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/mlx4/cq.h>
  35 #include <linux/slab.h>
  36 #include <linux/mlx4/qp.h>
  37 #include <linux/skbuff.h>
  38 #include <linux/if_ether.h>
  39 #include <linux/if_vlan.h>
  40 #include <linux/vmalloc.h>
  41
  42 #include "mlx4_en.h"
  43
  44
  45 static int mlx4_en_alloc_frag(struct mlx4_en_priv *priv,
  46                               struct mlx4_en_rx_desc *rx_desc,
  47                               struct page_frag *skb_frags,
  48                               struct mlx4_en_rx_alloc *ring_alloc,
  49                               int i)
  50 {
  51         struct mlx4_en_dev *mdev = priv->mdev;
  52         struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
  53         struct mlx4_en_rx_alloc *page_alloc = &ring_alloc[i];
  54         struct page *page;
  55         dma_addr_t dma;
  56
  57         if (page_alloc->offset == frag_info->last_offset) {
  58                 /* Allocate new page */
  59                 page = alloc_pages(GFP_ATOMIC | __GFP_COMP, MLX4_EN_ALLOC_ORDER);
  60                 if (!page)
  61                         return -ENOMEM;
  62
  63                 skb_frags[i].page = page_alloc->page;
  64                 skb_frags[i].offset = page_alloc->offset;
  65                 page_alloc->page = page;
  66                 page_alloc->offset = frag_info->frag_align;
  67         } else {
  68                 page = page_alloc->page;
  69                 get_page(page);
  70
  71                 skb_frags[i].page = page;
  72                 skb_frags[i].offset = page_alloc->offset;
  73                 page_alloc->offset += frag_info->frag_stride;
  74         }
  75         dma = pci_map_single(mdev->pdev, page_address(skb_frags[i].page) +
  76                              skb_frags[i].offset, frag_info->frag_size,
  77                              PCI_DMA_FROMDEVICE);
  78         rx_desc->data[i].addr = cpu_to_be64(dma);
  79         return 0;
  80 }
  81
  82 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
  83                                   struct mlx4_en_rx_ring *ring)
  84 {
  85         struct mlx4_en_rx_alloc *page_alloc;
  86         int i;
  87
  88         for (i = 0; i < priv->num_frags; i++) {
  89                 page_alloc = &ring->page_alloc[i];
  90                 page_alloc->page = alloc_pages(GFP_ATOMIC | __GFP_COMP,
  91                                                MLX4_EN_ALLOC_ORDER);
  92                 if (!page_alloc->page)
  93                         goto out;
  94
  95                 page_alloc->offset = priv->frag_info[i].frag_align;
  96                 en_dbg(DRV, priv, "Initialized allocator:%d with page:%p\n",
  97                        i, page_alloc->page);
  98         }
  99         return 0;
 100
 101 out:
 102         while (i--) {
 103                 page_alloc = &ring->page_alloc[i];
 104                 put_page(page_alloc->page);
 105                 page_alloc->page = NULL;
 106         }
 107         return -ENOMEM;
 108 }
 109
 110 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
 111                                       struct mlx4_en_rx_ring *ring)
 112 {
 113         struct mlx4_en_rx_alloc *page_alloc;
 114         int i;
 115
 116         for (i = 0; i < priv->num_frags; i++) {
 117                 page_alloc = &ring->page_alloc[i];
 118                 en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
 119                        i, page_count(page_alloc->page));
 120
 121                 put_page(page_alloc->page);
 122                 page_alloc->page = NULL;
 123         }
 124 }
 125
 126
 127 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
 128                                  struct mlx4_en_rx_ring *ring, int index)
 129 {
 130         struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
 131         struct skb_frag_struct *skb_frags = ring->rx_info +
 132                                             (index << priv->log_rx_info);
 133         int possible_frags;
 134         int i;
 135
 136         /* Set size and memtype fields */
 137         for (i = 0; i < priv->num_frags; i++) {
 138                 skb_frag_size_set(&skb_frags[i], priv->frag_info[i].frag_size);
 139                 rx_desc->data[i].byte_count =
 140                         cpu_to_be32(priv->frag_info[i].frag_size);
 141                 rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
 142         }
 143
 144         /* If the number of used fragments does not fill up the ring stride,
 145          * remaining (unused) fragments must be padded with null address/size
 146          * and a special memory key */
 147         possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
 148         for (i = priv->num_frags; i < possible_frags; i++) {
 149                 rx_desc->data[i].byte_count = 0;
 150                 rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
 151                 rx_desc->data[i].addr = 0;
 152         }
 153 }
 154
 155
 156 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
 157                                    struct mlx4_en_rx_ring *ring, int index)
 158 {
 159         struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
 160         struct page_frag *skb_frags = ring->rx_info +
 161                                       (index << priv->log_rx_info);
 162         int i;
 163
 164         for (i = 0; i < priv->num_frags; i++)
 165                 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, ring->page_alloc, i))
 166                         goto err;
 167
 168         return 0;
 169
 170 err:
 171         while (i--)
 172                 put_page(skb_frags[i].page);
 173         return -ENOMEM;
 174 }
 175
 176 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
 177 {
 178         *ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
 179 }
 180
 181 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
 182                                  struct mlx4_en_rx_ring *ring,
 183                                  int index)
 184 {
 185         struct mlx4_en_dev *mdev = priv->mdev;
 186         struct page_frag *skb_frags;
 187         struct mlx4_en_rx_desc *rx_desc = ring->buf + (index << ring->log_stride);
 188         dma_addr_t dma;
 189         int nr;
 190
 191         skb_frags = ring->rx_info + (index << priv->log_rx_info);
 192         for (nr = 0; nr < priv->num_frags; nr++) {
 193                 en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
 194                 dma = be64_to_cpu(rx_desc->data[nr].addr);
 195
 196                 en_dbg(DRV, priv, "Unmapping buffer at dma:0x%llx\n", (u64) dma);
 197                 pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
 198                                  PCI_DMA_FROMDEVICE);
 199                 put_page(skb_frags[nr].page);
 200         }
 201 }
 202
 203 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
 204 {
 205         struct mlx4_en_rx_ring *ring;
 206         int ring_ind;
 207         int buf_ind;
 208         int new_size;
 209
 210         for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
 211                 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 212                         ring = &priv->rx_ring[ring_ind];
 213
 214                         if (mlx4_en_prepare_rx_desc(priv, ring,
 215                                                     ring->actual_size)) {
 216                                 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
 217                                         en_err(priv, "Failed to allocate "
 218                                                      "enough rx buffers\n");
 219                                         return -ENOMEM;
 220                                 } else {
 221                                         new_size = rounddown_pow_of_two(ring->actual_size);
 222                                         en_warn(priv, "Only %d buffers allocated "
 223                                                       "reducing ring size to %d",
 224                                                 ring->actual_size, new_size);
 225                                         goto reduce_rings;
 226                                 }
 227                         }
 228                         ring->actual_size++;
 229                         ring->prod++;
 230                 }
 231         }
 232         return 0;
 233
 234 reduce_rings:
 235         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 236                 ring = &priv->rx_ring[ring_ind];
 237                 while (ring->actual_size > new_size) {
 238                         ring->actual_size--;
 239                         ring->prod--;
 240                         mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
 241                 }
 242         }
 243
 244         return 0;
 245 }
 246
 247 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
 248                                 struct mlx4_en_rx_ring *ring)
 249 {
 250         int index;
 251
 252         en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
 253                ring->cons, ring->prod);
 254
 255         /* Unmap and free Rx buffers */
 256         BUG_ON((u32) (ring->prod - ring->cons) > ring->actual_size);
 257         while (ring->cons != ring->prod) {
 258                 index = ring->cons & ring->size_mask;
 259                 en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
 260                 mlx4_en_free_rx_desc(priv, ring, index);
 261                 ++ring->cons;
 262         }
 263 }
 264
 265 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
 266                            struct mlx4_en_rx_ring *ring, u32 size, u16 stride)
 267 {
 268         struct mlx4_en_dev *mdev = priv->mdev;
 269         int err;
 270         int tmp;
 271
 272
 273         ring->prod = 0;
 274         ring->cons = 0;
 275         ring->size = size;
 276         ring->size_mask = size - 1;
 277         ring->stride = stride;
 278         ring->log_stride = ffs(ring->stride) - 1;
 279         ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
 280
 281         tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
 282                                         sizeof(struct skb_frag_struct));
 283         ring->rx_info = vmalloc(tmp);
 284         if (!ring->rx_info) {
 285                 en_err(priv, "Failed allocating rx_info ring\n");
 286                 return -ENOMEM;
 287         }
 288         en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
 289                  ring->rx_info, tmp);
 290
 291         err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
 292                                  ring->buf_size, 2 * PAGE_SIZE);
 293         if (err)
 294                 goto err_ring;
 295
 296         err = mlx4_en_map_buffer(&ring->wqres.buf);
 297         if (err) {
 298                 en_err(priv, "Failed to map RX buffer\n");
 299                 goto err_hwq;
 300         }
 301         ring->buf = ring->wqres.buf.direct.buf;
 302
 303         return 0;
 304
 305 err_hwq:
 306         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
 307 err_ring:
 308         vfree(ring->rx_info);
 309         ring->rx_info = NULL;
 310         return err;
 311 }
 312
 313 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
 314 {
 315         struct mlx4_en_rx_ring *ring;
 316         int i;
 317         int ring_ind;
 318         int err;
 319         int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
 320                                         DS_SIZE * priv->num_frags);
 321
 322         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 323                 ring = &priv->rx_ring[ring_ind];
 324
 325                 ring->prod = 0;
 326                 ring->cons = 0;
 327                 ring->actual_size = 0;
 328                 ring->cqn = priv->rx_cq[ring_ind].mcq.cqn;
 329
 330                 ring->stride = stride;
 331                 if (ring->stride <= TXBB_SIZE) {
 332                         /* Stamp first unused send wqe */
 333                         __be32 *ptr = (__be32 *)ring->buf;
 334                         __be32 stamp = cpu_to_be32(1 << STAMP_SHIFT);
 335                         *ptr = stamp;
 336                         /* Move pointer to start of rx section */
 337                         ring->buf += TXBB_SIZE;
 338                 }
 339
 340                 ring->log_stride = ffs(ring->stride) - 1;
 341                 ring->buf_size = ring->size * ring->stride;
 342
 343                 memset(ring->buf, 0, ring->buf_size);
 344                 mlx4_en_update_rx_prod_db(ring);
 345
 346                 /* Initailize all descriptors */
 347                 for (i = 0; i < ring->size; i++)
 348                         mlx4_en_init_rx_desc(priv, ring, i);
 349
 350                 /* Initialize page allocators */
 351                 err = mlx4_en_init_allocator(priv, ring);
 352                 if (err) {
 353                         en_err(priv, "Failed initializing ring allocator\n");
 354                         if (ring->stride <= TXBB_SIZE)
 355                                 ring->buf -= TXBB_SIZE;
 356                         ring_ind--;
 357                         goto err_allocator;
 358                 }
 359         }
 360         err = mlx4_en_fill_rx_buffers(priv);
 361         if (err)
 362                 goto err_buffers;
 363
 364         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
 365                 ring = &priv->rx_ring[ring_ind];
 366
 367                 ring->size_mask = ring->actual_size - 1;
 368                 mlx4_en_update_rx_prod_db(ring);
 369         }
 370
 371         return 0;
 372
 373 err_buffers:
 374         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
 375                 mlx4_en_free_rx_buf(priv, &priv->rx_ring[ring_ind]);
 376
 377         ring_ind = priv->rx_ring_num - 1;
 378 err_allocator:
 379         while (ring_ind >= 0) {
 380                 if (priv->rx_ring[ring_ind].stride <= TXBB_SIZE)
 381                         priv->rx_ring[ring_ind].buf -= TXBB_SIZE;
 382                 mlx4_en_destroy_allocator(priv, &priv->rx_ring[ring_ind]);
 383                 ring_ind--;
 384         }
 385         return err;
 386 }
 387
 388 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
 389                              struct mlx4_en_rx_ring *ring)
 390 {
 391         struct mlx4_en_dev *mdev = priv->mdev;
 392
 393         mlx4_en_unmap_buffer(&ring->wqres.buf);
 394         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size + TXBB_SIZE);
 395         vfree(ring->rx_info);
 396         ring->rx_info = NULL;
 397 }
 398
 399 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
 400                                 struct mlx4_en_rx_ring *ring)
 401 {
 402         mlx4_en_free_rx_buf(priv, ring);
 403         if (ring->stride <= TXBB_SIZE)
 404                 ring->buf -= TXBB_SIZE;
 405         mlx4_en_destroy_allocator(priv, ring);
 406 }
 407
 408
 409 /* Unmap a completed descriptor and free unused pages */
 410 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
 411                                     struct mlx4_en_rx_desc *rx_desc,
 412                                     struct page_frag *skb_frags,
 413                                     struct sk_buff *skb,
 414                                     struct mlx4_en_rx_alloc *page_alloc,
 415                                     int length)
 416 {
 417         struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
 418         struct mlx4_en_dev *mdev = priv->mdev;
 419         struct mlx4_en_frag_info *frag_info;
 420         int nr;
 421         dma_addr_t dma;
 422
 423         /* Collect used fragments while replacing them in the HW descirptors */
 424         for (nr = 0; nr < priv->num_frags; nr++) {
 425                 frag_info = &priv->frag_info[nr];
 426                 if (length <= frag_info->frag_prefix_size)
 427                         break;
 428
 429                 /* Save page reference in skb */
 430                 __skb_frag_set_page(&skb_frags_rx[nr], skb_frags[nr].page);
 431                 skb_frag_size_set(&skb_frags_rx[nr], skb_frags[nr].size);
 432                 skb_frags_rx[nr].page_offset = skb_frags[nr].offset;
 433                 skb->truesize += frag_info->frag_stride;
 434                 dma = be64_to_cpu(rx_desc->data[nr].addr);
 435
 436                 /* Allocate a replacement page */
 437                 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, page_alloc, nr))
 438                         goto fail;
 439
 440                 /* Unmap buffer */
 441                 pci_unmap_single(mdev->pdev, dma, skb_frag_size(&skb_frags_rx[nr]),
 442                                  PCI_DMA_FROMDEVICE);
 443         }
 444         /* Adjust size of last fragment to match actual length */
 445         if (nr > 0)
 446                 skb_frag_size_set(&skb_frags_rx[nr - 1],
 447                         length - priv->frag_info[nr - 1].frag_prefix_size);
 448         return nr;
 449
 450 fail:
 451         /* Drop all accumulated fragments (which have already been replaced in
 452          * the descriptor) of this packet; remaining fragments are reused... */
 453         while (nr > 0) {
 454                 nr--;
 455                 __skb_frag_unref(&skb_frags_rx[nr]);
 456         }
 457         return 0;
 458 }
 459
 460
 461 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
 462                                       struct mlx4_en_rx_desc *rx_desc,
 463                                       struct page_frag *skb_frags,
 464                                       struct mlx4_en_rx_alloc *page_alloc,
 465                                       unsigned int length)
 466 {
 467         struct mlx4_en_dev *mdev = priv->mdev;
 468         struct sk_buff *skb;
 469         void *va;
 470         int used_frags;
 471         dma_addr_t dma;
 472
 473         skb = dev_alloc_skb(SMALL_PACKET_SIZE + NET_IP_ALIGN);
 474         if (!skb) {
 475                 en_dbg(RX_ERR, priv, "Failed allocating skb\n");
 476                 return NULL;
 477         }
 478         skb->dev = priv->dev;
 479         skb_reserve(skb, NET_IP_ALIGN);
 480         skb->len = length;
 481
 482         /* Get pointer to first fragment so we could copy the headers into the
 483          * (linear part of the) skb */
 484         va = page_address(skb_frags[0].page) + skb_frags[0].offset;
 485
 486         if (length <= SMALL_PACKET_SIZE) {
 487                 /* We are copying all relevant data to the skb - temporarily
 488                  * synch buffers for the copy */
 489                 dma = be64_to_cpu(rx_desc->data[0].addr);
 490                 dma_sync_single_for_cpu(&mdev->pdev->dev, dma, length,
 491                                         DMA_FROM_DEVICE);
 492                 skb_copy_to_linear_data(skb, va, length);
 493                 dma_sync_single_for_device(&mdev->pdev->dev, dma, length,
 494                                            DMA_FROM_DEVICE);
 495                 skb->tail += length;
 496         } else {
 497
 498                 /* Move relevant fragments to skb */
 499                 used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, skb_frags,
 500                                                       skb, page_alloc, length);
 501                 if (unlikely(!used_frags)) {
 502                         kfree_skb(skb);
 503                         return NULL;
 504                 }
 505                 skb_shinfo(skb)->nr_frags = used_frags;
 506
 507                 /* Copy headers into the skb linear buffer */
 508                 memcpy(skb->data, va, HEADER_COPY_SIZE);
 509                 skb->tail += HEADER_COPY_SIZE;
 510
 511                 /* Skip headers in first fragment */
 512                 skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
 513
 514                 /* Adjust size of first fragment */
 515                 skb_frag_size_sub(&skb_shinfo(skb)->frags[0], HEADER_COPY_SIZE);
 516                 skb->data_len = length - HEADER_COPY_SIZE;
 517         }
 518         return skb;
 519 }
 520
 521 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
 522 {
 523         int i;
 524         int offset = ETH_HLEN;
 525
 526         for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
 527                 if (*(skb->data + offset) != (unsigned char) (i & 0xff))
 528                         goto out_loopback;
 529         }
 530         /* Loopback found */
 531         priv->loopback_ok = 1;
 532
 533 out_loopback:
 534         dev_kfree_skb_any(skb);
 535 }
 536
 537 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
 538 {
 539         struct mlx4_en_priv *priv = netdev_priv(dev);
 540         struct mlx4_cqe *cqe;
 541         struct mlx4_en_rx_ring *ring = &priv->rx_ring[cq->ring];
 542         struct page_frag *skb_frags;
 543         struct mlx4_en_rx_desc *rx_desc;
 544         struct sk_buff *skb;
 545         int index;
 546         int nr;
 547         unsigned int length;
 548         int polled = 0;
 549         int ip_summed;
 550
 551         if (!priv->port_up)
 552                 return 0;
 553
 554         /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
 555          * descriptor offset can be deduced from the CQE index instead of
 556          * reading 'cqe->index' */
 557         index = cq->mcq.cons_index & ring->size_mask;
 558         cqe = &cq->buf[index];
 559
 560         /* Process all completed CQEs */
 561         while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
 562                     cq->mcq.cons_index & cq->size)) {
 563
 564                 skb_frags = ring->rx_info + (index << priv->log_rx_info);
 565                 rx_desc = ring->buf + (index << ring->log_stride);
 566
 567                 /*
 568                  * make sure we read the CQE after we read the ownership bit
 569                  */
 570                 rmb();
 571
 572                 /* Drop packet on bad receive or bad checksum */
 573                 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
 574                                                 MLX4_CQE_OPCODE_ERROR)) {
 575                         en_err(priv, "CQE completed in error - vendor "
 576                                   "syndrom:%d syndrom:%d\n",
 577                                   ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
 578                                   ((struct mlx4_err_cqe *) cqe)->syndrome);
 579                         goto next;
 580                 }
 581                 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
 582                         en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
 583                         goto next;
 584                 }
 585
 586                 /*
 587                  * Packet is OK - process it.
 588                  */
 589                 length = be32_to_cpu(cqe->byte_cnt);
 590                 length -= ring->fcs_del;
 591                 ring->bytes += length;
 592                 ring->packets++;
 593
 594                 if (likely(dev->features & NETIF_F_RXCSUM)) {
 595                         if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
 596                             (cqe->checksum == cpu_to_be16(0xffff))) {
 597                                 ring->csum_ok++;
 598                                 /* This packet is eligible for LRO if it is:
 599                                  * - DIX Ethernet (type interpretation)
 600                                  * - TCP/IP (v4)
 601                                  * - without IP options
 602                                  * - not an IP fragment */
 603                                 if (dev->features & NETIF_F_GRO) {
 604                                         struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
 605                                         if (!gro_skb)
 606                                                 goto next;
 607
 608                                         nr = mlx4_en_complete_rx_desc(
 609                                                 priv, rx_desc,
 610                                                 skb_frags, gro_skb,
 611                                                 ring->page_alloc, length);
 612                                         if (!nr)
 613                                                 goto next;
 614
 615                                         skb_shinfo(gro_skb)->nr_frags = nr;
 616                                         gro_skb->len = length;
 617                                         gro_skb->data_len = length;
 618                                         gro_skb->ip_summed = CHECKSUM_UNNECESSARY;
 619
 620                                         if (cqe->vlan_my_qpn &
 621                                             cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK)) {
 622                                                 u16 vid = be16_to_cpu(cqe->sl_vid);
 623
 624                                                 __vlan_hwaccel_put_tag(gro_skb, vid);
 625                                         }
 626
 627                                         if (dev->features & NETIF_F_RXHASH)
 628                                                 gro_skb->rxhash = be32_to_cpu(cqe->immed_rss_invalid);
 629
 630                                         skb_record_rx_queue(gro_skb, cq->ring);
 631                                         napi_gro_frags(&cq->napi);
 632
 633                                         goto next;
 634                                 }
 635
 636                                 /* LRO not possible, complete processing here */
 637                                 ip_summed = CHECKSUM_UNNECESSARY;
 638                         } else {
 639                                 ip_summed = CHECKSUM_NONE;
 640                                 ring->csum_none++;
 641                         }
 642                 } else {
 643                         ip_summed = CHECKSUM_NONE;
 644                         ring->csum_none++;
 645                 }
 646
 647                 skb = mlx4_en_rx_skb(priv, rx_desc, skb_frags,
 648                                      ring->page_alloc, length);
 649                 if (!skb) {
 650                         priv->stats.rx_dropped++;
 651                         goto next;
 652                 }
 653
 654                 if (unlikely(priv->validate_loopback)) {
 655                         validate_loopback(priv, skb);
 656                         goto next;
 657                 }
 658
 659                 skb->ip_summed = ip_summed;
 660                 skb->protocol = eth_type_trans(skb, dev);
 661                 skb_record_rx_queue(skb, cq->ring);
 662
 663                 if (dev->features & NETIF_F_RXHASH)
 664                         skb->rxhash = be32_to_cpu(cqe->immed_rss_invalid);
 665
 666                 if (be32_to_cpu(cqe->vlan_my_qpn) &
 667                     MLX4_CQE_VLAN_PRESENT_MASK)
 668                         __vlan_hwaccel_put_tag(skb, be16_to_cpu(cqe->sl_vid));
 669
 670                 /* Push it up the stack */
 671                 netif_receive_skb(skb);
 672
 673 next:
 674                 ++cq->mcq.cons_index;
 675                 index = (cq->mcq.cons_index) & ring->size_mask;
 676                 cqe = &cq->buf[index];
 677                 if (++polled == budget) {
 678                         /* We are here because we reached the NAPI budget -
 679                          * flush only pending LRO sessions */
 680                         goto out;
 681                 }
 682         }
 683
 684 out:
 685         AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
 686         mlx4_cq_set_ci(&cq->mcq);
 687         wmb(); /* ensure HW sees CQ consumer before we post new buffers */
 688         ring->cons = cq->mcq.cons_index;
 689         ring->prod += polled; /* Polled descriptors were realocated in place */
 690         mlx4_en_update_rx_prod_db(ring);
 691         return polled;
 692 }
 693
 694
 695 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
 696 {
 697         struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
 698         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
 699
 700         if (priv->port_up)
 701                 napi_schedule(&cq->napi);
 702         else
 703                 mlx4_en_arm_cq(priv, cq);
 704 }
 705
 706 /* Rx CQ polling - called by NAPI */
 707 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
 708 {
 709         struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
 710         struct net_device *dev = cq->dev;
 711         struct mlx4_en_priv *priv = netdev_priv(dev);
 712         int done;
 713
 714         done = mlx4_en_process_rx_cq(dev, cq, budget);
 715
 716         /* If we used up all the quota - we're probably not done yet... */
 717         if (done == budget)
 718                 INC_PERF_COUNTER(priv->pstats.napi_quota);
 719         else {
 720                 /* Done for now */
 721                 napi_complete(napi);
 722                 mlx4_en_arm_cq(priv, cq);
 723         }
 724         return done;
 725 }
 726
 727
 728 /* Calculate the last offset position that accommodates a full fragment
 729  * (assuming fagment size = stride-align) */
 730 static int mlx4_en_last_alloc_offset(struct mlx4_en_priv *priv, u16 stride, u16 align)
 731 {
 732         u16 res = MLX4_EN_ALLOC_SIZE % stride;
 733         u16 offset = MLX4_EN_ALLOC_SIZE - stride - res + align;
 734
 735         en_dbg(DRV, priv, "Calculated last offset for stride:%d align:%d "
 736                             "res:%d offset:%d\n", stride, align, res, offset);
 737         return offset;
 738 }
 739
 740
 741 static int frag_sizes[] = {
 742         FRAG_SZ0,
 743         FRAG_SZ1,
 744         FRAG_SZ2,
 745         FRAG_SZ3
 746 };
 747
 748 void mlx4_en_calc_rx_buf(struct net_device *dev)
 749 {
 750         struct mlx4_en_priv *priv = netdev_priv(dev);
 751         int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
 752         int buf_size = 0;
 753         int i = 0;
 754
 755         while (buf_size < eff_mtu) {
 756                 priv->frag_info[i].frag_size =
 757                         (eff_mtu > buf_size + frag_sizes[i]) ?
 758                                 frag_sizes[i] : eff_mtu - buf_size;
 759                 priv->frag_info[i].frag_prefix_size = buf_size;
 760                 if (!i) {
 761                         priv->frag_info[i].frag_align = NET_IP_ALIGN;
 762                         priv->frag_info[i].frag_stride =
 763                                 ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
 764                 } else {
 765                         priv->frag_info[i].frag_align = 0;
 766                         priv->frag_info[i].frag_stride =
 767                                 ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
 768                 }
 769                 priv->frag_info[i].last_offset = mlx4_en_last_alloc_offset(
 770                                                 priv, priv->frag_info[i].frag_stride,
 771                                                 priv->frag_info[i].frag_align);
 772                 buf_size += priv->frag_info[i].frag_size;
 773                 i++;
 774         }
 775
 776         priv->num_frags = i;
 777         priv->rx_skb_size = eff_mtu;
 778         priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct skb_frag_struct));
 779
 780         en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
 781                   "num_frags:%d):\n", eff_mtu, priv->num_frags);
 782         for (i = 0; i < priv->num_frags; i++) {
 783                 en_dbg(DRV, priv, "  frag:%d - size:%d prefix:%d align:%d "
 784                                 "stride:%d last_offset:%d\n", i,
 785                                 priv->frag_info[i].frag_size,
 786                                 priv->frag_info[i].frag_prefix_size,
 787                                 priv->frag_info[i].frag_align,
 788                                 priv->frag_info[i].frag_stride,
 789                                 priv->frag_info[i].last_offset);
 790         }
 791 }
 792
 793 /* RSS related functions */
 794
 795 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
 796                                  struct mlx4_en_rx_ring *ring,
 797                                  enum mlx4_qp_state *state,
 798                                  struct mlx4_qp *qp)
 799 {
 800         struct mlx4_en_dev *mdev = priv->mdev;
 801         struct mlx4_qp_context *context;
 802         int err = 0;
 803
 804         context = kmalloc(sizeof *context , GFP_KERNEL);
 805         if (!context) {
 806                 en_err(priv, "Failed to allocate qp context\n");
 807                 return -ENOMEM;
 808         }
 809
 810         err = mlx4_qp_alloc(mdev->dev, qpn, qp);
 811         if (err) {
 812                 en_err(priv, "Failed to allocate qp #%x\n", qpn);
 813                 goto out;
 814         }
 815         qp->event = mlx4_en_sqp_event;
 816
 817         memset(context, 0, sizeof *context);
 818         mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
 819                                 qpn, ring->cqn, context);
 820         context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
 821
 822         /* Cancel FCS removal if FW allows */
 823         if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
 824                 context->param3 |= cpu_to_be32(1 << 29);
 825                 ring->fcs_del = ETH_FCS_LEN;
 826         } else
 827                 ring->fcs_del = 0;
 828
 829         err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
 830         if (err) {
 831                 mlx4_qp_remove(mdev->dev, qp);
 832                 mlx4_qp_free(mdev->dev, qp);
 833         }
 834         mlx4_en_update_rx_prod_db(ring);
 835 out:
 836         kfree(context);
 837         return err;
 838 }
 839
 840 /* Allocate rx qp's and configure them according to rss map */
 841 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
 842 {
 843         struct mlx4_en_dev *mdev = priv->mdev;
 844         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
 845         struct mlx4_qp_context context;
 846         struct mlx4_en_rss_context *rss_context;
 847         void *ptr;
 848         u8 rss_mask = 0x3f;
 849         int i, qpn;
 850         int err = 0;
 851         int good_qps = 0;
 852         static const u32 rsskey[10] = { 0xD181C62C, 0xF7F4DB5B, 0x1983A2FC,
 853                                 0x943E1ADB, 0xD9389E6B, 0xD1039C2C, 0xA74499AD,
 854                                 0x593D56D9, 0xF3253C06, 0x2ADC1FFC};
 855
 856         en_dbg(DRV, priv, "Configuring rss steering\n");
 857         err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
 858                                     priv->rx_ring_num,
 859                                     &rss_map->base_qpn);
 860         if (err) {
 861                 en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
 862                 return err;
 863         }
 864
 865         for (i = 0; i < priv->rx_ring_num; i++) {
 866                 qpn = rss_map->base_qpn + i;
 867                 err = mlx4_en_config_rss_qp(priv, qpn, &priv->rx_ring[i],
 868                                             &rss_map->state[i],
 869                                             &rss_map->qps[i]);
 870                 if (err)
 871                         goto rss_err;
 872
 873                 ++good_qps;
 874         }
 875
 876         /* Configure RSS indirection qp */
 877         err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
 878         if (err) {
 879                 en_err(priv, "Failed to allocate RSS indirection QP\n");
 880                 goto rss_err;
 881         }
 882         rss_map->indir_qp.event = mlx4_en_sqp_event;
 883         mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
 884                                 priv->rx_ring[0].cqn, &context);
 885
 886         ptr = ((void *) &context) + 0x3c;
 887         rss_context = ptr;
 888         rss_context->base_qpn = cpu_to_be32(ilog2(priv->rx_ring_num) << 24 |
 889                                             (rss_map->base_qpn));
 890         rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
 891         rss_context->flags = rss_mask;
 892         rss_context->hash_fn = 1;
 893         for (i = 0; i < 10; i++)
 894                 rss_context->rss_key[i] = rsskey[i];
 895
 896         if (priv->mdev->profile.udp_rss)
 897                 rss_context->base_qpn_udp = rss_context->default_qpn;
 898         err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
 899                                &rss_map->indir_qp, &rss_map->indir_state);
 900         if (err)
 901                 goto indir_err;
 902
 903         return 0;
 904
 905 indir_err:
 906         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
 907                        MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
 908         mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
 909         mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
 910 rss_err:
 911         for (i = 0; i < good_qps; i++) {
 912                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
 913                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
 914                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
 915                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
 916         }
 917         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
 918         return err;
 919 }
 920
 921 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
 922 {
 923         struct mlx4_en_dev *mdev = priv->mdev;
 924         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
 925         int i;
 926
 927         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
 928                        MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
 929         mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
 930         mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
 931
 932         for (i = 0; i < priv->rx_ring_num; i++) {
 933                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
 934                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
 935                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
 936                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
 937         }
 938         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
 939 }
 940
 941
 942
 943
 944