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         bool                    am_enabled;
  46 };
  47
  48 struct mlx5e_sq_param {
  49         u32                        sqc[MLX5_ST_SZ_DW(sqc)];
  50         struct mlx5_wq_param       wq;
  51         u16                        max_inline;
  52         u8                         min_inline_mode;
  53         bool                       icosq;
  54 };
  55
  56 struct mlx5e_cq_param {
  57         u32                        cqc[MLX5_ST_SZ_DW(cqc)];
  58         struct mlx5_wq_param       wq;
  59         u16                        eq_ix;
  60         u8                         cq_period_mode;
  61 };
  62
  63 struct mlx5e_channel_param {
  64         struct mlx5e_rq_param      rq;
  65         struct mlx5e_sq_param      sq;
  66         struct mlx5e_sq_param      icosq;
  67         struct mlx5e_cq_param      rx_cq;
  68         struct mlx5e_cq_param      tx_cq;
  69         struct mlx5e_cq_param      icosq_cq;
  70 };
  71
  72 static void mlx5e_update_carrier(struct mlx5e_priv *priv)
  73 {
  74         struct mlx5_core_dev *mdev = priv->mdev;
  75         u8 port_state;
  76
  77         port_state = mlx5_query_vport_state(mdev,
  78                 MLX5_QUERY_VPORT_STATE_IN_OP_MOD_VNIC_VPORT, 0);
  79
  80         if (port_state == VPORT_STATE_UP) {
  81                 netdev_info(priv->netdev, "Link up\n");
  82                 netif_carrier_on(priv->netdev);
  83         } else {
  84                 netdev_info(priv->netdev, "Link down\n");
  85                 netif_carrier_off(priv->netdev);
  86         }
  87 }
  88
  89 static void mlx5e_update_carrier_work(struct work_struct *work)
  90 {
  91         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
  92                                                update_carrier_work);
  93
  94         mutex_lock(&priv->state_lock);
  95         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
  96                 mlx5e_update_carrier(priv);
  97         mutex_unlock(&priv->state_lock);
  98 }
  99
 100 static void mlx5e_tx_timeout_work(struct work_struct *work)
 101 {
 102         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 103                                                tx_timeout_work);
 104         int err;
 105
 106         rtnl_lock();
 107         mutex_lock(&priv->state_lock);
 108         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
 109                 goto unlock;
 110         mlx5e_close_locked(priv->netdev);
 111         err = mlx5e_open_locked(priv->netdev);
 112         if (err)
 113                 netdev_err(priv->netdev, "mlx5e_open_locked failed recovering from a tx_timeout, err(%d).\n",
 114                            err);
 115 unlock:
 116         mutex_unlock(&priv->state_lock);
 117         rtnl_unlock();
 118 }
 119
 120 static void mlx5e_update_sw_counters(struct mlx5e_priv *priv)
 121 {
 122         struct mlx5e_sw_stats *s = &priv->stats.sw;
 123         struct mlx5e_rq_stats *rq_stats;
 124         struct mlx5e_sq_stats *sq_stats;
 125         u64 tx_offload_none = 0;
 126         int i, j;
 127
 128         memset(s, 0, sizeof(*s));
 129         for (i = 0; i < priv->params.num_channels; i++) {
 130                 rq_stats = &priv->channel[i]->rq.stats;
 131
 132                 s->rx_packets   += rq_stats->packets;
 133                 s->rx_bytes     += rq_stats->bytes;
 134                 s->rx_lro_packets += rq_stats->lro_packets;
 135                 s->rx_lro_bytes += rq_stats->lro_bytes;
 136                 s->rx_csum_none += rq_stats->csum_none;
 137                 s->rx_csum_complete += rq_stats->csum_complete;
 138                 s->rx_csum_unnecessary_inner += rq_stats->csum_unnecessary_inner;
 139                 s->rx_wqe_err   += rq_stats->wqe_err;
 140                 s->rx_mpwqe_filler += rq_stats->mpwqe_filler;
 141                 s->rx_mpwqe_frag   += rq_stats->mpwqe_frag;
 142                 s->rx_buff_alloc_err += rq_stats->buff_alloc_err;
 143                 s->rx_cqe_compress_blks += rq_stats->cqe_compress_blks;
 144                 s->rx_cqe_compress_pkts += rq_stats->cqe_compress_pkts;
 145
 146                 for (j = 0; j < priv->params.num_tc; j++) {
 147                         sq_stats = &priv->channel[i]->sq[j].stats;
 148
 149                         s->tx_packets           += sq_stats->packets;
 150                         s->tx_bytes             += sq_stats->bytes;
 151                         s->tx_tso_packets       += sq_stats->tso_packets;
 152                         s->tx_tso_bytes         += sq_stats->tso_bytes;
 153                         s->tx_tso_inner_packets += sq_stats->tso_inner_packets;
 154                         s->tx_tso_inner_bytes   += sq_stats->tso_inner_bytes;
 155                         s->tx_queue_stopped     += sq_stats->stopped;
 156                         s->tx_queue_wake        += sq_stats->wake;
 157                         s->tx_queue_dropped     += sq_stats->dropped;
 158                         s->tx_xmit_more         += sq_stats->xmit_more;
 159                         s->tx_csum_partial_inner += sq_stats->csum_partial_inner;
 160                         tx_offload_none         += sq_stats->csum_none;
 161                 }
 162         }
 163
 164         /* Update calculated offload counters */
 165         s->tx_csum_partial = s->tx_packets - tx_offload_none - s->tx_csum_partial_inner;
 166         s->rx_csum_unnecessary = s->rx_packets - s->rx_csum_none - s->rx_csum_complete;
 167
 168         s->link_down_events_phy = MLX5_GET(ppcnt_reg,
 169                                 priv->stats.pport.phy_counters,
 170                                 counter_set.phys_layer_cntrs.link_down_events);
 171 }
 172
 173 static void mlx5e_update_vport_counters(struct mlx5e_priv *priv)
 174 {
 175         int outlen = MLX5_ST_SZ_BYTES(query_vport_counter_out);
 176         u32 *out = (u32 *)priv->stats.vport.query_vport_out;
 177         u32 in[MLX5_ST_SZ_DW(query_vport_counter_in)] = {0};
 178         struct mlx5_core_dev *mdev = priv->mdev;
 179
 180         MLX5_SET(query_vport_counter_in, in, opcode,
 181                  MLX5_CMD_OP_QUERY_VPORT_COUNTER);
 182         MLX5_SET(query_vport_counter_in, in, op_mod, 0);
 183         MLX5_SET(query_vport_counter_in, in, other_vport, 0);
 184
 185         memset(out, 0, outlen);
 186         mlx5_cmd_exec(mdev, in, sizeof(in), out, outlen);
 187 }
 188
 189 static void mlx5e_update_pport_counters(struct mlx5e_priv *priv)
 190 {
 191         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
 192         struct mlx5_core_dev *mdev = priv->mdev;
 193         int sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
 194         int prio;
 195         void *out;
 196         u32 *in;
 197
 198         in = mlx5_vzalloc(sz);
 199         if (!in)
 200                 goto free_out;
 201
 202         MLX5_SET(ppcnt_reg, in, local_port, 1);
 203
 204         out = pstats->IEEE_802_3_counters;
 205         MLX5_SET(ppcnt_reg, in, grp, MLX5_IEEE_802_3_COUNTERS_GROUP);
 206         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 207
 208         out = pstats->RFC_2863_counters;
 209         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2863_COUNTERS_GROUP);
 210         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 211
 212         out = pstats->RFC_2819_counters;
 213         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2819_COUNTERS_GROUP);
 214         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 215
 216         out = pstats->phy_counters;
 217         MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_COUNTERS_GROUP);
 218         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 219
 220         MLX5_SET(ppcnt_reg, in, grp, MLX5_PER_PRIORITY_COUNTERS_GROUP);
 221         for (prio = 0; prio < NUM_PPORT_PRIO; prio++) {
 222                 out = pstats->per_prio_counters[prio];
 223                 MLX5_SET(ppcnt_reg, in, prio_tc, prio);
 224                 mlx5_core_access_reg(mdev, in, sz, out, sz,
 225                                      MLX5_REG_PPCNT, 0, 0);
 226         }
 227
 228 free_out:
 229         kvfree(in);
 230 }
 231
 232 static void mlx5e_update_q_counter(struct mlx5e_priv *priv)
 233 {
 234         struct mlx5e_qcounter_stats *qcnt = &priv->stats.qcnt;
 235
 236         if (!priv->q_counter)
 237                 return;
 238
 239         mlx5_core_query_out_of_buffer(priv->mdev, priv->q_counter,
 240                                       &qcnt->rx_out_of_buffer);
 241 }
 242
 243 void mlx5e_update_stats(struct mlx5e_priv *priv)
 244 {
 245         mlx5e_update_q_counter(priv);
 246         mlx5e_update_vport_counters(priv);
 247         mlx5e_update_pport_counters(priv);
 248         mlx5e_update_sw_counters(priv);
 249 }
 250
 251 void mlx5e_update_stats_work(struct work_struct *work)
 252 {
 253         struct delayed_work *dwork = to_delayed_work(work);
 254         struct mlx5e_priv *priv = container_of(dwork, struct mlx5e_priv,
 255                                                update_stats_work);
 256         mutex_lock(&priv->state_lock);
 257         if (test_bit(MLX5E_STATE_OPENED, &priv->state)) {
 258                 priv->profile->update_stats(priv);
 259                 queue_delayed_work(priv->wq, dwork,
 260                                    msecs_to_jiffies(MLX5E_UPDATE_STATS_INTERVAL));
 261         }
 262         mutex_unlock(&priv->state_lock);
 263 }
 264
 265 static void mlx5e_async_event(struct mlx5_core_dev *mdev, void *vpriv,
 266                               enum mlx5_dev_event event, unsigned long param)
 267 {
 268         struct mlx5e_priv *priv = vpriv;
 269
 270         if (!test_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLED, &priv->state))
 271                 return;
 272
 273         switch (event) {
 274         case MLX5_DEV_EVENT_PORT_UP:
 275         case MLX5_DEV_EVENT_PORT_DOWN:
 276                 queue_work(priv->wq, &priv->update_carrier_work);
 277                 break;
 278
 279         default:
 280                 break;
 281         }
 282 }
 283
 284 static void mlx5e_enable_async_events(struct mlx5e_priv *priv)
 285 {
 286         set_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLED, &priv->state);
 287 }
 288
 289 static void mlx5e_disable_async_events(struct mlx5e_priv *priv)
 290 {
 291         clear_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLED, &priv->state);
 292         synchronize_irq(mlx5_get_msix_vec(priv->mdev, MLX5_EQ_VEC_ASYNC));
 293 }
 294
 295 #define MLX5E_HW2SW_MTU(hwmtu) (hwmtu - (ETH_HLEN + VLAN_HLEN + ETH_FCS_LEN))
 296 #define MLX5E_SW2HW_MTU(swmtu) (swmtu + (ETH_HLEN + VLAN_HLEN + ETH_FCS_LEN))
 297
 298 static int mlx5e_create_rq(struct mlx5e_channel *c,
 299                            struct mlx5e_rq_param *param,
 300                            struct mlx5e_rq *rq)
 301 {
 302         struct mlx5e_priv *priv = c->priv;
 303         struct mlx5_core_dev *mdev = priv->mdev;
 304         void *rqc = param->rqc;
 305         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 306         u32 byte_count;
 307         int wq_sz;
 308         int err;
 309         int i;
 310
 311         param->wq.db_numa_node = cpu_to_node(c->cpu);
 312
 313         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
 314                                 &rq->wq_ctrl);
 315         if (err)
 316                 return err;
 317
 318         rq->wq.db = &rq->wq.db[MLX5_RCV_DBR];
 319
 320         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 321
 322         switch (priv->params.rq_wq_type) {
 323         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 324                 rq->wqe_info = kzalloc_node(wq_sz * sizeof(*rq->wqe_info),
 325                                             GFP_KERNEL, cpu_to_node(c->cpu));
 326                 if (!rq->wqe_info) {
 327                         err = -ENOMEM;
 328                         goto err_rq_wq_destroy;
 329                 }
 330                 rq->handle_rx_cqe = mlx5e_handle_rx_cqe_mpwrq;
 331                 rq->alloc_wqe = mlx5e_alloc_rx_mpwqe;
 332                 rq->dealloc_wqe = mlx5e_dealloc_rx_mpwqe;
 333
 334                 rq->mpwqe_mtt_offset = c->ix *
 335                         MLX5E_REQUIRED_MTTS(1, BIT(priv->params.log_rq_size));
 336
 337                 rq->mpwqe_stride_sz = BIT(priv->params.mpwqe_log_stride_sz);
 338                 rq->mpwqe_num_strides = BIT(priv->params.mpwqe_log_num_strides);
 339                 rq->wqe_sz = rq->mpwqe_stride_sz * rq->mpwqe_num_strides;
 340                 byte_count = rq->wqe_sz;
 341                 break;
 342         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 343                 rq->skb = kzalloc_node(wq_sz * sizeof(*rq->skb), GFP_KERNEL,
 344                                        cpu_to_node(c->cpu));
 345                 if (!rq->skb) {
 346                         err = -ENOMEM;
 347                         goto err_rq_wq_destroy;
 348                 }
 349                 rq->handle_rx_cqe = mlx5e_handle_rx_cqe;
 350                 rq->alloc_wqe = mlx5e_alloc_rx_wqe;
 351                 rq->dealloc_wqe = mlx5e_dealloc_rx_wqe;
 352
 353                 rq->wqe_sz = (priv->params.lro_en) ?
 354                                 priv->params.lro_wqe_sz :
 355                                 MLX5E_SW2HW_MTU(priv->netdev->mtu);
 356                 rq->wqe_sz = SKB_DATA_ALIGN(rq->wqe_sz);
 357                 byte_count = rq->wqe_sz;
 358                 byte_count |= MLX5_HW_START_PADDING;
 359         }
 360
 361         for (i = 0; i < wq_sz; i++) {
 362                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, i);
 363
 364                 wqe->data.byte_count = cpu_to_be32(byte_count);
 365         }
 366
 367         INIT_WORK(&rq->am.work, mlx5e_rx_am_work);
 368         rq->am.mode = priv->params.rx_cq_period_mode;
 369
 370         rq->wq_type = priv->params.rq_wq_type;
 371         rq->pdev    = c->pdev;
 372         rq->netdev  = c->netdev;
 373         rq->tstamp  = &priv->tstamp;
 374         rq->channel = c;
 375         rq->ix      = c->ix;
 376         rq->priv    = c->priv;
 377         rq->mkey_be = c->mkey_be;
 378         rq->umr_mkey_be = cpu_to_be32(c->priv->umr_mkey.key);
 379
 380         return 0;
 381
 382 err_rq_wq_destroy:
 383         mlx5_wq_destroy(&rq->wq_ctrl);
 384
 385         return err;
 386 }
 387
 388 static void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 389 {
 390         switch (rq->wq_type) {
 391         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 392                 kfree(rq->wqe_info);
 393                 break;
 394         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 395                 kfree(rq->skb);
 396         }
 397
 398         mlx5_wq_destroy(&rq->wq_ctrl);
 399 }
 400
 401 static int mlx5e_enable_rq(struct mlx5e_rq *rq, struct mlx5e_rq_param *param)
 402 {
 403         struct mlx5e_priv *priv = rq->priv;
 404         struct mlx5_core_dev *mdev = priv->mdev;
 405
 406         void *in;
 407         void *rqc;
 408         void *wq;
 409         int inlen;
 410         int err;
 411
 412         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 413                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 414         in = mlx5_vzalloc(inlen);
 415         if (!in)
 416                 return -ENOMEM;
 417
 418         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 419         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 420
 421         memcpy(rqc, param->rqc, sizeof(param->rqc));
 422
 423         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 424         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 425         MLX5_SET(rqc,  rqc, vsd, priv->params.vlan_strip_disable);
 426         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 427                                                 MLX5_ADAPTER_PAGE_SHIFT);
 428         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 429
 430         mlx5_fill_page_array(&rq->wq_ctrl.buf,
 431                              (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 432
 433         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 434
 435         kvfree(in);
 436
 437         return err;
 438 }
 439
 440 static int mlx5e_modify_rq_state(struct mlx5e_rq *rq, int curr_state,
 441                                  int next_state)
 442 {
 443         struct mlx5e_channel *c = rq->channel;
 444         struct mlx5e_priv *priv = c->priv;
 445         struct mlx5_core_dev *mdev = priv->mdev;
 446
 447         void *in;
 448         void *rqc;
 449         int inlen;
 450         int err;
 451
 452         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 453         in = mlx5_vzalloc(inlen);
 454         if (!in)
 455                 return -ENOMEM;
 456
 457         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 458
 459         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 460         MLX5_SET(rqc, rqc, state, next_state);
 461
 462         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 463
 464         kvfree(in);
 465
 466         return err;
 467 }
 468
 469 static int mlx5e_modify_rq_vsd(struct mlx5e_rq *rq, bool vsd)
 470 {
 471         struct mlx5e_channel *c = rq->channel;
 472         struct mlx5e_priv *priv = c->priv;
 473         struct mlx5_core_dev *mdev = priv->mdev;
 474
 475         void *in;
 476         void *rqc;
 477         int inlen;
 478         int err;
 479
 480         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 481         in = mlx5_vzalloc(inlen);
 482         if (!in)
 483                 return -ENOMEM;
 484
 485         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 486
 487         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 488         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 489                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
 490         MLX5_SET(rqc, rqc, vsd, vsd);
 491         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 492
 493         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 494
 495         kvfree(in);
 496
 497         return err;
 498 }
 499
 500 static void mlx5e_disable_rq(struct mlx5e_rq *rq)
 501 {
 502         mlx5_core_destroy_rq(rq->priv->mdev, rq->rqn);
 503 }
 504
 505 static int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq)
 506 {
 507         unsigned long exp_time = jiffies + msecs_to_jiffies(20000);
 508         struct mlx5e_channel *c = rq->channel;
 509         struct mlx5e_priv *priv = c->priv;
 510         struct mlx5_wq_ll *wq = &rq->wq;
 511
 512         while (time_before(jiffies, exp_time)) {
 513                 if (wq->cur_sz >= priv->params.min_rx_wqes)
 514                         return 0;
 515
 516                 msleep(20);
 517         }
 518
 519         return -ETIMEDOUT;
 520 }
 521
 522 static void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 523 {
 524         struct mlx5_wq_ll *wq = &rq->wq;
 525         struct mlx5e_rx_wqe *wqe;
 526         __be16 wqe_ix_be;
 527         u16 wqe_ix;
 528
 529         /* UMR WQE (if in progress) is always at wq->head */
 530         if (test_bit(MLX5E_RQ_STATE_UMR_WQE_IN_PROGRESS, &rq->state))
 531                 mlx5e_free_rx_fragmented_mpwqe(rq, &rq->wqe_info[wq->head]);
 532
 533         while (!mlx5_wq_ll_is_empty(wq)) {
 534                 wqe_ix_be = *wq->tail_next;
 535                 wqe_ix    = be16_to_cpu(wqe_ix_be);
 536                 wqe       = mlx5_wq_ll_get_wqe(&rq->wq, wqe_ix);
 537                 rq->dealloc_wqe(rq, wqe_ix);
 538                 mlx5_wq_ll_pop(&rq->wq, wqe_ix_be,
 539                                &wqe->next.next_wqe_index);
 540         }
 541 }
 542
 543 static int mlx5e_open_rq(struct mlx5e_channel *c,
 544                          struct mlx5e_rq_param *param,
 545                          struct mlx5e_rq *rq)
 546 {
 547         struct mlx5e_sq *sq = &c->icosq;
 548         u16 pi = sq->pc & sq->wq.sz_m1;
 549         int err;
 550
 551         err = mlx5e_create_rq(c, param, rq);
 552         if (err)
 553                 return err;
 554
 555         err = mlx5e_enable_rq(rq, param);
 556         if (err)
 557                 goto err_destroy_rq;
 558
 559         err = mlx5e_modify_rq_state(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 560         if (err)
 561                 goto err_disable_rq;
 562
 563         if (param->am_enabled)
 564                 set_bit(MLX5E_RQ_STATE_AM, &c->rq.state);
 565
 566         sq->ico_wqe_info[pi].opcode     = MLX5_OPCODE_NOP;
 567         sq->ico_wqe_info[pi].num_wqebbs = 1;
 568         mlx5e_send_nop(sq, true); /* trigger mlx5e_post_rx_wqes() */
 569
 570         return 0;
 571
 572 err_disable_rq:
 573         mlx5e_disable_rq(rq);
 574 err_destroy_rq:
 575         mlx5e_destroy_rq(rq);
 576
 577         return err;
 578 }
 579
 580 static void mlx5e_close_rq(struct mlx5e_rq *rq)
 581 {
 582         set_bit(MLX5E_RQ_STATE_FLUSH, &rq->state);
 583         napi_synchronize(&rq->channel->napi); /* prevent mlx5e_post_rx_wqes */
 584         cancel_work_sync(&rq->am.work);
 585
 586         mlx5e_disable_rq(rq);
 587         mlx5e_free_rx_descs(rq);
 588         mlx5e_destroy_rq(rq);
 589 }
 590
 591 static void mlx5e_free_sq_db(struct mlx5e_sq *sq)
 592 {
 593         kfree(sq->wqe_info);
 594         kfree(sq->dma_fifo);
 595         kfree(sq->skb);
 596 }
 597
 598 static int mlx5e_alloc_sq_db(struct mlx5e_sq *sq, int numa)
 599 {
 600         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 601         int df_sz = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
 602
 603         sq->skb = kzalloc_node(wq_sz * sizeof(*sq->skb), GFP_KERNEL, numa);
 604         sq->dma_fifo = kzalloc_node(df_sz * sizeof(*sq->dma_fifo), GFP_KERNEL,
 605                                     numa);
 606         sq->wqe_info = kzalloc_node(wq_sz * sizeof(*sq->wqe_info), GFP_KERNEL,
 607                                     numa);
 608
 609         if (!sq->skb || !sq->dma_fifo || !sq->wqe_info) {
 610                 mlx5e_free_sq_db(sq);
 611                 return -ENOMEM;
 612         }
 613
 614         sq->dma_fifo_mask = df_sz - 1;
 615
 616         return 0;
 617 }
 618
 619 static int mlx5e_create_sq(struct mlx5e_channel *c,
 620                            int tc,
 621                            struct mlx5e_sq_param *param,
 622                            struct mlx5e_sq *sq)
 623 {
 624         struct mlx5e_priv *priv = c->priv;
 625         struct mlx5_core_dev *mdev = priv->mdev;
 626
 627         void *sqc = param->sqc;
 628         void *sqc_wq = MLX5_ADDR_OF(sqc, sqc, wq);
 629         int err;
 630
 631         err = mlx5_alloc_map_uar(mdev, &sq->uar, !!MLX5_CAP_GEN(mdev, bf));
 632         if (err)
 633                 return err;
 634
 635         param->wq.db_numa_node = cpu_to_node(c->cpu);
 636
 637         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq,
 638                                  &sq->wq_ctrl);
 639         if (err)
 640                 goto err_unmap_free_uar;
 641
 642         sq->wq.db       = &sq->wq.db[MLX5_SND_DBR];
 643         if (sq->uar.bf_map) {
 644                 set_bit(MLX5E_SQ_STATE_BF_ENABLE, &sq->state);
 645                 sq->uar_map = sq->uar.bf_map;
 646         } else {
 647                 sq->uar_map = sq->uar.map;
 648         }
 649         sq->bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
 650         sq->max_inline  = param->max_inline;
 651         sq->min_inline_mode =
 652                 MLX5_CAP_ETH(mdev, wqe_inline_mode) == MLX5E_INLINE_MODE_VPORT_CONTEXT ?
 653                 param->min_inline_mode : 0;
 654
 655         err = mlx5e_alloc_sq_db(sq, cpu_to_node(c->cpu));
 656         if (err)
 657                 goto err_sq_wq_destroy;
 658
 659         if (param->icosq) {
 660                 u8 wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 661
 662                 sq->ico_wqe_info = kzalloc_node(sizeof(*sq->ico_wqe_info) *
 663                                                 wq_sz,
 664                                                 GFP_KERNEL,
 665                                                 cpu_to_node(c->cpu));
 666                 if (!sq->ico_wqe_info) {
 667                         err = -ENOMEM;
 668                         goto err_free_sq_db;
 669                 }
 670         } else {
 671                 int txq_ix;
 672
 673                 txq_ix = c->ix + tc * priv->params.num_channels;
 674                 sq->txq = netdev_get_tx_queue(priv->netdev, txq_ix);
 675                 priv->txq_to_sq_map[txq_ix] = sq;
 676         }
 677
 678         sq->pdev      = c->pdev;
 679         sq->tstamp    = &priv->tstamp;
 680         sq->mkey_be   = c->mkey_be;
 681         sq->channel   = c;
 682         sq->tc        = tc;
 683         sq->edge      = (sq->wq.sz_m1 + 1) - MLX5_SEND_WQE_MAX_WQEBBS;
 684         sq->bf_budget = MLX5E_SQ_BF_BUDGET;
 685
 686         return 0;
 687
 688 err_free_sq_db:
 689         mlx5e_free_sq_db(sq);
 690
 691 err_sq_wq_destroy:
 692         mlx5_wq_destroy(&sq->wq_ctrl);
 693
 694 err_unmap_free_uar:
 695         mlx5_unmap_free_uar(mdev, &sq->uar);
 696
 697         return err;
 698 }
 699
 700 static void mlx5e_destroy_sq(struct mlx5e_sq *sq)
 701 {
 702         struct mlx5e_channel *c = sq->channel;
 703         struct mlx5e_priv *priv = c->priv;
 704
 705         kfree(sq->ico_wqe_info);
 706         mlx5e_free_sq_db(sq);
 707         mlx5_wq_destroy(&sq->wq_ctrl);
 708         mlx5_unmap_free_uar(priv->mdev, &sq->uar);
 709 }
 710
 711 static int mlx5e_enable_sq(struct mlx5e_sq *sq, struct mlx5e_sq_param *param)
 712 {
 713         struct mlx5e_channel *c = sq->channel;
 714         struct mlx5e_priv *priv = c->priv;
 715         struct mlx5_core_dev *mdev = priv->mdev;
 716
 717         void *in;
 718         void *sqc;
 719         void *wq;
 720         int inlen;
 721         int err;
 722
 723         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
 724                 sizeof(u64) * sq->wq_ctrl.buf.npages;
 725         in = mlx5_vzalloc(inlen);
 726         if (!in)
 727                 return -ENOMEM;
 728
 729         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
 730         wq = MLX5_ADDR_OF(sqc, sqc, wq);
 731
 732         memcpy(sqc, param->sqc, sizeof(param->sqc));
 733
 734         MLX5_SET(sqc,  sqc, tis_num_0, param->icosq ? 0 : priv->tisn[sq->tc]);
 735         MLX5_SET(sqc,  sqc, cqn,                sq->cq.mcq.cqn);
 736         MLX5_SET(sqc,  sqc, min_wqe_inline_mode, sq->min_inline_mode);
 737         MLX5_SET(sqc,  sqc, state,              MLX5_SQC_STATE_RST);
 738         MLX5_SET(sqc,  sqc, tis_lst_sz,         param->icosq ? 0 : 1);
 739         MLX5_SET(sqc,  sqc, flush_in_error_en,  1);
 740
 741         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
 742         MLX5_SET(wq,   wq, uar_page,      sq->uar.index);
 743         MLX5_SET(wq,   wq, log_wq_pg_sz,  sq->wq_ctrl.buf.page_shift -
 744                                           MLX5_ADAPTER_PAGE_SHIFT);
 745         MLX5_SET64(wq, wq, dbr_addr,      sq->wq_ctrl.db.dma);
 746
 747         mlx5_fill_page_array(&sq->wq_ctrl.buf,
 748                              (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 749
 750         err = mlx5_core_create_sq(mdev, in, inlen, &sq->sqn);
 751
 752         kvfree(in);
 753
 754         return err;
 755 }
 756
 757 static int mlx5e_modify_sq(struct mlx5e_sq *sq, int curr_state,
 758                            int next_state, bool update_rl, int rl_index)
 759 {
 760         struct mlx5e_channel *c = sq->channel;
 761         struct mlx5e_priv *priv = c->priv;
 762         struct mlx5_core_dev *mdev = priv->mdev;
 763
 764         void *in;
 765         void *sqc;
 766         int inlen;
 767         int err;
 768
 769         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
 770         in = mlx5_vzalloc(inlen);
 771         if (!in)
 772                 return -ENOMEM;
 773
 774         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
 775
 776         MLX5_SET(modify_sq_in, in, sq_state, curr_state);
 777         MLX5_SET(sqc, sqc, state, next_state);
 778         if (update_rl && next_state == MLX5_SQC_STATE_RDY) {
 779                 MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
 780                 MLX5_SET(sqc,  sqc, packet_pacing_rate_limit_index, rl_index);
 781         }
 782
 783         err = mlx5_core_modify_sq(mdev, sq->sqn, in, inlen);
 784
 785         kvfree(in);
 786
 787         return err;
 788 }
 789
 790 static void mlx5e_disable_sq(struct mlx5e_sq *sq)
 791 {
 792         struct mlx5e_channel *c = sq->channel;
 793         struct mlx5e_priv *priv = c->priv;
 794         struct mlx5_core_dev *mdev = priv->mdev;
 795
 796         mlx5_core_destroy_sq(mdev, sq->sqn);
 797         if (sq->rate_limit)
 798                 mlx5_rl_remove_rate(mdev, sq->rate_limit);
 799 }
 800
 801 static int mlx5e_open_sq(struct mlx5e_channel *c,
 802                          int tc,
 803                          struct mlx5e_sq_param *param,
 804                          struct mlx5e_sq *sq)
 805 {
 806         int err;
 807
 808         err = mlx5e_create_sq(c, tc, param, sq);
 809         if (err)
 810                 return err;
 811
 812         err = mlx5e_enable_sq(sq, param);
 813         if (err)
 814                 goto err_destroy_sq;
 815
 816         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RST, MLX5_SQC_STATE_RDY,
 817                               false, 0);
 818         if (err)
 819                 goto err_disable_sq;
 820
 821         if (sq->txq) {
 822                 netdev_tx_reset_queue(sq->txq);
 823                 netif_tx_start_queue(sq->txq);
 824         }
 825
 826         return 0;
 827
 828 err_disable_sq:
 829         mlx5e_disable_sq(sq);
 830 err_destroy_sq:
 831         mlx5e_destroy_sq(sq);
 832
 833         return err;
 834 }
 835
 836 static inline void netif_tx_disable_queue(struct netdev_queue *txq)
 837 {
 838         __netif_tx_lock_bh(txq);
 839         netif_tx_stop_queue(txq);
 840         __netif_tx_unlock_bh(txq);
 841 }
 842
 843 static void mlx5e_close_sq(struct mlx5e_sq *sq)
 844 {
 845         set_bit(MLX5E_SQ_STATE_FLUSH, &sq->state);
 846         /* prevent netif_tx_wake_queue */
 847         napi_synchronize(&sq->channel->napi);
 848
 849         if (sq->txq) {
 850                 netif_tx_disable_queue(sq->txq);
 851
 852                 /* last doorbell out, godspeed .. */
 853                 if (mlx5e_sq_has_room_for(sq, 1))
 854                         mlx5e_send_nop(sq, true);
 855         }
 856
 857         mlx5e_disable_sq(sq);
 858         mlx5e_free_tx_descs(sq);
 859         mlx5e_destroy_sq(sq);
 860 }
 861
 862 static int mlx5e_create_cq(struct mlx5e_channel *c,
 863                            struct mlx5e_cq_param *param,
 864                            struct mlx5e_cq *cq)
 865 {
 866         struct mlx5e_priv *priv = c->priv;
 867         struct mlx5_core_dev *mdev = priv->mdev;
 868         struct mlx5_core_cq *mcq = &cq->mcq;
 869         int eqn_not_used;
 870         unsigned int irqn;
 871         int err;
 872         u32 i;
 873
 874         param->wq.buf_numa_node = cpu_to_node(c->cpu);
 875         param->wq.db_numa_node  = cpu_to_node(c->cpu);
 876         param->eq_ix   = c->ix;
 877
 878         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
 879                                &cq->wq_ctrl);
 880         if (err)
 881                 return err;
 882
 883         mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
 884
 885         cq->napi        = &c->napi;
 886
 887         mcq->cqe_sz     = 64;
 888         mcq->set_ci_db  = cq->wq_ctrl.db.db;
 889         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
 890         *mcq->set_ci_db = 0;
 891         *mcq->arm_db    = 0;
 892         mcq->vector     = param->eq_ix;
 893         mcq->comp       = mlx5e_completion_event;
 894         mcq->event      = mlx5e_cq_error_event;
 895         mcq->irqn       = irqn;
 896         mcq->uar        = &mdev->mlx5e_res.cq_uar;
 897
 898         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
 899                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
 900
 901                 cqe->op_own = 0xf1;
 902         }
 903
 904         cq->channel = c;
 905         cq->priv = priv;
 906
 907         return 0;
 908 }
 909
 910 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
 911 {
 912         mlx5_wq_destroy(&cq->wq_ctrl);
 913 }
 914
 915 static int mlx5e_enable_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
 916 {
 917         struct mlx5e_priv *priv = cq->priv;
 918         struct mlx5_core_dev *mdev = priv->mdev;
 919         struct mlx5_core_cq *mcq = &cq->mcq;
 920
 921         void *in;
 922         void *cqc;
 923         int inlen;
 924         unsigned int irqn_not_used;
 925         int eqn;
 926         int err;
 927
 928         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
 929                 sizeof(u64) * cq->wq_ctrl.buf.npages;
 930         in = mlx5_vzalloc(inlen);
 931         if (!in)
 932                 return -ENOMEM;
 933
 934         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
 935
 936         memcpy(cqc, param->cqc, sizeof(param->cqc));
 937
 938         mlx5_fill_page_array(&cq->wq_ctrl.buf,
 939                              (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
 940
 941         mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
 942
 943         MLX5_SET(cqc,   cqc, cq_period_mode, param->cq_period_mode);
 944         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
 945         MLX5_SET(cqc,   cqc, uar_page,      mcq->uar->index);
 946         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
 947                                             MLX5_ADAPTER_PAGE_SHIFT);
 948         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
 949
 950         err = mlx5_core_create_cq(mdev, mcq, in, inlen);
 951
 952         kvfree(in);
 953
 954         if (err)
 955                 return err;
 956
 957         mlx5e_cq_arm(cq);
 958
 959         return 0;
 960 }
 961
 962 static void mlx5e_disable_cq(struct mlx5e_cq *cq)
 963 {
 964         struct mlx5e_priv *priv = cq->priv;
 965         struct mlx5_core_dev *mdev = priv->mdev;
 966
 967         mlx5_core_destroy_cq(mdev, &cq->mcq);
 968 }
 969
 970 static int mlx5e_open_cq(struct mlx5e_channel *c,
 971                          struct mlx5e_cq_param *param,
 972                          struct mlx5e_cq *cq,
 973                          struct mlx5e_cq_moder moderation)
 974 {
 975         int err;
 976         struct mlx5e_priv *priv = c->priv;
 977         struct mlx5_core_dev *mdev = priv->mdev;
 978
 979         err = mlx5e_create_cq(c, param, cq);
 980         if (err)
 981                 return err;
 982
 983         err = mlx5e_enable_cq(cq, param);
 984         if (err)
 985                 goto err_destroy_cq;
 986
 987         if (MLX5_CAP_GEN(mdev, cq_moderation))
 988                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq,
 989                                                moderation.usec,
 990                                                moderation.pkts);
 991         return 0;
 992
 993 err_destroy_cq:
 994         mlx5e_destroy_cq(cq);
 995
 996         return err;
 997 }
 998
 999 static void mlx5e_close_cq(struct mlx5e_cq *cq)
1000 {
1001         mlx5e_disable_cq(cq);
1002         mlx5e_destroy_cq(cq);
1003 }
1004
1005 static int mlx5e_get_cpu(struct mlx5e_priv *priv, int ix)
1006 {
1007         return cpumask_first(priv->mdev->priv.irq_info[ix].mask);
1008 }
1009
1010 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1011                              struct mlx5e_channel_param *cparam)
1012 {
1013         struct mlx5e_priv *priv = c->priv;
1014         int err;
1015         int tc;
1016
1017         for (tc = 0; tc < c->num_tc; tc++) {
1018                 err = mlx5e_open_cq(c, &cparam->tx_cq, &c->sq[tc].cq,
1019                                     priv->params.tx_cq_moderation);
1020                 if (err)
1021                         goto err_close_tx_cqs;
1022         }
1023
1024         return 0;
1025
1026 err_close_tx_cqs:
1027         for (tc--; tc >= 0; tc--)
1028                 mlx5e_close_cq(&c->sq[tc].cq);
1029
1030         return err;
1031 }
1032
1033 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1034 {
1035         int tc;
1036
1037         for (tc = 0; tc < c->num_tc; tc++)
1038                 mlx5e_close_cq(&c->sq[tc].cq);
1039 }
1040
1041 static int mlx5e_open_sqs(struct mlx5e_channel *c,
1042                           struct mlx5e_channel_param *cparam)
1043 {
1044         int err;
1045         int tc;
1046
1047         for (tc = 0; tc < c->num_tc; tc++) {
1048                 err = mlx5e_open_sq(c, tc, &cparam->sq, &c->sq[tc]);
1049                 if (err)
1050                         goto err_close_sqs;
1051         }
1052
1053         return 0;
1054
1055 err_close_sqs:
1056         for (tc--; tc >= 0; tc--)
1057                 mlx5e_close_sq(&c->sq[tc]);
1058
1059         return err;
1060 }
1061
1062 static void mlx5e_close_sqs(struct mlx5e_channel *c)
1063 {
1064         int tc;
1065
1066         for (tc = 0; tc < c->num_tc; tc++)
1067                 mlx5e_close_sq(&c->sq[tc]);
1068 }
1069
1070 static void mlx5e_build_channeltc_to_txq_map(struct mlx5e_priv *priv, int ix)
1071 {
1072         int i;
1073
1074         for (i = 0; i < priv->profile->max_tc; i++)
1075                 priv->channeltc_to_txq_map[ix][i] =
1076                         ix + i * priv->params.num_channels;
1077 }
1078
1079 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1080                                 struct mlx5e_sq *sq, u32 rate)
1081 {
1082         struct mlx5e_priv *priv = netdev_priv(dev);
1083         struct mlx5_core_dev *mdev = priv->mdev;
1084         u16 rl_index = 0;
1085         int err;
1086
1087         if (rate == sq->rate_limit)
1088                 /* nothing to do */
1089                 return 0;
1090
1091         if (sq->rate_limit)
1092                 /* remove current rl index to free space to next ones */
1093                 mlx5_rl_remove_rate(mdev, sq->rate_limit);
1094
1095         sq->rate_limit = 0;
1096
1097         if (rate) {
1098                 err = mlx5_rl_add_rate(mdev, rate, &rl_index);
1099                 if (err) {
1100                         netdev_err(dev, "Failed configuring rate %u: %d\n",
1101                                    rate, err);
1102                         return err;
1103                 }
1104         }
1105
1106         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RDY,
1107                               MLX5_SQC_STATE_RDY, true, rl_index);
1108         if (err) {
1109                 netdev_err(dev, "Failed configuring rate %u: %d\n",
1110                            rate, err);
1111                 /* remove the rate from the table */
1112                 if (rate)
1113                         mlx5_rl_remove_rate(mdev, rate);
1114                 return err;
1115         }
1116
1117         sq->rate_limit = rate;
1118         return 0;
1119 }
1120
1121 static int mlx5e_set_tx_maxrate(struct net_device *dev, int index, u32 rate)
1122 {
1123         struct mlx5e_priv *priv = netdev_priv(dev);
1124         struct mlx5_core_dev *mdev = priv->mdev;
1125         struct mlx5e_sq *sq = priv->txq_to_sq_map[index];
1126         int err = 0;
1127
1128         if (!mlx5_rl_is_supported(mdev)) {
1129                 netdev_err(dev, "Rate limiting is not supported on this device\n");
1130                 return -EINVAL;
1131         }
1132
1133         /* rate is given in Mb/sec, HW config is in Kb/sec */
1134         rate = rate << 10;
1135
1136         /* Check whether rate in valid range, 0 is always valid */
1137         if (rate && !mlx5_rl_is_in_range(mdev, rate)) {
1138                 netdev_err(dev, "TX rate %u, is not in range\n", rate);
1139                 return -ERANGE;
1140         }
1141
1142         mutex_lock(&priv->state_lock);
1143         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
1144                 err = mlx5e_set_sq_maxrate(dev, sq, rate);
1145         if (!err)
1146                 priv->tx_rates[index] = rate;
1147         mutex_unlock(&priv->state_lock);
1148
1149         return err;
1150 }
1151
1152 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
1153                               struct mlx5e_channel_param *cparam,
1154                               struct mlx5e_channel **cp)
1155 {
1156         struct mlx5e_cq_moder icosq_cq_moder = {0, 0};
1157         struct net_device *netdev = priv->netdev;
1158         struct mlx5e_cq_moder rx_cq_profile;
1159         int cpu = mlx5e_get_cpu(priv, ix);
1160         struct mlx5e_channel *c;
1161         struct mlx5e_sq *sq;
1162         int err;
1163         int i;
1164
1165         c = kzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
1166         if (!c)
1167                 return -ENOMEM;
1168
1169         c->priv     = priv;
1170         c->ix       = ix;
1171         c->cpu      = cpu;
1172         c->pdev     = &priv->mdev->pdev->dev;
1173         c->netdev   = priv->netdev;
1174         c->mkey_be  = cpu_to_be32(priv->mdev->mlx5e_res.mkey.key);
1175         c->num_tc   = priv->params.num_tc;
1176
1177         if (priv->params.rx_am_enabled)
1178                 rx_cq_profile = mlx5e_am_get_def_profile(priv->params.rx_cq_period_mode);
1179         else
1180                 rx_cq_profile = priv->params.rx_cq_moderation;
1181
1182         mlx5e_build_channeltc_to_txq_map(priv, ix);
1183
1184         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
1185
1186         err = mlx5e_open_cq(c, &cparam->icosq_cq, &c->icosq.cq, icosq_cq_moder);
1187         if (err)
1188                 goto err_napi_del;
1189
1190         err = mlx5e_open_tx_cqs(c, cparam);
1191         if (err)
1192                 goto err_close_icosq_cq;
1193
1194         err = mlx5e_open_cq(c, &cparam->rx_cq, &c->rq.cq,
1195                             rx_cq_profile);
1196         if (err)
1197                 goto err_close_tx_cqs;
1198
1199         napi_enable(&c->napi);
1200
1201         err = mlx5e_open_sq(c, 0, &cparam->icosq, &c->icosq);
1202         if (err)
1203                 goto err_disable_napi;
1204
1205         err = mlx5e_open_sqs(c, cparam);
1206         if (err)
1207                 goto err_close_icosq;
1208
1209         for (i = 0; i < priv->params.num_tc; i++) {
1210                 u32 txq_ix = priv->channeltc_to_txq_map[ix][i];
1211
1212                 if (priv->tx_rates[txq_ix]) {
1213                         sq = priv->txq_to_sq_map[txq_ix];
1214                         mlx5e_set_sq_maxrate(priv->netdev, sq,
1215                                              priv->tx_rates[txq_ix]);
1216                 }
1217         }
1218
1219         err = mlx5e_open_rq(c, &cparam->rq, &c->rq);
1220         if (err)
1221                 goto err_close_sqs;
1222
1223         netif_set_xps_queue(netdev, get_cpu_mask(c->cpu), ix);
1224         *cp = c;
1225
1226         return 0;
1227
1228 err_close_sqs:
1229         mlx5e_close_sqs(c);
1230
1231 err_close_icosq:
1232         mlx5e_close_sq(&c->icosq);
1233
1234 err_disable_napi:
1235         napi_disable(&c->napi);
1236         mlx5e_close_cq(&c->rq.cq);
1237
1238 err_close_tx_cqs:
1239         mlx5e_close_tx_cqs(c);
1240
1241 err_close_icosq_cq:
1242         mlx5e_close_cq(&c->icosq.cq);
1243
1244 err_napi_del:
1245         netif_napi_del(&c->napi);
1246         napi_hash_del(&c->napi);
1247         kfree(c);
1248
1249         return err;
1250 }
1251
1252 static void mlx5e_close_channel(struct mlx5e_channel *c)
1253 {
1254         mlx5e_close_rq(&c->rq);
1255         mlx5e_close_sqs(c);
1256         mlx5e_close_sq(&c->icosq);
1257         napi_disable(&c->napi);
1258         mlx5e_close_cq(&c->rq.cq);
1259         mlx5e_close_tx_cqs(c);
1260         mlx5e_close_cq(&c->icosq.cq);
1261         netif_napi_del(&c->napi);
1262
1263         napi_hash_del(&c->napi);
1264         synchronize_rcu();
1265
1266         kfree(c);
1267 }
1268
1269 static void mlx5e_build_rq_param(struct mlx5e_priv *priv,
1270                                  struct mlx5e_rq_param *param)
1271 {
1272         void *rqc = param->rqc;
1273         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1274
1275         switch (priv->params.rq_wq_type) {
1276         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1277                 MLX5_SET(wq, wq, log_wqe_num_of_strides,
1278                          priv->params.mpwqe_log_num_strides - 9);
1279                 MLX5_SET(wq, wq, log_wqe_stride_size,
1280                          priv->params.mpwqe_log_stride_sz - 6);
1281                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
1282                 break;
1283         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1284                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1285         }
1286
1287         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1288         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1289         MLX5_SET(wq, wq, log_wq_sz,        priv->params.log_rq_size);
1290         MLX5_SET(wq, wq, pd,               priv->mdev->mlx5e_res.pdn);
1291         MLX5_SET(rqc, rqc, counter_set_id, priv->q_counter);
1292
1293         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1294         param->wq.linear = 1;
1295
1296         param->am_enabled = priv->params.rx_am_enabled;
1297 }
1298
1299 static void mlx5e_build_drop_rq_param(struct mlx5e_rq_param *param)
1300 {
1301         void *rqc = param->rqc;
1302         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1303
1304         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1305         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1306 }
1307
1308 static void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
1309                                         struct mlx5e_sq_param *param)
1310 {
1311         void *sqc = param->sqc;
1312         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1313
1314         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1315         MLX5_SET(wq, wq, pd,            priv->mdev->mlx5e_res.pdn);
1316
1317         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1318 }
1319
1320 static void mlx5e_build_sq_param(struct mlx5e_priv *priv,
1321                                  struct mlx5e_sq_param *param)
1322 {
1323         void *sqc = param->sqc;
1324         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1325
1326         mlx5e_build_sq_param_common(priv, param);
1327         MLX5_SET(wq, wq, log_wq_sz,     priv->params.log_sq_size);
1328
1329         param->max_inline = priv->params.tx_max_inline;
1330         param->min_inline_mode = priv->params.tx_min_inline_mode;
1331 }
1332
1333 static void mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
1334                                         struct mlx5e_cq_param *param)
1335 {
1336         void *cqc = param->cqc;
1337
1338         MLX5_SET(cqc, cqc, uar_page, priv->mdev->mlx5e_res.cq_uar.index);
1339 }
1340
1341 static void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
1342                                     struct mlx5e_cq_param *param)
1343 {
1344         void *cqc = param->cqc;
1345         u8 log_cq_size;
1346
1347         switch (priv->params.rq_wq_type) {
1348         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1349                 log_cq_size = priv->params.log_rq_size +
1350                         priv->params.mpwqe_log_num_strides;
1351                 break;
1352         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1353                 log_cq_size = priv->params.log_rq_size;
1354         }
1355
1356         MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
1357         if (priv->params.rx_cqe_compress) {
1358                 MLX5_SET(cqc, cqc, mini_cqe_res_format, MLX5_CQE_FORMAT_CSUM);
1359                 MLX5_SET(cqc, cqc, cqe_comp_en, 1);
1360         }
1361
1362         mlx5e_build_common_cq_param(priv, param);
1363
1364         param->cq_period_mode = priv->params.rx_cq_period_mode;
1365 }
1366
1367 static void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
1368                                     struct mlx5e_cq_param *param)
1369 {
1370         void *cqc = param->cqc;
1371
1372         MLX5_SET(cqc, cqc, log_cq_size, priv->params.log_sq_size);
1373
1374         mlx5e_build_common_cq_param(priv, param);
1375
1376         param->cq_period_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
1377 }
1378
1379 static void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
1380                                      struct mlx5e_cq_param *param,
1381                                      u8 log_wq_size)
1382 {
1383         void *cqc = param->cqc;
1384
1385         MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
1386
1387         mlx5e_build_common_cq_param(priv, param);
1388
1389         param->cq_period_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
1390 }
1391
1392 static void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
1393                                     struct mlx5e_sq_param *param,
1394                                     u8 log_wq_size)
1395 {
1396         void *sqc = param->sqc;
1397         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1398
1399         mlx5e_build_sq_param_common(priv, param);
1400
1401         MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
1402         MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(priv->mdev, reg_umr_sq));
1403
1404         param->icosq = true;
1405 }
1406
1407 static void mlx5e_build_channel_param(struct mlx5e_priv *priv, struct mlx5e_channel_param *cparam)
1408 {
1409         u8 icosq_log_wq_sz = MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
1410
1411         mlx5e_build_rq_param(priv, &cparam->rq);
1412         mlx5e_build_sq_param(priv, &cparam->sq);
1413         mlx5e_build_icosq_param(priv, &cparam->icosq, icosq_log_wq_sz);
1414         mlx5e_build_rx_cq_param(priv, &cparam->rx_cq);
1415         mlx5e_build_tx_cq_param(priv, &cparam->tx_cq);
1416         mlx5e_build_ico_cq_param(priv, &cparam->icosq_cq, icosq_log_wq_sz);
1417 }
1418
1419 static int mlx5e_open_channels(struct mlx5e_priv *priv)
1420 {
1421         struct mlx5e_channel_param *cparam;
1422         int nch = priv->params.num_channels;
1423         int err = -ENOMEM;
1424         int i;
1425         int j;
1426
1427         priv->channel = kcalloc(nch, sizeof(struct mlx5e_channel *),
1428                                 GFP_KERNEL);
1429
1430         priv->txq_to_sq_map = kcalloc(nch * priv->params.num_tc,
1431                                       sizeof(struct mlx5e_sq *), GFP_KERNEL);
1432
1433         cparam = kzalloc(sizeof(struct mlx5e_channel_param), GFP_KERNEL);
1434
1435         if (!priv->channel || !priv->txq_to_sq_map || !cparam)
1436                 goto err_free_txq_to_sq_map;
1437
1438         mlx5e_build_channel_param(priv, cparam);
1439
1440         for (i = 0; i < nch; i++) {
1441                 err = mlx5e_open_channel(priv, i, cparam, &priv->channel[i]);
1442                 if (err)
1443                         goto err_close_channels;
1444         }
1445
1446         for (j = 0; j < nch; j++) {
1447                 err = mlx5e_wait_for_min_rx_wqes(&priv->channel[j]->rq);
1448                 if (err)
1449                         goto err_close_channels;
1450         }
1451
1452         /* FIXME: This is a W/A for tx timeout watch dog false alarm when
1453          * polling for inactive tx queues.
1454          */
1455         netif_tx_start_all_queues(priv->netdev);
1456
1457         kfree(cparam);
1458         return 0;
1459
1460 err_close_channels:
1461         for (i--; i >= 0; i--)
1462                 mlx5e_close_channel(priv->channel[i]);
1463
1464 err_free_txq_to_sq_map:
1465         kfree(priv->txq_to_sq_map);
1466         kfree(priv->channel);
1467         kfree(cparam);
1468
1469         return err;
1470 }
1471
1472 static void mlx5e_close_channels(struct mlx5e_priv *priv)
1473 {
1474         int i;
1475
1476         /* FIXME: This is a W/A only for tx timeout watch dog false alarm when
1477          * polling for inactive tx queues.
1478          */
1479         netif_tx_stop_all_queues(priv->netdev);
1480         netif_tx_disable(priv->netdev);
1481
1482         for (i = 0; i < priv->params.num_channels; i++)
1483                 mlx5e_close_channel(priv->channel[i]);
1484
1485         kfree(priv->txq_to_sq_map);
1486         kfree(priv->channel);
1487 }
1488
1489 static int mlx5e_rx_hash_fn(int hfunc)
1490 {
1491         return (hfunc == ETH_RSS_HASH_TOP) ?
1492                MLX5_RX_HASH_FN_TOEPLITZ :
1493                MLX5_RX_HASH_FN_INVERTED_XOR8;
1494 }
1495
1496 static int mlx5e_bits_invert(unsigned long a, int size)
1497 {
1498         int inv = 0;
1499         int i;
1500
1501         for (i = 0; i < size; i++)
1502                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
1503
1504         return inv;
1505 }
1506
1507 static void mlx5e_fill_indir_rqt_rqns(struct mlx5e_priv *priv, void *rqtc)
1508 {
1509         int i;
1510
1511         for (i = 0; i < MLX5E_INDIR_RQT_SIZE; i++) {
1512                 int ix = i;
1513                 u32 rqn;
1514
1515                 if (priv->params.rss_hfunc == ETH_RSS_HASH_XOR)
1516                         ix = mlx5e_bits_invert(i, MLX5E_LOG_INDIR_RQT_SIZE);
1517
1518                 ix = priv->params.indirection_rqt[ix];
1519                 rqn = test_bit(MLX5E_STATE_OPENED, &priv->state) ?
1520                                 priv->channel[ix]->rq.rqn :
1521                                 priv->drop_rq.rqn;
1522                 MLX5_SET(rqtc, rqtc, rq_num[i], rqn);
1523         }
1524 }
1525
1526 static void mlx5e_fill_direct_rqt_rqn(struct mlx5e_priv *priv, void *rqtc,
1527                                       int ix)
1528 {
1529         u32 rqn = test_bit(MLX5E_STATE_OPENED, &priv->state) ?
1530                         priv->channel[ix]->rq.rqn :
1531                         priv->drop_rq.rqn;
1532
1533         MLX5_SET(rqtc, rqtc, rq_num[0], rqn);
1534 }
1535
1536 static int mlx5e_create_rqt(struct mlx5e_priv *priv, int sz,
1537                             int ix, struct mlx5e_rqt *rqt)
1538 {
1539         struct mlx5_core_dev *mdev = priv->mdev;
1540         void *rqtc;
1541         int inlen;
1542         int err;
1543         u32 *in;
1544
1545         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
1546         in = mlx5_vzalloc(inlen);
1547         if (!in)
1548                 return -ENOMEM;
1549
1550         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
1551
1552         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
1553         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
1554
1555         if (sz > 1) /* RSS */
1556                 mlx5e_fill_indir_rqt_rqns(priv, rqtc);
1557         else
1558                 mlx5e_fill_direct_rqt_rqn(priv, rqtc, ix);
1559
1560         err = mlx5_core_create_rqt(mdev, in, inlen, &rqt->rqtn);
1561         if (!err)
1562                 rqt->enabled = true;
1563
1564         kvfree(in);
1565         return err;
1566 }
1567
1568 void mlx5e_destroy_rqt(struct mlx5e_priv *priv, struct mlx5e_rqt *rqt)
1569 {
1570         rqt->enabled = false;
1571         mlx5_core_destroy_rqt(priv->mdev, rqt->rqtn);
1572 }
1573
1574 static int mlx5e_create_indirect_rqts(struct mlx5e_priv *priv)
1575 {
1576         struct mlx5e_rqt *rqt = &priv->indir_rqt;
1577
1578         return mlx5e_create_rqt(priv, MLX5E_INDIR_RQT_SIZE, 0, rqt);
1579 }
1580
1581 int mlx5e_create_direct_rqts(struct mlx5e_priv *priv)
1582 {
1583         struct mlx5e_rqt *rqt;
1584         int err;
1585         int ix;
1586
1587         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
1588                 rqt = &priv->direct_tir[ix].rqt;
1589                 err = mlx5e_create_rqt(priv, 1 /*size */, ix, rqt);
1590                 if (err)
1591                         goto err_destroy_rqts;
1592         }
1593
1594         return 0;
1595
1596 err_destroy_rqts:
1597         for (ix--; ix >= 0; ix--)
1598                 mlx5e_destroy_rqt(priv, &priv->direct_tir[ix].rqt);
1599
1600         return err;
1601 }
1602
1603 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, u32 rqtn, int sz, int ix)
1604 {
1605         struct mlx5_core_dev *mdev = priv->mdev;
1606         void *rqtc;
1607         int inlen;
1608         u32 *in;
1609         int err;
1610
1611         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
1612         in = mlx5_vzalloc(inlen);
1613         if (!in)
1614                 return -ENOMEM;
1615
1616         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
1617
1618         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
1619         if (sz > 1) /* RSS */
1620                 mlx5e_fill_indir_rqt_rqns(priv, rqtc);
1621         else
1622                 mlx5e_fill_direct_rqt_rqn(priv, rqtc, ix);
1623
1624         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
1625
1626         err = mlx5_core_modify_rqt(mdev, rqtn, in, inlen);
1627
1628         kvfree(in);
1629
1630         return err;
1631 }
1632
1633 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv)
1634 {
1635         u32 rqtn;
1636         int ix;
1637
1638         if (priv->indir_rqt.enabled) {
1639                 rqtn = priv->indir_rqt.rqtn;
1640                 mlx5e_redirect_rqt(priv, rqtn, MLX5E_INDIR_RQT_SIZE, 0);
1641         }
1642
1643         for (ix = 0; ix < priv->params.num_channels; ix++) {
1644                 if (!priv->direct_tir[ix].rqt.enabled)
1645                         continue;
1646                 rqtn = priv->direct_tir[ix].rqt.rqtn;
1647                 mlx5e_redirect_rqt(priv, rqtn, 1, ix);
1648         }
1649 }
1650
1651 static void mlx5e_build_tir_ctx_lro(void *tirc, struct mlx5e_priv *priv)
1652 {
1653         if (!priv->params.lro_en)
1654                 return;
1655
1656 #define ROUGH_MAX_L2_L3_HDR_SZ 256
1657
1658         MLX5_SET(tirc, tirc, lro_enable_mask,
1659                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
1660                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
1661         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
1662                  (priv->params.lro_wqe_sz -
1663                   ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
1664         MLX5_SET(tirc, tirc, lro_timeout_period_usecs,
1665                  MLX5_CAP_ETH(priv->mdev,
1666                               lro_timer_supported_periods[2]));
1667 }
1668
1669 void mlx5e_build_tir_ctx_hash(void *tirc, struct mlx5e_priv *priv)
1670 {
1671         MLX5_SET(tirc, tirc, rx_hash_fn,
1672                  mlx5e_rx_hash_fn(priv->params.rss_hfunc));
1673         if (priv->params.rss_hfunc == ETH_RSS_HASH_TOP) {
1674                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
1675                                              rx_hash_toeplitz_key);
1676                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
1677                                                rx_hash_toeplitz_key);
1678
1679                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
1680                 memcpy(rss_key, priv->params.toeplitz_hash_key, len);
1681         }
1682 }
1683
1684 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
1685 {
1686         struct mlx5_core_dev *mdev = priv->mdev;
1687
1688         void *in;
1689         void *tirc;
1690         int inlen;
1691         int err;
1692         int tt;
1693         int ix;
1694
1695         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
1696         in = mlx5_vzalloc(inlen);
1697         if (!in)
1698                 return -ENOMEM;
1699
1700         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
1701         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
1702
1703         mlx5e_build_tir_ctx_lro(tirc, priv);
1704
1705         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
1706                 err = mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in,
1707                                            inlen);
1708                 if (err)
1709                         goto free_in;
1710         }
1711
1712         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
1713                 err = mlx5_core_modify_tir(mdev, priv->direct_tir[ix].tirn,
1714                                            in, inlen);
1715                 if (err)
1716                         goto free_in;
1717         }
1718
1719 free_in:
1720         kvfree(in);
1721
1722         return err;
1723 }
1724
1725 static int mlx5e_set_mtu(struct mlx5e_priv *priv, u16 mtu)
1726 {
1727         struct mlx5_core_dev *mdev = priv->mdev;
1728         u16 hw_mtu = MLX5E_SW2HW_MTU(mtu);
1729         int err;
1730
1731         err = mlx5_set_port_mtu(mdev, hw_mtu, 1);
1732         if (err)
1733                 return err;
1734
1735         /* Update vport context MTU */
1736         mlx5_modify_nic_vport_mtu(mdev, hw_mtu);
1737         return 0;
1738 }
1739
1740 static void mlx5e_query_mtu(struct mlx5e_priv *priv, u16 *mtu)
1741 {
1742         struct mlx5_core_dev *mdev = priv->mdev;
1743         u16 hw_mtu = 0;
1744         int err;
1745
1746         err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
1747         if (err || !hw_mtu) /* fallback to port oper mtu */
1748                 mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
1749
1750         *mtu = MLX5E_HW2SW_MTU(hw_mtu);
1751 }
1752
1753 static int mlx5e_set_dev_port_mtu(struct net_device *netdev)
1754 {
1755         struct mlx5e_priv *priv = netdev_priv(netdev);
1756         u16 mtu;
1757         int err;
1758
1759         err = mlx5e_set_mtu(priv, netdev->mtu);
1760         if (err)
1761                 return err;
1762
1763         mlx5e_query_mtu(priv, &mtu);
1764         if (mtu != netdev->mtu)
1765                 netdev_warn(netdev, "%s: VPort MTU %d is different than netdev mtu %d\n",
1766                             __func__, mtu, netdev->mtu);
1767
1768         netdev->mtu = mtu;
1769         return 0;
1770 }
1771
1772 static void mlx5e_netdev_set_tcs(struct net_device *netdev)
1773 {
1774         struct mlx5e_priv *priv = netdev_priv(netdev);
1775         int nch = priv->params.num_channels;
1776         int ntc = priv->params.num_tc;
1777         int tc;
1778
1779         netdev_reset_tc(netdev);
1780
1781         if (ntc == 1)
1782                 return;
1783
1784         netdev_set_num_tc(netdev, ntc);
1785
1786         /* Map netdev TCs to offset 0
1787          * We have our own UP to TXQ mapping for QoS
1788          */
1789         for (tc = 0; tc < ntc; tc++)
1790                 netdev_set_tc_queue(netdev, tc, nch, 0);
1791 }
1792
1793 int mlx5e_open_locked(struct net_device *netdev)
1794 {
1795         struct mlx5e_priv *priv = netdev_priv(netdev);
1796         struct mlx5_core_dev *mdev = priv->mdev;
1797         int num_txqs;
1798         int err;
1799
1800         set_bit(MLX5E_STATE_OPENED, &priv->state);
1801
1802         mlx5e_netdev_set_tcs(netdev);
1803
1804         num_txqs = priv->params.num_channels * priv->params.num_tc;
1805         netif_set_real_num_tx_queues(netdev, num_txqs);
1806         netif_set_real_num_rx_queues(netdev, priv->params.num_channels);
1807
1808         err = mlx5e_open_channels(priv);
1809         if (err) {
1810                 netdev_err(netdev, "%s: mlx5e_open_channels failed, %d\n",
1811                            __func__, err);
1812                 goto err_clear_state_opened_flag;
1813         }
1814
1815         err = mlx5e_refresh_tirs_self_loopback_enable(priv->mdev);
1816         if (err) {
1817                 netdev_err(netdev, "%s: mlx5e_refresh_tirs_self_loopback_enable failed, %d\n",
1818                            __func__, err);
1819                 goto err_close_channels;
1820         }
1821
1822         mlx5e_redirect_rqts(priv);
1823         mlx5e_update_carrier(priv);
1824         mlx5e_timestamp_init(priv);
1825 #ifdef CONFIG_RFS_ACCEL
1826         priv->netdev->rx_cpu_rmap = priv->mdev->rmap;
1827 #endif
1828         if (priv->profile->update_stats)
1829                 queue_delayed_work(priv->wq, &priv->update_stats_work, 0);
1830
1831         if (MLX5_CAP_GEN(mdev, vport_group_manager)) {
1832                 err = mlx5e_add_sqs_fwd_rules(priv);
1833                 if (err)
1834                         goto err_close_channels;
1835         }
1836         return 0;
1837
1838 err_close_channels:
1839         mlx5e_close_channels(priv);
1840 err_clear_state_opened_flag:
1841         clear_bit(MLX5E_STATE_OPENED, &priv->state);
1842         return err;
1843 }
1844
1845 int mlx5e_open(struct net_device *netdev)
1846 {
1847         struct mlx5e_priv *priv = netdev_priv(netdev);
1848         int err;
1849
1850         mutex_lock(&priv->state_lock);
1851         err = mlx5e_open_locked(netdev);
1852         mutex_unlock(&priv->state_lock);
1853
1854         return err;
1855 }
1856
1857 int mlx5e_close_locked(struct net_device *netdev)
1858 {
1859         struct mlx5e_priv *priv = netdev_priv(netdev);
1860         struct mlx5_core_dev *mdev = priv->mdev;
1861
1862         /* May already be CLOSED in case a previous configuration operation
1863          * (e.g RX/TX queue size change) that involves close&open failed.
1864          */
1865         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
1866                 return 0;
1867
1868         clear_bit(MLX5E_STATE_OPENED, &priv->state);
1869
1870         if (MLX5_CAP_GEN(mdev, vport_group_manager))
1871                 mlx5e_remove_sqs_fwd_rules(priv);
1872
1873         mlx5e_timestamp_cleanup(priv);
1874         netif_carrier_off(priv->netdev);
1875         mlx5e_redirect_rqts(priv);
1876         mlx5e_close_channels(priv);
1877
1878         return 0;
1879 }
1880
1881 int mlx5e_close(struct net_device *netdev)
1882 {
1883         struct mlx5e_priv *priv = netdev_priv(netdev);
1884         int err;
1885
1886         if (!netif_device_present(netdev))
1887                 return -ENODEV;
1888
1889         mutex_lock(&priv->state_lock);
1890         err = mlx5e_close_locked(netdev);
1891         mutex_unlock(&priv->state_lock);
1892
1893         return err;
1894 }
1895
1896 static int mlx5e_create_drop_rq(struct mlx5e_priv *priv,
1897                                 struct mlx5e_rq *rq,
1898                                 struct mlx5e_rq_param *param)
1899 {
1900         struct mlx5_core_dev *mdev = priv->mdev;
1901         void *rqc = param->rqc;
1902         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
1903         int err;
1904
1905         param->wq.db_numa_node = param->wq.buf_numa_node;
1906
1907         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
1908                                 &rq->wq_ctrl);
1909         if (err)
1910                 return err;
1911
1912         rq->priv = priv;
1913
1914         return 0;
1915 }
1916
1917 static int mlx5e_create_drop_cq(struct mlx5e_priv *priv,
1918                                 struct mlx5e_cq *cq,
1919                                 struct mlx5e_cq_param *param)
1920 {
1921         struct mlx5_core_dev *mdev = priv->mdev;
1922         struct mlx5_core_cq *mcq = &cq->mcq;
1923         int eqn_not_used;
1924         unsigned int irqn;
1925         int err;
1926
1927         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1928                                &cq->wq_ctrl);
1929         if (err)
1930                 return err;
1931
1932         mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1933
1934         mcq->cqe_sz     = 64;
1935         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1936         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1937         *mcq->set_ci_db = 0;
1938         *mcq->arm_db    = 0;
1939         mcq->vector     = param->eq_ix;
1940         mcq->comp       = mlx5e_completion_event;
1941         mcq->event      = mlx5e_cq_error_event;
1942         mcq->irqn       = irqn;
1943         mcq->uar        = &mdev->mlx5e_res.cq_uar;
1944
1945         cq->priv = priv;
1946
1947         return 0;
1948 }
1949
1950 static int mlx5e_open_drop_rq(struct mlx5e_priv *priv)
1951 {
1952         struct mlx5e_cq_param cq_param;
1953         struct mlx5e_rq_param rq_param;
1954         struct mlx5e_rq *rq = &priv->drop_rq;
1955         struct mlx5e_cq *cq = &priv->drop_rq.cq;
1956         int err;
1957
1958         memset(&cq_param, 0, sizeof(cq_param));
1959         memset(&rq_param, 0, sizeof(rq_param));
1960         mlx5e_build_drop_rq_param(&rq_param);
1961
1962         err = mlx5e_create_drop_cq(priv, cq, &cq_param);
1963         if (err)
1964                 return err;
1965
1966         err = mlx5e_enable_cq(cq, &cq_param);
1967         if (err)
1968                 goto err_destroy_cq;
1969
1970         err = mlx5e_create_drop_rq(priv, rq, &rq_param);
1971         if (err)
1972                 goto err_disable_cq;
1973
1974         err = mlx5e_enable_rq(rq, &rq_param);
1975         if (err)
1976                 goto err_destroy_rq;
1977
1978         return 0;
1979
1980 err_destroy_rq:
1981         mlx5e_destroy_rq(&priv->drop_rq);
1982
1983 err_disable_cq:
1984         mlx5e_disable_cq(&priv->drop_rq.cq);
1985
1986 err_destroy_cq:
1987         mlx5e_destroy_cq(&priv->drop_rq.cq);
1988
1989         return err;
1990 }
1991
1992 static void mlx5e_close_drop_rq(struct mlx5e_priv *priv)
1993 {
1994         mlx5e_disable_rq(&priv->drop_rq);
1995         mlx5e_destroy_rq(&priv->drop_rq);
1996         mlx5e_disable_cq(&priv->drop_rq.cq);
1997         mlx5e_destroy_cq(&priv->drop_rq.cq);
1998 }
1999
2000 static int mlx5e_create_tis(struct mlx5e_priv *priv, int tc)
2001 {
2002         struct mlx5_core_dev *mdev = priv->mdev;
2003         u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
2004         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
2005
2006         MLX5_SET(tisc, tisc, prio, tc << 1);
2007         MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.td.tdn);
2008
2009         if (mlx5_lag_is_lacp_owner(mdev))
2010                 MLX5_SET(tisc, tisc, strict_lag_tx_port_affinity, 1);
2011
2012         return mlx5_core_create_tis(mdev, in, sizeof(in), &priv->tisn[tc]);
2013 }
2014
2015 static void mlx5e_destroy_tis(struct mlx5e_priv *priv, int tc)
2016 {
2017         mlx5_core_destroy_tis(priv->mdev, priv->tisn[tc]);
2018 }
2019
2020 int mlx5e_create_tises(struct mlx5e_priv *priv)
2021 {
2022         int err;
2023         int tc;
2024
2025         for (tc = 0; tc < priv->profile->max_tc; tc++) {
2026                 err = mlx5e_create_tis(priv, tc);
2027                 if (err)
2028                         goto err_close_tises;
2029         }
2030
2031         return 0;
2032
2033 err_close_tises:
2034         for (tc--; tc >= 0; tc--)
2035                 mlx5e_destroy_tis(priv, tc);
2036
2037         return err;
2038 }
2039
2040 void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
2041 {
2042         int tc;
2043
2044         for (tc = 0; tc < priv->profile->max_tc; tc++)
2045                 mlx5e_destroy_tis(priv, tc);
2046 }
2047
2048 static void mlx5e_build_indir_tir_ctx(struct mlx5e_priv *priv, u32 *tirc,
2049                                       enum mlx5e_traffic_types tt)
2050 {
2051         void *hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2052
2053         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2054
2055 #define MLX5_HASH_IP            (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2056                                  MLX5_HASH_FIELD_SEL_DST_IP)
2057
2058 #define MLX5_HASH_IP_L4PORTS    (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2059                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2060                                  MLX5_HASH_FIELD_SEL_L4_SPORT |\
2061                                  MLX5_HASH_FIELD_SEL_L4_DPORT)
2062
2063 #define MLX5_HASH_IP_IPSEC_SPI  (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2064                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2065                                  MLX5_HASH_FIELD_SEL_IPSEC_SPI)
2066
2067         mlx5e_build_tir_ctx_lro(tirc, priv);
2068
2069         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2070         MLX5_SET(tirc, tirc, indirect_table, priv->indir_rqt.rqtn);
2071         mlx5e_build_tir_ctx_hash(tirc, priv);
2072
2073         switch (tt) {
2074         case MLX5E_TT_IPV4_TCP:
2075                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2076                          MLX5_L3_PROT_TYPE_IPV4);
2077                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2078                          MLX5_L4_PROT_TYPE_TCP);
2079                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2080                          MLX5_HASH_IP_L4PORTS);
2081                 break;
2082
2083         case MLX5E_TT_IPV6_TCP:
2084                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2085                          MLX5_L3_PROT_TYPE_IPV6);
2086                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2087                          MLX5_L4_PROT_TYPE_TCP);
2088                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2089                          MLX5_HASH_IP_L4PORTS);
2090                 break;
2091
2092         case MLX5E_TT_IPV4_UDP:
2093                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2094                          MLX5_L3_PROT_TYPE_IPV4);
2095                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2096                          MLX5_L4_PROT_TYPE_UDP);
2097                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2098                          MLX5_HASH_IP_L4PORTS);
2099                 break;
2100
2101         case MLX5E_TT_IPV6_UDP:
2102                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2103                          MLX5_L3_PROT_TYPE_IPV6);
2104                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2105                          MLX5_L4_PROT_TYPE_UDP);
2106                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2107                          MLX5_HASH_IP_L4PORTS);
2108                 break;
2109
2110         case MLX5E_TT_IPV4_IPSEC_AH:
2111                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2112                          MLX5_L3_PROT_TYPE_IPV4);
2113                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2114                          MLX5_HASH_IP_IPSEC_SPI);
2115                 break;
2116
2117         case MLX5E_TT_IPV6_IPSEC_AH:
2118                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2119                          MLX5_L3_PROT_TYPE_IPV6);
2120                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2121                          MLX5_HASH_IP_IPSEC_SPI);
2122                 break;
2123
2124         case MLX5E_TT_IPV4_IPSEC_ESP:
2125                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2126                          MLX5_L3_PROT_TYPE_IPV4);
2127                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2128                          MLX5_HASH_IP_IPSEC_SPI);
2129                 break;
2130
2131         case MLX5E_TT_IPV6_IPSEC_ESP:
2132                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2133                          MLX5_L3_PROT_TYPE_IPV6);
2134                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2135                          MLX5_HASH_IP_IPSEC_SPI);
2136                 break;
2137
2138         case MLX5E_TT_IPV4:
2139                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2140                          MLX5_L3_PROT_TYPE_IPV4);
2141                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2142                          MLX5_HASH_IP);
2143                 break;
2144
2145         case MLX5E_TT_IPV6:
2146                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2147                          MLX5_L3_PROT_TYPE_IPV6);
2148                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2149                          MLX5_HASH_IP);
2150                 break;
2151         default:
2152                 WARN_ONCE(true,
2153                           "mlx5e_build_indir_tir_ctx: bad traffic type!\n");
2154         }
2155 }
2156
2157 static void mlx5e_build_direct_tir_ctx(struct mlx5e_priv *priv, u32 *tirc,
2158                                        u32 rqtn)
2159 {
2160         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2161
2162         mlx5e_build_tir_ctx_lro(tirc, priv);
2163
2164         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2165         MLX5_SET(tirc, tirc, indirect_table, rqtn);
2166         MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
2167 }
2168
2169 static int mlx5e_create_indirect_tirs(struct mlx5e_priv *priv)
2170 {
2171         struct mlx5e_tir *tir;
2172         void *tirc;
2173         int inlen;
2174         int err;
2175         u32 *in;
2176         int tt;
2177
2178         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2179         in = mlx5_vzalloc(inlen);
2180         if (!in)
2181                 return -ENOMEM;
2182
2183         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2184                 memset(in, 0, inlen);
2185                 tir = &priv->indir_tir[tt];
2186                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2187                 mlx5e_build_indir_tir_ctx(priv, tirc, tt);
2188                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2189                 if (err)
2190                         goto err_destroy_tirs;
2191         }
2192
2193         kvfree(in);
2194
2195         return 0;
2196
2197 err_destroy_tirs:
2198         for (tt--; tt >= 0; tt--)
2199                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[tt]);
2200
2201         kvfree(in);
2202
2203         return err;
2204 }
2205
2206 int mlx5e_create_direct_tirs(struct mlx5e_priv *priv)
2207 {
2208         int nch = priv->profile->max_nch(priv->mdev);
2209         struct mlx5e_tir *tir;
2210         void *tirc;
2211         int inlen;
2212         int err;
2213         u32 *in;
2214         int ix;
2215
2216         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2217         in = mlx5_vzalloc(inlen);
2218         if (!in)
2219                 return -ENOMEM;
2220
2221         for (ix = 0; ix < nch; ix++) {
2222                 memset(in, 0, inlen);
2223                 tir = &priv->direct_tir[ix];
2224                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2225                 mlx5e_build_direct_tir_ctx(priv, tirc,
2226                                            priv->direct_tir[ix].rqt.rqtn);
2227                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2228                 if (err)
2229                         goto err_destroy_ch_tirs;
2230         }
2231
2232         kvfree(in);
2233
2234         return 0;
2235
2236 err_destroy_ch_tirs:
2237         for (ix--; ix >= 0; ix--)
2238                 mlx5e_destroy_tir(priv->mdev, &priv->direct_tir[ix]);
2239
2240         kvfree(in);
2241
2242         return err;
2243 }
2244
2245 static void mlx5e_destroy_indirect_tirs(struct mlx5e_priv *priv)
2246 {
2247         int i;
2248
2249         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
2250                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[i]);
2251 }
2252
2253 void mlx5e_destroy_direct_tirs(struct mlx5e_priv *priv)
2254 {
2255         int nch = priv->profile->max_nch(priv->mdev);
2256         int i;
2257
2258         for (i = 0; i < nch; i++)
2259                 mlx5e_destroy_tir(priv->mdev, &priv->direct_tir[i]);
2260 }
2261
2262 int mlx5e_modify_rqs_vsd(struct mlx5e_priv *priv, bool vsd)
2263 {
2264         int err = 0;
2265         int i;
2266
2267         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
2268                 return 0;
2269
2270         for (i = 0; i < priv->params.num_channels; i++) {
2271                 err = mlx5e_modify_rq_vsd(&priv->channel[i]->rq, vsd);
2272                 if (err)
2273                         return err;
2274         }
2275
2276         return 0;
2277 }
2278
2279 static int mlx5e_setup_tc(struct net_device *netdev, u8 tc)
2280 {
2281         struct mlx5e_priv *priv = netdev_priv(netdev);
2282         bool was_opened;
2283         int err = 0;
2284
2285         if (tc && tc != MLX5E_MAX_NUM_TC)
2286                 return -EINVAL;
2287
2288         mutex_lock(&priv->state_lock);
2289
2290         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2291         if (was_opened)
2292                 mlx5e_close_locked(priv->netdev);
2293
2294         priv->params.num_tc = tc ? tc : 1;
2295
2296         if (was_opened)
2297                 err = mlx5e_open_locked(priv->netdev);
2298
2299         mutex_unlock(&priv->state_lock);
2300
2301         return err;
2302 }
2303
2304 static int mlx5e_ndo_setup_tc(struct net_device *dev, u32 handle,
2305                               __be16 proto, struct tc_to_netdev *tc)
2306 {
2307         struct mlx5e_priv *priv = netdev_priv(dev);
2308
2309         if (TC_H_MAJ(handle) != TC_H_MAJ(TC_H_INGRESS))
2310                 goto mqprio;
2311
2312         switch (tc->type) {
2313         case TC_SETUP_CLSFLOWER:
2314                 switch (tc->cls_flower->command) {
2315                 case TC_CLSFLOWER_REPLACE:
2316                         return mlx5e_configure_flower(priv, proto, tc->cls_flower);
2317                 case TC_CLSFLOWER_DESTROY:
2318                         return mlx5e_delete_flower(priv, tc->cls_flower);
2319                 case TC_CLSFLOWER_STATS:
2320                         return mlx5e_stats_flower(priv, tc->cls_flower);
2321                 }
2322         default:
2323                 return -EOPNOTSUPP;
2324         }
2325
2326 mqprio:
2327         if (tc->type != TC_SETUP_MQPRIO)
2328                 return -EINVAL;
2329
2330         return mlx5e_setup_tc(dev, tc->tc);
2331 }
2332
2333 struct rtnl_link_stats64 *
2334 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
2335 {
2336         struct mlx5e_priv *priv = netdev_priv(dev);
2337         struct mlx5e_sw_stats *sstats = &priv->stats.sw;
2338         struct mlx5e_vport_stats *vstats = &priv->stats.vport;
2339         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
2340
2341         stats->rx_packets = sstats->rx_packets;
2342         stats->rx_bytes   = sstats->rx_bytes;
2343         stats->tx_packets = sstats->tx_packets;
2344         stats->tx_bytes   = sstats->tx_bytes;
2345
2346         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
2347         stats->tx_dropped = sstats->tx_queue_dropped;
2348
2349         stats->rx_length_errors =
2350                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
2351                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
2352                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
2353         stats->rx_crc_errors =
2354                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
2355         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
2356         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
2357         stats->tx_carrier_errors =
2358                 PPORT_802_3_GET(pstats, a_symbol_error_during_carrier);
2359         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
2360                            stats->rx_frame_errors;
2361         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
2362
2363         /* vport multicast also counts packets that are dropped due to steering
2364          * or rx out of buffer
2365          */
2366         stats->multicast =
2367                 VPORT_COUNTER_GET(vstats, received_eth_multicast.packets);
2368
2369         return stats;
2370 }
2371
2372 static void mlx5e_set_rx_mode(struct net_device *dev)
2373 {
2374         struct mlx5e_priv *priv = netdev_priv(dev);
2375
2376         queue_work(priv->wq, &priv->set_rx_mode_work);
2377 }
2378
2379 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
2380 {
2381         struct mlx5e_priv *priv = netdev_priv(netdev);
2382         struct sockaddr *saddr = addr;
2383
2384         if (!is_valid_ether_addr(saddr->sa_data))
2385                 return -EADDRNOTAVAIL;
2386
2387         netif_addr_lock_bh(netdev);
2388         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
2389         netif_addr_unlock_bh(netdev);
2390
2391         queue_work(priv->wq, &priv->set_rx_mode_work);
2392
2393         return 0;
2394 }
2395
2396 #define MLX5E_SET_FEATURE(netdev, feature, enable)      \
2397         do {                                            \
2398                 if (enable)                             \
2399                         netdev->features |= feature;    \
2400                 else                                    \
2401                         netdev->features &= ~feature;   \
2402         } while (0)
2403
2404 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
2405
2406 static int set_feature_lro(struct net_device *netdev, bool enable)
2407 {
2408         struct mlx5e_priv *priv = netdev_priv(netdev);
2409         bool was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2410         int err;
2411
2412         mutex_lock(&priv->state_lock);
2413
2414         if (was_opened && (priv->params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST))
2415                 mlx5e_close_locked(priv->netdev);
2416
2417         priv->params.lro_en = enable;
2418         err = mlx5e_modify_tirs_lro(priv);
2419         if (err) {
2420                 netdev_err(netdev, "lro modify failed, %d\n", err);
2421                 priv->params.lro_en = !enable;
2422         }
2423
2424         if (was_opened && (priv->params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST))
2425                 mlx5e_open_locked(priv->netdev);
2426
2427         mutex_unlock(&priv->state_lock);
2428
2429         return err;
2430 }
2431
2432 static int set_feature_vlan_filter(struct net_device *netdev, bool enable)
2433 {
2434         struct mlx5e_priv *priv = netdev_priv(netdev);
2435
2436         if (enable)
2437                 mlx5e_enable_vlan_filter(priv);
2438         else
2439                 mlx5e_disable_vlan_filter(priv);
2440
2441         return 0;
2442 }
2443
2444 static int set_feature_tc_num_filters(struct net_device *netdev, bool enable)
2445 {
2446         struct mlx5e_priv *priv = netdev_priv(netdev);
2447
2448         if (!enable && mlx5e_tc_num_filters(priv)) {
2449                 netdev_err(netdev,
2450                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
2451                 return -EINVAL;
2452         }
2453
2454         return 0;
2455 }
2456
2457 static int set_feature_rx_all(struct net_device *netdev, bool enable)
2458 {
2459         struct mlx5e_priv *priv = netdev_priv(netdev);
2460         struct mlx5_core_dev *mdev = priv->mdev;
2461
2462         return mlx5_set_port_fcs(mdev, !enable);
2463 }
2464
2465 static int set_feature_rx_vlan(struct net_device *netdev, bool enable)
2466 {
2467         struct mlx5e_priv *priv = netdev_priv(netdev);
2468         int err;
2469
2470         mutex_lock(&priv->state_lock);
2471
2472         priv->params.vlan_strip_disable = !enable;
2473         err = mlx5e_modify_rqs_vsd(priv, !enable);
2474         if (err)
2475                 priv->params.vlan_strip_disable = enable;
2476
2477         mutex_unlock(&priv->state_lock);
2478
2479         return err;
2480 }
2481
2482 #ifdef CONFIG_RFS_ACCEL
2483 static int set_feature_arfs(struct net_device *netdev, bool enable)
2484 {
2485         struct mlx5e_priv *priv = netdev_priv(netdev);
2486         int err;
2487
2488         if (enable)
2489                 err = mlx5e_arfs_enable(priv);
2490         else
2491                 err = mlx5e_arfs_disable(priv);
2492
2493         return err;
2494 }
2495 #endif
2496
2497 static int mlx5e_handle_feature(struct net_device *netdev,
2498                                 netdev_features_t wanted_features,
2499                                 netdev_features_t feature,
2500                                 mlx5e_feature_handler feature_handler)
2501 {
2502         netdev_features_t changes = wanted_features ^ netdev->features;
2503         bool enable = !!(wanted_features & feature);
2504         int err;
2505
2506         if (!(changes & feature))
2507                 return 0;
2508
2509         err = feature_handler(netdev, enable);
2510         if (err) {
2511                 netdev_err(netdev, "%s feature 0x%llx failed err %d\n",
2512                            enable ? "Enable" : "Disable", feature, err);
2513                 return err;
2514         }
2515
2516         MLX5E_SET_FEATURE(netdev, feature, enable);
2517         return 0;
2518 }
2519
2520 static int mlx5e_set_features(struct net_device *netdev,
2521                               netdev_features_t features)
2522 {
2523         int err;
2524
2525         err  = mlx5e_handle_feature(netdev, features, NETIF_F_LRO,
2526                                     set_feature_lro);
2527         err |= mlx5e_handle_feature(netdev, features,
2528                                     NETIF_F_HW_VLAN_CTAG_FILTER,
2529                                     set_feature_vlan_filter);
2530         err |= mlx5e_handle_feature(netdev, features, NETIF_F_HW_TC,
2531                                     set_feature_tc_num_filters);
2532         err |= mlx5e_handle_feature(netdev, features, NETIF_F_RXALL,
2533                                     set_feature_rx_all);
2534         err |= mlx5e_handle_feature(netdev, features, NETIF_F_HW_VLAN_CTAG_RX,
2535                                     set_feature_rx_vlan);
2536 #ifdef CONFIG_RFS_ACCEL
2537         err |= mlx5e_handle_feature(netdev, features, NETIF_F_NTUPLE,
2538                                     set_feature_arfs);
2539 #endif
2540
2541         return err ? -EINVAL : 0;
2542 }
2543
2544 #define MXL5_HW_MIN_MTU 64
2545 #define MXL5E_MIN_MTU (MXL5_HW_MIN_MTU + ETH_FCS_LEN)
2546
2547 static int mlx5e_change_mtu(struct net_device *netdev, int new_mtu)
2548 {
2549         struct mlx5e_priv *priv = netdev_priv(netdev);
2550         struct mlx5_core_dev *mdev = priv->mdev;
2551         bool was_opened;
2552         u16 max_mtu;
2553         u16 min_mtu;
2554         int err = 0;
2555         bool reset;
2556
2557         mlx5_query_port_max_mtu(mdev, &max_mtu, 1);
2558
2559         max_mtu = MLX5E_HW2SW_MTU(max_mtu);
2560         min_mtu = MLX5E_HW2SW_MTU(MXL5E_MIN_MTU);
2561
2562         if (new_mtu > max_mtu || new_mtu < min_mtu) {
2563                 netdev_err(netdev,
2564                            "%s: Bad MTU (%d), valid range is: [%d..%d]\n",
2565                            __func__, new_mtu, min_mtu, max_mtu);
2566                 return -EINVAL;
2567         }
2568
2569         mutex_lock(&priv->state_lock);
2570
2571         reset = !priv->params.lro_en &&
2572                 (priv->params.rq_wq_type !=
2573                  MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
2574
2575         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2576         if (was_opened && reset)
2577                 mlx5e_close_locked(netdev);
2578
2579         netdev->mtu = new_mtu;
2580         mlx5e_set_dev_port_mtu(netdev);
2581
2582         if (was_opened && reset)
2583                 err = mlx5e_open_locked(netdev);
2584
2585         mutex_unlock(&priv->state_lock);
2586
2587         return err;
2588 }
2589
2590 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
2591 {
2592         switch (cmd) {
2593         case SIOCSHWTSTAMP:
2594                 return mlx5e_hwstamp_set(dev, ifr);
2595         case SIOCGHWTSTAMP:
2596                 return mlx5e_hwstamp_get(dev, ifr);
2597         default:
2598                 return -EOPNOTSUPP;
2599         }
2600 }
2601
2602 static int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
2603 {
2604         struct mlx5e_priv *priv = netdev_priv(dev);
2605         struct mlx5_core_dev *mdev = priv->mdev;
2606
2607         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
2608 }
2609
2610 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos)
2611 {
2612         struct mlx5e_priv *priv = netdev_priv(dev);
2613         struct mlx5_core_dev *mdev = priv->mdev;
2614
2615         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
2616                                            vlan, qos);
2617 }
2618
2619 static int mlx5e_set_vf_spoofchk(struct net_device *dev, int vf, bool setting)
2620 {
2621         struct mlx5e_priv *priv = netdev_priv(dev);
2622         struct mlx5_core_dev *mdev = priv->mdev;
2623
2624         return mlx5_eswitch_set_vport_spoofchk(mdev->priv.eswitch, vf + 1, setting);
2625 }
2626
2627 static int mlx5e_set_vf_trust(struct net_device *dev, int vf, bool setting)
2628 {
2629         struct mlx5e_priv *priv = netdev_priv(dev);
2630         struct mlx5_core_dev *mdev = priv->mdev;
2631
2632         return mlx5_eswitch_set_vport_trust(mdev->priv.eswitch, vf + 1, setting);
2633 }
2634 static int mlx5_vport_link2ifla(u8 esw_link)
2635 {
2636         switch (esw_link) {
2637         case MLX5_ESW_VPORT_ADMIN_STATE_DOWN:
2638                 return IFLA_VF_LINK_STATE_DISABLE;
2639         case MLX5_ESW_VPORT_ADMIN_STATE_UP:
2640                 return IFLA_VF_LINK_STATE_ENABLE;
2641         }
2642         return IFLA_VF_LINK_STATE_AUTO;
2643 }
2644
2645 static int mlx5_ifla_link2vport(u8 ifla_link)
2646 {
2647         switch (ifla_link) {
2648         case IFLA_VF_LINK_STATE_DISABLE:
2649                 return MLX5_ESW_VPORT_ADMIN_STATE_DOWN;
2650         case IFLA_VF_LINK_STATE_ENABLE:
2651                 return MLX5_ESW_VPORT_ADMIN_STATE_UP;
2652         }
2653         return MLX5_ESW_VPORT_ADMIN_STATE_AUTO;
2654 }
2655
2656 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
2657                                    int link_state)
2658 {
2659         struct mlx5e_priv *priv = netdev_priv(dev);
2660         struct mlx5_core_dev *mdev = priv->mdev;
2661
2662         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
2663                                             mlx5_ifla_link2vport(link_state));
2664 }
2665
2666 static int mlx5e_get_vf_config(struct net_device *dev,
2667                                int vf, struct ifla_vf_info *ivi)
2668 {
2669         struct mlx5e_priv *priv = netdev_priv(dev);
2670         struct mlx5_core_dev *mdev = priv->mdev;
2671         int err;
2672
2673         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
2674         if (err)
2675                 return err;
2676         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
2677         return 0;
2678 }
2679
2680 static int mlx5e_get_vf_stats(struct net_device *dev,
2681                               int vf, struct ifla_vf_stats *vf_stats)
2682 {
2683         struct mlx5e_priv *priv = netdev_priv(dev);
2684         struct mlx5_core_dev *mdev = priv->mdev;
2685
2686         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
2687                                             vf_stats);
2688 }
2689
2690 static void mlx5e_add_vxlan_port(struct net_device *netdev,
2691                                  struct udp_tunnel_info *ti)
2692 {
2693         struct mlx5e_priv *priv = netdev_priv(netdev);
2694
2695         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
2696                 return;
2697
2698         if (!mlx5e_vxlan_allowed(priv->mdev))
2699                 return;
2700
2701         mlx5e_vxlan_queue_work(priv, ti->sa_family, be16_to_cpu(ti->port), 1);
2702 }
2703
2704 static void mlx5e_del_vxlan_port(struct net_device *netdev,
2705                                  struct udp_tunnel_info *ti)
2706 {
2707         struct mlx5e_priv *priv = netdev_priv(netdev);
2708
2709         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
2710                 return;
2711
2712         if (!mlx5e_vxlan_allowed(priv->mdev))
2713                 return;
2714
2715         mlx5e_vxlan_queue_work(priv, ti->sa_family, be16_to_cpu(ti->port), 0);
2716 }
2717
2718 static netdev_features_t mlx5e_vxlan_features_check(struct mlx5e_priv *priv,
2719                                                     struct sk_buff *skb,
2720                                                     netdev_features_t features)
2721 {
2722         struct udphdr *udph;
2723         u16 proto;
2724         u16 port = 0;
2725
2726         switch (vlan_get_protocol(skb)) {
2727         case htons(ETH_P_IP):
2728                 proto = ip_hdr(skb)->protocol;
2729                 break;
2730         case htons(ETH_P_IPV6):
2731                 proto = ipv6_hdr(skb)->nexthdr;
2732                 break;
2733         default:
2734                 goto out;
2735         }
2736
2737         if (proto == IPPROTO_UDP) {
2738                 udph = udp_hdr(skb);
2739                 port = be16_to_cpu(udph->dest);
2740         }
2741
2742         /* Verify if UDP port is being offloaded by HW */
2743         if (port && mlx5e_vxlan_lookup_port(priv, port))
2744                 return features;
2745
2746 out:
2747         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
2748         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2749 }
2750
2751 static netdev_features_t mlx5e_features_check(struct sk_buff *skb,
2752                                               struct net_device *netdev,
2753                                               netdev_features_t features)
2754 {
2755         struct mlx5e_priv *priv = netdev_priv(netdev);
2756
2757         features = vlan_features_check(skb, features);
2758         features = vxlan_features_check(skb, features);
2759
2760         /* Validate if the tunneled packet is being offloaded by HW */
2761         if (skb->encapsulation &&
2762             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
2763                 return mlx5e_vxlan_features_check(priv, skb, features);
2764
2765         return features;
2766 }
2767
2768 static void mlx5e_tx_timeout(struct net_device *dev)
2769 {
2770         struct mlx5e_priv *priv = netdev_priv(dev);
2771         bool sched_work = false;
2772         int i;
2773
2774         netdev_err(dev, "TX timeout detected\n");
2775
2776         for (i = 0; i < priv->params.num_channels * priv->params.num_tc; i++) {
2777                 struct mlx5e_sq *sq = priv->txq_to_sq_map[i];
2778
2779                 if (!netif_xmit_stopped(netdev_get_tx_queue(dev, i)))
2780                         continue;
2781                 sched_work = true;
2782                 set_bit(MLX5E_SQ_STATE_FLUSH, &sq->state);
2783                 netdev_err(dev, "TX timeout on queue: %d, SQ: 0x%x, CQ: 0x%x, SQ Cons: 0x%x SQ Prod: 0x%x\n",
2784                            i, sq->sqn, sq->cq.mcq.cqn, sq->cc, sq->pc);
2785         }
2786
2787         if (sched_work && test_bit(MLX5E_STATE_OPENED, &priv->state))
2788                 schedule_work(&priv->tx_timeout_work);
2789 }
2790
2791 static const struct net_device_ops mlx5e_netdev_ops_basic = {
2792         .ndo_open                = mlx5e_open,
2793         .ndo_stop                = mlx5e_close,
2794         .ndo_start_xmit          = mlx5e_xmit,
2795         .ndo_setup_tc            = mlx5e_ndo_setup_tc,
2796         .ndo_select_queue        = mlx5e_select_queue,
2797         .ndo_get_stats64         = mlx5e_get_stats,
2798         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
2799         .ndo_set_mac_address     = mlx5e_set_mac,
2800         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
2801         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
2802         .ndo_set_features        = mlx5e_set_features,
2803         .ndo_change_mtu          = mlx5e_change_mtu,
2804         .ndo_do_ioctl            = mlx5e_ioctl,
2805         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
2806 #ifdef CONFIG_RFS_ACCEL
2807         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
2808 #endif
2809         .ndo_tx_timeout          = mlx5e_tx_timeout,
2810 };
2811
2812 static const struct net_device_ops mlx5e_netdev_ops_sriov = {
2813         .ndo_open                = mlx5e_open,
2814         .ndo_stop                = mlx5e_close,
2815         .ndo_start_xmit          = mlx5e_xmit,
2816         .ndo_setup_tc            = mlx5e_ndo_setup_tc,
2817         .ndo_select_queue        = mlx5e_select_queue,
2818         .ndo_get_stats64         = mlx5e_get_stats,
2819         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
2820         .ndo_set_mac_address     = mlx5e_set_mac,
2821         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
2822         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
2823         .ndo_set_features        = mlx5e_set_features,
2824         .ndo_change_mtu          = mlx5e_change_mtu,
2825         .ndo_do_ioctl            = mlx5e_ioctl,
2826         .ndo_udp_tunnel_add      = mlx5e_add_vxlan_port,
2827         .ndo_udp_tunnel_del      = mlx5e_del_vxlan_port,
2828         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
2829         .ndo_features_check      = mlx5e_features_check,
2830 #ifdef CONFIG_RFS_ACCEL
2831         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
2832 #endif
2833         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
2834         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
2835         .ndo_set_vf_spoofchk     = mlx5e_set_vf_spoofchk,
2836         .ndo_set_vf_trust        = mlx5e_set_vf_trust,
2837         .ndo_get_vf_config       = mlx5e_get_vf_config,
2838         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
2839         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
2840         .ndo_tx_timeout          = mlx5e_tx_timeout,
2841 };
2842
2843 static int mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
2844 {
2845         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
2846                 return -ENOTSUPP;
2847         if (!MLX5_CAP_GEN(mdev, eth_net_offloads) ||
2848             !MLX5_CAP_GEN(mdev, nic_flow_table) ||
2849             !MLX5_CAP_ETH(mdev, csum_cap) ||
2850             !MLX5_CAP_ETH(mdev, max_lso_cap) ||
2851             !MLX5_CAP_ETH(mdev, vlan_cap) ||
2852             !MLX5_CAP_ETH(mdev, rss_ind_tbl_cap) ||
2853             MLX5_CAP_FLOWTABLE(mdev,
2854                                flow_table_properties_nic_receive.max_ft_level)
2855                                < 3) {
2856                 mlx5_core_warn(mdev,
2857                                "Not creating net device, some required device capabilities are missing\n");
2858                 return -ENOTSUPP;
2859         }
2860         if (!MLX5_CAP_ETH(mdev, self_lb_en_modifiable))
2861                 mlx5_core_warn(mdev, "Self loop back prevention is not supported\n");
2862         if (!MLX5_CAP_GEN(mdev, cq_moderation))
2863                 mlx5_core_warn(mdev, "CQ modiration is not supported\n");
2864
2865         return 0;
2866 }
2867
2868 u16 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev)
2869 {
2870         int bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
2871
2872         return bf_buf_size -
2873                sizeof(struct mlx5e_tx_wqe) +
2874                2 /*sizeof(mlx5e_tx_wqe.inline_hdr_start)*/;
2875 }
2876
2877 #ifdef CONFIG_MLX5_CORE_EN_DCB
2878 static void mlx5e_ets_init(struct mlx5e_priv *priv)
2879 {
2880         int i;
2881
2882         priv->params.ets.ets_cap = mlx5_max_tc(priv->mdev) + 1;
2883         for (i = 0; i < priv->params.ets.ets_cap; i++) {
2884                 priv->params.ets.tc_tx_bw[i] = MLX5E_MAX_BW_ALLOC;
2885                 priv->params.ets.tc_tsa[i] = IEEE_8021QAZ_TSA_VENDOR;
2886                 priv->params.ets.prio_tc[i] = i;
2887         }
2888
2889         /* tclass[prio=0]=1, tclass[prio=1]=0, tclass[prio=i]=i (for i>1) */
2890         priv->params.ets.prio_tc[0] = 1;
2891         priv->params.ets.prio_tc[1] = 0;
2892 }
2893 #endif
2894
2895 void mlx5e_build_default_indir_rqt(struct mlx5_core_dev *mdev,
2896                                    u32 *indirection_rqt, int len,
2897                                    int num_channels)
2898 {
2899         int node = mdev->priv.numa_node;
2900         int node_num_of_cores;
2901         int i;
2902
2903         if (node == -1)
2904                 node = first_online_node;
2905
2906         node_num_of_cores = cpumask_weight(cpumask_of_node(node));
2907
2908         if (node_num_of_cores)
2909                 num_channels = min_t(int, num_channels, node_num_of_cores);
2910
2911         for (i = 0; i < len; i++)
2912                 indirection_rqt[i] = i % num_channels;
2913 }
2914
2915 static bool mlx5e_check_fragmented_striding_rq_cap(struct mlx5_core_dev *mdev)
2916 {
2917         return MLX5_CAP_GEN(mdev, striding_rq) &&
2918                 MLX5_CAP_GEN(mdev, umr_ptr_rlky) &&
2919                 MLX5_CAP_ETH(mdev, reg_umr_sq);
2920 }
2921
2922 static int mlx5e_get_pci_bw(struct mlx5_core_dev *mdev, u32 *pci_bw)
2923 {
2924         enum pcie_link_width width;
2925         enum pci_bus_speed speed;
2926         int err = 0;
2927
2928         err = pcie_get_minimum_link(mdev->pdev, &speed, &width);
2929         if (err)
2930                 return err;
2931
2932         if (speed == PCI_SPEED_UNKNOWN || width == PCIE_LNK_WIDTH_UNKNOWN)
2933                 return -EINVAL;
2934
2935         switch (speed) {
2936         case PCIE_SPEED_2_5GT:
2937                 *pci_bw = 2500 * width;
2938                 break;
2939         case PCIE_SPEED_5_0GT:
2940                 *pci_bw = 5000 * width;
2941                 break;
2942         case PCIE_SPEED_8_0GT:
2943                 *pci_bw = 8000 * width;
2944                 break;
2945         default:
2946                 return -EINVAL;
2947         }
2948
2949         return 0;
2950 }
2951
2952 static bool cqe_compress_heuristic(u32 link_speed, u32 pci_bw)
2953 {
2954         return (link_speed && pci_bw &&
2955                 (pci_bw < 40000) && (pci_bw < link_speed));
2956 }
2957
2958 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
2959 {
2960         params->rx_cq_period_mode = cq_period_mode;
2961
2962         params->rx_cq_moderation.pkts =
2963                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
2964         params->rx_cq_moderation.usec =
2965                         MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
2966
2967         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
2968                 params->rx_cq_moderation.usec =
2969                         MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE;
2970 }
2971
2972 static void mlx5e_query_min_inline(struct mlx5_core_dev *mdev,
2973                                    u8 *min_inline_mode)
2974 {
2975         switch (MLX5_CAP_ETH(mdev, wqe_inline_mode)) {
2976         case MLX5E_INLINE_MODE_L2:
2977                 *min_inline_mode = MLX5_INLINE_MODE_L2;
2978                 break;
2979         case MLX5E_INLINE_MODE_VPORT_CONTEXT:
2980                 mlx5_query_nic_vport_min_inline(mdev,
2981                                                 min_inline_mode);
2982                 break;
2983         case MLX5_INLINE_MODE_NOT_REQUIRED:
2984                 *min_inline_mode = MLX5_INLINE_MODE_NONE;
2985                 break;
2986         }
2987 }
2988
2989 static void mlx5e_build_nic_netdev_priv(struct mlx5_core_dev *mdev,
2990                                         struct net_device *netdev,
2991                                         const struct mlx5e_profile *profile,
2992                                         void *ppriv)
2993 {
2994         struct mlx5e_priv *priv = netdev_priv(netdev);
2995         u32 link_speed = 0;
2996         u32 pci_bw = 0;
2997         u8 cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
2998                                          MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
2999                                          MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
3000
3001         priv->params.log_sq_size           =
3002                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
3003         priv->params.rq_wq_type = mlx5e_check_fragmented_striding_rq_cap(mdev) ?
3004                 MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ :
3005                 MLX5_WQ_TYPE_LINKED_LIST;
3006
3007         /* set CQE compression */
3008         priv->params.rx_cqe_compress_admin = false;
3009         if (MLX5_CAP_GEN(mdev, cqe_compression) &&
3010             MLX5_CAP_GEN(mdev, vport_group_manager)) {
3011                 mlx5e_get_max_linkspeed(mdev, &link_speed);
3012                 mlx5e_get_pci_bw(mdev, &pci_bw);
3013                 mlx5_core_dbg(mdev, "Max link speed = %d, PCI BW = %d\n",
3014                               link_speed, pci_bw);
3015                 priv->params.rx_cqe_compress_admin =
3016                         cqe_compress_heuristic(link_speed, pci_bw);
3017         }
3018
3019         priv->params.rx_cqe_compress = priv->params.rx_cqe_compress_admin;
3020
3021         switch (priv->params.rq_wq_type) {
3022         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
3023                 priv->params.log_rq_size = MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE_MPW;
3024                 priv->params.mpwqe_log_stride_sz =
3025                         priv->params.rx_cqe_compress ?
3026                         MLX5_MPWRQ_LOG_STRIDE_SIZE_CQE_COMPRESS :
3027                         MLX5_MPWRQ_LOG_STRIDE_SIZE;
3028                 priv->params.mpwqe_log_num_strides = MLX5_MPWRQ_LOG_WQE_SZ -
3029                         priv->params.mpwqe_log_stride_sz;
3030                 priv->params.lro_en = true;
3031                 break;
3032         default: /* MLX5_WQ_TYPE_LINKED_LIST */
3033                 priv->params.log_rq_size = MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
3034         }
3035
3036         mlx5_core_info(mdev,
3037                        "MLX5E: StrdRq(%d) RqSz(%ld) StrdSz(%ld) RxCqeCmprss(%d)\n",
3038                        priv->params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ,
3039                        BIT(priv->params.log_rq_size),
3040                        BIT(priv->params.mpwqe_log_stride_sz),
3041                        priv->params.rx_cqe_compress_admin);
3042
3043         priv->params.min_rx_wqes = mlx5_min_rx_wqes(priv->params.rq_wq_type,
3044                                             BIT(priv->params.log_rq_size));
3045
3046         priv->params.rx_am_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
3047         mlx5e_set_rx_cq_mode_params(&priv->params, cq_period_mode);
3048
3049         priv->params.tx_cq_moderation.usec =
3050                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
3051         priv->params.tx_cq_moderation.pkts =
3052                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
3053         priv->params.tx_max_inline         = mlx5e_get_max_inline_cap(mdev);
3054         mlx5e_query_min_inline(mdev, &priv->params.tx_min_inline_mode);
3055         priv->params.num_tc                = 1;
3056         priv->params.rss_hfunc             = ETH_RSS_HASH_XOR;
3057
3058         netdev_rss_key_fill(priv->params.toeplitz_hash_key,
3059                             sizeof(priv->params.toeplitz_hash_key));
3060
3061         mlx5e_build_default_indir_rqt(mdev, priv->params.indirection_rqt,
3062                                       MLX5E_INDIR_RQT_SIZE, profile->max_nch(mdev));
3063
3064         priv->params.lro_wqe_sz            =
3065                 MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ;
3066
3067         /* Initialize pflags */
3068         MLX5E_SET_PRIV_FLAG(priv, MLX5E_PFLAG_RX_CQE_BASED_MODER,
3069                             priv->params.rx_cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
3070
3071         priv->mdev                         = mdev;
3072         priv->netdev                       = netdev;
3073         priv->params.num_channels          = profile->max_nch(mdev);
3074         priv->profile                      = profile;
3075         priv->ppriv                        = ppriv;
3076
3077 #ifdef CONFIG_MLX5_CORE_EN_DCB
3078         mlx5e_ets_init(priv);
3079 #endif
3080
3081         mutex_init(&priv->state_lock);
3082
3083         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
3084         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
3085         INIT_WORK(&priv->tx_timeout_work, mlx5e_tx_timeout_work);
3086         INIT_DELAYED_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
3087 }
3088
3089 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
3090 {
3091         struct mlx5e_priv *priv = netdev_priv(netdev);
3092
3093         mlx5_query_nic_vport_mac_address(priv->mdev, 0, netdev->dev_addr);
3094         if (is_zero_ether_addr(netdev->dev_addr) &&
3095             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
3096                 eth_hw_addr_random(netdev);
3097                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
3098         }
3099 }
3100
3101 static const struct switchdev_ops mlx5e_switchdev_ops = {
3102         .switchdev_port_attr_get        = mlx5e_attr_get,
3103 };
3104
3105 static void mlx5e_build_nic_netdev(struct net_device *netdev)
3106 {
3107         struct mlx5e_priv *priv = netdev_priv(netdev);
3108         struct mlx5_core_dev *mdev = priv->mdev;
3109         bool fcs_supported;
3110         bool fcs_enabled;
3111
3112         SET_NETDEV_DEV(netdev, &mdev->pdev->dev);
3113
3114         if (MLX5_CAP_GEN(mdev, vport_group_manager)) {
3115                 netdev->netdev_ops = &mlx5e_netdev_ops_sriov;
3116 #ifdef CONFIG_MLX5_CORE_EN_DCB
3117                 netdev->dcbnl_ops = &mlx5e_dcbnl_ops;
3118 #endif
3119         } else {
3120                 netdev->netdev_ops = &mlx5e_netdev_ops_basic;
3121         }
3122
3123         netdev->watchdog_timeo    = 15 * HZ;
3124
3125         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
3126
3127         netdev->vlan_features    |= NETIF_F_SG;
3128         netdev->vlan_features    |= NETIF_F_IP_CSUM;
3129         netdev->vlan_features    |= NETIF_F_IPV6_CSUM;
3130         netdev->vlan_features    |= NETIF_F_GRO;
3131         netdev->vlan_features    |= NETIF_F_TSO;
3132         netdev->vlan_features    |= NETIF_F_TSO6;
3133         netdev->vlan_features    |= NETIF_F_RXCSUM;
3134         netdev->vlan_features    |= NETIF_F_RXHASH;
3135
3136         if (!!MLX5_CAP_ETH(mdev, lro_cap))
3137                 netdev->vlan_features    |= NETIF_F_LRO;
3138
3139         netdev->hw_features       = netdev->vlan_features;
3140         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
3141         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
3142         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
3143
3144         if (mlx5e_vxlan_allowed(mdev)) {
3145                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL |
3146                                            NETIF_F_GSO_UDP_TUNNEL_CSUM |
3147                                            NETIF_F_GSO_PARTIAL;
3148                 netdev->hw_enc_features |= NETIF_F_IP_CSUM;
3149                 netdev->hw_enc_features |= NETIF_F_IPV6_CSUM;
3150                 netdev->hw_enc_features |= NETIF_F_TSO;
3151                 netdev->hw_enc_features |= NETIF_F_TSO6;
3152                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL;
3153                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL_CSUM |
3154                                            NETIF_F_GSO_PARTIAL;
3155                 netdev->gso_partial_features = NETIF_F_GSO_UDP_TUNNEL_CSUM;
3156         }
3157
3158         mlx5_query_port_fcs(mdev, &fcs_supported, &fcs_enabled);
3159
3160         if (fcs_supported)
3161                 netdev->hw_features |= NETIF_F_RXALL;
3162
3163         netdev->features          = netdev->hw_features;
3164         if (!priv->params.lro_en)
3165                 netdev->features  &= ~NETIF_F_LRO;
3166
3167         if (fcs_enabled)
3168                 netdev->features  &= ~NETIF_F_RXALL;
3169
3170 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
3171         if (FT_CAP(flow_modify_en) &&
3172             FT_CAP(modify_root) &&
3173             FT_CAP(identified_miss_table_mode) &&
3174             FT_CAP(flow_table_modify)) {
3175                 netdev->hw_features      |= NETIF_F_HW_TC;
3176 #ifdef CONFIG_RFS_ACCEL
3177                 netdev->hw_features      |= NETIF_F_NTUPLE;
3178 #endif
3179         }
3180
3181         netdev->features         |= NETIF_F_HIGHDMA;
3182
3183         netdev->priv_flags       |= IFF_UNICAST_FLT;
3184
3185         mlx5e_set_netdev_dev_addr(netdev);
3186
3187 #ifdef CONFIG_NET_SWITCHDEV
3188         if (MLX5_CAP_GEN(mdev, vport_group_manager))
3189                 netdev->switchdev_ops = &mlx5e_switchdev_ops;
3190 #endif
3191 }
3192
3193 static void mlx5e_create_q_counter(struct mlx5e_priv *priv)
3194 {
3195         struct mlx5_core_dev *mdev = priv->mdev;
3196         int err;
3197
3198         err = mlx5_core_alloc_q_counter(mdev, &priv->q_counter);
3199         if (err) {
3200                 mlx5_core_warn(mdev, "alloc queue counter failed, %d\n", err);
3201                 priv->q_counter = 0;
3202         }
3203 }
3204
3205 static void mlx5e_destroy_q_counter(struct mlx5e_priv *priv)
3206 {
3207         if (!priv->q_counter)
3208                 return;
3209
3210         mlx5_core_dealloc_q_counter(priv->mdev, priv->q_counter);
3211 }
3212
3213 static int mlx5e_create_umr_mkey(struct mlx5e_priv *priv)
3214 {
3215         struct mlx5_core_dev *mdev = priv->mdev;
3216         u64 npages = MLX5E_REQUIRED_MTTS(priv->profile->max_nch(mdev),
3217                                          BIT(MLX5E_PARAMS_MAXIMUM_LOG_RQ_SIZE_MPW));
3218         int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
3219         void *mkc;
3220         u32 *in;
3221         int err;
3222
3223         in = mlx5_vzalloc(inlen);
3224         if (!in)
3225                 return -ENOMEM;
3226
3227         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
3228
3229         npages = min_t(u32, ALIGN(U16_MAX, 4) * 2, npages);
3230
3231         MLX5_SET(mkc, mkc, free, 1);
3232         MLX5_SET(mkc, mkc, umr_en, 1);
3233         MLX5_SET(mkc, mkc, lw, 1);
3234         MLX5_SET(mkc, mkc, lr, 1);
3235         MLX5_SET(mkc, mkc, access_mode, MLX5_MKC_ACCESS_MODE_MTT);
3236
3237         MLX5_SET(mkc, mkc, qpn, 0xffffff);
3238         MLX5_SET(mkc, mkc, pd, mdev->mlx5e_res.pdn);
3239         MLX5_SET64(mkc, mkc, len, npages << PAGE_SHIFT);
3240         MLX5_SET(mkc, mkc, translations_octword_size,
3241                  MLX5_MTT_OCTW(npages));
3242         MLX5_SET(mkc, mkc, log_page_size, PAGE_SHIFT);
3243
3244         err = mlx5_core_create_mkey(mdev, &priv->umr_mkey, in, inlen);
3245
3246         kvfree(in);
3247         return err;
3248 }
3249
3250 static void mlx5e_nic_init(struct mlx5_core_dev *mdev,
3251                            struct net_device *netdev,
3252                            const struct mlx5e_profile *profile,
3253                            void *ppriv)
3254 {
3255         struct mlx5e_priv *priv = netdev_priv(netdev);
3256
3257         mlx5e_build_nic_netdev_priv(mdev, netdev, profile, ppriv);
3258         mlx5e_build_nic_netdev(netdev);
3259         mlx5e_vxlan_init(priv);
3260 }
3261
3262 static void mlx5e_nic_cleanup(struct mlx5e_priv *priv)
3263 {
3264         struct mlx5_core_dev *mdev = priv->mdev;
3265         struct mlx5_eswitch *esw = mdev->priv.eswitch;
3266
3267         mlx5e_vxlan_cleanup(priv);
3268
3269         if (MLX5_CAP_GEN(mdev, vport_group_manager))
3270                 mlx5_eswitch_unregister_vport_rep(esw, 0);
3271 }
3272
3273 static int mlx5e_init_nic_rx(struct mlx5e_priv *priv)
3274 {
3275         struct mlx5_core_dev *mdev = priv->mdev;
3276         int err;
3277         int i;
3278
3279         err = mlx5e_create_indirect_rqts(priv);
3280         if (err) {
3281                 mlx5_core_warn(mdev, "create indirect rqts failed, %d\n", err);
3282                 return err;
3283         }
3284
3285         err = mlx5e_create_direct_rqts(priv);
3286         if (err) {
3287                 mlx5_core_warn(mdev, "create direct rqts failed, %d\n", err);
3288                 goto err_destroy_indirect_rqts;
3289         }
3290
3291         err = mlx5e_create_indirect_tirs(priv);
3292         if (err) {
3293                 mlx5_core_warn(mdev, "create indirect tirs failed, %d\n", err);
3294                 goto err_destroy_direct_rqts;
3295         }
3296
3297         err = mlx5e_create_direct_tirs(priv);
3298         if (err) {
3299                 mlx5_core_warn(mdev, "create direct tirs failed, %d\n", err);
3300                 goto err_destroy_indirect_tirs;
3301         }
3302
3303         err = mlx5e_create_flow_steering(priv);
3304         if (err) {
3305                 mlx5_core_warn(mdev, "create flow steering failed, %d\n", err);
3306                 goto err_destroy_direct_tirs;
3307         }
3308
3309         err = mlx5e_tc_init(priv);
3310         if (err)
3311                 goto err_destroy_flow_steering;
3312
3313         return 0;
3314
3315 err_destroy_flow_steering:
3316         mlx5e_destroy_flow_steering(priv);
3317 err_destroy_direct_tirs:
3318         mlx5e_destroy_direct_tirs(priv);
3319 err_destroy_indirect_tirs:
3320         mlx5e_destroy_indirect_tirs(priv);
3321 err_destroy_direct_rqts:
3322         for (i = 0; i < priv->profile->max_nch(mdev); i++)
3323                 mlx5e_destroy_rqt(priv, &priv->direct_tir[i].rqt);
3324 err_destroy_indirect_rqts:
3325         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
3326         return err;
3327 }
3328
3329 static void mlx5e_cleanup_nic_rx(struct mlx5e_priv *priv)
3330 {
3331         int i;
3332
3333         mlx5e_tc_cleanup(priv);
3334         mlx5e_destroy_flow_steering(priv);
3335         mlx5e_destroy_direct_tirs(priv);
3336         mlx5e_destroy_indirect_tirs(priv);
3337         for (i = 0; i < priv->profile->max_nch(priv->mdev); i++)
3338                 mlx5e_destroy_rqt(priv, &priv->direct_tir[i].rqt);
3339         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
3340 }
3341
3342 static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
3343 {
3344         int err;
3345
3346         err = mlx5e_create_tises(priv);
3347         if (err) {
3348                 mlx5_core_warn(priv->mdev, "create tises failed, %d\n", err);
3349                 return err;
3350         }
3351
3352 #ifdef CONFIG_MLX5_CORE_EN_DCB
3353         mlx5e_dcbnl_ieee_setets_core(priv, &priv->params.ets);
3354 #endif
3355         return 0;
3356 }
3357
3358 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
3359 {
3360         struct net_device *netdev = priv->netdev;
3361         struct mlx5_core_dev *mdev = priv->mdev;
3362         struct mlx5_eswitch *esw = mdev->priv.eswitch;
3363         struct mlx5_eswitch_rep rep;
3364
3365         mlx5_lag_add(mdev, netdev);
3366
3367         if (mlx5e_vxlan_allowed(mdev)) {
3368                 rtnl_lock();
3369                 udp_tunnel_get_rx_info(netdev);
3370                 rtnl_unlock();
3371         }
3372
3373         mlx5e_enable_async_events(priv);
3374         queue_work(priv->wq, &priv->set_rx_mode_work);
3375
3376         if (MLX5_CAP_GEN(mdev, vport_group_manager)) {
3377                 mlx5_query_nic_vport_mac_address(mdev, 0, rep.hw_id);
3378                 rep.load = mlx5e_nic_rep_load;
3379                 rep.unload = mlx5e_nic_rep_unload;
3380                 rep.vport = 0;
3381                 rep.priv_data = priv;
3382                 mlx5_eswitch_register_vport_rep(esw, &rep);
3383         }
3384 }
3385
3386 static void mlx5e_nic_disable(struct mlx5e_priv *priv)
3387 {
3388         queue_work(priv->wq, &priv->set_rx_mode_work);
3389         mlx5e_disable_async_events(priv);
3390         mlx5_lag_remove(priv->mdev);
3391 }
3392
3393 static const struct mlx5e_profile mlx5e_nic_profile = {
3394         .init              = mlx5e_nic_init,
3395         .cleanup           = mlx5e_nic_cleanup,
3396         .init_rx           = mlx5e_init_nic_rx,
3397         .cleanup_rx        = mlx5e_cleanup_nic_rx,
3398         .init_tx           = mlx5e_init_nic_tx,
3399         .cleanup_tx        = mlx5e_cleanup_nic_tx,
3400         .enable            = mlx5e_nic_enable,
3401         .disable           = mlx5e_nic_disable,
3402         .update_stats      = mlx5e_update_stats,
3403         .max_nch           = mlx5e_get_max_num_channels,
3404         .max_tc            = MLX5E_MAX_NUM_TC,
3405 };
3406
3407 struct net_device *mlx5e_create_netdev(struct mlx5_core_dev *mdev,
3408                                        const struct mlx5e_profile *profile,
3409                                        void *ppriv)
3410 {
3411         int nch = profile->max_nch(mdev);
3412         struct net_device *netdev;
3413         struct mlx5e_priv *priv;
3414
3415         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv),
3416                                     nch * profile->max_tc,
3417                                     nch);
3418         if (!netdev) {
3419                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
3420                 return NULL;
3421         }
3422
3423         profile->init(mdev, netdev, profile, ppriv);
3424
3425         netif_carrier_off(netdev);
3426
3427         priv = netdev_priv(netdev);
3428
3429         priv->wq = create_singlethread_workqueue("mlx5e");
3430         if (!priv->wq)
3431                 goto err_cleanup_nic;
3432
3433         return netdev;
3434
3435 err_cleanup_nic:
3436         profile->cleanup(priv);
3437         free_netdev(netdev);
3438
3439         return NULL;
3440 }
3441
3442 int mlx5e_attach_netdev(struct mlx5_core_dev *mdev, struct net_device *netdev)
3443 {
3444         const struct mlx5e_profile *profile;
3445         struct mlx5e_priv *priv;
3446         int err;
3447
3448         priv = netdev_priv(netdev);
3449         profile = priv->profile;
3450         clear_bit(MLX5E_STATE_DESTROYING, &priv->state);
3451
3452         err = mlx5e_create_umr_mkey(priv);
3453         if (err) {
3454                 mlx5_core_err(mdev, "create umr mkey failed, %d\n", err);
3455                 goto out;
3456         }
3457
3458         err = profile->init_tx(priv);
3459         if (err)
3460                 goto err_destroy_umr_mkey;
3461
3462         err = mlx5e_open_drop_rq(priv);
3463         if (err) {
3464                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
3465                 goto err_cleanup_tx;
3466         }
3467
3468         err = profile->init_rx(priv);
3469         if (err)
3470                 goto err_close_drop_rq;
3471
3472         mlx5e_create_q_counter(priv);
3473
3474         mlx5e_init_l2_addr(priv);
3475
3476         mlx5e_set_dev_port_mtu(netdev);
3477
3478         if (profile->enable)
3479                 profile->enable(priv);
3480
3481         rtnl_lock();
3482         if (netif_running(netdev))
3483                 mlx5e_open(netdev);
3484         netif_device_attach(netdev);
3485         rtnl_unlock();
3486
3487         return 0;
3488
3489 err_close_drop_rq:
3490         mlx5e_close_drop_rq(priv);
3491
3492 err_cleanup_tx:
3493         profile->cleanup_tx(priv);
3494
3495 err_destroy_umr_mkey:
3496         mlx5_core_destroy_mkey(mdev, &priv->umr_mkey);
3497
3498 out:
3499         return err;
3500 }
3501
3502 static void mlx5e_register_vport_rep(struct mlx5_core_dev *mdev)
3503 {
3504         struct mlx5_eswitch *esw = mdev->priv.eswitch;
3505         int total_vfs = MLX5_TOTAL_VPORTS(mdev);
3506         int vport;
3507         u8 mac[ETH_ALEN];
3508
3509         if (!MLX5_CAP_GEN(mdev, vport_group_manager))
3510                 return;
3511
3512         mlx5_query_nic_vport_mac_address(mdev, 0, mac);
3513
3514         for (vport = 1; vport < total_vfs; vport++) {
3515                 struct mlx5_eswitch_rep rep;
3516
3517                 rep.load = mlx5e_vport_rep_load;
3518                 rep.unload = mlx5e_vport_rep_unload;
3519                 rep.vport = vport;
3520                 ether_addr_copy(rep.hw_id, mac);
3521                 mlx5_eswitch_register_vport_rep(esw, &rep);
3522         }
3523 }
3524
3525 void mlx5e_detach_netdev(struct mlx5_core_dev *mdev, struct net_device *netdev)
3526 {
3527         struct mlx5e_priv *priv = netdev_priv(netdev);
3528         const struct mlx5e_profile *profile = priv->profile;
3529
3530         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
3531         if (profile->disable)
3532                 profile->disable(priv);
3533
3534         flush_workqueue(priv->wq);
3535
3536         rtnl_lock();
3537         if (netif_running(netdev))
3538                 mlx5e_close(netdev);
3539         netif_device_detach(netdev);
3540         rtnl_unlock();
3541
3542         mlx5e_destroy_q_counter(priv);
3543         profile->cleanup_rx(priv);
3544         mlx5e_close_drop_rq(priv);
3545         profile->cleanup_tx(priv);
3546         mlx5_core_destroy_mkey(priv->mdev, &priv->umr_mkey);
3547         cancel_delayed_work_sync(&priv->update_stats_work);
3548 }
3549
3550 /* mlx5e_attach and mlx5e_detach scope should be only creating/destroying
3551  * hardware contexts and to connect it to the current netdev.
3552  */
3553 static int mlx5e_attach(struct mlx5_core_dev *mdev, void *vpriv)
3554 {
3555         struct mlx5e_priv *priv = vpriv;
3556         struct net_device *netdev = priv->netdev;
3557         int err;
3558
3559         if (netif_device_present(netdev))
3560                 return 0;
3561
3562         err = mlx5e_create_mdev_resources(mdev);
3563         if (err)
3564                 return err;
3565
3566         err = mlx5e_attach_netdev(mdev, netdev);
3567         if (err) {
3568                 mlx5e_destroy_mdev_resources(mdev);
3569                 return err;
3570         }
3571
3572         return 0;
3573 }
3574
3575 static void mlx5e_detach(struct mlx5_core_dev *mdev, void *vpriv)
3576 {
3577         struct mlx5e_priv *priv = vpriv;
3578         struct net_device *netdev = priv->netdev;
3579
3580         if (!netif_device_present(netdev))
3581                 return;
3582
3583         mlx5e_detach_netdev(mdev, netdev);
3584         mlx5e_destroy_mdev_resources(mdev);
3585 }
3586
3587 static void *mlx5e_add(struct mlx5_core_dev *mdev)
3588 {
3589         struct mlx5_eswitch *esw = mdev->priv.eswitch;
3590         int total_vfs = MLX5_TOTAL_VPORTS(mdev);
3591         void *ppriv = NULL;
3592         void *priv;
3593         int vport;
3594         int err;
3595         struct net_device *netdev;
3596
3597         err = mlx5e_check_required_hca_cap(mdev);
3598         if (err)
3599                 return NULL;
3600
3601         mlx5e_register_vport_rep(mdev);
3602
3603         if (MLX5_CAP_GEN(mdev, vport_group_manager))
3604                 ppriv = &esw->offloads.vport_reps[0];
3605
3606         netdev = mlx5e_create_netdev(mdev, &mlx5e_nic_profile, ppriv);
3607         if (!netdev) {
3608                 mlx5_core_err(mdev, "mlx5e_create_netdev failed\n");
3609                 goto err_unregister_reps;
3610         }
3611
3612         priv = netdev_priv(netdev);
3613
3614         err = mlx5e_attach(mdev, priv);
3615         if (err) {
3616                 mlx5_core_err(mdev, "mlx5e_attach failed, %d\n", err);
3617                 goto err_destroy_netdev;
3618         }
3619
3620         err = register_netdev(netdev);
3621         if (err) {
3622                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
3623                 goto err_detach;
3624         }
3625
3626         return priv;
3627
3628 err_detach:
3629         mlx5e_detach(mdev, priv);
3630
3631 err_destroy_netdev:
3632         mlx5e_destroy_netdev(mdev, priv);
3633
3634 err_unregister_reps:
3635         for (vport = 1; vport < total_vfs; vport++)
3636                 mlx5_eswitch_unregister_vport_rep(esw, vport);
3637
3638         return NULL;
3639 }
3640
3641 void mlx5e_destroy_netdev(struct mlx5_core_dev *mdev, struct mlx5e_priv *priv)
3642 {
3643         const struct mlx5e_profile *profile = priv->profile;
3644         struct net_device *netdev = priv->netdev;
3645
3646         unregister_netdev(netdev);
3647         destroy_workqueue(priv->wq);
3648         if (profile->cleanup)
3649                 profile->cleanup(priv);
3650         free_netdev(netdev);
3651 }
3652
3653 static void mlx5e_remove(struct mlx5_core_dev *mdev, void *vpriv)
3654 {
3655         struct mlx5_eswitch *esw = mdev->priv.eswitch;
3656         int total_vfs = MLX5_TOTAL_VPORTS(mdev);
3657         struct mlx5e_priv *priv = vpriv;
3658         int vport;
3659
3660         for (vport = 1; vport < total_vfs; vport++)
3661                 mlx5_eswitch_unregister_vport_rep(esw, vport);
3662
3663         mlx5e_detach(mdev, vpriv);
3664         mlx5e_destroy_netdev(mdev, priv);
3665 }
3666
3667 static void *mlx5e_get_netdev(void *vpriv)
3668 {
3669         struct mlx5e_priv *priv = vpriv;
3670
3671         return priv->netdev;
3672 }
3673
3674 static struct mlx5_interface mlx5e_interface = {
3675         .add       = mlx5e_add,
3676         .remove    = mlx5e_remove,
3677         .attach    = mlx5e_attach,
3678         .detach    = mlx5e_detach,
3679         .event     = mlx5e_async_event,
3680         .protocol  = MLX5_INTERFACE_PROTOCOL_ETH,
3681         .get_dev   = mlx5e_get_netdev,
3682 };
3683
3684 void mlx5e_init(void)
3685 {
3686         mlx5e_build_ptys2ethtool_map();
3687         mlx5_register_interface(&mlx5e_interface);
3688 }
3689
3690 void mlx5e_cleanup(void)
3691 {
3692         mlx5_unregister_interface(&mlx5e_interface);
3693 }