qede: Add support for ingress headroom
authorMintz, Yuval <Yuval.Mintz@cavium.com>
Fri, 7 Apr 2017 08:05:00 +0000 (11:05 +0300)
committerDavid S. Miller <davem@davemloft.net>
Fri, 7 Apr 2017 13:26:14 +0000 (06:26 -0700)
Driver currently doesn't support any headroom; The only 'available'
space it has in the head of the buffer is due to the placement
offset.
In order to allow [later] support of XDP adjustment of headroom,
modify the the ingress flow to properly handle a scenario where
the packets would have such.

Signed-off-by: Yuval Mintz <Yuval.Mintz@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/qlogic/qede/qede.h
drivers/net/ethernet/qlogic/qede/qede_fp.c
drivers/net/ethernet/qlogic/qede/qede_main.c

index 7ab2201a43b26cb325d30f7b0b1a7bdeb74f3101..5e7ad25db8adb9d25606bc61742298af8360bcb8 100644 (file)
@@ -313,21 +313,24 @@ struct qede_rx_queue {
        u8 data_direction;
        u8 rxq_id;
 
+       /* Used once per each NAPI run */
+       u16 num_rx_buffers;
+
+       u16 rx_headroom;
+
        u32 rx_buf_size;
        u32 rx_buf_seg_size;
 
-       u64 rcv_pkts;
-
        struct sw_rx_data *sw_rx_ring;
        struct qed_chain rx_bd_ring;
        struct qed_chain rx_comp_ring ____cacheline_aligned;
 
-       /* Used once per each NAPI run */
-       u16 num_rx_buffers;
-
        /* GRO */
        struct qede_agg_info tpa_info[ETH_TPA_MAX_AGGS_NUM];
 
+       /* Used once per each NAPI run */
+       u64 rcv_pkts;
+
        u64 rx_hw_errors;
        u64 rx_alloc_errors;
        u64 rx_ip_frags;
index c61cfcfbbd56ef29ea7d09154af2d732e32fc5dd..e382d4bdf4303a9f07dfa52fa31a5b9d491baa46 100644 (file)
@@ -87,7 +87,8 @@ int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy)
        rx_bd = (struct eth_rx_bd *)qed_chain_produce(&rxq->rx_bd_ring);
        WARN_ON(!rx_bd);
        rx_bd->addr.hi = cpu_to_le32(upper_32_bits(mapping));
-       rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping));
+       rx_bd->addr.lo = cpu_to_le32(lower_32_bits(mapping) +
+                                    rxq->rx_headroom);
 
        rxq->sw_rx_prod++;
        rxq->filled_buffers++;
@@ -509,7 +510,8 @@ static inline void qede_reuse_page(struct qede_rx_queue *rxq,
        new_mapping = curr_prod->mapping + curr_prod->page_offset;
 
        rx_bd_prod->addr.hi = cpu_to_le32(upper_32_bits(new_mapping));
-       rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping));
+       rx_bd_prod->addr.lo = cpu_to_le32(lower_32_bits(new_mapping) +
+                                         rxq->rx_headroom);
 
        rxq->sw_rx_prod++;
        curr_cons->data = NULL;
@@ -991,13 +993,14 @@ static bool qede_rx_xdp(struct qede_dev *edev,
                        struct qede_rx_queue *rxq,
                        struct bpf_prog *prog,
                        struct sw_rx_data *bd,
-                       struct eth_fast_path_rx_reg_cqe *cqe)
+                       struct eth_fast_path_rx_reg_cqe *cqe,
+                       u16 data_offset)
 {
        u16 len = le16_to_cpu(cqe->len_on_first_bd);
        struct xdp_buff xdp;
        enum xdp_action act;
 
-       xdp.data = page_address(bd->data) + cqe->placement_offset;
+       xdp.data = page_address(bd->data) + data_offset;
        xdp.data_end = xdp.data + len;
 
        /* Queues always have a full reset currently, so for the time
@@ -1026,7 +1029,7 @@ static bool qede_rx_xdp(struct qede_dev *edev,
                /* Now if there's a transmission problem, we'd still have to
                 * throw current buffer, as replacement was already allocated.
                 */
-               if (qede_xdp_xmit(edev, fp, bd, cqe->placement_offset, len)) {
+               if (qede_xdp_xmit(edev, fp, bd, data_offset, len)) {
                        dma_unmap_page(rxq->dev, bd->mapping,
                                       PAGE_SIZE, DMA_BIDIRECTIONAL);
                        __free_page(bd->data);
@@ -1053,7 +1056,7 @@ static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev,
                                            struct sw_rx_data *bd, u16 len,
                                            u16 pad)
 {
-       unsigned int offset = bd->page_offset;
+       unsigned int offset = bd->page_offset + pad;
        struct skb_frag_struct *frag;
        struct page *page = bd->data;
        unsigned int pull_len;
@@ -1070,7 +1073,7 @@ static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev,
         */
        if (len + pad <= edev->rx_copybreak) {
                memcpy(skb_put(skb, len),
-                      page_address(page) + pad + offset, len);
+                      page_address(page) + offset, len);
                qede_reuse_page(rxq, bd);
                goto out;
        }
@@ -1078,7 +1081,7 @@ static struct sk_buff *qede_rx_allocate_skb(struct qede_dev *edev,
        frag = &skb_shinfo(skb)->frags[0];
 
        skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
-                       page, pad + offset, len, rxq->rx_buf_seg_size);
+                       page, offset, len, rxq->rx_buf_seg_size);
 
        va = skb_frag_address(frag);
        pull_len = eth_get_headlen(va, QEDE_RX_HDR_SIZE);
@@ -1224,11 +1227,11 @@ static int qede_rx_process_cqe(struct qede_dev *edev,
 
        fp_cqe = &cqe->fast_path_regular;
        len = le16_to_cpu(fp_cqe->len_on_first_bd);
-       pad = fp_cqe->placement_offset;
+       pad = fp_cqe->placement_offset + rxq->rx_headroom;
 
        /* Run eBPF program if one is attached */
        if (xdp_prog)
-               if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe))
+               if (!qede_rx_xdp(edev, fp, rxq, xdp_prog, bd, fp_cqe, pad))
                        return 0;
 
        /* If this is an error packet then drop it */
index fa62c37dac7a2141661784ee7c4ef87ff1fa4fe0..91c3078c1ae45193d84fa0a9664b4d216a6b3011 100644 (file)
@@ -1188,8 +1188,9 @@ static int qede_alloc_mem_rxq(struct qede_dev *edev, struct qede_rx_queue *rxq)
 
        rxq->rx_buf_size = NET_IP_ALIGN + ETH_OVERHEAD + edev->ndev->mtu;
 
-       if (rxq->rx_buf_size > PAGE_SIZE)
-               rxq->rx_buf_size = PAGE_SIZE;
+       /* Make sure that the headroom and  payload fit in a single page */
+       if (rxq->rx_buf_size + rxq->rx_headroom > PAGE_SIZE)
+               rxq->rx_buf_size = PAGE_SIZE - rxq->rx_headroom;
 
        /* Segment size to spilt a page in multiple equal parts,
         * unless XDP is used in which case we'd use the entire page.