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/prefetch.h>
  34 #include <linux/ip.h>
  35 #include <linux/ipv6.h>
  36 #include <linux/tcp.h>
  37 #include <linux/bpf_trace.h>
  38 #include <net/busy_poll.h>
  39 #include "en.h"
  40 #include "en_tc.h"
  41 #include "eswitch.h"
  42 #include "en_rep.h"
  43 #include "ipoib/ipoib.h"
  44 #include "en_accel/ipsec_rxtx.h"
  45
  46 static inline bool mlx5e_rx_hw_stamp(struct mlx5e_tstamp *tstamp)
  47 {
  48         return tstamp->hwtstamp_config.rx_filter == HWTSTAMP_FILTER_ALL;
  49 }
  50
  51 static inline void mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cqcc,
  52                                        void *data)
  53 {
  54         u32 ci = cqcc & cq->wq.sz_m1;
  55
  56         memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, ci), sizeof(struct mlx5_cqe64));
  57 }
  58
  59 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
  60                                          struct mlx5e_cq *cq, u32 cqcc)
  61 {
  62         mlx5e_read_cqe_slot(cq, cqcc, &cq->title);
  63         cq->decmprs_left        = be32_to_cpu(cq->title.byte_cnt);
  64         cq->decmprs_wqe_counter = be16_to_cpu(cq->title.wqe_counter);
  65         rq->stats.cqe_compress_blks++;
  66 }
  67
  68 static inline void mlx5e_read_mini_arr_slot(struct mlx5e_cq *cq, u32 cqcc)
  69 {
  70         mlx5e_read_cqe_slot(cq, cqcc, cq->mini_arr);
  71         cq->mini_arr_idx = 0;
  72 }
  73
  74 static inline void mlx5e_cqes_update_owner(struct mlx5e_cq *cq, u32 cqcc, int n)
  75 {
  76         u8 op_own = (cqcc >> cq->wq.log_sz) & 1;
  77         u32 wq_sz = 1 << cq->wq.log_sz;
  78         u32 ci = cqcc & cq->wq.sz_m1;
  79         u32 ci_top = min_t(u32, wq_sz, ci + n);
  80
  81         for (; ci < ci_top; ci++, n--) {
  82                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
  83
  84                 cqe->op_own = op_own;
  85         }
  86
  87         if (unlikely(ci == wq_sz)) {
  88                 op_own = !op_own;
  89                 for (ci = 0; ci < n; ci++) {
  90                         struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
  91
  92                         cqe->op_own = op_own;
  93                 }
  94         }
  95 }
  96
  97 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
  98                                         struct mlx5e_cq *cq, u32 cqcc)
  99 {
 100         cq->title.byte_cnt     = cq->mini_arr[cq->mini_arr_idx].byte_cnt;
 101         cq->title.check_sum    = cq->mini_arr[cq->mini_arr_idx].checksum;
 102         cq->title.op_own      &= 0xf0;
 103         cq->title.op_own      |= 0x01 & (cqcc >> cq->wq.log_sz);
 104         cq->title.wqe_counter  = cpu_to_be16(cq->decmprs_wqe_counter);
 105
 106         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
 107                 cq->decmprs_wqe_counter +=
 108                         mpwrq_get_cqe_consumed_strides(&cq->title);
 109         else
 110                 cq->decmprs_wqe_counter =
 111                         (cq->decmprs_wqe_counter + 1) & rq->wq.sz_m1;
 112 }
 113
 114 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
 115                                                 struct mlx5e_cq *cq, u32 cqcc)
 116 {
 117         mlx5e_decompress_cqe(rq, cq, cqcc);
 118         cq->title.rss_hash_type   = 0;
 119         cq->title.rss_hash_result = 0;
 120 }
 121
 122 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
 123                                              struct mlx5e_cq *cq,
 124                                              int update_owner_only,
 125                                              int budget_rem)
 126 {
 127         u32 cqcc = cq->wq.cc + update_owner_only;
 128         u32 cqe_count;
 129         u32 i;
 130
 131         cqe_count = min_t(u32, cq->decmprs_left, budget_rem);
 132
 133         for (i = update_owner_only; i < cqe_count;
 134              i++, cq->mini_arr_idx++, cqcc++) {
 135                 if (cq->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
 136                         mlx5e_read_mini_arr_slot(cq, cqcc);
 137
 138                 mlx5e_decompress_cqe_no_hash(rq, cq, cqcc);
 139                 rq->handle_rx_cqe(rq, &cq->title);
 140         }
 141         mlx5e_cqes_update_owner(cq, cq->wq.cc, cqcc - cq->wq.cc);
 142         cq->wq.cc = cqcc;
 143         cq->decmprs_left -= cqe_count;
 144         rq->stats.cqe_compress_pkts += cqe_count;
 145
 146         return cqe_count;
 147 }
 148
 149 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
 150                                               struct mlx5e_cq *cq,
 151                                               int budget_rem)
 152 {
 153         mlx5e_read_title_slot(rq, cq, cq->wq.cc);
 154         mlx5e_read_mini_arr_slot(cq, cq->wq.cc + 1);
 155         mlx5e_decompress_cqe(rq, cq, cq->wq.cc);
 156         rq->handle_rx_cqe(rq, &cq->title);
 157         cq->mini_arr_idx++;
 158
 159         return mlx5e_decompress_cqes_cont(rq, cq, 1, budget_rem) - 1;
 160 }
 161
 162 #define RQ_PAGE_SIZE(rq) ((1 << rq->buff.page_order) << PAGE_SHIFT)
 163
 164 static inline bool mlx5e_page_is_reserved(struct page *page)
 165 {
 166         return page_is_pfmemalloc(page) || page_to_nid(page) != numa_node_id();
 167 }
 168
 169 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
 170                                       struct mlx5e_dma_info *dma_info)
 171 {
 172         struct mlx5e_page_cache *cache = &rq->page_cache;
 173         u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
 174
 175         if (tail_next == cache->head) {
 176                 rq->stats.cache_full++;
 177                 return false;
 178         }
 179
 180         if (unlikely(page_is_pfmemalloc(dma_info->page)))
 181                 return false;
 182
 183         cache->page_cache[cache->tail] = *dma_info;
 184         cache->tail = tail_next;
 185         return true;
 186 }
 187
 188 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
 189                                       struct mlx5e_dma_info *dma_info)
 190 {
 191         struct mlx5e_page_cache *cache = &rq->page_cache;
 192
 193         if (unlikely(cache->head == cache->tail)) {
 194                 rq->stats.cache_empty++;
 195                 return false;
 196         }
 197
 198         if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
 199                 rq->stats.cache_busy++;
 200                 return false;
 201         }
 202
 203         *dma_info = cache->page_cache[cache->head];
 204         cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
 205         rq->stats.cache_reuse++;
 206
 207         dma_sync_single_for_device(rq->pdev, dma_info->addr,
 208                                    RQ_PAGE_SIZE(rq),
 209                                    DMA_FROM_DEVICE);
 210         return true;
 211 }
 212
 213 static inline int mlx5e_page_alloc_mapped(struct mlx5e_rq *rq,
 214                                           struct mlx5e_dma_info *dma_info)
 215 {
 216         struct page *page;
 217
 218         if (mlx5e_rx_cache_get(rq, dma_info))
 219                 return 0;
 220
 221         page = dev_alloc_pages(rq->buff.page_order);
 222         if (unlikely(!page))
 223                 return -ENOMEM;
 224
 225         dma_info->page = page;
 226         dma_info->addr = dma_map_page(rq->pdev, page, 0,
 227                                       RQ_PAGE_SIZE(rq), rq->buff.map_dir);
 228         if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
 229                 put_page(page);
 230                 return -ENOMEM;
 231         }
 232
 233         return 0;
 234 }
 235
 236 void mlx5e_page_release(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info,
 237                         bool recycle)
 238 {
 239         if (likely(recycle) && mlx5e_rx_cache_put(rq, dma_info))
 240                 return;
 241
 242         dma_unmap_page(rq->pdev, dma_info->addr, RQ_PAGE_SIZE(rq),
 243                        rq->buff.map_dir);
 244         put_page(dma_info->page);
 245 }
 246
 247 static inline bool mlx5e_page_reuse(struct mlx5e_rq *rq,
 248                                     struct mlx5e_wqe_frag_info *wi)
 249 {
 250         return rq->wqe.page_reuse && wi->di.page &&
 251                 (wi->offset + rq->wqe.frag_sz <= RQ_PAGE_SIZE(rq)) &&
 252                 !mlx5e_page_is_reserved(wi->di.page);
 253 }
 254
 255 int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
 256 {
 257         struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
 258
 259         /* check if page exists, hence can be reused */
 260         if (!wi->di.page) {
 261                 if (unlikely(mlx5e_page_alloc_mapped(rq, &wi->di)))
 262                         return -ENOMEM;
 263                 wi->offset = 0;
 264         }
 265
 266         wqe->data.addr = cpu_to_be64(wi->di.addr + wi->offset +
 267                                      rq->rx_headroom);
 268         return 0;
 269 }
 270
 271 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
 272                                      struct mlx5e_wqe_frag_info *wi)
 273 {
 274         mlx5e_page_release(rq, &wi->di, true);
 275         wi->di.page = NULL;
 276 }
 277
 278 static inline void mlx5e_free_rx_wqe_reuse(struct mlx5e_rq *rq,
 279                                            struct mlx5e_wqe_frag_info *wi)
 280 {
 281         if (mlx5e_page_reuse(rq, wi)) {
 282                 rq->stats.page_reuse++;
 283                 return;
 284         }
 285
 286         mlx5e_free_rx_wqe(rq, wi);
 287 }
 288
 289 void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
 290 {
 291         struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
 292
 293         if (wi->di.page)
 294                 mlx5e_free_rx_wqe(rq, wi);
 295 }
 296
 297 static inline int mlx5e_mpwqe_strides_per_page(struct mlx5e_rq *rq)
 298 {
 299         return rq->mpwqe_num_strides >> MLX5_MPWRQ_WQE_PAGE_ORDER;
 300 }
 301
 302 static inline void mlx5e_add_skb_frag_mpwqe(struct mlx5e_rq *rq,
 303                                             struct sk_buff *skb,
 304                                             struct mlx5e_mpw_info *wi,
 305                                             u32 page_idx, u32 frag_offset,
 306                                             u32 len)
 307 {
 308         unsigned int truesize = ALIGN(len, rq->mpwqe_stride_sz);
 309
 310         dma_sync_single_for_cpu(rq->pdev,
 311                                 wi->umr.dma_info[page_idx].addr + frag_offset,
 312                                 len, DMA_FROM_DEVICE);
 313         wi->skbs_frags[page_idx]++;
 314         skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
 315                         wi->umr.dma_info[page_idx].page, frag_offset,
 316                         len, truesize);
 317 }
 318
 319 static inline void
 320 mlx5e_copy_skb_header_mpwqe(struct device *pdev,
 321                             struct sk_buff *skb,
 322                             struct mlx5e_mpw_info *wi,
 323                             u32 page_idx, u32 offset,
 324                             u32 headlen)
 325 {
 326         u16 headlen_pg = min_t(u32, headlen, PAGE_SIZE - offset);
 327         struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[page_idx];
 328         unsigned int len;
 329
 330          /* Aligning len to sizeof(long) optimizes memcpy performance */
 331         len = ALIGN(headlen_pg, sizeof(long));
 332         dma_sync_single_for_cpu(pdev, dma_info->addr + offset, len,
 333                                 DMA_FROM_DEVICE);
 334         skb_copy_to_linear_data_offset(skb, 0,
 335                                        page_address(dma_info->page) + offset,
 336                                        len);
 337         if (unlikely(offset + headlen > PAGE_SIZE)) {
 338                 dma_info++;
 339                 headlen_pg = len;
 340                 len = ALIGN(headlen - headlen_pg, sizeof(long));
 341                 dma_sync_single_for_cpu(pdev, dma_info->addr, len,
 342                                         DMA_FROM_DEVICE);
 343                 skb_copy_to_linear_data_offset(skb, headlen_pg,
 344                                                page_address(dma_info->page),
 345                                                len);
 346         }
 347 }
 348
 349 static inline void mlx5e_post_umr_wqe(struct mlx5e_rq *rq, u16 ix)
 350 {
 351         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 352         struct mlx5e_icosq *sq = &rq->channel->icosq;
 353         struct mlx5_wq_cyc *wq = &sq->wq;
 354         struct mlx5e_umr_wqe *wqe;
 355         u8 num_wqebbs = DIV_ROUND_UP(sizeof(*wqe), MLX5_SEND_WQE_BB);
 356         u16 pi;
 357
 358         /* fill sq edge with nops to avoid wqe wrap around */
 359         while ((pi = (sq->pc & wq->sz_m1)) > sq->edge) {
 360                 sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_NOP;
 361                 sq->db.ico_wqe[pi].num_wqebbs = 1;
 362                 mlx5e_post_nop(wq, sq->sqn, &sq->pc);
 363         }
 364
 365         wqe = mlx5_wq_cyc_get_wqe(wq, pi);
 366         memcpy(wqe, &wi->umr.wqe, sizeof(*wqe));
 367         wqe->ctrl.opmod_idx_opcode =
 368                 cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
 369                             MLX5_OPCODE_UMR);
 370
 371         sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_UMR;
 372         sq->db.ico_wqe[pi].num_wqebbs = num_wqebbs;
 373         sq->pc += num_wqebbs;
 374         mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &wqe->ctrl);
 375 }
 376
 377 static int mlx5e_alloc_rx_umr_mpwqe(struct mlx5e_rq *rq,
 378                                     struct mlx5e_rx_wqe *wqe,
 379                                     u16 ix)
 380 {
 381         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 382         u64 dma_offset = (u64)mlx5e_get_wqe_mtt_offset(rq, ix) << PAGE_SHIFT;
 383         int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
 384         int err;
 385         int i;
 386
 387         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++) {
 388                 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[i];
 389
 390                 err = mlx5e_page_alloc_mapped(rq, dma_info);
 391                 if (unlikely(err))
 392                         goto err_unmap;
 393                 wi->umr.mtt[i] = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
 394                 page_ref_add(dma_info->page, pg_strides);
 395                 wi->skbs_frags[i] = 0;
 396         }
 397
 398         wi->consumed_strides = 0;
 399         wqe->data.addr = cpu_to_be64(dma_offset);
 400
 401         return 0;
 402
 403 err_unmap:
 404         while (--i >= 0) {
 405                 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[i];
 406
 407                 page_ref_sub(dma_info->page, pg_strides);
 408                 mlx5e_page_release(rq, dma_info, true);
 409         }
 410
 411         return err;
 412 }
 413
 414 void mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
 415 {
 416         int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
 417         int i;
 418
 419         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++) {
 420                 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[i];
 421
 422                 page_ref_sub(dma_info->page, pg_strides - wi->skbs_frags[i]);
 423                 mlx5e_page_release(rq, dma_info, true);
 424         }
 425 }
 426
 427 void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
 428 {
 429         struct mlx5_wq_ll *wq = &rq->wq;
 430         struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
 431
 432         clear_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state);
 433
 434         if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state))) {
 435                 mlx5e_free_rx_mpwqe(rq, &rq->mpwqe.info[wq->head]);
 436                 return;
 437         }
 438
 439         mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
 440
 441         /* ensure wqes are visible to device before updating doorbell record */
 442         dma_wmb();
 443
 444         mlx5_wq_ll_update_db_record(wq);
 445 }
 446
 447 int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
 448 {
 449         int err;
 450
 451         err = mlx5e_alloc_rx_umr_mpwqe(rq, wqe, ix);
 452         if (unlikely(err))
 453                 return err;
 454         set_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state);
 455         mlx5e_post_umr_wqe(rq, ix);
 456         return -EBUSY;
 457 }
 458
 459 void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 460 {
 461         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 462
 463         mlx5e_free_rx_mpwqe(rq, wi);
 464 }
 465
 466 #define RQ_CANNOT_POST(rq) \
 467         (!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state) || \
 468          test_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state))
 469
 470 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
 471 {
 472         struct mlx5_wq_ll *wq = &rq->wq;
 473
 474         if (unlikely(RQ_CANNOT_POST(rq)))
 475                 return false;
 476
 477         while (!mlx5_wq_ll_is_full(wq)) {
 478                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
 479                 int err;
 480
 481                 err = rq->alloc_wqe(rq, wqe, wq->head);
 482                 if (err == -EBUSY)
 483                         return true;
 484                 if (unlikely(err)) {
 485                         rq->stats.buff_alloc_err++;
 486                         break;
 487                 }
 488
 489                 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
 490         }
 491
 492         /* ensure wqes are visible to device before updating doorbell record */
 493         dma_wmb();
 494
 495         mlx5_wq_ll_update_db_record(wq);
 496
 497         return !mlx5_wq_ll_is_full(wq);
 498 }
 499
 500 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
 501                                  u32 cqe_bcnt)
 502 {
 503         struct ethhdr   *eth = (struct ethhdr *)(skb->data);
 504         struct iphdr    *ipv4;
 505         struct ipv6hdr  *ipv6;
 506         struct tcphdr   *tcp;
 507         int network_depth = 0;
 508         __be16 proto;
 509         u16 tot_len;
 510
 511         u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
 512         int tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA  == l4_hdr_type) ||
 513                        (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
 514
 515         skb->mac_len = ETH_HLEN;
 516         proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
 517
 518         ipv4 = (struct iphdr *)(skb->data + network_depth);
 519         ipv6 = (struct ipv6hdr *)(skb->data + network_depth);
 520         tot_len = cqe_bcnt - network_depth;
 521
 522         if (proto == htons(ETH_P_IP)) {
 523                 tcp = (struct tcphdr *)(skb->data + network_depth +
 524                                         sizeof(struct iphdr));
 525                 ipv6 = NULL;
 526                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 527         } else {
 528                 tcp = (struct tcphdr *)(skb->data + network_depth +
 529                                         sizeof(struct ipv6hdr));
 530                 ipv4 = NULL;
 531                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 532         }
 533
 534         if (get_cqe_lro_tcppsh(cqe))
 535                 tcp->psh                = 1;
 536
 537         if (tcp_ack) {
 538                 tcp->ack                = 1;
 539                 tcp->ack_seq            = cqe->lro_ack_seq_num;
 540                 tcp->window             = cqe->lro_tcp_win;
 541         }
 542
 543         if (ipv4) {
 544                 ipv4->ttl               = cqe->lro_min_ttl;
 545                 ipv4->tot_len           = cpu_to_be16(tot_len);
 546                 ipv4->check             = 0;
 547                 ipv4->check             = ip_fast_csum((unsigned char *)ipv4,
 548                                                        ipv4->ihl);
 549         } else {
 550                 ipv6->hop_limit         = cqe->lro_min_ttl;
 551                 ipv6->payload_len       = cpu_to_be16(tot_len -
 552                                                       sizeof(struct ipv6hdr));
 553         }
 554 }
 555
 556 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
 557                                       struct sk_buff *skb)
 558 {
 559         u8 cht = cqe->rss_hash_type;
 560         int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
 561                  (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
 562                                             PKT_HASH_TYPE_NONE;
 563         skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
 564 }
 565
 566 static inline bool is_first_ethertype_ip(struct sk_buff *skb)
 567 {
 568         __be16 ethertype = ((struct ethhdr *)skb->data)->h_proto;
 569
 570         return (ethertype == htons(ETH_P_IP) || ethertype == htons(ETH_P_IPV6));
 571 }
 572
 573 static inline void mlx5e_handle_csum(struct net_device *netdev,
 574                                      struct mlx5_cqe64 *cqe,
 575                                      struct mlx5e_rq *rq,
 576                                      struct sk_buff *skb,
 577                                      bool   lro)
 578 {
 579         if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
 580                 goto csum_none;
 581
 582         if (lro) {
 583                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 584                 return;
 585         }
 586
 587         if (is_first_ethertype_ip(skb)) {
 588                 skb->ip_summed = CHECKSUM_COMPLETE;
 589                 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
 590                 rq->stats.csum_complete++;
 591                 return;
 592         }
 593
 594         if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
 595                    (cqe->hds_ip_ext & CQE_L4_OK))) {
 596                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 597                 if (cqe_is_tunneled(cqe)) {
 598                         skb->csum_level = 1;
 599                         skb->encapsulation = 1;
 600                         rq->stats.csum_unnecessary_inner++;
 601                 }
 602                 return;
 603         }
 604 csum_none:
 605         skb->ip_summed = CHECKSUM_NONE;
 606         rq->stats.csum_none++;
 607 }
 608
 609 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
 610                                       u32 cqe_bcnt,
 611                                       struct mlx5e_rq *rq,
 612                                       struct sk_buff *skb)
 613 {
 614         struct net_device *netdev = rq->netdev;
 615         struct mlx5e_tstamp *tstamp = rq->tstamp;
 616         int lro_num_seg;
 617
 618         lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
 619         if (lro_num_seg > 1) {
 620                 mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
 621                 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
 622                 /* Subtract one since we already counted this as one
 623                  * "regular" packet in mlx5e_complete_rx_cqe()
 624                  */
 625                 rq->stats.packets += lro_num_seg - 1;
 626                 rq->stats.lro_packets++;
 627                 rq->stats.lro_bytes += cqe_bcnt;
 628         }
 629
 630         if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
 631                 mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
 632
 633         skb_record_rx_queue(skb, rq->ix);
 634
 635         if (likely(netdev->features & NETIF_F_RXHASH))
 636                 mlx5e_skb_set_hash(cqe, skb);
 637
 638         if (cqe_has_vlan(cqe))
 639                 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 640                                        be16_to_cpu(cqe->vlan_info));
 641
 642         skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
 643
 644         mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
 645         skb->protocol = eth_type_trans(skb, netdev);
 646 }
 647
 648 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
 649                                          struct mlx5_cqe64 *cqe,
 650                                          u32 cqe_bcnt,
 651                                          struct sk_buff *skb)
 652 {
 653         rq->stats.packets++;
 654         rq->stats.bytes += cqe_bcnt;
 655         mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
 656 }
 657
 658 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_xdpsq *sq)
 659 {
 660         struct mlx5_wq_cyc *wq = &sq->wq;
 661         struct mlx5e_tx_wqe *wqe;
 662         u16 pi = (sq->pc - 1) & wq->sz_m1; /* last pi */
 663
 664         wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
 665
 666         mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &wqe->ctrl);
 667 }
 668
 669 static inline bool mlx5e_xmit_xdp_frame(struct mlx5e_rq *rq,
 670                                         struct mlx5e_dma_info *di,
 671                                         const struct xdp_buff *xdp)
 672 {
 673         struct mlx5e_xdpsq       *sq   = &rq->xdpsq;
 674         struct mlx5_wq_cyc       *wq   = &sq->wq;
 675         u16                       pi   = sq->pc & wq->sz_m1;
 676         struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
 677
 678         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
 679         struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
 680         struct mlx5_wqe_data_seg *dseg;
 681
 682         ptrdiff_t data_offset = xdp->data - xdp->data_hard_start;
 683         dma_addr_t dma_addr  = di->addr + data_offset;
 684         unsigned int dma_len = xdp->data_end - xdp->data;
 685
 686         prefetchw(wqe);
 687
 688         if (unlikely(dma_len < MLX5E_XDP_MIN_INLINE ||
 689                      MLX5E_SW2HW_MTU(rq->channel->priv, rq->netdev->mtu) < dma_len)) {
 690                 rq->stats.xdp_drop++;
 691                 return false;
 692         }
 693
 694         if (unlikely(!mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1))) {
 695                 if (sq->db.doorbell) {
 696                         /* SQ is full, ring doorbell */
 697                         mlx5e_xmit_xdp_doorbell(sq);
 698                         sq->db.doorbell = false;
 699                 }
 700                 rq->stats.xdp_tx_full++;
 701                 return false;
 702         }
 703
 704         dma_sync_single_for_device(sq->pdev, dma_addr, dma_len, PCI_DMA_TODEVICE);
 705
 706         cseg->fm_ce_se = 0;
 707
 708         dseg = (struct mlx5_wqe_data_seg *)eseg + 1;
 709
 710         /* copy the inline part if required */
 711         if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
 712                 memcpy(eseg->inline_hdr.start, xdp->data, MLX5E_XDP_MIN_INLINE);
 713                 eseg->inline_hdr.sz = cpu_to_be16(MLX5E_XDP_MIN_INLINE);
 714                 dma_len  -= MLX5E_XDP_MIN_INLINE;
 715                 dma_addr += MLX5E_XDP_MIN_INLINE;
 716                 dseg++;
 717         }
 718
 719         /* write the dma part */
 720         dseg->addr       = cpu_to_be64(dma_addr);
 721         dseg->byte_count = cpu_to_be32(dma_len);
 722
 723         cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_SEND);
 724
 725         /* move page to reference to sq responsibility,
 726          * and mark so it's not put back in page-cache.
 727          */
 728         rq->wqe.xdp_xmit = true;
 729         sq->db.di[pi] = *di;
 730         sq->pc++;
 731
 732         sq->db.doorbell = true;
 733
 734         rq->stats.xdp_tx++;
 735         return true;
 736 }
 737
 738 /* returns true if packet was consumed by xdp */
 739 static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq,
 740                                    struct mlx5e_dma_info *di,
 741                                    void *va, u16 *rx_headroom, u32 *len)
 742 {
 743         const struct bpf_prog *prog = READ_ONCE(rq->xdp_prog);
 744         struct xdp_buff xdp;
 745         u32 act;
 746
 747         if (!prog)
 748                 return false;
 749
 750         xdp.data = va + *rx_headroom;
 751         xdp.data_end = xdp.data + *len;
 752         xdp.data_hard_start = va;
 753
 754         act = bpf_prog_run_xdp(prog, &xdp);
 755         switch (act) {
 756         case XDP_PASS:
 757                 *rx_headroom = xdp.data - xdp.data_hard_start;
 758                 *len = xdp.data_end - xdp.data;
 759                 return false;
 760         case XDP_TX:
 761                 if (unlikely(!mlx5e_xmit_xdp_frame(rq, di, &xdp)))
 762                         trace_xdp_exception(rq->netdev, prog, act);
 763                 return true;
 764         default:
 765                 bpf_warn_invalid_xdp_action(act);
 766         case XDP_ABORTED:
 767                 trace_xdp_exception(rq->netdev, prog, act);
 768         case XDP_DROP:
 769                 rq->stats.xdp_drop++;
 770                 return true;
 771         }
 772 }
 773
 774 static inline
 775 struct sk_buff *skb_from_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
 776                              struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
 777 {
 778         struct mlx5e_dma_info *di = &wi->di;
 779         struct sk_buff *skb;
 780         void *va, *data;
 781         u16 rx_headroom = rq->rx_headroom;
 782         bool consumed;
 783         u32 frag_size;
 784
 785         va             = page_address(di->page) + wi->offset;
 786         data           = va + rx_headroom;
 787         frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
 788
 789         dma_sync_single_range_for_cpu(rq->pdev,
 790                                       di->addr + wi->offset,
 791                                       0, frag_size,
 792                                       DMA_FROM_DEVICE);
 793         prefetch(data);
 794         wi->offset += frag_size;
 795
 796         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
 797                 rq->stats.wqe_err++;
 798                 return NULL;
 799         }
 800
 801         rcu_read_lock();
 802         consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt);
 803         rcu_read_unlock();
 804         if (consumed)
 805                 return NULL; /* page/packet was consumed by XDP */
 806
 807         skb = build_skb(va, frag_size);
 808         if (unlikely(!skb)) {
 809                 rq->stats.buff_alloc_err++;
 810                 return NULL;
 811         }
 812
 813         /* queue up for recycling/reuse */
 814         page_ref_inc(di->page);
 815
 816         skb_reserve(skb, rx_headroom);
 817         skb_put(skb, cqe_bcnt);
 818
 819         return skb;
 820 }
 821
 822 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 823 {
 824         struct mlx5e_wqe_frag_info *wi;
 825         struct mlx5e_rx_wqe *wqe;
 826         __be16 wqe_counter_be;
 827         struct sk_buff *skb;
 828         u16 wqe_counter;
 829         u32 cqe_bcnt;
 830
 831         wqe_counter_be = cqe->wqe_counter;
 832         wqe_counter    = be16_to_cpu(wqe_counter_be);
 833         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
 834         wi             = &rq->wqe.frag_info[wqe_counter];
 835         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
 836
 837         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
 838         if (!skb) {
 839                 /* probably for XDP */
 840                 if (rq->wqe.xdp_xmit) {
 841                         wi->di.page = NULL;
 842                         rq->wqe.xdp_xmit = false;
 843                         /* do not return page to cache, it will be returned on XDP_TX completion */
 844                         goto wq_ll_pop;
 845                 }
 846                 /* probably an XDP_DROP, save the page-reuse checks */
 847                 mlx5e_free_rx_wqe(rq, wi);
 848                 goto wq_ll_pop;
 849         }
 850
 851         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 852         napi_gro_receive(rq->cq.napi, skb);
 853
 854         mlx5e_free_rx_wqe_reuse(rq, wi);
 855 wq_ll_pop:
 856         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
 857                        &wqe->next.next_wqe_index);
 858 }
 859
 860 void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 861 {
 862         struct net_device *netdev = rq->netdev;
 863         struct mlx5e_priv *priv = netdev_priv(netdev);
 864         struct mlx5e_rep_priv *rpriv  = priv->ppriv;
 865         struct mlx5_eswitch_rep *rep = rpriv->rep;
 866         struct mlx5e_wqe_frag_info *wi;
 867         struct mlx5e_rx_wqe *wqe;
 868         struct sk_buff *skb;
 869         __be16 wqe_counter_be;
 870         u16 wqe_counter;
 871         u32 cqe_bcnt;
 872
 873         wqe_counter_be = cqe->wqe_counter;
 874         wqe_counter    = be16_to_cpu(wqe_counter_be);
 875         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
 876         wi             = &rq->wqe.frag_info[wqe_counter];
 877         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
 878
 879         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
 880         if (!skb) {
 881                 if (rq->wqe.xdp_xmit) {
 882                         wi->di.page = NULL;
 883                         rq->wqe.xdp_xmit = false;
 884                         /* do not return page to cache, it will be returned on XDP_TX completion */
 885                         goto wq_ll_pop;
 886                 }
 887                 /* probably an XDP_DROP, save the page-reuse checks */
 888                 mlx5e_free_rx_wqe(rq, wi);
 889                 goto wq_ll_pop;
 890         }
 891
 892         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 893
 894         if (rep->vlan && skb_vlan_tag_present(skb))
 895                 skb_vlan_pop(skb);
 896
 897         napi_gro_receive(rq->cq.napi, skb);
 898
 899         mlx5e_free_rx_wqe_reuse(rq, wi);
 900 wq_ll_pop:
 901         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
 902                        &wqe->next.next_wqe_index);
 903 }
 904
 905 static inline void mlx5e_mpwqe_fill_rx_skb(struct mlx5e_rq *rq,
 906                                            struct mlx5_cqe64 *cqe,
 907                                            struct mlx5e_mpw_info *wi,
 908                                            u32 cqe_bcnt,
 909                                            struct sk_buff *skb)
 910 {
 911         u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
 912         u32 wqe_offset     = stride_ix * rq->mpwqe_stride_sz;
 913         u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
 914         u32 page_idx       = wqe_offset >> PAGE_SHIFT;
 915         u32 head_page_idx  = page_idx;
 916         u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
 917         u32 frag_offset    = head_offset + headlen;
 918         u16 byte_cnt       = cqe_bcnt - headlen;
 919
 920         if (unlikely(frag_offset >= PAGE_SIZE)) {
 921                 page_idx++;
 922                 frag_offset -= PAGE_SIZE;
 923         }
 924
 925         while (byte_cnt) {
 926                 u32 pg_consumed_bytes =
 927                         min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
 928
 929                 mlx5e_add_skb_frag_mpwqe(rq, skb, wi, page_idx, frag_offset,
 930                                          pg_consumed_bytes);
 931                 byte_cnt -= pg_consumed_bytes;
 932                 frag_offset = 0;
 933                 page_idx++;
 934         }
 935         /* copy header */
 936         mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, wi, head_page_idx,
 937                                     head_offset, headlen);
 938         /* skb linear part was allocated with headlen and aligned to long */
 939         skb->tail += headlen;
 940         skb->len  += headlen;
 941 }
 942
 943 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 944 {
 945         u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
 946         u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
 947         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
 948         struct mlx5e_rx_wqe  *wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_id);
 949         struct sk_buff *skb;
 950         u16 cqe_bcnt;
 951
 952         wi->consumed_strides += cstrides;
 953
 954         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
 955                 rq->stats.wqe_err++;
 956                 goto mpwrq_cqe_out;
 957         }
 958
 959         if (unlikely(mpwrq_is_filler_cqe(cqe))) {
 960                 rq->stats.mpwqe_filler++;
 961                 goto mpwrq_cqe_out;
 962         }
 963
 964         skb = napi_alloc_skb(rq->cq.napi,
 965                              ALIGN(MLX5_MPWRQ_SMALL_PACKET_THRESHOLD,
 966                                    sizeof(long)));
 967         if (unlikely(!skb)) {
 968                 rq->stats.buff_alloc_err++;
 969                 goto mpwrq_cqe_out;
 970         }
 971
 972         prefetchw(skb->data);
 973         cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
 974
 975         mlx5e_mpwqe_fill_rx_skb(rq, cqe, wi, cqe_bcnt, skb);
 976         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 977         napi_gro_receive(rq->cq.napi, skb);
 978
 979 mpwrq_cqe_out:
 980         if (likely(wi->consumed_strides < rq->mpwqe_num_strides))
 981                 return;
 982
 983         mlx5e_free_rx_mpwqe(rq, wi);
 984         mlx5_wq_ll_pop(&rq->wq, cqe->wqe_id, &wqe->next.next_wqe_index);
 985 }
 986
 987 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
 988 {
 989         struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
 990         struct mlx5e_xdpsq *xdpsq = &rq->xdpsq;
 991         int work_done = 0;
 992
 993         if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
 994                 return 0;
 995
 996         if (cq->decmprs_left)
 997                 work_done += mlx5e_decompress_cqes_cont(rq, cq, 0, budget);
 998
 999         for (; work_done < budget; work_done++) {
1000                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_cqe(&cq->wq);
1001
1002                 if (!cqe)
1003                         break;
1004
1005                 if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1006                         work_done +=
1007                                 mlx5e_decompress_cqes_start(rq, cq,
1008                                                             budget - work_done);
1009                         continue;
1010                 }
1011
1012                 mlx5_cqwq_pop(&cq->wq);
1013
1014                 rq->handle_rx_cqe(rq, cqe);
1015         }
1016
1017         if (xdpsq->db.doorbell) {
1018                 mlx5e_xmit_xdp_doorbell(xdpsq);
1019                 xdpsq->db.doorbell = false;
1020         }
1021
1022         mlx5_cqwq_update_db_record(&cq->wq);
1023
1024         /* ensure cq space is freed before enabling more cqes */
1025         wmb();
1026
1027         return work_done;
1028 }
1029
1030 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
1031 {
1032         struct mlx5e_xdpsq *sq;
1033         struct mlx5e_rq *rq;
1034         u16 sqcc;
1035         int i;
1036
1037         sq = container_of(cq, struct mlx5e_xdpsq, cq);
1038
1039         if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &sq->state)))
1040                 return false;
1041
1042         rq = container_of(sq, struct mlx5e_rq, xdpsq);
1043
1044         /* sq->cc must be updated only after mlx5_cqwq_update_db_record(),
1045          * otherwise a cq overrun may occur
1046          */
1047         sqcc = sq->cc;
1048
1049         for (i = 0; i < MLX5E_TX_CQ_POLL_BUDGET; i++) {
1050                 struct mlx5_cqe64 *cqe;
1051                 u16 wqe_counter;
1052                 bool last_wqe;
1053
1054                 cqe = mlx5_cqwq_get_cqe(&cq->wq);
1055                 if (!cqe)
1056                         break;
1057
1058                 mlx5_cqwq_pop(&cq->wq);
1059
1060                 wqe_counter = be16_to_cpu(cqe->wqe_counter);
1061
1062                 do {
1063                         struct mlx5e_dma_info *di;
1064                         u16 ci;
1065
1066                         last_wqe = (sqcc == wqe_counter);
1067
1068                         ci = sqcc & sq->wq.sz_m1;
1069                         di = &sq->db.di[ci];
1070
1071                         sqcc++;
1072                         /* Recycle RX page */
1073                         mlx5e_page_release(rq, di, true);
1074                 } while (!last_wqe);
1075         }
1076
1077         mlx5_cqwq_update_db_record(&cq->wq);
1078
1079         /* ensure cq space is freed before enabling more cqes */
1080         wmb();
1081
1082         sq->cc = sqcc;
1083         return (i == MLX5E_TX_CQ_POLL_BUDGET);
1084 }
1085
1086 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
1087 {
1088         struct mlx5e_rq *rq = container_of(sq, struct mlx5e_rq, xdpsq);
1089         struct mlx5e_dma_info *di;
1090         u16 ci;
1091
1092         while (sq->cc != sq->pc) {
1093                 ci = sq->cc & sq->wq.sz_m1;
1094                 di = &sq->db.di[ci];
1095                 sq->cc++;
1096
1097                 mlx5e_page_release(rq, di, false);
1098         }
1099 }
1100
1101 #ifdef CONFIG_MLX5_CORE_IPOIB
1102
1103 #define MLX5_IB_GRH_DGID_OFFSET 24
1104 #define MLX5_GID_SIZE           16
1105
1106 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1107                                          struct mlx5_cqe64 *cqe,
1108                                          u32 cqe_bcnt,
1109                                          struct sk_buff *skb)
1110 {
1111         struct net_device *netdev = rq->netdev;
1112         struct mlx5e_tstamp *tstamp = rq->tstamp;
1113         char *pseudo_header;
1114         u8 *dgid;
1115         u8 g;
1116
1117         g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
1118         dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1119         if ((!g) || dgid[0] != 0xff)
1120                 skb->pkt_type = PACKET_HOST;
1121         else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1122                 skb->pkt_type = PACKET_BROADCAST;
1123         else
1124                 skb->pkt_type = PACKET_MULTICAST;
1125
1126         /* TODO: IB/ipoib: Allow mcast packets from other VFs
1127          * 68996a6e760e5c74654723eeb57bf65628ae87f4
1128          */
1129
1130         skb_pull(skb, MLX5_IB_GRH_BYTES);
1131
1132         skb->protocol = *((__be16 *)(skb->data));
1133
1134         skb->ip_summed = CHECKSUM_COMPLETE;
1135         skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1136
1137         if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1138                 mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
1139
1140         skb_record_rx_queue(skb, rq->ix);
1141
1142         if (likely(netdev->features & NETIF_F_RXHASH))
1143                 mlx5e_skb_set_hash(cqe, skb);
1144
1145         /* 20 bytes of ipoib header and 4 for encap existing */
1146         pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1147         memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1148         skb_reset_mac_header(skb);
1149         skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1150
1151         skb->dev = netdev;
1152
1153         rq->stats.csum_complete++;
1154         rq->stats.packets++;
1155         rq->stats.bytes += cqe_bcnt;
1156 }
1157
1158 void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1159 {
1160         struct mlx5e_wqe_frag_info *wi;
1161         struct mlx5e_rx_wqe *wqe;
1162         __be16 wqe_counter_be;
1163         struct sk_buff *skb;
1164         u16 wqe_counter;
1165         u32 cqe_bcnt;
1166
1167         wqe_counter_be = cqe->wqe_counter;
1168         wqe_counter    = be16_to_cpu(wqe_counter_be);
1169         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1170         wi             = &rq->wqe.frag_info[wqe_counter];
1171         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1172
1173         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1174         if (!skb)
1175                 goto wq_free_wqe;
1176
1177         mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1178         napi_gro_receive(rq->cq.napi, skb);
1179
1180 wq_free_wqe:
1181         mlx5e_free_rx_wqe_reuse(rq, wi);
1182         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1183                        &wqe->next.next_wqe_index);
1184 }
1185
1186 #endif /* CONFIG_MLX5_CORE_IPOIB */
1187
1188 #ifdef CONFIG_MLX5_EN_IPSEC
1189
1190 void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1191 {
1192         struct mlx5e_wqe_frag_info *wi;
1193         struct mlx5e_rx_wqe *wqe;
1194         __be16 wqe_counter_be;
1195         struct sk_buff *skb;
1196         u16 wqe_counter;
1197         u32 cqe_bcnt;
1198
1199         wqe_counter_be = cqe->wqe_counter;
1200         wqe_counter    = be16_to_cpu(wqe_counter_be);
1201         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1202         wi             = &rq->wqe.frag_info[wqe_counter];
1203         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1204
1205         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1206         if (unlikely(!skb)) {
1207                 /* a DROP, save the page-reuse checks */
1208                 mlx5e_free_rx_wqe(rq, wi);
1209                 goto wq_ll_pop;
1210         }
1211         skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb);
1212         if (unlikely(!skb)) {
1213                 mlx5e_free_rx_wqe(rq, wi);
1214                 goto wq_ll_pop;
1215         }
1216
1217         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1218         napi_gro_receive(rq->cq.napi, skb);
1219
1220         mlx5e_free_rx_wqe_reuse(rq, wi);
1221 wq_ll_pop:
1222         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1223                        &wqe->next.next_wqe_index);
1224 }
1225
1226 #endif /* CONFIG_MLX5_EN_IPSEC */