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

   1 /*
   2  * Copyright (c) 2015-2016, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <net/tc_act/tc_gact.h>
  34 #include <net/pkt_cls.h>
  35 #include <linux/mlx5/fs.h>
  36 #include <net/vxlan.h>
  37 #include "en.h"
  38 #include "en_tc.h"
  39 #include "eswitch.h"
  40 #include "vxlan.h"
  41
  42 struct mlx5e_rq_param {
  43         u32                        rqc[MLX5_ST_SZ_DW(rqc)];
  44         struct mlx5_wq_param       wq;
  45 };
  46
  47 struct mlx5e_sq_param {
  48         u32                        sqc[MLX5_ST_SZ_DW(sqc)];
  49         struct mlx5_wq_param       wq;
  50         u16                        max_inline;
  51         bool                       icosq;
  52 };
  53
  54 struct mlx5e_cq_param {
  55         u32                        cqc[MLX5_ST_SZ_DW(cqc)];
  56         struct mlx5_wq_param       wq;
  57         u16                        eq_ix;
  58 };
  59
  60 struct mlx5e_channel_param {
  61         struct mlx5e_rq_param      rq;
  62         struct mlx5e_sq_param      sq;
  63         struct mlx5e_sq_param      icosq;
  64         struct mlx5e_cq_param      rx_cq;
  65         struct mlx5e_cq_param      tx_cq;
  66         struct mlx5e_cq_param      icosq_cq;
  67 };
  68
  69 static void mlx5e_update_carrier(struct mlx5e_priv *priv)
  70 {
  71         struct mlx5_core_dev *mdev = priv->mdev;
  72         u8 port_state;
  73
  74         port_state = mlx5_query_vport_state(mdev,
  75                 MLX5_QUERY_VPORT_STATE_IN_OP_MOD_VNIC_VPORT, 0);
  76
  77         if (port_state == VPORT_STATE_UP)
  78                 netif_carrier_on(priv->netdev);
  79         else
  80                 netif_carrier_off(priv->netdev);
  81 }
  82
  83 static void mlx5e_update_carrier_work(struct work_struct *work)
  84 {
  85         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
  86                                                update_carrier_work);
  87
  88         mutex_lock(&priv->state_lock);
  89         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
  90                 mlx5e_update_carrier(priv);
  91         mutex_unlock(&priv->state_lock);
  92 }
  93
  94 static void mlx5e_update_sw_counters(struct mlx5e_priv *priv)
  95 {
  96         struct mlx5e_sw_stats *s = &priv->stats.sw;
  97         struct mlx5e_rq_stats *rq_stats;
  98         struct mlx5e_sq_stats *sq_stats;
  99         u64 tx_offload_none = 0;
 100         int i, j;
 101
 102         memset(s, 0, sizeof(*s));
 103         for (i = 0; i < priv->params.num_channels; i++) {
 104                 rq_stats = &priv->channel[i]->rq.stats;
 105
 106                 s->rx_packets   += rq_stats->packets;
 107                 s->rx_bytes     += rq_stats->bytes;
 108                 s->lro_packets  += rq_stats->lro_packets;
 109                 s->lro_bytes    += rq_stats->lro_bytes;
 110                 s->rx_csum_none += rq_stats->csum_none;
 111                 s->rx_csum_sw   += rq_stats->csum_sw;
 112                 s->rx_wqe_err   += rq_stats->wqe_err;
 113                 s->rx_mpwqe_filler += rq_stats->mpwqe_filler;
 114                 s->rx_mpwqe_frag   += rq_stats->mpwqe_frag;
 115                 s->rx_buff_alloc_err += rq_stats->buff_alloc_err;
 116
 117                 for (j = 0; j < priv->params.num_tc; j++) {
 118                         sq_stats = &priv->channel[i]->sq[j].stats;
 119
 120                         s->tx_packets           += sq_stats->packets;
 121                         s->tx_bytes             += sq_stats->bytes;
 122                         s->tso_packets          += sq_stats->tso_packets;
 123                         s->tso_bytes            += sq_stats->tso_bytes;
 124                         s->tso_inner_packets    += sq_stats->tso_inner_packets;
 125                         s->tso_inner_bytes      += sq_stats->tso_inner_bytes;
 126                         s->tx_queue_stopped     += sq_stats->stopped;
 127                         s->tx_queue_wake        += sq_stats->wake;
 128                         s->tx_queue_dropped     += sq_stats->dropped;
 129                         s->tx_csum_inner        += sq_stats->csum_offload_inner;
 130                         tx_offload_none         += sq_stats->csum_offload_none;
 131                 }
 132         }
 133
 134         /* Update calculated offload counters */
 135         s->tx_csum_offload = s->tx_packets - tx_offload_none - s->tx_csum_inner;
 136         s->rx_csum_good    = s->rx_packets - s->rx_csum_none -
 137                              s->rx_csum_sw;
 138
 139         s->link_down_events = MLX5_GET(ppcnt_reg,
 140                                 priv->stats.pport.phy_counters,
 141                                 counter_set.phys_layer_cntrs.link_down_events);
 142 }
 143
 144 static void mlx5e_update_vport_counters(struct mlx5e_priv *priv)
 145 {
 146         int outlen = MLX5_ST_SZ_BYTES(query_vport_counter_out);
 147         u32 *out = (u32 *)priv->stats.vport.query_vport_out;
 148         u32 in[MLX5_ST_SZ_DW(query_vport_counter_in)];
 149         struct mlx5_core_dev *mdev = priv->mdev;
 150
 151         memset(in, 0, sizeof(in));
 152
 153         MLX5_SET(query_vport_counter_in, in, opcode,
 154                  MLX5_CMD_OP_QUERY_VPORT_COUNTER);
 155         MLX5_SET(query_vport_counter_in, in, op_mod, 0);
 156         MLX5_SET(query_vport_counter_in, in, other_vport, 0);
 157
 158         memset(out, 0, outlen);
 159
 160         mlx5_cmd_exec(mdev, in, sizeof(in), out, outlen);
 161 }
 162
 163 static void mlx5e_update_pport_counters(struct mlx5e_priv *priv)
 164 {
 165         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
 166         struct mlx5_core_dev *mdev = priv->mdev;
 167         int sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
 168         int prio;
 169         void *out;
 170         u32 *in;
 171
 172         in = mlx5_vzalloc(sz);
 173         if (!in)
 174                 goto free_out;
 175
 176         MLX5_SET(ppcnt_reg, in, local_port, 1);
 177
 178         out = pstats->IEEE_802_3_counters;
 179         MLX5_SET(ppcnt_reg, in, grp, MLX5_IEEE_802_3_COUNTERS_GROUP);
 180         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 181
 182         out = pstats->RFC_2863_counters;
 183         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2863_COUNTERS_GROUP);
 184         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 185
 186         out = pstats->RFC_2819_counters;
 187         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2819_COUNTERS_GROUP);
 188         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 189
 190         out = pstats->phy_counters;
 191         MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_COUNTERS_GROUP);
 192         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 193
 194         MLX5_SET(ppcnt_reg, in, grp, MLX5_PER_PRIORITY_COUNTERS_GROUP);
 195         for (prio = 0; prio < NUM_PPORT_PRIO; prio++) {
 196                 out = pstats->per_prio_counters[prio];
 197                 MLX5_SET(ppcnt_reg, in, prio_tc, prio);
 198                 mlx5_core_access_reg(mdev, in, sz, out, sz,
 199                                      MLX5_REG_PPCNT, 0, 0);
 200         }
 201
 202 free_out:
 203         kvfree(in);
 204 }
 205
 206 static void mlx5e_update_q_counter(struct mlx5e_priv *priv)
 207 {
 208         struct mlx5e_qcounter_stats *qcnt = &priv->stats.qcnt;
 209
 210         if (!priv->q_counter)
 211                 return;
 212
 213         mlx5_core_query_out_of_buffer(priv->mdev, priv->q_counter,
 214                                       &qcnt->rx_out_of_buffer);
 215 }
 216
 217 void mlx5e_update_stats(struct mlx5e_priv *priv)
 218 {
 219         mlx5e_update_q_counter(priv);
 220         mlx5e_update_vport_counters(priv);
 221         mlx5e_update_pport_counters(priv);
 222         mlx5e_update_sw_counters(priv);
 223 }
 224
 225 static void mlx5e_update_stats_work(struct work_struct *work)
 226 {
 227         struct delayed_work *dwork = to_delayed_work(work);
 228         struct mlx5e_priv *priv = container_of(dwork, struct mlx5e_priv,
 229                                                update_stats_work);
 230         mutex_lock(&priv->state_lock);
 231         if (test_bit(MLX5E_STATE_OPENED, &priv->state)) {
 232                 mlx5e_update_stats(priv);
 233                 schedule_delayed_work(dwork,
 234                                       msecs_to_jiffies(
 235                                               MLX5E_UPDATE_STATS_INTERVAL));
 236         }
 237         mutex_unlock(&priv->state_lock);
 238 }
 239
 240 static void mlx5e_async_event(struct mlx5_core_dev *mdev, void *vpriv,
 241                               enum mlx5_dev_event event, unsigned long param)
 242 {
 243         struct mlx5e_priv *priv = vpriv;
 244
 245         if (!test_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state))
 246                 return;
 247
 248         switch (event) {
 249         case MLX5_DEV_EVENT_PORT_UP:
 250         case MLX5_DEV_EVENT_PORT_DOWN:
 251                 schedule_work(&priv->update_carrier_work);
 252                 break;
 253
 254         default:
 255                 break;
 256         }
 257 }
 258
 259 static void mlx5e_enable_async_events(struct mlx5e_priv *priv)
 260 {
 261         set_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state);
 262 }
 263
 264 static void mlx5e_disable_async_events(struct mlx5e_priv *priv)
 265 {
 266         clear_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state);
 267         synchronize_irq(mlx5_get_msix_vec(priv->mdev, MLX5_EQ_VEC_ASYNC));
 268 }
 269
 270 #define MLX5E_HW2SW_MTU(hwmtu) (hwmtu - (ETH_HLEN + VLAN_HLEN + ETH_FCS_LEN))
 271 #define MLX5E_SW2HW_MTU(swmtu) (swmtu + (ETH_HLEN + VLAN_HLEN + ETH_FCS_LEN))
 272
 273 static int mlx5e_create_rq(struct mlx5e_channel *c,
 274                            struct mlx5e_rq_param *param,
 275                            struct mlx5e_rq *rq)
 276 {
 277         struct mlx5e_priv *priv = c->priv;
 278         struct mlx5_core_dev *mdev = priv->mdev;
 279         void *rqc = param->rqc;
 280         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 281         u32 byte_count;
 282         int wq_sz;
 283         int err;
 284         int i;
 285
 286         param->wq.db_numa_node = cpu_to_node(c->cpu);
 287
 288         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
 289                                 &rq->wq_ctrl);
 290         if (err)
 291                 return err;
 292
 293         rq->wq.db = &rq->wq.db[MLX5_RCV_DBR];
 294
 295         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 296
 297         switch (priv->params.rq_wq_type) {
 298         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 299                 rq->wqe_info = kzalloc_node(wq_sz * sizeof(*rq->wqe_info),
 300                                             GFP_KERNEL, cpu_to_node(c->cpu));
 301                 if (!rq->wqe_info) {
 302                         err = -ENOMEM;
 303                         goto err_rq_wq_destroy;
 304                 }
 305                 rq->handle_rx_cqe = mlx5e_handle_rx_cqe_mpwrq;
 306                 rq->alloc_wqe = mlx5e_alloc_rx_mpwqe;
 307
 308                 rq->wqe_sz = MLX5_MPWRQ_NUM_STRIDES * MLX5_MPWRQ_STRIDE_SIZE;
 309                 byte_count = rq->wqe_sz;
 310                 break;
 311         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 312                 rq->skb = kzalloc_node(wq_sz * sizeof(*rq->skb), GFP_KERNEL,
 313                                        cpu_to_node(c->cpu));
 314                 if (!rq->skb) {
 315                         err = -ENOMEM;
 316                         goto err_rq_wq_destroy;
 317                 }
 318                 rq->handle_rx_cqe = mlx5e_handle_rx_cqe;
 319                 rq->alloc_wqe = mlx5e_alloc_rx_wqe;
 320
 321                 rq->wqe_sz = (priv->params.lro_en) ?
 322                                 priv->params.lro_wqe_sz :
 323                                 MLX5E_SW2HW_MTU(priv->netdev->mtu);
 324                 rq->wqe_sz = SKB_DATA_ALIGN(rq->wqe_sz);
 325                 byte_count = rq->wqe_sz;
 326                 byte_count |= MLX5_HW_START_PADDING;
 327         }
 328
 329         for (i = 0; i < wq_sz; i++) {
 330                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, i);
 331
 332                 wqe->data.byte_count = cpu_to_be32(byte_count);
 333         }
 334
 335         rq->wq_type = priv->params.rq_wq_type;
 336         rq->pdev    = c->pdev;
 337         rq->netdev  = c->netdev;
 338         rq->tstamp  = &priv->tstamp;
 339         rq->channel = c;
 340         rq->ix      = c->ix;
 341         rq->priv    = c->priv;
 342         rq->mkey_be = c->mkey_be;
 343         rq->umr_mkey_be = cpu_to_be32(c->priv->umr_mkey.key);
 344
 345         return 0;
 346
 347 err_rq_wq_destroy:
 348         mlx5_wq_destroy(&rq->wq_ctrl);
 349
 350         return err;
 351 }
 352
 353 static void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 354 {
 355         switch (rq->wq_type) {
 356         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 357                 kfree(rq->wqe_info);
 358                 break;
 359         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 360                 kfree(rq->skb);
 361         }
 362
 363         mlx5_wq_destroy(&rq->wq_ctrl);
 364 }
 365
 366 static int mlx5e_enable_rq(struct mlx5e_rq *rq, struct mlx5e_rq_param *param)
 367 {
 368         struct mlx5e_priv *priv = rq->priv;
 369         struct mlx5_core_dev *mdev = priv->mdev;
 370
 371         void *in;
 372         void *rqc;
 373         void *wq;
 374         int inlen;
 375         int err;
 376
 377         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 378                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 379         in = mlx5_vzalloc(inlen);
 380         if (!in)
 381                 return -ENOMEM;
 382
 383         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 384         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 385
 386         memcpy(rqc, param->rqc, sizeof(param->rqc));
 387
 388         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 389         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 390         MLX5_SET(rqc,  rqc, flush_in_error_en,  1);
 391         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 392                                                 MLX5_ADAPTER_PAGE_SHIFT);
 393         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 394
 395         mlx5_fill_page_array(&rq->wq_ctrl.buf,
 396                              (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 397
 398         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 399
 400         kvfree(in);
 401
 402         return err;
 403 }
 404
 405 static int mlx5e_modify_rq(struct mlx5e_rq *rq, int curr_state, int next_state)
 406 {
 407         struct mlx5e_channel *c = rq->channel;
 408         struct mlx5e_priv *priv = c->priv;
 409         struct mlx5_core_dev *mdev = priv->mdev;
 410
 411         void *in;
 412         void *rqc;
 413         int inlen;
 414         int err;
 415
 416         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 417         in = mlx5_vzalloc(inlen);
 418         if (!in)
 419                 return -ENOMEM;
 420
 421         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 422
 423         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 424         MLX5_SET(rqc, rqc, state, next_state);
 425
 426         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 427
 428         kvfree(in);
 429
 430         return err;
 431 }
 432
 433 static void mlx5e_disable_rq(struct mlx5e_rq *rq)
 434 {
 435         mlx5_core_destroy_rq(rq->priv->mdev, rq->rqn);
 436 }
 437
 438 static int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq)
 439 {
 440         unsigned long exp_time = jiffies + msecs_to_jiffies(20000);
 441         struct mlx5e_channel *c = rq->channel;
 442         struct mlx5e_priv *priv = c->priv;
 443         struct mlx5_wq_ll *wq = &rq->wq;
 444
 445         while (time_before(jiffies, exp_time)) {
 446                 if (wq->cur_sz >= priv->params.min_rx_wqes)
 447                         return 0;
 448
 449                 msleep(20);
 450         }
 451
 452         return -ETIMEDOUT;
 453 }
 454
 455 static int mlx5e_open_rq(struct mlx5e_channel *c,
 456                          struct mlx5e_rq_param *param,
 457                          struct mlx5e_rq *rq)
 458 {
 459         struct mlx5e_sq *sq = &c->icosq;
 460         u16 pi = sq->pc & sq->wq.sz_m1;
 461         int err;
 462
 463         err = mlx5e_create_rq(c, param, rq);
 464         if (err)
 465                 return err;
 466
 467         err = mlx5e_enable_rq(rq, param);
 468         if (err)
 469                 goto err_destroy_rq;
 470
 471         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 472         if (err)
 473                 goto err_disable_rq;
 474
 475         set_bit(MLX5E_RQ_STATE_POST_WQES_ENABLE, &rq->state);
 476
 477         sq->ico_wqe_info[pi].opcode     = MLX5_OPCODE_NOP;
 478         sq->ico_wqe_info[pi].num_wqebbs = 1;
 479         mlx5e_send_nop(sq, true); /* trigger mlx5e_post_rx_wqes() */
 480
 481         return 0;
 482
 483 err_disable_rq:
 484         mlx5e_disable_rq(rq);
 485 err_destroy_rq:
 486         mlx5e_destroy_rq(rq);
 487
 488         return err;
 489 }
 490
 491 static void mlx5e_close_rq(struct mlx5e_rq *rq)
 492 {
 493         clear_bit(MLX5E_RQ_STATE_POST_WQES_ENABLE, &rq->state);
 494         napi_synchronize(&rq->channel->napi); /* prevent mlx5e_post_rx_wqes */
 495
 496         mlx5e_modify_rq(rq, MLX5_RQC_STATE_RDY, MLX5_RQC_STATE_ERR);
 497         while (!mlx5_wq_ll_is_empty(&rq->wq))
 498                 msleep(20);
 499
 500         /* avoid destroying rq before mlx5e_poll_rx_cq() is done with it */
 501         napi_synchronize(&rq->channel->napi);
 502
 503         mlx5e_disable_rq(rq);
 504         mlx5e_destroy_rq(rq);
 505 }
 506
 507 static void mlx5e_free_sq_db(struct mlx5e_sq *sq)
 508 {
 509         kfree(sq->wqe_info);
 510         kfree(sq->dma_fifo);
 511         kfree(sq->skb);
 512 }
 513
 514 static int mlx5e_alloc_sq_db(struct mlx5e_sq *sq, int numa)
 515 {
 516         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 517         int df_sz = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
 518
 519         sq->skb = kzalloc_node(wq_sz * sizeof(*sq->skb), GFP_KERNEL, numa);
 520         sq->dma_fifo = kzalloc_node(df_sz * sizeof(*sq->dma_fifo), GFP_KERNEL,
 521                                     numa);
 522         sq->wqe_info = kzalloc_node(wq_sz * sizeof(*sq->wqe_info), GFP_KERNEL,
 523                                     numa);
 524
 525         if (!sq->skb || !sq->dma_fifo || !sq->wqe_info) {
 526                 mlx5e_free_sq_db(sq);
 527                 return -ENOMEM;
 528         }
 529
 530         sq->dma_fifo_mask = df_sz - 1;
 531
 532         return 0;
 533 }
 534
 535 static int mlx5e_create_sq(struct mlx5e_channel *c,
 536                            int tc,
 537                            struct mlx5e_sq_param *param,
 538                            struct mlx5e_sq *sq)
 539 {
 540         struct mlx5e_priv *priv = c->priv;
 541         struct mlx5_core_dev *mdev = priv->mdev;
 542
 543         void *sqc = param->sqc;
 544         void *sqc_wq = MLX5_ADDR_OF(sqc, sqc, wq);
 545         int err;
 546
 547         err = mlx5_alloc_map_uar(mdev, &sq->uar, true);
 548         if (err)
 549                 return err;
 550
 551         param->wq.db_numa_node = cpu_to_node(c->cpu);
 552
 553         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq,
 554                                  &sq->wq_ctrl);
 555         if (err)
 556                 goto err_unmap_free_uar;
 557
 558         sq->wq.db       = &sq->wq.db[MLX5_SND_DBR];
 559         if (sq->uar.bf_map) {
 560                 set_bit(MLX5E_SQ_STATE_BF_ENABLE, &sq->state);
 561                 sq->uar_map = sq->uar.bf_map;
 562         } else {
 563                 sq->uar_map = sq->uar.map;
 564         }
 565         sq->bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
 566         sq->max_inline  = param->max_inline;
 567
 568         err = mlx5e_alloc_sq_db(sq, cpu_to_node(c->cpu));
 569         if (err)
 570                 goto err_sq_wq_destroy;
 571
 572         if (param->icosq) {
 573                 u8 wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 574
 575                 sq->ico_wqe_info = kzalloc_node(sizeof(*sq->ico_wqe_info) *
 576                                                 wq_sz,
 577                                                 GFP_KERNEL,
 578                                                 cpu_to_node(c->cpu));
 579                 if (!sq->ico_wqe_info) {
 580                         err = -ENOMEM;
 581                         goto err_free_sq_db;
 582                 }
 583         } else {
 584                 int txq_ix;
 585
 586                 txq_ix = c->ix + tc * priv->params.num_channels;
 587                 sq->txq = netdev_get_tx_queue(priv->netdev, txq_ix);
 588                 priv->txq_to_sq_map[txq_ix] = sq;
 589         }
 590
 591         sq->pdev      = c->pdev;
 592         sq->tstamp    = &priv->tstamp;
 593         sq->mkey_be   = c->mkey_be;
 594         sq->channel   = c;
 595         sq->tc        = tc;
 596         sq->edge      = (sq->wq.sz_m1 + 1) - MLX5_SEND_WQE_MAX_WQEBBS;
 597         sq->bf_budget = MLX5E_SQ_BF_BUDGET;
 598
 599         return 0;
 600
 601 err_free_sq_db:
 602         mlx5e_free_sq_db(sq);
 603
 604 err_sq_wq_destroy:
 605         mlx5_wq_destroy(&sq->wq_ctrl);
 606
 607 err_unmap_free_uar:
 608         mlx5_unmap_free_uar(mdev, &sq->uar);
 609
 610         return err;
 611 }
 612
 613 static void mlx5e_destroy_sq(struct mlx5e_sq *sq)
 614 {
 615         struct mlx5e_channel *c = sq->channel;
 616         struct mlx5e_priv *priv = c->priv;
 617
 618         kfree(sq->ico_wqe_info);
 619         mlx5e_free_sq_db(sq);
 620         mlx5_wq_destroy(&sq->wq_ctrl);
 621         mlx5_unmap_free_uar(priv->mdev, &sq->uar);
 622 }
 623
 624 static int mlx5e_enable_sq(struct mlx5e_sq *sq, struct mlx5e_sq_param *param)
 625 {
 626         struct mlx5e_channel *c = sq->channel;
 627         struct mlx5e_priv *priv = c->priv;
 628         struct mlx5_core_dev *mdev = priv->mdev;
 629
 630         void *in;
 631         void *sqc;
 632         void *wq;
 633         int inlen;
 634         int err;
 635
 636         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
 637                 sizeof(u64) * sq->wq_ctrl.buf.npages;
 638         in = mlx5_vzalloc(inlen);
 639         if (!in)
 640                 return -ENOMEM;
 641
 642         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
 643         wq = MLX5_ADDR_OF(sqc, sqc, wq);
 644
 645         memcpy(sqc, param->sqc, sizeof(param->sqc));
 646
 647         MLX5_SET(sqc,  sqc, tis_num_0, param->icosq ? 0 : priv->tisn[sq->tc]);
 648         MLX5_SET(sqc,  sqc, cqn,                sq->cq.mcq.cqn);
 649         MLX5_SET(sqc,  sqc, state,              MLX5_SQC_STATE_RST);
 650         MLX5_SET(sqc,  sqc, tis_lst_sz,         param->icosq ? 0 : 1);
 651         MLX5_SET(sqc,  sqc, flush_in_error_en,  1);
 652
 653         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
 654         MLX5_SET(wq,   wq, uar_page,      sq->uar.index);
 655         MLX5_SET(wq,   wq, log_wq_pg_sz,  sq->wq_ctrl.buf.page_shift -
 656                                           MLX5_ADAPTER_PAGE_SHIFT);
 657         MLX5_SET64(wq, wq, dbr_addr,      sq->wq_ctrl.db.dma);
 658
 659         mlx5_fill_page_array(&sq->wq_ctrl.buf,
 660                              (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 661
 662         err = mlx5_core_create_sq(mdev, in, inlen, &sq->sqn);
 663
 664         kvfree(in);
 665
 666         return err;
 667 }
 668
 669 static int mlx5e_modify_sq(struct mlx5e_sq *sq, int curr_state, int next_state)
 670 {
 671         struct mlx5e_channel *c = sq->channel;
 672         struct mlx5e_priv *priv = c->priv;
 673         struct mlx5_core_dev *mdev = priv->mdev;
 674
 675         void *in;
 676         void *sqc;
 677         int inlen;
 678         int err;
 679
 680         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
 681         in = mlx5_vzalloc(inlen);
 682         if (!in)
 683                 return -ENOMEM;
 684
 685         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
 686
 687         MLX5_SET(modify_sq_in, in, sq_state, curr_state);
 688         MLX5_SET(sqc, sqc, state, next_state);
 689
 690         err = mlx5_core_modify_sq(mdev, sq->sqn, in, inlen);
 691
 692         kvfree(in);
 693
 694         return err;
 695 }
 696
 697 static void mlx5e_disable_sq(struct mlx5e_sq *sq)
 698 {
 699         struct mlx5e_channel *c = sq->channel;
 700         struct mlx5e_priv *priv = c->priv;
 701         struct mlx5_core_dev *mdev = priv->mdev;
 702
 703         mlx5_core_destroy_sq(mdev, sq->sqn);
 704 }
 705
 706 static int mlx5e_open_sq(struct mlx5e_channel *c,
 707                          int tc,
 708                          struct mlx5e_sq_param *param,
 709                          struct mlx5e_sq *sq)
 710 {
 711         int err;
 712
 713         err = mlx5e_create_sq(c, tc, param, sq);
 714         if (err)
 715                 return err;
 716
 717         err = mlx5e_enable_sq(sq, param);
 718         if (err)
 719                 goto err_destroy_sq;
 720
 721         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RST, MLX5_SQC_STATE_RDY);
 722         if (err)
 723                 goto err_disable_sq;
 724
 725         if (sq->txq) {
 726                 set_bit(MLX5E_SQ_STATE_WAKE_TXQ_ENABLE, &sq->state);
 727                 netdev_tx_reset_queue(sq->txq);
 728                 netif_tx_start_queue(sq->txq);
 729         }
 730
 731         return 0;
 732
 733 err_disable_sq:
 734         mlx5e_disable_sq(sq);
 735 err_destroy_sq:
 736         mlx5e_destroy_sq(sq);
 737
 738         return err;
 739 }
 740
 741 static inline void netif_tx_disable_queue(struct netdev_queue *txq)
 742 {
 743         __netif_tx_lock_bh(txq);
 744         netif_tx_stop_queue(txq);
 745         __netif_tx_unlock_bh(txq);
 746 }
 747
 748 static void mlx5e_close_sq(struct mlx5e_sq *sq)
 749 {
 750         if (sq->txq) {
 751                 clear_bit(MLX5E_SQ_STATE_WAKE_TXQ_ENABLE, &sq->state);
 752                 /* prevent netif_tx_wake_queue */
 753                 napi_synchronize(&sq->channel->napi);
 754                 netif_tx_disable_queue(sq->txq);
 755
 756                 /* ensure hw is notified of all pending wqes */
 757                 if (mlx5e_sq_has_room_for(sq, 1))
 758                         mlx5e_send_nop(sq, true);
 759
 760                 mlx5e_modify_sq(sq, MLX5_SQC_STATE_RDY, MLX5_SQC_STATE_ERR);
 761         }
 762
 763         while (sq->cc != sq->pc) /* wait till sq is empty */
 764                 msleep(20);
 765
 766         /* avoid destroying sq before mlx5e_poll_tx_cq() is done with it */
 767         napi_synchronize(&sq->channel->napi);
 768
 769         mlx5e_disable_sq(sq);
 770         mlx5e_destroy_sq(sq);
 771 }
 772
 773 static int mlx5e_create_cq(struct mlx5e_channel *c,
 774                            struct mlx5e_cq_param *param,
 775                            struct mlx5e_cq *cq)
 776 {
 777         struct mlx5e_priv *priv = c->priv;
 778         struct mlx5_core_dev *mdev = priv->mdev;
 779         struct mlx5_core_cq *mcq = &cq->mcq;
 780         int eqn_not_used;
 781         unsigned int irqn;
 782         int err;
 783         u32 i;
 784
 785         param->wq.buf_numa_node = cpu_to_node(c->cpu);
 786         param->wq.db_numa_node  = cpu_to_node(c->cpu);
 787         param->eq_ix   = c->ix;
 788
 789         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
 790                                &cq->wq_ctrl);
 791         if (err)
 792                 return err;
 793
 794         mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
 795
 796         cq->napi        = &c->napi;
 797
 798         mcq->cqe_sz     = 64;
 799         mcq->set_ci_db  = cq->wq_ctrl.db.db;
 800         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
 801         *mcq->set_ci_db = 0;
 802         *mcq->arm_db    = 0;
 803         mcq->vector     = param->eq_ix;
 804         mcq->comp       = mlx5e_completion_event;
 805         mcq->event      = mlx5e_cq_error_event;
 806         mcq->irqn       = irqn;
 807         mcq->uar        = &priv->cq_uar;
 808
 809         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
 810                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
 811
 812                 cqe->op_own = 0xf1;
 813         }
 814
 815         cq->channel = c;
 816         cq->priv = priv;
 817
 818         return 0;
 819 }
 820
 821 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
 822 {
 823         mlx5_wq_destroy(&cq->wq_ctrl);
 824 }
 825
 826 static int mlx5e_enable_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
 827 {
 828         struct mlx5e_priv *priv = cq->priv;
 829         struct mlx5_core_dev *mdev = priv->mdev;
 830         struct mlx5_core_cq *mcq = &cq->mcq;
 831
 832         void *in;
 833         void *cqc;
 834         int inlen;
 835         unsigned int irqn_not_used;
 836         int eqn;
 837         int err;
 838
 839         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
 840                 sizeof(u64) * cq->wq_ctrl.buf.npages;
 841         in = mlx5_vzalloc(inlen);
 842         if (!in)
 843                 return -ENOMEM;
 844
 845         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
 846
 847         memcpy(cqc, param->cqc, sizeof(param->cqc));
 848
 849         mlx5_fill_page_array(&cq->wq_ctrl.buf,
 850                              (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
 851
 852         mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
 853
 854         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
 855         MLX5_SET(cqc,   cqc, uar_page,      mcq->uar->index);
 856         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
 857                                             MLX5_ADAPTER_PAGE_SHIFT);
 858         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
 859
 860         err = mlx5_core_create_cq(mdev, mcq, in, inlen);
 861
 862         kvfree(in);
 863
 864         if (err)
 865                 return err;
 866
 867         mlx5e_cq_arm(cq);
 868
 869         return 0;
 870 }
 871
 872 static void mlx5e_disable_cq(struct mlx5e_cq *cq)
 873 {
 874         struct mlx5e_priv *priv = cq->priv;
 875         struct mlx5_core_dev *mdev = priv->mdev;
 876
 877         mlx5_core_destroy_cq(mdev, &cq->mcq);
 878 }
 879
 880 static int mlx5e_open_cq(struct mlx5e_channel *c,
 881                          struct mlx5e_cq_param *param,
 882                          struct mlx5e_cq *cq,
 883                          u16 moderation_usecs,
 884                          u16 moderation_frames)
 885 {
 886         int err;
 887         struct mlx5e_priv *priv = c->priv;
 888         struct mlx5_core_dev *mdev = priv->mdev;
 889
 890         err = mlx5e_create_cq(c, param, cq);
 891         if (err)
 892                 return err;
 893
 894         err = mlx5e_enable_cq(cq, param);
 895         if (err)
 896                 goto err_destroy_cq;
 897
 898         if (MLX5_CAP_GEN(mdev, cq_moderation))
 899                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq,
 900                                                moderation_usecs,
 901                                                moderation_frames);
 902         return 0;
 903
 904 err_destroy_cq:
 905         mlx5e_destroy_cq(cq);
 906
 907         return err;
 908 }
 909
 910 static void mlx5e_close_cq(struct mlx5e_cq *cq)
 911 {
 912         mlx5e_disable_cq(cq);
 913         mlx5e_destroy_cq(cq);
 914 }
 915
 916 static int mlx5e_get_cpu(struct mlx5e_priv *priv, int ix)
 917 {
 918         return cpumask_first(priv->mdev->priv.irq_info[ix].mask);
 919 }
 920
 921 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
 922                              struct mlx5e_channel_param *cparam)
 923 {
 924         struct mlx5e_priv *priv = c->priv;
 925         int err;
 926         int tc;
 927
 928         for (tc = 0; tc < c->num_tc; tc++) {
 929                 err = mlx5e_open_cq(c, &cparam->tx_cq, &c->sq[tc].cq,
 930                                     priv->params.tx_cq_moderation_usec,
 931                                     priv->params.tx_cq_moderation_pkts);
 932                 if (err)
 933                         goto err_close_tx_cqs;
 934         }
 935
 936         return 0;
 937
 938 err_close_tx_cqs:
 939         for (tc--; tc >= 0; tc--)
 940                 mlx5e_close_cq(&c->sq[tc].cq);
 941
 942         return err;
 943 }
 944
 945 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
 946 {
 947         int tc;
 948
 949         for (tc = 0; tc < c->num_tc; tc++)
 950                 mlx5e_close_cq(&c->sq[tc].cq);
 951 }
 952
 953 static int mlx5e_open_sqs(struct mlx5e_channel *c,
 954                           struct mlx5e_channel_param *cparam)
 955 {
 956         int err;
 957         int tc;
 958
 959         for (tc = 0; tc < c->num_tc; tc++) {
 960                 err = mlx5e_open_sq(c, tc, &cparam->sq, &c->sq[tc]);
 961                 if (err)
 962                         goto err_close_sqs;
 963         }
 964
 965         return 0;
 966
 967 err_close_sqs:
 968         for (tc--; tc >= 0; tc--)
 969                 mlx5e_close_sq(&c->sq[tc]);
 970
 971         return err;
 972 }
 973
 974 static void mlx5e_close_sqs(struct mlx5e_channel *c)
 975 {
 976         int tc;
 977
 978         for (tc = 0; tc < c->num_tc; tc++)
 979                 mlx5e_close_sq(&c->sq[tc]);
 980 }
 981
 982 static void mlx5e_build_channeltc_to_txq_map(struct mlx5e_priv *priv, int ix)
 983 {
 984         int i;
 985
 986         for (i = 0; i < MLX5E_MAX_NUM_TC; i++)
 987                 priv->channeltc_to_txq_map[ix][i] =
 988                         ix + i * priv->params.num_channels;
 989 }
 990
 991 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
 992                               struct mlx5e_channel_param *cparam,
 993                               struct mlx5e_channel **cp)
 994 {
 995         struct net_device *netdev = priv->netdev;
 996         int cpu = mlx5e_get_cpu(priv, ix);
 997         struct mlx5e_channel *c;
 998         int err;
 999
1000         c = kzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
1001         if (!c)
1002                 return -ENOMEM;
1003
1004         c->priv     = priv;
1005         c->ix       = ix;
1006         c->cpu      = cpu;
1007         c->pdev     = &priv->mdev->pdev->dev;
1008         c->netdev   = priv->netdev;
1009         c->mkey_be  = cpu_to_be32(priv->mkey.key);
1010         c->num_tc   = priv->params.num_tc;
1011
1012         mlx5e_build_channeltc_to_txq_map(priv, ix);
1013
1014         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
1015
1016         err = mlx5e_open_cq(c, &cparam->icosq_cq, &c->icosq.cq, 0, 0);
1017         if (err)
1018                 goto err_napi_del;
1019
1020         err = mlx5e_open_tx_cqs(c, cparam);
1021         if (err)
1022                 goto err_close_icosq_cq;
1023
1024         err = mlx5e_open_cq(c, &cparam->rx_cq, &c->rq.cq,
1025                             priv->params.rx_cq_moderation_usec,
1026                             priv->params.rx_cq_moderation_pkts);
1027         if (err)
1028                 goto err_close_tx_cqs;
1029
1030         napi_enable(&c->napi);
1031
1032         err = mlx5e_open_sq(c, 0, &cparam->icosq, &c->icosq);
1033         if (err)
1034                 goto err_disable_napi;
1035
1036         err = mlx5e_open_sqs(c, cparam);
1037         if (err)
1038                 goto err_close_icosq;
1039
1040         err = mlx5e_open_rq(c, &cparam->rq, &c->rq);
1041         if (err)
1042                 goto err_close_sqs;
1043
1044         netif_set_xps_queue(netdev, get_cpu_mask(c->cpu), ix);
1045         *cp = c;
1046
1047         return 0;
1048
1049 err_close_sqs:
1050         mlx5e_close_sqs(c);
1051
1052 err_close_icosq:
1053         mlx5e_close_sq(&c->icosq);
1054
1055 err_disable_napi:
1056         napi_disable(&c->napi);
1057         mlx5e_close_cq(&c->rq.cq);
1058
1059 err_close_tx_cqs:
1060         mlx5e_close_tx_cqs(c);
1061
1062 err_close_icosq_cq:
1063         mlx5e_close_cq(&c->icosq.cq);
1064
1065 err_napi_del:
1066         netif_napi_del(&c->napi);
1067         napi_hash_del(&c->napi);
1068         kfree(c);
1069
1070         return err;
1071 }
1072
1073 static void mlx5e_close_channel(struct mlx5e_channel *c)
1074 {
1075         mlx5e_close_rq(&c->rq);
1076         mlx5e_close_sqs(c);
1077         mlx5e_close_sq(&c->icosq);
1078         napi_disable(&c->napi);
1079         mlx5e_close_cq(&c->rq.cq);
1080         mlx5e_close_tx_cqs(c);
1081         mlx5e_close_cq(&c->icosq.cq);
1082         netif_napi_del(&c->napi);
1083
1084         napi_hash_del(&c->napi);
1085         synchronize_rcu();
1086
1087         kfree(c);
1088 }
1089
1090 static void mlx5e_build_rq_param(struct mlx5e_priv *priv,
1091                                  struct mlx5e_rq_param *param)
1092 {
1093         void *rqc = param->rqc;
1094         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1095
1096         switch (priv->params.rq_wq_type) {
1097         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1098                 MLX5_SET(wq, wq, log_wqe_num_of_strides,
1099                          MLX5_MPWRQ_LOG_NUM_STRIDES - 9);
1100                 MLX5_SET(wq, wq, log_wqe_stride_size,
1101                          MLX5_MPWRQ_LOG_STRIDE_SIZE - 6);
1102                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
1103                 break;
1104         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1105                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1106         }
1107
1108         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1109         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1110         MLX5_SET(wq, wq, log_wq_sz,        priv->params.log_rq_size);
1111         MLX5_SET(wq, wq, pd,               priv->pdn);
1112         MLX5_SET(rqc, rqc, counter_set_id, priv->q_counter);
1113
1114         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1115         param->wq.linear = 1;
1116 }
1117
1118 static void mlx5e_build_drop_rq_param(struct mlx5e_rq_param *param)
1119 {
1120         void *rqc = param->rqc;
1121         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1122
1123         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1124         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1125 }
1126
1127 static void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
1128                                         struct mlx5e_sq_param *param)
1129 {
1130         void *sqc = param->sqc;
1131         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1132
1133         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1134         MLX5_SET(wq, wq, pd,            priv->pdn);
1135
1136         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1137 }
1138
1139 static void mlx5e_build_sq_param(struct mlx5e_priv *priv,
1140                                  struct mlx5e_sq_param *param)
1141 {
1142         void *sqc = param->sqc;
1143         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1144
1145         mlx5e_build_sq_param_common(priv, param);
1146         MLX5_SET(wq, wq, log_wq_sz,     priv->params.log_sq_size);
1147
1148         param->max_inline = priv->params.tx_max_inline;
1149 }
1150
1151 static void mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
1152                                         struct mlx5e_cq_param *param)
1153 {
1154         void *cqc = param->cqc;
1155
1156         MLX5_SET(cqc, cqc, uar_page, priv->cq_uar.index);
1157 }
1158
1159 static void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
1160                                     struct mlx5e_cq_param *param)
1161 {
1162         void *cqc = param->cqc;
1163         u8 log_cq_size;
1164
1165         switch (priv->params.rq_wq_type) {
1166         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1167                 log_cq_size = priv->params.log_rq_size +
1168                         MLX5_MPWRQ_LOG_NUM_STRIDES;
1169                 break;
1170         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1171                 log_cq_size = priv->params.log_rq_size;
1172         }
1173
1174         MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
1175
1176         mlx5e_build_common_cq_param(priv, param);
1177 }
1178
1179 static void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
1180                                     struct mlx5e_cq_param *param)
1181 {
1182         void *cqc = param->cqc;
1183
1184         MLX5_SET(cqc, cqc, log_cq_size, priv->params.log_sq_size);
1185
1186         mlx5e_build_common_cq_param(priv, param);
1187 }
1188
1189 static void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
1190                                      struct mlx5e_cq_param *param,
1191                                      u8 log_wq_size)
1192 {
1193         void *cqc = param->cqc;
1194
1195         MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
1196
1197         mlx5e_build_common_cq_param(priv, param);
1198 }
1199
1200 static void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
1201                                     struct mlx5e_sq_param *param,
1202                                     u8 log_wq_size)
1203 {
1204         void *sqc = param->sqc;
1205         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1206
1207         mlx5e_build_sq_param_common(priv, param);
1208
1209         MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
1210         MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(priv->mdev, reg_umr_sq));
1211
1212         param->icosq = true;
1213 }
1214
1215 static void mlx5e_build_channel_param(struct mlx5e_priv *priv,
1216                                       struct mlx5e_channel_param *cparam)
1217 {
1218         u8 icosq_log_wq_sz = MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
1219
1220         memset(cparam, 0, sizeof(*cparam));
1221
1222         mlx5e_build_rq_param(priv, &cparam->rq);
1223         mlx5e_build_sq_param(priv, &cparam->sq);
1224         mlx5e_build_icosq_param(priv, &cparam->icosq, icosq_log_wq_sz);
1225         mlx5e_build_rx_cq_param(priv, &cparam->rx_cq);
1226         mlx5e_build_tx_cq_param(priv, &cparam->tx_cq);
1227         mlx5e_build_ico_cq_param(priv, &cparam->icosq_cq, icosq_log_wq_sz);
1228 }
1229
1230 static int mlx5e_open_channels(struct mlx5e_priv *priv)
1231 {
1232         struct mlx5e_channel_param cparam;
1233         int nch = priv->params.num_channels;
1234         int err = -ENOMEM;
1235         int i;
1236         int j;
1237
1238         priv->channel = kcalloc(nch, sizeof(struct mlx5e_channel *),
1239                                 GFP_KERNEL);
1240
1241         priv->txq_to_sq_map = kcalloc(nch * priv->params.num_tc,
1242                                       sizeof(struct mlx5e_sq *), GFP_KERNEL);
1243
1244         if (!priv->channel || !priv->txq_to_sq_map)
1245                 goto err_free_txq_to_sq_map;
1246
1247         mlx5e_build_channel_param(priv, &cparam);
1248         for (i = 0; i < nch; i++) {
1249                 err = mlx5e_open_channel(priv, i, &cparam, &priv->channel[i]);
1250                 if (err)
1251                         goto err_close_channels;
1252         }
1253
1254         for (j = 0; j < nch; j++) {
1255                 err = mlx5e_wait_for_min_rx_wqes(&priv->channel[j]->rq);
1256                 if (err)
1257                         goto err_close_channels;
1258         }
1259
1260         return 0;
1261
1262 err_close_channels:
1263         for (i--; i >= 0; i--)
1264                 mlx5e_close_channel(priv->channel[i]);
1265
1266 err_free_txq_to_sq_map:
1267         kfree(priv->txq_to_sq_map);
1268         kfree(priv->channel);
1269
1270         return err;
1271 }
1272
1273 static void mlx5e_close_channels(struct mlx5e_priv *priv)
1274 {
1275         int i;
1276
1277         for (i = 0; i < priv->params.num_channels; i++)
1278                 mlx5e_close_channel(priv->channel[i]);
1279
1280         kfree(priv->txq_to_sq_map);
1281         kfree(priv->channel);
1282 }
1283
1284 static int mlx5e_rx_hash_fn(int hfunc)
1285 {
1286         return (hfunc == ETH_RSS_HASH_TOP) ?
1287                MLX5_RX_HASH_FN_TOEPLITZ :
1288                MLX5_RX_HASH_FN_INVERTED_XOR8;
1289 }
1290
1291 static int mlx5e_bits_invert(unsigned long a, int size)
1292 {
1293         int inv = 0;
1294         int i;
1295
1296         for (i = 0; i < size; i++)
1297                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
1298
1299         return inv;
1300 }
1301
1302 static void mlx5e_fill_indir_rqt_rqns(struct mlx5e_priv *priv, void *rqtc)
1303 {
1304         int i;
1305
1306         for (i = 0; i < MLX5E_INDIR_RQT_SIZE; i++) {
1307                 int ix = i;
1308
1309                 if (priv->params.rss_hfunc == ETH_RSS_HASH_XOR)
1310                         ix = mlx5e_bits_invert(i, MLX5E_LOG_INDIR_RQT_SIZE);
1311
1312                 ix = priv->params.indirection_rqt[ix];
1313                 MLX5_SET(rqtc, rqtc, rq_num[i],
1314                          test_bit(MLX5E_STATE_OPENED, &priv->state) ?
1315                          priv->channel[ix]->rq.rqn :
1316                          priv->drop_rq.rqn);
1317         }
1318 }
1319
1320 static void mlx5e_fill_rqt_rqns(struct mlx5e_priv *priv, void *rqtc,
1321                                 enum mlx5e_rqt_ix rqt_ix)
1322 {
1323
1324         switch (rqt_ix) {
1325         case MLX5E_INDIRECTION_RQT:
1326                 mlx5e_fill_indir_rqt_rqns(priv, rqtc);
1327
1328                 break;
1329
1330         default: /* MLX5E_SINGLE_RQ_RQT */
1331                 MLX5_SET(rqtc, rqtc, rq_num[0],
1332                          test_bit(MLX5E_STATE_OPENED, &priv->state) ?
1333                          priv->channel[0]->rq.rqn :
1334                          priv->drop_rq.rqn);
1335
1336                 break;
1337         }
1338 }
1339
1340 static int mlx5e_create_rqt(struct mlx5e_priv *priv, enum mlx5e_rqt_ix rqt_ix)
1341 {
1342         struct mlx5_core_dev *mdev = priv->mdev;
1343         u32 *in;
1344         void *rqtc;
1345         int inlen;
1346         int sz;
1347         int err;
1348
1349         sz = (rqt_ix == MLX5E_SINGLE_RQ_RQT) ? 1 : MLX5E_INDIR_RQT_SIZE;
1350
1351         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
1352         in = mlx5_vzalloc(inlen);
1353         if (!in)
1354                 return -ENOMEM;
1355
1356         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
1357
1358         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
1359         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
1360
1361         mlx5e_fill_rqt_rqns(priv, rqtc, rqt_ix);
1362
1363         err = mlx5_core_create_rqt(mdev, in, inlen, &priv->rqtn[rqt_ix]);
1364
1365         kvfree(in);
1366
1367         return err;
1368 }
1369
1370 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, enum mlx5e_rqt_ix rqt_ix)
1371 {
1372         struct mlx5_core_dev *mdev = priv->mdev;
1373         u32 *in;
1374         void *rqtc;
1375         int inlen;
1376         int sz;
1377         int err;
1378
1379         sz = (rqt_ix == MLX5E_SINGLE_RQ_RQT) ? 1 : MLX5E_INDIR_RQT_SIZE;
1380
1381         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
1382         in = mlx5_vzalloc(inlen);
1383         if (!in)
1384                 return -ENOMEM;
1385
1386         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
1387
1388         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
1389
1390         mlx5e_fill_rqt_rqns(priv, rqtc, rqt_ix);
1391
1392         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
1393
1394         err = mlx5_core_modify_rqt(mdev, priv->rqtn[rqt_ix], in, inlen);
1395
1396         kvfree(in);
1397
1398         return err;
1399 }
1400
1401 static void mlx5e_destroy_rqt(struct mlx5e_priv *priv, enum mlx5e_rqt_ix rqt_ix)
1402 {
1403         mlx5_core_destroy_rqt(priv->mdev, priv->rqtn[rqt_ix]);
1404 }
1405
1406 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv)
1407 {
1408         mlx5e_redirect_rqt(priv, MLX5E_INDIRECTION_RQT);
1409         mlx5e_redirect_rqt(priv, MLX5E_SINGLE_RQ_RQT);
1410 }
1411
1412 static void mlx5e_build_tir_ctx_lro(void *tirc, struct mlx5e_priv *priv)
1413 {
1414         if (!priv->params.lro_en)
1415                 return;
1416
1417 #define ROUGH_MAX_L2_L3_HDR_SZ 256
1418
1419         MLX5_SET(tirc, tirc, lro_enable_mask,
1420                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
1421                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
1422         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
1423                  (priv->params.lro_wqe_sz -
1424                   ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
1425         MLX5_SET(tirc, tirc, lro_timeout_period_usecs,
1426                  MLX5_CAP_ETH(priv->mdev,
1427                               lro_timer_supported_periods[2]));
1428 }
1429
1430 void mlx5e_build_tir_ctx_hash(void *tirc, struct mlx5e_priv *priv)
1431 {
1432         MLX5_SET(tirc, tirc, rx_hash_fn,
1433                  mlx5e_rx_hash_fn(priv->params.rss_hfunc));
1434         if (priv->params.rss_hfunc == ETH_RSS_HASH_TOP) {
1435                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
1436                                              rx_hash_toeplitz_key);
1437                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
1438                                                rx_hash_toeplitz_key);
1439
1440                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1441                 memcpy(rss_key, priv->params.toeplitz_hash_key, len);
1442         }
1443 }
1444
1445 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
1446 {
1447         struct mlx5_core_dev *mdev = priv->mdev;
1448
1449         void *in;
1450         void *tirc;
1451         int inlen;
1452         int err;
1453         int tt;
1454
1455         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
1456         in = mlx5_vzalloc(inlen);
1457         if (!in)
1458                 return -ENOMEM;
1459
1460         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
1461         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
1462
1463         mlx5e_build_tir_ctx_lro(tirc, priv);
1464
1465         for (tt = 0; tt < MLX5E_NUM_TT; tt++) {
1466                 err = mlx5_core_modify_tir(mdev, priv->tirn[tt], in, inlen);
1467                 if (err)
1468                         break;
1469         }
1470
1471         kvfree(in);
1472
1473         return err;
1474 }
1475
1476 static int mlx5e_refresh_tir_self_loopback_enable(struct mlx5_core_dev *mdev,
1477                                                   u32 tirn)
1478 {
1479         void *in;
1480         int inlen;
1481         int err;
1482
1483         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
1484         in = mlx5_vzalloc(inlen);
1485         if (!in)
1486                 return -ENOMEM;
1487
1488         MLX5_SET(modify_tir_in, in, bitmask.self_lb_en, 1);
1489
1490         err = mlx5_core_modify_tir(mdev, tirn, in, inlen);
1491
1492         kvfree(in);
1493
1494         return err;
1495 }
1496
1497 static int mlx5e_refresh_tirs_self_loopback_enable(struct mlx5e_priv *priv)
1498 {
1499         int err;
1500         int i;
1501
1502         for (i = 0; i < MLX5E_NUM_TT; i++) {
1503                 err = mlx5e_refresh_tir_self_loopback_enable(priv->mdev,
1504                                                              priv->tirn[i]);
1505                 if (err)
1506                         return err;
1507         }
1508
1509         return 0;
1510 }
1511
1512 static int mlx5e_set_dev_port_mtu(struct net_device *netdev)
1513 {
1514         struct mlx5e_priv *priv = netdev_priv(netdev);
1515         struct mlx5_core_dev *mdev = priv->mdev;
1516         int hw_mtu;
1517         int err;
1518
1519         err = mlx5_set_port_mtu(mdev, MLX5E_SW2HW_MTU(netdev->mtu), 1);
1520         if (err)
1521                 return err;
1522
1523         mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
1524
1525         if (MLX5E_HW2SW_MTU(hw_mtu) != netdev->mtu)
1526                 netdev_warn(netdev, "%s: Port MTU %d is different than netdev mtu %d\n",
1527                             __func__, MLX5E_HW2SW_MTU(hw_mtu), netdev->mtu);
1528
1529         netdev->mtu = MLX5E_HW2SW_MTU(hw_mtu);
1530         return 0;
1531 }
1532
1533 static void mlx5e_netdev_set_tcs(struct net_device *netdev)
1534 {
1535         struct mlx5e_priv *priv = netdev_priv(netdev);
1536         int nch = priv->params.num_channels;
1537         int ntc = priv->params.num_tc;
1538         int tc;
1539
1540         netdev_reset_tc(netdev);
1541
1542         if (ntc == 1)
1543                 return;
1544
1545         netdev_set_num_tc(netdev, ntc);
1546
1547         for (tc = 0; tc < ntc; tc++)
1548                 netdev_set_tc_queue(netdev, tc, nch, tc * nch);
1549 }
1550
1551 int mlx5e_open_locked(struct net_device *netdev)
1552 {
1553         struct mlx5e_priv *priv = netdev_priv(netdev);
1554         int num_txqs;
1555         int err;
1556
1557         set_bit(MLX5E_STATE_OPENED, &priv->state);
1558
1559         mlx5e_netdev_set_tcs(netdev);
1560
1561         num_txqs = priv->params.num_channels * priv->params.num_tc;
1562         netif_set_real_num_tx_queues(netdev, num_txqs);
1563         netif_set_real_num_rx_queues(netdev, priv->params.num_channels);
1564
1565         err = mlx5e_set_dev_port_mtu(netdev);
1566         if (err)
1567                 goto err_clear_state_opened_flag;
1568
1569         err = mlx5e_open_channels(priv);
1570         if (err) {
1571                 netdev_err(netdev, "%s: mlx5e_open_channels failed, %d\n",
1572                            __func__, err);
1573                 goto err_clear_state_opened_flag;
1574         }
1575
1576         err = mlx5e_refresh_tirs_self_loopback_enable(priv);
1577         if (err) {
1578                 netdev_err(netdev, "%s: mlx5e_refresh_tirs_self_loopback_enable failed, %d\n",
1579                            __func__, err);
1580                 goto err_close_channels;
1581         }
1582
1583         mlx5e_redirect_rqts(priv);
1584         mlx5e_update_carrier(priv);
1585         mlx5e_timestamp_init(priv);
1586
1587         schedule_delayed_work(&priv->update_stats_work, 0);
1588
1589         return 0;
1590
1591 err_close_channels:
1592         mlx5e_close_channels(priv);
1593 err_clear_state_opened_flag:
1594         clear_bit(MLX5E_STATE_OPENED, &priv->state);
1595         return err;
1596 }
1597
1598 static int mlx5e_open(struct net_device *netdev)
1599 {
1600         struct mlx5e_priv *priv = netdev_priv(netdev);
1601         int err;
1602
1603         mutex_lock(&priv->state_lock);
1604         err = mlx5e_open_locked(netdev);
1605         mutex_unlock(&priv->state_lock);
1606
1607         return err;
1608 }
1609
1610 int mlx5e_close_locked(struct net_device *netdev)
1611 {
1612         struct mlx5e_priv *priv = netdev_priv(netdev);
1613
1614         /* May already be CLOSED in case a previous configuration operation
1615          * (e.g RX/TX queue size change) that involves close&open failed.
1616          */
1617         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
1618                 return 0;
1619
1620         clear_bit(MLX5E_STATE_OPENED, &priv->state);
1621
1622         mlx5e_timestamp_cleanup(priv);
1623         netif_carrier_off(priv->netdev);
1624         mlx5e_redirect_rqts(priv);
1625         mlx5e_close_channels(priv);
1626
1627         return 0;
1628 }
1629
1630 static int mlx5e_close(struct net_device *netdev)
1631 {
1632         struct mlx5e_priv *priv = netdev_priv(netdev);
1633         int err;
1634
1635         mutex_lock(&priv->state_lock);
1636         err = mlx5e_close_locked(netdev);
1637         mutex_unlock(&priv->state_lock);
1638
1639         return err;
1640 }
1641
1642 static int mlx5e_create_drop_rq(struct mlx5e_priv *priv,
1643                                 struct mlx5e_rq *rq,
1644                                 struct mlx5e_rq_param *param)
1645 {
1646         struct mlx5_core_dev *mdev = priv->mdev;
1647         void *rqc = param->rqc;
1648         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
1649         int err;
1650
1651         param->wq.db_numa_node = param->wq.buf_numa_node;
1652
1653         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
1654                                 &rq->wq_ctrl);
1655         if (err)
1656                 return err;
1657
1658         rq->priv = priv;
1659
1660         return 0;
1661 }
1662
1663 static int mlx5e_create_drop_cq(struct mlx5e_priv *priv,
1664                                 struct mlx5e_cq *cq,
1665                                 struct mlx5e_cq_param *param)
1666 {
1667         struct mlx5_core_dev *mdev = priv->mdev;
1668         struct mlx5_core_cq *mcq = &cq->mcq;
1669         int eqn_not_used;
1670         unsigned int irqn;
1671         int err;
1672
1673         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1674                                &cq->wq_ctrl);
1675         if (err)
1676                 return err;
1677
1678         mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1679
1680         mcq->cqe_sz     = 64;
1681         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1682         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1683         *mcq->set_ci_db = 0;
1684         *mcq->arm_db    = 0;
1685         mcq->vector     = param->eq_ix;
1686         mcq->comp       = mlx5e_completion_event;
1687         mcq->event      = mlx5e_cq_error_event;
1688         mcq->irqn       = irqn;
1689         mcq->uar        = &priv->cq_uar;
1690
1691         cq->priv = priv;
1692
1693         return 0;
1694 }
1695
1696 static int mlx5e_open_drop_rq(struct mlx5e_priv *priv)
1697 {
1698         struct mlx5e_cq_param cq_param;
1699         struct mlx5e_rq_param rq_param;
1700         struct mlx5e_rq *rq = &priv->drop_rq;
1701         struct mlx5e_cq *cq = &priv->drop_rq.cq;
1702         int err;
1703
1704         memset(&cq_param, 0, sizeof(cq_param));
1705         memset(&rq_param, 0, sizeof(rq_param));
1706         mlx5e_build_drop_rq_param(&rq_param);
1707
1708         err = mlx5e_create_drop_cq(priv, cq, &cq_param);
1709         if (err)
1710                 return err;
1711
1712         err = mlx5e_enable_cq(cq, &cq_param);
1713         if (err)
1714                 goto err_destroy_cq;
1715
1716         err = mlx5e_create_drop_rq(priv, rq, &rq_param);
1717         if (err)
1718                 goto err_disable_cq;
1719
1720         err = mlx5e_enable_rq(rq, &rq_param);
1721         if (err)
1722                 goto err_destroy_rq;
1723
1724         return 0;
1725
1726 err_destroy_rq:
1727         mlx5e_destroy_rq(&priv->drop_rq);
1728
1729 err_disable_cq:
1730         mlx5e_disable_cq(&priv->drop_rq.cq);
1731
1732 err_destroy_cq:
1733         mlx5e_destroy_cq(&priv->drop_rq.cq);
1734
1735         return err;
1736 }
1737
1738 static void mlx5e_close_drop_rq(struct mlx5e_priv *priv)
1739 {
1740         mlx5e_disable_rq(&priv->drop_rq);
1741         mlx5e_destroy_rq(&priv->drop_rq);
1742         mlx5e_disable_cq(&priv->drop_rq.cq);
1743         mlx5e_destroy_cq(&priv->drop_rq.cq);
1744 }
1745
1746 static int mlx5e_create_tis(struct mlx5e_priv *priv, int tc)
1747 {
1748         struct mlx5_core_dev *mdev = priv->mdev;
1749         u32 in[MLX5_ST_SZ_DW(create_tis_in)];
1750         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
1751
1752         memset(in, 0, sizeof(in));
1753
1754         MLX5_SET(tisc, tisc, prio, tc << 1);
1755         MLX5_SET(tisc, tisc, transport_domain, priv->tdn);
1756
1757         return mlx5_core_create_tis(mdev, in, sizeof(in), &priv->tisn[tc]);
1758 }
1759
1760 static void mlx5e_destroy_tis(struct mlx5e_priv *priv, int tc)
1761 {
1762         mlx5_core_destroy_tis(priv->mdev, priv->tisn[tc]);
1763 }
1764
1765 static int mlx5e_create_tises(struct mlx5e_priv *priv)
1766 {
1767         int err;
1768         int tc;
1769
1770         for (tc = 0; tc < MLX5E_MAX_NUM_TC; tc++) {
1771                 err = mlx5e_create_tis(priv, tc);
1772                 if (err)
1773                         goto err_close_tises;
1774         }
1775
1776         return 0;
1777
1778 err_close_tises:
1779         for (tc--; tc >= 0; tc--)
1780                 mlx5e_destroy_tis(priv, tc);
1781
1782         return err;
1783 }
1784
1785 static void mlx5e_destroy_tises(struct mlx5e_priv *priv)
1786 {
1787         int tc;
1788
1789         for (tc = 0; tc < MLX5E_MAX_NUM_TC; tc++)
1790                 mlx5e_destroy_tis(priv, tc);
1791 }
1792
1793 static void mlx5e_build_tir_ctx(struct mlx5e_priv *priv, u32 *tirc, int tt)
1794 {
1795         void *hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
1796
1797         MLX5_SET(tirc, tirc, transport_domain, priv->tdn);
1798
1799 #define MLX5_HASH_IP            (MLX5_HASH_FIELD_SEL_SRC_IP   |\
1800                                  MLX5_HASH_FIELD_SEL_DST_IP)
1801
1802 #define MLX5_HASH_IP_L4PORTS    (MLX5_HASH_FIELD_SEL_SRC_IP   |\
1803                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
1804                                  MLX5_HASH_FIELD_SEL_L4_SPORT |\
1805                                  MLX5_HASH_FIELD_SEL_L4_DPORT)
1806
1807 #define MLX5_HASH_IP_IPSEC_SPI  (MLX5_HASH_FIELD_SEL_SRC_IP   |\
1808                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
1809                                  MLX5_HASH_FIELD_SEL_IPSEC_SPI)
1810
1811         mlx5e_build_tir_ctx_lro(tirc, priv);
1812
1813         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
1814
1815         switch (tt) {
1816         case MLX5E_TT_ANY:
1817                 MLX5_SET(tirc, tirc, indirect_table,
1818                          priv->rqtn[MLX5E_SINGLE_RQ_RQT]);
1819                 MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
1820                 break;
1821         default:
1822                 MLX5_SET(tirc, tirc, indirect_table,
1823                          priv->rqtn[MLX5E_INDIRECTION_RQT]);
1824                 mlx5e_build_tir_ctx_hash(tirc, priv);
1825                 break;
1826         }
1827
1828         switch (tt) {
1829         case MLX5E_TT_IPV4_TCP:
1830                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1831                          MLX5_L3_PROT_TYPE_IPV4);
1832                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1833                          MLX5_L4_PROT_TYPE_TCP);
1834                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1835                          MLX5_HASH_IP_L4PORTS);
1836                 break;
1837
1838         case MLX5E_TT_IPV6_TCP:
1839                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1840                          MLX5_L3_PROT_TYPE_IPV6);
1841                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1842                          MLX5_L4_PROT_TYPE_TCP);
1843                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1844                          MLX5_HASH_IP_L4PORTS);
1845                 break;
1846
1847         case MLX5E_TT_IPV4_UDP:
1848                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1849                          MLX5_L3_PROT_TYPE_IPV4);
1850                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1851                          MLX5_L4_PROT_TYPE_UDP);
1852                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1853                          MLX5_HASH_IP_L4PORTS);
1854                 break;
1855
1856         case MLX5E_TT_IPV6_UDP:
1857                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1858                          MLX5_L3_PROT_TYPE_IPV6);
1859                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
1860                          MLX5_L4_PROT_TYPE_UDP);
1861                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1862                          MLX5_HASH_IP_L4PORTS);
1863                 break;
1864
1865         case MLX5E_TT_IPV4_IPSEC_AH:
1866                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1867                          MLX5_L3_PROT_TYPE_IPV4);
1868                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1869                          MLX5_HASH_IP_IPSEC_SPI);
1870                 break;
1871
1872         case MLX5E_TT_IPV6_IPSEC_AH:
1873                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1874                          MLX5_L3_PROT_TYPE_IPV6);
1875                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1876                          MLX5_HASH_IP_IPSEC_SPI);
1877                 break;
1878
1879         case MLX5E_TT_IPV4_IPSEC_ESP:
1880                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1881                          MLX5_L3_PROT_TYPE_IPV4);
1882                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1883                          MLX5_HASH_IP_IPSEC_SPI);
1884                 break;
1885
1886         case MLX5E_TT_IPV6_IPSEC_ESP:
1887                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1888                          MLX5_L3_PROT_TYPE_IPV6);
1889                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1890                          MLX5_HASH_IP_IPSEC_SPI);
1891                 break;
1892
1893         case MLX5E_TT_IPV4:
1894                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1895                          MLX5_L3_PROT_TYPE_IPV4);
1896                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1897                          MLX5_HASH_IP);
1898                 break;
1899
1900         case MLX5E_TT_IPV6:
1901                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
1902                          MLX5_L3_PROT_TYPE_IPV6);
1903                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
1904                          MLX5_HASH_IP);
1905                 break;
1906         }
1907 }
1908
1909 static int mlx5e_create_tir(struct mlx5e_priv *priv, int tt)
1910 {
1911         struct mlx5_core_dev *mdev = priv->mdev;
1912         u32 *in;
1913         void *tirc;
1914         int inlen;
1915         int err;
1916
1917         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
1918         in = mlx5_vzalloc(inlen);
1919         if (!in)
1920                 return -ENOMEM;
1921
1922         tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
1923
1924         mlx5e_build_tir_ctx(priv, tirc, tt);
1925
1926         err = mlx5_core_create_tir(mdev, in, inlen, &priv->tirn[tt]);
1927
1928         kvfree(in);
1929
1930         return err;
1931 }
1932
1933 static void mlx5e_destroy_tir(struct mlx5e_priv *priv, int tt)
1934 {
1935         mlx5_core_destroy_tir(priv->mdev, priv->tirn[tt]);
1936 }
1937
1938 static int mlx5e_create_tirs(struct mlx5e_priv *priv)
1939 {
1940         int err;
1941         int i;
1942
1943         for (i = 0; i < MLX5E_NUM_TT; i++) {
1944                 err = mlx5e_create_tir(priv, i);
1945                 if (err)
1946                         goto err_destroy_tirs;
1947         }
1948
1949         return 0;
1950
1951 err_destroy_tirs:
1952         for (i--; i >= 0; i--)
1953                 mlx5e_destroy_tir(priv, i);
1954
1955         return err;
1956 }
1957
1958 static void mlx5e_destroy_tirs(struct mlx5e_priv *priv)
1959 {
1960         int i;
1961
1962         for (i = 0; i < MLX5E_NUM_TT; i++)
1963                 mlx5e_destroy_tir(priv, i);
1964 }
1965
1966 static int mlx5e_setup_tc(struct net_device *netdev, u8 tc)
1967 {
1968         struct mlx5e_priv *priv = netdev_priv(netdev);
1969         bool was_opened;
1970         int err = 0;
1971
1972         if (tc && tc != MLX5E_MAX_NUM_TC)
1973                 return -EINVAL;
1974
1975         mutex_lock(&priv->state_lock);
1976
1977         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
1978         if (was_opened)
1979                 mlx5e_close_locked(priv->netdev);
1980
1981         priv->params.num_tc = tc ? tc : 1;
1982
1983         if (was_opened)
1984                 err = mlx5e_open_locked(priv->netdev);
1985
1986         mutex_unlock(&priv->state_lock);
1987
1988         return err;
1989 }
1990
1991 static int mlx5e_ndo_setup_tc(struct net_device *dev, u32 handle,
1992                               __be16 proto, struct tc_to_netdev *tc)
1993 {
1994         struct mlx5e_priv *priv = netdev_priv(dev);
1995
1996         if (TC_H_MAJ(handle) != TC_H_MAJ(TC_H_INGRESS))
1997                 goto mqprio;
1998
1999         switch (tc->type) {
2000         case TC_SETUP_CLSFLOWER:
2001                 switch (tc->cls_flower->command) {
2002                 case TC_CLSFLOWER_REPLACE:
2003                         return mlx5e_configure_flower(priv, proto, tc->cls_flower);
2004                 case TC_CLSFLOWER_DESTROY:
2005                         return mlx5e_delete_flower(priv, tc->cls_flower);
2006                 }
2007         default:
2008                 return -EOPNOTSUPP;
2009         }
2010
2011 mqprio:
2012         if (tc->type != TC_SETUP_MQPRIO)
2013                 return -EINVAL;
2014
2015         return mlx5e_setup_tc(dev, tc->tc);
2016 }
2017
2018 static struct rtnl_link_stats64 *
2019 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
2020 {
2021         struct mlx5e_priv *priv = netdev_priv(dev);
2022         struct mlx5e_sw_stats *sstats = &priv->stats.sw;
2023         struct mlx5e_vport_stats *vstats = &priv->stats.vport;
2024         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
2025
2026         stats->rx_packets = sstats->rx_packets;
2027         stats->rx_bytes   = sstats->rx_bytes;
2028         stats->tx_packets = sstats->tx_packets;
2029         stats->tx_bytes   = sstats->tx_bytes;
2030
2031         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
2032         stats->tx_dropped = sstats->tx_queue_dropped;
2033
2034         stats->rx_length_errors =
2035                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
2036                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
2037                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
2038         stats->rx_crc_errors =
2039                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
2040         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
2041         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
2042         stats->tx_carrier_errors =
2043                 PPORT_802_3_GET(pstats, a_symbol_error_during_carrier);
2044         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
2045                            stats->rx_frame_errors;
2046         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
2047
2048         /* vport multicast also counts packets that are dropped due to steering
2049          * or rx out of buffer
2050          */
2051         stats->multicast =
2052                 VPORT_COUNTER_GET(vstats, received_eth_multicast.packets);
2053
2054         return stats;
2055 }
2056
2057 static void mlx5e_set_rx_mode(struct net_device *dev)
2058 {
2059         struct mlx5e_priv *priv = netdev_priv(dev);
2060
2061         schedule_work(&priv->set_rx_mode_work);
2062 }
2063
2064 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
2065 {
2066         struct mlx5e_priv *priv = netdev_priv(netdev);
2067         struct sockaddr *saddr = addr;
2068
2069         if (!is_valid_ether_addr(saddr->sa_data))
2070                 return -EADDRNOTAVAIL;
2071
2072         netif_addr_lock_bh(netdev);
2073         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
2074         netif_addr_unlock_bh(netdev);
2075
2076         schedule_work(&priv->set_rx_mode_work);
2077
2078         return 0;
2079 }
2080
2081 #define MLX5E_SET_FEATURE(netdev, feature, enable)      \
2082         do {                                            \
2083                 if (enable)                             \
2084                         netdev->features |= feature;    \
2085                 else                                    \
2086                         netdev->features &= ~feature;   \
2087         } while (0)
2088
2089 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
2090
2091 static int set_feature_lro(struct net_device *netdev, bool enable)
2092 {
2093         struct mlx5e_priv *priv = netdev_priv(netdev);
2094         bool was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2095         int err;
2096
2097         mutex_lock(&priv->state_lock);
2098
2099         if (was_opened && (priv->params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST))
2100                 mlx5e_close_locked(priv->netdev);
2101
2102         priv->params.lro_en = enable;
2103         err = mlx5e_modify_tirs_lro(priv);
2104         if (err) {
2105                 netdev_err(netdev, "lro modify failed, %d\n", err);
2106                 priv->params.lro_en = !enable;
2107         }
2108
2109         if (was_opened && (priv->params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST))
2110                 mlx5e_open_locked(priv->netdev);
2111
2112         mutex_unlock(&priv->state_lock);
2113
2114         return err;
2115 }
2116
2117 static int set_feature_vlan_filter(struct net_device *netdev, bool enable)
2118 {
2119         struct mlx5e_priv *priv = netdev_priv(netdev);
2120
2121         if (enable)
2122                 mlx5e_enable_vlan_filter(priv);
2123         else
2124                 mlx5e_disable_vlan_filter(priv);
2125
2126         return 0;
2127 }
2128
2129 static int set_feature_tc_num_filters(struct net_device *netdev, bool enable)
2130 {
2131         struct mlx5e_priv *priv = netdev_priv(netdev);
2132
2133         if (!enable && mlx5e_tc_num_filters(priv)) {
2134                 netdev_err(netdev,
2135                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
2136                 return -EINVAL;
2137         }
2138
2139         return 0;
2140 }
2141
2142 static int mlx5e_handle_feature(struct net_device *netdev,
2143                                 netdev_features_t wanted_features,
2144                                 netdev_features_t feature,
2145                                 mlx5e_feature_handler feature_handler)
2146 {
2147         netdev_features_t changes = wanted_features ^ netdev->features;
2148         bool enable = !!(wanted_features & feature);
2149         int err;
2150
2151         if (!(changes & feature))
2152                 return 0;
2153
2154         err = feature_handler(netdev, enable);
2155         if (err) {
2156                 netdev_err(netdev, "%s feature 0x%llx failed err %d\n",
2157                            enable ? "Enable" : "Disable", feature, err);
2158                 return err;
2159         }
2160
2161         MLX5E_SET_FEATURE(netdev, feature, enable);
2162         return 0;
2163 }
2164
2165 static int mlx5e_set_features(struct net_device *netdev,
2166                               netdev_features_t features)
2167 {
2168         int err;
2169
2170         err  = mlx5e_handle_feature(netdev, features, NETIF_F_LRO,
2171                                     set_feature_lro);
2172         err |= mlx5e_handle_feature(netdev, features,
2173                                     NETIF_F_HW_VLAN_CTAG_FILTER,
2174                                     set_feature_vlan_filter);
2175         err |= mlx5e_handle_feature(netdev, features, NETIF_F_HW_TC,
2176                                     set_feature_tc_num_filters);
2177
2178         return err ? -EINVAL : 0;
2179 }
2180
2181 static int mlx5e_change_mtu(struct net_device *netdev, int new_mtu)
2182 {
2183         struct mlx5e_priv *priv = netdev_priv(netdev);
2184         struct mlx5_core_dev *mdev = priv->mdev;
2185         bool was_opened;
2186         int max_mtu;
2187         int err = 0;
2188
2189         mlx5_query_port_max_mtu(mdev, &max_mtu, 1);
2190
2191         max_mtu = MLX5E_HW2SW_MTU(max_mtu);
2192
2193         if (new_mtu > max_mtu) {
2194                 netdev_err(netdev,
2195                            "%s: Bad MTU (%d) > (%d) Max\n",
2196                            __func__, new_mtu, max_mtu);
2197                 return -EINVAL;
2198         }
2199
2200         mutex_lock(&priv->state_lock);
2201
2202         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2203         if (was_opened)
2204                 mlx5e_close_locked(netdev);
2205
2206         netdev->mtu = new_mtu;
2207
2208         if (was_opened)
2209                 err = mlx5e_open_locked(netdev);
2210
2211         mutex_unlock(&priv->state_lock);
2212
2213         return err;
2214 }
2215
2216 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
2217 {
2218         switch (cmd) {
2219         case SIOCSHWTSTAMP:
2220                 return mlx5e_hwstamp_set(dev, ifr);
2221         case SIOCGHWTSTAMP:
2222                 return mlx5e_hwstamp_get(dev, ifr);
2223         default:
2224                 return -EOPNOTSUPP;
2225         }
2226 }
2227
2228 static int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
2229 {
2230         struct mlx5e_priv *priv = netdev_priv(dev);
2231         struct mlx5_core_dev *mdev = priv->mdev;
2232
2233         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
2234 }
2235
2236 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos)
2237 {
2238         struct mlx5e_priv *priv = netdev_priv(dev);
2239         struct mlx5_core_dev *mdev = priv->mdev;
2240
2241         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
2242                                            vlan, qos);
2243 }
2244
2245 static int mlx5_vport_link2ifla(u8 esw_link)
2246 {
2247         switch (esw_link) {
2248         case MLX5_ESW_VPORT_ADMIN_STATE_DOWN:
2249                 return IFLA_VF_LINK_STATE_DISABLE;
2250         case MLX5_ESW_VPORT_ADMIN_STATE_UP:
2251                 return IFLA_VF_LINK_STATE_ENABLE;
2252         }
2253         return IFLA_VF_LINK_STATE_AUTO;
2254 }
2255
2256 static int mlx5_ifla_link2vport(u8 ifla_link)
2257 {
2258         switch (ifla_link) {
2259         case IFLA_VF_LINK_STATE_DISABLE:
2260                 return MLX5_ESW_VPORT_ADMIN_STATE_DOWN;
2261         case IFLA_VF_LINK_STATE_ENABLE:
2262                 return MLX5_ESW_VPORT_ADMIN_STATE_UP;
2263         }
2264         return MLX5_ESW_VPORT_ADMIN_STATE_AUTO;
2265 }
2266
2267 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
2268                                    int link_state)
2269 {
2270         struct mlx5e_priv *priv = netdev_priv(dev);
2271         struct mlx5_core_dev *mdev = priv->mdev;
2272
2273         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
2274                                             mlx5_ifla_link2vport(link_state));
2275 }
2276
2277 static int mlx5e_get_vf_config(struct net_device *dev,
2278                                int vf, struct ifla_vf_info *ivi)
2279 {
2280         struct mlx5e_priv *priv = netdev_priv(dev);
2281         struct mlx5_core_dev *mdev = priv->mdev;
2282         int err;
2283
2284         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
2285         if (err)
2286                 return err;
2287         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
2288         return 0;
2289 }
2290
2291 static int mlx5e_get_vf_stats(struct net_device *dev,
2292                               int vf, struct ifla_vf_stats *vf_stats)
2293 {
2294         struct mlx5e_priv *priv = netdev_priv(dev);
2295         struct mlx5_core_dev *mdev = priv->mdev;
2296
2297         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
2298                                             vf_stats);
2299 }
2300
2301 static void mlx5e_add_vxlan_port(struct net_device *netdev,
2302                                  sa_family_t sa_family, __be16 port)
2303 {
2304         struct mlx5e_priv *priv = netdev_priv(netdev);
2305
2306         if (!mlx5e_vxlan_allowed(priv->mdev))
2307                 return;
2308
2309         mlx5e_vxlan_add_port(priv, be16_to_cpu(port));
2310 }
2311
2312 static void mlx5e_del_vxlan_port(struct net_device *netdev,
2313                                  sa_family_t sa_family, __be16 port)
2314 {
2315         struct mlx5e_priv *priv = netdev_priv(netdev);
2316
2317         if (!mlx5e_vxlan_allowed(priv->mdev))
2318                 return;
2319
2320         mlx5e_vxlan_del_port(priv, be16_to_cpu(port));
2321 }
2322
2323 static netdev_features_t mlx5e_vxlan_features_check(struct mlx5e_priv *priv,
2324                                                     struct sk_buff *skb,
2325                                                     netdev_features_t features)
2326 {
2327         struct udphdr *udph;
2328         u16 proto;
2329         u16 port = 0;
2330
2331         switch (vlan_get_protocol(skb)) {
2332         case htons(ETH_P_IP):
2333                 proto = ip_hdr(skb)->protocol;
2334                 break;
2335         case htons(ETH_P_IPV6):
2336                 proto = ipv6_hdr(skb)->nexthdr;
2337                 break;
2338         default:
2339                 goto out;
2340         }
2341
2342         if (proto == IPPROTO_UDP) {
2343                 udph = udp_hdr(skb);
2344                 port = be16_to_cpu(udph->dest);
2345         }
2346
2347         /* Verify if UDP port is being offloaded by HW */
2348         if (port && mlx5e_vxlan_lookup_port(priv, port))
2349                 return features;
2350
2351 out:
2352         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
2353         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2354 }
2355
2356 static netdev_features_t mlx5e_features_check(struct sk_buff *skb,
2357                                               struct net_device *netdev,
2358                                               netdev_features_t features)
2359 {
2360         struct mlx5e_priv *priv = netdev_priv(netdev);
2361
2362         features = vlan_features_check(skb, features);
2363         features = vxlan_features_check(skb, features);
2364
2365         /* Validate if the tunneled packet is being offloaded by HW */
2366         if (skb->encapsulation &&
2367             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
2368                 return mlx5e_vxlan_features_check(priv, skb, features);
2369
2370         return features;
2371 }
2372
2373 static const struct net_device_ops mlx5e_netdev_ops_basic = {
2374         .ndo_open                = mlx5e_open,
2375         .ndo_stop                = mlx5e_close,
2376         .ndo_start_xmit          = mlx5e_xmit,
2377         .ndo_setup_tc            = mlx5e_ndo_setup_tc,
2378         .ndo_select_queue        = mlx5e_select_queue,
2379         .ndo_get_stats64         = mlx5e_get_stats,
2380         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
2381         .ndo_set_mac_address     = mlx5e_set_mac,
2382         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
2383         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
2384         .ndo_set_features        = mlx5e_set_features,
2385         .ndo_change_mtu          = mlx5e_change_mtu,
2386         .ndo_do_ioctl            = mlx5e_ioctl,
2387 };
2388
2389 static const struct net_device_ops mlx5e_netdev_ops_sriov = {
2390         .ndo_open                = mlx5e_open,
2391         .ndo_stop                = mlx5e_close,
2392         .ndo_start_xmit          = mlx5e_xmit,
2393         .ndo_setup_tc            = mlx5e_ndo_setup_tc,
2394         .ndo_select_queue        = mlx5e_select_queue,
2395         .ndo_get_stats64         = mlx5e_get_stats,
2396         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
2397         .ndo_set_mac_address     = mlx5e_set_mac,
2398         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
2399         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
2400         .ndo_set_features        = mlx5e_set_features,
2401         .ndo_change_mtu          = mlx5e_change_mtu,
2402         .ndo_do_ioctl            = mlx5e_ioctl,
2403         .ndo_add_vxlan_port      = mlx5e_add_vxlan_port,
2404         .ndo_del_vxlan_port      = mlx5e_del_vxlan_port,
2405         .ndo_features_check      = mlx5e_features_check,
2406         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
2407         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
2408         .ndo_get_vf_config       = mlx5e_get_vf_config,
2409         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
2410         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
2411 };
2412
2413 static int mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
2414 {
2415         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
2416                 return -ENOTSUPP;
2417         if (!MLX5_CAP_GEN(mdev, eth_net_offloads) ||
2418             !MLX5_CAP_GEN(mdev, nic_flow_table) ||
2419             !MLX5_CAP_ETH(mdev, csum_cap) ||
2420             !MLX5_CAP_ETH(mdev, max_lso_cap) ||
2421             !MLX5_CAP_ETH(mdev, vlan_cap) ||
2422             !MLX5_CAP_ETH(mdev, rss_ind_tbl_cap) ||
2423             MLX5_CAP_FLOWTABLE(mdev,
2424                                flow_table_properties_nic_receive.max_ft_level)
2425                                < 3) {
2426                 mlx5_core_warn(mdev,
2427                                "Not creating net device, some required device capabilities are missing\n");
2428                 return -ENOTSUPP;
2429         }
2430         if (!MLX5_CAP_ETH(mdev, self_lb_en_modifiable))
2431                 mlx5_core_warn(mdev, "Self loop back prevention is not supported\n");
2432         if (!MLX5_CAP_GEN(mdev, cq_moderation))
2433                 mlx5_core_warn(mdev, "CQ modiration is not supported\n");
2434
2435         return 0;
2436 }
2437
2438 u16 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev)
2439 {
2440         int bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
2441
2442         return bf_buf_size -
2443                sizeof(struct mlx5e_tx_wqe) +
2444                2 /*sizeof(mlx5e_tx_wqe.inline_hdr_start)*/;
2445 }
2446
2447 #ifdef CONFIG_MLX5_CORE_EN_DCB
2448 static void mlx5e_ets_init(struct mlx5e_priv *priv)
2449 {
2450         int i;
2451
2452         priv->params.ets.ets_cap = mlx5_max_tc(priv->mdev) + 1;
2453         for (i = 0; i < priv->params.ets.ets_cap; i++) {
2454                 priv->params.ets.tc_tx_bw[i] = MLX5E_MAX_BW_ALLOC;
2455                 priv->params.ets.tc_tsa[i] = IEEE_8021QAZ_TSA_VENDOR;
2456                 priv->params.ets.prio_tc[i] = i;
2457         }
2458
2459         /* tclass[prio=0]=1, tclass[prio=1]=0, tclass[prio=i]=i (for i>1) */
2460         priv->params.ets.prio_tc[0] = 1;
2461         priv->params.ets.prio_tc[1] = 0;
2462 }
2463 #endif
2464
2465 void mlx5e_build_default_indir_rqt(struct mlx5_core_dev *mdev,
2466                                    u32 *indirection_rqt, int len,
2467                                    int num_channels)
2468 {
2469         int node = mdev->priv.numa_node;
2470         int node_num_of_cores;
2471         int i;
2472
2473         if (node == -1)
2474                 node = first_online_node;
2475
2476         node_num_of_cores = cpumask_weight(cpumask_of_node(node));
2477
2478         if (node_num_of_cores)
2479                 num_channels = min_t(int, num_channels, node_num_of_cores);
2480
2481         for (i = 0; i < len; i++)
2482                 indirection_rqt[i] = i % num_channels;
2483 }
2484
2485 static bool mlx5e_check_fragmented_striding_rq_cap(struct mlx5_core_dev *mdev)
2486 {
2487         return MLX5_CAP_GEN(mdev, striding_rq) &&
2488                 MLX5_CAP_GEN(mdev, umr_ptr_rlky) &&
2489                 MLX5_CAP_ETH(mdev, reg_umr_sq);
2490 }
2491
2492 static void mlx5e_build_netdev_priv(struct mlx5_core_dev *mdev,
2493                                     struct net_device *netdev,
2494                                     int num_channels)
2495 {
2496         struct mlx5e_priv *priv = netdev_priv(netdev);
2497
2498         priv->params.log_sq_size           =
2499                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
2500         priv->params.rq_wq_type = mlx5e_check_fragmented_striding_rq_cap(mdev) ?
2501                 MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ :
2502                 MLX5_WQ_TYPE_LINKED_LIST;
2503
2504         switch (priv->params.rq_wq_type) {
2505         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2506                 priv->params.log_rq_size = MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE_MPW;
2507                 priv->params.lro_en = true;
2508                 break;
2509         default: /* MLX5_WQ_TYPE_LINKED_LIST */
2510                 priv->params.log_rq_size = MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
2511         }
2512
2513         priv->params.min_rx_wqes = mlx5_min_rx_wqes(priv->params.rq_wq_type,
2514                                             BIT(priv->params.log_rq_size));
2515         priv->params.rx_cq_moderation_usec =
2516                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
2517         priv->params.rx_cq_moderation_pkts =
2518                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
2519         priv->params.tx_cq_moderation_usec =
2520                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
2521         priv->params.tx_cq_moderation_pkts =
2522                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
2523         priv->params.tx_max_inline         = mlx5e_get_max_inline_cap(mdev);
2524         priv->params.num_tc                = 1;
2525         priv->params.rss_hfunc             = ETH_RSS_HASH_XOR;
2526
2527         netdev_rss_key_fill(priv->params.toeplitz_hash_key,
2528                             sizeof(priv->params.toeplitz_hash_key));
2529
2530         mlx5e_build_default_indir_rqt(mdev, priv->params.indirection_rqt,
2531                                       MLX5E_INDIR_RQT_SIZE, num_channels);
2532
2533         priv->params.lro_wqe_sz            =
2534                 MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ;
2535
2536         priv->mdev                         = mdev;
2537         priv->netdev                       = netdev;
2538         priv->params.num_channels          = num_channels;
2539
2540 #ifdef CONFIG_MLX5_CORE_EN_DCB
2541         mlx5e_ets_init(priv);
2542 #endif
2543
2544         mutex_init(&priv->state_lock);
2545
2546         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
2547         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
2548         INIT_DELAYED_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
2549 }
2550
2551 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
2552 {
2553         struct mlx5e_priv *priv = netdev_priv(netdev);
2554
2555         mlx5_query_nic_vport_mac_address(priv->mdev, 0, netdev->dev_addr);
2556         if (is_zero_ether_addr(netdev->dev_addr) &&
2557             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
2558                 eth_hw_addr_random(netdev);
2559                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
2560         }
2561 }
2562
2563 static void mlx5e_build_netdev(struct net_device *netdev)
2564 {
2565         struct mlx5e_priv *priv = netdev_priv(netdev);
2566         struct mlx5_core_dev *mdev = priv->mdev;
2567
2568         SET_NETDEV_DEV(netdev, &mdev->pdev->dev);
2569
2570         if (MLX5_CAP_GEN(mdev, vport_group_manager)) {
2571                 netdev->netdev_ops = &mlx5e_netdev_ops_sriov;
2572 #ifdef CONFIG_MLX5_CORE_EN_DCB
2573                 netdev->dcbnl_ops = &mlx5e_dcbnl_ops;
2574 #endif
2575         } else {
2576                 netdev->netdev_ops = &mlx5e_netdev_ops_basic;
2577         }
2578
2579         netdev->watchdog_timeo    = 15 * HZ;
2580
2581         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
2582
2583         netdev->vlan_features    |= NETIF_F_SG;
2584         netdev->vlan_features    |= NETIF_F_IP_CSUM;
2585         netdev->vlan_features    |= NETIF_F_IPV6_CSUM;
2586         netdev->vlan_features    |= NETIF_F_GRO;
2587         netdev->vlan_features    |= NETIF_F_TSO;
2588         netdev->vlan_features    |= NETIF_F_TSO6;
2589         netdev->vlan_features    |= NETIF_F_RXCSUM;
2590         netdev->vlan_features    |= NETIF_F_RXHASH;
2591
2592         if (!!MLX5_CAP_ETH(mdev, lro_cap))
2593                 netdev->vlan_features    |= NETIF_F_LRO;
2594
2595         netdev->hw_features       = netdev->vlan_features;
2596         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
2597         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
2598         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
2599
2600         if (mlx5e_vxlan_allowed(mdev)) {
2601                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL;
2602                 netdev->hw_enc_features |= NETIF_F_IP_CSUM;
2603                 netdev->hw_enc_features |= NETIF_F_RXCSUM;
2604                 netdev->hw_enc_features |= NETIF_F_TSO;
2605                 netdev->hw_enc_features |= NETIF_F_TSO6;
2606                 netdev->hw_enc_features |= NETIF_F_RXHASH;
2607                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL;
2608         }
2609
2610         netdev->features          = netdev->hw_features;
2611         if (!priv->params.lro_en)
2612                 netdev->features  &= ~NETIF_F_LRO;
2613
2614 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
2615         if (FT_CAP(flow_modify_en) &&
2616             FT_CAP(modify_root) &&
2617             FT_CAP(identified_miss_table_mode) &&
2618             FT_CAP(flow_table_modify))
2619                 priv->netdev->hw_features      |= NETIF_F_HW_TC;
2620
2621         netdev->features         |= NETIF_F_HIGHDMA;
2622
2623         netdev->priv_flags       |= IFF_UNICAST_FLT;
2624
2625         mlx5e_set_netdev_dev_addr(netdev);
2626 }
2627
2628 static int mlx5e_create_mkey(struct mlx5e_priv *priv, u32 pdn,
2629                              struct mlx5_core_mkey *mkey)
2630 {
2631         struct mlx5_core_dev *mdev = priv->mdev;
2632         struct mlx5_create_mkey_mbox_in *in;
2633         int err;
2634
2635         in = mlx5_vzalloc(sizeof(*in));
2636         if (!in)
2637                 return -ENOMEM;
2638
2639         in->seg.flags = MLX5_PERM_LOCAL_WRITE |
2640                         MLX5_PERM_LOCAL_READ  |
2641                         MLX5_ACCESS_MODE_PA;
2642         in->seg.flags_pd = cpu_to_be32(pdn | MLX5_MKEY_LEN64);
2643         in->seg.qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
2644
2645         err = mlx5_core_create_mkey(mdev, mkey, in, sizeof(*in), NULL, NULL,
2646                                     NULL);
2647
2648         kvfree(in);
2649
2650         return err;
2651 }
2652
2653 static void mlx5e_create_q_counter(struct mlx5e_priv *priv)
2654 {
2655         struct mlx5_core_dev *mdev = priv->mdev;
2656         int err;
2657
2658         err = mlx5_core_alloc_q_counter(mdev, &priv->q_counter);
2659         if (err) {
2660                 mlx5_core_warn(mdev, "alloc queue counter failed, %d\n", err);
2661                 priv->q_counter = 0;
2662         }
2663 }
2664
2665 static void mlx5e_destroy_q_counter(struct mlx5e_priv *priv)
2666 {
2667         if (!priv->q_counter)
2668                 return;
2669
2670         mlx5_core_dealloc_q_counter(priv->mdev, priv->q_counter);
2671 }
2672
2673 static int mlx5e_create_umr_mkey(struct mlx5e_priv *priv)
2674 {
2675         struct mlx5_core_dev *mdev = priv->mdev;
2676         struct mlx5_create_mkey_mbox_in *in;
2677         struct mlx5_mkey_seg *mkc;
2678         int inlen = sizeof(*in);
2679         u64 npages =
2680                 mlx5e_get_max_num_channels(mdev) * MLX5_CHANNEL_MAX_NUM_MTTS;
2681         int err;
2682
2683         in = mlx5_vzalloc(inlen);
2684         if (!in)
2685                 return -ENOMEM;
2686
2687         mkc = &in->seg;
2688         mkc->status = MLX5_MKEY_STATUS_FREE;
2689         mkc->flags = MLX5_PERM_UMR_EN |
2690                      MLX5_PERM_LOCAL_READ |
2691                      MLX5_PERM_LOCAL_WRITE |
2692                      MLX5_ACCESS_MODE_MTT;
2693
2694         mkc->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
2695         mkc->flags_pd = cpu_to_be32(priv->pdn);
2696         mkc->len = cpu_to_be64(npages << PAGE_SHIFT);
2697         mkc->xlt_oct_size = cpu_to_be32(mlx5e_get_mtt_octw(npages));
2698         mkc->log2_page_size = PAGE_SHIFT;
2699
2700         err = mlx5_core_create_mkey(mdev, &priv->umr_mkey, in, inlen, NULL,
2701                                     NULL, NULL);
2702
2703         kvfree(in);
2704
2705         return err;
2706 }
2707
2708 static void *mlx5e_create_netdev(struct mlx5_core_dev *mdev)
2709 {
2710         struct net_device *netdev;
2711         struct mlx5e_priv *priv;
2712         int nch = mlx5e_get_max_num_channels(mdev);
2713         int err;
2714
2715         if (mlx5e_check_required_hca_cap(mdev))
2716                 return NULL;
2717
2718         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv),
2719                                     nch * MLX5E_MAX_NUM_TC,
2720                                     nch);
2721         if (!netdev) {
2722                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
2723                 return NULL;
2724         }
2725
2726         mlx5e_build_netdev_priv(mdev, netdev, nch);
2727         mlx5e_build_netdev(netdev);
2728
2729         netif_carrier_off(netdev);
2730
2731         priv = netdev_priv(netdev);
2732
2733         err = mlx5_alloc_map_uar(mdev, &priv->cq_uar, false);
2734         if (err) {
2735                 mlx5_core_err(mdev, "alloc_map uar failed, %d\n", err);
2736                 goto err_free_netdev;
2737         }
2738
2739         err = mlx5_core_alloc_pd(mdev, &priv->pdn);
2740         if (err) {
2741                 mlx5_core_err(mdev, "alloc pd failed, %d\n", err);
2742                 goto err_unmap_free_uar;
2743         }
2744
2745         err = mlx5_core_alloc_transport_domain(mdev, &priv->tdn);
2746         if (err) {
2747                 mlx5_core_err(mdev, "alloc td failed, %d\n", err);
2748                 goto err_dealloc_pd;
2749         }
2750
2751         err = mlx5e_create_mkey(priv, priv->pdn, &priv->mkey);
2752         if (err) {
2753                 mlx5_core_err(mdev, "create mkey failed, %d\n", err);
2754                 goto err_dealloc_transport_domain;
2755         }
2756
2757         err = mlx5e_create_umr_mkey(priv);
2758         if (err) {
2759                 mlx5_core_err(mdev, "create umr mkey failed, %d\n", err);
2760                 goto err_destroy_mkey;
2761         }
2762
2763         err = mlx5e_create_tises(priv);
2764         if (err) {
2765                 mlx5_core_warn(mdev, "create tises failed, %d\n", err);
2766                 goto err_destroy_umr_mkey;
2767         }
2768
2769         err = mlx5e_open_drop_rq(priv);
2770         if (err) {
2771                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
2772                 goto err_destroy_tises;
2773         }
2774
2775         err = mlx5e_create_rqt(priv, MLX5E_INDIRECTION_RQT);
2776         if (err) {
2777                 mlx5_core_warn(mdev, "create rqt(INDIR) failed, %d\n", err);
2778                 goto err_close_drop_rq;
2779         }
2780
2781         err = mlx5e_create_rqt(priv, MLX5E_SINGLE_RQ_RQT);
2782         if (err) {
2783                 mlx5_core_warn(mdev, "create rqt(SINGLE) failed, %d\n", err);
2784                 goto err_destroy_rqt_indir;
2785         }
2786
2787         err = mlx5e_create_tirs(priv);
2788         if (err) {
2789                 mlx5_core_warn(mdev, "create tirs failed, %d\n", err);
2790                 goto err_destroy_rqt_single;
2791         }
2792
2793         err = mlx5e_create_flow_tables(priv);
2794         if (err) {
2795                 mlx5_core_warn(mdev, "create flow tables failed, %d\n", err);
2796                 goto err_destroy_tirs;
2797         }
2798
2799         mlx5e_create_q_counter(priv);
2800
2801         mlx5e_init_eth_addr(priv);
2802
2803         mlx5e_vxlan_init(priv);
2804
2805         err = mlx5e_tc_init(priv);
2806         if (err)
2807                 goto err_dealloc_q_counters;
2808
2809 #ifdef CONFIG_MLX5_CORE_EN_DCB
2810         mlx5e_dcbnl_ieee_setets_core(priv, &priv->params.ets);
2811 #endif
2812
2813         err = register_netdev(netdev);
2814         if (err) {
2815                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
2816                 goto err_tc_cleanup;
2817         }
2818
2819         if (mlx5e_vxlan_allowed(mdev))
2820                 vxlan_get_rx_port(netdev);
2821
2822         mlx5e_enable_async_events(priv);
2823         schedule_work(&priv->set_rx_mode_work);
2824
2825         return priv;
2826
2827 err_tc_cleanup:
2828         mlx5e_tc_cleanup(priv);
2829
2830 err_dealloc_q_counters:
2831         mlx5e_destroy_q_counter(priv);
2832         mlx5e_destroy_flow_tables(priv);
2833
2834 err_destroy_tirs:
2835         mlx5e_destroy_tirs(priv);
2836
2837 err_destroy_rqt_single:
2838         mlx5e_destroy_rqt(priv, MLX5E_SINGLE_RQ_RQT);
2839
2840 err_destroy_rqt_indir:
2841         mlx5e_destroy_rqt(priv, MLX5E_INDIRECTION_RQT);
2842
2843 err_close_drop_rq:
2844         mlx5e_close_drop_rq(priv);
2845
2846 err_destroy_tises:
2847         mlx5e_destroy_tises(priv);
2848
2849 err_destroy_umr_mkey:
2850         mlx5_core_destroy_mkey(mdev, &priv->umr_mkey);
2851
2852 err_destroy_mkey:
2853         mlx5_core_destroy_mkey(mdev, &priv->mkey);
2854
2855 err_dealloc_transport_domain:
2856         mlx5_core_dealloc_transport_domain(mdev, priv->tdn);
2857
2858 err_dealloc_pd:
2859         mlx5_core_dealloc_pd(mdev, priv->pdn);
2860
2861 err_unmap_free_uar:
2862         mlx5_unmap_free_uar(mdev, &priv->cq_uar);
2863
2864 err_free_netdev:
2865         free_netdev(netdev);
2866
2867         return NULL;
2868 }
2869
2870 static void mlx5e_destroy_netdev(struct mlx5_core_dev *mdev, void *vpriv)
2871 {
2872         struct mlx5e_priv *priv = vpriv;
2873         struct net_device *netdev = priv->netdev;
2874
2875         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
2876
2877         schedule_work(&priv->set_rx_mode_work);
2878         mlx5e_disable_async_events(priv);
2879         flush_scheduled_work();
2880         unregister_netdev(netdev);
2881         mlx5e_tc_cleanup(priv);
2882         mlx5e_vxlan_cleanup(priv);
2883         mlx5e_destroy_q_counter(priv);
2884         mlx5e_destroy_flow_tables(priv);
2885         mlx5e_destroy_tirs(priv);
2886         mlx5e_destroy_rqt(priv, MLX5E_SINGLE_RQ_RQT);
2887         mlx5e_destroy_rqt(priv, MLX5E_INDIRECTION_RQT);
2888         mlx5e_close_drop_rq(priv);
2889         mlx5e_destroy_tises(priv);
2890         mlx5_core_destroy_mkey(priv->mdev, &priv->umr_mkey);
2891         mlx5_core_destroy_mkey(priv->mdev, &priv->mkey);
2892         mlx5_core_dealloc_transport_domain(priv->mdev, priv->tdn);
2893         mlx5_core_dealloc_pd(priv->mdev, priv->pdn);
2894         mlx5_unmap_free_uar(priv->mdev, &priv->cq_uar);
2895         free_netdev(netdev);
2896 }
2897
2898 static void *mlx5e_get_netdev(void *vpriv)
2899 {
2900         struct mlx5e_priv *priv = vpriv;
2901
2902         return priv->netdev;
2903 }
2904
2905 static struct mlx5_interface mlx5e_interface = {
2906         .add       = mlx5e_create_netdev,
2907         .remove    = mlx5e_destroy_netdev,
2908         .event     = mlx5e_async_event,
2909         .protocol  = MLX5_INTERFACE_PROTOCOL_ETH,
2910         .get_dev   = mlx5e_get_netdev,
2911 };
2912
2913 void mlx5e_init(void)
2914 {
2915         mlx5_register_interface(&mlx5e_interface);
2916 }
2917
2918 void mlx5e_cleanup(void)
2919 {
2920         mlx5_unregister_interface(&mlx5e_interface);
2921 }