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_drv.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/en_accel.h"
  49 #include "en_accel/tls.h"
  50 #include "accel/ipsec.h"
  51 #include "accel/tls.h"
  52 #include "lib/vxlan.h"
  53 #include "lib/clock.h"
  54 #include "en/port.h"
  55 #include "en/xdp.h"
  56 #include "lib/eq.h"
  57 #include "en/monitor_stats.h"
  58 #include "en/health.h"
  59 #include "en/params.h"
  60 #include "en/xsk/pool.h"
  61 #include "en/xsk/setup.h"
  62 #include "en/xsk/rx.h"
  63 #include "en/xsk/tx.h"
  64 #include "en/hv_vhca_stats.h"
  65 #include "en/devlink.h"
  66 #include "lib/mlx5.h"
  67 #include "en/ptp.h"
  68 #include "qos.h"
  69 #include "en/trap.h"
  70 #include "fpga/ipsec.h"
  71
  72 bool mlx5e_check_fragmented_striding_rq_cap(struct mlx5_core_dev *mdev)
  73 {
  74         bool striding_rq_umr = MLX5_CAP_GEN(mdev, striding_rq) &&
  75                 MLX5_CAP_GEN(mdev, umr_ptr_rlky) &&
  76                 MLX5_CAP_ETH(mdev, reg_umr_sq);
  77         u16 max_wqe_sz_cap = MLX5_CAP_GEN(mdev, max_wqe_sz_sq);
  78         bool inline_umr = MLX5E_UMR_WQE_INLINE_SZ <= max_wqe_sz_cap;
  79
  80         if (!striding_rq_umr)
  81                 return false;
  82         if (!inline_umr) {
  83                 mlx5_core_warn(mdev, "Cannot support Striding RQ: UMR WQE size (%d) exceeds maximum supported (%d).\n",
  84                                (int)MLX5E_UMR_WQE_INLINE_SZ, max_wqe_sz_cap);
  85                 return false;
  86         }
  87         return true;
  88 }
  89
  90 void mlx5e_init_rq_type_params(struct mlx5_core_dev *mdev,
  91                                struct mlx5e_params *params)
  92 {
  93         params->log_rq_mtu_frames = is_kdump_kernel() ?
  94                 MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE :
  95                 MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
  96
  97         mlx5_core_info(mdev, "MLX5E: StrdRq(%d) RqSz(%ld) StrdSz(%ld) RxCqeCmprss(%d)\n",
  98                        params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ,
  99                        params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ ?
 100                        BIT(mlx5e_mpwqe_get_log_rq_size(params, NULL)) :
 101                        BIT(params->log_rq_mtu_frames),
 102                        BIT(mlx5e_mpwqe_get_log_stride_size(mdev, params, NULL)),
 103                        MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS));
 104 }
 105
 106 bool mlx5e_striding_rq_possible(struct mlx5_core_dev *mdev,
 107                                 struct mlx5e_params *params)
 108 {
 109         if (!mlx5e_check_fragmented_striding_rq_cap(mdev))
 110                 return false;
 111
 112         if (mlx5_fpga_is_ipsec_device(mdev))
 113                 return false;
 114
 115         if (params->xdp_prog) {
 116                 /* XSK params are not considered here. If striding RQ is in use,
 117                  * and an XSK is being opened, mlx5e_rx_mpwqe_is_linear_skb will
 118                  * be called with the known XSK params.
 119                  */
 120                 if (!mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL))
 121                         return false;
 122         }
 123
 124         return true;
 125 }
 126
 127 void mlx5e_set_rq_type(struct mlx5_core_dev *mdev, struct mlx5e_params *params)
 128 {
 129         params->rq_wq_type = mlx5e_striding_rq_possible(mdev, params) &&
 130                 MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ) ?
 131                 MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ :
 132                 MLX5_WQ_TYPE_CYCLIC;
 133 }
 134
 135 void mlx5e_update_carrier(struct mlx5e_priv *priv)
 136 {
 137         struct mlx5_core_dev *mdev = priv->mdev;
 138         u8 port_state;
 139
 140         port_state = mlx5_query_vport_state(mdev,
 141                                             MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT,
 142                                             0);
 143
 144         if (port_state == VPORT_STATE_UP) {
 145                 netdev_info(priv->netdev, "Link up\n");
 146                 netif_carrier_on(priv->netdev);
 147         } else {
 148                 netdev_info(priv->netdev, "Link down\n");
 149                 netif_carrier_off(priv->netdev);
 150         }
 151 }
 152
 153 static void mlx5e_update_carrier_work(struct work_struct *work)
 154 {
 155         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 156                                                update_carrier_work);
 157
 158         mutex_lock(&priv->state_lock);
 159         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
 160                 if (priv->profile->update_carrier)
 161                         priv->profile->update_carrier(priv);
 162         mutex_unlock(&priv->state_lock);
 163 }
 164
 165 static void mlx5e_update_stats_work(struct work_struct *work)
 166 {
 167         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 168                                                update_stats_work);
 169
 170         mutex_lock(&priv->state_lock);
 171         priv->profile->update_stats(priv);
 172         mutex_unlock(&priv->state_lock);
 173 }
 174
 175 void mlx5e_queue_update_stats(struct mlx5e_priv *priv)
 176 {
 177         if (!priv->profile->update_stats)
 178                 return;
 179
 180         if (unlikely(test_bit(MLX5E_STATE_DESTROYING, &priv->state)))
 181                 return;
 182
 183         queue_work(priv->wq, &priv->update_stats_work);
 184 }
 185
 186 static int async_event(struct notifier_block *nb, unsigned long event, void *data)
 187 {
 188         struct mlx5e_priv *priv = container_of(nb, struct mlx5e_priv, events_nb);
 189         struct mlx5_eqe   *eqe = data;
 190
 191         if (event != MLX5_EVENT_TYPE_PORT_CHANGE)
 192                 return NOTIFY_DONE;
 193
 194         switch (eqe->sub_type) {
 195         case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
 196         case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
 197                 queue_work(priv->wq, &priv->update_carrier_work);
 198                 break;
 199         default:
 200                 return NOTIFY_DONE;
 201         }
 202
 203         return NOTIFY_OK;
 204 }
 205
 206 static void mlx5e_enable_async_events(struct mlx5e_priv *priv)
 207 {
 208         priv->events_nb.notifier_call = async_event;
 209         mlx5_notifier_register(priv->mdev, &priv->events_nb);
 210 }
 211
 212 static void mlx5e_disable_async_events(struct mlx5e_priv *priv)
 213 {
 214         mlx5_notifier_unregister(priv->mdev, &priv->events_nb);
 215 }
 216
 217 static int blocking_event(struct notifier_block *nb, unsigned long event, void *data)
 218 {
 219         struct mlx5e_priv *priv = container_of(nb, struct mlx5e_priv, blocking_events_nb);
 220         int err;
 221
 222         switch (event) {
 223         case MLX5_DRIVER_EVENT_TYPE_TRAP:
 224                 err = mlx5e_handle_trap_event(priv, data);
 225                 break;
 226         default:
 227                 netdev_warn(priv->netdev, "Sync event: Unknown event %ld\n", event);
 228                 err = -EINVAL;
 229         }
 230         return err;
 231 }
 232
 233 static void mlx5e_enable_blocking_events(struct mlx5e_priv *priv)
 234 {
 235         priv->blocking_events_nb.notifier_call = blocking_event;
 236         mlx5_blocking_notifier_register(priv->mdev, &priv->blocking_events_nb);
 237 }
 238
 239 static void mlx5e_disable_blocking_events(struct mlx5e_priv *priv)
 240 {
 241         mlx5_blocking_notifier_unregister(priv->mdev, &priv->blocking_events_nb);
 242 }
 243
 244 static inline void mlx5e_build_umr_wqe(struct mlx5e_rq *rq,
 245                                        struct mlx5e_icosq *sq,
 246                                        struct mlx5e_umr_wqe *wqe)
 247 {
 248         struct mlx5_wqe_ctrl_seg      *cseg = &wqe->ctrl;
 249         struct mlx5_wqe_umr_ctrl_seg *ucseg = &wqe->uctrl;
 250         u8 ds_cnt = DIV_ROUND_UP(MLX5E_UMR_WQE_INLINE_SZ, MLX5_SEND_WQE_DS);
 251
 252         cseg->qpn_ds    = cpu_to_be32((sq->sqn << MLX5_WQE_CTRL_QPN_SHIFT) |
 253                                       ds_cnt);
 254         cseg->umr_mkey  = rq->mkey_be;
 255
 256         ucseg->flags = MLX5_UMR_TRANSLATION_OFFSET_EN | MLX5_UMR_INLINE;
 257         ucseg->xlt_octowords =
 258                 cpu_to_be16(MLX5_MTT_OCTW(MLX5_MPWRQ_PAGES_PER_WQE));
 259         ucseg->mkey_mask     = cpu_to_be64(MLX5_MKEY_MASK_FREE);
 260 }
 261
 262 static int mlx5e_rq_alloc_mpwqe_info(struct mlx5e_rq *rq,
 263                                      struct mlx5e_channel *c)
 264 {
 265         int wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 266
 267         rq->mpwqe.info = kvzalloc_node(array_size(wq_sz,
 268                                                   sizeof(*rq->mpwqe.info)),
 269                                        GFP_KERNEL, cpu_to_node(c->cpu));
 270         if (!rq->mpwqe.info)
 271                 return -ENOMEM;
 272
 273         mlx5e_build_umr_wqe(rq, &c->icosq, &rq->mpwqe.umr_wqe);
 274
 275         return 0;
 276 }
 277
 278 static int mlx5e_create_umr_mkey(struct mlx5_core_dev *mdev,
 279                                  u64 npages, u8 page_shift,
 280                                  struct mlx5_core_mkey *umr_mkey,
 281                                  dma_addr_t filler_addr)
 282 {
 283         struct mlx5_mtt *mtt;
 284         int inlen;
 285         void *mkc;
 286         u32 *in;
 287         int err;
 288         int i;
 289
 290         inlen = MLX5_ST_SZ_BYTES(create_mkey_in) + sizeof(*mtt) * npages;
 291
 292         in = kvzalloc(inlen, GFP_KERNEL);
 293         if (!in)
 294                 return -ENOMEM;
 295
 296         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 297
 298         MLX5_SET(mkc, mkc, free, 1);
 299         MLX5_SET(mkc, mkc, umr_en, 1);
 300         MLX5_SET(mkc, mkc, lw, 1);
 301         MLX5_SET(mkc, mkc, lr, 1);
 302         MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
 303         mlx5e_mkey_set_relaxed_ordering(mdev, mkc);
 304         MLX5_SET(mkc, mkc, qpn, 0xffffff);
 305         MLX5_SET(mkc, mkc, pd, mdev->mlx5e_res.pdn);
 306         MLX5_SET64(mkc, mkc, len, npages << page_shift);
 307         MLX5_SET(mkc, mkc, translations_octword_size,
 308                  MLX5_MTT_OCTW(npages));
 309         MLX5_SET(mkc, mkc, log_page_size, page_shift);
 310         MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
 311                  MLX5_MTT_OCTW(npages));
 312
 313         /* Initialize the mkey with all MTTs pointing to a default
 314          * page (filler_addr). When the channels are activated, UMR
 315          * WQEs will redirect the RX WQEs to the actual memory from
 316          * the RQ's pool, while the gaps (wqe_overflow) remain mapped
 317          * to the default page.
 318          */
 319         mtt = MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
 320         for (i = 0 ; i < npages ; i++)
 321                 mtt[i].ptag = cpu_to_be64(filler_addr);
 322
 323         err = mlx5_core_create_mkey(mdev, umr_mkey, in, inlen);
 324
 325         kvfree(in);
 326         return err;
 327 }
 328
 329 static int mlx5e_create_rq_umr_mkey(struct mlx5_core_dev *mdev, struct mlx5e_rq *rq)
 330 {
 331         u64 num_mtts = MLX5E_REQUIRED_MTTS(mlx5_wq_ll_get_size(&rq->mpwqe.wq));
 332
 333         return mlx5e_create_umr_mkey(mdev, num_mtts, PAGE_SHIFT, &rq->umr_mkey,
 334                                      rq->wqe_overflow.addr);
 335 }
 336
 337 static inline u64 mlx5e_get_mpwqe_offset(struct mlx5e_rq *rq, u16 wqe_ix)
 338 {
 339         return (wqe_ix << MLX5E_LOG_ALIGNED_MPWQE_PPW) << PAGE_SHIFT;
 340 }
 341
 342 static void mlx5e_init_frags_partition(struct mlx5e_rq *rq)
 343 {
 344         struct mlx5e_wqe_frag_info next_frag = {};
 345         struct mlx5e_wqe_frag_info *prev = NULL;
 346         int i;
 347
 348         next_frag.di = &rq->wqe.di[0];
 349
 350         for (i = 0; i < mlx5_wq_cyc_get_size(&rq->wqe.wq); i++) {
 351                 struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
 352                 struct mlx5e_wqe_frag_info *frag =
 353                         &rq->wqe.frags[i << rq->wqe.info.log_num_frags];
 354                 int f;
 355
 356                 for (f = 0; f < rq->wqe.info.num_frags; f++, frag++) {
 357                         if (next_frag.offset + frag_info[f].frag_stride > PAGE_SIZE) {
 358                                 next_frag.di++;
 359                                 next_frag.offset = 0;
 360                                 if (prev)
 361                                         prev->last_in_page = true;
 362                         }
 363                         *frag = next_frag;
 364
 365                         /* prepare next */
 366                         next_frag.offset += frag_info[f].frag_stride;
 367                         prev = frag;
 368                 }
 369         }
 370
 371         if (prev)
 372                 prev->last_in_page = true;
 373 }
 374
 375 int mlx5e_init_di_list(struct mlx5e_rq *rq, int wq_sz, int node)
 376 {
 377         int len = wq_sz << rq->wqe.info.log_num_frags;
 378
 379         rq->wqe.di = kvzalloc_node(array_size(len, sizeof(*rq->wqe.di)), GFP_KERNEL, node);
 380         if (!rq->wqe.di)
 381                 return -ENOMEM;
 382
 383         mlx5e_init_frags_partition(rq);
 384
 385         return 0;
 386 }
 387
 388 void mlx5e_free_di_list(struct mlx5e_rq *rq)
 389 {
 390         kvfree(rq->wqe.di);
 391 }
 392
 393 static void mlx5e_rq_err_cqe_work(struct work_struct *recover_work)
 394 {
 395         struct mlx5e_rq *rq = container_of(recover_work, struct mlx5e_rq, recover_work);
 396
 397         mlx5e_reporter_rq_cqe_err(rq);
 398 }
 399
 400 static int mlx5e_alloc_mpwqe_rq_drop_page(struct mlx5e_rq *rq)
 401 {
 402         rq->wqe_overflow.page = alloc_page(GFP_KERNEL);
 403         if (!rq->wqe_overflow.page)
 404                 return -ENOMEM;
 405
 406         rq->wqe_overflow.addr = dma_map_page(rq->pdev, rq->wqe_overflow.page, 0,
 407                                              PAGE_SIZE, rq->buff.map_dir);
 408         if (dma_mapping_error(rq->pdev, rq->wqe_overflow.addr)) {
 409                 __free_page(rq->wqe_overflow.page);
 410                 return -ENOMEM;
 411         }
 412         return 0;
 413 }
 414
 415 static void mlx5e_free_mpwqe_rq_drop_page(struct mlx5e_rq *rq)
 416 {
 417          dma_unmap_page(rq->pdev, rq->wqe_overflow.addr, PAGE_SIZE,
 418                         rq->buff.map_dir);
 419          __free_page(rq->wqe_overflow.page);
 420 }
 421
 422 static int mlx5e_alloc_rq(struct mlx5e_channel *c,
 423                           struct mlx5e_params *params,
 424                           struct mlx5e_xsk_param *xsk,
 425                           struct xsk_buff_pool *xsk_pool,
 426                           struct mlx5e_rq_param *rqp,
 427                           struct mlx5e_rq *rq)
 428 {
 429         struct page_pool_params pp_params = { 0 };
 430         struct mlx5_core_dev *mdev = c->mdev;
 431         void *rqc = rqp->rqc;
 432         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 433         u32 rq_xdp_ix;
 434         u32 pool_size;
 435         int wq_sz;
 436         int err;
 437         int i;
 438
 439         rqp->wq.db_numa_node = cpu_to_node(c->cpu);
 440
 441         rq->wq_type = params->rq_wq_type;
 442         rq->pdev    = c->pdev;
 443         rq->netdev  = c->netdev;
 444         rq->priv    = c->priv;
 445         rq->tstamp  = c->tstamp;
 446         rq->clock   = &mdev->clock;
 447         rq->icosq   = &c->icosq;
 448         rq->ix      = c->ix;
 449         rq->mdev    = mdev;
 450         rq->hw_mtu  = MLX5E_SW2HW_MTU(params, params->sw_mtu);
 451         rq->xdpsq   = &c->rq_xdpsq;
 452         rq->xsk_pool = xsk_pool;
 453         rq->ptp_cyc2time = mlx5_is_real_time_rq(mdev) ?
 454                            mlx5_real_time_cyc2time :
 455                            mlx5_timecounter_cyc2time;
 456
 457         if (rq->xsk_pool)
 458                 rq->stats = &c->priv->channel_stats[c->ix].xskrq;
 459         else
 460                 rq->stats = &c->priv->channel_stats[c->ix].rq;
 461         INIT_WORK(&rq->recover_work, mlx5e_rq_err_cqe_work);
 462
 463         if (params->xdp_prog)
 464                 bpf_prog_inc(params->xdp_prog);
 465         RCU_INIT_POINTER(rq->xdp_prog, params->xdp_prog);
 466
 467         rq_xdp_ix = rq->ix;
 468         if (xsk)
 469                 rq_xdp_ix += params->num_channels * MLX5E_RQ_GROUP_XSK;
 470         err = xdp_rxq_info_reg(&rq->xdp_rxq, rq->netdev, rq_xdp_ix, 0);
 471         if (err < 0)
 472                 goto err_rq_xdp_prog;
 473
 474         rq->buff.map_dir = params->xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
 475         rq->buff.headroom = mlx5e_get_rq_headroom(mdev, params, xsk);
 476         pool_size = 1 << params->log_rq_mtu_frames;
 477
 478         switch (rq->wq_type) {
 479         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 480                 err = mlx5_wq_ll_create(mdev, &rqp->wq, rqc_wq, &rq->mpwqe.wq,
 481                                         &rq->wq_ctrl);
 482                 if (err)
 483                         goto err_rq_xdp;
 484
 485                 err = mlx5e_alloc_mpwqe_rq_drop_page(rq);
 486                 if (err)
 487                         goto err_rq_wq_destroy;
 488
 489                 rq->mpwqe.wq.db = &rq->mpwqe.wq.db[MLX5_RCV_DBR];
 490
 491                 wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 492
 493                 pool_size = MLX5_MPWRQ_PAGES_PER_WQE <<
 494                         mlx5e_mpwqe_get_log_rq_size(params, xsk);
 495
 496                 rq->mpwqe.log_stride_sz = mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
 497                 rq->mpwqe.num_strides =
 498                         BIT(mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk));
 499
 500                 rq->buff.frame0_sz = (1 << rq->mpwqe.log_stride_sz);
 501
 502                 err = mlx5e_create_rq_umr_mkey(mdev, rq);
 503                 if (err)
 504                         goto err_rq_drop_page;
 505                 rq->mkey_be = cpu_to_be32(rq->umr_mkey.key);
 506
 507                 err = mlx5e_rq_alloc_mpwqe_info(rq, c);
 508                 if (err)
 509                         goto err_rq_mkey;
 510                 break;
 511         default: /* MLX5_WQ_TYPE_CYCLIC */
 512                 err = mlx5_wq_cyc_create(mdev, &rqp->wq, rqc_wq, &rq->wqe.wq,
 513                                          &rq->wq_ctrl);
 514                 if (err)
 515                         goto err_rq_xdp;
 516
 517                 rq->wqe.wq.db = &rq->wqe.wq.db[MLX5_RCV_DBR];
 518
 519                 wq_sz = mlx5_wq_cyc_get_size(&rq->wqe.wq);
 520
 521                 rq->wqe.info = rqp->frags_info;
 522                 rq->buff.frame0_sz = rq->wqe.info.arr[0].frag_stride;
 523
 524                 rq->wqe.frags =
 525                         kvzalloc_node(array_size(sizeof(*rq->wqe.frags),
 526                                         (wq_sz << rq->wqe.info.log_num_frags)),
 527                                       GFP_KERNEL, cpu_to_node(c->cpu));
 528                 if (!rq->wqe.frags) {
 529                         err = -ENOMEM;
 530                         goto err_rq_wq_destroy;
 531                 }
 532
 533                 err = mlx5e_init_di_list(rq, wq_sz, cpu_to_node(c->cpu));
 534                 if (err)
 535                         goto err_rq_frags;
 536
 537                 rq->mkey_be = c->mkey_be;
 538         }
 539
 540         err = mlx5e_rq_set_handlers(rq, params, xsk);
 541         if (err)
 542                 goto err_free_by_rq_type;
 543
 544         if (xsk) {
 545                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
 546                                                  MEM_TYPE_XSK_BUFF_POOL, NULL);
 547                 xsk_pool_set_rxq_info(rq->xsk_pool, &rq->xdp_rxq);
 548         } else {
 549                 /* Create a page_pool and register it with rxq */
 550                 pp_params.order     = 0;
 551                 pp_params.flags     = 0; /* No-internal DMA mapping in page_pool */
 552                 pp_params.pool_size = pool_size;
 553                 pp_params.nid       = cpu_to_node(c->cpu);
 554                 pp_params.dev       = c->pdev;
 555                 pp_params.dma_dir   = rq->buff.map_dir;
 556
 557                 /* page_pool can be used even when there is no rq->xdp_prog,
 558                  * given page_pool does not handle DMA mapping there is no
 559                  * required state to clear. And page_pool gracefully handle
 560                  * elevated refcnt.
 561                  */
 562                 rq->page_pool = page_pool_create(&pp_params);
 563                 if (IS_ERR(rq->page_pool)) {
 564                         err = PTR_ERR(rq->page_pool);
 565                         rq->page_pool = NULL;
 566                         goto err_free_by_rq_type;
 567                 }
 568                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
 569                                                  MEM_TYPE_PAGE_POOL, rq->page_pool);
 570         }
 571         if (err)
 572                 goto err_free_by_rq_type;
 573
 574         for (i = 0; i < wq_sz; i++) {
 575                 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 576                         struct mlx5e_rx_wqe_ll *wqe =
 577                                 mlx5_wq_ll_get_wqe(&rq->mpwqe.wq, i);
 578                         u32 byte_count =
 579                                 rq->mpwqe.num_strides << rq->mpwqe.log_stride_sz;
 580                         u64 dma_offset = mlx5e_get_mpwqe_offset(rq, i);
 581
 582                         wqe->data[0].addr = cpu_to_be64(dma_offset + rq->buff.headroom);
 583                         wqe->data[0].byte_count = cpu_to_be32(byte_count);
 584                         wqe->data[0].lkey = rq->mkey_be;
 585                 } else {
 586                         struct mlx5e_rx_wqe_cyc *wqe =
 587                                 mlx5_wq_cyc_get_wqe(&rq->wqe.wq, i);
 588                         int f;
 589
 590                         for (f = 0; f < rq->wqe.info.num_frags; f++) {
 591                                 u32 frag_size = rq->wqe.info.arr[f].frag_size |
 592                                         MLX5_HW_START_PADDING;
 593
 594                                 wqe->data[f].byte_count = cpu_to_be32(frag_size);
 595                                 wqe->data[f].lkey = rq->mkey_be;
 596                         }
 597                         /* check if num_frags is not a pow of two */
 598                         if (rq->wqe.info.num_frags < (1 << rq->wqe.info.log_num_frags)) {
 599                                 wqe->data[f].byte_count = 0;
 600                                 wqe->data[f].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
 601                                 wqe->data[f].addr = 0;
 602                         }
 603                 }
 604         }
 605
 606         INIT_WORK(&rq->dim.work, mlx5e_rx_dim_work);
 607
 608         switch (params->rx_cq_moderation.cq_period_mode) {
 609         case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
 610                 rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
 611                 break;
 612         case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
 613         default:
 614                 rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
 615         }
 616
 617         rq->page_cache.head = 0;
 618         rq->page_cache.tail = 0;
 619
 620         return 0;
 621
 622 err_free_by_rq_type:
 623         switch (rq->wq_type) {
 624         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 625                 kvfree(rq->mpwqe.info);
 626 err_rq_mkey:
 627                 mlx5_core_destroy_mkey(mdev, &rq->umr_mkey);
 628 err_rq_drop_page:
 629                 mlx5e_free_mpwqe_rq_drop_page(rq);
 630                 break;
 631         default: /* MLX5_WQ_TYPE_CYCLIC */
 632                 mlx5e_free_di_list(rq);
 633 err_rq_frags:
 634                 kvfree(rq->wqe.frags);
 635         }
 636 err_rq_wq_destroy:
 637         mlx5_wq_destroy(&rq->wq_ctrl);
 638 err_rq_xdp:
 639         xdp_rxq_info_unreg(&rq->xdp_rxq);
 640 err_rq_xdp_prog:
 641         if (params->xdp_prog)
 642                 bpf_prog_put(params->xdp_prog);
 643
 644         return err;
 645 }
 646
 647 static void mlx5e_free_rq(struct mlx5e_rq *rq)
 648 {
 649         struct bpf_prog *old_prog;
 650         int i;
 651
 652         old_prog = rcu_dereference_protected(rq->xdp_prog,
 653                                              lockdep_is_held(&rq->priv->state_lock));
 654         if (old_prog)
 655                 bpf_prog_put(old_prog);
 656
 657         switch (rq->wq_type) {
 658         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 659                 kvfree(rq->mpwqe.info);
 660                 mlx5_core_destroy_mkey(rq->mdev, &rq->umr_mkey);
 661                 mlx5e_free_mpwqe_rq_drop_page(rq);
 662                 break;
 663         default: /* MLX5_WQ_TYPE_CYCLIC */
 664                 kvfree(rq->wqe.frags);
 665                 mlx5e_free_di_list(rq);
 666         }
 667
 668         for (i = rq->page_cache.head; i != rq->page_cache.tail;
 669              i = (i + 1) & (MLX5E_CACHE_SIZE - 1)) {
 670                 struct mlx5e_dma_info *dma_info = &rq->page_cache.page_cache[i];
 671
 672                 /* With AF_XDP, page_cache is not used, so this loop is not
 673                  * entered, and it's safe to call mlx5e_page_release_dynamic
 674                  * directly.
 675                  */
 676                 mlx5e_page_release_dynamic(rq, dma_info, false);
 677         }
 678
 679         xdp_rxq_info_unreg(&rq->xdp_rxq);
 680         page_pool_destroy(rq->page_pool);
 681         mlx5_wq_destroy(&rq->wq_ctrl);
 682 }
 683
 684 int mlx5e_create_rq(struct mlx5e_rq *rq, struct mlx5e_rq_param *param)
 685 {
 686         struct mlx5_core_dev *mdev = rq->mdev;
 687         u8 ts_format;
 688         void *in;
 689         void *rqc;
 690         void *wq;
 691         int inlen;
 692         int err;
 693
 694         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 695                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 696         in = kvzalloc(inlen, GFP_KERNEL);
 697         if (!in)
 698                 return -ENOMEM;
 699
 700         ts_format = mlx5_is_real_time_rq(mdev) ?
 701                     MLX5_RQC_TIMESTAMP_FORMAT_REAL_TIME :
 702                     MLX5_RQC_TIMESTAMP_FORMAT_FREE_RUNNING;
 703         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 704         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 705
 706         memcpy(rqc, param->rqc, sizeof(param->rqc));
 707
 708         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 709         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 710         MLX5_SET(rqc,  rqc, ts_format,          ts_format);
 711         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 712                                                 MLX5_ADAPTER_PAGE_SHIFT);
 713         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 714
 715         mlx5_fill_page_frag_array(&rq->wq_ctrl.buf,
 716                                   (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 717
 718         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 719
 720         kvfree(in);
 721
 722         return err;
 723 }
 724
 725 int mlx5e_modify_rq_state(struct mlx5e_rq *rq, int curr_state, int next_state)
 726 {
 727         struct mlx5_core_dev *mdev = rq->mdev;
 728
 729         void *in;
 730         void *rqc;
 731         int inlen;
 732         int err;
 733
 734         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 735         in = kvzalloc(inlen, GFP_KERNEL);
 736         if (!in)
 737                 return -ENOMEM;
 738
 739         if (curr_state == MLX5_RQC_STATE_RST && next_state == MLX5_RQC_STATE_RDY)
 740                 mlx5e_rqwq_reset(rq);
 741
 742         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 743
 744         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 745         MLX5_SET(rqc, rqc, state, next_state);
 746
 747         err = mlx5_core_modify_rq(mdev, rq->rqn, in);
 748
 749         kvfree(in);
 750
 751         return err;
 752 }
 753
 754 static int mlx5e_modify_rq_scatter_fcs(struct mlx5e_rq *rq, bool enable)
 755 {
 756         struct mlx5_core_dev *mdev = rq->mdev;
 757
 758         void *in;
 759         void *rqc;
 760         int inlen;
 761         int err;
 762
 763         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 764         in = kvzalloc(inlen, GFP_KERNEL);
 765         if (!in)
 766                 return -ENOMEM;
 767
 768         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 769
 770         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 771         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 772                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_SCATTER_FCS);
 773         MLX5_SET(rqc, rqc, scatter_fcs, enable);
 774         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 775
 776         err = mlx5_core_modify_rq(mdev, rq->rqn, in);
 777
 778         kvfree(in);
 779
 780         return err;
 781 }
 782
 783 static int mlx5e_modify_rq_vsd(struct mlx5e_rq *rq, bool vsd)
 784 {
 785         struct mlx5_core_dev *mdev = rq->mdev;
 786         void *in;
 787         void *rqc;
 788         int inlen;
 789         int err;
 790
 791         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 792         in = kvzalloc(inlen, GFP_KERNEL);
 793         if (!in)
 794                 return -ENOMEM;
 795
 796         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 797
 798         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 799         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 800                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
 801         MLX5_SET(rqc, rqc, vsd, vsd);
 802         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 803
 804         err = mlx5_core_modify_rq(mdev, rq->rqn, in);
 805
 806         kvfree(in);
 807
 808         return err;
 809 }
 810
 811 void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 812 {
 813         mlx5_core_destroy_rq(rq->mdev, rq->rqn);
 814 }
 815
 816 int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq, int wait_time)
 817 {
 818         unsigned long exp_time = jiffies + msecs_to_jiffies(wait_time);
 819
 820         u16 min_wqes = mlx5_min_rx_wqes(rq->wq_type, mlx5e_rqwq_get_size(rq));
 821
 822         do {
 823                 if (mlx5e_rqwq_get_cur_sz(rq) >= min_wqes)
 824                         return 0;
 825
 826                 msleep(20);
 827         } while (time_before(jiffies, exp_time));
 828
 829         netdev_warn(rq->netdev, "Failed to get min RX wqes on Channel[%d] RQN[0x%x] wq cur_sz(%d) min_rx_wqes(%d)\n",
 830                     rq->ix, rq->rqn, mlx5e_rqwq_get_cur_sz(rq), min_wqes);
 831
 832         mlx5e_reporter_rx_timeout(rq);
 833         return -ETIMEDOUT;
 834 }
 835
 836 void mlx5e_free_rx_in_progress_descs(struct mlx5e_rq *rq)
 837 {
 838         struct mlx5_wq_ll *wq;
 839         u16 head;
 840         int i;
 841
 842         if (rq->wq_type != MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
 843                 return;
 844
 845         wq = &rq->mpwqe.wq;
 846         head = wq->head;
 847
 848         /* Outstanding UMR WQEs (in progress) start at wq->head */
 849         for (i = 0; i < rq->mpwqe.umr_in_progress; i++) {
 850                 rq->dealloc_wqe(rq, head);
 851                 head = mlx5_wq_ll_get_wqe_next_ix(wq, head);
 852         }
 853
 854         rq->mpwqe.actual_wq_head = wq->head;
 855         rq->mpwqe.umr_in_progress = 0;
 856         rq->mpwqe.umr_completed = 0;
 857 }
 858
 859 void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 860 {
 861         __be16 wqe_ix_be;
 862         u16 wqe_ix;
 863
 864         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 865                 struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
 866
 867                 mlx5e_free_rx_in_progress_descs(rq);
 868
 869                 while (!mlx5_wq_ll_is_empty(wq)) {
 870                         struct mlx5e_rx_wqe_ll *wqe;
 871
 872                         wqe_ix_be = *wq->tail_next;
 873                         wqe_ix    = be16_to_cpu(wqe_ix_be);
 874                         wqe       = mlx5_wq_ll_get_wqe(wq, wqe_ix);
 875                         rq->dealloc_wqe(rq, wqe_ix);
 876                         mlx5_wq_ll_pop(wq, wqe_ix_be,
 877                                        &wqe->next.next_wqe_index);
 878                 }
 879         } else {
 880                 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
 881
 882                 while (!mlx5_wq_cyc_is_empty(wq)) {
 883                         wqe_ix = mlx5_wq_cyc_get_tail(wq);
 884                         rq->dealloc_wqe(rq, wqe_ix);
 885                         mlx5_wq_cyc_pop(wq);
 886                 }
 887         }
 888
 889 }
 890
 891 int mlx5e_open_rq(struct mlx5e_channel *c, struct mlx5e_params *params,
 892                   struct mlx5e_rq_param *param, struct mlx5e_xsk_param *xsk,
 893                   struct xsk_buff_pool *xsk_pool, struct mlx5e_rq *rq)
 894 {
 895         int err;
 896
 897         err = mlx5e_alloc_rq(c, params, xsk, xsk_pool, param, rq);
 898         if (err)
 899                 return err;
 900
 901         err = mlx5e_create_rq(rq, param);
 902         if (err)
 903                 goto err_free_rq;
 904
 905         err = mlx5e_modify_rq_state(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 906         if (err)
 907                 goto err_destroy_rq;
 908
 909         if (mlx5e_is_tls_on(c->priv) && !mlx5_accel_is_ktls_device(c->mdev))
 910                 __set_bit(MLX5E_RQ_STATE_FPGA_TLS, &c->rq.state); /* must be FPGA */
 911
 912         if (MLX5_CAP_ETH(c->mdev, cqe_checksum_full))
 913                 __set_bit(MLX5E_RQ_STATE_CSUM_FULL, &c->rq.state);
 914
 915         if (params->rx_dim_enabled)
 916                 __set_bit(MLX5E_RQ_STATE_AM, &c->rq.state);
 917
 918         /* We disable csum_complete when XDP is enabled since
 919          * XDP programs might manipulate packets which will render
 920          * skb->checksum incorrect.
 921          */
 922         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_NO_CSUM_COMPLETE) || c->xdp)
 923                 __set_bit(MLX5E_RQ_STATE_NO_CSUM_COMPLETE, &c->rq.state);
 924
 925         /* For CQE compression on striding RQ, use stride index provided by
 926          * HW if capability is supported.
 927          */
 928         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ) &&
 929             MLX5_CAP_GEN(c->mdev, mini_cqe_resp_stride_index))
 930                 __set_bit(MLX5E_RQ_STATE_MINI_CQE_HW_STRIDX, &c->rq.state);
 931
 932         return 0;
 933
 934 err_destroy_rq:
 935         mlx5e_destroy_rq(rq);
 936 err_free_rq:
 937         mlx5e_free_rq(rq);
 938
 939         return err;
 940 }
 941
 942 void mlx5e_activate_rq(struct mlx5e_rq *rq)
 943 {
 944         set_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 945         mlx5e_trigger_irq(rq->icosq);
 946 }
 947
 948 void mlx5e_deactivate_rq(struct mlx5e_rq *rq)
 949 {
 950         clear_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 951         synchronize_net(); /* Sync with NAPI to prevent mlx5e_post_rx_wqes. */
 952 }
 953
 954 void mlx5e_close_rq(struct mlx5e_rq *rq)
 955 {
 956         cancel_work_sync(&rq->dim.work);
 957         cancel_work_sync(&rq->icosq->recover_work);
 958         cancel_work_sync(&rq->recover_work);
 959         mlx5e_destroy_rq(rq);
 960         mlx5e_free_rx_descs(rq);
 961         mlx5e_free_rq(rq);
 962 }
 963
 964 static void mlx5e_free_xdpsq_db(struct mlx5e_xdpsq *sq)
 965 {
 966         kvfree(sq->db.xdpi_fifo.xi);
 967         kvfree(sq->db.wqe_info);
 968 }
 969
 970 static int mlx5e_alloc_xdpsq_fifo(struct mlx5e_xdpsq *sq, int numa)
 971 {
 972         struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo;
 973         int wq_sz        = mlx5_wq_cyc_get_size(&sq->wq);
 974         int dsegs_per_wq = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
 975
 976         xdpi_fifo->xi = kvzalloc_node(sizeof(*xdpi_fifo->xi) * dsegs_per_wq,
 977                                       GFP_KERNEL, numa);
 978         if (!xdpi_fifo->xi)
 979                 return -ENOMEM;
 980
 981         xdpi_fifo->pc   = &sq->xdpi_fifo_pc;
 982         xdpi_fifo->cc   = &sq->xdpi_fifo_cc;
 983         xdpi_fifo->mask = dsegs_per_wq - 1;
 984
 985         return 0;
 986 }
 987
 988 static int mlx5e_alloc_xdpsq_db(struct mlx5e_xdpsq *sq, int numa)
 989 {
 990         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 991         int err;
 992
 993         sq->db.wqe_info = kvzalloc_node(sizeof(*sq->db.wqe_info) * wq_sz,
 994                                         GFP_KERNEL, numa);
 995         if (!sq->db.wqe_info)
 996                 return -ENOMEM;
 997
 998         err = mlx5e_alloc_xdpsq_fifo(sq, numa);
 999         if (err) {
1000                 mlx5e_free_xdpsq_db(sq);
1001                 return err;
1002         }
1003
1004         return 0;
1005 }
1006
1007 static int mlx5e_alloc_xdpsq(struct mlx5e_channel *c,
1008                              struct mlx5e_params *params,
1009                              struct xsk_buff_pool *xsk_pool,
1010                              struct mlx5e_sq_param *param,
1011                              struct mlx5e_xdpsq *sq,
1012                              bool is_redirect)
1013 {
1014         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1015         struct mlx5_core_dev *mdev = c->mdev;
1016         struct mlx5_wq_cyc *wq = &sq->wq;
1017         int err;
1018
1019         sq->pdev      = c->pdev;
1020         sq->mkey_be   = c->mkey_be;
1021         sq->channel   = c;
1022         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1023         sq->min_inline_mode = params->tx_min_inline_mode;
1024         sq->hw_mtu    = MLX5E_SW2HW_MTU(params, params->sw_mtu);
1025         sq->xsk_pool  = xsk_pool;
1026
1027         sq->stats = sq->xsk_pool ?
1028                 &c->priv->channel_stats[c->ix].xsksq :
1029                 is_redirect ?
1030                         &c->priv->channel_stats[c->ix].xdpsq :
1031                         &c->priv->channel_stats[c->ix].rq_xdpsq;
1032
1033         param->wq.db_numa_node = cpu_to_node(c->cpu);
1034         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
1035         if (err)
1036                 return err;
1037         wq->db = &wq->db[MLX5_SND_DBR];
1038
1039         err = mlx5e_alloc_xdpsq_db(sq, cpu_to_node(c->cpu));
1040         if (err)
1041                 goto err_sq_wq_destroy;
1042
1043         return 0;
1044
1045 err_sq_wq_destroy:
1046         mlx5_wq_destroy(&sq->wq_ctrl);
1047
1048         return err;
1049 }
1050
1051 static void mlx5e_free_xdpsq(struct mlx5e_xdpsq *sq)
1052 {
1053         mlx5e_free_xdpsq_db(sq);
1054         mlx5_wq_destroy(&sq->wq_ctrl);
1055 }
1056
1057 static void mlx5e_free_icosq_db(struct mlx5e_icosq *sq)
1058 {
1059         kvfree(sq->db.wqe_info);
1060 }
1061
1062 static int mlx5e_alloc_icosq_db(struct mlx5e_icosq *sq, int numa)
1063 {
1064         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1065         size_t size;
1066
1067         size = array_size(wq_sz, sizeof(*sq->db.wqe_info));
1068         sq->db.wqe_info = kvzalloc_node(size, GFP_KERNEL, numa);
1069         if (!sq->db.wqe_info)
1070                 return -ENOMEM;
1071
1072         return 0;
1073 }
1074
1075 static void mlx5e_icosq_err_cqe_work(struct work_struct *recover_work)
1076 {
1077         struct mlx5e_icosq *sq = container_of(recover_work, struct mlx5e_icosq,
1078                                               recover_work);
1079
1080         mlx5e_reporter_icosq_cqe_err(sq);
1081 }
1082
1083 static int mlx5e_alloc_icosq(struct mlx5e_channel *c,
1084                              struct mlx5e_sq_param *param,
1085                              struct mlx5e_icosq *sq)
1086 {
1087         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1088         struct mlx5_core_dev *mdev = c->mdev;
1089         struct mlx5_wq_cyc *wq = &sq->wq;
1090         int err;
1091
1092         sq->channel   = c;
1093         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1094
1095         param->wq.db_numa_node = cpu_to_node(c->cpu);
1096         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
1097         if (err)
1098                 return err;
1099         wq->db = &wq->db[MLX5_SND_DBR];
1100
1101         err = mlx5e_alloc_icosq_db(sq, cpu_to_node(c->cpu));
1102         if (err)
1103                 goto err_sq_wq_destroy;
1104
1105         INIT_WORK(&sq->recover_work, mlx5e_icosq_err_cqe_work);
1106
1107         return 0;
1108
1109 err_sq_wq_destroy:
1110         mlx5_wq_destroy(&sq->wq_ctrl);
1111
1112         return err;
1113 }
1114
1115 static void mlx5e_free_icosq(struct mlx5e_icosq *sq)
1116 {
1117         mlx5e_free_icosq_db(sq);
1118         mlx5_wq_destroy(&sq->wq_ctrl);
1119 }
1120
1121 void mlx5e_free_txqsq_db(struct mlx5e_txqsq *sq)
1122 {
1123         kvfree(sq->db.wqe_info);
1124         kvfree(sq->db.skb_fifo.fifo);
1125         kvfree(sq->db.dma_fifo);
1126 }
1127
1128 int mlx5e_alloc_txqsq_db(struct mlx5e_txqsq *sq, int numa)
1129 {
1130         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1131         int df_sz = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
1132
1133         sq->db.dma_fifo = kvzalloc_node(array_size(df_sz,
1134                                                    sizeof(*sq->db.dma_fifo)),
1135                                         GFP_KERNEL, numa);
1136         sq->db.skb_fifo.fifo = kvzalloc_node(array_size(df_sz,
1137                                                         sizeof(*sq->db.skb_fifo.fifo)),
1138                                         GFP_KERNEL, numa);
1139         sq->db.wqe_info = kvzalloc_node(array_size(wq_sz,
1140                                                    sizeof(*sq->db.wqe_info)),
1141                                         GFP_KERNEL, numa);
1142         if (!sq->db.dma_fifo || !sq->db.skb_fifo.fifo || !sq->db.wqe_info) {
1143                 mlx5e_free_txqsq_db(sq);
1144                 return -ENOMEM;
1145         }
1146
1147         sq->dma_fifo_mask = df_sz - 1;
1148
1149         sq->db.skb_fifo.pc   = &sq->skb_fifo_pc;
1150         sq->db.skb_fifo.cc   = &sq->skb_fifo_cc;
1151         sq->db.skb_fifo.mask = df_sz - 1;
1152
1153         return 0;
1154 }
1155
1156 static int mlx5e_alloc_txqsq(struct mlx5e_channel *c,
1157                              int txq_ix,
1158                              struct mlx5e_params *params,
1159                              struct mlx5e_sq_param *param,
1160                              struct mlx5e_txqsq *sq,
1161                              int tc)
1162 {
1163         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1164         struct mlx5_core_dev *mdev = c->mdev;
1165         struct mlx5_wq_cyc *wq = &sq->wq;
1166         int err;
1167
1168         sq->pdev      = c->pdev;
1169         sq->tstamp    = c->tstamp;
1170         sq->clock     = &mdev->clock;
1171         sq->mkey_be   = c->mkey_be;
1172         sq->netdev    = c->netdev;
1173         sq->mdev      = c->mdev;
1174         sq->priv      = c->priv;
1175         sq->ch_ix     = c->ix;
1176         sq->txq_ix    = txq_ix;
1177         sq->uar_map   = mdev->mlx5e_res.bfreg.map;
1178         sq->min_inline_mode = params->tx_min_inline_mode;
1179         sq->hw_mtu    = MLX5E_SW2HW_MTU(params, params->sw_mtu);
1180         INIT_WORK(&sq->recover_work, mlx5e_tx_err_cqe_work);
1181         if (!MLX5_CAP_ETH(mdev, wqe_vlan_insert))
1182                 set_bit(MLX5E_SQ_STATE_VLAN_NEED_L2_INLINE, &sq->state);
1183         if (MLX5_IPSEC_DEV(c->priv->mdev))
1184                 set_bit(MLX5E_SQ_STATE_IPSEC, &sq->state);
1185         if (mlx5_accel_is_tls_device(c->priv->mdev))
1186                 set_bit(MLX5E_SQ_STATE_TLS, &sq->state);
1187         if (param->is_mpw)
1188                 set_bit(MLX5E_SQ_STATE_MPWQE, &sq->state);
1189         sq->stop_room = param->stop_room;
1190         sq->ptp_cyc2time = mlx5_is_real_time_sq(mdev) ?
1191                            mlx5_real_time_cyc2time :
1192                            mlx5_timecounter_cyc2time;
1193
1194         param->wq.db_numa_node = cpu_to_node(c->cpu);
1195         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
1196         if (err)
1197                 return err;
1198         wq->db    = &wq->db[MLX5_SND_DBR];
1199
1200         err = mlx5e_alloc_txqsq_db(sq, cpu_to_node(c->cpu));
1201         if (err)
1202                 goto err_sq_wq_destroy;
1203
1204         INIT_WORK(&sq->dim.work, mlx5e_tx_dim_work);
1205         sq->dim.mode = params->tx_cq_moderation.cq_period_mode;
1206
1207         return 0;
1208
1209 err_sq_wq_destroy:
1210         mlx5_wq_destroy(&sq->wq_ctrl);
1211
1212         return err;
1213 }
1214
1215 void mlx5e_free_txqsq(struct mlx5e_txqsq *sq)
1216 {
1217         mlx5e_free_txqsq_db(sq);
1218         mlx5_wq_destroy(&sq->wq_ctrl);
1219 }
1220
1221 static int mlx5e_create_sq(struct mlx5_core_dev *mdev,
1222                            struct mlx5e_sq_param *param,
1223                            struct mlx5e_create_sq_param *csp,
1224                            u32 *sqn)
1225 {
1226         u8 ts_format;
1227         void *in;
1228         void *sqc;
1229         void *wq;
1230         int inlen;
1231         int err;
1232
1233         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1234                 sizeof(u64) * csp->wq_ctrl->buf.npages;
1235         in = kvzalloc(inlen, GFP_KERNEL);
1236         if (!in)
1237                 return -ENOMEM;
1238
1239         ts_format = mlx5_is_real_time_sq(mdev) ?
1240                     MLX5_SQC_TIMESTAMP_FORMAT_REAL_TIME :
1241                     MLX5_SQC_TIMESTAMP_FORMAT_FREE_RUNNING;
1242         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1243         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1244
1245         memcpy(sqc, param->sqc, sizeof(param->sqc));
1246         MLX5_SET(sqc,  sqc, tis_lst_sz, csp->tis_lst_sz);
1247         MLX5_SET(sqc,  sqc, tis_num_0, csp->tisn);
1248         MLX5_SET(sqc,  sqc, cqn, csp->cqn);
1249         MLX5_SET(sqc,  sqc, ts_cqe_to_dest_cqn, csp->ts_cqe_to_dest_cqn);
1250         MLX5_SET(sqc,  sqc, ts_format, ts_format);
1251
1252
1253         if (MLX5_CAP_ETH(mdev, wqe_inline_mode) == MLX5_CAP_INLINE_MODE_VPORT_CONTEXT)
1254                 MLX5_SET(sqc,  sqc, min_wqe_inline_mode, csp->min_inline_mode);
1255
1256         MLX5_SET(sqc,  sqc, state, MLX5_SQC_STATE_RST);
1257         MLX5_SET(sqc,  sqc, flush_in_error_en, 1);
1258
1259         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
1260         MLX5_SET(wq,   wq, uar_page,      mdev->mlx5e_res.bfreg.index);
1261         MLX5_SET(wq,   wq, log_wq_pg_sz,  csp->wq_ctrl->buf.page_shift -
1262                                           MLX5_ADAPTER_PAGE_SHIFT);
1263         MLX5_SET64(wq, wq, dbr_addr,      csp->wq_ctrl->db.dma);
1264
1265         mlx5_fill_page_frag_array(&csp->wq_ctrl->buf,
1266                                   (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
1267
1268         err = mlx5_core_create_sq(mdev, in, inlen, sqn);
1269
1270         kvfree(in);
1271
1272         return err;
1273 }
1274
1275 int mlx5e_modify_sq(struct mlx5_core_dev *mdev, u32 sqn,
1276                     struct mlx5e_modify_sq_param *p)
1277 {
1278         u64 bitmask = 0;
1279         void *in;
1280         void *sqc;
1281         int inlen;
1282         int err;
1283
1284         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1285         in = kvzalloc(inlen, GFP_KERNEL);
1286         if (!in)
1287                 return -ENOMEM;
1288
1289         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1290
1291         MLX5_SET(modify_sq_in, in, sq_state, p->curr_state);
1292         MLX5_SET(sqc, sqc, state, p->next_state);
1293         if (p->rl_update && p->next_state == MLX5_SQC_STATE_RDY) {
1294                 bitmask |= 1;
1295                 MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, p->rl_index);
1296         }
1297         if (p->qos_update && p->next_state == MLX5_SQC_STATE_RDY) {
1298                 bitmask |= 1 << 2;
1299                 MLX5_SET(sqc, sqc, qos_queue_group_id, p->qos_queue_group_id);
1300         }
1301         MLX5_SET64(modify_sq_in, in, modify_bitmask, bitmask);
1302
1303         err = mlx5_core_modify_sq(mdev, sqn, in);
1304
1305         kvfree(in);
1306
1307         return err;
1308 }
1309
1310 static void mlx5e_destroy_sq(struct mlx5_core_dev *mdev, u32 sqn)
1311 {
1312         mlx5_core_destroy_sq(mdev, sqn);
1313 }
1314
1315 int mlx5e_create_sq_rdy(struct mlx5_core_dev *mdev,
1316                         struct mlx5e_sq_param *param,
1317                         struct mlx5e_create_sq_param *csp,
1318                         u16 qos_queue_group_id,
1319                         u32 *sqn)
1320 {
1321         struct mlx5e_modify_sq_param msp = {0};
1322         int err;
1323
1324         err = mlx5e_create_sq(mdev, param, csp, sqn);
1325         if (err)
1326                 return err;
1327
1328         msp.curr_state = MLX5_SQC_STATE_RST;
1329         msp.next_state = MLX5_SQC_STATE_RDY;
1330         if (qos_queue_group_id) {
1331                 msp.qos_update = true;
1332                 msp.qos_queue_group_id = qos_queue_group_id;
1333         }
1334         err = mlx5e_modify_sq(mdev, *sqn, &msp);
1335         if (err)
1336                 mlx5e_destroy_sq(mdev, *sqn);
1337
1338         return err;
1339 }
1340
1341 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1342                                 struct mlx5e_txqsq *sq, u32 rate);
1343
1344 int mlx5e_open_txqsq(struct mlx5e_channel *c, u32 tisn, int txq_ix,
1345                      struct mlx5e_params *params, struct mlx5e_sq_param *param,
1346                      struct mlx5e_txqsq *sq, int tc, u16 qos_queue_group_id, u16 qos_qid)
1347 {
1348         struct mlx5e_create_sq_param csp = {};
1349         u32 tx_rate;
1350         int err;
1351
1352         err = mlx5e_alloc_txqsq(c, txq_ix, params, param, sq, tc);
1353         if (err)
1354                 return err;
1355
1356         if (qos_queue_group_id)
1357                 sq->stats = c->priv->htb.qos_sq_stats[qos_qid];
1358         else
1359                 sq->stats = &c->priv->channel_stats[c->ix].sq[tc];
1360
1361         csp.tisn            = tisn;
1362         csp.tis_lst_sz      = 1;
1363         csp.cqn             = sq->cq.mcq.cqn;
1364         csp.wq_ctrl         = &sq->wq_ctrl;
1365         csp.min_inline_mode = sq->min_inline_mode;
1366         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, qos_queue_group_id, &sq->sqn);
1367         if (err)
1368                 goto err_free_txqsq;
1369
1370         tx_rate = c->priv->tx_rates[sq->txq_ix];
1371         if (tx_rate)
1372                 mlx5e_set_sq_maxrate(c->netdev, sq, tx_rate);
1373
1374         if (params->tx_dim_enabled)
1375                 sq->state |= BIT(MLX5E_SQ_STATE_AM);
1376
1377         return 0;
1378
1379 err_free_txqsq:
1380         mlx5e_free_txqsq(sq);
1381
1382         return err;
1383 }
1384
1385 void mlx5e_activate_txqsq(struct mlx5e_txqsq *sq)
1386 {
1387         sq->txq = netdev_get_tx_queue(sq->netdev, sq->txq_ix);
1388         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1389         netdev_tx_reset_queue(sq->txq);
1390         netif_tx_start_queue(sq->txq);
1391 }
1392
1393 void mlx5e_tx_disable_queue(struct netdev_queue *txq)
1394 {
1395         __netif_tx_lock_bh(txq);
1396         netif_tx_stop_queue(txq);
1397         __netif_tx_unlock_bh(txq);
1398 }
1399
1400 void mlx5e_deactivate_txqsq(struct mlx5e_txqsq *sq)
1401 {
1402         struct mlx5_wq_cyc *wq = &sq->wq;
1403
1404         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1405         synchronize_net(); /* Sync with NAPI to prevent netif_tx_wake_queue. */
1406
1407         mlx5e_tx_disable_queue(sq->txq);
1408
1409         /* last doorbell out, godspeed .. */
1410         if (mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1)) {
1411                 u16 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
1412                 struct mlx5e_tx_wqe *nop;
1413
1414                 sq->db.wqe_info[pi] = (struct mlx5e_tx_wqe_info) {
1415                         .num_wqebbs = 1,
1416                 };
1417
1418                 nop = mlx5e_post_nop(wq, sq->sqn, &sq->pc);
1419                 mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &nop->ctrl);
1420         }
1421 }
1422
1423 void mlx5e_close_txqsq(struct mlx5e_txqsq *sq)
1424 {
1425         struct mlx5_core_dev *mdev = sq->mdev;
1426         struct mlx5_rate_limit rl = {0};
1427
1428         cancel_work_sync(&sq->dim.work);
1429         cancel_work_sync(&sq->recover_work);
1430         mlx5e_destroy_sq(mdev, sq->sqn);
1431         if (sq->rate_limit) {
1432                 rl.rate = sq->rate_limit;
1433                 mlx5_rl_remove_rate(mdev, &rl);
1434         }
1435         mlx5e_free_txqsq_descs(sq);
1436         mlx5e_free_txqsq(sq);
1437 }
1438
1439 void mlx5e_tx_err_cqe_work(struct work_struct *recover_work)
1440 {
1441         struct mlx5e_txqsq *sq = container_of(recover_work, struct mlx5e_txqsq,
1442                                               recover_work);
1443
1444         mlx5e_reporter_tx_err_cqe(sq);
1445 }
1446
1447 int mlx5e_open_icosq(struct mlx5e_channel *c, struct mlx5e_params *params,
1448                      struct mlx5e_sq_param *param, struct mlx5e_icosq *sq)
1449 {
1450         struct mlx5e_create_sq_param csp = {};
1451         int err;
1452
1453         err = mlx5e_alloc_icosq(c, param, sq);
1454         if (err)
1455                 return err;
1456
1457         csp.cqn             = sq->cq.mcq.cqn;
1458         csp.wq_ctrl         = &sq->wq_ctrl;
1459         csp.min_inline_mode = params->tx_min_inline_mode;
1460         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, 0, &sq->sqn);
1461         if (err)
1462                 goto err_free_icosq;
1463
1464         return 0;
1465
1466 err_free_icosq:
1467         mlx5e_free_icosq(sq);
1468
1469         return err;
1470 }
1471
1472 void mlx5e_activate_icosq(struct mlx5e_icosq *icosq)
1473 {
1474         set_bit(MLX5E_SQ_STATE_ENABLED, &icosq->state);
1475 }
1476
1477 void mlx5e_deactivate_icosq(struct mlx5e_icosq *icosq)
1478 {
1479         clear_bit(MLX5E_SQ_STATE_ENABLED, &icosq->state);
1480         synchronize_net(); /* Sync with NAPI. */
1481 }
1482
1483 void mlx5e_close_icosq(struct mlx5e_icosq *sq)
1484 {
1485         struct mlx5e_channel *c = sq->channel;
1486
1487         mlx5e_destroy_sq(c->mdev, sq->sqn);
1488         mlx5e_free_icosq_descs(sq);
1489         mlx5e_free_icosq(sq);
1490 }
1491
1492 int mlx5e_open_xdpsq(struct mlx5e_channel *c, struct mlx5e_params *params,
1493                      struct mlx5e_sq_param *param, struct xsk_buff_pool *xsk_pool,
1494                      struct mlx5e_xdpsq *sq, bool is_redirect)
1495 {
1496         struct mlx5e_create_sq_param csp = {};
1497         int err;
1498
1499         err = mlx5e_alloc_xdpsq(c, params, xsk_pool, param, sq, is_redirect);
1500         if (err)
1501                 return err;
1502
1503         csp.tis_lst_sz      = 1;
1504         csp.tisn            = c->priv->tisn[c->lag_port][0]; /* tc = 0 */
1505         csp.cqn             = sq->cq.mcq.cqn;
1506         csp.wq_ctrl         = &sq->wq_ctrl;
1507         csp.min_inline_mode = sq->min_inline_mode;
1508         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1509         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, 0, &sq->sqn);
1510         if (err)
1511                 goto err_free_xdpsq;
1512
1513         mlx5e_set_xmit_fp(sq, param->is_mpw);
1514
1515         if (!param->is_mpw) {
1516                 unsigned int ds_cnt = MLX5E_XDP_TX_DS_COUNT;
1517                 unsigned int inline_hdr_sz = 0;
1518                 int i;
1519
1520                 if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
1521                         inline_hdr_sz = MLX5E_XDP_MIN_INLINE;
1522                         ds_cnt++;
1523                 }
1524
1525                 /* Pre initialize fixed WQE fields */
1526                 for (i = 0; i < mlx5_wq_cyc_get_size(&sq->wq); i++) {
1527                         struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(&sq->wq, i);
1528                         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
1529                         struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
1530                         struct mlx5_wqe_data_seg *dseg;
1531
1532                         sq->db.wqe_info[i] = (struct mlx5e_xdp_wqe_info) {
1533                                 .num_wqebbs = 1,
1534                                 .num_pkts   = 1,
1535                         };
1536
1537                         cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
1538                         eseg->inline_hdr.sz = cpu_to_be16(inline_hdr_sz);
1539
1540                         dseg = (struct mlx5_wqe_data_seg *)cseg + (ds_cnt - 1);
1541                         dseg->lkey = sq->mkey_be;
1542                 }
1543         }
1544
1545         return 0;
1546
1547 err_free_xdpsq:
1548         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1549         mlx5e_free_xdpsq(sq);
1550
1551         return err;
1552 }
1553
1554 void mlx5e_close_xdpsq(struct mlx5e_xdpsq *sq)
1555 {
1556         struct mlx5e_channel *c = sq->channel;
1557
1558         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1559         synchronize_net(); /* Sync with NAPI. */
1560
1561         mlx5e_destroy_sq(c->mdev, sq->sqn);
1562         mlx5e_free_xdpsq_descs(sq);
1563         mlx5e_free_xdpsq(sq);
1564 }
1565
1566 static int mlx5e_alloc_cq_common(struct mlx5e_priv *priv,
1567                                  struct mlx5e_cq_param *param,
1568                                  struct mlx5e_cq *cq)
1569 {
1570         struct mlx5_core_dev *mdev = priv->mdev;
1571         struct mlx5_core_cq *mcq = &cq->mcq;
1572         int eqn_not_used;
1573         unsigned int irqn;
1574         int err;
1575         u32 i;
1576
1577         err = mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1578         if (err)
1579                 return err;
1580
1581         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1582                                &cq->wq_ctrl);
1583         if (err)
1584                 return err;
1585
1586         mcq->cqe_sz     = 64;
1587         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1588         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1589         *mcq->set_ci_db = 0;
1590         *mcq->arm_db    = 0;
1591         mcq->vector     = param->eq_ix;
1592         mcq->comp       = mlx5e_completion_event;
1593         mcq->event      = mlx5e_cq_error_event;
1594         mcq->irqn       = irqn;
1595
1596         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1597                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1598
1599                 cqe->op_own = 0xf1;
1600         }
1601
1602         cq->mdev = mdev;
1603         cq->netdev = priv->netdev;
1604         cq->priv = priv;
1605
1606         return 0;
1607 }
1608
1609 static int mlx5e_alloc_cq(struct mlx5e_priv *priv,
1610                           struct mlx5e_cq_param *param,
1611                           struct mlx5e_create_cq_param *ccp,
1612                           struct mlx5e_cq *cq)
1613 {
1614         int err;
1615
1616         param->wq.buf_numa_node = ccp->node;
1617         param->wq.db_numa_node  = ccp->node;
1618         param->eq_ix            = ccp->ix;
1619
1620         err = mlx5e_alloc_cq_common(priv, param, cq);
1621
1622         cq->napi     = ccp->napi;
1623         cq->ch_stats = ccp->ch_stats;
1624
1625         return err;
1626 }
1627
1628 static void mlx5e_free_cq(struct mlx5e_cq *cq)
1629 {
1630         mlx5_wq_destroy(&cq->wq_ctrl);
1631 }
1632
1633 static int mlx5e_create_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
1634 {
1635         u32 out[MLX5_ST_SZ_DW(create_cq_out)];
1636         struct mlx5_core_dev *mdev = cq->mdev;
1637         struct mlx5_core_cq *mcq = &cq->mcq;
1638
1639         void *in;
1640         void *cqc;
1641         int inlen;
1642         unsigned int irqn_not_used;
1643         int eqn;
1644         int err;
1645
1646         err = mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
1647         if (err)
1648                 return err;
1649
1650         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1651                 sizeof(u64) * cq->wq_ctrl.buf.npages;
1652         in = kvzalloc(inlen, GFP_KERNEL);
1653         if (!in)
1654                 return -ENOMEM;
1655
1656         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1657
1658         memcpy(cqc, param->cqc, sizeof(param->cqc));
1659
1660         mlx5_fill_page_frag_array(&cq->wq_ctrl.buf,
1661                                   (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
1662
1663         MLX5_SET(cqc,   cqc, cq_period_mode, param->cq_period_mode);
1664         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
1665         MLX5_SET(cqc,   cqc, uar_page,      mdev->priv.uar->index);
1666         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
1667                                             MLX5_ADAPTER_PAGE_SHIFT);
1668         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
1669
1670         err = mlx5_core_create_cq(mdev, mcq, in, inlen, out, sizeof(out));
1671
1672         kvfree(in);
1673
1674         if (err)
1675                 return err;
1676
1677         mlx5e_cq_arm(cq);
1678
1679         return 0;
1680 }
1681
1682 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
1683 {
1684         mlx5_core_destroy_cq(cq->mdev, &cq->mcq);
1685 }
1686
1687 int mlx5e_open_cq(struct mlx5e_priv *priv, struct dim_cq_moder moder,
1688                   struct mlx5e_cq_param *param, struct mlx5e_create_cq_param *ccp,
1689                   struct mlx5e_cq *cq)
1690 {
1691         struct mlx5_core_dev *mdev = priv->mdev;
1692         int err;
1693
1694         err = mlx5e_alloc_cq(priv, param, ccp, cq);
1695         if (err)
1696                 return err;
1697
1698         err = mlx5e_create_cq(cq, param);
1699         if (err)
1700                 goto err_free_cq;
1701
1702         if (MLX5_CAP_GEN(mdev, cq_moderation))
1703                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq, moder.usec, moder.pkts);
1704         return 0;
1705
1706 err_free_cq:
1707         mlx5e_free_cq(cq);
1708
1709         return err;
1710 }
1711
1712 void mlx5e_close_cq(struct mlx5e_cq *cq)
1713 {
1714         mlx5e_destroy_cq(cq);
1715         mlx5e_free_cq(cq);
1716 }
1717
1718 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1719                              struct mlx5e_params *params,
1720                              struct mlx5e_create_cq_param *ccp,
1721                              struct mlx5e_channel_param *cparam)
1722 {
1723         int err;
1724         int tc;
1725
1726         for (tc = 0; tc < c->num_tc; tc++) {
1727                 err = mlx5e_open_cq(c->priv, params->tx_cq_moderation, &cparam->txq_sq.cqp,
1728                                     ccp, &c->sq[tc].cq);
1729                 if (err)
1730                         goto err_close_tx_cqs;
1731         }
1732
1733         return 0;
1734
1735 err_close_tx_cqs:
1736         for (tc--; tc >= 0; tc--)
1737                 mlx5e_close_cq(&c->sq[tc].cq);
1738
1739         return err;
1740 }
1741
1742 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1743 {
1744         int tc;
1745
1746         for (tc = 0; tc < c->num_tc; tc++)
1747                 mlx5e_close_cq(&c->sq[tc].cq);
1748 }
1749
1750 static int mlx5e_open_sqs(struct mlx5e_channel *c,
1751                           struct mlx5e_params *params,
1752                           struct mlx5e_channel_param *cparam)
1753 {
1754         int err, tc;
1755
1756         for (tc = 0; tc < params->num_tc; tc++) {
1757                 int txq_ix = c->ix + tc * params->num_channels;
1758
1759                 err = mlx5e_open_txqsq(c, c->priv->tisn[c->lag_port][tc], txq_ix,
1760                                        params, &cparam->txq_sq, &c->sq[tc], tc, 0, 0);
1761                 if (err)
1762                         goto err_close_sqs;
1763         }
1764
1765         return 0;
1766
1767 err_close_sqs:
1768         for (tc--; tc >= 0; tc--)
1769                 mlx5e_close_txqsq(&c->sq[tc]);
1770
1771         return err;
1772 }
1773
1774 static void mlx5e_close_sqs(struct mlx5e_channel *c)
1775 {
1776         int tc;
1777
1778         for (tc = 0; tc < c->num_tc; tc++)
1779                 mlx5e_close_txqsq(&c->sq[tc]);
1780 }
1781
1782 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1783                                 struct mlx5e_txqsq *sq, u32 rate)
1784 {
1785         struct mlx5e_priv *priv = netdev_priv(dev);
1786         struct mlx5_core_dev *mdev = priv->mdev;
1787         struct mlx5e_modify_sq_param msp = {0};
1788         struct mlx5_rate_limit rl = {0};
1789         u16 rl_index = 0;
1790         int err;
1791
1792         if (rate == sq->rate_limit)
1793                 /* nothing to do */
1794                 return 0;
1795
1796         if (sq->rate_limit) {
1797                 rl.rate = sq->rate_limit;
1798                 /* remove current rl index to free space to next ones */
1799                 mlx5_rl_remove_rate(mdev, &rl);
1800         }
1801
1802         sq->rate_limit = 0;
1803
1804         if (rate) {
1805                 rl.rate = rate;
1806                 err = mlx5_rl_add_rate(mdev, &rl_index, &rl);
1807                 if (err) {
1808                         netdev_err(dev, "Failed configuring rate %u: %d\n",
1809                                    rate, err);
1810                         return err;
1811                 }
1812         }
1813
1814         msp.curr_state = MLX5_SQC_STATE_RDY;
1815         msp.next_state = MLX5_SQC_STATE_RDY;
1816         msp.rl_index   = rl_index;
1817         msp.rl_update  = true;
1818         err = mlx5e_modify_sq(mdev, sq->sqn, &msp);
1819         if (err) {
1820                 netdev_err(dev, "Failed configuring rate %u: %d\n",
1821                            rate, err);
1822                 /* remove the rate from the table */
1823                 if (rate)
1824                         mlx5_rl_remove_rate(mdev, &rl);
1825                 return err;
1826         }
1827
1828         sq->rate_limit = rate;
1829         return 0;
1830 }
1831
1832 static int mlx5e_set_tx_maxrate(struct net_device *dev, int index, u32 rate)
1833 {
1834         struct mlx5e_priv *priv = netdev_priv(dev);
1835         struct mlx5_core_dev *mdev = priv->mdev;
1836         struct mlx5e_txqsq *sq = priv->txq2sq[index];
1837         int err = 0;
1838
1839         if (!mlx5_rl_is_supported(mdev)) {
1840                 netdev_err(dev, "Rate limiting is not supported on this device\n");
1841                 return -EINVAL;
1842         }
1843
1844         /* rate is given in Mb/sec, HW config is in Kb/sec */
1845         rate = rate << 10;
1846
1847         /* Check whether rate in valid range, 0 is always valid */
1848         if (rate && !mlx5_rl_is_in_range(mdev, rate)) {
1849                 netdev_err(dev, "TX rate %u, is not in range\n", rate);
1850                 return -ERANGE;
1851         }
1852
1853         mutex_lock(&priv->state_lock);
1854         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
1855                 err = mlx5e_set_sq_maxrate(dev, sq, rate);
1856         if (!err)
1857                 priv->tx_rates[index] = rate;
1858         mutex_unlock(&priv->state_lock);
1859
1860         return err;
1861 }
1862
1863 void mlx5e_build_create_cq_param(struct mlx5e_create_cq_param *ccp, struct mlx5e_channel *c)
1864 {
1865         *ccp = (struct mlx5e_create_cq_param) {
1866                 .napi = &c->napi,
1867                 .ch_stats = c->stats,
1868                 .node = cpu_to_node(c->cpu),
1869                 .ix = c->ix,
1870         };
1871 }
1872
1873 static int mlx5e_open_queues(struct mlx5e_channel *c,
1874                              struct mlx5e_params *params,
1875                              struct mlx5e_channel_param *cparam)
1876 {
1877         struct dim_cq_moder icocq_moder = {0, 0};
1878         struct mlx5e_create_cq_param ccp;
1879         int err;
1880
1881         mlx5e_build_create_cq_param(&ccp, c);
1882
1883         err = mlx5e_open_cq(c->priv, icocq_moder, &cparam->async_icosq.cqp, &ccp,
1884                             &c->async_icosq.cq);
1885         if (err)
1886                 return err;
1887
1888         err = mlx5e_open_cq(c->priv, icocq_moder, &cparam->icosq.cqp, &ccp,
1889                             &c->icosq.cq);
1890         if (err)
1891                 goto err_close_async_icosq_cq;
1892
1893         err = mlx5e_open_tx_cqs(c, params, &ccp, cparam);
1894         if (err)
1895                 goto err_close_icosq_cq;
1896
1897         err = mlx5e_open_cq(c->priv, params->tx_cq_moderation, &cparam->xdp_sq.cqp, &ccp,
1898                             &c->xdpsq.cq);
1899         if (err)
1900                 goto err_close_tx_cqs;
1901
1902         err = mlx5e_open_cq(c->priv, params->rx_cq_moderation, &cparam->rq.cqp, &ccp,
1903                             &c->rq.cq);
1904         if (err)
1905                 goto err_close_xdp_tx_cqs;
1906
1907         err = c->xdp ? mlx5e_open_cq(c->priv, params->tx_cq_moderation, &cparam->xdp_sq.cqp,
1908                                      &ccp, &c->rq_xdpsq.cq) : 0;
1909         if (err)
1910                 goto err_close_rx_cq;
1911
1912         spin_lock_init(&c->async_icosq_lock);
1913
1914         err = mlx5e_open_icosq(c, params, &cparam->async_icosq, &c->async_icosq);
1915         if (err)
1916                 goto err_close_xdpsq_cq;
1917
1918         err = mlx5e_open_icosq(c, params, &cparam->icosq, &c->icosq);
1919         if (err)
1920                 goto err_close_async_icosq;
1921
1922         err = mlx5e_open_sqs(c, params, cparam);
1923         if (err)
1924                 goto err_close_icosq;
1925
1926         if (c->xdp) {
1927                 err = mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, NULL,
1928                                        &c->rq_xdpsq, false);
1929                 if (err)
1930                         goto err_close_sqs;
1931         }
1932
1933         err = mlx5e_open_rq(c, params, &cparam->rq, NULL, NULL, &c->rq);
1934         if (err)
1935                 goto err_close_xdp_sq;
1936
1937         err = mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, NULL, &c->xdpsq, true);
1938         if (err)
1939                 goto err_close_rq;
1940
1941         return 0;
1942
1943 err_close_rq:
1944         mlx5e_close_rq(&c->rq);
1945
1946 err_close_xdp_sq:
1947         if (c->xdp)
1948                 mlx5e_close_xdpsq(&c->rq_xdpsq);
1949
1950 err_close_sqs:
1951         mlx5e_close_sqs(c);
1952
1953 err_close_icosq:
1954         mlx5e_close_icosq(&c->icosq);
1955
1956 err_close_async_icosq:
1957         mlx5e_close_icosq(&c->async_icosq);
1958
1959 err_close_xdpsq_cq:
1960         if (c->xdp)
1961                 mlx5e_close_cq(&c->rq_xdpsq.cq);
1962
1963 err_close_rx_cq:
1964         mlx5e_close_cq(&c->rq.cq);
1965
1966 err_close_xdp_tx_cqs:
1967         mlx5e_close_cq(&c->xdpsq.cq);
1968
1969 err_close_tx_cqs:
1970         mlx5e_close_tx_cqs(c);
1971
1972 err_close_icosq_cq:
1973         mlx5e_close_cq(&c->icosq.cq);
1974
1975 err_close_async_icosq_cq:
1976         mlx5e_close_cq(&c->async_icosq.cq);
1977
1978         return err;
1979 }
1980
1981 static void mlx5e_close_queues(struct mlx5e_channel *c)
1982 {
1983         mlx5e_close_xdpsq(&c->xdpsq);
1984         mlx5e_close_rq(&c->rq);
1985         if (c->xdp)
1986                 mlx5e_close_xdpsq(&c->rq_xdpsq);
1987         mlx5e_close_sqs(c);
1988         mlx5e_close_icosq(&c->icosq);
1989         mlx5e_close_icosq(&c->async_icosq);
1990         if (c->xdp)
1991                 mlx5e_close_cq(&c->rq_xdpsq.cq);
1992         mlx5e_close_cq(&c->rq.cq);
1993         mlx5e_close_cq(&c->xdpsq.cq);
1994         mlx5e_close_tx_cqs(c);
1995         mlx5e_close_cq(&c->icosq.cq);
1996         mlx5e_close_cq(&c->async_icosq.cq);
1997 }
1998
1999 static u8 mlx5e_enumerate_lag_port(struct mlx5_core_dev *mdev, int ix)
2000 {
2001         u16 port_aff_bias = mlx5_core_is_pf(mdev) ? 0 : MLX5_CAP_GEN(mdev, vhca_id);
2002
2003         return (ix + port_aff_bias) % mlx5e_get_num_lag_ports(mdev);
2004 }
2005
2006 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
2007                               struct mlx5e_params *params,
2008                               struct mlx5e_channel_param *cparam,
2009                               struct xsk_buff_pool *xsk_pool,
2010                               struct mlx5e_channel **cp)
2011 {
2012         int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(priv->mdev, ix));
2013         struct net_device *netdev = priv->netdev;
2014         struct mlx5e_xsk_param xsk;
2015         struct mlx5e_channel *c;
2016         unsigned int irq;
2017         int err;
2018         int eqn;
2019
2020         err = mlx5_vector2eqn(priv->mdev, ix, &eqn, &irq);
2021         if (err)
2022                 return err;
2023
2024         c = kvzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
2025         if (!c)
2026                 return -ENOMEM;
2027
2028         c->priv     = priv;
2029         c->mdev     = priv->mdev;
2030         c->tstamp   = &priv->tstamp;
2031         c->ix       = ix;
2032         c->cpu      = cpu;
2033         c->pdev     = mlx5_core_dma_dev(priv->mdev);
2034         c->netdev   = priv->netdev;
2035         c->mkey_be  = cpu_to_be32(priv->mdev->mlx5e_res.mkey.key);
2036         c->num_tc   = params->num_tc;
2037         c->xdp      = !!params->xdp_prog;
2038         c->stats    = &priv->channel_stats[ix].ch;
2039         c->aff_mask = irq_get_effective_affinity_mask(irq);
2040         c->lag_port = mlx5e_enumerate_lag_port(priv->mdev, ix);
2041
2042         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
2043
2044         err = mlx5e_open_queues(c, params, cparam);
2045         if (unlikely(err))
2046                 goto err_napi_del;
2047
2048         if (xsk_pool) {
2049                 mlx5e_build_xsk_param(xsk_pool, &xsk);
2050                 err = mlx5e_open_xsk(priv, params, &xsk, xsk_pool, c);
2051                 if (unlikely(err))
2052                         goto err_close_queues;
2053         }
2054
2055         *cp = c;
2056
2057         return 0;
2058
2059 err_close_queues:
2060         mlx5e_close_queues(c);
2061
2062 err_napi_del:
2063         netif_napi_del(&c->napi);
2064
2065         kvfree(c);
2066
2067         return err;
2068 }
2069
2070 static void mlx5e_activate_channel(struct mlx5e_channel *c)
2071 {
2072         int tc;
2073
2074         napi_enable(&c->napi);
2075
2076         for (tc = 0; tc < c->num_tc; tc++)
2077                 mlx5e_activate_txqsq(&c->sq[tc]);
2078         mlx5e_activate_icosq(&c->icosq);
2079         mlx5e_activate_icosq(&c->async_icosq);
2080         mlx5e_activate_rq(&c->rq);
2081
2082         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2083                 mlx5e_activate_xsk(c);
2084 }
2085
2086 static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
2087 {
2088         int tc;
2089
2090         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2091                 mlx5e_deactivate_xsk(c);
2092
2093         mlx5e_deactivate_rq(&c->rq);
2094         mlx5e_deactivate_icosq(&c->async_icosq);
2095         mlx5e_deactivate_icosq(&c->icosq);
2096         for (tc = 0; tc < c->num_tc; tc++)
2097                 mlx5e_deactivate_txqsq(&c->sq[tc]);
2098         mlx5e_qos_deactivate_queues(c);
2099
2100         napi_disable(&c->napi);
2101 }
2102
2103 static void mlx5e_close_channel(struct mlx5e_channel *c)
2104 {
2105         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2106                 mlx5e_close_xsk(c);
2107         mlx5e_close_queues(c);
2108         mlx5e_qos_close_queues(c);
2109         netif_napi_del(&c->napi);
2110
2111         kvfree(c);
2112 }
2113
2114 #define DEFAULT_FRAG_SIZE (2048)
2115
2116 static void mlx5e_build_rq_frags_info(struct mlx5_core_dev *mdev,
2117                                       struct mlx5e_params *params,
2118                                       struct mlx5e_xsk_param *xsk,
2119                                       struct mlx5e_rq_frags_info *info)
2120 {
2121         u32 byte_count = MLX5E_SW2HW_MTU(params, params->sw_mtu);
2122         int frag_size_max = DEFAULT_FRAG_SIZE;
2123         u32 buf_size = 0;
2124         int i;
2125
2126         if (mlx5_fpga_is_ipsec_device(mdev))
2127                 byte_count += MLX5E_METADATA_ETHER_LEN;
2128
2129         if (mlx5e_rx_is_linear_skb(params, xsk)) {
2130                 int frag_stride;
2131
2132                 frag_stride = mlx5e_rx_get_linear_frag_sz(params, xsk);
2133                 frag_stride = roundup_pow_of_two(frag_stride);
2134
2135                 info->arr[0].frag_size = byte_count;
2136                 info->arr[0].frag_stride = frag_stride;
2137                 info->num_frags = 1;
2138                 info->wqe_bulk = PAGE_SIZE / frag_stride;
2139                 goto out;
2140         }
2141
2142         if (byte_count > PAGE_SIZE +
2143             (MLX5E_MAX_RX_FRAGS - 1) * frag_size_max)
2144                 frag_size_max = PAGE_SIZE;
2145
2146         i = 0;
2147         while (buf_size < byte_count) {
2148                 int frag_size = byte_count - buf_size;
2149
2150                 if (i < MLX5E_MAX_RX_FRAGS - 1)
2151                         frag_size = min(frag_size, frag_size_max);
2152
2153                 info->arr[i].frag_size = frag_size;
2154                 info->arr[i].frag_stride = roundup_pow_of_two(frag_size);
2155
2156                 buf_size += frag_size;
2157                 i++;
2158         }
2159         info->num_frags = i;
2160         /* number of different wqes sharing a page */
2161         info->wqe_bulk = 1 + (info->num_frags % 2);
2162
2163 out:
2164         info->wqe_bulk = max_t(u8, info->wqe_bulk, 8);
2165         info->log_num_frags = order_base_2(info->num_frags);
2166 }
2167
2168 static inline u8 mlx5e_get_rqwq_log_stride(u8 wq_type, int ndsegs)
2169 {
2170         int sz = sizeof(struct mlx5_wqe_data_seg) * ndsegs;
2171
2172         switch (wq_type) {
2173         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2174                 sz += sizeof(struct mlx5e_rx_wqe_ll);
2175                 break;
2176         default: /* MLX5_WQ_TYPE_CYCLIC */
2177                 sz += sizeof(struct mlx5e_rx_wqe_cyc);
2178         }
2179
2180         return order_base_2(sz);
2181 }
2182
2183 static u8 mlx5e_get_rq_log_wq_sz(void *rqc)
2184 {
2185         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2186
2187         return MLX5_GET(wq, wq, log_wq_sz);
2188 }
2189
2190 void mlx5e_build_rq_param(struct mlx5e_priv *priv,
2191                           struct mlx5e_params *params,
2192                           struct mlx5e_xsk_param *xsk,
2193                           struct mlx5e_rq_param *param)
2194 {
2195         struct mlx5_core_dev *mdev = priv->mdev;
2196         void *rqc = param->rqc;
2197         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2198         int ndsegs = 1;
2199
2200         switch (params->rq_wq_type) {
2201         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2202                 MLX5_SET(wq, wq, log_wqe_num_of_strides,
2203                          mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk) -
2204                          MLX5_MPWQE_LOG_NUM_STRIDES_BASE);
2205                 MLX5_SET(wq, wq, log_wqe_stride_size,
2206                          mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk) -
2207                          MLX5_MPWQE_LOG_STRIDE_SZ_BASE);
2208                 MLX5_SET(wq, wq, log_wq_sz, mlx5e_mpwqe_get_log_rq_size(params, xsk));
2209                 break;
2210         default: /* MLX5_WQ_TYPE_CYCLIC */
2211                 MLX5_SET(wq, wq, log_wq_sz, params->log_rq_mtu_frames);
2212                 mlx5e_build_rq_frags_info(mdev, params, xsk, &param->frags_info);
2213                 ndsegs = param->frags_info.num_frags;
2214         }
2215
2216         MLX5_SET(wq, wq, wq_type,          params->rq_wq_type);
2217         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
2218         MLX5_SET(wq, wq, log_wq_stride,
2219                  mlx5e_get_rqwq_log_stride(params->rq_wq_type, ndsegs));
2220         MLX5_SET(wq, wq, pd,               mdev->mlx5e_res.pdn);
2221         MLX5_SET(rqc, rqc, counter_set_id, priv->q_counter);
2222         MLX5_SET(rqc, rqc, vsd,            params->vlan_strip_disable);
2223         MLX5_SET(rqc, rqc, scatter_fcs,    params->scatter_fcs_en);
2224
2225         param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
2226         mlx5e_build_rx_cq_param(priv, params, xsk, &param->cqp);
2227 }
2228
2229 static void mlx5e_build_drop_rq_param(struct mlx5e_priv *priv,
2230                                       struct mlx5e_rq_param *param)
2231 {
2232         struct mlx5_core_dev *mdev = priv->mdev;
2233         void *rqc = param->rqc;
2234         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2235
2236         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
2237         MLX5_SET(wq, wq, log_wq_stride,
2238                  mlx5e_get_rqwq_log_stride(MLX5_WQ_TYPE_CYCLIC, 1));
2239         MLX5_SET(rqc, rqc, counter_set_id, priv->drop_rq_q_counter);
2240
2241         param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
2242 }
2243
2244 void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
2245                                  struct mlx5e_sq_param *param)
2246 {
2247         void *sqc = param->sqc;
2248         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2249
2250         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
2251         MLX5_SET(wq, wq, pd,            priv->mdev->mlx5e_res.pdn);
2252
2253         param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(priv->mdev));
2254 }
2255
2256 void mlx5e_build_sq_param(struct mlx5e_priv *priv, struct mlx5e_params *params,
2257                           struct mlx5e_sq_param *param)
2258 {
2259         void *sqc = param->sqc;
2260         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2261         bool allow_swp;
2262
2263         allow_swp = mlx5_geneve_tx_allowed(priv->mdev) ||
2264                     !!MLX5_IPSEC_DEV(priv->mdev);
2265         mlx5e_build_sq_param_common(priv, param);
2266         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
2267         MLX5_SET(sqc, sqc, allow_swp, allow_swp);
2268         param->is_mpw = MLX5E_GET_PFLAG(params, MLX5E_PFLAG_SKB_TX_MPWQE);
2269         param->stop_room = mlx5e_calc_sq_stop_room(priv->mdev, params);
2270         mlx5e_build_tx_cq_param(priv, params, &param->cqp);
2271 }
2272
2273 static void mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
2274                                         struct mlx5e_cq_param *param)
2275 {
2276         void *cqc = param->cqc;
2277
2278         MLX5_SET(cqc, cqc, uar_page, priv->mdev->priv.uar->index);
2279         if (MLX5_CAP_GEN(priv->mdev, cqe_128_always) && cache_line_size() >= 128)
2280                 MLX5_SET(cqc, cqc, cqe_sz, CQE_STRIDE_128_PAD);
2281 }
2282
2283 void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
2284                              struct mlx5e_params *params,
2285                              struct mlx5e_xsk_param *xsk,
2286                              struct mlx5e_cq_param *param)
2287 {
2288         struct mlx5_core_dev *mdev = priv->mdev;
2289         bool hw_stridx = false;
2290         void *cqc = param->cqc;
2291         u8 log_cq_size;
2292
2293         switch (params->rq_wq_type) {
2294         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2295                 log_cq_size = mlx5e_mpwqe_get_log_rq_size(params, xsk) +
2296                         mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk);
2297                 hw_stridx = MLX5_CAP_GEN(mdev, mini_cqe_resp_stride_index);
2298                 break;
2299         default: /* MLX5_WQ_TYPE_CYCLIC */
2300                 log_cq_size = params->log_rq_mtu_frames;
2301         }
2302
2303         MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
2304         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
2305                 MLX5_SET(cqc, cqc, mini_cqe_res_format, hw_stridx ?
2306                          MLX5_CQE_FORMAT_CSUM_STRIDX : MLX5_CQE_FORMAT_CSUM);
2307                 MLX5_SET(cqc, cqc, cqe_comp_en, 1);
2308         }
2309
2310         mlx5e_build_common_cq_param(priv, param);
2311         param->cq_period_mode = params->rx_cq_moderation.cq_period_mode;
2312 }
2313
2314 void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
2315                              struct mlx5e_params *params,
2316                              struct mlx5e_cq_param *param)
2317 {
2318         void *cqc = param->cqc;
2319
2320         MLX5_SET(cqc, cqc, log_cq_size, params->log_sq_size);
2321
2322         mlx5e_build_common_cq_param(priv, param);
2323         param->cq_period_mode = params->tx_cq_moderation.cq_period_mode;
2324 }
2325
2326 void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
2327                               u8 log_wq_size,
2328                               struct mlx5e_cq_param *param)
2329 {
2330         void *cqc = param->cqc;
2331
2332         MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
2333
2334         mlx5e_build_common_cq_param(priv, param);
2335
2336         param->cq_period_mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
2337 }
2338
2339 void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
2340                              u8 log_wq_size,
2341                              struct mlx5e_sq_param *param)
2342 {
2343         void *sqc = param->sqc;
2344         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2345
2346         mlx5e_build_sq_param_common(priv, param);
2347
2348         MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
2349         MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(priv->mdev, reg_umr_sq));
2350         mlx5e_build_ico_cq_param(priv, log_wq_size, &param->cqp);
2351 }
2352
2353 void mlx5e_build_xdpsq_param(struct mlx5e_priv *priv,
2354                              struct mlx5e_params *params,
2355                              struct mlx5e_sq_param *param)
2356 {
2357         void *sqc = param->sqc;
2358         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2359
2360         mlx5e_build_sq_param_common(priv, param);
2361         MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
2362         param->is_mpw = MLX5E_GET_PFLAG(params, MLX5E_PFLAG_XDP_TX_MPWQE);
2363         mlx5e_build_tx_cq_param(priv, params, &param->cqp);
2364 }
2365
2366 static u8 mlx5e_build_icosq_log_wq_sz(struct mlx5e_params *params,
2367                                       struct mlx5e_rq_param *rqp)
2368 {
2369         switch (params->rq_wq_type) {
2370         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
2371                 return order_base_2(MLX5E_UMR_WQEBBS) +
2372                         mlx5e_get_rq_log_wq_sz(rqp->rqc);
2373         default: /* MLX5_WQ_TYPE_CYCLIC */
2374                 return MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
2375         }
2376 }
2377
2378 static u8 mlx5e_build_async_icosq_log_wq_sz(struct net_device *netdev)
2379 {
2380         if (netdev->hw_features & NETIF_F_HW_TLS_RX)
2381                 return MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
2382
2383         return MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
2384 }
2385
2386 static void mlx5e_build_channel_param(struct mlx5e_priv *priv,
2387                                       struct mlx5e_params *params,
2388                                       struct mlx5e_channel_param *cparam)
2389 {
2390         u8 icosq_log_wq_sz, async_icosq_log_wq_sz;
2391
2392         mlx5e_build_rq_param(priv, params, NULL, &cparam->rq);
2393
2394         icosq_log_wq_sz = mlx5e_build_icosq_log_wq_sz(params, &cparam->rq);
2395         async_icosq_log_wq_sz = mlx5e_build_async_icosq_log_wq_sz(priv->netdev);
2396
2397         mlx5e_build_sq_param(priv, params, &cparam->txq_sq);
2398         mlx5e_build_xdpsq_param(priv, params, &cparam->xdp_sq);
2399         mlx5e_build_icosq_param(priv, icosq_log_wq_sz, &cparam->icosq);
2400         mlx5e_build_icosq_param(priv, async_icosq_log_wq_sz, &cparam->async_icosq);
2401 }
2402
2403 int mlx5e_open_channels(struct mlx5e_priv *priv,
2404                         struct mlx5e_channels *chs)
2405 {
2406         struct mlx5e_channel_param *cparam;
2407         int err = -ENOMEM;
2408         int i;
2409
2410         chs->num = chs->params.num_channels;
2411
2412         chs->c = kcalloc(chs->num, sizeof(struct mlx5e_channel *), GFP_KERNEL);
2413         cparam = kvzalloc(sizeof(struct mlx5e_channel_param), GFP_KERNEL);
2414         if (!chs->c || !cparam)
2415                 goto err_free;
2416
2417         mlx5e_build_channel_param(priv, &chs->params, cparam);
2418         for (i = 0; i < chs->num; i++) {
2419                 struct xsk_buff_pool *xsk_pool = NULL;
2420
2421                 if (chs->params.xdp_prog)
2422                         xsk_pool = mlx5e_xsk_get_pool(&chs->params, chs->params.xsk, i);
2423
2424                 err = mlx5e_open_channel(priv, i, &chs->params, cparam, xsk_pool, &chs->c[i]);
2425                 if (err)
2426                         goto err_close_channels;
2427         }
2428
2429         if (MLX5E_GET_PFLAG(&chs->params, MLX5E_PFLAG_TX_PORT_TS)) {
2430                 err = mlx5e_port_ptp_open(priv, &chs->params, chs->c[0]->lag_port,
2431                                           &chs->port_ptp);
2432                 if (err)
2433                         goto err_close_channels;
2434         }
2435
2436         err = mlx5e_qos_open_queues(priv, chs);
2437         if (err)
2438                 goto err_close_ptp;
2439
2440         mlx5e_health_channels_update(priv);
2441         kvfree(cparam);
2442         return 0;
2443
2444 err_close_ptp:
2445         if (chs->port_ptp)
2446                 mlx5e_port_ptp_close(chs->port_ptp);
2447
2448 err_close_channels:
2449         for (i--; i >= 0; i--)
2450                 mlx5e_close_channel(chs->c[i]);
2451
2452 err_free:
2453         kfree(chs->c);
2454         kvfree(cparam);
2455         chs->num = 0;
2456         return err;
2457 }
2458
2459 static void mlx5e_activate_channels(struct mlx5e_channels *chs)
2460 {
2461         int i;
2462
2463         for (i = 0; i < chs->num; i++)
2464                 mlx5e_activate_channel(chs->c[i]);
2465
2466         if (chs->port_ptp)
2467                 mlx5e_ptp_activate_channel(chs->port_ptp);
2468 }
2469
2470 #define MLX5E_RQ_WQES_TIMEOUT 20000 /* msecs */
2471
2472 static int mlx5e_wait_channels_min_rx_wqes(struct mlx5e_channels *chs)
2473 {
2474         int err = 0;
2475         int i;
2476
2477         for (i = 0; i < chs->num; i++) {
2478                 int timeout = err ? 0 : MLX5E_RQ_WQES_TIMEOUT;
2479
2480                 err |= mlx5e_wait_for_min_rx_wqes(&chs->c[i]->rq, timeout);
2481
2482                 /* Don't wait on the XSK RQ, because the newer xdpsock sample
2483                  * doesn't provide any Fill Ring entries at the setup stage.
2484                  */
2485         }
2486
2487         return err ? -ETIMEDOUT : 0;
2488 }
2489
2490 static void mlx5e_deactivate_channels(struct mlx5e_channels *chs)
2491 {
2492         int i;
2493
2494         if (chs->port_ptp)
2495                 mlx5e_ptp_deactivate_channel(chs->port_ptp);
2496
2497         for (i = 0; i < chs->num; i++)
2498                 mlx5e_deactivate_channel(chs->c[i]);
2499 }
2500
2501 void mlx5e_close_channels(struct mlx5e_channels *chs)
2502 {
2503         int i;
2504
2505         if (chs->port_ptp)
2506                 mlx5e_port_ptp_close(chs->port_ptp);
2507
2508         for (i = 0; i < chs->num; i++)
2509                 mlx5e_close_channel(chs->c[i]);
2510
2511         kfree(chs->c);
2512         chs->num = 0;
2513 }
2514
2515 static int
2516 mlx5e_create_rqt(struct mlx5e_priv *priv, int sz, struct mlx5e_rqt *rqt)
2517 {
2518         struct mlx5_core_dev *mdev = priv->mdev;
2519         void *rqtc;
2520         int inlen;
2521         int err;
2522         u32 *in;
2523         int i;
2524
2525         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2526         in = kvzalloc(inlen, GFP_KERNEL);
2527         if (!in)
2528                 return -ENOMEM;
2529
2530         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2531
2532         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2533         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2534
2535         for (i = 0; i < sz; i++)
2536                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->drop_rq.rqn);
2537
2538         err = mlx5_core_create_rqt(mdev, in, inlen, &rqt->rqtn);
2539         if (!err)
2540                 rqt->enabled = true;
2541
2542         kvfree(in);
2543         return err;
2544 }
2545
2546 void mlx5e_destroy_rqt(struct mlx5e_priv *priv, struct mlx5e_rqt *rqt)
2547 {
2548         rqt->enabled = false;
2549         mlx5_core_destroy_rqt(priv->mdev, rqt->rqtn);
2550 }
2551
2552 int mlx5e_create_indirect_rqt(struct mlx5e_priv *priv)
2553 {
2554         struct mlx5e_rqt *rqt = &priv->indir_rqt;
2555         int err;
2556
2557         err = mlx5e_create_rqt(priv, MLX5E_INDIR_RQT_SIZE, rqt);
2558         if (err)
2559                 mlx5_core_warn(priv->mdev, "create indirect rqts failed, %d\n", err);
2560         return err;
2561 }
2562
2563 int mlx5e_create_direct_rqts(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
2564 {
2565         int err;
2566         int ix;
2567
2568         for (ix = 0; ix < priv->max_nch; ix++) {
2569                 err = mlx5e_create_rqt(priv, 1 /*size */, &tirs[ix].rqt);
2570                 if (unlikely(err))
2571                         goto err_destroy_rqts;
2572         }
2573
2574         return 0;
2575
2576 err_destroy_rqts:
2577         mlx5_core_warn(priv->mdev, "create rqts failed, %d\n", err);
2578         for (ix--; ix >= 0; ix--)
2579                 mlx5e_destroy_rqt(priv, &tirs[ix].rqt);
2580
2581         return err;
2582 }
2583
2584 void mlx5e_destroy_direct_rqts(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
2585 {
2586         int i;
2587
2588         for (i = 0; i < priv->max_nch; i++)
2589                 mlx5e_destroy_rqt(priv, &tirs[i].rqt);
2590 }
2591
2592 static int mlx5e_rx_hash_fn(int hfunc)
2593 {
2594         return (hfunc == ETH_RSS_HASH_TOP) ?
2595                MLX5_RX_HASH_FN_TOEPLITZ :
2596                MLX5_RX_HASH_FN_INVERTED_XOR8;
2597 }
2598
2599 int mlx5e_bits_invert(unsigned long a, int size)
2600 {
2601         int inv = 0;
2602         int i;
2603
2604         for (i = 0; i < size; i++)
2605                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
2606
2607         return inv;
2608 }
2609
2610 static void mlx5e_fill_rqt_rqns(struct mlx5e_priv *priv, int sz,
2611                                 struct mlx5e_redirect_rqt_param rrp, void *rqtc)
2612 {
2613         int i;
2614
2615         for (i = 0; i < sz; i++) {
2616                 u32 rqn;
2617
2618                 if (rrp.is_rss) {
2619                         int ix = i;
2620
2621                         if (rrp.rss.hfunc == ETH_RSS_HASH_XOR)
2622                                 ix = mlx5e_bits_invert(i, ilog2(sz));
2623
2624                         ix = priv->rss_params.indirection_rqt[ix];
2625                         rqn = rrp.rss.channels->c[ix]->rq.rqn;
2626                 } else {
2627                         rqn = rrp.rqn;
2628                 }
2629                 MLX5_SET(rqtc, rqtc, rq_num[i], rqn);
2630         }
2631 }
2632
2633 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, u32 rqtn, int sz,
2634                        struct mlx5e_redirect_rqt_param rrp)
2635 {
2636         struct mlx5_core_dev *mdev = priv->mdev;
2637         void *rqtc;
2638         int inlen;
2639         u32 *in;
2640         int err;
2641
2642         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
2643         in = kvzalloc(inlen, GFP_KERNEL);
2644         if (!in)
2645                 return -ENOMEM;
2646
2647         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
2648
2649         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2650         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
2651         mlx5e_fill_rqt_rqns(priv, sz, rrp, rqtc);
2652         err = mlx5_core_modify_rqt(mdev, rqtn, in, inlen);
2653
2654         kvfree(in);
2655         return err;
2656 }
2657
2658 static u32 mlx5e_get_direct_rqn(struct mlx5e_priv *priv, int ix,
2659                                 struct mlx5e_redirect_rqt_param rrp)
2660 {
2661         if (!rrp.is_rss)
2662                 return rrp.rqn;
2663
2664         if (ix >= rrp.rss.channels->num)
2665                 return priv->drop_rq.rqn;
2666
2667         return rrp.rss.channels->c[ix]->rq.rqn;
2668 }
2669
2670 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv,
2671                                 struct mlx5e_redirect_rqt_param rrp)
2672 {
2673         u32 rqtn;
2674         int ix;
2675
2676         if (priv->indir_rqt.enabled) {
2677                 /* RSS RQ table */
2678                 rqtn = priv->indir_rqt.rqtn;
2679                 mlx5e_redirect_rqt(priv, rqtn, MLX5E_INDIR_RQT_SIZE, rrp);
2680         }
2681
2682         for (ix = 0; ix < priv->max_nch; ix++) {
2683                 struct mlx5e_redirect_rqt_param direct_rrp = {
2684                         .is_rss = false,
2685                         {
2686                                 .rqn    = mlx5e_get_direct_rqn(priv, ix, rrp)
2687                         },
2688                 };
2689
2690                 /* Direct RQ Tables */
2691                 if (!priv->direct_tir[ix].rqt.enabled)
2692                         continue;
2693
2694                 rqtn = priv->direct_tir[ix].rqt.rqtn;
2695                 mlx5e_redirect_rqt(priv, rqtn, 1, direct_rrp);
2696         }
2697 }
2698
2699 static void mlx5e_redirect_rqts_to_channels(struct mlx5e_priv *priv,
2700                                             struct mlx5e_channels *chs)
2701 {
2702         struct mlx5e_redirect_rqt_param rrp = {
2703                 .is_rss        = true,
2704                 {
2705                         .rss = {
2706                                 .channels  = chs,
2707                                 .hfunc     = priv->rss_params.hfunc,
2708                         }
2709                 },
2710         };
2711
2712         mlx5e_redirect_rqts(priv, rrp);
2713 }
2714
2715 static void mlx5e_redirect_rqts_to_drop(struct mlx5e_priv *priv)
2716 {
2717         struct mlx5e_redirect_rqt_param drop_rrp = {
2718                 .is_rss = false,
2719                 {
2720                         .rqn = priv->drop_rq.rqn,
2721                 },
2722         };
2723
2724         mlx5e_redirect_rqts(priv, drop_rrp);
2725 }
2726
2727 static const struct mlx5e_tirc_config tirc_default_config[MLX5E_NUM_INDIR_TIRS] = {
2728         [MLX5E_TT_IPV4_TCP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2729                                 .l4_prot_type = MLX5_L4_PROT_TYPE_TCP,
2730                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2731         },
2732         [MLX5E_TT_IPV6_TCP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2733                                 .l4_prot_type = MLX5_L4_PROT_TYPE_TCP,
2734                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2735         },
2736         [MLX5E_TT_IPV4_UDP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2737                                 .l4_prot_type = MLX5_L4_PROT_TYPE_UDP,
2738                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2739         },
2740         [MLX5E_TT_IPV6_UDP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2741                                 .l4_prot_type = MLX5_L4_PROT_TYPE_UDP,
2742                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2743         },
2744         [MLX5E_TT_IPV4_IPSEC_AH] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2745                                      .l4_prot_type = 0,
2746                                      .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2747         },
2748         [MLX5E_TT_IPV6_IPSEC_AH] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2749                                      .l4_prot_type = 0,
2750                                      .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2751         },
2752         [MLX5E_TT_IPV4_IPSEC_ESP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2753                                       .l4_prot_type = 0,
2754                                       .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2755         },
2756         [MLX5E_TT_IPV6_IPSEC_ESP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2757                                       .l4_prot_type = 0,
2758                                       .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2759         },
2760         [MLX5E_TT_IPV4] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2761                             .l4_prot_type = 0,
2762                             .rx_hash_fields = MLX5_HASH_IP,
2763         },
2764         [MLX5E_TT_IPV6] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2765                             .l4_prot_type = 0,
2766                             .rx_hash_fields = MLX5_HASH_IP,
2767         },
2768 };
2769
2770 struct mlx5e_tirc_config mlx5e_tirc_get_default_config(enum mlx5e_traffic_types tt)
2771 {
2772         return tirc_default_config[tt];
2773 }
2774
2775 static void mlx5e_build_tir_ctx_lro(struct mlx5e_params *params, void *tirc)
2776 {
2777         if (!params->lro_en)
2778                 return;
2779
2780 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2781
2782         MLX5_SET(tirc, tirc, lro_enable_mask,
2783                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2784                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2785         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
2786                  (MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ - ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2787         MLX5_SET(tirc, tirc, lro_timeout_period_usecs, params->lro_timeout);
2788 }
2789
2790 void mlx5e_build_indir_tir_ctx_hash(struct mlx5e_rss_params *rss_params,
2791                                     const struct mlx5e_tirc_config *ttconfig,
2792                                     void *tirc, bool inner)
2793 {
2794         void *hfso = inner ? MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner) :
2795                              MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2796
2797         MLX5_SET(tirc, tirc, rx_hash_fn, mlx5e_rx_hash_fn(rss_params->hfunc));
2798         if (rss_params->hfunc == ETH_RSS_HASH_TOP) {
2799                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
2800                                              rx_hash_toeplitz_key);
2801                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
2802                                                rx_hash_toeplitz_key);
2803
2804                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2805                 memcpy(rss_key, rss_params->toeplitz_hash_key, len);
2806         }
2807         MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2808                  ttconfig->l3_prot_type);
2809         MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2810                  ttconfig->l4_prot_type);
2811         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2812                  ttconfig->rx_hash_fields);
2813 }
2814
2815 static void mlx5e_update_rx_hash_fields(struct mlx5e_tirc_config *ttconfig,
2816                                         enum mlx5e_traffic_types tt,
2817                                         u32 rx_hash_fields)
2818 {
2819         *ttconfig                = tirc_default_config[tt];
2820         ttconfig->rx_hash_fields = rx_hash_fields;
2821 }
2822
2823 void mlx5e_modify_tirs_hash(struct mlx5e_priv *priv, void *in)
2824 {
2825         void *tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2826         struct mlx5e_rss_params *rss = &priv->rss_params;
2827         struct mlx5_core_dev *mdev = priv->mdev;
2828         int ctxlen = MLX5_ST_SZ_BYTES(tirc);
2829         struct mlx5e_tirc_config ttconfig;
2830         int tt;
2831
2832         MLX5_SET(modify_tir_in, in, bitmask.hash, 1);
2833
2834         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2835                 memset(tirc, 0, ctxlen);
2836                 mlx5e_update_rx_hash_fields(&ttconfig, tt,
2837                                             rss->rx_hash_fields[tt]);
2838                 mlx5e_build_indir_tir_ctx_hash(rss, &ttconfig, tirc, false);
2839                 mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in);
2840         }
2841
2842         /* Verify inner tirs resources allocated */
2843         if (!priv->inner_indir_tir[0].tirn)
2844                 return;
2845
2846         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2847                 memset(tirc, 0, ctxlen);
2848                 mlx5e_update_rx_hash_fields(&ttconfig, tt,
2849                                             rss->rx_hash_fields[tt]);
2850                 mlx5e_build_indir_tir_ctx_hash(rss, &ttconfig, tirc, true);
2851                 mlx5_core_modify_tir(mdev, priv->inner_indir_tir[tt].tirn, in);
2852         }
2853 }
2854
2855 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
2856 {
2857         struct mlx5_core_dev *mdev = priv->mdev;
2858
2859         void *in;
2860         void *tirc;
2861         int inlen;
2862         int err;
2863         int tt;
2864         int ix;
2865
2866         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
2867         in = kvzalloc(inlen, GFP_KERNEL);
2868         if (!in)
2869                 return -ENOMEM;
2870
2871         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
2872         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2873
2874         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2875
2876         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2877                 err = mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in);
2878                 if (err)
2879                         goto free_in;
2880         }
2881
2882         for (ix = 0; ix < priv->max_nch; ix++) {
2883                 err = mlx5_core_modify_tir(mdev, priv->direct_tir[ix].tirn, in);
2884                 if (err)
2885                         goto free_in;
2886         }
2887
2888 free_in:
2889         kvfree(in);
2890
2891         return err;
2892 }
2893
2894 static MLX5E_DEFINE_PREACTIVATE_WRAPPER_CTX(mlx5e_modify_tirs_lro);
2895
2896 static int mlx5e_set_mtu(struct mlx5_core_dev *mdev,
2897                          struct mlx5e_params *params, u16 mtu)
2898 {
2899         u16 hw_mtu = MLX5E_SW2HW_MTU(params, mtu);
2900         int err;
2901
2902         err = mlx5_set_port_mtu(mdev, hw_mtu, 1);
2903         if (err)
2904                 return err;
2905
2906         /* Update vport context MTU */
2907         mlx5_modify_nic_vport_mtu(mdev, hw_mtu);
2908         return 0;
2909 }
2910
2911 static void mlx5e_query_mtu(struct mlx5_core_dev *mdev,
2912                             struct mlx5e_params *params, u16 *mtu)
2913 {
2914         u16 hw_mtu = 0;
2915         int err;
2916
2917         err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2918         if (err || !hw_mtu) /* fallback to port oper mtu */
2919                 mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
2920
2921         *mtu = MLX5E_HW2SW_MTU(params, hw_mtu);
2922 }
2923
2924 int mlx5e_set_dev_port_mtu(struct mlx5e_priv *priv)
2925 {
2926         struct mlx5e_params *params = &priv->channels.params;
2927         struct net_device *netdev = priv->netdev;
2928         struct mlx5_core_dev *mdev = priv->mdev;
2929         u16 mtu;
2930         int err;
2931
2932         err = mlx5e_set_mtu(mdev, params, params->sw_mtu);
2933         if (err)
2934                 return err;
2935
2936         mlx5e_query_mtu(mdev, params, &mtu);
2937         if (mtu != params->sw_mtu)
2938                 netdev_warn(netdev, "%s: VPort MTU %d is different than netdev mtu %d\n",
2939                             __func__, mtu, params->sw_mtu);
2940
2941         params->sw_mtu = mtu;
2942         return 0;
2943 }
2944
2945 MLX5E_DEFINE_PREACTIVATE_WRAPPER_CTX(mlx5e_set_dev_port_mtu);
2946
2947 void mlx5e_set_netdev_mtu_boundaries(struct mlx5e_priv *priv)
2948 {
2949         struct mlx5e_params *params = &priv->channels.params;
2950         struct net_device *netdev   = priv->netdev;
2951         struct mlx5_core_dev *mdev  = priv->mdev;
2952         u16 max_mtu;
2953
2954         /* MTU range: 68 - hw-specific max */
2955         netdev->min_mtu = ETH_MIN_MTU;
2956
2957         mlx5_query_port_max_mtu(mdev, &max_mtu, 1);
2958         netdev->max_mtu = min_t(unsigned int, MLX5E_HW2SW_MTU(params, max_mtu),
2959                                 ETH_MAX_MTU);
2960 }
2961
2962 static void mlx5e_netdev_set_tcs(struct net_device *netdev, u16 nch, u8 ntc)
2963 {
2964         int tc;
2965
2966         netdev_reset_tc(netdev);
2967
2968         if (ntc == 1)
2969                 return;
2970
2971         netdev_set_num_tc(netdev, ntc);
2972
2973         /* Map netdev TCs to offset 0
2974          * We have our own UP to TXQ mapping for QoS
2975          */
2976         for (tc = 0; tc < ntc; tc++)
2977                 netdev_set_tc_queue(netdev, tc, nch, 0);
2978 }
2979
2980 int mlx5e_update_tx_netdev_queues(struct mlx5e_priv *priv)
2981 {
2982         int qos_queues, nch, ntc, num_txqs, err;
2983
2984         qos_queues = mlx5e_qos_cur_leaf_nodes(priv);
2985
2986         nch = priv->channels.params.num_channels;
2987         ntc = priv->channels.params.num_tc;
2988         num_txqs = nch * ntc + qos_queues;
2989         if (MLX5E_GET_PFLAG(&priv->channels.params, MLX5E_PFLAG_TX_PORT_TS))
2990                 num_txqs += ntc;
2991
2992         mlx5e_dbg(DRV, priv, "Setting num_txqs %d\n", num_txqs);
2993         err = netif_set_real_num_tx_queues(priv->netdev, num_txqs);
2994         if (err)
2995                 netdev_warn(priv->netdev, "netif_set_real_num_tx_queues failed, %d\n", err);
2996
2997         return err;
2998 }
2999
3000 static int mlx5e_update_netdev_queues(struct mlx5e_priv *priv)
3001 {
3002         struct net_device *netdev = priv->netdev;
3003         int old_num_txqs, old_ntc;
3004         int num_rxqs, nch, ntc;
3005         int err;
3006
3007         old_num_txqs = netdev->real_num_tx_queues;
3008         old_ntc = netdev->num_tc;
3009
3010         nch = priv->channels.params.num_channels;
3011         ntc = priv->channels.params.num_tc;
3012         num_rxqs = nch * priv->profile->rq_groups;
3013
3014         mlx5e_netdev_set_tcs(netdev, nch, ntc);
3015
3016         err = mlx5e_update_tx_netdev_queues(priv);
3017         if (err)
3018                 goto err_tcs;
3019         err = netif_set_real_num_rx_queues(netdev, num_rxqs);
3020         if (err) {
3021                 netdev_warn(netdev, "netif_set_real_num_rx_queues failed, %d\n", err);
3022                 goto err_txqs;
3023         }
3024
3025         return 0;
3026
3027 err_txqs:
3028         /* netif_set_real_num_rx_queues could fail only when nch increased. Only
3029          * one of nch and ntc is changed in this function. That means, the call
3030          * to netif_set_real_num_tx_queues below should not fail, because it
3031          * decreases the number of TX queues.
3032          */
3033         WARN_ON_ONCE(netif_set_real_num_tx_queues(netdev, old_num_txqs));
3034
3035 err_tcs:
3036         mlx5e_netdev_set_tcs(netdev, old_num_txqs / old_ntc, old_ntc);
3037         return err;
3038 }
3039
3040 static void mlx5e_set_default_xps_cpumasks(struct mlx5e_priv *priv,
3041                                            struct mlx5e_params *params)
3042 {
3043         struct mlx5_core_dev *mdev = priv->mdev;
3044         int num_comp_vectors, ix, irq;
3045
3046         num_comp_vectors = mlx5_comp_vectors_count(mdev);
3047
3048         for (ix = 0; ix < params->num_channels; ix++) {
3049                 cpumask_clear(priv->scratchpad.cpumask);
3050
3051                 for (irq = ix; irq < num_comp_vectors; irq += params->num_channels) {
3052                         int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(mdev, irq));
3053
3054                         cpumask_set_cpu(cpu, priv->scratchpad.cpumask);
3055                 }
3056
3057                 netif_set_xps_queue(priv->netdev, priv->scratchpad.cpumask, ix);
3058         }
3059 }
3060
3061 int mlx5e_num_channels_changed(struct mlx5e_priv *priv)
3062 {
3063         u16 count = priv->channels.params.num_channels;
3064         int err;
3065
3066         err = mlx5e_update_netdev_queues(priv);
3067         if (err)
3068                 return err;
3069
3070         mlx5e_set_default_xps_cpumasks(priv, &priv->channels.params);
3071
3072         if (!netif_is_rxfh_configured(priv->netdev))
3073                 mlx5e_build_default_indir_rqt(priv->rss_params.indirection_rqt,
3074                                               MLX5E_INDIR_RQT_SIZE, count);
3075
3076         return 0;
3077 }
3078
3079 MLX5E_DEFINE_PREACTIVATE_WRAPPER_CTX(mlx5e_num_channels_changed);
3080
3081 static void mlx5e_build_txq_maps(struct mlx5e_priv *priv)
3082 {
3083         int i, ch, tc, num_tc;
3084
3085         ch = priv->channels.num;
3086         num_tc = priv->channels.params.num_tc;
3087
3088         for (i = 0; i < ch; i++) {
3089                 for (tc = 0; tc < num_tc; tc++) {
3090                         struct mlx5e_channel *c = priv->channels.c[i];
3091                         struct mlx5e_txqsq *sq = &c->sq[tc];
3092
3093                         priv->txq2sq[sq->txq_ix] = sq;
3094                         priv->channel_tc2realtxq[i][tc] = i + tc * ch;
3095                 }
3096         }
3097
3098         if (!priv->channels.port_ptp)
3099                 return;
3100
3101         for (tc = 0; tc < num_tc; tc++) {
3102                 struct mlx5e_port_ptp *c = priv->channels.port_ptp;
3103                 struct mlx5e_txqsq *sq = &c->ptpsq[tc].txqsq;
3104
3105                 priv->txq2sq[sq->txq_ix] = sq;
3106                 priv->port_ptp_tc2realtxq[tc] = priv->num_tc_x_num_ch + tc;
3107         }
3108 }
3109
3110 static void mlx5e_update_num_tc_x_num_ch(struct mlx5e_priv *priv)
3111 {
3112         /* Sync with mlx5e_select_queue. */
3113         WRITE_ONCE(priv->num_tc_x_num_ch,
3114                    priv->channels.params.num_tc * priv->channels.num);
3115 }
3116
3117 void mlx5e_activate_priv_channels(struct mlx5e_priv *priv)
3118 {
3119         mlx5e_update_num_tc_x_num_ch(priv);
3120         mlx5e_build_txq_maps(priv);
3121         mlx5e_activate_channels(&priv->channels);
3122         mlx5e_qos_activate_queues(priv);
3123         mlx5e_xdp_tx_enable(priv);
3124         netif_tx_start_all_queues(priv->netdev);
3125
3126         if (mlx5e_is_vport_rep(priv))
3127                 mlx5e_add_sqs_fwd_rules(priv);
3128
3129         mlx5e_wait_channels_min_rx_wqes(&priv->channels);
3130         mlx5e_redirect_rqts_to_channels(priv, &priv->channels);
3131
3132         mlx5e_xsk_redirect_rqts_to_channels(priv, &priv->channels);
3133 }
3134
3135 void mlx5e_deactivate_priv_channels(struct mlx5e_priv *priv)
3136 {
3137         mlx5e_xsk_redirect_rqts_to_drop(priv, &priv->channels);
3138
3139         mlx5e_redirect_rqts_to_drop(priv);
3140
3141         if (mlx5e_is_vport_rep(priv))
3142                 mlx5e_remove_sqs_fwd_rules(priv);
3143
3144         /* FIXME: This is a W/A only for tx timeout watch dog false alarm when
3145          * polling for inactive tx queues.
3146          */
3147         netif_tx_stop_all_queues(priv->netdev);
3148         netif_tx_disable(priv->netdev);
3149         mlx5e_xdp_tx_disable(priv);
3150         mlx5e_deactivate_channels(&priv->channels);
3151 }
3152
3153 static int mlx5e_switch_priv_channels(struct mlx5e_priv *priv,
3154                                       struct mlx5e_channels *new_chs,
3155                                       mlx5e_fp_preactivate preactivate,
3156                                       void *context)
3157 {
3158         struct net_device *netdev = priv->netdev;
3159         struct mlx5e_channels old_chs;
3160         int carrier_ok;
3161         int err = 0;
3162
3163         carrier_ok = netif_carrier_ok(netdev);
3164         netif_carrier_off(netdev);
3165
3166         mlx5e_deactivate_priv_channels(priv);
3167
3168         old_chs = priv->channels;
3169         priv->channels = *new_chs;
3170
3171         /* New channels are ready to roll, call the preactivate hook if needed
3172          * to modify HW settings or update kernel parameters.
3173          */
3174         if (preactivate) {
3175                 err = preactivate(priv, context);
3176                 if (err) {
3177                         priv->channels = old_chs;
3178                         goto out;
3179                 }
3180         }
3181
3182         mlx5e_close_channels(&old_chs);
3183         priv->profile->update_rx(priv);
3184
3185 out:
3186         mlx5e_activate_priv_channels(priv);
3187
3188         /* return carrier back if needed */
3189         if (carrier_ok)
3190                 netif_carrier_on(netdev);
3191
3192         return err;
3193 }
3194
3195 int mlx5e_safe_switch_channels(struct mlx5e_priv *priv,
3196                                struct mlx5e_channels *new_chs,
3197                                mlx5e_fp_preactivate preactivate,
3198                                void *context)
3199 {
3200         int err;
3201
3202         err = mlx5e_open_channels(priv, new_chs);
3203         if (err)
3204                 return err;
3205
3206         err = mlx5e_switch_priv_channels(priv, new_chs, preactivate, context);
3207         if (err)
3208                 goto err_close;
3209
3210         return 0;
3211
3212 err_close:
3213         mlx5e_close_channels(new_chs);
3214
3215         return err;
3216 }
3217
3218 int mlx5e_safe_reopen_channels(struct mlx5e_priv *priv)
3219 {
3220         struct mlx5e_channels new_channels = {};
3221
3222         new_channels.params = priv->channels.params;
3223         return mlx5e_safe_switch_channels(priv, &new_channels, NULL, NULL);
3224 }
3225
3226 void mlx5e_timestamp_init(struct mlx5e_priv *priv)
3227 {
3228         priv->tstamp.tx_type   = HWTSTAMP_TX_OFF;
3229         priv->tstamp.rx_filter = HWTSTAMP_FILTER_NONE;
3230 }
3231
3232 static void mlx5e_modify_admin_state(struct mlx5_core_dev *mdev,
3233                                      enum mlx5_port_status state)
3234 {
3235         struct mlx5_eswitch *esw = mdev->priv.eswitch;
3236         int vport_admin_state;
3237
3238         mlx5_set_port_admin_status(mdev, state);
3239
3240         if (mlx5_eswitch_mode(mdev) == MLX5_ESWITCH_OFFLOADS ||
3241             !MLX5_CAP_GEN(mdev, uplink_follow))
3242                 return;
3243
3244         if (state == MLX5_PORT_UP)
3245                 vport_admin_state = MLX5_VPORT_ADMIN_STATE_AUTO;
3246         else
3247                 vport_admin_state = MLX5_VPORT_ADMIN_STATE_DOWN;
3248
3249         mlx5_eswitch_set_vport_state(esw, MLX5_VPORT_UPLINK, vport_admin_state);
3250 }
3251
3252 int mlx5e_open_locked(struct net_device *netdev)
3253 {
3254         struct mlx5e_priv *priv = netdev_priv(netdev);
3255         int err;
3256
3257         set_bit(MLX5E_STATE_OPENED, &priv->state);
3258
3259         err = mlx5e_open_channels(priv, &priv->channels);
3260         if (err)
3261                 goto err_clear_state_opened_flag;
3262
3263         priv->profile->update_rx(priv);
3264         mlx5e_activate_priv_channels(priv);
3265         mlx5e_apply_traps(priv, true);
3266         if (priv->profile->update_carrier)
3267                 priv->profile->update_carrier(priv);
3268
3269         mlx5e_queue_update_stats(priv);
3270         return 0;
3271
3272 err_clear_state_opened_flag:
3273         clear_bit(MLX5E_STATE_OPENED, &priv->state);
3274         return err;
3275 }
3276
3277 int mlx5e_open(struct net_device *netdev)
3278 {
3279         struct mlx5e_priv *priv = netdev_priv(netdev);
3280         int err;
3281
3282         mutex_lock(&priv->state_lock);
3283         err = mlx5e_open_locked(netdev);
3284         if (!err)
3285                 mlx5e_modify_admin_state(priv->mdev, MLX5_PORT_UP);
3286         mutex_unlock(&priv->state_lock);
3287
3288         return err;
3289 }
3290
3291 int mlx5e_close_locked(struct net_device *netdev)
3292 {
3293         struct mlx5e_priv *priv = netdev_priv(netdev);
3294
3295         /* May already be CLOSED in case a previous configuration operation
3296          * (e.g RX/TX queue size change) that involves close&open failed.
3297          */
3298         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3299                 return 0;
3300
3301         mlx5e_apply_traps(priv, false);
3302         clear_bit(MLX5E_STATE_OPENED, &priv->state);
3303
3304         netif_carrier_off(priv->netdev);
3305         mlx5e_deactivate_priv_channels(priv);
3306         mlx5e_close_channels(&priv->channels);
3307
3308         return 0;
3309 }
3310
3311 int mlx5e_close(struct net_device *netdev)
3312 {
3313         struct mlx5e_priv *priv = netdev_priv(netdev);
3314         int err;
3315
3316         if (!netif_device_present(netdev))
3317                 return -ENODEV;
3318
3319         mutex_lock(&priv->state_lock);
3320         mlx5e_modify_admin_state(priv->mdev, MLX5_PORT_DOWN);
3321         err = mlx5e_close_locked(netdev);
3322         mutex_unlock(&priv->state_lock);
3323
3324         return err;
3325 }
3326
3327 static void mlx5e_free_drop_rq(struct mlx5e_rq *rq)
3328 {
3329         mlx5_wq_destroy(&rq->wq_ctrl);
3330 }
3331
3332 static int mlx5e_alloc_drop_rq(struct mlx5_core_dev *mdev,
3333                                struct mlx5e_rq *rq,
3334                                struct mlx5e_rq_param *param)
3335 {
3336         void *rqc = param->rqc;
3337         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
3338         int err;
3339
3340         param->wq.db_numa_node = param->wq.buf_numa_node;
3341
3342         err = mlx5_wq_cyc_create(mdev, &param->wq, rqc_wq, &rq->wqe.wq,
3343                                  &rq->wq_ctrl);
3344         if (err)
3345                 return err;
3346
3347         /* Mark as unused given "Drop-RQ" packets never reach XDP */
3348         xdp_rxq_info_unused(&rq->xdp_rxq);
3349
3350         rq->mdev = mdev;
3351
3352         return 0;
3353 }
3354
3355 static int mlx5e_alloc_drop_cq(struct mlx5e_priv *priv,
3356                                struct mlx5e_cq *cq,
3357                                struct mlx5e_cq_param *param)
3358 {
3359         struct mlx5_core_dev *mdev = priv->mdev;
3360
3361         param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
3362         param->wq.db_numa_node  = dev_to_node(mlx5_core_dma_dev(mdev));
3363
3364         return mlx5e_alloc_cq_common(priv, param, cq);
3365 }
3366
3367 int mlx5e_open_drop_rq(struct mlx5e_priv *priv,
3368                        struct mlx5e_rq *drop_rq)
3369 {
3370         struct mlx5_core_dev *mdev = priv->mdev;
3371         struct mlx5e_cq_param cq_param = {};
3372         struct mlx5e_rq_param rq_param = {};
3373         struct mlx5e_cq *cq = &drop_rq->cq;
3374         int err;
3375
3376         mlx5e_build_drop_rq_param(priv, &rq_param);
3377
3378         err = mlx5e_alloc_drop_cq(priv, cq, &cq_param);
3379         if (err)
3380                 return err;
3381
3382         err = mlx5e_create_cq(cq, &cq_param);
3383         if (err)
3384                 goto err_free_cq;
3385
3386         err = mlx5e_alloc_drop_rq(mdev, drop_rq, &rq_param);
3387         if (err)
3388                 goto err_destroy_cq;
3389
3390         err = mlx5e_create_rq(drop_rq, &rq_param);
3391         if (err)
3392                 goto err_free_rq;
3393
3394         err = mlx5e_modify_rq_state(drop_rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
3395         if (err)
3396                 mlx5_core_warn(priv->mdev, "modify_rq_state failed, rx_if_down_packets won't be counted %d\n", err);
3397
3398         return 0;
3399
3400 err_free_rq:
3401         mlx5e_free_drop_rq(drop_rq);
3402
3403 err_destroy_cq:
3404         mlx5e_destroy_cq(cq);
3405
3406 err_free_cq:
3407         mlx5e_free_cq(cq);
3408
3409         return err;
3410 }
3411
3412 void mlx5e_close_drop_rq(struct mlx5e_rq *drop_rq)
3413 {
3414         mlx5e_destroy_rq(drop_rq);
3415         mlx5e_free_drop_rq(drop_rq);
3416         mlx5e_destroy_cq(&drop_rq->cq);
3417         mlx5e_free_cq(&drop_rq->cq);
3418 }
3419
3420 int mlx5e_create_tis(struct mlx5_core_dev *mdev, void *in, u32 *tisn)
3421 {
3422         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
3423
3424         MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.td.tdn);
3425
3426         if (MLX5_GET(tisc, tisc, tls_en))
3427                 MLX5_SET(tisc, tisc, pd, mdev->mlx5e_res.pdn);
3428
3429         if (mlx5_lag_is_lacp_owner(mdev))
3430                 MLX5_SET(tisc, tisc, strict_lag_tx_port_affinity, 1);
3431
3432         return mlx5_core_create_tis(mdev, in, tisn);
3433 }
3434
3435 void mlx5e_destroy_tis(struct mlx5_core_dev *mdev, u32 tisn)
3436 {
3437         mlx5_core_destroy_tis(mdev, tisn);
3438 }
3439
3440 void mlx5e_destroy_tises(struct mlx5e_priv *priv)
3441 {
3442         int tc, i;
3443
3444         for (i = 0; i < mlx5e_get_num_lag_ports(priv->mdev); i++)
3445                 for (tc = 0; tc < priv->profile->max_tc; tc++)
3446                         mlx5e_destroy_tis(priv->mdev, priv->tisn[i][tc]);
3447 }
3448
3449 static bool mlx5e_lag_should_assign_affinity(struct mlx5_core_dev *mdev)
3450 {
3451         return MLX5_CAP_GEN(mdev, lag_tx_port_affinity) && mlx5e_get_num_lag_ports(mdev) > 1;
3452 }
3453
3454 int mlx5e_create_tises(struct mlx5e_priv *priv)
3455 {
3456         int tc, i;
3457         int err;
3458
3459         for (i = 0; i < mlx5e_get_num_lag_ports(priv->mdev); i++) {
3460                 for (tc = 0; tc < priv->profile->max_tc; tc++) {
3461                         u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
3462                         void *tisc;
3463
3464                         tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
3465
3466                         MLX5_SET(tisc, tisc, prio, tc << 1);
3467
3468                         if (mlx5e_lag_should_assign_affinity(priv->mdev))
3469                                 MLX5_SET(tisc, tisc, lag_tx_port_affinity, i + 1);
3470
3471                         err = mlx5e_create_tis(priv->mdev, in, &priv->tisn[i][tc]);
3472                         if (err)
3473                                 goto err_close_tises;
3474                 }
3475         }
3476
3477         return 0;
3478
3479 err_close_tises:
3480         for (; i >= 0; i--) {
3481                 for (tc--; tc >= 0; tc--)
3482                         mlx5e_destroy_tis(priv->mdev, priv->tisn[i][tc]);
3483                 tc = priv->profile->max_tc;
3484         }
3485
3486         return err;
3487 }
3488
3489 static void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
3490 {
3491         mlx5e_destroy_tises(priv);
3492 }
3493
3494 static void mlx5e_build_indir_tir_ctx_common(struct mlx5e_priv *priv,
3495                                              u32 rqtn, u32 *tirc)
3496 {
3497         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.td.tdn);
3498         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
3499         MLX5_SET(tirc, tirc, indirect_table, rqtn);
3500         MLX5_SET(tirc, tirc, tunneled_offload_en,
3501                  priv->channels.params.tunneled_offload_en);
3502
3503         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
3504 }
3505
3506 static void mlx5e_build_indir_tir_ctx(struct mlx5e_priv *priv,
3507                                       enum mlx5e_traffic_types tt,
3508                                       u32 *tirc)
3509 {
3510         mlx5e_build_indir_tir_ctx_common(priv, priv->indir_rqt.rqtn, tirc);
3511         mlx5e_build_indir_tir_ctx_hash(&priv->rss_params,
3512                                        &tirc_default_config[tt], tirc, false);
3513 }
3514
3515 static void mlx5e_build_direct_tir_ctx(struct mlx5e_priv *priv, u32 rqtn, u32 *tirc)
3516 {
3517         mlx5e_build_indir_tir_ctx_common(priv, rqtn, tirc);
3518         MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
3519 }
3520
3521 static void mlx5e_build_inner_indir_tir_ctx(struct mlx5e_priv *priv,
3522                                             enum mlx5e_traffic_types tt,
3523                                             u32 *tirc)
3524 {
3525         mlx5e_build_indir_tir_ctx_common(priv, priv->indir_rqt.rqtn, tirc);
3526         mlx5e_build_indir_tir_ctx_hash(&priv->rss_params,
3527                                        &tirc_default_config[tt], tirc, true);
3528 }
3529
3530 int mlx5e_create_indirect_tirs(struct mlx5e_priv *priv, bool inner_ttc)
3531 {
3532         struct mlx5e_tir *tir;
3533         void *tirc;
3534         int inlen;
3535         int i = 0;
3536         int err;
3537         u32 *in;
3538         int tt;
3539
3540         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
3541         in = kvzalloc(inlen, GFP_KERNEL);
3542         if (!in)
3543                 return -ENOMEM;
3544
3545         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
3546                 memset(in, 0, inlen);
3547                 tir = &priv->indir_tir[tt];
3548                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3549                 mlx5e_build_indir_tir_ctx(priv, tt, tirc);
3550                 err = mlx5e_create_tir(priv->mdev, tir, in);
3551                 if (err) {
3552                         mlx5_core_warn(priv->mdev, "create indirect tirs failed, %d\n", err);
3553                         goto err_destroy_inner_tirs;
3554                 }
3555         }
3556
3557         if (!inner_ttc || !mlx5e_tunnel_inner_ft_supported(priv->mdev))
3558                 goto out;
3559
3560         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++) {
3561                 memset(in, 0, inlen);
3562                 tir = &priv->inner_indir_tir[i];
3563                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3564                 mlx5e_build_inner_indir_tir_ctx(priv, i, tirc);
3565                 err = mlx5e_create_tir(priv->mdev, tir, in);
3566                 if (err) {
3567                         mlx5_core_warn(priv->mdev, "create inner indirect tirs failed, %d\n", err);
3568                         goto err_destroy_inner_tirs;
3569                 }
3570         }
3571
3572 out:
3573         kvfree(in);
3574
3575         return 0;
3576
3577 err_destroy_inner_tirs:
3578         for (i--; i >= 0; i--)
3579                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3580
3581         for (tt--; tt >= 0; tt--)
3582                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[tt]);
3583
3584         kvfree(in);
3585
3586         return err;
3587 }
3588
3589 int mlx5e_create_direct_tirs(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
3590 {
3591         struct mlx5e_tir *tir;
3592         void *tirc;
3593         int inlen;
3594         int err = 0;
3595         u32 *in;
3596         int ix;
3597
3598         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
3599         in = kvzalloc(inlen, GFP_KERNEL);
3600         if (!in)
3601                 return -ENOMEM;
3602
3603         for (ix = 0; ix < priv->max_nch; ix++) {
3604                 memset(in, 0, inlen);
3605                 tir = &tirs[ix];
3606                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3607                 mlx5e_build_direct_tir_ctx(priv, tir->rqt.rqtn, tirc);
3608                 err = mlx5e_create_tir(priv->mdev, tir, in);
3609                 if (unlikely(err))
3610                         goto err_destroy_ch_tirs;
3611         }
3612
3613         goto out;
3614
3615 err_destroy_ch_tirs:
3616         mlx5_core_warn(priv->mdev, "create tirs failed, %d\n", err);
3617         for (ix--; ix >= 0; ix--)
3618                 mlx5e_destroy_tir(priv->mdev, &tirs[ix]);
3619
3620 out:
3621         kvfree(in);
3622
3623         return err;
3624 }
3625
3626 void mlx5e_destroy_indirect_tirs(struct mlx5e_priv *priv)
3627 {
3628         int i;
3629
3630         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3631                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[i]);
3632
3633         /* Verify inner tirs resources allocated */
3634         if (!priv->inner_indir_tir[0].tirn)
3635                 return;
3636
3637         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3638                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3639 }
3640
3641 void mlx5e_destroy_direct_tirs(struct mlx5e_priv *priv, struct mlx5e_tir *tirs)
3642 {
3643         int i;
3644
3645         for (i = 0; i < priv->max_nch; i++)
3646                 mlx5e_destroy_tir(priv->mdev, &tirs[i]);
3647 }
3648
3649 static int mlx5e_modify_channels_scatter_fcs(struct mlx5e_channels *chs, bool enable)
3650 {
3651         int err = 0;
3652         int i;
3653
3654         for (i = 0; i < chs->num; i++) {
3655                 err = mlx5e_modify_rq_scatter_fcs(&chs->c[i]->rq, enable);
3656                 if (err)
3657                         return err;
3658         }
3659
3660         return 0;
3661 }
3662
3663 static int mlx5e_modify_channels_vsd(struct mlx5e_channels *chs, bool vsd)
3664 {
3665         int err = 0;
3666         int i;
3667
3668         for (i = 0; i < chs->num; i++) {
3669                 err = mlx5e_modify_rq_vsd(&chs->c[i]->rq, vsd);
3670                 if (err)
3671                         return err;
3672         }
3673
3674         return 0;
3675 }
3676
3677 static int mlx5e_setup_tc_mqprio(struct mlx5e_priv *priv,
3678                                  struct tc_mqprio_qopt *mqprio)
3679 {
3680         struct mlx5e_channels new_channels = {};
3681         u8 tc = mqprio->num_tc;
3682         int err = 0;
3683
3684         mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
3685
3686         if (tc && tc != MLX5E_MAX_NUM_TC)
3687                 return -EINVAL;
3688
3689         mutex_lock(&priv->state_lock);
3690
3691         /* MQPRIO is another toplevel qdisc that can't be attached
3692          * simultaneously with the offloaded HTB.
3693          */
3694         if (WARN_ON(priv->htb.maj_id)) {
3695                 err = -EINVAL;
3696                 goto out;
3697         }
3698
3699         new_channels.params = priv->channels.params;
3700         new_channels.params.num_tc = tc ? tc : 1;
3701
3702         if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
3703                 struct mlx5e_params old_params;
3704
3705                 old_params = priv->channels.params;
3706                 priv->channels.params = new_channels.params;
3707                 err = mlx5e_num_channels_changed(priv);
3708                 if (err)
3709                         priv->channels.params = old_params;
3710
3711                 goto out;
3712         }
3713
3714         err = mlx5e_safe_switch_channels(priv, &new_channels,
3715                                          mlx5e_num_channels_changed_ctx, NULL);
3716
3717 out:
3718         priv->max_opened_tc = max_t(u8, priv->max_opened_tc,
3719                                     priv->channels.params.num_tc);
3720         mutex_unlock(&priv->state_lock);
3721         return err;
3722 }
3723
3724 static int mlx5e_setup_tc_htb(struct mlx5e_priv *priv, struct tc_htb_qopt_offload *htb)
3725 {
3726         int res;
3727
3728         switch (htb->command) {
3729         case TC_HTB_CREATE:
3730                 return mlx5e_htb_root_add(priv, htb->parent_classid, htb->classid,
3731                                           htb->extack);
3732         case TC_HTB_DESTROY:
3733                 return mlx5e_htb_root_del(priv);
3734         case TC_HTB_LEAF_ALLOC_QUEUE:
3735                 res = mlx5e_htb_leaf_alloc_queue(priv, htb->classid, htb->parent_classid,
3736                                                  htb->rate, htb->ceil, htb->extack);
3737                 if (res < 0)
3738                         return res;
3739                 htb->qid = res;
3740                 return 0;
3741         case TC_HTB_LEAF_TO_INNER:
3742                 return mlx5e_htb_leaf_to_inner(priv, htb->parent_classid, htb->classid,
3743                                                htb->rate, htb->ceil, htb->extack);
3744         case TC_HTB_LEAF_DEL:
3745                 return mlx5e_htb_leaf_del(priv, htb->classid, &htb->moved_qid, &htb->qid,
3746                                           htb->extack);
3747         case TC_HTB_LEAF_DEL_LAST:
3748         case TC_HTB_LEAF_DEL_LAST_FORCE:
3749                 return mlx5e_htb_leaf_del_last(priv, htb->classid,
3750                                                htb->command == TC_HTB_LEAF_DEL_LAST_FORCE,
3751                                                htb->extack);
3752         case TC_HTB_NODE_MODIFY:
3753                 return mlx5e_htb_node_modify(priv, htb->classid, htb->rate, htb->ceil,
3754                                              htb->extack);
3755         case TC_HTB_LEAF_QUERY_QUEUE:
3756                 res = mlx5e_get_txq_by_classid(priv, htb->classid);
3757                 if (res < 0)
3758                         return res;
3759                 htb->qid = res;
3760                 return 0;
3761         default:
3762                 return -EOPNOTSUPP;
3763         }
3764 }
3765
3766 static LIST_HEAD(mlx5e_block_cb_list);
3767
3768 static int mlx5e_setup_tc(struct net_device *dev, enum tc_setup_type type,
3769                           void *type_data)
3770 {
3771         struct mlx5e_priv *priv = netdev_priv(dev);
3772         int err;
3773
3774         switch (type) {
3775         case TC_SETUP_BLOCK: {
3776                 struct flow_block_offload *f = type_data;
3777
3778                 f->unlocked_driver_cb = true;
3779                 return flow_block_cb_setup_simple(type_data,
3780                                                   &mlx5e_block_cb_list,
3781                                                   mlx5e_setup_tc_block_cb,
3782                                                   priv, priv, true);
3783         }
3784         case TC_SETUP_QDISC_MQPRIO:
3785                 return mlx5e_setup_tc_mqprio(priv, type_data);
3786         case TC_SETUP_QDISC_HTB:
3787                 mutex_lock(&priv->state_lock);
3788                 err = mlx5e_setup_tc_htb(priv, type_data);
3789                 mutex_unlock(&priv->state_lock);
3790                 return err;
3791         default:
3792                 return -EOPNOTSUPP;
3793         }
3794 }
3795
3796 void mlx5e_fold_sw_stats64(struct mlx5e_priv *priv, struct rtnl_link_stats64 *s)
3797 {
3798         int i;
3799
3800         for (i = 0; i < priv->max_nch; i++) {
3801                 struct mlx5e_channel_stats *channel_stats = &priv->channel_stats[i];
3802                 struct mlx5e_rq_stats *xskrq_stats = &channel_stats->xskrq;
3803                 struct mlx5e_rq_stats *rq_stats = &channel_stats->rq;
3804                 int j;
3805
3806                 s->rx_packets   += rq_stats->packets + xskrq_stats->packets;
3807                 s->rx_bytes     += rq_stats->bytes + xskrq_stats->bytes;
3808                 s->multicast    += rq_stats->mcast_packets + xskrq_stats->mcast_packets;
3809
3810                 for (j = 0; j < priv->max_opened_tc; j++) {
3811                         struct mlx5e_sq_stats *sq_stats = &channel_stats->sq[j];
3812
3813                         s->tx_packets    += sq_stats->packets;
3814                         s->tx_bytes      += sq_stats->bytes;
3815                         s->tx_dropped    += sq_stats->dropped;
3816                 }
3817         }
3818 }
3819
3820 void
3821 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
3822 {
3823         struct mlx5e_priv *priv = netdev_priv(dev);
3824         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
3825
3826         /* In switchdev mode, monitor counters doesn't monitor
3827          * rx/tx stats of 802_3. The update stats mechanism
3828          * should keep the 802_3 layout counters updated
3829          */
3830         if (!mlx5e_monitor_counter_supported(priv) ||
3831             mlx5e_is_uplink_rep(priv)) {
3832                 /* update HW stats in background for next time */
3833                 mlx5e_queue_update_stats(priv);
3834         }
3835
3836         if (mlx5e_is_uplink_rep(priv)) {
3837                 stats->rx_packets = PPORT_802_3_GET(pstats, a_frames_received_ok);
3838                 stats->rx_bytes   = PPORT_802_3_GET(pstats, a_octets_received_ok);
3839                 stats->tx_packets = PPORT_802_3_GET(pstats, a_frames_transmitted_ok);
3840                 stats->tx_bytes   = PPORT_802_3_GET(pstats, a_octets_transmitted_ok);
3841         } else {
3842                 mlx5e_fold_sw_stats64(priv, stats);
3843         }
3844
3845         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
3846
3847         stats->rx_length_errors =
3848                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
3849                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
3850                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
3851         stats->rx_crc_errors =
3852                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
3853         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
3854         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
3855         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
3856                            stats->rx_frame_errors;
3857         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
3858 }
3859
3860 static void mlx5e_nic_set_rx_mode(struct mlx5e_priv *priv)
3861 {
3862         if (mlx5e_is_uplink_rep(priv))
3863                 return; /* no rx mode for uplink rep */
3864
3865         queue_work(priv->wq, &priv->set_rx_mode_work);
3866 }
3867
3868 static void mlx5e_set_rx_mode(struct net_device *dev)
3869 {
3870         struct mlx5e_priv *priv = netdev_priv(dev);
3871
3872         mlx5e_nic_set_rx_mode(priv);
3873 }
3874
3875 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
3876 {
3877         struct mlx5e_priv *priv = netdev_priv(netdev);
3878         struct sockaddr *saddr = addr;
3879
3880         if (!is_valid_ether_addr(saddr->sa_data))
3881                 return -EADDRNOTAVAIL;
3882
3883         netif_addr_lock_bh(netdev);
3884         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
3885         netif_addr_unlock_bh(netdev);
3886
3887         mlx5e_nic_set_rx_mode(priv);
3888
3889         return 0;
3890 }
3891
3892 #define MLX5E_SET_FEATURE(features, feature, enable)    \
3893         do {                                            \
3894                 if (enable)                             \
3895                         *features |= feature;           \
3896                 else                                    \
3897                         *features &= ~feature;          \
3898         } while (0)
3899
3900 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
3901
3902 static int set_feature_lro(struct net_device *netdev, bool enable)
3903 {
3904         struct mlx5e_priv *priv = netdev_priv(netdev);
3905         struct mlx5_core_dev *mdev = priv->mdev;
3906         struct mlx5e_channels new_channels = {};
3907         struct mlx5e_params *cur_params;
3908         int err = 0;
3909         bool reset;
3910
3911         mutex_lock(&priv->state_lock);
3912
3913         if (enable && priv->xsk.refcnt) {
3914                 netdev_warn(netdev, "LRO is incompatible with AF_XDP (%u XSKs are active)\n",
3915                             priv->xsk.refcnt);
3916                 err = -EINVAL;
3917                 goto out;
3918         }
3919
3920         cur_params = &priv->channels.params;
3921         if (enable && !MLX5E_GET_PFLAG(cur_params, MLX5E_PFLAG_RX_STRIDING_RQ)) {
3922                 netdev_warn(netdev, "can't set LRO with legacy RQ\n");
3923                 err = -EINVAL;
3924                 goto out;
3925         }
3926
3927         reset = test_bit(MLX5E_STATE_OPENED, &priv->state);
3928
3929         new_channels.params = *cur_params;
3930         new_channels.params.lro_en = enable;
3931
3932         if (cur_params->rq_wq_type != MLX5_WQ_TYPE_CYCLIC) {
3933                 if (mlx5e_rx_mpwqe_is_linear_skb(mdev, cur_params, NULL) ==
3934                     mlx5e_rx_mpwqe_is_linear_skb(mdev, &new_channels.params, NULL))
3935                         reset = false;
3936         }
3937
3938         if (!reset) {
3939                 struct mlx5e_params old_params;
3940
3941                 old_params = *cur_params;
3942                 *cur_params = new_channels.params;
3943                 err = mlx5e_modify_tirs_lro(priv);
3944                 if (err)
3945                         *cur_params = old_params;
3946                 goto out;
3947         }
3948
3949         err = mlx5e_safe_switch_channels(priv, &new_channels,
3950                                          mlx5e_modify_tirs_lro_ctx, NULL);
3951 out:
3952         mutex_unlock(&priv->state_lock);
3953         return err;
3954 }
3955
3956 static int set_feature_cvlan_filter(struct net_device *netdev, bool enable)
3957 {
3958         struct mlx5e_priv *priv = netdev_priv(netdev);
3959
3960         if (enable)
3961                 mlx5e_enable_cvlan_filter(priv);
3962         else
3963                 mlx5e_disable_cvlan_filter(priv);
3964
3965         return 0;
3966 }
3967
3968 static int set_feature_hw_tc(struct net_device *netdev, bool enable)
3969 {
3970         struct mlx5e_priv *priv = netdev_priv(netdev);
3971
3972 #if IS_ENABLED(CONFIG_MLX5_CLS_ACT)
3973         if (!enable && mlx5e_tc_num_filters(priv, MLX5_TC_FLAG(NIC_OFFLOAD))) {
3974                 netdev_err(netdev,
3975                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
3976                 return -EINVAL;
3977         }
3978 #endif
3979
3980         if (!enable && priv->htb.maj_id) {
3981                 netdev_err(netdev, "Active HTB offload, can't turn hw_tc_offload off\n");
3982                 return -EINVAL;
3983         }
3984
3985         return 0;
3986 }
3987
3988 static int set_feature_rx_all(struct net_device *netdev, bool enable)
3989 {
3990         struct mlx5e_priv *priv = netdev_priv(netdev);
3991         struct mlx5_core_dev *mdev = priv->mdev;
3992
3993         return mlx5_set_port_fcs(mdev, !enable);
3994 }
3995
3996 static int set_feature_rx_fcs(struct net_device *netdev, bool enable)
3997 {
3998         struct mlx5e_priv *priv = netdev_priv(netdev);
3999         int err;
4000
4001         mutex_lock(&priv->state_lock);
4002
4003         priv->channels.params.scatter_fcs_en = enable;
4004         err = mlx5e_modify_channels_scatter_fcs(&priv->channels, enable);
4005         if (err)
4006                 priv->channels.params.scatter_fcs_en = !enable;
4007
4008         mutex_unlock(&priv->state_lock);
4009
4010         return err;
4011 }
4012
4013 static int set_feature_rx_vlan(struct net_device *netdev, bool enable)
4014 {
4015         struct mlx5e_priv *priv = netdev_priv(netdev);
4016         int err = 0;
4017
4018         mutex_lock(&priv->state_lock);
4019
4020         priv->channels.params.vlan_strip_disable = !enable;
4021         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
4022                 goto unlock;
4023
4024         err = mlx5e_modify_channels_vsd(&priv->channels, !enable);
4025         if (err)
4026                 priv->channels.params.vlan_strip_disable = enable;
4027
4028 unlock:
4029         mutex_unlock(&priv->state_lock);
4030
4031         return err;
4032 }
4033
4034 #ifdef CONFIG_MLX5_EN_ARFS
4035 static int set_feature_arfs(struct net_device *netdev, bool enable)
4036 {
4037         struct mlx5e_priv *priv = netdev_priv(netdev);
4038         int err;
4039
4040         if (enable)
4041                 err = mlx5e_arfs_enable(priv);
4042         else
4043                 err = mlx5e_arfs_disable(priv);
4044
4045         return err;
4046 }
4047 #endif
4048
4049 static int mlx5e_handle_feature(struct net_device *netdev,
4050                                 netdev_features_t *features,
4051                                 netdev_features_t wanted_features,
4052                                 netdev_features_t feature,
4053                                 mlx5e_feature_handler feature_handler)
4054 {
4055         netdev_features_t changes = wanted_features ^ netdev->features;
4056         bool enable = !!(wanted_features & feature);
4057         int err;
4058
4059         if (!(changes & feature))
4060                 return 0;
4061
4062         err = feature_handler(netdev, enable);
4063         if (err) {
4064                 netdev_err(netdev, "%s feature %pNF failed, err %d\n",
4065                            enable ? "Enable" : "Disable", &feature, err);
4066                 return err;
4067         }
4068
4069         MLX5E_SET_FEATURE(features, feature, enable);
4070         return 0;
4071 }
4072
4073 int mlx5e_set_features(struct net_device *netdev, netdev_features_t features)
4074 {
4075         netdev_features_t oper_features = netdev->features;
4076         int err = 0;
4077
4078 #define MLX5E_HANDLE_FEATURE(feature, handler) \
4079         mlx5e_handle_feature(netdev, &oper_features, features, feature, handler)
4080
4081         err |= MLX5E_HANDLE_FEATURE(NETIF_F_LRO, set_feature_lro);
4082         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_VLAN_CTAG_FILTER,
4083                                     set_feature_cvlan_filter);
4084         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_TC, set_feature_hw_tc);
4085         err |= MLX5E_HANDLE_FEATURE(NETIF_F_RXALL, set_feature_rx_all);
4086         err |= MLX5E_HANDLE_FEATURE(NETIF_F_RXFCS, set_feature_rx_fcs);
4087         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_VLAN_CTAG_RX, set_feature_rx_vlan);
4088 #ifdef CONFIG_MLX5_EN_ARFS
4089         err |= MLX5E_HANDLE_FEATURE(NETIF_F_NTUPLE, set_feature_arfs);
4090 #endif
4091         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_TLS_RX, mlx5e_ktls_set_feature_rx);
4092
4093         if (err) {
4094                 netdev->features = oper_features;
4095                 return -EINVAL;
4096         }
4097
4098         return 0;
4099 }
4100
4101 static netdev_features_t mlx5e_fix_features(struct net_device *netdev,
4102                                             netdev_features_t features)
4103 {
4104         struct mlx5e_priv *priv = netdev_priv(netdev);
4105         struct mlx5e_params *params;
4106
4107         mutex_lock(&priv->state_lock);
4108         params = &priv->channels.params;
4109         if (!bitmap_empty(priv->fs.vlan.active_svlans, VLAN_N_VID)) {
4110                 /* HW strips the outer C-tag header, this is a problem
4111                  * for S-tag traffic.
4112                  */
4113                 features &= ~NETIF_F_HW_VLAN_CTAG_RX;
4114                 if (!params->vlan_strip_disable)
4115                         netdev_warn(netdev, "Dropping C-tag vlan stripping offload due to S-tag vlan\n");
4116         }
4117
4118         if (!MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ)) {
4119                 if (features & NETIF_F_LRO) {
4120                         netdev_warn(netdev, "Disabling LRO, not supported in legacy RQ\n");
4121                         features &= ~NETIF_F_LRO;
4122                 }
4123         }
4124
4125         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
4126                 features &= ~NETIF_F_RXHASH;
4127                 if (netdev->features & NETIF_F_RXHASH)
4128                         netdev_warn(netdev, "Disabling rxhash, not supported when CQE compress is active\n");
4129         }
4130
4131         mutex_unlock(&priv->state_lock);
4132
4133         return features;
4134 }
4135
4136 static bool mlx5e_xsk_validate_mtu(struct net_device *netdev,
4137                                    struct mlx5e_channels *chs,
4138                                    struct mlx5e_params *new_params,
4139                                    struct mlx5_core_dev *mdev)
4140 {
4141         u16 ix;
4142
4143         for (ix = 0; ix < chs->params.num_channels; ix++) {
4144                 struct xsk_buff_pool *xsk_pool =
4145                         mlx5e_xsk_get_pool(&chs->params, chs->params.xsk, ix);
4146                 struct mlx5e_xsk_param xsk;
4147
4148                 if (!xsk_pool)
4149                         continue;
4150
4151                 mlx5e_build_xsk_param(xsk_pool, &xsk);
4152
4153                 if (!mlx5e_validate_xsk_param(new_params, &xsk, mdev)) {
4154                         u32 hr = mlx5e_get_linear_rq_headroom(new_params, &xsk);
4155                         int max_mtu_frame, max_mtu_page, max_mtu;
4156
4157                         /* Two criteria must be met:
4158                          * 1. HW MTU + all headrooms <= XSK frame size.
4159                          * 2. Size of SKBs allocated on XDP_PASS <= PAGE_SIZE.
4160                          */
4161                         max_mtu_frame = MLX5E_HW2SW_MTU(new_params, xsk.chunk_size - hr);
4162                         max_mtu_page = mlx5e_xdp_max_mtu(new_params, &xsk);
4163                         max_mtu = min(max_mtu_frame, max_mtu_page);
4164
4165                         netdev_err(netdev, "MTU %d is too big for an XSK running on channel %u. Try MTU <= %d\n",
4166                                    new_params->sw_mtu, ix, max_mtu);
4167                         return false;
4168                 }
4169         }
4170
4171         return true;
4172 }
4173
4174 int mlx5e_change_mtu(struct net_device *netdev, int new_mtu,
4175                      mlx5e_fp_preactivate preactivate)
4176 {
4177         struct mlx5e_priv *priv = netdev_priv(netdev);
4178         struct mlx5e_channels new_channels = {};
4179         struct mlx5e_params *params;
4180         int err = 0;
4181         bool reset;
4182
4183         mutex_lock(&priv->state_lock);
4184
4185         params = &priv->channels.params;
4186
4187         reset = !params->lro_en;
4188         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
4189
4190         new_channels.params = *params;
4191         new_channels.params.sw_mtu = new_mtu;
4192         err = mlx5e_validate_params(priv, &new_channels.params);
4193         if (err)
4194                 goto out;
4195
4196         if (params->xdp_prog &&
4197             !mlx5e_rx_is_linear_skb(&new_channels.params, NULL)) {
4198                 netdev_err(netdev, "MTU(%d) > %d is not allowed while XDP enabled\n",
4199                            new_mtu, mlx5e_xdp_max_mtu(params, NULL));
4200                 err = -EINVAL;
4201                 goto out;
4202         }
4203
4204         if (priv->xsk.refcnt &&
4205             !mlx5e_xsk_validate_mtu(netdev, &priv->channels,
4206                                     &new_channels.params, priv->mdev)) {
4207                 err = -EINVAL;
4208                 goto out;
4209         }
4210
4211         if (params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
4212                 bool is_linear = mlx5e_rx_mpwqe_is_linear_skb(priv->mdev,
4213                                                               &new_channels.params,
4214                                                               NULL);
4215                 u8 ppw_old = mlx5e_mpwqe_log_pkts_per_wqe(params, NULL);
4216                 u8 ppw_new = mlx5e_mpwqe_log_pkts_per_wqe(&new_channels.params, NULL);
4217
4218                 /* If XSK is active, XSK RQs are linear. */
4219                 is_linear |= priv->xsk.refcnt;
4220
4221                 /* Always reset in linear mode - hw_mtu is used in data path. */
4222                 reset = reset && (is_linear || (ppw_old != ppw_new));
4223         }
4224
4225         if (!reset) {
4226                 unsigned int old_mtu = params->sw_mtu;
4227
4228                 params->sw_mtu = new_mtu;
4229                 if (preactivate) {
4230                         err = preactivate(priv, NULL);
4231                         if (err) {
4232                                 params->sw_mtu = old_mtu;
4233                                 goto out;
4234                         }
4235                 }
4236                 netdev->mtu = params->sw_mtu;
4237                 goto out;
4238         }
4239
4240         err = mlx5e_safe_switch_channels(priv, &new_channels, preactivate, NULL);
4241         if (err)
4242                 goto out;
4243
4244         netdev->mtu = new_channels.params.sw_mtu;
4245
4246 out:
4247         mutex_unlock(&priv->state_lock);
4248         return err;
4249 }
4250
4251 static int mlx5e_change_nic_mtu(struct net_device *netdev, int new_mtu)
4252 {
4253         return mlx5e_change_mtu(netdev, new_mtu, mlx5e_set_dev_port_mtu_ctx);
4254 }
4255
4256 int mlx5e_hwstamp_set(struct mlx5e_priv *priv, struct ifreq *ifr)
4257 {
4258         struct hwtstamp_config config;
4259         int err;
4260
4261         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz) ||
4262             (mlx5_clock_get_ptp_index(priv->mdev) == -1))
4263                 return -EOPNOTSUPP;
4264
4265         if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
4266                 return -EFAULT;
4267
4268         /* TX HW timestamp */
4269         switch (config.tx_type) {
4270         case HWTSTAMP_TX_OFF:
4271         case HWTSTAMP_TX_ON:
4272                 break;
4273         default:
4274                 return -ERANGE;
4275         }
4276
4277         mutex_lock(&priv->state_lock);
4278         /* RX HW timestamp */
4279         switch (config.rx_filter) {
4280         case HWTSTAMP_FILTER_NONE:
4281                 /* Reset CQE compression to Admin default */
4282                 mlx5e_modify_rx_cqe_compression_locked(priv, priv->channels.params.rx_cqe_compress_def);
4283                 break;
4284         case HWTSTAMP_FILTER_ALL:
4285         case HWTSTAMP_FILTER_SOME:
4286         case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
4287         case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
4288         case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
4289         case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
4290         case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
4291         case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
4292         case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
4293         case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
4294         case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
4295         case HWTSTAMP_FILTER_PTP_V2_EVENT:
4296         case HWTSTAMP_FILTER_PTP_V2_SYNC:
4297         case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
4298         case HWTSTAMP_FILTER_NTP_ALL:
4299                 /* Disable CQE compression */
4300                 if (MLX5E_GET_PFLAG(&priv->channels.params, MLX5E_PFLAG_RX_CQE_COMPRESS))
4301                         netdev_warn(priv->netdev, "Disabling RX cqe compression\n");
4302                 err = mlx5e_modify_rx_cqe_compression_locked(priv, false);
4303                 if (err) {
4304                         netdev_err(priv->netdev, "Failed disabling cqe compression err=%d\n", err);
4305                         mutex_unlock(&priv->state_lock);
4306                         return err;
4307                 }
4308                 config.rx_filter = HWTSTAMP_FILTER_ALL;
4309                 break;
4310         default:
4311                 mutex_unlock(&priv->state_lock);
4312                 return -ERANGE;
4313         }
4314
4315         memcpy(&priv->tstamp, &config, sizeof(config));
4316         mutex_unlock(&priv->state_lock);
4317
4318         /* might need to fix some features */
4319         netdev_update_features(priv->netdev);
4320
4321         return copy_to_user(ifr->ifr_data, &config,
4322                             sizeof(config)) ? -EFAULT : 0;
4323 }
4324
4325 int mlx5e_hwstamp_get(struct mlx5e_priv *priv, struct ifreq *ifr)
4326 {
4327         struct hwtstamp_config *cfg = &priv->tstamp;
4328
4329         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
4330                 return -EOPNOTSUPP;
4331
4332         return copy_to_user(ifr->ifr_data, cfg, sizeof(*cfg)) ? -EFAULT : 0;
4333 }
4334
4335 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
4336 {
4337         struct mlx5e_priv *priv = netdev_priv(dev);
4338
4339         switch (cmd) {
4340         case SIOCSHWTSTAMP:
4341                 return mlx5e_hwstamp_set(priv, ifr);
4342         case SIOCGHWTSTAMP:
4343                 return mlx5e_hwstamp_get(priv, ifr);
4344         default:
4345                 return -EOPNOTSUPP;
4346         }
4347 }
4348
4349 #ifdef CONFIG_MLX5_ESWITCH
4350 int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
4351 {
4352         struct mlx5e_priv *priv = netdev_priv(dev);
4353         struct mlx5_core_dev *mdev = priv->mdev;
4354
4355         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
4356 }
4357
4358 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos,
4359                              __be16 vlan_proto)
4360 {
4361         struct mlx5e_priv *priv = netdev_priv(dev);
4362         struct mlx5_core_dev *mdev = priv->mdev;
4363
4364         if (vlan_proto != htons(ETH_P_8021Q))
4365                 return -EPROTONOSUPPORT;
4366
4367         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
4368                                            vlan, qos);
4369 }
4370
4371 static int mlx5e_set_vf_spoofchk(struct net_device *dev, int vf, bool setting)
4372 {
4373         struct mlx5e_priv *priv = netdev_priv(dev);
4374         struct mlx5_core_dev *mdev = priv->mdev;
4375
4376         return mlx5_eswitch_set_vport_spoofchk(mdev->priv.eswitch, vf + 1, setting);
4377 }
4378
4379 static int mlx5e_set_vf_trust(struct net_device *dev, int vf, bool setting)
4380 {
4381         struct mlx5e_priv *priv = netdev_priv(dev);
4382         struct mlx5_core_dev *mdev = priv->mdev;
4383
4384         return mlx5_eswitch_set_vport_trust(mdev->priv.eswitch, vf + 1, setting);
4385 }
4386
4387 int mlx5e_set_vf_rate(struct net_device *dev, int vf, int min_tx_rate,
4388                       int max_tx_rate)
4389 {
4390         struct mlx5e_priv *priv = netdev_priv(dev);
4391         struct mlx5_core_dev *mdev = priv->mdev;
4392
4393         return mlx5_eswitch_set_vport_rate(mdev->priv.eswitch, vf + 1,
4394                                            max_tx_rate, min_tx_rate);
4395 }
4396
4397 static int mlx5_vport_link2ifla(u8 esw_link)
4398 {
4399         switch (esw_link) {
4400         case MLX5_VPORT_ADMIN_STATE_DOWN:
4401                 return IFLA_VF_LINK_STATE_DISABLE;
4402         case MLX5_VPORT_ADMIN_STATE_UP:
4403                 return IFLA_VF_LINK_STATE_ENABLE;
4404         }
4405         return IFLA_VF_LINK_STATE_AUTO;
4406 }
4407
4408 static int mlx5_ifla_link2vport(u8 ifla_link)
4409 {
4410         switch (ifla_link) {
4411         case IFLA_VF_LINK_STATE_DISABLE:
4412                 return MLX5_VPORT_ADMIN_STATE_DOWN;
4413         case IFLA_VF_LINK_STATE_ENABLE:
4414                 return MLX5_VPORT_ADMIN_STATE_UP;
4415         }
4416         return MLX5_VPORT_ADMIN_STATE_AUTO;
4417 }
4418
4419 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
4420                                    int link_state)
4421 {
4422         struct mlx5e_priv *priv = netdev_priv(dev);
4423         struct mlx5_core_dev *mdev = priv->mdev;
4424
4425         if (mlx5e_is_uplink_rep(priv))
4426                 return -EOPNOTSUPP;
4427
4428         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
4429                                             mlx5_ifla_link2vport(link_state));
4430 }
4431
4432 int mlx5e_get_vf_config(struct net_device *dev,
4433                         int vf, struct ifla_vf_info *ivi)
4434 {
4435         struct mlx5e_priv *priv = netdev_priv(dev);
4436         struct mlx5_core_dev *mdev = priv->mdev;
4437         int err;
4438
4439         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
4440         if (err)
4441                 return err;
4442         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
4443         return 0;
4444 }
4445
4446 int mlx5e_get_vf_stats(struct net_device *dev,
4447                        int vf, struct ifla_vf_stats *vf_stats)
4448 {
4449         struct mlx5e_priv *priv = netdev_priv(dev);
4450         struct mlx5_core_dev *mdev = priv->mdev;
4451
4452         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
4453                                             vf_stats);
4454 }
4455
4456 static bool
4457 mlx5e_has_offload_stats(const struct net_device *dev, int attr_id)
4458 {
4459         struct mlx5e_priv *priv = netdev_priv(dev);
4460
4461         if (!mlx5e_is_uplink_rep(priv))
4462                 return false;
4463
4464         return mlx5e_rep_has_offload_stats(dev, attr_id);
4465 }
4466
4467 static int
4468 mlx5e_get_offload_stats(int attr_id, const struct net_device *dev,
4469                         void *sp)
4470 {
4471         struct mlx5e_priv *priv = netdev_priv(dev);
4472
4473         if (!mlx5e_is_uplink_rep(priv))
4474                 return -EOPNOTSUPP;
4475
4476         return mlx5e_rep_get_offload_stats(attr_id, dev, sp);
4477 }
4478 #endif
4479
4480 static bool mlx5e_tunnel_proto_supported_tx(struct mlx5_core_dev *mdev, u8 proto_type)
4481 {
4482         switch (proto_type) {
4483         case IPPROTO_GRE:
4484                 return MLX5_CAP_ETH(mdev, tunnel_stateless_gre);
4485         case IPPROTO_IPIP:
4486         case IPPROTO_IPV6:
4487                 return (MLX5_CAP_ETH(mdev, tunnel_stateless_ip_over_ip) ||
4488                         MLX5_CAP_ETH(mdev, tunnel_stateless_ip_over_ip_tx));
4489         default:
4490                 return false;
4491         }
4492 }
4493
4494 static bool mlx5e_gre_tunnel_inner_proto_offload_supported(struct mlx5_core_dev *mdev,
4495                                                            struct sk_buff *skb)
4496 {
4497         switch (skb->inner_protocol) {
4498         case htons(ETH_P_IP):
4499         case htons(ETH_P_IPV6):
4500         case htons(ETH_P_TEB):
4501                 return true;
4502         case htons(ETH_P_MPLS_UC):
4503         case htons(ETH_P_MPLS_MC):
4504                 return MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre);
4505         }
4506         return false;
4507 }
4508
4509 static netdev_features_t mlx5e_tunnel_features_check(struct mlx5e_priv *priv,
4510                                                      struct sk_buff *skb,
4511                                                      netdev_features_t features)
4512 {
4513         unsigned int offset = 0;
4514         struct udphdr *udph;
4515         u8 proto;
4516         u16 port;
4517
4518         switch (vlan_get_protocol(skb)) {
4519         case htons(ETH_P_IP):
4520                 proto = ip_hdr(skb)->protocol;
4521                 break;
4522         case htons(ETH_P_IPV6):
4523                 proto = ipv6_find_hdr(skb, &offset, -1, NULL, NULL);
4524                 break;
4525         default:
4526                 goto out;
4527         }
4528
4529         switch (proto) {
4530         case IPPROTO_GRE:
4531                 if (mlx5e_gre_tunnel_inner_proto_offload_supported(priv->mdev, skb))
4532                         return features;
4533                 break;
4534         case IPPROTO_IPIP:
4535         case IPPROTO_IPV6:
4536                 if (mlx5e_tunnel_proto_supported_tx(priv->mdev, IPPROTO_IPIP))
4537                         return features;
4538                 break;
4539         case IPPROTO_UDP:
4540                 udph = udp_hdr(skb);
4541                 port = be16_to_cpu(udph->dest);
4542
4543                 /* Verify if UDP port is being offloaded by HW */
4544                 if (mlx5_vxlan_lookup_port(priv->mdev->vxlan, port))
4545                         return features;
4546
4547 #if IS_ENABLED(CONFIG_GENEVE)
4548                 /* Support Geneve offload for default UDP port */
4549                 if (port == GENEVE_UDP_PORT && mlx5_geneve_tx_allowed(priv->mdev))
4550                         return features;
4551 #endif
4552         }
4553
4554 out:
4555         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
4556         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
4557 }
4558
4559 netdev_features_t mlx5e_features_check(struct sk_buff *skb,
4560                                        struct net_device *netdev,
4561                                        netdev_features_t features)
4562 {
4563         struct mlx5e_priv *priv = netdev_priv(netdev);
4564
4565         features = vlan_features_check(skb, features);
4566         features = vxlan_features_check(skb, features);
4567
4568         if (mlx5e_ipsec_feature_check(skb, netdev, features))
4569                 return features;
4570
4571         /* Validate if the tunneled packet is being offloaded by HW */
4572         if (skb->encapsulation &&
4573             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
4574                 return mlx5e_tunnel_features_check(priv, skb, features);
4575
4576         return features;
4577 }
4578
4579 static void mlx5e_tx_timeout_work(struct work_struct *work)
4580 {
4581         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
4582                                                tx_timeout_work);
4583         struct net_device *netdev = priv->netdev;
4584         int i;
4585
4586         rtnl_lock();
4587         mutex_lock(&priv->state_lock);
4588
4589         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
4590                 goto unlock;
4591
4592         for (i = 0; i < netdev->real_num_tx_queues; i++) {
4593                 struct netdev_queue *dev_queue =
4594                         netdev_get_tx_queue(netdev, i);
4595                 struct mlx5e_txqsq *sq = priv->txq2sq[i];
4596
4597                 if (!netif_xmit_stopped(dev_queue))
4598                         continue;
4599
4600                 if (mlx5e_reporter_tx_timeout(sq))
4601                 /* break if tried to reopened channels */
4602                         break;
4603         }
4604
4605 unlock:
4606         mutex_unlock(&priv->state_lock);
4607         rtnl_unlock();
4608 }
4609
4610 static void mlx5e_tx_timeout(struct net_device *dev, unsigned int txqueue)
4611 {
4612         struct mlx5e_priv *priv = netdev_priv(dev);
4613
4614         netdev_err(dev, "TX timeout detected\n");
4615         queue_work(priv->wq, &priv->tx_timeout_work);
4616 }
4617
4618 static int mlx5e_xdp_allowed(struct mlx5e_priv *priv, struct bpf_prog *prog)
4619 {
4620         struct net_device *netdev = priv->netdev;
4621         struct mlx5e_channels new_channels = {};
4622
4623         if (priv->channels.params.lro_en) {
4624                 netdev_warn(netdev, "can't set XDP while LRO is on, disable LRO first\n");
4625                 return -EINVAL;
4626         }
4627
4628         if (mlx5_fpga_is_ipsec_device(priv->mdev)) {
4629                 netdev_warn(netdev,
4630                             "XDP is not available on Innova cards with IPsec support\n");
4631                 return -EINVAL;
4632         }
4633
4634         new_channels.params = priv->channels.params;
4635         new_channels.params.xdp_prog = prog;
4636
4637         /* No XSK params: AF_XDP can't be enabled yet at the point of setting
4638          * the XDP program.
4639          */
4640         if (!mlx5e_rx_is_linear_skb(&new_channels.params, NULL)) {
4641                 netdev_warn(netdev, "XDP is not allowed with MTU(%d) > %d\n",
4642                             new_channels.params.sw_mtu,
4643                             mlx5e_xdp_max_mtu(&new_channels.params, NULL));
4644                 return -EINVAL;
4645         }
4646
4647         return 0;
4648 }
4649
4650 static void mlx5e_rq_replace_xdp_prog(struct mlx5e_rq *rq, struct bpf_prog *prog)
4651 {
4652         struct bpf_prog *old_prog;
4653
4654         old_prog = rcu_replace_pointer(rq->xdp_prog, prog,
4655                                        lockdep_is_held(&rq->priv->state_lock));
4656         if (old_prog)
4657                 bpf_prog_put(old_prog);
4658 }
4659
4660 static int mlx5e_xdp_set(struct net_device *netdev, struct bpf_prog *prog)
4661 {
4662         struct mlx5e_priv *priv = netdev_priv(netdev);
4663         struct bpf_prog *old_prog;
4664         bool reset, was_opened;
4665         int err = 0;
4666         int i;
4667
4668         mutex_lock(&priv->state_lock);
4669
4670         if (prog) {
4671                 err = mlx5e_xdp_allowed(priv, prog);
4672                 if (err)
4673                         goto unlock;
4674         }
4675
4676         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
4677         /* no need for full reset when exchanging programs */
4678         reset = (!priv->channels.params.xdp_prog || !prog);
4679
4680         if (was_opened && !reset)
4681                 /* num_channels is invariant here, so we can take the
4682                  * batched reference right upfront.
4683                  */
4684                 bpf_prog_add(prog, priv->channels.num);
4685
4686         if (was_opened && reset) {
4687                 struct mlx5e_channels new_channels = {};
4688
4689                 new_channels.params = priv->channels.params;
4690                 new_channels.params.xdp_prog = prog;
4691                 mlx5e_set_rq_type(priv->mdev, &new_channels.params);
4692                 old_prog = priv->channels.params.xdp_prog;
4693
4694                 err = mlx5e_safe_switch_channels(priv, &new_channels, NULL, NULL);
4695                 if (err)
4696                         goto unlock;
4697         } else {
4698                 /* exchange programs, extra prog reference we got from caller
4699                  * as long as we don't fail from this point onwards.
4700                  */
4701                 old_prog = xchg(&priv->channels.params.xdp_prog, prog);
4702         }
4703
4704         if (old_prog)
4705                 bpf_prog_put(old_prog);
4706
4707         if (!was_opened && reset) /* change RQ type according to priv->xdp_prog */
4708                 mlx5e_set_rq_type(priv->mdev, &priv->channels.params);
4709
4710         if (!was_opened || reset)
4711                 goto unlock;
4712
4713         /* exchanging programs w/o reset, we update ref counts on behalf
4714          * of the channels RQs here.
4715          */
4716         for (i = 0; i < priv->channels.num; i++) {
4717                 struct mlx5e_channel *c = priv->channels.c[i];
4718
4719                 mlx5e_rq_replace_xdp_prog(&c->rq, prog);
4720                 if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
4721                         mlx5e_rq_replace_xdp_prog(&c->xskrq, prog);
4722         }
4723
4724 unlock:
4725         mutex_unlock(&priv->state_lock);
4726         return err;
4727 }
4728
4729 static int mlx5e_xdp(struct net_device *dev, struct netdev_bpf *xdp)
4730 {
4731         switch (xdp->command) {
4732         case XDP_SETUP_PROG:
4733                 return mlx5e_xdp_set(dev, xdp->prog);
4734         case XDP_SETUP_XSK_POOL:
4735                 return mlx5e_xsk_setup_pool(dev, xdp->xsk.pool,
4736                                             xdp->xsk.queue_id);
4737         default:
4738                 return -EINVAL;
4739         }
4740 }
4741
4742 #ifdef CONFIG_MLX5_ESWITCH
4743 static int mlx5e_bridge_getlink(struct sk_buff *skb, u32 pid, u32 seq,
4744                                 struct net_device *dev, u32 filter_mask,
4745                                 int nlflags)
4746 {
4747         struct mlx5e_priv *priv = netdev_priv(dev);
4748         struct mlx5_core_dev *mdev = priv->mdev;
4749         u8 mode, setting;
4750         int err;
4751
4752         err = mlx5_eswitch_get_vepa(mdev->priv.eswitch, &setting);
4753         if (err)
4754                 return err;
4755         mode = setting ? BRIDGE_MODE_VEPA : BRIDGE_MODE_VEB;
4756         return ndo_dflt_bridge_getlink(skb, pid, seq, dev,
4757                                        mode,
4758                                        0, 0, nlflags, filter_mask, NULL);
4759 }
4760
4761 static int mlx5e_bridge_setlink(struct net_device *dev, struct nlmsghdr *nlh,
4762                                 u16 flags, struct netlink_ext_ack *extack)
4763 {
4764         struct mlx5e_priv *priv = netdev_priv(dev);
4765         struct mlx5_core_dev *mdev = priv->mdev;
4766         struct nlattr *attr, *br_spec;
4767         u16 mode = BRIDGE_MODE_UNDEF;
4768         u8 setting;
4769         int rem;
4770
4771         br_spec = nlmsg_find_attr(nlh, sizeof(struct ifinfomsg), IFLA_AF_SPEC);
4772         if (!br_spec)
4773                 return -EINVAL;
4774
4775         nla_for_each_nested(attr, br_spec, rem) {
4776                 if (nla_type(attr) != IFLA_BRIDGE_MODE)
4777                         continue;
4778
4779                 if (nla_len(attr) < sizeof(mode))
4780                         return -EINVAL;
4781
4782                 mode = nla_get_u16(attr);
4783                 if (mode > BRIDGE_MODE_VEPA)
4784                         return -EINVAL;
4785
4786                 break;
4787         }
4788
4789         if (mode == BRIDGE_MODE_UNDEF)
4790                 return -EINVAL;
4791
4792         setting = (mode == BRIDGE_MODE_VEPA) ?  1 : 0;
4793         return mlx5_eswitch_set_vepa(mdev->priv.eswitch, setting);
4794 }
4795 #endif
4796
4797 const struct net_device_ops mlx5e_netdev_ops = {
4798         .ndo_open                = mlx5e_open,
4799         .ndo_stop                = mlx5e_close,
4800         .ndo_start_xmit          = mlx5e_xmit,
4801         .ndo_setup_tc            = mlx5e_setup_tc,
4802         .ndo_select_queue        = mlx5e_select_queue,
4803         .ndo_get_stats64         = mlx5e_get_stats,
4804         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
4805         .ndo_set_mac_address     = mlx5e_set_mac,
4806         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
4807         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
4808         .ndo_set_features        = mlx5e_set_features,
4809         .ndo_fix_features        = mlx5e_fix_features,
4810         .ndo_change_mtu          = mlx5e_change_nic_mtu,
4811         .ndo_do_ioctl            = mlx5e_ioctl,
4812         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
4813         .ndo_features_check      = mlx5e_features_check,
4814         .ndo_tx_timeout          = mlx5e_tx_timeout,
4815         .ndo_bpf                 = mlx5e_xdp,
4816         .ndo_xdp_xmit            = mlx5e_xdp_xmit,
4817         .ndo_xsk_wakeup          = mlx5e_xsk_wakeup,
4818 #ifdef CONFIG_MLX5_EN_ARFS
4819         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
4820 #endif
4821 #ifdef CONFIG_MLX5_ESWITCH
4822         .ndo_bridge_setlink      = mlx5e_bridge_setlink,
4823         .ndo_bridge_getlink      = mlx5e_bridge_getlink,
4824
4825         /* SRIOV E-Switch NDOs */
4826         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
4827         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
4828         .ndo_set_vf_spoofchk     = mlx5e_set_vf_spoofchk,
4829         .ndo_set_vf_trust        = mlx5e_set_vf_trust,
4830         .ndo_set_vf_rate         = mlx5e_set_vf_rate,
4831         .ndo_get_vf_config       = mlx5e_get_vf_config,
4832         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
4833         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
4834         .ndo_has_offload_stats   = mlx5e_has_offload_stats,
4835         .ndo_get_offload_stats   = mlx5e_get_offload_stats,
4836 #endif
4837         .ndo_get_devlink_port    = mlx5e_get_devlink_port,
4838 };
4839
4840 void mlx5e_build_default_indir_rqt(u32 *indirection_rqt, int len,
4841                                    int num_channels)
4842 {
4843         int i;
4844
4845         for (i = 0; i < len; i++)
4846                 indirection_rqt[i] = i % num_channels;
4847 }
4848
4849 static bool slow_pci_heuristic(struct mlx5_core_dev *mdev)
4850 {
4851         u32 link_speed = 0;
4852         u32 pci_bw = 0;
4853
4854         mlx5e_port_max_linkspeed(mdev, &link_speed);
4855         pci_bw = pcie_bandwidth_available(mdev->pdev, NULL, NULL, NULL);
4856         mlx5_core_dbg_once(mdev, "Max link speed = %d, PCI BW = %d\n",
4857                            link_speed, pci_bw);
4858
4859 #define MLX5E_SLOW_PCI_RATIO (2)
4860
4861         return link_speed && pci_bw &&
4862                 link_speed > MLX5E_SLOW_PCI_RATIO * pci_bw;
4863 }
4864
4865 static struct dim_cq_moder mlx5e_get_def_tx_moderation(u8 cq_period_mode)
4866 {
4867         struct dim_cq_moder moder;
4868
4869         moder.cq_period_mode = cq_period_mode;
4870         moder.pkts = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
4871         moder.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
4872         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4873                 moder.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE;
4874
4875         return moder;
4876 }
4877
4878 static struct dim_cq_moder mlx5e_get_def_rx_moderation(u8 cq_period_mode)
4879 {
4880         struct dim_cq_moder moder;
4881
4882         moder.cq_period_mode = cq_period_mode;
4883         moder.pkts = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
4884         moder.usec = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
4885         if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
4886                 moder.usec = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE;
4887
4888         return moder;
4889 }
4890
4891 static u8 mlx5_to_net_dim_cq_period_mode(u8 cq_period_mode)
4892 {
4893         return cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE ?
4894                 DIM_CQ_PERIOD_MODE_START_FROM_CQE :
4895                 DIM_CQ_PERIOD_MODE_START_FROM_EQE;
4896 }
4897
4898 void mlx5e_reset_tx_moderation(struct mlx5e_params *params, u8 cq_period_mode)
4899 {
4900         if (params->tx_dim_enabled) {
4901                 u8 dim_period_mode = mlx5_to_net_dim_cq_period_mode(cq_period_mode);
4902
4903                 params->tx_cq_moderation = net_dim_get_def_tx_moderation(dim_period_mode);
4904         } else {
4905                 params->tx_cq_moderation = mlx5e_get_def_tx_moderation(cq_period_mode);
4906         }
4907 }
4908
4909 void mlx5e_reset_rx_moderation(struct mlx5e_params *params, u8 cq_period_mode)
4910 {
4911         if (params->rx_dim_enabled) {
4912                 u8 dim_period_mode = mlx5_to_net_dim_cq_period_mode(cq_period_mode);
4913
4914                 params->rx_cq_moderation = net_dim_get_def_rx_moderation(dim_period_mode);
4915         } else {
4916                 params->rx_cq_moderation = mlx5e_get_def_rx_moderation(cq_period_mode);
4917         }
4918 }
4919
4920 void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4921 {
4922         mlx5e_reset_tx_moderation(params, cq_period_mode);
4923         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_TX_CQE_BASED_MODER,
4924                         params->tx_cq_moderation.cq_period_mode ==
4925                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4926 }
4927
4928 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
4929 {
4930         mlx5e_reset_rx_moderation(params, cq_period_mode);
4931         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_BASED_MODER,
4932                         params->rx_cq_moderation.cq_period_mode ==
4933                                 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
4934 }
4935
4936 static u32 mlx5e_choose_lro_timeout(struct mlx5_core_dev *mdev, u32 wanted_timeout)
4937 {
4938         int i;
4939
4940         /* The supported periods are organized in ascending order */
4941         for (i = 0; i < MLX5E_LRO_TIMEOUT_ARR_SIZE - 1; i++)
4942                 if (MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]) >= wanted_timeout)
4943                         break;
4944
4945         return MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]);
4946 }
4947
4948 void mlx5e_build_rq_params(struct mlx5_core_dev *mdev,
4949                            struct mlx5e_params *params)
4950 {
4951         /* Prefer Striding RQ, unless any of the following holds:
4952          * - Striding RQ configuration is not possible/supported.
4953          * - Slow PCI heuristic.
4954          * - Legacy RQ would use linear SKB while Striding RQ would use non-linear.
4955          *
4956          * No XSK params: checking the availability of striding RQ in general.
4957          */
4958         if (!slow_pci_heuristic(mdev) &&
4959             mlx5e_striding_rq_possible(mdev, params) &&
4960             (mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL) ||
4961              !mlx5e_rx_is_linear_skb(params, NULL)))
4962                 MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ, true);
4963         mlx5e_set_rq_type(mdev, params);
4964         mlx5e_init_rq_type_params(mdev, params);
4965 }
4966
4967 void mlx5e_build_rss_params(struct mlx5e_rss_params *rss_params,
4968                             u16 num_channels)
4969 {
4970         enum mlx5e_traffic_types tt;
4971
4972         rss_params->hfunc = ETH_RSS_HASH_TOP;
4973         netdev_rss_key_fill(rss_params->toeplitz_hash_key,
4974                             sizeof(rss_params->toeplitz_hash_key));
4975         mlx5e_build_default_indir_rqt(rss_params->indirection_rqt,
4976                                       MLX5E_INDIR_RQT_SIZE, num_channels);
4977         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++)
4978                 rss_params->rx_hash_fields[tt] =
4979                         tirc_default_config[tt].rx_hash_fields;
4980 }
4981
4982 void mlx5e_build_nic_params(struct mlx5e_priv *priv, struct mlx5e_xsk *xsk, u16 mtu)
4983 {
4984         struct mlx5e_rss_params *rss_params = &priv->rss_params;
4985         struct mlx5e_params *params = &priv->channels.params;
4986         struct mlx5_core_dev *mdev = priv->mdev;
4987         u8 rx_cq_period_mode;
4988
4989         priv->max_nch = mlx5e_calc_max_nch(priv, priv->profile);
4990
4991         params->sw_mtu = mtu;
4992         params->hard_mtu = MLX5E_ETH_HARD_MTU;
4993         params->num_channels = min_t(unsigned int, MLX5E_MAX_NUM_CHANNELS / 2,
4994                                      priv->max_nch);
4995         params->num_tc       = 1;
4996
4997         /* SQ */
4998         params->log_sq_size = is_kdump_kernel() ?
4999                 MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE :
5000                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
5001         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_SKB_TX_MPWQE,
5002                         MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe));
5003
5004         /* XDP SQ */
5005         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_XDP_TX_MPWQE,
5006                         MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe));
5007
5008         /* set CQE compression */
5009         params->rx_cqe_compress_def = false;
5010         if (MLX5_CAP_GEN(mdev, cqe_compression) &&
5011             MLX5_CAP_GEN(mdev, vport_group_manager))
5012                 params->rx_cqe_compress_def = slow_pci_heuristic(mdev);
5013
5014         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS, params->rx_cqe_compress_def);
5015         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_NO_CSUM_COMPLETE, false);
5016
5017         /* RQ */
5018         mlx5e_build_rq_params(mdev, params);
5019
5020         /* HW LRO */
5021         if (MLX5_CAP_ETH(mdev, lro_cap) &&
5022             params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
5023                 /* No XSK params: checking the availability of striding RQ in general. */
5024                 if (!mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL))
5025                         params->lro_en = !slow_pci_heuristic(mdev);
5026         }
5027         params->lro_timeout = mlx5e_choose_lro_timeout(mdev, MLX5E_DEFAULT_LRO_TIMEOUT);
5028
5029         /* CQ moderation params */
5030         rx_cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
5031                         MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
5032                         MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
5033         params->rx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
5034         params->tx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
5035         mlx5e_set_rx_cq_mode_params(params, rx_cq_period_mode);
5036         mlx5e_set_tx_cq_mode_params(params, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
5037
5038         /* TX inline */
5039         mlx5_query_min_inline(mdev, &params->tx_min_inline_mode);
5040
5041         /* RSS */
5042         mlx5e_build_rss_params(rss_params, params->num_channels);
5043         params->tunneled_offload_en =
5044                 mlx5e_tunnel_inner_ft_supported(mdev);
5045
5046         /* AF_XDP */
5047         params->xsk = xsk;
5048
5049         /* Do not update netdev->features directly in here
5050          * on mlx5e_attach_netdev() we will call mlx5e_update_features()
5051          * To update netdev->features please modify mlx5e_fix_features()
5052          */
5053 }
5054
5055 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
5056 {
5057         struct mlx5e_priv *priv = netdev_priv(netdev);
5058
5059         mlx5_query_mac_address(priv->mdev, netdev->dev_addr);
5060         if (is_zero_ether_addr(netdev->dev_addr) &&
5061             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
5062                 eth_hw_addr_random(netdev);
5063                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
5064         }
5065 }
5066
5067 static int mlx5e_vxlan_set_port(struct net_device *netdev, unsigned int table,
5068                                 unsigned int entry, struct udp_tunnel_info *ti)
5069 {
5070         struct mlx5e_priv *priv = netdev_priv(netdev);
5071
5072         return mlx5_vxlan_add_port(priv->mdev->vxlan, ntohs(ti->port));
5073 }
5074
5075 static int mlx5e_vxlan_unset_port(struct net_device *netdev, unsigned int table,
5076                                   unsigned int entry, struct udp_tunnel_info *ti)
5077 {
5078         struct mlx5e_priv *priv = netdev_priv(netdev);
5079
5080         return mlx5_vxlan_del_port(priv->mdev->vxlan, ntohs(ti->port));
5081 }
5082
5083 void mlx5e_vxlan_set_netdev_info(struct mlx5e_priv *priv)
5084 {
5085         if (!mlx5_vxlan_allowed(priv->mdev->vxlan))
5086                 return;
5087
5088         priv->nic_info.set_port = mlx5e_vxlan_set_port;
5089         priv->nic_info.unset_port = mlx5e_vxlan_unset_port;
5090         priv->nic_info.flags = UDP_TUNNEL_NIC_INFO_MAY_SLEEP |
5091                                 UDP_TUNNEL_NIC_INFO_STATIC_IANA_VXLAN;
5092         priv->nic_info.tables[0].tunnel_types = UDP_TUNNEL_TYPE_VXLAN;
5093         /* Don't count the space hard-coded to the IANA port */
5094         priv->nic_info.tables[0].n_entries =
5095                 mlx5_vxlan_max_udp_ports(priv->mdev) - 1;
5096
5097         priv->netdev->udp_tunnel_nic_info = &priv->nic_info;
5098 }
5099
5100 static bool mlx5e_tunnel_any_tx_proto_supported(struct mlx5_core_dev *mdev)
5101 {
5102         int tt;
5103
5104         for (tt = 0; tt < MLX5E_NUM_TUNNEL_TT; tt++) {
5105                 if (mlx5e_tunnel_proto_supported_tx(mdev, mlx5e_get_proto_by_tunnel_type(tt)))
5106                         return true;
5107         }
5108         return (mlx5_vxlan_allowed(mdev->vxlan) || mlx5_geneve_tx_allowed(mdev));
5109 }
5110
5111 static void mlx5e_build_nic_netdev(struct net_device *netdev)
5112 {
5113         struct mlx5e_priv *priv = netdev_priv(netdev);
5114         struct mlx5_core_dev *mdev = priv->mdev;
5115         bool fcs_supported;
5116         bool fcs_enabled;
5117
5118         SET_NETDEV_DEV(netdev, mdev->device);
5119
5120         netdev->netdev_ops = &mlx5e_netdev_ops;
5121
5122         mlx5e_dcbnl_build_netdev(netdev);
5123
5124         netdev->watchdog_timeo    = 15 * HZ;
5125
5126         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
5127
5128         netdev->vlan_features    |= NETIF_F_SG;
5129         netdev->vlan_features    |= NETIF_F_HW_CSUM;
5130         netdev->vlan_features    |= NETIF_F_GRO;
5131         netdev->vlan_features    |= NETIF_F_TSO;
5132         netdev->vlan_features    |= NETIF_F_TSO6;
5133         netdev->vlan_features    |= NETIF_F_RXCSUM;
5134         netdev->vlan_features    |= NETIF_F_RXHASH;
5135
5136         netdev->mpls_features    |= NETIF_F_SG;
5137         netdev->mpls_features    |= NETIF_F_HW_CSUM;
5138         netdev->mpls_features    |= NETIF_F_TSO;
5139         netdev->mpls_features    |= NETIF_F_TSO6;
5140
5141         netdev->hw_enc_features  |= NETIF_F_HW_VLAN_CTAG_TX;
5142         netdev->hw_enc_features  |= NETIF_F_HW_VLAN_CTAG_RX;
5143
5144         if (!!MLX5_CAP_ETH(mdev, lro_cap) &&
5145             mlx5e_check_fragmented_striding_rq_cap(mdev))
5146                 netdev->vlan_features    |= NETIF_F_LRO;
5147
5148         netdev->hw_features       = netdev->vlan_features;
5149         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
5150         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
5151         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
5152         netdev->hw_features      |= NETIF_F_HW_VLAN_STAG_TX;
5153
5154         if (mlx5e_tunnel_any_tx_proto_supported(mdev)) {
5155                 netdev->hw_enc_features |= NETIF_F_HW_CSUM;
5156                 netdev->hw_enc_features |= NETIF_F_TSO;
5157                 netdev->hw_enc_features |= NETIF_F_TSO6;
5158                 netdev->hw_enc_features |= NETIF_F_GSO_PARTIAL;
5159         }
5160
5161         if (mlx5_vxlan_allowed(mdev->vxlan) || mlx5_geneve_tx_allowed(mdev)) {
5162                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL |
5163                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
5164                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL |
5165                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
5166                 netdev->gso_partial_features = NETIF_F_GSO_UDP_TUNNEL_CSUM;
5167                 netdev->vlan_features |= NETIF_F_GSO_UDP_TUNNEL |
5168                                          NETIF_F_GSO_UDP_TUNNEL_CSUM;
5169         }
5170
5171         if (mlx5e_tunnel_proto_supported_tx(mdev, IPPROTO_GRE)) {
5172                 netdev->hw_features     |= NETIF_F_GSO_GRE |
5173                                            NETIF_F_GSO_GRE_CSUM;
5174                 netdev->hw_enc_features |= NETIF_F_GSO_GRE |
5175                                            NETIF_F_GSO_GRE_CSUM;
5176                 netdev->gso_partial_features |= NETIF_F_GSO_GRE |
5177                                                 NETIF_F_GSO_GRE_CSUM;
5178         }
5179
5180         if (mlx5e_tunnel_proto_supported_tx(mdev, IPPROTO_IPIP)) {
5181                 netdev->hw_features |= NETIF_F_GSO_IPXIP4 |
5182                                        NETIF_F_GSO_IPXIP6;
5183                 netdev->hw_enc_features |= NETIF_F_GSO_IPXIP4 |
5184                                            NETIF_F_GSO_IPXIP6;
5185                 netdev->gso_partial_features |= NETIF_F_GSO_IPXIP4 |
5186                                                 NETIF_F_GSO_IPXIP6;
5187         }
5188
5189         netdev->hw_features                      |= NETIF_F_GSO_PARTIAL;
5190         netdev->gso_partial_features             |= NETIF_F_GSO_UDP_L4;
5191         netdev->hw_features                      |= NETIF_F_GSO_UDP_L4;
5192         netdev->features                         |= NETIF_F_GSO_UDP_L4;
5193
5194         mlx5_query_port_fcs(mdev, &fcs_supported, &fcs_enabled);
5195
5196         if (fcs_supported)
5197                 netdev->hw_features |= NETIF_F_RXALL;
5198
5199         if (MLX5_CAP_ETH(mdev, scatter_fcs))
5200                 netdev->hw_features |= NETIF_F_RXFCS;
5201
5202         netdev->features          = netdev->hw_features;
5203
5204         /* Defaults */
5205         if (fcs_enabled)
5206                 netdev->features  &= ~NETIF_F_RXALL;
5207         netdev->features  &= ~NETIF_F_LRO;
5208         netdev->features  &= ~NETIF_F_RXFCS;
5209
5210 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
5211         if (FT_CAP(flow_modify_en) &&
5212             FT_CAP(modify_root) &&
5213             FT_CAP(identified_miss_table_mode) &&
5214             FT_CAP(flow_table_modify)) {
5215 #if IS_ENABLED(CONFIG_MLX5_CLS_ACT)
5216                 netdev->hw_features      |= NETIF_F_HW_TC;
5217 #endif
5218 #ifdef CONFIG_MLX5_EN_ARFS
5219                 netdev->hw_features      |= NETIF_F_NTUPLE;
5220 #endif
5221         }
5222         if (mlx5_qos_is_supported(mdev))
5223                 netdev->features |= NETIF_F_HW_TC;
5224
5225         netdev->features         |= NETIF_F_HIGHDMA;
5226         netdev->features         |= NETIF_F_HW_VLAN_STAG_FILTER;
5227
5228         netdev->priv_flags       |= IFF_UNICAST_FLT;
5229
5230         mlx5e_set_netdev_dev_addr(netdev);
5231         mlx5e_ipsec_build_netdev(priv);
5232         mlx5e_tls_build_netdev(priv);
5233 }
5234
5235 void mlx5e_create_q_counters(struct mlx5e_priv *priv)
5236 {
5237         u32 out[MLX5_ST_SZ_DW(alloc_q_counter_out)] = {};
5238         u32 in[MLX5_ST_SZ_DW(alloc_q_counter_in)] = {};
5239         struct mlx5_core_dev *mdev = priv->mdev;
5240         int err;
5241
5242         MLX5_SET(alloc_q_counter_in, in, opcode, MLX5_CMD_OP_ALLOC_Q_COUNTER);
5243         err = mlx5_cmd_exec_inout(mdev, alloc_q_counter, in, out);
5244         if (!err)
5245                 priv->q_counter =
5246                         MLX5_GET(alloc_q_counter_out, out, counter_set_id);
5247
5248         err = mlx5_cmd_exec_inout(mdev, alloc_q_counter, in, out);
5249         if (!err)
5250                 priv->drop_rq_q_counter =
5251                         MLX5_GET(alloc_q_counter_out, out, counter_set_id);
5252 }
5253
5254 void mlx5e_destroy_q_counters(struct mlx5e_priv *priv)
5255 {
5256         u32 in[MLX5_ST_SZ_DW(dealloc_q_counter_in)] = {};
5257
5258         MLX5_SET(dealloc_q_counter_in, in, opcode,
5259                  MLX5_CMD_OP_DEALLOC_Q_COUNTER);
5260         if (priv->q_counter) {
5261                 MLX5_SET(dealloc_q_counter_in, in, counter_set_id,
5262                          priv->q_counter);
5263                 mlx5_cmd_exec_in(priv->mdev, dealloc_q_counter, in);
5264         }
5265
5266         if (priv->drop_rq_q_counter) {
5267                 MLX5_SET(dealloc_q_counter_in, in, counter_set_id,
5268                          priv->drop_rq_q_counter);
5269                 mlx5_cmd_exec_in(priv->mdev, dealloc_q_counter, in);
5270         }
5271 }
5272
5273 static int mlx5e_nic_init(struct mlx5_core_dev *mdev,
5274                           struct net_device *netdev)
5275 {
5276         struct mlx5e_priv *priv = netdev_priv(netdev);
5277         int err;
5278
5279         mlx5e_build_nic_params(priv, &priv->xsk, netdev->mtu);
5280         mlx5e_vxlan_set_netdev_info(priv);
5281
5282         mlx5e_timestamp_init(priv);
5283
5284         err = mlx5e_ipsec_init(priv);
5285         if (err)
5286                 mlx5_core_err(mdev, "IPSec initialization failed, %d\n", err);
5287
5288         err = mlx5e_tls_init(priv);
5289         if (err)
5290                 mlx5_core_err(mdev, "TLS initialization failed, %d\n", err);
5291
5292         err = mlx5e_devlink_port_register(priv);
5293         if (err)
5294                 mlx5_core_err(mdev, "mlx5e_devlink_port_register failed, %d\n", err);
5295
5296         mlx5e_health_create_reporters(priv);
5297
5298         return 0;
5299 }
5300
5301 static void mlx5e_nic_cleanup(struct mlx5e_priv *priv)
5302 {
5303         mlx5e_health_destroy_reporters(priv);
5304         mlx5e_devlink_port_unregister(priv);
5305         mlx5e_tls_cleanup(priv);
5306         mlx5e_ipsec_cleanup(priv);
5307 }
5308
5309 static int mlx5e_init_nic_rx(struct mlx5e_priv *priv)
5310 {
5311         struct mlx5_core_dev *mdev = priv->mdev;
5312         int err;
5313
5314         mlx5e_create_q_counters(priv);
5315
5316         err = mlx5e_open_drop_rq(priv, &priv->drop_rq);
5317         if (err) {
5318                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
5319                 goto err_destroy_q_counters;
5320         }
5321
5322         err = mlx5e_create_indirect_rqt(priv);
5323         if (err)
5324                 goto err_close_drop_rq;
5325
5326         err = mlx5e_create_direct_rqts(priv, priv->direct_tir);
5327         if (err)
5328                 goto err_destroy_indirect_rqts;
5329
5330         err = mlx5e_create_indirect_tirs(priv, true);
5331         if (err)
5332                 goto err_destroy_direct_rqts;
5333
5334         err = mlx5e_create_direct_tirs(priv, priv->direct_tir);
5335         if (err)
5336                 goto err_destroy_indirect_tirs;
5337
5338         err = mlx5e_create_direct_rqts(priv, priv->xsk_tir);
5339         if (unlikely(err))
5340                 goto err_destroy_direct_tirs;
5341
5342         err = mlx5e_create_direct_tirs(priv, priv->xsk_tir);
5343         if (unlikely(err))
5344                 goto err_destroy_xsk_rqts;
5345
5346         err = mlx5e_create_flow_steering(priv);
5347         if (err) {
5348                 mlx5_core_warn(mdev, "create flow steering failed, %d\n", err);
5349                 goto err_destroy_xsk_tirs;
5350         }
5351
5352         err = mlx5e_tc_nic_init(priv);
5353         if (err)
5354                 goto err_destroy_flow_steering;
5355
5356         err = mlx5e_accel_init_rx(priv);
5357         if (err)
5358                 goto err_tc_nic_cleanup;
5359
5360 #ifdef CONFIG_MLX5_EN_ARFS
5361         priv->netdev->rx_cpu_rmap =  mlx5_eq_table_get_rmap(priv->mdev);
5362 #endif
5363
5364         return 0;
5365
5366 err_tc_nic_cleanup:
5367         mlx5e_tc_nic_cleanup(priv);
5368 err_destroy_flow_steering:
5369         mlx5e_destroy_flow_steering(priv);
5370 err_destroy_xsk_tirs:
5371         mlx5e_destroy_direct_tirs(priv, priv->xsk_tir);
5372 err_destroy_xsk_rqts:
5373         mlx5e_destroy_direct_rqts(priv, priv->xsk_tir);
5374 err_destroy_direct_tirs:
5375         mlx5e_destroy_direct_tirs(priv, priv->direct_tir);
5376 err_destroy_indirect_tirs:
5377         mlx5e_destroy_indirect_tirs(priv);
5378 err_destroy_direct_rqts:
5379         mlx5e_destroy_direct_rqts(priv, priv->direct_tir);
5380 err_destroy_indirect_rqts:
5381         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
5382 err_close_drop_rq:
5383         mlx5e_close_drop_rq(&priv->drop_rq);
5384 err_destroy_q_counters:
5385         mlx5e_destroy_q_counters(priv);
5386         return err;
5387 }
5388
5389 static void mlx5e_cleanup_nic_rx(struct mlx5e_priv *priv)
5390 {
5391         mlx5e_accel_cleanup_rx(priv);
5392         mlx5e_tc_nic_cleanup(priv);
5393         mlx5e_destroy_flow_steering(priv);
5394         mlx5e_destroy_direct_tirs(priv, priv->xsk_tir);
5395         mlx5e_destroy_direct_rqts(priv, priv->xsk_tir);
5396         mlx5e_destroy_direct_tirs(priv, priv->direct_tir);
5397         mlx5e_destroy_indirect_tirs(priv);
5398         mlx5e_destroy_direct_rqts(priv, priv->direct_tir);
5399         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
5400         mlx5e_close_drop_rq(&priv->drop_rq);
5401         mlx5e_destroy_q_counters(priv);
5402 }
5403
5404 static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
5405 {
5406         int err;
5407
5408         err = mlx5e_create_tises(priv);
5409         if (err) {
5410                 mlx5_core_warn(priv->mdev, "create tises failed, %d\n", err);
5411                 return err;
5412         }
5413
5414         mlx5e_dcbnl_initialize(priv);
5415         return 0;
5416 }
5417
5418 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
5419 {
5420         struct net_device *netdev = priv->netdev;
5421         struct mlx5_core_dev *mdev = priv->mdev;
5422
5423         mlx5e_init_l2_addr(priv);
5424
5425         /* Marking the link as currently not needed by the Driver */
5426         if (!netif_running(netdev))
5427                 mlx5e_modify_admin_state(mdev, MLX5_PORT_DOWN);
5428
5429         mlx5e_set_netdev_mtu_boundaries(priv);
5430         mlx5e_set_dev_port_mtu(priv);
5431
5432         mlx5_lag_add(mdev, netdev);
5433
5434         mlx5e_enable_async_events(priv);
5435         mlx5e_enable_blocking_events(priv);
5436         if (mlx5e_monitor_counter_supported(priv))
5437                 mlx5e_monitor_counter_init(priv);
5438
5439         mlx5e_hv_vhca_stats_create(priv);
5440         if (netdev->reg_state != NETREG_REGISTERED)
5441                 return;
5442         mlx5e_dcbnl_init_app(priv);
5443
5444         mlx5e_nic_set_rx_mode(priv);
5445
5446         rtnl_lock();
5447         if (netif_running(netdev))
5448                 mlx5e_open(netdev);
5449         udp_tunnel_nic_reset_ntf(priv->netdev);
5450         netif_device_attach(netdev);
5451         rtnl_unlock();
5452 }
5453
5454 static void mlx5e_nic_disable(struct mlx5e_priv *priv)
5455 {
5456         struct mlx5_core_dev *mdev = priv->mdev;
5457
5458         if (priv->netdev->reg_state == NETREG_REGISTERED)
5459                 mlx5e_dcbnl_delete_app(priv);
5460
5461         rtnl_lock();
5462         if (netif_running(priv->netdev))
5463                 mlx5e_close(priv->netdev);
5464         netif_device_detach(priv->netdev);
5465         rtnl_unlock();
5466
5467         mlx5e_nic_set_rx_mode(priv);
5468
5469         mlx5e_hv_vhca_stats_destroy(priv);
5470         if (mlx5e_monitor_counter_supported(priv))
5471                 mlx5e_monitor_counter_cleanup(priv);
5472
5473         mlx5e_disable_blocking_events(priv);
5474         if (priv->en_trap) {
5475                 mlx5e_deactivate_trap(priv);
5476                 mlx5e_close_trap(priv->en_trap);
5477                 priv->en_trap = NULL;
5478         }
5479         mlx5e_disable_async_events(priv);
5480         mlx5_lag_remove(mdev);
5481         mlx5_vxlan_reset_to_default(mdev->vxlan);
5482 }
5483
5484 int mlx5e_update_nic_rx(struct mlx5e_priv *priv)
5485 {
5486         return mlx5e_refresh_tirs(priv, false, false);
5487 }
5488
5489 static const struct mlx5e_profile mlx5e_nic_profile = {
5490         .init              = mlx5e_nic_init,
5491         .cleanup           = mlx5e_nic_cleanup,
5492         .init_rx           = mlx5e_init_nic_rx,
5493         .cleanup_rx        = mlx5e_cleanup_nic_rx,
5494         .init_tx           = mlx5e_init_nic_tx,
5495         .cleanup_tx        = mlx5e_cleanup_nic_tx,
5496         .enable            = mlx5e_nic_enable,
5497         .disable           = mlx5e_nic_disable,
5498         .update_rx         = mlx5e_update_nic_rx,
5499         .update_stats      = mlx5e_stats_update_ndo_stats,
5500         .update_carrier    = mlx5e_update_carrier,
5501         .rx_handlers       = &mlx5e_rx_handlers_nic,
5502         .max_tc            = MLX5E_MAX_NUM_TC,
5503         .rq_groups         = MLX5E_NUM_RQ_GROUPS(XSK),
5504         .stats_grps        = mlx5e_nic_stats_grps,
5505         .stats_grps_num    = mlx5e_nic_stats_grps_num,
5506 };
5507
5508 /* mlx5e generic netdev management API (move to en_common.c) */
5509 int mlx5e_priv_init(struct mlx5e_priv *priv,
5510                     struct net_device *netdev,
5511                     struct mlx5_core_dev *mdev)
5512 {
5513         memset(priv, 0, sizeof(*priv));
5514
5515         /* priv init */
5516         priv->mdev        = mdev;
5517         priv->netdev      = netdev;
5518         priv->msglevel    = MLX5E_MSG_LEVEL;
5519         priv->max_opened_tc = 1;
5520
5521         if (!alloc_cpumask_var(&priv->scratchpad.cpumask, GFP_KERNEL))
5522                 return -ENOMEM;
5523
5524         mutex_init(&priv->state_lock);
5525         hash_init(priv->htb.qos_tc2node);
5526         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
5527         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
5528         INIT_WORK(&priv->tx_timeout_work, mlx5e_tx_timeout_work);
5529         INIT_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
5530
5531         priv->wq = create_singlethread_workqueue("mlx5e");
5532         if (!priv->wq)
5533                 goto err_free_cpumask;
5534
5535         return 0;
5536
5537 err_free_cpumask:
5538         free_cpumask_var(priv->scratchpad.cpumask);
5539
5540         return -ENOMEM;
5541 }
5542
5543 void mlx5e_priv_cleanup(struct mlx5e_priv *priv)
5544 {
5545         int i;
5546
5547         destroy_workqueue(priv->wq);
5548         free_cpumask_var(priv->scratchpad.cpumask);
5549
5550         for (i = 0; i < priv->htb.max_qos_sqs; i++)
5551                 kfree(priv->htb.qos_sq_stats[i]);
5552         kvfree(priv->htb.qos_sq_stats);
5553 }
5554
5555 struct net_device *
5556 mlx5e_create_netdev(struct mlx5_core_dev *mdev, unsigned int txqs, unsigned int rxqs)
5557 {
5558         struct net_device *netdev;
5559         int err;
5560
5561         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv), txqs, rxqs);
5562         if (!netdev) {
5563                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
5564                 return NULL;
5565         }
5566
5567         err = mlx5e_priv_init(netdev_priv(netdev), netdev, mdev);
5568         if (err) {
5569                 mlx5_core_err(mdev, "mlx5e_priv_init failed, err=%d\n", err);
5570                 goto err_free_netdev;
5571         }
5572
5573         netif_carrier_off(netdev);
5574         dev_net_set(netdev, mlx5_core_net(mdev));
5575
5576         return netdev;
5577
5578 err_free_netdev:
5579         free_netdev(netdev);
5580
5581         return NULL;
5582 }
5583
5584 static void mlx5e_update_features(struct net_device *netdev)
5585 {
5586         if (netdev->reg_state != NETREG_REGISTERED)
5587                 return; /* features will be updated on netdev registration */
5588
5589         rtnl_lock();
5590         netdev_update_features(netdev);
5591         rtnl_unlock();
5592 }
5593
5594 int mlx5e_attach_netdev(struct mlx5e_priv *priv)
5595 {
5596         const bool take_rtnl = priv->netdev->reg_state == NETREG_REGISTERED;
5597         const struct mlx5e_profile *profile = priv->profile;
5598         int max_nch;
5599         int err;
5600
5601         clear_bit(MLX5E_STATE_DESTROYING, &priv->state);
5602
5603         /* max number of channels may have changed */
5604         max_nch = mlx5e_get_max_num_channels(priv->mdev);
5605         if (priv->channels.params.num_channels > max_nch) {
5606                 mlx5_core_warn(priv->mdev, "MLX5E: Reducing number of channels to %d\n", max_nch);
5607                 /* Reducing the number of channels - RXFH has to be reset, and
5608                  * mlx5e_num_channels_changed below will build the RQT.
5609                  */
5610                 priv->netdev->priv_flags &= ~IFF_RXFH_CONFIGURED;
5611                 priv->channels.params.num_channels = max_nch;
5612         }
5613         /* 1. Set the real number of queues in the kernel the first time.
5614          * 2. Set our default XPS cpumask.
5615          * 3. Build the RQT.
5616          *
5617          * rtnl_lock is required by netif_set_real_num_*_queues in case the
5618          * netdev has been registered by this point (if this function was called
5619          * in the reload or resume flow).
5620          */
5621         if (take_rtnl)
5622                 rtnl_lock();
5623         err = mlx5e_num_channels_changed(priv);
5624         if (take_rtnl)
5625                 rtnl_unlock();
5626         if (err)
5627                 goto out;
5628
5629         err = profile->init_tx(priv);
5630         if (err)
5631                 goto out;
5632
5633         err = profile->init_rx(priv);
5634         if (err)
5635                 goto err_cleanup_tx;
5636
5637         if (profile->enable)
5638                 profile->enable(priv);
5639
5640         mlx5e_update_features(priv->netdev);
5641
5642         return 0;
5643
5644 err_cleanup_tx:
5645         profile->cleanup_tx(priv);
5646
5647 out:
5648         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
5649         cancel_work_sync(&priv->update_stats_work);
5650         return err;
5651 }
5652
5653 void mlx5e_detach_netdev(struct mlx5e_priv *priv)
5654 {
5655         const struct mlx5e_profile *profile = priv->profile;
5656
5657         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
5658
5659         if (profile->disable)
5660                 profile->disable(priv);
5661         flush_workqueue(priv->wq);
5662
5663         profile->cleanup_rx(priv);
5664         profile->cleanup_tx(priv);
5665         cancel_work_sync(&priv->update_stats_work);
5666 }
5667
5668 static int
5669 mlx5e_netdev_attach_profile(struct mlx5e_priv *priv,
5670                             const struct mlx5e_profile *new_profile, void *new_ppriv)
5671 {
5672         struct net_device *netdev = priv->netdev;
5673         struct mlx5_core_dev *mdev = priv->mdev;
5674         int err;
5675
5676         err = mlx5e_priv_init(priv, netdev, mdev);
5677         if (err) {
5678                 mlx5_core_err(mdev, "mlx5e_priv_init failed, err=%d\n", err);
5679                 return err;
5680         }
5681         netif_carrier_off(netdev);
5682         priv->profile = new_profile;
5683         priv->ppriv = new_ppriv;
5684         err = new_profile->init(priv->mdev, priv->netdev);
5685         if (err)
5686                 return err;
5687         err = mlx5e_attach_netdev(priv);
5688         if (err)
5689                 new_profile->cleanup(priv);
5690         return err;
5691 }
5692
5693 int mlx5e_netdev_change_profile(struct mlx5e_priv *priv,
5694                                 const struct mlx5e_profile *new_profile, void *new_ppriv)
5695 {
5696         unsigned int new_max_nch = mlx5e_calc_max_nch(priv, new_profile);
5697         const struct mlx5e_profile *orig_profile = priv->profile;
5698         void *orig_ppriv = priv->ppriv;
5699         int err, rollback_err;
5700
5701         /* sanity */
5702         if (new_max_nch != priv->max_nch) {
5703                 netdev_warn(priv->netdev,
5704                             "%s: Replacing profile with different max channels\n",
5705                             __func__);
5706                 return -EINVAL;
5707         }
5708
5709         /* cleanup old profile */
5710         mlx5e_detach_netdev(priv);
5711         priv->profile->cleanup(priv);
5712         mlx5e_priv_cleanup(priv);
5713
5714         err = mlx5e_netdev_attach_profile(priv, new_profile, new_ppriv);
5715         if (err) { /* roll back to original profile */
5716                 netdev_warn(priv->netdev, "%s: new profile init failed, %d\n",
5717                             __func__, err);
5718                 goto rollback;
5719         }
5720
5721         return 0;
5722
5723 rollback:
5724         rollback_err = mlx5e_netdev_attach_profile(priv, orig_profile, orig_ppriv);
5725         if (rollback_err) {
5726                 netdev_err(priv->netdev,
5727                            "%s: failed to rollback to orig profile, %d\n",
5728                            __func__, rollback_err);
5729         }
5730         return err;
5731 }
5732
5733 void mlx5e_destroy_netdev(struct mlx5e_priv *priv)
5734 {
5735         struct net_device *netdev = priv->netdev;
5736
5737         mlx5e_priv_cleanup(priv);
5738         free_netdev(netdev);
5739 }
5740
5741 static int mlx5e_resume(struct auxiliary_device *adev)
5742 {
5743         struct mlx5_adev *edev = container_of(adev, struct mlx5_adev, adev);
5744         struct mlx5e_priv *priv = dev_get_drvdata(&adev->dev);
5745         struct net_device *netdev = priv->netdev;
5746         struct mlx5_core_dev *mdev = edev->mdev;
5747         int err;
5748
5749         if (netif_device_present(netdev))
5750                 return 0;
5751
5752         err = mlx5e_create_mdev_resources(mdev);
5753         if (err)
5754                 return err;
5755
5756         err = mlx5e_attach_netdev(priv);
5757         if (err) {
5758                 mlx5e_destroy_mdev_resources(mdev);
5759                 return err;
5760         }
5761
5762         return 0;
5763 }
5764
5765 static int mlx5e_suspend(struct auxiliary_device *adev, pm_message_t state)
5766 {
5767         struct mlx5e_priv *priv = dev_get_drvdata(&adev->dev);
5768         struct net_device *netdev = priv->netdev;
5769         struct mlx5_core_dev *mdev = priv->mdev;
5770
5771         if (!netif_device_present(netdev))
5772                 return -ENODEV;
5773
5774         mlx5e_detach_netdev(priv);
5775         mlx5e_destroy_mdev_resources(mdev);
5776         return 0;
5777 }
5778
5779 static int mlx5e_probe(struct auxiliary_device *adev,
5780                        const struct auxiliary_device_id *id)
5781 {
5782         struct mlx5_adev *edev = container_of(adev, struct mlx5_adev, adev);
5783         const struct mlx5e_profile *profile = &mlx5e_nic_profile;
5784         struct mlx5_core_dev *mdev = edev->mdev;
5785         struct net_device *netdev;
5786         pm_message_t state = {};
5787         unsigned int txqs, rxqs, ptp_txqs = 0;
5788         struct mlx5e_priv *priv;
5789         int qos_sqs = 0;
5790         int err;
5791         int nch;
5792
5793         if (MLX5_CAP_GEN(mdev, ts_cqe_to_dest_cqn))
5794                 ptp_txqs = profile->max_tc;
5795
5796         if (mlx5_qos_is_supported(mdev))
5797                 qos_sqs = mlx5e_qos_max_leaf_nodes(mdev);
5798
5799         nch = mlx5e_get_max_num_channels(mdev);
5800         txqs = nch * profile->max_tc + ptp_txqs + qos_sqs;
5801         rxqs = nch * profile->rq_groups;
5802         netdev = mlx5e_create_netdev(mdev, txqs, rxqs);
5803         if (!netdev) {
5804                 mlx5_core_err(mdev, "mlx5e_create_netdev failed\n");
5805                 return -ENOMEM;
5806         }
5807
5808         mlx5e_build_nic_netdev(netdev);
5809
5810         priv = netdev_priv(netdev);
5811         dev_set_drvdata(&adev->dev, priv);
5812
5813         priv->profile = profile;
5814         priv->ppriv = NULL;
5815         err = profile->init(mdev, netdev);
5816         if (err) {
5817                 mlx5_core_err(mdev, "mlx5e_nic_profile init failed, %d\n", err);
5818                 goto err_destroy_netdev;
5819         }
5820
5821         err = mlx5e_resume(adev);
5822         if (err) {
5823                 mlx5_core_err(mdev, "mlx5e_resume failed, %d\n", err);
5824                 goto err_profile_cleanup;
5825         }
5826
5827         err = register_netdev(netdev);
5828         if (err) {
5829                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
5830                 goto err_resume;
5831         }
5832
5833         mlx5e_devlink_port_type_eth_set(priv);
5834
5835         mlx5e_dcbnl_init_app(priv);
5836         return 0;
5837
5838 err_resume:
5839         mlx5e_suspend(adev, state);
5840 err_profile_cleanup:
5841         profile->cleanup(priv);
5842 err_destroy_netdev:
5843         mlx5e_destroy_netdev(priv);
5844         return err;
5845 }
5846
5847 static void mlx5e_remove(struct auxiliary_device *adev)
5848 {
5849         struct mlx5e_priv *priv = dev_get_drvdata(&adev->dev);
5850         pm_message_t state = {};
5851
5852         mlx5e_dcbnl_delete_app(priv);
5853         unregister_netdev(priv->netdev);
5854         mlx5e_suspend(adev, state);
5855         priv->profile->cleanup(priv);
5856         mlx5e_destroy_netdev(priv);
5857 }
5858
5859 static const struct auxiliary_device_id mlx5e_id_table[] = {
5860         { .name = MLX5_ADEV_NAME ".eth", },
5861         {},
5862 };
5863
5864 MODULE_DEVICE_TABLE(auxiliary, mlx5e_id_table);
5865
5866 static struct auxiliary_driver mlx5e_driver = {
5867         .name = "eth",
5868         .probe = mlx5e_probe,
5869         .remove = mlx5e_remove,
5870         .suspend = mlx5e_suspend,
5871         .resume = mlx5e_resume,
5872         .id_table = mlx5e_id_table,
5873 };
5874
5875 int mlx5e_init(void)
5876 {
5877         int ret;
5878
5879         mlx5e_ipsec_build_inverse_table();
5880         mlx5e_build_ptys2ethtool_map();
5881         ret = mlx5e_rep_init();
5882         if (ret)
5883                 return ret;
5884
5885         ret = auxiliary_driver_register(&mlx5e_driver);
5886         if (ret)
5887                 mlx5e_rep_cleanup();
5888         return ret;
5889 }
5890
5891 void mlx5e_cleanup(void)
5892 {
5893         auxiliary_driver_unregister(&mlx5e_driver);
5894         mlx5e_rep_cleanup();
5895 }