drivers/net/ethernet/mellanox/mlx5/core/en_rx.c

   1 /*
   2  * Copyright (c) 2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/ip.h>
  34 #include <linux/ipv6.h>
  35 #include <linux/tcp.h>
  36 #include <net/busy_poll.h>
  37 #include "en.h"
  38 #include "en_tc.h"
  39
  40 static inline bool mlx5e_rx_hw_stamp(struct mlx5e_tstamp *tstamp)
  41 {
  42         return tstamp->hwtstamp_config.rx_filter == HWTSTAMP_FILTER_ALL;
  43 }
  44
  45 static inline int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq,
  46                                      struct mlx5e_rx_wqe *wqe, u16 ix)
  47 {
  48         struct sk_buff *skb;
  49         dma_addr_t dma_addr;
  50
  51         skb = netdev_alloc_skb(rq->netdev, rq->wqe_sz);
  52         if (unlikely(!skb))
  53                 return -ENOMEM;
  54
  55         dma_addr = dma_map_single(rq->pdev,
  56                                   /* hw start padding */
  57                                   skb->data,
  58                                   /* hw end padding */
  59                                   rq->wqe_sz,
  60                                   DMA_FROM_DEVICE);
  61
  62         if (unlikely(dma_mapping_error(rq->pdev, dma_addr)))
  63                 goto err_free_skb;
  64
  65         skb_reserve(skb, MLX5E_NET_IP_ALIGN);
  66
  67         *((dma_addr_t *)skb->cb) = dma_addr;
  68         wqe->data.addr = cpu_to_be64(dma_addr + MLX5E_NET_IP_ALIGN);
  69
  70         rq->skb[ix] = skb;
  71
  72         return 0;
  73
  74 err_free_skb:
  75         dev_kfree_skb(skb);
  76
  77         return -ENOMEM;
  78 }
  79
  80 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
  81 {
  82         struct mlx5_wq_ll *wq = &rq->wq;
  83
  84         if (unlikely(!test_bit(MLX5E_RQ_STATE_POST_WQES_ENABLE, &rq->state)))
  85                 return false;
  86
  87         while (!mlx5_wq_ll_is_full(wq)) {
  88                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
  89
  90                 if (unlikely(mlx5e_alloc_rx_wqe(rq, wqe, wq->head)))
  91                         break;
  92
  93                 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
  94         }
  95
  96         /* ensure wqes are visible to device before updating doorbell record */
  97         dma_wmb();
  98
  99         mlx5_wq_ll_update_db_record(wq);
 100
 101         return !mlx5_wq_ll_is_full(wq);
 102 }
 103
 104 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe)
 105 {
 106         struct ethhdr   *eth    = (struct ethhdr *)(skb->data);
 107         struct iphdr    *ipv4   = (struct iphdr *)(skb->data + ETH_HLEN);
 108         struct ipv6hdr  *ipv6   = (struct ipv6hdr *)(skb->data + ETH_HLEN);
 109         struct tcphdr   *tcp;
 110
 111         u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
 112         int tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA  == l4_hdr_type) ||
 113                        (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
 114
 115         u16 tot_len = be32_to_cpu(cqe->byte_cnt) - ETH_HLEN;
 116
 117         if (eth->h_proto == htons(ETH_P_IP)) {
 118                 tcp = (struct tcphdr *)(skb->data + ETH_HLEN +
 119                                         sizeof(struct iphdr));
 120                 ipv6 = NULL;
 121                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 122         } else {
 123                 tcp = (struct tcphdr *)(skb->data + ETH_HLEN +
 124                                         sizeof(struct ipv6hdr));
 125                 ipv4 = NULL;
 126                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 127         }
 128
 129         if (get_cqe_lro_tcppsh(cqe))
 130                 tcp->psh                = 1;
 131
 132         if (tcp_ack) {
 133                 tcp->ack                = 1;
 134                 tcp->ack_seq            = cqe->lro_ack_seq_num;
 135                 tcp->window             = cqe->lro_tcp_win;
 136         }
 137
 138         if (ipv4) {
 139                 ipv4->ttl               = cqe->lro_min_ttl;
 140                 ipv4->tot_len           = cpu_to_be16(tot_len);
 141                 ipv4->check             = 0;
 142                 ipv4->check             = ip_fast_csum((unsigned char *)ipv4,
 143                                                        ipv4->ihl);
 144         } else {
 145                 ipv6->hop_limit         = cqe->lro_min_ttl;
 146                 ipv6->payload_len       = cpu_to_be16(tot_len -
 147                                                       sizeof(struct ipv6hdr));
 148         }
 149 }
 150
 151 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
 152                                       struct sk_buff *skb)
 153 {
 154         u8 cht = cqe->rss_hash_type;
 155         int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
 156                  (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
 157                                             PKT_HASH_TYPE_NONE;
 158         skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
 159 }
 160
 161 static inline bool is_first_ethertype_ip(struct sk_buff *skb)
 162 {
 163         __be16 ethertype = ((struct ethhdr *)skb->data)->h_proto;
 164
 165         return (ethertype == htons(ETH_P_IP) || ethertype == htons(ETH_P_IPV6));
 166 }
 167
 168 static inline void mlx5e_handle_csum(struct net_device *netdev,
 169                                      struct mlx5_cqe64 *cqe,
 170                                      struct mlx5e_rq *rq,
 171                                      struct sk_buff *skb,
 172                                      bool   lro)
 173 {
 174         if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
 175                 goto csum_none;
 176
 177         if (lro) {
 178                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 179         } else if (likely(is_first_ethertype_ip(skb))) {
 180                 skb->ip_summed = CHECKSUM_COMPLETE;
 181                 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
 182                 rq->stats.csum_sw++;
 183         } else {
 184                 goto csum_none;
 185         }
 186
 187         return;
 188
 189 csum_none:
 190         skb->ip_summed = CHECKSUM_NONE;
 191         rq->stats.csum_none++;
 192 }
 193
 194 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
 195                                       struct mlx5e_rq *rq,
 196                                       struct sk_buff *skb)
 197 {
 198         struct net_device *netdev = rq->netdev;
 199         u32 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
 200         struct mlx5e_tstamp *tstamp = rq->tstamp;
 201         int lro_num_seg;
 202
 203         skb_put(skb, cqe_bcnt);
 204
 205         lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
 206         if (lro_num_seg > 1) {
 207                 mlx5e_lro_update_hdr(skb, cqe);
 208                 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
 209                 rq->stats.lro_packets++;
 210                 rq->stats.lro_bytes += cqe_bcnt;
 211         }
 212
 213         if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
 214                 mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
 215
 216         mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
 217
 218         skb->protocol = eth_type_trans(skb, netdev);
 219
 220         skb_record_rx_queue(skb, rq->ix);
 221
 222         if (likely(netdev->features & NETIF_F_RXHASH))
 223                 mlx5e_skb_set_hash(cqe, skb);
 224
 225         if (cqe_has_vlan(cqe))
 226                 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 227                                        be16_to_cpu(cqe->vlan_info));
 228
 229         skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
 230 }
 231
 232 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
 233 {
 234         struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
 235         int work_done;
 236
 237         for (work_done = 0; work_done < budget; work_done++) {
 238                 struct mlx5e_rx_wqe *wqe;
 239                 struct mlx5_cqe64 *cqe;
 240                 struct sk_buff *skb;
 241                 __be16 wqe_counter_be;
 242                 u16 wqe_counter;
 243
 244                 cqe = mlx5e_get_cqe(cq);
 245                 if (!cqe)
 246                         break;
 247
 248                 mlx5_cqwq_pop(&cq->wq);
 249
 250                 wqe_counter_be = cqe->wqe_counter;
 251                 wqe_counter    = be16_to_cpu(wqe_counter_be);
 252                 wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
 253                 skb            = rq->skb[wqe_counter];
 254                 prefetch(skb->data);
 255                 rq->skb[wqe_counter] = NULL;
 256
 257                 dma_unmap_single(rq->pdev,
 258                                  *((dma_addr_t *)skb->cb),
 259                                  rq->wqe_sz,
 260                                  DMA_FROM_DEVICE);
 261
 262                 if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
 263                         rq->stats.wqe_err++;
 264                         dev_kfree_skb(skb);
 265                         goto wq_ll_pop;
 266                 }
 267
 268                 mlx5e_build_rx_skb(cqe, rq, skb);
 269                 rq->stats.packets++;
 270                 rq->stats.bytes += be32_to_cpu(cqe->byte_cnt);
 271                 napi_gro_receive(cq->napi, skb);
 272
 273 wq_ll_pop:
 274                 mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
 275                                &wqe->next.next_wqe_index);
 276         }
 277
 278         mlx5_cqwq_update_db_record(&cq->wq);
 279
 280         /* ensure cq space is freed before enabling more cqes */
 281         wmb();
 282
 283         return work_done;
 284 }