drivers/net/ethernet/mellanox/mlx4/en_tx.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 <asm/page.h>
  35 #include <linux/mlx4/cq.h>
  36 #include <linux/slab.h>
  37 #include <linux/mlx4/qp.h>
  38 #include <linux/skbuff.h>
  39 #include <linux/if_vlan.h>
  40 #include <linux/vmalloc.h>
  41 #include <linux/tcp.h>
  42 #include <linux/moduleparam.h>
  43
  44 #include "mlx4_en.h"
  45
  46 enum {
  47         MAX_INLINE = 104, /* 128 - 16 - 4 - 4 */
  48         MAX_BF = 256,
  49 };
  50
  51 static int inline_thold __read_mostly = MAX_INLINE;
  52
  53 module_param_named(inline_thold, inline_thold, int, 0444);
  54 MODULE_PARM_DESC(inline_thold, "threshold for using inline data");
  55
  56 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
  57                            struct mlx4_en_tx_ring *ring, int qpn, u32 size,
  58                            u16 stride)
  59 {
  60         struct mlx4_en_dev *mdev = priv->mdev;
  61         int tmp;
  62         int err;
  63
  64         ring->size = size;
  65         ring->size_mask = size - 1;
  66         ring->stride = stride;
  67
  68         inline_thold = min(inline_thold, MAX_INLINE);
  69
  70         tmp = size * sizeof(struct mlx4_en_tx_info);
  71         ring->tx_info = vmalloc(tmp);
  72         if (!ring->tx_info)
  73                 return -ENOMEM;
  74
  75         en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
  76                  ring->tx_info, tmp);
  77
  78         ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
  79         if (!ring->bounce_buf) {
  80                 err = -ENOMEM;
  81                 goto err_tx;
  82         }
  83         ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
  84
  85         err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
  86                                  2 * PAGE_SIZE);
  87         if (err) {
  88                 en_err(priv, "Failed allocating hwq resources\n");
  89                 goto err_bounce;
  90         }
  91
  92         err = mlx4_en_map_buffer(&ring->wqres.buf);
  93         if (err) {
  94                 en_err(priv, "Failed to map TX buffer\n");
  95                 goto err_hwq_res;
  96         }
  97
  98         ring->buf = ring->wqres.buf.direct.buf;
  99
 100         en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
 101                "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
 102                ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
 103
 104         ring->qpn = qpn;
 105         err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp);
 106         if (err) {
 107                 en_err(priv, "Failed allocating qp %d\n", ring->qpn);
 108                 goto err_map;
 109         }
 110         ring->qp.event = mlx4_en_sqp_event;
 111
 112         err = mlx4_bf_alloc(mdev->dev, &ring->bf);
 113         if (err) {
 114                 en_dbg(DRV, priv, "working without blueflame (%d)", err);
 115                 ring->bf.uar = &mdev->priv_uar;
 116                 ring->bf.uar->map = mdev->uar_map;
 117                 ring->bf_enabled = false;
 118         } else
 119                 ring->bf_enabled = true;
 120
 121         ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
 122
 123         return 0;
 124
 125 err_map:
 126         mlx4_en_unmap_buffer(&ring->wqres.buf);
 127 err_hwq_res:
 128         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
 129 err_bounce:
 130         kfree(ring->bounce_buf);
 131         ring->bounce_buf = NULL;
 132 err_tx:
 133         vfree(ring->tx_info);
 134         ring->tx_info = NULL;
 135         return err;
 136 }
 137
 138 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
 139                              struct mlx4_en_tx_ring *ring)
 140 {
 141         struct mlx4_en_dev *mdev = priv->mdev;
 142         en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
 143
 144         if (ring->bf_enabled)
 145                 mlx4_bf_free(mdev->dev, &ring->bf);
 146         mlx4_qp_remove(mdev->dev, &ring->qp);
 147         mlx4_qp_free(mdev->dev, &ring->qp);
 148         mlx4_en_unmap_buffer(&ring->wqres.buf);
 149         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
 150         kfree(ring->bounce_buf);
 151         ring->bounce_buf = NULL;
 152         vfree(ring->tx_info);
 153         ring->tx_info = NULL;
 154 }
 155
 156 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
 157                              struct mlx4_en_tx_ring *ring,
 158                              int cq, int user_prio)
 159 {
 160         struct mlx4_en_dev *mdev = priv->mdev;
 161         int err;
 162
 163         ring->cqn = cq;
 164         ring->prod = 0;
 165         ring->cons = 0xffffffff;
 166         ring->last_nr_txbb = 1;
 167         ring->poll_cnt = 0;
 168         memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
 169         memset(ring->buf, 0, ring->buf_size);
 170
 171         ring->qp_state = MLX4_QP_STATE_RST;
 172         ring->doorbell_qpn = ring->qp.qpn << 8;
 173
 174         mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
 175                                 ring->cqn, user_prio, &ring->context);
 176         if (ring->bf_enabled)
 177                 ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
 178
 179         err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
 180                                &ring->qp, &ring->qp_state);
 181
 182         return err;
 183 }
 184
 185 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
 186                                 struct mlx4_en_tx_ring *ring)
 187 {
 188         struct mlx4_en_dev *mdev = priv->mdev;
 189
 190         mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
 191                        MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
 192 }
 193
 194 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
 195                               struct mlx4_en_tx_ring *ring, int index,
 196                               u8 owner)
 197 {
 198         __be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
 199         struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
 200         struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
 201         void *end = ring->buf + ring->buf_size;
 202         __be32 *ptr = (__be32 *)tx_desc;
 203         int i;
 204
 205         /* Optimize the common case when there are no wraparounds */
 206         if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
 207                 /* Stamp the freed descriptor */
 208                 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
 209                      i += STAMP_STRIDE) {
 210                         *ptr = stamp;
 211                         ptr += STAMP_DWORDS;
 212                 }
 213         } else {
 214                 /* Stamp the freed descriptor */
 215                 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
 216                      i += STAMP_STRIDE) {
 217                         *ptr = stamp;
 218                         ptr += STAMP_DWORDS;
 219                         if ((void *)ptr >= end) {
 220                                 ptr = ring->buf;
 221                                 stamp ^= cpu_to_be32(0x80000000);
 222                         }
 223                 }
 224         }
 225 }
 226
 227
 228 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
 229                                 struct mlx4_en_tx_ring *ring,
 230                                 int index, u8 owner, u64 timestamp)
 231 {
 232         struct mlx4_en_dev *mdev = priv->mdev;
 233         struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
 234         struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
 235         struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
 236         struct sk_buff *skb = tx_info->skb;
 237         struct skb_frag_struct *frag;
 238         void *end = ring->buf + ring->buf_size;
 239         int frags = skb_shinfo(skb)->nr_frags;
 240         int i;
 241         struct skb_shared_hwtstamps hwts;
 242
 243         if (timestamp) {
 244                 mlx4_en_fill_hwtstamps(mdev, &hwts, timestamp);
 245                 skb_tstamp_tx(skb, &hwts);
 246         }
 247
 248         /* Optimize the common case when there are no wraparounds */
 249         if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
 250                 if (!tx_info->inl) {
 251                         if (tx_info->linear) {
 252                                 dma_unmap_single(priv->ddev,
 253                                         (dma_addr_t) be64_to_cpu(data->addr),
 254                                          be32_to_cpu(data->byte_count),
 255                                          PCI_DMA_TODEVICE);
 256                                 ++data;
 257                         }
 258
 259                         for (i = 0; i < frags; i++) {
 260                                 frag = &skb_shinfo(skb)->frags[i];
 261                                 dma_unmap_page(priv->ddev,
 262                                         (dma_addr_t) be64_to_cpu(data[i].addr),
 263                                         skb_frag_size(frag), PCI_DMA_TODEVICE);
 264                         }
 265                 }
 266         } else {
 267                 if (!tx_info->inl) {
 268                         if ((void *) data >= end) {
 269                                 data = ring->buf + ((void *)data - end);
 270                         }
 271
 272                         if (tx_info->linear) {
 273                                 dma_unmap_single(priv->ddev,
 274                                         (dma_addr_t) be64_to_cpu(data->addr),
 275                                          be32_to_cpu(data->byte_count),
 276                                          PCI_DMA_TODEVICE);
 277                                 ++data;
 278                         }
 279
 280                         for (i = 0; i < frags; i++) {
 281                                 /* Check for wraparound before unmapping */
 282                                 if ((void *) data >= end)
 283                                         data = ring->buf;
 284                                 frag = &skb_shinfo(skb)->frags[i];
 285                                 dma_unmap_page(priv->ddev,
 286                                         (dma_addr_t) be64_to_cpu(data->addr),
 287                                          skb_frag_size(frag), PCI_DMA_TODEVICE);
 288                                 ++data;
 289                         }
 290                 }
 291         }
 292         dev_kfree_skb_any(skb);
 293         return tx_info->nr_txbb;
 294 }
 295
 296
 297 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
 298 {
 299         struct mlx4_en_priv *priv = netdev_priv(dev);
 300         int cnt = 0;
 301
 302         /* Skip last polled descriptor */
 303         ring->cons += ring->last_nr_txbb;
 304         en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
 305                  ring->cons, ring->prod);
 306
 307         if ((u32) (ring->prod - ring->cons) > ring->size) {
 308                 if (netif_msg_tx_err(priv))
 309                         en_warn(priv, "Tx consumer passed producer!\n");
 310                 return 0;
 311         }
 312
 313         while (ring->cons != ring->prod) {
 314                 ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
 315                                                 ring->cons & ring->size_mask,
 316                                                 !!(ring->cons & ring->size), 0);
 317                 ring->cons += ring->last_nr_txbb;
 318                 cnt++;
 319         }
 320
 321         netdev_tx_reset_queue(ring->tx_queue);
 322
 323         if (cnt)
 324                 en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
 325
 326         return cnt;
 327 }
 328
 329 static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
 330 {
 331         struct mlx4_en_priv *priv = netdev_priv(dev);
 332         struct mlx4_cq *mcq = &cq->mcq;
 333         struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
 334         struct mlx4_cqe *cqe;
 335         u16 index;
 336         u16 new_index, ring_index, stamp_index;
 337         u32 txbbs_skipped = 0;
 338         u32 txbbs_stamp = 0;
 339         u32 cons_index = mcq->cons_index;
 340         int size = cq->size;
 341         u32 size_mask = ring->size_mask;
 342         struct mlx4_cqe *buf = cq->buf;
 343         u32 packets = 0;
 344         u32 bytes = 0;
 345         int factor = priv->cqe_factor;
 346         u64 timestamp = 0;
 347
 348         if (!priv->port_up)
 349                 return;
 350
 351         index = cons_index & size_mask;
 352         cqe = &buf[(index << factor) + factor];
 353         ring_index = ring->cons & size_mask;
 354         stamp_index = ring_index;
 355
 356         /* Process all completed CQEs */
 357         while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
 358                         cons_index & size)) {
 359                 /*
 360                  * make sure we read the CQE after we read the
 361                  * ownership bit
 362                  */
 363                 rmb();
 364
 365                 /* Skip over last polled CQE */
 366                 new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
 367
 368                 do {
 369                         txbbs_skipped += ring->last_nr_txbb;
 370                         ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
 371                         if (ring->tx_info[ring_index].ts_requested)
 372                                 timestamp = mlx4_en_get_cqe_ts(cqe);
 373
 374                         /* free next descriptor */
 375                         ring->last_nr_txbb = mlx4_en_free_tx_desc(
 376                                         priv, ring, ring_index,
 377                                         !!((ring->cons + txbbs_skipped) &
 378                                         ring->size), timestamp);
 379
 380                         mlx4_en_stamp_wqe(priv, ring, stamp_index,
 381                                           !!((ring->cons + txbbs_stamp) &
 382                                                 ring->size));
 383                         stamp_index = ring_index;
 384                         txbbs_stamp = txbbs_skipped;
 385                         packets++;
 386                         bytes += ring->tx_info[ring_index].nr_bytes;
 387                 } while (ring_index != new_index);
 388
 389                 ++cons_index;
 390                 index = cons_index & size_mask;
 391                 cqe = &buf[(index << factor) + factor];
 392         }
 393
 394
 395         /*
 396          * To prevent CQ overflow we first update CQ consumer and only then
 397          * the ring consumer.
 398          */
 399         mcq->cons_index = cons_index;
 400         mlx4_cq_set_ci(mcq);
 401         wmb();
 402         ring->cons += txbbs_skipped;
 403         netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
 404
 405         /*
 406          * Wakeup Tx queue if this stopped, and at least 1 packet
 407          * was completed
 408          */
 409         if (netif_tx_queue_stopped(ring->tx_queue) && txbbs_skipped > 0) {
 410                 netif_tx_wake_queue(ring->tx_queue);
 411                 priv->port_stats.wake_queue++;
 412         }
 413 }
 414
 415 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
 416 {
 417         struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
 418         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
 419
 420         mlx4_en_process_tx_cq(cq->dev, cq);
 421         mlx4_en_arm_cq(priv, cq);
 422 }
 423
 424
 425 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
 426                                                       struct mlx4_en_tx_ring *ring,
 427                                                       u32 index,
 428                                                       unsigned int desc_size)
 429 {
 430         u32 copy = (ring->size - index) * TXBB_SIZE;
 431         int i;
 432
 433         for (i = desc_size - copy - 4; i >= 0; i -= 4) {
 434                 if ((i & (TXBB_SIZE - 1)) == 0)
 435                         wmb();
 436
 437                 *((u32 *) (ring->buf + i)) =
 438                         *((u32 *) (ring->bounce_buf + copy + i));
 439         }
 440
 441         for (i = copy - 4; i >= 4 ; i -= 4) {
 442                 if ((i & (TXBB_SIZE - 1)) == 0)
 443                         wmb();
 444
 445                 *((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
 446                         *((u32 *) (ring->bounce_buf + i));
 447         }
 448
 449         /* Return real descriptor location */
 450         return ring->buf + index * TXBB_SIZE;
 451 }
 452
 453 static int is_inline(struct sk_buff *skb, void **pfrag)
 454 {
 455         void *ptr;
 456
 457         if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
 458                 if (skb_shinfo(skb)->nr_frags == 1) {
 459                         ptr = skb_frag_address_safe(&skb_shinfo(skb)->frags[0]);
 460                         if (unlikely(!ptr))
 461                                 return 0;
 462
 463                         if (pfrag)
 464                                 *pfrag = ptr;
 465
 466                         return 1;
 467                 } else if (unlikely(skb_shinfo(skb)->nr_frags))
 468                         return 0;
 469                 else
 470                         return 1;
 471         }
 472
 473         return 0;
 474 }
 475
 476 static int inline_size(struct sk_buff *skb)
 477 {
 478         if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
 479             <= MLX4_INLINE_ALIGN)
 480                 return ALIGN(skb->len + CTRL_SIZE +
 481                              sizeof(struct mlx4_wqe_inline_seg), 16);
 482         else
 483                 return ALIGN(skb->len + CTRL_SIZE + 2 *
 484                              sizeof(struct mlx4_wqe_inline_seg), 16);
 485 }
 486
 487 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
 488                          int *lso_header_size)
 489 {
 490         struct mlx4_en_priv *priv = netdev_priv(dev);
 491         int real_size;
 492
 493         if (skb_is_gso(skb)) {
 494                 *lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
 495                 real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
 496                         ALIGN(*lso_header_size + 4, DS_SIZE);
 497                 if (unlikely(*lso_header_size != skb_headlen(skb))) {
 498                         /* We add a segment for the skb linear buffer only if
 499                          * it contains data */
 500                         if (*lso_header_size < skb_headlen(skb))
 501                                 real_size += DS_SIZE;
 502                         else {
 503                                 if (netif_msg_tx_err(priv))
 504                                         en_warn(priv, "Non-linear headers\n");
 505                                 return 0;
 506                         }
 507                 }
 508         } else {
 509                 *lso_header_size = 0;
 510                 if (!is_inline(skb, NULL))
 511                         real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
 512                 else
 513                         real_size = inline_size(skb);
 514         }
 515
 516         return real_size;
 517 }
 518
 519 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
 520                              int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
 521 {
 522         struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
 523         int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
 524
 525         if (skb->len <= spc) {
 526                 inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
 527                 skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
 528                 if (skb_shinfo(skb)->nr_frags)
 529                         memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
 530                                skb_frag_size(&skb_shinfo(skb)->frags[0]));
 531
 532         } else {
 533                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
 534                 if (skb_headlen(skb) <= spc) {
 535                         skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
 536                         if (skb_headlen(skb) < spc) {
 537                                 memcpy(((void *)(inl + 1)) + skb_headlen(skb),
 538                                         fragptr, spc - skb_headlen(skb));
 539                                 fragptr +=  spc - skb_headlen(skb);
 540                         }
 541                         inl = (void *) (inl + 1) + spc;
 542                         memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
 543                 } else {
 544                         skb_copy_from_linear_data(skb, inl + 1, spc);
 545                         inl = (void *) (inl + 1) + spc;
 546                         skb_copy_from_linear_data_offset(skb, spc, inl + 1,
 547                                         skb_headlen(skb) - spc);
 548                         if (skb_shinfo(skb)->nr_frags)
 549                                 memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
 550                                         fragptr, skb_frag_size(&skb_shinfo(skb)->frags[0]));
 551                 }
 552
 553                 wmb();
 554                 inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
 555         }
 556 }
 557
 558 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
 559 {
 560         struct mlx4_en_priv *priv = netdev_priv(dev);
 561         u16 rings_p_up = priv->num_tx_rings_p_up;
 562         u8 up = 0;
 563
 564         if (dev->num_tc)
 565                 return skb_tx_hash(dev, skb);
 566
 567         if (vlan_tx_tag_present(skb))
 568                 up = vlan_tx_tag_get(skb) >> VLAN_PRIO_SHIFT;
 569
 570         return __skb_tx_hash(dev, skb, rings_p_up) + up * rings_p_up;
 571 }
 572
 573 static void mlx4_bf_copy(void __iomem *dst, unsigned long *src, unsigned bytecnt)
 574 {
 575         __iowrite64_copy(dst, src, bytecnt / 8);
 576 }
 577
 578 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
 579 {
 580         struct mlx4_en_priv *priv = netdev_priv(dev);
 581         struct mlx4_en_dev *mdev = priv->mdev;
 582         struct mlx4_en_tx_ring *ring;
 583         struct mlx4_en_tx_desc *tx_desc;
 584         struct mlx4_wqe_data_seg *data;
 585         struct skb_frag_struct *frag;
 586         struct mlx4_en_tx_info *tx_info;
 587         struct ethhdr *ethh;
 588         int tx_ind = 0;
 589         int nr_txbb;
 590         int desc_size;
 591         int real_size;
 592         dma_addr_t dma;
 593         u32 index, bf_index;
 594         __be32 op_own;
 595         u16 vlan_tag = 0;
 596         int i;
 597         int lso_header_size;
 598         void *fragptr;
 599         bool bounce = false;
 600
 601         if (!priv->port_up)
 602                 goto tx_drop;
 603
 604         real_size = get_real_size(skb, dev, &lso_header_size);
 605         if (unlikely(!real_size))
 606                 goto tx_drop;
 607
 608         /* Align descriptor to TXBB size */
 609         desc_size = ALIGN(real_size, TXBB_SIZE);
 610         nr_txbb = desc_size / TXBB_SIZE;
 611         if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
 612                 if (netif_msg_tx_err(priv))
 613                         en_warn(priv, "Oversized header or SG list\n");
 614                 goto tx_drop;
 615         }
 616
 617         tx_ind = skb->queue_mapping;
 618         ring = &priv->tx_ring[tx_ind];
 619         if (vlan_tx_tag_present(skb))
 620                 vlan_tag = vlan_tx_tag_get(skb);
 621
 622         /* Check available TXBBs And 2K spare for prefetch */
 623         if (unlikely(((int)(ring->prod - ring->cons)) >
 624                      ring->size - HEADROOM - MAX_DESC_TXBBS)) {
 625                 /* every full Tx ring stops queue */
 626                 netif_tx_stop_queue(ring->tx_queue);
 627                 priv->port_stats.queue_stopped++;
 628
 629                 /* If queue was emptied after the if, and before the
 630                  * stop_queue - need to wake the queue, or else it will remain
 631                  * stopped forever.
 632                  * Need a memory barrier to make sure ring->cons was not
 633                  * updated before queue was stopped.
 634                  */
 635                 wmb();
 636
 637                 if (unlikely(((int)(ring->prod - ring->cons)) <=
 638                              ring->size - HEADROOM - MAX_DESC_TXBBS)) {
 639                         netif_tx_wake_queue(ring->tx_queue);
 640                         priv->port_stats.wake_queue++;
 641                 } else {
 642                         return NETDEV_TX_BUSY;
 643                 }
 644         }
 645
 646         /* Track current inflight packets for performance analysis */
 647         AVG_PERF_COUNTER(priv->pstats.inflight_avg,
 648                          (u32) (ring->prod - ring->cons - 1));
 649
 650         /* Packet is good - grab an index and transmit it */
 651         index = ring->prod & ring->size_mask;
 652         bf_index = ring->prod;
 653
 654         /* See if we have enough space for whole descriptor TXBB for setting
 655          * SW ownership on next descriptor; if not, use a bounce buffer. */
 656         if (likely(index + nr_txbb <= ring->size))
 657                 tx_desc = ring->buf + index * TXBB_SIZE;
 658         else {
 659                 tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
 660                 bounce = true;
 661         }
 662
 663         /* Save skb in tx_info ring */
 664         tx_info = &ring->tx_info[index];
 665         tx_info->skb = skb;
 666         tx_info->nr_txbb = nr_txbb;
 667
 668         /*
 669          * For timestamping add flag to skb_shinfo and
 670          * set flag for further reference
 671          */
 672         if (ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
 673             skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
 674                 skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
 675                 tx_info->ts_requested = 1;
 676         }
 677
 678         /* Prepare ctrl segement apart opcode+ownership, which depends on
 679          * whether LSO is used */
 680         tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
 681         tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
 682                 !!vlan_tx_tag_present(skb);
 683         tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
 684         tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
 685         if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
 686                 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
 687                                                          MLX4_WQE_CTRL_TCP_UDP_CSUM);
 688                 ring->tx_csum++;
 689         }
 690
 691         if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
 692                 /* Copy dst mac address to wqe. This allows loopback in eSwitch,
 693                  * so that VFs and PF can communicate with each other
 694                  */
 695                 ethh = (struct ethhdr *)skb->data;
 696                 tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
 697                 tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
 698         }
 699
 700         /* Handle LSO (TSO) packets */
 701         if (lso_header_size) {
 702                 /* Mark opcode as LSO */
 703                 op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
 704                         ((ring->prod & ring->size) ?
 705                                 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
 706
 707                 /* Fill in the LSO prefix */
 708                 tx_desc->lso.mss_hdr_size = cpu_to_be32(
 709                         skb_shinfo(skb)->gso_size << 16 | lso_header_size);
 710
 711                 /* Copy headers;
 712                  * note that we already verified that it is linear */
 713                 memcpy(tx_desc->lso.header, skb->data, lso_header_size);
 714                 data = ((void *) &tx_desc->lso +
 715                         ALIGN(lso_header_size + 4, DS_SIZE));
 716
 717                 priv->port_stats.tso_packets++;
 718                 i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
 719                         !!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
 720                 tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
 721                 ring->packets += i;
 722         } else {
 723                 /* Normal (Non LSO) packet */
 724                 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
 725                         ((ring->prod & ring->size) ?
 726                          cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
 727                 data = &tx_desc->data;
 728                 tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
 729                 ring->packets++;
 730
 731         }
 732         ring->bytes += tx_info->nr_bytes;
 733         netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
 734         AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
 735
 736
 737         /* valid only for none inline segments */
 738         tx_info->data_offset = (void *) data - (void *) tx_desc;
 739
 740         tx_info->linear = (lso_header_size < skb_headlen(skb) && !is_inline(skb, NULL)) ? 1 : 0;
 741         data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
 742
 743         if (!is_inline(skb, &fragptr)) {
 744                 /* Map fragments */
 745                 for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
 746                         frag = &skb_shinfo(skb)->frags[i];
 747                         dma = skb_frag_dma_map(priv->ddev, frag,
 748                                                0, skb_frag_size(frag),
 749                                                DMA_TO_DEVICE);
 750                         data->addr = cpu_to_be64(dma);
 751                         data->lkey = cpu_to_be32(mdev->mr.key);
 752                         wmb();
 753                         data->byte_count = cpu_to_be32(skb_frag_size(frag));
 754                         --data;
 755                 }
 756
 757                 /* Map linear part */
 758                 if (tx_info->linear) {
 759                         dma = dma_map_single(priv->ddev, skb->data + lso_header_size,
 760                                              skb_headlen(skb) - lso_header_size, PCI_DMA_TODEVICE);
 761                         data->addr = cpu_to_be64(dma);
 762                         data->lkey = cpu_to_be32(mdev->mr.key);
 763                         wmb();
 764                         data->byte_count = cpu_to_be32(skb_headlen(skb) - lso_header_size);
 765                 }
 766                 tx_info->inl = 0;
 767         } else {
 768                 build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
 769                 tx_info->inl = 1;
 770         }
 771
 772         ring->prod += nr_txbb;
 773
 774         /* If we used a bounce buffer then copy descriptor back into place */
 775         if (bounce)
 776                 tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
 777
 778         skb_tx_timestamp(skb);
 779
 780         if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tx_tag_present(skb)) {
 781                 *(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
 782                 op_own |= htonl((bf_index & 0xffff) << 8);
 783                 /* Ensure new descirptor hits memory
 784                 * before setting ownership of this descriptor to HW */
 785                 wmb();
 786                 tx_desc->ctrl.owner_opcode = op_own;
 787
 788                 wmb();
 789
 790                 mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned long *) &tx_desc->ctrl,
 791                      desc_size);
 792
 793                 wmb();
 794
 795                 ring->bf.offset ^= ring->bf.buf_size;
 796         } else {
 797                 /* Ensure new descirptor hits memory
 798                 * before setting ownership of this descriptor to HW */
 799                 wmb();
 800                 tx_desc->ctrl.owner_opcode = op_own;
 801                 wmb();
 802                 iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
 803         }
 804
 805         return NETDEV_TX_OK;
 806
 807 tx_drop:
 808         dev_kfree_skb_any(skb);
 809         priv->stats.tx_dropped++;
 810         return NETDEV_TX_OK;
 811 }
 812