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/ip6_checksum.h>
  37 #include <net/page_pool.h>
  38 #include <net/inet_ecn.h>
  39 #include "en.h"
  40 #include "en/txrx.h"
  41 #include "en_tc.h"
  42 #include "eswitch.h"
  43 #include "en_rep.h"
  44 #include "en/rep/tc.h"
  45 #include "ipoib/ipoib.h"
  46 #include "accel/ipsec.h"
  47 #include "fpga/ipsec.h"
  48 #include "en_accel/ipsec_rxtx.h"
  49 #include "en_accel/tls_rxtx.h"
  50 #include "en/xdp.h"
  51 #include "en/xsk/rx.h"
  52 #include "en/health.h"
  53 #include "en/params.h"
  54 #include "devlink.h"
  55 #include "en/devlink.h"
  56
  57 static struct sk_buff *
  58 mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
  59                                 u16 cqe_bcnt, u32 head_offset, u32 page_idx);
  60 static struct sk_buff *
  61 mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
  62                                    u16 cqe_bcnt, u32 head_offset, u32 page_idx);
  63 static void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe);
  64 static void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe);
  65
  66 const struct mlx5e_rx_handlers mlx5e_rx_handlers_nic = {
  67         .handle_rx_cqe       = mlx5e_handle_rx_cqe,
  68         .handle_rx_cqe_mpwqe = mlx5e_handle_rx_cqe_mpwrq,
  69 };
  70
  71 static inline bool mlx5e_rx_hw_stamp(struct hwtstamp_config *config)
  72 {
  73         return config->rx_filter == HWTSTAMP_FILTER_ALL;
  74 }
  75
  76 static inline void mlx5e_read_cqe_slot(struct mlx5_cqwq *wq,
  77                                        u32 cqcc, void *data)
  78 {
  79         u32 ci = mlx5_cqwq_ctr2ix(wq, cqcc);
  80
  81         memcpy(data, mlx5_cqwq_get_wqe(wq, ci), sizeof(struct mlx5_cqe64));
  82 }
  83
  84 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
  85                                          struct mlx5_cqwq *wq,
  86                                          u32 cqcc)
  87 {
  88         struct mlx5e_cq_decomp *cqd = &rq->cqd;
  89         struct mlx5_cqe64 *title = &cqd->title;
  90
  91         mlx5e_read_cqe_slot(wq, cqcc, title);
  92         cqd->left        = be32_to_cpu(title->byte_cnt);
  93         cqd->wqe_counter = be16_to_cpu(title->wqe_counter);
  94         rq->stats->cqe_compress_blks++;
  95 }
  96
  97 static inline void mlx5e_read_mini_arr_slot(struct mlx5_cqwq *wq,
  98                                             struct mlx5e_cq_decomp *cqd,
  99                                             u32 cqcc)
 100 {
 101         mlx5e_read_cqe_slot(wq, cqcc, cqd->mini_arr);
 102         cqd->mini_arr_idx = 0;
 103 }
 104
 105 static inline void mlx5e_cqes_update_owner(struct mlx5_cqwq *wq, int n)
 106 {
 107         u32 cqcc   = wq->cc;
 108         u8  op_own = mlx5_cqwq_get_ctr_wrap_cnt(wq, cqcc) & 1;
 109         u32 ci     = mlx5_cqwq_ctr2ix(wq, cqcc);
 110         u32 wq_sz  = mlx5_cqwq_get_size(wq);
 111         u32 ci_top = min_t(u32, wq_sz, ci + n);
 112
 113         for (; ci < ci_top; ci++, n--) {
 114                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(wq, ci);
 115
 116                 cqe->op_own = op_own;
 117         }
 118
 119         if (unlikely(ci == wq_sz)) {
 120                 op_own = !op_own;
 121                 for (ci = 0; ci < n; ci++) {
 122                         struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(wq, ci);
 123
 124                         cqe->op_own = op_own;
 125                 }
 126         }
 127 }
 128
 129 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
 130                                         struct mlx5_cqwq *wq,
 131                                         u32 cqcc)
 132 {
 133         struct mlx5e_cq_decomp *cqd = &rq->cqd;
 134         struct mlx5_mini_cqe8 *mini_cqe = &cqd->mini_arr[cqd->mini_arr_idx];
 135         struct mlx5_cqe64 *title = &cqd->title;
 136
 137         title->byte_cnt     = mini_cqe->byte_cnt;
 138         title->check_sum    = mini_cqe->checksum;
 139         title->op_own      &= 0xf0;
 140         title->op_own      |= 0x01 & (cqcc >> wq->fbc.log_sz);
 141
 142         /* state bit set implies linked-list striding RQ wq type and
 143          * HW stride index capability supported
 144          */
 145         if (test_bit(MLX5E_RQ_STATE_MINI_CQE_HW_STRIDX, &rq->state)) {
 146                 title->wqe_counter = mini_cqe->stridx;
 147                 return;
 148         }
 149
 150         /* HW stride index capability not supported */
 151         title->wqe_counter = cpu_to_be16(cqd->wqe_counter);
 152         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
 153                 cqd->wqe_counter += mpwrq_get_cqe_consumed_strides(title);
 154         else
 155                 cqd->wqe_counter =
 156                         mlx5_wq_cyc_ctr2ix(&rq->wqe.wq, cqd->wqe_counter + 1);
 157 }
 158
 159 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
 160                                                 struct mlx5_cqwq *wq,
 161                                                 u32 cqcc)
 162 {
 163         struct mlx5e_cq_decomp *cqd = &rq->cqd;
 164
 165         mlx5e_decompress_cqe(rq, wq, cqcc);
 166         cqd->title.rss_hash_type   = 0;
 167         cqd->title.rss_hash_result = 0;
 168 }
 169
 170 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
 171                                              struct mlx5_cqwq *wq,
 172                                              int update_owner_only,
 173                                              int budget_rem)
 174 {
 175         struct mlx5e_cq_decomp *cqd = &rq->cqd;
 176         u32 cqcc = wq->cc + update_owner_only;
 177         u32 cqe_count;
 178         u32 i;
 179
 180         cqe_count = min_t(u32, cqd->left, budget_rem);
 181
 182         for (i = update_owner_only; i < cqe_count;
 183              i++, cqd->mini_arr_idx++, cqcc++) {
 184                 if (cqd->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
 185                         mlx5e_read_mini_arr_slot(wq, cqd, cqcc);
 186
 187                 mlx5e_decompress_cqe_no_hash(rq, wq, cqcc);
 188                 INDIRECT_CALL_2(rq->handle_rx_cqe, mlx5e_handle_rx_cqe_mpwrq,
 189                                 mlx5e_handle_rx_cqe, rq, &cqd->title);
 190         }
 191         mlx5e_cqes_update_owner(wq, cqcc - wq->cc);
 192         wq->cc = cqcc;
 193         cqd->left -= cqe_count;
 194         rq->stats->cqe_compress_pkts += cqe_count;
 195
 196         return cqe_count;
 197 }
 198
 199 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
 200                                               struct mlx5_cqwq *wq,
 201                                               int budget_rem)
 202 {
 203         struct mlx5e_cq_decomp *cqd = &rq->cqd;
 204         u32 cc = wq->cc;
 205
 206         mlx5e_read_title_slot(rq, wq, cc);
 207         mlx5e_read_mini_arr_slot(wq, cqd, cc + 1);
 208         mlx5e_decompress_cqe(rq, wq, cc);
 209         INDIRECT_CALL_2(rq->handle_rx_cqe, mlx5e_handle_rx_cqe_mpwrq,
 210                         mlx5e_handle_rx_cqe, rq, &cqd->title);
 211         cqd->mini_arr_idx++;
 212
 213         return mlx5e_decompress_cqes_cont(rq, wq, 1, budget_rem) - 1;
 214 }
 215
 216 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
 217                                       struct mlx5e_dma_info *dma_info)
 218 {
 219         struct mlx5e_page_cache *cache = &rq->page_cache;
 220         u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
 221         struct mlx5e_rq_stats *stats = rq->stats;
 222
 223         if (tail_next == cache->head) {
 224                 stats->cache_full++;
 225                 return false;
 226         }
 227
 228         if (!dev_page_is_reusable(dma_info->page)) {
 229                 stats->cache_waive++;
 230                 return false;
 231         }
 232
 233         cache->page_cache[cache->tail] = *dma_info;
 234         cache->tail = tail_next;
 235         return true;
 236 }
 237
 238 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
 239                                       struct mlx5e_dma_info *dma_info)
 240 {
 241         struct mlx5e_page_cache *cache = &rq->page_cache;
 242         struct mlx5e_rq_stats *stats = rq->stats;
 243
 244         if (unlikely(cache->head == cache->tail)) {
 245                 stats->cache_empty++;
 246                 return false;
 247         }
 248
 249         if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
 250                 stats->cache_busy++;
 251                 return false;
 252         }
 253
 254         *dma_info = cache->page_cache[cache->head];
 255         cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
 256         stats->cache_reuse++;
 257
 258         dma_sync_single_for_device(rq->pdev, dma_info->addr,
 259                                    PAGE_SIZE,
 260                                    DMA_FROM_DEVICE);
 261         return true;
 262 }
 263
 264 static inline int mlx5e_page_alloc_pool(struct mlx5e_rq *rq,
 265                                         struct mlx5e_dma_info *dma_info)
 266 {
 267         if (mlx5e_rx_cache_get(rq, dma_info))
 268                 return 0;
 269
 270         dma_info->page = page_pool_dev_alloc_pages(rq->page_pool);
 271         if (unlikely(!dma_info->page))
 272                 return -ENOMEM;
 273
 274         dma_info->addr = dma_map_page(rq->pdev, dma_info->page, 0,
 275                                       PAGE_SIZE, rq->buff.map_dir);
 276         if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
 277                 page_pool_recycle_direct(rq->page_pool, dma_info->page);
 278                 dma_info->page = NULL;
 279                 return -ENOMEM;
 280         }
 281
 282         return 0;
 283 }
 284
 285 static inline int mlx5e_page_alloc(struct mlx5e_rq *rq,
 286                                    struct mlx5e_dma_info *dma_info)
 287 {
 288         if (rq->xsk_pool)
 289                 return mlx5e_xsk_page_alloc_pool(rq, dma_info);
 290         else
 291                 return mlx5e_page_alloc_pool(rq, dma_info);
 292 }
 293
 294 void mlx5e_page_dma_unmap(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info)
 295 {
 296         dma_unmap_page(rq->pdev, dma_info->addr, PAGE_SIZE, rq->buff.map_dir);
 297 }
 298
 299 void mlx5e_page_release_dynamic(struct mlx5e_rq *rq,
 300                                 struct mlx5e_dma_info *dma_info,
 301                                 bool recycle)
 302 {
 303         if (likely(recycle)) {
 304                 if (mlx5e_rx_cache_put(rq, dma_info))
 305                         return;
 306
 307                 mlx5e_page_dma_unmap(rq, dma_info);
 308                 page_pool_recycle_direct(rq->page_pool, dma_info->page);
 309         } else {
 310                 mlx5e_page_dma_unmap(rq, dma_info);
 311                 page_pool_release_page(rq->page_pool, dma_info->page);
 312                 put_page(dma_info->page);
 313         }
 314 }
 315
 316 static inline void mlx5e_page_release(struct mlx5e_rq *rq,
 317                                       struct mlx5e_dma_info *dma_info,
 318                                       bool recycle)
 319 {
 320         if (rq->xsk_pool)
 321                 /* The `recycle` parameter is ignored, and the page is always
 322                  * put into the Reuse Ring, because there is no way to return
 323                  * the page to the userspace when the interface goes down.
 324                  */
 325                 xsk_buff_free(dma_info->xsk);
 326         else
 327                 mlx5e_page_release_dynamic(rq, dma_info, recycle);
 328 }
 329
 330 static inline int mlx5e_get_rx_frag(struct mlx5e_rq *rq,
 331                                     struct mlx5e_wqe_frag_info *frag)
 332 {
 333         int err = 0;
 334
 335         if (!frag->offset)
 336                 /* On first frag (offset == 0), replenish page (dma_info actually).
 337                  * Other frags that point to the same dma_info (with a different
 338                  * offset) should just use the new one without replenishing again
 339                  * by themselves.
 340                  */
 341                 err = mlx5e_page_alloc(rq, frag->di);
 342
 343         return err;
 344 }
 345
 346 static inline void mlx5e_put_rx_frag(struct mlx5e_rq *rq,
 347                                      struct mlx5e_wqe_frag_info *frag,
 348                                      bool recycle)
 349 {
 350         if (frag->last_in_page)
 351                 mlx5e_page_release(rq, frag->di, recycle);
 352 }
 353
 354 static inline struct mlx5e_wqe_frag_info *get_frag(struct mlx5e_rq *rq, u16 ix)
 355 {
 356         return &rq->wqe.frags[ix << rq->wqe.info.log_num_frags];
 357 }
 358
 359 static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe_cyc *wqe,
 360                               u16 ix)
 361 {
 362         struct mlx5e_wqe_frag_info *frag = get_frag(rq, ix);
 363         int err;
 364         int i;
 365
 366         for (i = 0; i < rq->wqe.info.num_frags; i++, frag++) {
 367                 err = mlx5e_get_rx_frag(rq, frag);
 368                 if (unlikely(err))
 369                         goto free_frags;
 370
 371                 wqe->data[i].addr = cpu_to_be64(frag->di->addr +
 372                                                 frag->offset + rq->buff.headroom);
 373         }
 374
 375         return 0;
 376
 377 free_frags:
 378         while (--i >= 0)
 379                 mlx5e_put_rx_frag(rq, --frag, true);
 380
 381         return err;
 382 }
 383
 384 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
 385                                      struct mlx5e_wqe_frag_info *wi,
 386                                      bool recycle)
 387 {
 388         int i;
 389
 390         for (i = 0; i < rq->wqe.info.num_frags; i++, wi++)
 391                 mlx5e_put_rx_frag(rq, wi, recycle);
 392 }
 393
 394 static void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
 395 {
 396         struct mlx5e_wqe_frag_info *wi = get_frag(rq, ix);
 397
 398         mlx5e_free_rx_wqe(rq, wi, false);
 399 }
 400
 401 static int mlx5e_alloc_rx_wqes(struct mlx5e_rq *rq, u16 ix, u8 wqe_bulk)
 402 {
 403         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
 404         int err;
 405         int i;
 406
 407         if (rq->xsk_pool) {
 408                 int pages_desired = wqe_bulk << rq->wqe.info.log_num_frags;
 409
 410                 /* Check in advance that we have enough frames, instead of
 411                  * allocating one-by-one, failing and moving frames to the
 412                  * Reuse Ring.
 413                  */
 414                 if (unlikely(!xsk_buff_can_alloc(rq->xsk_pool, pages_desired)))
 415                         return -ENOMEM;
 416         }
 417
 418         for (i = 0; i < wqe_bulk; i++) {
 419                 struct mlx5e_rx_wqe_cyc *wqe = mlx5_wq_cyc_get_wqe(wq, ix + i);
 420
 421                 err = mlx5e_alloc_rx_wqe(rq, wqe, ix + i);
 422                 if (unlikely(err))
 423                         goto free_wqes;
 424         }
 425
 426         return 0;
 427
 428 free_wqes:
 429         while (--i >= 0)
 430                 mlx5e_dealloc_rx_wqe(rq, ix + i);
 431
 432         return err;
 433 }
 434
 435 static inline void
 436 mlx5e_add_skb_frag(struct mlx5e_rq *rq, struct sk_buff *skb,
 437                    struct mlx5e_dma_info *di, u32 frag_offset, u32 len,
 438                    unsigned int truesize)
 439 {
 440         dma_sync_single_for_cpu(rq->pdev,
 441                                 di->addr + frag_offset,
 442                                 len, DMA_FROM_DEVICE);
 443         page_ref_inc(di->page);
 444         skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
 445                         di->page, frag_offset, len, truesize);
 446 }
 447
 448 static inline void
 449 mlx5e_copy_skb_header(struct device *pdev, struct sk_buff *skb,
 450                       struct mlx5e_dma_info *dma_info,
 451                       int offset_from, u32 headlen)
 452 {
 453         const void *from = page_address(dma_info->page) + offset_from;
 454         /* Aligning len to sizeof(long) optimizes memcpy performance */
 455         unsigned int len = ALIGN(headlen, sizeof(long));
 456
 457         dma_sync_single_for_cpu(pdev, dma_info->addr + offset_from, len,
 458                                 DMA_FROM_DEVICE);
 459         skb_copy_to_linear_data(skb, from, len);
 460 }
 461
 462 static void
 463 mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi, bool recycle)
 464 {
 465         bool no_xdp_xmit;
 466         struct mlx5e_dma_info *dma_info = wi->umr.dma_info;
 467         int i;
 468
 469         /* A common case for AF_XDP. */
 470         if (bitmap_full(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE))
 471                 return;
 472
 473         no_xdp_xmit = bitmap_empty(wi->xdp_xmit_bitmap,
 474                                    MLX5_MPWRQ_PAGES_PER_WQE);
 475
 476         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++)
 477                 if (no_xdp_xmit || !test_bit(i, wi->xdp_xmit_bitmap))
 478                         mlx5e_page_release(rq, &dma_info[i], recycle);
 479 }
 480
 481 static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq, u8 n)
 482 {
 483         struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
 484
 485         do {
 486                 u16 next_wqe_index = mlx5_wq_ll_get_wqe_next_ix(wq, wq->head);
 487
 488                 mlx5_wq_ll_push(wq, next_wqe_index);
 489         } while (--n);
 490
 491         /* ensure wqes are visible to device before updating doorbell record */
 492         dma_wmb();
 493
 494         mlx5_wq_ll_update_db_record(wq);
 495 }
 496
 497 static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 498 {
 499         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 500         struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
 501         struct mlx5e_icosq *sq = rq->icosq;
 502         struct mlx5_wq_cyc *wq = &sq->wq;
 503         struct mlx5e_umr_wqe *umr_wqe;
 504         u16 pi;
 505         int err;
 506         int i;
 507
 508         /* Check in advance that we have enough frames, instead of allocating
 509          * one-by-one, failing and moving frames to the Reuse Ring.
 510          */
 511         if (rq->xsk_pool &&
 512             unlikely(!xsk_buff_can_alloc(rq->xsk_pool, MLX5_MPWRQ_PAGES_PER_WQE))) {
 513                 err = -ENOMEM;
 514                 goto err;
 515         }
 516
 517         pi = mlx5e_icosq_get_next_pi(sq, MLX5E_UMR_WQEBBS);
 518         umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
 519         memcpy(umr_wqe, &rq->mpwqe.umr_wqe, offsetof(struct mlx5e_umr_wqe, inline_mtts));
 520
 521         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
 522                 err = mlx5e_page_alloc(rq, dma_info);
 523                 if (unlikely(err))
 524                         goto err_unmap;
 525                 umr_wqe->inline_mtts[i].ptag = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
 526         }
 527
 528         bitmap_zero(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
 529         wi->consumed_strides = 0;
 530
 531         umr_wqe->ctrl.opmod_idx_opcode =
 532                 cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
 533                             MLX5_OPCODE_UMR);
 534         umr_wqe->uctrl.xlt_offset =
 535                 cpu_to_be16(MLX5_ALIGNED_MTTS_OCTW(MLX5E_REQUIRED_MTTS(ix)));
 536
 537         sq->db.wqe_info[pi] = (struct mlx5e_icosq_wqe_info) {
 538                 .wqe_type   = MLX5E_ICOSQ_WQE_UMR_RX,
 539                 .num_wqebbs = MLX5E_UMR_WQEBBS,
 540                 .umr.rq     = rq,
 541         };
 542
 543         sq->pc += MLX5E_UMR_WQEBBS;
 544
 545         sq->doorbell_cseg = &umr_wqe->ctrl;
 546
 547         return 0;
 548
 549 err_unmap:
 550         while (--i >= 0) {
 551                 dma_info--;
 552                 mlx5e_page_release(rq, dma_info, true);
 553         }
 554
 555 err:
 556         rq->stats->buff_alloc_err++;
 557
 558         return err;
 559 }
 560
 561 static void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 562 {
 563         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 564         /* Don't recycle, this function is called on rq/netdev close */
 565         mlx5e_free_rx_mpwqe(rq, wi, false);
 566 }
 567
 568 INDIRECT_CALLABLE_SCOPE bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
 569 {
 570         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
 571         u8 wqe_bulk;
 572         int err;
 573
 574         if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
 575                 return false;
 576
 577         wqe_bulk = rq->wqe.info.wqe_bulk;
 578
 579         if (mlx5_wq_cyc_missing(wq) < wqe_bulk)
 580                 return false;
 581
 582         if (rq->page_pool)
 583                 page_pool_nid_changed(rq->page_pool, numa_mem_id());
 584
 585         do {
 586                 u16 head = mlx5_wq_cyc_get_head(wq);
 587
 588                 err = mlx5e_alloc_rx_wqes(rq, head, wqe_bulk);
 589                 if (unlikely(err)) {
 590                         rq->stats->buff_alloc_err++;
 591                         break;
 592                 }
 593
 594                 mlx5_wq_cyc_push_n(wq, wqe_bulk);
 595         } while (mlx5_wq_cyc_missing(wq) >= wqe_bulk);
 596
 597         /* ensure wqes are visible to device before updating doorbell record */
 598         dma_wmb();
 599
 600         mlx5_wq_cyc_update_db_record(wq);
 601
 602         return !!err;
 603 }
 604
 605 void mlx5e_free_icosq_descs(struct mlx5e_icosq *sq)
 606 {
 607         u16 sqcc;
 608
 609         sqcc = sq->cc;
 610
 611         while (sqcc != sq->pc) {
 612                 struct mlx5e_icosq_wqe_info *wi;
 613                 u16 ci;
 614
 615                 ci = mlx5_wq_cyc_ctr2ix(&sq->wq, sqcc);
 616                 wi = &sq->db.wqe_info[ci];
 617                 sqcc += wi->num_wqebbs;
 618 #ifdef CONFIG_MLX5_EN_TLS
 619                 switch (wi->wqe_type) {
 620                 case MLX5E_ICOSQ_WQE_SET_PSV_TLS:
 621                         mlx5e_ktls_handle_ctx_completion(wi);
 622                         break;
 623                 case MLX5E_ICOSQ_WQE_GET_PSV_TLS:
 624                         mlx5e_ktls_handle_get_psv_completion(wi, sq);
 625                         break;
 626                 }
 627 #endif
 628         }
 629         sq->cc = sqcc;
 630 }
 631
 632 int mlx5e_poll_ico_cq(struct mlx5e_cq *cq)
 633 {
 634         struct mlx5e_icosq *sq = container_of(cq, struct mlx5e_icosq, cq);
 635         struct mlx5_cqe64 *cqe;
 636         u16 sqcc;
 637         int i;
 638
 639         if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &sq->state)))
 640                 return 0;
 641
 642         cqe = mlx5_cqwq_get_cqe(&cq->wq);
 643         if (likely(!cqe))
 644                 return 0;
 645
 646         /* sq->cc must be updated only after mlx5_cqwq_update_db_record(),
 647          * otherwise a cq overrun may occur
 648          */
 649         sqcc = sq->cc;
 650
 651         i = 0;
 652         do {
 653                 u16 wqe_counter;
 654                 bool last_wqe;
 655
 656                 mlx5_cqwq_pop(&cq->wq);
 657
 658                 wqe_counter = be16_to_cpu(cqe->wqe_counter);
 659
 660                 do {
 661                         struct mlx5e_icosq_wqe_info *wi;
 662                         u16 ci;
 663
 664                         last_wqe = (sqcc == wqe_counter);
 665
 666                         ci = mlx5_wq_cyc_ctr2ix(&sq->wq, sqcc);
 667                         wi = &sq->db.wqe_info[ci];
 668                         sqcc += wi->num_wqebbs;
 669
 670                         if (last_wqe && unlikely(get_cqe_opcode(cqe) != MLX5_CQE_REQ)) {
 671                                 netdev_WARN_ONCE(cq->netdev,
 672                                                  "Bad OP in ICOSQ CQE: 0x%x\n",
 673                                                  get_cqe_opcode(cqe));
 674                                 mlx5e_dump_error_cqe(&sq->cq, sq->sqn,
 675                                                      (struct mlx5_err_cqe *)cqe);
 676                                 mlx5_wq_cyc_wqe_dump(&sq->wq, ci, wi->num_wqebbs);
 677                                 if (!test_and_set_bit(MLX5E_SQ_STATE_RECOVERING, &sq->state))
 678                                         queue_work(cq->priv->wq, &sq->recover_work);
 679                                 break;
 680                         }
 681
 682                         switch (wi->wqe_type) {
 683                         case MLX5E_ICOSQ_WQE_UMR_RX:
 684                                 wi->umr.rq->mpwqe.umr_completed++;
 685                                 break;
 686                         case MLX5E_ICOSQ_WQE_NOP:
 687                                 break;
 688 #ifdef CONFIG_MLX5_EN_TLS
 689                         case MLX5E_ICOSQ_WQE_UMR_TLS:
 690                                 break;
 691                         case MLX5E_ICOSQ_WQE_SET_PSV_TLS:
 692                                 mlx5e_ktls_handle_ctx_completion(wi);
 693                                 break;
 694                         case MLX5E_ICOSQ_WQE_GET_PSV_TLS:
 695                                 mlx5e_ktls_handle_get_psv_completion(wi, sq);
 696                                 break;
 697 #endif
 698                         default:
 699                                 netdev_WARN_ONCE(cq->netdev,
 700                                                  "Bad WQE type in ICOSQ WQE info: 0x%x\n",
 701                                                  wi->wqe_type);
 702                         }
 703                 } while (!last_wqe);
 704         } while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
 705
 706         sq->cc = sqcc;
 707
 708         mlx5_cqwq_update_db_record(&cq->wq);
 709
 710         return i;
 711 }
 712
 713 INDIRECT_CALLABLE_SCOPE bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
 714 {
 715         struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
 716         u8  umr_completed = rq->mpwqe.umr_completed;
 717         struct mlx5e_icosq *sq = rq->icosq;
 718         int alloc_err = 0;
 719         u8  missing, i;
 720         u16 head;
 721
 722         if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
 723                 return false;
 724
 725         if (umr_completed) {
 726                 mlx5e_post_rx_mpwqe(rq, umr_completed);
 727                 rq->mpwqe.umr_in_progress -= umr_completed;
 728                 rq->mpwqe.umr_completed = 0;
 729         }
 730
 731         missing = mlx5_wq_ll_missing(wq) - rq->mpwqe.umr_in_progress;
 732
 733         if (unlikely(rq->mpwqe.umr_in_progress > rq->mpwqe.umr_last_bulk))
 734                 rq->stats->congst_umr++;
 735
 736 #define UMR_WQE_BULK (2)
 737         if (likely(missing < UMR_WQE_BULK))
 738                 return false;
 739
 740         if (rq->page_pool)
 741                 page_pool_nid_changed(rq->page_pool, numa_mem_id());
 742
 743         head = rq->mpwqe.actual_wq_head;
 744         i = missing;
 745         do {
 746                 alloc_err = mlx5e_alloc_rx_mpwqe(rq, head);
 747
 748                 if (unlikely(alloc_err))
 749                         break;
 750                 head = mlx5_wq_ll_get_wqe_next_ix(wq, head);
 751         } while (--i);
 752
 753         rq->mpwqe.umr_last_bulk    = missing - i;
 754         if (sq->doorbell_cseg) {
 755                 mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, sq->doorbell_cseg);
 756                 sq->doorbell_cseg = NULL;
 757         }
 758
 759         rq->mpwqe.umr_in_progress += rq->mpwqe.umr_last_bulk;
 760         rq->mpwqe.actual_wq_head   = head;
 761
 762         /* If XSK Fill Ring doesn't have enough frames, report the error, so
 763          * that one of the actions can be performed:
 764          * 1. If need_wakeup is used, signal that the application has to kick
 765          * the driver when it refills the Fill Ring.
 766          * 2. Otherwise, busy poll by rescheduling the NAPI poll.
 767          */
 768         if (unlikely(alloc_err == -ENOMEM && rq->xsk_pool))
 769                 return true;
 770
 771         return false;
 772 }
 773
 774 static void mlx5e_lro_update_tcp_hdr(struct mlx5_cqe64 *cqe, struct tcphdr *tcp)
 775 {
 776         u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
 777         u8 tcp_ack     = (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA) ||
 778                          (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA);
 779
 780         tcp->check                      = 0;
 781         tcp->psh                        = get_cqe_lro_tcppsh(cqe);
 782
 783         if (tcp_ack) {
 784                 tcp->ack                = 1;
 785                 tcp->ack_seq            = cqe->lro_ack_seq_num;
 786                 tcp->window             = cqe->lro_tcp_win;
 787         }
 788 }
 789
 790 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
 791                                  u32 cqe_bcnt)
 792 {
 793         struct ethhdr   *eth = (struct ethhdr *)(skb->data);
 794         struct tcphdr   *tcp;
 795         int network_depth = 0;
 796         __wsum check;
 797         __be16 proto;
 798         u16 tot_len;
 799         void *ip_p;
 800
 801         proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
 802
 803         tot_len = cqe_bcnt - network_depth;
 804         ip_p = skb->data + network_depth;
 805
 806         if (proto == htons(ETH_P_IP)) {
 807                 struct iphdr *ipv4 = ip_p;
 808
 809                 tcp = ip_p + sizeof(struct iphdr);
 810                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 811
 812                 ipv4->ttl               = cqe->lro_min_ttl;
 813                 ipv4->tot_len           = cpu_to_be16(tot_len);
 814                 ipv4->check             = 0;
 815                 ipv4->check             = ip_fast_csum((unsigned char *)ipv4,
 816                                                        ipv4->ihl);
 817
 818                 mlx5e_lro_update_tcp_hdr(cqe, tcp);
 819                 check = csum_partial(tcp, tcp->doff * 4,
 820                                      csum_unfold((__force __sum16)cqe->check_sum));
 821                 /* Almost done, don't forget the pseudo header */
 822                 tcp->check = csum_tcpudp_magic(ipv4->saddr, ipv4->daddr,
 823                                                tot_len - sizeof(struct iphdr),
 824                                                IPPROTO_TCP, check);
 825         } else {
 826                 u16 payload_len = tot_len - sizeof(struct ipv6hdr);
 827                 struct ipv6hdr *ipv6 = ip_p;
 828
 829                 tcp = ip_p + sizeof(struct ipv6hdr);
 830                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 831
 832                 ipv6->hop_limit         = cqe->lro_min_ttl;
 833                 ipv6->payload_len       = cpu_to_be16(payload_len);
 834
 835                 mlx5e_lro_update_tcp_hdr(cqe, tcp);
 836                 check = csum_partial(tcp, tcp->doff * 4,
 837                                      csum_unfold((__force __sum16)cqe->check_sum));
 838                 /* Almost done, don't forget the pseudo header */
 839                 tcp->check = csum_ipv6_magic(&ipv6->saddr, &ipv6->daddr, payload_len,
 840                                              IPPROTO_TCP, check);
 841         }
 842 }
 843
 844 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
 845                                       struct sk_buff *skb)
 846 {
 847         u8 cht = cqe->rss_hash_type;
 848         int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
 849                  (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
 850                                             PKT_HASH_TYPE_NONE;
 851         skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
 852 }
 853
 854 static inline bool is_last_ethertype_ip(struct sk_buff *skb, int *network_depth,
 855                                         __be16 *proto)
 856 {
 857         *proto = ((struct ethhdr *)skb->data)->h_proto;
 858         *proto = __vlan_get_protocol(skb, *proto, network_depth);
 859
 860         if (*proto == htons(ETH_P_IP))
 861                 return pskb_may_pull(skb, *network_depth + sizeof(struct iphdr));
 862
 863         if (*proto == htons(ETH_P_IPV6))
 864                 return pskb_may_pull(skb, *network_depth + sizeof(struct ipv6hdr));
 865
 866         return false;
 867 }
 868
 869 static inline void mlx5e_enable_ecn(struct mlx5e_rq *rq, struct sk_buff *skb)
 870 {
 871         int network_depth = 0;
 872         __be16 proto;
 873         void *ip;
 874         int rc;
 875
 876         if (unlikely(!is_last_ethertype_ip(skb, &network_depth, &proto)))
 877                 return;
 878
 879         ip = skb->data + network_depth;
 880         rc = ((proto == htons(ETH_P_IP)) ? IP_ECN_set_ce((struct iphdr *)ip) :
 881                                          IP6_ECN_set_ce(skb, (struct ipv6hdr *)ip));
 882
 883         rq->stats->ecn_mark += !!rc;
 884 }
 885
 886 static u8 get_ip_proto(struct sk_buff *skb, int network_depth, __be16 proto)
 887 {
 888         void *ip_p = skb->data + network_depth;
 889
 890         return (proto == htons(ETH_P_IP)) ? ((struct iphdr *)ip_p)->protocol :
 891                                             ((struct ipv6hdr *)ip_p)->nexthdr;
 892 }
 893
 894 #define short_frame(size) ((size) <= ETH_ZLEN + ETH_FCS_LEN)
 895
 896 #define MAX_PADDING 8
 897
 898 static void
 899 tail_padding_csum_slow(struct sk_buff *skb, int offset, int len,
 900                        struct mlx5e_rq_stats *stats)
 901 {
 902         stats->csum_complete_tail_slow++;
 903         skb->csum = csum_block_add(skb->csum,
 904                                    skb_checksum(skb, offset, len, 0),
 905                                    offset);
 906 }
 907
 908 static void
 909 tail_padding_csum(struct sk_buff *skb, int offset,
 910                   struct mlx5e_rq_stats *stats)
 911 {
 912         u8 tail_padding[MAX_PADDING];
 913         int len = skb->len - offset;
 914         void *tail;
 915
 916         if (unlikely(len > MAX_PADDING)) {
 917                 tail_padding_csum_slow(skb, offset, len, stats);
 918                 return;
 919         }
 920
 921         tail = skb_header_pointer(skb, offset, len, tail_padding);
 922         if (unlikely(!tail)) {
 923                 tail_padding_csum_slow(skb, offset, len, stats);
 924                 return;
 925         }
 926
 927         stats->csum_complete_tail++;
 928         skb->csum = csum_block_add(skb->csum, csum_partial(tail, len, 0), offset);
 929 }
 930
 931 static void
 932 mlx5e_skb_csum_fixup(struct sk_buff *skb, int network_depth, __be16 proto,
 933                      struct mlx5e_rq_stats *stats)
 934 {
 935         struct ipv6hdr *ip6;
 936         struct iphdr   *ip4;
 937         int pkt_len;
 938
 939         /* Fixup vlan headers, if any */
 940         if (network_depth > ETH_HLEN)
 941                 /* CQE csum is calculated from the IP header and does
 942                  * not cover VLAN headers (if present). This will add
 943                  * the checksum manually.
 944                  */
 945                 skb->csum = csum_partial(skb->data + ETH_HLEN,
 946                                          network_depth - ETH_HLEN,
 947                                          skb->csum);
 948
 949         /* Fixup tail padding, if any */
 950         switch (proto) {
 951         case htons(ETH_P_IP):
 952                 ip4 = (struct iphdr *)(skb->data + network_depth);
 953                 pkt_len = network_depth + ntohs(ip4->tot_len);
 954                 break;
 955         case htons(ETH_P_IPV6):
 956                 ip6 = (struct ipv6hdr *)(skb->data + network_depth);
 957                 pkt_len = network_depth + sizeof(*ip6) + ntohs(ip6->payload_len);
 958                 break;
 959         default:
 960                 return;
 961         }
 962
 963         if (likely(pkt_len >= skb->len))
 964                 return;
 965
 966         tail_padding_csum(skb, pkt_len, stats);
 967 }
 968
 969 static inline void mlx5e_handle_csum(struct net_device *netdev,
 970                                      struct mlx5_cqe64 *cqe,
 971                                      struct mlx5e_rq *rq,
 972                                      struct sk_buff *skb,
 973                                      bool   lro)
 974 {
 975         struct mlx5e_rq_stats *stats = rq->stats;
 976         int network_depth = 0;
 977         __be16 proto;
 978
 979         if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
 980                 goto csum_none;
 981
 982         if (lro) {
 983                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 984                 stats->csum_unnecessary++;
 985                 return;
 986         }
 987
 988         /* True when explicitly set via priv flag, or XDP prog is loaded */
 989         if (test_bit(MLX5E_RQ_STATE_NO_CSUM_COMPLETE, &rq->state))
 990                 goto csum_unnecessary;
 991
 992         /* CQE csum doesn't cover padding octets in short ethernet
 993          * frames. And the pad field is appended prior to calculating
 994          * and appending the FCS field.
 995          *
 996          * Detecting these padded frames requires to verify and parse
 997          * IP headers, so we simply force all those small frames to be
 998          * CHECKSUM_UNNECESSARY even if they are not padded.
 999          */
1000         if (short_frame(skb->len))
1001                 goto csum_unnecessary;
1002
1003         if (likely(is_last_ethertype_ip(skb, &network_depth, &proto))) {
1004                 if (unlikely(get_ip_proto(skb, network_depth, proto) == IPPROTO_SCTP))
1005                         goto csum_unnecessary;
1006
1007                 stats->csum_complete++;
1008                 skb->ip_summed = CHECKSUM_COMPLETE;
1009                 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1010
1011                 if (test_bit(MLX5E_RQ_STATE_CSUM_FULL, &rq->state))
1012                         return; /* CQE csum covers all received bytes */
1013
1014                 /* csum might need some fixups ...*/
1015                 mlx5e_skb_csum_fixup(skb, network_depth, proto, stats);
1016                 return;
1017         }
1018
1019 csum_unnecessary:
1020         if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
1021                    (cqe->hds_ip_ext & CQE_L4_OK))) {
1022                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1023                 if (cqe_is_tunneled(cqe)) {
1024                         skb->csum_level = 1;
1025                         skb->encapsulation = 1;
1026                         stats->csum_unnecessary_inner++;
1027                         return;
1028                 }
1029                 stats->csum_unnecessary++;
1030                 return;
1031         }
1032 csum_none:
1033         skb->ip_summed = CHECKSUM_NONE;
1034         stats->csum_none++;
1035 }
1036
1037 #define MLX5E_CE_BIT_MASK 0x80
1038
1039 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
1040                                       u32 cqe_bcnt,
1041                                       struct mlx5e_rq *rq,
1042                                       struct sk_buff *skb)
1043 {
1044         u8 lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
1045         struct mlx5e_rq_stats *stats = rq->stats;
1046         struct net_device *netdev = rq->netdev;
1047
1048         skb->mac_len = ETH_HLEN;
1049
1050         mlx5e_tls_handle_rx_skb(rq, skb, cqe, &cqe_bcnt);
1051
1052         if (unlikely(mlx5_ipsec_is_rx_flow(cqe)))
1053                 mlx5e_ipsec_offload_handle_rx_skb(netdev, skb, cqe);
1054
1055         if (lro_num_seg > 1) {
1056                 mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
1057                 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
1058                 /* Subtract one since we already counted this as one
1059                  * "regular" packet in mlx5e_complete_rx_cqe()
1060                  */
1061                 stats->packets += lro_num_seg - 1;
1062                 stats->lro_packets++;
1063                 stats->lro_bytes += cqe_bcnt;
1064         }
1065
1066         if (unlikely(mlx5e_rx_hw_stamp(rq->tstamp)))
1067                 skb_hwtstamps(skb)->hwtstamp = mlx5e_cqe_ts_to_ns(rq->ptp_cyc2time,
1068                                                                   rq->clock, get_cqe_ts(cqe));
1069         skb_record_rx_queue(skb, rq->ix);
1070
1071         if (likely(netdev->features & NETIF_F_RXHASH))
1072                 mlx5e_skb_set_hash(cqe, skb);
1073
1074         if (cqe_has_vlan(cqe)) {
1075                 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
1076                                        be16_to_cpu(cqe->vlan_info));
1077                 stats->removed_vlan_packets++;
1078         }
1079
1080         skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
1081
1082         mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
1083         /* checking CE bit in cqe - MSB in ml_path field */
1084         if (unlikely(cqe->ml_path & MLX5E_CE_BIT_MASK))
1085                 mlx5e_enable_ecn(rq, skb);
1086
1087         skb->protocol = eth_type_trans(skb, netdev);
1088
1089         if (unlikely(mlx5e_skb_is_multicast(skb)))
1090                 stats->mcast_packets++;
1091 }
1092
1093 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
1094                                          struct mlx5_cqe64 *cqe,
1095                                          u32 cqe_bcnt,
1096                                          struct sk_buff *skb)
1097 {
1098         struct mlx5e_rq_stats *stats = rq->stats;
1099
1100         stats->packets++;
1101         stats->bytes += cqe_bcnt;
1102         mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
1103 }
1104
1105 static inline
1106 struct sk_buff *mlx5e_build_linear_skb(struct mlx5e_rq *rq, void *va,
1107                                        u32 frag_size, u16 headroom,
1108                                        u32 cqe_bcnt)
1109 {
1110         struct sk_buff *skb = build_skb(va, frag_size);
1111
1112         if (unlikely(!skb)) {
1113                 rq->stats->buff_alloc_err++;
1114                 return NULL;
1115         }
1116
1117         skb_reserve(skb, headroom);
1118         skb_put(skb, cqe_bcnt);
1119
1120         return skb;
1121 }
1122
1123 static void mlx5e_fill_xdp_buff(struct mlx5e_rq *rq, void *va, u16 headroom,
1124                                 u32 len, struct xdp_buff *xdp)
1125 {
1126         xdp_init_buff(xdp, rq->buff.frame0_sz, &rq->xdp_rxq);
1127         xdp_prepare_buff(xdp, va, headroom, len, false);
1128 }
1129
1130 static struct sk_buff *
1131 mlx5e_skb_from_cqe_linear(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
1132                           struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
1133 {
1134         struct mlx5e_dma_info *di = wi->di;
1135         u16 rx_headroom = rq->buff.headroom;
1136         struct xdp_buff xdp;
1137         struct sk_buff *skb;
1138         void *va, *data;
1139         u32 frag_size;
1140
1141         va             = page_address(di->page) + wi->offset;
1142         data           = va + rx_headroom;
1143         frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
1144
1145         dma_sync_single_range_for_cpu(rq->pdev, di->addr, wi->offset,
1146                                       frag_size, DMA_FROM_DEVICE);
1147         net_prefetchw(va); /* xdp_frame data area */
1148         net_prefetch(data);
1149
1150         mlx5e_fill_xdp_buff(rq, va, rx_headroom, cqe_bcnt, &xdp);
1151         if (mlx5e_xdp_handle(rq, di, &cqe_bcnt, &xdp))
1152                 return NULL; /* page/packet was consumed by XDP */
1153
1154         rx_headroom = xdp.data - xdp.data_hard_start;
1155         frag_size = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
1156         skb = mlx5e_build_linear_skb(rq, va, frag_size, rx_headroom, cqe_bcnt);
1157         if (unlikely(!skb))
1158                 return NULL;
1159
1160         /* queue up for recycling/reuse */
1161         page_ref_inc(di->page);
1162
1163         return skb;
1164 }
1165
1166 static struct sk_buff *
1167 mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
1168                              struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
1169 {
1170         struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
1171         struct mlx5e_wqe_frag_info *head_wi = wi;
1172         u16 headlen      = min_t(u32, MLX5E_RX_MAX_HEAD, cqe_bcnt);
1173         u16 frag_headlen = headlen;
1174         u16 byte_cnt     = cqe_bcnt - headlen;
1175         struct sk_buff *skb;
1176
1177         /* XDP is not supported in this configuration, as incoming packets
1178          * might spread among multiple pages.
1179          */
1180         skb = napi_alloc_skb(rq->cq.napi,
1181                              ALIGN(MLX5E_RX_MAX_HEAD, sizeof(long)));
1182         if (unlikely(!skb)) {
1183                 rq->stats->buff_alloc_err++;
1184                 return NULL;
1185         }
1186
1187         net_prefetchw(skb->data);
1188
1189         while (byte_cnt) {
1190                 u16 frag_consumed_bytes =
1191                         min_t(u16, frag_info->frag_size - frag_headlen, byte_cnt);
1192
1193                 mlx5e_add_skb_frag(rq, skb, wi->di, wi->offset + frag_headlen,
1194                                    frag_consumed_bytes, frag_info->frag_stride);
1195                 byte_cnt -= frag_consumed_bytes;
1196                 frag_headlen = 0;
1197                 frag_info++;
1198                 wi++;
1199         }
1200
1201         /* copy header */
1202         mlx5e_copy_skb_header(rq->pdev, skb, head_wi->di, head_wi->offset, headlen);
1203         /* skb linear part was allocated with headlen and aligned to long */
1204         skb->tail += headlen;
1205         skb->len  += headlen;
1206
1207         return skb;
1208 }
1209
1210 static void trigger_report(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1211 {
1212         struct mlx5_err_cqe *err_cqe = (struct mlx5_err_cqe *)cqe;
1213         struct mlx5e_priv *priv = rq->priv;
1214
1215         if (cqe_syndrome_needs_recover(err_cqe->syndrome) &&
1216             !test_and_set_bit(MLX5E_RQ_STATE_RECOVERING, &rq->state)) {
1217                 mlx5e_dump_error_cqe(&rq->cq, rq->rqn, err_cqe);
1218                 queue_work(priv->wq, &rq->recover_work);
1219         }
1220 }
1221
1222 static void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1223 {
1224         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1225         struct mlx5e_wqe_frag_info *wi;
1226         struct sk_buff *skb;
1227         u32 cqe_bcnt;
1228         u16 ci;
1229
1230         ci       = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1231         wi       = get_frag(rq, ci);
1232         cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1233
1234         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1235                 trigger_report(rq, cqe);
1236                 rq->stats->wqe_err++;
1237                 goto free_wqe;
1238         }
1239
1240         skb = INDIRECT_CALL_2(rq->wqe.skb_from_cqe,
1241                               mlx5e_skb_from_cqe_linear,
1242                               mlx5e_skb_from_cqe_nonlinear,
1243                               rq, cqe, wi, cqe_bcnt);
1244         if (!skb) {
1245                 /* probably for XDP */
1246                 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)) {
1247                         /* do not return page to cache,
1248                          * it will be returned on XDP_TX completion.
1249                          */
1250                         goto wq_cyc_pop;
1251                 }
1252                 goto free_wqe;
1253         }
1254
1255         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1256
1257         if (mlx5e_cqe_regb_chain(cqe))
1258                 if (!mlx5e_tc_update_skb(cqe, skb)) {
1259                         dev_kfree_skb_any(skb);
1260                         goto free_wqe;
1261                 }
1262
1263         napi_gro_receive(rq->cq.napi, skb);
1264
1265 free_wqe:
1266         mlx5e_free_rx_wqe(rq, wi, true);
1267 wq_cyc_pop:
1268         mlx5_wq_cyc_pop(wq);
1269 }
1270
1271 #ifdef CONFIG_MLX5_ESWITCH
1272 static void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1273 {
1274         struct net_device *netdev = rq->netdev;
1275         struct mlx5e_priv *priv = netdev_priv(netdev);
1276         struct mlx5e_rep_priv *rpriv  = priv->ppriv;
1277         struct mlx5_eswitch_rep *rep = rpriv->rep;
1278         struct mlx5e_tc_update_priv tc_priv = {};
1279         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1280         struct mlx5e_wqe_frag_info *wi;
1281         struct sk_buff *skb;
1282         u32 cqe_bcnt;
1283         u16 ci;
1284
1285         ci       = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1286         wi       = get_frag(rq, ci);
1287         cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1288
1289         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1290                 rq->stats->wqe_err++;
1291                 goto free_wqe;
1292         }
1293
1294         skb = INDIRECT_CALL_2(rq->wqe.skb_from_cqe,
1295                               mlx5e_skb_from_cqe_linear,
1296                               mlx5e_skb_from_cqe_nonlinear,
1297                               rq, cqe, wi, cqe_bcnt);
1298         if (!skb) {
1299                 /* probably for XDP */
1300                 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)) {
1301                         /* do not return page to cache,
1302                          * it will be returned on XDP_TX completion.
1303                          */
1304                         goto wq_cyc_pop;
1305                 }
1306                 goto free_wqe;
1307         }
1308
1309         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1310
1311         if (rep->vlan && skb_vlan_tag_present(skb))
1312                 skb_vlan_pop(skb);
1313
1314         if (unlikely(!mlx5_ipsec_is_rx_flow(cqe) &&
1315                      !mlx5e_rep_tc_update_skb(cqe, skb, &tc_priv))) {
1316                 dev_kfree_skb_any(skb);
1317                 goto free_wqe;
1318         }
1319
1320         napi_gro_receive(rq->cq.napi, skb);
1321
1322         mlx5_rep_tc_post_napi_receive(&tc_priv);
1323
1324 free_wqe:
1325         mlx5e_free_rx_wqe(rq, wi, true);
1326 wq_cyc_pop:
1327         mlx5_wq_cyc_pop(wq);
1328 }
1329
1330 static void mlx5e_handle_rx_cqe_mpwrq_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1331 {
1332         u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
1333         u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
1334         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
1335         u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
1336         u32 wqe_offset     = stride_ix << rq->mpwqe.log_stride_sz;
1337         u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
1338         u32 page_idx       = wqe_offset >> PAGE_SHIFT;
1339         struct mlx5e_tc_update_priv tc_priv = {};
1340         struct mlx5e_rx_wqe_ll *wqe;
1341         struct mlx5_wq_ll *wq;
1342         struct sk_buff *skb;
1343         u16 cqe_bcnt;
1344
1345         wi->consumed_strides += cstrides;
1346
1347         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1348                 trigger_report(rq, cqe);
1349                 rq->stats->wqe_err++;
1350                 goto mpwrq_cqe_out;
1351         }
1352
1353         if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1354                 struct mlx5e_rq_stats *stats = rq->stats;
1355
1356                 stats->mpwqe_filler_cqes++;
1357                 stats->mpwqe_filler_strides += cstrides;
1358                 goto mpwrq_cqe_out;
1359         }
1360
1361         cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1362
1363         skb = INDIRECT_CALL_2(rq->mpwqe.skb_from_cqe_mpwrq,
1364                               mlx5e_skb_from_cqe_mpwrq_linear,
1365                               mlx5e_skb_from_cqe_mpwrq_nonlinear,
1366                               rq, wi, cqe_bcnt, head_offset, page_idx);
1367         if (!skb)
1368                 goto mpwrq_cqe_out;
1369
1370         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1371
1372         if (unlikely(!mlx5_ipsec_is_rx_flow(cqe) &&
1373                      !mlx5e_rep_tc_update_skb(cqe, skb, &tc_priv))) {
1374                 dev_kfree_skb_any(skb);
1375                 goto mpwrq_cqe_out;
1376         }
1377
1378         napi_gro_receive(rq->cq.napi, skb);
1379
1380         mlx5_rep_tc_post_napi_receive(&tc_priv);
1381
1382 mpwrq_cqe_out:
1383         if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1384                 return;
1385
1386         wq  = &rq->mpwqe.wq;
1387         wqe = mlx5_wq_ll_get_wqe(wq, wqe_id);
1388         mlx5e_free_rx_mpwqe(rq, wi, true);
1389         mlx5_wq_ll_pop(wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1390 }
1391
1392 const struct mlx5e_rx_handlers mlx5e_rx_handlers_rep = {
1393         .handle_rx_cqe       = mlx5e_handle_rx_cqe_rep,
1394         .handle_rx_cqe_mpwqe = mlx5e_handle_rx_cqe_mpwrq_rep,
1395 };
1396 #endif
1397
1398 static struct sk_buff *
1399 mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
1400                                    u16 cqe_bcnt, u32 head_offset, u32 page_idx)
1401 {
1402         u16 headlen = min_t(u16, MLX5E_RX_MAX_HEAD, cqe_bcnt);
1403         struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
1404         u32 frag_offset    = head_offset + headlen;
1405         u32 byte_cnt       = cqe_bcnt - headlen;
1406         struct mlx5e_dma_info *head_di = di;
1407         struct sk_buff *skb;
1408
1409         skb = napi_alloc_skb(rq->cq.napi,
1410                              ALIGN(MLX5E_RX_MAX_HEAD, sizeof(long)));
1411         if (unlikely(!skb)) {
1412                 rq->stats->buff_alloc_err++;
1413                 return NULL;
1414         }
1415
1416         net_prefetchw(skb->data);
1417
1418         if (unlikely(frag_offset >= PAGE_SIZE)) {
1419                 di++;
1420                 frag_offset -= PAGE_SIZE;
1421         }
1422
1423         while (byte_cnt) {
1424                 u32 pg_consumed_bytes =
1425                         min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
1426                 unsigned int truesize =
1427                         ALIGN(pg_consumed_bytes, BIT(rq->mpwqe.log_stride_sz));
1428
1429                 mlx5e_add_skb_frag(rq, skb, di, frag_offset,
1430                                    pg_consumed_bytes, truesize);
1431                 byte_cnt -= pg_consumed_bytes;
1432                 frag_offset = 0;
1433                 di++;
1434         }
1435         /* copy header */
1436         mlx5e_copy_skb_header(rq->pdev, skb, head_di, head_offset, headlen);
1437         /* skb linear part was allocated with headlen and aligned to long */
1438         skb->tail += headlen;
1439         skb->len  += headlen;
1440
1441         return skb;
1442 }
1443
1444 static struct sk_buff *
1445 mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
1446                                 u16 cqe_bcnt, u32 head_offset, u32 page_idx)
1447 {
1448         struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
1449         u16 rx_headroom = rq->buff.headroom;
1450         u32 cqe_bcnt32 = cqe_bcnt;
1451         struct xdp_buff xdp;
1452         struct sk_buff *skb;
1453         void *va, *data;
1454         u32 frag_size;
1455
1456         /* Check packet size. Note LRO doesn't use linear SKB */
1457         if (unlikely(cqe_bcnt > rq->hw_mtu)) {
1458                 rq->stats->oversize_pkts_sw_drop++;
1459                 return NULL;
1460         }
1461
1462         va             = page_address(di->page) + head_offset;
1463         data           = va + rx_headroom;
1464         frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt32);
1465
1466         dma_sync_single_range_for_cpu(rq->pdev, di->addr, head_offset,
1467                                       frag_size, DMA_FROM_DEVICE);
1468         net_prefetchw(va); /* xdp_frame data area */
1469         net_prefetch(data);
1470
1471         mlx5e_fill_xdp_buff(rq, va, rx_headroom, cqe_bcnt32, &xdp);
1472         if (mlx5e_xdp_handle(rq, di, &cqe_bcnt32, &xdp)) {
1473                 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags))
1474                         __set_bit(page_idx, wi->xdp_xmit_bitmap); /* non-atomic */
1475                 return NULL; /* page/packet was consumed by XDP */
1476         }
1477
1478         rx_headroom = xdp.data - xdp.data_hard_start;
1479         frag_size = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt32);
1480         skb = mlx5e_build_linear_skb(rq, va, frag_size, rx_headroom, cqe_bcnt32);
1481         if (unlikely(!skb))
1482                 return NULL;
1483
1484         /* queue up for recycling/reuse */
1485         page_ref_inc(di->page);
1486
1487         return skb;
1488 }
1489
1490 static void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1491 {
1492         u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
1493         u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
1494         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
1495         u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
1496         u32 wqe_offset     = stride_ix << rq->mpwqe.log_stride_sz;
1497         u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
1498         u32 page_idx       = wqe_offset >> PAGE_SHIFT;
1499         struct mlx5e_rx_wqe_ll *wqe;
1500         struct mlx5_wq_ll *wq;
1501         struct sk_buff *skb;
1502         u16 cqe_bcnt;
1503
1504         wi->consumed_strides += cstrides;
1505
1506         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1507                 trigger_report(rq, cqe);
1508                 rq->stats->wqe_err++;
1509                 goto mpwrq_cqe_out;
1510         }
1511
1512         if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1513                 struct mlx5e_rq_stats *stats = rq->stats;
1514
1515                 stats->mpwqe_filler_cqes++;
1516                 stats->mpwqe_filler_strides += cstrides;
1517                 goto mpwrq_cqe_out;
1518         }
1519
1520         cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1521
1522         skb = INDIRECT_CALL_2(rq->mpwqe.skb_from_cqe_mpwrq,
1523                               mlx5e_skb_from_cqe_mpwrq_linear,
1524                               mlx5e_skb_from_cqe_mpwrq_nonlinear,
1525                               rq, wi, cqe_bcnt, head_offset, page_idx);
1526         if (!skb)
1527                 goto mpwrq_cqe_out;
1528
1529         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1530
1531         if (mlx5e_cqe_regb_chain(cqe))
1532                 if (!mlx5e_tc_update_skb(cqe, skb)) {
1533                         dev_kfree_skb_any(skb);
1534                         goto mpwrq_cqe_out;
1535                 }
1536
1537         napi_gro_receive(rq->cq.napi, skb);
1538
1539 mpwrq_cqe_out:
1540         if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1541                 return;
1542
1543         wq  = &rq->mpwqe.wq;
1544         wqe = mlx5_wq_ll_get_wqe(wq, wqe_id);
1545         mlx5e_free_rx_mpwqe(rq, wi, true);
1546         mlx5_wq_ll_pop(wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1547 }
1548
1549 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
1550 {
1551         struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
1552         struct mlx5_cqwq *cqwq = &cq->wq;
1553         struct mlx5_cqe64 *cqe;
1554         int work_done = 0;
1555
1556         if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
1557                 return 0;
1558
1559         if (rq->cqd.left) {
1560                 work_done += mlx5e_decompress_cqes_cont(rq, cqwq, 0, budget);
1561                 if (work_done >= budget)
1562                         goto out;
1563         }
1564
1565         cqe = mlx5_cqwq_get_cqe(cqwq);
1566         if (!cqe) {
1567                 if (unlikely(work_done))
1568                         goto out;
1569                 return 0;
1570         }
1571
1572         do {
1573                 if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1574                         work_done +=
1575                                 mlx5e_decompress_cqes_start(rq, cqwq,
1576                                                             budget - work_done);
1577                         continue;
1578                 }
1579
1580                 mlx5_cqwq_pop(cqwq);
1581
1582                 INDIRECT_CALL_2(rq->handle_rx_cqe, mlx5e_handle_rx_cqe_mpwrq,
1583                                 mlx5e_handle_rx_cqe, rq, cqe);
1584         } while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(cqwq)));
1585
1586 out:
1587         if (rcu_access_pointer(rq->xdp_prog))
1588                 mlx5e_xdp_rx_poll_complete(rq);
1589
1590         mlx5_cqwq_update_db_record(cqwq);
1591
1592         /* ensure cq space is freed before enabling more cqes */
1593         wmb();
1594
1595         return work_done;
1596 }
1597
1598 #ifdef CONFIG_MLX5_CORE_IPOIB
1599
1600 #define MLX5_IB_GRH_SGID_OFFSET 8
1601 #define MLX5_IB_GRH_DGID_OFFSET 24
1602 #define MLX5_GID_SIZE           16
1603
1604 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1605                                          struct mlx5_cqe64 *cqe,
1606                                          u32 cqe_bcnt,
1607                                          struct sk_buff *skb)
1608 {
1609         struct hwtstamp_config *tstamp;
1610         struct mlx5e_rq_stats *stats;
1611         struct net_device *netdev;
1612         struct mlx5e_priv *priv;
1613         char *pseudo_header;
1614         u32 flags_rqpn;
1615         u32 qpn;
1616         u8 *dgid;
1617         u8 g;
1618
1619         qpn = be32_to_cpu(cqe->sop_drop_qpn) & 0xffffff;
1620         netdev = mlx5i_pkey_get_netdev(rq->netdev, qpn);
1621
1622         /* No mapping present, cannot process SKB. This might happen if a child
1623          * interface is going down while having unprocessed CQEs on parent RQ
1624          */
1625         if (unlikely(!netdev)) {
1626                 /* TODO: add drop counters support */
1627                 skb->dev = NULL;
1628                 pr_warn_once("Unable to map QPN %u to dev - dropping skb\n", qpn);
1629                 return;
1630         }
1631
1632         priv = mlx5i_epriv(netdev);
1633         tstamp = &priv->tstamp;
1634         stats = &priv->channel_stats[rq->ix].rq;
1635
1636         flags_rqpn = be32_to_cpu(cqe->flags_rqpn);
1637         g = (flags_rqpn >> 28) & 3;
1638         dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1639         if ((!g) || dgid[0] != 0xff)
1640                 skb->pkt_type = PACKET_HOST;
1641         else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1642                 skb->pkt_type = PACKET_BROADCAST;
1643         else
1644                 skb->pkt_type = PACKET_MULTICAST;
1645
1646         /* Drop packets that this interface sent, ie multicast packets
1647          * that the HCA has replicated.
1648          */
1649         if (g && (qpn == (flags_rqpn & 0xffffff)) &&
1650             (memcmp(netdev->dev_addr + 4, skb->data + MLX5_IB_GRH_SGID_OFFSET,
1651                     MLX5_GID_SIZE) == 0)) {
1652                 skb->dev = NULL;
1653                 return;
1654         }
1655
1656         skb_pull(skb, MLX5_IB_GRH_BYTES);
1657
1658         skb->protocol = *((__be16 *)(skb->data));
1659
1660         if (netdev->features & NETIF_F_RXCSUM) {
1661                 skb->ip_summed = CHECKSUM_COMPLETE;
1662                 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1663                 stats->csum_complete++;
1664         } else {
1665                 skb->ip_summed = CHECKSUM_NONE;
1666                 stats->csum_none++;
1667         }
1668
1669         if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1670                 skb_hwtstamps(skb)->hwtstamp = mlx5e_cqe_ts_to_ns(rq->ptp_cyc2time,
1671                                                                   rq->clock, get_cqe_ts(cqe));
1672         skb_record_rx_queue(skb, rq->ix);
1673
1674         if (likely(netdev->features & NETIF_F_RXHASH))
1675                 mlx5e_skb_set_hash(cqe, skb);
1676
1677         /* 20 bytes of ipoib header and 4 for encap existing */
1678         pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1679         memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1680         skb_reset_mac_header(skb);
1681         skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1682
1683         skb->dev = netdev;
1684
1685         stats->packets++;
1686         stats->bytes += cqe_bcnt;
1687 }
1688
1689 static void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1690 {
1691         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1692         struct mlx5e_wqe_frag_info *wi;
1693         struct sk_buff *skb;
1694         u32 cqe_bcnt;
1695         u16 ci;
1696
1697         ci       = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1698         wi       = get_frag(rq, ci);
1699         cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1700
1701         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1702                 rq->stats->wqe_err++;
1703                 goto wq_free_wqe;
1704         }
1705
1706         skb = INDIRECT_CALL_2(rq->wqe.skb_from_cqe,
1707                               mlx5e_skb_from_cqe_linear,
1708                               mlx5e_skb_from_cqe_nonlinear,
1709                               rq, cqe, wi, cqe_bcnt);
1710         if (!skb)
1711                 goto wq_free_wqe;
1712
1713         mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1714         if (unlikely(!skb->dev)) {
1715                 dev_kfree_skb_any(skb);
1716                 goto wq_free_wqe;
1717         }
1718         napi_gro_receive(rq->cq.napi, skb);
1719
1720 wq_free_wqe:
1721         mlx5e_free_rx_wqe(rq, wi, true);
1722         mlx5_wq_cyc_pop(wq);
1723 }
1724
1725 const struct mlx5e_rx_handlers mlx5i_rx_handlers = {
1726         .handle_rx_cqe       = mlx5i_handle_rx_cqe,
1727         .handle_rx_cqe_mpwqe = NULL, /* Not supported */
1728 };
1729 #endif /* CONFIG_MLX5_CORE_IPOIB */
1730
1731 #ifdef CONFIG_MLX5_EN_IPSEC
1732
1733 static void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1734 {
1735         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1736         struct mlx5e_wqe_frag_info *wi;
1737         struct sk_buff *skb;
1738         u32 cqe_bcnt;
1739         u16 ci;
1740
1741         ci       = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1742         wi       = get_frag(rq, ci);
1743         cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1744
1745         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1746                 rq->stats->wqe_err++;
1747                 goto wq_free_wqe;
1748         }
1749
1750         skb = INDIRECT_CALL_2(rq->wqe.skb_from_cqe,
1751                               mlx5e_skb_from_cqe_linear,
1752                               mlx5e_skb_from_cqe_nonlinear,
1753                               rq, cqe, wi, cqe_bcnt);
1754         if (unlikely(!skb)) /* a DROP, save the page-reuse checks */
1755                 goto wq_free_wqe;
1756
1757         skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb, &cqe_bcnt);
1758         if (unlikely(!skb))
1759                 goto wq_free_wqe;
1760
1761         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1762         napi_gro_receive(rq->cq.napi, skb);
1763
1764 wq_free_wqe:
1765         mlx5e_free_rx_wqe(rq, wi, true);
1766         mlx5_wq_cyc_pop(wq);
1767 }
1768
1769 #endif /* CONFIG_MLX5_EN_IPSEC */
1770
1771 int mlx5e_rq_set_handlers(struct mlx5e_rq *rq, struct mlx5e_params *params, bool xsk)
1772 {
1773         struct net_device *netdev = rq->netdev;
1774         struct mlx5_core_dev *mdev = rq->mdev;
1775         struct mlx5e_priv *priv = rq->priv;
1776
1777         switch (rq->wq_type) {
1778         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1779                 rq->mpwqe.skb_from_cqe_mpwrq = xsk ?
1780                         mlx5e_xsk_skb_from_cqe_mpwrq_linear :
1781                         mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL) ?
1782                                 mlx5e_skb_from_cqe_mpwrq_linear :
1783                                 mlx5e_skb_from_cqe_mpwrq_nonlinear;
1784                 rq->post_wqes = mlx5e_post_rx_mpwqes;
1785                 rq->dealloc_wqe = mlx5e_dealloc_rx_mpwqe;
1786
1787                 rq->handle_rx_cqe = priv->profile->rx_handlers->handle_rx_cqe_mpwqe;
1788                 if (mlx5_fpga_is_ipsec_device(mdev)) {
1789                         netdev_err(netdev, "MPWQE RQ with Innova IPSec offload not supported\n");
1790                         return -EINVAL;
1791                 }
1792                 if (!rq->handle_rx_cqe) {
1793                         netdev_err(netdev, "RX handler of MPWQE RQ is not set\n");
1794                         return -EINVAL;
1795                 }
1796                 break;
1797         default: /* MLX5_WQ_TYPE_CYCLIC */
1798                 rq->wqe.skb_from_cqe = xsk ?
1799                         mlx5e_xsk_skb_from_cqe_linear :
1800                         mlx5e_rx_is_linear_skb(params, NULL) ?
1801                                 mlx5e_skb_from_cqe_linear :
1802                                 mlx5e_skb_from_cqe_nonlinear;
1803                 rq->post_wqes = mlx5e_post_rx_wqes;
1804                 rq->dealloc_wqe = mlx5e_dealloc_rx_wqe;
1805
1806 #ifdef CONFIG_MLX5_EN_IPSEC
1807                 if ((mlx5_fpga_ipsec_device_caps(mdev) & MLX5_ACCEL_IPSEC_CAP_DEVICE) &&
1808                     priv->ipsec)
1809                         rq->handle_rx_cqe = mlx5e_ipsec_handle_rx_cqe;
1810                 else
1811 #endif
1812                         rq->handle_rx_cqe = priv->profile->rx_handlers->handle_rx_cqe;
1813                 if (!rq->handle_rx_cqe) {
1814                         netdev_err(netdev, "RX handler of RQ is not set\n");
1815                         return -EINVAL;
1816                 }
1817         }
1818
1819         return 0;
1820 }
1821
1822 static void mlx5e_trap_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1823 {
1824         struct mlx5e_priv *priv = netdev_priv(rq->netdev);
1825         struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1826         struct mlx5e_wqe_frag_info *wi;
1827         struct devlink_port *dl_port;
1828         struct sk_buff *skb;
1829         u32 cqe_bcnt;
1830         u16 trap_id;
1831         u16 ci;
1832
1833         trap_id  = get_cqe_flow_tag(cqe);
1834         ci       = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1835         wi       = get_frag(rq, ci);
1836         cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1837
1838         if (unlikely(MLX5E_RX_ERR_CQE(cqe))) {
1839                 rq->stats->wqe_err++;
1840                 goto free_wqe;
1841         }
1842
1843         skb = mlx5e_skb_from_cqe_nonlinear(rq, cqe, wi, cqe_bcnt);
1844         if (!skb)
1845                 goto free_wqe;
1846
1847         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1848         skb_push(skb, ETH_HLEN);
1849
1850         dl_port = mlx5e_devlink_get_dl_port(priv);
1851         mlx5_devlink_trap_report(rq->mdev, trap_id, skb, dl_port);
1852         dev_kfree_skb_any(skb);
1853
1854 free_wqe:
1855         mlx5e_free_rx_wqe(rq, wi, false);
1856         mlx5_wq_cyc_pop(wq);
1857 }
1858
1859 void mlx5e_rq_set_trap_handlers(struct mlx5e_rq *rq, struct mlx5e_params *params)
1860 {
1861         rq->wqe.skb_from_cqe = mlx5e_rx_is_linear_skb(params, NULL) ?
1862                                mlx5e_skb_from_cqe_linear :
1863                                mlx5e_skb_from_cqe_nonlinear;
1864         rq->post_wqes = mlx5e_post_rx_wqes;
1865         rq->dealloc_wqe = mlx5e_dealloc_rx_wqe;
1866         rq->handle_rx_cqe = mlx5e_trap_handle_rx_cqe;
1867 }