena_timer_service, 0);
        mod_timer(&adapter->timer_service, round_jiffies(jiffies + HZ));
 
-       dev_info(&pdev->dev, "%s found at mem %lx, mac addr %pM Queues %d\n",
+       if (ena_dev->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_HOST)
+               queue_type_str = "Regular";
+       else
+               queue_type_str = "Low Latency";
+
+       dev_info(&pdev->dev,
+                "%s found at mem %lx, mac addr %pM Queues %d, Placement policy: %s\n",
                 DEVICE_NAME, (long)pci_resource_start(pdev, 0),
-                netdev->dev_addr, io_queue_num);
+                netdev->dev_addr, io_queue_num, queue_type_str);
 
        set_bit(ENA_FLAG_DEVICE_RUNNING, &adapter->flags);
 
index 7c7ae56c52cfddb11a8e1b096738dfc499ea57dc..4fa7d2fda4750d906758b92d1b0b905e2061b36d 100644 (file)
@@ -151,6 +151,9 @@ struct ena_tx_buffer {
        /* num of buffers used by this skb */
        u32 num_of_bufs;
 
+       /* Indicate if bufs[0] map the linear data of the skb. */
+       u8 map_linear_data;
+
        /* Used for detect missing tx packets to limit the number of prints */
        u32 print_once;
        /* Save the last jiffies to detect missing tx packets
@@ -186,6 +189,7 @@ struct ena_stats_tx {
        u64 tx_poll;
        u64 doorbells;
        u64 bad_req_id;
+       u64 llq_buffer_copy;
        u64 missed_tx;
 };
 
@@ -257,6 +261,8 @@ struct ena_ring {
                struct ena_stats_tx tx_stats;
                struct ena_stats_rx rx_stats;
        };
+
+       u8 *push_buf_intermediate_buf;
        int empty_rx_queue;
 } ____cacheline_aligned;