drivers/net/ethernet/mellanox/mlx4/en_rx.c

   1 /*
   2  * Copyright (c) 2007 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
  34 #include <linux/bpf.h>
  35 #include <linux/bpf_trace.h>
  36 #include <linux/mlx4/cq.h>
  37 #include <linux/slab.h>
  38 #include <linux/mlx4/qp.h>
  39 #include <linux/skbuff.h>
  40 #include <linux/rculist.h>
  41 #include <linux/if_ether.h>
  42 #include <linux/if_vlan.h>
  43 #include <linux/vmalloc.h>
  44 #include <linux/irq.h>
  45
  46 #include <net/ip.h>
  47 #if IS_ENABLED(CONFIG_IPV6)
  48 #include <net/ip6_checksum.h>
  49 #endif
  50
  51 #include "mlx4_en.h"
  52
  53 static int mlx4_alloc_page(struct mlx4_en_priv *priv,
  54                            struct mlx4_en_rx_alloc *frag,
  55                            gfp_t gfp)
  56 {
  57         struct page *page;
  58         dma_addr_t dma;
  59
  60         page = alloc_page(gfp);
  61         if (unlikely(!page))
  62                 return -ENOMEM;
  63         dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE, priv->dma_dir);
  64         if (unlikely(dma_mapping_error(priv->ddev, dma))) {
  65                 __free_page(page);
  66                 return -ENOMEM;
  67         }
  68         frag->page = page;
  69         frag->dma = dma;
  70         frag->page_offset = priv->rx_headroom;
  71         return 0;
  72 }
  73
  74 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
  75                                struct mlx4_en_rx_ring *ring,
  76                                struct mlx4_en_rx_desc *rx_desc,
  77                                struct mlx4_en_rx_alloc *frags,
  78                                gfp_t gfp)
  79 {
  80         int i;
  81
  82         for (i = 0; i < priv->num_frags; i++, frags++) {
  83                 if (!frags->page) {
  84                         if (mlx4_alloc_page(priv, frags, gfp))
  85                                 return -ENOMEM;
  86                         ring->rx_alloc_pages++;
  87                 }
  88                 rx_desc->data[i].addr = cpu_to_be64(frags->dma +
  89                                                     frags->page_offset);
  90         }
  91         return 0;
  92 }
  93
  94 static void mlx4_en_free_frag(const struct mlx4_en_priv *priv,
  95                               struct mlx4_en_rx_alloc *frag)
  96 {
  97         if (frag->page) {
  98                 dma_unmap_page(priv->ddev, frag->dma,
  99                                PAGE_SIZE, priv->dma_dir);
 100                 __free_page(frag->page);
 101         }
 102         /* We need to clear all fields, otherwise a change of priv->log_rx_info
 103          * could lead to see garbage later in frag->page.
 104          */
 105         memset(frag, 0, sizeof(*frag));
 106 }
 107
 108 static void mlx4_en_init_rx_desc(const struct mlx4_en_priv *priv,
 109                                  struct mlx4_en_rx_ring *ring, int index)
 110 {
 111         struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
 112         int possible_frags;
 113         int i;
 114
 115         /* Set size and memtype fields */
 116         for (i = 0; i < priv->num_frags; i++) {
 117                 rx_desc->data[i].byte_count =
 118                         cpu_to_be32(priv->frag_info[i].frag_size);
 119                 rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
 120         }
 121
 122         /* If the number of used fragments does not fill up the ring stride,
 123          * remaining (unused) fragments must be padded with null address/size
 124          * and a special memory key */
 125         possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
 126         for (i = priv->num_frags; i < possible_frags; i++) {
 127                 rx_desc->data[i].byte_count = 0;
 128                 rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
 129                 rx_desc->data[i].addr = 0;
 130         }
 131 }
 132
 133 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
 134                                    struct mlx4_en_rx_ring *ring, int index,
 135                                    gfp_t gfp)
 136 {
 137         struct mlx4_en_rx_desc *rx_desc = ring->buf +
 138                 (index << ring->log_stride);
 139         struct mlx4_en_rx_alloc *frags = ring->rx_info +
 140                                         (index << priv->log_rx_info);
 141         if (likely(ring->page_cache.index > 0)) {
 142                 /* XDP uses a single page per frame */
 143                 if (!frags->page) {
 144                         ring->page_cache.index--;
 145                         frags->page = ring->page_cache.buf[ring->page_cache.index].page;
 146                         frags->dma  = ring->page_cache.buf[ring->page_cache.index].dma;
 147                 }
 148                 frags->page_offset = XDP_PACKET_HEADROOM;
 149                 rx_desc->data[0].addr = cpu_to_be64(frags->dma +
 150                                                     XDP_PACKET_HEADROOM);
 151                 return 0;
 152         }
 153
 154         return mlx4_en_alloc_frags(priv, ring, rx_desc, frags, gfp);
 155 }
 156
 157 static bool mlx4_en_is_ring_empty(const struct mlx4_en_rx_ring *ring)
 158 {
 159         return ring->prod == ring->cons;
 160 }
 161
 162 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
 163 {
 164         *ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
 165 }
 166
 167 /* slow path */
 168 static void mlx4_en_free_rx_desc(const struct mlx4_en_priv *priv,
 169                                  struct mlx4_en_rx_ring *ring,
 170                                  int index)
 171 {
 172         struct mlx4_en_rx_alloc *frags;
 173         int nr;
 174
 175         frags = ring->rx_info + (index << priv->log_rx_info);
 176         for (nr = 0; nr < priv->num_frags; nr++) {
 177                 en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
 178                 mlx4_en_free_frag(priv, frags + nr);
 179         }
 180 }
 181
 182 /* Function not in fast-path */
 183 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
 184 {
 185         struct mlx4_en_rx_ring *ring;
 186         int ring_ind;
 187         int buf_ind;
 188         int new_size;
 189
 190         for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
 191                 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 192                         ring = priv->rx_ring[ring_ind];
 193
 194                         if (mlx4_en_prepare_rx_desc(priv, ring,
 195                                                     ring->actual_size,
 196                                                     GFP_KERNEL)) {
 197                                 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
 198                                         en_err(priv, "Failed to allocate enough rx buffers\n");
 199                                         return -ENOMEM;
 200                                 } else {
 201                                         new_size = rounddown_pow_of_two(ring->actual_size);
 202                                         en_warn(priv, "Only %d buffers allocated reducing ring size to %d\n",
 203                                                 ring->actual_size, new_size);
 204                                         goto reduce_rings;
 205                                 }
 206                         }
 207                         ring->actual_size++;
 208                         ring->prod++;
 209                 }
 210         }
 211         return 0;
 212
 213 reduce_rings:
 214         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 215                 ring = priv->rx_ring[ring_ind];
 216                 while (ring->actual_size > new_size) {
 217                         ring->actual_size--;
 218                         ring->prod--;
 219                         mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
 220                 }
 221         }
 222
 223         return 0;
 224 }
 225
 226 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
 227                                 struct mlx4_en_rx_ring *ring)
 228 {
 229         int index;
 230
 231         en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
 232                ring->cons, ring->prod);
 233
 234         /* Unmap and free Rx buffers */
 235         for (index = 0; index < ring->size; index++) {
 236                 en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
 237                 mlx4_en_free_rx_desc(priv, ring, index);
 238         }
 239         ring->cons = 0;
 240         ring->prod = 0;
 241 }
 242
 243 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
 244 {
 245         int i;
 246         int num_of_eqs;
 247         int num_rx_rings;
 248         struct mlx4_dev *dev = mdev->dev;
 249
 250         mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
 251                 num_of_eqs = max_t(int, MIN_RX_RINGS,
 252                                    min_t(int,
 253                                          mlx4_get_eqs_per_port(mdev->dev, i),
 254                                          DEF_RX_RINGS));
 255
 256                 num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
 257                         min_t(int, num_of_eqs, num_online_cpus());
 258                 mdev->profile.prof[i].rx_ring_num =
 259                         rounddown_pow_of_two(num_rx_rings);
 260         }
 261 }
 262
 263 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
 264                            struct mlx4_en_rx_ring **pring,
 265                            u32 size, u16 stride, int node, int queue_index)
 266 {
 267         struct mlx4_en_dev *mdev = priv->mdev;
 268         struct mlx4_en_rx_ring *ring;
 269         int err = -ENOMEM;
 270         int tmp;
 271
 272         ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
 273         if (!ring) {
 274                 en_err(priv, "Failed to allocate RX ring structure\n");
 275                 return -ENOMEM;
 276         }
 277
 278         ring->prod = 0;
 279         ring->cons = 0;
 280         ring->size = size;
 281         ring->size_mask = size - 1;
 282         ring->stride = stride;
 283         ring->log_stride = ffs(ring->stride) - 1;
 284         ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
 285
 286         if (xdp_rxq_info_reg(&ring->xdp_rxq, priv->dev, queue_index, 0) < 0)
 287                 goto err_ring;
 288
 289         tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
 290                                         sizeof(struct mlx4_en_rx_alloc));
 291         ring->rx_info = kvzalloc_node(tmp, GFP_KERNEL, node);
 292         if (!ring->rx_info) {
 293                 err = -ENOMEM;
 294                 goto err_xdp_info;
 295         }
 296
 297         en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
 298                  ring->rx_info, tmp);
 299
 300         /* Allocate HW buffers on provided NUMA node */
 301         set_dev_node(&mdev->dev->persist->pdev->dev, node);
 302         err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
 303         set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
 304         if (err)
 305                 goto err_info;
 306
 307         ring->buf = ring->wqres.buf.direct.buf;
 308
 309         ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
 310
 311         *pring = ring;
 312         return 0;
 313
 314 err_info:
 315         kvfree(ring->rx_info);
 316         ring->rx_info = NULL;
 317 err_xdp_info:
 318         xdp_rxq_info_unreg(&ring->xdp_rxq);
 319 err_ring:
 320         kfree(ring);
 321         *pring = NULL;
 322
 323         return err;
 324 }
 325
 326 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
 327 {
 328         struct mlx4_en_rx_ring *ring;
 329         int i;
 330         int ring_ind;
 331         int err;
 332         int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
 333                                         DS_SIZE * priv->num_frags);
 334
 335         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 336                 ring = priv->rx_ring[ring_ind];
 337
 338                 ring->prod = 0;
 339                 ring->cons = 0;
 340                 ring->actual_size = 0;
 341                 ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
 342
 343                 ring->stride = stride;
 344                 if (ring->stride <= TXBB_SIZE) {
 345                         /* Stamp first unused send wqe */
 346                         __be32 *ptr = (__be32 *)ring->buf;
 347                         __be32 stamp = cpu_to_be32(1 << STAMP_SHIFT);
 348                         *ptr = stamp;
 349                         /* Move pointer to start of rx section */
 350                         ring->buf += TXBB_SIZE;
 351                 }
 352
 353                 ring->log_stride = ffs(ring->stride) - 1;
 354                 ring->buf_size = ring->size * ring->stride;
 355
 356                 memset(ring->buf, 0, ring->buf_size);
 357                 mlx4_en_update_rx_prod_db(ring);
 358
 359                 /* Initialize all descriptors */
 360                 for (i = 0; i < ring->size; i++)
 361                         mlx4_en_init_rx_desc(priv, ring, i);
 362         }
 363         err = mlx4_en_fill_rx_buffers(priv);
 364         if (err)
 365                 goto err_buffers;
 366
 367         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 368                 ring = priv->rx_ring[ring_ind];
 369
 370                 ring->size_mask = ring->actual_size - 1;
 371                 mlx4_en_update_rx_prod_db(ring);
 372         }
 373
 374         return 0;
 375
 376 err_buffers:
 377         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
 378                 mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
 379
 380         ring_ind = priv->rx_ring_num - 1;
 381         while (ring_ind >= 0) {
 382                 if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
 383                         priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
 384                 ring_ind--;
 385         }
 386         return err;
 387 }
 388
 389 /* We recover from out of memory by scheduling our napi poll
 390  * function (mlx4_en_process_cq), which tries to allocate
 391  * all missing RX buffers (call to mlx4_en_refill_rx_buffers).
 392  */
 393 void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
 394 {
 395         int ring;
 396
 397         if (!priv->port_up)
 398                 return;
 399
 400         for (ring = 0; ring < priv->rx_ring_num; ring++) {
 401                 if (mlx4_en_is_ring_empty(priv->rx_ring[ring])) {
 402                         local_bh_disable();
 403                         napi_schedule(&priv->rx_cq[ring]->napi);
 404                         local_bh_enable();
 405                 }
 406         }
 407 }
 408
 409 /* When the rx ring is running in page-per-packet mode, a released frame can go
 410  * directly into a small cache, to avoid unmapping or touching the page
 411  * allocator. In bpf prog performance scenarios, buffers are either forwarded
 412  * or dropped, never converted to skbs, so every page can come directly from
 413  * this cache when it is sized to be a multiple of the napi budget.
 414  */
 415 bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
 416                         struct mlx4_en_rx_alloc *frame)
 417 {
 418         struct mlx4_en_page_cache *cache = &ring->page_cache;
 419
 420         if (cache->index >= MLX4_EN_CACHE_SIZE)
 421                 return false;
 422
 423         cache->buf[cache->index].page = frame->page;
 424         cache->buf[cache->index].dma = frame->dma;
 425         cache->index++;
 426         return true;
 427 }
 428
 429 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
 430                              struct mlx4_en_rx_ring **pring,
 431                              u32 size, u16 stride)
 432 {
 433         struct mlx4_en_dev *mdev = priv->mdev;
 434         struct mlx4_en_rx_ring *ring = *pring;
 435         struct bpf_prog *old_prog;
 436
 437         old_prog = rcu_dereference_protected(
 438                                         ring->xdp_prog,
 439                                         lockdep_is_held(&mdev->state_lock));
 440         if (old_prog)
 441                 bpf_prog_put(old_prog);
 442         xdp_rxq_info_unreg(&ring->xdp_rxq);
 443         mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
 444         kvfree(ring->rx_info);
 445         ring->rx_info = NULL;
 446         kfree(ring);
 447         *pring = NULL;
 448 }
 449
 450 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
 451                                 struct mlx4_en_rx_ring *ring)
 452 {
 453         int i;
 454
 455         for (i = 0; i < ring->page_cache.index; i++) {
 456                 dma_unmap_page(priv->ddev, ring->page_cache.buf[i].dma,
 457                                PAGE_SIZE, priv->dma_dir);
 458                 put_page(ring->page_cache.buf[i].page);
 459         }
 460         ring->page_cache.index = 0;
 461         mlx4_en_free_rx_buf(priv, ring);
 462         if (ring->stride <= TXBB_SIZE)
 463                 ring->buf -= TXBB_SIZE;
 464 }
 465
 466
 467 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 468                                     struct mlx4_en_rx_alloc *frags,
 469                                     struct sk_buff *skb,
 470                                     int length)
 471 {
 472         const struct mlx4_en_frag_info *frag_info = priv->frag_info;
 473         unsigned int truesize = 0;
 474         bool release = true;
 475         int nr, frag_size;
 476         struct page *page;
 477         dma_addr_t dma;
 478
 479         /* Collect used fragments while replacing them in the HW descriptors */
 480         for (nr = 0;; frags++) {
 481                 frag_size = min_t(int, length, frag_info->frag_size);
 482
 483                 page = frags->page;
 484                 if (unlikely(!page))
 485                         goto fail;
 486
 487                 dma = frags->dma;
 488                 dma_sync_single_range_for_cpu(priv->ddev, dma, frags->page_offset,
 489                                               frag_size, priv->dma_dir);
 490
 491                 __skb_fill_page_desc(skb, nr, page, frags->page_offset,
 492                                      frag_size);
 493
 494                 truesize += frag_info->frag_stride;
 495                 if (frag_info->frag_stride == PAGE_SIZE / 2) {
 496                         frags->page_offset ^= PAGE_SIZE / 2;
 497                         release = page_count(page) != 1 ||
 498                                   page_is_pfmemalloc(page) ||
 499                                   page_to_nid(page) != numa_mem_id();
 500                 } else if (!priv->rx_headroom) {
 501                         /* rx_headroom for non XDP setup is always 0.
 502                          * When XDP is set, the above condition will
 503                          * guarantee page is always released.
 504                          */
 505                         u32 sz_align = ALIGN(frag_size, SMP_CACHE_BYTES);
 506
 507                         frags->page_offset += sz_align;
 508                         release = frags->page_offset + frag_info->frag_size > PAGE_SIZE;
 509                 }
 510                 if (release) {
 511                         dma_unmap_page(priv->ddev, dma, PAGE_SIZE, priv->dma_dir);
 512                         frags->page = NULL;
 513                 } else {
 514                         page_ref_inc(page);
 515                 }
 516
 517                 nr++;
 518                 length -= frag_size;
 519                 if (!length)
 520                         break;
 521                 frag_info++;
 522         }
 523         skb->truesize += truesize;
 524         return nr;
 525
 526 fail:
 527         while (nr > 0) {
 528                 nr--;
 529                 __skb_frag_unref(skb_shinfo(skb)->frags + nr, false);
 530         }
 531         return 0;
 532 }
 533
 534 static void validate_loopback(struct mlx4_en_priv *priv, void *va)
 535 {
 536         const unsigned char *data = va + ETH_HLEN;
 537         int i;
 538
 539         for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++) {
 540                 if (data[i] != (unsigned char)i)
 541                         return;
 542         }
 543         /* Loopback found */
 544         priv->loopback_ok = 1;
 545 }
 546
 547 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
 548                                       struct mlx4_en_rx_ring *ring)
 549 {
 550         u32 missing = ring->actual_size - (ring->prod - ring->cons);
 551
 552         /* Try to batch allocations, but not too much. */
 553         if (missing < 8)
 554                 return;
 555         do {
 556                 if (mlx4_en_prepare_rx_desc(priv, ring,
 557                                             ring->prod & ring->size_mask,
 558                                             GFP_ATOMIC | __GFP_MEMALLOC))
 559                         break;
 560                 ring->prod++;
 561         } while (likely(--missing));
 562
 563         mlx4_en_update_rx_prod_db(ring);
 564 }
 565
 566 /* When hardware doesn't strip the vlan, we need to calculate the checksum
 567  * over it and add it to the hardware's checksum calculation
 568  */
 569 static inline __wsum get_fixed_vlan_csum(__wsum hw_checksum,
 570                                          struct vlan_hdr *vlanh)
 571 {
 572         return csum_add(hw_checksum, *(__wsum *)vlanh);
 573 }
 574
 575 /* Although the stack expects checksum which doesn't include the pseudo
 576  * header, the HW adds it. To address that, we are subtracting the pseudo
 577  * header checksum from the checksum value provided by the HW.
 578  */
 579 static int get_fixed_ipv4_csum(__wsum hw_checksum, struct sk_buff *skb,
 580                                struct iphdr *iph)
 581 {
 582         __u16 length_for_csum = 0;
 583         __wsum csum_pseudo_header = 0;
 584         __u8 ipproto = iph->protocol;
 585
 586         if (unlikely(ipproto == IPPROTO_SCTP))
 587                 return -1;
 588
 589         length_for_csum = (be16_to_cpu(iph->tot_len) - (iph->ihl << 2));
 590         csum_pseudo_header = csum_tcpudp_nofold(iph->saddr, iph->daddr,
 591                                                 length_for_csum, ipproto, 0);
 592         skb->csum = csum_sub(hw_checksum, csum_pseudo_header);
 593         return 0;
 594 }
 595
 596 #if IS_ENABLED(CONFIG_IPV6)
 597 /* In IPv6 packets, hw_checksum lacks 6 bytes from IPv6 header:
 598  * 4 first bytes : priority, version, flow_lbl
 599  * and 2 additional bytes : nexthdr, hop_limit.
 600  */
 601 static int get_fixed_ipv6_csum(__wsum hw_checksum, struct sk_buff *skb,
 602                                struct ipv6hdr *ipv6h)
 603 {
 604         __u8 nexthdr = ipv6h->nexthdr;
 605         __wsum temp;
 606
 607         if (unlikely(nexthdr == IPPROTO_FRAGMENT ||
 608                      nexthdr == IPPROTO_HOPOPTS ||
 609                      nexthdr == IPPROTO_SCTP))
 610                 return -1;
 611
 612         /* priority, version, flow_lbl */
 613         temp = csum_add(hw_checksum, *(__wsum *)ipv6h);
 614         /* nexthdr and hop_limit */
 615         skb->csum = csum_add(temp, (__force __wsum)*(__be16 *)&ipv6h->nexthdr);
 616         return 0;
 617 }
 618 #endif
 619
 620 #define short_frame(size) ((size) <= ETH_ZLEN + ETH_FCS_LEN)
 621
 622 /* We reach this function only after checking that any of
 623  * the (IPv4 | IPv6) bits are set in cqe->status.
 624  */
 625 static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
 626                       netdev_features_t dev_features)
 627 {
 628         __wsum hw_checksum = 0;
 629         void *hdr;
 630
 631         /* CQE csum doesn't cover padding octets in short ethernet
 632          * frames. And the pad field is appended prior to calculating
 633          * and appending the FCS field.
 634          *
 635          * Detecting these padded frames requires to verify and parse
 636          * IP headers, so we simply force all those small frames to skip
 637          * checksum complete.
 638          */
 639         if (short_frame(skb->len))
 640                 return -EINVAL;
 641
 642         hdr = (u8 *)va + sizeof(struct ethhdr);
 643         hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
 644
 645         if (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK) &&
 646             !(dev_features & NETIF_F_HW_VLAN_CTAG_RX)) {
 647                 hw_checksum = get_fixed_vlan_csum(hw_checksum, hdr);
 648                 hdr += sizeof(struct vlan_hdr);
 649         }
 650
 651 #if IS_ENABLED(CONFIG_IPV6)
 652         if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV6))
 653                 return get_fixed_ipv6_csum(hw_checksum, skb, hdr);
 654 #endif
 655         return get_fixed_ipv4_csum(hw_checksum, skb, hdr);
 656 }
 657
 658 #if IS_ENABLED(CONFIG_IPV6)
 659 #define MLX4_CQE_STATUS_IP_ANY (MLX4_CQE_STATUS_IPV4 | MLX4_CQE_STATUS_IPV6)
 660 #else
 661 #define MLX4_CQE_STATUS_IP_ANY (MLX4_CQE_STATUS_IPV4)
 662 #endif
 663
 664 struct mlx4_en_xdp_buff {
 665         struct xdp_buff xdp;
 666         struct mlx4_cqe *cqe;
 667         struct mlx4_en_dev *mdev;
 668         struct mlx4_en_rx_ring *ring;
 669         struct net_device *dev;
 670 };
 671
 672 int mlx4_en_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
 673 {
 674         struct mlx4_en_xdp_buff *_ctx = (void *)ctx;
 675
 676         if (unlikely(_ctx->ring->hwtstamp_rx_filter != HWTSTAMP_FILTER_ALL))
 677                 return -ENODATA;
 678
 679         *timestamp = mlx4_en_get_hwtstamp(_ctx->mdev,
 680                                           mlx4_en_get_cqe_ts(_ctx->cqe));
 681         return 0;
 682 }
 683
 684 int mlx4_en_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash,
 685                         enum xdp_rss_hash_type *rss_type)
 686 {
 687         struct mlx4_en_xdp_buff *_ctx = (void *)ctx;
 688         struct mlx4_cqe *cqe = _ctx->cqe;
 689         enum xdp_rss_hash_type xht = 0;
 690         __be16 status;
 691
 692         if (unlikely(!(_ctx->dev->features & NETIF_F_RXHASH)))
 693                 return -ENODATA;
 694
 695         *hash = be32_to_cpu(cqe->immed_rss_invalid);
 696         status = cqe->status;
 697         if (status & cpu_to_be16(MLX4_CQE_STATUS_TCP))
 698                 xht = XDP_RSS_L4_TCP;
 699         if (status & cpu_to_be16(MLX4_CQE_STATUS_UDP))
 700                 xht = XDP_RSS_L4_UDP;
 701         if (status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 | MLX4_CQE_STATUS_IPV4F))
 702                 xht |= XDP_RSS_L3_IPV4;
 703         if (status & cpu_to_be16(MLX4_CQE_STATUS_IPV6)) {
 704                 xht |= XDP_RSS_L3_IPV6;
 705                 if (cqe->ipv6_ext_mask)
 706                         xht |= XDP_RSS_L3_DYNHDR;
 707         }
 708         *rss_type = xht;
 709
 710         return 0;
 711 }
 712
 713 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
 714 {
 715         struct mlx4_en_priv *priv = netdev_priv(dev);
 716         struct mlx4_en_xdp_buff mxbuf = {};
 717         int factor = priv->cqe_factor;
 718         struct mlx4_en_rx_ring *ring;
 719         struct bpf_prog *xdp_prog;
 720         int cq_ring = cq->ring;
 721         bool doorbell_pending;
 722         bool xdp_redir_flush;
 723         struct mlx4_cqe *cqe;
 724         int polled = 0;
 725         int index;
 726
 727         if (unlikely(!priv->port_up || budget <= 0))
 728                 return 0;
 729
 730         ring = priv->rx_ring[cq_ring];
 731
 732         xdp_prog = rcu_dereference_bh(ring->xdp_prog);
 733         xdp_init_buff(&mxbuf.xdp, priv->frag_info[0].frag_stride, &ring->xdp_rxq);
 734         doorbell_pending = false;
 735         xdp_redir_flush = false;
 736
 737         /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
 738          * descriptor offset can be deduced from the CQE index instead of
 739          * reading 'cqe->index' */
 740         index = cq->mcq.cons_index & ring->size_mask;
 741         cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
 742
 743         /* Process all completed CQEs */
 744         while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
 745                     cq->mcq.cons_index & cq->size)) {
 746                 struct mlx4_en_rx_alloc *frags;
 747                 enum pkt_hash_types hash_type;
 748                 struct sk_buff *skb;
 749                 unsigned int length;
 750                 int ip_summed;
 751                 void *va;
 752                 int nr;
 753
 754                 frags = ring->rx_info + (index << priv->log_rx_info);
 755                 va = page_address(frags[0].page) + frags[0].page_offset;
 756                 net_prefetchw(va);
 757                 /*
 758                  * make sure we read the CQE after we read the ownership bit
 759                  */
 760                 dma_rmb();
 761
 762                 /* Drop packet on bad receive or bad checksum */
 763                 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
 764                                                 MLX4_CQE_OPCODE_ERROR)) {
 765                         en_err(priv, "CQE completed in error - vendor syndrome:%d syndrome:%d\n",
 766                                ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
 767                                ((struct mlx4_err_cqe *)cqe)->syndrome);
 768                         goto next;
 769                 }
 770                 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
 771                         en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
 772                         goto next;
 773                 }
 774
 775                 /* Check if we need to drop the packet if SRIOV is not enabled
 776                  * and not performing the selftest or flb disabled
 777                  */
 778                 if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
 779                         const struct ethhdr *ethh = va;
 780                         dma_addr_t dma;
 781                         /* Get pointer to first fragment since we haven't
 782                          * skb yet and cast it to ethhdr struct
 783                          */
 784                         dma = frags[0].dma + frags[0].page_offset;
 785                         dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
 786                                                 DMA_FROM_DEVICE);
 787
 788                         if (is_multicast_ether_addr(ethh->h_dest)) {
 789                                 struct mlx4_mac_entry *entry;
 790                                 struct hlist_head *bucket;
 791                                 unsigned int mac_hash;
 792
 793                                 /* Drop the packet, since HW loopback-ed it */
 794                                 mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
 795                                 bucket = &priv->mac_hash[mac_hash];
 796                                 hlist_for_each_entry_rcu_bh(entry, bucket, hlist) {
 797                                         if (ether_addr_equal_64bits(entry->mac,
 798                                                                     ethh->h_source))
 799                                                 goto next;
 800                                 }
 801                         }
 802                 }
 803
 804                 if (unlikely(priv->validate_loopback)) {
 805                         validate_loopback(priv, va);
 806                         goto next;
 807                 }
 808
 809                 /*
 810                  * Packet is OK - process it.
 811                  */
 812                 length = be32_to_cpu(cqe->byte_cnt);
 813                 length -= ring->fcs_del;
 814
 815                 /* A bpf program gets first chance to drop the packet. It may
 816                  * read bytes but not past the end of the frag.
 817                  */
 818                 if (xdp_prog) {
 819                         dma_addr_t dma;
 820                         void *orig_data;
 821                         u32 act;
 822
 823                         dma = frags[0].dma + frags[0].page_offset;
 824                         dma_sync_single_for_cpu(priv->ddev, dma,
 825                                                 priv->frag_info[0].frag_size,
 826                                                 DMA_FROM_DEVICE);
 827
 828                         xdp_prepare_buff(&mxbuf.xdp, va - frags[0].page_offset,
 829                                          frags[0].page_offset, length, true);
 830                         orig_data = mxbuf.xdp.data;
 831                         mxbuf.cqe = cqe;
 832                         mxbuf.mdev = priv->mdev;
 833                         mxbuf.ring = ring;
 834                         mxbuf.dev = dev;
 835
 836                         act = bpf_prog_run_xdp(xdp_prog, &mxbuf.xdp);
 837
 838                         length = mxbuf.xdp.data_end - mxbuf.xdp.data;
 839                         if (mxbuf.xdp.data != orig_data) {
 840                                 frags[0].page_offset = mxbuf.xdp.data -
 841                                         mxbuf.xdp.data_hard_start;
 842                                 va = mxbuf.xdp.data;
 843                         }
 844
 845                         switch (act) {
 846                         case XDP_PASS:
 847                                 break;
 848                         case XDP_REDIRECT:
 849                                 if (likely(!xdp_do_redirect(dev, &mxbuf.xdp, xdp_prog))) {
 850                                         ring->xdp_redirect++;
 851                                         xdp_redir_flush = true;
 852                                         frags[0].page = NULL;
 853                                         goto next;
 854                                 }
 855                                 ring->xdp_redirect_fail++;
 856                                 trace_xdp_exception(dev, xdp_prog, act);
 857                                 goto xdp_drop_no_cnt;
 858                         case XDP_TX:
 859                                 if (likely(!mlx4_en_xmit_frame(ring, frags, priv,
 860                                                         length, cq_ring,
 861                                                         &doorbell_pending))) {
 862                                         frags[0].page = NULL;
 863                                         goto next;
 864                                 }
 865                                 trace_xdp_exception(dev, xdp_prog, act);
 866                                 goto xdp_drop_no_cnt; /* Drop on xmit failure */
 867                         default:
 868                                 bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
 869                                 fallthrough;
 870                         case XDP_ABORTED:
 871                                 trace_xdp_exception(dev, xdp_prog, act);
 872                                 fallthrough;
 873                         case XDP_DROP:
 874                                 ring->xdp_drop++;
 875 xdp_drop_no_cnt:
 876                                 goto next;
 877                         }
 878                 }
 879
 880                 ring->bytes += length;
 881                 ring->packets++;
 882
 883                 skb = napi_get_frags(&cq->napi);
 884                 if (unlikely(!skb))
 885                         goto next;
 886
 887                 if (unlikely(ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL)) {
 888                         u64 timestamp = mlx4_en_get_cqe_ts(cqe);
 889
 890                         mlx4_en_fill_hwtstamps(priv->mdev, skb_hwtstamps(skb),
 891                                                timestamp);
 892                 }
 893                 skb_record_rx_queue(skb, cq_ring);
 894
 895                 if (likely(dev->features & NETIF_F_RXCSUM)) {
 896                         /* TODO: For IP non TCP/UDP packets when csum complete is
 897                          * not an option (not supported or any other reason) we can
 898                          * actually check cqe IPOK status bit and report
 899                          * CHECKSUM_UNNECESSARY rather than CHECKSUM_NONE
 900                          */
 901                         if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
 902                                                        MLX4_CQE_STATUS_UDP)) &&
 903                             (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
 904                             cqe->checksum == cpu_to_be16(0xffff)) {
 905                                 bool l2_tunnel;
 906
 907                                 l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
 908                                         (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
 909                                 ip_summed = CHECKSUM_UNNECESSARY;
 910                                 hash_type = PKT_HASH_TYPE_L4;
 911                                 if (l2_tunnel)
 912                                         skb->csum_level = 1;
 913                                 ring->csum_ok++;
 914                         } else {
 915                                 if (!(priv->flags & MLX4_EN_FLAG_RX_CSUM_NON_TCP_UDP &&
 916                                       (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IP_ANY))))
 917                                         goto csum_none;
 918                                 if (check_csum(cqe, skb, va, dev->features))
 919                                         goto csum_none;
 920                                 ip_summed = CHECKSUM_COMPLETE;
 921                                 hash_type = PKT_HASH_TYPE_L3;
 922                                 ring->csum_complete++;
 923                         }
 924                 } else {
 925 csum_none:
 926                         ip_summed = CHECKSUM_NONE;
 927                         hash_type = PKT_HASH_TYPE_L3;
 928                         ring->csum_none++;
 929                 }
 930                 skb->ip_summed = ip_summed;
 931                 if (dev->features & NETIF_F_RXHASH)
 932                         skb_set_hash(skb,
 933                                      be32_to_cpu(cqe->immed_rss_invalid),
 934                                      hash_type);
 935
 936                 if ((cqe->vlan_my_qpn &
 937                      cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
 938                     (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
 939                         __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
 940                                                be16_to_cpu(cqe->sl_vid));
 941                 else if ((cqe->vlan_my_qpn &
 942                           cpu_to_be32(MLX4_CQE_SVLAN_PRESENT_MASK)) &&
 943                          (dev->features & NETIF_F_HW_VLAN_STAG_RX))
 944                         __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021AD),
 945                                                be16_to_cpu(cqe->sl_vid));
 946
 947                 nr = mlx4_en_complete_rx_desc(priv, frags, skb, length);
 948                 if (likely(nr)) {
 949                         skb_shinfo(skb)->nr_frags = nr;
 950                         skb->len = length;
 951                         skb->data_len = length;
 952                         napi_gro_frags(&cq->napi);
 953                 } else {
 954                         __vlan_hwaccel_clear_tag(skb);
 955                         skb_clear_hash(skb);
 956                 }
 957 next:
 958                 ++cq->mcq.cons_index;
 959                 index = (cq->mcq.cons_index) & ring->size_mask;
 960                 cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
 961                 if (unlikely(++polled == budget))
 962                         break;
 963         }
 964
 965         if (xdp_redir_flush)
 966                 xdp_do_flush();
 967
 968         if (likely(polled)) {
 969                 if (doorbell_pending) {
 970                         priv->tx_cq[TX_XDP][cq_ring]->xdp_busy = true;
 971                         mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq_ring]);
 972                 }
 973
 974                 mlx4_cq_set_ci(&cq->mcq);
 975                 wmb(); /* ensure HW sees CQ consumer before we post new buffers */
 976                 ring->cons = cq->mcq.cons_index;
 977         }
 978
 979         mlx4_en_refill_rx_buffers(priv, ring);
 980
 981         return polled;
 982 }
 983
 984
 985 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
 986 {
 987         struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
 988         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
 989
 990         if (likely(priv->port_up))
 991                 napi_schedule_irqoff(&cq->napi);
 992         else
 993                 mlx4_en_arm_cq(priv, cq);
 994 }
 995
 996 /* Rx CQ polling - called by NAPI */
 997 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
 998 {
 999         struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
1000         struct net_device *dev = cq->dev;
1001         struct mlx4_en_priv *priv = netdev_priv(dev);
1002         struct mlx4_en_cq *xdp_tx_cq = NULL;
1003         bool clean_complete = true;
1004         int done;
1005
1006         if (!budget)
1007                 return 0;
1008
1009         if (priv->tx_ring_num[TX_XDP]) {
1010                 xdp_tx_cq = priv->tx_cq[TX_XDP][cq->ring];
1011                 if (xdp_tx_cq->xdp_busy) {
1012                         clean_complete = mlx4_en_process_tx_cq(dev, xdp_tx_cq,
1013                                                                budget) < budget;
1014                         xdp_tx_cq->xdp_busy = !clean_complete;
1015                 }
1016         }
1017
1018         done = mlx4_en_process_rx_cq(dev, cq, budget);
1019
1020         /* If we used up all the quota - we're probably not done yet... */
1021         if (done == budget || !clean_complete) {
1022                 int cpu_curr;
1023
1024                 /* in case we got here because of !clean_complete */
1025                 done = budget;
1026
1027                 cpu_curr = smp_processor_id();
1028
1029                 if (likely(cpumask_test_cpu(cpu_curr, cq->aff_mask)))
1030                         return budget;
1031
1032                 /* Current cpu is not according to smp_irq_affinity -
1033                  * probably affinity changed. Need to stop this NAPI
1034                  * poll, and restart it on the right CPU.
1035                  * Try to avoid returning a too small value (like 0),
1036                  * to not fool net_rx_action() and its netdev_budget
1037                  */
1038                 if (done)
1039                         done--;
1040         }
1041         /* Done for now */
1042         if (likely(napi_complete_done(napi, done)))
1043                 mlx4_en_arm_cq(priv, cq);
1044         return done;
1045 }
1046
1047 void mlx4_en_calc_rx_buf(struct net_device *dev)
1048 {
1049         struct mlx4_en_priv *priv = netdev_priv(dev);
1050         int eff_mtu = MLX4_EN_EFF_MTU(dev->mtu);
1051         int i = 0;
1052
1053         /* bpf requires buffers to be set up as 1 packet per page.
1054          * This only works when num_frags == 1.
1055          */
1056         if (priv->tx_ring_num[TX_XDP]) {
1057                 priv->frag_info[0].frag_size = eff_mtu;
1058                 /* This will gain efficient xdp frame recycling at the
1059                  * expense of more costly truesize accounting
1060                  */
1061                 priv->frag_info[0].frag_stride = PAGE_SIZE;
1062                 priv->dma_dir = DMA_BIDIRECTIONAL;
1063                 priv->rx_headroom = XDP_PACKET_HEADROOM;
1064                 i = 1;
1065         } else {
1066                 int frag_size_max = 2048, buf_size = 0;
1067
1068                 /* should not happen, right ? */
1069                 if (eff_mtu > PAGE_SIZE + (MLX4_EN_MAX_RX_FRAGS - 1) * 2048)
1070                         frag_size_max = PAGE_SIZE;
1071
1072                 while (buf_size < eff_mtu) {
1073                         int frag_stride, frag_size = eff_mtu - buf_size;
1074                         int pad, nb;
1075
1076                         if (i < MLX4_EN_MAX_RX_FRAGS - 1)
1077                                 frag_size = min(frag_size, frag_size_max);
1078
1079                         priv->frag_info[i].frag_size = frag_size;
1080                         frag_stride = ALIGN(frag_size, SMP_CACHE_BYTES);
1081                         /* We can only pack 2 1536-bytes frames in on 4K page
1082                          * Therefore, each frame would consume more bytes (truesize)
1083                          */
1084                         nb = PAGE_SIZE / frag_stride;
1085                         pad = (PAGE_SIZE - nb * frag_stride) / nb;
1086                         pad &= ~(SMP_CACHE_BYTES - 1);
1087                         priv->frag_info[i].frag_stride = frag_stride + pad;
1088
1089                         buf_size += frag_size;
1090                         i++;
1091                 }
1092                 priv->dma_dir = DMA_FROM_DEVICE;
1093                 priv->rx_headroom = 0;
1094         }
1095
1096         priv->num_frags = i;
1097         priv->rx_skb_size = eff_mtu;
1098         priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
1099
1100         en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d num_frags:%d):\n",
1101                eff_mtu, priv->num_frags);
1102         for (i = 0; i < priv->num_frags; i++) {
1103                 en_dbg(DRV,
1104                        priv,
1105                        "  frag:%d - size:%d stride:%d\n",
1106                        i,
1107                        priv->frag_info[i].frag_size,
1108                        priv->frag_info[i].frag_stride);
1109         }
1110 }
1111
1112 /* RSS related functions */
1113
1114 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
1115                                  struct mlx4_en_rx_ring *ring,
1116                                  enum mlx4_qp_state *state,
1117                                  struct mlx4_qp *qp)
1118 {
1119         struct mlx4_en_dev *mdev = priv->mdev;
1120         struct mlx4_qp_context *context;
1121         int err = 0;
1122
1123         context = kzalloc(sizeof(*context), GFP_KERNEL);
1124         if (!context)
1125                 return -ENOMEM;
1126
1127         err = mlx4_qp_alloc(mdev->dev, qpn, qp);
1128         if (err) {
1129                 en_err(priv, "Failed to allocate qp #%x\n", qpn);
1130                 goto out;
1131         }
1132         qp->event = mlx4_en_sqp_event;
1133
1134         mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
1135                                 qpn, ring->cqn, -1, context);
1136         context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
1137
1138         /* Cancel FCS removal if FW allows */
1139         if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
1140                 context->param3 |= cpu_to_be32(1 << 29);
1141                 if (priv->dev->features & NETIF_F_RXFCS)
1142                         ring->fcs_del = 0;
1143                 else
1144                         ring->fcs_del = ETH_FCS_LEN;
1145         } else
1146                 ring->fcs_del = 0;
1147
1148         err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
1149         if (err) {
1150                 mlx4_qp_remove(mdev->dev, qp);
1151                 mlx4_qp_free(mdev->dev, qp);
1152         }
1153         mlx4_en_update_rx_prod_db(ring);
1154 out:
1155         kfree(context);
1156         return err;
1157 }
1158
1159 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1160 {
1161         int err;
1162         u32 qpn;
1163
1164         err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn,
1165                                     MLX4_RESERVE_A0_QP,
1166                                     MLX4_RES_USAGE_DRIVER);
1167         if (err) {
1168                 en_err(priv, "Failed reserving drop qpn\n");
1169                 return err;
1170         }
1171         err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp);
1172         if (err) {
1173                 en_err(priv, "Failed allocating drop qp\n");
1174                 mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1175                 return err;
1176         }
1177
1178         return 0;
1179 }
1180
1181 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1182 {
1183         u32 qpn;
1184
1185         qpn = priv->drop_qp.qpn;
1186         mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1187         mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1188         mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1189 }
1190
1191 /* Allocate rx qp's and configure them according to rss map */
1192 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1193 {
1194         struct mlx4_en_dev *mdev = priv->mdev;
1195         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1196         struct mlx4_qp_context context;
1197         struct mlx4_rss_context *rss_context;
1198         int rss_rings;
1199         void *ptr;
1200         u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1201                         MLX4_RSS_TCP_IPV6);
1202         int i, qpn;
1203         int err = 0;
1204         int good_qps = 0;
1205         u8 flags;
1206
1207         en_dbg(DRV, priv, "Configuring rss steering\n");
1208
1209         flags = priv->rx_ring_num == 1 ? MLX4_RESERVE_A0_QP : 0;
1210         err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1211                                     priv->rx_ring_num,
1212                                     &rss_map->base_qpn, flags,
1213                                     MLX4_RES_USAGE_DRIVER);
1214         if (err) {
1215                 en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1216                 return err;
1217         }
1218
1219         for (i = 0; i < priv->rx_ring_num; i++) {
1220                 qpn = rss_map->base_qpn + i;
1221                 err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1222                                             &rss_map->state[i],
1223                                             &rss_map->qps[i]);
1224                 if (err)
1225                         goto rss_err;
1226
1227                 ++good_qps;
1228         }
1229
1230         if (priv->rx_ring_num == 1) {
1231                 rss_map->indir_qp = &rss_map->qps[0];
1232                 priv->base_qpn = rss_map->indir_qp->qpn;
1233                 en_info(priv, "Optimized Non-RSS steering\n");
1234                 return 0;
1235         }
1236
1237         rss_map->indir_qp = kzalloc(sizeof(*rss_map->indir_qp), GFP_KERNEL);
1238         if (!rss_map->indir_qp) {
1239                 err = -ENOMEM;
1240                 goto rss_err;
1241         }
1242
1243         /* Configure RSS indirection qp */
1244         err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, rss_map->indir_qp);
1245         if (err) {
1246                 en_err(priv, "Failed to allocate RSS indirection QP\n");
1247                 goto qp_alloc_err;
1248         }
1249
1250         rss_map->indir_qp->event = mlx4_en_sqp_event;
1251         mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1252                                 priv->rx_ring[0]->cqn, -1, &context);
1253
1254         if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1255                 rss_rings = priv->rx_ring_num;
1256         else
1257                 rss_rings = priv->prof->rss_rings;
1258
1259         ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1260                                         + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1261         rss_context = ptr;
1262         rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1263                                             (rss_map->base_qpn));
1264         rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1265         if (priv->mdev->profile.udp_rss) {
1266                 rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1267                 rss_context->base_qpn_udp = rss_context->default_qpn;
1268         }
1269
1270         if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1271                 en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1272                 rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1273         }
1274
1275         rss_context->flags = rss_mask;
1276         rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1277         if (priv->rss_hash_fn == ETH_RSS_HASH_XOR) {
1278                 rss_context->hash_fn = MLX4_RSS_HASH_XOR;
1279         } else if (priv->rss_hash_fn == ETH_RSS_HASH_TOP) {
1280                 rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1281                 memcpy(rss_context->rss_key, priv->rss_key,
1282                        MLX4_EN_RSS_KEY_SIZE);
1283         } else {
1284                 en_err(priv, "Unknown RSS hash function requested\n");
1285                 err = -EINVAL;
1286                 goto indir_err;
1287         }
1288
1289         err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1290                                rss_map->indir_qp, &rss_map->indir_state);
1291         if (err)
1292                 goto indir_err;
1293
1294         return 0;
1295
1296 indir_err:
1297         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1298                        MLX4_QP_STATE_RST, NULL, 0, 0, rss_map->indir_qp);
1299         mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
1300         mlx4_qp_free(mdev->dev, rss_map->indir_qp);
1301 qp_alloc_err:
1302         kfree(rss_map->indir_qp);
1303         rss_map->indir_qp = NULL;
1304 rss_err:
1305         for (i = 0; i < good_qps; i++) {
1306                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1307                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1308                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1309                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1310         }
1311         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1312         return err;
1313 }
1314
1315 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1316 {
1317         struct mlx4_en_dev *mdev = priv->mdev;
1318         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1319         int i;
1320
1321         if (priv->rx_ring_num > 1) {
1322                 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1323                                MLX4_QP_STATE_RST, NULL, 0, 0,
1324                                rss_map->indir_qp);
1325                 mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
1326                 mlx4_qp_free(mdev->dev, rss_map->indir_qp);
1327                 kfree(rss_map->indir_qp);
1328                 rss_map->indir_qp = NULL;
1329         }
1330
1331         for (i = 0; i < priv->rx_ring_num; i++) {
1332                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1333                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1334                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1335                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1336         }
1337         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1338 }