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

   1 /*
   2  * Copyright (c) 2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/ip.h>
  34 #include <linux/ipv6.h>
  35 #include <linux/tcp.h>
  36 #include <net/busy_poll.h>
  37 #include "en.h"
  38 #include "en_tc.h"
  39
  40 static inline bool mlx5e_rx_hw_stamp(struct mlx5e_tstamp *tstamp)
  41 {
  42         return tstamp->hwtstamp_config.rx_filter == HWTSTAMP_FILTER_ALL;
  43 }
  44
  45 static inline void mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cqcc,
  46                                        void *data)
  47 {
  48         u32 ci = cqcc & cq->wq.sz_m1;
  49
  50         memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, ci), sizeof(struct mlx5_cqe64));
  51 }
  52
  53 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
  54                                          struct mlx5e_cq *cq, u32 cqcc)
  55 {
  56         mlx5e_read_cqe_slot(cq, cqcc, &cq->title);
  57         cq->decmprs_left        = be32_to_cpu(cq->title.byte_cnt);
  58         cq->decmprs_wqe_counter = be16_to_cpu(cq->title.wqe_counter);
  59         rq->stats.cqe_compress_blks++;
  60 }
  61
  62 static inline void mlx5e_read_mini_arr_slot(struct mlx5e_cq *cq, u32 cqcc)
  63 {
  64         mlx5e_read_cqe_slot(cq, cqcc, cq->mini_arr);
  65         cq->mini_arr_idx = 0;
  66 }
  67
  68 static inline void mlx5e_cqes_update_owner(struct mlx5e_cq *cq, u32 cqcc, int n)
  69 {
  70         u8 op_own = (cqcc >> cq->wq.log_sz) & 1;
  71         u32 wq_sz = 1 << cq->wq.log_sz;
  72         u32 ci = cqcc & cq->wq.sz_m1;
  73         u32 ci_top = min_t(u32, wq_sz, ci + n);
  74
  75         for (; ci < ci_top; ci++, n--) {
  76                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
  77
  78                 cqe->op_own = op_own;
  79         }
  80
  81         if (unlikely(ci == wq_sz)) {
  82                 op_own = !op_own;
  83                 for (ci = 0; ci < n; ci++) {
  84                         struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
  85
  86                         cqe->op_own = op_own;
  87                 }
  88         }
  89 }
  90
  91 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
  92                                         struct mlx5e_cq *cq, u32 cqcc)
  93 {
  94         u16 wqe_cnt_step;
  95
  96         cq->title.byte_cnt     = cq->mini_arr[cq->mini_arr_idx].byte_cnt;
  97         cq->title.check_sum    = cq->mini_arr[cq->mini_arr_idx].checksum;
  98         cq->title.op_own      &= 0xf0;
  99         cq->title.op_own      |= 0x01 & (cqcc >> cq->wq.log_sz);
 100         cq->title.wqe_counter  = cpu_to_be16(cq->decmprs_wqe_counter);
 101
 102         wqe_cnt_step =
 103                 rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ ?
 104                 mpwrq_get_cqe_consumed_strides(&cq->title) : 1;
 105         cq->decmprs_wqe_counter =
 106                 (cq->decmprs_wqe_counter + wqe_cnt_step) & rq->wq.sz_m1;
 107 }
 108
 109 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
 110                                                 struct mlx5e_cq *cq, u32 cqcc)
 111 {
 112         mlx5e_decompress_cqe(rq, cq, cqcc);
 113         cq->title.rss_hash_type   = 0;
 114         cq->title.rss_hash_result = 0;
 115 }
 116
 117 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
 118                                              struct mlx5e_cq *cq,
 119                                              int update_owner_only,
 120                                              int budget_rem)
 121 {
 122         u32 cqcc = cq->wq.cc + update_owner_only;
 123         u32 cqe_count;
 124         u32 i;
 125
 126         cqe_count = min_t(u32, cq->decmprs_left, budget_rem);
 127
 128         for (i = update_owner_only; i < cqe_count;
 129              i++, cq->mini_arr_idx++, cqcc++) {
 130                 if (cq->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
 131                         mlx5e_read_mini_arr_slot(cq, cqcc);
 132
 133                 mlx5e_decompress_cqe_no_hash(rq, cq, cqcc);
 134                 rq->handle_rx_cqe(rq, &cq->title);
 135         }
 136         mlx5e_cqes_update_owner(cq, cq->wq.cc, cqcc - cq->wq.cc);
 137         cq->wq.cc = cqcc;
 138         cq->decmprs_left -= cqe_count;
 139         rq->stats.cqe_compress_pkts += cqe_count;
 140
 141         return cqe_count;
 142 }
 143
 144 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
 145                                               struct mlx5e_cq *cq,
 146                                               int budget_rem)
 147 {
 148         mlx5e_read_title_slot(rq, cq, cq->wq.cc);
 149         mlx5e_read_mini_arr_slot(cq, cq->wq.cc + 1);
 150         mlx5e_decompress_cqe(rq, cq, cq->wq.cc);
 151         rq->handle_rx_cqe(rq, &cq->title);
 152         cq->mini_arr_idx++;
 153
 154         return mlx5e_decompress_cqes_cont(rq, cq, 1, budget_rem) - 1;
 155 }
 156
 157 void mlx5e_modify_rx_cqe_compression(struct mlx5e_priv *priv, bool val)
 158 {
 159         bool was_opened;
 160
 161         if (!MLX5_CAP_GEN(priv->mdev, cqe_compression))
 162                 return;
 163
 164         mutex_lock(&priv->state_lock);
 165
 166         if (priv->params.rx_cqe_compress == val)
 167                 goto unlock;
 168
 169         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
 170         if (was_opened)
 171                 mlx5e_close_locked(priv->netdev);
 172
 173         priv->params.rx_cqe_compress = val;
 174
 175         if (was_opened)
 176                 mlx5e_open_locked(priv->netdev);
 177
 178 unlock:
 179         mutex_unlock(&priv->state_lock);
 180 }
 181
 182 #define RQ_PAGE_SIZE(rq) ((1 << rq->buff.page_order) << PAGE_SHIFT)
 183
 184 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
 185                                       struct mlx5e_dma_info *dma_info)
 186 {
 187         struct mlx5e_page_cache *cache = &rq->page_cache;
 188         u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
 189
 190         if (tail_next == cache->head) {
 191                 rq->stats.cache_full++;
 192                 return false;
 193         }
 194
 195         cache->page_cache[cache->tail] = *dma_info;
 196         cache->tail = tail_next;
 197         return true;
 198 }
 199
 200 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
 201                                       struct mlx5e_dma_info *dma_info)
 202 {
 203         struct mlx5e_page_cache *cache = &rq->page_cache;
 204
 205         if (unlikely(cache->head == cache->tail)) {
 206                 rq->stats.cache_empty++;
 207                 return false;
 208         }
 209
 210         if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
 211                 rq->stats.cache_busy++;
 212                 return false;
 213         }
 214
 215         *dma_info = cache->page_cache[cache->head];
 216         cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
 217         rq->stats.cache_reuse++;
 218
 219         dma_sync_single_for_device(rq->pdev, dma_info->addr,
 220                                    RQ_PAGE_SIZE(rq),
 221                                    DMA_FROM_DEVICE);
 222         return true;
 223 }
 224
 225 static inline int mlx5e_page_alloc_mapped(struct mlx5e_rq *rq,
 226                                           struct mlx5e_dma_info *dma_info)
 227 {
 228         struct page *page;
 229
 230         if (mlx5e_rx_cache_get(rq, dma_info))
 231                 return 0;
 232
 233         page = dev_alloc_pages(rq->buff.page_order);
 234         if (unlikely(!page))
 235                 return -ENOMEM;
 236
 237         dma_info->page = page;
 238         dma_info->addr = dma_map_page(rq->pdev, page, 0,
 239                                       RQ_PAGE_SIZE(rq), rq->buff.map_dir);
 240         if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
 241                 put_page(page);
 242                 return -ENOMEM;
 243         }
 244
 245         return 0;
 246 }
 247
 248 void mlx5e_page_release(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info,
 249                         bool recycle)
 250 {
 251         if (likely(recycle) && mlx5e_rx_cache_put(rq, dma_info))
 252                 return;
 253
 254         dma_unmap_page(rq->pdev, dma_info->addr, RQ_PAGE_SIZE(rq),
 255                        rq->buff.map_dir);
 256         put_page(dma_info->page);
 257 }
 258
 259 int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
 260 {
 261         struct mlx5e_dma_info *di = &rq->dma_info[ix];
 262
 263         if (unlikely(mlx5e_page_alloc_mapped(rq, di)))
 264                 return -ENOMEM;
 265
 266         wqe->data.addr = cpu_to_be64(di->addr + MLX5_RX_HEADROOM);
 267         return 0;
 268 }
 269
 270 void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
 271 {
 272         struct mlx5e_dma_info *di = &rq->dma_info[ix];
 273
 274         mlx5e_page_release(rq, di, true);
 275 }
 276
 277 static inline int mlx5e_mpwqe_strides_per_page(struct mlx5e_rq *rq)
 278 {
 279         return rq->mpwqe_num_strides >> MLX5_MPWRQ_WQE_PAGE_ORDER;
 280 }
 281
 282 static inline void mlx5e_add_skb_frag_mpwqe(struct mlx5e_rq *rq,
 283                                             struct sk_buff *skb,
 284                                             struct mlx5e_mpw_info *wi,
 285                                             u32 page_idx, u32 frag_offset,
 286                                             u32 len)
 287 {
 288         unsigned int truesize = ALIGN(len, rq->mpwqe_stride_sz);
 289
 290         dma_sync_single_for_cpu(rq->pdev,
 291                                 wi->umr.dma_info[page_idx].addr + frag_offset,
 292                                 len, DMA_FROM_DEVICE);
 293         wi->skbs_frags[page_idx]++;
 294         skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
 295                         wi->umr.dma_info[page_idx].page, frag_offset,
 296                         len, truesize);
 297 }
 298
 299 static inline void
 300 mlx5e_copy_skb_header_mpwqe(struct device *pdev,
 301                             struct sk_buff *skb,
 302                             struct mlx5e_mpw_info *wi,
 303                             u32 page_idx, u32 offset,
 304                             u32 headlen)
 305 {
 306         u16 headlen_pg = min_t(u32, headlen, PAGE_SIZE - offset);
 307         struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[page_idx];
 308         unsigned int len;
 309
 310          /* Aligning len to sizeof(long) optimizes memcpy performance */
 311         len = ALIGN(headlen_pg, sizeof(long));
 312         dma_sync_single_for_cpu(pdev, dma_info->addr + offset, len,
 313                                 DMA_FROM_DEVICE);
 314         skb_copy_to_linear_data_offset(skb, 0,
 315                                        page_address(dma_info->page) + offset,
 316                                        len);
 317         if (unlikely(offset + headlen > PAGE_SIZE)) {
 318                 dma_info++;
 319                 headlen_pg = len;
 320                 len = ALIGN(headlen - headlen_pg, sizeof(long));
 321                 dma_sync_single_for_cpu(pdev, dma_info->addr, len,
 322                                         DMA_FROM_DEVICE);
 323                 skb_copy_to_linear_data_offset(skb, headlen_pg,
 324                                                page_address(dma_info->page),
 325                                                len);
 326         }
 327 }
 328
 329 static inline void mlx5e_post_umr_wqe(struct mlx5e_rq *rq, u16 ix)
 330 {
 331         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 332         struct mlx5e_sq *sq = &rq->channel->icosq;
 333         struct mlx5_wq_cyc *wq = &sq->wq;
 334         struct mlx5e_umr_wqe *wqe;
 335         u8 num_wqebbs = DIV_ROUND_UP(sizeof(*wqe), MLX5_SEND_WQE_BB);
 336         u16 pi;
 337
 338         /* fill sq edge with nops to avoid wqe wrap around */
 339         while ((pi = (sq->pc & wq->sz_m1)) > sq->edge) {
 340                 sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_NOP;
 341                 sq->db.ico_wqe[pi].num_wqebbs = 1;
 342                 mlx5e_send_nop(sq, true);
 343         }
 344
 345         wqe = mlx5_wq_cyc_get_wqe(wq, pi);
 346         memcpy(wqe, &wi->umr.wqe, sizeof(*wqe));
 347         wqe->ctrl.opmod_idx_opcode =
 348                 cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
 349                             MLX5_OPCODE_UMR);
 350
 351         sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_UMR;
 352         sq->db.ico_wqe[pi].num_wqebbs = num_wqebbs;
 353         sq->pc += num_wqebbs;
 354         mlx5e_tx_notify_hw(sq, &wqe->ctrl, 0);
 355 }
 356
 357 static int mlx5e_alloc_rx_umr_mpwqe(struct mlx5e_rq *rq,
 358                                     struct mlx5e_rx_wqe *wqe,
 359                                     u16 ix)
 360 {
 361         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 362         u64 dma_offset = (u64)mlx5e_get_wqe_mtt_offset(rq, ix) << PAGE_SHIFT;
 363         int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
 364         int err;
 365         int i;
 366
 367         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++) {
 368                 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[i];
 369
 370                 err = mlx5e_page_alloc_mapped(rq, dma_info);
 371                 if (unlikely(err))
 372                         goto err_unmap;
 373                 wi->umr.mtt[i] = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
 374                 page_ref_add(dma_info->page, pg_strides);
 375                 wi->skbs_frags[i] = 0;
 376         }
 377
 378         wi->consumed_strides = 0;
 379         wqe->data.addr = cpu_to_be64(dma_offset);
 380
 381         return 0;
 382
 383 err_unmap:
 384         while (--i >= 0) {
 385                 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[i];
 386
 387                 page_ref_sub(dma_info->page, pg_strides);
 388                 mlx5e_page_release(rq, dma_info, true);
 389         }
 390
 391         return err;
 392 }
 393
 394 void mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
 395 {
 396         int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
 397         int i;
 398
 399         for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++) {
 400                 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[i];
 401
 402                 page_ref_sub(dma_info->page, pg_strides - wi->skbs_frags[i]);
 403                 mlx5e_page_release(rq, dma_info, true);
 404         }
 405 }
 406
 407 void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
 408 {
 409         struct mlx5_wq_ll *wq = &rq->wq;
 410         struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
 411
 412         clear_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state);
 413
 414         if (unlikely(test_bit(MLX5E_RQ_STATE_FLUSH, &rq->state))) {
 415                 mlx5e_free_rx_mpwqe(rq, &rq->mpwqe.info[wq->head]);
 416                 return;
 417         }
 418
 419         mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
 420
 421         /* ensure wqes are visible to device before updating doorbell record */
 422         dma_wmb();
 423
 424         mlx5_wq_ll_update_db_record(wq);
 425 }
 426
 427 int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
 428 {
 429         int err;
 430
 431         err = mlx5e_alloc_rx_umr_mpwqe(rq, wqe, ix);
 432         if (unlikely(err))
 433                 return err;
 434         set_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state);
 435         mlx5e_post_umr_wqe(rq, ix);
 436         return -EBUSY;
 437 }
 438
 439 void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
 440 {
 441         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 442
 443         mlx5e_free_rx_mpwqe(rq, wi);
 444 }
 445
 446 #define RQ_CANNOT_POST(rq) \
 447         (test_bit(MLX5E_RQ_STATE_FLUSH, &rq->state) || \
 448          test_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state))
 449
 450 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
 451 {
 452         struct mlx5_wq_ll *wq = &rq->wq;
 453
 454         if (unlikely(RQ_CANNOT_POST(rq)))
 455                 return false;
 456
 457         while (!mlx5_wq_ll_is_full(wq)) {
 458                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
 459                 int err;
 460
 461                 err = rq->alloc_wqe(rq, wqe, wq->head);
 462                 if (err == -EBUSY)
 463                         return true;
 464                 if (unlikely(err)) {
 465                         rq->stats.buff_alloc_err++;
 466                         break;
 467                 }
 468
 469                 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
 470         }
 471
 472         /* ensure wqes are visible to device before updating doorbell record */
 473         dma_wmb();
 474
 475         mlx5_wq_ll_update_db_record(wq);
 476
 477         return !mlx5_wq_ll_is_full(wq);
 478 }
 479
 480 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
 481                                  u32 cqe_bcnt)
 482 {
 483         struct ethhdr   *eth = (struct ethhdr *)(skb->data);
 484         struct iphdr    *ipv4;
 485         struct ipv6hdr  *ipv6;
 486         struct tcphdr   *tcp;
 487         int network_depth = 0;
 488         __be16 proto;
 489         u16 tot_len;
 490
 491         u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
 492         int tcp_ack = ((CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA  == l4_hdr_type) ||
 493                        (CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA == l4_hdr_type));
 494
 495         skb->mac_len = ETH_HLEN;
 496         proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
 497
 498         ipv4 = (struct iphdr *)(skb->data + network_depth);
 499         ipv6 = (struct ipv6hdr *)(skb->data + network_depth);
 500         tot_len = cqe_bcnt - network_depth;
 501
 502         if (proto == htons(ETH_P_IP)) {
 503                 tcp = (struct tcphdr *)(skb->data + network_depth +
 504                                         sizeof(struct iphdr));
 505                 ipv6 = NULL;
 506                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 507         } else {
 508                 tcp = (struct tcphdr *)(skb->data + network_depth +
 509                                         sizeof(struct ipv6hdr));
 510                 ipv4 = NULL;
 511                 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 512         }
 513
 514         if (get_cqe_lro_tcppsh(cqe))
 515                 tcp->psh                = 1;
 516
 517         if (tcp_ack) {
 518                 tcp->ack                = 1;
 519                 tcp->ack_seq            = cqe->lro_ack_seq_num;
 520                 tcp->window             = cqe->lro_tcp_win;
 521         }
 522
 523         if (ipv4) {
 524                 ipv4->ttl               = cqe->lro_min_ttl;
 525                 ipv4->tot_len           = cpu_to_be16(tot_len);
 526                 ipv4->check             = 0;
 527                 ipv4->check             = ip_fast_csum((unsigned char *)ipv4,
 528                                                        ipv4->ihl);
 529         } else {
 530                 ipv6->hop_limit         = cqe->lro_min_ttl;
 531                 ipv6->payload_len       = cpu_to_be16(tot_len -
 532                                                       sizeof(struct ipv6hdr));
 533         }
 534 }
 535
 536 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
 537                                       struct sk_buff *skb)
 538 {
 539         u8 cht = cqe->rss_hash_type;
 540         int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
 541                  (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
 542                                             PKT_HASH_TYPE_NONE;
 543         skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
 544 }
 545
 546 static inline bool is_first_ethertype_ip(struct sk_buff *skb)
 547 {
 548         __be16 ethertype = ((struct ethhdr *)skb->data)->h_proto;
 549
 550         return (ethertype == htons(ETH_P_IP) || ethertype == htons(ETH_P_IPV6));
 551 }
 552
 553 static inline void mlx5e_handle_csum(struct net_device *netdev,
 554                                      struct mlx5_cqe64 *cqe,
 555                                      struct mlx5e_rq *rq,
 556                                      struct sk_buff *skb,
 557                                      bool   lro)
 558 {
 559         if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
 560                 goto csum_none;
 561
 562         if (lro) {
 563                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 564                 return;
 565         }
 566
 567         if (is_first_ethertype_ip(skb)) {
 568                 skb->ip_summed = CHECKSUM_COMPLETE;
 569                 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
 570                 rq->stats.csum_complete++;
 571                 return;
 572         }
 573
 574         if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
 575                    (cqe->hds_ip_ext & CQE_L4_OK))) {
 576                 skb->ip_summed = CHECKSUM_UNNECESSARY;
 577                 if (cqe_is_tunneled(cqe)) {
 578                         skb->csum_level = 1;
 579                         skb->encapsulation = 1;
 580                         rq->stats.csum_unnecessary_inner++;
 581                 }
 582                 return;
 583         }
 584 csum_none:
 585         skb->ip_summed = CHECKSUM_NONE;
 586         rq->stats.csum_none++;
 587 }
 588
 589 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
 590                                       u32 cqe_bcnt,
 591                                       struct mlx5e_rq *rq,
 592                                       struct sk_buff *skb)
 593 {
 594         struct net_device *netdev = rq->netdev;
 595         struct mlx5e_tstamp *tstamp = rq->tstamp;
 596         int lro_num_seg;
 597
 598         lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
 599         if (lro_num_seg > 1) {
 600                 mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
 601                 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
 602                 rq->stats.lro_packets++;
 603                 rq->stats.lro_bytes += cqe_bcnt;
 604         }
 605
 606         if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
 607                 mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
 608
 609         skb_record_rx_queue(skb, rq->ix);
 610
 611         if (likely(netdev->features & NETIF_F_RXHASH))
 612                 mlx5e_skb_set_hash(cqe, skb);
 613
 614         if (cqe_has_vlan(cqe))
 615                 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 616                                        be16_to_cpu(cqe->vlan_info));
 617
 618         skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
 619
 620         mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
 621         skb->protocol = eth_type_trans(skb, netdev);
 622 }
 623
 624 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
 625                                          struct mlx5_cqe64 *cqe,
 626                                          u32 cqe_bcnt,
 627                                          struct sk_buff *skb)
 628 {
 629         rq->stats.packets++;
 630         rq->stats.bytes += cqe_bcnt;
 631         mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
 632         napi_gro_receive(rq->cq.napi, skb);
 633 }
 634
 635 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_sq *sq)
 636 {
 637         struct mlx5_wq_cyc *wq = &sq->wq;
 638         struct mlx5e_tx_wqe *wqe;
 639         u16 pi = (sq->pc - MLX5E_XDP_TX_WQEBBS) & wq->sz_m1; /* last pi */
 640
 641         wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
 642
 643         wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
 644         mlx5e_tx_notify_hw(sq, &wqe->ctrl, 0);
 645 }
 646
 647 static inline void mlx5e_xmit_xdp_frame(struct mlx5e_rq *rq,
 648                                         struct mlx5e_dma_info *di,
 649                                         unsigned int data_offset,
 650                                         int len)
 651 {
 652         struct mlx5e_sq          *sq   = &rq->channel->xdp_sq;
 653         struct mlx5_wq_cyc       *wq   = &sq->wq;
 654         u16                      pi    = sq->pc & wq->sz_m1;
 655         struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
 656         struct mlx5e_sq_wqe_info *wi   = &sq->db.xdp.wqe_info[pi];
 657
 658         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
 659         struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
 660         struct mlx5_wqe_data_seg *dseg;
 661
 662         dma_addr_t dma_addr  = di->addr + data_offset + MLX5E_XDP_MIN_INLINE;
 663         unsigned int dma_len = len - MLX5E_XDP_MIN_INLINE;
 664         void *data           = page_address(di->page) + data_offset;
 665
 666         if (unlikely(!mlx5e_sq_has_room_for(sq, MLX5E_XDP_TX_WQEBBS))) {
 667                 if (sq->db.xdp.doorbell) {
 668                         /* SQ is full, ring doorbell */
 669                         mlx5e_xmit_xdp_doorbell(sq);
 670                         sq->db.xdp.doorbell = false;
 671                 }
 672                 rq->stats.xdp_tx_full++;
 673                 mlx5e_page_release(rq, di, true);
 674                 return;
 675         }
 676
 677         dma_sync_single_for_device(sq->pdev, dma_addr, dma_len,
 678                                    PCI_DMA_TODEVICE);
 679
 680         memset(wqe, 0, sizeof(*wqe));
 681
 682         /* copy the inline part */
 683         memcpy(eseg->inline_hdr_start, data, MLX5E_XDP_MIN_INLINE);
 684         eseg->inline_hdr_sz = cpu_to_be16(MLX5E_XDP_MIN_INLINE);
 685
 686         dseg = (struct mlx5_wqe_data_seg *)cseg + (MLX5E_XDP_TX_DS_COUNT - 1);
 687
 688         /* write the dma part */
 689         dseg->addr       = cpu_to_be64(dma_addr);
 690         dseg->byte_count = cpu_to_be32(dma_len);
 691         dseg->lkey       = sq->mkey_be;
 692
 693         cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_SEND);
 694         cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | MLX5E_XDP_TX_DS_COUNT);
 695
 696         sq->db.xdp.di[pi] = *di;
 697         wi->opcode     = MLX5_OPCODE_SEND;
 698         wi->num_wqebbs = MLX5E_XDP_TX_WQEBBS;
 699         sq->pc += MLX5E_XDP_TX_WQEBBS;
 700
 701         sq->db.xdp.doorbell = true;
 702         rq->stats.xdp_tx++;
 703 }
 704
 705 /* returns true if packet was consumed by xdp */
 706 static inline bool mlx5e_xdp_handle(struct mlx5e_rq *rq,
 707                                     const struct bpf_prog *prog,
 708                                     struct mlx5e_dma_info *di,
 709                                     void *data, u16 len)
 710 {
 711         struct xdp_buff xdp;
 712         u32 act;
 713
 714         if (!prog)
 715                 return false;
 716
 717         xdp.data = data;
 718         xdp.data_end = xdp.data + len;
 719         act = bpf_prog_run_xdp(prog, &xdp);
 720         switch (act) {
 721         case XDP_PASS:
 722                 return false;
 723         case XDP_TX:
 724                 mlx5e_xmit_xdp_frame(rq, di, MLX5_RX_HEADROOM, len);
 725                 return true;
 726         default:
 727                 bpf_warn_invalid_xdp_action(act);
 728         case XDP_ABORTED:
 729         case XDP_DROP:
 730                 rq->stats.xdp_drop++;
 731                 mlx5e_page_release(rq, di, true);
 732                 return true;
 733         }
 734 }
 735
 736 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 737 {
 738         struct bpf_prog *xdp_prog = READ_ONCE(rq->xdp_prog);
 739         struct mlx5e_dma_info *di;
 740         struct mlx5e_rx_wqe *wqe;
 741         __be16 wqe_counter_be;
 742         struct sk_buff *skb;
 743         u16 wqe_counter;
 744         void *va, *data;
 745         u32 cqe_bcnt;
 746
 747         wqe_counter_be = cqe->wqe_counter;
 748         wqe_counter    = be16_to_cpu(wqe_counter_be);
 749         wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
 750         di             = &rq->dma_info[wqe_counter];
 751         va             = page_address(di->page);
 752         data           = va + MLX5_RX_HEADROOM;
 753
 754         dma_sync_single_range_for_cpu(rq->pdev,
 755                                       di->addr,
 756                                       MLX5_RX_HEADROOM,
 757                                       rq->buff.wqe_sz,
 758                                       DMA_FROM_DEVICE);
 759         prefetch(data);
 760         cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
 761
 762         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
 763                 rq->stats.wqe_err++;
 764                 mlx5e_page_release(rq, di, true);
 765                 goto wq_ll_pop;
 766         }
 767
 768         if (mlx5e_xdp_handle(rq, xdp_prog, di, data, cqe_bcnt))
 769                 goto wq_ll_pop; /* page/packet was consumed by XDP */
 770
 771         skb = build_skb(va, RQ_PAGE_SIZE(rq));
 772         if (unlikely(!skb)) {
 773                 rq->stats.buff_alloc_err++;
 774                 mlx5e_page_release(rq, di, true);
 775                 goto wq_ll_pop;
 776         }
 777
 778         /* queue up for recycling ..*/
 779         page_ref_inc(di->page);
 780         mlx5e_page_release(rq, di, true);
 781
 782         skb_reserve(skb, MLX5_RX_HEADROOM);
 783         skb_put(skb, cqe_bcnt);
 784
 785         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 786
 787 wq_ll_pop:
 788         mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
 789                        &wqe->next.next_wqe_index);
 790 }
 791
 792 static inline void mlx5e_mpwqe_fill_rx_skb(struct mlx5e_rq *rq,
 793                                            struct mlx5_cqe64 *cqe,
 794                                            struct mlx5e_mpw_info *wi,
 795                                            u32 cqe_bcnt,
 796                                            struct sk_buff *skb)
 797 {
 798         u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
 799         u32 wqe_offset     = stride_ix * rq->mpwqe_stride_sz;
 800         u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
 801         u32 page_idx       = wqe_offset >> PAGE_SHIFT;
 802         u32 head_page_idx  = page_idx;
 803         u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
 804         u32 frag_offset    = head_offset + headlen;
 805         u16 byte_cnt       = cqe_bcnt - headlen;
 806
 807         if (unlikely(frag_offset >= PAGE_SIZE)) {
 808                 page_idx++;
 809                 frag_offset -= PAGE_SIZE;
 810         }
 811
 812         while (byte_cnt) {
 813                 u32 pg_consumed_bytes =
 814                         min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
 815
 816                 mlx5e_add_skb_frag_mpwqe(rq, skb, wi, page_idx, frag_offset,
 817                                          pg_consumed_bytes);
 818                 byte_cnt -= pg_consumed_bytes;
 819                 frag_offset = 0;
 820                 page_idx++;
 821         }
 822         /* copy header */
 823         mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, wi, head_page_idx,
 824                                     head_offset, headlen);
 825         /* skb linear part was allocated with headlen and aligned to long */
 826         skb->tail += headlen;
 827         skb->len  += headlen;
 828 }
 829
 830 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
 831 {
 832         u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
 833         u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
 834         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
 835         struct mlx5e_rx_wqe  *wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_id);
 836         struct sk_buff *skb;
 837         u16 cqe_bcnt;
 838
 839         wi->consumed_strides += cstrides;
 840
 841         if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
 842                 rq->stats.wqe_err++;
 843                 goto mpwrq_cqe_out;
 844         }
 845
 846         if (unlikely(mpwrq_is_filler_cqe(cqe))) {
 847                 rq->stats.mpwqe_filler++;
 848                 goto mpwrq_cqe_out;
 849         }
 850
 851         skb = napi_alloc_skb(rq->cq.napi,
 852                              ALIGN(MLX5_MPWRQ_SMALL_PACKET_THRESHOLD,
 853                                    sizeof(long)));
 854         if (unlikely(!skb)) {
 855                 rq->stats.buff_alloc_err++;
 856                 goto mpwrq_cqe_out;
 857         }
 858
 859         prefetch(skb->data);
 860         cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
 861
 862         mlx5e_mpwqe_fill_rx_skb(rq, cqe, wi, cqe_bcnt, skb);
 863         mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
 864
 865 mpwrq_cqe_out:
 866         if (likely(wi->consumed_strides < rq->mpwqe_num_strides))
 867                 return;
 868
 869         mlx5e_free_rx_mpwqe(rq, wi);
 870         mlx5_wq_ll_pop(&rq->wq, cqe->wqe_id, &wqe->next.next_wqe_index);
 871 }
 872
 873 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
 874 {
 875         struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
 876         struct mlx5e_sq *xdp_sq = &rq->channel->xdp_sq;
 877         int work_done = 0;
 878
 879         if (unlikely(test_bit(MLX5E_RQ_STATE_FLUSH, &rq->state)))
 880                 return 0;
 881
 882         if (cq->decmprs_left)
 883                 work_done += mlx5e_decompress_cqes_cont(rq, cq, 0, budget);
 884
 885         for (; work_done < budget; work_done++) {
 886                 struct mlx5_cqe64 *cqe = mlx5e_get_cqe(cq);
 887
 888                 if (!cqe)
 889                         break;
 890
 891                 if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
 892                         work_done +=
 893                                 mlx5e_decompress_cqes_start(rq, cq,
 894                                                             budget - work_done);
 895                         continue;
 896                 }
 897
 898                 mlx5_cqwq_pop(&cq->wq);
 899
 900                 rq->handle_rx_cqe(rq, cqe);
 901         }
 902
 903         if (xdp_sq->db.xdp.doorbell) {
 904                 mlx5e_xmit_xdp_doorbell(xdp_sq);
 905                 xdp_sq->db.xdp.doorbell = false;
 906         }
 907
 908         mlx5_cqwq_update_db_record(&cq->wq);
 909
 910         /* ensure cq space is freed before enabling more cqes */
 911         wmb();
 912
 913         return work_done;
 914 }