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 <net/ip6_checksum.h>
  40 #include "en.h"
  41 #include "en_tc.h"
  42 #include "eswitch.h"
  43 #include "en_rep.h"
  44 #include "ipoib/ipoib.h"
  45 #include "en_accel/ipsec_rxtx.h"
  46 #include "lib/clock.h"
  47
  48 static inline bool mlx5e_rx_hw_stamp(struct hwtstamp_config *config)
  49 {
  50         return config->rx_filter == HWTSTAMP_FILTER_ALL;
  51 }
  52
  53 static inline void mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cqcc,
  54                                        void *data)
  55 {
  56         u32 ci = cqcc & cq->wq.fbc.sz_m1;
  57
  58         memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, ci), sizeof(struct mlx5_cqe64));
  59 }
  60
  61 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
  62                                          struct mlx5e_cq *cq, u32 cqcc)
  63 {
  64         mlx5e_read_cqe_slot(cq, cqcc, &cq->title);
  65         cq->decmprs_left        = be32_to_cpu(cq->title.byte_cnt);
  66         cq->decmprs_wqe_counter = be16_to_cpu(cq->title.wqe_counter);
  67         rq->stats.cqe_compress_blks++;
  68 }
  69
  70 static inline void mlx5e_read_mini_arr_slot(struct mlx5e_cq *cq, u32 cqcc)
  71 {
  72         mlx5e_read_cqe_slot(cq, cqcc, cq->mini_arr);
  73         cq->mini_arr_idx = 0;
  74 }
  75
  76 static inline void mlx5e_cqes_update_owner(struct mlx5e_cq *cq, u32 cqcc, int n)
  77 {
  78         struct mlx5_frag_buf_ctrl *fbc = &cq->wq.fbc;
  79         u8 op_own = (cqcc >> fbc->log_sz) & 1;
  80         u32 wq_sz = 1 << fbc->log_sz;
  81         u32 ci = cqcc & fbc->sz_m1;
  82         u32 ci_top = min_t(u32, wq_sz, ci + n);
  83
  84         for (; ci < ci_top; ci++, n--) {
  85                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
  86
  87                 cqe->op_own = op_own;
  88         }
  89
  90         if (unlikely(ci == wq_sz)) {
  91                 op_own = !op_own;
  92                 for (ci = 0; ci < n; ci++) {
  93                         struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
  94
  95                         cqe->op_own = op_own;
  96                 }
  97         }
  98 }
  99
 100 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
 101                                         struct mlx5e_cq *cq, u32 cqcc)
 102 {
 103         cq->title.byte_cnt     = cq->mini_arr[cq->mini_arr_idx].byte_cnt;
 104         cq->title.check_sum    = cq->mini_arr[cq->mini_arr_idx].checksum;
 105         cq->title.op_own      &= 0xf0;
 106         cq->title.op_own      |= 0x01 & (cqcc >> cq->wq.fbc.log_sz);
 107         cq->title.wqe_counter  = cpu_to_be16(cq->decmprs_wqe_counter);
 108
 109         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
 110                 cq->decmprs_wqe_counter +=
 111                         mpwrq_get_cqe_consumed_strides(&cq->title);
 112         else
 113                 cq->decmprs_wqe_counter =
 114                         (cq->decmprs_wqe_counter + 1) & rq->wq.sz_m1;
 115 }
 116
 117 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
 118                                                 struct mlx5e_cq *cq, u32 cqcc)
 119 {
 120         mlx5e_decompress_cqe(rq, cq, cqcc);
 121         cq->title.rss_hash_type   = 0;
 122         cq->title.rss_hash_result = 0;
 123 }
 124
 125 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
 126                                              struct mlx5e_cq *cq,
 127                                              int update_owner_only,
 128                                              int budget_rem)
 129 {
 130         u32 cqcc = cq->wq.cc + update_owner_only;
 131         u32 cqe_count;
 132         u32 i;
 133
 134         cqe_count = min_t(u32, cq->decmprs_left, budget_rem);
 135
 136         for (i = update_owner_only; i < cqe_count;
 137              i++, cq->mini_arr_idx++, cqcc++) {
 138                 if (cq->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
 139                         mlx5e_read_mini_arr_slot(cq, cqcc);
 140
 141                 mlx5e_decompress_cqe_no_hash(rq, cq, cqcc);
 142                 rq->handle_rx_cqe(rq, &cq->title);
 143         }
 144         mlx5e_cqes_update_owner(cq, cq->wq.cc, cqcc - cq->wq.cc);
 145         cq->wq.cc = cqcc;
 146         cq->decmprs_left -= cqe_count;
 147         rq->stats.cqe_compress_pkts += cqe_count;
 148
 149         return cqe_count;
 150 }
 151
 152 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
 153                                               struct mlx5e_cq *cq,
 154                                               int budget_rem)
 155 {
 156         mlx5e_read_title_slot(rq, cq, cq->wq.cc);
 157         mlx5e_read_mini_arr_slot(cq, cq->wq.cc + 1);
 158         mlx5e_decompress_cqe(rq, cq, cq->wq.cc);
 159         rq->handle_rx_cqe(rq, &cq->title);
 160         cq->mini_arr_idx++;
 161
 162         return mlx5e_decompress_cqes_cont(rq, cq, 1, budget_rem) - 1;
 163 }
 164
 165 #define RQ_PAGE_SIZE(rq) ((1 << rq->buff.page_order) << PAGE_SHIFT)
 166
 167 static inline bool mlx5e_page_is_reserved(struct page *page)
 168 {
 169         return page_is_pfmemalloc(page) || page_to_nid(page) != numa_mem_id();
 170 }
 171
 172 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
 173                                       struct mlx5e_dma_info *dma_info)
 174 {
 175         struct mlx5e_page_cache *cache = &rq->page_cache;
 176         u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
 177
 178         if (tail_next == cache->head) {
 179                 rq->stats.cache_full++;
 180                 return false;
 181         }
 182
 183         if (unlikely(mlx5e_page_is_reserved(dma_info->page))) {
 184                 rq->stats.cache_waive++;
 185                 return false;
 186         }
 187
 188         cache->page_cache[cache->tail] = *dma_info;
 189         cache->tail = tail_next;
 190         return true;
 191 }
 192
 193 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
 194                                       struct mlx5e_dma_info *dma_info)
 195 {
 196         struct mlx5e_page_cache *cache = &rq->page_cache;
 197
 198         if (unlikely(cache->head == cache->tail)) {
 199                 rq->stats.cache_empty++;
 200                 return false;
 201         }
 202
 203         if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
 204                 rq->stats.cache_busy++;
 205                 return false;
 206         }
 207
 208         *dma_info = cache->page_cache[cache->head];
 209         cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
 210         rq->stats.cache_reuse++;
 211
 212         dma_sync_single_for_device(rq->pdev, dma_info->addr,
 213                                    RQ_PAGE_SIZE(rq),
 214                                    DMA_FROM_DEVICE);
 215         return true;
 216 }
 217
 218 static inline int mlx5e_page_alloc_mapped(struct mlx5e_rq *rq,
 219                                           struct mlx5e_dma_info *dma_info)
 220 {
 221         if (mlx5e_rx_cache_get(rq, dma_info))
 222                 return 0;
 223
 224         dma_info->page = dev_alloc_pages(rq->buff.page_order);
 225         if (unlikely(!dma_info->page))
 226                 return -ENOMEM;
 227
 228         dma_info->addr = dma_map_page(rq->pdev, dma_info->page, 0,
 229                                       RQ_PAGE_SIZE(rq), rq->buff.map_dir);
 230         if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
 231                 put_page(dma_info->page);
 232                 dma_info->page = NULL;
 233                 return -ENOMEM;
 234         }
 235
 236         return 0;
 237 }
 238
 239 void mlx5e_page_release(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info,
 240                         bool recycle)
 241 {
 242         if (likely(recycle) && mlx5e_rx_cache_put(rq, dma_info))
 243                 return;
 244
 245         dma_unmap_page(rq->pdev, dma_info->addr, RQ_PAGE_SIZE(rq),
 246                        rq->buff.map_dir);
 247         put_page(dma_info->page);
 248 }
 249
 250 static inline bool mlx5e_page_reuse(struct mlx5e_rq *rq,
 251                                     struct mlx5e_wqe_frag_info *wi)
 252 {
 253         return rq->wqe.page_reuse && wi->di.page &&
 254                 (wi->offset + rq->wqe.frag_sz <= RQ_PAGE_SIZE(rq)) &&
 255                 !mlx5e_page_is_reserved(wi->di.page);
 256 }
 257
 258 static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
 259 {
 260         struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
 261
 262         /* check if page exists, hence can be reused */
 263         if (!wi->di.page) {
 264                 if (unlikely(mlx5e_page_alloc_mapped(rq, &wi->di)))
 265                         return -ENOMEM;
 266                 wi->offset = 0;
 267         }
 268
 269         wqe->data.addr = cpu_to_be64(wi->di.addr + wi->offset + rq->buff.headroom);
 270         return 0;
 271 }
 272
 273 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
 274                                      struct mlx5e_wqe_frag_info *wi)
 275 {
 276         mlx5e_page_release(rq, &wi->di, true);
 277         wi->di.page = NULL;
 278 }
 279
 280 static inline void mlx5e_free_rx_wqe_reuse(struct mlx5e_rq *rq,
 281                                            struct mlx5e_wqe_frag_info *wi)
 282 {
 283         if (mlx5e_page_reuse(rq, wi)) {
 284                 rq->stats.page_reuse++;
 285                 return;
 286         }
 287
 288         mlx5e_free_rx_wqe(rq, wi);
 289 }
 290
 291 void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
 292 {
 293         struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
 294
 295         if (wi->di.page)
 296                 mlx5e_free_rx_wqe(rq, wi);
 297 }
 298
 299 static inline void mlx5e_add_skb_frag_mpwqe(struct mlx5e_rq *rq,
 300                                             struct sk_buff *skb,
 301                                             struct mlx5e_dma_info *di,
 302                                             u32 frag_offset, u32 len)
 303 {
 304         unsigned int truesize = ALIGN(len, BIT(rq->mpwqe.log_stride_sz));
 305
 306         dma_sync_single_for_cpu(rq->pdev,
 307                                 di->addr + frag_offset,
 308                                 len, DMA_FROM_DEVICE);
 309         page_ref_inc(di->page);
 310         skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
 311                         di->page, frag_offset, len, truesize);
 312 }
 313
 314 static inline void
 315 mlx5e_copy_skb_header_mpwqe(struct device *pdev,
 316                             struct sk_buff *skb,
 317                             struct mlx5e_dma_info *dma_info,
 318                             u32 offset, u32 headlen)
 319 {
 320         u16 headlen_pg = min_t(u32, headlen, PAGE_SIZE - offset);
 321         unsigned int len;
 322
 323          /* Aligning len to sizeof(long) optimizes memcpy performance */
 324         len = ALIGN(headlen_pg, sizeof(long));
 325         dma_sync_single_for_cpu(pdev, dma_info->addr + offset, len,
 326                                 DMA_FROM_DEVICE);
 327         skb_copy_to_linear_data(skb, page_address(dma_info->page) + offset, len);
 328
 329         if (unlikely(offset + headlen > PAGE_SIZE)) {
 330                 dma_info++;
 331                 headlen_pg = len;
 332                 len = ALIGN(headlen - headlen_pg, sizeof(long));
 333                 dma_sync_single_for_cpu(pdev, dma_info->addr, len,
 334                                         DMA_FROM_DEVICE);
 335                 skb_copy_to_linear_data_offset(skb, headlen_pg,
 336                                                page_address(dma_info->page),
 337                                                len);
 338         }
 339 }
 340
 341 void mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
 342 {
 343         const bool no_xdp_xmit =
 344                 bitmap_empty(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
 345         struct mlx5e_dma_info *dma_info = wi->umr.dma_info;
 346         int i;
 347
 348         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++)
 349                 if (no_xdp_xmit || !test_bit(i, wi->xdp_xmit_bitmap))
 350                         mlx5e_page_release(rq, &dma_info[i], true);
 351 }
 352
 353 static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
 354 {
 355         struct mlx5_wq_ll *wq = &rq->wq;
 356         struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
 357
 358         rq->mpwqe.umr_in_progress = false;
 359
 360         mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
 361
 362         /* ensure wqes are visible to device before updating doorbell record */
 363         dma_wmb();
 364
 365         mlx5_wq_ll_update_db_record(wq);
 366 }
 367
 368 static inline u16 mlx5e_icosq_wrap_cnt(struct mlx5e_icosq *sq)
 369 {
 370         return sq->pc >> MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
 371 }
 372
 373 static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 374 {
 375         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 376         struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
 377         struct mlx5e_icosq *sq = &rq->channel->icosq;
 378         struct mlx5_wq_cyc *wq = &sq->wq;
 379         struct mlx5e_umr_wqe *umr_wqe;
 380         u16 xlt_offset = ix << (MLX5E_LOG_ALIGNED_MPWQE_PPW - 1);
 381         int err;
 382         u16 pi;
 383         int i;
 384
 385         /* fill sq edge with nops to avoid wqe wrap around */
 386         while ((pi = (sq->pc & wq->sz_m1)) > sq->edge) {
 387                 sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_NOP;
 388                 mlx5e_post_nop(wq, sq->sqn, &sq->pc);
 389         }
 390
 391         umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
 392         if (unlikely(mlx5e_icosq_wrap_cnt(sq) < 2))
 393                 memcpy(umr_wqe, &rq->mpwqe.umr_wqe,
 394                        offsetof(struct mlx5e_umr_wqe, inline_mtts));
 395
 396         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
 397                 err = mlx5e_page_alloc_mapped(rq, dma_info);
 398                 if (unlikely(err))
 399                         goto err_unmap;
 400                 umr_wqe->inline_mtts[i].ptag = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
 401         }
 402
 403         bitmap_zero(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
 404         wi->consumed_strides = 0;
 405
 406         rq->mpwqe.umr_in_progress = true;
 407
 408         umr_wqe->ctrl.opmod_idx_opcode =
 409                 cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
 410                             MLX5_OPCODE_UMR);
 411         umr_wqe->uctrl.xlt_offset = cpu_to_be16(xlt_offset);
 412
 413         sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_UMR;
 414         sq->pc += MLX5E_UMR_WQEBBS;
 415         mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &umr_wqe->ctrl);
 416
 417         return 0;
 418
 419 err_unmap:
 420         while (--i >= 0) {
 421                 dma_info--;
 422                 mlx5e_page_release(rq, dma_info, true);
 423         }
 424         rq->stats.buff_alloc_err++;
 425
 426         return err;
 427 }
 428
 429 void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 430 {
 431         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 432
 433         mlx5e_free_rx_mpwqe(rq, wi);
 434 }
 435
 436 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
 437 {
 438         struct mlx5_wq_ll *wq = &rq->wq;
 439         int err;
 440
 441         if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
 442                 return false;
 443
 444         if (mlx5_wq_ll_is_full(wq))
 445                 return false;
 446
 447         do {
 448                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
 449
 450                 err = mlx5e_alloc_rx_wqe(rq, wqe, wq->head);
 451                 if (unlikely(err)) {
 452                         rq->stats.buff_alloc_err++;
 453                         break;
 454                 }
 455
 456                 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
 457         } while (!mlx5_wq_ll_is_full(wq));
 458
 459         /* ensure wqes are visible to device before updating doorbell record */
 460         dma_wmb();
 461
 462         mlx5_wq_ll_update_db_record(wq);
 463
 464         return !!err;
 465 }
 466
 467 static inline void mlx5e_poll_ico_single_cqe(struct mlx5e_cq *cq,
 468                                              struct mlx5e_icosq *sq,
 469                                              struct mlx5e_rq *rq,
 470                                              struct mlx5_cqe64 *cqe)
 471 {
 472         struct mlx5_wq_cyc *wq = &sq->wq;
 473         u16 ci = be16_to_cpu(cqe->wqe_counter) & wq->sz_m1;
 474         struct mlx5e_sq_wqe_info *icowi = &sq->db.ico_wqe[ci];
 475
 476         mlx5_cqwq_pop(&cq->wq);
 477
 478         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_REQ)) {
 479                 netdev_WARN_ONCE(cq->channel->netdev,
 480                                  "Bad OP in ICOSQ CQE: 0x%x\n", cqe->op_own);
 481                 return;
 482         }
 483
 484         if (likely(icowi->opcode == MLX5_OPCODE_UMR)) {
 485                 mlx5e_post_rx_mpwqe(rq);
 486                 return;
 487         }
 488
 489         if (unlikely(icowi->opcode != MLX5_OPCODE_NOP))
 490                 netdev_WARN_ONCE(cq->channel->netdev,
 491                                  "Bad OPCODE in ICOSQ WQE info: 0x%x\n", icowi->opcode);
 492 }
 493
 494 static void mlx5e_poll_ico_cq(struct mlx5e_cq *cq, struct mlx5e_rq *rq)
 495 {
 496         struct mlx5e_icosq *sq = container_of(cq, struct mlx5e_icosq, cq);
 497         struct mlx5_cqe64 *cqe;
 498
 499         if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
 500                 return;
 501
 502         cqe = mlx5_cqwq_get_cqe(&cq->wq);
 503         if (likely(!cqe))
 504                 return;
 505
 506         /* by design, there's only a single cqe */
 507         mlx5e_poll_ico_single_cqe(cq, sq, rq, cqe);
 508
 509         mlx5_cqwq_update_db_record(&cq->wq);
 510 }
 511
 512 bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
 513 {
 514         struct mlx5_wq_ll *wq = &rq->wq;
 515
 516         if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
 517                 return false;
 518
 519         mlx5e_poll_ico_cq(&rq->channel->icosq.cq, rq);
 520
 521         if (mlx5_wq_ll_is_full(wq))
 522                 return false;
 523
 524         if (!rq->mpwqe.umr_in_progress)
 525                 mlx5e_alloc_rx_mpwqe(rq, wq->head);
 526
 527         return false;
 528 }
 529
 530 static void mlx5e_lro_update_tcp_hdr(struct mlx5_cqe64 *cqe, struct tcphdr *tcp)
 531 {
 532         u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
 533         u8 tcp_ack     = (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA) ||
 534                          (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA);
 535
 536         tcp->check                      = 0;
 537         tcp->psh                        = get_cqe_lro_tcppsh(cqe);
 538
 539         if (tcp_ack) {
 540                 tcp->ack                = 1;
 541                 tcp->ack_seq            = cqe->lro_ack_seq_num;
 542                 tcp->window             = cqe->lro_tcp_win;
 543         }
 544 }
 545
 546 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
 547                                  u32 cqe_bcnt)
 548 {
 549         struct ethhdr   *eth = (struct ethhdr *)(skb->data);
 550         struct tcphdr   *tcp;
 551         int network_depth = 0;
 552         __wsum check;
 553         __be16 proto;
 554         u16 tot_len;
 555         void *ip_p;
 556
 557         proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
 558
 559         tot_len = cqe_bcnt - network_depth;
 560         ip_p = skb->data + network_depth;
 561
 562         if (proto == htons(ETH_P_IP)) {
 563                 struct iphdr *ipv4 = ip_p;
 564
 565                 tcp = ip_p + sizeof(struct iphdr);
 566                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 567
 568                 ipv4->ttl               = cqe->lro_min_ttl;
 569                 ipv4->tot_len           = cpu_to_be16(tot_len);
 570                 ipv4->check             = 0;
 571                 ipv4->check             = ip_fast_csum((unsigned char *)ipv4,
 572                                                        ipv4->ihl);
 573
 574                 mlx5e_lro_update_tcp_hdr(cqe, tcp);
 575                 check = csum_partial(tcp, tcp->doff * 4,
 576                                      csum_unfold((__force __sum16)cqe->check_sum));
 577                 /* Almost done, don't forget the pseudo header */
 578                 tcp->check = csum_tcpudp_magic(ipv4->saddr, ipv4->daddr,
 579                                                tot_len - sizeof(struct iphdr),
 580                                                IPPROTO_TCP, check);
 581         } else {
 582                 u16 payload_len = tot_len - sizeof(struct ipv6hdr);
 583                 struct ipv6hdr *ipv6 = ip_p;
 584
 585                 tcp = ip_p + sizeof(struct ipv6hdr);
 586                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 587
 588                 ipv6->hop_limit         = cqe->lro_min_ttl;
 589                 ipv6->payload_len       = cpu_to_be16(payload_len);
 590
 591                 mlx5e_lro_update_tcp_hdr(cqe, tcp);
 592                 check = csum_partial(tcp, tcp->doff * 4,
 593                                      csum_unfold((__force __sum16)cqe->check_sum));
 594                 /* Almost done, don't forget the pseudo header */
 595                 tcp->check = csum_ipv6_magic(&ipv6->saddr, &ipv6->daddr, payload_len,
 596                                              IPPROTO_TCP, check);
 597         }
 598 }
 599
 600 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
 601                                       struct sk_buff *skb)
 602 {
 603         u8 cht = cqe->rss_hash_type;
 604         int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
 605                  (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
 606                                             PKT_HASH_TYPE_NONE;
 607         skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
 608 }
 609
 610 static inline bool is_last_ethertype_ip(struct sk_buff *skb, int *network_depth)
 611 {
 612         __be16 ethertype = ((struct ethhdr *)skb->data)->h_proto;
 613
 614         ethertype = __vlan_get_protocol(skb, ethertype, network_depth);
 615         return (ethertype == htons(ETH_P_IP) || ethertype == htons(ETH_P_IPV6));
 616 }
 617
 618 static __be32 mlx5e_get_fcs(struct sk_buff *skb)
 619 {
 620         int last_frag_sz, bytes_in_prev, nr_frags;
 621         u8 *fcs_p1, *fcs_p2;
 622         skb_frag_t *last_frag;
 623         __be32 fcs_bytes;
 624
 625         if (!skb_is_nonlinear(skb))
 626                 return *(__be32 *)(skb->data + skb->len - ETH_FCS_LEN);
 627
 628         nr_frags = skb_shinfo(skb)->nr_frags;
 629         last_frag = &skb_shinfo(skb)->frags[nr_frags - 1];
 630         last_frag_sz = skb_frag_size(last_frag);
 631
 632         /* If all FCS data is in last frag */
 633         if (last_frag_sz >= ETH_FCS_LEN)
 634                 return *(__be32 *)(skb_frag_address(last_frag) +
 635                                    last_frag_sz - ETH_FCS_LEN);
 636
 637         fcs_p2 = (u8 *)skb_frag_address(last_frag);
 638         bytes_in_prev = ETH_FCS_LEN - last_frag_sz;
 639
 640         /* Find where the other part of the FCS is - Linear or another frag */
 641         if (nr_frags == 1) {
 642                 fcs_p1 = skb_tail_pointer(skb);
 643         } else {
 644                 skb_frag_t *prev_frag = &skb_shinfo(skb)->frags[nr_frags - 2];
 645
 646                 fcs_p1 = skb_frag_address(prev_frag) +
 647                             skb_frag_size(prev_frag);
 648         }
 649         fcs_p1 -= bytes_in_prev;
 650
 651         memcpy(&fcs_bytes, fcs_p1, bytes_in_prev);
 652         memcpy(((u8 *)&fcs_bytes) + bytes_in_prev, fcs_p2, last_frag_sz);
 653
 654         return fcs_bytes;
 655 }
 656
 657 static inline void mlx5e_handle_csum(struct net_device *netdev,
 658                                      struct mlx5_cqe64 *cqe,
 659                                      struct mlx5e_rq *rq,
 660                                      struct sk_buff *skb,
 661                                      bool   lro)
 662 {
 663         int network_depth = 0;
 664
 665         if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
 666                 goto csum_none;
 667
 668         if (lro) {
 669                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 670                 rq->stats.csum_unnecessary++;
 671                 return;
 672         }
 673
 674         if (likely(is_last_ethertype_ip(skb, &network_depth))) {
 675                 skb->ip_summed = CHECKSUM_COMPLETE;
 676                 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
 677                 if (network_depth > ETH_HLEN)
 678                         /* CQE csum is calculated from the IP header and does
 679                          * not cover VLAN headers (if present). This will add
 680                          * the checksum manually.
 681                          */
 682                         skb->csum = csum_partial(skb->data + ETH_HLEN,
 683                                                  network_depth - ETH_HLEN,
 684                                                  skb->csum);
 685                 if (unlikely(netdev->features & NETIF_F_RXFCS))
 686                         skb->csum = csum_add(skb->csum,
 687                                              (__force __wsum)mlx5e_get_fcs(skb));
 688                 rq->stats.csum_complete++;
 689                 return;
 690         }
 691
 692         if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
 693                    (cqe->hds_ip_ext & CQE_L4_OK))) {
 694                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 695                 if (cqe_is_tunneled(cqe)) {
 696                         skb->csum_level = 1;
 697                         skb->encapsulation = 1;
 698                         rq->stats.csum_unnecessary_inner++;
 699                         return;
 700                 }
 701                 rq->stats.csum_unnecessary++;
 702                 return;
 703         }
 704 csum_none:
 705         skb->ip_summed = CHECKSUM_NONE;
 706         rq->stats.csum_none++;
 707 }
 708
 709 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
 710                                       u32 cqe_bcnt,
 711                                       struct mlx5e_rq *rq,
 712                                       struct sk_buff *skb)
 713 {
 714         struct net_device *netdev = rq->netdev;
 715         int lro_num_seg;
 716
 717         skb->mac_len = ETH_HLEN;
 718         lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
 719         if (lro_num_seg > 1) {
 720                 mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
 721                 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
 722                 /* Subtract one since we already counted this as one
 723                  * "regular" packet in mlx5e_complete_rx_cqe()
 724                  */
 725                 rq->stats.packets += lro_num_seg - 1;
 726                 rq->stats.lro_packets++;
 727                 rq->stats.lro_bytes += cqe_bcnt;
 728         }
 729
 730         if (unlikely(mlx5e_rx_hw_stamp(rq->tstamp)))
 731                 skb_hwtstamps(skb)->hwtstamp =
 732                                 mlx5_timecounter_cyc2time(rq->clock, get_cqe_ts(cqe));
 733
 734         skb_record_rx_queue(skb, rq->ix);
 735
 736         if (likely(netdev->features & NETIF_F_RXHASH))
 737                 mlx5e_skb_set_hash(cqe, skb);
 738
 739         if (cqe_has_vlan(cqe)) {
 740                 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 741                                        be16_to_cpu(cqe->vlan_info));
 742                 rq->stats.removed_vlan_packets++;
 743         }
 744
 745         skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
 746
 747         mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
 748         skb->protocol = eth_type_trans(skb, netdev);
 749 }
 750
 751 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
 752                                          struct mlx5_cqe64 *cqe,
 753                                          u32 cqe_bcnt,
 754                                          struct sk_buff *skb)
 755 {
 756         rq->stats.packets++;
 757         rq->stats.bytes += cqe_bcnt;
 758         mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
 759 }
 760
 761 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_xdpsq *sq)
 762 {
 763         struct mlx5_wq_cyc *wq = &sq->wq;
 764         struct mlx5e_tx_wqe *wqe;
 765         u16 pi = (sq->pc - 1) & wq->sz_m1; /* last pi */
 766
 767         wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
 768
 769         mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &wqe->ctrl);
 770 }
 771
 772 static inline bool mlx5e_xmit_xdp_frame(struct mlx5e_rq *rq,
 773                                         struct mlx5e_dma_info *di,
 774                                         const struct xdp_buff *xdp)
 775 {
 776         struct mlx5e_xdpsq       *sq   = &rq->xdpsq;
 777         struct mlx5_wq_cyc       *wq   = &sq->wq;
 778         u16                       pi   = sq->pc & wq->sz_m1;
 779         struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
 780
 781         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
 782         struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
 783         struct mlx5_wqe_data_seg *dseg;
 784
 785         ptrdiff_t data_offset = xdp->data - xdp->data_hard_start;
 786         dma_addr_t dma_addr  = di->addr + data_offset;
 787         unsigned int dma_len = xdp->data_end - xdp->data;
 788
 789         prefetchw(wqe);
 790
 791         if (unlikely(dma_len < MLX5E_XDP_MIN_INLINE || rq->hw_mtu < dma_len)) {
 792                 rq->stats.xdp_drop++;
 793                 return false;
 794         }
 795
 796         if (unlikely(!mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1))) {
 797                 if (sq->db.doorbell) {
 798                         /* SQ is full, ring doorbell */
 799                         mlx5e_xmit_xdp_doorbell(sq);
 800                         sq->db.doorbell = false;
 801                 }
 802                 rq->stats.xdp_tx_full++;
 803                 return false;
 804         }
 805
 806         dma_sync_single_for_device(sq->pdev, dma_addr, dma_len, PCI_DMA_TODEVICE);
 807
 808         cseg->fm_ce_se = 0;
 809
 810         dseg = (struct mlx5_wqe_data_seg *)eseg + 1;
 811
 812         /* copy the inline part if required */
 813         if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
 814                 memcpy(eseg->inline_hdr.start, xdp->data, MLX5E_XDP_MIN_INLINE);
 815                 eseg->inline_hdr.sz = cpu_to_be16(MLX5E_XDP_MIN_INLINE);
 816                 dma_len  -= MLX5E_XDP_MIN_INLINE;
 817                 dma_addr += MLX5E_XDP_MIN_INLINE;
 818                 dseg++;
 819         }
 820
 821         /* write the dma part */
 822         dseg->addr       = cpu_to_be64(dma_addr);
 823         dseg->byte_count = cpu_to_be32(dma_len);
 824
 825         cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_SEND);
 826
 827         /* move page to reference to sq responsibility,
 828          * and mark so it's not put back in page-cache.
 829          */
 830         __set_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags); /* non-atomic */
 831         sq->db.di[pi] = *di;
 832         sq->pc++;
 833
 834         sq->db.doorbell = true;
 835
 836         rq->stats.xdp_tx++;
 837         return true;
 838 }
 839
 840 /* returns true if packet was consumed by xdp */
 841 static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq,
 842                                    struct mlx5e_dma_info *di,
 843                                    void *va, u16 *rx_headroom, u32 *len)
 844 {
 845         const struct bpf_prog *prog = READ_ONCE(rq->xdp_prog);
 846         struct xdp_buff xdp;
 847         u32 act;
 848
 849         if (!prog)
 850                 return false;
 851
 852         xdp.data = va + *rx_headroom;
 853         xdp_set_data_meta_invalid(&xdp);
 854         xdp.data_end = xdp.data + *len;
 855         xdp.data_hard_start = va;
 856         xdp.rxq = &rq->xdp_rxq;
 857
 858         act = bpf_prog_run_xdp(prog, &xdp);
 859         switch (act) {
 860         case XDP_PASS:
 861                 *rx_headroom = xdp.data - xdp.data_hard_start;
 862                 *len = xdp.data_end - xdp.data;
 863                 return false;
 864         case XDP_TX:
 865                 if (unlikely(!mlx5e_xmit_xdp_frame(rq, di, &xdp)))
 866                         trace_xdp_exception(rq->netdev, prog, act);
 867                 return true;
 868         default:
 869                 bpf_warn_invalid_xdp_action(act);
 870         case XDP_ABORTED:
 871                 trace_xdp_exception(rq->netdev, prog, act);
 872         case XDP_DROP:
 873                 rq->stats.xdp_drop++;
 874                 return true;
 875         }
 876 }
 877
 878 static inline
 879 struct sk_buff *mlx5e_build_linear_skb(struct mlx5e_rq *rq, void *va,
 880                                        u32 frag_size, u16 headroom,
 881                                        u32 cqe_bcnt)
 882 {
 883         struct sk_buff *skb = build_skb(va, frag_size);
 884
 885         if (unlikely(!skb)) {
 886                 rq->stats.buff_alloc_err++;
 887                 return NULL;
 888         }
 889
 890         skb_reserve(skb, headroom);
 891         skb_put(skb, cqe_bcnt);
 892
 893         return skb;
 894 }
 895
 896 static inline
 897 struct sk_buff *skb_from_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
 898                              struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
 899 {
 900         struct mlx5e_dma_info *di = &wi->di;
 901         u16 rx_headroom = rq->buff.headroom;
 902         struct sk_buff *skb;
 903         void *va, *data;
 904         bool consumed;
 905         u32 frag_size;
 906
 907         va             = page_address(di->page) + wi->offset;
 908         data           = va + rx_headroom;
 909         frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
 910
 911         dma_sync_single_range_for_cpu(rq->pdev, di->addr, wi->offset,
 912                                       frag_size, DMA_FROM_DEVICE);
 913         prefetch(data);
 914         wi->offset += frag_size;
 915
 916         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
 917                 rq->stats.wqe_err++;
 918                 return NULL;
 919         }
 920
 921         rcu_read_lock();
 922         consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt);
 923         rcu_read_unlock();
 924         if (consumed)
 925                 return NULL; /* page/packet was consumed by XDP */
 926
 927         skb = mlx5e_build_linear_skb(rq, va, frag_size, rx_headroom, cqe_bcnt);
 928         if (unlikely(!skb))
 929                 return NULL;
 930
 931         /* queue up for recycling/reuse */
 932         page_ref_inc(di->page);
 933
 934         return skb;
 935 }
 936
 937 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 938 {
 939         struct mlx5e_wqe_frag_info *wi;
 940         struct mlx5e_rx_wqe *wqe;
 941         __be16 wqe_counter_be;
 942         struct sk_buff *skb;
 943         u16 wqe_counter;
 944         u32 cqe_bcnt;
 945
 946         wqe_counter_be = cqe->wqe_counter;
 947         wqe_counter    = be16_to_cpu(wqe_counter_be);
 948         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
 949         wi             = &rq->wqe.frag_info[wqe_counter];
 950         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
 951
 952         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
 953         if (!skb) {
 954                 /* probably for XDP */
 955                 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)) {
 956                         wi->di.page = NULL;
 957                         /* do not return page to cache, it will be returned on XDP_TX completion */
 958                         goto wq_ll_pop;
 959                 }
 960                 /* probably an XDP_DROP, save the page-reuse checks */
 961                 mlx5e_free_rx_wqe(rq, wi);
 962                 goto wq_ll_pop;
 963         }
 964
 965         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 966         napi_gro_receive(rq->cq.napi, skb);
 967
 968         mlx5e_free_rx_wqe_reuse(rq, wi);
 969 wq_ll_pop:
 970         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
 971                        &wqe->next.next_wqe_index);
 972 }
 973
 974 #ifdef CONFIG_MLX5_ESWITCH
 975 void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 976 {
 977         struct net_device *netdev = rq->netdev;
 978         struct mlx5e_priv *priv = netdev_priv(netdev);
 979         struct mlx5e_rep_priv *rpriv  = priv->ppriv;
 980         struct mlx5_eswitch_rep *rep = rpriv->rep;
 981         struct mlx5e_wqe_frag_info *wi;
 982         struct mlx5e_rx_wqe *wqe;
 983         struct sk_buff *skb;
 984         __be16 wqe_counter_be;
 985         u16 wqe_counter;
 986         u32 cqe_bcnt;
 987
 988         wqe_counter_be = cqe->wqe_counter;
 989         wqe_counter    = be16_to_cpu(wqe_counter_be);
 990         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
 991         wi             = &rq->wqe.frag_info[wqe_counter];
 992         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
 993
 994         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
 995         if (!skb) {
 996                 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)) {
 997                         wi->di.page = NULL;
 998                         /* do not return page to cache, it will be returned on XDP_TX completion */
 999                         goto wq_ll_pop;
1000                 }
1001                 /* probably an XDP_DROP, save the page-reuse checks */
1002                 mlx5e_free_rx_wqe(rq, wi);
1003                 goto wq_ll_pop;
1004         }
1005
1006         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1007
1008         if (rep->vlan && skb_vlan_tag_present(skb))
1009                 skb_vlan_pop(skb);
1010
1011         napi_gro_receive(rq->cq.napi, skb);
1012
1013         mlx5e_free_rx_wqe_reuse(rq, wi);
1014 wq_ll_pop:
1015         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1016                        &wqe->next.next_wqe_index);
1017 }
1018 #endif
1019
1020 struct sk_buff *
1021 mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
1022                                    u16 cqe_bcnt, u32 head_offset, u32 page_idx)
1023 {
1024         u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
1025         struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
1026         u32 frag_offset    = head_offset + headlen;
1027         u32 byte_cnt       = cqe_bcnt - headlen;
1028         struct mlx5e_dma_info *head_di = di;
1029         struct sk_buff *skb;
1030
1031         skb = napi_alloc_skb(rq->cq.napi,
1032                              ALIGN(MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, sizeof(long)));
1033         if (unlikely(!skb)) {
1034                 rq->stats.buff_alloc_err++;
1035                 return NULL;
1036         }
1037
1038         prefetchw(skb->data);
1039
1040         if (unlikely(frag_offset >= PAGE_SIZE)) {
1041                 di++;
1042                 frag_offset -= PAGE_SIZE;
1043         }
1044
1045         while (byte_cnt) {
1046                 u32 pg_consumed_bytes =
1047                         min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
1048
1049                 mlx5e_add_skb_frag_mpwqe(rq, skb, di, frag_offset,
1050                                          pg_consumed_bytes);
1051                 byte_cnt -= pg_consumed_bytes;
1052                 frag_offset = 0;
1053                 di++;
1054         }
1055         /* copy header */
1056         mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, head_di,
1057                                     head_offset, headlen);
1058         /* skb linear part was allocated with headlen and aligned to long */
1059         skb->tail += headlen;
1060         skb->len  += headlen;
1061
1062         return skb;
1063 }
1064
1065 struct sk_buff *
1066 mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
1067                                 u16 cqe_bcnt, u32 head_offset, u32 page_idx)
1068 {
1069         struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
1070         u16 rx_headroom = rq->buff.headroom;
1071         u32 cqe_bcnt32 = cqe_bcnt;
1072         struct sk_buff *skb;
1073         void *va, *data;
1074         u32 frag_size;
1075         bool consumed;
1076
1077         va             = page_address(di->page) + head_offset;
1078         data           = va + rx_headroom;
1079         frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt32);
1080
1081         dma_sync_single_range_for_cpu(rq->pdev, di->addr, head_offset,
1082                                       frag_size, DMA_FROM_DEVICE);
1083         prefetch(data);
1084
1085         rcu_read_lock();
1086         consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt32);
1087         rcu_read_unlock();
1088         if (consumed) {
1089                 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags))
1090                         __set_bit(page_idx, wi->xdp_xmit_bitmap); /* non-atomic */
1091                 return NULL; /* page/packet was consumed by XDP */
1092         }
1093
1094         skb = mlx5e_build_linear_skb(rq, va, frag_size, rx_headroom, cqe_bcnt32);
1095         if (unlikely(!skb))
1096                 return NULL;
1097
1098         /* queue up for recycling/reuse */
1099         page_ref_inc(di->page);
1100
1101         return skb;
1102 }
1103
1104 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1105 {
1106         u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
1107         u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
1108         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
1109         u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
1110         u32 wqe_offset     = stride_ix << rq->mpwqe.log_stride_sz;
1111         u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
1112         u32 page_idx       = wqe_offset >> PAGE_SHIFT;
1113         struct mlx5e_rx_wqe *wqe;
1114         struct sk_buff *skb;
1115         u16 cqe_bcnt;
1116
1117         wi->consumed_strides += cstrides;
1118
1119         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
1120                 rq->stats.wqe_err++;
1121                 goto mpwrq_cqe_out;
1122         }
1123
1124         if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1125                 rq->stats.mpwqe_filler++;
1126                 goto mpwrq_cqe_out;
1127         }
1128
1129         cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1130
1131         skb = rq->mpwqe.skb_from_cqe_mpwrq(rq, wi, cqe_bcnt, head_offset,
1132                                            page_idx);
1133         if (!skb)
1134                 goto mpwrq_cqe_out;
1135
1136         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1137         napi_gro_receive(rq->cq.napi, skb);
1138
1139 mpwrq_cqe_out:
1140         if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1141                 return;
1142
1143         wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_id);
1144         mlx5e_free_rx_mpwqe(rq, wi);
1145         mlx5_wq_ll_pop(&rq->wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1146 }
1147
1148 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
1149 {
1150         struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
1151         struct mlx5e_xdpsq *xdpsq;
1152         struct mlx5_cqe64 *cqe;
1153         int work_done = 0;
1154
1155         if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
1156                 return 0;
1157
1158         if (cq->decmprs_left)
1159                 work_done += mlx5e_decompress_cqes_cont(rq, cq, 0, budget);
1160
1161         cqe = mlx5_cqwq_get_cqe(&cq->wq);
1162         if (!cqe)
1163                 return 0;
1164
1165         xdpsq = &rq->xdpsq;
1166
1167         do {
1168                 if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1169                         work_done +=
1170                                 mlx5e_decompress_cqes_start(rq, cq,
1171                                                             budget - work_done);
1172                         continue;
1173                 }
1174
1175                 mlx5_cqwq_pop(&cq->wq);
1176
1177                 rq->handle_rx_cqe(rq, cqe);
1178         } while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1179
1180         if (xdpsq->db.doorbell) {
1181                 mlx5e_xmit_xdp_doorbell(xdpsq);
1182                 xdpsq->db.doorbell = false;
1183         }
1184
1185         mlx5_cqwq_update_db_record(&cq->wq);
1186
1187         /* ensure cq space is freed before enabling more cqes */
1188         wmb();
1189
1190         return work_done;
1191 }
1192
1193 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
1194 {
1195         struct mlx5e_xdpsq *sq;
1196         struct mlx5_cqe64 *cqe;
1197         struct mlx5e_rq *rq;
1198         u16 sqcc;
1199         int i;
1200
1201         sq = container_of(cq, struct mlx5e_xdpsq, cq);
1202
1203         if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
1204                 return false;
1205
1206         cqe = mlx5_cqwq_get_cqe(&cq->wq);
1207         if (!cqe)
1208                 return false;
1209
1210         rq = container_of(sq, struct mlx5e_rq, xdpsq);
1211
1212         /* sq->cc must be updated only after mlx5_cqwq_update_db_record(),
1213          * otherwise a cq overrun may occur
1214          */
1215         sqcc = sq->cc;
1216
1217         i = 0;
1218         do {
1219                 u16 wqe_counter;
1220                 bool last_wqe;
1221
1222                 mlx5_cqwq_pop(&cq->wq);
1223
1224                 wqe_counter = be16_to_cpu(cqe->wqe_counter);
1225
1226                 do {
1227                         struct mlx5e_dma_info *di;
1228                         u16 ci;
1229
1230                         last_wqe = (sqcc == wqe_counter);
1231
1232                         ci = sqcc & sq->wq.sz_m1;
1233                         di = &sq->db.di[ci];
1234
1235                         sqcc++;
1236                         /* Recycle RX page */
1237                         mlx5e_page_release(rq, di, true);
1238                 } while (!last_wqe);
1239         } while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1240
1241         mlx5_cqwq_update_db_record(&cq->wq);
1242
1243         /* ensure cq space is freed before enabling more cqes */
1244         wmb();
1245
1246         sq->cc = sqcc;
1247         return (i == MLX5E_TX_CQ_POLL_BUDGET);
1248 }
1249
1250 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
1251 {
1252         struct mlx5e_rq *rq = container_of(sq, struct mlx5e_rq, xdpsq);
1253         struct mlx5e_dma_info *di;
1254         u16 ci;
1255
1256         while (sq->cc != sq->pc) {
1257                 ci = sq->cc & sq->wq.sz_m1;
1258                 di = &sq->db.di[ci];
1259                 sq->cc++;
1260
1261                 mlx5e_page_release(rq, di, false);
1262         }
1263 }
1264
1265 #ifdef CONFIG_MLX5_CORE_IPOIB
1266
1267 #define MLX5_IB_GRH_DGID_OFFSET 24
1268 #define MLX5_GID_SIZE           16
1269
1270 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1271                                          struct mlx5_cqe64 *cqe,
1272                                          u32 cqe_bcnt,
1273                                          struct sk_buff *skb)
1274 {
1275         struct hwtstamp_config *tstamp;
1276         struct net_device *netdev;
1277         struct mlx5e_priv *priv;
1278         char *pseudo_header;
1279         u32 qpn;
1280         u8 *dgid;
1281         u8 g;
1282
1283         qpn = be32_to_cpu(cqe->sop_drop_qpn) & 0xffffff;
1284         netdev = mlx5i_pkey_get_netdev(rq->netdev, qpn);
1285
1286         /* No mapping present, cannot process SKB. This might happen if a child
1287          * interface is going down while having unprocessed CQEs on parent RQ
1288          */
1289         if (unlikely(!netdev)) {
1290                 /* TODO: add drop counters support */
1291                 skb->dev = NULL;
1292                 pr_warn_once("Unable to map QPN %u to dev - dropping skb\n", qpn);
1293                 return;
1294         }
1295
1296         priv = mlx5i_epriv(netdev);
1297         tstamp = &priv->tstamp;
1298
1299         g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
1300         dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1301         if ((!g) || dgid[0] != 0xff)
1302                 skb->pkt_type = PACKET_HOST;
1303         else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1304                 skb->pkt_type = PACKET_BROADCAST;
1305         else
1306                 skb->pkt_type = PACKET_MULTICAST;
1307
1308         /* TODO: IB/ipoib: Allow mcast packets from other VFs
1309          * 68996a6e760e5c74654723eeb57bf65628ae87f4
1310          */
1311
1312         skb_pull(skb, MLX5_IB_GRH_BYTES);
1313
1314         skb->protocol = *((__be16 *)(skb->data));
1315
1316         skb->ip_summed = CHECKSUM_COMPLETE;
1317         skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1318
1319         if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1320                 skb_hwtstamps(skb)->hwtstamp =
1321                                 mlx5_timecounter_cyc2time(rq->clock, get_cqe_ts(cqe));
1322
1323         skb_record_rx_queue(skb, rq->ix);
1324
1325         if (likely(netdev->features & NETIF_F_RXHASH))
1326                 mlx5e_skb_set_hash(cqe, skb);
1327
1328         /* 20 bytes of ipoib header and 4 for encap existing */
1329         pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1330         memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1331         skb_reset_mac_header(skb);
1332         skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1333
1334         skb->dev = netdev;
1335
1336         rq->stats.csum_complete++;
1337         rq->stats.packets++;
1338         rq->stats.bytes += cqe_bcnt;
1339 }
1340
1341 void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1342 {
1343         struct mlx5e_wqe_frag_info *wi;
1344         struct mlx5e_rx_wqe *wqe;
1345         __be16 wqe_counter_be;
1346         struct sk_buff *skb;
1347         u16 wqe_counter;
1348         u32 cqe_bcnt;
1349
1350         wqe_counter_be = cqe->wqe_counter;
1351         wqe_counter    = be16_to_cpu(wqe_counter_be);
1352         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1353         wi             = &rq->wqe.frag_info[wqe_counter];
1354         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1355
1356         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1357         if (!skb)
1358                 goto wq_free_wqe;
1359
1360         mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1361         if (unlikely(!skb->dev)) {
1362                 dev_kfree_skb_any(skb);
1363                 goto wq_free_wqe;
1364         }
1365         napi_gro_receive(rq->cq.napi, skb);
1366
1367 wq_free_wqe:
1368         mlx5e_free_rx_wqe_reuse(rq, wi);
1369         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1370                        &wqe->next.next_wqe_index);
1371 }
1372
1373 #endif /* CONFIG_MLX5_CORE_IPOIB */
1374
1375 #ifdef CONFIG_MLX5_EN_IPSEC
1376
1377 void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1378 {
1379         struct mlx5e_wqe_frag_info *wi;
1380         struct mlx5e_rx_wqe *wqe;
1381         __be16 wqe_counter_be;
1382         struct sk_buff *skb;
1383         u16 wqe_counter;
1384         u32 cqe_bcnt;
1385
1386         wqe_counter_be = cqe->wqe_counter;
1387         wqe_counter    = be16_to_cpu(wqe_counter_be);
1388         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1389         wi             = &rq->wqe.frag_info[wqe_counter];
1390         cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1391
1392         skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1393         if (unlikely(!skb)) {
1394                 /* a DROP, save the page-reuse checks */
1395                 mlx5e_free_rx_wqe(rq, wi);
1396                 goto wq_ll_pop;
1397         }
1398         skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb);
1399         if (unlikely(!skb)) {
1400                 mlx5e_free_rx_wqe(rq, wi);
1401                 goto wq_ll_pop;
1402         }
1403
1404         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1405         napi_gro_receive(rq->cq.napi, skb);
1406
1407         mlx5e_free_rx_wqe_reuse(rq, wi);
1408 wq_ll_pop:
1409         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1410                        &wqe->next.next_wqe_index);
1411 }
1412
1413 #endif /* CONFIG_MLX5_EN_IPSEC */