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 <linux/bpf.h>
  38 #include "eswitch.h"
  39 #include "en.h"
  40 #include "en_tc.h"
  41 #include "en_rep.h"
  42 #include "en_accel/ipsec.h"
  43 #include "en_accel/ipsec_rxtx.h"
  44 #include "accel/ipsec.h"
  45 #include "vxlan.h"
  46
  47 struct mlx5e_rq_param {
  48         u32                     rqc[MLX5_ST_SZ_DW(rqc)];
  49         struct mlx5_wq_param    wq;
  50 };
  51
  52 struct mlx5e_sq_param {
  53         u32                        sqc[MLX5_ST_SZ_DW(sqc)];
  54         struct mlx5_wq_param       wq;
  55 };
  56
  57 struct mlx5e_cq_param {
  58         u32                        cqc[MLX5_ST_SZ_DW(cqc)];
  59         struct mlx5_wq_param       wq;
  60         u16                        eq_ix;
  61         u8                         cq_period_mode;
  62 };
  63
  64 struct mlx5e_channel_param {
  65         struct mlx5e_rq_param      rq;
  66         struct mlx5e_sq_param      sq;
  67         struct mlx5e_sq_param      xdp_sq;
  68         struct mlx5e_sq_param      icosq;
  69         struct mlx5e_cq_param      rx_cq;
  70         struct mlx5e_cq_param      tx_cq;
  71         struct mlx5e_cq_param      icosq_cq;
  72 };
  73
  74 static bool mlx5e_check_fragmented_striding_rq_cap(struct mlx5_core_dev *mdev)
  75 {
  76         return MLX5_CAP_GEN(mdev, striding_rq) &&
  77                 MLX5_CAP_GEN(mdev, umr_ptr_rlky) &&
  78                 MLX5_CAP_ETH(mdev, reg_umr_sq);
  79 }
  80
  81 void mlx5e_init_rq_type_params(struct mlx5_core_dev *mdev,
  82                                struct mlx5e_params *params, u8 rq_type)
  83 {
  84         params->rq_wq_type = rq_type;
  85         params->lro_wqe_sz = MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ;
  86         switch (params->rq_wq_type) {
  87         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
  88                 params->log_rq_size = is_kdump_kernel() ?
  89                         MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE_MPW :
  90                         MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE_MPW;
  91                 params->mpwqe_log_stride_sz = MLX5E_MPWQE_STRIDE_SZ(mdev,
  92                         MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS));
  93                 params->mpwqe_log_num_strides = MLX5_MPWRQ_LOG_WQE_SZ -
  94                         params->mpwqe_log_stride_sz;
  95                 break;
  96         default: /* MLX5_WQ_TYPE_LINKED_LIST */
  97                 params->log_rq_size = is_kdump_kernel() ?
  98                         MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE :
  99                         MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
 100                 params->rq_headroom = params->xdp_prog ?
 101                         XDP_PACKET_HEADROOM : MLX5_RX_HEADROOM;
 102                 params->rq_headroom += NET_IP_ALIGN;
 103
 104                 /* Extra room needed for build_skb */
 105                 params->lro_wqe_sz -= params->rq_headroom +
 106                         SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 107         }
 108
 109         mlx5_core_info(mdev, "MLX5E: StrdRq(%d) RqSz(%ld) StrdSz(%ld) RxCqeCmprss(%d)\n",
 110                        params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ,
 111                        BIT(params->log_rq_size),
 112                        BIT(params->mpwqe_log_stride_sz),
 113                        MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS));
 114 }
 115
 116 static void mlx5e_set_rq_params(struct mlx5_core_dev *mdev,
 117                                 struct mlx5e_params *params)
 118 {
 119         u8 rq_type = mlx5e_check_fragmented_striding_rq_cap(mdev) &&
 120                     !params->xdp_prog && !MLX5_IPSEC_DEV(mdev) ?
 121                     MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ :
 122                     MLX5_WQ_TYPE_LINKED_LIST;
 123         mlx5e_init_rq_type_params(mdev, params, rq_type);
 124 }
 125
 126 static void mlx5e_update_carrier(struct mlx5e_priv *priv)
 127 {
 128         struct mlx5_core_dev *mdev = priv->mdev;
 129         u8 port_state;
 130
 131         port_state = mlx5_query_vport_state(mdev,
 132                                             MLX5_QUERY_VPORT_STATE_IN_OP_MOD_VNIC_VPORT,
 133                                             0);
 134
 135         if (port_state == VPORT_STATE_UP) {
 136                 netdev_info(priv->netdev, "Link up\n");
 137                 netif_carrier_on(priv->netdev);
 138         } else {
 139                 netdev_info(priv->netdev, "Link down\n");
 140                 netif_carrier_off(priv->netdev);
 141         }
 142 }
 143
 144 static void mlx5e_update_carrier_work(struct work_struct *work)
 145 {
 146         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 147                                                update_carrier_work);
 148
 149         mutex_lock(&priv->state_lock);
 150         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
 151                 if (priv->profile->update_carrier)
 152                         priv->profile->update_carrier(priv);
 153         mutex_unlock(&priv->state_lock);
 154 }
 155
 156 static void mlx5e_tx_timeout_work(struct work_struct *work)
 157 {
 158         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 159                                                tx_timeout_work);
 160         int err;
 161
 162         rtnl_lock();
 163         mutex_lock(&priv->state_lock);
 164         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
 165                 goto unlock;
 166         mlx5e_close_locked(priv->netdev);
 167         err = mlx5e_open_locked(priv->netdev);
 168         if (err)
 169                 netdev_err(priv->netdev, "mlx5e_open_locked failed recovering from a tx_timeout, err(%d).\n",
 170                            err);
 171 unlock:
 172         mutex_unlock(&priv->state_lock);
 173         rtnl_unlock();
 174 }
 175
 176 static void mlx5e_update_sw_counters(struct mlx5e_priv *priv)
 177 {
 178         struct mlx5e_sw_stats temp, *s = &temp;
 179         struct mlx5e_rq_stats *rq_stats;
 180         struct mlx5e_sq_stats *sq_stats;
 181         int i, j;
 182
 183         memset(s, 0, sizeof(*s));
 184         for (i = 0; i < priv->channels.num; i++) {
 185                 struct mlx5e_channel *c = priv->channels.c[i];
 186
 187                 rq_stats = &c->rq.stats;
 188
 189                 s->rx_packets   += rq_stats->packets;
 190                 s->rx_bytes     += rq_stats->bytes;
 191                 s->rx_lro_packets += rq_stats->lro_packets;
 192                 s->rx_lro_bytes += rq_stats->lro_bytes;
 193                 s->rx_removed_vlan_packets += rq_stats->removed_vlan_packets;
 194                 s->rx_csum_none += rq_stats->csum_none;
 195                 s->rx_csum_complete += rq_stats->csum_complete;
 196                 s->rx_csum_unnecessary += rq_stats->csum_unnecessary;
 197                 s->rx_csum_unnecessary_inner += rq_stats->csum_unnecessary_inner;
 198                 s->rx_xdp_drop += rq_stats->xdp_drop;
 199                 s->rx_xdp_tx += rq_stats->xdp_tx;
 200                 s->rx_xdp_tx_full += rq_stats->xdp_tx_full;
 201                 s->rx_wqe_err   += rq_stats->wqe_err;
 202                 s->rx_mpwqe_filler += rq_stats->mpwqe_filler;
 203                 s->rx_buff_alloc_err += rq_stats->buff_alloc_err;
 204                 s->rx_cqe_compress_blks += rq_stats->cqe_compress_blks;
 205                 s->rx_cqe_compress_pkts += rq_stats->cqe_compress_pkts;
 206                 s->rx_page_reuse  += rq_stats->page_reuse;
 207                 s->rx_cache_reuse += rq_stats->cache_reuse;
 208                 s->rx_cache_full  += rq_stats->cache_full;
 209                 s->rx_cache_empty += rq_stats->cache_empty;
 210                 s->rx_cache_busy  += rq_stats->cache_busy;
 211                 s->rx_cache_waive += rq_stats->cache_waive;
 212
 213                 for (j = 0; j < priv->channels.params.num_tc; j++) {
 214                         sq_stats = &c->sq[j].stats;
 215
 216                         s->tx_packets           += sq_stats->packets;
 217                         s->tx_bytes             += sq_stats->bytes;
 218                         s->tx_tso_packets       += sq_stats->tso_packets;
 219                         s->tx_tso_bytes         += sq_stats->tso_bytes;
 220                         s->tx_tso_inner_packets += sq_stats->tso_inner_packets;
 221                         s->tx_tso_inner_bytes   += sq_stats->tso_inner_bytes;
 222                         s->tx_added_vlan_packets += sq_stats->added_vlan_packets;
 223                         s->tx_queue_stopped     += sq_stats->stopped;
 224                         s->tx_queue_wake        += sq_stats->wake;
 225                         s->tx_queue_dropped     += sq_stats->dropped;
 226                         s->tx_xmit_more         += sq_stats->xmit_more;
 227                         s->tx_csum_partial_inner += sq_stats->csum_partial_inner;
 228                         s->tx_csum_none         += sq_stats->csum_none;
 229                         s->tx_csum_partial      += sq_stats->csum_partial;
 230                 }
 231         }
 232
 233         s->link_down_events_phy = MLX5_GET(ppcnt_reg,
 234                                 priv->stats.pport.phy_counters,
 235                                 counter_set.phys_layer_cntrs.link_down_events);
 236         memcpy(&priv->stats.sw, s, sizeof(*s));
 237 }
 238
 239 static void mlx5e_update_vport_counters(struct mlx5e_priv *priv)
 240 {
 241         int outlen = MLX5_ST_SZ_BYTES(query_vport_counter_out);
 242         u32 *out = (u32 *)priv->stats.vport.query_vport_out;
 243         u32 in[MLX5_ST_SZ_DW(query_vport_counter_in)] = {0};
 244         struct mlx5_core_dev *mdev = priv->mdev;
 245
 246         MLX5_SET(query_vport_counter_in, in, opcode,
 247                  MLX5_CMD_OP_QUERY_VPORT_COUNTER);
 248         MLX5_SET(query_vport_counter_in, in, op_mod, 0);
 249         MLX5_SET(query_vport_counter_in, in, other_vport, 0);
 250
 251         mlx5_cmd_exec(mdev, in, sizeof(in), out, outlen);
 252 }
 253
 254 static void mlx5e_update_pport_counters(struct mlx5e_priv *priv, bool full)
 255 {
 256         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
 257         struct mlx5_core_dev *mdev = priv->mdev;
 258         u32 in[MLX5_ST_SZ_DW(ppcnt_reg)] = {0};
 259         int sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
 260         int prio;
 261         void *out;
 262
 263         MLX5_SET(ppcnt_reg, in, local_port, 1);
 264
 265         out = pstats->IEEE_802_3_counters;
 266         MLX5_SET(ppcnt_reg, in, grp, MLX5_IEEE_802_3_COUNTERS_GROUP);
 267         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 268
 269         if (!full)
 270                 return;
 271
 272         out = pstats->RFC_2863_counters;
 273         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2863_COUNTERS_GROUP);
 274         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 275
 276         out = pstats->RFC_2819_counters;
 277         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2819_COUNTERS_GROUP);
 278         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 279
 280         out = pstats->phy_counters;
 281         MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_COUNTERS_GROUP);
 282         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 283
 284         if (MLX5_CAP_PCAM_FEATURE(mdev, ppcnt_statistical_group)) {
 285                 out = pstats->phy_statistical_counters;
 286                 MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_STATISTICAL_GROUP);
 287                 mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 288         }
 289
 290         if (MLX5_CAP_PCAM_FEATURE(mdev, rx_buffer_fullness_counters)) {
 291                 out = pstats->eth_ext_counters;
 292                 MLX5_SET(ppcnt_reg, in, grp, MLX5_ETHERNET_EXTENDED_COUNTERS_GROUP);
 293                 mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 294         }
 295
 296         MLX5_SET(ppcnt_reg, in, grp, MLX5_PER_PRIORITY_COUNTERS_GROUP);
 297         for (prio = 0; prio < NUM_PPORT_PRIO; prio++) {
 298                 out = pstats->per_prio_counters[prio];
 299                 MLX5_SET(ppcnt_reg, in, prio_tc, prio);
 300                 mlx5_core_access_reg(mdev, in, sz, out, sz,
 301                                      MLX5_REG_PPCNT, 0, 0);
 302         }
 303 }
 304
 305 static void mlx5e_update_q_counter(struct mlx5e_priv *priv)
 306 {
 307         struct mlx5e_qcounter_stats *qcnt = &priv->stats.qcnt;
 308         u32 out[MLX5_ST_SZ_DW(query_q_counter_out)];
 309         int err;
 310
 311         if (!priv->q_counter)
 312                 return;
 313
 314         err = mlx5_core_query_q_counter(priv->mdev, priv->q_counter, 0, out, sizeof(out));
 315         if (err)
 316                 return;
 317
 318         qcnt->rx_out_of_buffer = MLX5_GET(query_q_counter_out, out, out_of_buffer);
 319 }
 320
 321 static void mlx5e_update_pcie_counters(struct mlx5e_priv *priv)
 322 {
 323         struct mlx5e_pcie_stats *pcie_stats = &priv->stats.pcie;
 324         struct mlx5_core_dev *mdev = priv->mdev;
 325         u32 in[MLX5_ST_SZ_DW(mpcnt_reg)] = {0};
 326         int sz = MLX5_ST_SZ_BYTES(mpcnt_reg);
 327         void *out;
 328
 329         if (!MLX5_CAP_MCAM_FEATURE(mdev, pcie_performance_group))
 330                 return;
 331
 332         out = pcie_stats->pcie_perf_counters;
 333         MLX5_SET(mpcnt_reg, in, grp, MLX5_PCIE_PERFORMANCE_COUNTERS_GROUP);
 334         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_MPCNT, 0, 0);
 335 }
 336
 337 void mlx5e_update_stats(struct mlx5e_priv *priv, bool full)
 338 {
 339         if (full) {
 340                 mlx5e_update_pcie_counters(priv);
 341                 mlx5e_ipsec_update_stats(priv);
 342         }
 343         mlx5e_update_pport_counters(priv, full);
 344         mlx5e_update_vport_counters(priv);
 345         mlx5e_update_q_counter(priv);
 346         mlx5e_update_sw_counters(priv);
 347 }
 348
 349 static void mlx5e_update_ndo_stats(struct mlx5e_priv *priv)
 350 {
 351         mlx5e_update_stats(priv, false);
 352 }
 353
 354 void mlx5e_update_stats_work(struct work_struct *work)
 355 {
 356         struct delayed_work *dwork = to_delayed_work(work);
 357         struct mlx5e_priv *priv = container_of(dwork, struct mlx5e_priv,
 358                                                update_stats_work);
 359         mutex_lock(&priv->state_lock);
 360         if (test_bit(MLX5E_STATE_OPENED, &priv->state)) {
 361                 priv->profile->update_stats(priv);
 362                 queue_delayed_work(priv->wq, dwork,
 363                                    msecs_to_jiffies(MLX5E_UPDATE_STATS_INTERVAL));
 364         }
 365         mutex_unlock(&priv->state_lock);
 366 }
 367
 368 static void mlx5e_async_event(struct mlx5_core_dev *mdev, void *vpriv,
 369                               enum mlx5_dev_event event, unsigned long param)
 370 {
 371         struct mlx5e_priv *priv = vpriv;
 372
 373         if (!test_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLED, &priv->state))
 374                 return;
 375
 376         switch (event) {
 377         case MLX5_DEV_EVENT_PORT_UP:
 378         case MLX5_DEV_EVENT_PORT_DOWN:
 379                 queue_work(priv->wq, &priv->update_carrier_work);
 380                 break;
 381         default:
 382                 break;
 383         }
 384 }
 385
 386 static void mlx5e_enable_async_events(struct mlx5e_priv *priv)
 387 {
 388         set_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLED, &priv->state);
 389 }
 390
 391 static void mlx5e_disable_async_events(struct mlx5e_priv *priv)
 392 {
 393         clear_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLED, &priv->state);
 394         synchronize_irq(pci_irq_vector(priv->mdev->pdev, MLX5_EQ_VEC_ASYNC));
 395 }
 396
 397 static inline int mlx5e_get_wqe_mtt_sz(void)
 398 {
 399         /* UMR copies MTTs in units of MLX5_UMR_MTT_ALIGNMENT bytes.
 400          * To avoid copying garbage after the mtt array, we allocate
 401          * a little more.
 402          */
 403         return ALIGN(MLX5_MPWRQ_PAGES_PER_WQE * sizeof(__be64),
 404                      MLX5_UMR_MTT_ALIGNMENT);
 405 }
 406
 407 static inline void mlx5e_build_umr_wqe(struct mlx5e_rq *rq,
 408                                        struct mlx5e_icosq *sq,
 409                                        struct mlx5e_umr_wqe *wqe,
 410                                        u16 ix)
 411 {
 412         struct mlx5_wqe_ctrl_seg      *cseg = &wqe->ctrl;
 413         struct mlx5_wqe_umr_ctrl_seg *ucseg = &wqe->uctrl;
 414         struct mlx5_wqe_data_seg      *dseg = &wqe->data;
 415         struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
 416         u8 ds_cnt = DIV_ROUND_UP(sizeof(*wqe), MLX5_SEND_WQE_DS);
 417         u32 umr_wqe_mtt_offset = mlx5e_get_wqe_mtt_offset(rq, ix);
 418
 419         cseg->qpn_ds    = cpu_to_be32((sq->sqn << MLX5_WQE_CTRL_QPN_SHIFT) |
 420                                       ds_cnt);
 421         cseg->fm_ce_se  = MLX5_WQE_CTRL_CQ_UPDATE;
 422         cseg->imm       = rq->mkey_be;
 423
 424         ucseg->flags = MLX5_UMR_TRANSLATION_OFFSET_EN;
 425         ucseg->xlt_octowords =
 426                 cpu_to_be16(MLX5_MTT_OCTW(MLX5_MPWRQ_PAGES_PER_WQE));
 427         ucseg->bsf_octowords =
 428                 cpu_to_be16(MLX5_MTT_OCTW(umr_wqe_mtt_offset));
 429         ucseg->mkey_mask     = cpu_to_be64(MLX5_MKEY_MASK_FREE);
 430
 431         dseg->lkey = sq->mkey_be;
 432         dseg->addr = cpu_to_be64(wi->umr.mtt_addr);
 433 }
 434
 435 static int mlx5e_rq_alloc_mpwqe_info(struct mlx5e_rq *rq,
 436                                      struct mlx5e_channel *c)
 437 {
 438         int wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 439         int mtt_sz = mlx5e_get_wqe_mtt_sz();
 440         int mtt_alloc = mtt_sz + MLX5_UMR_ALIGN - 1;
 441         int i;
 442
 443         rq->mpwqe.info = kzalloc_node(wq_sz * sizeof(*rq->mpwqe.info),
 444                                       GFP_KERNEL, cpu_to_node(c->cpu));
 445         if (!rq->mpwqe.info)
 446                 goto err_out;
 447
 448         /* We allocate more than mtt_sz as we will align the pointer */
 449         rq->mpwqe.mtt_no_align = kzalloc_node(mtt_alloc * wq_sz, GFP_KERNEL,
 450                                         cpu_to_node(c->cpu));
 451         if (unlikely(!rq->mpwqe.mtt_no_align))
 452                 goto err_free_wqe_info;
 453
 454         for (i = 0; i < wq_sz; i++) {
 455                 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[i];
 456
 457                 wi->umr.mtt = PTR_ALIGN(rq->mpwqe.mtt_no_align + i * mtt_alloc,
 458                                         MLX5_UMR_ALIGN);
 459                 wi->umr.mtt_addr = dma_map_single(c->pdev, wi->umr.mtt, mtt_sz,
 460                                                   PCI_DMA_TODEVICE);
 461                 if (unlikely(dma_mapping_error(c->pdev, wi->umr.mtt_addr)))
 462                         goto err_unmap_mtts;
 463
 464                 mlx5e_build_umr_wqe(rq, &c->icosq, &wi->umr.wqe, i);
 465         }
 466
 467         return 0;
 468
 469 err_unmap_mtts:
 470         while (--i >= 0) {
 471                 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[i];
 472
 473                 dma_unmap_single(c->pdev, wi->umr.mtt_addr, mtt_sz,
 474                                  PCI_DMA_TODEVICE);
 475         }
 476         kfree(rq->mpwqe.mtt_no_align);
 477 err_free_wqe_info:
 478         kfree(rq->mpwqe.info);
 479
 480 err_out:
 481         return -ENOMEM;
 482 }
 483
 484 static void mlx5e_rq_free_mpwqe_info(struct mlx5e_rq *rq)
 485 {
 486         int wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 487         int mtt_sz = mlx5e_get_wqe_mtt_sz();
 488         int i;
 489
 490         for (i = 0; i < wq_sz; i++) {
 491                 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[i];
 492
 493                 dma_unmap_single(rq->pdev, wi->umr.mtt_addr, mtt_sz,
 494                                  PCI_DMA_TODEVICE);
 495         }
 496         kfree(rq->mpwqe.mtt_no_align);
 497         kfree(rq->mpwqe.info);
 498 }
 499
 500 static int mlx5e_create_umr_mkey(struct mlx5_core_dev *mdev,
 501                                  u64 npages, u8 page_shift,
 502                                  struct mlx5_core_mkey *umr_mkey)
 503 {
 504         int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
 505         void *mkc;
 506         u32 *in;
 507         int err;
 508
 509         if (!MLX5E_VALID_NUM_MTTS(npages))
 510                 return -EINVAL;
 511
 512         in = kvzalloc(inlen, GFP_KERNEL);
 513         if (!in)
 514                 return -ENOMEM;
 515
 516         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 517
 518         MLX5_SET(mkc, mkc, free, 1);
 519         MLX5_SET(mkc, mkc, umr_en, 1);
 520         MLX5_SET(mkc, mkc, lw, 1);
 521         MLX5_SET(mkc, mkc, lr, 1);
 522         MLX5_SET(mkc, mkc, access_mode, MLX5_MKC_ACCESS_MODE_MTT);
 523
 524         MLX5_SET(mkc, mkc, qpn, 0xffffff);
 525         MLX5_SET(mkc, mkc, pd, mdev->mlx5e_res.pdn);
 526         MLX5_SET64(mkc, mkc, len, npages << page_shift);
 527         MLX5_SET(mkc, mkc, translations_octword_size,
 528                  MLX5_MTT_OCTW(npages));
 529         MLX5_SET(mkc, mkc, log_page_size, page_shift);
 530
 531         err = mlx5_core_create_mkey(mdev, umr_mkey, in, inlen);
 532
 533         kvfree(in);
 534         return err;
 535 }
 536
 537 static int mlx5e_create_rq_umr_mkey(struct mlx5_core_dev *mdev, struct mlx5e_rq *rq)
 538 {
 539         u64 num_mtts = MLX5E_REQUIRED_MTTS(mlx5_wq_ll_get_size(&rq->wq));
 540
 541         return mlx5e_create_umr_mkey(mdev, num_mtts, PAGE_SHIFT, &rq->umr_mkey);
 542 }
 543
 544 static int mlx5e_alloc_rq(struct mlx5e_channel *c,
 545                           struct mlx5e_params *params,
 546                           struct mlx5e_rq_param *rqp,
 547                           struct mlx5e_rq *rq)
 548 {
 549         struct mlx5_core_dev *mdev = c->mdev;
 550         void *rqc = rqp->rqc;
 551         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 552         u32 byte_count;
 553         int npages;
 554         int wq_sz;
 555         int err;
 556         int i;
 557
 558         rqp->wq.db_numa_node = cpu_to_node(c->cpu);
 559
 560         err = mlx5_wq_ll_create(mdev, &rqp->wq, rqc_wq, &rq->wq,
 561                                 &rq->wq_ctrl);
 562         if (err)
 563                 return err;
 564
 565         rq->wq.db = &rq->wq.db[MLX5_RCV_DBR];
 566
 567         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 568
 569         rq->wq_type = params->rq_wq_type;
 570         rq->pdev    = c->pdev;
 571         rq->netdev  = c->netdev;
 572         rq->tstamp  = c->tstamp;
 573         rq->clock   = &mdev->clock;
 574         rq->channel = c;
 575         rq->ix      = c->ix;
 576         rq->mdev    = mdev;
 577
 578         rq->xdp_prog = params->xdp_prog ? bpf_prog_inc(params->xdp_prog) : NULL;
 579         if (IS_ERR(rq->xdp_prog)) {
 580                 err = PTR_ERR(rq->xdp_prog);
 581                 rq->xdp_prog = NULL;
 582                 goto err_rq_wq_destroy;
 583         }
 584
 585         err = xdp_rxq_info_reg(&rq->xdp_rxq, rq->netdev, rq->ix);
 586         if (err < 0)
 587                 goto err_rq_wq_destroy;
 588
 589         rq->buff.map_dir = rq->xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
 590         rq->buff.headroom = params->rq_headroom;
 591
 592         switch (rq->wq_type) {
 593         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 594
 595                 rq->post_wqes = mlx5e_post_rx_mpwqes;
 596                 rq->dealloc_wqe = mlx5e_dealloc_rx_mpwqe;
 597
 598                 rq->handle_rx_cqe = c->priv->profile->rx_handlers.handle_rx_cqe_mpwqe;
 599 #ifdef CONFIG_MLX5_EN_IPSEC
 600                 if (MLX5_IPSEC_DEV(mdev)) {
 601                         err = -EINVAL;
 602                         netdev_err(c->netdev, "MPWQE RQ with IPSec offload not supported\n");
 603                         goto err_rq_wq_destroy;
 604                 }
 605 #endif
 606                 if (!rq->handle_rx_cqe) {
 607                         err = -EINVAL;
 608                         netdev_err(c->netdev, "RX handler of MPWQE RQ is not set, err %d\n", err);
 609                         goto err_rq_wq_destroy;
 610                 }
 611
 612                 rq->mpwqe.log_stride_sz = params->mpwqe_log_stride_sz;
 613                 rq->mpwqe.num_strides = BIT(params->mpwqe_log_num_strides);
 614
 615                 byte_count = rq->mpwqe.num_strides << rq->mpwqe.log_stride_sz;
 616
 617                 err = mlx5e_create_rq_umr_mkey(mdev, rq);
 618                 if (err)
 619                         goto err_rq_wq_destroy;
 620                 rq->mkey_be = cpu_to_be32(rq->umr_mkey.key);
 621
 622                 err = mlx5e_rq_alloc_mpwqe_info(rq, c);
 623                 if (err)
 624                         goto err_destroy_umr_mkey;
 625                 break;
 626         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 627                 rq->wqe.frag_info =
 628                         kzalloc_node(wq_sz * sizeof(*rq->wqe.frag_info),
 629                                      GFP_KERNEL, cpu_to_node(c->cpu));
 630                 if (!rq->wqe.frag_info) {
 631                         err = -ENOMEM;
 632                         goto err_rq_wq_destroy;
 633                 }
 634                 rq->post_wqes = mlx5e_post_rx_wqes;
 635                 rq->dealloc_wqe = mlx5e_dealloc_rx_wqe;
 636
 637 #ifdef CONFIG_MLX5_EN_IPSEC
 638                 if (c->priv->ipsec)
 639                         rq->handle_rx_cqe = mlx5e_ipsec_handle_rx_cqe;
 640                 else
 641 #endif
 642                         rq->handle_rx_cqe = c->priv->profile->rx_handlers.handle_rx_cqe;
 643                 if (!rq->handle_rx_cqe) {
 644                         kfree(rq->wqe.frag_info);
 645                         err = -EINVAL;
 646                         netdev_err(c->netdev, "RX handler of RQ is not set, err %d\n", err);
 647                         goto err_rq_wq_destroy;
 648                 }
 649
 650                 byte_count = params->lro_en  ?
 651                                 params->lro_wqe_sz :
 652                                 MLX5E_SW2HW_MTU(c->priv, c->netdev->mtu);
 653 #ifdef CONFIG_MLX5_EN_IPSEC
 654                 if (MLX5_IPSEC_DEV(mdev))
 655                         byte_count += MLX5E_METADATA_ETHER_LEN;
 656 #endif
 657                 rq->wqe.page_reuse = !params->xdp_prog && !params->lro_en;
 658
 659                 /* calc the required page order */
 660                 rq->wqe.frag_sz = MLX5_SKB_FRAG_SZ(rq->buff.headroom + byte_count);
 661                 npages = DIV_ROUND_UP(rq->wqe.frag_sz, PAGE_SIZE);
 662                 rq->buff.page_order = order_base_2(npages);
 663
 664                 byte_count |= MLX5_HW_START_PADDING;
 665                 rq->mkey_be = c->mkey_be;
 666         }
 667
 668         for (i = 0; i < wq_sz; i++) {
 669                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, i);
 670
 671                 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 672                         u64 dma_offset = (u64)mlx5e_get_wqe_mtt_offset(rq, i) << PAGE_SHIFT;
 673
 674                         wqe->data.addr = cpu_to_be64(dma_offset);
 675                 }
 676
 677                 wqe->data.byte_count = cpu_to_be32(byte_count);
 678                 wqe->data.lkey = rq->mkey_be;
 679         }
 680
 681         INIT_WORK(&rq->dim.work, mlx5e_rx_dim_work);
 682
 683         switch (params->rx_cq_moderation.cq_period_mode) {
 684         case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
 685                 rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE;
 686                 break;
 687         case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
 688         default:
 689                 rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
 690         }
 691
 692         rq->page_cache.head = 0;
 693         rq->page_cache.tail = 0;
 694
 695         return 0;
 696
 697 err_destroy_umr_mkey:
 698         mlx5_core_destroy_mkey(mdev, &rq->umr_mkey);
 699
 700 err_rq_wq_destroy:
 701         if (rq->xdp_prog)
 702                 bpf_prog_put(rq->xdp_prog);
 703         xdp_rxq_info_unreg(&rq->xdp_rxq);
 704         mlx5_wq_destroy(&rq->wq_ctrl);
 705
 706         return err;
 707 }
 708
 709 static void mlx5e_free_rq(struct mlx5e_rq *rq)
 710 {
 711         int i;
 712
 713         if (rq->xdp_prog)
 714                 bpf_prog_put(rq->xdp_prog);
 715
 716         xdp_rxq_info_unreg(&rq->xdp_rxq);
 717
 718         switch (rq->wq_type) {
 719         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 720                 mlx5e_rq_free_mpwqe_info(rq);
 721                 mlx5_core_destroy_mkey(rq->mdev, &rq->umr_mkey);
 722                 break;
 723         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 724                 kfree(rq->wqe.frag_info);
 725         }
 726
 727         for (i = rq->page_cache.head; i != rq->page_cache.tail;
 728              i = (i + 1) & (MLX5E_CACHE_SIZE - 1)) {
 729                 struct mlx5e_dma_info *dma_info = &rq->page_cache.page_cache[i];
 730
 731                 mlx5e_page_release(rq, dma_info, false);
 732         }
 733         mlx5_wq_destroy(&rq->wq_ctrl);
 734 }
 735
 736 static int mlx5e_create_rq(struct mlx5e_rq *rq,
 737                            struct mlx5e_rq_param *param)
 738 {
 739         struct mlx5_core_dev *mdev = rq->mdev;
 740
 741         void *in;
 742         void *rqc;
 743         void *wq;
 744         int inlen;
 745         int err;
 746
 747         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 748                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 749         in = kvzalloc(inlen, GFP_KERNEL);
 750         if (!in)
 751                 return -ENOMEM;
 752
 753         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 754         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 755
 756         memcpy(rqc, param->rqc, sizeof(param->rqc));
 757
 758         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 759         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 760         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 761                                                 MLX5_ADAPTER_PAGE_SHIFT);
 762         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 763
 764         mlx5_fill_page_array(&rq->wq_ctrl.buf,
 765                              (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 766
 767         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 768
 769         kvfree(in);
 770
 771         return err;
 772 }
 773
 774 static int mlx5e_modify_rq_state(struct mlx5e_rq *rq, int curr_state,
 775                                  int next_state)
 776 {
 777         struct mlx5e_channel *c = rq->channel;
 778         struct mlx5_core_dev *mdev = c->mdev;
 779
 780         void *in;
 781         void *rqc;
 782         int inlen;
 783         int err;
 784
 785         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 786         in = kvzalloc(inlen, GFP_KERNEL);
 787         if (!in)
 788                 return -ENOMEM;
 789
 790         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 791
 792         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 793         MLX5_SET(rqc, rqc, state, next_state);
 794
 795         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 796
 797         kvfree(in);
 798
 799         return err;
 800 }
 801
 802 static int mlx5e_modify_rq_scatter_fcs(struct mlx5e_rq *rq, bool enable)
 803 {
 804         struct mlx5e_channel *c = rq->channel;
 805         struct mlx5e_priv *priv = c->priv;
 806         struct mlx5_core_dev *mdev = priv->mdev;
 807
 808         void *in;
 809         void *rqc;
 810         int inlen;
 811         int err;
 812
 813         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 814         in = kvzalloc(inlen, GFP_KERNEL);
 815         if (!in)
 816                 return -ENOMEM;
 817
 818         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 819
 820         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 821         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 822                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_SCATTER_FCS);
 823         MLX5_SET(rqc, rqc, scatter_fcs, enable);
 824         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 825
 826         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 827
 828         kvfree(in);
 829
 830         return err;
 831 }
 832
 833 static int mlx5e_modify_rq_vsd(struct mlx5e_rq *rq, bool vsd)
 834 {
 835         struct mlx5e_channel *c = rq->channel;
 836         struct mlx5_core_dev *mdev = c->mdev;
 837         void *in;
 838         void *rqc;
 839         int inlen;
 840         int err;
 841
 842         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 843         in = kvzalloc(inlen, GFP_KERNEL);
 844         if (!in)
 845                 return -ENOMEM;
 846
 847         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 848
 849         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 850         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 851                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
 852         MLX5_SET(rqc, rqc, vsd, vsd);
 853         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 854
 855         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 856
 857         kvfree(in);
 858
 859         return err;
 860 }
 861
 862 static void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 863 {
 864         mlx5_core_destroy_rq(rq->mdev, rq->rqn);
 865 }
 866
 867 static int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq)
 868 {
 869         unsigned long exp_time = jiffies + msecs_to_jiffies(20000);
 870         struct mlx5e_channel *c = rq->channel;
 871
 872         struct mlx5_wq_ll *wq = &rq->wq;
 873         u16 min_wqes = mlx5_min_rx_wqes(rq->wq_type, mlx5_wq_ll_get_size(wq));
 874
 875         while (time_before(jiffies, exp_time)) {
 876                 if (wq->cur_sz >= min_wqes)
 877                         return 0;
 878
 879                 msleep(20);
 880         }
 881
 882         netdev_warn(c->netdev, "Failed to get min RX wqes on RQN[0x%x] wq cur_sz(%d) min_rx_wqes(%d)\n",
 883                     rq->rqn, wq->cur_sz, min_wqes);
 884         return -ETIMEDOUT;
 885 }
 886
 887 static void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 888 {
 889         struct mlx5_wq_ll *wq = &rq->wq;
 890         struct mlx5e_rx_wqe *wqe;
 891         __be16 wqe_ix_be;
 892         u16 wqe_ix;
 893
 894         /* UMR WQE (if in progress) is always at wq->head */
 895         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ &&
 896             rq->mpwqe.umr_in_progress)
 897                 mlx5e_free_rx_mpwqe(rq, &rq->mpwqe.info[wq->head]);
 898
 899         while (!mlx5_wq_ll_is_empty(wq)) {
 900                 wqe_ix_be = *wq->tail_next;
 901                 wqe_ix    = be16_to_cpu(wqe_ix_be);
 902                 wqe       = mlx5_wq_ll_get_wqe(&rq->wq, wqe_ix);
 903                 rq->dealloc_wqe(rq, wqe_ix);
 904                 mlx5_wq_ll_pop(&rq->wq, wqe_ix_be,
 905                                &wqe->next.next_wqe_index);
 906         }
 907
 908         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST && rq->wqe.page_reuse) {
 909                 /* Clean outstanding pages on handled WQEs that decided to do page-reuse,
 910                  * but yet to be re-posted.
 911                  */
 912                 int wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 913
 914                 for (wqe_ix = 0; wqe_ix < wq_sz; wqe_ix++)
 915                         rq->dealloc_wqe(rq, wqe_ix);
 916         }
 917 }
 918
 919 static int mlx5e_open_rq(struct mlx5e_channel *c,
 920                          struct mlx5e_params *params,
 921                          struct mlx5e_rq_param *param,
 922                          struct mlx5e_rq *rq)
 923 {
 924         int err;
 925
 926         err = mlx5e_alloc_rq(c, params, param, rq);
 927         if (err)
 928                 return err;
 929
 930         err = mlx5e_create_rq(rq, param);
 931         if (err)
 932                 goto err_free_rq;
 933
 934         err = mlx5e_modify_rq_state(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 935         if (err)
 936                 goto err_destroy_rq;
 937
 938         if (params->rx_dim_enabled)
 939                 c->rq.state |= BIT(MLX5E_RQ_STATE_AM);
 940
 941         return 0;
 942
 943 err_destroy_rq:
 944         mlx5e_destroy_rq(rq);
 945 err_free_rq:
 946         mlx5e_free_rq(rq);
 947
 948         return err;
 949 }
 950
 951 static void mlx5e_activate_rq(struct mlx5e_rq *rq)
 952 {
 953         struct mlx5e_icosq *sq = &rq->channel->icosq;
 954         u16 pi = sq->pc & sq->wq.sz_m1;
 955         struct mlx5e_tx_wqe *nopwqe;
 956
 957         set_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 958         sq->db.ico_wqe[pi].opcode     = MLX5_OPCODE_NOP;
 959         nopwqe = mlx5e_post_nop(&sq->wq, sq->sqn, &sq->pc);
 960         mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &nopwqe->ctrl);
 961 }
 962
 963 static void mlx5e_deactivate_rq(struct mlx5e_rq *rq)
 964 {
 965         clear_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 966         napi_synchronize(&rq->channel->napi); /* prevent mlx5e_post_rx_wqes */
 967 }
 968
 969 static void mlx5e_close_rq(struct mlx5e_rq *rq)
 970 {
 971         cancel_work_sync(&rq->dim.work);
 972         mlx5e_destroy_rq(rq);
 973         mlx5e_free_rx_descs(rq);
 974         mlx5e_free_rq(rq);
 975 }
 976
 977 static void mlx5e_free_xdpsq_db(struct mlx5e_xdpsq *sq)
 978 {
 979         kfree(sq->db.di);
 980 }
 981
 982 static int mlx5e_alloc_xdpsq_db(struct mlx5e_xdpsq *sq, int numa)
 983 {
 984         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 985
 986         sq->db.di = kzalloc_node(sizeof(*sq->db.di) * wq_sz,
 987                                      GFP_KERNEL, numa);
 988         if (!sq->db.di) {
 989                 mlx5e_free_xdpsq_db(sq);
 990                 return -ENOMEM;
 991         }
 992
 993         return 0;
 994 }
 995
 996 static int mlx5e_alloc_xdpsq(struct mlx5e_channel *c,
 997                              struct mlx5e_params *params,
 998                              struct mlx5e_sq_param *param,
 999                              struct mlx5e_xdpsq *sq)
1000 {
1001         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1002         struct mlx5_core_dev *mdev = c->mdev;
1003         int err;
1004
1005         sq->pdev      = c->pdev;
1006         sq->mkey_be   = c->mkey_be;
1007         sq->channel   = c;
1008         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1009         sq->min_inline_mode = params->tx_min_inline_mode;
1010
1011         param->wq.db_numa_node = cpu_to_node(c->cpu);
1012         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq, &sq->wq_ctrl);
1013         if (err)
1014                 return err;
1015         sq->wq.db = &sq->wq.db[MLX5_SND_DBR];
1016
1017         err = mlx5e_alloc_xdpsq_db(sq, cpu_to_node(c->cpu));
1018         if (err)
1019                 goto err_sq_wq_destroy;
1020
1021         return 0;
1022
1023 err_sq_wq_destroy:
1024         mlx5_wq_destroy(&sq->wq_ctrl);
1025
1026         return err;
1027 }
1028
1029 static void mlx5e_free_xdpsq(struct mlx5e_xdpsq *sq)
1030 {
1031         mlx5e_free_xdpsq_db(sq);
1032         mlx5_wq_destroy(&sq->wq_ctrl);
1033 }
1034
1035 static void mlx5e_free_icosq_db(struct mlx5e_icosq *sq)
1036 {
1037         kfree(sq->db.ico_wqe);
1038 }
1039
1040 static int mlx5e_alloc_icosq_db(struct mlx5e_icosq *sq, int numa)
1041 {
1042         u8 wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1043
1044         sq->db.ico_wqe = kzalloc_node(sizeof(*sq->db.ico_wqe) * wq_sz,
1045                                       GFP_KERNEL, numa);
1046         if (!sq->db.ico_wqe)
1047                 return -ENOMEM;
1048
1049         return 0;
1050 }
1051
1052 static int mlx5e_alloc_icosq(struct mlx5e_channel *c,
1053                              struct mlx5e_sq_param *param,
1054                              struct mlx5e_icosq *sq)
1055 {
1056         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1057         struct mlx5_core_dev *mdev = c->mdev;
1058         int err;
1059
1060         sq->mkey_be   = c->mkey_be;
1061         sq->channel   = c;
1062         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1063
1064         param->wq.db_numa_node = cpu_to_node(c->cpu);
1065         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq, &sq->wq_ctrl);
1066         if (err)
1067                 return err;
1068         sq->wq.db = &sq->wq.db[MLX5_SND_DBR];
1069
1070         err = mlx5e_alloc_icosq_db(sq, cpu_to_node(c->cpu));
1071         if (err)
1072                 goto err_sq_wq_destroy;
1073
1074         sq->edge = (sq->wq.sz_m1 + 1) - MLX5E_ICOSQ_MAX_WQEBBS;
1075
1076         return 0;
1077
1078 err_sq_wq_destroy:
1079         mlx5_wq_destroy(&sq->wq_ctrl);
1080
1081         return err;
1082 }
1083
1084 static void mlx5e_free_icosq(struct mlx5e_icosq *sq)
1085 {
1086         mlx5e_free_icosq_db(sq);
1087         mlx5_wq_destroy(&sq->wq_ctrl);
1088 }
1089
1090 static void mlx5e_free_txqsq_db(struct mlx5e_txqsq *sq)
1091 {
1092         kfree(sq->db.wqe_info);
1093         kfree(sq->db.dma_fifo);
1094 }
1095
1096 static int mlx5e_alloc_txqsq_db(struct mlx5e_txqsq *sq, int numa)
1097 {
1098         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1099         int df_sz = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
1100
1101         sq->db.dma_fifo = kzalloc_node(df_sz * sizeof(*sq->db.dma_fifo),
1102                                            GFP_KERNEL, numa);
1103         sq->db.wqe_info = kzalloc_node(wq_sz * sizeof(*sq->db.wqe_info),
1104                                            GFP_KERNEL, numa);
1105         if (!sq->db.dma_fifo || !sq->db.wqe_info) {
1106                 mlx5e_free_txqsq_db(sq);
1107                 return -ENOMEM;
1108         }
1109
1110         sq->dma_fifo_mask = df_sz - 1;
1111
1112         return 0;
1113 }
1114
1115 static int mlx5e_alloc_txqsq(struct mlx5e_channel *c,
1116                              int txq_ix,
1117                              struct mlx5e_params *params,
1118                              struct mlx5e_sq_param *param,
1119                              struct mlx5e_txqsq *sq)
1120 {
1121         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1122         struct mlx5_core_dev *mdev = c->mdev;
1123         int err;
1124
1125         sq->pdev      = c->pdev;
1126         sq->tstamp    = c->tstamp;
1127         sq->clock     = &mdev->clock;
1128         sq->mkey_be   = c->mkey_be;
1129         sq->channel   = c;
1130         sq->txq_ix    = txq_ix;
1131         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1132         sq->max_inline      = params->tx_max_inline;
1133         sq->min_inline_mode = params->tx_min_inline_mode;
1134         if (MLX5_IPSEC_DEV(c->priv->mdev))
1135                 set_bit(MLX5E_SQ_STATE_IPSEC, &sq->state);
1136
1137         param->wq.db_numa_node = cpu_to_node(c->cpu);
1138         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq, &sq->wq_ctrl);
1139         if (err)
1140                 return err;
1141         sq->wq.db    = &sq->wq.db[MLX5_SND_DBR];
1142
1143         err = mlx5e_alloc_txqsq_db(sq, cpu_to_node(c->cpu));
1144         if (err)
1145                 goto err_sq_wq_destroy;
1146
1147         sq->edge = (sq->wq.sz_m1 + 1) - MLX5_SEND_WQE_MAX_WQEBBS;
1148
1149         return 0;
1150
1151 err_sq_wq_destroy:
1152         mlx5_wq_destroy(&sq->wq_ctrl);
1153
1154         return err;
1155 }
1156
1157 static void mlx5e_free_txqsq(struct mlx5e_txqsq *sq)
1158 {
1159         mlx5e_free_txqsq_db(sq);
1160         mlx5_wq_destroy(&sq->wq_ctrl);
1161 }
1162
1163 struct mlx5e_create_sq_param {
1164         struct mlx5_wq_ctrl        *wq_ctrl;
1165         u32                         cqn;
1166         u32                         tisn;
1167         u8                          tis_lst_sz;
1168         u8                          min_inline_mode;
1169 };
1170
1171 static int mlx5e_create_sq(struct mlx5_core_dev *mdev,
1172                            struct mlx5e_sq_param *param,
1173                            struct mlx5e_create_sq_param *csp,
1174                            u32 *sqn)
1175 {
1176         void *in;
1177         void *sqc;
1178         void *wq;
1179         int inlen;
1180         int err;
1181
1182         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1183                 sizeof(u64) * csp->wq_ctrl->buf.npages;
1184         in = kvzalloc(inlen, GFP_KERNEL);
1185         if (!in)
1186                 return -ENOMEM;
1187
1188         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1189         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1190
1191         memcpy(sqc, param->sqc, sizeof(param->sqc));
1192         MLX5_SET(sqc,  sqc, tis_lst_sz, csp->tis_lst_sz);
1193         MLX5_SET(sqc,  sqc, tis_num_0, csp->tisn);
1194         MLX5_SET(sqc,  sqc, cqn, csp->cqn);
1195
1196         if (MLX5_CAP_ETH(mdev, wqe_inline_mode) == MLX5_CAP_INLINE_MODE_VPORT_CONTEXT)
1197                 MLX5_SET(sqc,  sqc, min_wqe_inline_mode, csp->min_inline_mode);
1198
1199         MLX5_SET(sqc,  sqc, state, MLX5_SQC_STATE_RST);
1200
1201         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
1202         MLX5_SET(wq,   wq, uar_page,      mdev->mlx5e_res.bfreg.index);
1203         MLX5_SET(wq,   wq, log_wq_pg_sz,  csp->wq_ctrl->buf.page_shift -
1204                                           MLX5_ADAPTER_PAGE_SHIFT);
1205         MLX5_SET64(wq, wq, dbr_addr,      csp->wq_ctrl->db.dma);
1206
1207         mlx5_fill_page_array(&csp->wq_ctrl->buf, (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
1208
1209         err = mlx5_core_create_sq(mdev, in, inlen, sqn);
1210
1211         kvfree(in);
1212
1213         return err;
1214 }
1215
1216 struct mlx5e_modify_sq_param {
1217         int curr_state;
1218         int next_state;
1219         bool rl_update;
1220         int rl_index;
1221 };
1222
1223 static int mlx5e_modify_sq(struct mlx5_core_dev *mdev, u32 sqn,
1224                            struct mlx5e_modify_sq_param *p)
1225 {
1226         void *in;
1227         void *sqc;
1228         int inlen;
1229         int err;
1230
1231         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1232         in = kvzalloc(inlen, GFP_KERNEL);
1233         if (!in)
1234                 return -ENOMEM;
1235
1236         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1237
1238         MLX5_SET(modify_sq_in, in, sq_state, p->curr_state);
1239         MLX5_SET(sqc, sqc, state, p->next_state);
1240         if (p->rl_update && p->next_state == MLX5_SQC_STATE_RDY) {
1241                 MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
1242                 MLX5_SET(sqc,  sqc, packet_pacing_rate_limit_index, p->rl_index);
1243         }
1244
1245         err = mlx5_core_modify_sq(mdev, sqn, in, inlen);
1246
1247         kvfree(in);
1248
1249         return err;
1250 }
1251
1252 static void mlx5e_destroy_sq(struct mlx5_core_dev *mdev, u32 sqn)
1253 {
1254         mlx5_core_destroy_sq(mdev, sqn);
1255 }
1256
1257 static int mlx5e_create_sq_rdy(struct mlx5_core_dev *mdev,
1258                                struct mlx5e_sq_param *param,
1259                                struct mlx5e_create_sq_param *csp,
1260                                u32 *sqn)
1261 {
1262         struct mlx5e_modify_sq_param msp = {0};
1263         int err;
1264
1265         err = mlx5e_create_sq(mdev, param, csp, sqn);
1266         if (err)
1267                 return err;
1268
1269         msp.curr_state = MLX5_SQC_STATE_RST;
1270         msp.next_state = MLX5_SQC_STATE_RDY;
1271         err = mlx5e_modify_sq(mdev, *sqn, &msp);
1272         if (err)
1273                 mlx5e_destroy_sq(mdev, *sqn);
1274
1275         return err;
1276 }
1277
1278 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1279                                 struct mlx5e_txqsq *sq, u32 rate);
1280
1281 static int mlx5e_open_txqsq(struct mlx5e_channel *c,
1282                             u32 tisn,
1283                             int txq_ix,
1284                             struct mlx5e_params *params,
1285                             struct mlx5e_sq_param *param,
1286                             struct mlx5e_txqsq *sq)
1287 {
1288         struct mlx5e_create_sq_param csp = {};
1289         u32 tx_rate;
1290         int err;
1291
1292         err = mlx5e_alloc_txqsq(c, txq_ix, params, param, sq);
1293         if (err)
1294                 return err;
1295
1296         csp.tisn            = tisn;
1297         csp.tis_lst_sz      = 1;
1298         csp.cqn             = sq->cq.mcq.cqn;
1299         csp.wq_ctrl         = &sq->wq_ctrl;
1300         csp.min_inline_mode = sq->min_inline_mode;
1301         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1302         if (err)
1303                 goto err_free_txqsq;
1304
1305         tx_rate = c->priv->tx_rates[sq->txq_ix];
1306         if (tx_rate)
1307                 mlx5e_set_sq_maxrate(c->netdev, sq, tx_rate);
1308
1309         return 0;
1310
1311 err_free_txqsq:
1312         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1313         mlx5e_free_txqsq(sq);
1314
1315         return err;
1316 }
1317
1318 static void mlx5e_activate_txqsq(struct mlx5e_txqsq *sq)
1319 {
1320         sq->txq = netdev_get_tx_queue(sq->channel->netdev, sq->txq_ix);
1321         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1322         netdev_tx_reset_queue(sq->txq);
1323         netif_tx_start_queue(sq->txq);
1324 }
1325
1326 static inline void netif_tx_disable_queue(struct netdev_queue *txq)
1327 {
1328         __netif_tx_lock_bh(txq);
1329         netif_tx_stop_queue(txq);
1330         __netif_tx_unlock_bh(txq);
1331 }
1332
1333 static void mlx5e_deactivate_txqsq(struct mlx5e_txqsq *sq)
1334 {
1335         struct mlx5e_channel *c = sq->channel;
1336
1337         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1338         /* prevent netif_tx_wake_queue */
1339         napi_synchronize(&c->napi);
1340
1341         netif_tx_disable_queue(sq->txq);
1342
1343         /* last doorbell out, godspeed .. */
1344         if (mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc, 1)) {
1345                 struct mlx5e_tx_wqe *nop;
1346
1347                 sq->db.wqe_info[(sq->pc & sq->wq.sz_m1)].skb = NULL;
1348                 nop = mlx5e_post_nop(&sq->wq, sq->sqn, &sq->pc);
1349                 mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &nop->ctrl);
1350         }
1351 }
1352
1353 static void mlx5e_close_txqsq(struct mlx5e_txqsq *sq)
1354 {
1355         struct mlx5e_channel *c = sq->channel;
1356         struct mlx5_core_dev *mdev = c->mdev;
1357
1358         mlx5e_destroy_sq(mdev, sq->sqn);
1359         if (sq->rate_limit)
1360                 mlx5_rl_remove_rate(mdev, sq->rate_limit);
1361         mlx5e_free_txqsq_descs(sq);
1362         mlx5e_free_txqsq(sq);
1363 }
1364
1365 static int mlx5e_open_icosq(struct mlx5e_channel *c,
1366                             struct mlx5e_params *params,
1367                             struct mlx5e_sq_param *param,
1368                             struct mlx5e_icosq *sq)
1369 {
1370         struct mlx5e_create_sq_param csp = {};
1371         int err;
1372
1373         err = mlx5e_alloc_icosq(c, param, sq);
1374         if (err)
1375                 return err;
1376
1377         csp.cqn             = sq->cq.mcq.cqn;
1378         csp.wq_ctrl         = &sq->wq_ctrl;
1379         csp.min_inline_mode = params->tx_min_inline_mode;
1380         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1381         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1382         if (err)
1383                 goto err_free_icosq;
1384
1385         return 0;
1386
1387 err_free_icosq:
1388         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1389         mlx5e_free_icosq(sq);
1390
1391         return err;
1392 }
1393
1394 static void mlx5e_close_icosq(struct mlx5e_icosq *sq)
1395 {
1396         struct mlx5e_channel *c = sq->channel;
1397
1398         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1399         napi_synchronize(&c->napi);
1400
1401         mlx5e_destroy_sq(c->mdev, sq->sqn);
1402         mlx5e_free_icosq(sq);
1403 }
1404
1405 static int mlx5e_open_xdpsq(struct mlx5e_channel *c,
1406                             struct mlx5e_params *params,
1407                             struct mlx5e_sq_param *param,
1408                             struct mlx5e_xdpsq *sq)
1409 {
1410         unsigned int ds_cnt = MLX5E_XDP_TX_DS_COUNT;
1411         struct mlx5e_create_sq_param csp = {};
1412         unsigned int inline_hdr_sz = 0;
1413         int err;
1414         int i;
1415
1416         err = mlx5e_alloc_xdpsq(c, params, param, sq);
1417         if (err)
1418                 return err;
1419
1420         csp.tis_lst_sz      = 1;
1421         csp.tisn            = c->priv->tisn[0]; /* tc = 0 */
1422         csp.cqn             = sq->cq.mcq.cqn;
1423         csp.wq_ctrl         = &sq->wq_ctrl;
1424         csp.min_inline_mode = sq->min_inline_mode;
1425         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1426         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1427         if (err)
1428                 goto err_free_xdpsq;
1429
1430         if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
1431                 inline_hdr_sz = MLX5E_XDP_MIN_INLINE;
1432                 ds_cnt++;
1433         }
1434
1435         /* Pre initialize fixed WQE fields */
1436         for (i = 0; i < mlx5_wq_cyc_get_size(&sq->wq); i++) {
1437                 struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(&sq->wq, i);
1438                 struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
1439                 struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
1440                 struct mlx5_wqe_data_seg *dseg;
1441
1442                 cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
1443                 eseg->inline_hdr.sz = cpu_to_be16(inline_hdr_sz);
1444
1445                 dseg = (struct mlx5_wqe_data_seg *)cseg + (ds_cnt - 1);
1446                 dseg->lkey = sq->mkey_be;
1447         }
1448
1449         return 0;
1450
1451 err_free_xdpsq:
1452         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1453         mlx5e_free_xdpsq(sq);
1454
1455         return err;
1456 }
1457
1458 static void mlx5e_close_xdpsq(struct mlx5e_xdpsq *sq)
1459 {
1460         struct mlx5e_channel *c = sq->channel;
1461
1462         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1463         napi_synchronize(&c->napi);
1464
1465         mlx5e_destroy_sq(c->mdev, sq->sqn);
1466         mlx5e_free_xdpsq_descs(sq);
1467         mlx5e_free_xdpsq(sq);
1468 }
1469
1470 static int mlx5e_alloc_cq_common(struct mlx5_core_dev *mdev,
1471                                  struct mlx5e_cq_param *param,
1472                                  struct mlx5e_cq *cq)
1473 {
1474         struct mlx5_core_cq *mcq = &cq->mcq;
1475         int eqn_not_used;
1476         unsigned int irqn;
1477         int err;
1478         u32 i;
1479
1480         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1481                                &cq->wq_ctrl);
1482         if (err)
1483                 return err;
1484
1485         mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1486
1487         mcq->cqe_sz     = 64;
1488         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1489         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1490         *mcq->set_ci_db = 0;
1491         *mcq->arm_db    = 0;
1492         mcq->vector     = param->eq_ix;
1493         mcq->comp       = mlx5e_completion_event;
1494         mcq->event      = mlx5e_cq_error_event;
1495         mcq->irqn       = irqn;
1496
1497         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1498                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1499
1500                 cqe->op_own = 0xf1;
1501         }
1502
1503         cq->mdev = mdev;
1504
1505         return 0;
1506 }
1507
1508 static int mlx5e_alloc_cq(struct mlx5e_channel *c,
1509                           struct mlx5e_cq_param *param,
1510                           struct mlx5e_cq *cq)
1511 {
1512         struct mlx5_core_dev *mdev = c->priv->mdev;
1513         int err;
1514
1515         param->wq.buf_numa_node = cpu_to_node(c->cpu);
1516         param->wq.db_numa_node  = cpu_to_node(c->cpu);
1517         param->eq_ix   = c->ix;
1518
1519         err = mlx5e_alloc_cq_common(mdev, param, cq);
1520
1521         cq->napi    = &c->napi;
1522         cq->channel = c;
1523
1524         return err;
1525 }
1526
1527 static void mlx5e_free_cq(struct mlx5e_cq *cq)
1528 {
1529         mlx5_cqwq_destroy(&cq->wq_ctrl);
1530 }
1531
1532 static int mlx5e_create_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
1533 {
1534         struct mlx5_core_dev *mdev = cq->mdev;
1535         struct mlx5_core_cq *mcq = &cq->mcq;
1536
1537         void *in;
1538         void *cqc;
1539         int inlen;
1540         unsigned int irqn_not_used;
1541         int eqn;
1542         int err;
1543
1544         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1545                 sizeof(u64) * cq->wq_ctrl.frag_buf.npages;
1546         in = kvzalloc(inlen, GFP_KERNEL);
1547         if (!in)
1548                 return -ENOMEM;
1549
1550         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1551
1552         memcpy(cqc, param->cqc, sizeof(param->cqc));
1553
1554         mlx5_fill_page_frag_array(&cq->wq_ctrl.frag_buf,
1555                                   (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
1556
1557         mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
1558
1559         MLX5_SET(cqc,   cqc, cq_period_mode, param->cq_period_mode);
1560         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
1561         MLX5_SET(cqc,   cqc, uar_page,      mdev->priv.uar->index);
1562         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.frag_buf.page_shift -
1563                                             MLX5_ADAPTER_PAGE_SHIFT);
1564         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
1565
1566         err = mlx5_core_create_cq(mdev, mcq, in, inlen);
1567
1568         kvfree(in);
1569
1570         if (err)
1571                 return err;
1572
1573         mlx5e_cq_arm(cq);
1574
1575         return 0;
1576 }
1577
1578 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
1579 {
1580         mlx5_core_destroy_cq(cq->mdev, &cq->mcq);
1581 }
1582
1583 static int mlx5e_open_cq(struct mlx5e_channel *c,
1584                          struct net_dim_cq_moder moder,
1585                          struct mlx5e_cq_param *param,
1586                          struct mlx5e_cq *cq)
1587 {
1588         struct mlx5_core_dev *mdev = c->mdev;
1589         int err;
1590
1591         err = mlx5e_alloc_cq(c, param, cq);
1592         if (err)
1593                 return err;
1594
1595         err = mlx5e_create_cq(cq, param);
1596         if (err)
1597                 goto err_free_cq;
1598
1599         if (MLX5_CAP_GEN(mdev, cq_moderation))
1600                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq, moder.usec, moder.pkts);
1601         return 0;
1602
1603 err_free_cq:
1604         mlx5e_free_cq(cq);
1605
1606         return err;
1607 }
1608
1609 static void mlx5e_close_cq(struct mlx5e_cq *cq)
1610 {
1611         mlx5e_destroy_cq(cq);
1612         mlx5e_free_cq(cq);
1613 }
1614
1615 static int mlx5e_get_cpu(struct mlx5e_priv *priv, int ix)
1616 {
1617         return cpumask_first(priv->mdev->priv.irq_info[ix].mask);
1618 }
1619
1620 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1621                              struct mlx5e_params *params,
1622                              struct mlx5e_channel_param *cparam)
1623 {
1624         int err;
1625         int tc;
1626
1627         for (tc = 0; tc < c->num_tc; tc++) {
1628                 err = mlx5e_open_cq(c, params->tx_cq_moderation,
1629                                     &cparam->tx_cq, &c->sq[tc].cq);
1630                 if (err)
1631                         goto err_close_tx_cqs;
1632         }
1633
1634         return 0;
1635
1636 err_close_tx_cqs:
1637         for (tc--; tc >= 0; tc--)
1638                 mlx5e_close_cq(&c->sq[tc].cq);
1639
1640         return err;
1641 }
1642
1643 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1644 {
1645         int tc;
1646
1647         for (tc = 0; tc < c->num_tc; tc++)
1648                 mlx5e_close_cq(&c->sq[tc].cq);
1649 }
1650
1651 static int mlx5e_open_sqs(struct mlx5e_channel *c,
1652                           struct mlx5e_params *params,
1653                           struct mlx5e_channel_param *cparam)
1654 {
1655         int err;
1656         int tc;
1657
1658         for (tc = 0; tc < params->num_tc; tc++) {
1659                 int txq_ix = c->ix + tc * params->num_channels;
1660
1661                 err = mlx5e_open_txqsq(c, c->priv->tisn[tc], txq_ix,
1662                                        params, &cparam->sq, &c->sq[tc]);
1663                 if (err)
1664                         goto err_close_sqs;
1665         }
1666
1667         return 0;
1668
1669 err_close_sqs:
1670         for (tc--; tc >= 0; tc--)
1671                 mlx5e_close_txqsq(&c->sq[tc]);
1672
1673         return err;
1674 }
1675
1676 static void mlx5e_close_sqs(struct mlx5e_channel *c)
1677 {
1678         int tc;
1679
1680         for (tc = 0; tc < c->num_tc; tc++)
1681                 mlx5e_close_txqsq(&c->sq[tc]);
1682 }
1683
1684 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1685                                 struct mlx5e_txqsq *sq, u32 rate)
1686 {
1687         struct mlx5e_priv *priv = netdev_priv(dev);
1688         struct mlx5_core_dev *mdev = priv->mdev;
1689         struct mlx5e_modify_sq_param msp = {0};
1690         u16 rl_index = 0;
1691         int err;
1692
1693         if (rate == sq->rate_limit)
1694                 /* nothing to do */
1695                 return 0;
1696
1697         if (sq->rate_limit)
1698                 /* remove current rl index to free space to next ones */
1699                 mlx5_rl_remove_rate(mdev, sq->rate_limit);
1700
1701         sq->rate_limit = 0;
1702
1703         if (rate) {
1704                 err = mlx5_rl_add_rate(mdev, rate, &rl_index);
1705                 if (err) {
1706                         netdev_err(dev, "Failed configuring rate %u: %d\n",
1707                                    rate, err);
1708                         return err;
1709                 }
1710         }
1711
1712         msp.curr_state = MLX5_SQC_STATE_RDY;
1713         msp.next_state = MLX5_SQC_STATE_RDY;
1714         msp.rl_index   = rl_index;
1715         msp.rl_update  = true;
1716         err = mlx5e_modify_sq(mdev, sq->sqn, &msp);
1717         if (err) {
1718                 netdev_err(dev, "Failed configuring rate %u: %d\n",
1719                            rate, err);
1720                 /* remove the rate from the table */
1721                 if (rate)
1722                         mlx5_rl_remove_rate(mdev, rate);
1723                 return err;
1724         }
1725
1726         sq->rate_limit = rate;
1727         return 0;
1728 }
1729
1730 static int mlx5e_set_tx_maxrate(struct net_device *dev, int index, u32 rate)
1731 {
1732         struct mlx5e_priv *priv = netdev_priv(dev);
1733         struct mlx5_core_dev *mdev = priv->mdev;
1734         struct mlx5e_txqsq *sq = priv->txq2sq[index];
1735         int err = 0;
1736
1737         if (!mlx5_rl_is_supported(mdev)) {
1738                 netdev_err(dev, "Rate limiting is not supported on this device\n");
1739                 return -EINVAL;
1740         }
1741
1742         /* rate is given in Mb/sec, HW config is in Kb/sec */
1743         rate = rate << 10;
1744
1745         /* Check whether rate in valid range, 0 is always valid */
1746         if (rate && !mlx5_rl_is_in_range(mdev, rate)) {
1747                 netdev_err(dev, "TX rate %u, is not in range\n", rate);
1748                 return -ERANGE;
1749         }
1750
1751         mutex_lock(&priv->state_lock);
1752         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
1753                 err = mlx5e_set_sq_maxrate(dev, sq, rate);
1754         if (!err)
1755                 priv->tx_rates[index] = rate;
1756         mutex_unlock(&priv->state_lock);
1757
1758         return err;
1759 }
1760
1761 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
1762                               struct mlx5e_params *params,
1763                               struct mlx5e_channel_param *cparam,
1764                               struct mlx5e_channel **cp)
1765 {
1766         struct net_dim_cq_moder icocq_moder = {0, 0};
1767         struct net_device *netdev = priv->netdev;
1768         int cpu = mlx5e_get_cpu(priv, ix);
1769         struct mlx5e_channel *c;
1770         unsigned int irq;
1771         int err;
1772         int eqn;
1773
1774         c = kzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
1775         if (!c)
1776                 return -ENOMEM;
1777
1778         c->priv     = priv;
1779         c->mdev     = priv->mdev;
1780         c->tstamp   = &priv->tstamp;
1781         c->ix       = ix;
1782         c->cpu      = cpu;
1783         c->pdev     = &priv->mdev->pdev->dev;
1784         c->netdev   = priv->netdev;
1785         c->mkey_be  = cpu_to_be32(priv->mdev->mlx5e_res.mkey.key);
1786         c->num_tc   = params->num_tc;
1787         c->xdp      = !!params->xdp_prog;
1788
1789         mlx5_vector2eqn(priv->mdev, ix, &eqn, &irq);
1790         c->irq_desc = irq_to_desc(irq);
1791
1792         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
1793
1794         err = mlx5e_open_cq(c, icocq_moder, &cparam->icosq_cq, &c->icosq.cq);
1795         if (err)
1796                 goto err_napi_del;
1797
1798         err = mlx5e_open_tx_cqs(c, params, cparam);
1799         if (err)
1800                 goto err_close_icosq_cq;
1801
1802         err = mlx5e_open_cq(c, params->rx_cq_moderation, &cparam->rx_cq, &c->rq.cq);
1803         if (err)
1804                 goto err_close_tx_cqs;
1805
1806         /* XDP SQ CQ params are same as normal TXQ sq CQ params */
1807         err = c->xdp ? mlx5e_open_cq(c, params->tx_cq_moderation,
1808                                      &cparam->tx_cq, &c->rq.xdpsq.cq) : 0;
1809         if (err)
1810                 goto err_close_rx_cq;
1811
1812         napi_enable(&c->napi);
1813
1814         err = mlx5e_open_icosq(c, params, &cparam->icosq, &c->icosq);
1815         if (err)
1816                 goto err_disable_napi;
1817
1818         err = mlx5e_open_sqs(c, params, cparam);
1819         if (err)
1820                 goto err_close_icosq;
1821
1822         err = c->xdp ? mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, &c->rq.xdpsq) : 0;
1823         if (err)
1824                 goto err_close_sqs;
1825
1826         err = mlx5e_open_rq(c, params, &cparam->rq, &c->rq);
1827         if (err)
1828                 goto err_close_xdp_sq;
1829
1830         *cp = c;
1831
1832         return 0;
1833 err_close_xdp_sq:
1834         if (c->xdp)
1835                 mlx5e_close_xdpsq(&c->rq.xdpsq);
1836
1837 err_close_sqs:
1838         mlx5e_close_sqs(c);
1839
1840 err_close_icosq:
1841         mlx5e_close_icosq(&c->icosq);
1842
1843 err_disable_napi:
1844         napi_disable(&c->napi);
1845         if (c->xdp)
1846                 mlx5e_close_cq(&c->rq.xdpsq.cq);
1847
1848 err_close_rx_cq:
1849         mlx5e_close_cq(&c->rq.cq);
1850
1851 err_close_tx_cqs:
1852         mlx5e_close_tx_cqs(c);
1853
1854 err_close_icosq_cq:
1855         mlx5e_close_cq(&c->icosq.cq);
1856
1857 err_napi_del:
1858         netif_napi_del(&c->napi);
1859         kfree(c);
1860
1861         return err;
1862 }
1863
1864 static void mlx5e_activate_channel(struct mlx5e_channel *c)
1865 {
1866         int tc;
1867
1868         for (tc = 0; tc < c->num_tc; tc++)
1869                 mlx5e_activate_txqsq(&c->sq[tc]);
1870         mlx5e_activate_rq(&c->rq);
1871         netif_set_xps_queue(c->netdev, get_cpu_mask(c->cpu), c->ix);
1872 }
1873
1874 static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
1875 {
1876         int tc;
1877
1878         mlx5e_deactivate_rq(&c->rq);
1879         for (tc = 0; tc < c->num_tc; tc++)
1880                 mlx5e_deactivate_txqsq(&c->sq[tc]);
1881 }
1882
1883 static void mlx5e_close_channel(struct mlx5e_channel *c)
1884 {
1885         mlx5e_close_rq(&c->rq);
1886         if (c->xdp)
1887                 mlx5e_close_xdpsq(&c->rq.xdpsq);
1888         mlx5e_close_sqs(c);
1889         mlx5e_close_icosq(&c->icosq);
1890         napi_disable(&c->napi);
1891         if (c->xdp)
1892                 mlx5e_close_cq(&c->rq.xdpsq.cq);
1893         mlx5e_close_cq(&c->rq.cq);
1894         mlx5e_close_tx_cqs(c);
1895         mlx5e_close_cq(&c->icosq.cq);
1896         netif_napi_del(&c->napi);
1897
1898         kfree(c);
1899 }
1900
1901 static void mlx5e_build_rq_param(struct mlx5e_priv *priv,
1902                                  struct mlx5e_params *params,
1903                                  struct mlx5e_rq_param *param)
1904 {
1905         void *rqc = param->rqc;
1906         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1907
1908         switch (params->rq_wq_type) {
1909         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1910                 MLX5_SET(wq, wq, log_wqe_num_of_strides, params->mpwqe_log_num_strides - 9);
1911                 MLX5_SET(wq, wq, log_wqe_stride_size, params->mpwqe_log_stride_sz - 6);
1912                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
1913                 break;
1914         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1915                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1916         }
1917
1918         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1919         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1920         MLX5_SET(wq, wq, log_wq_sz,        params->log_rq_size);
1921         MLX5_SET(wq, wq, pd,               priv->mdev->mlx5e_res.pdn);
1922         MLX5_SET(rqc, rqc, counter_set_id, priv->q_counter);
1923         MLX5_SET(rqc, rqc, vsd,            params->vlan_strip_disable);
1924         MLX5_SET(rqc, rqc, scatter_fcs,    params->scatter_fcs_en);
1925
1926         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1927         param->wq.linear = 1;
1928 }
1929
1930 static void mlx5e_build_drop_rq_param(struct mlx5e_rq_param *param)
1931 {
1932         void *rqc = param->rqc;
1933         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1934
1935         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1936         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1937 }
1938
1939 static void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
1940                                         struct mlx5e_sq_param *param)
1941 {
1942         void *sqc = param->sqc;
1943         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1944
1945         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1946         MLX5_SET(wq, wq, pd,            priv->mdev->mlx5e_res.pdn);
1947
1948         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1949 }
1950
1951 static void mlx5e_build_sq_param(struct mlx5e_priv *priv,
1952                                  struct mlx5e_params *params,
1953                                  struct mlx5e_sq_param *param)
1954 {
1955         void *sqc = param->sqc;
1956         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1957
1958         mlx5e_build_sq_param_common(priv, param);
1959         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
1960         MLX5_SET(sqc, sqc, allow_swp, !!MLX5_IPSEC_DEV(priv->mdev));
1961 }
1962
1963 static void mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
1964                                         struct mlx5e_cq_param *param)
1965 {
1966         void *cqc = param->cqc;
1967
1968         MLX5_SET(cqc, cqc, uar_page, priv->mdev->priv.uar->index);
1969 }
1970
1971 static void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
1972                                     struct mlx5e_params *params,
1973                                     struct mlx5e_cq_param *param)
1974 {
1975         void *cqc = param->cqc;
1976         u8 log_cq_size;
1977
1978         switch (params->rq_wq_type) {
1979         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1980                 log_cq_size = params->log_rq_size + params->mpwqe_log_num_strides;
1981                 break;
1982         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1983                 log_cq_size = params->log_rq_size;
1984         }
1985
1986         MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
1987         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
1988                 MLX5_SET(cqc, cqc, mini_cqe_res_format, MLX5_CQE_FORMAT_CSUM);
1989                 MLX5_SET(cqc, cqc, cqe_comp_en, 1);
1990         }
1991
1992         mlx5e_build_common_cq_param(priv, param);
1993         param->cq_period_mode = params->rx_cq_moderation.cq_period_mode;
1994 }
1995
1996 static void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
1997                                     struct mlx5e_params *params,
1998                                     struct mlx5e_cq_param *param)
1999 {
2000         void *cqc = param->cqc;
2001
2002         MLX5_SET(cqc, cqc, log_cq_size, params->log_sq_size);
2003
2004         mlx5e_build_common_cq_param(priv, param);
2005         param->cq_period_mode = params->tx_cq_moderation.cq_period_mode;
2006 }
2007
2008 static void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
2009                                      u8 log_wq_size,
2010                                      struct mlx5e_cq_param *param)
2011 {
2012         void *cqc = param->cqc;
2013
2014         MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
2015
2016         mlx5e_build_common_cq_param(priv, param);
2017
2018         param->cq_period_mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
2019 }
2020
2021 static void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
2022                                     u8 log_wq_size,
2023                                     struct mlx5e_sq_param *param)
2024 {
2025         void *sqc = param->sqc;
2026         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2027
2028         mlx5e_build_sq_param_common(priv, param);
2029
2030         MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
2031         MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(priv->mdev, reg_umr_sq));
2032 }
2033
2034 static void mlx5e_build_xdpsq_param(struct mlx5e_priv *priv,
2035                                     struct mlx5e_params *params,
2036                                     struct mlx5e_sq_param *param)
2037 {
2038         void *sqc = param->sqc;
2039         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2040
2041         mlx5e_build_sq_param_common(priv, param);
2042         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
2043 }
2044
2045 static void mlx5e_build_channel_param(struct mlx5e_priv *priv,
2046                                       struct mlx5e_params *params,
2047                                       struct mlx5e_channel_param *cparam)
2048 {
2049         u8 icosq_log_wq_sz = MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
2050
2051         mlx5e_build_rq_param(priv, params, &cparam->rq);
2052         mlx5e_build_sq_param(priv, params, &cparam->sq);
2053         mlx5e_build_xdpsq_param(priv, params, &cparam->xdp_sq);
2054         mlx5e_build_icosq_param(priv, icosq_log_wq_sz, &cparam->icosq);
2055         mlx5e_build_rx_cq_param(priv, params, &cparam->rx_cq);
2056         mlx5e_build_tx_cq_param(priv, params, &cparam->tx_cq);
2057         mlx5e_build_ico_cq_param(priv, icosq_log_wq_sz, &cparam->icosq_cq);
2058 }
2059
2060 int mlx5e_open_channels(struct mlx5e_priv *priv,
2061                         struct mlx5e_channels *chs)
2062 {
2063         struct mlx5e_channel_param *cparam;
2064         int err = -ENOMEM;
2065         int i;
2066
2067         chs->num = chs->params.num_channels;
2068
2069         chs->c = kcalloc(chs->num, sizeof(struct mlx5e_channel *), GFP_KERNEL);
2070         cparam = kzalloc(sizeof(struct mlx5e_channel_param), GFP_KERNEL);
2071         if (!chs->c || !cparam)
2072                 goto err_free;
2073
2074         mlx5e_build_channel_param(priv, &chs->params, cparam);
2075         for (i = 0; i < chs->num; i++) {
2076                 err = mlx5e_open_channel(priv, i, &chs->params, cparam, &chs->c[i]);
2077                 if (err)
2078                         goto err_close_channels;
2079         }
2080
2081         kfree(cparam);
2082         return 0;
2083
2084 err_close_channels:
2085         for (i--; i >= 0; i--)
2086                 mlx5e_close_channel(chs->c[i]);
2087
2088 err_free:
2089         kfree(chs->c);
2090         kfree(cparam);
2091         chs->num = 0;
2092         return err;
2093 }
2094
2095 static void mlx5e_activate_channels(struct mlx5e_channels *chs)
2096 {
2097         int i;
2098
2099         for (i = 0; i < chs->num; i++)
2100                 mlx5e_activate_channel(chs->c[i]);
2101 }
2102
2103 static int mlx5e_wait_channels_min_rx_wqes(struct mlx5e_channels *chs)
2104 {
2105         int err = 0;
2106         int i;
2107
2108         for (i = 0; i < chs->num; i++) {
2109                 err = mlx5e_wait_for_min_rx_wqes(&chs->c[i]->rq);
2110                 if (err)
2111                         break;
2112         }
2113
2114         return err;
2115 }
2116
2117 static void mlx5e_deactivate_channels(struct mlx5e_channels *chs)
2118 {
2119         int i;
2120
2121         for (i = 0; i < chs->num; i++)
2122                 mlx5e_deactivate_channel(chs->c[i]);
2123 }
2124
2125 void mlx5e_close_channels(struct mlx5e_channels *chs)
2126 {
2127         int i;
2128
2129         for (i = 0; i < chs->num; i++)
2130                 mlx5e_close_channel(chs->c[i]);
2131
2132         kfree(chs->c);
2133         chs->num = 0;
2134 }
2135
2136 static int
2137 mlx5e_create_rqt(struct mlx5e_priv *priv, int sz, struct mlx5e_rqt *rqt)
2138 {
2139         struct mlx5_core_dev *mdev = priv->mdev;
2140         void *rqtc;
2141         int inlen;
2142         int err;
2143         u32 *in;
2144         int i;
2145
2146         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2147         in = kvzalloc(inlen, GFP_KERNEL);
2148         if (!in)
2149                 return -ENOMEM;
2150
2151         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2152
2153         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2154         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2155
2156         for (i = 0; i < sz; i++)
2157                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->drop_rq.rqn);
2158
2159         err = mlx5_core_create_rqt(mdev, in, inlen, &rqt->rqtn);
2160         if (!err)
2161                 rqt->enabled = true;
2162
2163         kvfree(in);
2164         return err;
2165 }
2166
2167 void mlx5e_destroy_rqt(struct mlx5e_priv *priv, struct mlx5e_rqt *rqt)
2168 {
2169         rqt->enabled = false;
2170         mlx5_core_destroy_rqt(priv->mdev, rqt->rqtn);
2171 }
2172
2173 int mlx5e_create_indirect_rqt(struct mlx5e_priv *priv)
2174 {
2175         struct mlx5e_rqt *rqt = &priv->indir_rqt;
2176         int err;
2177
2178         err = mlx5e_create_rqt(priv, MLX5E_INDIR_RQT_SIZE, rqt);
2179         if (err)
2180                 mlx5_core_warn(priv->mdev, "create indirect rqts failed, %d\n", err);
2181         return err;
2182 }
2183
2184 int mlx5e_create_direct_rqts(struct mlx5e_priv *priv)
2185 {
2186         struct mlx5e_rqt *rqt;
2187         int err;
2188         int ix;
2189
2190         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
2191                 rqt = &priv->direct_tir[ix].rqt;
2192                 err = mlx5e_create_rqt(priv, 1 /*size */, rqt);
2193                 if (err)
2194                         goto err_destroy_rqts;
2195         }
2196
2197         return 0;
2198
2199 err_destroy_rqts:
2200         mlx5_core_warn(priv->mdev, "create direct rqts failed, %d\n", err);
2201         for (ix--; ix >= 0; ix--)
2202                 mlx5e_destroy_rqt(priv, &priv->direct_tir[ix].rqt);
2203
2204         return err;
2205 }
2206
2207 void mlx5e_destroy_direct_rqts(struct mlx5e_priv *priv)
2208 {
2209         int i;
2210
2211         for (i = 0; i < priv->profile->max_nch(priv->mdev); i++)
2212                 mlx5e_destroy_rqt(priv, &priv->direct_tir[i].rqt);
2213 }
2214
2215 static int mlx5e_rx_hash_fn(int hfunc)
2216 {
2217         return (hfunc == ETH_RSS_HASH_TOP) ?
2218                MLX5_RX_HASH_FN_TOEPLITZ :
2219                MLX5_RX_HASH_FN_INVERTED_XOR8;
2220 }
2221
2222 static int mlx5e_bits_invert(unsigned long a, int size)
2223 {
2224         int inv = 0;
2225         int i;
2226
2227         for (i = 0; i < size; i++)
2228                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
2229
2230         return inv;
2231 }
2232
2233 static void mlx5e_fill_rqt_rqns(struct mlx5e_priv *priv, int sz,
2234                                 struct mlx5e_redirect_rqt_param rrp, void *rqtc)
2235 {
2236         int i;
2237
2238         for (i = 0; i < sz; i++) {
2239                 u32 rqn;
2240
2241                 if (rrp.is_rss) {
2242                         int ix = i;
2243
2244                         if (rrp.rss.hfunc == ETH_RSS_HASH_XOR)
2245                                 ix = mlx5e_bits_invert(i, ilog2(sz));
2246
2247                         ix = priv->channels.params.indirection_rqt[ix];
2248                         rqn = rrp.rss.channels->c[ix]->rq.rqn;
2249                 } else {
2250                         rqn = rrp.rqn;
2251                 }
2252                 MLX5_SET(rqtc, rqtc, rq_num[i], rqn);
2253         }
2254 }
2255
2256 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, u32 rqtn, int sz,
2257                        struct mlx5e_redirect_rqt_param rrp)
2258 {
2259         struct mlx5_core_dev *mdev = priv->mdev;
2260         void *rqtc;
2261         int inlen;
2262         u32 *in;
2263         int err;
2264
2265         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
2266         in = kvzalloc(inlen, GFP_KERNEL);
2267         if (!in)
2268                 return -ENOMEM;
2269
2270         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
2271
2272         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2273         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
2274         mlx5e_fill_rqt_rqns(priv, sz, rrp, rqtc);
2275         err = mlx5_core_modify_rqt(mdev, rqtn, in, inlen);
2276
2277         kvfree(in);
2278         return err;
2279 }
2280
2281 static u32 mlx5e_get_direct_rqn(struct mlx5e_priv *priv, int ix,
2282                                 struct mlx5e_redirect_rqt_param rrp)
2283 {
2284         if (!rrp.is_rss)
2285                 return rrp.rqn;
2286
2287         if (ix >= rrp.rss.channels->num)
2288                 return priv->drop_rq.rqn;
2289
2290         return rrp.rss.channels->c[ix]->rq.rqn;
2291 }
2292
2293 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv,
2294                                 struct mlx5e_redirect_rqt_param rrp)
2295 {
2296         u32 rqtn;
2297         int ix;
2298
2299         if (priv->indir_rqt.enabled) {
2300                 /* RSS RQ table */
2301                 rqtn = priv->indir_rqt.rqtn;
2302                 mlx5e_redirect_rqt(priv, rqtn, MLX5E_INDIR_RQT_SIZE, rrp);
2303         }
2304
2305         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
2306                 struct mlx5e_redirect_rqt_param direct_rrp = {
2307                         .is_rss = false,
2308                         {
2309                                 .rqn    = mlx5e_get_direct_rqn(priv, ix, rrp)
2310                         },
2311                 };
2312
2313                 /* Direct RQ Tables */
2314                 if (!priv->direct_tir[ix].rqt.enabled)
2315                         continue;
2316
2317                 rqtn = priv->direct_tir[ix].rqt.rqtn;
2318                 mlx5e_redirect_rqt(priv, rqtn, 1, direct_rrp);
2319         }
2320 }
2321
2322 static void mlx5e_redirect_rqts_to_channels(struct mlx5e_priv *priv,
2323                                             struct mlx5e_channels *chs)
2324 {
2325         struct mlx5e_redirect_rqt_param rrp = {
2326                 .is_rss        = true,
2327                 {
2328                         .rss = {
2329                                 .channels  = chs,
2330                                 .hfunc     = chs->params.rss_hfunc,
2331                         }
2332                 },
2333         };
2334
2335         mlx5e_redirect_rqts(priv, rrp);
2336 }
2337
2338 static void mlx5e_redirect_rqts_to_drop(struct mlx5e_priv *priv)
2339 {
2340         struct mlx5e_redirect_rqt_param drop_rrp = {
2341                 .is_rss = false,
2342                 {
2343                         .rqn = priv->drop_rq.rqn,
2344                 },
2345         };
2346
2347         mlx5e_redirect_rqts(priv, drop_rrp);
2348 }
2349
2350 static void mlx5e_build_tir_ctx_lro(struct mlx5e_params *params, void *tirc)
2351 {
2352         if (!params->lro_en)
2353                 return;
2354
2355 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2356
2357         MLX5_SET(tirc, tirc, lro_enable_mask,
2358                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2359                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2360         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
2361                  (params->lro_wqe_sz - ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2362         MLX5_SET(tirc, tirc, lro_timeout_period_usecs, params->lro_timeout);
2363 }
2364
2365 void mlx5e_build_indir_tir_ctx_hash(struct mlx5e_params *params,
2366                                     enum mlx5e_traffic_types tt,
2367                                     void *tirc, bool inner)
2368 {
2369         void *hfso = inner ? MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner) :
2370                              MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2371
2372 #define MLX5_HASH_IP            (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2373                                  MLX5_HASH_FIELD_SEL_DST_IP)
2374
2375 #define MLX5_HASH_IP_L4PORTS    (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2376                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2377                                  MLX5_HASH_FIELD_SEL_L4_SPORT |\
2378                                  MLX5_HASH_FIELD_SEL_L4_DPORT)
2379
2380 #define MLX5_HASH_IP_IPSEC_SPI  (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2381                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2382                                  MLX5_HASH_FIELD_SEL_IPSEC_SPI)
2383
2384         MLX5_SET(tirc, tirc, rx_hash_fn, mlx5e_rx_hash_fn(params->rss_hfunc));
2385         if (params->rss_hfunc == ETH_RSS_HASH_TOP) {
2386                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
2387                                              rx_hash_toeplitz_key);
2388                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
2389                                                rx_hash_toeplitz_key);
2390
2391                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2392                 memcpy(rss_key, params->toeplitz_hash_key, len);
2393         }
2394
2395         switch (tt) {
2396         case MLX5E_TT_IPV4_TCP:
2397                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2398                          MLX5_L3_PROT_TYPE_IPV4);
2399                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2400                          MLX5_L4_PROT_TYPE_TCP);
2401                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2402                          MLX5_HASH_IP_L4PORTS);
2403                 break;
2404
2405         case MLX5E_TT_IPV6_TCP:
2406                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2407                          MLX5_L3_PROT_TYPE_IPV6);
2408                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2409                          MLX5_L4_PROT_TYPE_TCP);
2410                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2411                          MLX5_HASH_IP_L4PORTS);
2412                 break;
2413
2414         case MLX5E_TT_IPV4_UDP:
2415                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2416                          MLX5_L3_PROT_TYPE_IPV4);
2417                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2418                          MLX5_L4_PROT_TYPE_UDP);
2419                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2420                          MLX5_HASH_IP_L4PORTS);
2421                 break;
2422
2423         case MLX5E_TT_IPV6_UDP:
2424                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2425                          MLX5_L3_PROT_TYPE_IPV6);
2426                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2427                          MLX5_L4_PROT_TYPE_UDP);
2428                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2429                          MLX5_HASH_IP_L4PORTS);
2430                 break;
2431
2432         case MLX5E_TT_IPV4_IPSEC_AH:
2433                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2434                          MLX5_L3_PROT_TYPE_IPV4);
2435                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2436                          MLX5_HASH_IP_IPSEC_SPI);
2437                 break;
2438
2439         case MLX5E_TT_IPV6_IPSEC_AH:
2440                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2441                          MLX5_L3_PROT_TYPE_IPV6);
2442                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2443                          MLX5_HASH_IP_IPSEC_SPI);
2444                 break;
2445
2446         case MLX5E_TT_IPV4_IPSEC_ESP:
2447                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2448                          MLX5_L3_PROT_TYPE_IPV4);
2449                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2450                          MLX5_HASH_IP_IPSEC_SPI);
2451                 break;
2452
2453         case MLX5E_TT_IPV6_IPSEC_ESP:
2454                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2455                          MLX5_L3_PROT_TYPE_IPV6);
2456                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2457                          MLX5_HASH_IP_IPSEC_SPI);
2458                 break;
2459
2460         case MLX5E_TT_IPV4:
2461                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2462                          MLX5_L3_PROT_TYPE_IPV4);
2463                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2464                          MLX5_HASH_IP);
2465                 break;
2466
2467         case MLX5E_TT_IPV6:
2468                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2469                          MLX5_L3_PROT_TYPE_IPV6);
2470                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2471                          MLX5_HASH_IP);
2472                 break;
2473         default:
2474                 WARN_ONCE(true, "%s: bad traffic type!\n", __func__);
2475         }
2476 }
2477
2478 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
2479 {
2480         struct mlx5_core_dev *mdev = priv->mdev;
2481
2482         void *in;
2483         void *tirc;
2484         int inlen;
2485         int err;
2486         int tt;
2487         int ix;
2488
2489         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
2490         in = kvzalloc(inlen, GFP_KERNEL);
2491         if (!in)
2492                 return -ENOMEM;
2493
2494         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
2495         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2496
2497         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2498
2499         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2500                 err = mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in,
2501                                            inlen);
2502                 if (err)
2503                         goto free_in;
2504         }
2505
2506         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
2507                 err = mlx5_core_modify_tir(mdev, priv->direct_tir[ix].tirn,
2508                                            in, inlen);
2509                 if (err)
2510                         goto free_in;
2511         }
2512
2513 free_in:
2514         kvfree(in);
2515
2516         return err;
2517 }
2518
2519 static void mlx5e_build_inner_indir_tir_ctx(struct mlx5e_priv *priv,
2520                                             enum mlx5e_traffic_types tt,
2521                                             u32 *tirc)
2522 {
2523         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2524
2525         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2526
2527         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2528         MLX5_SET(tirc, tirc, indirect_table, priv->indir_rqt.rqtn);
2529         MLX5_SET(tirc, tirc, tunneled_offload_en, 0x1);
2530
2531         mlx5e_build_indir_tir_ctx_hash(&priv->channels.params, tt, tirc, true);
2532 }
2533
2534 static int mlx5e_set_mtu(struct mlx5e_priv *priv, u16 mtu)
2535 {
2536         struct mlx5_core_dev *mdev = priv->mdev;
2537         u16 hw_mtu = MLX5E_SW2HW_MTU(priv, mtu);
2538         int err;
2539
2540         err = mlx5_set_port_mtu(mdev, hw_mtu, 1);
2541         if (err)
2542                 return err;
2543
2544         /* Update vport context MTU */
2545         mlx5_modify_nic_vport_mtu(mdev, hw_mtu);
2546         return 0;
2547 }
2548
2549 static void mlx5e_query_mtu(struct mlx5e_priv *priv, u16 *mtu)
2550 {
2551         struct mlx5_core_dev *mdev = priv->mdev;
2552         u16 hw_mtu = 0;
2553         int err;
2554
2555         err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2556         if (err || !hw_mtu) /* fallback to port oper mtu */
2557                 mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
2558
2559         *mtu = MLX5E_HW2SW_MTU(priv, hw_mtu);
2560 }
2561
2562 static int mlx5e_set_dev_port_mtu(struct mlx5e_priv *priv)
2563 {
2564         struct net_device *netdev = priv->netdev;
2565         u16 mtu;
2566         int err;
2567
2568         err = mlx5e_set_mtu(priv, netdev->mtu);
2569         if (err)
2570                 return err;
2571
2572         mlx5e_query_mtu(priv, &mtu);
2573         if (mtu != netdev->mtu)
2574                 netdev_warn(netdev, "%s: VPort MTU %d is different than netdev mtu %d\n",
2575                             __func__, mtu, netdev->mtu);
2576
2577         netdev->mtu = mtu;
2578         return 0;
2579 }
2580
2581 static void mlx5e_netdev_set_tcs(struct net_device *netdev)
2582 {
2583         struct mlx5e_priv *priv = netdev_priv(netdev);
2584         int nch = priv->channels.params.num_channels;
2585         int ntc = priv->channels.params.num_tc;
2586         int tc;
2587
2588         netdev_reset_tc(netdev);
2589
2590         if (ntc == 1)
2591                 return;
2592
2593         netdev_set_num_tc(netdev, ntc);
2594
2595         /* Map netdev TCs to offset 0
2596          * We have our own UP to TXQ mapping for QoS
2597          */
2598         for (tc = 0; tc < ntc; tc++)
2599                 netdev_set_tc_queue(netdev, tc, nch, 0);
2600 }
2601
2602 static void mlx5e_build_channels_tx_maps(struct mlx5e_priv *priv)
2603 {
2604         struct mlx5e_channel *c;
2605         struct mlx5e_txqsq *sq;
2606         int i, tc;
2607
2608         for (i = 0; i < priv->channels.num; i++)
2609                 for (tc = 0; tc < priv->profile->max_tc; tc++)
2610                         priv->channel_tc2txq[i][tc] = i + tc * priv->channels.num;
2611
2612         for (i = 0; i < priv->channels.num; i++) {
2613                 c = priv->channels.c[i];
2614                 for (tc = 0; tc < c->num_tc; tc++) {
2615                         sq = &c->sq[tc];
2616                         priv->txq2sq[sq->txq_ix] = sq;
2617                 }
2618         }
2619 }
2620
2621 void mlx5e_activate_priv_channels(struct mlx5e_priv *priv)
2622 {
2623         int num_txqs = priv->channels.num * priv->channels.params.num_tc;
2624         struct net_device *netdev = priv->netdev;
2625
2626         mlx5e_netdev_set_tcs(netdev);
2627         netif_set_real_num_tx_queues(netdev, num_txqs);
2628         netif_set_real_num_rx_queues(netdev, priv->channels.num);
2629
2630         mlx5e_build_channels_tx_maps(priv);
2631         mlx5e_activate_channels(&priv->channels);
2632         netif_tx_start_all_queues(priv->netdev);
2633
2634         if (MLX5_VPORT_MANAGER(priv->mdev))
2635                 mlx5e_add_sqs_fwd_rules(priv);
2636
2637         mlx5e_wait_channels_min_rx_wqes(&priv->channels);
2638         mlx5e_redirect_rqts_to_channels(priv, &priv->channels);
2639 }
2640
2641 void mlx5e_deactivate_priv_channels(struct mlx5e_priv *priv)
2642 {
2643         mlx5e_redirect_rqts_to_drop(priv);
2644
2645         if (MLX5_VPORT_MANAGER(priv->mdev))
2646                 mlx5e_remove_sqs_fwd_rules(priv);
2647
2648         /* FIXME: This is a W/A only for tx timeout watch dog false alarm when
2649          * polling for inactive tx queues.
2650          */
2651         netif_tx_stop_all_queues(priv->netdev);
2652         netif_tx_disable(priv->netdev);
2653         mlx5e_deactivate_channels(&priv->channels);
2654 }
2655
2656 void mlx5e_switch_priv_channels(struct mlx5e_priv *priv,
2657                                 struct mlx5e_channels *new_chs,
2658                                 mlx5e_fp_hw_modify hw_modify)
2659 {
2660         struct net_device *netdev = priv->netdev;
2661         int new_num_txqs;
2662         int carrier_ok;
2663         new_num_txqs = new_chs->num * new_chs->params.num_tc;
2664
2665         carrier_ok = netif_carrier_ok(netdev);
2666         netif_carrier_off(netdev);
2667
2668         if (new_num_txqs < netdev->real_num_tx_queues)
2669                 netif_set_real_num_tx_queues(netdev, new_num_txqs);
2670
2671         mlx5e_deactivate_priv_channels(priv);
2672         mlx5e_close_channels(&priv->channels);
2673
2674         priv->channels = *new_chs;
2675
2676         /* New channels are ready to roll, modify HW settings if needed */
2677         if (hw_modify)
2678                 hw_modify(priv);
2679
2680         mlx5e_refresh_tirs(priv, false);
2681         mlx5e_activate_priv_channels(priv);
2682
2683         /* return carrier back if needed */
2684         if (carrier_ok)
2685                 netif_carrier_on(netdev);
2686 }
2687
2688 void mlx5e_timestamp_set(struct mlx5e_priv *priv)
2689 {
2690         priv->tstamp.tx_type   = HWTSTAMP_TX_OFF;
2691         priv->tstamp.rx_filter = HWTSTAMP_FILTER_NONE;
2692 }
2693
2694 int mlx5e_open_locked(struct net_device *netdev)
2695 {
2696         struct mlx5e_priv *priv = netdev_priv(netdev);
2697         int err;
2698
2699         set_bit(MLX5E_STATE_OPENED, &priv->state);
2700
2701         err = mlx5e_open_channels(priv, &priv->channels);
2702         if (err)
2703                 goto err_clear_state_opened_flag;
2704
2705         mlx5e_refresh_tirs(priv, false);
2706         mlx5e_activate_priv_channels(priv);
2707         if (priv->profile->update_carrier)
2708                 priv->profile->update_carrier(priv);
2709         mlx5e_timestamp_set(priv);
2710
2711         if (priv->profile->update_stats)
2712                 queue_delayed_work(priv->wq, &priv->update_stats_work, 0);
2713
2714         return 0;
2715
2716 err_clear_state_opened_flag:
2717         clear_bit(MLX5E_STATE_OPENED, &priv->state);
2718         return err;
2719 }
2720
2721 int mlx5e_open(struct net_device *netdev)
2722 {
2723         struct mlx5e_priv *priv = netdev_priv(netdev);
2724         int err;
2725
2726         mutex_lock(&priv->state_lock);
2727         err = mlx5e_open_locked(netdev);
2728         if (!err)
2729                 mlx5_set_port_admin_status(priv->mdev, MLX5_PORT_UP);
2730         mutex_unlock(&priv->state_lock);
2731
2732         return err;
2733 }
2734
2735 int mlx5e_close_locked(struct net_device *netdev)
2736 {
2737         struct mlx5e_priv *priv = netdev_priv(netdev);
2738
2739         /* May already be CLOSED in case a previous configuration operation
2740          * (e.g RX/TX queue size change) that involves close&open failed.
2741          */
2742         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
2743                 return 0;
2744
2745         clear_bit(MLX5E_STATE_OPENED, &priv->state);
2746
2747         netif_carrier_off(priv->netdev);
2748         mlx5e_deactivate_priv_channels(priv);
2749         mlx5e_close_channels(&priv->channels);
2750
2751         return 0;
2752 }
2753
2754 int mlx5e_close(struct net_device *netdev)
2755 {
2756         struct mlx5e_priv *priv = netdev_priv(netdev);
2757         int err;
2758
2759         if (!netif_device_present(netdev))
2760                 return -ENODEV;
2761
2762         mutex_lock(&priv->state_lock);
2763         mlx5_set_port_admin_status(priv->mdev, MLX5_PORT_DOWN);
2764         err = mlx5e_close_locked(netdev);
2765         mutex_unlock(&priv->state_lock);
2766
2767         return err;
2768 }
2769
2770 static int mlx5e_alloc_drop_rq(struct mlx5_core_dev *mdev,
2771                                struct mlx5e_rq *rq,
2772                                struct mlx5e_rq_param *param)
2773 {
2774         void *rqc = param->rqc;
2775         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
2776         int err;
2777
2778         param->wq.db_numa_node = param->wq.buf_numa_node;
2779
2780         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
2781                                 &rq->wq_ctrl);
2782         if (err)
2783                 return err;
2784
2785         /* Mark as unused given "Drop-RQ" packets never reach XDP */
2786         xdp_rxq_info_unused(&rq->xdp_rxq);
2787
2788         rq->mdev = mdev;
2789
2790         return 0;
2791 }
2792
2793 static int mlx5e_alloc_drop_cq(struct mlx5_core_dev *mdev,
2794                                struct mlx5e_cq *cq,
2795                                struct mlx5e_cq_param *param)
2796 {
2797         return mlx5e_alloc_cq_common(mdev, param, cq);
2798 }
2799
2800 static int mlx5e_open_drop_rq(struct mlx5_core_dev *mdev,
2801                               struct mlx5e_rq *drop_rq)
2802 {
2803         struct mlx5e_cq_param cq_param = {};
2804         struct mlx5e_rq_param rq_param = {};
2805         struct mlx5e_cq *cq = &drop_rq->cq;
2806         int err;
2807
2808         mlx5e_build_drop_rq_param(&rq_param);
2809
2810         err = mlx5e_alloc_drop_cq(mdev, cq, &cq_param);
2811         if (err)
2812                 return err;
2813
2814         err = mlx5e_create_cq(cq, &cq_param);
2815         if (err)
2816                 goto err_free_cq;
2817
2818         err = mlx5e_alloc_drop_rq(mdev, drop_rq, &rq_param);
2819         if (err)
2820                 goto err_destroy_cq;
2821
2822         err = mlx5e_create_rq(drop_rq, &rq_param);
2823         if (err)
2824                 goto err_free_rq;
2825
2826         return 0;
2827
2828 err_free_rq:
2829         mlx5e_free_rq(drop_rq);
2830
2831 err_destroy_cq:
2832         mlx5e_destroy_cq(cq);
2833
2834 err_free_cq:
2835         mlx5e_free_cq(cq);
2836
2837         return err;
2838 }
2839
2840 static void mlx5e_close_drop_rq(struct mlx5e_rq *drop_rq)
2841 {
2842         mlx5e_destroy_rq(drop_rq);
2843         mlx5e_free_rq(drop_rq);
2844         mlx5e_destroy_cq(&drop_rq->cq);
2845         mlx5e_free_cq(&drop_rq->cq);
2846 }
2847
2848 int mlx5e_create_tis(struct mlx5_core_dev *mdev, int tc,
2849                      u32 underlay_qpn, u32 *tisn)
2850 {
2851         u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
2852         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
2853
2854         MLX5_SET(tisc, tisc, prio, tc << 1);
2855         MLX5_SET(tisc, tisc, underlay_qpn, underlay_qpn);
2856         MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.td.tdn);
2857
2858         if (mlx5_lag_is_lacp_owner(mdev))
2859                 MLX5_SET(tisc, tisc, strict_lag_tx_port_affinity, 1);
2860
2861         return mlx5_core_create_tis(mdev, in, sizeof(in), tisn);
2862 }
2863
2864 void mlx5e_destroy_tis(struct mlx5_core_dev *mdev, u32 tisn)
2865 {
2866         mlx5_core_destroy_tis(mdev, tisn);
2867 }
2868
2869 int mlx5e_create_tises(struct mlx5e_priv *priv)
2870 {
2871         int err;
2872         int tc;
2873
2874         for (tc = 0; tc < priv->profile->max_tc; tc++) {
2875                 err = mlx5e_create_tis(priv->mdev, tc, 0, &priv->tisn[tc]);
2876                 if (err)
2877                         goto err_close_tises;
2878         }
2879
2880         return 0;
2881
2882 err_close_tises:
2883         for (tc--; tc >= 0; tc--)
2884                 mlx5e_destroy_tis(priv->mdev, priv->tisn[tc]);
2885
2886         return err;
2887 }
2888
2889 void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
2890 {
2891         int tc;
2892
2893         for (tc = 0; tc < priv->profile->max_tc; tc++)
2894                 mlx5e_destroy_tis(priv->mdev, priv->tisn[tc]);
2895 }
2896
2897 static void mlx5e_build_indir_tir_ctx(struct mlx5e_priv *priv,
2898                                       enum mlx5e_traffic_types tt,
2899                                       u32 *tirc)
2900 {
2901         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2902
2903         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2904
2905         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2906         MLX5_SET(tirc, tirc, indirect_table, priv->indir_rqt.rqtn);
2907         mlx5e_build_indir_tir_ctx_hash(&priv->channels.params, tt, tirc, false);
2908 }
2909
2910 static void mlx5e_build_direct_tir_ctx(struct mlx5e_priv *priv, u32 rqtn, u32 *tirc)
2911 {
2912         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2913
2914         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2915
2916         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2917         MLX5_SET(tirc, tirc, indirect_table, rqtn);
2918         MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
2919 }
2920
2921 int mlx5e_create_indirect_tirs(struct mlx5e_priv *priv)
2922 {
2923         struct mlx5e_tir *tir;
2924         void *tirc;
2925         int inlen;
2926         int i = 0;
2927         int err;
2928         u32 *in;
2929         int tt;
2930
2931         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2932         in = kvzalloc(inlen, GFP_KERNEL);
2933         if (!in)
2934                 return -ENOMEM;
2935
2936         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2937                 memset(in, 0, inlen);
2938                 tir = &priv->indir_tir[tt];
2939                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2940                 mlx5e_build_indir_tir_ctx(priv, tt, tirc);
2941                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2942                 if (err) {
2943                         mlx5_core_warn(priv->mdev, "create indirect tirs failed, %d\n", err);
2944                         goto err_destroy_inner_tirs;
2945                 }
2946         }
2947
2948         if (!mlx5e_tunnel_inner_ft_supported(priv->mdev))
2949                 goto out;
2950
2951         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++) {
2952                 memset(in, 0, inlen);
2953                 tir = &priv->inner_indir_tir[i];
2954                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2955                 mlx5e_build_inner_indir_tir_ctx(priv, i, tirc);
2956                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2957                 if (err) {
2958                         mlx5_core_warn(priv->mdev, "create inner indirect tirs failed, %d\n", err);
2959                         goto err_destroy_inner_tirs;
2960                 }
2961         }
2962
2963 out:
2964         kvfree(in);
2965
2966         return 0;
2967
2968 err_destroy_inner_tirs:
2969         for (i--; i >= 0; i--)
2970                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
2971
2972         for (tt--; tt >= 0; tt--)
2973                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[tt]);
2974
2975         kvfree(in);
2976
2977         return err;
2978 }
2979
2980 int mlx5e_create_direct_tirs(struct mlx5e_priv *priv)
2981 {
2982         int nch = priv->profile->max_nch(priv->mdev);
2983         struct mlx5e_tir *tir;
2984         void *tirc;
2985         int inlen;
2986         int err;
2987         u32 *in;
2988         int ix;
2989
2990         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2991         in = kvzalloc(inlen, GFP_KERNEL);
2992         if (!in)
2993                 return -ENOMEM;
2994
2995         for (ix = 0; ix < nch; ix++) {
2996                 memset(in, 0, inlen);
2997                 tir = &priv->direct_tir[ix];
2998                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2999                 mlx5e_build_direct_tir_ctx(priv, priv->direct_tir[ix].rqt.rqtn, tirc);
3000                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
3001                 if (err)
3002                         goto err_destroy_ch_tirs;
3003         }
3004
3005         kvfree(in);
3006
3007         return 0;
3008
3009 err_destroy_ch_tirs:
3010         mlx5_core_warn(priv->mdev, "create direct tirs failed, %d\n", err);
3011         for (ix--; ix >= 0; ix--)
3012                 mlx5e_destroy_tir(priv->mdev, &priv->direct_tir[ix]);
3013
3014         kvfree(in);
3015
3016         return err;
3017 }
3018
3019 void mlx5e_destroy_indirect_tirs(struct mlx5e_priv *priv)
3020 {
3021         int i;
3022
3023         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3024                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[i]);
3025
3026         if (!mlx5e_tunnel_inner_ft_supported(priv->mdev))
3027                 return;
3028
3029         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3030                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3031 }
3032
3033 void mlx5e_destroy_direct_tirs(struct mlx5e_priv *priv)
3034 {
3035         int nch = priv->profile->max_nch(priv->mdev);
3036         int i;
3037
3038         for (i = 0; i < nch; i++)
3039                 mlx5e_destroy_tir(priv->mdev, &priv->direct_tir[i]);
3040 }
3041
3042 static int mlx5e_modify_channels_scatter_fcs(struct mlx5e_channels *chs, bool enable)
3043 {
3044         int err = 0;
3045         int i;
3046
3047         for (i = 0; i < chs->num; i++) {
3048                 err = mlx5e_modify_rq_scatter_fcs(&chs->c[i]->rq, enable);
3049                 if (err)
3050                         return err;
3051         }
3052
3053         return 0;
3054 }
3055
3056 static int mlx5e_modify_channels_vsd(struct mlx5e_channels *chs, bool vsd)
3057 {
3058         int err = 0;
3059         int i;
3060
3061         for (i = 0; i < chs->num; i++) {
3062                 err = mlx5e_modify_rq_vsd(&chs->c[i]->rq, vsd);
3063                 if (err)
3064                         return err;
3065         }
3066
3067         return 0;
3068 }
3069
3070 static int mlx5e_setup_tc_mqprio(struct net_device *netdev,
3071                                  struct tc_mqprio_qopt *mqprio)
3072 {
3073         struct mlx5e_priv *priv = netdev_priv(netdev);
3074         struct mlx5e_channels new_channels = {};
3075         u8 tc = mqprio->num_tc;
3076         int err = 0;
3077
3078         mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
3079
3080         if (tc && tc != MLX5E_MAX_NUM_TC)
3081                 return -EINVAL;
3082
3083         mutex_lock(&priv->state_lock);
3084
3085         new_channels.params = priv->channels.params;
3086         new_channels.params.num_tc = tc ? tc : 1;
3087
3088         if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
3089                 priv->channels.params = new_channels.params;
3090                 goto out;
3091         }
3092
3093         err = mlx5e_open_channels(priv, &new_channels);
3094         if (err)
3095                 goto out;
3096
3097         mlx5e_switch_priv_channels(priv, &new_channels, NULL);
3098 out:
3099         mutex_unlock(&priv->state_lock);
3100         return err;
3101 }
3102
3103 #ifdef CONFIG_MLX5_ESWITCH
3104 static int mlx5e_setup_tc_cls_flower(struct mlx5e_priv *priv,
3105                                      struct tc_cls_flower_offload *cls_flower)
3106 {
3107         if (cls_flower->common.chain_index)
3108                 return -EOPNOTSUPP;
3109
3110         switch (cls_flower->command) {
3111         case TC_CLSFLOWER_REPLACE:
3112                 return mlx5e_configure_flower(priv, cls_flower);
3113         case TC_CLSFLOWER_DESTROY:
3114                 return mlx5e_delete_flower(priv, cls_flower);
3115         case TC_CLSFLOWER_STATS:
3116                 return mlx5e_stats_flower(priv, cls_flower);
3117         default:
3118                 return -EOPNOTSUPP;
3119         }
3120 }
3121
3122 int mlx5e_setup_tc_block_cb(enum tc_setup_type type, void *type_data,
3123                             void *cb_priv)
3124 {
3125         struct mlx5e_priv *priv = cb_priv;
3126
3127         if (!tc_can_offload(priv->netdev))
3128                 return -EOPNOTSUPP;
3129
3130         switch (type) {
3131         case TC_SETUP_CLSFLOWER:
3132                 return mlx5e_setup_tc_cls_flower(priv, type_data);
3133         default:
3134                 return -EOPNOTSUPP;
3135         }
3136 }
3137
3138 static int mlx5e_setup_tc_block(struct net_device *dev,
3139                                 struct tc_block_offload *f)
3140 {
3141         struct mlx5e_priv *priv = netdev_priv(dev);
3142
3143         if (f->binder_type != TCF_BLOCK_BINDER_TYPE_CLSACT_INGRESS)
3144                 return -EOPNOTSUPP;
3145
3146         switch (f->command) {
3147         case TC_BLOCK_BIND:
3148                 return tcf_block_cb_register(f->block, mlx5e_setup_tc_block_cb,
3149                                              priv, priv);
3150         case TC_BLOCK_UNBIND:
3151                 tcf_block_cb_unregister(f->block, mlx5e_setup_tc_block_cb,
3152                                         priv);
3153                 return 0;
3154         default:
3155                 return -EOPNOTSUPP;
3156         }
3157 }
3158 #endif
3159
3160 int mlx5e_setup_tc(struct net_device *dev, enum tc_setup_type type,
3161                    void *type_data)
3162 {
3163         switch (type) {
3164 #ifdef CONFIG_MLX5_ESWITCH
3165         case TC_SETUP_BLOCK:
3166                 return mlx5e_setup_tc_block(dev, type_data);
3167 #endif
3168         case TC_SETUP_QDISC_MQPRIO:
3169                 return mlx5e_setup_tc_mqprio(dev, type_data);
3170         default:
3171                 return -EOPNOTSUPP;
3172         }
3173 }
3174
3175 static void
3176 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
3177 {
3178         struct mlx5e_priv *priv = netdev_priv(dev);
3179         struct mlx5e_sw_stats *sstats = &priv->stats.sw;
3180         struct mlx5e_vport_stats *vstats = &priv->stats.vport;
3181         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
3182
3183         if (mlx5e_is_uplink_rep(priv)) {
3184                 stats->rx_packets = PPORT_802_3_GET(pstats, a_frames_received_ok);
3185                 stats->rx_bytes   = PPORT_802_3_GET(pstats, a_octets_received_ok);
3186                 stats->tx_packets = PPORT_802_3_GET(pstats, a_frames_transmitted_ok);
3187                 stats->tx_bytes   = PPORT_802_3_GET(pstats, a_octets_transmitted_ok);
3188         } else {
3189                 stats->rx_packets = sstats->rx_packets;
3190                 stats->rx_bytes   = sstats->rx_bytes;
3191                 stats->tx_packets = sstats->tx_packets;
3192                 stats->tx_bytes   = sstats->tx_bytes;
3193                 stats->tx_dropped = sstats->tx_queue_dropped;
3194         }
3195
3196         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
3197
3198         stats->rx_length_errors =
3199                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
3200                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
3201                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
3202         stats->rx_crc_errors =
3203                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
3204         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
3205         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
3206         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
3207                            stats->rx_frame_errors;
3208         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
3209
3210         /* vport multicast also counts packets that are dropped due to steering
3211          * or rx out of buffer
3212          */
3213         stats->multicast =
3214                 VPORT_COUNTER_GET(vstats, received_eth_multicast.packets);
3215 }
3216
3217 static void mlx5e_set_rx_mode(struct net_device *dev)
3218 {
3219         struct mlx5e_priv *priv = netdev_priv(dev);
3220
3221         queue_work(priv->wq, &priv->set_rx_mode_work);
3222 }
3223
3224 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
3225 {
3226         struct mlx5e_priv *priv = netdev_priv(netdev);
3227         struct sockaddr *saddr = addr;
3228
3229         if (!is_valid_ether_addr(saddr->sa_data))
3230                 return -EADDRNOTAVAIL;
3231
3232         netif_addr_lock_bh(netdev);
3233         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
3234         netif_addr_unlock_bh(netdev);
3235
3236         queue_work(priv->wq, &priv->set_rx_mode_work);
3237
3238         return 0;
3239 }
3240
3241 #define MLX5E_SET_FEATURE(netdev, feature, enable)      \
3242         do {                                            \
3243                 if (enable)                             \
3244                         netdev->features |= feature;    \
3245                 else                                    \
3246                         netdev->features &= ~feature;   \
3247         } while (0)
3248
3249 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
3250
3251 static int set_feature_lro(struct net_device *netdev, bool enable)
3252 {
3253         struct mlx5e_priv *priv = netdev_priv(netdev);
3254         struct mlx5e_channels new_channels = {};
3255         int err = 0;
3256         bool reset;
3257
3258         mutex_lock(&priv->state_lock);
3259
3260         reset = (priv->channels.params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST);
3261         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
3262
3263         new_channels.params = priv->channels.params;
3264         new_channels.params.lro_en = enable;
3265
3266         if (!reset) {
3267                 priv->channels.params = new_channels.params;
3268                 err = mlx5e_modify_tirs_lro(priv);
3269                 goto out;
3270         }
3271
3272         err = mlx5e_open_channels(priv, &new_channels);
3273         if (err)
3274                 goto out;
3275
3276         mlx5e_switch_priv_channels(priv, &new_channels, mlx5e_modify_tirs_lro);
3277 out:
3278         mutex_unlock(&priv->state_lock);
3279         return err;
3280 }
3281
3282 static int set_feature_cvlan_filter(struct net_device *netdev, bool enable)
3283 {
3284         struct mlx5e_priv *priv = netdev_priv(netdev);
3285
3286         if (enable)
3287                 mlx5e_enable_cvlan_filter(priv);
3288         else
3289                 mlx5e_disable_cvlan_filter(priv);
3290
3291         return 0;
3292 }
3293
3294 static int set_feature_tc_num_filters(struct net_device *netdev, bool enable)
3295 {
3296         struct mlx5e_priv *priv = netdev_priv(netdev);
3297
3298         if (!enable && mlx5e_tc_num_filters(priv)) {
3299                 netdev_err(netdev,
3300                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
3301                 return -EINVAL;
3302         }
3303
3304         return 0;
3305 }
3306
3307 static int set_feature_rx_all(struct net_device *netdev, bool enable)
3308 {
3309         struct mlx5e_priv *priv = netdev_priv(netdev);
3310         struct mlx5_core_dev *mdev = priv->mdev;
3311
3312         return mlx5_set_port_fcs(mdev, !enable);
3313 }
3314
3315 static int set_feature_rx_fcs(struct net_device *netdev, bool enable)
3316 {
3317         struct mlx5e_priv *priv = netdev_priv(netdev);
3318         int err;
3319
3320         mutex_lock(&priv->state_lock);
3321
3322         priv->channels.params.scatter_fcs_en = enable;
3323         err = mlx5e_modify_channels_scatter_fcs(&priv->channels, enable);
3324         if (err)
3325                 priv->channels.params.scatter_fcs_en = !enable;
3326
3327         mutex_unlock(&priv->state_lock);
3328
3329         return err;
3330 }
3331
3332 static int set_feature_rx_vlan(struct net_device *netdev, bool enable)
3333 {
3334         struct mlx5e_priv *priv = netdev_priv(netdev);
3335         int err = 0;
3336
3337         mutex_lock(&priv->state_lock);
3338
3339         priv->channels.params.vlan_strip_disable = !enable;
3340         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3341                 goto unlock;
3342
3343         err = mlx5e_modify_channels_vsd(&priv->channels, !enable);
3344         if (err)
3345                 priv->channels.params.vlan_strip_disable = enable;
3346
3347 unlock:
3348         mutex_unlock(&priv->state_lock);
3349
3350         return err;
3351 }
3352
3353 #ifdef CONFIG_RFS_ACCEL
3354 static int set_feature_arfs(struct net_device *netdev, bool enable)
3355 {
3356         struct mlx5e_priv *priv = netdev_priv(netdev);
3357         int err;
3358
3359         if (enable)
3360                 err = mlx5e_arfs_enable(priv);
3361         else
3362                 err = mlx5e_arfs_disable(priv);
3363
3364         return err;
3365 }
3366 #endif
3367
3368 static int mlx5e_handle_feature(struct net_device *netdev,
3369                                 netdev_features_t wanted_features,
3370                                 netdev_features_t feature,
3371                                 mlx5e_feature_handler feature_handler)
3372 {
3373         netdev_features_t changes = wanted_features ^ netdev->features;
3374         bool enable = !!(wanted_features & feature);
3375         int err;
3376
3377         if (!(changes & feature))
3378                 return 0;
3379
3380         err = feature_handler(netdev, enable);
3381         if (err) {
3382                 netdev_err(netdev, "%s feature %pNF failed, err %d\n",
3383                            enable ? "Enable" : "Disable", &feature, err);
3384                 return err;
3385         }
3386
3387         MLX5E_SET_FEATURE(netdev, feature, enable);
3388         return 0;
3389 }
3390
3391 static int mlx5e_set_features(struct net_device *netdev,
3392                               netdev_features_t features)
3393 {
3394         int err;
3395
3396         err  = mlx5e_handle_feature(netdev, features, NETIF_F_LRO,
3397                                     set_feature_lro);
3398         err |= mlx5e_handle_feature(netdev, features,
3399                                     NETIF_F_HW_VLAN_CTAG_FILTER,
3400                                     set_feature_cvlan_filter);
3401         err |= mlx5e_handle_feature(netdev, features, NETIF_F_HW_TC,
3402                                     set_feature_tc_num_filters);
3403         err |= mlx5e_handle_feature(netdev, features, NETIF_F_RXALL,
3404                                     set_feature_rx_all);
3405         err |= mlx5e_handle_feature(netdev, features, NETIF_F_RXFCS,
3406                                     set_feature_rx_fcs);
3407         err |= mlx5e_handle_feature(netdev, features, NETIF_F_HW_VLAN_CTAG_RX,
3408                                     set_feature_rx_vlan);
3409 #ifdef CONFIG_RFS_ACCEL
3410         err |= mlx5e_handle_feature(netdev, features, NETIF_F_NTUPLE,
3411                                     set_feature_arfs);
3412 #endif
3413
3414         return err ? -EINVAL : 0;
3415 }
3416
3417 static netdev_features_t mlx5e_fix_features(struct net_device *netdev,
3418                                             netdev_features_t features)
3419 {
3420         struct mlx5e_priv *priv = netdev_priv(netdev);
3421
3422         mutex_lock(&priv->state_lock);
3423         if (!bitmap_empty(priv->fs.vlan.active_svlans, VLAN_N_VID)) {
3424                 /* HW strips the outer C-tag header, this is a problem
3425                  * for S-tag traffic.
3426                  */
3427                 features &= ~NETIF_F_HW_VLAN_CTAG_RX;
3428                 if (!priv->channels.params.vlan_strip_disable)
3429                         netdev_warn(netdev, "Dropping C-tag vlan stripping offload due to S-tag vlan\n");
3430         }
3431         mutex_unlock(&priv->state_lock);
3432
3433         return features;
3434 }
3435
3436 static int mlx5e_change_mtu(struct net_device *netdev, int new_mtu)
3437 {
3438         struct mlx5e_priv *priv = netdev_priv(netdev);
3439         struct mlx5e_channels new_channels = {};
3440         int curr_mtu;
3441         int err = 0;
3442         bool reset;
3443
3444         mutex_lock(&priv->state_lock);
3445
3446         reset = !priv->channels.params.lro_en &&
3447                 (priv->channels.params.rq_wq_type !=
3448                  MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
3449
3450         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
3451
3452         curr_mtu    = netdev->mtu;
3453         netdev->mtu = new_mtu;
3454
3455         if (!reset) {
3456                 mlx5e_set_dev_port_mtu(priv);
3457                 goto out;
3458         }
3459
3460         new_channels.params = priv->channels.params;
3461         err = mlx5e_open_channels(priv, &new_channels);
3462         if (err) {
3463                 netdev->mtu = curr_mtu;
3464                 goto out;
3465         }
3466
3467         mlx5e_switch_priv_channels(priv, &new_channels, mlx5e_set_dev_port_mtu);
3468
3469 out:
3470         mutex_unlock(&priv->state_lock);
3471         return err;
3472 }
3473
3474 int mlx5e_hwstamp_set(struct mlx5e_priv *priv, struct ifreq *ifr)
3475 {
3476         struct hwtstamp_config config;
3477         int err;
3478
3479         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
3480                 return -EOPNOTSUPP;
3481
3482         if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
3483                 return -EFAULT;
3484
3485         /* TX HW timestamp */
3486         switch (config.tx_type) {
3487         case HWTSTAMP_TX_OFF:
3488         case HWTSTAMP_TX_ON:
3489                 break;
3490         default:
3491                 return -ERANGE;
3492         }
3493
3494         mutex_lock(&priv->state_lock);
3495         /* RX HW timestamp */
3496         switch (config.rx_filter) {
3497         case HWTSTAMP_FILTER_NONE:
3498                 /* Reset CQE compression to Admin default */
3499                 mlx5e_modify_rx_cqe_compression_locked(priv, priv->channels.params.rx_cqe_compress_def);
3500                 break;
3501         case HWTSTAMP_FILTER_ALL:
3502         case HWTSTAMP_FILTER_SOME:
3503         case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
3504         case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
3505         case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
3506         case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
3507         case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
3508         case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
3509         case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
3510         case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
3511         case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
3512         case HWTSTAMP_FILTER_PTP_V2_EVENT:
3513         case HWTSTAMP_FILTER_PTP_V2_SYNC:
3514         case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
3515         case HWTSTAMP_FILTER_NTP_ALL:
3516                 /* Disable CQE compression */
3517                 netdev_warn(priv->netdev, "Disabling cqe compression");
3518                 err = mlx5e_modify_rx_cqe_compression_locked(priv, false);
3519                 if (err) {
3520                         netdev_err(priv->netdev, "Failed disabling cqe compression err=%d\n", err);
3521                         mutex_unlock(&priv->state_lock);
3522                         return err;
3523                 }
3524                 config.rx_filter = HWTSTAMP_FILTER_ALL;
3525                 break;
3526         default:
3527                 mutex_unlock(&priv->state_lock);
3528                 return -ERANGE;
3529         }
3530
3531         memcpy(&priv->tstamp, &config, sizeof(config));
3532         mutex_unlock(&priv->state_lock);
3533
3534         return copy_to_user(ifr->ifr_data, &config,
3535                             sizeof(config)) ? -EFAULT : 0;
3536 }
3537
3538 int mlx5e_hwstamp_get(struct mlx5e_priv *priv, struct ifreq *ifr)
3539 {
3540         struct hwtstamp_config *cfg = &priv->tstamp;
3541
3542         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
3543                 return -EOPNOTSUPP;
3544
3545         return copy_to_user(ifr->ifr_data, cfg, sizeof(*cfg)) ? -EFAULT : 0;
3546 }
3547
3548 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
3549 {
3550         struct mlx5e_priv *priv = netdev_priv(dev);
3551
3552         switch (cmd) {
3553         case SIOCSHWTSTAMP:
3554                 return mlx5e_hwstamp_set(priv, ifr);
3555         case SIOCGHWTSTAMP:
3556                 return mlx5e_hwstamp_get(priv, ifr);
3557         default:
3558                 return -EOPNOTSUPP;
3559         }
3560 }
3561
3562 #ifdef CONFIG_MLX5_ESWITCH
3563 static int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
3564 {
3565         struct mlx5e_priv *priv = netdev_priv(dev);
3566         struct mlx5_core_dev *mdev = priv->mdev;
3567
3568         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
3569 }
3570
3571 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos,
3572                              __be16 vlan_proto)
3573 {
3574         struct mlx5e_priv *priv = netdev_priv(dev);
3575         struct mlx5_core_dev *mdev = priv->mdev;
3576
3577         if (vlan_proto != htons(ETH_P_8021Q))
3578                 return -EPROTONOSUPPORT;
3579
3580         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
3581                                            vlan, qos);
3582 }
3583
3584 static int mlx5e_set_vf_spoofchk(struct net_device *dev, int vf, bool setting)
3585 {
3586         struct mlx5e_priv *priv = netdev_priv(dev);
3587         struct mlx5_core_dev *mdev = priv->mdev;
3588
3589         return mlx5_eswitch_set_vport_spoofchk(mdev->priv.eswitch, vf + 1, setting);
3590 }
3591
3592 static int mlx5e_set_vf_trust(struct net_device *dev, int vf, bool setting)
3593 {
3594         struct mlx5e_priv *priv = netdev_priv(dev);
3595         struct mlx5_core_dev *mdev = priv->mdev;
3596
3597         return mlx5_eswitch_set_vport_trust(mdev->priv.eswitch, vf + 1, setting);
3598 }
3599
3600 static int mlx5e_set_vf_rate(struct net_device *dev, int vf, int min_tx_rate,
3601                              int max_tx_rate)
3602 {
3603         struct mlx5e_priv *priv = netdev_priv(dev);
3604         struct mlx5_core_dev *mdev = priv->mdev;
3605
3606         return mlx5_eswitch_set_vport_rate(mdev->priv.eswitch, vf + 1,
3607                                            max_tx_rate, min_tx_rate);
3608 }
3609
3610 static int mlx5_vport_link2ifla(u8 esw_link)
3611 {
3612         switch (esw_link) {
3613         case MLX5_ESW_VPORT_ADMIN_STATE_DOWN:
3614                 return IFLA_VF_LINK_STATE_DISABLE;
3615         case MLX5_ESW_VPORT_ADMIN_STATE_UP:
3616                 return IFLA_VF_LINK_STATE_ENABLE;
3617         }
3618         return IFLA_VF_LINK_STATE_AUTO;
3619 }
3620
3621 static int mlx5_ifla_link2vport(u8 ifla_link)
3622 {
3623         switch (ifla_link) {
3624         case IFLA_VF_LINK_STATE_DISABLE:
3625                 return MLX5_ESW_VPORT_ADMIN_STATE_DOWN;
3626         case IFLA_VF_LINK_STATE_ENABLE:
3627                 return MLX5_ESW_VPORT_ADMIN_STATE_UP;
3628         }
3629         return MLX5_ESW_VPORT_ADMIN_STATE_AUTO;
3630 }
3631
3632 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
3633                                    int link_state)
3634 {
3635         struct mlx5e_priv *priv = netdev_priv(dev);
3636         struct mlx5_core_dev *mdev = priv->mdev;
3637
3638         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
3639                                             mlx5_ifla_link2vport(link_state));
3640 }
3641
3642 static int mlx5e_get_vf_config(struct net_device *dev,
3643                                int vf, struct ifla_vf_info *ivi)
3644 {
3645         struct mlx5e_priv *priv = netdev_priv(dev);
3646         struct mlx5_core_dev *mdev = priv->mdev;
3647         int err;
3648
3649         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
3650         if (err)
3651                 return err;
3652         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
3653         return 0;
3654 }
3655
3656 static int mlx5e_get_vf_stats(struct net_device *dev,
3657                               int vf, struct ifla_vf_stats *vf_stats)
3658 {
3659         struct mlx5e_priv *priv = netdev_priv(dev);
3660         struct mlx5_core_dev *mdev = priv->mdev;
3661
3662         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
3663                                             vf_stats);
3664 }
3665 #endif
3666
3667 static void mlx5e_add_vxlan_port(struct net_device *netdev,
3668                                  struct udp_tunnel_info *ti)
3669 {
3670         struct mlx5e_priv *priv = netdev_priv(netdev);
3671
3672         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3673                 return;
3674
3675         if (!mlx5e_vxlan_allowed(priv->mdev))
3676                 return;
3677
3678         mlx5e_vxlan_queue_work(priv, ti->sa_family, be16_to_cpu(ti->port), 1);
3679 }
3680
3681 static void mlx5e_del_vxlan_port(struct net_device *netdev,
3682                                  struct udp_tunnel_info *ti)
3683 {
3684         struct mlx5e_priv *priv = netdev_priv(netdev);
3685
3686         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3687                 return;
3688
3689         if (!mlx5e_vxlan_allowed(priv->mdev))
3690                 return;
3691
3692         mlx5e_vxlan_queue_work(priv, ti->sa_family, be16_to_cpu(ti->port), 0);
3693 }
3694
3695 static netdev_features_t mlx5e_tunnel_features_check(struct mlx5e_priv *priv,
3696                                                      struct sk_buff *skb,
3697                                                      netdev_features_t features)
3698 {
3699         unsigned int offset = 0;
3700         struct udphdr *udph;
3701         u8 proto;
3702         u16 port;
3703
3704         switch (vlan_get_protocol(skb)) {
3705         case htons(ETH_P_IP):
3706                 proto = ip_hdr(skb)->protocol;
3707                 break;
3708         case htons(ETH_P_IPV6):
3709                 proto = ipv6_find_hdr(skb, &offset, -1, NULL, NULL);
3710                 break;
3711         default:
3712                 goto out;
3713         }
3714
3715         switch (proto) {
3716         case IPPROTO_GRE:
3717                 return features;
3718         case IPPROTO_UDP:
3719                 udph = udp_hdr(skb);
3720                 port = be16_to_cpu(udph->dest);
3721
3722                 /* Verify if UDP port is being offloaded by HW */
3723                 if (mlx5e_vxlan_lookup_port(priv, port))
3724                         return features;
3725         }
3726
3727 out:
3728         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
3729         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3730 }
3731
3732 static netdev_features_t mlx5e_features_check(struct sk_buff *skb,
3733                                               struct net_device *netdev,
3734                                               netdev_features_t features)
3735 {
3736         struct mlx5e_priv *priv = netdev_priv(netdev);
3737
3738         features = vlan_features_check(skb, features);
3739         features = vxlan_features_check(skb, features);
3740
3741 #ifdef CONFIG_MLX5_EN_IPSEC
3742         if (mlx5e_ipsec_feature_check(skb, netdev, features))
3743                 return features;
3744 #endif
3745
3746         /* Validate if the tunneled packet is being offloaded by HW */
3747         if (skb->encapsulation &&
3748             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
3749                 return mlx5e_tunnel_features_check(priv, skb, features);
3750
3751         return features;
3752 }
3753
3754 static void mlx5e_tx_timeout(struct net_device *dev)
3755 {
3756         struct mlx5e_priv *priv = netdev_priv(dev);
3757         bool sched_work = false;
3758         int i;
3759
3760         netdev_err(dev, "TX timeout detected\n");
3761
3762         for (i = 0; i < priv->channels.num * priv->channels.params.num_tc; i++) {
3763                 struct mlx5e_txqsq *sq = priv->txq2sq[i];
3764
3765                 if (!netif_xmit_stopped(netdev_get_tx_queue(dev, i)))
3766                         continue;
3767                 sched_work = true;
3768                 clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
3769                 netdev_err(dev, "TX timeout on queue: %d, SQ: 0x%x, CQ: 0x%x, SQ Cons: 0x%x SQ Prod: 0x%x\n",
3770                            i, sq->sqn, sq->cq.mcq.cqn, sq->cc, sq->pc);
3771         }
3772
3773         if (sched_work && test_bit(MLX5E_STATE_OPENED, &priv->state))
3774                 schedule_work(&priv->tx_timeout_work);
3775 }
3776
3777 static int mlx5e_xdp_set(struct net_device *netdev, struct bpf_prog *prog)
3778 {
3779         struct mlx5e_priv *priv = netdev_priv(netdev);
3780         struct bpf_prog *old_prog;
3781         int err = 0;
3782         bool reset, was_opened;
3783         int i;
3784
3785         mutex_lock(&priv->state_lock);
3786
3787         if ((netdev->features & NETIF_F_LRO) && prog) {
3788                 netdev_warn(netdev, "can't set XDP while LRO is on, disable LRO first\n");
3789                 err = -EINVAL;
3790                 goto unlock;
3791         }
3792
3793         if ((netdev->features & NETIF_F_HW_ESP) && prog) {
3794                 netdev_warn(netdev, "can't set XDP with IPSec offload\n");
3795                 err = -EINVAL;
3796                 goto unlock;
3797         }
3798
3799         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
3800         /* no need for full reset when exchanging programs */
3801         reset = (!priv->channels.params.xdp_prog || !prog);
3802
3803         if (was_opened && reset)
3804                 mlx5e_close_locked(netdev);
3805         if (was_opened && !reset) {
3806                 /* num_channels is invariant here, so we can take the
3807                  * batched reference right upfront.
3808                  */
3809                 prog = bpf_prog_add(prog, priv->channels.num);
3810                 if (IS_ERR(prog)) {
3811                         err = PTR_ERR(prog);
3812                         goto unlock;
3813                 }
3814         }
3815
3816         /* exchange programs, extra prog reference we got from caller
3817          * as long as we don't fail from this point onwards.
3818          */
3819         old_prog = xchg(&priv->channels.params.xdp_prog, prog);
3820         if (old_prog)
3821                 bpf_prog_put(old_prog);
3822
3823         if (reset) /* change RQ type according to priv->xdp_prog */
3824                 mlx5e_set_rq_params(priv->mdev, &priv->channels.params);
3825
3826         if (was_opened && reset)
3827                 mlx5e_open_locked(netdev);
3828
3829         if (!test_bit(MLX5E_STATE_OPENED, &priv->state) || reset)
3830                 goto unlock;
3831
3832         /* exchanging programs w/o reset, we update ref counts on behalf
3833          * of the channels RQs here.
3834          */
3835         for (i = 0; i < priv->channels.num; i++) {
3836                 struct mlx5e_channel *c = priv->channels.c[i];
3837
3838                 clear_bit(MLX5E_RQ_STATE_ENABLED, &c->rq.state);
3839                 napi_synchronize(&c->napi);
3840                 /* prevent mlx5e_poll_rx_cq from accessing rq->xdp_prog */
3841
3842                 old_prog = xchg(&c->rq.xdp_prog, prog);
3843
3844                 set_bit(MLX5E_RQ_STATE_ENABLED, &c->rq.state);
3845                 /* napi_schedule in case we have missed anything */
3846                 napi_schedule(&c->napi);
3847
3848                 if (old_prog)
3849                         bpf_prog_put(old_prog);
3850         }
3851
3852 unlock:
3853         mutex_unlock(&priv->state_lock);
3854         return err;
3855 }
3856
3857 static u32 mlx5e_xdp_query(struct net_device *dev)
3858 {
3859         struct mlx5e_priv *priv = netdev_priv(dev);
3860         const struct bpf_prog *xdp_prog;
3861         u32 prog_id = 0;
3862
3863         mutex_lock(&priv->state_lock);
3864         xdp_prog = priv->channels.params.xdp_prog;
3865         if (xdp_prog)
3866                 prog_id = xdp_prog->aux->id;
3867         mutex_unlock(&priv->state_lock);
3868
3869         return prog_id;
3870 }
3871
3872 static int mlx5e_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3873 {
3874         switch (xdp->command) {
3875         case XDP_SETUP_PROG:
3876                 return mlx5e_xdp_set(dev, xdp->prog);
3877         case XDP_QUERY_PROG:
3878                 xdp->prog_id = mlx5e_xdp_query(dev);
3879                 xdp->prog_attached = !!xdp->prog_id;
3880                 return 0;
3881         default:
3882                 return -EINVAL;
3883         }
3884 }
3885
3886 #ifdef CONFIG_NET_POLL_CONTROLLER
3887 /* Fake "interrupt" called by netpoll (eg netconsole) to send skbs without
3888  * reenabling interrupts.
3889  */
3890 static void mlx5e_netpoll(struct net_device *dev)
3891 {
3892         struct mlx5e_priv *priv = netdev_priv(dev);
3893         struct mlx5e_channels *chs = &priv->channels;
3894
3895         int i;
3896
3897         for (i = 0; i < chs->num; i++)
3898                 napi_schedule(&chs->c[i]->napi);
3899 }
3900 #endif
3901
3902 static const struct net_device_ops mlx5e_netdev_ops = {
3903         .ndo_open                = mlx5e_open,
3904         .ndo_stop                = mlx5e_close,
3905         .ndo_start_xmit          = mlx5e_xmit,
3906         .ndo_setup_tc            = mlx5e_setup_tc,
3907         .ndo_select_queue        = mlx5e_select_queue,
3908         .ndo_get_stats64         = mlx5e_get_stats,
3909         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
3910         .ndo_set_mac_address     = mlx5e_set_mac,
3911         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
3912         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
3913         .ndo_set_features        = mlx5e_set_features,
3914         .ndo_fix_features        = mlx5e_fix_features,
3915         .ndo_change_mtu          = mlx5e_change_mtu,
3916         .ndo_do_ioctl            = mlx5e_ioctl,
3917         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
3918         .ndo_udp_tunnel_add      = mlx5e_add_vxlan_port,
3919         .ndo_udp_tunnel_del      = mlx5e_del_vxlan_port,
3920         .ndo_features_check      = mlx5e_features_check,
3921 #ifdef CONFIG_RFS_ACCEL
3922         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
3923 #endif
3924         .ndo_tx_timeout          = mlx5e_tx_timeout,
3925         .ndo_bpf                 = mlx5e_xdp,
3926 #ifdef CONFIG_NET_POLL_CONTROLLER
3927         .ndo_poll_controller     = mlx5e_netpoll,
3928 #endif
3929 #ifdef CONFIG_MLX5_ESWITCH
3930         /* SRIOV E-Switch NDOs */
3931         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
3932         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
3933         .ndo_set_vf_spoofchk     = mlx5e_set_vf_spoofchk,
3934         .ndo_set_vf_trust        = mlx5e_set_vf_trust,
3935         .ndo_set_vf_rate         = mlx5e_set_vf_rate,
3936         .ndo_get_vf_config       = mlx5e_get_vf_config,
3937         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
3938         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
3939         .ndo_has_offload_stats   = mlx5e_has_offload_stats,
3940         .ndo_get_offload_stats   = mlx5e_get_offload_stats,
3941 #endif
3942 };
3943
3944 static int mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
3945 {
3946         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
3947                 return -EOPNOTSUPP;
3948         if (!MLX5_CAP_GEN(mdev, eth_net_offloads) ||
3949             !MLX5_CAP_GEN(mdev, nic_flow_table) ||
3950             !MLX5_CAP_ETH(mdev, csum_cap) ||
3951             !MLX5_CAP_ETH(mdev, max_lso_cap) ||
3952             !MLX5_CAP_ETH(mdev, vlan_cap) ||
3953             !MLX5_CAP_ETH(mdev, rss_ind_tbl_cap) ||
3954             MLX5_CAP_FLOWTABLE(mdev,
3955                                flow_table_properties_nic_receive.max_ft_level)
3956                                < 3) {
3957                 mlx5_core_warn(mdev,
3958                                "Not creating net device, some required device capabilities are missing\n");
3959                 return -EOPNOTSUPP;
3960         }
3961         if (!MLX5_CAP_ETH(mdev, self_lb_en_modifiable))
3962                 mlx5_core_warn(mdev, "Self loop back prevention is not supported\n");
3963         if (!MLX5_CAP_GEN(mdev, cq_moderation))
3964                 mlx5_core_warn(mdev, "CQ moderation is not supported\n");
3965
3966         return 0;
3967 }
3968
3969 u16 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev)
3970 {
3971         int bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
3972
3973         return bf_buf_size -
3974                sizeof(struct mlx5e_tx_wqe) +
3975                2 /*sizeof(mlx5e_tx_wqe.inline_hdr_start)*/;
3976 }
3977
3978 void mlx5e_build_default_indir_rqt(u32 *indirection_rqt, int len,
3979                                    int num_channels)
3980 {
3981         int i;
3982
3983         for (i = 0; i < len; i++)
3984                 indirection_rqt[i] = i % num_channels;
3985 }
3986
3987 static int mlx5e_get_pci_bw(struct mlx5_core_dev *mdev, u32 *pci_bw)
3988 {
3989         enum pcie_link_width width;
3990         enum pci_bus_speed speed;
3991         int err = 0;
3992
3993         err = pcie_get_minimum_link(mdev->pdev, &speed, &width);
3994         if (err)
3995                 return err;
3996
3997         if (speed == PCI_SPEED_UNKNOWN || width == PCIE_LNK_WIDTH_UNKNOWN)
3998                 return -EINVAL;
3999
4000         switch (speed) {
4001         case PCIE_SPEED_2_5GT:
4002                 *pci_bw = 2500 * width;
4003                 break;
4004         case PCIE_SPEED_5_0GT:
4005                 *pci_bw = 5000 * width;
4006                 break;
4007         case PCIE_SPEED_8_0GT:
4008                 *pci_bw = 8000 * width;
4009                 break;
4010         default:
4011                 return -EINVAL;
4012         }
4013
4014         return 0;
4015 }
4016
4017 static bool cqe_compress_heuristic(u32 link_speed, u32 pci_bw)
4018 {
4019         return (link_speed && pci_bw &&
4020                 (pci_bw < 40000) && (pci_bw < link_speed));
4021 }
4022
4023 static bool hw_lro_heuristic(u32 link_speed, u32 pci_bw)
4024 {
4025         return !(link_speed && pci_bw &&
4026                  (pci_bw <= 16000) && (pci_bw < link_speed));
4027 }
4028
4029 void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4030 {
4031         params->tx_cq_moderation.cq_period_mode = cq_period_mode;
4032
4033         params->tx_cq_moderation.pkts =
4034                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
4035         params->tx_cq_moderation.usec =
4036                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
4037
4038         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4039                 params->tx_cq_moderation.usec =
4040                         MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE;
4041
4042         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_TX_CQE_BASED_MODER,
4043                         params->tx_cq_moderation.cq_period_mode ==
4044                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4045 }
4046
4047 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4048 {
4049         params->rx_cq_moderation.cq_period_mode = cq_period_mode;
4050
4051         params->rx_cq_moderation.pkts =
4052                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
4053         params->rx_cq_moderation.usec =
4054                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
4055
4056         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4057                 params->rx_cq_moderation.usec =
4058                         MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE;
4059
4060         if (params->rx_dim_enabled) {
4061                 switch (cq_period_mode) {
4062                 case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
4063                         params->rx_cq_moderation =
4064                                 net_dim_get_def_profile(NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE);
4065                         break;
4066                 case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
4067                 default:
4068                         params->rx_cq_moderation =
4069                                 net_dim_get_def_profile(NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE);
4070                 }
4071         }
4072
4073         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_BASED_MODER,
4074                         params->rx_cq_moderation.cq_period_mode ==
4075                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4076 }
4077
4078 u32 mlx5e_choose_lro_timeout(struct mlx5_core_dev *mdev, u32 wanted_timeout)
4079 {
4080         int i;
4081
4082         /* The supported periods are organized in ascending order */
4083         for (i = 0; i < MLX5E_LRO_TIMEOUT_ARR_SIZE - 1; i++)
4084                 if (MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]) >= wanted_timeout)
4085                         break;
4086
4087         return MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]);
4088 }
4089
4090 void mlx5e_build_nic_params(struct mlx5_core_dev *mdev,
4091                             struct mlx5e_params *params,
4092                             u16 max_channels)
4093 {
4094         u8 cq_period_mode = 0;
4095         u32 link_speed = 0;
4096         u32 pci_bw = 0;
4097
4098         params->num_channels = max_channels;
4099         params->num_tc       = 1;
4100
4101         mlx5e_get_max_linkspeed(mdev, &link_speed);
4102         mlx5e_get_pci_bw(mdev, &pci_bw);
4103         mlx5_core_dbg(mdev, "Max link speed = %d, PCI BW = %d\n",
4104                       link_speed, pci_bw);
4105
4106         /* SQ */
4107         params->log_sq_size = is_kdump_kernel() ?
4108                 MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE :
4109                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
4110
4111         /* set CQE compression */
4112         params->rx_cqe_compress_def = false;
4113         if (MLX5_CAP_GEN(mdev, cqe_compression) &&
4114             MLX5_CAP_GEN(mdev, vport_group_manager))
4115                 params->rx_cqe_compress_def = cqe_compress_heuristic(link_speed, pci_bw);
4116
4117         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS, params->rx_cqe_compress_def);
4118
4119         /* RQ */
4120         mlx5e_set_rq_params(mdev, params);
4121
4122         /* HW LRO */
4123
4124         /* TODO: && MLX5_CAP_ETH(mdev, lro_cap) */
4125         if (params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
4126                 params->lro_en = hw_lro_heuristic(link_speed, pci_bw);
4127         params->lro_timeout = mlx5e_choose_lro_timeout(mdev, MLX5E_DEFAULT_LRO_TIMEOUT);
4128
4129         /* CQ moderation params */
4130         cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
4131                         MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
4132                         MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
4133         params->rx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
4134         mlx5e_set_rx_cq_mode_params(params, cq_period_mode);
4135         mlx5e_set_tx_cq_mode_params(params, cq_period_mode);
4136
4137         /* TX inline */
4138         params->tx_max_inline = mlx5e_get_max_inline_cap(mdev);
4139         params->tx_min_inline_mode = mlx5e_params_calculate_tx_min_inline(mdev);
4140
4141         /* RSS */
4142         params->rss_hfunc = ETH_RSS_HASH_XOR;
4143         netdev_rss_key_fill(params->toeplitz_hash_key, sizeof(params->toeplitz_hash_key));
4144         mlx5e_build_default_indir_rqt(params->indirection_rqt,
4145                                       MLX5E_INDIR_RQT_SIZE, max_channels);
4146 }
4147
4148 static void mlx5e_build_nic_netdev_priv(struct mlx5_core_dev *mdev,
4149                                         struct net_device *netdev,
4150                                         const struct mlx5e_profile *profile,
4151                                         void *ppriv)
4152 {
4153         struct mlx5e_priv *priv = netdev_priv(netdev);
4154
4155         priv->mdev        = mdev;
4156         priv->netdev      = netdev;
4157         priv->profile     = profile;
4158         priv->ppriv       = ppriv;
4159         priv->msglevel    = MLX5E_MSG_LEVEL;
4160         priv->hard_mtu = MLX5E_ETH_HARD_MTU;
4161
4162         mlx5e_build_nic_params(mdev, &priv->channels.params, profile->max_nch(mdev));
4163
4164         mutex_init(&priv->state_lock);
4165
4166         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
4167         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
4168         INIT_WORK(&priv->tx_timeout_work, mlx5e_tx_timeout_work);
4169         INIT_DELAYED_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
4170 }
4171
4172 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
4173 {
4174         struct mlx5e_priv *priv = netdev_priv(netdev);
4175
4176         mlx5_query_nic_vport_mac_address(priv->mdev, 0, netdev->dev_addr);
4177         if (is_zero_ether_addr(netdev->dev_addr) &&
4178             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
4179                 eth_hw_addr_random(netdev);
4180                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
4181         }
4182 }
4183
4184 #if IS_ENABLED(CONFIG_NET_SWITCHDEV) && IS_ENABLED(CONFIG_MLX5_ESWITCH)
4185 static const struct switchdev_ops mlx5e_switchdev_ops = {
4186         .switchdev_port_attr_get        = mlx5e_attr_get,
4187 };
4188 #endif
4189
4190 static void mlx5e_build_nic_netdev(struct net_device *netdev)
4191 {
4192         struct mlx5e_priv *priv = netdev_priv(netdev);
4193         struct mlx5_core_dev *mdev = priv->mdev;
4194         bool fcs_supported;
4195         bool fcs_enabled;
4196
4197         SET_NETDEV_DEV(netdev, &mdev->pdev->dev);
4198
4199         netdev->netdev_ops = &mlx5e_netdev_ops;
4200
4201 #ifdef CONFIG_MLX5_CORE_EN_DCB
4202         if (MLX5_CAP_GEN(mdev, vport_group_manager) && MLX5_CAP_GEN(mdev, qos))
4203                 netdev->dcbnl_ops = &mlx5e_dcbnl_ops;
4204 #endif
4205
4206         netdev->watchdog_timeo    = 15 * HZ;
4207
4208         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
4209
4210         netdev->vlan_features    |= NETIF_F_SG;
4211         netdev->vlan_features    |= NETIF_F_IP_CSUM;
4212         netdev->vlan_features    |= NETIF_F_IPV6_CSUM;
4213         netdev->vlan_features    |= NETIF_F_GRO;
4214         netdev->vlan_features    |= NETIF_F_TSO;
4215         netdev->vlan_features    |= NETIF_F_TSO6;
4216         netdev->vlan_features    |= NETIF_F_RXCSUM;
4217         netdev->vlan_features    |= NETIF_F_RXHASH;
4218
4219         if (!!MLX5_CAP_ETH(mdev, lro_cap))
4220                 netdev->vlan_features    |= NETIF_F_LRO;
4221
4222         netdev->hw_features       = netdev->vlan_features;
4223         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
4224         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
4225         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
4226         netdev->hw_features      |= NETIF_F_HW_VLAN_STAG_TX;
4227
4228         if (mlx5e_vxlan_allowed(mdev) || MLX5_CAP_ETH(mdev, tunnel_stateless_gre)) {
4229                 netdev->hw_features     |= NETIF_F_GSO_PARTIAL;
4230                 netdev->hw_enc_features |= NETIF_F_IP_CSUM;
4231                 netdev->hw_enc_features |= NETIF_F_IPV6_CSUM;
4232                 netdev->hw_enc_features |= NETIF_F_TSO;
4233                 netdev->hw_enc_features |= NETIF_F_TSO6;
4234                 netdev->hw_enc_features |= NETIF_F_GSO_PARTIAL;
4235         }
4236
4237         if (mlx5e_vxlan_allowed(mdev)) {
4238                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL |
4239                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4240                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL |
4241                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4242                 netdev->gso_partial_features = NETIF_F_GSO_UDP_TUNNEL_CSUM;
4243         }
4244
4245         if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre)) {
4246                 netdev->hw_features     |= NETIF_F_GSO_GRE |
4247                                            NETIF_F_GSO_GRE_CSUM;
4248                 netdev->hw_enc_features |= NETIF_F_GSO_GRE |
4249                                            NETIF_F_GSO_GRE_CSUM;
4250                 netdev->gso_partial_features |= NETIF_F_GSO_GRE |
4251                                                 NETIF_F_GSO_GRE_CSUM;
4252         }
4253
4254         mlx5_query_port_fcs(mdev, &fcs_supported, &fcs_enabled);
4255
4256         if (fcs_supported)
4257                 netdev->hw_features |= NETIF_F_RXALL;
4258
4259         if (MLX5_CAP_ETH(mdev, scatter_fcs))
4260                 netdev->hw_features |= NETIF_F_RXFCS;
4261
4262         netdev->features          = netdev->hw_features;
4263         if (!priv->channels.params.lro_en)
4264                 netdev->features  &= ~NETIF_F_LRO;
4265
4266         if (fcs_enabled)
4267                 netdev->features  &= ~NETIF_F_RXALL;
4268
4269         if (!priv->channels.params.scatter_fcs_en)
4270                 netdev->features  &= ~NETIF_F_RXFCS;
4271
4272 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
4273         if (FT_CAP(flow_modify_en) &&
4274             FT_CAP(modify_root) &&
4275             FT_CAP(identified_miss_table_mode) &&
4276             FT_CAP(flow_table_modify)) {
4277                 netdev->hw_features      |= NETIF_F_HW_TC;
4278 #ifdef CONFIG_RFS_ACCEL
4279                 netdev->hw_features      |= NETIF_F_NTUPLE;
4280 #endif
4281         }
4282
4283         netdev->features         |= NETIF_F_HIGHDMA;
4284         netdev->features         |= NETIF_F_HW_VLAN_STAG_FILTER;
4285
4286         netdev->priv_flags       |= IFF_UNICAST_FLT;
4287
4288         mlx5e_set_netdev_dev_addr(netdev);
4289
4290 #if IS_ENABLED(CONFIG_NET_SWITCHDEV) && IS_ENABLED(CONFIG_MLX5_ESWITCH)
4291         if (MLX5_VPORT_MANAGER(mdev))
4292                 netdev->switchdev_ops = &mlx5e_switchdev_ops;
4293 #endif
4294
4295         mlx5e_ipsec_build_netdev(priv);
4296 }
4297
4298 static void mlx5e_create_q_counter(struct mlx5e_priv *priv)
4299 {
4300         struct mlx5_core_dev *mdev = priv->mdev;
4301         int err;
4302
4303         err = mlx5_core_alloc_q_counter(mdev, &priv->q_counter);
4304         if (err) {
4305                 mlx5_core_warn(mdev, "alloc queue counter failed, %d\n", err);
4306                 priv->q_counter = 0;
4307         }
4308 }
4309
4310 static void mlx5e_destroy_q_counter(struct mlx5e_priv *priv)
4311 {
4312         if (!priv->q_counter)
4313                 return;
4314
4315         mlx5_core_dealloc_q_counter(priv->mdev, priv->q_counter);
4316 }
4317
4318 static void mlx5e_nic_init(struct mlx5_core_dev *mdev,
4319                            struct net_device *netdev,
4320                            const struct mlx5e_profile *profile,
4321                            void *ppriv)
4322 {
4323         struct mlx5e_priv *priv = netdev_priv(netdev);
4324         int err;
4325
4326         mlx5e_build_nic_netdev_priv(mdev, netdev, profile, ppriv);
4327         err = mlx5e_ipsec_init(priv);
4328         if (err)
4329                 mlx5_core_err(mdev, "IPSec initialization failed, %d\n", err);
4330         mlx5e_build_nic_netdev(netdev);
4331         mlx5e_vxlan_init(priv);
4332 }
4333
4334 static void mlx5e_nic_cleanup(struct mlx5e_priv *priv)
4335 {
4336         mlx5e_ipsec_cleanup(priv);
4337         mlx5e_vxlan_cleanup(priv);
4338 }
4339
4340 static int mlx5e_init_nic_rx(struct mlx5e_priv *priv)
4341 {
4342         struct mlx5_core_dev *mdev = priv->mdev;
4343         int err;
4344
4345         err = mlx5e_create_indirect_rqt(priv);
4346         if (err)
4347                 return err;
4348
4349         err = mlx5e_create_direct_rqts(priv);
4350         if (err)
4351                 goto err_destroy_indirect_rqts;
4352
4353         err = mlx5e_create_indirect_tirs(priv);
4354         if (err)
4355                 goto err_destroy_direct_rqts;
4356
4357         err = mlx5e_create_direct_tirs(priv);
4358         if (err)
4359                 goto err_destroy_indirect_tirs;
4360
4361         err = mlx5e_create_flow_steering(priv);
4362         if (err) {
4363                 mlx5_core_warn(mdev, "create flow steering failed, %d\n", err);
4364                 goto err_destroy_direct_tirs;
4365         }
4366
4367         err = mlx5e_tc_init(priv);
4368         if (err)
4369                 goto err_destroy_flow_steering;
4370
4371         return 0;
4372
4373 err_destroy_flow_steering:
4374         mlx5e_destroy_flow_steering(priv);
4375 err_destroy_direct_tirs:
4376         mlx5e_destroy_direct_tirs(priv);
4377 err_destroy_indirect_tirs:
4378         mlx5e_destroy_indirect_tirs(priv);
4379 err_destroy_direct_rqts:
4380         mlx5e_destroy_direct_rqts(priv);
4381 err_destroy_indirect_rqts:
4382         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
4383         return err;
4384 }
4385
4386 static void mlx5e_cleanup_nic_rx(struct mlx5e_priv *priv)
4387 {
4388         mlx5e_tc_cleanup(priv);
4389         mlx5e_destroy_flow_steering(priv);
4390         mlx5e_destroy_direct_tirs(priv);
4391         mlx5e_destroy_indirect_tirs(priv);
4392         mlx5e_destroy_direct_rqts(priv);
4393         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
4394 }
4395
4396 static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
4397 {
4398         int err;
4399
4400         err = mlx5e_create_tises(priv);
4401         if (err) {
4402                 mlx5_core_warn(priv->mdev, "create tises failed, %d\n", err);
4403                 return err;
4404         }
4405
4406 #ifdef CONFIG_MLX5_CORE_EN_DCB
4407         mlx5e_dcbnl_initialize(priv);
4408 #endif
4409         return 0;
4410 }
4411
4412 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
4413 {
4414         struct net_device *netdev = priv->netdev;
4415         struct mlx5_core_dev *mdev = priv->mdev;
4416         u16 max_mtu;
4417
4418         mlx5e_init_l2_addr(priv);
4419
4420         /* Marking the link as currently not needed by the Driver */
4421         if (!netif_running(netdev))
4422                 mlx5_set_port_admin_status(mdev, MLX5_PORT_DOWN);
4423
4424         /* MTU range: 68 - hw-specific max */
4425         netdev->min_mtu = ETH_MIN_MTU;
4426         mlx5_query_port_max_mtu(priv->mdev, &max_mtu, 1);
4427         netdev->max_mtu = MLX5E_HW2SW_MTU(priv, max_mtu);
4428         mlx5e_set_dev_port_mtu(priv);
4429
4430         mlx5_lag_add(mdev, netdev);
4431
4432         mlx5e_enable_async_events(priv);
4433
4434         if (MLX5_VPORT_MANAGER(priv->mdev))
4435                 mlx5e_register_vport_reps(priv);
4436
4437         if (netdev->reg_state != NETREG_REGISTERED)
4438                 return;
4439 #ifdef CONFIG_MLX5_CORE_EN_DCB
4440         mlx5e_dcbnl_init_app(priv);
4441 #endif
4442         /* Device already registered: sync netdev system state */
4443         if (mlx5e_vxlan_allowed(mdev)) {
4444                 rtnl_lock();
4445                 udp_tunnel_get_rx_info(netdev);
4446                 rtnl_unlock();
4447         }
4448
4449         queue_work(priv->wq, &priv->set_rx_mode_work);
4450
4451         rtnl_lock();
4452         if (netif_running(netdev))
4453                 mlx5e_open(netdev);
4454         netif_device_attach(netdev);
4455         rtnl_unlock();
4456 }
4457
4458 static void mlx5e_nic_disable(struct mlx5e_priv *priv)
4459 {
4460         struct mlx5_core_dev *mdev = priv->mdev;
4461
4462 #ifdef CONFIG_MLX5_CORE_EN_DCB
4463         if (priv->netdev->reg_state == NETREG_REGISTERED)
4464                 mlx5e_dcbnl_delete_app(priv);
4465 #endif
4466
4467         rtnl_lock();
4468         if (netif_running(priv->netdev))
4469                 mlx5e_close(priv->netdev);
4470         netif_device_detach(priv->netdev);
4471         rtnl_unlock();
4472
4473         queue_work(priv->wq, &priv->set_rx_mode_work);
4474
4475         if (MLX5_VPORT_MANAGER(priv->mdev))
4476                 mlx5e_unregister_vport_reps(priv);
4477
4478         mlx5e_disable_async_events(priv);
4479         mlx5_lag_remove(mdev);
4480 }
4481
4482 static const struct mlx5e_profile mlx5e_nic_profile = {
4483         .init              = mlx5e_nic_init,
4484         .cleanup           = mlx5e_nic_cleanup,
4485         .init_rx           = mlx5e_init_nic_rx,
4486         .cleanup_rx        = mlx5e_cleanup_nic_rx,
4487         .init_tx           = mlx5e_init_nic_tx,
4488         .cleanup_tx        = mlx5e_cleanup_nic_tx,
4489         .enable            = mlx5e_nic_enable,
4490         .disable           = mlx5e_nic_disable,
4491         .update_stats      = mlx5e_update_ndo_stats,
4492         .max_nch           = mlx5e_get_max_num_channels,
4493         .update_carrier    = mlx5e_update_carrier,
4494         .rx_handlers.handle_rx_cqe       = mlx5e_handle_rx_cqe,
4495         .rx_handlers.handle_rx_cqe_mpwqe = mlx5e_handle_rx_cqe_mpwrq,
4496         .max_tc            = MLX5E_MAX_NUM_TC,
4497 };
4498
4499 /* mlx5e generic netdev management API (move to en_common.c) */
4500
4501 struct net_device *mlx5e_create_netdev(struct mlx5_core_dev *mdev,
4502                                        const struct mlx5e_profile *profile,
4503                                        void *ppriv)
4504 {
4505         int nch = profile->max_nch(mdev);
4506         struct net_device *netdev;
4507         struct mlx5e_priv *priv;
4508
4509         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv),
4510                                     nch * profile->max_tc,
4511                                     nch);
4512         if (!netdev) {
4513                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
4514                 return NULL;
4515         }
4516
4517 #ifdef CONFIG_RFS_ACCEL
4518         netdev->rx_cpu_rmap = mdev->rmap;
4519 #endif
4520
4521         profile->init(mdev, netdev, profile, ppriv);
4522
4523         netif_carrier_off(netdev);
4524
4525         priv = netdev_priv(netdev);
4526
4527         priv->wq = create_singlethread_workqueue("mlx5e");
4528         if (!priv->wq)
4529                 goto err_cleanup_nic;
4530
4531         return netdev;
4532
4533 err_cleanup_nic:
4534         if (profile->cleanup)
4535                 profile->cleanup(priv);
4536         free_netdev(netdev);
4537
4538         return NULL;
4539 }
4540
4541 int mlx5e_attach_netdev(struct mlx5e_priv *priv)
4542 {
4543         struct mlx5_core_dev *mdev = priv->mdev;
4544         const struct mlx5e_profile *profile;
4545         int err;
4546
4547         profile = priv->profile;
4548         clear_bit(MLX5E_STATE_DESTROYING, &priv->state);
4549
4550         err = profile->init_tx(priv);
4551         if (err)
4552                 goto out;
4553
4554         err = mlx5e_open_drop_rq(mdev, &priv->drop_rq);
4555         if (err) {
4556                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
4557                 goto err_cleanup_tx;
4558         }
4559
4560         err = profile->init_rx(priv);
4561         if (err)
4562                 goto err_close_drop_rq;
4563
4564         mlx5e_create_q_counter(priv);
4565
4566         if (profile->enable)
4567                 profile->enable(priv);
4568
4569         return 0;
4570
4571 err_close_drop_rq:
4572         mlx5e_close_drop_rq(&priv->drop_rq);
4573
4574 err_cleanup_tx:
4575         profile->cleanup_tx(priv);
4576
4577 out:
4578         return err;
4579 }
4580
4581 void mlx5e_detach_netdev(struct mlx5e_priv *priv)
4582 {
4583         const struct mlx5e_profile *profile = priv->profile;
4584
4585         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
4586
4587         if (profile->disable)
4588                 profile->disable(priv);
4589         flush_workqueue(priv->wq);
4590
4591         mlx5e_destroy_q_counter(priv);
4592         profile->cleanup_rx(priv);
4593         mlx5e_close_drop_rq(&priv->drop_rq);
4594         profile->cleanup_tx(priv);
4595         cancel_delayed_work_sync(&priv->update_stats_work);
4596 }
4597
4598 void mlx5e_destroy_netdev(struct mlx5e_priv *priv)
4599 {
4600         const struct mlx5e_profile *profile = priv->profile;
4601         struct net_device *netdev = priv->netdev;
4602
4603         destroy_workqueue(priv->wq);
4604         if (profile->cleanup)
4605                 profile->cleanup(priv);
4606         free_netdev(netdev);
4607 }
4608
4609 /* mlx5e_attach and mlx5e_detach scope should be only creating/destroying
4610  * hardware contexts and to connect it to the current netdev.
4611  */
4612 static int mlx5e_attach(struct mlx5_core_dev *mdev, void *vpriv)
4613 {
4614         struct mlx5e_priv *priv = vpriv;
4615         struct net_device *netdev = priv->netdev;
4616         int err;
4617
4618         if (netif_device_present(netdev))
4619                 return 0;
4620
4621         err = mlx5e_create_mdev_resources(mdev);
4622         if (err)
4623                 return err;
4624
4625         err = mlx5e_attach_netdev(priv);
4626         if (err) {
4627                 mlx5e_destroy_mdev_resources(mdev);
4628                 return err;
4629         }
4630
4631         return 0;
4632 }
4633
4634 static void mlx5e_detach(struct mlx5_core_dev *mdev, void *vpriv)
4635 {
4636         struct mlx5e_priv *priv = vpriv;
4637         struct net_device *netdev = priv->netdev;
4638
4639         if (!netif_device_present(netdev))
4640                 return;
4641
4642         mlx5e_detach_netdev(priv);
4643         mlx5e_destroy_mdev_resources(mdev);
4644 }
4645
4646 static void *mlx5e_add(struct mlx5_core_dev *mdev)
4647 {
4648         struct net_device *netdev;
4649         void *rpriv = NULL;
4650         void *priv;
4651         int err;
4652
4653         err = mlx5e_check_required_hca_cap(mdev);
4654         if (err)
4655                 return NULL;
4656
4657 #ifdef CONFIG_MLX5_ESWITCH
4658         if (MLX5_VPORT_MANAGER(mdev)) {
4659                 rpriv = mlx5e_alloc_nic_rep_priv(mdev);
4660                 if (!rpriv) {
4661                         mlx5_core_warn(mdev, "Failed to alloc NIC rep priv data\n");
4662                         return NULL;
4663                 }
4664         }
4665 #endif
4666
4667         netdev = mlx5e_create_netdev(mdev, &mlx5e_nic_profile, rpriv);
4668         if (!netdev) {
4669                 mlx5_core_err(mdev, "mlx5e_create_netdev failed\n");
4670                 goto err_free_rpriv;
4671         }
4672
4673         priv = netdev_priv(netdev);
4674
4675         err = mlx5e_attach(mdev, priv);
4676         if (err) {
4677                 mlx5_core_err(mdev, "mlx5e_attach failed, %d\n", err);
4678                 goto err_destroy_netdev;
4679         }
4680
4681         err = register_netdev(netdev);
4682         if (err) {
4683                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
4684                 goto err_detach;
4685         }
4686
4687 #ifdef CONFIG_MLX5_CORE_EN_DCB
4688         mlx5e_dcbnl_init_app(priv);
4689 #endif
4690         return priv;
4691
4692 err_detach:
4693         mlx5e_detach(mdev, priv);
4694 err_destroy_netdev:
4695         mlx5e_destroy_netdev(priv);
4696 err_free_rpriv:
4697         kfree(rpriv);
4698         return NULL;
4699 }
4700
4701 static void mlx5e_remove(struct mlx5_core_dev *mdev, void *vpriv)
4702 {
4703         struct mlx5e_priv *priv = vpriv;
4704         void *ppriv = priv->ppriv;
4705
4706 #ifdef CONFIG_MLX5_CORE_EN_DCB
4707         mlx5e_dcbnl_delete_app(priv);
4708 #endif
4709         unregister_netdev(priv->netdev);
4710         mlx5e_detach(mdev, vpriv);
4711         mlx5e_destroy_netdev(priv);
4712         kfree(ppriv);
4713 }
4714
4715 static void *mlx5e_get_netdev(void *vpriv)
4716 {
4717         struct mlx5e_priv *priv = vpriv;
4718
4719         return priv->netdev;
4720 }
4721
4722 static struct mlx5_interface mlx5e_interface = {
4723         .add       = mlx5e_add,
4724         .remove    = mlx5e_remove,
4725         .attach    = mlx5e_attach,
4726         .detach    = mlx5e_detach,
4727         .event     = mlx5e_async_event,
4728         .protocol  = MLX5_INTERFACE_PROTOCOL_ETH,
4729         .get_dev   = mlx5e_get_netdev,
4730 };
4731
4732 void mlx5e_init(void)
4733 {
4734         mlx5e_ipsec_build_inverse_table();
4735         mlx5e_build_ptys2ethtool_map();
4736         mlx5_register_interface(&mlx5e_interface);
4737 }
4738
4739 void mlx5e_cleanup(void)
4740 {
4741         mlx5_unregister_interface(&mlx5e_interface);
4742 }