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 <net/geneve.h>
  38 #include <linux/bpf.h>
  39 #include <linux/if_bridge.h>
  40 #include <net/page_pool.h>
  41 #include <net/xdp_sock.h>
  42 #include "eswitch.h"
  43 #include "en.h"
  44 #include "en/txrx.h"
  45 #include "en_tc.h"
  46 #include "en_rep.h"
  47 #include "en_accel/ipsec.h"
  48 #include "en_accel/ipsec_rxtx.h"
  49 #include "en_accel/en_accel.h"
  50 #include "en_accel/tls.h"
  51 #include "accel/ipsec.h"
  52 #include "accel/tls.h"
  53 #include "lib/vxlan.h"
  54 #include "lib/clock.h"
  55 #include "en/port.h"
  56 #include "en/xdp.h"
  57 #include "lib/eq.h"
  58 #include "en/monitor_stats.h"
  59 #include "en/health.h"
  60 #include "en/params.h"
  61 #include "en/xsk/umem.h"
  62 #include "en/xsk/setup.h"
  63 #include "en/xsk/rx.h"
  64 #include "en/xsk/tx.h"
  65 #include "en/hv_vhca_stats.h"
  66 #include "lib/mlx5.h"
  67
  68
  69 bool mlx5e_check_fragmented_striding_rq_cap(struct mlx5_core_dev *mdev)
  70 {
  71         bool striding_rq_umr = MLX5_CAP_GEN(mdev, striding_rq) &&
  72                 MLX5_CAP_GEN(mdev, umr_ptr_rlky) &&
  73                 MLX5_CAP_ETH(mdev, reg_umr_sq);
  74         u16 max_wqe_sz_cap = MLX5_CAP_GEN(mdev, max_wqe_sz_sq);
  75         bool inline_umr = MLX5E_UMR_WQE_INLINE_SZ <= max_wqe_sz_cap;
  76
  77         if (!striding_rq_umr)
  78                 return false;
  79         if (!inline_umr) {
  80                 mlx5_core_warn(mdev, "Cannot support Striding RQ: UMR WQE size (%d) exceeds maximum supported (%d).\n",
  81                                (int)MLX5E_UMR_WQE_INLINE_SZ, max_wqe_sz_cap);
  82                 return false;
  83         }
  84         return true;
  85 }
  86
  87 void mlx5e_init_rq_type_params(struct mlx5_core_dev *mdev,
  88                                struct mlx5e_params *params)
  89 {
  90         params->log_rq_mtu_frames = is_kdump_kernel() ?
  91                 MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE :
  92                 MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
  93
  94         mlx5_core_info(mdev, "MLX5E: StrdRq(%d) RqSz(%ld) StrdSz(%ld) RxCqeCmprss(%d)\n",
  95                        params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ,
  96                        params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ ?
  97                        BIT(mlx5e_mpwqe_get_log_rq_size(params, NULL)) :
  98                        BIT(params->log_rq_mtu_frames),
  99                        BIT(mlx5e_mpwqe_get_log_stride_size(mdev, params, NULL)),
 100                        MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS));
 101 }
 102
 103 bool mlx5e_striding_rq_possible(struct mlx5_core_dev *mdev,
 104                                 struct mlx5e_params *params)
 105 {
 106         if (!mlx5e_check_fragmented_striding_rq_cap(mdev))
 107                 return false;
 108
 109         if (MLX5_IPSEC_DEV(mdev))
 110                 return false;
 111
 112         if (params->xdp_prog) {
 113                 /* XSK params are not considered here. If striding RQ is in use,
 114                  * and an XSK is being opened, mlx5e_rx_mpwqe_is_linear_skb will
 115                  * be called with the known XSK params.
 116                  */
 117                 if (!mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL))
 118                         return false;
 119         }
 120
 121         return true;
 122 }
 123
 124 void mlx5e_set_rq_type(struct mlx5_core_dev *mdev, struct mlx5e_params *params)
 125 {
 126         params->rq_wq_type = mlx5e_striding_rq_possible(mdev, params) &&
 127                 MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ) ?
 128                 MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ :
 129                 MLX5_WQ_TYPE_CYCLIC;
 130 }
 131
 132 void mlx5e_update_carrier(struct mlx5e_priv *priv)
 133 {
 134         struct mlx5_core_dev *mdev = priv->mdev;
 135         u8 port_state;
 136
 137         port_state = mlx5_query_vport_state(mdev,
 138                                             MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT,
 139                                             0);
 140
 141         if (port_state == VPORT_STATE_UP) {
 142                 netdev_info(priv->netdev, "Link up\n");
 143                 netif_carrier_on(priv->netdev);
 144         } else {
 145                 netdev_info(priv->netdev, "Link down\n");
 146                 netif_carrier_off(priv->netdev);
 147         }
 148 }
 149
 150 static void mlx5e_update_carrier_work(struct work_struct *work)
 151 {
 152         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 153                                                update_carrier_work);
 154
 155         mutex_lock(&priv->state_lock);
 156         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
 157                 if (priv->profile->update_carrier)
 158                         priv->profile->update_carrier(priv);
 159         mutex_unlock(&priv->state_lock);
 160 }
 161
 162 void mlx5e_update_ndo_stats(struct mlx5e_priv *priv)
 163 {
 164         int i;
 165
 166         for (i = mlx5e_nic_stats_grps_num(priv) - 1; i >= 0; i--)
 167                 if (mlx5e_nic_stats_grps[i]->update_stats_mask &
 168                     MLX5E_NDO_UPDATE_STATS)
 169                         mlx5e_nic_stats_grps[i]->update_stats(priv);
 170 }
 171
 172 static void mlx5e_update_stats_work(struct work_struct *work)
 173 {
 174         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 175                                                update_stats_work);
 176
 177         mutex_lock(&priv->state_lock);
 178         priv->profile->update_stats(priv);
 179         mutex_unlock(&priv->state_lock);
 180 }
 181
 182 void mlx5e_queue_update_stats(struct mlx5e_priv *priv)
 183 {
 184         if (!priv->profile->update_stats)
 185                 return;
 186
 187         if (unlikely(test_bit(MLX5E_STATE_DESTROYING, &priv->state)))
 188                 return;
 189
 190         queue_work(priv->wq, &priv->update_stats_work);
 191 }
 192
 193 static int async_event(struct notifier_block *nb, unsigned long event, void *data)
 194 {
 195         struct mlx5e_priv *priv = container_of(nb, struct mlx5e_priv, events_nb);
 196         struct mlx5_eqe   *eqe = data;
 197
 198         if (event != MLX5_EVENT_TYPE_PORT_CHANGE)
 199                 return NOTIFY_DONE;
 200
 201         switch (eqe->sub_type) {
 202         case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
 203         case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
 204                 queue_work(priv->wq, &priv->update_carrier_work);
 205                 break;
 206         default:
 207                 return NOTIFY_DONE;
 208         }
 209
 210         return NOTIFY_OK;
 211 }
 212
 213 static void mlx5e_enable_async_events(struct mlx5e_priv *priv)
 214 {
 215         priv->events_nb.notifier_call = async_event;
 216         mlx5_notifier_register(priv->mdev, &priv->events_nb);
 217 }
 218
 219 static void mlx5e_disable_async_events(struct mlx5e_priv *priv)
 220 {
 221         mlx5_notifier_unregister(priv->mdev, &priv->events_nb);
 222 }
 223
 224 static inline void mlx5e_build_umr_wqe(struct mlx5e_rq *rq,
 225                                        struct mlx5e_icosq *sq,
 226                                        struct mlx5e_umr_wqe *wqe)
 227 {
 228         struct mlx5_wqe_ctrl_seg      *cseg = &wqe->ctrl;
 229         struct mlx5_wqe_umr_ctrl_seg *ucseg = &wqe->uctrl;
 230         u8 ds_cnt = DIV_ROUND_UP(MLX5E_UMR_WQE_INLINE_SZ, MLX5_SEND_WQE_DS);
 231
 232         cseg->qpn_ds    = cpu_to_be32((sq->sqn << MLX5_WQE_CTRL_QPN_SHIFT) |
 233                                       ds_cnt);
 234         cseg->fm_ce_se  = MLX5_WQE_CTRL_CQ_UPDATE;
 235         cseg->imm       = rq->mkey_be;
 236
 237         ucseg->flags = MLX5_UMR_TRANSLATION_OFFSET_EN | MLX5_UMR_INLINE;
 238         ucseg->xlt_octowords =
 239                 cpu_to_be16(MLX5_MTT_OCTW(MLX5_MPWRQ_PAGES_PER_WQE));
 240         ucseg->mkey_mask     = cpu_to_be64(MLX5_MKEY_MASK_FREE);
 241 }
 242
 243 static int mlx5e_rq_alloc_mpwqe_info(struct mlx5e_rq *rq,
 244                                      struct mlx5e_channel *c)
 245 {
 246         int wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 247
 248         rq->mpwqe.info = kvzalloc_node(array_size(wq_sz,
 249                                                   sizeof(*rq->mpwqe.info)),
 250                                        GFP_KERNEL, cpu_to_node(c->cpu));
 251         if (!rq->mpwqe.info)
 252                 return -ENOMEM;
 253
 254         mlx5e_build_umr_wqe(rq, &c->icosq, &rq->mpwqe.umr_wqe);
 255
 256         return 0;
 257 }
 258
 259 static int mlx5e_create_umr_mkey(struct mlx5_core_dev *mdev,
 260                                  u64 npages, u8 page_shift,
 261                                  struct mlx5_core_mkey *umr_mkey)
 262 {
 263         int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
 264         void *mkc;
 265         u32 *in;
 266         int err;
 267
 268         in = kvzalloc(inlen, GFP_KERNEL);
 269         if (!in)
 270                 return -ENOMEM;
 271
 272         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 273
 274         MLX5_SET(mkc, mkc, free, 1);
 275         MLX5_SET(mkc, mkc, umr_en, 1);
 276         MLX5_SET(mkc, mkc, lw, 1);
 277         MLX5_SET(mkc, mkc, lr, 1);
 278         MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
 279
 280         MLX5_SET(mkc, mkc, qpn, 0xffffff);
 281         MLX5_SET(mkc, mkc, pd, mdev->mlx5e_res.pdn);
 282         MLX5_SET64(mkc, mkc, len, npages << page_shift);
 283         MLX5_SET(mkc, mkc, translations_octword_size,
 284                  MLX5_MTT_OCTW(npages));
 285         MLX5_SET(mkc, mkc, log_page_size, page_shift);
 286
 287         err = mlx5_core_create_mkey(mdev, umr_mkey, in, inlen);
 288
 289         kvfree(in);
 290         return err;
 291 }
 292
 293 static int mlx5e_create_rq_umr_mkey(struct mlx5_core_dev *mdev, struct mlx5e_rq *rq)
 294 {
 295         u64 num_mtts = MLX5E_REQUIRED_MTTS(mlx5_wq_ll_get_size(&rq->mpwqe.wq));
 296
 297         return mlx5e_create_umr_mkey(mdev, num_mtts, PAGE_SHIFT, &rq->umr_mkey);
 298 }
 299
 300 static inline u64 mlx5e_get_mpwqe_offset(struct mlx5e_rq *rq, u16 wqe_ix)
 301 {
 302         return (wqe_ix << MLX5E_LOG_ALIGNED_MPWQE_PPW) << PAGE_SHIFT;
 303 }
 304
 305 static void mlx5e_init_frags_partition(struct mlx5e_rq *rq)
 306 {
 307         struct mlx5e_wqe_frag_info next_frag = {};
 308         struct mlx5e_wqe_frag_info *prev = NULL;
 309         int i;
 310
 311         next_frag.di = &rq->wqe.di[0];
 312
 313         for (i = 0; i < mlx5_wq_cyc_get_size(&rq->wqe.wq); i++) {
 314                 struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
 315                 struct mlx5e_wqe_frag_info *frag =
 316                         &rq->wqe.frags[i << rq->wqe.info.log_num_frags];
 317                 int f;
 318
 319                 for (f = 0; f < rq->wqe.info.num_frags; f++, frag++) {
 320                         if (next_frag.offset + frag_info[f].frag_stride > PAGE_SIZE) {
 321                                 next_frag.di++;
 322                                 next_frag.offset = 0;
 323                                 if (prev)
 324                                         prev->last_in_page = true;
 325                         }
 326                         *frag = next_frag;
 327
 328                         /* prepare next */
 329                         next_frag.offset += frag_info[f].frag_stride;
 330                         prev = frag;
 331                 }
 332         }
 333
 334         if (prev)
 335                 prev->last_in_page = true;
 336 }
 337
 338 static int mlx5e_init_di_list(struct mlx5e_rq *rq,
 339                               int wq_sz, int cpu)
 340 {
 341         int len = wq_sz << rq->wqe.info.log_num_frags;
 342
 343         rq->wqe.di = kvzalloc_node(array_size(len, sizeof(*rq->wqe.di)),
 344                                    GFP_KERNEL, cpu_to_node(cpu));
 345         if (!rq->wqe.di)
 346                 return -ENOMEM;
 347
 348         mlx5e_init_frags_partition(rq);
 349
 350         return 0;
 351 }
 352
 353 static void mlx5e_free_di_list(struct mlx5e_rq *rq)
 354 {
 355         kvfree(rq->wqe.di);
 356 }
 357
 358 static void mlx5e_rq_err_cqe_work(struct work_struct *recover_work)
 359 {
 360         struct mlx5e_rq *rq = container_of(recover_work, struct mlx5e_rq, recover_work);
 361
 362         mlx5e_reporter_rq_cqe_err(rq);
 363 }
 364
 365 static int mlx5e_alloc_rq(struct mlx5e_channel *c,
 366                           struct mlx5e_params *params,
 367                           struct mlx5e_xsk_param *xsk,
 368                           struct xdp_umem *umem,
 369                           struct mlx5e_rq_param *rqp,
 370                           struct mlx5e_rq *rq)
 371 {
 372         struct page_pool_params pp_params = { 0 };
 373         struct mlx5_core_dev *mdev = c->mdev;
 374         void *rqc = rqp->rqc;
 375         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 376         u32 num_xsk_frames = 0;
 377         u32 rq_xdp_ix;
 378         u32 pool_size;
 379         int wq_sz;
 380         int err;
 381         int i;
 382
 383         rqp->wq.db_numa_node = cpu_to_node(c->cpu);
 384
 385         rq->wq_type = params->rq_wq_type;
 386         rq->pdev    = c->pdev;
 387         rq->netdev  = c->netdev;
 388         rq->tstamp  = c->tstamp;
 389         rq->clock   = &mdev->clock;
 390         rq->channel = c;
 391         rq->ix      = c->ix;
 392         rq->mdev    = mdev;
 393         rq->hw_mtu  = MLX5E_SW2HW_MTU(params, params->sw_mtu);
 394         rq->xdpsq   = &c->rq_xdpsq;
 395         rq->umem    = umem;
 396
 397         if (rq->umem)
 398                 rq->stats = &c->priv->channel_stats[c->ix].xskrq;
 399         else
 400                 rq->stats = &c->priv->channel_stats[c->ix].rq;
 401         INIT_WORK(&rq->recover_work, mlx5e_rq_err_cqe_work);
 402
 403         if (params->xdp_prog)
 404                 bpf_prog_inc(params->xdp_prog);
 405         rq->xdp_prog = params->xdp_prog;
 406
 407         rq_xdp_ix = rq->ix;
 408         if (xsk)
 409                 rq_xdp_ix += params->num_channels * MLX5E_RQ_GROUP_XSK;
 410         err = xdp_rxq_info_reg(&rq->xdp_rxq, rq->netdev, rq_xdp_ix);
 411         if (err < 0)
 412                 goto err_rq_wq_destroy;
 413
 414         rq->buff.map_dir = rq->xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
 415         rq->buff.headroom = mlx5e_get_rq_headroom(mdev, params, xsk);
 416         rq->buff.umem_headroom = xsk ? xsk->headroom : 0;
 417         pool_size = 1 << params->log_rq_mtu_frames;
 418
 419         switch (rq->wq_type) {
 420         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 421                 err = mlx5_wq_ll_create(mdev, &rqp->wq, rqc_wq, &rq->mpwqe.wq,
 422                                         &rq->wq_ctrl);
 423                 if (err)
 424                         return err;
 425
 426                 rq->mpwqe.wq.db = &rq->mpwqe.wq.db[MLX5_RCV_DBR];
 427
 428                 wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 429
 430                 if (xsk)
 431                         num_xsk_frames = wq_sz <<
 432                                 mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk);
 433
 434                 pool_size = MLX5_MPWRQ_PAGES_PER_WQE <<
 435                         mlx5e_mpwqe_get_log_rq_size(params, xsk);
 436
 437                 rq->post_wqes = mlx5e_post_rx_mpwqes;
 438                 rq->dealloc_wqe = mlx5e_dealloc_rx_mpwqe;
 439
 440                 rq->handle_rx_cqe = c->priv->profile->rx_handlers.handle_rx_cqe_mpwqe;
 441 #ifdef CONFIG_MLX5_EN_IPSEC
 442                 if (MLX5_IPSEC_DEV(mdev)) {
 443                         err = -EINVAL;
 444                         netdev_err(c->netdev, "MPWQE RQ with IPSec offload not supported\n");
 445                         goto err_rq_wq_destroy;
 446                 }
 447 #endif
 448                 if (!rq->handle_rx_cqe) {
 449                         err = -EINVAL;
 450                         netdev_err(c->netdev, "RX handler of MPWQE RQ is not set, err %d\n", err);
 451                         goto err_rq_wq_destroy;
 452                 }
 453
 454                 rq->mpwqe.skb_from_cqe_mpwrq = xsk ?
 455                         mlx5e_xsk_skb_from_cqe_mpwrq_linear :
 456                         mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL) ?
 457                                 mlx5e_skb_from_cqe_mpwrq_linear :
 458                                 mlx5e_skb_from_cqe_mpwrq_nonlinear;
 459
 460                 rq->mpwqe.log_stride_sz = mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
 461                 rq->mpwqe.num_strides =
 462                         BIT(mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk));
 463
 464                 err = mlx5e_create_rq_umr_mkey(mdev, rq);
 465                 if (err)
 466                         goto err_rq_wq_destroy;
 467                 rq->mkey_be = cpu_to_be32(rq->umr_mkey.key);
 468
 469                 err = mlx5e_rq_alloc_mpwqe_info(rq, c);
 470                 if (err)
 471                         goto err_free;
 472                 break;
 473         default: /* MLX5_WQ_TYPE_CYCLIC */
 474                 err = mlx5_wq_cyc_create(mdev, &rqp->wq, rqc_wq, &rq->wqe.wq,
 475                                          &rq->wq_ctrl);
 476                 if (err)
 477                         return err;
 478
 479                 rq->wqe.wq.db = &rq->wqe.wq.db[MLX5_RCV_DBR];
 480
 481                 wq_sz = mlx5_wq_cyc_get_size(&rq->wqe.wq);
 482
 483                 if (xsk)
 484                         num_xsk_frames = wq_sz << rq->wqe.info.log_num_frags;
 485
 486                 rq->wqe.info = rqp->frags_info;
 487                 rq->wqe.frags =
 488                         kvzalloc_node(array_size(sizeof(*rq->wqe.frags),
 489                                         (wq_sz << rq->wqe.info.log_num_frags)),
 490                                       GFP_KERNEL, cpu_to_node(c->cpu));
 491                 if (!rq->wqe.frags) {
 492                         err = -ENOMEM;
 493                         goto err_free;
 494                 }
 495
 496                 err = mlx5e_init_di_list(rq, wq_sz, c->cpu);
 497                 if (err)
 498                         goto err_free;
 499
 500                 rq->post_wqes = mlx5e_post_rx_wqes;
 501                 rq->dealloc_wqe = mlx5e_dealloc_rx_wqe;
 502
 503 #ifdef CONFIG_MLX5_EN_IPSEC
 504                 if (c->priv->ipsec)
 505                         rq->handle_rx_cqe = mlx5e_ipsec_handle_rx_cqe;
 506                 else
 507 #endif
 508                         rq->handle_rx_cqe = c->priv->profile->rx_handlers.handle_rx_cqe;
 509                 if (!rq->handle_rx_cqe) {
 510                         err = -EINVAL;
 511                         netdev_err(c->netdev, "RX handler of RQ is not set, err %d\n", err);
 512                         goto err_free;
 513                 }
 514
 515                 rq->wqe.skb_from_cqe = xsk ?
 516                         mlx5e_xsk_skb_from_cqe_linear :
 517                         mlx5e_rx_is_linear_skb(params, NULL) ?
 518                                 mlx5e_skb_from_cqe_linear :
 519                                 mlx5e_skb_from_cqe_nonlinear;
 520                 rq->mkey_be = c->mkey_be;
 521         }
 522
 523         if (xsk) {
 524                 err = mlx5e_xsk_resize_reuseq(umem, num_xsk_frames);
 525                 if (unlikely(err)) {
 526                         mlx5_core_err(mdev, "Unable to allocate the Reuse Ring for %u frames\n",
 527                                       num_xsk_frames);
 528                         goto err_free;
 529                 }
 530
 531                 rq->zca.free = mlx5e_xsk_zca_free;
 532                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
 533                                                  MEM_TYPE_ZERO_COPY,
 534                                                  &rq->zca);
 535         } else {
 536                 /* Create a page_pool and register it with rxq */
 537                 pp_params.order     = 0;
 538                 pp_params.flags     = 0; /* No-internal DMA mapping in page_pool */
 539                 pp_params.pool_size = pool_size;
 540                 pp_params.nid       = cpu_to_node(c->cpu);
 541                 pp_params.dev       = c->pdev;
 542                 pp_params.dma_dir   = rq->buff.map_dir;
 543
 544                 /* page_pool can be used even when there is no rq->xdp_prog,
 545                  * given page_pool does not handle DMA mapping there is no
 546                  * required state to clear. And page_pool gracefully handle
 547                  * elevated refcnt.
 548                  */
 549                 rq->page_pool = page_pool_create(&pp_params);
 550                 if (IS_ERR(rq->page_pool)) {
 551                         err = PTR_ERR(rq->page_pool);
 552                         rq->page_pool = NULL;
 553                         goto err_free;
 554                 }
 555                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
 556                                                  MEM_TYPE_PAGE_POOL, rq->page_pool);
 557         }
 558         if (err)
 559                 goto err_free;
 560
 561         for (i = 0; i < wq_sz; i++) {
 562                 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 563                         struct mlx5e_rx_wqe_ll *wqe =
 564                                 mlx5_wq_ll_get_wqe(&rq->mpwqe.wq, i);
 565                         u32 byte_count =
 566                                 rq->mpwqe.num_strides << rq->mpwqe.log_stride_sz;
 567                         u64 dma_offset = mlx5e_get_mpwqe_offset(rq, i);
 568
 569                         wqe->data[0].addr = cpu_to_be64(dma_offset + rq->buff.headroom);
 570                         wqe->data[0].byte_count = cpu_to_be32(byte_count);
 571                         wqe->data[0].lkey = rq->mkey_be;
 572                 } else {
 573                         struct mlx5e_rx_wqe_cyc *wqe =
 574                                 mlx5_wq_cyc_get_wqe(&rq->wqe.wq, i);
 575                         int f;
 576
 577                         for (f = 0; f < rq->wqe.info.num_frags; f++) {
 578                                 u32 frag_size = rq->wqe.info.arr[f].frag_size |
 579                                         MLX5_HW_START_PADDING;
 580
 581                                 wqe->data[f].byte_count = cpu_to_be32(frag_size);
 582                                 wqe->data[f].lkey = rq->mkey_be;
 583                         }
 584                         /* check if num_frags is not a pow of two */
 585                         if (rq->wqe.info.num_frags < (1 << rq->wqe.info.log_num_frags)) {
 586                                 wqe->data[f].byte_count = 0;
 587                                 wqe->data[f].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
 588                                 wqe->data[f].addr = 0;
 589                         }
 590                 }
 591         }
 592
 593         INIT_WORK(&rq->dim.work, mlx5e_rx_dim_work);
 594
 595         switch (params->rx_cq_moderation.cq_period_mode) {
 596         case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
 597                 rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
 598                 break;
 599         case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
 600         default:
 601                 rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
 602         }
 603
 604         rq->page_cache.head = 0;
 605         rq->page_cache.tail = 0;
 606
 607         return 0;
 608
 609 err_free:
 610         switch (rq->wq_type) {
 611         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 612                 kvfree(rq->mpwqe.info);
 613                 mlx5_core_destroy_mkey(mdev, &rq->umr_mkey);
 614                 break;
 615         default: /* MLX5_WQ_TYPE_CYCLIC */
 616                 kvfree(rq->wqe.frags);
 617                 mlx5e_free_di_list(rq);
 618         }
 619
 620 err_rq_wq_destroy:
 621         if (rq->xdp_prog)
 622                 bpf_prog_put(rq->xdp_prog);
 623         xdp_rxq_info_unreg(&rq->xdp_rxq);
 624         page_pool_destroy(rq->page_pool);
 625         mlx5_wq_destroy(&rq->wq_ctrl);
 626
 627         return err;
 628 }
 629
 630 static void mlx5e_free_rq(struct mlx5e_rq *rq)
 631 {
 632         int i;
 633
 634         if (rq->xdp_prog)
 635                 bpf_prog_put(rq->xdp_prog);
 636
 637         switch (rq->wq_type) {
 638         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 639                 kvfree(rq->mpwqe.info);
 640                 mlx5_core_destroy_mkey(rq->mdev, &rq->umr_mkey);
 641                 break;
 642         default: /* MLX5_WQ_TYPE_CYCLIC */
 643                 kvfree(rq->wqe.frags);
 644                 mlx5e_free_di_list(rq);
 645         }
 646
 647         for (i = rq->page_cache.head; i != rq->page_cache.tail;
 648              i = (i + 1) & (MLX5E_CACHE_SIZE - 1)) {
 649                 struct mlx5e_dma_info *dma_info = &rq->page_cache.page_cache[i];
 650
 651                 /* With AF_XDP, page_cache is not used, so this loop is not
 652                  * entered, and it's safe to call mlx5e_page_release_dynamic
 653                  * directly.
 654                  */
 655                 mlx5e_page_release_dynamic(rq, dma_info, false);
 656         }
 657
 658         xdp_rxq_info_unreg(&rq->xdp_rxq);
 659         page_pool_destroy(rq->page_pool);
 660         mlx5_wq_destroy(&rq->wq_ctrl);
 661 }
 662
 663 static int mlx5e_create_rq(struct mlx5e_rq *rq,
 664                            struct mlx5e_rq_param *param)
 665 {
 666         struct mlx5_core_dev *mdev = rq->mdev;
 667
 668         void *in;
 669         void *rqc;
 670         void *wq;
 671         int inlen;
 672         int err;
 673
 674         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 675                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 676         in = kvzalloc(inlen, GFP_KERNEL);
 677         if (!in)
 678                 return -ENOMEM;
 679
 680         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 681         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 682
 683         memcpy(rqc, param->rqc, sizeof(param->rqc));
 684
 685         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 686         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 687         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 688                                                 MLX5_ADAPTER_PAGE_SHIFT);
 689         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 690
 691         mlx5_fill_page_frag_array(&rq->wq_ctrl.buf,
 692                                   (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 693
 694         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 695
 696         kvfree(in);
 697
 698         return err;
 699 }
 700
 701 int mlx5e_modify_rq_state(struct mlx5e_rq *rq, int curr_state, int next_state)
 702 {
 703         struct mlx5_core_dev *mdev = rq->mdev;
 704
 705         void *in;
 706         void *rqc;
 707         int inlen;
 708         int err;
 709
 710         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 711         in = kvzalloc(inlen, GFP_KERNEL);
 712         if (!in)
 713                 return -ENOMEM;
 714
 715         if (curr_state == MLX5_RQC_STATE_RST && next_state == MLX5_RQC_STATE_RDY)
 716                 mlx5e_rqwq_reset(rq);
 717
 718         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 719
 720         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 721         MLX5_SET(rqc, rqc, state, next_state);
 722
 723         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 724
 725         kvfree(in);
 726
 727         return err;
 728 }
 729
 730 static int mlx5e_modify_rq_scatter_fcs(struct mlx5e_rq *rq, bool enable)
 731 {
 732         struct mlx5e_channel *c = rq->channel;
 733         struct mlx5e_priv *priv = c->priv;
 734         struct mlx5_core_dev *mdev = priv->mdev;
 735
 736         void *in;
 737         void *rqc;
 738         int inlen;
 739         int err;
 740
 741         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 742         in = kvzalloc(inlen, GFP_KERNEL);
 743         if (!in)
 744                 return -ENOMEM;
 745
 746         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 747
 748         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 749         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 750                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_SCATTER_FCS);
 751         MLX5_SET(rqc, rqc, scatter_fcs, enable);
 752         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 753
 754         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 755
 756         kvfree(in);
 757
 758         return err;
 759 }
 760
 761 static int mlx5e_modify_rq_vsd(struct mlx5e_rq *rq, bool vsd)
 762 {
 763         struct mlx5e_channel *c = rq->channel;
 764         struct mlx5_core_dev *mdev = c->mdev;
 765         void *in;
 766         void *rqc;
 767         int inlen;
 768         int err;
 769
 770         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 771         in = kvzalloc(inlen, GFP_KERNEL);
 772         if (!in)
 773                 return -ENOMEM;
 774
 775         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 776
 777         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 778         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 779                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
 780         MLX5_SET(rqc, rqc, vsd, vsd);
 781         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 782
 783         err = mlx5_core_modify_rq(mdev, rq->rqn, in, inlen);
 784
 785         kvfree(in);
 786
 787         return err;
 788 }
 789
 790 static void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 791 {
 792         mlx5_core_destroy_rq(rq->mdev, rq->rqn);
 793 }
 794
 795 int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq, int wait_time)
 796 {
 797         unsigned long exp_time = jiffies + msecs_to_jiffies(wait_time);
 798         struct mlx5e_channel *c = rq->channel;
 799
 800         u16 min_wqes = mlx5_min_rx_wqes(rq->wq_type, mlx5e_rqwq_get_size(rq));
 801
 802         do {
 803                 if (mlx5e_rqwq_get_cur_sz(rq) >= min_wqes)
 804                         return 0;
 805
 806                 msleep(20);
 807         } while (time_before(jiffies, exp_time));
 808
 809         netdev_warn(c->netdev, "Failed to get min RX wqes on Channel[%d] RQN[0x%x] wq cur_sz(%d) min_rx_wqes(%d)\n",
 810                     c->ix, rq->rqn, mlx5e_rqwq_get_cur_sz(rq), min_wqes);
 811
 812         mlx5e_reporter_rx_timeout(rq);
 813         return -ETIMEDOUT;
 814 }
 815
 816 void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 817 {
 818         __be16 wqe_ix_be;
 819         u16 wqe_ix;
 820
 821         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 822                 struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
 823                 u16 head = wq->head;
 824                 int i;
 825
 826                 /* Outstanding UMR WQEs (in progress) start at wq->head */
 827                 for (i = 0; i < rq->mpwqe.umr_in_progress; i++) {
 828                         rq->dealloc_wqe(rq, head);
 829                         head = mlx5_wq_ll_get_wqe_next_ix(wq, head);
 830                 }
 831
 832                 while (!mlx5_wq_ll_is_empty(wq)) {
 833                         struct mlx5e_rx_wqe_ll *wqe;
 834
 835                         wqe_ix_be = *wq->tail_next;
 836                         wqe_ix    = be16_to_cpu(wqe_ix_be);
 837                         wqe       = mlx5_wq_ll_get_wqe(wq, wqe_ix);
 838                         rq->dealloc_wqe(rq, wqe_ix);
 839                         mlx5_wq_ll_pop(wq, wqe_ix_be,
 840                                        &wqe->next.next_wqe_index);
 841                 }
 842         } else {
 843                 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
 844
 845                 while (!mlx5_wq_cyc_is_empty(wq)) {
 846                         wqe_ix = mlx5_wq_cyc_get_tail(wq);
 847                         rq->dealloc_wqe(rq, wqe_ix);
 848                         mlx5_wq_cyc_pop(wq);
 849                 }
 850         }
 851
 852 }
 853
 854 int mlx5e_open_rq(struct mlx5e_channel *c, struct mlx5e_params *params,
 855                   struct mlx5e_rq_param *param, struct mlx5e_xsk_param *xsk,
 856                   struct xdp_umem *umem, struct mlx5e_rq *rq)
 857 {
 858         int err;
 859
 860         err = mlx5e_alloc_rq(c, params, xsk, umem, param, rq);
 861         if (err)
 862                 return err;
 863
 864         err = mlx5e_create_rq(rq, param);
 865         if (err)
 866                 goto err_free_rq;
 867
 868         err = mlx5e_modify_rq_state(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 869         if (err)
 870                 goto err_destroy_rq;
 871
 872         if (MLX5_CAP_ETH(c->mdev, cqe_checksum_full))
 873                 __set_bit(MLX5E_RQ_STATE_CSUM_FULL, &c->rq.state);
 874
 875         if (params->rx_dim_enabled)
 876                 __set_bit(MLX5E_RQ_STATE_AM, &c->rq.state);
 877
 878         /* We disable csum_complete when XDP is enabled since
 879          * XDP programs might manipulate packets which will render
 880          * skb->checksum incorrect.
 881          */
 882         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_NO_CSUM_COMPLETE) || c->xdp)
 883                 __set_bit(MLX5E_RQ_STATE_NO_CSUM_COMPLETE, &c->rq.state);
 884
 885         return 0;
 886
 887 err_destroy_rq:
 888         mlx5e_destroy_rq(rq);
 889 err_free_rq:
 890         mlx5e_free_rq(rq);
 891
 892         return err;
 893 }
 894
 895 void mlx5e_activate_rq(struct mlx5e_rq *rq)
 896 {
 897         set_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 898         mlx5e_trigger_irq(&rq->channel->icosq);
 899 }
 900
 901 void mlx5e_deactivate_rq(struct mlx5e_rq *rq)
 902 {
 903         clear_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 904         napi_synchronize(&rq->channel->napi); /* prevent mlx5e_post_rx_wqes */
 905 }
 906
 907 void mlx5e_close_rq(struct mlx5e_rq *rq)
 908 {
 909         cancel_work_sync(&rq->dim.work);
 910         cancel_work_sync(&rq->channel->icosq.recover_work);
 911         cancel_work_sync(&rq->recover_work);
 912         mlx5e_destroy_rq(rq);
 913         mlx5e_free_rx_descs(rq);
 914         mlx5e_free_rq(rq);
 915 }
 916
 917 static void mlx5e_free_xdpsq_db(struct mlx5e_xdpsq *sq)
 918 {
 919         kvfree(sq->db.xdpi_fifo.xi);
 920         kvfree(sq->db.wqe_info);
 921 }
 922
 923 static int mlx5e_alloc_xdpsq_fifo(struct mlx5e_xdpsq *sq, int numa)
 924 {
 925         struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo;
 926         int wq_sz        = mlx5_wq_cyc_get_size(&sq->wq);
 927         int dsegs_per_wq = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
 928
 929         xdpi_fifo->xi = kvzalloc_node(sizeof(*xdpi_fifo->xi) * dsegs_per_wq,
 930                                       GFP_KERNEL, numa);
 931         if (!xdpi_fifo->xi)
 932                 return -ENOMEM;
 933
 934         xdpi_fifo->pc   = &sq->xdpi_fifo_pc;
 935         xdpi_fifo->cc   = &sq->xdpi_fifo_cc;
 936         xdpi_fifo->mask = dsegs_per_wq - 1;
 937
 938         return 0;
 939 }
 940
 941 static int mlx5e_alloc_xdpsq_db(struct mlx5e_xdpsq *sq, int numa)
 942 {
 943         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 944         int err;
 945
 946         sq->db.wqe_info = kvzalloc_node(sizeof(*sq->db.wqe_info) * wq_sz,
 947                                         GFP_KERNEL, numa);
 948         if (!sq->db.wqe_info)
 949                 return -ENOMEM;
 950
 951         err = mlx5e_alloc_xdpsq_fifo(sq, numa);
 952         if (err) {
 953                 mlx5e_free_xdpsq_db(sq);
 954                 return err;
 955         }
 956
 957         return 0;
 958 }
 959
 960 static int mlx5e_alloc_xdpsq(struct mlx5e_channel *c,
 961                              struct mlx5e_params *params,
 962                              struct xdp_umem *umem,
 963                              struct mlx5e_sq_param *param,
 964                              struct mlx5e_xdpsq *sq,
 965                              bool is_redirect)
 966 {
 967         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
 968         struct mlx5_core_dev *mdev = c->mdev;
 969         struct mlx5_wq_cyc *wq = &sq->wq;
 970         int err;
 971
 972         sq->pdev      = c->pdev;
 973         sq->mkey_be   = c->mkey_be;
 974         sq->channel   = c;
 975         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
 976         sq->min_inline_mode = params->tx_min_inline_mode;
 977         sq->hw_mtu    = MLX5E_SW2HW_MTU(params, params->sw_mtu);
 978         sq->umem      = umem;
 979
 980         sq->stats = sq->umem ?
 981                 &c->priv->channel_stats[c->ix].xsksq :
 982                 is_redirect ?
 983                         &c->priv->channel_stats[c->ix].xdpsq :
 984                         &c->priv->channel_stats[c->ix].rq_xdpsq;
 985
 986         param->wq.db_numa_node = cpu_to_node(c->cpu);
 987         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
 988         if (err)
 989                 return err;
 990         wq->db = &wq->db[MLX5_SND_DBR];
 991
 992         err = mlx5e_alloc_xdpsq_db(sq, cpu_to_node(c->cpu));
 993         if (err)
 994                 goto err_sq_wq_destroy;
 995
 996         return 0;
 997
 998 err_sq_wq_destroy:
 999         mlx5_wq_destroy(&sq->wq_ctrl);
1000
1001         return err;
1002 }
1003
1004 static void mlx5e_free_xdpsq(struct mlx5e_xdpsq *sq)
1005 {
1006         mlx5e_free_xdpsq_db(sq);
1007         mlx5_wq_destroy(&sq->wq_ctrl);
1008 }
1009
1010 static void mlx5e_free_icosq_db(struct mlx5e_icosq *sq)
1011 {
1012         kvfree(sq->db.ico_wqe);
1013 }
1014
1015 static int mlx5e_alloc_icosq_db(struct mlx5e_icosq *sq, int numa)
1016 {
1017         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1018
1019         sq->db.ico_wqe = kvzalloc_node(array_size(wq_sz,
1020                                                   sizeof(*sq->db.ico_wqe)),
1021                                        GFP_KERNEL, numa);
1022         if (!sq->db.ico_wqe)
1023                 return -ENOMEM;
1024
1025         return 0;
1026 }
1027
1028 static void mlx5e_icosq_err_cqe_work(struct work_struct *recover_work)
1029 {
1030         struct mlx5e_icosq *sq = container_of(recover_work, struct mlx5e_icosq,
1031                                               recover_work);
1032
1033         mlx5e_reporter_icosq_cqe_err(sq);
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         struct mlx5_wq_cyc *wq = &sq->wq;
1043         int err;
1044
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, wq, &sq->wq_ctrl);
1050         if (err)
1051                 return err;
1052         wq->db = &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         INIT_WORK(&sq->recover_work, mlx5e_icosq_err_cqe_work);
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         kvfree(sq->db.wqe_info);
1077         kvfree(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 = kvzalloc_node(array_size(df_sz,
1086                                                    sizeof(*sq->db.dma_fifo)),
1087                                         GFP_KERNEL, numa);
1088         sq->db.wqe_info = kvzalloc_node(array_size(wq_sz,
1089                                                    sizeof(*sq->db.wqe_info)),
1090                                         GFP_KERNEL, numa);
1091         if (!sq->db.dma_fifo || !sq->db.wqe_info) {
1092                 mlx5e_free_txqsq_db(sq);
1093                 return -ENOMEM;
1094         }
1095
1096         sq->dma_fifo_mask = df_sz - 1;
1097
1098         return 0;
1099 }
1100
1101 static void mlx5e_tx_err_cqe_work(struct work_struct *recover_work);
1102 static int mlx5e_alloc_txqsq(struct mlx5e_channel *c,
1103                              int txq_ix,
1104                              struct mlx5e_params *params,
1105                              struct mlx5e_sq_param *param,
1106                              struct mlx5e_txqsq *sq,
1107                              int tc)
1108 {
1109         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1110         struct mlx5_core_dev *mdev = c->mdev;
1111         struct mlx5_wq_cyc *wq = &sq->wq;
1112         int err;
1113
1114         sq->pdev      = c->pdev;
1115         sq->tstamp    = c->tstamp;
1116         sq->clock     = &mdev->clock;
1117         sq->mkey_be   = c->mkey_be;
1118         sq->channel   = c;
1119         sq->ch_ix     = c->ix;
1120         sq->txq_ix    = txq_ix;
1121         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1122         sq->min_inline_mode = params->tx_min_inline_mode;
1123         sq->hw_mtu    = MLX5E_SW2HW_MTU(params, params->sw_mtu);
1124         sq->stats     = &c->priv->channel_stats[c->ix].sq[tc];
1125         sq->stop_room = MLX5E_SQ_STOP_ROOM;
1126         INIT_WORK(&sq->recover_work, mlx5e_tx_err_cqe_work);
1127         if (!MLX5_CAP_ETH(mdev, wqe_vlan_insert))
1128                 set_bit(MLX5E_SQ_STATE_VLAN_NEED_L2_INLINE, &sq->state);
1129         if (MLX5_IPSEC_DEV(c->priv->mdev))
1130                 set_bit(MLX5E_SQ_STATE_IPSEC, &sq->state);
1131 #ifdef CONFIG_MLX5_EN_TLS
1132         if (mlx5_accel_is_tls_device(c->priv->mdev)) {
1133                 set_bit(MLX5E_SQ_STATE_TLS, &sq->state);
1134                 sq->stop_room += MLX5E_SQ_TLS_ROOM +
1135                         mlx5e_ktls_dumps_num_wqebbs(sq, MAX_SKB_FRAGS,
1136                                                     TLS_MAX_PAYLOAD_SIZE);
1137         }
1138 #endif
1139
1140         param->wq.db_numa_node = cpu_to_node(c->cpu);
1141         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
1142         if (err)
1143                 return err;
1144         wq->db    = &wq->db[MLX5_SND_DBR];
1145
1146         err = mlx5e_alloc_txqsq_db(sq, cpu_to_node(c->cpu));
1147         if (err)
1148                 goto err_sq_wq_destroy;
1149
1150         INIT_WORK(&sq->dim.work, mlx5e_tx_dim_work);
1151         sq->dim.mode = params->tx_cq_moderation.cq_period_mode;
1152
1153         return 0;
1154
1155 err_sq_wq_destroy:
1156         mlx5_wq_destroy(&sq->wq_ctrl);
1157
1158         return err;
1159 }
1160
1161 static void mlx5e_free_txqsq(struct mlx5e_txqsq *sq)
1162 {
1163         mlx5e_free_txqsq_db(sq);
1164         mlx5_wq_destroy(&sq->wq_ctrl);
1165 }
1166
1167 struct mlx5e_create_sq_param {
1168         struct mlx5_wq_ctrl        *wq_ctrl;
1169         u32                         cqn;
1170         u32                         tisn;
1171         u8                          tis_lst_sz;
1172         u8                          min_inline_mode;
1173 };
1174
1175 static int mlx5e_create_sq(struct mlx5_core_dev *mdev,
1176                            struct mlx5e_sq_param *param,
1177                            struct mlx5e_create_sq_param *csp,
1178                            u32 *sqn)
1179 {
1180         void *in;
1181         void *sqc;
1182         void *wq;
1183         int inlen;
1184         int err;
1185
1186         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1187                 sizeof(u64) * csp->wq_ctrl->buf.npages;
1188         in = kvzalloc(inlen, GFP_KERNEL);
1189         if (!in)
1190                 return -ENOMEM;
1191
1192         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1193         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1194
1195         memcpy(sqc, param->sqc, sizeof(param->sqc));
1196         MLX5_SET(sqc,  sqc, tis_lst_sz, csp->tis_lst_sz);
1197         MLX5_SET(sqc,  sqc, tis_num_0, csp->tisn);
1198         MLX5_SET(sqc,  sqc, cqn, csp->cqn);
1199
1200         if (MLX5_CAP_ETH(mdev, wqe_inline_mode) == MLX5_CAP_INLINE_MODE_VPORT_CONTEXT)
1201                 MLX5_SET(sqc,  sqc, min_wqe_inline_mode, csp->min_inline_mode);
1202
1203         MLX5_SET(sqc,  sqc, state, MLX5_SQC_STATE_RST);
1204         MLX5_SET(sqc,  sqc, flush_in_error_en, 1);
1205
1206         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
1207         MLX5_SET(wq,   wq, uar_page,      mdev->mlx5e_res.bfreg.index);
1208         MLX5_SET(wq,   wq, log_wq_pg_sz,  csp->wq_ctrl->buf.page_shift -
1209                                           MLX5_ADAPTER_PAGE_SHIFT);
1210         MLX5_SET64(wq, wq, dbr_addr,      csp->wq_ctrl->db.dma);
1211
1212         mlx5_fill_page_frag_array(&csp->wq_ctrl->buf,
1213                                   (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
1214
1215         err = mlx5_core_create_sq(mdev, in, inlen, sqn);
1216
1217         kvfree(in);
1218
1219         return err;
1220 }
1221
1222 int mlx5e_modify_sq(struct mlx5_core_dev *mdev, u32 sqn,
1223                     struct mlx5e_modify_sq_param *p)
1224 {
1225         void *in;
1226         void *sqc;
1227         int inlen;
1228         int err;
1229
1230         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1231         in = kvzalloc(inlen, GFP_KERNEL);
1232         if (!in)
1233                 return -ENOMEM;
1234
1235         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1236
1237         MLX5_SET(modify_sq_in, in, sq_state, p->curr_state);
1238         MLX5_SET(sqc, sqc, state, p->next_state);
1239         if (p->rl_update && p->next_state == MLX5_SQC_STATE_RDY) {
1240                 MLX5_SET64(modify_sq_in, in, modify_bitmask, 1);
1241                 MLX5_SET(sqc,  sqc, packet_pacing_rate_limit_index, p->rl_index);
1242         }
1243
1244         err = mlx5_core_modify_sq(mdev, sqn, in, inlen);
1245
1246         kvfree(in);
1247
1248         return err;
1249 }
1250
1251 static void mlx5e_destroy_sq(struct mlx5_core_dev *mdev, u32 sqn)
1252 {
1253         mlx5_core_destroy_sq(mdev, sqn);
1254 }
1255
1256 static int mlx5e_create_sq_rdy(struct mlx5_core_dev *mdev,
1257                                struct mlx5e_sq_param *param,
1258                                struct mlx5e_create_sq_param *csp,
1259                                u32 *sqn)
1260 {
1261         struct mlx5e_modify_sq_param msp = {0};
1262         int err;
1263
1264         err = mlx5e_create_sq(mdev, param, csp, sqn);
1265         if (err)
1266                 return err;
1267
1268         msp.curr_state = MLX5_SQC_STATE_RST;
1269         msp.next_state = MLX5_SQC_STATE_RDY;
1270         err = mlx5e_modify_sq(mdev, *sqn, &msp);
1271         if (err)
1272                 mlx5e_destroy_sq(mdev, *sqn);
1273
1274         return err;
1275 }
1276
1277 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1278                                 struct mlx5e_txqsq *sq, u32 rate);
1279
1280 static int mlx5e_open_txqsq(struct mlx5e_channel *c,
1281                             u32 tisn,
1282                             int txq_ix,
1283                             struct mlx5e_params *params,
1284                             struct mlx5e_sq_param *param,
1285                             struct mlx5e_txqsq *sq,
1286                             int tc)
1287 {
1288         struct mlx5e_create_sq_param csp = {};
1289         u32 tx_rate;
1290         int err;
1291
1292         err = mlx5e_alloc_txqsq(c, txq_ix, params, param, sq, tc);
1293         if (err)
1294                 return err;
1295
1296         csp.tisn            = tisn;
1297         csp.tis_lst_sz      = 1;
1298         csp.cqn             = sq->cq.mcq.cqn;
1299         csp.wq_ctrl         = &sq->wq_ctrl;
1300         csp.min_inline_mode = sq->min_inline_mode;
1301         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1302         if (err)
1303                 goto err_free_txqsq;
1304
1305         tx_rate = c->priv->tx_rates[sq->txq_ix];
1306         if (tx_rate)
1307                 mlx5e_set_sq_maxrate(c->netdev, sq, tx_rate);
1308
1309         if (params->tx_dim_enabled)
1310                 sq->state |= BIT(MLX5E_SQ_STATE_AM);
1311
1312         return 0;
1313
1314 err_free_txqsq:
1315         mlx5e_free_txqsq(sq);
1316
1317         return err;
1318 }
1319
1320 void mlx5e_activate_txqsq(struct mlx5e_txqsq *sq)
1321 {
1322         sq->txq = netdev_get_tx_queue(sq->channel->netdev, sq->txq_ix);
1323         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1324         netdev_tx_reset_queue(sq->txq);
1325         netif_tx_start_queue(sq->txq);
1326 }
1327
1328 void mlx5e_tx_disable_queue(struct netdev_queue *txq)
1329 {
1330         __netif_tx_lock_bh(txq);
1331         netif_tx_stop_queue(txq);
1332         __netif_tx_unlock_bh(txq);
1333 }
1334
1335 static void mlx5e_deactivate_txqsq(struct mlx5e_txqsq *sq)
1336 {
1337         struct mlx5e_channel *c = sq->channel;
1338         struct mlx5_wq_cyc *wq = &sq->wq;
1339
1340         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1341         /* prevent netif_tx_wake_queue */
1342         napi_synchronize(&c->napi);
1343
1344         mlx5e_tx_disable_queue(sq->txq);
1345
1346         /* last doorbell out, godspeed .. */
1347         if (mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1)) {
1348                 u16 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
1349                 struct mlx5e_tx_wqe_info *wi;
1350                 struct mlx5e_tx_wqe *nop;
1351
1352                 wi = &sq->db.wqe_info[pi];
1353
1354                 memset(wi, 0, sizeof(*wi));
1355                 wi->num_wqebbs = 1;
1356                 nop = mlx5e_post_nop(wq, sq->sqn, &sq->pc);
1357                 mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &nop->ctrl);
1358         }
1359 }
1360
1361 static void mlx5e_close_txqsq(struct mlx5e_txqsq *sq)
1362 {
1363         struct mlx5e_channel *c = sq->channel;
1364         struct mlx5_core_dev *mdev = c->mdev;
1365         struct mlx5_rate_limit rl = {0};
1366
1367         cancel_work_sync(&sq->dim.work);
1368         cancel_work_sync(&sq->recover_work);
1369         mlx5e_destroy_sq(mdev, sq->sqn);
1370         if (sq->rate_limit) {
1371                 rl.rate = sq->rate_limit;
1372                 mlx5_rl_remove_rate(mdev, &rl);
1373         }
1374         mlx5e_free_txqsq_descs(sq);
1375         mlx5e_free_txqsq(sq);
1376 }
1377
1378 static void mlx5e_tx_err_cqe_work(struct work_struct *recover_work)
1379 {
1380         struct mlx5e_txqsq *sq = container_of(recover_work, struct mlx5e_txqsq,
1381                                               recover_work);
1382
1383         mlx5e_reporter_tx_err_cqe(sq);
1384 }
1385
1386 int mlx5e_open_icosq(struct mlx5e_channel *c, struct mlx5e_params *params,
1387                      struct mlx5e_sq_param *param, struct mlx5e_icosq *sq)
1388 {
1389         struct mlx5e_create_sq_param csp = {};
1390         int err;
1391
1392         err = mlx5e_alloc_icosq(c, param, sq);
1393         if (err)
1394                 return err;
1395
1396         csp.cqn             = sq->cq.mcq.cqn;
1397         csp.wq_ctrl         = &sq->wq_ctrl;
1398         csp.min_inline_mode = params->tx_min_inline_mode;
1399         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1400         if (err)
1401                 goto err_free_icosq;
1402
1403         return 0;
1404
1405 err_free_icosq:
1406         mlx5e_free_icosq(sq);
1407
1408         return err;
1409 }
1410
1411 void mlx5e_activate_icosq(struct mlx5e_icosq *icosq)
1412 {
1413         set_bit(MLX5E_SQ_STATE_ENABLED, &icosq->state);
1414 }
1415
1416 void mlx5e_deactivate_icosq(struct mlx5e_icosq *icosq)
1417 {
1418         struct mlx5e_channel *c = icosq->channel;
1419
1420         clear_bit(MLX5E_SQ_STATE_ENABLED, &icosq->state);
1421         napi_synchronize(&c->napi);
1422 }
1423
1424 void mlx5e_close_icosq(struct mlx5e_icosq *sq)
1425 {
1426         struct mlx5e_channel *c = sq->channel;
1427
1428         mlx5e_destroy_sq(c->mdev, sq->sqn);
1429         mlx5e_free_icosq(sq);
1430 }
1431
1432 int mlx5e_open_xdpsq(struct mlx5e_channel *c, struct mlx5e_params *params,
1433                      struct mlx5e_sq_param *param, struct xdp_umem *umem,
1434                      struct mlx5e_xdpsq *sq, bool is_redirect)
1435 {
1436         struct mlx5e_create_sq_param csp = {};
1437         int err;
1438
1439         err = mlx5e_alloc_xdpsq(c, params, umem, param, sq, is_redirect);
1440         if (err)
1441                 return err;
1442
1443         csp.tis_lst_sz      = 1;
1444         csp.tisn            = c->priv->tisn[c->lag_port][0]; /* tc = 0 */
1445         csp.cqn             = sq->cq.mcq.cqn;
1446         csp.wq_ctrl         = &sq->wq_ctrl;
1447         csp.min_inline_mode = sq->min_inline_mode;
1448         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1449         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, &sq->sqn);
1450         if (err)
1451                 goto err_free_xdpsq;
1452
1453         mlx5e_set_xmit_fp(sq, param->is_mpw);
1454
1455         if (!param->is_mpw) {
1456                 unsigned int ds_cnt = MLX5E_XDP_TX_DS_COUNT;
1457                 unsigned int inline_hdr_sz = 0;
1458                 int i;
1459
1460                 if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
1461                         inline_hdr_sz = MLX5E_XDP_MIN_INLINE;
1462                         ds_cnt++;
1463                 }
1464
1465                 /* Pre initialize fixed WQE fields */
1466                 for (i = 0; i < mlx5_wq_cyc_get_size(&sq->wq); i++) {
1467                         struct mlx5e_xdp_wqe_info *wi  = &sq->db.wqe_info[i];
1468                         struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(&sq->wq, i);
1469                         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
1470                         struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
1471                         struct mlx5_wqe_data_seg *dseg;
1472
1473                         cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
1474                         eseg->inline_hdr.sz = cpu_to_be16(inline_hdr_sz);
1475
1476                         dseg = (struct mlx5_wqe_data_seg *)cseg + (ds_cnt - 1);
1477                         dseg->lkey = sq->mkey_be;
1478
1479                         wi->num_wqebbs = 1;
1480                         wi->num_pkts   = 1;
1481                 }
1482         }
1483
1484         return 0;
1485
1486 err_free_xdpsq:
1487         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1488         mlx5e_free_xdpsq(sq);
1489
1490         return err;
1491 }
1492
1493 void mlx5e_close_xdpsq(struct mlx5e_xdpsq *sq)
1494 {
1495         struct mlx5e_channel *c = sq->channel;
1496
1497         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1498         napi_synchronize(&c->napi);
1499
1500         mlx5e_destroy_sq(c->mdev, sq->sqn);
1501         mlx5e_free_xdpsq_descs(sq);
1502         mlx5e_free_xdpsq(sq);
1503 }
1504
1505 static int mlx5e_alloc_cq_common(struct mlx5_core_dev *mdev,
1506                                  struct mlx5e_cq_param *param,
1507                                  struct mlx5e_cq *cq)
1508 {
1509         struct mlx5_core_cq *mcq = &cq->mcq;
1510         int eqn_not_used;
1511         unsigned int irqn;
1512         int err;
1513         u32 i;
1514
1515         err = mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1516         if (err)
1517                 return err;
1518
1519         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1520                                &cq->wq_ctrl);
1521         if (err)
1522                 return err;
1523
1524         mcq->cqe_sz     = 64;
1525         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1526         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1527         *mcq->set_ci_db = 0;
1528         *mcq->arm_db    = 0;
1529         mcq->vector     = param->eq_ix;
1530         mcq->comp       = mlx5e_completion_event;
1531         mcq->event      = mlx5e_cq_error_event;
1532         mcq->irqn       = irqn;
1533
1534         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1535                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1536
1537                 cqe->op_own = 0xf1;
1538         }
1539
1540         cq->mdev = mdev;
1541
1542         return 0;
1543 }
1544
1545 static int mlx5e_alloc_cq(struct mlx5e_channel *c,
1546                           struct mlx5e_cq_param *param,
1547                           struct mlx5e_cq *cq)
1548 {
1549         struct mlx5_core_dev *mdev = c->priv->mdev;
1550         int err;
1551
1552         param->wq.buf_numa_node = cpu_to_node(c->cpu);
1553         param->wq.db_numa_node  = cpu_to_node(c->cpu);
1554         param->eq_ix   = c->ix;
1555
1556         err = mlx5e_alloc_cq_common(mdev, param, cq);
1557
1558         cq->napi    = &c->napi;
1559         cq->channel = c;
1560
1561         return err;
1562 }
1563
1564 static void mlx5e_free_cq(struct mlx5e_cq *cq)
1565 {
1566         mlx5_wq_destroy(&cq->wq_ctrl);
1567 }
1568
1569 static int mlx5e_create_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
1570 {
1571         u32 out[MLX5_ST_SZ_DW(create_cq_out)];
1572         struct mlx5_core_dev *mdev = cq->mdev;
1573         struct mlx5_core_cq *mcq = &cq->mcq;
1574
1575         void *in;
1576         void *cqc;
1577         int inlen;
1578         unsigned int irqn_not_used;
1579         int eqn;
1580         int err;
1581
1582         err = mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
1583         if (err)
1584                 return err;
1585
1586         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1587                 sizeof(u64) * cq->wq_ctrl.buf.npages;
1588         in = kvzalloc(inlen, GFP_KERNEL);
1589         if (!in)
1590                 return -ENOMEM;
1591
1592         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1593
1594         memcpy(cqc, param->cqc, sizeof(param->cqc));
1595
1596         mlx5_fill_page_frag_array(&cq->wq_ctrl.buf,
1597                                   (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
1598
1599         MLX5_SET(cqc,   cqc, cq_period_mode, param->cq_period_mode);
1600         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
1601         MLX5_SET(cqc,   cqc, uar_page,      mdev->priv.uar->index);
1602         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
1603                                             MLX5_ADAPTER_PAGE_SHIFT);
1604         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
1605
1606         err = mlx5_core_create_cq(mdev, mcq, in, inlen, out, sizeof(out));
1607
1608         kvfree(in);
1609
1610         if (err)
1611                 return err;
1612
1613         mlx5e_cq_arm(cq);
1614
1615         return 0;
1616 }
1617
1618 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
1619 {
1620         mlx5_core_destroy_cq(cq->mdev, &cq->mcq);
1621 }
1622
1623 int mlx5e_open_cq(struct mlx5e_channel *c, struct dim_cq_moder moder,
1624                   struct mlx5e_cq_param *param, struct mlx5e_cq *cq)
1625 {
1626         struct mlx5_core_dev *mdev = c->mdev;
1627         int err;
1628
1629         err = mlx5e_alloc_cq(c, param, cq);
1630         if (err)
1631                 return err;
1632
1633         err = mlx5e_create_cq(cq, param);
1634         if (err)
1635                 goto err_free_cq;
1636
1637         if (MLX5_CAP_GEN(mdev, cq_moderation))
1638                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq, moder.usec, moder.pkts);
1639         return 0;
1640
1641 err_free_cq:
1642         mlx5e_free_cq(cq);
1643
1644         return err;
1645 }
1646
1647 void mlx5e_close_cq(struct mlx5e_cq *cq)
1648 {
1649         mlx5e_destroy_cq(cq);
1650         mlx5e_free_cq(cq);
1651 }
1652
1653 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1654                              struct mlx5e_params *params,
1655                              struct mlx5e_channel_param *cparam)
1656 {
1657         int err;
1658         int tc;
1659
1660         for (tc = 0; tc < c->num_tc; tc++) {
1661                 err = mlx5e_open_cq(c, params->tx_cq_moderation,
1662                                     &cparam->tx_cq, &c->sq[tc].cq);
1663                 if (err)
1664                         goto err_close_tx_cqs;
1665         }
1666
1667         return 0;
1668
1669 err_close_tx_cqs:
1670         for (tc--; tc >= 0; tc--)
1671                 mlx5e_close_cq(&c->sq[tc].cq);
1672
1673         return err;
1674 }
1675
1676 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1677 {
1678         int tc;
1679
1680         for (tc = 0; tc < c->num_tc; tc++)
1681                 mlx5e_close_cq(&c->sq[tc].cq);
1682 }
1683
1684 static int mlx5e_open_sqs(struct mlx5e_channel *c,
1685                           struct mlx5e_params *params,
1686                           struct mlx5e_channel_param *cparam)
1687 {
1688         int err, tc;
1689
1690         for (tc = 0; tc < params->num_tc; tc++) {
1691                 int txq_ix = c->ix + tc * params->num_channels;
1692
1693                 err = mlx5e_open_txqsq(c, c->priv->tisn[c->lag_port][tc], txq_ix,
1694                                        params, &cparam->sq, &c->sq[tc], tc);
1695                 if (err)
1696                         goto err_close_sqs;
1697         }
1698
1699         return 0;
1700
1701 err_close_sqs:
1702         for (tc--; tc >= 0; tc--)
1703                 mlx5e_close_txqsq(&c->sq[tc]);
1704
1705         return err;
1706 }
1707
1708 static void mlx5e_close_sqs(struct mlx5e_channel *c)
1709 {
1710         int tc;
1711
1712         for (tc = 0; tc < c->num_tc; tc++)
1713                 mlx5e_close_txqsq(&c->sq[tc]);
1714 }
1715
1716 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1717                                 struct mlx5e_txqsq *sq, u32 rate)
1718 {
1719         struct mlx5e_priv *priv = netdev_priv(dev);
1720         struct mlx5_core_dev *mdev = priv->mdev;
1721         struct mlx5e_modify_sq_param msp = {0};
1722         struct mlx5_rate_limit rl = {0};
1723         u16 rl_index = 0;
1724         int err;
1725
1726         if (rate == sq->rate_limit)
1727                 /* nothing to do */
1728                 return 0;
1729
1730         if (sq->rate_limit) {
1731                 rl.rate = sq->rate_limit;
1732                 /* remove current rl index to free space to next ones */
1733                 mlx5_rl_remove_rate(mdev, &rl);
1734         }
1735
1736         sq->rate_limit = 0;
1737
1738         if (rate) {
1739                 rl.rate = rate;
1740                 err = mlx5_rl_add_rate(mdev, &rl_index, &rl);
1741                 if (err) {
1742                         netdev_err(dev, "Failed configuring rate %u: %d\n",
1743                                    rate, err);
1744                         return err;
1745                 }
1746         }
1747
1748         msp.curr_state = MLX5_SQC_STATE_RDY;
1749         msp.next_state = MLX5_SQC_STATE_RDY;
1750         msp.rl_index   = rl_index;
1751         msp.rl_update  = true;
1752         err = mlx5e_modify_sq(mdev, sq->sqn, &msp);
1753         if (err) {
1754                 netdev_err(dev, "Failed configuring rate %u: %d\n",
1755                            rate, err);
1756                 /* remove the rate from the table */
1757                 if (rate)
1758                         mlx5_rl_remove_rate(mdev, &rl);
1759                 return err;
1760         }
1761
1762         sq->rate_limit = rate;
1763         return 0;
1764 }
1765
1766 static int mlx5e_set_tx_maxrate(struct net_device *dev, int index, u32 rate)
1767 {
1768         struct mlx5e_priv *priv = netdev_priv(dev);
1769         struct mlx5_core_dev *mdev = priv->mdev;
1770         struct mlx5e_txqsq *sq = priv->txq2sq[index];
1771         int err = 0;
1772
1773         if (!mlx5_rl_is_supported(mdev)) {
1774                 netdev_err(dev, "Rate limiting is not supported on this device\n");
1775                 return -EINVAL;
1776         }
1777
1778         /* rate is given in Mb/sec, HW config is in Kb/sec */
1779         rate = rate << 10;
1780
1781         /* Check whether rate in valid range, 0 is always valid */
1782         if (rate && !mlx5_rl_is_in_range(mdev, rate)) {
1783                 netdev_err(dev, "TX rate %u, is not in range\n", rate);
1784                 return -ERANGE;
1785         }
1786
1787         mutex_lock(&priv->state_lock);
1788         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
1789                 err = mlx5e_set_sq_maxrate(dev, sq, rate);
1790         if (!err)
1791                 priv->tx_rates[index] = rate;
1792         mutex_unlock(&priv->state_lock);
1793
1794         return err;
1795 }
1796
1797 static int mlx5e_alloc_xps_cpumask(struct mlx5e_channel *c,
1798                                    struct mlx5e_params *params)
1799 {
1800         int num_comp_vectors = mlx5_comp_vectors_count(c->mdev);
1801         int irq;
1802
1803         if (!zalloc_cpumask_var(&c->xps_cpumask, GFP_KERNEL))
1804                 return -ENOMEM;
1805
1806         for (irq = c->ix; irq < num_comp_vectors; irq += params->num_channels) {
1807                 int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(c->mdev, irq));
1808
1809                 cpumask_set_cpu(cpu, c->xps_cpumask);
1810         }
1811
1812         return 0;
1813 }
1814
1815 static void mlx5e_free_xps_cpumask(struct mlx5e_channel *c)
1816 {
1817         free_cpumask_var(c->xps_cpumask);
1818 }
1819
1820 static int mlx5e_open_queues(struct mlx5e_channel *c,
1821                              struct mlx5e_params *params,
1822                              struct mlx5e_channel_param *cparam)
1823 {
1824         struct dim_cq_moder icocq_moder = {0, 0};
1825         int err;
1826
1827         err = mlx5e_open_cq(c, icocq_moder, &cparam->icosq_cq, &c->icosq.cq);
1828         if (err)
1829                 return err;
1830
1831         err = mlx5e_open_tx_cqs(c, params, cparam);
1832         if (err)
1833                 goto err_close_icosq_cq;
1834
1835         err = mlx5e_open_cq(c, params->tx_cq_moderation, &cparam->tx_cq, &c->xdpsq.cq);
1836         if (err)
1837                 goto err_close_tx_cqs;
1838
1839         err = mlx5e_open_cq(c, params->rx_cq_moderation, &cparam->rx_cq, &c->rq.cq);
1840         if (err)
1841                 goto err_close_xdp_tx_cqs;
1842
1843         /* XDP SQ CQ params are same as normal TXQ sq CQ params */
1844         err = c->xdp ? mlx5e_open_cq(c, params->tx_cq_moderation,
1845                                      &cparam->tx_cq, &c->rq_xdpsq.cq) : 0;
1846         if (err)
1847                 goto err_close_rx_cq;
1848
1849         napi_enable(&c->napi);
1850
1851         err = mlx5e_open_icosq(c, params, &cparam->icosq, &c->icosq);
1852         if (err)
1853                 goto err_disable_napi;
1854
1855         err = mlx5e_open_sqs(c, params, cparam);
1856         if (err)
1857                 goto err_close_icosq;
1858
1859         if (c->xdp) {
1860                 err = mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, NULL,
1861                                        &c->rq_xdpsq, false);
1862                 if (err)
1863                         goto err_close_sqs;
1864         }
1865
1866         err = mlx5e_open_rq(c, params, &cparam->rq, NULL, NULL, &c->rq);
1867         if (err)
1868                 goto err_close_xdp_sq;
1869
1870         err = mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, NULL, &c->xdpsq, true);
1871         if (err)
1872                 goto err_close_rq;
1873
1874         return 0;
1875
1876 err_close_rq:
1877         mlx5e_close_rq(&c->rq);
1878
1879 err_close_xdp_sq:
1880         if (c->xdp)
1881                 mlx5e_close_xdpsq(&c->rq_xdpsq);
1882
1883 err_close_sqs:
1884         mlx5e_close_sqs(c);
1885
1886 err_close_icosq:
1887         mlx5e_close_icosq(&c->icosq);
1888
1889 err_disable_napi:
1890         napi_disable(&c->napi);
1891
1892         if (c->xdp)
1893                 mlx5e_close_cq(&c->rq_xdpsq.cq);
1894
1895 err_close_rx_cq:
1896         mlx5e_close_cq(&c->rq.cq);
1897
1898 err_close_xdp_tx_cqs:
1899         mlx5e_close_cq(&c->xdpsq.cq);
1900
1901 err_close_tx_cqs:
1902         mlx5e_close_tx_cqs(c);
1903
1904 err_close_icosq_cq:
1905         mlx5e_close_cq(&c->icosq.cq);
1906
1907         return err;
1908 }
1909
1910 static void mlx5e_close_queues(struct mlx5e_channel *c)
1911 {
1912         mlx5e_close_xdpsq(&c->xdpsq);
1913         mlx5e_close_rq(&c->rq);
1914         if (c->xdp)
1915                 mlx5e_close_xdpsq(&c->rq_xdpsq);
1916         mlx5e_close_sqs(c);
1917         mlx5e_close_icosq(&c->icosq);
1918         napi_disable(&c->napi);
1919         if (c->xdp)
1920                 mlx5e_close_cq(&c->rq_xdpsq.cq);
1921         mlx5e_close_cq(&c->rq.cq);
1922         mlx5e_close_cq(&c->xdpsq.cq);
1923         mlx5e_close_tx_cqs(c);
1924         mlx5e_close_cq(&c->icosq.cq);
1925 }
1926
1927 static u8 mlx5e_enumerate_lag_port(struct mlx5_core_dev *mdev, int ix)
1928 {
1929         u16 port_aff_bias = mlx5_core_is_pf(mdev) ? 0 : MLX5_CAP_GEN(mdev, vhca_id);
1930
1931         return (ix + port_aff_bias) % mlx5e_get_num_lag_ports(mdev);
1932 }
1933
1934 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
1935                               struct mlx5e_params *params,
1936                               struct mlx5e_channel_param *cparam,
1937                               struct xdp_umem *umem,
1938                               struct mlx5e_channel **cp)
1939 {
1940         int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(priv->mdev, ix));
1941         struct net_device *netdev = priv->netdev;
1942         struct mlx5e_xsk_param xsk;
1943         struct mlx5e_channel *c;
1944         unsigned int irq;
1945         int err;
1946         int eqn;
1947
1948         err = mlx5_vector2eqn(priv->mdev, ix, &eqn, &irq);
1949         if (err)
1950                 return err;
1951
1952         c = kvzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
1953         if (!c)
1954                 return -ENOMEM;
1955
1956         c->priv     = priv;
1957         c->mdev     = priv->mdev;
1958         c->tstamp   = &priv->tstamp;
1959         c->ix       = ix;
1960         c->cpu      = cpu;
1961         c->pdev     = priv->mdev->device;
1962         c->netdev   = priv->netdev;
1963         c->mkey_be  = cpu_to_be32(priv->mdev->mlx5e_res.mkey.key);
1964         c->num_tc   = params->num_tc;
1965         c->xdp      = !!params->xdp_prog;
1966         c->stats    = &priv->channel_stats[ix].ch;
1967         c->irq_desc = irq_to_desc(irq);
1968         c->lag_port = mlx5e_enumerate_lag_port(priv->mdev, ix);
1969
1970         err = mlx5e_alloc_xps_cpumask(c, params);
1971         if (err)
1972                 goto err_free_channel;
1973
1974         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
1975
1976         err = mlx5e_open_queues(c, params, cparam);
1977         if (unlikely(err))
1978                 goto err_napi_del;
1979
1980         if (umem) {
1981                 mlx5e_build_xsk_param(umem, &xsk);
1982                 err = mlx5e_open_xsk(priv, params, &xsk, umem, c);
1983                 if (unlikely(err))
1984                         goto err_close_queues;
1985         }
1986
1987         *cp = c;
1988
1989         return 0;
1990
1991 err_close_queues:
1992         mlx5e_close_queues(c);
1993
1994 err_napi_del:
1995         netif_napi_del(&c->napi);
1996         mlx5e_free_xps_cpumask(c);
1997
1998 err_free_channel:
1999         kvfree(c);
2000
2001         return err;
2002 }
2003
2004 static void mlx5e_activate_channel(struct mlx5e_channel *c)
2005 {
2006         int tc;
2007
2008         for (tc = 0; tc < c->num_tc; tc++)
2009                 mlx5e_activate_txqsq(&c->sq[tc]);
2010         mlx5e_activate_icosq(&c->icosq);
2011         mlx5e_activate_rq(&c->rq);
2012         netif_set_xps_queue(c->netdev, c->xps_cpumask, c->ix);
2013
2014         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2015                 mlx5e_activate_xsk(c);
2016 }
2017
2018 static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
2019 {
2020         int tc;
2021
2022         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2023                 mlx5e_deactivate_xsk(c);
2024
2025         mlx5e_deactivate_rq(&c->rq);
2026         mlx5e_deactivate_icosq(&c->icosq);
2027         for (tc = 0; tc < c->num_tc; tc++)
2028                 mlx5e_deactivate_txqsq(&c->sq[tc]);
2029 }
2030
2031 static void mlx5e_close_channel(struct mlx5e_channel *c)
2032 {
2033         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2034                 mlx5e_close_xsk(c);
2035         mlx5e_close_queues(c);
2036         netif_napi_del(&c->napi);
2037         mlx5e_free_xps_cpumask(c);
2038
2039         kvfree(c);
2040 }
2041
2042 #define DEFAULT_FRAG_SIZE (2048)
2043
2044 static void mlx5e_build_rq_frags_info(struct mlx5_core_dev *mdev,
2045                                       struct mlx5e_params *params,
2046                                       struct mlx5e_xsk_param *xsk,
2047                                       struct mlx5e_rq_frags_info *info)
2048 {
2049         u32 byte_count = MLX5E_SW2HW_MTU(params, params->sw_mtu);
2050         int frag_size_max = DEFAULT_FRAG_SIZE;
2051         u32 buf_size = 0;
2052         int i;
2053
2054 #ifdef CONFIG_MLX5_EN_IPSEC
2055         if (MLX5_IPSEC_DEV(mdev))
2056                 byte_count += MLX5E_METADATA_ETHER_LEN;
2057 #endif
2058
2059         if (mlx5e_rx_is_linear_skb(params, xsk)) {
2060                 int frag_stride;
2061
2062                 frag_stride = mlx5e_rx_get_linear_frag_sz(params, xsk);
2063                 frag_stride = roundup_pow_of_two(frag_stride);
2064
2065                 info->arr[0].frag_size = byte_count;
2066                 info->arr[0].frag_stride = frag_stride;
2067                 info->num_frags = 1;
2068                 info->wqe_bulk = PAGE_SIZE / frag_stride;
2069                 goto out;
2070         }
2071
2072         if (byte_count > PAGE_SIZE +
2073             (MLX5E_MAX_RX_FRAGS - 1) * frag_size_max)
2074                 frag_size_max = PAGE_SIZE;
2075
2076         i = 0;
2077         while (buf_size < byte_count) {
2078                 int frag_size = byte_count - buf_size;
2079
2080                 if (i < MLX5E_MAX_RX_FRAGS - 1)
2081                         frag_size = min(frag_size, frag_size_max);
2082
2083                 info->arr[i].frag_size = frag_size;
2084                 info->arr[i].frag_stride = roundup_pow_of_two(frag_size);
2085
2086                 buf_size += frag_size;
2087                 i++;
2088         }
2089         info->num_frags = i;
2090         /* number of different wqes sharing a page */
2091         info->wqe_bulk = 1 + (info->num_frags % 2);
2092
2093 out:
2094         info->wqe_bulk = max_t(u8, info->wqe_bulk, 8);
2095         info->log_num_frags = order_base_2(info->num_frags);
2096 }
2097
2098 static inline u8 mlx5e_get_rqwq_log_stride(u8 wq_type, int ndsegs)
2099 {
2100         int sz = sizeof(struct mlx5_wqe_data_seg) * ndsegs;
2101
2102         switch (wq_type) {
2103         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2104                 sz += sizeof(struct mlx5e_rx_wqe_ll);
2105                 break;
2106         default: /* MLX5_WQ_TYPE_CYCLIC */
2107                 sz += sizeof(struct mlx5e_rx_wqe_cyc);
2108         }
2109
2110         return order_base_2(sz);
2111 }
2112
2113 static u8 mlx5e_get_rq_log_wq_sz(void *rqc)
2114 {
2115         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2116
2117         return MLX5_GET(wq, wq, log_wq_sz);
2118 }
2119
2120 void mlx5e_build_rq_param(struct mlx5e_priv *priv,
2121                           struct mlx5e_params *params,
2122                           struct mlx5e_xsk_param *xsk,
2123                           struct mlx5e_rq_param *param)
2124 {
2125         struct mlx5_core_dev *mdev = priv->mdev;
2126         void *rqc = param->rqc;
2127         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2128         int ndsegs = 1;
2129
2130         switch (params->rq_wq_type) {
2131         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2132                 MLX5_SET(wq, wq, log_wqe_num_of_strides,
2133                          mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk) -
2134                          MLX5_MPWQE_LOG_NUM_STRIDES_BASE);
2135                 MLX5_SET(wq, wq, log_wqe_stride_size,
2136                          mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk) -
2137                          MLX5_MPWQE_LOG_STRIDE_SZ_BASE);
2138                 MLX5_SET(wq, wq, log_wq_sz, mlx5e_mpwqe_get_log_rq_size(params, xsk));
2139                 break;
2140         default: /* MLX5_WQ_TYPE_CYCLIC */
2141                 MLX5_SET(wq, wq, log_wq_sz, params->log_rq_mtu_frames);
2142                 mlx5e_build_rq_frags_info(mdev, params, xsk, &param->frags_info);
2143                 ndsegs = param->frags_info.num_frags;
2144         }
2145
2146         MLX5_SET(wq, wq, wq_type,          params->rq_wq_type);
2147         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
2148         MLX5_SET(wq, wq, log_wq_stride,
2149                  mlx5e_get_rqwq_log_stride(params->rq_wq_type, ndsegs));
2150         MLX5_SET(wq, wq, pd,               mdev->mlx5e_res.pdn);
2151         MLX5_SET(rqc, rqc, counter_set_id, priv->q_counter);
2152         MLX5_SET(rqc, rqc, vsd,            params->vlan_strip_disable);
2153         MLX5_SET(rqc, rqc, scatter_fcs,    params->scatter_fcs_en);
2154
2155         param->wq.buf_numa_node = dev_to_node(mdev->device);
2156 }
2157
2158 static void mlx5e_build_drop_rq_param(struct mlx5e_priv *priv,
2159                                       struct mlx5e_rq_param *param)
2160 {
2161         struct mlx5_core_dev *mdev = priv->mdev;
2162         void *rqc = param->rqc;
2163         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2164
2165         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
2166         MLX5_SET(wq, wq, log_wq_stride,
2167                  mlx5e_get_rqwq_log_stride(MLX5_WQ_TYPE_CYCLIC, 1));
2168         MLX5_SET(rqc, rqc, counter_set_id, priv->drop_rq_q_counter);
2169
2170         param->wq.buf_numa_node = dev_to_node(mdev->device);
2171 }
2172
2173 void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
2174                                  struct mlx5e_sq_param *param)
2175 {
2176         void *sqc = param->sqc;
2177         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2178
2179         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
2180         MLX5_SET(wq, wq, pd,            priv->mdev->mlx5e_res.pdn);
2181
2182         param->wq.buf_numa_node = dev_to_node(priv->mdev->device);
2183 }
2184
2185 static void mlx5e_build_sq_param(struct mlx5e_priv *priv,
2186                                  struct mlx5e_params *params,
2187                                  struct mlx5e_sq_param *param)
2188 {
2189         void *sqc = param->sqc;
2190         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2191         bool allow_swp;
2192
2193         allow_swp = mlx5_geneve_tx_allowed(priv->mdev) ||
2194                     !!MLX5_IPSEC_DEV(priv->mdev);
2195         mlx5e_build_sq_param_common(priv, param);
2196         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
2197         MLX5_SET(sqc, sqc, allow_swp, allow_swp);
2198 }
2199
2200 static void mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
2201                                         struct mlx5e_cq_param *param)
2202 {
2203         void *cqc = param->cqc;
2204
2205         MLX5_SET(cqc, cqc, uar_page, priv->mdev->priv.uar->index);
2206         if (MLX5_CAP_GEN(priv->mdev, cqe_128_always) && cache_line_size() >= 128)
2207                 MLX5_SET(cqc, cqc, cqe_sz, CQE_STRIDE_128_PAD);
2208 }
2209
2210 void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
2211                              struct mlx5e_params *params,
2212                              struct mlx5e_xsk_param *xsk,
2213                              struct mlx5e_cq_param *param)
2214 {
2215         struct mlx5_core_dev *mdev = priv->mdev;
2216         void *cqc = param->cqc;
2217         u8 log_cq_size;
2218
2219         switch (params->rq_wq_type) {
2220         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2221                 log_cq_size = mlx5e_mpwqe_get_log_rq_size(params, xsk) +
2222                         mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk);
2223                 break;
2224         default: /* MLX5_WQ_TYPE_CYCLIC */
2225                 log_cq_size = params->log_rq_mtu_frames;
2226         }
2227
2228         MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
2229         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
2230                 MLX5_SET(cqc, cqc, mini_cqe_res_format, MLX5_CQE_FORMAT_CSUM);
2231                 MLX5_SET(cqc, cqc, cqe_comp_en, 1);
2232         }
2233
2234         mlx5e_build_common_cq_param(priv, param);
2235         param->cq_period_mode = params->rx_cq_moderation.cq_period_mode;
2236 }
2237
2238 void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
2239                              struct mlx5e_params *params,
2240                              struct mlx5e_cq_param *param)
2241 {
2242         void *cqc = param->cqc;
2243
2244         MLX5_SET(cqc, cqc, log_cq_size, params->log_sq_size);
2245
2246         mlx5e_build_common_cq_param(priv, param);
2247         param->cq_period_mode = params->tx_cq_moderation.cq_period_mode;
2248 }
2249
2250 void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
2251                               u8 log_wq_size,
2252                               struct mlx5e_cq_param *param)
2253 {
2254         void *cqc = param->cqc;
2255
2256         MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
2257
2258         mlx5e_build_common_cq_param(priv, param);
2259
2260         param->cq_period_mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
2261 }
2262
2263 void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
2264                              u8 log_wq_size,
2265                              struct mlx5e_sq_param *param)
2266 {
2267         void *sqc = param->sqc;
2268         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2269
2270         mlx5e_build_sq_param_common(priv, param);
2271
2272         MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
2273         MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(priv->mdev, reg_umr_sq));
2274 }
2275
2276 void mlx5e_build_xdpsq_param(struct mlx5e_priv *priv,
2277                              struct mlx5e_params *params,
2278                              struct mlx5e_sq_param *param)
2279 {
2280         void *sqc = param->sqc;
2281         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2282
2283         mlx5e_build_sq_param_common(priv, param);
2284         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
2285         param->is_mpw = MLX5E_GET_PFLAG(params, MLX5E_PFLAG_XDP_TX_MPWQE);
2286 }
2287
2288 static u8 mlx5e_build_icosq_log_wq_sz(struct mlx5e_params *params,
2289                                       struct mlx5e_rq_param *rqp)
2290 {
2291         switch (params->rq_wq_type) {
2292         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2293                 return order_base_2(MLX5E_UMR_WQEBBS) +
2294                         mlx5e_get_rq_log_wq_sz(rqp->rqc);
2295         default: /* MLX5_WQ_TYPE_CYCLIC */
2296                 return MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
2297         }
2298 }
2299
2300 static void mlx5e_build_channel_param(struct mlx5e_priv *priv,
2301                                       struct mlx5e_params *params,
2302                                       struct mlx5e_channel_param *cparam)
2303 {
2304         u8 icosq_log_wq_sz;
2305
2306         mlx5e_build_rq_param(priv, params, NULL, &cparam->rq);
2307
2308         icosq_log_wq_sz = mlx5e_build_icosq_log_wq_sz(params, &cparam->rq);
2309
2310         mlx5e_build_sq_param(priv, params, &cparam->sq);
2311         mlx5e_build_xdpsq_param(priv, params, &cparam->xdp_sq);
2312         mlx5e_build_icosq_param(priv, icosq_log_wq_sz, &cparam->icosq);
2313         mlx5e_build_rx_cq_param(priv, params, NULL, &cparam->rx_cq);
2314         mlx5e_build_tx_cq_param(priv, params, &cparam->tx_cq);
2315         mlx5e_build_ico_cq_param(priv, icosq_log_wq_sz, &cparam->icosq_cq);
2316 }
2317
2318 int mlx5e_open_channels(struct mlx5e_priv *priv,
2319                         struct mlx5e_channels *chs)
2320 {
2321         struct mlx5e_channel_param *cparam;
2322         int err = -ENOMEM;
2323         int i;
2324
2325         chs->num = chs->params.num_channels;
2326
2327         chs->c = kcalloc(chs->num, sizeof(struct mlx5e_channel *), GFP_KERNEL);
2328         cparam = kvzalloc(sizeof(struct mlx5e_channel_param), GFP_KERNEL);
2329         if (!chs->c || !cparam)
2330                 goto err_free;
2331
2332         mlx5e_build_channel_param(priv, &chs->params, cparam);
2333         for (i = 0; i < chs->num; i++) {
2334                 struct xdp_umem *umem = NULL;
2335
2336                 if (chs->params.xdp_prog)
2337                         umem = mlx5e_xsk_get_umem(&chs->params, chs->params.xsk, i);
2338
2339                 err = mlx5e_open_channel(priv, i, &chs->params, cparam, umem, &chs->c[i]);
2340                 if (err)
2341                         goto err_close_channels;
2342         }
2343
2344         mlx5e_health_channels_update(priv);
2345         kvfree(cparam);
2346         return 0;
2347
2348 err_close_channels:
2349         for (i--; i >= 0; i--)
2350                 mlx5e_close_channel(chs->c[i]);
2351
2352 err_free:
2353         kfree(chs->c);
2354         kvfree(cparam);
2355         chs->num = 0;
2356         return err;
2357 }
2358
2359 static void mlx5e_activate_channels(struct mlx5e_channels *chs)
2360 {
2361         int i;
2362
2363         for (i = 0; i < chs->num; i++)
2364                 mlx5e_activate_channel(chs->c[i]);
2365 }
2366
2367 #define MLX5E_RQ_WQES_TIMEOUT 20000 /* msecs */
2368
2369 static int mlx5e_wait_channels_min_rx_wqes(struct mlx5e_channels *chs)
2370 {
2371         int err = 0;
2372         int i;
2373
2374         for (i = 0; i < chs->num; i++) {
2375                 int timeout = err ? 0 : MLX5E_RQ_WQES_TIMEOUT;
2376
2377                 err |= mlx5e_wait_for_min_rx_wqes(&chs->c[i]->rq, timeout);
2378
2379                 /* Don't wait on the XSK RQ, because the newer xdpsock sample
2380                  * doesn't provide any Fill Ring entries at the setup stage.
2381                  */
2382         }
2383
2384         return err ? -ETIMEDOUT : 0;
2385 }
2386
2387 static void mlx5e_deactivate_channels(struct mlx5e_channels *chs)
2388 {
2389         int i;
2390
2391         for (i = 0; i < chs->num; i++)
2392                 mlx5e_deactivate_channel(chs->c[i]);
2393 }
2394
2395 void mlx5e_close_channels(struct mlx5e_channels *chs)
2396 {
2397         int i;
2398
2399         for (i = 0; i < chs->num; i++)
2400                 mlx5e_close_channel(chs->c[i]);
2401
2402         kfree(chs->c);
2403         chs->num = 0;
2404 }
2405
2406 static int
2407 mlx5e_create_rqt(struct mlx5e_priv *priv, int sz, struct mlx5e_rqt *rqt)
2408 {
2409         struct mlx5_core_dev *mdev = priv->mdev;
2410         void *rqtc;
2411         int inlen;
2412         int err;
2413         u32 *in;
2414         int i;
2415
2416         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2417         in = kvzalloc(inlen, GFP_KERNEL);
2418         if (!in)
2419                 return -ENOMEM;
2420
2421         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2422
2423         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2424         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2425
2426         for (i = 0; i < sz; i++)
2427                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->drop_rq.rqn);
2428
2429         err = mlx5_core_create_rqt(mdev, in, inlen, &rqt->rqtn);
2430         if (!err)
2431                 rqt->enabled = true;
2432
2433         kvfree(in);
2434         return err;
2435 }
2436
2437 void mlx5e_destroy_rqt(struct mlx5e_priv *priv, struct mlx5e_rqt *rqt)
2438 {
2439         rqt->enabled = false;
2440         mlx5_core_destroy_rqt(priv->mdev, rqt->rqtn);
2441 }
2442
2443 int mlx5e_create_indirect_rqt(struct mlx5e_priv *priv)
2444 {
2445         struct mlx5e_rqt *rqt = &priv->indir_rqt;
2446         int err;
2447
2448         err = mlx5e_create_rqt(priv, MLX5E_INDIR_RQT_SIZE, rqt);
2449         if (err)
2450                 mlx5_core_warn(priv->mdev, "create indirect rqts failed, %d\n", err);
2451         return err;
2452 }
2453
2454 int mlx5e_create_direct_rqts(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
2455 {
2456         int err;
2457         int ix;
2458
2459         for (ix = 0; ix < priv->max_nch; ix++) {
2460                 err = mlx5e_create_rqt(priv, 1 /*size */, &tirs[ix].rqt);
2461                 if (unlikely(err))
2462                         goto err_destroy_rqts;
2463         }
2464
2465         return 0;
2466
2467 err_destroy_rqts:
2468         mlx5_core_warn(priv->mdev, "create rqts failed, %d\n", err);
2469         for (ix--; ix >= 0; ix--)
2470                 mlx5e_destroy_rqt(priv, &tirs[ix].rqt);
2471
2472         return err;
2473 }
2474
2475 void mlx5e_destroy_direct_rqts(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
2476 {
2477         int i;
2478
2479         for (i = 0; i < priv->max_nch; i++)
2480                 mlx5e_destroy_rqt(priv, &tirs[i].rqt);
2481 }
2482
2483 static int mlx5e_rx_hash_fn(int hfunc)
2484 {
2485         return (hfunc == ETH_RSS_HASH_TOP) ?
2486                MLX5_RX_HASH_FN_TOEPLITZ :
2487                MLX5_RX_HASH_FN_INVERTED_XOR8;
2488 }
2489
2490 int mlx5e_bits_invert(unsigned long a, int size)
2491 {
2492         int inv = 0;
2493         int i;
2494
2495         for (i = 0; i < size; i++)
2496                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
2497
2498         return inv;
2499 }
2500
2501 static void mlx5e_fill_rqt_rqns(struct mlx5e_priv *priv, int sz,
2502                                 struct mlx5e_redirect_rqt_param rrp, void *rqtc)
2503 {
2504         int i;
2505
2506         for (i = 0; i < sz; i++) {
2507                 u32 rqn;
2508
2509                 if (rrp.is_rss) {
2510                         int ix = i;
2511
2512                         if (rrp.rss.hfunc == ETH_RSS_HASH_XOR)
2513                                 ix = mlx5e_bits_invert(i, ilog2(sz));
2514
2515                         ix = priv->rss_params.indirection_rqt[ix];
2516                         rqn = rrp.rss.channels->c[ix]->rq.rqn;
2517                 } else {
2518                         rqn = rrp.rqn;
2519                 }
2520                 MLX5_SET(rqtc, rqtc, rq_num[i], rqn);
2521         }
2522 }
2523
2524 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, u32 rqtn, int sz,
2525                        struct mlx5e_redirect_rqt_param rrp)
2526 {
2527         struct mlx5_core_dev *mdev = priv->mdev;
2528         void *rqtc;
2529         int inlen;
2530         u32 *in;
2531         int err;
2532
2533         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
2534         in = kvzalloc(inlen, GFP_KERNEL);
2535         if (!in)
2536                 return -ENOMEM;
2537
2538         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
2539
2540         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2541         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
2542         mlx5e_fill_rqt_rqns(priv, sz, rrp, rqtc);
2543         err = mlx5_core_modify_rqt(mdev, rqtn, in, inlen);
2544
2545         kvfree(in);
2546         return err;
2547 }
2548
2549 static u32 mlx5e_get_direct_rqn(struct mlx5e_priv *priv, int ix,
2550                                 struct mlx5e_redirect_rqt_param rrp)
2551 {
2552         if (!rrp.is_rss)
2553                 return rrp.rqn;
2554
2555         if (ix >= rrp.rss.channels->num)
2556                 return priv->drop_rq.rqn;
2557
2558         return rrp.rss.channels->c[ix]->rq.rqn;
2559 }
2560
2561 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv,
2562                                 struct mlx5e_redirect_rqt_param rrp)
2563 {
2564         u32 rqtn;
2565         int ix;
2566
2567         if (priv->indir_rqt.enabled) {
2568                 /* RSS RQ table */
2569                 rqtn = priv->indir_rqt.rqtn;
2570                 mlx5e_redirect_rqt(priv, rqtn, MLX5E_INDIR_RQT_SIZE, rrp);
2571         }
2572
2573         for (ix = 0; ix < priv->max_nch; ix++) {
2574                 struct mlx5e_redirect_rqt_param direct_rrp = {
2575                         .is_rss = false,
2576                         {
2577                                 .rqn    = mlx5e_get_direct_rqn(priv, ix, rrp)
2578                         },
2579                 };
2580
2581                 /* Direct RQ Tables */
2582                 if (!priv->direct_tir[ix].rqt.enabled)
2583                         continue;
2584
2585                 rqtn = priv->direct_tir[ix].rqt.rqtn;
2586                 mlx5e_redirect_rqt(priv, rqtn, 1, direct_rrp);
2587         }
2588 }
2589
2590 static void mlx5e_redirect_rqts_to_channels(struct mlx5e_priv *priv,
2591                                             struct mlx5e_channels *chs)
2592 {
2593         struct mlx5e_redirect_rqt_param rrp = {
2594                 .is_rss        = true,
2595                 {
2596                         .rss = {
2597                                 .channels  = chs,
2598                                 .hfunc     = priv->rss_params.hfunc,
2599                         }
2600                 },
2601         };
2602
2603         mlx5e_redirect_rqts(priv, rrp);
2604 }
2605
2606 static void mlx5e_redirect_rqts_to_drop(struct mlx5e_priv *priv)
2607 {
2608         struct mlx5e_redirect_rqt_param drop_rrp = {
2609                 .is_rss = false,
2610                 {
2611                         .rqn = priv->drop_rq.rqn,
2612                 },
2613         };
2614
2615         mlx5e_redirect_rqts(priv, drop_rrp);
2616 }
2617
2618 static const struct mlx5e_tirc_config tirc_default_config[MLX5E_NUM_INDIR_TIRS] = {
2619         [MLX5E_TT_IPV4_TCP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2620                                 .l4_prot_type = MLX5_L4_PROT_TYPE_TCP,
2621                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2622         },
2623         [MLX5E_TT_IPV6_TCP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2624                                 .l4_prot_type = MLX5_L4_PROT_TYPE_TCP,
2625                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2626         },
2627         [MLX5E_TT_IPV4_UDP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2628                                 .l4_prot_type = MLX5_L4_PROT_TYPE_UDP,
2629                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2630         },
2631         [MLX5E_TT_IPV6_UDP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2632                                 .l4_prot_type = MLX5_L4_PROT_TYPE_UDP,
2633                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2634         },
2635         [MLX5E_TT_IPV4_IPSEC_AH] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2636                                      .l4_prot_type = 0,
2637                                      .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2638         },
2639         [MLX5E_TT_IPV6_IPSEC_AH] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2640                                      .l4_prot_type = 0,
2641                                      .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2642         },
2643         [MLX5E_TT_IPV4_IPSEC_ESP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2644                                       .l4_prot_type = 0,
2645                                       .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2646         },
2647         [MLX5E_TT_IPV6_IPSEC_ESP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2648                                       .l4_prot_type = 0,
2649                                       .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2650         },
2651         [MLX5E_TT_IPV4] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2652                             .l4_prot_type = 0,
2653                             .rx_hash_fields = MLX5_HASH_IP,
2654         },
2655         [MLX5E_TT_IPV6] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2656                             .l4_prot_type = 0,
2657                             .rx_hash_fields = MLX5_HASH_IP,
2658         },
2659 };
2660
2661 struct mlx5e_tirc_config mlx5e_tirc_get_default_config(enum mlx5e_traffic_types tt)
2662 {
2663         return tirc_default_config[tt];
2664 }
2665
2666 static void mlx5e_build_tir_ctx_lro(struct mlx5e_params *params, void *tirc)
2667 {
2668         if (!params->lro_en)
2669                 return;
2670
2671 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2672
2673         MLX5_SET(tirc, tirc, lro_enable_mask,
2674                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2675                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2676         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
2677                  (MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ - ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2678         MLX5_SET(tirc, tirc, lro_timeout_period_usecs, params->lro_timeout);
2679 }
2680
2681 void mlx5e_build_indir_tir_ctx_hash(struct mlx5e_rss_params *rss_params,
2682                                     const struct mlx5e_tirc_config *ttconfig,
2683                                     void *tirc, bool inner)
2684 {
2685         void *hfso = inner ? MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner) :
2686                              MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2687
2688         MLX5_SET(tirc, tirc, rx_hash_fn, mlx5e_rx_hash_fn(rss_params->hfunc));
2689         if (rss_params->hfunc == ETH_RSS_HASH_TOP) {
2690                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
2691                                              rx_hash_toeplitz_key);
2692                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
2693                                                rx_hash_toeplitz_key);
2694
2695                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2696                 memcpy(rss_key, rss_params->toeplitz_hash_key, len);
2697         }
2698         MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2699                  ttconfig->l3_prot_type);
2700         MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2701                  ttconfig->l4_prot_type);
2702         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2703                  ttconfig->rx_hash_fields);
2704 }
2705
2706 static void mlx5e_update_rx_hash_fields(struct mlx5e_tirc_config *ttconfig,
2707                                         enum mlx5e_traffic_types tt,
2708                                         u32 rx_hash_fields)
2709 {
2710         *ttconfig                = tirc_default_config[tt];
2711         ttconfig->rx_hash_fields = rx_hash_fields;
2712 }
2713
2714 void mlx5e_modify_tirs_hash(struct mlx5e_priv *priv, void *in, int inlen)
2715 {
2716         void *tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2717         struct mlx5e_rss_params *rss = &priv->rss_params;
2718         struct mlx5_core_dev *mdev = priv->mdev;
2719         int ctxlen = MLX5_ST_SZ_BYTES(tirc);
2720         struct mlx5e_tirc_config ttconfig;
2721         int tt;
2722
2723         MLX5_SET(modify_tir_in, in, bitmask.hash, 1);
2724
2725         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2726                 memset(tirc, 0, ctxlen);
2727                 mlx5e_update_rx_hash_fields(&ttconfig, tt,
2728                                             rss->rx_hash_fields[tt]);
2729                 mlx5e_build_indir_tir_ctx_hash(rss, &ttconfig, tirc, false);
2730                 mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in, inlen);
2731         }
2732
2733         if (!mlx5e_tunnel_inner_ft_supported(priv->mdev))
2734                 return;
2735
2736         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2737                 memset(tirc, 0, ctxlen);
2738                 mlx5e_update_rx_hash_fields(&ttconfig, tt,
2739                                             rss->rx_hash_fields[tt]);
2740                 mlx5e_build_indir_tir_ctx_hash(rss, &ttconfig, tirc, true);
2741                 mlx5_core_modify_tir(mdev, priv->inner_indir_tir[tt].tirn, in,
2742                                      inlen);
2743         }
2744 }
2745
2746 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
2747 {
2748         struct mlx5_core_dev *mdev = priv->mdev;
2749
2750         void *in;
2751         void *tirc;
2752         int inlen;
2753         int err;
2754         int tt;
2755         int ix;
2756
2757         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
2758         in = kvzalloc(inlen, GFP_KERNEL);
2759         if (!in)
2760                 return -ENOMEM;
2761
2762         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
2763         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2764
2765         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2766
2767         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2768                 err = mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in,
2769                                            inlen);
2770                 if (err)
2771                         goto free_in;
2772         }
2773
2774         for (ix = 0; ix < priv->max_nch; ix++) {
2775                 err = mlx5_core_modify_tir(mdev, priv->direct_tir[ix].tirn,
2776                                            in, inlen);
2777                 if (err)
2778                         goto free_in;
2779         }
2780
2781 free_in:
2782         kvfree(in);
2783
2784         return err;
2785 }
2786
2787 static int mlx5e_set_mtu(struct mlx5_core_dev *mdev,
2788                          struct mlx5e_params *params, u16 mtu)
2789 {
2790         u16 hw_mtu = MLX5E_SW2HW_MTU(params, mtu);
2791         int err;
2792
2793         err = mlx5_set_port_mtu(mdev, hw_mtu, 1);
2794         if (err)
2795                 return err;
2796
2797         /* Update vport context MTU */
2798         mlx5_modify_nic_vport_mtu(mdev, hw_mtu);
2799         return 0;
2800 }
2801
2802 static void mlx5e_query_mtu(struct mlx5_core_dev *mdev,
2803                             struct mlx5e_params *params, u16 *mtu)
2804 {
2805         u16 hw_mtu = 0;
2806         int err;
2807
2808         err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2809         if (err || !hw_mtu) /* fallback to port oper mtu */
2810                 mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
2811
2812         *mtu = MLX5E_HW2SW_MTU(params, hw_mtu);
2813 }
2814
2815 int mlx5e_set_dev_port_mtu(struct mlx5e_priv *priv)
2816 {
2817         struct mlx5e_params *params = &priv->channels.params;
2818         struct net_device *netdev = priv->netdev;
2819         struct mlx5_core_dev *mdev = priv->mdev;
2820         u16 mtu;
2821         int err;
2822
2823         err = mlx5e_set_mtu(mdev, params, params->sw_mtu);
2824         if (err)
2825                 return err;
2826
2827         mlx5e_query_mtu(mdev, params, &mtu);
2828         if (mtu != params->sw_mtu)
2829                 netdev_warn(netdev, "%s: VPort MTU %d is different than netdev mtu %d\n",
2830                             __func__, mtu, params->sw_mtu);
2831
2832         params->sw_mtu = mtu;
2833         return 0;
2834 }
2835
2836 void mlx5e_set_netdev_mtu_boundaries(struct mlx5e_priv *priv)
2837 {
2838         struct mlx5e_params *params = &priv->channels.params;
2839         struct net_device *netdev   = priv->netdev;
2840         struct mlx5_core_dev *mdev  = priv->mdev;
2841         u16 max_mtu;
2842
2843         /* MTU range: 68 - hw-specific max */
2844         netdev->min_mtu = ETH_MIN_MTU;
2845
2846         mlx5_query_port_max_mtu(mdev, &max_mtu, 1);
2847         netdev->max_mtu = min_t(unsigned int, MLX5E_HW2SW_MTU(params, max_mtu),
2848                                 ETH_MAX_MTU);
2849 }
2850
2851 static void mlx5e_netdev_set_tcs(struct net_device *netdev)
2852 {
2853         struct mlx5e_priv *priv = netdev_priv(netdev);
2854         int nch = priv->channels.params.num_channels;
2855         int ntc = priv->channels.params.num_tc;
2856         int tc;
2857
2858         netdev_reset_tc(netdev);
2859
2860         if (ntc == 1)
2861                 return;
2862
2863         netdev_set_num_tc(netdev, ntc);
2864
2865         /* Map netdev TCs to offset 0
2866          * We have our own UP to TXQ mapping for QoS
2867          */
2868         for (tc = 0; tc < ntc; tc++)
2869                 netdev_set_tc_queue(netdev, tc, nch, 0);
2870 }
2871
2872 static void mlx5e_update_netdev_queues(struct mlx5e_priv *priv)
2873 {
2874         int num_txqs = priv->channels.num * priv->channels.params.num_tc;
2875         int num_rxqs = priv->channels.num * priv->profile->rq_groups;
2876         struct net_device *netdev = priv->netdev;
2877
2878         mlx5e_netdev_set_tcs(netdev);
2879         netif_set_real_num_tx_queues(netdev, num_txqs);
2880         netif_set_real_num_rx_queues(netdev, num_rxqs);
2881 }
2882
2883 static void mlx5e_build_txq_maps(struct mlx5e_priv *priv)
2884 {
2885         int i, ch;
2886
2887         ch = priv->channels.num;
2888
2889         for (i = 0; i < ch; i++) {
2890                 int tc;
2891
2892                 for (tc = 0; tc < priv->channels.params.num_tc; tc++) {
2893                         struct mlx5e_channel *c = priv->channels.c[i];
2894                         struct mlx5e_txqsq *sq = &c->sq[tc];
2895
2896                         priv->txq2sq[sq->txq_ix] = sq;
2897                         priv->channel_tc2realtxq[i][tc] = i + tc * ch;
2898                 }
2899         }
2900 }
2901
2902 void mlx5e_activate_priv_channels(struct mlx5e_priv *priv)
2903 {
2904         mlx5e_update_netdev_queues(priv);
2905
2906         mlx5e_build_txq_maps(priv);
2907         mlx5e_activate_channels(&priv->channels);
2908         mlx5e_xdp_tx_enable(priv);
2909         netif_tx_start_all_queues(priv->netdev);
2910
2911         if (mlx5e_is_vport_rep(priv))
2912                 mlx5e_add_sqs_fwd_rules(priv);
2913
2914         mlx5e_wait_channels_min_rx_wqes(&priv->channels);
2915         mlx5e_redirect_rqts_to_channels(priv, &priv->channels);
2916
2917         mlx5e_xsk_redirect_rqts_to_channels(priv, &priv->channels);
2918 }
2919
2920 void mlx5e_deactivate_priv_channels(struct mlx5e_priv *priv)
2921 {
2922         mlx5e_xsk_redirect_rqts_to_drop(priv, &priv->channels);
2923
2924         mlx5e_redirect_rqts_to_drop(priv);
2925
2926         if (mlx5e_is_vport_rep(priv))
2927                 mlx5e_remove_sqs_fwd_rules(priv);
2928
2929         /* FIXME: This is a W/A only for tx timeout watch dog false alarm when
2930          * polling for inactive tx queues.
2931          */
2932         netif_tx_stop_all_queues(priv->netdev);
2933         netif_tx_disable(priv->netdev);
2934         mlx5e_xdp_tx_disable(priv);
2935         mlx5e_deactivate_channels(&priv->channels);
2936 }
2937
2938 static void mlx5e_switch_priv_channels(struct mlx5e_priv *priv,
2939                                        struct mlx5e_channels *new_chs,
2940                                        mlx5e_fp_preactivate preactivate)
2941 {
2942         struct net_device *netdev = priv->netdev;
2943         int new_num_txqs;
2944         int carrier_ok;
2945
2946         new_num_txqs = new_chs->num * new_chs->params.num_tc;
2947
2948         carrier_ok = netif_carrier_ok(netdev);
2949         netif_carrier_off(netdev);
2950
2951         if (new_num_txqs < netdev->real_num_tx_queues)
2952                 netif_set_real_num_tx_queues(netdev, new_num_txqs);
2953
2954         mlx5e_deactivate_priv_channels(priv);
2955         mlx5e_close_channels(&priv->channels);
2956
2957         priv->channels = *new_chs;
2958
2959         /* New channels are ready to roll, call the preactivate hook if needed
2960          * to modify HW settings or update kernel parameters.
2961          */
2962         if (preactivate)
2963                 preactivate(priv);
2964
2965         priv->profile->update_rx(priv);
2966         mlx5e_activate_priv_channels(priv);
2967
2968         /* return carrier back if needed */
2969         if (carrier_ok)
2970                 netif_carrier_on(netdev);
2971 }
2972
2973 int mlx5e_safe_switch_channels(struct mlx5e_priv *priv,
2974                                struct mlx5e_channels *new_chs,
2975                                mlx5e_fp_preactivate preactivate)
2976 {
2977         int err;
2978
2979         err = mlx5e_open_channels(priv, new_chs);
2980         if (err)
2981                 return err;
2982
2983         mlx5e_switch_priv_channels(priv, new_chs, preactivate);
2984         return 0;
2985 }
2986
2987 int mlx5e_safe_reopen_channels(struct mlx5e_priv *priv)
2988 {
2989         struct mlx5e_channels new_channels = {};
2990
2991         new_channels.params = priv->channels.params;
2992         return mlx5e_safe_switch_channels(priv, &new_channels, NULL);
2993 }
2994
2995 void mlx5e_timestamp_init(struct mlx5e_priv *priv)
2996 {
2997         priv->tstamp.tx_type   = HWTSTAMP_TX_OFF;
2998         priv->tstamp.rx_filter = HWTSTAMP_FILTER_NONE;
2999 }
3000
3001 int mlx5e_open_locked(struct net_device *netdev)
3002 {
3003         struct mlx5e_priv *priv = netdev_priv(netdev);
3004         int err;
3005
3006         set_bit(MLX5E_STATE_OPENED, &priv->state);
3007
3008         err = mlx5e_open_channels(priv, &priv->channels);
3009         if (err)
3010                 goto err_clear_state_opened_flag;
3011
3012         priv->profile->update_rx(priv);
3013         mlx5e_activate_priv_channels(priv);
3014         if (priv->profile->update_carrier)
3015                 priv->profile->update_carrier(priv);
3016
3017         mlx5e_queue_update_stats(priv);
3018         return 0;
3019
3020 err_clear_state_opened_flag:
3021         clear_bit(MLX5E_STATE_OPENED, &priv->state);
3022         return err;
3023 }
3024
3025 int mlx5e_open(struct net_device *netdev)
3026 {
3027         struct mlx5e_priv *priv = netdev_priv(netdev);
3028         int err;
3029
3030         mutex_lock(&priv->state_lock);
3031         err = mlx5e_open_locked(netdev);
3032         if (!err)
3033                 mlx5_set_port_admin_status(priv->mdev, MLX5_PORT_UP);
3034         mutex_unlock(&priv->state_lock);
3035
3036         if (mlx5_vxlan_allowed(priv->mdev->vxlan))
3037                 udp_tunnel_get_rx_info(netdev);
3038
3039         return err;
3040 }
3041
3042 int mlx5e_close_locked(struct net_device *netdev)
3043 {
3044         struct mlx5e_priv *priv = netdev_priv(netdev);
3045
3046         /* May already be CLOSED in case a previous configuration operation
3047          * (e.g RX/TX queue size change) that involves close&open failed.
3048          */
3049         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3050                 return 0;
3051
3052         clear_bit(MLX5E_STATE_OPENED, &priv->state);
3053
3054         netif_carrier_off(priv->netdev);
3055         mlx5e_deactivate_priv_channels(priv);
3056         mlx5e_close_channels(&priv->channels);
3057
3058         return 0;
3059 }
3060
3061 int mlx5e_close(struct net_device *netdev)
3062 {
3063         struct mlx5e_priv *priv = netdev_priv(netdev);
3064         int err;
3065
3066         if (!netif_device_present(netdev))
3067                 return -ENODEV;
3068
3069         mutex_lock(&priv->state_lock);
3070         mlx5_set_port_admin_status(priv->mdev, MLX5_PORT_DOWN);
3071         err = mlx5e_close_locked(netdev);
3072         mutex_unlock(&priv->state_lock);
3073
3074         return err;
3075 }
3076
3077 static int mlx5e_alloc_drop_rq(struct mlx5_core_dev *mdev,
3078                                struct mlx5e_rq *rq,
3079                                struct mlx5e_rq_param *param)
3080 {
3081         void *rqc = param->rqc;
3082         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
3083         int err;
3084
3085         param->wq.db_numa_node = param->wq.buf_numa_node;
3086
3087         err = mlx5_wq_cyc_create(mdev, &param->wq, rqc_wq, &rq->wqe.wq,
3088                                  &rq->wq_ctrl);
3089         if (err)
3090                 return err;
3091
3092         /* Mark as unused given "Drop-RQ" packets never reach XDP */
3093         xdp_rxq_info_unused(&rq->xdp_rxq);
3094
3095         rq->mdev = mdev;
3096
3097         return 0;
3098 }
3099
3100 static int mlx5e_alloc_drop_cq(struct mlx5_core_dev *mdev,
3101                                struct mlx5e_cq *cq,
3102                                struct mlx5e_cq_param *param)
3103 {
3104         param->wq.buf_numa_node = dev_to_node(mdev->device);
3105         param->wq.db_numa_node  = dev_to_node(mdev->device);
3106
3107         return mlx5e_alloc_cq_common(mdev, param, cq);
3108 }
3109
3110 int mlx5e_open_drop_rq(struct mlx5e_priv *priv,
3111                        struct mlx5e_rq *drop_rq)
3112 {
3113         struct mlx5_core_dev *mdev = priv->mdev;
3114         struct mlx5e_cq_param cq_param = {};
3115         struct mlx5e_rq_param rq_param = {};
3116         struct mlx5e_cq *cq = &drop_rq->cq;
3117         int err;
3118
3119         mlx5e_build_drop_rq_param(priv, &rq_param);
3120
3121         err = mlx5e_alloc_drop_cq(mdev, cq, &cq_param);
3122         if (err)
3123                 return err;
3124
3125         err = mlx5e_create_cq(cq, &cq_param);
3126         if (err)
3127                 goto err_free_cq;
3128
3129         err = mlx5e_alloc_drop_rq(mdev, drop_rq, &rq_param);
3130         if (err)
3131                 goto err_destroy_cq;
3132
3133         err = mlx5e_create_rq(drop_rq, &rq_param);
3134         if (err)
3135                 goto err_free_rq;
3136
3137         err = mlx5e_modify_rq_state(drop_rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
3138         if (err)
3139                 mlx5_core_warn(priv->mdev, "modify_rq_state failed, rx_if_down_packets won't be counted %d\n", err);
3140
3141         return 0;
3142
3143 err_free_rq:
3144         mlx5e_free_rq(drop_rq);
3145
3146 err_destroy_cq:
3147         mlx5e_destroy_cq(cq);
3148
3149 err_free_cq:
3150         mlx5e_free_cq(cq);
3151
3152         return err;
3153 }
3154
3155 void mlx5e_close_drop_rq(struct mlx5e_rq *drop_rq)
3156 {
3157         mlx5e_destroy_rq(drop_rq);
3158         mlx5e_free_rq(drop_rq);
3159         mlx5e_destroy_cq(&drop_rq->cq);
3160         mlx5e_free_cq(&drop_rq->cq);
3161 }
3162
3163 int mlx5e_create_tis(struct mlx5_core_dev *mdev, void *in, u32 *tisn)
3164 {
3165         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
3166
3167         MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.td.tdn);
3168
3169         if (MLX5_GET(tisc, tisc, tls_en))
3170                 MLX5_SET(tisc, tisc, pd, mdev->mlx5e_res.pdn);
3171
3172         if (mlx5_lag_is_lacp_owner(mdev))
3173                 MLX5_SET(tisc, tisc, strict_lag_tx_port_affinity, 1);
3174
3175         return mlx5_core_create_tis(mdev, in, MLX5_ST_SZ_BYTES(create_tis_in), tisn);
3176 }
3177
3178 void mlx5e_destroy_tis(struct mlx5_core_dev *mdev, u32 tisn)
3179 {
3180         mlx5_core_destroy_tis(mdev, tisn);
3181 }
3182
3183 void mlx5e_destroy_tises(struct mlx5e_priv *priv)
3184 {
3185         int tc, i;
3186
3187         for (i = 0; i < mlx5e_get_num_lag_ports(priv->mdev); i++)
3188                 for (tc = 0; tc < priv->profile->max_tc; tc++)
3189                         mlx5e_destroy_tis(priv->mdev, priv->tisn[i][tc]);
3190 }
3191
3192 static bool mlx5e_lag_should_assign_affinity(struct mlx5_core_dev *mdev)
3193 {
3194         return MLX5_CAP_GEN(mdev, lag_tx_port_affinity) && mlx5e_get_num_lag_ports(mdev) > 1;
3195 }
3196
3197 int mlx5e_create_tises(struct mlx5e_priv *priv)
3198 {
3199         int tc, i;
3200         int err;
3201
3202         for (i = 0; i < mlx5e_get_num_lag_ports(priv->mdev); i++) {
3203                 for (tc = 0; tc < priv->profile->max_tc; tc++) {
3204                         u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
3205                         void *tisc;
3206
3207                         tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
3208
3209                         MLX5_SET(tisc, tisc, prio, tc << 1);
3210
3211                         if (mlx5e_lag_should_assign_affinity(priv->mdev))
3212                                 MLX5_SET(tisc, tisc, lag_tx_port_affinity, i + 1);
3213
3214                         err = mlx5e_create_tis(priv->mdev, in, &priv->tisn[i][tc]);
3215                         if (err)
3216                                 goto err_close_tises;
3217                 }
3218         }
3219
3220         return 0;
3221
3222 err_close_tises:
3223         for (; i >= 0; i--) {
3224                 for (tc--; tc >= 0; tc--)
3225                         mlx5e_destroy_tis(priv->mdev, priv->tisn[i][tc]);
3226                 tc = priv->profile->max_tc;
3227         }
3228
3229         return err;
3230 }
3231
3232 static void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
3233 {
3234         mlx5e_destroy_tises(priv);
3235 }
3236
3237 static void mlx5e_build_indir_tir_ctx_common(struct mlx5e_priv *priv,
3238                                              u32 rqtn, u32 *tirc)
3239 {
3240         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
3241         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
3242         MLX5_SET(tirc, tirc, indirect_table, rqtn);
3243         MLX5_SET(tirc, tirc, tunneled_offload_en,
3244                  priv->channels.params.tunneled_offload_en);
3245
3246         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
3247 }
3248
3249 static void mlx5e_build_indir_tir_ctx(struct mlx5e_priv *priv,
3250                                       enum mlx5e_traffic_types tt,
3251                                       u32 *tirc)
3252 {
3253         mlx5e_build_indir_tir_ctx_common(priv, priv->indir_rqt.rqtn, tirc);
3254         mlx5e_build_indir_tir_ctx_hash(&priv->rss_params,
3255                                        &tirc_default_config[tt], tirc, false);
3256 }
3257
3258 static void mlx5e_build_direct_tir_ctx(struct mlx5e_priv *priv, u32 rqtn, u32 *tirc)
3259 {
3260         mlx5e_build_indir_tir_ctx_common(priv, rqtn, tirc);
3261         MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
3262 }
3263
3264 static void mlx5e_build_inner_indir_tir_ctx(struct mlx5e_priv *priv,
3265                                             enum mlx5e_traffic_types tt,
3266                                             u32 *tirc)
3267 {
3268         mlx5e_build_indir_tir_ctx_common(priv, priv->indir_rqt.rqtn, tirc);
3269         mlx5e_build_indir_tir_ctx_hash(&priv->rss_params,
3270                                        &tirc_default_config[tt], tirc, true);
3271 }
3272
3273 int mlx5e_create_indirect_tirs(struct mlx5e_priv *priv, bool inner_ttc)
3274 {
3275         struct mlx5e_tir *tir;
3276         void *tirc;
3277         int inlen;
3278         int i = 0;
3279         int err;
3280         u32 *in;
3281         int tt;
3282
3283         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
3284         in = kvzalloc(inlen, GFP_KERNEL);
3285         if (!in)
3286                 return -ENOMEM;
3287
3288         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
3289                 memset(in, 0, inlen);
3290                 tir = &priv->indir_tir[tt];
3291                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3292                 mlx5e_build_indir_tir_ctx(priv, tt, tirc);
3293                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
3294                 if (err) {
3295                         mlx5_core_warn(priv->mdev, "create indirect tirs failed, %d\n", err);
3296                         goto err_destroy_inner_tirs;
3297                 }
3298         }
3299
3300         if (!inner_ttc || !mlx5e_tunnel_inner_ft_supported(priv->mdev))
3301                 goto out;
3302
3303         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++) {
3304                 memset(in, 0, inlen);
3305                 tir = &priv->inner_indir_tir[i];
3306                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3307                 mlx5e_build_inner_indir_tir_ctx(priv, i, tirc);
3308                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
3309                 if (err) {
3310                         mlx5_core_warn(priv->mdev, "create inner indirect tirs failed, %d\n", err);
3311                         goto err_destroy_inner_tirs;
3312                 }
3313         }
3314
3315 out:
3316         kvfree(in);
3317
3318         return 0;
3319
3320 err_destroy_inner_tirs:
3321         for (i--; i >= 0; i--)
3322                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3323
3324         for (tt--; tt >= 0; tt--)
3325                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[tt]);
3326
3327         kvfree(in);
3328
3329         return err;
3330 }
3331
3332 int mlx5e_create_direct_tirs(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
3333 {
3334         struct mlx5e_tir *tir;
3335         void *tirc;
3336         int inlen;
3337         int err = 0;
3338         u32 *in;
3339         int ix;
3340
3341         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
3342         in = kvzalloc(inlen, GFP_KERNEL);
3343         if (!in)
3344                 return -ENOMEM;
3345
3346         for (ix = 0; ix < priv->max_nch; ix++) {
3347                 memset(in, 0, inlen);
3348                 tir = &tirs[ix];
3349                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3350                 mlx5e_build_direct_tir_ctx(priv, tir->rqt.rqtn, tirc);
3351                 err = mlx5e_create_tir(priv->mdev, tir, in, inlen);
3352                 if (unlikely(err))
3353                         goto err_destroy_ch_tirs;
3354         }
3355
3356         goto out;
3357
3358 err_destroy_ch_tirs:
3359         mlx5_core_warn(priv->mdev, "create tirs failed, %d\n", err);
3360         for (ix--; ix >= 0; ix--)
3361                 mlx5e_destroy_tir(priv->mdev, &tirs[ix]);
3362
3363 out:
3364         kvfree(in);
3365
3366         return err;
3367 }
3368
3369 void mlx5e_destroy_indirect_tirs(struct mlx5e_priv *priv, bool inner_ttc)
3370 {
3371         int i;
3372
3373         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3374                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[i]);
3375
3376         if (!inner_ttc || !mlx5e_tunnel_inner_ft_supported(priv->mdev))
3377                 return;
3378
3379         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3380                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3381 }
3382
3383 void mlx5e_destroy_direct_tirs(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
3384 {
3385         int i;
3386
3387         for (i = 0; i < priv->max_nch; i++)
3388                 mlx5e_destroy_tir(priv->mdev, &tirs[i]);
3389 }
3390
3391 static int mlx5e_modify_channels_scatter_fcs(struct mlx5e_channels *chs, bool enable)
3392 {
3393         int err = 0;
3394         int i;
3395
3396         for (i = 0; i < chs->num; i++) {
3397                 err = mlx5e_modify_rq_scatter_fcs(&chs->c[i]->rq, enable);
3398                 if (err)
3399                         return err;
3400         }
3401
3402         return 0;
3403 }
3404
3405 static int mlx5e_modify_channels_vsd(struct mlx5e_channels *chs, bool vsd)
3406 {
3407         int err = 0;
3408         int i;
3409
3410         for (i = 0; i < chs->num; i++) {
3411                 err = mlx5e_modify_rq_vsd(&chs->c[i]->rq, vsd);
3412                 if (err)
3413                         return err;
3414         }
3415
3416         return 0;
3417 }
3418
3419 static int mlx5e_setup_tc_mqprio(struct mlx5e_priv *priv,
3420                                  struct tc_mqprio_qopt *mqprio)
3421 {
3422         struct mlx5e_channels new_channels = {};
3423         u8 tc = mqprio->num_tc;
3424         int err = 0;
3425
3426         mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
3427
3428         if (tc && tc != MLX5E_MAX_NUM_TC)
3429                 return -EINVAL;
3430
3431         mutex_lock(&priv->state_lock);
3432
3433         new_channels.params = priv->channels.params;
3434         new_channels.params.num_tc = tc ? tc : 1;
3435
3436         if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
3437                 priv->channels.params = new_channels.params;
3438                 goto out;
3439         }
3440
3441         err = mlx5e_safe_switch_channels(priv, &new_channels, NULL);
3442         if (err)
3443                 goto out;
3444
3445         priv->max_opened_tc = max_t(u8, priv->max_opened_tc,
3446                                     new_channels.params.num_tc);
3447 out:
3448         mutex_unlock(&priv->state_lock);
3449         return err;
3450 }
3451
3452 #ifdef CONFIG_MLX5_ESWITCH
3453 static int mlx5e_setup_tc_cls_flower(struct mlx5e_priv *priv,
3454                                      struct flow_cls_offload *cls_flower,
3455                                      unsigned long flags)
3456 {
3457         switch (cls_flower->command) {
3458         case FLOW_CLS_REPLACE:
3459                 return mlx5e_configure_flower(priv->netdev, priv, cls_flower,
3460                                               flags);
3461         case FLOW_CLS_DESTROY:
3462                 return mlx5e_delete_flower(priv->netdev, priv, cls_flower,
3463                                            flags);
3464         case FLOW_CLS_STATS:
3465                 return mlx5e_stats_flower(priv->netdev, priv, cls_flower,
3466                                           flags);
3467         default:
3468                 return -EOPNOTSUPP;
3469         }
3470 }
3471
3472 static int mlx5e_setup_tc_block_cb(enum tc_setup_type type, void *type_data,
3473                                    void *cb_priv)
3474 {
3475         unsigned long flags = MLX5_TC_FLAG(INGRESS) | MLX5_TC_FLAG(NIC_OFFLOAD);
3476         struct mlx5e_priv *priv = cb_priv;
3477
3478         switch (type) {
3479         case TC_SETUP_CLSFLOWER:
3480                 return mlx5e_setup_tc_cls_flower(priv, type_data, flags);
3481         default:
3482                 return -EOPNOTSUPP;
3483         }
3484 }
3485 #endif
3486
3487 static LIST_HEAD(mlx5e_block_cb_list);
3488
3489 static int mlx5e_setup_tc(struct net_device *dev, enum tc_setup_type type,
3490                           void *type_data)
3491 {
3492         struct mlx5e_priv *priv = netdev_priv(dev);
3493
3494         switch (type) {
3495 #ifdef CONFIG_MLX5_ESWITCH
3496         case TC_SETUP_BLOCK: {
3497                 struct flow_block_offload *f = type_data;
3498
3499                 f->unlocked_driver_cb = true;
3500                 return flow_block_cb_setup_simple(type_data,
3501                                                   &mlx5e_block_cb_list,
3502                                                   mlx5e_setup_tc_block_cb,
3503                                                   priv, priv, true);
3504         }
3505 #endif
3506         case TC_SETUP_QDISC_MQPRIO:
3507                 return mlx5e_setup_tc_mqprio(priv, type_data);
3508         default:
3509                 return -EOPNOTSUPP;
3510         }
3511 }
3512
3513 void mlx5e_fold_sw_stats64(struct mlx5e_priv *priv, struct rtnl_link_stats64 *s)
3514 {
3515         int i;
3516
3517         for (i = 0; i < priv->max_nch; i++) {
3518                 struct mlx5e_channel_stats *channel_stats = &priv->channel_stats[i];
3519                 struct mlx5e_rq_stats *xskrq_stats = &channel_stats->xskrq;
3520                 struct mlx5e_rq_stats *rq_stats = &channel_stats->rq;
3521                 int j;
3522
3523                 s->rx_packets   += rq_stats->packets + xskrq_stats->packets;
3524                 s->rx_bytes     += rq_stats->bytes + xskrq_stats->bytes;
3525
3526                 for (j = 0; j < priv->max_opened_tc; j++) {
3527                         struct mlx5e_sq_stats *sq_stats = &channel_stats->sq[j];
3528
3529                         s->tx_packets    += sq_stats->packets;
3530                         s->tx_bytes      += sq_stats->bytes;
3531                         s->tx_dropped    += sq_stats->dropped;
3532                 }
3533         }
3534 }
3535
3536 void
3537 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
3538 {
3539         struct mlx5e_priv *priv = netdev_priv(dev);
3540         struct mlx5e_vport_stats *vstats = &priv->stats.vport;
3541         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
3542
3543         if (!mlx5e_monitor_counter_supported(priv)) {
3544                 /* update HW stats in background for next time */
3545                 mlx5e_queue_update_stats(priv);
3546         }
3547
3548         if (mlx5e_is_uplink_rep(priv)) {
3549                 stats->rx_packets = PPORT_802_3_GET(pstats, a_frames_received_ok);
3550                 stats->rx_bytes   = PPORT_802_3_GET(pstats, a_octets_received_ok);
3551                 stats->tx_packets = PPORT_802_3_GET(pstats, a_frames_transmitted_ok);
3552                 stats->tx_bytes   = PPORT_802_3_GET(pstats, a_octets_transmitted_ok);
3553         } else {
3554                 mlx5e_fold_sw_stats64(priv, stats);
3555         }
3556
3557         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
3558
3559         stats->rx_length_errors =
3560                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
3561                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
3562                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
3563         stats->rx_crc_errors =
3564                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
3565         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
3566         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
3567         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
3568                            stats->rx_frame_errors;
3569         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
3570
3571         /* vport multicast also counts packets that are dropped due to steering
3572          * or rx out of buffer
3573          */
3574         stats->multicast =
3575                 VPORT_COUNTER_GET(vstats, received_eth_multicast.packets);
3576 }
3577
3578 static void mlx5e_set_rx_mode(struct net_device *dev)
3579 {
3580         struct mlx5e_priv *priv = netdev_priv(dev);
3581
3582         queue_work(priv->wq, &priv->set_rx_mode_work);
3583 }
3584
3585 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
3586 {
3587         struct mlx5e_priv *priv = netdev_priv(netdev);
3588         struct sockaddr *saddr = addr;
3589
3590         if (!is_valid_ether_addr(saddr->sa_data))
3591                 return -EADDRNOTAVAIL;
3592
3593         netif_addr_lock_bh(netdev);
3594         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
3595         netif_addr_unlock_bh(netdev);
3596
3597         queue_work(priv->wq, &priv->set_rx_mode_work);
3598
3599         return 0;
3600 }
3601
3602 #define MLX5E_SET_FEATURE(features, feature, enable)    \
3603         do {                                            \
3604                 if (enable)                             \
3605                         *features |= feature;           \
3606                 else                                    \
3607                         *features &= ~feature;          \
3608         } while (0)
3609
3610 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
3611
3612 static int set_feature_lro(struct net_device *netdev, bool enable)
3613 {
3614         struct mlx5e_priv *priv = netdev_priv(netdev);
3615         struct mlx5_core_dev *mdev = priv->mdev;
3616         struct mlx5e_channels new_channels = {};
3617         struct mlx5e_params *old_params;
3618         int err = 0;
3619         bool reset;
3620
3621         mutex_lock(&priv->state_lock);
3622
3623         if (enable && priv->xsk.refcnt) {
3624                 netdev_warn(netdev, "LRO is incompatible with AF_XDP (%hu XSKs are active)\n",
3625                             priv->xsk.refcnt);
3626                 err = -EINVAL;
3627                 goto out;
3628         }
3629
3630         old_params = &priv->channels.params;
3631         if (enable && !MLX5E_GET_PFLAG(old_params, MLX5E_PFLAG_RX_STRIDING_RQ)) {
3632                 netdev_warn(netdev, "can't set LRO with legacy RQ\n");
3633                 err = -EINVAL;
3634                 goto out;
3635         }
3636
3637         reset = test_bit(MLX5E_STATE_OPENED, &priv->state);
3638
3639         new_channels.params = *old_params;
3640         new_channels.params.lro_en = enable;
3641
3642         if (old_params->rq_wq_type != MLX5_WQ_TYPE_CYCLIC) {
3643                 if (mlx5e_rx_mpwqe_is_linear_skb(mdev, old_params, NULL) ==
3644                     mlx5e_rx_mpwqe_is_linear_skb(mdev, &new_channels.params, NULL))
3645                         reset = false;
3646         }
3647
3648         if (!reset) {
3649                 *old_params = new_channels.params;
3650                 err = mlx5e_modify_tirs_lro(priv);
3651                 goto out;
3652         }
3653
3654         err = mlx5e_safe_switch_channels(priv, &new_channels, mlx5e_modify_tirs_lro);
3655 out:
3656         mutex_unlock(&priv->state_lock);
3657         return err;
3658 }
3659
3660 static int set_feature_cvlan_filter(struct net_device *netdev, bool enable)
3661 {
3662         struct mlx5e_priv *priv = netdev_priv(netdev);
3663
3664         if (enable)
3665                 mlx5e_enable_cvlan_filter(priv);
3666         else
3667                 mlx5e_disable_cvlan_filter(priv);
3668
3669         return 0;
3670 }
3671
3672 #ifdef CONFIG_MLX5_ESWITCH
3673 static int set_feature_tc_num_filters(struct net_device *netdev, bool enable)
3674 {
3675         struct mlx5e_priv *priv = netdev_priv(netdev);
3676
3677         if (!enable && mlx5e_tc_num_filters(priv, MLX5_TC_FLAG(NIC_OFFLOAD))) {
3678                 netdev_err(netdev,
3679                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
3680                 return -EINVAL;
3681         }
3682
3683         return 0;
3684 }
3685 #endif
3686
3687 static int set_feature_rx_all(struct net_device *netdev, bool enable)
3688 {
3689         struct mlx5e_priv *priv = netdev_priv(netdev);
3690         struct mlx5_core_dev *mdev = priv->mdev;
3691
3692         return mlx5_set_port_fcs(mdev, !enable);
3693 }
3694
3695 static int set_feature_rx_fcs(struct net_device *netdev, bool enable)
3696 {
3697         struct mlx5e_priv *priv = netdev_priv(netdev);
3698         int err;
3699
3700         mutex_lock(&priv->state_lock);
3701
3702         priv->channels.params.scatter_fcs_en = enable;
3703         err = mlx5e_modify_channels_scatter_fcs(&priv->channels, enable);
3704         if (err)
3705                 priv->channels.params.scatter_fcs_en = !enable;
3706
3707         mutex_unlock(&priv->state_lock);
3708
3709         return err;
3710 }
3711
3712 static int set_feature_rx_vlan(struct net_device *netdev, bool enable)
3713 {
3714         struct mlx5e_priv *priv = netdev_priv(netdev);
3715         int err = 0;
3716
3717         mutex_lock(&priv->state_lock);
3718
3719         priv->channels.params.vlan_strip_disable = !enable;
3720         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3721                 goto unlock;
3722
3723         err = mlx5e_modify_channels_vsd(&priv->channels, !enable);
3724         if (err)
3725                 priv->channels.params.vlan_strip_disable = enable;
3726
3727 unlock:
3728         mutex_unlock(&priv->state_lock);
3729
3730         return err;
3731 }
3732
3733 #ifdef CONFIG_MLX5_EN_ARFS
3734 static int set_feature_arfs(struct net_device *netdev, bool enable)
3735 {
3736         struct mlx5e_priv *priv = netdev_priv(netdev);
3737         int err;
3738
3739         if (enable)
3740                 err = mlx5e_arfs_enable(priv);
3741         else
3742                 err = mlx5e_arfs_disable(priv);
3743
3744         return err;
3745 }
3746 #endif
3747
3748 static int mlx5e_handle_feature(struct net_device *netdev,
3749                                 netdev_features_t *features,
3750                                 netdev_features_t wanted_features,
3751                                 netdev_features_t feature,
3752                                 mlx5e_feature_handler feature_handler)
3753 {
3754         netdev_features_t changes = wanted_features ^ netdev->features;
3755         bool enable = !!(wanted_features & feature);
3756         int err;
3757
3758         if (!(changes & feature))
3759                 return 0;
3760
3761         err = feature_handler(netdev, enable);
3762         if (err) {
3763                 netdev_err(netdev, "%s feature %pNF failed, err %d\n",
3764                            enable ? "Enable" : "Disable", &feature, err);
3765                 return err;
3766         }
3767
3768         MLX5E_SET_FEATURE(features, feature, enable);
3769         return 0;
3770 }
3771
3772 int mlx5e_set_features(struct net_device *netdev, netdev_features_t features)
3773 {
3774         netdev_features_t oper_features = netdev->features;
3775         int err = 0;
3776
3777 #define MLX5E_HANDLE_FEATURE(feature, handler) \
3778         mlx5e_handle_feature(netdev, &oper_features, features, feature, handler)
3779
3780         err |= MLX5E_HANDLE_FEATURE(NETIF_F_LRO, set_feature_lro);
3781         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_VLAN_CTAG_FILTER,
3782                                     set_feature_cvlan_filter);
3783 #ifdef CONFIG_MLX5_ESWITCH
3784         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_TC, set_feature_tc_num_filters);
3785 #endif
3786         err |= MLX5E_HANDLE_FEATURE(NETIF_F_RXALL, set_feature_rx_all);
3787         err |= MLX5E_HANDLE_FEATURE(NETIF_F_RXFCS, set_feature_rx_fcs);
3788         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_VLAN_CTAG_RX, set_feature_rx_vlan);
3789 #ifdef CONFIG_MLX5_EN_ARFS
3790         err |= MLX5E_HANDLE_FEATURE(NETIF_F_NTUPLE, set_feature_arfs);
3791 #endif
3792
3793         if (err) {
3794                 netdev->features = oper_features;
3795                 return -EINVAL;
3796         }
3797
3798         return 0;
3799 }
3800
3801 static netdev_features_t mlx5e_fix_features(struct net_device *netdev,
3802                                             netdev_features_t features)
3803 {
3804         struct mlx5e_priv *priv = netdev_priv(netdev);
3805         struct mlx5e_params *params;
3806
3807         mutex_lock(&priv->state_lock);
3808         params = &priv->channels.params;
3809         if (!bitmap_empty(priv->fs.vlan.active_svlans, VLAN_N_VID)) {
3810                 /* HW strips the outer C-tag header, this is a problem
3811                  * for S-tag traffic.
3812                  */
3813                 features &= ~NETIF_F_HW_VLAN_CTAG_RX;
3814                 if (!params->vlan_strip_disable)
3815                         netdev_warn(netdev, "Dropping C-tag vlan stripping offload due to S-tag vlan\n");
3816         }
3817         if (!MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ)) {
3818                 if (features & NETIF_F_LRO) {
3819                         netdev_warn(netdev, "Disabling LRO, not supported in legacy RQ\n");
3820                         features &= ~NETIF_F_LRO;
3821                 }
3822         }
3823
3824         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
3825                 features &= ~NETIF_F_RXHASH;
3826                 if (netdev->features & NETIF_F_RXHASH)
3827                         netdev_warn(netdev, "Disabling rxhash, not supported when CQE compress is active\n");
3828         }
3829
3830         mutex_unlock(&priv->state_lock);
3831
3832         return features;
3833 }
3834
3835 static bool mlx5e_xsk_validate_mtu(struct net_device *netdev,
3836                                    struct mlx5e_channels *chs,
3837                                    struct mlx5e_params *new_params,
3838                                    struct mlx5_core_dev *mdev)
3839 {
3840         u16 ix;
3841
3842         for (ix = 0; ix < chs->params.num_channels; ix++) {
3843                 struct xdp_umem *umem = mlx5e_xsk_get_umem(&chs->params, chs->params.xsk, ix);
3844                 struct mlx5e_xsk_param xsk;
3845
3846                 if (!umem)
3847                         continue;
3848
3849                 mlx5e_build_xsk_param(umem, &xsk);
3850
3851                 if (!mlx5e_validate_xsk_param(new_params, &xsk, mdev)) {
3852                         u32 hr = mlx5e_get_linear_rq_headroom(new_params, &xsk);
3853                         int max_mtu_frame, max_mtu_page, max_mtu;
3854
3855                         /* Two criteria must be met:
3856                          * 1. HW MTU + all headrooms <= XSK frame size.
3857                          * 2. Size of SKBs allocated on XDP_PASS <= PAGE_SIZE.
3858                          */
3859                         max_mtu_frame = MLX5E_HW2SW_MTU(new_params, xsk.chunk_size - hr);
3860                         max_mtu_page = mlx5e_xdp_max_mtu(new_params, &xsk);
3861                         max_mtu = min(max_mtu_frame, max_mtu_page);
3862
3863                         netdev_err(netdev, "MTU %d is too big for an XSK running on channel %hu. Try MTU <= %d\n",
3864                                    new_params->sw_mtu, ix, max_mtu);
3865                         return false;
3866                 }
3867         }
3868
3869         return true;
3870 }
3871
3872 int mlx5e_change_mtu(struct net_device *netdev, int new_mtu,
3873                      change_hw_mtu_cb set_mtu_cb)
3874 {
3875         struct mlx5e_priv *priv = netdev_priv(netdev);
3876         struct mlx5e_channels new_channels = {};
3877         struct mlx5e_params *params;
3878         int err = 0;
3879         bool reset;
3880
3881         mutex_lock(&priv->state_lock);
3882
3883         params = &priv->channels.params;
3884
3885         reset = !params->lro_en;
3886         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
3887
3888         new_channels.params = *params;
3889         new_channels.params.sw_mtu = new_mtu;
3890
3891         if (params->xdp_prog &&
3892             !mlx5e_rx_is_linear_skb(&new_channels.params, NULL)) {
3893                 netdev_err(netdev, "MTU(%d) > %d is not allowed while XDP enabled\n",
3894                            new_mtu, mlx5e_xdp_max_mtu(params, NULL));
3895                 err = -EINVAL;
3896                 goto out;
3897         }
3898
3899         if (priv->xsk.refcnt &&
3900             !mlx5e_xsk_validate_mtu(netdev, &priv->channels,
3901                                     &new_channels.params, priv->mdev)) {
3902                 err = -EINVAL;
3903                 goto out;
3904         }
3905
3906         if (params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
3907                 bool is_linear = mlx5e_rx_mpwqe_is_linear_skb(priv->mdev,
3908                                                               &new_channels.params,
3909                                                               NULL);
3910                 u8 ppw_old = mlx5e_mpwqe_log_pkts_per_wqe(params, NULL);
3911                 u8 ppw_new = mlx5e_mpwqe_log_pkts_per_wqe(&new_channels.params, NULL);
3912
3913                 /* If XSK is active, XSK RQs are linear. */
3914                 is_linear |= priv->xsk.refcnt;
3915
3916                 /* Always reset in linear mode - hw_mtu is used in data path. */
3917                 reset = reset && (is_linear || (ppw_old != ppw_new));
3918         }
3919
3920         if (!reset) {
3921                 params->sw_mtu = new_mtu;
3922                 if (set_mtu_cb)
3923                         set_mtu_cb(priv);
3924                 netdev->mtu = params->sw_mtu;
3925                 goto out;
3926         }
3927
3928         err = mlx5e_safe_switch_channels(priv, &new_channels, set_mtu_cb);
3929         if (err)
3930                 goto out;
3931
3932         netdev->mtu = new_channels.params.sw_mtu;
3933
3934 out:
3935         mutex_unlock(&priv->state_lock);
3936         return err;
3937 }
3938
3939 static int mlx5e_change_nic_mtu(struct net_device *netdev, int new_mtu)
3940 {
3941         return mlx5e_change_mtu(netdev, new_mtu, mlx5e_set_dev_port_mtu);
3942 }
3943
3944 int mlx5e_hwstamp_set(struct mlx5e_priv *priv, struct ifreq *ifr)
3945 {
3946         struct hwtstamp_config config;
3947         int err;
3948
3949         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz) ||
3950             (mlx5_clock_get_ptp_index(priv->mdev) == -1))
3951                 return -EOPNOTSUPP;
3952
3953         if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
3954                 return -EFAULT;
3955
3956         /* TX HW timestamp */
3957         switch (config.tx_type) {
3958         case HWTSTAMP_TX_OFF:
3959         case HWTSTAMP_TX_ON:
3960                 break;
3961         default:
3962                 return -ERANGE;
3963         }
3964
3965         mutex_lock(&priv->state_lock);
3966         /* RX HW timestamp */
3967         switch (config.rx_filter) {
3968         case HWTSTAMP_FILTER_NONE:
3969                 /* Reset CQE compression to Admin default */
3970                 mlx5e_modify_rx_cqe_compression_locked(priv, priv->channels.params.rx_cqe_compress_def);
3971                 break;
3972         case HWTSTAMP_FILTER_ALL:
3973         case HWTSTAMP_FILTER_SOME:
3974         case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
3975         case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
3976         case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
3977         case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
3978         case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
3979         case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
3980         case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
3981         case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
3982         case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
3983         case HWTSTAMP_FILTER_PTP_V2_EVENT:
3984         case HWTSTAMP_FILTER_PTP_V2_SYNC:
3985         case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
3986         case HWTSTAMP_FILTER_NTP_ALL:
3987                 /* Disable CQE compression */
3988                 if (MLX5E_GET_PFLAG(&priv->channels.params, MLX5E_PFLAG_RX_CQE_COMPRESS))
3989                         netdev_warn(priv->netdev, "Disabling RX cqe compression\n");
3990                 err = mlx5e_modify_rx_cqe_compression_locked(priv, false);
3991                 if (err) {
3992                         netdev_err(priv->netdev, "Failed disabling cqe compression err=%d\n", err);
3993                         mutex_unlock(&priv->state_lock);
3994                         return err;
3995                 }
3996                 config.rx_filter = HWTSTAMP_FILTER_ALL;
3997                 break;
3998         default:
3999                 mutex_unlock(&priv->state_lock);
4000                 return -ERANGE;
4001         }
4002
4003         memcpy(&priv->tstamp, &config, sizeof(config));
4004         mutex_unlock(&priv->state_lock);
4005
4006         /* might need to fix some features */
4007         netdev_update_features(priv->netdev);
4008
4009         return copy_to_user(ifr->ifr_data, &config,
4010                             sizeof(config)) ? -EFAULT : 0;
4011 }
4012
4013 int mlx5e_hwstamp_get(struct mlx5e_priv *priv, struct ifreq *ifr)
4014 {
4015         struct hwtstamp_config *cfg = &priv->tstamp;
4016
4017         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
4018                 return -EOPNOTSUPP;
4019
4020         return copy_to_user(ifr->ifr_data, cfg, sizeof(*cfg)) ? -EFAULT : 0;
4021 }
4022
4023 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
4024 {
4025         struct mlx5e_priv *priv = netdev_priv(dev);
4026
4027         switch (cmd) {
4028         case SIOCSHWTSTAMP:
4029                 return mlx5e_hwstamp_set(priv, ifr);
4030         case SIOCGHWTSTAMP:
4031                 return mlx5e_hwstamp_get(priv, ifr);
4032         default:
4033                 return -EOPNOTSUPP;
4034         }
4035 }
4036
4037 #ifdef CONFIG_MLX5_ESWITCH
4038 int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
4039 {
4040         struct mlx5e_priv *priv = netdev_priv(dev);
4041         struct mlx5_core_dev *mdev = priv->mdev;
4042
4043         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
4044 }
4045
4046 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos,
4047                              __be16 vlan_proto)
4048 {
4049         struct mlx5e_priv *priv = netdev_priv(dev);
4050         struct mlx5_core_dev *mdev = priv->mdev;
4051
4052         if (vlan_proto != htons(ETH_P_8021Q))
4053                 return -EPROTONOSUPPORT;
4054
4055         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
4056                                            vlan, qos);
4057 }
4058
4059 static int mlx5e_set_vf_spoofchk(struct net_device *dev, int vf, bool setting)
4060 {
4061         struct mlx5e_priv *priv = netdev_priv(dev);
4062         struct mlx5_core_dev *mdev = priv->mdev;
4063
4064         return mlx5_eswitch_set_vport_spoofchk(mdev->priv.eswitch, vf + 1, setting);
4065 }
4066
4067 static int mlx5e_set_vf_trust(struct net_device *dev, int vf, bool setting)
4068 {
4069         struct mlx5e_priv *priv = netdev_priv(dev);
4070         struct mlx5_core_dev *mdev = priv->mdev;
4071
4072         return mlx5_eswitch_set_vport_trust(mdev->priv.eswitch, vf + 1, setting);
4073 }
4074
4075 int mlx5e_set_vf_rate(struct net_device *dev, int vf, int min_tx_rate,
4076                       int max_tx_rate)
4077 {
4078         struct mlx5e_priv *priv = netdev_priv(dev);
4079         struct mlx5_core_dev *mdev = priv->mdev;
4080
4081         return mlx5_eswitch_set_vport_rate(mdev->priv.eswitch, vf + 1,
4082                                            max_tx_rate, min_tx_rate);
4083 }
4084
4085 static int mlx5_vport_link2ifla(u8 esw_link)
4086 {
4087         switch (esw_link) {
4088         case MLX5_VPORT_ADMIN_STATE_DOWN:
4089                 return IFLA_VF_LINK_STATE_DISABLE;
4090         case MLX5_VPORT_ADMIN_STATE_UP:
4091                 return IFLA_VF_LINK_STATE_ENABLE;
4092         }
4093         return IFLA_VF_LINK_STATE_AUTO;
4094 }
4095
4096 static int mlx5_ifla_link2vport(u8 ifla_link)
4097 {
4098         switch (ifla_link) {
4099         case IFLA_VF_LINK_STATE_DISABLE:
4100                 return MLX5_VPORT_ADMIN_STATE_DOWN;
4101         case IFLA_VF_LINK_STATE_ENABLE:
4102                 return MLX5_VPORT_ADMIN_STATE_UP;
4103         }
4104         return MLX5_VPORT_ADMIN_STATE_AUTO;
4105 }
4106
4107 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
4108                                    int link_state)
4109 {
4110         struct mlx5e_priv *priv = netdev_priv(dev);
4111         struct mlx5_core_dev *mdev = priv->mdev;
4112
4113         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
4114                                             mlx5_ifla_link2vport(link_state));
4115 }
4116
4117 int mlx5e_get_vf_config(struct net_device *dev,
4118                         int vf, struct ifla_vf_info *ivi)
4119 {
4120         struct mlx5e_priv *priv = netdev_priv(dev);
4121         struct mlx5_core_dev *mdev = priv->mdev;
4122         int err;
4123
4124         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
4125         if (err)
4126                 return err;
4127         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
4128         return 0;
4129 }
4130
4131 int mlx5e_get_vf_stats(struct net_device *dev,
4132                        int vf, struct ifla_vf_stats *vf_stats)
4133 {
4134         struct mlx5e_priv *priv = netdev_priv(dev);
4135         struct mlx5_core_dev *mdev = priv->mdev;
4136
4137         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
4138                                             vf_stats);
4139 }
4140 #endif
4141
4142 struct mlx5e_vxlan_work {
4143         struct work_struct      work;
4144         struct mlx5e_priv       *priv;
4145         u16                     port;
4146 };
4147
4148 static void mlx5e_vxlan_add_work(struct work_struct *work)
4149 {
4150         struct mlx5e_vxlan_work *vxlan_work =
4151                 container_of(work, struct mlx5e_vxlan_work, work);
4152         struct mlx5e_priv *priv = vxlan_work->priv;
4153         u16 port = vxlan_work->port;
4154
4155         mutex_lock(&priv->state_lock);
4156         mlx5_vxlan_add_port(priv->mdev->vxlan, port);
4157         mutex_unlock(&priv->state_lock);
4158
4159         kfree(vxlan_work);
4160 }
4161
4162 static void mlx5e_vxlan_del_work(struct work_struct *work)
4163 {
4164         struct mlx5e_vxlan_work *vxlan_work =
4165                 container_of(work, struct mlx5e_vxlan_work, work);
4166         struct mlx5e_priv *priv         = vxlan_work->priv;
4167         u16 port = vxlan_work->port;
4168
4169         mutex_lock(&priv->state_lock);
4170         mlx5_vxlan_del_port(priv->mdev->vxlan, port);
4171         mutex_unlock(&priv->state_lock);
4172         kfree(vxlan_work);
4173 }
4174
4175 static void mlx5e_vxlan_queue_work(struct mlx5e_priv *priv, u16 port, int add)
4176 {
4177         struct mlx5e_vxlan_work *vxlan_work;
4178
4179         vxlan_work = kmalloc(sizeof(*vxlan_work), GFP_ATOMIC);
4180         if (!vxlan_work)
4181                 return;
4182
4183         if (add)
4184                 INIT_WORK(&vxlan_work->work, mlx5e_vxlan_add_work);
4185         else
4186                 INIT_WORK(&vxlan_work->work, mlx5e_vxlan_del_work);
4187
4188         vxlan_work->priv = priv;
4189         vxlan_work->port = port;
4190         queue_work(priv->wq, &vxlan_work->work);
4191 }
4192
4193 void mlx5e_add_vxlan_port(struct net_device *netdev, struct udp_tunnel_info *ti)
4194 {
4195         struct mlx5e_priv *priv = netdev_priv(netdev);
4196
4197         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
4198                 return;
4199
4200         if (!mlx5_vxlan_allowed(priv->mdev->vxlan))
4201                 return;
4202
4203         mlx5e_vxlan_queue_work(priv, be16_to_cpu(ti->port), 1);
4204 }
4205
4206 void mlx5e_del_vxlan_port(struct net_device *netdev, struct udp_tunnel_info *ti)
4207 {
4208         struct mlx5e_priv *priv = netdev_priv(netdev);
4209
4210         if (ti->type != UDP_TUNNEL_TYPE_VXLAN)
4211                 return;
4212
4213         if (!mlx5_vxlan_allowed(priv->mdev->vxlan))
4214                 return;
4215
4216         mlx5e_vxlan_queue_work(priv, be16_to_cpu(ti->port), 0);
4217 }
4218
4219 static netdev_features_t mlx5e_tunnel_features_check(struct mlx5e_priv *priv,
4220                                                      struct sk_buff *skb,
4221                                                      netdev_features_t features)
4222 {
4223         unsigned int offset = 0;
4224         struct udphdr *udph;
4225         u8 proto;
4226         u16 port;
4227
4228         switch (vlan_get_protocol(skb)) {
4229         case htons(ETH_P_IP):
4230                 proto = ip_hdr(skb)->protocol;
4231                 break;
4232         case htons(ETH_P_IPV6):
4233                 proto = ipv6_find_hdr(skb, &offset, -1, NULL, NULL);
4234                 break;
4235         default:
4236                 goto out;
4237         }
4238
4239         switch (proto) {
4240         case IPPROTO_GRE:
4241                 return features;
4242         case IPPROTO_IPIP:
4243         case IPPROTO_IPV6:
4244                 if (mlx5e_tunnel_proto_supported(priv->mdev, IPPROTO_IPIP))
4245                         return features;
4246                 break;
4247         case IPPROTO_UDP:
4248                 udph = udp_hdr(skb);
4249                 port = be16_to_cpu(udph->dest);
4250
4251                 /* Verify if UDP port is being offloaded by HW */
4252                 if (mlx5_vxlan_lookup_port(priv->mdev->vxlan, port))
4253                         return features;
4254
4255 #if IS_ENABLED(CONFIG_GENEVE)
4256                 /* Support Geneve offload for default UDP port */
4257                 if (port == GENEVE_UDP_PORT && mlx5_geneve_tx_allowed(priv->mdev))
4258                         return features;
4259 #endif
4260         }
4261
4262 out:
4263         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
4264         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
4265 }
4266
4267 netdev_features_t mlx5e_features_check(struct sk_buff *skb,
4268                                        struct net_device *netdev,
4269                                        netdev_features_t features)
4270 {
4271         struct mlx5e_priv *priv = netdev_priv(netdev);
4272
4273         features = vlan_features_check(skb, features);
4274         features = vxlan_features_check(skb, features);
4275
4276 #ifdef CONFIG_MLX5_EN_IPSEC
4277         if (mlx5e_ipsec_feature_check(skb, netdev, features))
4278                 return features;
4279 #endif
4280
4281         /* Validate if the tunneled packet is being offloaded by HW */
4282         if (skb->encapsulation &&
4283             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
4284                 return mlx5e_tunnel_features_check(priv, skb, features);
4285
4286         return features;
4287 }
4288
4289 static void mlx5e_tx_timeout_work(struct work_struct *work)
4290 {
4291         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
4292                                                tx_timeout_work);
4293         bool report_failed = false;
4294         int err;
4295         int i;
4296
4297         rtnl_lock();
4298         mutex_lock(&priv->state_lock);
4299
4300         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
4301                 goto unlock;
4302
4303         for (i = 0; i < priv->channels.num * priv->channels.params.num_tc; i++) {
4304                 struct netdev_queue *dev_queue =
4305                         netdev_get_tx_queue(priv->netdev, i);
4306                 struct mlx5e_txqsq *sq = priv->txq2sq[i];
4307
4308                 if (!netif_xmit_stopped(dev_queue))
4309                         continue;
4310
4311                 if (mlx5e_reporter_tx_timeout(sq))
4312                         report_failed = true;
4313         }
4314
4315         if (!report_failed)
4316                 goto unlock;
4317
4318         err = mlx5e_safe_reopen_channels(priv);
4319         if (err)
4320                 netdev_err(priv->netdev,
4321                            "mlx5e_safe_reopen_channels failed recovering from a tx_timeout, err(%d).\n",
4322                            err);
4323
4324 unlock:
4325         mutex_unlock(&priv->state_lock);
4326         rtnl_unlock();
4327 }
4328
4329 static void mlx5e_tx_timeout(struct net_device *dev, unsigned int txqueue)
4330 {
4331         struct mlx5e_priv *priv = netdev_priv(dev);
4332
4333         netdev_err(dev, "TX timeout detected\n");
4334         queue_work(priv->wq, &priv->tx_timeout_work);
4335 }
4336
4337 static int mlx5e_xdp_allowed(struct mlx5e_priv *priv, struct bpf_prog *prog)
4338 {
4339         struct net_device *netdev = priv->netdev;
4340         struct mlx5e_channels new_channels = {};
4341
4342         if (priv->channels.params.lro_en) {
4343                 netdev_warn(netdev, "can't set XDP while LRO is on, disable LRO first\n");
4344                 return -EINVAL;
4345         }
4346
4347         if (MLX5_IPSEC_DEV(priv->mdev)) {
4348                 netdev_warn(netdev, "can't set XDP with IPSec offload\n");
4349                 return -EINVAL;
4350         }
4351
4352         new_channels.params = priv->channels.params;
4353         new_channels.params.xdp_prog = prog;
4354
4355         /* No XSK params: AF_XDP can't be enabled yet at the point of setting
4356          * the XDP program.
4357          */
4358         if (!mlx5e_rx_is_linear_skb(&new_channels.params, NULL)) {
4359                 netdev_warn(netdev, "XDP is not allowed with MTU(%d) > %d\n",
4360                             new_channels.params.sw_mtu,
4361                             mlx5e_xdp_max_mtu(&new_channels.params, NULL));
4362                 return -EINVAL;
4363         }
4364
4365         return 0;
4366 }
4367
4368 static int mlx5e_xdp_set(struct net_device *netdev, struct bpf_prog *prog)
4369 {
4370         struct mlx5e_priv *priv = netdev_priv(netdev);
4371         struct bpf_prog *old_prog;
4372         bool reset, was_opened;
4373         int err = 0;
4374         int i;
4375
4376         mutex_lock(&priv->state_lock);
4377
4378         if (prog) {
4379                 err = mlx5e_xdp_allowed(priv, prog);
4380                 if (err)
4381                         goto unlock;
4382         }
4383
4384         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
4385         /* no need for full reset when exchanging programs */
4386         reset = (!priv->channels.params.xdp_prog || !prog);
4387
4388         if (was_opened && !reset)
4389                 /* num_channels is invariant here, so we can take the
4390                  * batched reference right upfront.
4391                  */
4392                 bpf_prog_add(prog, priv->channels.num);
4393
4394         if (was_opened && reset) {
4395                 struct mlx5e_channels new_channels = {};
4396
4397                 new_channels.params = priv->channels.params;
4398                 new_channels.params.xdp_prog = prog;
4399                 mlx5e_set_rq_type(priv->mdev, &new_channels.params);
4400                 old_prog = priv->channels.params.xdp_prog;
4401
4402                 err = mlx5e_safe_switch_channels(priv, &new_channels, NULL);
4403                 if (err)
4404                         goto unlock;
4405         } else {
4406                 /* exchange programs, extra prog reference we got from caller
4407                  * as long as we don't fail from this point onwards.
4408                  */
4409                 old_prog = xchg(&priv->channels.params.xdp_prog, prog);
4410         }
4411
4412         if (old_prog)
4413                 bpf_prog_put(old_prog);
4414
4415         if (!was_opened && reset) /* change RQ type according to priv->xdp_prog */
4416                 mlx5e_set_rq_type(priv->mdev, &priv->channels.params);
4417
4418         if (!was_opened || reset)
4419                 goto unlock;
4420
4421         /* exchanging programs w/o reset, we update ref counts on behalf
4422          * of the channels RQs here.
4423          */
4424         for (i = 0; i < priv->channels.num; i++) {
4425                 struct mlx5e_channel *c = priv->channels.c[i];
4426                 bool xsk_open = test_bit(MLX5E_CHANNEL_STATE_XSK, c->state);
4427
4428                 clear_bit(MLX5E_RQ_STATE_ENABLED, &c->rq.state);
4429                 if (xsk_open)
4430                         clear_bit(MLX5E_RQ_STATE_ENABLED, &c->xskrq.state);
4431                 napi_synchronize(&c->napi);
4432                 /* prevent mlx5e_poll_rx_cq from accessing rq->xdp_prog */
4433
4434                 old_prog = xchg(&c->rq.xdp_prog, prog);
4435                 if (old_prog)
4436                         bpf_prog_put(old_prog);
4437
4438                 if (xsk_open) {
4439                         old_prog = xchg(&c->xskrq.xdp_prog, prog);
4440                         if (old_prog)
4441                                 bpf_prog_put(old_prog);
4442                 }
4443
4444                 set_bit(MLX5E_RQ_STATE_ENABLED, &c->rq.state);
4445                 if (xsk_open)
4446                         set_bit(MLX5E_RQ_STATE_ENABLED, &c->xskrq.state);
4447                 /* napi_schedule in case we have missed anything */
4448                 napi_schedule(&c->napi);
4449         }
4450
4451 unlock:
4452         mutex_unlock(&priv->state_lock);
4453         return err;
4454 }
4455
4456 static u32 mlx5e_xdp_query(struct net_device *dev)
4457 {
4458         struct mlx5e_priv *priv = netdev_priv(dev);
4459         const struct bpf_prog *xdp_prog;
4460         u32 prog_id = 0;
4461
4462         mutex_lock(&priv->state_lock);
4463         xdp_prog = priv->channels.params.xdp_prog;
4464         if (xdp_prog)
4465                 prog_id = xdp_prog->aux->id;
4466         mutex_unlock(&priv->state_lock);
4467
4468         return prog_id;
4469 }
4470
4471 static int mlx5e_xdp(struct net_device *dev, struct netdev_bpf *xdp)
4472 {
4473         switch (xdp->command) {
4474         case XDP_SETUP_PROG:
4475                 return mlx5e_xdp_set(dev, xdp->prog);
4476         case XDP_QUERY_PROG:
4477                 xdp->prog_id = mlx5e_xdp_query(dev);
4478                 return 0;
4479         case XDP_SETUP_XSK_UMEM:
4480                 return mlx5e_xsk_setup_umem(dev, xdp->xsk.umem,
4481                                             xdp->xsk.queue_id);
4482         default:
4483                 return -EINVAL;
4484         }
4485 }
4486
4487 #ifdef CONFIG_MLX5_ESWITCH
4488 static int mlx5e_bridge_getlink(struct sk_buff *skb, u32 pid, u32 seq,
4489                                 struct net_device *dev, u32 filter_mask,
4490                                 int nlflags)
4491 {
4492         struct mlx5e_priv *priv = netdev_priv(dev);
4493         struct mlx5_core_dev *mdev = priv->mdev;
4494         u8 mode, setting;
4495         int err;
4496
4497         err = mlx5_eswitch_get_vepa(mdev->priv.eswitch, &setting);
4498         if (err)
4499                 return err;
4500         mode = setting ? BRIDGE_MODE_VEPA : BRIDGE_MODE_VEB;
4501         return ndo_dflt_bridge_getlink(skb, pid, seq, dev,
4502                                        mode,
4503                                        0, 0, nlflags, filter_mask, NULL);
4504 }
4505
4506 static int mlx5e_bridge_setlink(struct net_device *dev, struct nlmsghdr *nlh,
4507                                 u16 flags, struct netlink_ext_ack *extack)
4508 {
4509         struct mlx5e_priv *priv = netdev_priv(dev);
4510         struct mlx5_core_dev *mdev = priv->mdev;
4511         struct nlattr *attr, *br_spec;
4512         u16 mode = BRIDGE_MODE_UNDEF;
4513         u8 setting;
4514         int rem;
4515
4516         br_spec = nlmsg_find_attr(nlh, sizeof(struct ifinfomsg), IFLA_AF_SPEC);
4517         if (!br_spec)
4518                 return -EINVAL;
4519
4520         nla_for_each_nested(attr, br_spec, rem) {
4521                 if (nla_type(attr) != IFLA_BRIDGE_MODE)
4522                         continue;
4523
4524                 if (nla_len(attr) < sizeof(mode))
4525                         return -EINVAL;
4526
4527                 mode = nla_get_u16(attr);
4528                 if (mode > BRIDGE_MODE_VEPA)
4529                         return -EINVAL;
4530
4531                 break;
4532         }
4533
4534         if (mode == BRIDGE_MODE_UNDEF)
4535                 return -EINVAL;
4536
4537         setting = (mode == BRIDGE_MODE_VEPA) ?  1 : 0;
4538         return mlx5_eswitch_set_vepa(mdev->priv.eswitch, setting);
4539 }
4540 #endif
4541
4542 const struct net_device_ops mlx5e_netdev_ops = {
4543         .ndo_open                = mlx5e_open,
4544         .ndo_stop                = mlx5e_close,
4545         .ndo_start_xmit          = mlx5e_xmit,
4546         .ndo_setup_tc            = mlx5e_setup_tc,
4547         .ndo_select_queue        = mlx5e_select_queue,
4548         .ndo_get_stats64         = mlx5e_get_stats,
4549         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
4550         .ndo_set_mac_address     = mlx5e_set_mac,
4551         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
4552         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
4553         .ndo_set_features        = mlx5e_set_features,
4554         .ndo_fix_features        = mlx5e_fix_features,
4555         .ndo_change_mtu          = mlx5e_change_nic_mtu,
4556         .ndo_do_ioctl            = mlx5e_ioctl,
4557         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
4558         .ndo_udp_tunnel_add      = mlx5e_add_vxlan_port,
4559         .ndo_udp_tunnel_del      = mlx5e_del_vxlan_port,
4560         .ndo_features_check      = mlx5e_features_check,
4561         .ndo_tx_timeout          = mlx5e_tx_timeout,
4562         .ndo_bpf                 = mlx5e_xdp,
4563         .ndo_xdp_xmit            = mlx5e_xdp_xmit,
4564         .ndo_xsk_wakeup          = mlx5e_xsk_wakeup,
4565 #ifdef CONFIG_MLX5_EN_ARFS
4566         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
4567 #endif
4568 #ifdef CONFIG_MLX5_ESWITCH
4569         .ndo_bridge_setlink      = mlx5e_bridge_setlink,
4570         .ndo_bridge_getlink      = mlx5e_bridge_getlink,
4571
4572         /* SRIOV E-Switch NDOs */
4573         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
4574         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
4575         .ndo_set_vf_spoofchk     = mlx5e_set_vf_spoofchk,
4576         .ndo_set_vf_trust        = mlx5e_set_vf_trust,
4577         .ndo_set_vf_rate         = mlx5e_set_vf_rate,
4578         .ndo_get_vf_config       = mlx5e_get_vf_config,
4579         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
4580         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
4581 #endif
4582 };
4583
4584 static int mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
4585 {
4586         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
4587                 return -EOPNOTSUPP;
4588         if (!MLX5_CAP_GEN(mdev, eth_net_offloads) ||
4589             !MLX5_CAP_GEN(mdev, nic_flow_table) ||
4590             !MLX5_CAP_ETH(mdev, csum_cap) ||
4591             !MLX5_CAP_ETH(mdev, max_lso_cap) ||
4592             !MLX5_CAP_ETH(mdev, vlan_cap) ||
4593             !MLX5_CAP_ETH(mdev, rss_ind_tbl_cap) ||
4594             MLX5_CAP_FLOWTABLE(mdev,
4595                                flow_table_properties_nic_receive.max_ft_level)
4596                                < 3) {
4597                 mlx5_core_warn(mdev,
4598                                "Not creating net device, some required device capabilities are missing\n");
4599                 return -EOPNOTSUPP;
4600         }
4601         if (!MLX5_CAP_ETH(mdev, self_lb_en_modifiable))
4602                 mlx5_core_warn(mdev, "Self loop back prevention is not supported\n");
4603         if (!MLX5_CAP_GEN(mdev, cq_moderation))
4604                 mlx5_core_warn(mdev, "CQ moderation is not supported\n");
4605
4606         return 0;
4607 }
4608
4609 void mlx5e_build_default_indir_rqt(u32 *indirection_rqt, int len,
4610                                    int num_channels)
4611 {
4612         int i;
4613
4614         for (i = 0; i < len; i++)
4615                 indirection_rqt[i] = i % num_channels;
4616 }
4617
4618 static bool slow_pci_heuristic(struct mlx5_core_dev *mdev)
4619 {
4620         u32 link_speed = 0;
4621         u32 pci_bw = 0;
4622
4623         mlx5e_port_max_linkspeed(mdev, &link_speed);
4624         pci_bw = pcie_bandwidth_available(mdev->pdev, NULL, NULL, NULL);
4625         mlx5_core_dbg_once(mdev, "Max link speed = %d, PCI BW = %d\n",
4626                            link_speed, pci_bw);
4627
4628 #define MLX5E_SLOW_PCI_RATIO (2)
4629
4630         return link_speed && pci_bw &&
4631                 link_speed > MLX5E_SLOW_PCI_RATIO * pci_bw;
4632 }
4633
4634 static struct dim_cq_moder mlx5e_get_def_tx_moderation(u8 cq_period_mode)
4635 {
4636         struct dim_cq_moder moder;
4637
4638         moder.cq_period_mode = cq_period_mode;
4639         moder.pkts = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
4640         moder.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
4641         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4642                 moder.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE;
4643
4644         return moder;
4645 }
4646
4647 static struct dim_cq_moder mlx5e_get_def_rx_moderation(u8 cq_period_mode)
4648 {
4649         struct dim_cq_moder moder;
4650
4651         moder.cq_period_mode = cq_period_mode;
4652         moder.pkts = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
4653         moder.usec = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
4654         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4655                 moder.usec = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE;
4656
4657         return moder;
4658 }
4659
4660 static u8 mlx5_to_net_dim_cq_period_mode(u8 cq_period_mode)
4661 {
4662         return cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE ?
4663                 DIM_CQ_PERIOD_MODE_START_FROM_CQE :
4664                 DIM_CQ_PERIOD_MODE_START_FROM_EQE;
4665 }
4666
4667 void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4668 {
4669         if (params->tx_dim_enabled) {
4670                 u8 dim_period_mode = mlx5_to_net_dim_cq_period_mode(cq_period_mode);
4671
4672                 params->tx_cq_moderation = net_dim_get_def_tx_moderation(dim_period_mode);
4673         } else {
4674                 params->tx_cq_moderation = mlx5e_get_def_tx_moderation(cq_period_mode);
4675         }
4676
4677         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_TX_CQE_BASED_MODER,
4678                         params->tx_cq_moderation.cq_period_mode ==
4679                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4680 }
4681
4682 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4683 {
4684         if (params->rx_dim_enabled) {
4685                 u8 dim_period_mode = mlx5_to_net_dim_cq_period_mode(cq_period_mode);
4686
4687                 params->rx_cq_moderation = net_dim_get_def_rx_moderation(dim_period_mode);
4688         } else {
4689                 params->rx_cq_moderation = mlx5e_get_def_rx_moderation(cq_period_mode);
4690         }
4691
4692         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_BASED_MODER,
4693                         params->rx_cq_moderation.cq_period_mode ==
4694                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4695 }
4696
4697 static u32 mlx5e_choose_lro_timeout(struct mlx5_core_dev *mdev, u32 wanted_timeout)
4698 {
4699         int i;
4700
4701         /* The supported periods are organized in ascending order */
4702         for (i = 0; i < MLX5E_LRO_TIMEOUT_ARR_SIZE - 1; i++)
4703                 if (MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]) >= wanted_timeout)
4704                         break;
4705
4706         return MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]);
4707 }
4708
4709 void mlx5e_build_rq_params(struct mlx5_core_dev *mdev,
4710                            struct mlx5e_params *params)
4711 {
4712         /* Prefer Striding RQ, unless any of the following holds:
4713          * - Striding RQ configuration is not possible/supported.
4714          * - Slow PCI heuristic.
4715          * - Legacy RQ would use linear SKB while Striding RQ would use non-linear.
4716          *
4717          * No XSK params: checking the availability of striding RQ in general.
4718          */
4719         if (!slow_pci_heuristic(mdev) &&
4720             mlx5e_striding_rq_possible(mdev, params) &&
4721             (mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL) ||
4722              !mlx5e_rx_is_linear_skb(params, NULL)))
4723                 MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ, true);
4724         mlx5e_set_rq_type(mdev, params);
4725         mlx5e_init_rq_type_params(mdev, params);
4726 }
4727
4728 void mlx5e_build_rss_params(struct mlx5e_rss_params *rss_params,
4729                             u16 num_channels)
4730 {
4731         enum mlx5e_traffic_types tt;
4732
4733         rss_params->hfunc = ETH_RSS_HASH_TOP;
4734         netdev_rss_key_fill(rss_params->toeplitz_hash_key,
4735                             sizeof(rss_params->toeplitz_hash_key));
4736         mlx5e_build_default_indir_rqt(rss_params->indirection_rqt,
4737                                       MLX5E_INDIR_RQT_SIZE, num_channels);
4738         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++)
4739                 rss_params->rx_hash_fields[tt] =
4740                         tirc_default_config[tt].rx_hash_fields;
4741 }
4742
4743 void mlx5e_build_nic_params(struct mlx5e_priv *priv,
4744                             struct mlx5e_xsk *xsk,
4745                             struct mlx5e_rss_params *rss_params,
4746                             struct mlx5e_params *params,
4747                             u16 mtu)
4748 {
4749         struct mlx5_core_dev *mdev = priv->mdev;
4750         u8 rx_cq_period_mode;
4751
4752         params->sw_mtu = mtu;
4753         params->hard_mtu = MLX5E_ETH_HARD_MTU;
4754         params->num_channels = min_t(unsigned int, MLX5E_MAX_NUM_CHANNELS / 2,
4755                                      priv->max_nch);
4756         params->num_tc       = 1;
4757
4758         /* SQ */
4759         params->log_sq_size = is_kdump_kernel() ?
4760                 MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE :
4761                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
4762
4763         /* XDP SQ */
4764         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_XDP_TX_MPWQE,
4765                         MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe));
4766
4767         /* set CQE compression */
4768         params->rx_cqe_compress_def = false;
4769         if (MLX5_CAP_GEN(mdev, cqe_compression) &&
4770             MLX5_CAP_GEN(mdev, vport_group_manager))
4771                 params->rx_cqe_compress_def = slow_pci_heuristic(mdev);
4772
4773         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS, params->rx_cqe_compress_def);
4774         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_NO_CSUM_COMPLETE, false);
4775
4776         /* RQ */
4777         mlx5e_build_rq_params(mdev, params);
4778
4779         /* HW LRO */
4780         if (MLX5_CAP_ETH(mdev, lro_cap) &&
4781             params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
4782                 /* No XSK params: checking the availability of striding RQ in general. */
4783                 if (!mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL))
4784                         params->lro_en = !slow_pci_heuristic(mdev);
4785         }
4786         params->lro_timeout = mlx5e_choose_lro_timeout(mdev, MLX5E_DEFAULT_LRO_TIMEOUT);
4787
4788         /* CQ moderation params */
4789         rx_cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
4790                         MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
4791                         MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
4792         params->rx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
4793         params->tx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
4794         mlx5e_set_rx_cq_mode_params(params, rx_cq_period_mode);
4795         mlx5e_set_tx_cq_mode_params(params, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
4796
4797         /* TX inline */
4798         mlx5_query_min_inline(mdev, &params->tx_min_inline_mode);
4799
4800         /* RSS */
4801         mlx5e_build_rss_params(rss_params, params->num_channels);
4802         params->tunneled_offload_en =
4803                 mlx5e_tunnel_inner_ft_supported(mdev);
4804
4805         /* AF_XDP */
4806         params->xsk = xsk;
4807 }
4808
4809 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
4810 {
4811         struct mlx5e_priv *priv = netdev_priv(netdev);
4812
4813         mlx5_query_mac_address(priv->mdev, netdev->dev_addr);
4814         if (is_zero_ether_addr(netdev->dev_addr) &&
4815             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
4816                 eth_hw_addr_random(netdev);
4817                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
4818         }
4819 }
4820
4821 static void mlx5e_build_nic_netdev(struct net_device *netdev)
4822 {
4823         struct mlx5e_priv *priv = netdev_priv(netdev);
4824         struct mlx5_core_dev *mdev = priv->mdev;
4825         bool fcs_supported;
4826         bool fcs_enabled;
4827
4828         SET_NETDEV_DEV(netdev, mdev->device);
4829
4830         netdev->netdev_ops = &mlx5e_netdev_ops;
4831
4832 #ifdef CONFIG_MLX5_CORE_EN_DCB
4833         if (MLX5_CAP_GEN(mdev, vport_group_manager) && MLX5_CAP_GEN(mdev, qos))
4834                 netdev->dcbnl_ops = &mlx5e_dcbnl_ops;
4835 #endif
4836
4837         netdev->watchdog_timeo    = 15 * HZ;
4838
4839         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
4840
4841         netdev->vlan_features    |= NETIF_F_SG;
4842         netdev->vlan_features    |= NETIF_F_HW_CSUM;
4843         netdev->vlan_features    |= NETIF_F_GRO;
4844         netdev->vlan_features    |= NETIF_F_TSO;
4845         netdev->vlan_features    |= NETIF_F_TSO6;
4846         netdev->vlan_features    |= NETIF_F_RXCSUM;
4847         netdev->vlan_features    |= NETIF_F_RXHASH;
4848
4849         netdev->mpls_features    |= NETIF_F_SG;
4850         netdev->mpls_features    |= NETIF_F_HW_CSUM;
4851         netdev->mpls_features    |= NETIF_F_TSO;
4852         netdev->mpls_features    |= NETIF_F_TSO6;
4853
4854         netdev->hw_enc_features  |= NETIF_F_HW_VLAN_CTAG_TX;
4855         netdev->hw_enc_features  |= NETIF_F_HW_VLAN_CTAG_RX;
4856
4857         if (!!MLX5_CAP_ETH(mdev, lro_cap) &&
4858             mlx5e_check_fragmented_striding_rq_cap(mdev))
4859                 netdev->vlan_features    |= NETIF_F_LRO;
4860
4861         netdev->hw_features       = netdev->vlan_features;
4862         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
4863         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
4864         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
4865         netdev->hw_features      |= NETIF_F_HW_VLAN_STAG_TX;
4866
4867         if (mlx5_vxlan_allowed(mdev->vxlan) || mlx5_geneve_tx_allowed(mdev) ||
4868             mlx5e_any_tunnel_proto_supported(mdev)) {
4869                 netdev->hw_enc_features |= NETIF_F_HW_CSUM;
4870                 netdev->hw_enc_features |= NETIF_F_TSO;
4871                 netdev->hw_enc_features |= NETIF_F_TSO6;
4872                 netdev->hw_enc_features |= NETIF_F_GSO_PARTIAL;
4873         }
4874
4875         if (mlx5_vxlan_allowed(mdev->vxlan) || mlx5_geneve_tx_allowed(mdev)) {
4876                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL |
4877                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4878                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL |
4879                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4880                 netdev->gso_partial_features = NETIF_F_GSO_UDP_TUNNEL_CSUM;
4881                 netdev->vlan_features |= NETIF_F_GSO_UDP_TUNNEL |
4882                                          NETIF_F_GSO_UDP_TUNNEL_CSUM;
4883         }
4884
4885         if (mlx5e_tunnel_proto_supported(mdev, IPPROTO_GRE)) {
4886                 netdev->hw_features     |= NETIF_F_GSO_GRE |
4887                                            NETIF_F_GSO_GRE_CSUM;
4888                 netdev->hw_enc_features |= NETIF_F_GSO_GRE |
4889                                            NETIF_F_GSO_GRE_CSUM;
4890                 netdev->gso_partial_features |= NETIF_F_GSO_GRE |
4891                                                 NETIF_F_GSO_GRE_CSUM;
4892         }
4893
4894         if (mlx5e_tunnel_proto_supported(mdev, IPPROTO_IPIP)) {
4895                 netdev->hw_features |= NETIF_F_GSO_IPXIP4 |
4896                                        NETIF_F_GSO_IPXIP6;
4897                 netdev->hw_enc_features |= NETIF_F_GSO_IPXIP4 |
4898                                            NETIF_F_GSO_IPXIP6;
4899                 netdev->gso_partial_features |= NETIF_F_GSO_IPXIP4 |
4900                                                 NETIF_F_GSO_IPXIP6;
4901         }
4902
4903         netdev->hw_features                      |= NETIF_F_GSO_PARTIAL;
4904         netdev->gso_partial_features             |= NETIF_F_GSO_UDP_L4;
4905         netdev->hw_features                      |= NETIF_F_GSO_UDP_L4;
4906         netdev->features                         |= NETIF_F_GSO_UDP_L4;
4907
4908         mlx5_query_port_fcs(mdev, &fcs_supported, &fcs_enabled);
4909
4910         if (fcs_supported)
4911                 netdev->hw_features |= NETIF_F_RXALL;
4912
4913         if (MLX5_CAP_ETH(mdev, scatter_fcs))
4914                 netdev->hw_features |= NETIF_F_RXFCS;
4915
4916         netdev->features          = netdev->hw_features;
4917         if (!priv->channels.params.lro_en)
4918                 netdev->features  &= ~NETIF_F_LRO;
4919
4920         if (fcs_enabled)
4921                 netdev->features  &= ~NETIF_F_RXALL;
4922
4923         if (!priv->channels.params.scatter_fcs_en)
4924                 netdev->features  &= ~NETIF_F_RXFCS;
4925
4926         /* prefere CQE compression over rxhash */
4927         if (MLX5E_GET_PFLAG(&priv->channels.params, MLX5E_PFLAG_RX_CQE_COMPRESS))
4928                 netdev->features &= ~NETIF_F_RXHASH;
4929
4930 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
4931         if (FT_CAP(flow_modify_en) &&
4932             FT_CAP(modify_root) &&
4933             FT_CAP(identified_miss_table_mode) &&
4934             FT_CAP(flow_table_modify)) {
4935 #ifdef CONFIG_MLX5_ESWITCH
4936                 netdev->hw_features      |= NETIF_F_HW_TC;
4937 #endif
4938 #ifdef CONFIG_MLX5_EN_ARFS
4939                 netdev->hw_features      |= NETIF_F_NTUPLE;
4940 #endif
4941         }
4942
4943         netdev->features         |= NETIF_F_HIGHDMA;
4944         netdev->features         |= NETIF_F_HW_VLAN_STAG_FILTER;
4945
4946         netdev->priv_flags       |= IFF_UNICAST_FLT;
4947
4948         mlx5e_set_netdev_dev_addr(netdev);
4949         mlx5e_ipsec_build_netdev(priv);
4950         mlx5e_tls_build_netdev(priv);
4951 }
4952
4953 void mlx5e_create_q_counters(struct mlx5e_priv *priv)
4954 {
4955         struct mlx5_core_dev *mdev = priv->mdev;
4956         int err;
4957
4958         err = mlx5_core_alloc_q_counter(mdev, &priv->q_counter);
4959         if (err) {
4960                 mlx5_core_warn(mdev, "alloc queue counter failed, %d\n", err);
4961                 priv->q_counter = 0;
4962         }
4963
4964         err = mlx5_core_alloc_q_counter(mdev, &priv->drop_rq_q_counter);
4965         if (err) {
4966                 mlx5_core_warn(mdev, "alloc drop RQ counter failed, %d\n", err);
4967                 priv->drop_rq_q_counter = 0;
4968         }
4969 }
4970
4971 void mlx5e_destroy_q_counters(struct mlx5e_priv *priv)
4972 {
4973         if (priv->q_counter)
4974                 mlx5_core_dealloc_q_counter(priv->mdev, priv->q_counter);
4975
4976         if (priv->drop_rq_q_counter)
4977                 mlx5_core_dealloc_q_counter(priv->mdev, priv->drop_rq_q_counter);
4978 }
4979
4980 static int mlx5e_nic_init(struct mlx5_core_dev *mdev,
4981                           struct net_device *netdev,
4982                           const struct mlx5e_profile *profile,
4983                           void *ppriv)
4984 {
4985         struct mlx5e_priv *priv = netdev_priv(netdev);
4986         struct mlx5e_rss_params *rss = &priv->rss_params;
4987         int err;
4988
4989         err = mlx5e_netdev_init(netdev, priv, mdev, profile, ppriv);
4990         if (err)
4991                 return err;
4992
4993         mlx5e_build_nic_params(priv, &priv->xsk, rss, &priv->channels.params,
4994                                netdev->mtu);
4995
4996         mlx5e_timestamp_init(priv);
4997
4998         err = mlx5e_ipsec_init(priv);
4999         if (err)
5000                 mlx5_core_err(mdev, "IPSec initialization failed, %d\n", err);
5001         err = mlx5e_tls_init(priv);
5002         if (err)
5003                 mlx5_core_err(mdev, "TLS initialization failed, %d\n", err);
5004         mlx5e_build_nic_netdev(netdev);
5005         mlx5e_health_create_reporters(priv);
5006
5007         return 0;
5008 }
5009
5010 static void mlx5e_nic_cleanup(struct mlx5e_priv *priv)
5011 {
5012         mlx5e_health_destroy_reporters(priv);
5013         mlx5e_tls_cleanup(priv);
5014         mlx5e_ipsec_cleanup(priv);
5015         mlx5e_netdev_cleanup(priv->netdev, priv);
5016 }
5017
5018 static int mlx5e_init_nic_rx(struct mlx5e_priv *priv)
5019 {
5020         struct mlx5_core_dev *mdev = priv->mdev;
5021         int err;
5022
5023         mlx5e_create_q_counters(priv);
5024
5025         err = mlx5e_open_drop_rq(priv, &priv->drop_rq);
5026         if (err) {
5027                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
5028                 goto err_destroy_q_counters;
5029         }
5030
5031         err = mlx5e_create_indirect_rqt(priv);
5032         if (err)
5033                 goto err_close_drop_rq;
5034
5035         err = mlx5e_create_direct_rqts(priv, priv->direct_tir);
5036         if (err)
5037                 goto err_destroy_indirect_rqts;
5038
5039         err = mlx5e_create_indirect_tirs(priv, true);
5040         if (err)
5041                 goto err_destroy_direct_rqts;
5042
5043         err = mlx5e_create_direct_tirs(priv, priv->direct_tir);
5044         if (err)
5045                 goto err_destroy_indirect_tirs;
5046
5047         err = mlx5e_create_direct_rqts(priv, priv->xsk_tir);
5048         if (unlikely(err))
5049                 goto err_destroy_direct_tirs;
5050
5051         err = mlx5e_create_direct_tirs(priv, priv->xsk_tir);
5052         if (unlikely(err))
5053                 goto err_destroy_xsk_rqts;
5054
5055         err = mlx5e_create_flow_steering(priv);
5056         if (err) {
5057                 mlx5_core_warn(mdev, "create flow steering failed, %d\n", err);
5058                 goto err_destroy_xsk_tirs;
5059         }
5060
5061         err = mlx5e_tc_nic_init(priv);
5062         if (err)
5063                 goto err_destroy_flow_steering;
5064
5065         return 0;
5066
5067 err_destroy_flow_steering:
5068         mlx5e_destroy_flow_steering(priv);
5069 err_destroy_xsk_tirs:
5070         mlx5e_destroy_direct_tirs(priv, priv->xsk_tir);
5071 err_destroy_xsk_rqts:
5072         mlx5e_destroy_direct_rqts(priv, priv->xsk_tir);
5073 err_destroy_direct_tirs:
5074         mlx5e_destroy_direct_tirs(priv, priv->direct_tir);
5075 err_destroy_indirect_tirs:
5076         mlx5e_destroy_indirect_tirs(priv, true);
5077 err_destroy_direct_rqts:
5078         mlx5e_destroy_direct_rqts(priv, priv->direct_tir);
5079 err_destroy_indirect_rqts:
5080         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
5081 err_close_drop_rq:
5082         mlx5e_close_drop_rq(&priv->drop_rq);
5083 err_destroy_q_counters:
5084         mlx5e_destroy_q_counters(priv);
5085         return err;
5086 }
5087
5088 static void mlx5e_cleanup_nic_rx(struct mlx5e_priv *priv)
5089 {
5090         mlx5e_tc_nic_cleanup(priv);
5091         mlx5e_destroy_flow_steering(priv);
5092         mlx5e_destroy_direct_tirs(priv, priv->xsk_tir);
5093         mlx5e_destroy_direct_rqts(priv, priv->xsk_tir);
5094         mlx5e_destroy_direct_tirs(priv, priv->direct_tir);
5095         mlx5e_destroy_indirect_tirs(priv, true);
5096         mlx5e_destroy_direct_rqts(priv, priv->direct_tir);
5097         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
5098         mlx5e_close_drop_rq(&priv->drop_rq);
5099         mlx5e_destroy_q_counters(priv);
5100 }
5101
5102 static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
5103 {
5104         int err;
5105
5106         err = mlx5e_create_tises(priv);
5107         if (err) {
5108                 mlx5_core_warn(priv->mdev, "create tises failed, %d\n", err);
5109                 return err;
5110         }
5111
5112 #ifdef CONFIG_MLX5_CORE_EN_DCB
5113         mlx5e_dcbnl_initialize(priv);
5114 #endif
5115         return 0;
5116 }
5117
5118 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
5119 {
5120         struct net_device *netdev = priv->netdev;
5121         struct mlx5_core_dev *mdev = priv->mdev;
5122
5123         mlx5e_init_l2_addr(priv);
5124
5125         /* Marking the link as currently not needed by the Driver */
5126         if (!netif_running(netdev))
5127                 mlx5_set_port_admin_status(mdev, MLX5_PORT_DOWN);
5128
5129         mlx5e_set_netdev_mtu_boundaries(priv);
5130         mlx5e_set_dev_port_mtu(priv);
5131
5132         mlx5_lag_add(mdev, netdev);
5133
5134         mlx5e_enable_async_events(priv);
5135         if (mlx5e_monitor_counter_supported(priv))
5136                 mlx5e_monitor_counter_init(priv);
5137
5138         mlx5e_hv_vhca_stats_create(priv);
5139         if (netdev->reg_state != NETREG_REGISTERED)
5140                 return;
5141 #ifdef CONFIG_MLX5_CORE_EN_DCB
5142         mlx5e_dcbnl_init_app(priv);
5143 #endif
5144
5145         queue_work(priv->wq, &priv->set_rx_mode_work);
5146
5147         rtnl_lock();
5148         if (netif_running(netdev))
5149                 mlx5e_open(netdev);
5150         netif_device_attach(netdev);
5151         rtnl_unlock();
5152 }
5153
5154 static void mlx5e_nic_disable(struct mlx5e_priv *priv)
5155 {
5156         struct net_device *netdev = priv->netdev;
5157         struct mlx5_core_dev *mdev = priv->mdev;
5158
5159 #ifdef CONFIG_MLX5_CORE_EN_DCB
5160         if (priv->netdev->reg_state == NETREG_REGISTERED)
5161                 mlx5e_dcbnl_delete_app(priv);
5162 #endif
5163
5164         rtnl_lock();
5165         if (netif_running(priv->netdev))
5166                 mlx5e_close(priv->netdev);
5167         netif_device_detach(priv->netdev);
5168         rtnl_unlock();
5169
5170         queue_work(priv->wq, &priv->set_rx_mode_work);
5171
5172         mlx5e_hv_vhca_stats_destroy(priv);
5173         if (mlx5e_monitor_counter_supported(priv))
5174                 mlx5e_monitor_counter_cleanup(priv);
5175
5176         mlx5e_disable_async_events(priv);
5177         mlx5_lag_remove(mdev, netdev);
5178 }
5179
5180 int mlx5e_update_nic_rx(struct mlx5e_priv *priv)
5181 {
5182         return mlx5e_refresh_tirs(priv, false);
5183 }
5184
5185 static const struct mlx5e_profile mlx5e_nic_profile = {
5186         .init              = mlx5e_nic_init,
5187         .cleanup           = mlx5e_nic_cleanup,
5188         .init_rx           = mlx5e_init_nic_rx,
5189         .cleanup_rx        = mlx5e_cleanup_nic_rx,
5190         .init_tx           = mlx5e_init_nic_tx,
5191         .cleanup_tx        = mlx5e_cleanup_nic_tx,
5192         .enable            = mlx5e_nic_enable,
5193         .disable           = mlx5e_nic_disable,
5194         .update_rx         = mlx5e_update_nic_rx,
5195         .update_stats      = mlx5e_update_ndo_stats,
5196         .update_carrier    = mlx5e_update_carrier,
5197         .rx_handlers.handle_rx_cqe       = mlx5e_handle_rx_cqe,
5198         .rx_handlers.handle_rx_cqe_mpwqe = mlx5e_handle_rx_cqe_mpwrq,
5199         .max_tc            = MLX5E_MAX_NUM_TC,
5200         .rq_groups         = MLX5E_NUM_RQ_GROUPS(XSK),
5201         .stats_grps        = mlx5e_nic_stats_grps,
5202         .stats_grps_num    = mlx5e_nic_stats_grps_num,
5203 };
5204
5205 /* mlx5e generic netdev management API (move to en_common.c) */
5206
5207 /* mlx5e_netdev_init/cleanup must be called from profile->init/cleanup callbacks */
5208 int mlx5e_netdev_init(struct net_device *netdev,
5209                       struct mlx5e_priv *priv,
5210                       struct mlx5_core_dev *mdev,
5211                       const struct mlx5e_profile *profile,
5212                       void *ppriv)
5213 {
5214         /* priv init */
5215         priv->mdev        = mdev;
5216         priv->netdev      = netdev;
5217         priv->profile     = profile;
5218         priv->ppriv       = ppriv;
5219         priv->msglevel    = MLX5E_MSG_LEVEL;
5220         priv->max_nch     = netdev->num_rx_queues / max_t(u8, profile->rq_groups, 1);
5221         priv->max_opened_tc = 1;
5222
5223         mutex_init(&priv->state_lock);
5224         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
5225         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
5226         INIT_WORK(&priv->tx_timeout_work, mlx5e_tx_timeout_work);
5227         INIT_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
5228
5229         priv->wq = create_singlethread_workqueue("mlx5e");
5230         if (!priv->wq)
5231                 return -ENOMEM;
5232
5233         /* netdev init */
5234         netif_carrier_off(netdev);
5235
5236 #ifdef CONFIG_MLX5_EN_ARFS
5237         netdev->rx_cpu_rmap =  mlx5_eq_table_get_rmap(mdev);
5238 #endif
5239
5240         return 0;
5241 }
5242
5243 void mlx5e_netdev_cleanup(struct net_device *netdev, struct mlx5e_priv *priv)
5244 {
5245         destroy_workqueue(priv->wq);
5246 }
5247
5248 struct net_device *mlx5e_create_netdev(struct mlx5_core_dev *mdev,
5249                                        const struct mlx5e_profile *profile,
5250                                        int nch,
5251                                        void *ppriv)
5252 {
5253         struct net_device *netdev;
5254         int err;
5255
5256         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv),
5257                                     nch * profile->max_tc,
5258                                     nch * profile->rq_groups);
5259         if (!netdev) {
5260                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
5261                 return NULL;
5262         }
5263
5264         err = profile->init(mdev, netdev, profile, ppriv);
5265         if (err) {
5266                 mlx5_core_err(mdev, "failed to init mlx5e profile %d\n", err);
5267                 goto err_free_netdev;
5268         }
5269
5270         return netdev;
5271
5272 err_free_netdev:
5273         free_netdev(netdev);
5274
5275         return NULL;
5276 }
5277
5278 int mlx5e_attach_netdev(struct mlx5e_priv *priv)
5279 {
5280         const struct mlx5e_profile *profile;
5281         int max_nch;
5282         int err;
5283
5284         profile = priv->profile;
5285         clear_bit(MLX5E_STATE_DESTROYING, &priv->state);
5286
5287         /* max number of channels may have changed */
5288         max_nch = mlx5e_get_max_num_channels(priv->mdev);
5289         if (priv->channels.params.num_channels > max_nch) {
5290                 mlx5_core_warn(priv->mdev, "MLX5E: Reducing number of channels to %d\n", max_nch);
5291                 priv->channels.params.num_channels = max_nch;
5292                 mlx5e_build_default_indir_rqt(priv->rss_params.indirection_rqt,
5293                                               MLX5E_INDIR_RQT_SIZE, max_nch);
5294         }
5295
5296         err = profile->init_tx(priv);
5297         if (err)
5298                 goto out;
5299
5300         err = profile->init_rx(priv);
5301         if (err)
5302                 goto err_cleanup_tx;
5303
5304         if (profile->enable)
5305                 profile->enable(priv);
5306
5307         return 0;
5308
5309 err_cleanup_tx:
5310         profile->cleanup_tx(priv);
5311
5312 out:
5313         return err;
5314 }
5315
5316 void mlx5e_detach_netdev(struct mlx5e_priv *priv)
5317 {
5318         const struct mlx5e_profile *profile = priv->profile;
5319
5320         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
5321
5322         if (profile->disable)
5323                 profile->disable(priv);
5324         flush_workqueue(priv->wq);
5325
5326         profile->cleanup_rx(priv);
5327         profile->cleanup_tx(priv);
5328         cancel_work_sync(&priv->update_stats_work);
5329 }
5330
5331 void mlx5e_destroy_netdev(struct mlx5e_priv *priv)
5332 {
5333         const struct mlx5e_profile *profile = priv->profile;
5334         struct net_device *netdev = priv->netdev;
5335
5336         if (profile->cleanup)
5337                 profile->cleanup(priv);
5338         free_netdev(netdev);
5339 }
5340
5341 /* mlx5e_attach and mlx5e_detach scope should be only creating/destroying
5342  * hardware contexts and to connect it to the current netdev.
5343  */
5344 static int mlx5e_attach(struct mlx5_core_dev *mdev, void *vpriv)
5345 {
5346         struct mlx5e_priv *priv = vpriv;
5347         struct net_device *netdev = priv->netdev;
5348         int err;
5349
5350         if (netif_device_present(netdev))
5351                 return 0;
5352
5353         err = mlx5e_create_mdev_resources(mdev);
5354         if (err)
5355                 return err;
5356
5357         err = mlx5e_attach_netdev(priv);
5358         if (err) {
5359                 mlx5e_destroy_mdev_resources(mdev);
5360                 return err;
5361         }
5362
5363         return 0;
5364 }
5365
5366 static void mlx5e_detach(struct mlx5_core_dev *mdev, void *vpriv)
5367 {
5368         struct mlx5e_priv *priv = vpriv;
5369         struct net_device *netdev = priv->netdev;
5370
5371 #ifdef CONFIG_MLX5_ESWITCH
5372         if (MLX5_ESWITCH_MANAGER(mdev) && vpriv == mdev)
5373                 return;
5374 #endif
5375
5376         if (!netif_device_present(netdev))
5377                 return;
5378
5379         mlx5e_detach_netdev(priv);
5380         mlx5e_destroy_mdev_resources(mdev);
5381 }
5382
5383 static void *mlx5e_add(struct mlx5_core_dev *mdev)
5384 {
5385         struct net_device *netdev;
5386         void *priv;
5387         int err;
5388         int nch;
5389
5390         err = mlx5e_check_required_hca_cap(mdev);
5391         if (err)
5392                 return NULL;
5393
5394 #ifdef CONFIG_MLX5_ESWITCH
5395         if (MLX5_ESWITCH_MANAGER(mdev) &&
5396             mlx5_eswitch_mode(mdev->priv.eswitch) == MLX5_ESWITCH_OFFLOADS) {
5397                 mlx5e_rep_register_vport_reps(mdev);
5398                 return mdev;
5399         }
5400 #endif
5401
5402         nch = mlx5e_get_max_num_channels(mdev);
5403         netdev = mlx5e_create_netdev(mdev, &mlx5e_nic_profile, nch, NULL);
5404         if (!netdev) {
5405                 mlx5_core_err(mdev, "mlx5e_create_netdev failed\n");
5406                 return NULL;
5407         }
5408
5409         dev_net_set(netdev, mlx5_core_net(mdev));
5410         priv = netdev_priv(netdev);
5411
5412         err = mlx5e_attach(mdev, priv);
5413         if (err) {
5414                 mlx5_core_err(mdev, "mlx5e_attach failed, %d\n", err);
5415                 goto err_destroy_netdev;
5416         }
5417
5418         err = register_netdev(netdev);
5419         if (err) {
5420                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
5421                 goto err_detach;
5422         }
5423
5424 #ifdef CONFIG_MLX5_CORE_EN_DCB
5425         mlx5e_dcbnl_init_app(priv);
5426 #endif
5427         return priv;
5428
5429 err_detach:
5430         mlx5e_detach(mdev, priv);
5431 err_destroy_netdev:
5432         mlx5e_destroy_netdev(priv);
5433         return NULL;
5434 }
5435
5436 static void mlx5e_remove(struct mlx5_core_dev *mdev, void *vpriv)
5437 {
5438         struct mlx5e_priv *priv;
5439
5440 #ifdef CONFIG_MLX5_ESWITCH
5441         if (MLX5_ESWITCH_MANAGER(mdev) && vpriv == mdev) {
5442                 mlx5e_rep_unregister_vport_reps(mdev);
5443                 return;
5444         }
5445 #endif
5446         priv = vpriv;
5447 #ifdef CONFIG_MLX5_CORE_EN_DCB
5448         mlx5e_dcbnl_delete_app(priv);
5449 #endif
5450         unregister_netdev(priv->netdev);
5451         mlx5e_detach(mdev, vpriv);
5452         mlx5e_destroy_netdev(priv);
5453 }
5454
5455 static struct mlx5_interface mlx5e_interface = {
5456         .add       = mlx5e_add,
5457         .remove    = mlx5e_remove,
5458         .attach    = mlx5e_attach,
5459         .detach    = mlx5e_detach,
5460         .protocol  = MLX5_INTERFACE_PROTOCOL_ETH,
5461 };
5462
5463 void mlx5e_init(void)
5464 {
5465         mlx5e_ipsec_build_inverse_table();
5466         mlx5e_build_ptys2ethtool_map();
5467         mlx5_register_interface(&mlx5e_interface);
5468 }
5469
5470 void mlx5e_cleanup(void)
5471 {
5472         mlx5_unregister_interface(&mlx5e_interface);
5473 }