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         rq->buff.map_dir = rq->xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
 586         rq->buff.headroom = params->rq_headroom;
 587
 588         switch (rq->wq_type) {
 589         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 590
 591                 rq->post_wqes = mlx5e_post_rx_mpwqes;
 592                 rq->dealloc_wqe = mlx5e_dealloc_rx_mpwqe;
 593
 594                 rq->handle_rx_cqe = c->priv->profile->rx_handlers.handle_rx_cqe_mpwqe;
 595 #ifdef CONFIG_MLX5_EN_IPSEC
 596                 if (MLX5_IPSEC_DEV(mdev)) {
 597                         err = -EINVAL;
 598                         netdev_err(c->netdev, "MPWQE RQ with IPSec offload not supported\n");
 599                         goto err_rq_wq_destroy;
 600                 }
 601 #endif
 602                 if (!rq->handle_rx_cqe) {
 603                         err = -EINVAL;
 604                         netdev_err(c->netdev, "RX handler of MPWQE RQ is not set, err %d\n", err);
 605                         goto err_rq_wq_destroy;
 606                 }
 607
 608                 rq->mpwqe.log_stride_sz = params->mpwqe_log_stride_sz;
 609                 rq->mpwqe.num_strides = BIT(params->mpwqe_log_num_strides);
 610
 611                 byte_count = rq->mpwqe.num_strides << rq->mpwqe.log_stride_sz;
 612
 613                 err = mlx5e_create_rq_umr_mkey(mdev, rq);
 614                 if (err)
 615                         goto err_rq_wq_destroy;
 616                 rq->mkey_be = cpu_to_be32(rq->umr_mkey.key);
 617
 618                 err = mlx5e_rq_alloc_mpwqe_info(rq, c);
 619                 if (err)
 620                         goto err_destroy_umr_mkey;
 621                 break;
 622         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 623                 rq->wqe.frag_info =
 624                         kzalloc_node(wq_sz * sizeof(*rq->wqe.frag_info),
 625                                      GFP_KERNEL, cpu_to_node(c->cpu));
 626                 if (!rq->wqe.frag_info) {
 627                         err = -ENOMEM;
 628                         goto err_rq_wq_destroy;
 629                 }
 630                 rq->post_wqes = mlx5e_post_rx_wqes;
 631                 rq->dealloc_wqe = mlx5e_dealloc_rx_wqe;
 632
 633 #ifdef CONFIG_MLX5_EN_IPSEC
 634                 if (c->priv->ipsec)
 635                         rq->handle_rx_cqe = mlx5e_ipsec_handle_rx_cqe;
 636                 else
 637 #endif
 638                         rq->handle_rx_cqe = c->priv->profile->rx_handlers.handle_rx_cqe;
 639                 if (!rq->handle_rx_cqe) {
 640                         kfree(rq->wqe.frag_info);
 641                         err = -EINVAL;
 642                         netdev_err(c->netdev, "RX handler of RQ is not set, err %d\n", err);
 643                         goto err_rq_wq_destroy;
 644                 }
 645
 646                 byte_count = params->lro_en  ?
 647                                 params->lro_wqe_sz :
 648                                 MLX5E_SW2HW_MTU(c->priv, c->netdev->mtu);
 649 #ifdef CONFIG_MLX5_EN_IPSEC
 650                 if (MLX5_IPSEC_DEV(mdev))
 651                         byte_count += MLX5E_METADATA_ETHER_LEN;
 652 #endif
 653                 rq->wqe.page_reuse = !params->xdp_prog && !params->lro_en;
 654
 655                 /* calc the required page order */
 656                 rq->wqe.frag_sz = MLX5_SKB_FRAG_SZ(rq->buff.headroom + byte_count);
 657                 npages = DIV_ROUND_UP(rq->wqe.frag_sz, PAGE_SIZE);
 658                 rq->buff.page_order = order_base_2(npages);
 659
 660                 byte_count |= MLX5_HW_START_PADDING;
 661                 rq->mkey_be = c->mkey_be;
 662         }
 663
 664         for (i = 0; i < wq_sz; i++) {
 665                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, i);
 666
 667                 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 668                         u64 dma_offset = (u64)mlx5e_get_wqe_mtt_offset(rq, i) << PAGE_SHIFT;
 669
 670                         wqe->data.addr = cpu_to_be64(dma_offset);
 671                 }
 672
 673                 wqe->data.byte_count = cpu_to_be32(byte_count);
 674                 wqe->data.lkey = rq->mkey_be;
 675         }
 676
 677         INIT_WORK(&rq->am.work, mlx5e_rx_am_work);
 678         rq->am.mode = params->rx_cq_moderation.cq_period_mode;
 679         rq->page_cache.head = 0;
 680         rq->page_cache.tail = 0;
 681
 682         return 0;
 683
 684 err_destroy_umr_mkey:
 685         mlx5_core_destroy_mkey(mdev, &rq->umr_mkey);
 686
 687 err_rq_wq_destroy:
 688         if (rq->xdp_prog)
 689                 bpf_prog_put(rq->xdp_prog);
 690         mlx5_wq_destroy(&rq->wq_ctrl);
 691
 692         return err;
 693 }
 694
 695 static void mlx5e_free_rq(struct mlx5e_rq *rq)
 696 {
 697         int i;
 698
 699         if (rq->xdp_prog)
 700                 bpf_prog_put(rq->xdp_prog);
 701
 702         switch (rq->wq_type) {
 703         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 704                 mlx5e_rq_free_mpwqe_info(rq);
 705                 mlx5_core_destroy_mkey(rq->mdev, &rq->umr_mkey);
 706                 break;
 707         default: /* MLX5_WQ_TYPE_LINKED_LIST */
 708                 kfree(rq->wqe.frag_info);
 709         }
 710
 711         for (i = rq->page_cache.head; i != rq->page_cache.tail;
 712              i = (i + 1) & (MLX5E_CACHE_SIZE - 1)) {
 713                 struct mlx5e_dma_info *dma_info = &rq->page_cache.page_cache[i];
 714
 715                 mlx5e_page_release(rq, dma_info, false);
 716         }
 717         mlx5_wq_destroy(&rq->wq_ctrl);
 718 }
 719
 720 static int mlx5e_create_rq(struct mlx5e_rq *rq,
 721                            struct mlx5e_rq_param *param)
 722 {
 723         struct mlx5_core_dev *mdev = rq->mdev;
 724
 725         void *in;
 726         void *rqc;
 727         void *wq;
 728         int inlen;
 729         int err;
 730
 731         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 732                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 733         in = kvzalloc(inlen, GFP_KERNEL);
 734         if (!in)
 735                 return -ENOMEM;
 736
 737         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 738         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 739
 740         memcpy(rqc, param->rqc, sizeof(param->rqc));
 741
 742         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 743         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 744         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 745                                                 MLX5_ADAPTER_PAGE_SHIFT);
 746         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 747
 748         mlx5_fill_page_array(&rq->wq_ctrl.buf,
 749                              (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 750
 751         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 752
 753         kvfree(in);
 754
 755         return err;
 756 }
 757
 758 static int mlx5e_modify_rq_state(struct mlx5e_rq *rq, int curr_state,
 759                                  int next_state)
 760 {
 761         struct mlx5e_channel *c = rq->channel;
 762         struct mlx5_core_dev *mdev = c->mdev;
 763
 764         void *in;
 765         void *rqc;
 766         int inlen;
 767         int err;
 768
 769         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 770         in = kvzalloc(inlen, GFP_KERNEL);
 771         if (!in)
 772                 return -ENOMEM;
 773
 774         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 775
 776         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 777         MLX5_SET(rqc, rqc, state, next_state);
 778
 779         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 780
 781         kvfree(in);
 782
 783         return err;
 784 }
 785
 786 static int mlx5e_modify_rq_scatter_fcs(struct mlx5e_rq *rq, bool enable)
 787 {
 788         struct mlx5e_channel *c = rq->channel;
 789         struct mlx5e_priv *priv = c->priv;
 790         struct mlx5_core_dev *mdev = priv->mdev;
 791
 792         void *in;
 793         void *rqc;
 794         int inlen;
 795         int err;
 796
 797         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 798         in = kvzalloc(inlen, GFP_KERNEL);
 799         if (!in)
 800                 return -ENOMEM;
 801
 802         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 803
 804         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 805         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 806                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_SCATTER_FCS);
 807         MLX5_SET(rqc, rqc, scatter_fcs, enable);
 808         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 809
 810         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 811
 812         kvfree(in);
 813
 814         return err;
 815 }
 816
 817 static int mlx5e_modify_rq_vsd(struct mlx5e_rq *rq, bool vsd)
 818 {
 819         struct mlx5e_channel *c = rq->channel;
 820         struct mlx5_core_dev *mdev = c->mdev;
 821         void *in;
 822         void *rqc;
 823         int inlen;
 824         int err;
 825
 826         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 827         in = kvzalloc(inlen, GFP_KERNEL);
 828         if (!in)
 829                 return -ENOMEM;
 830
 831         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 832
 833         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 834         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 835                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
 836         MLX5_SET(rqc, rqc, vsd, vsd);
 837         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 838
 839         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 840
 841         kvfree(in);
 842
 843         return err;
 844 }
 845
 846 static void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 847 {
 848         mlx5_core_destroy_rq(rq->mdev, rq->rqn);
 849 }
 850
 851 static int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq)
 852 {
 853         unsigned long exp_time = jiffies + msecs_to_jiffies(20000);
 854         struct mlx5e_channel *c = rq->channel;
 855
 856         struct mlx5_wq_ll *wq = &rq->wq;
 857         u16 min_wqes = mlx5_min_rx_wqes(rq->wq_type, mlx5_wq_ll_get_size(wq));
 858
 859         while (time_before(jiffies, exp_time)) {
 860                 if (wq->cur_sz >= min_wqes)
 861                         return 0;
 862
 863                 msleep(20);
 864         }
 865
 866         netdev_warn(c->netdev, "Failed to get min RX wqes on RQN[0x%x] wq cur_sz(%d) min_rx_wqes(%d)\n",
 867                     rq->rqn, wq->cur_sz, min_wqes);
 868         return -ETIMEDOUT;
 869 }
 870
 871 static void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 872 {
 873         struct mlx5_wq_ll *wq = &rq->wq;
 874         struct mlx5e_rx_wqe *wqe;
 875         __be16 wqe_ix_be;
 876         u16 wqe_ix;
 877
 878         /* UMR WQE (if in progress) is always at wq->head */
 879         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ &&
 880             rq->mpwqe.umr_in_progress)
 881                 mlx5e_free_rx_mpwqe(rq, &rq->mpwqe.info[wq->head]);
 882
 883         while (!mlx5_wq_ll_is_empty(wq)) {
 884                 wqe_ix_be = *wq->tail_next;
 885                 wqe_ix    = be16_to_cpu(wqe_ix_be);
 886                 wqe       = mlx5_wq_ll_get_wqe(&rq->wq, wqe_ix);
 887                 rq->dealloc_wqe(rq, wqe_ix);
 888                 mlx5_wq_ll_pop(&rq->wq, wqe_ix_be,
 889                                &wqe->next.next_wqe_index);
 890         }
 891
 892         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST && rq->wqe.page_reuse) {
 893                 /* Clean outstanding pages on handled WQEs that decided to do page-reuse,
 894                  * but yet to be re-posted.
 895                  */
 896                 int wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 897
 898                 for (wqe_ix = 0; wqe_ix < wq_sz; wqe_ix++)
 899                         rq->dealloc_wqe(rq, wqe_ix);
 900         }
 901 }
 902
 903 static int mlx5e_open_rq(struct mlx5e_channel *c,
 904                          struct mlx5e_params *params,
 905                          struct mlx5e_rq_param *param,
 906                          struct mlx5e_rq *rq)
 907 {
 908         int err;
 909
 910         err = mlx5e_alloc_rq(c, params, param, rq);
 911         if (err)
 912                 return err;
 913
 914         err = mlx5e_create_rq(rq, param);
 915         if (err)
 916                 goto err_free_rq;
 917
 918         err = mlx5e_modify_rq_state(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 919         if (err)
 920                 goto err_destroy_rq;
 921
 922         if (params->rx_am_enabled)
 923                 c->rq.state |= BIT(MLX5E_RQ_STATE_AM);
 924
 925         return 0;
 926
 927 err_destroy_rq:
 928         mlx5e_destroy_rq(rq);
 929 err_free_rq:
 930         mlx5e_free_rq(rq);
 931
 932         return err;
 933 }
 934
 935 static void mlx5e_activate_rq(struct mlx5e_rq *rq)
 936 {
 937         struct mlx5e_icosq *sq = &rq->channel->icosq;
 938         u16 pi = sq->pc & sq->wq.sz_m1;
 939         struct mlx5e_tx_wqe *nopwqe;
 940
 941         set_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 942         sq->db.ico_wqe[pi].opcode     = MLX5_OPCODE_NOP;
 943         nopwqe = mlx5e_post_nop(&sq->wq, sq->sqn, &sq->pc);
 944         mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &nopwqe->ctrl);
 945 }
 946
 947 static void mlx5e_deactivate_rq(struct mlx5e_rq *rq)
 948 {
 949         clear_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 950         napi_synchronize(&rq->channel->napi); /* prevent mlx5e_post_rx_wqes */
 951 }
 952
 953 static void mlx5e_close_rq(struct mlx5e_rq *rq)
 954 {
 955         cancel_work_sync(&rq->am.work);
 956         mlx5e_destroy_rq(rq);
 957         mlx5e_free_rx_descs(rq);
 958         mlx5e_free_rq(rq);
 959 }
 960
 961 static void mlx5e_free_xdpsq_db(struct mlx5e_xdpsq *sq)
 962 {
 963         kfree(sq->db.di);
 964 }
 965
 966 static int mlx5e_alloc_xdpsq_db(struct mlx5e_xdpsq *sq, int numa)
 967 {
 968         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 969
 970         sq->db.di = kzalloc_node(sizeof(*sq->db.di) * wq_sz,
 971                                      GFP_KERNEL, numa);
 972         if (!sq->db.di) {
 973                 mlx5e_free_xdpsq_db(sq);
 974                 return -ENOMEM;
 975         }
 976
 977         return 0;
 978 }
 979
 980 static int mlx5e_alloc_xdpsq(struct mlx5e_channel *c,
 981                              struct mlx5e_params *params,
 982                              struct mlx5e_sq_param *param,
 983                              struct mlx5e_xdpsq *sq)
 984 {
 985         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
 986         struct mlx5_core_dev *mdev = c->mdev;
 987         int err;
 988
 989         sq->pdev      = c->pdev;
 990         sq->mkey_be   = c->mkey_be;
 991         sq->channel   = c;
 992         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
 993         sq->min_inline_mode = params->tx_min_inline_mode;
 994
 995         param->wq.db_numa_node = cpu_to_node(c->cpu);
 996         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq, &sq->wq_ctrl);
 997         if (err)
 998                 return err;
 999         sq->wq.db = &sq->wq.db[MLX5_SND_DBR];
1000
1001         err = mlx5e_alloc_xdpsq_db(sq, cpu_to_node(c->cpu));
1002         if (err)
1003                 goto err_sq_wq_destroy;
1004
1005         return 0;
1006
1007 err_sq_wq_destroy:
1008         mlx5_wq_destroy(&sq->wq_ctrl);
1009
1010         return err;
1011 }
1012
1013 static void mlx5e_free_xdpsq(struct mlx5e_xdpsq *sq)
1014 {
1015         mlx5e_free_xdpsq_db(sq);
1016         mlx5_wq_destroy(&sq->wq_ctrl);
1017 }
1018
1019 static void mlx5e_free_icosq_db(struct mlx5e_icosq *sq)
1020 {
1021         kfree(sq->db.ico_wqe);
1022 }
1023
1024 static int mlx5e_alloc_icosq_db(struct mlx5e_icosq *sq, int numa)
1025 {
1026         u8 wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1027
1028         sq->db.ico_wqe = kzalloc_node(sizeof(*sq->db.ico_wqe) * wq_sz,
1029                                       GFP_KERNEL, numa);
1030         if (!sq->db.ico_wqe)
1031                 return -ENOMEM;
1032
1033         return 0;
1034 }
1035
1036 static int mlx5e_alloc_icosq(struct mlx5e_channel *c,
1037                              struct mlx5e_sq_param *param,
1038                              struct mlx5e_icosq *sq)
1039 {
1040         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1041         struct mlx5_core_dev *mdev = c->mdev;
1042         int err;
1043
1044         sq->mkey_be   = c->mkey_be;
1045         sq->channel   = c;
1046         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1047
1048         param->wq.db_numa_node = cpu_to_node(c->cpu);
1049         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq, &sq->wq_ctrl);
1050         if (err)
1051                 return err;
1052         sq->wq.db = &sq->wq.db[MLX5_SND_DBR];
1053
1054         err = mlx5e_alloc_icosq_db(sq, cpu_to_node(c->cpu));
1055         if (err)
1056                 goto err_sq_wq_destroy;
1057
1058         sq->edge = (sq->wq.sz_m1 + 1) - MLX5E_ICOSQ_MAX_WQEBBS;
1059
1060         return 0;
1061
1062 err_sq_wq_destroy:
1063         mlx5_wq_destroy(&sq->wq_ctrl);
1064
1065         return err;
1066 }
1067
1068 static void mlx5e_free_icosq(struct mlx5e_icosq *sq)
1069 {
1070         mlx5e_free_icosq_db(sq);
1071         mlx5_wq_destroy(&sq->wq_ctrl);
1072 }
1073
1074 static void mlx5e_free_txqsq_db(struct mlx5e_txqsq *sq)
1075 {
1076         kfree(sq->db.wqe_info);
1077         kfree(sq->db.dma_fifo);
1078 }
1079
1080 static int mlx5e_alloc_txqsq_db(struct mlx5e_txqsq *sq, int numa)
1081 {
1082         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1083         int df_sz = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
1084
1085         sq->db.dma_fifo = kzalloc_node(df_sz * sizeof(*sq->db.dma_fifo),
1086                                            GFP_KERNEL, numa);
1087         sq->db.wqe_info = kzalloc_node(wq_sz * sizeof(*sq->db.wqe_info),
1088                                            GFP_KERNEL, numa);
1089         if (!sq->db.dma_fifo || !sq->db.wqe_info) {
1090                 mlx5e_free_txqsq_db(sq);
1091                 return -ENOMEM;
1092         }
1093
1094         sq->dma_fifo_mask = df_sz - 1;
1095
1096         return 0;
1097 }
1098
1099 static int mlx5e_alloc_txqsq(struct mlx5e_channel *c,
1100                              int txq_ix,
1101                              struct mlx5e_params *params,
1102                              struct mlx5e_sq_param *param,
1103                              struct mlx5e_txqsq *sq)
1104 {
1105         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1106         struct mlx5_core_dev *mdev = c->mdev;
1107         int err;
1108
1109         sq->pdev      = c->pdev;
1110         sq->tstamp    = c->tstamp;
1111         sq->clock     = &mdev->clock;
1112         sq->mkey_be   = c->mkey_be;
1113         sq->channel   = c;
1114         sq->txq_ix    = txq_ix;
1115         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1116         sq->max_inline      = params->tx_max_inline;
1117         sq->min_inline_mode = params->tx_min_inline_mode;
1118         if (MLX5_IPSEC_DEV(c->priv->mdev))
1119                 set_bit(MLX5E_SQ_STATE_IPSEC, &sq->state);
1120
1121         param->wq.db_numa_node = cpu_to_node(c->cpu);
1122         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq, &sq->wq_ctrl);
1123         if (err)
1124                 return err;
1125         sq->wq.db    = &sq->wq.db[MLX5_SND_DBR];
1126
1127         err = mlx5e_alloc_txqsq_db(sq, cpu_to_node(c->cpu));
1128         if (err)
1129                 goto err_sq_wq_destroy;
1130
1131         sq->edge = (sq->wq.sz_m1 + 1) - MLX5_SEND_WQE_MAX_WQEBBS;
1132
1133         return 0;
1134
1135 err_sq_wq_destroy:
1136         mlx5_wq_destroy(&sq->wq_ctrl);
1137
1138         return err;
1139 }
1140
1141 static void mlx5e_free_txqsq(struct mlx5e_txqsq *sq)
1142 {
1143         mlx5e_free_txqsq_db(sq);
1144         mlx5_wq_destroy(&sq->wq_ctrl);
1145 }
1146
1147 struct mlx5e_create_sq_param {
1148         struct mlx5_wq_ctrl        *wq_ctrl;
1149         u32                         cqn;
1150         u32                         tisn;
1151         u8                          tis_lst_sz;
1152         u8                          min_inline_mode;
1153 };
1154
1155 static int mlx5e_create_sq(struct mlx5_core_dev *mdev,
1156                            struct mlx5e_sq_param *param,
1157                            struct mlx5e_create_sq_param *csp,
1158                            u32 *sqn)
1159 {
1160         void *in;
1161         void *sqc;
1162         void *wq;
1163         int inlen;
1164         int err;
1165
1166         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1167                 sizeof(u64) * csp->wq_ctrl->buf.npages;
1168         in = kvzalloc(inlen, GFP_KERNEL);
1169         if (!in)
1170                 return -ENOMEM;
1171
1172         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1173         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1174
1175         memcpy(sqc, param->sqc, sizeof(param->sqc));
1176         MLX5_SET(sqc,  sqc, tis_lst_sz, csp->tis_lst_sz);
1177         MLX5_SET(sqc,  sqc, tis_num_0, csp->tisn);
1178         MLX5_SET(sqc,  sqc, cqn, csp->cqn);
1179
1180         if (MLX5_CAP_ETH(mdev, wqe_inline_mode) == MLX5_CAP_INLINE_MODE_VPORT_CONTEXT)
1181                 MLX5_SET(sqc,  sqc, min_wqe_inline_mode, csp->min_inline_mode);
1182
1183         MLX5_SET(sqc,  sqc, state, MLX5_SQC_STATE_RST);
1184
1185         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
1186         MLX5_SET(wq,   wq, uar_page,      mdev->mlx5e_res.bfreg.index);
1187         MLX5_SET(wq,   wq, log_wq_pg_sz,  csp->wq_ctrl->buf.page_shift -
1188                                           MLX5_ADAPTER_PAGE_SHIFT);
1189         MLX5_SET64(wq, wq, dbr_addr,      csp->wq_ctrl->db.dma);
1190
1191         mlx5_fill_page_array(&csp->wq_ctrl->buf, (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
1192
1193         err = mlx5_core_create_sq(mdev, in, inlen, sqn);
1194
1195         kvfree(in);
1196
1197         return err;
1198 }
1199
1200 struct mlx5e_modify_sq_param {
1201         int curr_state;
1202         int next_state;
1203         bool rl_update;
1204         int rl_index;
1205 };
1206
1207 static int mlx5e_modify_sq(struct mlx5_core_dev *mdev, u32 sqn,
1208                            struct mlx5e_modify_sq_param *p)
1209 {
1210         void *in;
1211         void *sqc;
1212         int inlen;
1213         int err;
1214
1215         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1216         in = kvzalloc(inlen, GFP_KERNEL);
1217         if (!in)
1218                 return -ENOMEM;
1219
1220         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1221
1222         MLX5_SET(modify_sq_in, in, sq_state, p->curr_state);
1223         MLX5_SET(sqc, sqc, state, p->next_state);
1224         if (p->rl_update && p->next_state == MLX5_SQC_STATE_RDY) {
1225                 MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
1226                 MLX5_SET(sqc,  sqc, packet_pacing_rate_limit_index, p->rl_index);
1227         }
1228
1229         err = mlx5_core_modify_sq(mdev, sqn, in, inlen);
1230
1231         kvfree(in);
1232
1233         return err;
1234 }
1235
1236 static void mlx5e_destroy_sq(struct mlx5_core_dev *mdev, u32 sqn)
1237 {
1238         mlx5_core_destroy_sq(mdev, sqn);
1239 }
1240
1241 static int mlx5e_create_sq_rdy(struct mlx5_core_dev *mdev,
1242                                struct mlx5e_sq_param *param,
1243                                struct mlx5e_create_sq_param *csp,
1244                                u32 *sqn)
1245 {
1246         struct mlx5e_modify_sq_param msp = {0};
1247         int err;
1248
1249         err = mlx5e_create_sq(mdev, param, csp, sqn);
1250         if (err)
1251                 return err;
1252
1253         msp.curr_state = MLX5_SQC_STATE_RST;
1254         msp.next_state = MLX5_SQC_STATE_RDY;
1255         err = mlx5e_modify_sq(mdev, *sqn, &msp);
1256         if (err)
1257                 mlx5e_destroy_sq(mdev, *sqn);
1258
1259         return err;
1260 }
1261
1262 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1263                                 struct mlx5e_txqsq *sq, u32 rate);
1264
1265 static int mlx5e_open_txqsq(struct mlx5e_channel *c,
1266                             u32 tisn,
1267                             int txq_ix,
1268                             struct mlx5e_params *params,
1269                             struct mlx5e_sq_param *param,
1270                             struct mlx5e_txqsq *sq)
1271 {
1272         struct mlx5e_create_sq_param csp = {};
1273         u32 tx_rate;
1274         int err;
1275
1276         err = mlx5e_alloc_txqsq(c, txq_ix, params, param, sq);
1277         if (err)
1278                 return err;
1279
1280         csp.tisn            = tisn;
1281         csp.tis_lst_sz      = 1;
1282         csp.cqn             = sq->cq.mcq.cqn;
1283         csp.wq_ctrl         = &sq->wq_ctrl;
1284         csp.min_inline_mode = sq->min_inline_mode;
1285         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1286         if (err)
1287                 goto err_free_txqsq;
1288
1289         tx_rate = c->priv->tx_rates[sq->txq_ix];
1290         if (tx_rate)
1291                 mlx5e_set_sq_maxrate(c->netdev, sq, tx_rate);
1292
1293         return 0;
1294
1295 err_free_txqsq:
1296         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1297         mlx5e_free_txqsq(sq);
1298
1299         return err;
1300 }
1301
1302 static void mlx5e_activate_txqsq(struct mlx5e_txqsq *sq)
1303 {
1304         sq->txq = netdev_get_tx_queue(sq->channel->netdev, sq->txq_ix);
1305         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1306         netdev_tx_reset_queue(sq->txq);
1307         netif_tx_start_queue(sq->txq);
1308 }
1309
1310 static inline void netif_tx_disable_queue(struct netdev_queue *txq)
1311 {
1312         __netif_tx_lock_bh(txq);
1313         netif_tx_stop_queue(txq);
1314         __netif_tx_unlock_bh(txq);
1315 }
1316
1317 static void mlx5e_deactivate_txqsq(struct mlx5e_txqsq *sq)
1318 {
1319         struct mlx5e_channel *c = sq->channel;
1320
1321         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1322         /* prevent netif_tx_wake_queue */
1323         napi_synchronize(&c->napi);
1324
1325         netif_tx_disable_queue(sq->txq);
1326
1327         /* last doorbell out, godspeed .. */
1328         if (mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc, 1)) {
1329                 struct mlx5e_tx_wqe *nop;
1330
1331                 sq->db.wqe_info[(sq->pc & sq->wq.sz_m1)].skb = NULL;
1332                 nop = mlx5e_post_nop(&sq->wq, sq->sqn, &sq->pc);
1333                 mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &nop->ctrl);
1334         }
1335 }
1336
1337 static void mlx5e_close_txqsq(struct mlx5e_txqsq *sq)
1338 {
1339         struct mlx5e_channel *c = sq->channel;
1340         struct mlx5_core_dev *mdev = c->mdev;
1341
1342         mlx5e_destroy_sq(mdev, sq->sqn);
1343         if (sq->rate_limit)
1344                 mlx5_rl_remove_rate(mdev, sq->rate_limit);
1345         mlx5e_free_txqsq_descs(sq);
1346         mlx5e_free_txqsq(sq);
1347 }
1348
1349 static int mlx5e_open_icosq(struct mlx5e_channel *c,
1350                             struct mlx5e_params *params,
1351                             struct mlx5e_sq_param *param,
1352                             struct mlx5e_icosq *sq)
1353 {
1354         struct mlx5e_create_sq_param csp = {};
1355         int err;
1356
1357         err = mlx5e_alloc_icosq(c, param, sq);
1358         if (err)
1359                 return err;
1360
1361         csp.cqn             = sq->cq.mcq.cqn;
1362         csp.wq_ctrl         = &sq->wq_ctrl;
1363         csp.min_inline_mode = params->tx_min_inline_mode;
1364         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1365         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1366         if (err)
1367                 goto err_free_icosq;
1368
1369         return 0;
1370
1371 err_free_icosq:
1372         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1373         mlx5e_free_icosq(sq);
1374
1375         return err;
1376 }
1377
1378 static void mlx5e_close_icosq(struct mlx5e_icosq *sq)
1379 {
1380         struct mlx5e_channel *c = sq->channel;
1381
1382         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1383         napi_synchronize(&c->napi);
1384
1385         mlx5e_destroy_sq(c->mdev, sq->sqn);
1386         mlx5e_free_icosq(sq);
1387 }
1388
1389 static int mlx5e_open_xdpsq(struct mlx5e_channel *c,
1390                             struct mlx5e_params *params,
1391                             struct mlx5e_sq_param *param,
1392                             struct mlx5e_xdpsq *sq)
1393 {
1394         unsigned int ds_cnt = MLX5E_XDP_TX_DS_COUNT;
1395         struct mlx5e_create_sq_param csp = {};
1396         unsigned int inline_hdr_sz = 0;
1397         int err;
1398         int i;
1399
1400         err = mlx5e_alloc_xdpsq(c, params, param, sq);
1401         if (err)
1402                 return err;
1403
1404         csp.tis_lst_sz      = 1;
1405         csp.tisn            = c->priv->tisn[0]; /* tc = 0 */
1406         csp.cqn             = sq->cq.mcq.cqn;
1407         csp.wq_ctrl         = &sq->wq_ctrl;
1408         csp.min_inline_mode = sq->min_inline_mode;
1409         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1410         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1411         if (err)
1412                 goto err_free_xdpsq;
1413
1414         if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
1415                 inline_hdr_sz = MLX5E_XDP_MIN_INLINE;
1416                 ds_cnt++;
1417         }
1418
1419         /* Pre initialize fixed WQE fields */
1420         for (i = 0; i < mlx5_wq_cyc_get_size(&sq->wq); i++) {
1421                 struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(&sq->wq, i);
1422                 struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
1423                 struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
1424                 struct mlx5_wqe_data_seg *dseg;
1425
1426                 cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
1427                 eseg->inline_hdr.sz = cpu_to_be16(inline_hdr_sz);
1428
1429                 dseg = (struct mlx5_wqe_data_seg *)cseg + (ds_cnt - 1);
1430                 dseg->lkey = sq->mkey_be;
1431         }
1432
1433         return 0;
1434
1435 err_free_xdpsq:
1436         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1437         mlx5e_free_xdpsq(sq);
1438
1439         return err;
1440 }
1441
1442 static void mlx5e_close_xdpsq(struct mlx5e_xdpsq *sq)
1443 {
1444         struct mlx5e_channel *c = sq->channel;
1445
1446         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1447         napi_synchronize(&c->napi);
1448
1449         mlx5e_destroy_sq(c->mdev, sq->sqn);
1450         mlx5e_free_xdpsq_descs(sq);
1451         mlx5e_free_xdpsq(sq);
1452 }
1453
1454 static int mlx5e_alloc_cq_common(struct mlx5_core_dev *mdev,
1455                                  struct mlx5e_cq_param *param,
1456                                  struct mlx5e_cq *cq)
1457 {
1458         struct mlx5_core_cq *mcq = &cq->mcq;
1459         int eqn_not_used;
1460         unsigned int irqn;
1461         int err;
1462         u32 i;
1463
1464         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1465                                &cq->wq_ctrl);
1466         if (err)
1467                 return err;
1468
1469         mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1470
1471         mcq->cqe_sz     = 64;
1472         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1473         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1474         *mcq->set_ci_db = 0;
1475         *mcq->arm_db    = 0;
1476         mcq->vector     = param->eq_ix;
1477         mcq->comp       = mlx5e_completion_event;
1478         mcq->event      = mlx5e_cq_error_event;
1479         mcq->irqn       = irqn;
1480
1481         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1482                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1483
1484                 cqe->op_own = 0xf1;
1485         }
1486
1487         cq->mdev = mdev;
1488
1489         return 0;
1490 }
1491
1492 static int mlx5e_alloc_cq(struct mlx5e_channel *c,
1493                           struct mlx5e_cq_param *param,
1494                           struct mlx5e_cq *cq)
1495 {
1496         struct mlx5_core_dev *mdev = c->priv->mdev;
1497         int err;
1498
1499         param->wq.buf_numa_node = cpu_to_node(c->cpu);
1500         param->wq.db_numa_node  = cpu_to_node(c->cpu);
1501         param->eq_ix   = c->ix;
1502
1503         err = mlx5e_alloc_cq_common(mdev, param, cq);
1504
1505         cq->napi    = &c->napi;
1506         cq->channel = c;
1507
1508         return err;
1509 }
1510
1511 static void mlx5e_free_cq(struct mlx5e_cq *cq)
1512 {
1513         mlx5_cqwq_destroy(&cq->wq_ctrl);
1514 }
1515
1516 static int mlx5e_create_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
1517 {
1518         struct mlx5_core_dev *mdev = cq->mdev;
1519         struct mlx5_core_cq *mcq = &cq->mcq;
1520
1521         void *in;
1522         void *cqc;
1523         int inlen;
1524         unsigned int irqn_not_used;
1525         int eqn;
1526         int err;
1527
1528         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1529                 sizeof(u64) * cq->wq_ctrl.frag_buf.npages;
1530         in = kvzalloc(inlen, GFP_KERNEL);
1531         if (!in)
1532                 return -ENOMEM;
1533
1534         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1535
1536         memcpy(cqc, param->cqc, sizeof(param->cqc));
1537
1538         mlx5_fill_page_frag_array(&cq->wq_ctrl.frag_buf,
1539                                   (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
1540
1541         mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
1542
1543         MLX5_SET(cqc,   cqc, cq_period_mode, param->cq_period_mode);
1544         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
1545         MLX5_SET(cqc,   cqc, uar_page,      mdev->priv.uar->index);
1546         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.frag_buf.page_shift -
1547                                             MLX5_ADAPTER_PAGE_SHIFT);
1548         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
1549
1550         err = mlx5_core_create_cq(mdev, mcq, in, inlen);
1551
1552         kvfree(in);
1553
1554         if (err)
1555                 return err;
1556
1557         mlx5e_cq_arm(cq);
1558
1559         return 0;
1560 }
1561
1562 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
1563 {
1564         mlx5_core_destroy_cq(cq->mdev, &cq->mcq);
1565 }
1566
1567 static int mlx5e_open_cq(struct mlx5e_channel *c,
1568                          struct mlx5e_cq_moder moder,
1569                          struct mlx5e_cq_param *param,
1570                          struct mlx5e_cq *cq)
1571 {
1572         struct mlx5_core_dev *mdev = c->mdev;
1573         int err;
1574
1575         err = mlx5e_alloc_cq(c, param, cq);
1576         if (err)
1577                 return err;
1578
1579         err = mlx5e_create_cq(cq, param);
1580         if (err)
1581                 goto err_free_cq;
1582
1583         if (MLX5_CAP_GEN(mdev, cq_moderation))
1584                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq, moder.usec, moder.pkts);
1585         return 0;
1586
1587 err_free_cq:
1588         mlx5e_free_cq(cq);
1589
1590         return err;
1591 }
1592
1593 static void mlx5e_close_cq(struct mlx5e_cq *cq)
1594 {
1595         mlx5e_destroy_cq(cq);
1596         mlx5e_free_cq(cq);
1597 }
1598
1599 static int mlx5e_get_cpu(struct mlx5e_priv *priv, int ix)
1600 {
1601         return cpumask_first(priv->mdev->priv.irq_info[ix].mask);
1602 }
1603
1604 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1605                              struct mlx5e_params *params,
1606                              struct mlx5e_channel_param *cparam)
1607 {
1608         int err;
1609         int tc;
1610
1611         for (tc = 0; tc < c->num_tc; tc++) {
1612                 err = mlx5e_open_cq(c, params->tx_cq_moderation,
1613                                     &cparam->tx_cq, &c->sq[tc].cq);
1614                 if (err)
1615                         goto err_close_tx_cqs;
1616         }
1617
1618         return 0;
1619
1620 err_close_tx_cqs:
1621         for (tc--; tc >= 0; tc--)
1622                 mlx5e_close_cq(&c->sq[tc].cq);
1623
1624         return err;
1625 }
1626
1627 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1628 {
1629         int tc;
1630
1631         for (tc = 0; tc < c->num_tc; tc++)
1632                 mlx5e_close_cq(&c->sq[tc].cq);
1633 }
1634
1635 static int mlx5e_open_sqs(struct mlx5e_channel *c,
1636                           struct mlx5e_params *params,
1637                           struct mlx5e_channel_param *cparam)
1638 {
1639         int err;
1640         int tc;
1641
1642         for (tc = 0; tc < params->num_tc; tc++) {
1643                 int txq_ix = c->ix + tc * params->num_channels;
1644
1645                 err = mlx5e_open_txqsq(c, c->priv->tisn[tc], txq_ix,
1646                                        params, &cparam->sq, &c->sq[tc]);
1647                 if (err)
1648                         goto err_close_sqs;
1649         }
1650
1651         return 0;
1652
1653 err_close_sqs:
1654         for (tc--; tc >= 0; tc--)
1655                 mlx5e_close_txqsq(&c->sq[tc]);
1656
1657         return err;
1658 }
1659
1660 static void mlx5e_close_sqs(struct mlx5e_channel *c)
1661 {
1662         int tc;
1663
1664         for (tc = 0; tc < c->num_tc; tc++)
1665                 mlx5e_close_txqsq(&c->sq[tc]);
1666 }
1667
1668 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1669                                 struct mlx5e_txqsq *sq, u32 rate)
1670 {
1671         struct mlx5e_priv *priv = netdev_priv(dev);
1672         struct mlx5_core_dev *mdev = priv->mdev;
1673         struct mlx5e_modify_sq_param msp = {0};
1674         u16 rl_index = 0;
1675         int err;
1676
1677         if (rate == sq->rate_limit)
1678                 /* nothing to do */
1679                 return 0;
1680
1681         if (sq->rate_limit)
1682                 /* remove current rl index to free space to next ones */
1683                 mlx5_rl_remove_rate(mdev, sq->rate_limit);
1684
1685         sq->rate_limit = 0;
1686
1687         if (rate) {
1688                 err = mlx5_rl_add_rate(mdev, rate, &rl_index);
1689                 if (err) {
1690                         netdev_err(dev, "Failed configuring rate %u: %d\n",
1691                                    rate, err);
1692                         return err;
1693                 }
1694         }
1695
1696         msp.curr_state = MLX5_SQC_STATE_RDY;
1697         msp.next_state = MLX5_SQC_STATE_RDY;
1698         msp.rl_index   = rl_index;
1699         msp.rl_update  = true;
1700         err = mlx5e_modify_sq(mdev, sq->sqn, &msp);
1701         if (err) {
1702                 netdev_err(dev, "Failed configuring rate %u: %d\n",
1703                            rate, err);
1704                 /* remove the rate from the table */
1705                 if (rate)
1706                         mlx5_rl_remove_rate(mdev, rate);
1707                 return err;
1708         }
1709
1710         sq->rate_limit = rate;
1711         return 0;
1712 }
1713
1714 static int mlx5e_set_tx_maxrate(struct net_device *dev, int index, u32 rate)
1715 {
1716         struct mlx5e_priv *priv = netdev_priv(dev);
1717         struct mlx5_core_dev *mdev = priv->mdev;
1718         struct mlx5e_txqsq *sq = priv->txq2sq[index];
1719         int err = 0;
1720
1721         if (!mlx5_rl_is_supported(mdev)) {
1722                 netdev_err(dev, "Rate limiting is not supported on this device\n");
1723                 return -EINVAL;
1724         }
1725
1726         /* rate is given in Mb/sec, HW config is in Kb/sec */
1727         rate = rate << 10;
1728
1729         /* Check whether rate in valid range, 0 is always valid */
1730         if (rate && !mlx5_rl_is_in_range(mdev, rate)) {
1731                 netdev_err(dev, "TX rate %u, is not in range\n", rate);
1732                 return -ERANGE;
1733         }
1734
1735         mutex_lock(&priv->state_lock);
1736         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
1737                 err = mlx5e_set_sq_maxrate(dev, sq, rate);
1738         if (!err)
1739                 priv->tx_rates[index] = rate;
1740         mutex_unlock(&priv->state_lock);
1741
1742         return err;
1743 }
1744
1745 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
1746                               struct mlx5e_params *params,
1747                               struct mlx5e_channel_param *cparam,
1748                               struct mlx5e_channel **cp)
1749 {
1750         struct mlx5e_cq_moder icocq_moder = {0, 0};
1751         struct net_device *netdev = priv->netdev;
1752         int cpu = mlx5e_get_cpu(priv, ix);
1753         struct mlx5e_channel *c;
1754         unsigned int irq;
1755         int err;
1756         int eqn;
1757
1758         c = kzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
1759         if (!c)
1760                 return -ENOMEM;
1761
1762         c->priv     = priv;
1763         c->mdev     = priv->mdev;
1764         c->tstamp   = &priv->tstamp;
1765         c->ix       = ix;
1766         c->cpu      = cpu;
1767         c->pdev     = &priv->mdev->pdev->dev;
1768         c->netdev   = priv->netdev;
1769         c->mkey_be  = cpu_to_be32(priv->mdev->mlx5e_res.mkey.key);
1770         c->num_tc   = params->num_tc;
1771         c->xdp      = !!params->xdp_prog;
1772
1773         mlx5_vector2eqn(priv->mdev, ix, &eqn, &irq);
1774         c->irq_desc = irq_to_desc(irq);
1775
1776         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
1777
1778         err = mlx5e_open_cq(c, icocq_moder, &cparam->icosq_cq, &c->icosq.cq);
1779         if (err)
1780                 goto err_napi_del;
1781
1782         err = mlx5e_open_tx_cqs(c, params, cparam);
1783         if (err)
1784                 goto err_close_icosq_cq;
1785
1786         err = mlx5e_open_cq(c, params->rx_cq_moderation, &cparam->rx_cq, &c->rq.cq);
1787         if (err)
1788                 goto err_close_tx_cqs;
1789
1790         /* XDP SQ CQ params are same as normal TXQ sq CQ params */
1791         err = c->xdp ? mlx5e_open_cq(c, params->tx_cq_moderation,
1792                                      &cparam->tx_cq, &c->rq.xdpsq.cq) : 0;
1793         if (err)
1794                 goto err_close_rx_cq;
1795
1796         napi_enable(&c->napi);
1797
1798         err = mlx5e_open_icosq(c, params, &cparam->icosq, &c->icosq);
1799         if (err)
1800                 goto err_disable_napi;
1801
1802         err = mlx5e_open_sqs(c, params, cparam);
1803         if (err)
1804                 goto err_close_icosq;
1805
1806         err = c->xdp ? mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, &c->rq.xdpsq) : 0;
1807         if (err)
1808                 goto err_close_sqs;
1809
1810         err = mlx5e_open_rq(c, params, &cparam->rq, &c->rq);
1811         if (err)
1812                 goto err_close_xdp_sq;
1813
1814         *cp = c;
1815
1816         return 0;
1817 err_close_xdp_sq:
1818         if (c->xdp)
1819                 mlx5e_close_xdpsq(&c->rq.xdpsq);
1820
1821 err_close_sqs:
1822         mlx5e_close_sqs(c);
1823
1824 err_close_icosq:
1825         mlx5e_close_icosq(&c->icosq);
1826
1827 err_disable_napi:
1828         napi_disable(&c->napi);
1829         if (c->xdp)
1830                 mlx5e_close_cq(&c->rq.xdpsq.cq);
1831
1832 err_close_rx_cq:
1833         mlx5e_close_cq(&c->rq.cq);
1834
1835 err_close_tx_cqs:
1836         mlx5e_close_tx_cqs(c);
1837
1838 err_close_icosq_cq:
1839         mlx5e_close_cq(&c->icosq.cq);
1840
1841 err_napi_del:
1842         netif_napi_del(&c->napi);
1843         kfree(c);
1844
1845         return err;
1846 }
1847
1848 static void mlx5e_activate_channel(struct mlx5e_channel *c)
1849 {
1850         int tc;
1851
1852         for (tc = 0; tc < c->num_tc; tc++)
1853                 mlx5e_activate_txqsq(&c->sq[tc]);
1854         mlx5e_activate_rq(&c->rq);
1855         netif_set_xps_queue(c->netdev, get_cpu_mask(c->cpu), c->ix);
1856 }
1857
1858 static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
1859 {
1860         int tc;
1861
1862         mlx5e_deactivate_rq(&c->rq);
1863         for (tc = 0; tc < c->num_tc; tc++)
1864                 mlx5e_deactivate_txqsq(&c->sq[tc]);
1865 }
1866
1867 static void mlx5e_close_channel(struct mlx5e_channel *c)
1868 {
1869         mlx5e_close_rq(&c->rq);
1870         if (c->xdp)
1871                 mlx5e_close_xdpsq(&c->rq.xdpsq);
1872         mlx5e_close_sqs(c);
1873         mlx5e_close_icosq(&c->icosq);
1874         napi_disable(&c->napi);
1875         if (c->xdp)
1876                 mlx5e_close_cq(&c->rq.xdpsq.cq);
1877         mlx5e_close_cq(&c->rq.cq);
1878         mlx5e_close_tx_cqs(c);
1879         mlx5e_close_cq(&c->icosq.cq);
1880         netif_napi_del(&c->napi);
1881
1882         kfree(c);
1883 }
1884
1885 static void mlx5e_build_rq_param(struct mlx5e_priv *priv,
1886                                  struct mlx5e_params *params,
1887                                  struct mlx5e_rq_param *param)
1888 {
1889         void *rqc = param->rqc;
1890         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1891
1892         switch (params->rq_wq_type) {
1893         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1894                 MLX5_SET(wq, wq, log_wqe_num_of_strides, params->mpwqe_log_num_strides - 9);
1895                 MLX5_SET(wq, wq, log_wqe_stride_size, params->mpwqe_log_stride_sz - 6);
1896                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
1897                 break;
1898         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1899                 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1900         }
1901
1902         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1903         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1904         MLX5_SET(wq, wq, log_wq_sz,        params->log_rq_size);
1905         MLX5_SET(wq, wq, pd,               priv->mdev->mlx5e_res.pdn);
1906         MLX5_SET(rqc, rqc, counter_set_id, priv->q_counter);
1907         MLX5_SET(rqc, rqc, vsd,            params->vlan_strip_disable);
1908         MLX5_SET(rqc, rqc, scatter_fcs,    params->scatter_fcs_en);
1909
1910         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1911         param->wq.linear = 1;
1912 }
1913
1914 static void mlx5e_build_drop_rq_param(struct mlx5e_rq_param *param)
1915 {
1916         void *rqc = param->rqc;
1917         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1918
1919         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1920         MLX5_SET(wq, wq, log_wq_stride,    ilog2(sizeof(struct mlx5e_rx_wqe)));
1921 }
1922
1923 static void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
1924                                         struct mlx5e_sq_param *param)
1925 {
1926         void *sqc = param->sqc;
1927         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1928
1929         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1930         MLX5_SET(wq, wq, pd,            priv->mdev->mlx5e_res.pdn);
1931
1932         param->wq.buf_numa_node = dev_to_node(&priv->mdev->pdev->dev);
1933 }
1934
1935 static void mlx5e_build_sq_param(struct mlx5e_priv *priv,
1936                                  struct mlx5e_params *params,
1937                                  struct mlx5e_sq_param *param)
1938 {
1939         void *sqc = param->sqc;
1940         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1941
1942         mlx5e_build_sq_param_common(priv, param);
1943         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
1944         MLX5_SET(sqc, sqc, allow_swp, !!MLX5_IPSEC_DEV(priv->mdev));
1945 }
1946
1947 static void mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
1948                                         struct mlx5e_cq_param *param)
1949 {
1950         void *cqc = param->cqc;
1951
1952         MLX5_SET(cqc, cqc, uar_page, priv->mdev->priv.uar->index);
1953 }
1954
1955 static void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
1956                                     struct mlx5e_params *params,
1957                                     struct mlx5e_cq_param *param)
1958 {
1959         void *cqc = param->cqc;
1960         u8 log_cq_size;
1961
1962         switch (params->rq_wq_type) {
1963         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
1964                 log_cq_size = params->log_rq_size + params->mpwqe_log_num_strides;
1965                 break;
1966         default: /* MLX5_WQ_TYPE_LINKED_LIST */
1967                 log_cq_size = params->log_rq_size;
1968         }
1969
1970         MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
1971         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
1972                 MLX5_SET(cqc, cqc, mini_cqe_res_format, MLX5_CQE_FORMAT_CSUM);
1973                 MLX5_SET(cqc, cqc, cqe_comp_en, 1);
1974         }
1975
1976         mlx5e_build_common_cq_param(priv, param);
1977         param->cq_period_mode = params->rx_cq_moderation.cq_period_mode;
1978 }
1979
1980 static void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
1981                                     struct mlx5e_params *params,
1982                                     struct mlx5e_cq_param *param)
1983 {
1984         void *cqc = param->cqc;
1985
1986         MLX5_SET(cqc, cqc, log_cq_size, params->log_sq_size);
1987
1988         mlx5e_build_common_cq_param(priv, param);
1989         param->cq_period_mode = params->tx_cq_moderation.cq_period_mode;
1990 }
1991
1992 static void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
1993                                      u8 log_wq_size,
1994                                      struct mlx5e_cq_param *param)
1995 {
1996         void *cqc = param->cqc;
1997
1998         MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
1999
2000         mlx5e_build_common_cq_param(priv, param);
2001
2002         param->cq_period_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
2003 }
2004
2005 static void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
2006                                     u8 log_wq_size,
2007                                     struct mlx5e_sq_param *param)
2008 {
2009         void *sqc = param->sqc;
2010         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2011
2012         mlx5e_build_sq_param_common(priv, param);
2013
2014         MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
2015         MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(priv->mdev, reg_umr_sq));
2016 }
2017
2018 static void mlx5e_build_xdpsq_param(struct mlx5e_priv *priv,
2019                                     struct mlx5e_params *params,
2020                                     struct mlx5e_sq_param *param)
2021 {
2022         void *sqc = param->sqc;
2023         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2024
2025         mlx5e_build_sq_param_common(priv, param);
2026         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
2027 }
2028
2029 static void mlx5e_build_channel_param(struct mlx5e_priv *priv,
2030                                       struct mlx5e_params *params,
2031                                       struct mlx5e_channel_param *cparam)
2032 {
2033         u8 icosq_log_wq_sz = MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
2034
2035         mlx5e_build_rq_param(priv, params, &cparam->rq);
2036         mlx5e_build_sq_param(priv, params, &cparam->sq);
2037         mlx5e_build_xdpsq_param(priv, params, &cparam->xdp_sq);
2038         mlx5e_build_icosq_param(priv, icosq_log_wq_sz, &cparam->icosq);
2039         mlx5e_build_rx_cq_param(priv, params, &cparam->rx_cq);
2040         mlx5e_build_tx_cq_param(priv, params, &cparam->tx_cq);
2041         mlx5e_build_ico_cq_param(priv, icosq_log_wq_sz, &cparam->icosq_cq);
2042 }
2043
2044 int mlx5e_open_channels(struct mlx5e_priv *priv,
2045                         struct mlx5e_channels *chs)
2046 {
2047         struct mlx5e_channel_param *cparam;
2048         int err = -ENOMEM;
2049         int i;
2050
2051         chs->num = chs->params.num_channels;
2052
2053         chs->c = kcalloc(chs->num, sizeof(struct mlx5e_channel *), GFP_KERNEL);
2054         cparam = kzalloc(sizeof(struct mlx5e_channel_param), GFP_KERNEL);
2055         if (!chs->c || !cparam)
2056                 goto err_free;
2057
2058         mlx5e_build_channel_param(priv, &chs->params, cparam);
2059         for (i = 0; i < chs->num; i++) {
2060                 err = mlx5e_open_channel(priv, i, &chs->params, cparam, &chs->c[i]);
2061                 if (err)
2062                         goto err_close_channels;
2063         }
2064
2065         kfree(cparam);
2066         return 0;
2067
2068 err_close_channels:
2069         for (i--; i >= 0; i--)
2070                 mlx5e_close_channel(chs->c[i]);
2071
2072 err_free:
2073         kfree(chs->c);
2074         kfree(cparam);
2075         chs->num = 0;
2076         return err;
2077 }
2078
2079 static void mlx5e_activate_channels(struct mlx5e_channels *chs)
2080 {
2081         int i;
2082
2083         for (i = 0; i < chs->num; i++)
2084                 mlx5e_activate_channel(chs->c[i]);
2085 }
2086
2087 static int mlx5e_wait_channels_min_rx_wqes(struct mlx5e_channels *chs)
2088 {
2089         int err = 0;
2090         int i;
2091
2092         for (i = 0; i < chs->num; i++) {
2093                 err = mlx5e_wait_for_min_rx_wqes(&chs->c[i]->rq);
2094                 if (err)
2095                         break;
2096         }
2097
2098         return err;
2099 }
2100
2101 static void mlx5e_deactivate_channels(struct mlx5e_channels *chs)
2102 {
2103         int i;
2104
2105         for (i = 0; i < chs->num; i++)
2106                 mlx5e_deactivate_channel(chs->c[i]);
2107 }
2108
2109 void mlx5e_close_channels(struct mlx5e_channels *chs)
2110 {
2111         int i;
2112
2113         for (i = 0; i < chs->num; i++)
2114                 mlx5e_close_channel(chs->c[i]);
2115
2116         kfree(chs->c);
2117         chs->num = 0;
2118 }
2119
2120 static int
2121 mlx5e_create_rqt(struct mlx5e_priv *priv, int sz, struct mlx5e_rqt *rqt)
2122 {
2123         struct mlx5_core_dev *mdev = priv->mdev;
2124         void *rqtc;
2125         int inlen;
2126         int err;
2127         u32 *in;
2128         int i;
2129
2130         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2131         in = kvzalloc(inlen, GFP_KERNEL);
2132         if (!in)
2133                 return -ENOMEM;
2134
2135         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2136
2137         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2138         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2139
2140         for (i = 0; i < sz; i++)
2141                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->drop_rq.rqn);
2142
2143         err = mlx5_core_create_rqt(mdev, in, inlen, &rqt->rqtn);
2144         if (!err)
2145                 rqt->enabled = true;
2146
2147         kvfree(in);
2148         return err;
2149 }
2150
2151 void mlx5e_destroy_rqt(struct mlx5e_priv *priv, struct mlx5e_rqt *rqt)
2152 {
2153         rqt->enabled = false;
2154         mlx5_core_destroy_rqt(priv->mdev, rqt->rqtn);
2155 }
2156
2157 int mlx5e_create_indirect_rqt(struct mlx5e_priv *priv)
2158 {
2159         struct mlx5e_rqt *rqt = &priv->indir_rqt;
2160         int err;
2161
2162         err = mlx5e_create_rqt(priv, MLX5E_INDIR_RQT_SIZE, rqt);
2163         if (err)
2164                 mlx5_core_warn(priv->mdev, "create indirect rqts failed, %d\n", err);
2165         return err;
2166 }
2167
2168 int mlx5e_create_direct_rqts(struct mlx5e_priv *priv)
2169 {
2170         struct mlx5e_rqt *rqt;
2171         int err;
2172         int ix;
2173
2174         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
2175                 rqt = &priv->direct_tir[ix].rqt;
2176                 err = mlx5e_create_rqt(priv, 1 /*size */, rqt);
2177                 if (err)
2178                         goto err_destroy_rqts;
2179         }
2180
2181         return 0;
2182
2183 err_destroy_rqts:
2184         mlx5_core_warn(priv->mdev, "create direct rqts failed, %d\n", err);
2185         for (ix--; ix >= 0; ix--)
2186                 mlx5e_destroy_rqt(priv, &priv->direct_tir[ix].rqt);
2187
2188         return err;
2189 }
2190
2191 void mlx5e_destroy_direct_rqts(struct mlx5e_priv *priv)
2192 {
2193         int i;
2194
2195         for (i = 0; i < priv->profile->max_nch(priv->mdev); i++)
2196                 mlx5e_destroy_rqt(priv, &priv->direct_tir[i].rqt);
2197 }
2198
2199 static int mlx5e_rx_hash_fn(int hfunc)
2200 {
2201         return (hfunc == ETH_RSS_HASH_TOP) ?
2202                MLX5_RX_HASH_FN_TOEPLITZ :
2203                MLX5_RX_HASH_FN_INVERTED_XOR8;
2204 }
2205
2206 static int mlx5e_bits_invert(unsigned long a, int size)
2207 {
2208         int inv = 0;
2209         int i;
2210
2211         for (i = 0; i < size; i++)
2212                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
2213
2214         return inv;
2215 }
2216
2217 static void mlx5e_fill_rqt_rqns(struct mlx5e_priv *priv, int sz,
2218                                 struct mlx5e_redirect_rqt_param rrp, void *rqtc)
2219 {
2220         int i;
2221
2222         for (i = 0; i < sz; i++) {
2223                 u32 rqn;
2224
2225                 if (rrp.is_rss) {
2226                         int ix = i;
2227
2228                         if (rrp.rss.hfunc == ETH_RSS_HASH_XOR)
2229                                 ix = mlx5e_bits_invert(i, ilog2(sz));
2230
2231                         ix = priv->channels.params.indirection_rqt[ix];
2232                         rqn = rrp.rss.channels->c[ix]->rq.rqn;
2233                 } else {
2234                         rqn = rrp.rqn;
2235                 }
2236                 MLX5_SET(rqtc, rqtc, rq_num[i], rqn);
2237         }
2238 }
2239
2240 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, u32 rqtn, int sz,
2241                        struct mlx5e_redirect_rqt_param rrp)
2242 {
2243         struct mlx5_core_dev *mdev = priv->mdev;
2244         void *rqtc;
2245         int inlen;
2246         u32 *in;
2247         int err;
2248
2249         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
2250         in = kvzalloc(inlen, GFP_KERNEL);
2251         if (!in)
2252                 return -ENOMEM;
2253
2254         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
2255
2256         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2257         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
2258         mlx5e_fill_rqt_rqns(priv, sz, rrp, rqtc);
2259         err = mlx5_core_modify_rqt(mdev, rqtn, in, inlen);
2260
2261         kvfree(in);
2262         return err;
2263 }
2264
2265 static u32 mlx5e_get_direct_rqn(struct mlx5e_priv *priv, int ix,
2266                                 struct mlx5e_redirect_rqt_param rrp)
2267 {
2268         if (!rrp.is_rss)
2269                 return rrp.rqn;
2270
2271         if (ix >= rrp.rss.channels->num)
2272                 return priv->drop_rq.rqn;
2273
2274         return rrp.rss.channels->c[ix]->rq.rqn;
2275 }
2276
2277 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv,
2278                                 struct mlx5e_redirect_rqt_param rrp)
2279 {
2280         u32 rqtn;
2281         int ix;
2282
2283         if (priv->indir_rqt.enabled) {
2284                 /* RSS RQ table */
2285                 rqtn = priv->indir_rqt.rqtn;
2286                 mlx5e_redirect_rqt(priv, rqtn, MLX5E_INDIR_RQT_SIZE, rrp);
2287         }
2288
2289         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
2290                 struct mlx5e_redirect_rqt_param direct_rrp = {
2291                         .is_rss = false,
2292                         {
2293                                 .rqn    = mlx5e_get_direct_rqn(priv, ix, rrp)
2294                         },
2295                 };
2296
2297                 /* Direct RQ Tables */
2298                 if (!priv->direct_tir[ix].rqt.enabled)
2299                         continue;
2300
2301                 rqtn = priv->direct_tir[ix].rqt.rqtn;
2302                 mlx5e_redirect_rqt(priv, rqtn, 1, direct_rrp);
2303         }
2304 }
2305
2306 static void mlx5e_redirect_rqts_to_channels(struct mlx5e_priv *priv,
2307                                             struct mlx5e_channels *chs)
2308 {
2309         struct mlx5e_redirect_rqt_param rrp = {
2310                 .is_rss        = true,
2311                 {
2312                         .rss = {
2313                                 .channels  = chs,
2314                                 .hfunc     = chs->params.rss_hfunc,
2315                         }
2316                 },
2317         };
2318
2319         mlx5e_redirect_rqts(priv, rrp);
2320 }
2321
2322 static void mlx5e_redirect_rqts_to_drop(struct mlx5e_priv *priv)
2323 {
2324         struct mlx5e_redirect_rqt_param drop_rrp = {
2325                 .is_rss = false,
2326                 {
2327                         .rqn = priv->drop_rq.rqn,
2328                 },
2329         };
2330
2331         mlx5e_redirect_rqts(priv, drop_rrp);
2332 }
2333
2334 static void mlx5e_build_tir_ctx_lro(struct mlx5e_params *params, void *tirc)
2335 {
2336         if (!params->lro_en)
2337                 return;
2338
2339 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2340
2341         MLX5_SET(tirc, tirc, lro_enable_mask,
2342                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2343                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2344         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
2345                  (params->lro_wqe_sz - ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2346         MLX5_SET(tirc, tirc, lro_timeout_period_usecs, params->lro_timeout);
2347 }
2348
2349 void mlx5e_build_indir_tir_ctx_hash(struct mlx5e_params *params,
2350                                     enum mlx5e_traffic_types tt,
2351                                     void *tirc, bool inner)
2352 {
2353         void *hfso = inner ? MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner) :
2354                              MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2355
2356 #define MLX5_HASH_IP            (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2357                                  MLX5_HASH_FIELD_SEL_DST_IP)
2358
2359 #define MLX5_HASH_IP_L4PORTS    (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2360                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2361                                  MLX5_HASH_FIELD_SEL_L4_SPORT |\
2362                                  MLX5_HASH_FIELD_SEL_L4_DPORT)
2363
2364 #define MLX5_HASH_IP_IPSEC_SPI  (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2365                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2366                                  MLX5_HASH_FIELD_SEL_IPSEC_SPI)
2367
2368         MLX5_SET(tirc, tirc, rx_hash_fn, mlx5e_rx_hash_fn(params->rss_hfunc));
2369         if (params->rss_hfunc == ETH_RSS_HASH_TOP) {
2370                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
2371                                              rx_hash_toeplitz_key);
2372                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
2373                                                rx_hash_toeplitz_key);
2374
2375                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2376                 memcpy(rss_key, params->toeplitz_hash_key, len);
2377         }
2378
2379         switch (tt) {
2380         case MLX5E_TT_IPV4_TCP:
2381                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2382                          MLX5_L3_PROT_TYPE_IPV4);
2383                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2384                          MLX5_L4_PROT_TYPE_TCP);
2385                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2386                          MLX5_HASH_IP_L4PORTS);
2387                 break;
2388
2389         case MLX5E_TT_IPV6_TCP:
2390                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2391                          MLX5_L3_PROT_TYPE_IPV6);
2392                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2393                          MLX5_L4_PROT_TYPE_TCP);
2394                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2395                          MLX5_HASH_IP_L4PORTS);
2396                 break;
2397
2398         case MLX5E_TT_IPV4_UDP:
2399                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2400                          MLX5_L3_PROT_TYPE_IPV4);
2401                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2402                          MLX5_L4_PROT_TYPE_UDP);
2403                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2404                          MLX5_HASH_IP_L4PORTS);
2405                 break;
2406
2407         case MLX5E_TT_IPV6_UDP:
2408                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2409                          MLX5_L3_PROT_TYPE_IPV6);
2410                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2411                          MLX5_L4_PROT_TYPE_UDP);
2412                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2413                          MLX5_HASH_IP_L4PORTS);
2414                 break;
2415
2416         case MLX5E_TT_IPV4_IPSEC_AH:
2417                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2418                          MLX5_L3_PROT_TYPE_IPV4);
2419                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2420                          MLX5_HASH_IP_IPSEC_SPI);
2421                 break;
2422
2423         case MLX5E_TT_IPV6_IPSEC_AH:
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, selected_fields,
2427                          MLX5_HASH_IP_IPSEC_SPI);
2428                 break;
2429
2430         case MLX5E_TT_IPV4_IPSEC_ESP:
2431                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2432                          MLX5_L3_PROT_TYPE_IPV4);
2433                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2434                          MLX5_HASH_IP_IPSEC_SPI);
2435                 break;
2436
2437         case MLX5E_TT_IPV6_IPSEC_ESP:
2438                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2439                          MLX5_L3_PROT_TYPE_IPV6);
2440                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2441                          MLX5_HASH_IP_IPSEC_SPI);
2442                 break;
2443
2444         case MLX5E_TT_IPV4:
2445                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2446                          MLX5_L3_PROT_TYPE_IPV4);
2447                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2448                          MLX5_HASH_IP);
2449                 break;
2450
2451         case MLX5E_TT_IPV6:
2452                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2453                          MLX5_L3_PROT_TYPE_IPV6);
2454                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2455                          MLX5_HASH_IP);
2456                 break;
2457         default:
2458                 WARN_ONCE(true, "%s: bad traffic type!\n", __func__);
2459         }
2460 }
2461
2462 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
2463 {
2464         struct mlx5_core_dev *mdev = priv->mdev;
2465
2466         void *in;
2467         void *tirc;
2468         int inlen;
2469         int err;
2470         int tt;
2471         int ix;
2472
2473         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
2474         in = kvzalloc(inlen, GFP_KERNEL);
2475         if (!in)
2476                 return -ENOMEM;
2477
2478         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
2479         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2480
2481         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2482
2483         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2484                 err = mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in,
2485                                            inlen);
2486                 if (err)
2487                         goto free_in;
2488         }
2489
2490         for (ix = 0; ix < priv->profile->max_nch(priv->mdev); ix++) {
2491                 err = mlx5_core_modify_tir(mdev, priv->direct_tir[ix].tirn,
2492                                            in, inlen);
2493                 if (err)
2494                         goto free_in;
2495         }
2496
2497 free_in:
2498         kvfree(in);
2499
2500         return err;
2501 }
2502
2503 static void mlx5e_build_inner_indir_tir_ctx(struct mlx5e_priv *priv,
2504                                             enum mlx5e_traffic_types tt,
2505                                             u32 *tirc)
2506 {
2507         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2508
2509         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2510
2511         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2512         MLX5_SET(tirc, tirc, indirect_table, priv->indir_rqt.rqtn);
2513         MLX5_SET(tirc, tirc, tunneled_offload_en, 0x1);
2514
2515         mlx5e_build_indir_tir_ctx_hash(&priv->channels.params, tt, tirc, true);
2516 }
2517
2518 static int mlx5e_set_mtu(struct mlx5e_priv *priv, u16 mtu)
2519 {
2520         struct mlx5_core_dev *mdev = priv->mdev;
2521         u16 hw_mtu = MLX5E_SW2HW_MTU(priv, mtu);
2522         int err;
2523
2524         err = mlx5_set_port_mtu(mdev, hw_mtu, 1);
2525         if (err)
2526                 return err;
2527
2528         /* Update vport context MTU */
2529         mlx5_modify_nic_vport_mtu(mdev, hw_mtu);
2530         return 0;
2531 }
2532
2533 static void mlx5e_query_mtu(struct mlx5e_priv *priv, u16 *mtu)
2534 {
2535         struct mlx5_core_dev *mdev = priv->mdev;
2536         u16 hw_mtu = 0;
2537         int err;
2538
2539         err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2540         if (err || !hw_mtu) /* fallback to port oper mtu */
2541                 mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
2542
2543         *mtu = MLX5E_HW2SW_MTU(priv, hw_mtu);
2544 }
2545
2546 static int mlx5e_set_dev_port_mtu(struct mlx5e_priv *priv)
2547 {
2548         struct net_device *netdev = priv->netdev;
2549         u16 mtu;
2550         int err;
2551
2552         err = mlx5e_set_mtu(priv, netdev->mtu);
2553         if (err)
2554                 return err;
2555
2556         mlx5e_query_mtu(priv, &mtu);
2557         if (mtu != netdev->mtu)
2558                 netdev_warn(netdev, "%s: VPort MTU %d is different than netdev mtu %d\n",
2559                             __func__, mtu, netdev->mtu);
2560
2561         netdev->mtu = mtu;
2562         return 0;
2563 }
2564
2565 static void mlx5e_netdev_set_tcs(struct net_device *netdev)
2566 {
2567         struct mlx5e_priv *priv = netdev_priv(netdev);
2568         int nch = priv->channels.params.num_channels;
2569         int ntc = priv->channels.params.num_tc;
2570         int tc;
2571
2572         netdev_reset_tc(netdev);
2573
2574         if (ntc == 1)
2575                 return;
2576
2577         netdev_set_num_tc(netdev, ntc);
2578
2579         /* Map netdev TCs to offset 0
2580          * We have our own UP to TXQ mapping for QoS
2581          */
2582         for (tc = 0; tc < ntc; tc++)
2583                 netdev_set_tc_queue(netdev, tc, nch, 0);
2584 }
2585
2586 static void mlx5e_build_channels_tx_maps(struct mlx5e_priv *priv)
2587 {
2588         struct mlx5e_channel *c;
2589         struct mlx5e_txqsq *sq;
2590         int i, tc;
2591
2592         for (i = 0; i < priv->channels.num; i++)
2593                 for (tc = 0; tc < priv->profile->max_tc; tc++)
2594                         priv->channel_tc2txq[i][tc] = i + tc * priv->channels.num;
2595
2596         for (i = 0; i < priv->channels.num; i++) {
2597                 c = priv->channels.c[i];
2598                 for (tc = 0; tc < c->num_tc; tc++) {
2599                         sq = &c->sq[tc];
2600                         priv->txq2sq[sq->txq_ix] = sq;
2601                 }
2602         }
2603 }
2604
2605 void mlx5e_activate_priv_channels(struct mlx5e_priv *priv)
2606 {
2607         int num_txqs = priv->channels.num * priv->channels.params.num_tc;
2608         struct net_device *netdev = priv->netdev;
2609
2610         mlx5e_netdev_set_tcs(netdev);
2611         netif_set_real_num_tx_queues(netdev, num_txqs);
2612         netif_set_real_num_rx_queues(netdev, priv->channels.num);
2613
2614         mlx5e_build_channels_tx_maps(priv);
2615         mlx5e_activate_channels(&priv->channels);
2616         netif_tx_start_all_queues(priv->netdev);
2617
2618         if (MLX5_VPORT_MANAGER(priv->mdev))
2619                 mlx5e_add_sqs_fwd_rules(priv);
2620
2621         mlx5e_wait_channels_min_rx_wqes(&priv->channels);
2622         mlx5e_redirect_rqts_to_channels(priv, &priv->channels);
2623 }
2624
2625 void mlx5e_deactivate_priv_channels(struct mlx5e_priv *priv)
2626 {
2627         mlx5e_redirect_rqts_to_drop(priv);
2628
2629         if (MLX5_VPORT_MANAGER(priv->mdev))
2630                 mlx5e_remove_sqs_fwd_rules(priv);
2631
2632         /* FIXME: This is a W/A only for tx timeout watch dog false alarm when
2633          * polling for inactive tx queues.
2634          */
2635         netif_tx_stop_all_queues(priv->netdev);
2636         netif_tx_disable(priv->netdev);
2637         mlx5e_deactivate_channels(&priv->channels);
2638 }
2639
2640 void mlx5e_switch_priv_channels(struct mlx5e_priv *priv,
2641                                 struct mlx5e_channels *new_chs,
2642                                 mlx5e_fp_hw_modify hw_modify)
2643 {
2644         struct net_device *netdev = priv->netdev;
2645         int new_num_txqs;
2646         int carrier_ok;
2647         new_num_txqs = new_chs->num * new_chs->params.num_tc;
2648
2649         carrier_ok = netif_carrier_ok(netdev);
2650         netif_carrier_off(netdev);
2651
2652         if (new_num_txqs < netdev->real_num_tx_queues)
2653                 netif_set_real_num_tx_queues(netdev, new_num_txqs);
2654
2655         mlx5e_deactivate_priv_channels(priv);
2656         mlx5e_close_channels(&priv->channels);
2657
2658         priv->channels = *new_chs;
2659
2660         /* New channels are ready to roll, modify HW settings if needed */
2661         if (hw_modify)
2662                 hw_modify(priv);
2663
2664         mlx5e_refresh_tirs(priv, false);
2665         mlx5e_activate_priv_channels(priv);
2666
2667         /* return carrier back if needed */
2668         if (carrier_ok)
2669                 netif_carrier_on(netdev);
2670 }
2671
2672 void mlx5e_timestamp_init(struct mlx5e_priv *priv)
2673 {
2674         priv->tstamp.tx_type   = HWTSTAMP_TX_OFF;
2675         priv->tstamp.rx_filter = HWTSTAMP_FILTER_NONE;
2676 }
2677
2678 int mlx5e_open_locked(struct net_device *netdev)
2679 {
2680         struct mlx5e_priv *priv = netdev_priv(netdev);
2681         int err;
2682
2683         set_bit(MLX5E_STATE_OPENED, &priv->state);
2684
2685         err = mlx5e_open_channels(priv, &priv->channels);
2686         if (err)
2687                 goto err_clear_state_opened_flag;
2688
2689         mlx5e_refresh_tirs(priv, false);
2690         mlx5e_activate_priv_channels(priv);
2691         if (priv->profile->update_carrier)
2692                 priv->profile->update_carrier(priv);
2693
2694         if (priv->profile->update_stats)
2695                 queue_delayed_work(priv->wq, &priv->update_stats_work, 0);
2696
2697         return 0;
2698
2699 err_clear_state_opened_flag:
2700         clear_bit(MLX5E_STATE_OPENED, &priv->state);
2701         return err;
2702 }
2703
2704 int mlx5e_open(struct net_device *netdev)
2705 {
2706         struct mlx5e_priv *priv = netdev_priv(netdev);
2707         int err;
2708
2709         mutex_lock(&priv->state_lock);
2710         err = mlx5e_open_locked(netdev);
2711         if (!err)
2712                 mlx5_set_port_admin_status(priv->mdev, MLX5_PORT_UP);
2713         mutex_unlock(&priv->state_lock);
2714
2715         return err;
2716 }
2717
2718 int mlx5e_close_locked(struct net_device *netdev)
2719 {
2720         struct mlx5e_priv *priv = netdev_priv(netdev);
2721
2722         /* May already be CLOSED in case a previous configuration operation
2723          * (e.g RX/TX queue size change) that involves close&open failed.
2724          */
2725         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
2726                 return 0;
2727
2728         clear_bit(MLX5E_STATE_OPENED, &priv->state);
2729
2730         netif_carrier_off(priv->netdev);
2731         mlx5e_deactivate_priv_channels(priv);
2732         mlx5e_close_channels(&priv->channels);
2733
2734         return 0;
2735 }
2736
2737 int mlx5e_close(struct net_device *netdev)
2738 {
2739         struct mlx5e_priv *priv = netdev_priv(netdev);
2740         int err;
2741
2742         if (!netif_device_present(netdev))
2743                 return -ENODEV;
2744
2745         mutex_lock(&priv->state_lock);
2746         mlx5_set_port_admin_status(priv->mdev, MLX5_PORT_DOWN);
2747         err = mlx5e_close_locked(netdev);
2748         mutex_unlock(&priv->state_lock);
2749
2750         return err;
2751 }
2752
2753 static int mlx5e_alloc_drop_rq(struct mlx5_core_dev *mdev,
2754                                struct mlx5e_rq *rq,
2755                                struct mlx5e_rq_param *param)
2756 {
2757         void *rqc = param->rqc;
2758         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
2759         int err;
2760
2761         param->wq.db_numa_node = param->wq.buf_numa_node;
2762
2763         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
2764                                 &rq->wq_ctrl);
2765         if (err)
2766                 return err;
2767
2768         rq->mdev = mdev;
2769
2770         return 0;
2771 }
2772
2773 static int mlx5e_alloc_drop_cq(struct mlx5_core_dev *mdev,
2774                                struct mlx5e_cq *cq,
2775                                struct mlx5e_cq_param *param)
2776 {
2777         return mlx5e_alloc_cq_common(mdev, param, cq);
2778 }
2779
2780 static int mlx5e_open_drop_rq(struct mlx5_core_dev *mdev,
2781                               struct mlx5e_rq *drop_rq)
2782 {
2783         struct mlx5e_cq_param cq_param = {};
2784         struct mlx5e_rq_param rq_param = {};
2785         struct mlx5e_cq *cq = &drop_rq->cq;
2786         int err;
2787
2788         mlx5e_build_drop_rq_param(&rq_param);
2789
2790         err = mlx5e_alloc_drop_cq(mdev, cq, &cq_param);
2791         if (err)
2792                 return err;
2793
2794         err = mlx5e_create_cq(cq, &cq_param);
2795         if (err)
2796                 goto err_free_cq;
2797
2798         err = mlx5e_alloc_drop_rq(mdev, drop_rq, &rq_param);
2799         if (err)
2800                 goto err_destroy_cq;
2801
2802         err = mlx5e_create_rq(drop_rq, &rq_param);
2803         if (err)
2804                 goto err_free_rq;
2805
2806         return 0;
2807
2808 err_free_rq:
2809         mlx5e_free_rq(drop_rq);
2810
2811 err_destroy_cq:
2812         mlx5e_destroy_cq(cq);
2813
2814 err_free_cq:
2815         mlx5e_free_cq(cq);
2816
2817         return err;
2818 }
2819
2820 static void mlx5e_close_drop_rq(struct mlx5e_rq *drop_rq)
2821 {
2822         mlx5e_destroy_rq(drop_rq);
2823         mlx5e_free_rq(drop_rq);
2824         mlx5e_destroy_cq(&drop_rq->cq);
2825         mlx5e_free_cq(&drop_rq->cq);
2826 }
2827
2828 int mlx5e_create_tis(struct mlx5_core_dev *mdev, int tc,
2829                      u32 underlay_qpn, u32 *tisn)
2830 {
2831         u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {0};
2832         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
2833
2834         MLX5_SET(tisc, tisc, prio, tc << 1);
2835         MLX5_SET(tisc, tisc, underlay_qpn, underlay_qpn);
2836         MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.td.tdn);
2837
2838         if (mlx5_lag_is_lacp_owner(mdev))
2839                 MLX5_SET(tisc, tisc, strict_lag_tx_port_affinity, 1);
2840
2841         return mlx5_core_create_tis(mdev, in, sizeof(in), tisn);
2842 }
2843
2844 void mlx5e_destroy_tis(struct mlx5_core_dev *mdev, u32 tisn)
2845 {
2846         mlx5_core_destroy_tis(mdev, tisn);
2847 }
2848
2849 int mlx5e_create_tises(struct mlx5e_priv *priv)
2850 {
2851         int err;
2852         int tc;
2853
2854         for (tc = 0; tc < priv->profile->max_tc; tc++) {
2855                 err = mlx5e_create_tis(priv->mdev, tc, 0, &priv->tisn[tc]);
2856                 if (err)
2857                         goto err_close_tises;
2858         }
2859
2860         return 0;
2861
2862 err_close_tises:
2863         for (tc--; tc >= 0; tc--)
2864                 mlx5e_destroy_tis(priv->mdev, priv->tisn[tc]);
2865
2866         return err;
2867 }
2868
2869 void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
2870 {
2871         int tc;
2872
2873         for (tc = 0; tc < priv->profile->max_tc; tc++)
2874                 mlx5e_destroy_tis(priv->mdev, priv->tisn[tc]);
2875 }
2876
2877 static void mlx5e_build_indir_tir_ctx(struct mlx5e_priv *priv,
2878                                       enum mlx5e_traffic_types tt,
2879                                       u32 *tirc)
2880 {
2881         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2882
2883         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2884
2885         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2886         MLX5_SET(tirc, tirc, indirect_table, priv->indir_rqt.rqtn);
2887         mlx5e_build_indir_tir_ctx_hash(&priv->channels.params, tt, tirc, false);
2888 }
2889
2890 static void mlx5e_build_direct_tir_ctx(struct mlx5e_priv *priv, u32 rqtn, u32 *tirc)
2891 {
2892         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
2893
2894         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2895
2896         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
2897         MLX5_SET(tirc, tirc, indirect_table, rqtn);
2898         MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
2899 }
2900
2901 int mlx5e_create_indirect_tirs(struct mlx5e_priv *priv)
2902 {
2903         struct mlx5e_tir *tir;
2904         void *tirc;
2905         int inlen;
2906         int i = 0;
2907         int err;
2908         u32 *in;
2909         int tt;
2910
2911         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2912         in = kvzalloc(inlen, GFP_KERNEL);
2913         if (!in)
2914                 return -ENOMEM;
2915
2916         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2917                 memset(in, 0, inlen);
2918                 tir = &priv->indir_tir[tt];
2919                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2920                 mlx5e_build_indir_tir_ctx(priv, tt, tirc);
2921                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2922                 if (err) {
2923                         mlx5_core_warn(priv->mdev, "create indirect tirs failed, %d\n", err);
2924                         goto err_destroy_inner_tirs;
2925                 }
2926         }
2927
2928         if (!mlx5e_tunnel_inner_ft_supported(priv->mdev))
2929                 goto out;
2930
2931         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++) {
2932                 memset(in, 0, inlen);
2933                 tir = &priv->inner_indir_tir[i];
2934                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2935                 mlx5e_build_inner_indir_tir_ctx(priv, i, tirc);
2936                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2937                 if (err) {
2938                         mlx5_core_warn(priv->mdev, "create inner indirect tirs failed, %d\n", err);
2939                         goto err_destroy_inner_tirs;
2940                 }
2941         }
2942
2943 out:
2944         kvfree(in);
2945
2946         return 0;
2947
2948 err_destroy_inner_tirs:
2949         for (i--; i >= 0; i--)
2950                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
2951
2952         for (tt--; tt >= 0; tt--)
2953                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[tt]);
2954
2955         kvfree(in);
2956
2957         return err;
2958 }
2959
2960 int mlx5e_create_direct_tirs(struct mlx5e_priv *priv)
2961 {
2962         int nch = priv->profile->max_nch(priv->mdev);
2963         struct mlx5e_tir *tir;
2964         void *tirc;
2965         int inlen;
2966         int err;
2967         u32 *in;
2968         int ix;
2969
2970         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2971         in = kvzalloc(inlen, GFP_KERNEL);
2972         if (!in)
2973                 return -ENOMEM;
2974
2975         for (ix = 0; ix < nch; ix++) {
2976                 memset(in, 0, inlen);
2977                 tir = &priv->direct_tir[ix];
2978                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
2979                 mlx5e_build_direct_tir_ctx(priv, priv->direct_tir[ix].rqt.rqtn, tirc);
2980                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
2981                 if (err)
2982                         goto err_destroy_ch_tirs;
2983         }
2984
2985         kvfree(in);
2986
2987         return 0;
2988
2989 err_destroy_ch_tirs:
2990         mlx5_core_warn(priv->mdev, "create direct tirs failed, %d\n", err);
2991         for (ix--; ix >= 0; ix--)
2992                 mlx5e_destroy_tir(priv->mdev, &priv->direct_tir[ix]);
2993
2994         kvfree(in);
2995
2996         return err;
2997 }
2998
2999 void mlx5e_destroy_indirect_tirs(struct mlx5e_priv *priv)
3000 {
3001         int i;
3002
3003         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3004                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[i]);
3005
3006         if (!mlx5e_tunnel_inner_ft_supported(priv->mdev))
3007                 return;
3008
3009         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3010                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3011 }
3012
3013 void mlx5e_destroy_direct_tirs(struct mlx5e_priv *priv)
3014 {
3015         int nch = priv->profile->max_nch(priv->mdev);
3016         int i;
3017
3018         for (i = 0; i < nch; i++)
3019                 mlx5e_destroy_tir(priv->mdev, &priv->direct_tir[i]);
3020 }
3021
3022 static int mlx5e_modify_channels_scatter_fcs(struct mlx5e_channels *chs, bool enable)
3023 {
3024         int err = 0;
3025         int i;
3026
3027         for (i = 0; i < chs->num; i++) {
3028                 err = mlx5e_modify_rq_scatter_fcs(&chs->c[i]->rq, enable);
3029                 if (err)
3030                         return err;
3031         }
3032
3033         return 0;
3034 }
3035
3036 static int mlx5e_modify_channels_vsd(struct mlx5e_channels *chs, bool vsd)
3037 {
3038         int err = 0;
3039         int i;
3040
3041         for (i = 0; i < chs->num; i++) {
3042                 err = mlx5e_modify_rq_vsd(&chs->c[i]->rq, vsd);
3043                 if (err)
3044                         return err;
3045         }
3046
3047         return 0;
3048 }
3049
3050 static int mlx5e_setup_tc_mqprio(struct net_device *netdev,
3051                                  struct tc_mqprio_qopt *mqprio)
3052 {
3053         struct mlx5e_priv *priv = netdev_priv(netdev);
3054         struct mlx5e_channels new_channels = {};
3055         u8 tc = mqprio->num_tc;
3056         int err = 0;
3057
3058         mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
3059
3060         if (tc && tc != MLX5E_MAX_NUM_TC)
3061                 return -EINVAL;
3062
3063         mutex_lock(&priv->state_lock);
3064
3065         new_channels.params = priv->channels.params;
3066         new_channels.params.num_tc = tc ? tc : 1;
3067
3068         if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
3069                 priv->channels.params = new_channels.params;
3070                 goto out;
3071         }
3072
3073         err = mlx5e_open_channels(priv, &new_channels);
3074         if (err)
3075                 goto out;
3076
3077         mlx5e_switch_priv_channels(priv, &new_channels, NULL);
3078 out:
3079         mutex_unlock(&priv->state_lock);
3080         return err;
3081 }
3082
3083 #ifdef CONFIG_MLX5_ESWITCH
3084 static int mlx5e_setup_tc_cls_flower(struct mlx5e_priv *priv,
3085                                      struct tc_cls_flower_offload *cls_flower)
3086 {
3087         if (cls_flower->common.chain_index)
3088                 return -EOPNOTSUPP;
3089
3090         switch (cls_flower->command) {
3091         case TC_CLSFLOWER_REPLACE:
3092                 return mlx5e_configure_flower(priv, cls_flower);
3093         case TC_CLSFLOWER_DESTROY:
3094                 return mlx5e_delete_flower(priv, cls_flower);
3095         case TC_CLSFLOWER_STATS:
3096                 return mlx5e_stats_flower(priv, cls_flower);
3097         default:
3098                 return -EOPNOTSUPP;
3099         }
3100 }
3101
3102 int mlx5e_setup_tc_block_cb(enum tc_setup_type type, void *type_data,
3103                             void *cb_priv)
3104 {
3105         struct mlx5e_priv *priv = cb_priv;
3106
3107         if (!tc_can_offload(priv->netdev))
3108                 return -EOPNOTSUPP;
3109
3110         switch (type) {
3111         case TC_SETUP_CLSFLOWER:
3112                 return mlx5e_setup_tc_cls_flower(priv, type_data);
3113         default:
3114                 return -EOPNOTSUPP;
3115         }
3116 }
3117
3118 static int mlx5e_setup_tc_block(struct net_device *dev,
3119                                 struct tc_block_offload *f)
3120 {
3121         struct mlx5e_priv *priv = netdev_priv(dev);
3122
3123         if (f->binder_type != TCF_BLOCK_BINDER_TYPE_CLSACT_INGRESS)
3124                 return -EOPNOTSUPP;
3125
3126         switch (f->command) {
3127         case TC_BLOCK_BIND:
3128                 return tcf_block_cb_register(f->block, mlx5e_setup_tc_block_cb,
3129                                              priv, priv);
3130         case TC_BLOCK_UNBIND:
3131                 tcf_block_cb_unregister(f->block, mlx5e_setup_tc_block_cb,
3132                                         priv);
3133                 return 0;
3134         default:
3135                 return -EOPNOTSUPP;
3136         }
3137 }
3138 #endif
3139
3140 int mlx5e_setup_tc(struct net_device *dev, enum tc_setup_type type,
3141                    void *type_data)
3142 {
3143         switch (type) {
3144 #ifdef CONFIG_MLX5_ESWITCH
3145         case TC_SETUP_BLOCK:
3146                 return mlx5e_setup_tc_block(dev, type_data);
3147 #endif
3148         case TC_SETUP_QDISC_MQPRIO:
3149                 return mlx5e_setup_tc_mqprio(dev, type_data);
3150         default:
3151                 return -EOPNOTSUPP;
3152         }
3153 }
3154
3155 static void
3156 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
3157 {
3158         struct mlx5e_priv *priv = netdev_priv(dev);
3159         struct mlx5e_sw_stats *sstats = &priv->stats.sw;
3160         struct mlx5e_vport_stats *vstats = &priv->stats.vport;
3161         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
3162
3163         if (mlx5e_is_uplink_rep(priv)) {
3164                 stats->rx_packets = PPORT_802_3_GET(pstats, a_frames_received_ok);
3165                 stats->rx_bytes   = PPORT_802_3_GET(pstats, a_octets_received_ok);
3166                 stats->tx_packets = PPORT_802_3_GET(pstats, a_frames_transmitted_ok);
3167                 stats->tx_bytes   = PPORT_802_3_GET(pstats, a_octets_transmitted_ok);
3168         } else {
3169                 stats->rx_packets = sstats->rx_packets;
3170                 stats->rx_bytes   = sstats->rx_bytes;
3171                 stats->tx_packets = sstats->tx_packets;
3172                 stats->tx_bytes   = sstats->tx_bytes;
3173                 stats->tx_dropped = sstats->tx_queue_dropped;
3174         }
3175
3176         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
3177
3178         stats->rx_length_errors =
3179                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
3180                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
3181                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
3182         stats->rx_crc_errors =
3183                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
3184         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
3185         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
3186         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
3187                            stats->rx_frame_errors;
3188         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
3189
3190         /* vport multicast also counts packets that are dropped due to steering
3191          * or rx out of buffer
3192          */
3193         stats->multicast =
3194                 VPORT_COUNTER_GET(vstats, received_eth_multicast.packets);
3195 }
3196
3197 static void mlx5e_set_rx_mode(struct net_device *dev)
3198 {
3199         struct mlx5e_priv *priv = netdev_priv(dev);
3200
3201         queue_work(priv->wq, &priv->set_rx_mode_work);
3202 }
3203
3204 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
3205 {
3206         struct mlx5e_priv *priv = netdev_priv(netdev);
3207         struct sockaddr *saddr = addr;
3208
3209         if (!is_valid_ether_addr(saddr->sa_data))
3210                 return -EADDRNOTAVAIL;
3211
3212         netif_addr_lock_bh(netdev);
3213         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
3214         netif_addr_unlock_bh(netdev);
3215
3216         queue_work(priv->wq, &priv->set_rx_mode_work);
3217
3218         return 0;
3219 }
3220
3221 #define MLX5E_SET_FEATURE(features, feature, enable)    \
3222         do {                                            \
3223                 if (enable)                             \
3224                         *features |= feature;           \
3225                 else                                    \
3226                         *features &= ~feature;          \
3227         } while (0)
3228
3229 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
3230
3231 static int set_feature_lro(struct net_device *netdev, bool enable)
3232 {
3233         struct mlx5e_priv *priv = netdev_priv(netdev);
3234         struct mlx5e_channels new_channels = {};
3235         int err = 0;
3236         bool reset;
3237
3238         mutex_lock(&priv->state_lock);
3239
3240         reset = (priv->channels.params.rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST);
3241         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
3242
3243         new_channels.params = priv->channels.params;
3244         new_channels.params.lro_en = enable;
3245
3246         if (!reset) {
3247                 priv->channels.params = new_channels.params;
3248                 err = mlx5e_modify_tirs_lro(priv);
3249                 goto out;
3250         }
3251
3252         err = mlx5e_open_channels(priv, &new_channels);
3253         if (err)
3254                 goto out;
3255
3256         mlx5e_switch_priv_channels(priv, &new_channels, mlx5e_modify_tirs_lro);
3257 out:
3258         mutex_unlock(&priv->state_lock);
3259         return err;
3260 }
3261
3262 static int set_feature_cvlan_filter(struct net_device *netdev, bool enable)
3263 {
3264         struct mlx5e_priv *priv = netdev_priv(netdev);
3265
3266         if (enable)
3267                 mlx5e_enable_cvlan_filter(priv);
3268         else
3269                 mlx5e_disable_cvlan_filter(priv);
3270
3271         return 0;
3272 }
3273
3274 static int set_feature_tc_num_filters(struct net_device *netdev, bool enable)
3275 {
3276         struct mlx5e_priv *priv = netdev_priv(netdev);
3277
3278         if (!enable && mlx5e_tc_num_filters(priv)) {
3279                 netdev_err(netdev,
3280                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
3281                 return -EINVAL;
3282         }
3283
3284         return 0;
3285 }
3286
3287 static int set_feature_rx_all(struct net_device *netdev, bool enable)
3288 {
3289         struct mlx5e_priv *priv = netdev_priv(netdev);
3290         struct mlx5_core_dev *mdev = priv->mdev;
3291
3292         return mlx5_set_port_fcs(mdev, !enable);
3293 }
3294
3295 static int set_feature_rx_fcs(struct net_device *netdev, bool enable)
3296 {
3297         struct mlx5e_priv *priv = netdev_priv(netdev);
3298         int err;
3299
3300         mutex_lock(&priv->state_lock);
3301
3302         priv->channels.params.scatter_fcs_en = enable;
3303         err = mlx5e_modify_channels_scatter_fcs(&priv->channels, enable);
3304         if (err)
3305                 priv->channels.params.scatter_fcs_en = !enable;
3306
3307         mutex_unlock(&priv->state_lock);
3308
3309         return err;
3310 }
3311
3312 static int set_feature_rx_vlan(struct net_device *netdev, bool enable)
3313 {
3314         struct mlx5e_priv *priv = netdev_priv(netdev);
3315         int err = 0;
3316
3317         mutex_lock(&priv->state_lock);
3318
3319         priv->channels.params.vlan_strip_disable = !enable;
3320         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3321                 goto unlock;
3322
3323         err = mlx5e_modify_channels_vsd(&priv->channels, !enable);
3324         if (err)
3325                 priv->channels.params.vlan_strip_disable = enable;
3326
3327 unlock:
3328         mutex_unlock(&priv->state_lock);
3329
3330         return err;
3331 }
3332
3333 #ifdef CONFIG_RFS_ACCEL
3334 static int set_feature_arfs(struct net_device *netdev, bool enable)
3335 {
3336         struct mlx5e_priv *priv = netdev_priv(netdev);
3337         int err;
3338
3339         if (enable)
3340                 err = mlx5e_arfs_enable(priv);
3341         else
3342                 err = mlx5e_arfs_disable(priv);
3343
3344         return err;
3345 }
3346 #endif
3347
3348 static int mlx5e_handle_feature(struct net_device *netdev,
3349                                 netdev_features_t *features,
3350                                 netdev_features_t wanted_features,
3351                                 netdev_features_t feature,
3352                                 mlx5e_feature_handler feature_handler)
3353 {
3354         netdev_features_t changes = wanted_features ^ netdev->features;
3355         bool enable = !!(wanted_features & feature);
3356         int err;
3357
3358         if (!(changes & feature))
3359                 return 0;
3360
3361         err = feature_handler(netdev, enable);
3362         if (err) {
3363                 netdev_err(netdev, "%s feature %pNF failed, err %d\n",
3364                            enable ? "Enable" : "Disable", &feature, err);
3365                 return err;
3366         }
3367
3368         MLX5E_SET_FEATURE(features, feature, enable);
3369         return 0;
3370 }
3371
3372 static int mlx5e_set_features(struct net_device *netdev,
3373                               netdev_features_t features)
3374 {
3375         netdev_features_t oper_features = netdev->features;
3376         int err;
3377
3378         err  = mlx5e_handle_feature(netdev, &oper_features, features,
3379                                     NETIF_F_LRO, set_feature_lro);
3380         err |= mlx5e_handle_feature(netdev, &oper_features, features,
3381                                     NETIF_F_HW_VLAN_CTAG_FILTER,
3382                                     set_feature_cvlan_filter);
3383         err |= mlx5e_handle_feature(netdev, &oper_features, features,
3384                                     NETIF_F_HW_TC, set_feature_tc_num_filters);
3385         err |= mlx5e_handle_feature(netdev, &oper_features, features,
3386                                     NETIF_F_RXALL, set_feature_rx_all);
3387         err |= mlx5e_handle_feature(netdev, &oper_features, features,
3388                                     NETIF_F_RXFCS, set_feature_rx_fcs);
3389         err |= mlx5e_handle_feature(netdev, &oper_features, features,
3390                                     NETIF_F_HW_VLAN_CTAG_RX, set_feature_rx_vlan);
3391 #ifdef CONFIG_RFS_ACCEL
3392         err |= mlx5e_handle_feature(netdev, &oper_features, features,
3393                                     NETIF_F_NTUPLE, set_feature_arfs);
3394 #endif
3395
3396         if (err) {
3397                 netdev->features = oper_features;
3398                 return -EINVAL;
3399         }
3400
3401         return 0;
3402 }
3403
3404 static netdev_features_t mlx5e_fix_features(struct net_device *netdev,
3405                                             netdev_features_t features)
3406 {
3407         struct mlx5e_priv *priv = netdev_priv(netdev);
3408
3409         mutex_lock(&priv->state_lock);
3410         if (!bitmap_empty(priv->fs.vlan.active_svlans, VLAN_N_VID)) {
3411                 /* HW strips the outer C-tag header, this is a problem
3412                  * for S-tag traffic.
3413                  */
3414                 features &= ~NETIF_F_HW_VLAN_CTAG_RX;
3415                 if (!priv->channels.params.vlan_strip_disable)
3416                         netdev_warn(netdev, "Dropping C-tag vlan stripping offload due to S-tag vlan\n");
3417         }
3418         mutex_unlock(&priv->state_lock);
3419
3420         return features;
3421 }
3422
3423 static int mlx5e_change_mtu(struct net_device *netdev, int new_mtu)
3424 {
3425         struct mlx5e_priv *priv = netdev_priv(netdev);
3426         struct mlx5e_channels new_channels = {};
3427         int curr_mtu;
3428         int err = 0;
3429         bool reset;
3430
3431         mutex_lock(&priv->state_lock);
3432
3433         reset = !priv->channels.params.lro_en &&
3434                 (priv->channels.params.rq_wq_type !=
3435                  MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ);
3436
3437         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
3438
3439         curr_mtu    = netdev->mtu;
3440         netdev->mtu = new_mtu;
3441
3442         if (!reset) {
3443                 mlx5e_set_dev_port_mtu(priv);
3444                 goto out;
3445         }
3446
3447         new_channels.params = priv->channels.params;
3448         err = mlx5e_open_channels(priv, &new_channels);
3449         if (err) {
3450                 netdev->mtu = curr_mtu;
3451                 goto out;
3452         }
3453
3454         mlx5e_switch_priv_channels(priv, &new_channels, mlx5e_set_dev_port_mtu);
3455
3456 out:
3457         mutex_unlock(&priv->state_lock);
3458         return err;
3459 }
3460
3461 int mlx5e_hwstamp_set(struct mlx5e_priv *priv, struct ifreq *ifr)
3462 {
3463         struct hwtstamp_config config;
3464         int err;
3465
3466         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
3467                 return -EOPNOTSUPP;
3468
3469         if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
3470                 return -EFAULT;
3471
3472         /* TX HW timestamp */
3473         switch (config.tx_type) {
3474         case HWTSTAMP_TX_OFF:
3475         case HWTSTAMP_TX_ON:
3476                 break;
3477         default:
3478                 return -ERANGE;
3479         }
3480
3481         mutex_lock(&priv->state_lock);
3482         /* RX HW timestamp */
3483         switch (config.rx_filter) {
3484         case HWTSTAMP_FILTER_NONE:
3485                 /* Reset CQE compression to Admin default */
3486                 mlx5e_modify_rx_cqe_compression_locked(priv, priv->channels.params.rx_cqe_compress_def);
3487                 break;
3488         case HWTSTAMP_FILTER_ALL:
3489         case HWTSTAMP_FILTER_SOME:
3490         case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
3491         case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
3492         case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
3493         case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
3494         case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
3495         case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
3496         case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
3497         case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
3498         case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
3499         case HWTSTAMP_FILTER_PTP_V2_EVENT:
3500         case HWTSTAMP_FILTER_PTP_V2_SYNC:
3501         case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
3502         case HWTSTAMP_FILTER_NTP_ALL:
3503                 /* Disable CQE compression */
3504                 netdev_warn(priv->netdev, "Disabling cqe compression");
3505                 err = mlx5e_modify_rx_cqe_compression_locked(priv, false);
3506                 if (err) {
3507                         netdev_err(priv->netdev, "Failed disabling cqe compression err=%d\n", err);
3508                         mutex_unlock(&priv->state_lock);
3509                         return err;
3510                 }
3511                 config.rx_filter = HWTSTAMP_FILTER_ALL;
3512                 break;
3513         default:
3514                 mutex_unlock(&priv->state_lock);
3515                 return -ERANGE;
3516         }
3517
3518         memcpy(&priv->tstamp, &config, sizeof(config));
3519         mutex_unlock(&priv->state_lock);
3520
3521         return copy_to_user(ifr->ifr_data, &config,
3522                             sizeof(config)) ? -EFAULT : 0;
3523 }
3524
3525 int mlx5e_hwstamp_get(struct mlx5e_priv *priv, struct ifreq *ifr)
3526 {
3527         struct hwtstamp_config *cfg = &priv->tstamp;
3528
3529         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
3530                 return -EOPNOTSUPP;
3531
3532         return copy_to_user(ifr->ifr_data, cfg, sizeof(*cfg)) ? -EFAULT : 0;
3533 }
3534
3535 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
3536 {
3537         struct mlx5e_priv *priv = netdev_priv(dev);
3538
3539         switch (cmd) {
3540         case SIOCSHWTSTAMP:
3541                 return mlx5e_hwstamp_set(priv, ifr);
3542         case SIOCGHWTSTAMP:
3543                 return mlx5e_hwstamp_get(priv, ifr);
3544         default:
3545                 return -EOPNOTSUPP;
3546         }
3547 }
3548
3549 #ifdef CONFIG_MLX5_ESWITCH
3550 static int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
3551 {
3552         struct mlx5e_priv *priv = netdev_priv(dev);
3553         struct mlx5_core_dev *mdev = priv->mdev;
3554
3555         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
3556 }
3557
3558 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos,
3559                              __be16 vlan_proto)
3560 {
3561         struct mlx5e_priv *priv = netdev_priv(dev);
3562         struct mlx5_core_dev *mdev = priv->mdev;
3563
3564         if (vlan_proto != htons(ETH_P_8021Q))
3565                 return -EPROTONOSUPPORT;
3566
3567         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
3568                                            vlan, qos);
3569 }
3570
3571 static int mlx5e_set_vf_spoofchk(struct net_device *dev, int vf, bool setting)
3572 {
3573         struct mlx5e_priv *priv = netdev_priv(dev);
3574         struct mlx5_core_dev *mdev = priv->mdev;
3575
3576         return mlx5_eswitch_set_vport_spoofchk(mdev->priv.eswitch, vf + 1, setting);
3577 }
3578
3579 static int mlx5e_set_vf_trust(struct net_device *dev, int vf, bool setting)
3580 {
3581         struct mlx5e_priv *priv = netdev_priv(dev);
3582         struct mlx5_core_dev *mdev = priv->mdev;
3583
3584         return mlx5_eswitch_set_vport_trust(mdev->priv.eswitch, vf + 1, setting);
3585 }
3586
3587 static int mlx5e_set_vf_rate(struct net_device *dev, int vf, int min_tx_rate,
3588                              int max_tx_rate)
3589 {
3590         struct mlx5e_priv *priv = netdev_priv(dev);
3591         struct mlx5_core_dev *mdev = priv->mdev;
3592
3593         return mlx5_eswitch_set_vport_rate(mdev->priv.eswitch, vf + 1,
3594                                            max_tx_rate, min_tx_rate);
3595 }
3596
3597 static int mlx5_vport_link2ifla(u8 esw_link)
3598 {
3599         switch (esw_link) {
3600         case MLX5_ESW_VPORT_ADMIN_STATE_DOWN:
3601                 return IFLA_VF_LINK_STATE_DISABLE;
3602         case MLX5_ESW_VPORT_ADMIN_STATE_UP:
3603                 return IFLA_VF_LINK_STATE_ENABLE;
3604         }
3605         return IFLA_VF_LINK_STATE_AUTO;
3606 }
3607
3608 static int mlx5_ifla_link2vport(u8 ifla_link)
3609 {
3610         switch (ifla_link) {
3611         case IFLA_VF_LINK_STATE_DISABLE:
3612                 return MLX5_ESW_VPORT_ADMIN_STATE_DOWN;
3613         case IFLA_VF_LINK_STATE_ENABLE:
3614                 return MLX5_ESW_VPORT_ADMIN_STATE_UP;
3615         }
3616         return MLX5_ESW_VPORT_ADMIN_STATE_AUTO;
3617 }
3618
3619 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
3620                                    int link_state)
3621 {
3622         struct mlx5e_priv *priv = netdev_priv(dev);
3623         struct mlx5_core_dev *mdev = priv->mdev;
3624
3625         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
3626                                             mlx5_ifla_link2vport(link_state));
3627 }
3628
3629 static int mlx5e_get_vf_config(struct net_device *dev,
3630                                int vf, struct ifla_vf_info *ivi)
3631 {
3632         struct mlx5e_priv *priv = netdev_priv(dev);
3633         struct mlx5_core_dev *mdev = priv->mdev;
3634         int err;
3635
3636         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
3637         if (err)
3638                 return err;
3639         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
3640         return 0;
3641 }
3642
3643 static int mlx5e_get_vf_stats(struct net_device *dev,
3644                               int vf, struct ifla_vf_stats *vf_stats)
3645 {
3646         struct mlx5e_priv *priv = netdev_priv(dev);
3647         struct mlx5_core_dev *mdev = priv->mdev;
3648
3649         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
3650                                             vf_stats);
3651 }
3652 #endif
3653
3654 static void mlx5e_add_vxlan_port(struct net_device *netdev,
3655                                  struct udp_tunnel_info *ti)
3656 {
3657         struct mlx5e_priv *priv = netdev_priv(netdev);
3658
3659         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3660                 return;
3661
3662         if (!mlx5e_vxlan_allowed(priv->mdev))
3663                 return;
3664
3665         mlx5e_vxlan_queue_work(priv, ti->sa_family, be16_to_cpu(ti->port), 1);
3666 }
3667
3668 static void mlx5e_del_vxlan_port(struct net_device *netdev,
3669                                  struct udp_tunnel_info *ti)
3670 {
3671         struct mlx5e_priv *priv = netdev_priv(netdev);
3672
3673         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
3674                 return;
3675
3676         if (!mlx5e_vxlan_allowed(priv->mdev))
3677                 return;
3678
3679         mlx5e_vxlan_queue_work(priv, ti->sa_family, be16_to_cpu(ti->port), 0);
3680 }
3681
3682 static netdev_features_t mlx5e_tunnel_features_check(struct mlx5e_priv *priv,
3683                                                      struct sk_buff *skb,
3684                                                      netdev_features_t features)
3685 {
3686         unsigned int offset = 0;
3687         struct udphdr *udph;
3688         u8 proto;
3689         u16 port;
3690
3691         switch (vlan_get_protocol(skb)) {
3692         case htons(ETH_P_IP):
3693                 proto = ip_hdr(skb)->protocol;
3694                 break;
3695         case htons(ETH_P_IPV6):
3696                 proto = ipv6_find_hdr(skb, &offset, -1, NULL, NULL);
3697                 break;
3698         default:
3699                 goto out;
3700         }
3701
3702         switch (proto) {
3703         case IPPROTO_GRE:
3704                 return features;
3705         case IPPROTO_UDP:
3706                 udph = udp_hdr(skb);
3707                 port = be16_to_cpu(udph->dest);
3708
3709                 /* Verify if UDP port is being offloaded by HW */
3710                 if (mlx5e_vxlan_lookup_port(priv, port))
3711                         return features;
3712         }
3713
3714 out:
3715         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
3716         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3717 }
3718
3719 static netdev_features_t mlx5e_features_check(struct sk_buff *skb,
3720                                               struct net_device *netdev,
3721                                               netdev_features_t features)
3722 {
3723         struct mlx5e_priv *priv = netdev_priv(netdev);
3724
3725         features = vlan_features_check(skb, features);
3726         features = vxlan_features_check(skb, features);
3727
3728 #ifdef CONFIG_MLX5_EN_IPSEC
3729         if (mlx5e_ipsec_feature_check(skb, netdev, features))
3730                 return features;
3731 #endif
3732
3733         /* Validate if the tunneled packet is being offloaded by HW */
3734         if (skb->encapsulation &&
3735             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
3736                 return mlx5e_tunnel_features_check(priv, skb, features);
3737
3738         return features;
3739 }
3740
3741 static void mlx5e_tx_timeout(struct net_device *dev)
3742 {
3743         struct mlx5e_priv *priv = netdev_priv(dev);
3744         bool sched_work = false;
3745         int i;
3746
3747         netdev_err(dev, "TX timeout detected\n");
3748
3749         for (i = 0; i < priv->channels.num * priv->channels.params.num_tc; i++) {
3750                 struct mlx5e_txqsq *sq = priv->txq2sq[i];
3751
3752                 if (!netif_xmit_stopped(netdev_get_tx_queue(dev, i)))
3753                         continue;
3754                 sched_work = true;
3755                 clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
3756                 netdev_err(dev, "TX timeout on queue: %d, SQ: 0x%x, CQ: 0x%x, SQ Cons: 0x%x SQ Prod: 0x%x\n",
3757                            i, sq->sqn, sq->cq.mcq.cqn, sq->cc, sq->pc);
3758         }
3759
3760         if (sched_work && test_bit(MLX5E_STATE_OPENED, &priv->state))
3761                 schedule_work(&priv->tx_timeout_work);
3762 }
3763
3764 static int mlx5e_xdp_set(struct net_device *netdev, struct bpf_prog *prog)
3765 {
3766         struct mlx5e_priv *priv = netdev_priv(netdev);
3767         struct bpf_prog *old_prog;
3768         int err = 0;
3769         bool reset, was_opened;
3770         int i;
3771
3772         mutex_lock(&priv->state_lock);
3773
3774         if ((netdev->features & NETIF_F_LRO) && prog) {
3775                 netdev_warn(netdev, "can't set XDP while LRO is on, disable LRO first\n");
3776                 err = -EINVAL;
3777                 goto unlock;
3778         }
3779
3780         if ((netdev->features & NETIF_F_HW_ESP) && prog) {
3781                 netdev_warn(netdev, "can't set XDP with IPSec offload\n");
3782                 err = -EINVAL;
3783                 goto unlock;
3784         }
3785
3786         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
3787         /* no need for full reset when exchanging programs */
3788         reset = (!priv->channels.params.xdp_prog || !prog);
3789
3790         if (was_opened && reset)
3791                 mlx5e_close_locked(netdev);
3792         if (was_opened && !reset) {
3793                 /* num_channels is invariant here, so we can take the
3794                  * batched reference right upfront.
3795                  */
3796                 prog = bpf_prog_add(prog, priv->channels.num);
3797                 if (IS_ERR(prog)) {
3798                         err = PTR_ERR(prog);
3799                         goto unlock;
3800                 }
3801         }
3802
3803         /* exchange programs, extra prog reference we got from caller
3804          * as long as we don't fail from this point onwards.
3805          */
3806         old_prog = xchg(&priv->channels.params.xdp_prog, prog);
3807         if (old_prog)
3808                 bpf_prog_put(old_prog);
3809
3810         if (reset) /* change RQ type according to priv->xdp_prog */
3811                 mlx5e_set_rq_params(priv->mdev, &priv->channels.params);
3812
3813         if (was_opened && reset)
3814                 mlx5e_open_locked(netdev);
3815
3816         if (!test_bit(MLX5E_STATE_OPENED, &priv->state) || reset)
3817                 goto unlock;
3818
3819         /* exchanging programs w/o reset, we update ref counts on behalf
3820          * of the channels RQs here.
3821          */
3822         for (i = 0; i < priv->channels.num; i++) {
3823                 struct mlx5e_channel *c = priv->channels.c[i];
3824
3825                 clear_bit(MLX5E_RQ_STATE_ENABLED, &c->rq.state);
3826                 napi_synchronize(&c->napi);
3827                 /* prevent mlx5e_poll_rx_cq from accessing rq->xdp_prog */
3828
3829                 old_prog = xchg(&c->rq.xdp_prog, prog);
3830
3831                 set_bit(MLX5E_RQ_STATE_ENABLED, &c->rq.state);
3832                 /* napi_schedule in case we have missed anything */
3833                 napi_schedule(&c->napi);
3834
3835                 if (old_prog)
3836                         bpf_prog_put(old_prog);
3837         }
3838
3839 unlock:
3840         mutex_unlock(&priv->state_lock);
3841         return err;
3842 }
3843
3844 static u32 mlx5e_xdp_query(struct net_device *dev)
3845 {
3846         struct mlx5e_priv *priv = netdev_priv(dev);
3847         const struct bpf_prog *xdp_prog;
3848         u32 prog_id = 0;
3849
3850         mutex_lock(&priv->state_lock);
3851         xdp_prog = priv->channels.params.xdp_prog;
3852         if (xdp_prog)
3853                 prog_id = xdp_prog->aux->id;
3854         mutex_unlock(&priv->state_lock);
3855
3856         return prog_id;
3857 }
3858
3859 static int mlx5e_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3860 {
3861         switch (xdp->command) {
3862         case XDP_SETUP_PROG:
3863                 return mlx5e_xdp_set(dev, xdp->prog);
3864         case XDP_QUERY_PROG:
3865                 xdp->prog_id = mlx5e_xdp_query(dev);
3866                 xdp->prog_attached = !!xdp->prog_id;
3867                 return 0;
3868         default:
3869                 return -EINVAL;
3870         }
3871 }
3872
3873 #ifdef CONFIG_NET_POLL_CONTROLLER
3874 /* Fake "interrupt" called by netpoll (eg netconsole) to send skbs without
3875  * reenabling interrupts.
3876  */
3877 static void mlx5e_netpoll(struct net_device *dev)
3878 {
3879         struct mlx5e_priv *priv = netdev_priv(dev);
3880         struct mlx5e_channels *chs = &priv->channels;
3881
3882         int i;
3883
3884         for (i = 0; i < chs->num; i++)
3885                 napi_schedule(&chs->c[i]->napi);
3886 }
3887 #endif
3888
3889 static const struct net_device_ops mlx5e_netdev_ops = {
3890         .ndo_open                = mlx5e_open,
3891         .ndo_stop                = mlx5e_close,
3892         .ndo_start_xmit          = mlx5e_xmit,
3893         .ndo_setup_tc            = mlx5e_setup_tc,
3894         .ndo_select_queue        = mlx5e_select_queue,
3895         .ndo_get_stats64         = mlx5e_get_stats,
3896         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
3897         .ndo_set_mac_address     = mlx5e_set_mac,
3898         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
3899         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
3900         .ndo_set_features        = mlx5e_set_features,
3901         .ndo_fix_features        = mlx5e_fix_features,
3902         .ndo_change_mtu          = mlx5e_change_mtu,
3903         .ndo_do_ioctl            = mlx5e_ioctl,
3904         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
3905         .ndo_udp_tunnel_add      = mlx5e_add_vxlan_port,
3906         .ndo_udp_tunnel_del      = mlx5e_del_vxlan_port,
3907         .ndo_features_check      = mlx5e_features_check,
3908 #ifdef CONFIG_RFS_ACCEL
3909         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
3910 #endif
3911         .ndo_tx_timeout          = mlx5e_tx_timeout,
3912         .ndo_bpf                 = mlx5e_xdp,
3913 #ifdef CONFIG_NET_POLL_CONTROLLER
3914         .ndo_poll_controller     = mlx5e_netpoll,
3915 #endif
3916 #ifdef CONFIG_MLX5_ESWITCH
3917         /* SRIOV E-Switch NDOs */
3918         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
3919         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
3920         .ndo_set_vf_spoofchk     = mlx5e_set_vf_spoofchk,
3921         .ndo_set_vf_trust        = mlx5e_set_vf_trust,
3922         .ndo_set_vf_rate         = mlx5e_set_vf_rate,
3923         .ndo_get_vf_config       = mlx5e_get_vf_config,
3924         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
3925         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
3926         .ndo_has_offload_stats   = mlx5e_has_offload_stats,
3927         .ndo_get_offload_stats   = mlx5e_get_offload_stats,
3928 #endif
3929 };
3930
3931 static int mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
3932 {
3933         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
3934                 return -EOPNOTSUPP;
3935         if (!MLX5_CAP_GEN(mdev, eth_net_offloads) ||
3936             !MLX5_CAP_GEN(mdev, nic_flow_table) ||
3937             !MLX5_CAP_ETH(mdev, csum_cap) ||
3938             !MLX5_CAP_ETH(mdev, max_lso_cap) ||
3939             !MLX5_CAP_ETH(mdev, vlan_cap) ||
3940             !MLX5_CAP_ETH(mdev, rss_ind_tbl_cap) ||
3941             MLX5_CAP_FLOWTABLE(mdev,
3942                                flow_table_properties_nic_receive.max_ft_level)
3943                                < 3) {
3944                 mlx5_core_warn(mdev,
3945                                "Not creating net device, some required device capabilities are missing\n");
3946                 return -EOPNOTSUPP;
3947         }
3948         if (!MLX5_CAP_ETH(mdev, self_lb_en_modifiable))
3949                 mlx5_core_warn(mdev, "Self loop back prevention is not supported\n");
3950         if (!MLX5_CAP_GEN(mdev, cq_moderation))
3951                 mlx5_core_warn(mdev, "CQ moderation is not supported\n");
3952
3953         return 0;
3954 }
3955
3956 u16 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev)
3957 {
3958         int bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
3959
3960         return bf_buf_size -
3961                sizeof(struct mlx5e_tx_wqe) +
3962                2 /*sizeof(mlx5e_tx_wqe.inline_hdr_start)*/;
3963 }
3964
3965 void mlx5e_build_default_indir_rqt(u32 *indirection_rqt, int len,
3966                                    int num_channels)
3967 {
3968         int i;
3969
3970         for (i = 0; i < len; i++)
3971                 indirection_rqt[i] = i % num_channels;
3972 }
3973
3974 static int mlx5e_get_pci_bw(struct mlx5_core_dev *mdev, u32 *pci_bw)
3975 {
3976         enum pcie_link_width width;
3977         enum pci_bus_speed speed;
3978         int err = 0;
3979
3980         err = pcie_get_minimum_link(mdev->pdev, &speed, &width);
3981         if (err)
3982                 return err;
3983
3984         if (speed == PCI_SPEED_UNKNOWN || width == PCIE_LNK_WIDTH_UNKNOWN)
3985                 return -EINVAL;
3986
3987         switch (speed) {
3988         case PCIE_SPEED_2_5GT:
3989                 *pci_bw = 2500 * width;
3990                 break;
3991         case PCIE_SPEED_5_0GT:
3992                 *pci_bw = 5000 * width;
3993                 break;
3994         case PCIE_SPEED_8_0GT:
3995                 *pci_bw = 8000 * width;
3996                 break;
3997         default:
3998                 return -EINVAL;
3999         }
4000
4001         return 0;
4002 }
4003
4004 static bool cqe_compress_heuristic(u32 link_speed, u32 pci_bw)
4005 {
4006         return (link_speed && pci_bw &&
4007                 (pci_bw < 40000) && (pci_bw < link_speed));
4008 }
4009
4010 static bool hw_lro_heuristic(u32 link_speed, u32 pci_bw)
4011 {
4012         return !(link_speed && pci_bw &&
4013                  (pci_bw <= 16000) && (pci_bw < link_speed));
4014 }
4015
4016 void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4017 {
4018         params->tx_cq_moderation.cq_period_mode = cq_period_mode;
4019
4020         params->tx_cq_moderation.pkts =
4021                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
4022         params->tx_cq_moderation.usec =
4023                 MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
4024
4025         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4026                 params->tx_cq_moderation.usec =
4027                         MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE;
4028
4029         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_TX_CQE_BASED_MODER,
4030                         params->tx_cq_moderation.cq_period_mode ==
4031                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4032 }
4033
4034 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4035 {
4036         params->rx_cq_moderation.cq_period_mode = cq_period_mode;
4037
4038         params->rx_cq_moderation.pkts =
4039                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
4040         params->rx_cq_moderation.usec =
4041                 MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
4042
4043         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4044                 params->rx_cq_moderation.usec =
4045                         MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE;
4046
4047         if (params->rx_am_enabled)
4048                 params->rx_cq_moderation =
4049                         mlx5e_am_get_def_profile(cq_period_mode);
4050
4051         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_BASED_MODER,
4052                         params->rx_cq_moderation.cq_period_mode ==
4053                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4054 }
4055
4056 u32 mlx5e_choose_lro_timeout(struct mlx5_core_dev *mdev, u32 wanted_timeout)
4057 {
4058         int i;
4059
4060         /* The supported periods are organized in ascending order */
4061         for (i = 0; i < MLX5E_LRO_TIMEOUT_ARR_SIZE - 1; i++)
4062                 if (MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]) >= wanted_timeout)
4063                         break;
4064
4065         return MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]);
4066 }
4067
4068 void mlx5e_build_nic_params(struct mlx5_core_dev *mdev,
4069                             struct mlx5e_params *params,
4070                             u16 max_channels)
4071 {
4072         u8 cq_period_mode = 0;
4073         u32 link_speed = 0;
4074         u32 pci_bw = 0;
4075
4076         params->num_channels = max_channels;
4077         params->num_tc       = 1;
4078
4079         mlx5e_get_max_linkspeed(mdev, &link_speed);
4080         mlx5e_get_pci_bw(mdev, &pci_bw);
4081         mlx5_core_dbg(mdev, "Max link speed = %d, PCI BW = %d\n",
4082                       link_speed, pci_bw);
4083
4084         /* SQ */
4085         params->log_sq_size = is_kdump_kernel() ?
4086                 MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE :
4087                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
4088
4089         /* set CQE compression */
4090         params->rx_cqe_compress_def = false;
4091         if (MLX5_CAP_GEN(mdev, cqe_compression) &&
4092             MLX5_CAP_GEN(mdev, vport_group_manager))
4093                 params->rx_cqe_compress_def = cqe_compress_heuristic(link_speed, pci_bw);
4094
4095         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS, params->rx_cqe_compress_def);
4096
4097         /* RQ */
4098         mlx5e_set_rq_params(mdev, params);
4099
4100         /* HW LRO */
4101
4102         /* TODO: && MLX5_CAP_ETH(mdev, lro_cap) */
4103         if (params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
4104                 params->lro_en = hw_lro_heuristic(link_speed, pci_bw);
4105         params->lro_timeout = mlx5e_choose_lro_timeout(mdev, MLX5E_DEFAULT_LRO_TIMEOUT);
4106
4107         /* CQ moderation params */
4108         cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
4109                         MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
4110                         MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
4111         params->rx_am_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
4112         mlx5e_set_rx_cq_mode_params(params, cq_period_mode);
4113         mlx5e_set_tx_cq_mode_params(params, cq_period_mode);
4114
4115         /* TX inline */
4116         params->tx_max_inline = mlx5e_get_max_inline_cap(mdev);
4117         params->tx_min_inline_mode = mlx5e_params_calculate_tx_min_inline(mdev);
4118
4119         /* RSS */
4120         params->rss_hfunc = ETH_RSS_HASH_XOR;
4121         netdev_rss_key_fill(params->toeplitz_hash_key, sizeof(params->toeplitz_hash_key));
4122         mlx5e_build_default_indir_rqt(params->indirection_rqt,
4123                                       MLX5E_INDIR_RQT_SIZE, max_channels);
4124 }
4125
4126 static void mlx5e_build_nic_netdev_priv(struct mlx5_core_dev *mdev,
4127                                         struct net_device *netdev,
4128                                         const struct mlx5e_profile *profile,
4129                                         void *ppriv)
4130 {
4131         struct mlx5e_priv *priv = netdev_priv(netdev);
4132
4133         priv->mdev        = mdev;
4134         priv->netdev      = netdev;
4135         priv->profile     = profile;
4136         priv->ppriv       = ppriv;
4137         priv->msglevel    = MLX5E_MSG_LEVEL;
4138         priv->hard_mtu = MLX5E_ETH_HARD_MTU;
4139
4140         mlx5e_build_nic_params(mdev, &priv->channels.params, profile->max_nch(mdev));
4141
4142         mutex_init(&priv->state_lock);
4143
4144         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
4145         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
4146         INIT_WORK(&priv->tx_timeout_work, mlx5e_tx_timeout_work);
4147         INIT_DELAYED_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
4148
4149         mlx5e_timestamp_init(priv);
4150 }
4151
4152 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
4153 {
4154         struct mlx5e_priv *priv = netdev_priv(netdev);
4155
4156         mlx5_query_nic_vport_mac_address(priv->mdev, 0, netdev->dev_addr);
4157         if (is_zero_ether_addr(netdev->dev_addr) &&
4158             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
4159                 eth_hw_addr_random(netdev);
4160                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
4161         }
4162 }
4163
4164 #if IS_ENABLED(CONFIG_NET_SWITCHDEV) && IS_ENABLED(CONFIG_MLX5_ESWITCH)
4165 static const struct switchdev_ops mlx5e_switchdev_ops = {
4166         .switchdev_port_attr_get        = mlx5e_attr_get,
4167 };
4168 #endif
4169
4170 static void mlx5e_build_nic_netdev(struct net_device *netdev)
4171 {
4172         struct mlx5e_priv *priv = netdev_priv(netdev);
4173         struct mlx5_core_dev *mdev = priv->mdev;
4174         bool fcs_supported;
4175         bool fcs_enabled;
4176
4177         SET_NETDEV_DEV(netdev, &mdev->pdev->dev);
4178
4179         netdev->netdev_ops = &mlx5e_netdev_ops;
4180
4181 #ifdef CONFIG_MLX5_CORE_EN_DCB
4182         if (MLX5_CAP_GEN(mdev, vport_group_manager) && MLX5_CAP_GEN(mdev, qos))
4183                 netdev->dcbnl_ops = &mlx5e_dcbnl_ops;
4184 #endif
4185
4186         netdev->watchdog_timeo    = 15 * HZ;
4187
4188         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
4189
4190         netdev->vlan_features    |= NETIF_F_SG;
4191         netdev->vlan_features    |= NETIF_F_IP_CSUM;
4192         netdev->vlan_features    |= NETIF_F_IPV6_CSUM;
4193         netdev->vlan_features    |= NETIF_F_GRO;
4194         netdev->vlan_features    |= NETIF_F_TSO;
4195         netdev->vlan_features    |= NETIF_F_TSO6;
4196         netdev->vlan_features    |= NETIF_F_RXCSUM;
4197         netdev->vlan_features    |= NETIF_F_RXHASH;
4198
4199         if (!!MLX5_CAP_ETH(mdev, lro_cap))
4200                 netdev->vlan_features    |= NETIF_F_LRO;
4201
4202         netdev->hw_features       = netdev->vlan_features;
4203         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
4204         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
4205         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
4206         netdev->hw_features      |= NETIF_F_HW_VLAN_STAG_TX;
4207
4208         if (mlx5e_vxlan_allowed(mdev) || MLX5_CAP_ETH(mdev, tunnel_stateless_gre)) {
4209                 netdev->hw_features     |= NETIF_F_GSO_PARTIAL;
4210                 netdev->hw_enc_features |= NETIF_F_IP_CSUM;
4211                 netdev->hw_enc_features |= NETIF_F_IPV6_CSUM;
4212                 netdev->hw_enc_features |= NETIF_F_TSO;
4213                 netdev->hw_enc_features |= NETIF_F_TSO6;
4214                 netdev->hw_enc_features |= NETIF_F_GSO_PARTIAL;
4215         }
4216
4217         if (mlx5e_vxlan_allowed(mdev)) {
4218                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL |
4219                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4220                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL |
4221                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4222                 netdev->gso_partial_features = NETIF_F_GSO_UDP_TUNNEL_CSUM;
4223         }
4224
4225         if (MLX5_CAP_ETH(mdev, tunnel_stateless_gre)) {
4226                 netdev->hw_features     |= NETIF_F_GSO_GRE |
4227                                            NETIF_F_GSO_GRE_CSUM;
4228                 netdev->hw_enc_features |= NETIF_F_GSO_GRE |
4229                                            NETIF_F_GSO_GRE_CSUM;
4230                 netdev->gso_partial_features |= NETIF_F_GSO_GRE |
4231                                                 NETIF_F_GSO_GRE_CSUM;
4232         }
4233
4234         mlx5_query_port_fcs(mdev, &fcs_supported, &fcs_enabled);
4235
4236         if (fcs_supported)
4237                 netdev->hw_features |= NETIF_F_RXALL;
4238
4239         if (MLX5_CAP_ETH(mdev, scatter_fcs))
4240                 netdev->hw_features |= NETIF_F_RXFCS;
4241
4242         netdev->features          = netdev->hw_features;
4243         if (!priv->channels.params.lro_en)
4244                 netdev->features  &= ~NETIF_F_LRO;
4245
4246         if (fcs_enabled)
4247                 netdev->features  &= ~NETIF_F_RXALL;
4248
4249         if (!priv->channels.params.scatter_fcs_en)
4250                 netdev->features  &= ~NETIF_F_RXFCS;
4251
4252 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
4253         if (FT_CAP(flow_modify_en) &&
4254             FT_CAP(modify_root) &&
4255             FT_CAP(identified_miss_table_mode) &&
4256             FT_CAP(flow_table_modify)) {
4257                 netdev->hw_features      |= NETIF_F_HW_TC;
4258 #ifdef CONFIG_RFS_ACCEL
4259                 netdev->hw_features      |= NETIF_F_NTUPLE;
4260 #endif
4261         }
4262
4263         netdev->features         |= NETIF_F_HIGHDMA;
4264         netdev->features         |= NETIF_F_HW_VLAN_STAG_FILTER;
4265
4266         netdev->priv_flags       |= IFF_UNICAST_FLT;
4267
4268         mlx5e_set_netdev_dev_addr(netdev);
4269
4270 #if IS_ENABLED(CONFIG_NET_SWITCHDEV) && IS_ENABLED(CONFIG_MLX5_ESWITCH)
4271         if (MLX5_VPORT_MANAGER(mdev))
4272                 netdev->switchdev_ops = &mlx5e_switchdev_ops;
4273 #endif
4274
4275         mlx5e_ipsec_build_netdev(priv);
4276 }
4277
4278 static void mlx5e_create_q_counter(struct mlx5e_priv *priv)
4279 {
4280         struct mlx5_core_dev *mdev = priv->mdev;
4281         int err;
4282
4283         err = mlx5_core_alloc_q_counter(mdev, &priv->q_counter);
4284         if (err) {
4285                 mlx5_core_warn(mdev, "alloc queue counter failed, %d\n", err);
4286                 priv->q_counter = 0;
4287         }
4288 }
4289
4290 static void mlx5e_destroy_q_counter(struct mlx5e_priv *priv)
4291 {
4292         if (!priv->q_counter)
4293                 return;
4294
4295         mlx5_core_dealloc_q_counter(priv->mdev, priv->q_counter);
4296 }
4297
4298 static void mlx5e_nic_init(struct mlx5_core_dev *mdev,
4299                            struct net_device *netdev,
4300                            const struct mlx5e_profile *profile,
4301                            void *ppriv)
4302 {
4303         struct mlx5e_priv *priv = netdev_priv(netdev);
4304         int err;
4305
4306         mlx5e_build_nic_netdev_priv(mdev, netdev, profile, ppriv);
4307         err = mlx5e_ipsec_init(priv);
4308         if (err)
4309                 mlx5_core_err(mdev, "IPSec initialization failed, %d\n", err);
4310         mlx5e_build_nic_netdev(netdev);
4311         mlx5e_vxlan_init(priv);
4312 }
4313
4314 static void mlx5e_nic_cleanup(struct mlx5e_priv *priv)
4315 {
4316         mlx5e_ipsec_cleanup(priv);
4317         mlx5e_vxlan_cleanup(priv);
4318
4319         if (priv->channels.params.xdp_prog)
4320                 bpf_prog_put(priv->channels.params.xdp_prog);
4321 }
4322
4323 static int mlx5e_init_nic_rx(struct mlx5e_priv *priv)
4324 {
4325         struct mlx5_core_dev *mdev = priv->mdev;
4326         int err;
4327
4328         err = mlx5e_create_indirect_rqt(priv);
4329         if (err)
4330                 return err;
4331
4332         err = mlx5e_create_direct_rqts(priv);
4333         if (err)
4334                 goto err_destroy_indirect_rqts;
4335
4336         err = mlx5e_create_indirect_tirs(priv);
4337         if (err)
4338                 goto err_destroy_direct_rqts;
4339
4340         err = mlx5e_create_direct_tirs(priv);
4341         if (err)
4342                 goto err_destroy_indirect_tirs;
4343
4344         err = mlx5e_create_flow_steering(priv);
4345         if (err) {
4346                 mlx5_core_warn(mdev, "create flow steering failed, %d\n", err);
4347                 goto err_destroy_direct_tirs;
4348         }
4349
4350         err = mlx5e_tc_init(priv);
4351         if (err)
4352                 goto err_destroy_flow_steering;
4353
4354         return 0;
4355
4356 err_destroy_flow_steering:
4357         mlx5e_destroy_flow_steering(priv);
4358 err_destroy_direct_tirs:
4359         mlx5e_destroy_direct_tirs(priv);
4360 err_destroy_indirect_tirs:
4361         mlx5e_destroy_indirect_tirs(priv);
4362 err_destroy_direct_rqts:
4363         mlx5e_destroy_direct_rqts(priv);
4364 err_destroy_indirect_rqts:
4365         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
4366         return err;
4367 }
4368
4369 static void mlx5e_cleanup_nic_rx(struct mlx5e_priv *priv)
4370 {
4371         mlx5e_tc_cleanup(priv);
4372         mlx5e_destroy_flow_steering(priv);
4373         mlx5e_destroy_direct_tirs(priv);
4374         mlx5e_destroy_indirect_tirs(priv);
4375         mlx5e_destroy_direct_rqts(priv);
4376         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
4377 }
4378
4379 static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
4380 {
4381         int err;
4382
4383         err = mlx5e_create_tises(priv);
4384         if (err) {
4385                 mlx5_core_warn(priv->mdev, "create tises failed, %d\n", err);
4386                 return err;
4387         }
4388
4389 #ifdef CONFIG_MLX5_CORE_EN_DCB
4390         mlx5e_dcbnl_initialize(priv);
4391 #endif
4392         return 0;
4393 }
4394
4395 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
4396 {
4397         struct net_device *netdev = priv->netdev;
4398         struct mlx5_core_dev *mdev = priv->mdev;
4399         u16 max_mtu;
4400
4401         mlx5e_init_l2_addr(priv);
4402
4403         /* Marking the link as currently not needed by the Driver */
4404         if (!netif_running(netdev))
4405                 mlx5_set_port_admin_status(mdev, MLX5_PORT_DOWN);
4406
4407         /* MTU range: 68 - hw-specific max */
4408         netdev->min_mtu = ETH_MIN_MTU;
4409         mlx5_query_port_max_mtu(priv->mdev, &max_mtu, 1);
4410         netdev->max_mtu = MLX5E_HW2SW_MTU(priv, max_mtu);
4411         mlx5e_set_dev_port_mtu(priv);
4412
4413         mlx5_lag_add(mdev, netdev);
4414
4415         mlx5e_enable_async_events(priv);
4416
4417         if (MLX5_VPORT_MANAGER(priv->mdev))
4418                 mlx5e_register_vport_reps(priv);
4419
4420         if (netdev->reg_state != NETREG_REGISTERED)
4421                 return;
4422 #ifdef CONFIG_MLX5_CORE_EN_DCB
4423         mlx5e_dcbnl_init_app(priv);
4424 #endif
4425         /* Device already registered: sync netdev system state */
4426         if (mlx5e_vxlan_allowed(mdev)) {
4427                 rtnl_lock();
4428                 udp_tunnel_get_rx_info(netdev);
4429                 rtnl_unlock();
4430         }
4431
4432         queue_work(priv->wq, &priv->set_rx_mode_work);
4433
4434         rtnl_lock();
4435         if (netif_running(netdev))
4436                 mlx5e_open(netdev);
4437         netif_device_attach(netdev);
4438         rtnl_unlock();
4439 }
4440
4441 static void mlx5e_nic_disable(struct mlx5e_priv *priv)
4442 {
4443         struct mlx5_core_dev *mdev = priv->mdev;
4444
4445 #ifdef CONFIG_MLX5_CORE_EN_DCB
4446         if (priv->netdev->reg_state == NETREG_REGISTERED)
4447                 mlx5e_dcbnl_delete_app(priv);
4448 #endif
4449
4450         rtnl_lock();
4451         if (netif_running(priv->netdev))
4452                 mlx5e_close(priv->netdev);
4453         netif_device_detach(priv->netdev);
4454         rtnl_unlock();
4455
4456         queue_work(priv->wq, &priv->set_rx_mode_work);
4457
4458         if (MLX5_VPORT_MANAGER(priv->mdev))
4459                 mlx5e_unregister_vport_reps(priv);
4460
4461         mlx5e_disable_async_events(priv);
4462         mlx5_lag_remove(mdev);
4463 }
4464
4465 static const struct mlx5e_profile mlx5e_nic_profile = {
4466         .init              = mlx5e_nic_init,
4467         .cleanup           = mlx5e_nic_cleanup,
4468         .init_rx           = mlx5e_init_nic_rx,
4469         .cleanup_rx        = mlx5e_cleanup_nic_rx,
4470         .init_tx           = mlx5e_init_nic_tx,
4471         .cleanup_tx        = mlx5e_cleanup_nic_tx,
4472         .enable            = mlx5e_nic_enable,
4473         .disable           = mlx5e_nic_disable,
4474         .update_stats      = mlx5e_update_ndo_stats,
4475         .max_nch           = mlx5e_get_max_num_channels,
4476         .update_carrier    = mlx5e_update_carrier,
4477         .rx_handlers.handle_rx_cqe       = mlx5e_handle_rx_cqe,
4478         .rx_handlers.handle_rx_cqe_mpwqe = mlx5e_handle_rx_cqe_mpwrq,
4479         .max_tc            = MLX5E_MAX_NUM_TC,
4480 };
4481
4482 /* mlx5e generic netdev management API (move to en_common.c) */
4483
4484 struct net_device *mlx5e_create_netdev(struct mlx5_core_dev *mdev,
4485                                        const struct mlx5e_profile *profile,
4486                                        void *ppriv)
4487 {
4488         int nch = profile->max_nch(mdev);
4489         struct net_device *netdev;
4490         struct mlx5e_priv *priv;
4491
4492         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv),
4493                                     nch * profile->max_tc,
4494                                     nch);
4495         if (!netdev) {
4496                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
4497                 return NULL;
4498         }
4499
4500 #ifdef CONFIG_RFS_ACCEL
4501         netdev->rx_cpu_rmap = mdev->rmap;
4502 #endif
4503
4504         profile->init(mdev, netdev, profile, ppriv);
4505
4506         netif_carrier_off(netdev);
4507
4508         priv = netdev_priv(netdev);
4509
4510         priv->wq = create_singlethread_workqueue("mlx5e");
4511         if (!priv->wq)
4512                 goto err_cleanup_nic;
4513
4514         return netdev;
4515
4516 err_cleanup_nic:
4517         if (profile->cleanup)
4518                 profile->cleanup(priv);
4519         free_netdev(netdev);
4520
4521         return NULL;
4522 }
4523
4524 int mlx5e_attach_netdev(struct mlx5e_priv *priv)
4525 {
4526         struct mlx5_core_dev *mdev = priv->mdev;
4527         const struct mlx5e_profile *profile;
4528         int err;
4529
4530         profile = priv->profile;
4531         clear_bit(MLX5E_STATE_DESTROYING, &priv->state);
4532
4533         err = profile->init_tx(priv);
4534         if (err)
4535                 goto out;
4536
4537         err = mlx5e_open_drop_rq(mdev, &priv->drop_rq);
4538         if (err) {
4539                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
4540                 goto err_cleanup_tx;
4541         }
4542
4543         err = profile->init_rx(priv);
4544         if (err)
4545                 goto err_close_drop_rq;
4546
4547         mlx5e_create_q_counter(priv);
4548
4549         if (profile->enable)
4550                 profile->enable(priv);
4551
4552         return 0;
4553
4554 err_close_drop_rq:
4555         mlx5e_close_drop_rq(&priv->drop_rq);
4556
4557 err_cleanup_tx:
4558         profile->cleanup_tx(priv);
4559
4560 out:
4561         return err;
4562 }
4563
4564 void mlx5e_detach_netdev(struct mlx5e_priv *priv)
4565 {
4566         const struct mlx5e_profile *profile = priv->profile;
4567
4568         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
4569
4570         if (profile->disable)
4571                 profile->disable(priv);
4572         flush_workqueue(priv->wq);
4573
4574         mlx5e_destroy_q_counter(priv);
4575         profile->cleanup_rx(priv);
4576         mlx5e_close_drop_rq(&priv->drop_rq);
4577         profile->cleanup_tx(priv);
4578         cancel_delayed_work_sync(&priv->update_stats_work);
4579 }
4580
4581 void mlx5e_destroy_netdev(struct mlx5e_priv *priv)
4582 {
4583         const struct mlx5e_profile *profile = priv->profile;
4584         struct net_device *netdev = priv->netdev;
4585
4586         destroy_workqueue(priv->wq);
4587         if (profile->cleanup)
4588                 profile->cleanup(priv);
4589         free_netdev(netdev);
4590 }
4591
4592 /* mlx5e_attach and mlx5e_detach scope should be only creating/destroying
4593  * hardware contexts and to connect it to the current netdev.
4594  */
4595 static int mlx5e_attach(struct mlx5_core_dev *mdev, void *vpriv)
4596 {
4597         struct mlx5e_priv *priv = vpriv;
4598         struct net_device *netdev = priv->netdev;
4599         int err;
4600
4601         if (netif_device_present(netdev))
4602                 return 0;
4603
4604         err = mlx5e_create_mdev_resources(mdev);
4605         if (err)
4606                 return err;
4607
4608         err = mlx5e_attach_netdev(priv);
4609         if (err) {
4610                 mlx5e_destroy_mdev_resources(mdev);
4611                 return err;
4612         }
4613
4614         return 0;
4615 }
4616
4617 static void mlx5e_detach(struct mlx5_core_dev *mdev, void *vpriv)
4618 {
4619         struct mlx5e_priv *priv = vpriv;
4620         struct net_device *netdev = priv->netdev;
4621
4622         if (!netif_device_present(netdev))
4623                 return;
4624
4625         mlx5e_detach_netdev(priv);
4626         mlx5e_destroy_mdev_resources(mdev);
4627 }
4628
4629 static void *mlx5e_add(struct mlx5_core_dev *mdev)
4630 {
4631         struct net_device *netdev;
4632         void *rpriv = NULL;
4633         void *priv;
4634         int err;
4635
4636         err = mlx5e_check_required_hca_cap(mdev);
4637         if (err)
4638                 return NULL;
4639
4640 #ifdef CONFIG_MLX5_ESWITCH
4641         if (MLX5_VPORT_MANAGER(mdev)) {
4642                 rpriv = mlx5e_alloc_nic_rep_priv(mdev);
4643                 if (!rpriv) {
4644                         mlx5_core_warn(mdev, "Failed to alloc NIC rep priv data\n");
4645                         return NULL;
4646                 }
4647         }
4648 #endif
4649
4650         netdev = mlx5e_create_netdev(mdev, &mlx5e_nic_profile, rpriv);
4651         if (!netdev) {
4652                 mlx5_core_err(mdev, "mlx5e_create_netdev failed\n");
4653                 goto err_free_rpriv;
4654         }
4655
4656         priv = netdev_priv(netdev);
4657
4658         err = mlx5e_attach(mdev, priv);
4659         if (err) {
4660                 mlx5_core_err(mdev, "mlx5e_attach failed, %d\n", err);
4661                 goto err_destroy_netdev;
4662         }
4663
4664         err = register_netdev(netdev);
4665         if (err) {
4666                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
4667                 goto err_detach;
4668         }
4669
4670 #ifdef CONFIG_MLX5_CORE_EN_DCB
4671         mlx5e_dcbnl_init_app(priv);
4672 #endif
4673         return priv;
4674
4675 err_detach:
4676         mlx5e_detach(mdev, priv);
4677 err_destroy_netdev:
4678         mlx5e_destroy_netdev(priv);
4679 err_free_rpriv:
4680         kfree(rpriv);
4681         return NULL;
4682 }
4683
4684 static void mlx5e_remove(struct mlx5_core_dev *mdev, void *vpriv)
4685 {
4686         struct mlx5e_priv *priv = vpriv;
4687         void *ppriv = priv->ppriv;
4688
4689 #ifdef CONFIG_MLX5_CORE_EN_DCB
4690         mlx5e_dcbnl_delete_app(priv);
4691 #endif
4692         unregister_netdev(priv->netdev);
4693         mlx5e_detach(mdev, vpriv);
4694         mlx5e_destroy_netdev(priv);
4695         kfree(ppriv);
4696 }
4697
4698 static void *mlx5e_get_netdev(void *vpriv)
4699 {
4700         struct mlx5e_priv *priv = vpriv;
4701
4702         return priv->netdev;
4703 }
4704
4705 static struct mlx5_interface mlx5e_interface = {
4706         .add       = mlx5e_add,
4707         .remove    = mlx5e_remove,
4708         .attach    = mlx5e_attach,
4709         .detach    = mlx5e_detach,
4710         .event     = mlx5e_async_event,
4711         .protocol  = MLX5_INTERFACE_PROTOCOL_ETH,
4712         .get_dev   = mlx5e_get_netdev,
4713 };
4714
4715 void mlx5e_init(void)
4716 {
4717         mlx5e_ipsec_build_inverse_table();
4718         mlx5e_build_ptys2ethtool_map();
4719         mlx5_register_interface(&mlx5e_interface);
4720 }
4721
4722 void mlx5e_cleanup(void)
4723 {
4724         mlx5_unregister_interface(&mlx5e_interface);
4725 }