target/linux/generic/patches-3.18/077-02-bgmac-implement-GRO-and-use-build_skb.patch

   1 From: Felix Fietkau <nbd@nbd.name>
   2 Date: Mon, 23 Mar 2015 02:41:25 +0100
   3 Subject: [PATCH] bgmac: implement GRO and use build_skb
   4
   5 This improves performance for routing and local rx
   6
   7 Signed-off-by: Felix Fietkau <nbd@nbd.name>
   8 ---
   9
  10 --- a/drivers/net/ethernet/broadcom/bgmac.c
  11 +++ b/drivers/net/ethernet/broadcom/bgmac.c
  12 @@ -276,31 +276,31 @@ static int bgmac_dma_rx_skb_for_slot(str
  13                                      struct bgmac_slot_info *slot)
  14  {
  15         struct device *dma_dev = bgmac->core->dma_dev;
  16 -       struct sk_buff *skb;
  17         dma_addr_t dma_addr;
  18         struct bgmac_rx_header *rx;
  19 +       void *buf;
  20
  21         /* Alloc skb */
  22 -       skb = netdev_alloc_skb(bgmac->net_dev, BGMAC_RX_BUF_SIZE);
  23 -       if (!skb)
  24 +       buf = netdev_alloc_frag(BGMAC_RX_ALLOC_SIZE);
  25 +       if (!buf)
  26                 return -ENOMEM;
  27
  28         /* Poison - if everything goes fine, hardware will overwrite it */
  29 -       rx = (struct bgmac_rx_header *)skb->data;
  30 +       rx = buf;
  31         rx->len = cpu_to_le16(0xdead);
  32         rx->flags = cpu_to_le16(0xbeef);
  33
  34         /* Map skb for the DMA */
  35 -       dma_addr = dma_map_single(dma_dev, skb->data,
  36 -                                 BGMAC_RX_BUF_SIZE, DMA_FROM_DEVICE);
  37 +       dma_addr = dma_map_single(dma_dev, buf, BGMAC_RX_BUF_SIZE,
  38 +                                 DMA_FROM_DEVICE);
  39         if (dma_mapping_error(dma_dev, dma_addr)) {
  40                 bgmac_err(bgmac, "DMA mapping error\n");
  41 -               dev_kfree_skb(skb);
  42 +               put_page(virt_to_head_page(buf));
  43                 return -ENOMEM;
  44         }
  45
  46         /* Update the slot */
  47 -       slot->skb = skb;
  48 +       slot->buf = buf;
  49         slot->dma_addr = dma_addr;
  50
  51         return 0;
  52 @@ -343,8 +343,9 @@ static int bgmac_dma_rx_read(struct bgma
  53         while (ring->start != ring->end) {
  54                 struct device *dma_dev = bgmac->core->dma_dev;
  55                 struct bgmac_slot_info *slot = &ring->slots[ring->start];
  56 -               struct sk_buff *skb = slot->skb;
  57 -               struct bgmac_rx_header *rx;
  58 +               struct bgmac_rx_header *rx = slot->buf;
  59 +               struct sk_buff *skb;
  60 +               void *buf = slot->buf;
  61                 u16 len, flags;
  62
  63                 /* Unmap buffer to make it accessible to the CPU */
  64 @@ -352,7 +353,6 @@ static int bgmac_dma_rx_read(struct bgma
  65                                         BGMAC_RX_BUF_SIZE, DMA_FROM_DEVICE);
  66
  67                 /* Get info from the header */
  68 -               rx = (struct bgmac_rx_header *)skb->data;
  69                 len = le16_to_cpu(rx->len);
  70                 flags = le16_to_cpu(rx->flags);
  71
  72 @@ -393,12 +393,13 @@ static int bgmac_dma_rx_read(struct bgma
  73                         dma_unmap_single(dma_dev, old_dma_addr,
  74                                          BGMAC_RX_BUF_SIZE, DMA_FROM_DEVICE);
  75
  76 +                       skb = build_skb(buf, BGMAC_RX_ALLOC_SIZE);
  77                         skb_put(skb, BGMAC_RX_FRAME_OFFSET + len);
  78                         skb_pull(skb, BGMAC_RX_FRAME_OFFSET);
  79
  80                         skb_checksum_none_assert(skb);
  81                         skb->protocol = eth_type_trans(skb, bgmac->net_dev);
  82 -                       netif_receive_skb(skb);
  83 +                       napi_gro_receive(&bgmac->napi, skb);
  84                         handled++;
  85                 } while (0);
  86
  87 @@ -434,12 +435,11 @@ static bool bgmac_dma_unaligned(struct b
  88         return false;
  89  }
  90
  91 -static void bgmac_dma_ring_free(struct bgmac *bgmac,
  92 -                               struct bgmac_dma_ring *ring)
  93 +static void bgmac_dma_tx_ring_free(struct bgmac *bgmac,
  94 +                                  struct bgmac_dma_ring *ring)
  95  {
  96         struct device *dma_dev = bgmac->core->dma_dev;
  97         struct bgmac_slot_info *slot;
  98 -       int size;
  99         int i;
 100
 101         for (i = 0; i < ring->num_slots; i++) {
 102 @@ -451,23 +451,55 @@ static void bgmac_dma_ring_free(struct b
 103                         dev_kfree_skb(slot->skb);
 104                 }
 105         }
 106 +}
 107 +
 108 +static void bgmac_dma_rx_ring_free(struct bgmac *bgmac,
 109 +                                  struct bgmac_dma_ring *ring)
 110 +{
 111 +       struct device *dma_dev = bgmac->core->dma_dev;
 112 +       struct bgmac_slot_info *slot;
 113 +       int i;
 114 +
 115 +       for (i = 0; i < ring->num_slots; i++) {
 116 +               slot = &ring->slots[i];
 117 +               if (!slot->buf)
 118 +                       continue;
 119
 120 -       if (ring->cpu_base) {
 121 -               /* Free ring of descriptors */
 122 -               size = ring->num_slots * sizeof(struct bgmac_dma_desc);
 123 -               dma_free_coherent(dma_dev, size, ring->cpu_base,
 124 -                                 ring->dma_base);
 125 +               if (slot->dma_addr)
 126 +                       dma_unmap_single(dma_dev, slot->dma_addr,
 127 +                                        BGMAC_RX_BUF_SIZE,
 128 +                                        DMA_FROM_DEVICE);
 129 +               put_page(virt_to_head_page(slot->buf));
 130         }
 131  }
 132
 133 +static void bgmac_dma_ring_desc_free(struct bgmac *bgmac,
 134 +                                    struct bgmac_dma_ring *ring)
 135 +{
 136 +       struct device *dma_dev = bgmac->core->dma_dev;
 137 +       int size;
 138 +
 139 +       if (!ring->cpu_base)
 140 +           return;
 141 +
 142 +       /* Free ring of descriptors */
 143 +       size = ring->num_slots * sizeof(struct bgmac_dma_desc);
 144 +       dma_free_coherent(dma_dev, size, ring->cpu_base,
 145 +                         ring->dma_base);
 146 +}
 147 +
 148  static void bgmac_dma_free(struct bgmac *bgmac)
 149  {
 150         int i;
 151
 152 -       for (i = 0; i < BGMAC_MAX_TX_RINGS; i++)
 153 -               bgmac_dma_ring_free(bgmac, &bgmac->tx_ring[i]);
 154 -       for (i = 0; i < BGMAC_MAX_RX_RINGS; i++)
 155 -               bgmac_dma_ring_free(bgmac, &bgmac->rx_ring[i]);
 156 +       for (i = 0; i < BGMAC_MAX_TX_RINGS; i++) {
 157 +               bgmac_dma_tx_ring_free(bgmac, &bgmac->tx_ring[i]);
 158 +               bgmac_dma_ring_desc_free(bgmac, &bgmac->tx_ring[i]);
 159 +       }
 160 +       for (i = 0; i < BGMAC_MAX_RX_RINGS; i++) {
 161 +               bgmac_dma_rx_ring_free(bgmac, &bgmac->rx_ring[i]);
 162 +               bgmac_dma_ring_desc_free(bgmac, &bgmac->rx_ring[i]);
 163 +       }
 164  }
 165
 166  static int bgmac_dma_alloc(struct bgmac *bgmac)
 167 --- a/drivers/net/ethernet/broadcom/bgmac.h
 168 +++ b/drivers/net/ethernet/broadcom/bgmac.h
 169 @@ -362,6 +362,8 @@
 170  #define BGMAC_RX_FRAME_OFFSET                  30              /* There are 2 unused bytes between header and real data */
 171  #define BGMAC_RX_MAX_FRAME_SIZE                        1536            /* Copied from b44/tg3 */
 172  #define BGMAC_RX_BUF_SIZE                      (BGMAC_RX_FRAME_OFFSET + BGMAC_RX_MAX_FRAME_SIZE)
 173 +#define BGMAC_RX_ALLOC_SIZE                    (SKB_DATA_ALIGN(BGMAC_RX_BUF_SIZE) + \
 174 +                                                SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
 175
 176  #define BGMAC_BFL_ENETROBO                     0x0010          /* has ephy roboswitch spi */
 177  #define BGMAC_BFL_ENETADM                      0x0080          /* has ADMtek switch */
 178 @@ -383,7 +385,10 @@
 179  #define ETHER_MAX_LEN   1518
 180
 181  struct bgmac_slot_info {
 182 -       struct sk_buff *skb;
 183 +       union {
 184 +               struct sk_buff *skb;
 185 +               void *buf;
 186 +       };
 187         dma_addr_t dma_addr;
 188  };
 189