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_update_carrier(struct mlx5e_priv *priv)
  91 {
  92         struct mlx5_core_dev *mdev = priv->mdev;
  93         u8 port_state;
  94
  95         port_state = mlx5_query_vport_state(mdev,
  96                                             MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT,
  97                                             0);
  98
  99         if (port_state == VPORT_STATE_UP) {
 100                 netdev_info(priv->netdev, "Link up\n");
 101                 netif_carrier_on(priv->netdev);
 102         } else {
 103                 netdev_info(priv->netdev, "Link down\n");
 104                 netif_carrier_off(priv->netdev);
 105         }
 106 }
 107
 108 static void mlx5e_update_carrier_work(struct work_struct *work)
 109 {
 110         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 111                                                update_carrier_work);
 112
 113         mutex_lock(&priv->state_lock);
 114         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
 115                 if (priv->profile->update_carrier)
 116                         priv->profile->update_carrier(priv);
 117         mutex_unlock(&priv->state_lock);
 118 }
 119
 120 static void mlx5e_update_stats_work(struct work_struct *work)
 121 {
 122         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 123                                                update_stats_work);
 124
 125         mutex_lock(&priv->state_lock);
 126         priv->profile->update_stats(priv);
 127         mutex_unlock(&priv->state_lock);
 128 }
 129
 130 void mlx5e_queue_update_stats(struct mlx5e_priv *priv)
 131 {
 132         if (!priv->profile->update_stats)
 133                 return;
 134
 135         if (unlikely(test_bit(MLX5E_STATE_DESTROYING, &priv->state)))
 136                 return;
 137
 138         queue_work(priv->wq, &priv->update_stats_work);
 139 }
 140
 141 static int async_event(struct notifier_block *nb, unsigned long event, void *data)
 142 {
 143         struct mlx5e_priv *priv = container_of(nb, struct mlx5e_priv, events_nb);
 144         struct mlx5_eqe   *eqe = data;
 145
 146         if (event != MLX5_EVENT_TYPE_PORT_CHANGE)
 147                 return NOTIFY_DONE;
 148
 149         switch (eqe->sub_type) {
 150         case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
 151         case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
 152                 queue_work(priv->wq, &priv->update_carrier_work);
 153                 break;
 154         default:
 155                 return NOTIFY_DONE;
 156         }
 157
 158         return NOTIFY_OK;
 159 }
 160
 161 static void mlx5e_enable_async_events(struct mlx5e_priv *priv)
 162 {
 163         priv->events_nb.notifier_call = async_event;
 164         mlx5_notifier_register(priv->mdev, &priv->events_nb);
 165 }
 166
 167 static void mlx5e_disable_async_events(struct mlx5e_priv *priv)
 168 {
 169         mlx5_notifier_unregister(priv->mdev, &priv->events_nb);
 170 }
 171
 172 static int blocking_event(struct notifier_block *nb, unsigned long event, void *data)
 173 {
 174         struct mlx5e_priv *priv = container_of(nb, struct mlx5e_priv, blocking_events_nb);
 175         int err;
 176
 177         switch (event) {
 178         case MLX5_DRIVER_EVENT_TYPE_TRAP:
 179                 err = mlx5e_handle_trap_event(priv, data);
 180                 break;
 181         default:
 182                 netdev_warn(priv->netdev, "Sync event: Unknown event %ld\n", event);
 183                 err = -EINVAL;
 184         }
 185         return err;
 186 }
 187
 188 static void mlx5e_enable_blocking_events(struct mlx5e_priv *priv)
 189 {
 190         priv->blocking_events_nb.notifier_call = blocking_event;
 191         mlx5_blocking_notifier_register(priv->mdev, &priv->blocking_events_nb);
 192 }
 193
 194 static void mlx5e_disable_blocking_events(struct mlx5e_priv *priv)
 195 {
 196         mlx5_blocking_notifier_unregister(priv->mdev, &priv->blocking_events_nb);
 197 }
 198
 199 static inline void mlx5e_build_umr_wqe(struct mlx5e_rq *rq,
 200                                        struct mlx5e_icosq *sq,
 201                                        struct mlx5e_umr_wqe *wqe)
 202 {
 203         struct mlx5_wqe_ctrl_seg      *cseg = &wqe->ctrl;
 204         struct mlx5_wqe_umr_ctrl_seg *ucseg = &wqe->uctrl;
 205         u8 ds_cnt = DIV_ROUND_UP(MLX5E_UMR_WQE_INLINE_SZ, MLX5_SEND_WQE_DS);
 206
 207         cseg->qpn_ds    = cpu_to_be32((sq->sqn << MLX5_WQE_CTRL_QPN_SHIFT) |
 208                                       ds_cnt);
 209         cseg->umr_mkey  = rq->mkey_be;
 210
 211         ucseg->flags = MLX5_UMR_TRANSLATION_OFFSET_EN | MLX5_UMR_INLINE;
 212         ucseg->xlt_octowords =
 213                 cpu_to_be16(MLX5_MTT_OCTW(MLX5_MPWRQ_PAGES_PER_WQE));
 214         ucseg->mkey_mask     = cpu_to_be64(MLX5_MKEY_MASK_FREE);
 215 }
 216
 217 static int mlx5e_rq_alloc_mpwqe_info(struct mlx5e_rq *rq, int node)
 218 {
 219         int wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 220
 221         rq->mpwqe.info = kvzalloc_node(array_size(wq_sz,
 222                                                   sizeof(*rq->mpwqe.info)),
 223                                        GFP_KERNEL, node);
 224         if (!rq->mpwqe.info)
 225                 return -ENOMEM;
 226
 227         mlx5e_build_umr_wqe(rq, rq->icosq, &rq->mpwqe.umr_wqe);
 228
 229         return 0;
 230 }
 231
 232 static int mlx5e_create_umr_mkey(struct mlx5_core_dev *mdev,
 233                                  u64 npages, u8 page_shift,
 234                                  struct mlx5_core_mkey *umr_mkey,
 235                                  dma_addr_t filler_addr)
 236 {
 237         struct mlx5_mtt *mtt;
 238         int inlen;
 239         void *mkc;
 240         u32 *in;
 241         int err;
 242         int i;
 243
 244         inlen = MLX5_ST_SZ_BYTES(create_mkey_in) + sizeof(*mtt) * npages;
 245
 246         in = kvzalloc(inlen, GFP_KERNEL);
 247         if (!in)
 248                 return -ENOMEM;
 249
 250         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 251
 252         MLX5_SET(mkc, mkc, free, 1);
 253         MLX5_SET(mkc, mkc, umr_en, 1);
 254         MLX5_SET(mkc, mkc, lw, 1);
 255         MLX5_SET(mkc, mkc, lr, 1);
 256         MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT);
 257         mlx5e_mkey_set_relaxed_ordering(mdev, mkc);
 258         MLX5_SET(mkc, mkc, qpn, 0xffffff);
 259         MLX5_SET(mkc, mkc, pd, mdev->mlx5e_res.hw_objs.pdn);
 260         MLX5_SET64(mkc, mkc, len, npages << page_shift);
 261         MLX5_SET(mkc, mkc, translations_octword_size,
 262                  MLX5_MTT_OCTW(npages));
 263         MLX5_SET(mkc, mkc, log_page_size, page_shift);
 264         MLX5_SET(create_mkey_in, in, translations_octword_actual_size,
 265                  MLX5_MTT_OCTW(npages));
 266
 267         /* Initialize the mkey with all MTTs pointing to a default
 268          * page (filler_addr). When the channels are activated, UMR
 269          * WQEs will redirect the RX WQEs to the actual memory from
 270          * the RQ's pool, while the gaps (wqe_overflow) remain mapped
 271          * to the default page.
 272          */
 273         mtt = MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt);
 274         for (i = 0 ; i < npages ; i++)
 275                 mtt[i].ptag = cpu_to_be64(filler_addr);
 276
 277         err = mlx5_core_create_mkey(mdev, umr_mkey, in, inlen);
 278
 279         kvfree(in);
 280         return err;
 281 }
 282
 283 static int mlx5e_create_rq_umr_mkey(struct mlx5_core_dev *mdev, struct mlx5e_rq *rq)
 284 {
 285         u64 num_mtts = MLX5E_REQUIRED_MTTS(mlx5_wq_ll_get_size(&rq->mpwqe.wq));
 286
 287         return mlx5e_create_umr_mkey(mdev, num_mtts, PAGE_SHIFT, &rq->umr_mkey,
 288                                      rq->wqe_overflow.addr);
 289 }
 290
 291 static u64 mlx5e_get_mpwqe_offset(u16 wqe_ix)
 292 {
 293         return MLX5E_REQUIRED_MTTS(wqe_ix) << PAGE_SHIFT;
 294 }
 295
 296 static void mlx5e_init_frags_partition(struct mlx5e_rq *rq)
 297 {
 298         struct mlx5e_wqe_frag_info next_frag = {};
 299         struct mlx5e_wqe_frag_info *prev = NULL;
 300         int i;
 301
 302         next_frag.di = &rq->wqe.di[0];
 303
 304         for (i = 0; i < mlx5_wq_cyc_get_size(&rq->wqe.wq); i++) {
 305                 struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
 306                 struct mlx5e_wqe_frag_info *frag =
 307                         &rq->wqe.frags[i << rq->wqe.info.log_num_frags];
 308                 int f;
 309
 310                 for (f = 0; f < rq->wqe.info.num_frags; f++, frag++) {
 311                         if (next_frag.offset + frag_info[f].frag_stride > PAGE_SIZE) {
 312                                 next_frag.di++;
 313                                 next_frag.offset = 0;
 314                                 if (prev)
 315                                         prev->last_in_page = true;
 316                         }
 317                         *frag = next_frag;
 318
 319                         /* prepare next */
 320                         next_frag.offset += frag_info[f].frag_stride;
 321                         prev = frag;
 322                 }
 323         }
 324
 325         if (prev)
 326                 prev->last_in_page = true;
 327 }
 328
 329 int mlx5e_init_di_list(struct mlx5e_rq *rq, int wq_sz, int node)
 330 {
 331         int len = wq_sz << rq->wqe.info.log_num_frags;
 332
 333         rq->wqe.di = kvzalloc_node(array_size(len, sizeof(*rq->wqe.di)), GFP_KERNEL, node);
 334         if (!rq->wqe.di)
 335                 return -ENOMEM;
 336
 337         mlx5e_init_frags_partition(rq);
 338
 339         return 0;
 340 }
 341
 342 void mlx5e_free_di_list(struct mlx5e_rq *rq)
 343 {
 344         kvfree(rq->wqe.di);
 345 }
 346
 347 static void mlx5e_rq_err_cqe_work(struct work_struct *recover_work)
 348 {
 349         struct mlx5e_rq *rq = container_of(recover_work, struct mlx5e_rq, recover_work);
 350
 351         mlx5e_reporter_rq_cqe_err(rq);
 352 }
 353
 354 static int mlx5e_alloc_mpwqe_rq_drop_page(struct mlx5e_rq *rq)
 355 {
 356         rq->wqe_overflow.page = alloc_page(GFP_KERNEL);
 357         if (!rq->wqe_overflow.page)
 358                 return -ENOMEM;
 359
 360         rq->wqe_overflow.addr = dma_map_page(rq->pdev, rq->wqe_overflow.page, 0,
 361                                              PAGE_SIZE, rq->buff.map_dir);
 362         if (dma_mapping_error(rq->pdev, rq->wqe_overflow.addr)) {
 363                 __free_page(rq->wqe_overflow.page);
 364                 return -ENOMEM;
 365         }
 366         return 0;
 367 }
 368
 369 static void mlx5e_free_mpwqe_rq_drop_page(struct mlx5e_rq *rq)
 370 {
 371          dma_unmap_page(rq->pdev, rq->wqe_overflow.addr, PAGE_SIZE,
 372                         rq->buff.map_dir);
 373          __free_page(rq->wqe_overflow.page);
 374 }
 375
 376 static int mlx5e_init_rxq_rq(struct mlx5e_channel *c, struct mlx5e_params *params,
 377                              struct mlx5e_rq *rq)
 378 {
 379         struct mlx5_core_dev *mdev = c->mdev;
 380         int err;
 381
 382         rq->wq_type      = params->rq_wq_type;
 383         rq->pdev         = c->pdev;
 384         rq->netdev       = c->netdev;
 385         rq->priv         = c->priv;
 386         rq->tstamp       = c->tstamp;
 387         rq->clock        = &mdev->clock;
 388         rq->icosq        = &c->icosq;
 389         rq->ix           = c->ix;
 390         rq->mdev         = mdev;
 391         rq->hw_mtu       = MLX5E_SW2HW_MTU(params, params->sw_mtu);
 392         rq->xdpsq        = &c->rq_xdpsq;
 393         rq->stats        = &c->priv->channel_stats[c->ix].rq;
 394         rq->ptp_cyc2time = mlx5_rq_ts_translator(mdev);
 395         err = mlx5e_rq_set_handlers(rq, params, NULL);
 396         if (err)
 397                 return err;
 398
 399         return xdp_rxq_info_reg(&rq->xdp_rxq, rq->netdev, rq->ix, 0);
 400 }
 401
 402 static int mlx5e_alloc_rq(struct mlx5e_params *params,
 403                           struct mlx5e_xsk_param *xsk,
 404                           struct mlx5e_rq_param *rqp,
 405                           int node, struct mlx5e_rq *rq)
 406 {
 407         struct page_pool_params pp_params = { 0 };
 408         struct mlx5_core_dev *mdev = rq->mdev;
 409         void *rqc = rqp->rqc;
 410         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 411         u32 pool_size;
 412         int wq_sz;
 413         int err;
 414         int i;
 415
 416         rqp->wq.db_numa_node = node;
 417         INIT_WORK(&rq->recover_work, mlx5e_rq_err_cqe_work);
 418
 419         if (params->xdp_prog)
 420                 bpf_prog_inc(params->xdp_prog);
 421         RCU_INIT_POINTER(rq->xdp_prog, params->xdp_prog);
 422
 423         rq->buff.map_dir = params->xdp_prog ? DMA_BIDIRECTIONAL : DMA_FROM_DEVICE;
 424         rq->buff.headroom = mlx5e_get_rq_headroom(mdev, params, xsk);
 425         pool_size = 1 << params->log_rq_mtu_frames;
 426
 427         switch (rq->wq_type) {
 428         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 429                 err = mlx5_wq_ll_create(mdev, &rqp->wq, rqc_wq, &rq->mpwqe.wq,
 430                                         &rq->wq_ctrl);
 431                 if (err)
 432                         goto err_rq_xdp_prog;
 433
 434                 err = mlx5e_alloc_mpwqe_rq_drop_page(rq);
 435                 if (err)
 436                         goto err_rq_wq_destroy;
 437
 438                 rq->mpwqe.wq.db = &rq->mpwqe.wq.db[MLX5_RCV_DBR];
 439
 440                 wq_sz = mlx5_wq_ll_get_size(&rq->mpwqe.wq);
 441
 442                 pool_size = MLX5_MPWRQ_PAGES_PER_WQE <<
 443                         mlx5e_mpwqe_get_log_rq_size(params, xsk);
 444
 445                 rq->mpwqe.log_stride_sz = mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
 446                 rq->mpwqe.num_strides =
 447                         BIT(mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk));
 448
 449                 rq->buff.frame0_sz = (1 << rq->mpwqe.log_stride_sz);
 450
 451                 err = mlx5e_create_rq_umr_mkey(mdev, rq);
 452                 if (err)
 453                         goto err_rq_drop_page;
 454                 rq->mkey_be = cpu_to_be32(rq->umr_mkey.key);
 455
 456                 err = mlx5e_rq_alloc_mpwqe_info(rq, node);
 457                 if (err)
 458                         goto err_rq_mkey;
 459                 break;
 460         default: /* MLX5_WQ_TYPE_CYCLIC */
 461                 err = mlx5_wq_cyc_create(mdev, &rqp->wq, rqc_wq, &rq->wqe.wq,
 462                                          &rq->wq_ctrl);
 463                 if (err)
 464                         goto err_rq_xdp_prog;
 465
 466                 rq->wqe.wq.db = &rq->wqe.wq.db[MLX5_RCV_DBR];
 467
 468                 wq_sz = mlx5_wq_cyc_get_size(&rq->wqe.wq);
 469
 470                 rq->wqe.info = rqp->frags_info;
 471                 rq->buff.frame0_sz = rq->wqe.info.arr[0].frag_stride;
 472
 473                 rq->wqe.frags =
 474                         kvzalloc_node(array_size(sizeof(*rq->wqe.frags),
 475                                         (wq_sz << rq->wqe.info.log_num_frags)),
 476                                       GFP_KERNEL, node);
 477                 if (!rq->wqe.frags) {
 478                         err = -ENOMEM;
 479                         goto err_rq_wq_destroy;
 480                 }
 481
 482                 err = mlx5e_init_di_list(rq, wq_sz, node);
 483                 if (err)
 484                         goto err_rq_frags;
 485
 486                 rq->mkey_be = cpu_to_be32(mdev->mlx5e_res.hw_objs.mkey.key);
 487         }
 488
 489         if (xsk) {
 490                 err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
 491                                                  MEM_TYPE_XSK_BUFF_POOL, NULL);
 492                 xsk_pool_set_rxq_info(rq->xsk_pool, &rq->xdp_rxq);
 493         } else {
 494                 /* Create a page_pool and register it with rxq */
 495                 pp_params.order     = 0;
 496                 pp_params.flags     = 0; /* No-internal DMA mapping in page_pool */
 497                 pp_params.pool_size = pool_size;
 498                 pp_params.nid       = node;
 499                 pp_params.dev       = rq->pdev;
 500                 pp_params.dma_dir   = rq->buff.map_dir;
 501
 502                 /* page_pool can be used even when there is no rq->xdp_prog,
 503                  * given page_pool does not handle DMA mapping there is no
 504                  * required state to clear. And page_pool gracefully handle
 505                  * elevated refcnt.
 506                  */
 507                 rq->page_pool = page_pool_create(&pp_params);
 508                 if (IS_ERR(rq->page_pool)) {
 509                         err = PTR_ERR(rq->page_pool);
 510                         rq->page_pool = NULL;
 511                         goto err_free_by_rq_type;
 512                 }
 513                 if (xdp_rxq_info_is_reg(&rq->xdp_rxq))
 514                         err = xdp_rxq_info_reg_mem_model(&rq->xdp_rxq,
 515                                                          MEM_TYPE_PAGE_POOL, rq->page_pool);
 516         }
 517         if (err)
 518                 goto err_free_by_rq_type;
 519
 520         for (i = 0; i < wq_sz; i++) {
 521                 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 522                         struct mlx5e_rx_wqe_ll *wqe =
 523                                 mlx5_wq_ll_get_wqe(&rq->mpwqe.wq, i);
 524                         u32 byte_count =
 525                                 rq->mpwqe.num_strides << rq->mpwqe.log_stride_sz;
 526                         u64 dma_offset = mlx5e_get_mpwqe_offset(i);
 527
 528                         wqe->data[0].addr = cpu_to_be64(dma_offset + rq->buff.headroom);
 529                         wqe->data[0].byte_count = cpu_to_be32(byte_count);
 530                         wqe->data[0].lkey = rq->mkey_be;
 531                 } else {
 532                         struct mlx5e_rx_wqe_cyc *wqe =
 533                                 mlx5_wq_cyc_get_wqe(&rq->wqe.wq, i);
 534                         int f;
 535
 536                         for (f = 0; f < rq->wqe.info.num_frags; f++) {
 537                                 u32 frag_size = rq->wqe.info.arr[f].frag_size |
 538                                         MLX5_HW_START_PADDING;
 539
 540                                 wqe->data[f].byte_count = cpu_to_be32(frag_size);
 541                                 wqe->data[f].lkey = rq->mkey_be;
 542                         }
 543                         /* check if num_frags is not a pow of two */
 544                         if (rq->wqe.info.num_frags < (1 << rq->wqe.info.log_num_frags)) {
 545                                 wqe->data[f].byte_count = 0;
 546                                 wqe->data[f].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
 547                                 wqe->data[f].addr = 0;
 548                         }
 549                 }
 550         }
 551
 552         INIT_WORK(&rq->dim.work, mlx5e_rx_dim_work);
 553
 554         switch (params->rx_cq_moderation.cq_period_mode) {
 555         case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
 556                 rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_CQE;
 557                 break;
 558         case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
 559         default:
 560                 rq->dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
 561         }
 562
 563         rq->page_cache.head = 0;
 564         rq->page_cache.tail = 0;
 565
 566         return 0;
 567
 568 err_free_by_rq_type:
 569         switch (rq->wq_type) {
 570         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 571                 kvfree(rq->mpwqe.info);
 572 err_rq_mkey:
 573                 mlx5_core_destroy_mkey(mdev, &rq->umr_mkey);
 574 err_rq_drop_page:
 575                 mlx5e_free_mpwqe_rq_drop_page(rq);
 576                 break;
 577         default: /* MLX5_WQ_TYPE_CYCLIC */
 578                 mlx5e_free_di_list(rq);
 579 err_rq_frags:
 580                 kvfree(rq->wqe.frags);
 581         }
 582 err_rq_wq_destroy:
 583         mlx5_wq_destroy(&rq->wq_ctrl);
 584 err_rq_xdp_prog:
 585         if (params->xdp_prog)
 586                 bpf_prog_put(params->xdp_prog);
 587
 588         return err;
 589 }
 590
 591 static void mlx5e_free_rq(struct mlx5e_rq *rq)
 592 {
 593         struct bpf_prog *old_prog;
 594         int i;
 595
 596         if (xdp_rxq_info_is_reg(&rq->xdp_rxq)) {
 597                 old_prog = rcu_dereference_protected(rq->xdp_prog,
 598                                                      lockdep_is_held(&rq->priv->state_lock));
 599                 if (old_prog)
 600                         bpf_prog_put(old_prog);
 601         }
 602
 603         switch (rq->wq_type) {
 604         case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
 605                 kvfree(rq->mpwqe.info);
 606                 mlx5_core_destroy_mkey(rq->mdev, &rq->umr_mkey);
 607                 mlx5e_free_mpwqe_rq_drop_page(rq);
 608                 break;
 609         default: /* MLX5_WQ_TYPE_CYCLIC */
 610                 kvfree(rq->wqe.frags);
 611                 mlx5e_free_di_list(rq);
 612         }
 613
 614         for (i = rq->page_cache.head; i != rq->page_cache.tail;
 615              i = (i + 1) & (MLX5E_CACHE_SIZE - 1)) {
 616                 struct mlx5e_dma_info *dma_info = &rq->page_cache.page_cache[i];
 617
 618                 /* With AF_XDP, page_cache is not used, so this loop is not
 619                  * entered, and it's safe to call mlx5e_page_release_dynamic
 620                  * directly.
 621                  */
 622                 mlx5e_page_release_dynamic(rq, dma_info, false);
 623         }
 624
 625         xdp_rxq_info_unreg(&rq->xdp_rxq);
 626         page_pool_destroy(rq->page_pool);
 627         mlx5_wq_destroy(&rq->wq_ctrl);
 628 }
 629
 630 int mlx5e_create_rq(struct mlx5e_rq *rq, struct mlx5e_rq_param *param)
 631 {
 632         struct mlx5_core_dev *mdev = rq->mdev;
 633         u8 ts_format;
 634         void *in;
 635         void *rqc;
 636         void *wq;
 637         int inlen;
 638         int err;
 639
 640         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 641                 sizeof(u64) * rq->wq_ctrl.buf.npages;
 642         in = kvzalloc(inlen, GFP_KERNEL);
 643         if (!in)
 644                 return -ENOMEM;
 645
 646         ts_format = mlx5_is_real_time_rq(mdev) ?
 647                     MLX5_RQC_TIMESTAMP_FORMAT_REAL_TIME :
 648                     MLX5_RQC_TIMESTAMP_FORMAT_FREE_RUNNING;
 649         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 650         wq  = MLX5_ADDR_OF(rqc, rqc, wq);
 651
 652         memcpy(rqc, param->rqc, sizeof(param->rqc));
 653
 654         MLX5_SET(rqc,  rqc, cqn,                rq->cq.mcq.cqn);
 655         MLX5_SET(rqc,  rqc, state,              MLX5_RQC_STATE_RST);
 656         MLX5_SET(rqc,  rqc, ts_format,          ts_format);
 657         MLX5_SET(wq,   wq,  log_wq_pg_sz,       rq->wq_ctrl.buf.page_shift -
 658                                                 MLX5_ADAPTER_PAGE_SHIFT);
 659         MLX5_SET64(wq, wq,  dbr_addr,           rq->wq_ctrl.db.dma);
 660
 661         mlx5_fill_page_frag_array(&rq->wq_ctrl.buf,
 662                                   (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
 663
 664         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 665
 666         kvfree(in);
 667
 668         return err;
 669 }
 670
 671 int mlx5e_modify_rq_state(struct mlx5e_rq *rq, int curr_state, int next_state)
 672 {
 673         struct mlx5_core_dev *mdev = rq->mdev;
 674
 675         void *in;
 676         void *rqc;
 677         int inlen;
 678         int err;
 679
 680         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 681         in = kvzalloc(inlen, GFP_KERNEL);
 682         if (!in)
 683                 return -ENOMEM;
 684
 685         if (curr_state == MLX5_RQC_STATE_RST && next_state == MLX5_RQC_STATE_RDY)
 686                 mlx5e_rqwq_reset(rq);
 687
 688         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 689
 690         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 691         MLX5_SET(rqc, rqc, state, next_state);
 692
 693         err = mlx5_core_modify_rq(mdev, rq->rqn, in);
 694
 695         kvfree(in);
 696
 697         return err;
 698 }
 699
 700 static int mlx5e_modify_rq_scatter_fcs(struct mlx5e_rq *rq, bool enable)
 701 {
 702         struct mlx5_core_dev *mdev = rq->mdev;
 703
 704         void *in;
 705         void *rqc;
 706         int inlen;
 707         int err;
 708
 709         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 710         in = kvzalloc(inlen, GFP_KERNEL);
 711         if (!in)
 712                 return -ENOMEM;
 713
 714         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 715
 716         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 717         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 718                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_SCATTER_FCS);
 719         MLX5_SET(rqc, rqc, scatter_fcs, enable);
 720         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 721
 722         err = mlx5_core_modify_rq(mdev, rq->rqn, in);
 723
 724         kvfree(in);
 725
 726         return err;
 727 }
 728
 729 static int mlx5e_modify_rq_vsd(struct mlx5e_rq *rq, bool vsd)
 730 {
 731         struct mlx5_core_dev *mdev = rq->mdev;
 732         void *in;
 733         void *rqc;
 734         int inlen;
 735         int err;
 736
 737         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 738         in = kvzalloc(inlen, GFP_KERNEL);
 739         if (!in)
 740                 return -ENOMEM;
 741
 742         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 743
 744         MLX5_SET(modify_rq_in, in, rq_state, MLX5_RQC_STATE_RDY);
 745         MLX5_SET64(modify_rq_in, in, modify_bitmask,
 746                    MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD);
 747         MLX5_SET(rqc, rqc, vsd, vsd);
 748         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RDY);
 749
 750         err = mlx5_core_modify_rq(mdev, rq->rqn, in);
 751
 752         kvfree(in);
 753
 754         return err;
 755 }
 756
 757 void mlx5e_destroy_rq(struct mlx5e_rq *rq)
 758 {
 759         mlx5_core_destroy_rq(rq->mdev, rq->rqn);
 760 }
 761
 762 int mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq, int wait_time)
 763 {
 764         unsigned long exp_time = jiffies + msecs_to_jiffies(wait_time);
 765
 766         u16 min_wqes = mlx5_min_rx_wqes(rq->wq_type, mlx5e_rqwq_get_size(rq));
 767
 768         do {
 769                 if (mlx5e_rqwq_get_cur_sz(rq) >= min_wqes)
 770                         return 0;
 771
 772                 msleep(20);
 773         } while (time_before(jiffies, exp_time));
 774
 775         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",
 776                     rq->ix, rq->rqn, mlx5e_rqwq_get_cur_sz(rq), min_wqes);
 777
 778         mlx5e_reporter_rx_timeout(rq);
 779         return -ETIMEDOUT;
 780 }
 781
 782 void mlx5e_free_rx_in_progress_descs(struct mlx5e_rq *rq)
 783 {
 784         struct mlx5_wq_ll *wq;
 785         u16 head;
 786         int i;
 787
 788         if (rq->wq_type != MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
 789                 return;
 790
 791         wq = &rq->mpwqe.wq;
 792         head = wq->head;
 793
 794         /* Outstanding UMR WQEs (in progress) start at wq->head */
 795         for (i = 0; i < rq->mpwqe.umr_in_progress; i++) {
 796                 rq->dealloc_wqe(rq, head);
 797                 head = mlx5_wq_ll_get_wqe_next_ix(wq, head);
 798         }
 799
 800         rq->mpwqe.actual_wq_head = wq->head;
 801         rq->mpwqe.umr_in_progress = 0;
 802         rq->mpwqe.umr_completed = 0;
 803 }
 804
 805 void mlx5e_free_rx_descs(struct mlx5e_rq *rq)
 806 {
 807         __be16 wqe_ix_be;
 808         u16 wqe_ix;
 809
 810         if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
 811                 struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
 812
 813                 mlx5e_free_rx_in_progress_descs(rq);
 814
 815                 while (!mlx5_wq_ll_is_empty(wq)) {
 816                         struct mlx5e_rx_wqe_ll *wqe;
 817
 818                         wqe_ix_be = *wq->tail_next;
 819                         wqe_ix    = be16_to_cpu(wqe_ix_be);
 820                         wqe       = mlx5_wq_ll_get_wqe(wq, wqe_ix);
 821                         rq->dealloc_wqe(rq, wqe_ix);
 822                         mlx5_wq_ll_pop(wq, wqe_ix_be,
 823                                        &wqe->next.next_wqe_index);
 824                 }
 825         } else {
 826                 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
 827
 828                 while (!mlx5_wq_cyc_is_empty(wq)) {
 829                         wqe_ix = mlx5_wq_cyc_get_tail(wq);
 830                         rq->dealloc_wqe(rq, wqe_ix);
 831                         mlx5_wq_cyc_pop(wq);
 832                 }
 833         }
 834
 835 }
 836
 837 int mlx5e_open_rq(struct mlx5e_params *params, struct mlx5e_rq_param *param,
 838                   struct mlx5e_xsk_param *xsk, int node,
 839                   struct mlx5e_rq *rq)
 840 {
 841         struct mlx5_core_dev *mdev = rq->mdev;
 842         int err;
 843
 844         err = mlx5e_alloc_rq(params, xsk, param, node, rq);
 845         if (err)
 846                 return err;
 847
 848         err = mlx5e_create_rq(rq, param);
 849         if (err)
 850                 goto err_free_rq;
 851
 852         err = mlx5e_modify_rq_state(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
 853         if (err)
 854                 goto err_destroy_rq;
 855
 856         if (mlx5e_is_tls_on(rq->priv) && !mlx5_accel_is_ktls_device(mdev))
 857                 __set_bit(MLX5E_RQ_STATE_FPGA_TLS, &rq->state); /* must be FPGA */
 858
 859         if (MLX5_CAP_ETH(mdev, cqe_checksum_full))
 860                 __set_bit(MLX5E_RQ_STATE_CSUM_FULL, &rq->state);
 861
 862         if (params->rx_dim_enabled)
 863                 __set_bit(MLX5E_RQ_STATE_AM, &rq->state);
 864
 865         /* We disable csum_complete when XDP is enabled since
 866          * XDP programs might manipulate packets which will render
 867          * skb->checksum incorrect.
 868          */
 869         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_NO_CSUM_COMPLETE) || params->xdp_prog)
 870                 __set_bit(MLX5E_RQ_STATE_NO_CSUM_COMPLETE, &rq->state);
 871
 872         /* For CQE compression on striding RQ, use stride index provided by
 873          * HW if capability is supported.
 874          */
 875         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ) &&
 876             MLX5_CAP_GEN(mdev, mini_cqe_resp_stride_index))
 877                 __set_bit(MLX5E_RQ_STATE_MINI_CQE_HW_STRIDX, &rq->state);
 878
 879         return 0;
 880
 881 err_destroy_rq:
 882         mlx5e_destroy_rq(rq);
 883 err_free_rq:
 884         mlx5e_free_rq(rq);
 885
 886         return err;
 887 }
 888
 889 void mlx5e_activate_rq(struct mlx5e_rq *rq)
 890 {
 891         set_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 892         if (rq->icosq)
 893                 mlx5e_trigger_irq(rq->icosq);
 894         else
 895                 napi_schedule(rq->cq.napi);
 896 }
 897
 898 void mlx5e_deactivate_rq(struct mlx5e_rq *rq)
 899 {
 900         clear_bit(MLX5E_RQ_STATE_ENABLED, &rq->state);
 901         synchronize_net(); /* Sync with NAPI to prevent mlx5e_post_rx_wqes. */
 902 }
 903
 904 void mlx5e_close_rq(struct mlx5e_rq *rq)
 905 {
 906         cancel_work_sync(&rq->dim.work);
 907         if (rq->icosq)
 908                 cancel_work_sync(&rq->icosq->recover_work);
 909         cancel_work_sync(&rq->recover_work);
 910         mlx5e_destroy_rq(rq);
 911         mlx5e_free_rx_descs(rq);
 912         mlx5e_free_rq(rq);
 913 }
 914
 915 static void mlx5e_free_xdpsq_db(struct mlx5e_xdpsq *sq)
 916 {
 917         kvfree(sq->db.xdpi_fifo.xi);
 918         kvfree(sq->db.wqe_info);
 919 }
 920
 921 static int mlx5e_alloc_xdpsq_fifo(struct mlx5e_xdpsq *sq, int numa)
 922 {
 923         struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo;
 924         int wq_sz        = mlx5_wq_cyc_get_size(&sq->wq);
 925         int dsegs_per_wq = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
 926
 927         xdpi_fifo->xi = kvzalloc_node(sizeof(*xdpi_fifo->xi) * dsegs_per_wq,
 928                                       GFP_KERNEL, numa);
 929         if (!xdpi_fifo->xi)
 930                 return -ENOMEM;
 931
 932         xdpi_fifo->pc   = &sq->xdpi_fifo_pc;
 933         xdpi_fifo->cc   = &sq->xdpi_fifo_cc;
 934         xdpi_fifo->mask = dsegs_per_wq - 1;
 935
 936         return 0;
 937 }
 938
 939 static int mlx5e_alloc_xdpsq_db(struct mlx5e_xdpsq *sq, int numa)
 940 {
 941         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
 942         int err;
 943
 944         sq->db.wqe_info = kvzalloc_node(sizeof(*sq->db.wqe_info) * wq_sz,
 945                                         GFP_KERNEL, numa);
 946         if (!sq->db.wqe_info)
 947                 return -ENOMEM;
 948
 949         err = mlx5e_alloc_xdpsq_fifo(sq, numa);
 950         if (err) {
 951                 mlx5e_free_xdpsq_db(sq);
 952                 return err;
 953         }
 954
 955         return 0;
 956 }
 957
 958 static int mlx5e_alloc_xdpsq(struct mlx5e_channel *c,
 959                              struct mlx5e_params *params,
 960                              struct xsk_buff_pool *xsk_pool,
 961                              struct mlx5e_sq_param *param,
 962                              struct mlx5e_xdpsq *sq,
 963                              bool is_redirect)
 964 {
 965         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
 966         struct mlx5_core_dev *mdev = c->mdev;
 967         struct mlx5_wq_cyc *wq = &sq->wq;
 968         int err;
 969
 970         sq->pdev      = c->pdev;
 971         sq->mkey_be   = c->mkey_be;
 972         sq->channel   = c;
 973         sq->uar_map   = mdev->mlx5e_res.hw_objs.bfreg.map;
 974         sq->min_inline_mode = params->tx_min_inline_mode;
 975         sq->hw_mtu    = MLX5E_SW2HW_MTU(params, params->sw_mtu);
 976         sq->xsk_pool  = xsk_pool;
 977
 978         sq->stats = sq->xsk_pool ?
 979                 &c->priv->channel_stats[c->ix].xsksq :
 980                 is_redirect ?
 981                         &c->priv->channel_stats[c->ix].xdpsq :
 982                         &c->priv->channel_stats[c->ix].rq_xdpsq;
 983
 984         param->wq.db_numa_node = cpu_to_node(c->cpu);
 985         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
 986         if (err)
 987                 return err;
 988         wq->db = &wq->db[MLX5_SND_DBR];
 989
 990         err = mlx5e_alloc_xdpsq_db(sq, cpu_to_node(c->cpu));
 991         if (err)
 992                 goto err_sq_wq_destroy;
 993
 994         return 0;
 995
 996 err_sq_wq_destroy:
 997         mlx5_wq_destroy(&sq->wq_ctrl);
 998
 999         return err;
1000 }
1001
1002 static void mlx5e_free_xdpsq(struct mlx5e_xdpsq *sq)
1003 {
1004         mlx5e_free_xdpsq_db(sq);
1005         mlx5_wq_destroy(&sq->wq_ctrl);
1006 }
1007
1008 static void mlx5e_free_icosq_db(struct mlx5e_icosq *sq)
1009 {
1010         kvfree(sq->db.wqe_info);
1011 }
1012
1013 static int mlx5e_alloc_icosq_db(struct mlx5e_icosq *sq, int numa)
1014 {
1015         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1016         size_t size;
1017
1018         size = array_size(wq_sz, sizeof(*sq->db.wqe_info));
1019         sq->db.wqe_info = kvzalloc_node(size, GFP_KERNEL, numa);
1020         if (!sq->db.wqe_info)
1021                 return -ENOMEM;
1022
1023         return 0;
1024 }
1025
1026 static void mlx5e_icosq_err_cqe_work(struct work_struct *recover_work)
1027 {
1028         struct mlx5e_icosq *sq = container_of(recover_work, struct mlx5e_icosq,
1029                                               recover_work);
1030
1031         mlx5e_reporter_icosq_cqe_err(sq);
1032 }
1033
1034 static int mlx5e_alloc_icosq(struct mlx5e_channel *c,
1035                              struct mlx5e_sq_param *param,
1036                              struct mlx5e_icosq *sq)
1037 {
1038         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1039         struct mlx5_core_dev *mdev = c->mdev;
1040         struct mlx5_wq_cyc *wq = &sq->wq;
1041         int err;
1042
1043         sq->channel   = c;
1044         sq->uar_map   = mdev->mlx5e_res.hw_objs.bfreg.map;
1045         sq->reserved_room = param->stop_room;
1046
1047         param->wq.db_numa_node = cpu_to_node(c->cpu);
1048         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
1049         if (err)
1050                 return err;
1051         wq->db = &wq->db[MLX5_SND_DBR];
1052
1053         err = mlx5e_alloc_icosq_db(sq, cpu_to_node(c->cpu));
1054         if (err)
1055                 goto err_sq_wq_destroy;
1056
1057         INIT_WORK(&sq->recover_work, mlx5e_icosq_err_cqe_work);
1058
1059         return 0;
1060
1061 err_sq_wq_destroy:
1062         mlx5_wq_destroy(&sq->wq_ctrl);
1063
1064         return err;
1065 }
1066
1067 static void mlx5e_free_icosq(struct mlx5e_icosq *sq)
1068 {
1069         mlx5e_free_icosq_db(sq);
1070         mlx5_wq_destroy(&sq->wq_ctrl);
1071 }
1072
1073 void mlx5e_free_txqsq_db(struct mlx5e_txqsq *sq)
1074 {
1075         kvfree(sq->db.wqe_info);
1076         kvfree(sq->db.skb_fifo.fifo);
1077         kvfree(sq->db.dma_fifo);
1078 }
1079
1080 int mlx5e_alloc_txqsq_db(struct mlx5e_txqsq *sq, int numa)
1081 {
1082         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1083         int df_sz = wq_sz * MLX5_SEND_WQEBB_NUM_DS;
1084
1085         sq->db.dma_fifo = kvzalloc_node(array_size(df_sz,
1086                                                    sizeof(*sq->db.dma_fifo)),
1087                                         GFP_KERNEL, numa);
1088         sq->db.skb_fifo.fifo = kvzalloc_node(array_size(df_sz,
1089                                                         sizeof(*sq->db.skb_fifo.fifo)),
1090                                         GFP_KERNEL, numa);
1091         sq->db.wqe_info = kvzalloc_node(array_size(wq_sz,
1092                                                    sizeof(*sq->db.wqe_info)),
1093                                         GFP_KERNEL, numa);
1094         if (!sq->db.dma_fifo || !sq->db.skb_fifo.fifo || !sq->db.wqe_info) {
1095                 mlx5e_free_txqsq_db(sq);
1096                 return -ENOMEM;
1097         }
1098
1099         sq->dma_fifo_mask = df_sz - 1;
1100
1101         sq->db.skb_fifo.pc   = &sq->skb_fifo_pc;
1102         sq->db.skb_fifo.cc   = &sq->skb_fifo_cc;
1103         sq->db.skb_fifo.mask = df_sz - 1;
1104
1105         return 0;
1106 }
1107
1108 static int mlx5e_alloc_txqsq(struct mlx5e_channel *c,
1109                              int txq_ix,
1110                              struct mlx5e_params *params,
1111                              struct mlx5e_sq_param *param,
1112                              struct mlx5e_txqsq *sq,
1113                              int tc)
1114 {
1115         void *sqc_wq               = MLX5_ADDR_OF(sqc, param->sqc, wq);
1116         struct mlx5_core_dev *mdev = c->mdev;
1117         struct mlx5_wq_cyc *wq = &sq->wq;
1118         int err;
1119
1120         sq->pdev      = c->pdev;
1121         sq->tstamp    = c->tstamp;
1122         sq->clock     = &mdev->clock;
1123         sq->mkey_be   = c->mkey_be;
1124         sq->netdev    = c->netdev;
1125         sq->mdev      = c->mdev;
1126         sq->priv      = c->priv;
1127         sq->ch_ix     = c->ix;
1128         sq->txq_ix    = txq_ix;
1129         sq->uar_map   = mdev->mlx5e_res.hw_objs.bfreg.map;
1130         sq->min_inline_mode = params->tx_min_inline_mode;
1131         sq->hw_mtu    = MLX5E_SW2HW_MTU(params, params->sw_mtu);
1132         INIT_WORK(&sq->recover_work, mlx5e_tx_err_cqe_work);
1133         if (!MLX5_CAP_ETH(mdev, wqe_vlan_insert))
1134                 set_bit(MLX5E_SQ_STATE_VLAN_NEED_L2_INLINE, &sq->state);
1135         if (MLX5_IPSEC_DEV(c->priv->mdev))
1136                 set_bit(MLX5E_SQ_STATE_IPSEC, &sq->state);
1137         if (param->is_mpw)
1138                 set_bit(MLX5E_SQ_STATE_MPWQE, &sq->state);
1139         sq->stop_room = param->stop_room;
1140         sq->ptp_cyc2time = mlx5_sq_ts_translator(mdev);
1141
1142         param->wq.db_numa_node = cpu_to_node(c->cpu);
1143         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, wq, &sq->wq_ctrl);
1144         if (err)
1145                 return err;
1146         wq->db    = &wq->db[MLX5_SND_DBR];
1147
1148         err = mlx5e_alloc_txqsq_db(sq, cpu_to_node(c->cpu));
1149         if (err)
1150                 goto err_sq_wq_destroy;
1151
1152         INIT_WORK(&sq->dim.work, mlx5e_tx_dim_work);
1153         sq->dim.mode = params->tx_cq_moderation.cq_period_mode;
1154
1155         return 0;
1156
1157 err_sq_wq_destroy:
1158         mlx5_wq_destroy(&sq->wq_ctrl);
1159
1160         return err;
1161 }
1162
1163 void mlx5e_free_txqsq(struct mlx5e_txqsq *sq)
1164 {
1165         mlx5e_free_txqsq_db(sq);
1166         mlx5_wq_destroy(&sq->wq_ctrl);
1167 }
1168
1169 static int mlx5e_create_sq(struct mlx5_core_dev *mdev,
1170                            struct mlx5e_sq_param *param,
1171                            struct mlx5e_create_sq_param *csp,
1172                            u32 *sqn)
1173 {
1174         u8 ts_format;
1175         void *in;
1176         void *sqc;
1177         void *wq;
1178         int inlen;
1179         int err;
1180
1181         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1182                 sizeof(u64) * csp->wq_ctrl->buf.npages;
1183         in = kvzalloc(inlen, GFP_KERNEL);
1184         if (!in)
1185                 return -ENOMEM;
1186
1187         ts_format = mlx5_is_real_time_sq(mdev) ?
1188                     MLX5_SQC_TIMESTAMP_FORMAT_REAL_TIME :
1189                     MLX5_SQC_TIMESTAMP_FORMAT_FREE_RUNNING;
1190         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1191         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1192
1193         memcpy(sqc, param->sqc, sizeof(param->sqc));
1194         MLX5_SET(sqc,  sqc, tis_lst_sz, csp->tis_lst_sz);
1195         MLX5_SET(sqc,  sqc, tis_num_0, csp->tisn);
1196         MLX5_SET(sqc,  sqc, cqn, csp->cqn);
1197         MLX5_SET(sqc,  sqc, ts_cqe_to_dest_cqn, csp->ts_cqe_to_dest_cqn);
1198         MLX5_SET(sqc,  sqc, ts_format, ts_format);
1199
1200
1201         if (MLX5_CAP_ETH(mdev, wqe_inline_mode) == MLX5_CAP_INLINE_MODE_VPORT_CONTEXT)
1202                 MLX5_SET(sqc,  sqc, min_wqe_inline_mode, csp->min_inline_mode);
1203
1204         MLX5_SET(sqc,  sqc, state, MLX5_SQC_STATE_RST);
1205         MLX5_SET(sqc,  sqc, flush_in_error_en, 1);
1206
1207         MLX5_SET(wq,   wq, wq_type,       MLX5_WQ_TYPE_CYCLIC);
1208         MLX5_SET(wq,   wq, uar_page,      mdev->mlx5e_res.hw_objs.bfreg.index);
1209         MLX5_SET(wq,   wq, log_wq_pg_sz,  csp->wq_ctrl->buf.page_shift -
1210                                           MLX5_ADAPTER_PAGE_SHIFT);
1211         MLX5_SET64(wq, wq, dbr_addr,      csp->wq_ctrl->db.dma);
1212
1213         mlx5_fill_page_frag_array(&csp->wq_ctrl->buf,
1214                                   (__be64 *)MLX5_ADDR_OF(wq, wq, pas));
1215
1216         err = mlx5_core_create_sq(mdev, in, inlen, sqn);
1217
1218         kvfree(in);
1219
1220         return err;
1221 }
1222
1223 int mlx5e_modify_sq(struct mlx5_core_dev *mdev, u32 sqn,
1224                     struct mlx5e_modify_sq_param *p)
1225 {
1226         u64 bitmask = 0;
1227         void *in;
1228         void *sqc;
1229         int inlen;
1230         int err;
1231
1232         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1233         in = kvzalloc(inlen, GFP_KERNEL);
1234         if (!in)
1235                 return -ENOMEM;
1236
1237         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1238
1239         MLX5_SET(modify_sq_in, in, sq_state, p->curr_state);
1240         MLX5_SET(sqc, sqc, state, p->next_state);
1241         if (p->rl_update && p->next_state == MLX5_SQC_STATE_RDY) {
1242                 bitmask |= 1;
1243                 MLX5_SET(sqc, sqc, packet_pacing_rate_limit_index, p->rl_index);
1244         }
1245         if (p->qos_update && p->next_state == MLX5_SQC_STATE_RDY) {
1246                 bitmask |= 1 << 2;
1247                 MLX5_SET(sqc, sqc, qos_queue_group_id, p->qos_queue_group_id);
1248         }
1249         MLX5_SET64(modify_sq_in, in, modify_bitmask, bitmask);
1250
1251         err = mlx5_core_modify_sq(mdev, sqn, in);
1252
1253         kvfree(in);
1254
1255         return err;
1256 }
1257
1258 static void mlx5e_destroy_sq(struct mlx5_core_dev *mdev, u32 sqn)
1259 {
1260         mlx5_core_destroy_sq(mdev, sqn);
1261 }
1262
1263 int mlx5e_create_sq_rdy(struct mlx5_core_dev *mdev,
1264                         struct mlx5e_sq_param *param,
1265                         struct mlx5e_create_sq_param *csp,
1266                         u16 qos_queue_group_id,
1267                         u32 *sqn)
1268 {
1269         struct mlx5e_modify_sq_param msp = {0};
1270         int err;
1271
1272         err = mlx5e_create_sq(mdev, param, csp, sqn);
1273         if (err)
1274                 return err;
1275
1276         msp.curr_state = MLX5_SQC_STATE_RST;
1277         msp.next_state = MLX5_SQC_STATE_RDY;
1278         if (qos_queue_group_id) {
1279                 msp.qos_update = true;
1280                 msp.qos_queue_group_id = qos_queue_group_id;
1281         }
1282         err = mlx5e_modify_sq(mdev, *sqn, &msp);
1283         if (err)
1284                 mlx5e_destroy_sq(mdev, *sqn);
1285
1286         return err;
1287 }
1288
1289 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1290                                 struct mlx5e_txqsq *sq, u32 rate);
1291
1292 int mlx5e_open_txqsq(struct mlx5e_channel *c, u32 tisn, int txq_ix,
1293                      struct mlx5e_params *params, struct mlx5e_sq_param *param,
1294                      struct mlx5e_txqsq *sq, int tc, u16 qos_queue_group_id, u16 qos_qid)
1295 {
1296         struct mlx5e_create_sq_param csp = {};
1297         u32 tx_rate;
1298         int err;
1299
1300         err = mlx5e_alloc_txqsq(c, txq_ix, params, param, sq, tc);
1301         if (err)
1302                 return err;
1303
1304         if (qos_queue_group_id)
1305                 sq->stats = c->priv->htb.qos_sq_stats[qos_qid];
1306         else
1307                 sq->stats = &c->priv->channel_stats[c->ix].sq[tc];
1308
1309         csp.tisn            = tisn;
1310         csp.tis_lst_sz      = 1;
1311         csp.cqn             = sq->cq.mcq.cqn;
1312         csp.wq_ctrl         = &sq->wq_ctrl;
1313         csp.min_inline_mode = sq->min_inline_mode;
1314         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, qos_queue_group_id, &sq->sqn);
1315         if (err)
1316                 goto err_free_txqsq;
1317
1318         tx_rate = c->priv->tx_rates[sq->txq_ix];
1319         if (tx_rate)
1320                 mlx5e_set_sq_maxrate(c->netdev, sq, tx_rate);
1321
1322         if (params->tx_dim_enabled)
1323                 sq->state |= BIT(MLX5E_SQ_STATE_AM);
1324
1325         return 0;
1326
1327 err_free_txqsq:
1328         mlx5e_free_txqsq(sq);
1329
1330         return err;
1331 }
1332
1333 void mlx5e_activate_txqsq(struct mlx5e_txqsq *sq)
1334 {
1335         sq->txq = netdev_get_tx_queue(sq->netdev, sq->txq_ix);
1336         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1337         netdev_tx_reset_queue(sq->txq);
1338         netif_tx_start_queue(sq->txq);
1339 }
1340
1341 void mlx5e_tx_disable_queue(struct netdev_queue *txq)
1342 {
1343         __netif_tx_lock_bh(txq);
1344         netif_tx_stop_queue(txq);
1345         __netif_tx_unlock_bh(txq);
1346 }
1347
1348 void mlx5e_deactivate_txqsq(struct mlx5e_txqsq *sq)
1349 {
1350         struct mlx5_wq_cyc *wq = &sq->wq;
1351
1352         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1353         synchronize_net(); /* Sync with NAPI to prevent netif_tx_wake_queue. */
1354
1355         mlx5e_tx_disable_queue(sq->txq);
1356
1357         /* last doorbell out, godspeed .. */
1358         if (mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1)) {
1359                 u16 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
1360                 struct mlx5e_tx_wqe *nop;
1361
1362                 sq->db.wqe_info[pi] = (struct mlx5e_tx_wqe_info) {
1363                         .num_wqebbs = 1,
1364                 };
1365
1366                 nop = mlx5e_post_nop(wq, sq->sqn, &sq->pc);
1367                 mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &nop->ctrl);
1368         }
1369 }
1370
1371 void mlx5e_close_txqsq(struct mlx5e_txqsq *sq)
1372 {
1373         struct mlx5_core_dev *mdev = sq->mdev;
1374         struct mlx5_rate_limit rl = {0};
1375
1376         cancel_work_sync(&sq->dim.work);
1377         cancel_work_sync(&sq->recover_work);
1378         mlx5e_destroy_sq(mdev, sq->sqn);
1379         if (sq->rate_limit) {
1380                 rl.rate = sq->rate_limit;
1381                 mlx5_rl_remove_rate(mdev, &rl);
1382         }
1383         mlx5e_free_txqsq_descs(sq);
1384         mlx5e_free_txqsq(sq);
1385 }
1386
1387 void mlx5e_tx_err_cqe_work(struct work_struct *recover_work)
1388 {
1389         struct mlx5e_txqsq *sq = container_of(recover_work, struct mlx5e_txqsq,
1390                                               recover_work);
1391
1392         mlx5e_reporter_tx_err_cqe(sq);
1393 }
1394
1395 int mlx5e_open_icosq(struct mlx5e_channel *c, struct mlx5e_params *params,
1396                      struct mlx5e_sq_param *param, struct mlx5e_icosq *sq)
1397 {
1398         struct mlx5e_create_sq_param csp = {};
1399         int err;
1400
1401         err = mlx5e_alloc_icosq(c, param, sq);
1402         if (err)
1403                 return err;
1404
1405         csp.cqn             = sq->cq.mcq.cqn;
1406         csp.wq_ctrl         = &sq->wq_ctrl;
1407         csp.min_inline_mode = params->tx_min_inline_mode;
1408         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, 0, &sq->sqn);
1409         if (err)
1410                 goto err_free_icosq;
1411
1412         if (param->is_tls) {
1413                 sq->ktls_resync = mlx5e_ktls_rx_resync_create_resp_list();
1414                 if (IS_ERR(sq->ktls_resync)) {
1415                         err = PTR_ERR(sq->ktls_resync);
1416                         goto err_destroy_icosq;
1417                 }
1418         }
1419         return 0;
1420
1421 err_destroy_icosq:
1422         mlx5e_destroy_sq(c->mdev, sq->sqn);
1423 err_free_icosq:
1424         mlx5e_free_icosq(sq);
1425
1426         return err;
1427 }
1428
1429 void mlx5e_activate_icosq(struct mlx5e_icosq *icosq)
1430 {
1431         set_bit(MLX5E_SQ_STATE_ENABLED, &icosq->state);
1432 }
1433
1434 void mlx5e_deactivate_icosq(struct mlx5e_icosq *icosq)
1435 {
1436         clear_bit(MLX5E_SQ_STATE_ENABLED, &icosq->state);
1437         synchronize_net(); /* Sync with NAPI. */
1438 }
1439
1440 void mlx5e_close_icosq(struct mlx5e_icosq *sq)
1441 {
1442         struct mlx5e_channel *c = sq->channel;
1443
1444         if (sq->ktls_resync)
1445                 mlx5e_ktls_rx_resync_destroy_resp_list(sq->ktls_resync);
1446         mlx5e_destroy_sq(c->mdev, sq->sqn);
1447         mlx5e_free_icosq_descs(sq);
1448         mlx5e_free_icosq(sq);
1449 }
1450
1451 int mlx5e_open_xdpsq(struct mlx5e_channel *c, struct mlx5e_params *params,
1452                      struct mlx5e_sq_param *param, struct xsk_buff_pool *xsk_pool,
1453                      struct mlx5e_xdpsq *sq, bool is_redirect)
1454 {
1455         struct mlx5e_create_sq_param csp = {};
1456         int err;
1457
1458         err = mlx5e_alloc_xdpsq(c, params, xsk_pool, param, sq, is_redirect);
1459         if (err)
1460                 return err;
1461
1462         csp.tis_lst_sz      = 1;
1463         csp.tisn            = c->priv->tisn[c->lag_port][0]; /* tc = 0 */
1464         csp.cqn             = sq->cq.mcq.cqn;
1465         csp.wq_ctrl         = &sq->wq_ctrl;
1466         csp.min_inline_mode = sq->min_inline_mode;
1467         set_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1468         err = mlx5e_create_sq_rdy(c->mdev, param, &csp, 0, &sq->sqn);
1469         if (err)
1470                 goto err_free_xdpsq;
1471
1472         mlx5e_set_xmit_fp(sq, param->is_mpw);
1473
1474         if (!param->is_mpw) {
1475                 unsigned int ds_cnt = MLX5E_XDP_TX_DS_COUNT;
1476                 unsigned int inline_hdr_sz = 0;
1477                 int i;
1478
1479                 if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
1480                         inline_hdr_sz = MLX5E_XDP_MIN_INLINE;
1481                         ds_cnt++;
1482                 }
1483
1484                 /* Pre initialize fixed WQE fields */
1485                 for (i = 0; i < mlx5_wq_cyc_get_size(&sq->wq); i++) {
1486                         struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(&sq->wq, i);
1487                         struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
1488                         struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
1489                         struct mlx5_wqe_data_seg *dseg;
1490
1491                         sq->db.wqe_info[i] = (struct mlx5e_xdp_wqe_info) {
1492                                 .num_wqebbs = 1,
1493                                 .num_pkts   = 1,
1494                         };
1495
1496                         cseg->qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
1497                         eseg->inline_hdr.sz = cpu_to_be16(inline_hdr_sz);
1498
1499                         dseg = (struct mlx5_wqe_data_seg *)cseg + (ds_cnt - 1);
1500                         dseg->lkey = sq->mkey_be;
1501                 }
1502         }
1503
1504         return 0;
1505
1506 err_free_xdpsq:
1507         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1508         mlx5e_free_xdpsq(sq);
1509
1510         return err;
1511 }
1512
1513 void mlx5e_close_xdpsq(struct mlx5e_xdpsq *sq)
1514 {
1515         struct mlx5e_channel *c = sq->channel;
1516
1517         clear_bit(MLX5E_SQ_STATE_ENABLED, &sq->state);
1518         synchronize_net(); /* Sync with NAPI. */
1519
1520         mlx5e_destroy_sq(c->mdev, sq->sqn);
1521         mlx5e_free_xdpsq_descs(sq);
1522         mlx5e_free_xdpsq(sq);
1523 }
1524
1525 static int mlx5e_alloc_cq_common(struct mlx5e_priv *priv,
1526                                  struct mlx5e_cq_param *param,
1527                                  struct mlx5e_cq *cq)
1528 {
1529         struct mlx5_core_dev *mdev = priv->mdev;
1530         struct mlx5_core_cq *mcq = &cq->mcq;
1531         int eqn_not_used;
1532         unsigned int irqn;
1533         int err;
1534         u32 i;
1535
1536         err = mlx5_vector2eqn(mdev, param->eq_ix, &eqn_not_used, &irqn);
1537         if (err)
1538                 return err;
1539
1540         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1541                                &cq->wq_ctrl);
1542         if (err)
1543                 return err;
1544
1545         mcq->cqe_sz     = 64;
1546         mcq->set_ci_db  = cq->wq_ctrl.db.db;
1547         mcq->arm_db     = cq->wq_ctrl.db.db + 1;
1548         *mcq->set_ci_db = 0;
1549         *mcq->arm_db    = 0;
1550         mcq->vector     = param->eq_ix;
1551         mcq->comp       = mlx5e_completion_event;
1552         mcq->event      = mlx5e_cq_error_event;
1553         mcq->irqn       = irqn;
1554
1555         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1556                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1557
1558                 cqe->op_own = 0xf1;
1559         }
1560
1561         cq->mdev = mdev;
1562         cq->netdev = priv->netdev;
1563         cq->priv = priv;
1564
1565         return 0;
1566 }
1567
1568 static int mlx5e_alloc_cq(struct mlx5e_priv *priv,
1569                           struct mlx5e_cq_param *param,
1570                           struct mlx5e_create_cq_param *ccp,
1571                           struct mlx5e_cq *cq)
1572 {
1573         int err;
1574
1575         param->wq.buf_numa_node = ccp->node;
1576         param->wq.db_numa_node  = ccp->node;
1577         param->eq_ix            = ccp->ix;
1578
1579         err = mlx5e_alloc_cq_common(priv, param, cq);
1580
1581         cq->napi     = ccp->napi;
1582         cq->ch_stats = ccp->ch_stats;
1583
1584         return err;
1585 }
1586
1587 static void mlx5e_free_cq(struct mlx5e_cq *cq)
1588 {
1589         mlx5_wq_destroy(&cq->wq_ctrl);
1590 }
1591
1592 static int mlx5e_create_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param)
1593 {
1594         u32 out[MLX5_ST_SZ_DW(create_cq_out)];
1595         struct mlx5_core_dev *mdev = cq->mdev;
1596         struct mlx5_core_cq *mcq = &cq->mcq;
1597
1598         void *in;
1599         void *cqc;
1600         int inlen;
1601         unsigned int irqn_not_used;
1602         int eqn;
1603         int err;
1604
1605         err = mlx5_vector2eqn(mdev, param->eq_ix, &eqn, &irqn_not_used);
1606         if (err)
1607                 return err;
1608
1609         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1610                 sizeof(u64) * cq->wq_ctrl.buf.npages;
1611         in = kvzalloc(inlen, GFP_KERNEL);
1612         if (!in)
1613                 return -ENOMEM;
1614
1615         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1616
1617         memcpy(cqc, param->cqc, sizeof(param->cqc));
1618
1619         mlx5_fill_page_frag_array(&cq->wq_ctrl.buf,
1620                                   (__be64 *)MLX5_ADDR_OF(create_cq_in, in, pas));
1621
1622         MLX5_SET(cqc,   cqc, cq_period_mode, param->cq_period_mode);
1623         MLX5_SET(cqc,   cqc, c_eqn,         eqn);
1624         MLX5_SET(cqc,   cqc, uar_page,      mdev->priv.uar->index);
1625         MLX5_SET(cqc,   cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
1626                                             MLX5_ADAPTER_PAGE_SHIFT);
1627         MLX5_SET64(cqc, cqc, dbr_addr,      cq->wq_ctrl.db.dma);
1628
1629         err = mlx5_core_create_cq(mdev, mcq, in, inlen, out, sizeof(out));
1630
1631         kvfree(in);
1632
1633         if (err)
1634                 return err;
1635
1636         mlx5e_cq_arm(cq);
1637
1638         return 0;
1639 }
1640
1641 static void mlx5e_destroy_cq(struct mlx5e_cq *cq)
1642 {
1643         mlx5_core_destroy_cq(cq->mdev, &cq->mcq);
1644 }
1645
1646 int mlx5e_open_cq(struct mlx5e_priv *priv, struct dim_cq_moder moder,
1647                   struct mlx5e_cq_param *param, struct mlx5e_create_cq_param *ccp,
1648                   struct mlx5e_cq *cq)
1649 {
1650         struct mlx5_core_dev *mdev = priv->mdev;
1651         int err;
1652
1653         err = mlx5e_alloc_cq(priv, param, ccp, cq);
1654         if (err)
1655                 return err;
1656
1657         err = mlx5e_create_cq(cq, param);
1658         if (err)
1659                 goto err_free_cq;
1660
1661         if (MLX5_CAP_GEN(mdev, cq_moderation))
1662                 mlx5_core_modify_cq_moderation(mdev, &cq->mcq, moder.usec, moder.pkts);
1663         return 0;
1664
1665 err_free_cq:
1666         mlx5e_free_cq(cq);
1667
1668         return err;
1669 }
1670
1671 void mlx5e_close_cq(struct mlx5e_cq *cq)
1672 {
1673         mlx5e_destroy_cq(cq);
1674         mlx5e_free_cq(cq);
1675 }
1676
1677 static int mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1678                              struct mlx5e_params *params,
1679                              struct mlx5e_create_cq_param *ccp,
1680                              struct mlx5e_channel_param *cparam)
1681 {
1682         int err;
1683         int tc;
1684
1685         for (tc = 0; tc < c->num_tc; tc++) {
1686                 err = mlx5e_open_cq(c->priv, params->tx_cq_moderation, &cparam->txq_sq.cqp,
1687                                     ccp, &c->sq[tc].cq);
1688                 if (err)
1689                         goto err_close_tx_cqs;
1690         }
1691
1692         return 0;
1693
1694 err_close_tx_cqs:
1695         for (tc--; tc >= 0; tc--)
1696                 mlx5e_close_cq(&c->sq[tc].cq);
1697
1698         return err;
1699 }
1700
1701 static void mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1702 {
1703         int tc;
1704
1705         for (tc = 0; tc < c->num_tc; tc++)
1706                 mlx5e_close_cq(&c->sq[tc].cq);
1707 }
1708
1709 static int mlx5e_open_sqs(struct mlx5e_channel *c,
1710                           struct mlx5e_params *params,
1711                           struct mlx5e_channel_param *cparam)
1712 {
1713         int err, tc;
1714
1715         for (tc = 0; tc < params->num_tc; tc++) {
1716                 int txq_ix = c->ix + tc * params->num_channels;
1717
1718                 err = mlx5e_open_txqsq(c, c->priv->tisn[c->lag_port][tc], txq_ix,
1719                                        params, &cparam->txq_sq, &c->sq[tc], tc, 0, 0);
1720                 if (err)
1721                         goto err_close_sqs;
1722         }
1723
1724         return 0;
1725
1726 err_close_sqs:
1727         for (tc--; tc >= 0; tc--)
1728                 mlx5e_close_txqsq(&c->sq[tc]);
1729
1730         return err;
1731 }
1732
1733 static void mlx5e_close_sqs(struct mlx5e_channel *c)
1734 {
1735         int tc;
1736
1737         for (tc = 0; tc < c->num_tc; tc++)
1738                 mlx5e_close_txqsq(&c->sq[tc]);
1739 }
1740
1741 static int mlx5e_set_sq_maxrate(struct net_device *dev,
1742                                 struct mlx5e_txqsq *sq, u32 rate)
1743 {
1744         struct mlx5e_priv *priv = netdev_priv(dev);
1745         struct mlx5_core_dev *mdev = priv->mdev;
1746         struct mlx5e_modify_sq_param msp = {0};
1747         struct mlx5_rate_limit rl = {0};
1748         u16 rl_index = 0;
1749         int err;
1750
1751         if (rate == sq->rate_limit)
1752                 /* nothing to do */
1753                 return 0;
1754
1755         if (sq->rate_limit) {
1756                 rl.rate = sq->rate_limit;
1757                 /* remove current rl index to free space to next ones */
1758                 mlx5_rl_remove_rate(mdev, &rl);
1759         }
1760
1761         sq->rate_limit = 0;
1762
1763         if (rate) {
1764                 rl.rate = rate;
1765                 err = mlx5_rl_add_rate(mdev, &rl_index, &rl);
1766                 if (err) {
1767                         netdev_err(dev, "Failed configuring rate %u: %d\n",
1768                                    rate, err);
1769                         return err;
1770                 }
1771         }
1772
1773         msp.curr_state = MLX5_SQC_STATE_RDY;
1774         msp.next_state = MLX5_SQC_STATE_RDY;
1775         msp.rl_index   = rl_index;
1776         msp.rl_update  = true;
1777         err = mlx5e_modify_sq(mdev, sq->sqn, &msp);
1778         if (err) {
1779                 netdev_err(dev, "Failed configuring rate %u: %d\n",
1780                            rate, err);
1781                 /* remove the rate from the table */
1782                 if (rate)
1783                         mlx5_rl_remove_rate(mdev, &rl);
1784                 return err;
1785         }
1786
1787         sq->rate_limit = rate;
1788         return 0;
1789 }
1790
1791 static int mlx5e_set_tx_maxrate(struct net_device *dev, int index, u32 rate)
1792 {
1793         struct mlx5e_priv *priv = netdev_priv(dev);
1794         struct mlx5_core_dev *mdev = priv->mdev;
1795         struct mlx5e_txqsq *sq = priv->txq2sq[index];
1796         int err = 0;
1797
1798         if (!mlx5_rl_is_supported(mdev)) {
1799                 netdev_err(dev, "Rate limiting is not supported on this device\n");
1800                 return -EINVAL;
1801         }
1802
1803         /* rate is given in Mb/sec, HW config is in Kb/sec */
1804         rate = rate << 10;
1805
1806         /* Check whether rate in valid range, 0 is always valid */
1807         if (rate && !mlx5_rl_is_in_range(mdev, rate)) {
1808                 netdev_err(dev, "TX rate %u, is not in range\n", rate);
1809                 return -ERANGE;
1810         }
1811
1812         mutex_lock(&priv->state_lock);
1813         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
1814                 err = mlx5e_set_sq_maxrate(dev, sq, rate);
1815         if (!err)
1816                 priv->tx_rates[index] = rate;
1817         mutex_unlock(&priv->state_lock);
1818
1819         return err;
1820 }
1821
1822 static int mlx5e_open_rxq_rq(struct mlx5e_channel *c, struct mlx5e_params *params,
1823                              struct mlx5e_rq_param *rq_params)
1824 {
1825         int err;
1826
1827         err = mlx5e_init_rxq_rq(c, params, &c->rq);
1828         if (err)
1829                 return err;
1830
1831         return mlx5e_open_rq(params, rq_params, NULL, cpu_to_node(c->cpu), &c->rq);
1832 }
1833
1834 static int mlx5e_open_queues(struct mlx5e_channel *c,
1835                              struct mlx5e_params *params,
1836                              struct mlx5e_channel_param *cparam)
1837 {
1838         struct dim_cq_moder icocq_moder = {0, 0};
1839         struct mlx5e_create_cq_param ccp;
1840         int err;
1841
1842         mlx5e_build_create_cq_param(&ccp, c);
1843
1844         err = mlx5e_open_cq(c->priv, icocq_moder, &cparam->async_icosq.cqp, &ccp,
1845                             &c->async_icosq.cq);
1846         if (err)
1847                 return err;
1848
1849         err = mlx5e_open_cq(c->priv, icocq_moder, &cparam->icosq.cqp, &ccp,
1850                             &c->icosq.cq);
1851         if (err)
1852                 goto err_close_async_icosq_cq;
1853
1854         err = mlx5e_open_tx_cqs(c, params, &ccp, cparam);
1855         if (err)
1856                 goto err_close_icosq_cq;
1857
1858         err = mlx5e_open_cq(c->priv, params->tx_cq_moderation, &cparam->xdp_sq.cqp, &ccp,
1859                             &c->xdpsq.cq);
1860         if (err)
1861                 goto err_close_tx_cqs;
1862
1863         err = mlx5e_open_cq(c->priv, params->rx_cq_moderation, &cparam->rq.cqp, &ccp,
1864                             &c->rq.cq);
1865         if (err)
1866                 goto err_close_xdp_tx_cqs;
1867
1868         err = c->xdp ? mlx5e_open_cq(c->priv, params->tx_cq_moderation, &cparam->xdp_sq.cqp,
1869                                      &ccp, &c->rq_xdpsq.cq) : 0;
1870         if (err)
1871                 goto err_close_rx_cq;
1872
1873         spin_lock_init(&c->async_icosq_lock);
1874
1875         err = mlx5e_open_icosq(c, params, &cparam->async_icosq, &c->async_icosq);
1876         if (err)
1877                 goto err_close_xdpsq_cq;
1878
1879         err = mlx5e_open_icosq(c, params, &cparam->icosq, &c->icosq);
1880         if (err)
1881                 goto err_close_async_icosq;
1882
1883         err = mlx5e_open_sqs(c, params, cparam);
1884         if (err)
1885                 goto err_close_icosq;
1886
1887         if (c->xdp) {
1888                 err = mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, NULL,
1889                                        &c->rq_xdpsq, false);
1890                 if (err)
1891                         goto err_close_sqs;
1892         }
1893
1894         err = mlx5e_open_rxq_rq(c, params, &cparam->rq);
1895         if (err)
1896                 goto err_close_xdp_sq;
1897
1898         err = mlx5e_open_xdpsq(c, params, &cparam->xdp_sq, NULL, &c->xdpsq, true);
1899         if (err)
1900                 goto err_close_rq;
1901
1902         return 0;
1903
1904 err_close_rq:
1905         mlx5e_close_rq(&c->rq);
1906
1907 err_close_xdp_sq:
1908         if (c->xdp)
1909                 mlx5e_close_xdpsq(&c->rq_xdpsq);
1910
1911 err_close_sqs:
1912         mlx5e_close_sqs(c);
1913
1914 err_close_icosq:
1915         mlx5e_close_icosq(&c->icosq);
1916
1917 err_close_async_icosq:
1918         mlx5e_close_icosq(&c->async_icosq);
1919
1920 err_close_xdpsq_cq:
1921         if (c->xdp)
1922                 mlx5e_close_cq(&c->rq_xdpsq.cq);
1923
1924 err_close_rx_cq:
1925         mlx5e_close_cq(&c->rq.cq);
1926
1927 err_close_xdp_tx_cqs:
1928         mlx5e_close_cq(&c->xdpsq.cq);
1929
1930 err_close_tx_cqs:
1931         mlx5e_close_tx_cqs(c);
1932
1933 err_close_icosq_cq:
1934         mlx5e_close_cq(&c->icosq.cq);
1935
1936 err_close_async_icosq_cq:
1937         mlx5e_close_cq(&c->async_icosq.cq);
1938
1939         return err;
1940 }
1941
1942 static void mlx5e_close_queues(struct mlx5e_channel *c)
1943 {
1944         mlx5e_close_xdpsq(&c->xdpsq);
1945         mlx5e_close_rq(&c->rq);
1946         if (c->xdp)
1947                 mlx5e_close_xdpsq(&c->rq_xdpsq);
1948         mlx5e_close_sqs(c);
1949         mlx5e_close_icosq(&c->icosq);
1950         mlx5e_close_icosq(&c->async_icosq);
1951         if (c->xdp)
1952                 mlx5e_close_cq(&c->rq_xdpsq.cq);
1953         mlx5e_close_cq(&c->rq.cq);
1954         mlx5e_close_cq(&c->xdpsq.cq);
1955         mlx5e_close_tx_cqs(c);
1956         mlx5e_close_cq(&c->icosq.cq);
1957         mlx5e_close_cq(&c->async_icosq.cq);
1958 }
1959
1960 static u8 mlx5e_enumerate_lag_port(struct mlx5_core_dev *mdev, int ix)
1961 {
1962         u16 port_aff_bias = mlx5_core_is_pf(mdev) ? 0 : MLX5_CAP_GEN(mdev, vhca_id);
1963
1964         return (ix + port_aff_bias) % mlx5e_get_num_lag_ports(mdev);
1965 }
1966
1967 static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
1968                               struct mlx5e_params *params,
1969                               struct mlx5e_channel_param *cparam,
1970                               struct xsk_buff_pool *xsk_pool,
1971                               struct mlx5e_channel **cp)
1972 {
1973         int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(priv->mdev, ix));
1974         struct net_device *netdev = priv->netdev;
1975         struct mlx5e_xsk_param xsk;
1976         struct mlx5e_channel *c;
1977         unsigned int irq;
1978         int err;
1979         int eqn;
1980
1981         err = mlx5_vector2eqn(priv->mdev, ix, &eqn, &irq);
1982         if (err)
1983                 return err;
1984
1985         c = kvzalloc_node(sizeof(*c), GFP_KERNEL, cpu_to_node(cpu));
1986         if (!c)
1987                 return -ENOMEM;
1988
1989         c->priv     = priv;
1990         c->mdev     = priv->mdev;
1991         c->tstamp   = &priv->tstamp;
1992         c->ix       = ix;
1993         c->cpu      = cpu;
1994         c->pdev     = mlx5_core_dma_dev(priv->mdev);
1995         c->netdev   = priv->netdev;
1996         c->mkey_be  = cpu_to_be32(priv->mdev->mlx5e_res.hw_objs.mkey.key);
1997         c->num_tc   = params->num_tc;
1998         c->xdp      = !!params->xdp_prog;
1999         c->stats    = &priv->channel_stats[ix].ch;
2000         c->aff_mask = irq_get_effective_affinity_mask(irq);
2001         c->lag_port = mlx5e_enumerate_lag_port(priv->mdev, ix);
2002
2003         netif_napi_add(netdev, &c->napi, mlx5e_napi_poll, 64);
2004
2005         err = mlx5e_open_queues(c, params, cparam);
2006         if (unlikely(err))
2007                 goto err_napi_del;
2008
2009         if (xsk_pool) {
2010                 mlx5e_build_xsk_param(xsk_pool, &xsk);
2011                 err = mlx5e_open_xsk(priv, params, &xsk, xsk_pool, c);
2012                 if (unlikely(err))
2013                         goto err_close_queues;
2014         }
2015
2016         *cp = c;
2017
2018         return 0;
2019
2020 err_close_queues:
2021         mlx5e_close_queues(c);
2022
2023 err_napi_del:
2024         netif_napi_del(&c->napi);
2025
2026         kvfree(c);
2027
2028         return err;
2029 }
2030
2031 static void mlx5e_activate_channel(struct mlx5e_channel *c)
2032 {
2033         int tc;
2034
2035         napi_enable(&c->napi);
2036
2037         for (tc = 0; tc < c->num_tc; tc++)
2038                 mlx5e_activate_txqsq(&c->sq[tc]);
2039         mlx5e_activate_icosq(&c->icosq);
2040         mlx5e_activate_icosq(&c->async_icosq);
2041         mlx5e_activate_rq(&c->rq);
2042
2043         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2044                 mlx5e_activate_xsk(c);
2045 }
2046
2047 static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
2048 {
2049         int tc;
2050
2051         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2052                 mlx5e_deactivate_xsk(c);
2053
2054         mlx5e_deactivate_rq(&c->rq);
2055         mlx5e_deactivate_icosq(&c->async_icosq);
2056         mlx5e_deactivate_icosq(&c->icosq);
2057         for (tc = 0; tc < c->num_tc; tc++)
2058                 mlx5e_deactivate_txqsq(&c->sq[tc]);
2059         mlx5e_qos_deactivate_queues(c);
2060
2061         napi_disable(&c->napi);
2062 }
2063
2064 static void mlx5e_close_channel(struct mlx5e_channel *c)
2065 {
2066         if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
2067                 mlx5e_close_xsk(c);
2068         mlx5e_close_queues(c);
2069         mlx5e_qos_close_queues(c);
2070         netif_napi_del(&c->napi);
2071
2072         kvfree(c);
2073 }
2074
2075 int mlx5e_open_channels(struct mlx5e_priv *priv,
2076                         struct mlx5e_channels *chs)
2077 {
2078         struct mlx5e_channel_param *cparam;
2079         int err = -ENOMEM;
2080         int i;
2081
2082         chs->num = chs->params.num_channels;
2083
2084         chs->c = kcalloc(chs->num, sizeof(struct mlx5e_channel *), GFP_KERNEL);
2085         cparam = kvzalloc(sizeof(struct mlx5e_channel_param), GFP_KERNEL);
2086         if (!chs->c || !cparam)
2087                 goto err_free;
2088
2089         mlx5e_build_channel_param(priv->mdev, &chs->params, priv->q_counter, cparam);
2090         for (i = 0; i < chs->num; i++) {
2091                 struct xsk_buff_pool *xsk_pool = NULL;
2092
2093                 if (chs->params.xdp_prog)
2094                         xsk_pool = mlx5e_xsk_get_pool(&chs->params, chs->params.xsk, i);
2095
2096                 err = mlx5e_open_channel(priv, i, &chs->params, cparam, xsk_pool, &chs->c[i]);
2097                 if (err)
2098                         goto err_close_channels;
2099         }
2100
2101         if (MLX5E_GET_PFLAG(&chs->params, MLX5E_PFLAG_TX_PORT_TS) || chs->params.ptp_rx) {
2102                 err = mlx5e_ptp_open(priv, &chs->params, chs->c[0]->lag_port, &chs->ptp);
2103                 if (err)
2104                         goto err_close_channels;
2105         }
2106
2107         err = mlx5e_qos_open_queues(priv, chs);
2108         if (err)
2109                 goto err_close_ptp;
2110
2111         mlx5e_health_channels_update(priv);
2112         kvfree(cparam);
2113         return 0;
2114
2115 err_close_ptp:
2116         if (chs->ptp)
2117                 mlx5e_ptp_close(chs->ptp);
2118
2119 err_close_channels:
2120         for (i--; i >= 0; i--)
2121                 mlx5e_close_channel(chs->c[i]);
2122
2123 err_free:
2124         kfree(chs->c);
2125         kvfree(cparam);
2126         chs->num = 0;
2127         return err;
2128 }
2129
2130 static void mlx5e_activate_channels(struct mlx5e_channels *chs)
2131 {
2132         int i;
2133
2134         for (i = 0; i < chs->num; i++)
2135                 mlx5e_activate_channel(chs->c[i]);
2136
2137         if (chs->ptp)
2138                 mlx5e_ptp_activate_channel(chs->ptp);
2139 }
2140
2141 #define MLX5E_RQ_WQES_TIMEOUT 20000 /* msecs */
2142
2143 static int mlx5e_wait_channels_min_rx_wqes(struct mlx5e_channels *chs)
2144 {
2145         int err = 0;
2146         int i;
2147
2148         for (i = 0; i < chs->num; i++) {
2149                 int timeout = err ? 0 : MLX5E_RQ_WQES_TIMEOUT;
2150
2151                 err |= mlx5e_wait_for_min_rx_wqes(&chs->c[i]->rq, timeout);
2152
2153                 /* Don't wait on the XSK RQ, because the newer xdpsock sample
2154                  * doesn't provide any Fill Ring entries at the setup stage.
2155                  */
2156         }
2157
2158         return err ? -ETIMEDOUT : 0;
2159 }
2160
2161 static void mlx5e_deactivate_channels(struct mlx5e_channels *chs)
2162 {
2163         int i;
2164
2165         if (chs->ptp)
2166                 mlx5e_ptp_deactivate_channel(chs->ptp);
2167
2168         for (i = 0; i < chs->num; i++)
2169                 mlx5e_deactivate_channel(chs->c[i]);
2170 }
2171
2172 void mlx5e_close_channels(struct mlx5e_channels *chs)
2173 {
2174         int i;
2175
2176         if (chs->ptp) {
2177                 mlx5e_ptp_close(chs->ptp);
2178                 chs->ptp = NULL;
2179         }
2180         for (i = 0; i < chs->num; i++)
2181                 mlx5e_close_channel(chs->c[i]);
2182
2183         kfree(chs->c);
2184         chs->num = 0;
2185 }
2186
2187 static int
2188 mlx5e_create_rqt(struct mlx5e_priv *priv, int sz, struct mlx5e_rqt *rqt)
2189 {
2190         struct mlx5_core_dev *mdev = priv->mdev;
2191         void *rqtc;
2192         int inlen;
2193         int err;
2194         u32 *in;
2195         int i;
2196
2197         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2198         in = kvzalloc(inlen, GFP_KERNEL);
2199         if (!in)
2200                 return -ENOMEM;
2201
2202         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2203
2204         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2205         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2206
2207         for (i = 0; i < sz; i++)
2208                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->drop_rq.rqn);
2209
2210         err = mlx5_core_create_rqt(mdev, in, inlen, &rqt->rqtn);
2211         if (!err)
2212                 rqt->enabled = true;
2213
2214         kvfree(in);
2215         return err;
2216 }
2217
2218 void mlx5e_destroy_rqt(struct mlx5e_priv *priv, struct mlx5e_rqt *rqt)
2219 {
2220         rqt->enabled = false;
2221         mlx5_core_destroy_rqt(priv->mdev, rqt->rqtn);
2222 }
2223
2224 int mlx5e_create_indirect_rqt(struct mlx5e_priv *priv)
2225 {
2226         struct mlx5e_rqt *rqt = &priv->indir_rqt;
2227         int err;
2228
2229         err = mlx5e_create_rqt(priv, MLX5E_INDIR_RQT_SIZE, rqt);
2230         if (err)
2231                 mlx5_core_warn(priv->mdev, "create indirect rqts failed, %d\n", err);
2232         return err;
2233 }
2234
2235 int mlx5e_create_direct_rqts(struct mlx5e_priv *priv, struct mlx5e_tir *tirs, int n)
2236 {
2237         int err;
2238         int ix;
2239
2240         for (ix = 0; ix < n; ix++) {
2241                 err = mlx5e_create_rqt(priv, 1 /*size */, &tirs[ix].rqt);
2242                 if (unlikely(err))
2243                         goto err_destroy_rqts;
2244         }
2245
2246         return 0;
2247
2248 err_destroy_rqts:
2249         mlx5_core_warn(priv->mdev, "create rqts failed, %d\n", err);
2250         for (ix--; ix >= 0; ix--)
2251                 mlx5e_destroy_rqt(priv, &tirs[ix].rqt);
2252
2253         return err;
2254 }
2255
2256 void mlx5e_destroy_direct_rqts(struct mlx5e_priv *priv, struct mlx5e_tir *tirs, int n)
2257 {
2258         int i;
2259
2260         for (i = 0; i < n; i++)
2261                 mlx5e_destroy_rqt(priv, &tirs[i].rqt);
2262 }
2263
2264 static int mlx5e_rx_hash_fn(int hfunc)
2265 {
2266         return (hfunc == ETH_RSS_HASH_TOP) ?
2267                MLX5_RX_HASH_FN_TOEPLITZ :
2268                MLX5_RX_HASH_FN_INVERTED_XOR8;
2269 }
2270
2271 int mlx5e_bits_invert(unsigned long a, int size)
2272 {
2273         int inv = 0;
2274         int i;
2275
2276         for (i = 0; i < size; i++)
2277                 inv |= (test_bit(size - i - 1, &a) ? 1 : 0) << i;
2278
2279         return inv;
2280 }
2281
2282 static void mlx5e_fill_rqt_rqns(struct mlx5e_priv *priv, int sz,
2283                                 struct mlx5e_redirect_rqt_param rrp, void *rqtc)
2284 {
2285         int i;
2286
2287         for (i = 0; i < sz; i++) {
2288                 u32 rqn;
2289
2290                 if (rrp.is_rss) {
2291                         int ix = i;
2292
2293                         if (rrp.rss.hfunc == ETH_RSS_HASH_XOR)
2294                                 ix = mlx5e_bits_invert(i, ilog2(sz));
2295
2296                         ix = priv->rss_params.indirection_rqt[ix];
2297                         rqn = rrp.rss.channels->c[ix]->rq.rqn;
2298                 } else {
2299                         rqn = rrp.rqn;
2300                 }
2301                 MLX5_SET(rqtc, rqtc, rq_num[i], rqn);
2302         }
2303 }
2304
2305 int mlx5e_redirect_rqt(struct mlx5e_priv *priv, u32 rqtn, int sz,
2306                        struct mlx5e_redirect_rqt_param rrp)
2307 {
2308         struct mlx5_core_dev *mdev = priv->mdev;
2309         void *rqtc;
2310         int inlen;
2311         u32 *in;
2312         int err;
2313
2314         inlen = MLX5_ST_SZ_BYTES(modify_rqt_in) + sizeof(u32) * sz;
2315         in = kvzalloc(inlen, GFP_KERNEL);
2316         if (!in)
2317                 return -ENOMEM;
2318
2319         rqtc = MLX5_ADDR_OF(modify_rqt_in, in, ctx);
2320
2321         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2322         MLX5_SET(modify_rqt_in, in, bitmask.rqn_list, 1);
2323         mlx5e_fill_rqt_rqns(priv, sz, rrp, rqtc);
2324         err = mlx5_core_modify_rqt(mdev, rqtn, in, inlen);
2325
2326         kvfree(in);
2327         return err;
2328 }
2329
2330 static u32 mlx5e_get_direct_rqn(struct mlx5e_priv *priv, int ix,
2331                                 struct mlx5e_redirect_rqt_param rrp)
2332 {
2333         if (!rrp.is_rss)
2334                 return rrp.rqn;
2335
2336         if (ix >= rrp.rss.channels->num)
2337                 return priv->drop_rq.rqn;
2338
2339         return rrp.rss.channels->c[ix]->rq.rqn;
2340 }
2341
2342 static void mlx5e_redirect_rqts(struct mlx5e_priv *priv,
2343                                 struct mlx5e_redirect_rqt_param rrp,
2344                                 struct mlx5e_redirect_rqt_param *ptp_rrp)
2345 {
2346         u32 rqtn;
2347         int ix;
2348
2349         if (priv->indir_rqt.enabled) {
2350                 /* RSS RQ table */
2351                 rqtn = priv->indir_rqt.rqtn;
2352                 mlx5e_redirect_rqt(priv, rqtn, MLX5E_INDIR_RQT_SIZE, rrp);
2353         }
2354
2355         for (ix = 0; ix < priv->max_nch; ix++) {
2356                 struct mlx5e_redirect_rqt_param direct_rrp = {
2357                         .is_rss = false,
2358                         {
2359                                 .rqn    = mlx5e_get_direct_rqn(priv, ix, rrp)
2360                         },
2361                 };
2362
2363                 /* Direct RQ Tables */
2364                 if (!priv->direct_tir[ix].rqt.enabled)
2365                         continue;
2366
2367                 rqtn = priv->direct_tir[ix].rqt.rqtn;
2368                 mlx5e_redirect_rqt(priv, rqtn, 1, direct_rrp);
2369         }
2370         if (ptp_rrp) {
2371                 rqtn = priv->ptp_tir.rqt.rqtn;
2372                 mlx5e_redirect_rqt(priv, rqtn, 1, *ptp_rrp);
2373         }
2374 }
2375
2376 static void mlx5e_redirect_rqts_to_channels(struct mlx5e_priv *priv,
2377                                             struct mlx5e_channels *chs)
2378 {
2379         bool rx_ptp_support = priv->profile->rx_ptp_support;
2380         struct mlx5e_redirect_rqt_param *ptp_rrp_p = NULL;
2381         struct mlx5e_redirect_rqt_param rrp = {
2382                 .is_rss        = true,
2383                 {
2384                         .rss = {
2385                                 .channels  = chs,
2386                                 .hfunc     = priv->rss_params.hfunc,
2387                         }
2388                 },
2389         };
2390         struct mlx5e_redirect_rqt_param ptp_rrp;
2391
2392         if (rx_ptp_support) {
2393                 u32 ptp_rqn;
2394
2395                 ptp_rrp.is_rss = false;
2396                 ptp_rrp.rqn = mlx5e_ptp_get_rqn(priv->channels.ptp, &ptp_rqn) ?
2397                               priv->drop_rq.rqn : ptp_rqn;
2398                 ptp_rrp_p = &ptp_rrp;
2399         }
2400         mlx5e_redirect_rqts(priv, rrp, ptp_rrp_p);
2401 }
2402
2403 static void mlx5e_redirect_rqts_to_drop(struct mlx5e_priv *priv)
2404 {
2405         bool rx_ptp_support = priv->profile->rx_ptp_support;
2406         struct mlx5e_redirect_rqt_param drop_rrp = {
2407                 .is_rss = false,
2408                 {
2409                         .rqn = priv->drop_rq.rqn,
2410                 },
2411         };
2412
2413         mlx5e_redirect_rqts(priv, drop_rrp, rx_ptp_support ? &drop_rrp : NULL);
2414 }
2415
2416 static const struct mlx5e_tirc_config tirc_default_config[MLX5E_NUM_INDIR_TIRS] = {
2417         [MLX5E_TT_IPV4_TCP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2418                                 .l4_prot_type = MLX5_L4_PROT_TYPE_TCP,
2419                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2420         },
2421         [MLX5E_TT_IPV6_TCP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2422                                 .l4_prot_type = MLX5_L4_PROT_TYPE_TCP,
2423                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2424         },
2425         [MLX5E_TT_IPV4_UDP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2426                                 .l4_prot_type = MLX5_L4_PROT_TYPE_UDP,
2427                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2428         },
2429         [MLX5E_TT_IPV6_UDP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2430                                 .l4_prot_type = MLX5_L4_PROT_TYPE_UDP,
2431                                 .rx_hash_fields = MLX5_HASH_IP_L4PORTS,
2432         },
2433         [MLX5E_TT_IPV4_IPSEC_AH] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2434                                      .l4_prot_type = 0,
2435                                      .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2436         },
2437         [MLX5E_TT_IPV6_IPSEC_AH] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2438                                      .l4_prot_type = 0,
2439                                      .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2440         },
2441         [MLX5E_TT_IPV4_IPSEC_ESP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2442                                       .l4_prot_type = 0,
2443                                       .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2444         },
2445         [MLX5E_TT_IPV6_IPSEC_ESP] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2446                                       .l4_prot_type = 0,
2447                                       .rx_hash_fields = MLX5_HASH_IP_IPSEC_SPI,
2448         },
2449         [MLX5E_TT_IPV4] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV4,
2450                             .l4_prot_type = 0,
2451                             .rx_hash_fields = MLX5_HASH_IP,
2452         },
2453         [MLX5E_TT_IPV6] = { .l3_prot_type = MLX5_L3_PROT_TYPE_IPV6,
2454                             .l4_prot_type = 0,
2455                             .rx_hash_fields = MLX5_HASH_IP,
2456         },
2457 };
2458
2459 struct mlx5e_tirc_config mlx5e_tirc_get_default_config(enum mlx5e_traffic_types tt)
2460 {
2461         return tirc_default_config[tt];
2462 }
2463
2464 static void mlx5e_build_tir_ctx_lro(struct mlx5e_params *params, void *tirc)
2465 {
2466         if (!params->lro_en)
2467                 return;
2468
2469 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2470
2471         MLX5_SET(tirc, tirc, lro_enable_mask,
2472                  MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2473                  MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2474         MLX5_SET(tirc, tirc, lro_max_ip_payload_size,
2475                  (MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ - ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2476         MLX5_SET(tirc, tirc, lro_timeout_period_usecs, params->lro_timeout);
2477 }
2478
2479 void mlx5e_build_indir_tir_ctx_hash(struct mlx5e_rss_params *rss_params,
2480                                     const struct mlx5e_tirc_config *ttconfig,
2481                                     void *tirc, bool inner)
2482 {
2483         void *hfso = inner ? MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_inner) :
2484                              MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2485
2486         MLX5_SET(tirc, tirc, rx_hash_fn, mlx5e_rx_hash_fn(rss_params->hfunc));
2487         if (rss_params->hfunc == ETH_RSS_HASH_TOP) {
2488                 void *rss_key = MLX5_ADDR_OF(tirc, tirc,
2489                                              rx_hash_toeplitz_key);
2490                 size_t len = MLX5_FLD_SZ_BYTES(tirc,
2491                                                rx_hash_toeplitz_key);
2492
2493                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2494                 memcpy(rss_key, rss_params->toeplitz_hash_key, len);
2495         }
2496         MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2497                  ttconfig->l3_prot_type);
2498         MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2499                  ttconfig->l4_prot_type);
2500         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2501                  ttconfig->rx_hash_fields);
2502 }
2503
2504 static void mlx5e_update_rx_hash_fields(struct mlx5e_tirc_config *ttconfig,
2505                                         enum mlx5e_traffic_types tt,
2506                                         u32 rx_hash_fields)
2507 {
2508         *ttconfig                = tirc_default_config[tt];
2509         ttconfig->rx_hash_fields = rx_hash_fields;
2510 }
2511
2512 void mlx5e_modify_tirs_hash(struct mlx5e_priv *priv, void *in)
2513 {
2514         void *tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2515         struct mlx5e_rss_params *rss = &priv->rss_params;
2516         struct mlx5_core_dev *mdev = priv->mdev;
2517         int ctxlen = MLX5_ST_SZ_BYTES(tirc);
2518         struct mlx5e_tirc_config ttconfig;
2519         int tt;
2520
2521         MLX5_SET(modify_tir_in, in, bitmask.hash, 1);
2522
2523         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2524                 memset(tirc, 0, ctxlen);
2525                 mlx5e_update_rx_hash_fields(&ttconfig, tt,
2526                                             rss->rx_hash_fields[tt]);
2527                 mlx5e_build_indir_tir_ctx_hash(rss, &ttconfig, tirc, false);
2528                 mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in);
2529         }
2530
2531         /* Verify inner tirs resources allocated */
2532         if (!priv->inner_indir_tir[0].tirn)
2533                 return;
2534
2535         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2536                 memset(tirc, 0, ctxlen);
2537                 mlx5e_update_rx_hash_fields(&ttconfig, tt,
2538                                             rss->rx_hash_fields[tt]);
2539                 mlx5e_build_indir_tir_ctx_hash(rss, &ttconfig, tirc, true);
2540                 mlx5_core_modify_tir(mdev, priv->inner_indir_tir[tt].tirn, in);
2541         }
2542 }
2543
2544 static int mlx5e_modify_tirs_lro(struct mlx5e_priv *priv)
2545 {
2546         struct mlx5_core_dev *mdev = priv->mdev;
2547
2548         void *in;
2549         void *tirc;
2550         int inlen;
2551         int err;
2552         int tt;
2553         int ix;
2554
2555         inlen = MLX5_ST_SZ_BYTES(modify_tir_in);
2556         in = kvzalloc(inlen, GFP_KERNEL);
2557         if (!in)
2558                 return -ENOMEM;
2559
2560         MLX5_SET(modify_tir_in, in, bitmask.lro, 1);
2561         tirc = MLX5_ADDR_OF(modify_tir_in, in, ctx);
2562
2563         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
2564
2565         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
2566                 err = mlx5_core_modify_tir(mdev, priv->indir_tir[tt].tirn, in);
2567                 if (err)
2568                         goto free_in;
2569         }
2570
2571         for (ix = 0; ix < priv->max_nch; ix++) {
2572                 err = mlx5_core_modify_tir(mdev, priv->direct_tir[ix].tirn, in);
2573                 if (err)
2574                         goto free_in;
2575         }
2576
2577 free_in:
2578         kvfree(in);
2579
2580         return err;
2581 }
2582
2583 static MLX5E_DEFINE_PREACTIVATE_WRAPPER_CTX(mlx5e_modify_tirs_lro);
2584
2585 static int mlx5e_set_mtu(struct mlx5_core_dev *mdev,
2586                          struct mlx5e_params *params, u16 mtu)
2587 {
2588         u16 hw_mtu = MLX5E_SW2HW_MTU(params, mtu);
2589         int err;
2590
2591         err = mlx5_set_port_mtu(mdev, hw_mtu, 1);
2592         if (err)
2593                 return err;
2594
2595         /* Update vport context MTU */
2596         mlx5_modify_nic_vport_mtu(mdev, hw_mtu);
2597         return 0;
2598 }
2599
2600 static void mlx5e_query_mtu(struct mlx5_core_dev *mdev,
2601                             struct mlx5e_params *params, u16 *mtu)
2602 {
2603         u16 hw_mtu = 0;
2604         int err;
2605
2606         err = mlx5_query_nic_vport_mtu(mdev, &hw_mtu);
2607         if (err || !hw_mtu) /* fallback to port oper mtu */
2608                 mlx5_query_port_oper_mtu(mdev, &hw_mtu, 1);
2609
2610         *mtu = MLX5E_HW2SW_MTU(params, hw_mtu);
2611 }
2612
2613 int mlx5e_set_dev_port_mtu(struct mlx5e_priv *priv)
2614 {
2615         struct mlx5e_params *params = &priv->channels.params;
2616         struct net_device *netdev = priv->netdev;
2617         struct mlx5_core_dev *mdev = priv->mdev;
2618         u16 mtu;
2619         int err;
2620
2621         err = mlx5e_set_mtu(mdev, params, params->sw_mtu);
2622         if (err)
2623                 return err;
2624
2625         mlx5e_query_mtu(mdev, params, &mtu);
2626         if (mtu != params->sw_mtu)
2627                 netdev_warn(netdev, "%s: VPort MTU %d is different than netdev mtu %d\n",
2628                             __func__, mtu, params->sw_mtu);
2629
2630         params->sw_mtu = mtu;
2631         return 0;
2632 }
2633
2634 MLX5E_DEFINE_PREACTIVATE_WRAPPER_CTX(mlx5e_set_dev_port_mtu);
2635
2636 void mlx5e_set_netdev_mtu_boundaries(struct mlx5e_priv *priv)
2637 {
2638         struct mlx5e_params *params = &priv->channels.params;
2639         struct net_device *netdev   = priv->netdev;
2640         struct mlx5_core_dev *mdev  = priv->mdev;
2641         u16 max_mtu;
2642
2643         /* MTU range: 68 - hw-specific max */
2644         netdev->min_mtu = ETH_MIN_MTU;
2645
2646         mlx5_query_port_max_mtu(mdev, &max_mtu, 1);
2647         netdev->max_mtu = min_t(unsigned int, MLX5E_HW2SW_MTU(params, max_mtu),
2648                                 ETH_MAX_MTU);
2649 }
2650
2651 static void mlx5e_netdev_set_tcs(struct net_device *netdev, u16 nch, u8 ntc)
2652 {
2653         int tc;
2654
2655         netdev_reset_tc(netdev);
2656
2657         if (ntc == 1)
2658                 return;
2659
2660         netdev_set_num_tc(netdev, ntc);
2661
2662         /* Map netdev TCs to offset 0
2663          * We have our own UP to TXQ mapping for QoS
2664          */
2665         for (tc = 0; tc < ntc; tc++)
2666                 netdev_set_tc_queue(netdev, tc, nch, 0);
2667 }
2668
2669 int mlx5e_update_tx_netdev_queues(struct mlx5e_priv *priv)
2670 {
2671         int qos_queues, nch, ntc, num_txqs, err;
2672
2673         qos_queues = mlx5e_qos_cur_leaf_nodes(priv);
2674
2675         nch = priv->channels.params.num_channels;
2676         ntc = priv->channels.params.num_tc;
2677         num_txqs = nch * ntc + qos_queues;
2678         if (MLX5E_GET_PFLAG(&priv->channels.params, MLX5E_PFLAG_TX_PORT_TS))
2679                 num_txqs += ntc;
2680
2681         mlx5e_dbg(DRV, priv, "Setting num_txqs %d\n", num_txqs);
2682         err = netif_set_real_num_tx_queues(priv->netdev, num_txqs);
2683         if (err)
2684                 netdev_warn(priv->netdev, "netif_set_real_num_tx_queues failed, %d\n", err);
2685
2686         return err;
2687 }
2688
2689 static int mlx5e_update_netdev_queues(struct mlx5e_priv *priv)
2690 {
2691         struct net_device *netdev = priv->netdev;
2692         int old_num_txqs, old_ntc;
2693         int num_rxqs, nch, ntc;
2694         int err;
2695
2696         old_num_txqs = netdev->real_num_tx_queues;
2697         old_ntc = netdev->num_tc;
2698
2699         nch = priv->channels.params.num_channels;
2700         ntc = priv->channels.params.num_tc;
2701         num_rxqs = nch * priv->profile->rq_groups;
2702         if (priv->channels.params.ptp_rx)
2703                 num_rxqs++;
2704
2705         mlx5e_netdev_set_tcs(netdev, nch, ntc);
2706
2707         err = mlx5e_update_tx_netdev_queues(priv);
2708         if (err)
2709                 goto err_tcs;
2710         err = netif_set_real_num_rx_queues(netdev, num_rxqs);
2711         if (err) {
2712                 netdev_warn(netdev, "netif_set_real_num_rx_queues failed, %d\n", err);
2713                 goto err_txqs;
2714         }
2715
2716         return 0;
2717
2718 err_txqs:
2719         /* netif_set_real_num_rx_queues could fail only when nch increased. Only
2720          * one of nch and ntc is changed in this function. That means, the call
2721          * to netif_set_real_num_tx_queues below should not fail, because it
2722          * decreases the number of TX queues.
2723          */
2724         WARN_ON_ONCE(netif_set_real_num_tx_queues(netdev, old_num_txqs));
2725
2726 err_tcs:
2727         mlx5e_netdev_set_tcs(netdev, old_num_txqs / old_ntc, old_ntc);
2728         return err;
2729 }
2730
2731 static void mlx5e_set_default_xps_cpumasks(struct mlx5e_priv *priv,
2732                                            struct mlx5e_params *params)
2733 {
2734         struct mlx5_core_dev *mdev = priv->mdev;
2735         int num_comp_vectors, ix, irq;
2736
2737         num_comp_vectors = mlx5_comp_vectors_count(mdev);
2738
2739         for (ix = 0; ix < params->num_channels; ix++) {
2740                 cpumask_clear(priv->scratchpad.cpumask);
2741
2742                 for (irq = ix; irq < num_comp_vectors; irq += params->num_channels) {
2743                         int cpu = cpumask_first(mlx5_comp_irq_get_affinity_mask(mdev, irq));
2744
2745                         cpumask_set_cpu(cpu, priv->scratchpad.cpumask);
2746                 }
2747
2748                 netif_set_xps_queue(priv->netdev, priv->scratchpad.cpumask, ix);
2749         }
2750 }
2751
2752 int mlx5e_num_channels_changed(struct mlx5e_priv *priv)
2753 {
2754         u16 count = priv->channels.params.num_channels;
2755         int err;
2756
2757         err = mlx5e_update_netdev_queues(priv);
2758         if (err)
2759                 return err;
2760
2761         mlx5e_set_default_xps_cpumasks(priv, &priv->channels.params);
2762
2763         if (!netif_is_rxfh_configured(priv->netdev))
2764                 mlx5e_build_default_indir_rqt(priv->rss_params.indirection_rqt,
2765                                               MLX5E_INDIR_RQT_SIZE, count);
2766
2767         return 0;
2768 }
2769
2770 MLX5E_DEFINE_PREACTIVATE_WRAPPER_CTX(mlx5e_num_channels_changed);
2771
2772 static void mlx5e_build_txq_maps(struct mlx5e_priv *priv)
2773 {
2774         int i, ch, tc, num_tc;
2775
2776         ch = priv->channels.num;
2777         num_tc = priv->channels.params.num_tc;
2778
2779         for (i = 0; i < ch; i++) {
2780                 for (tc = 0; tc < num_tc; tc++) {
2781                         struct mlx5e_channel *c = priv->channels.c[i];
2782                         struct mlx5e_txqsq *sq = &c->sq[tc];
2783
2784                         priv->txq2sq[sq->txq_ix] = sq;
2785                         priv->channel_tc2realtxq[i][tc] = i + tc * ch;
2786                 }
2787         }
2788
2789         if (!priv->channels.ptp)
2790                 return;
2791
2792         if (!test_bit(MLX5E_PTP_STATE_TX, priv->channels.ptp->state))
2793                 return;
2794
2795         for (tc = 0; tc < num_tc; tc++) {
2796                 struct mlx5e_ptp *c = priv->channels.ptp;
2797                 struct mlx5e_txqsq *sq = &c->ptpsq[tc].txqsq;
2798
2799                 priv->txq2sq[sq->txq_ix] = sq;
2800                 priv->port_ptp_tc2realtxq[tc] = priv->num_tc_x_num_ch + tc;
2801         }
2802 }
2803
2804 static void mlx5e_update_num_tc_x_num_ch(struct mlx5e_priv *priv)
2805 {
2806         /* Sync with mlx5e_select_queue. */
2807         WRITE_ONCE(priv->num_tc_x_num_ch,
2808                    priv->channels.params.num_tc * priv->channels.num);
2809 }
2810
2811 void mlx5e_activate_priv_channels(struct mlx5e_priv *priv)
2812 {
2813         mlx5e_update_num_tc_x_num_ch(priv);
2814         mlx5e_build_txq_maps(priv);
2815         mlx5e_activate_channels(&priv->channels);
2816         mlx5e_qos_activate_queues(priv);
2817         mlx5e_xdp_tx_enable(priv);
2818         netif_tx_start_all_queues(priv->netdev);
2819
2820         if (mlx5e_is_vport_rep(priv))
2821                 mlx5e_add_sqs_fwd_rules(priv);
2822
2823         mlx5e_wait_channels_min_rx_wqes(&priv->channels);
2824         mlx5e_redirect_rqts_to_channels(priv, &priv->channels);
2825
2826         mlx5e_xsk_redirect_rqts_to_channels(priv, &priv->channels);
2827 }
2828
2829 void mlx5e_deactivate_priv_channels(struct mlx5e_priv *priv)
2830 {
2831         mlx5e_xsk_redirect_rqts_to_drop(priv, &priv->channels);
2832
2833         mlx5e_redirect_rqts_to_drop(priv);
2834
2835         if (mlx5e_is_vport_rep(priv))
2836                 mlx5e_remove_sqs_fwd_rules(priv);
2837
2838         /* FIXME: This is a W/A only for tx timeout watch dog false alarm when
2839          * polling for inactive tx queues.
2840          */
2841         netif_tx_stop_all_queues(priv->netdev);
2842         netif_tx_disable(priv->netdev);
2843         mlx5e_xdp_tx_disable(priv);
2844         mlx5e_deactivate_channels(&priv->channels);
2845 }
2846
2847 static int mlx5e_switch_priv_channels(struct mlx5e_priv *priv,
2848                                       struct mlx5e_channels *new_chs,
2849                                       mlx5e_fp_preactivate preactivate,
2850                                       void *context)
2851 {
2852         struct net_device *netdev = priv->netdev;
2853         struct mlx5e_channels old_chs;
2854         int carrier_ok;
2855         bool opened;
2856         int err = 0;
2857
2858         opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2859         if (opened) {
2860                 carrier_ok = netif_carrier_ok(netdev);
2861                 netif_carrier_off(netdev);
2862
2863                 mlx5e_deactivate_priv_channels(priv);
2864         }
2865
2866         old_chs = priv->channels;
2867         priv->channels = *new_chs;
2868
2869         /* New channels are ready to roll, call the preactivate hook if needed
2870          * to modify HW settings or update kernel parameters.
2871          */
2872         if (preactivate) {
2873                 err = preactivate(priv, context);
2874                 if (err) {
2875                         priv->channels = old_chs;
2876                         goto out;
2877                 }
2878         }
2879
2880         if (opened) {
2881                 mlx5e_close_channels(&old_chs);
2882                 priv->profile->update_rx(priv);
2883         }
2884
2885 out:
2886         if (opened) {
2887                 mlx5e_activate_priv_channels(priv);
2888
2889                 /* return carrier back if needed */
2890                 if (carrier_ok)
2891                         netif_carrier_on(netdev);
2892         }
2893
2894         return err;
2895 }
2896
2897 int mlx5e_safe_switch_channels(struct mlx5e_priv *priv,
2898                                struct mlx5e_channels *new_chs,
2899                                mlx5e_fp_preactivate preactivate,
2900                                void *context)
2901 {
2902         bool opened;
2903         int err;
2904
2905         opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2906
2907         if (opened) {
2908                 err = mlx5e_open_channels(priv, new_chs);
2909                 if (err)
2910                         return err;
2911         }
2912
2913         err = mlx5e_switch_priv_channels(priv, new_chs, preactivate, context);
2914         if (err)
2915                 goto err_close;
2916
2917         return 0;
2918
2919 err_close:
2920         if (opened)
2921                 mlx5e_close_channels(new_chs);
2922
2923         return err;
2924 }
2925
2926 int mlx5e_safe_reopen_channels(struct mlx5e_priv *priv)
2927 {
2928         struct mlx5e_channels new_channels = {};
2929
2930         new_channels.params = priv->channels.params;
2931         return mlx5e_safe_switch_channels(priv, &new_channels, NULL, NULL);
2932 }
2933
2934 void mlx5e_timestamp_init(struct mlx5e_priv *priv)
2935 {
2936         priv->tstamp.tx_type   = HWTSTAMP_TX_OFF;
2937         priv->tstamp.rx_filter = HWTSTAMP_FILTER_NONE;
2938 }
2939
2940 static void mlx5e_modify_admin_state(struct mlx5_core_dev *mdev,
2941                                      enum mlx5_port_status state)
2942 {
2943         struct mlx5_eswitch *esw = mdev->priv.eswitch;
2944         int vport_admin_state;
2945
2946         mlx5_set_port_admin_status(mdev, state);
2947
2948         if (mlx5_eswitch_mode(mdev) == MLX5_ESWITCH_OFFLOADS ||
2949             !MLX5_CAP_GEN(mdev, uplink_follow))
2950                 return;
2951
2952         if (state == MLX5_PORT_UP)
2953                 vport_admin_state = MLX5_VPORT_ADMIN_STATE_AUTO;
2954         else
2955                 vport_admin_state = MLX5_VPORT_ADMIN_STATE_DOWN;
2956
2957         mlx5_eswitch_set_vport_state(esw, MLX5_VPORT_UPLINK, vport_admin_state);
2958 }
2959
2960 int mlx5e_open_locked(struct net_device *netdev)
2961 {
2962         struct mlx5e_priv *priv = netdev_priv(netdev);
2963         int err;
2964
2965         set_bit(MLX5E_STATE_OPENED, &priv->state);
2966
2967         err = mlx5e_open_channels(priv, &priv->channels);
2968         if (err)
2969                 goto err_clear_state_opened_flag;
2970
2971         priv->profile->update_rx(priv);
2972         mlx5e_activate_priv_channels(priv);
2973         mlx5e_apply_traps(priv, true);
2974         if (priv->profile->update_carrier)
2975                 priv->profile->update_carrier(priv);
2976
2977         mlx5e_queue_update_stats(priv);
2978         return 0;
2979
2980 err_clear_state_opened_flag:
2981         clear_bit(MLX5E_STATE_OPENED, &priv->state);
2982         return err;
2983 }
2984
2985 int mlx5e_open(struct net_device *netdev)
2986 {
2987         struct mlx5e_priv *priv = netdev_priv(netdev);
2988         int err;
2989
2990         mutex_lock(&priv->state_lock);
2991         err = mlx5e_open_locked(netdev);
2992         if (!err)
2993                 mlx5e_modify_admin_state(priv->mdev, MLX5_PORT_UP);
2994         mutex_unlock(&priv->state_lock);
2995
2996         return err;
2997 }
2998
2999 int mlx5e_close_locked(struct net_device *netdev)
3000 {
3001         struct mlx5e_priv *priv = netdev_priv(netdev);
3002
3003         /* May already be CLOSED in case a previous configuration operation
3004          * (e.g RX/TX queue size change) that involves close&open failed.
3005          */
3006         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3007                 return 0;
3008
3009         mlx5e_apply_traps(priv, false);
3010         clear_bit(MLX5E_STATE_OPENED, &priv->state);
3011
3012         netif_carrier_off(priv->netdev);
3013         mlx5e_deactivate_priv_channels(priv);
3014         mlx5e_close_channels(&priv->channels);
3015
3016         return 0;
3017 }
3018
3019 int mlx5e_close(struct net_device *netdev)
3020 {
3021         struct mlx5e_priv *priv = netdev_priv(netdev);
3022         int err;
3023
3024         if (!netif_device_present(netdev))
3025                 return -ENODEV;
3026
3027         mutex_lock(&priv->state_lock);
3028         mlx5e_modify_admin_state(priv->mdev, MLX5_PORT_DOWN);
3029         err = mlx5e_close_locked(netdev);
3030         mutex_unlock(&priv->state_lock);
3031
3032         return err;
3033 }
3034
3035 static void mlx5e_free_drop_rq(struct mlx5e_rq *rq)
3036 {
3037         mlx5_wq_destroy(&rq->wq_ctrl);
3038 }
3039
3040 static int mlx5e_alloc_drop_rq(struct mlx5_core_dev *mdev,
3041                                struct mlx5e_rq *rq,
3042                                struct mlx5e_rq_param *param)
3043 {
3044         void *rqc = param->rqc;
3045         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
3046         int err;
3047
3048         param->wq.db_numa_node = param->wq.buf_numa_node;
3049
3050         err = mlx5_wq_cyc_create(mdev, &param->wq, rqc_wq, &rq->wqe.wq,
3051                                  &rq->wq_ctrl);
3052         if (err)
3053                 return err;
3054
3055         /* Mark as unused given "Drop-RQ" packets never reach XDP */
3056         xdp_rxq_info_unused(&rq->xdp_rxq);
3057
3058         rq->mdev = mdev;
3059
3060         return 0;
3061 }
3062
3063 static int mlx5e_alloc_drop_cq(struct mlx5e_priv *priv,
3064                                struct mlx5e_cq *cq,
3065                                struct mlx5e_cq_param *param)
3066 {
3067         struct mlx5_core_dev *mdev = priv->mdev;
3068
3069         param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
3070         param->wq.db_numa_node  = dev_to_node(mlx5_core_dma_dev(mdev));
3071
3072         return mlx5e_alloc_cq_common(priv, param, cq);
3073 }
3074
3075 int mlx5e_open_drop_rq(struct mlx5e_priv *priv,
3076                        struct mlx5e_rq *drop_rq)
3077 {
3078         struct mlx5_core_dev *mdev = priv->mdev;
3079         struct mlx5e_cq_param cq_param = {};
3080         struct mlx5e_rq_param rq_param = {};
3081         struct mlx5e_cq *cq = &drop_rq->cq;
3082         int err;
3083
3084         mlx5e_build_drop_rq_param(mdev, priv->drop_rq_q_counter, &rq_param);
3085
3086         err = mlx5e_alloc_drop_cq(priv, cq, &cq_param);
3087         if (err)
3088                 return err;
3089
3090         err = mlx5e_create_cq(cq, &cq_param);
3091         if (err)
3092                 goto err_free_cq;
3093
3094         err = mlx5e_alloc_drop_rq(mdev, drop_rq, &rq_param);
3095         if (err)
3096                 goto err_destroy_cq;
3097
3098         err = mlx5e_create_rq(drop_rq, &rq_param);
3099         if (err)
3100                 goto err_free_rq;
3101
3102         err = mlx5e_modify_rq_state(drop_rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
3103         if (err)
3104                 mlx5_core_warn(priv->mdev, "modify_rq_state failed, rx_if_down_packets won't be counted %d\n", err);
3105
3106         return 0;
3107
3108 err_free_rq:
3109         mlx5e_free_drop_rq(drop_rq);
3110
3111 err_destroy_cq:
3112         mlx5e_destroy_cq(cq);
3113
3114 err_free_cq:
3115         mlx5e_free_cq(cq);
3116
3117         return err;
3118 }
3119
3120 void mlx5e_close_drop_rq(struct mlx5e_rq *drop_rq)
3121 {
3122         mlx5e_destroy_rq(drop_rq);
3123         mlx5e_free_drop_rq(drop_rq);
3124         mlx5e_destroy_cq(&drop_rq->cq);
3125         mlx5e_free_cq(&drop_rq->cq);
3126 }
3127
3128 int mlx5e_create_tis(struct mlx5_core_dev *mdev, void *in, u32 *tisn)
3129 {
3130         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
3131
3132         MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.hw_objs.td.tdn);
3133
3134         if (MLX5_GET(tisc, tisc, tls_en))
3135                 MLX5_SET(tisc, tisc, pd, mdev->mlx5e_res.hw_objs.pdn);
3136
3137         if (mlx5_lag_is_lacp_owner(mdev))
3138                 MLX5_SET(tisc, tisc, strict_lag_tx_port_affinity, 1);
3139
3140         return mlx5_core_create_tis(mdev, in, tisn);
3141 }
3142
3143 void mlx5e_destroy_tis(struct mlx5_core_dev *mdev, u32 tisn)
3144 {
3145         mlx5_core_destroy_tis(mdev, tisn);
3146 }
3147
3148 void mlx5e_destroy_tises(struct mlx5e_priv *priv)
3149 {
3150         int tc, i;
3151
3152         for (i = 0; i < mlx5e_get_num_lag_ports(priv->mdev); i++)
3153                 for (tc = 0; tc < priv->profile->max_tc; tc++)
3154                         mlx5e_destroy_tis(priv->mdev, priv->tisn[i][tc]);
3155 }
3156
3157 static bool mlx5e_lag_should_assign_affinity(struct mlx5_core_dev *mdev)
3158 {
3159         return MLX5_CAP_GEN(mdev, lag_tx_port_affinity) && mlx5e_get_num_lag_ports(mdev) > 1;
3160 }
3161
3162 int mlx5e_create_tises(struct mlx5e_priv *priv)
3163 {
3164         int tc, i;
3165         int err;
3166
3167         for (i = 0; i < mlx5e_get_num_lag_ports(priv->mdev); i++) {
3168                 for (tc = 0; tc < priv->profile->max_tc; tc++) {
3169                         u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
3170                         void *tisc;
3171
3172                         tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
3173
3174                         MLX5_SET(tisc, tisc, prio, tc << 1);
3175
3176                         if (mlx5e_lag_should_assign_affinity(priv->mdev))
3177                                 MLX5_SET(tisc, tisc, lag_tx_port_affinity, i + 1);
3178
3179                         err = mlx5e_create_tis(priv->mdev, in, &priv->tisn[i][tc]);
3180                         if (err)
3181                                 goto err_close_tises;
3182                 }
3183         }
3184
3185         return 0;
3186
3187 err_close_tises:
3188         for (; i >= 0; i--) {
3189                 for (tc--; tc >= 0; tc--)
3190                         mlx5e_destroy_tis(priv->mdev, priv->tisn[i][tc]);
3191                 tc = priv->profile->max_tc;
3192         }
3193
3194         return err;
3195 }
3196
3197 static void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
3198 {
3199         mlx5e_destroy_tises(priv);
3200 }
3201
3202 static void mlx5e_build_indir_tir_ctx_common(struct mlx5e_priv *priv,
3203                                              u32 rqtn, u32 *tirc)
3204 {
3205         MLX5_SET(tirc, tirc, transport_domain, priv->mdev->mlx5e_res.hw_objs.td.tdn);
3206         MLX5_SET(tirc, tirc, disp_type, MLX5_TIRC_DISP_TYPE_INDIRECT);
3207         MLX5_SET(tirc, tirc, indirect_table, rqtn);
3208         MLX5_SET(tirc, tirc, tunneled_offload_en,
3209                  priv->channels.params.tunneled_offload_en);
3210
3211         mlx5e_build_tir_ctx_lro(&priv->channels.params, tirc);
3212 }
3213
3214 static void mlx5e_build_indir_tir_ctx(struct mlx5e_priv *priv,
3215                                       enum mlx5e_traffic_types tt,
3216                                       u32 *tirc)
3217 {
3218         mlx5e_build_indir_tir_ctx_common(priv, priv->indir_rqt.rqtn, tirc);
3219         mlx5e_build_indir_tir_ctx_hash(&priv->rss_params,
3220                                        &tirc_default_config[tt], tirc, false);
3221 }
3222
3223 static void mlx5e_build_direct_tir_ctx(struct mlx5e_priv *priv, u32 rqtn, u32 *tirc)
3224 {
3225         mlx5e_build_indir_tir_ctx_common(priv, rqtn, tirc);
3226         MLX5_SET(tirc, tirc, rx_hash_fn, MLX5_RX_HASH_FN_INVERTED_XOR8);
3227 }
3228
3229 static void mlx5e_build_inner_indir_tir_ctx(struct mlx5e_priv *priv,
3230                                             enum mlx5e_traffic_types tt,
3231                                             u32 *tirc)
3232 {
3233         mlx5e_build_indir_tir_ctx_common(priv, priv->indir_rqt.rqtn, tirc);
3234         mlx5e_build_indir_tir_ctx_hash(&priv->rss_params,
3235                                        &tirc_default_config[tt], tirc, true);
3236 }
3237
3238 int mlx5e_create_indirect_tirs(struct mlx5e_priv *priv, bool inner_ttc)
3239 {
3240         struct mlx5e_tir *tir;
3241         void *tirc;
3242         int inlen;
3243         int i = 0;
3244         int err;
3245         u32 *in;
3246         int tt;
3247
3248         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
3249         in = kvzalloc(inlen, GFP_KERNEL);
3250         if (!in)
3251                 return -ENOMEM;
3252
3253         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++) {
3254                 memset(in, 0, inlen);
3255                 tir = &priv->indir_tir[tt];
3256                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3257                 mlx5e_build_indir_tir_ctx(priv, tt, tirc);
3258                 err = mlx5e_create_tir(priv->mdev, tir, in);
3259                 if (err) {
3260                         mlx5_core_warn(priv->mdev, "create indirect tirs failed, %d\n", err);
3261                         goto err_destroy_inner_tirs;
3262                 }
3263         }
3264
3265         if (!inner_ttc || !mlx5e_tunnel_inner_ft_supported(priv->mdev))
3266                 goto out;
3267
3268         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++) {
3269                 memset(in, 0, inlen);
3270                 tir = &priv->inner_indir_tir[i];
3271                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3272                 mlx5e_build_inner_indir_tir_ctx(priv, i, tirc);
3273                 err = mlx5e_create_tir(priv->mdev, tir, in);
3274                 if (err) {
3275                         mlx5_core_warn(priv->mdev, "create inner indirect tirs failed, %d\n", err);
3276                         goto err_destroy_inner_tirs;
3277                 }
3278         }
3279
3280 out:
3281         kvfree(in);
3282
3283         return 0;
3284
3285 err_destroy_inner_tirs:
3286         for (i--; i >= 0; i--)
3287                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3288
3289         for (tt--; tt >= 0; tt--)
3290                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[tt]);
3291
3292         kvfree(in);
3293
3294         return err;
3295 }
3296
3297 int mlx5e_create_direct_tirs(struct mlx5e_priv *priv, struct mlx5e_tir *tirs, int n)
3298 {
3299         struct mlx5e_tir *tir;
3300         void *tirc;
3301         int inlen;
3302         int err = 0;
3303         u32 *in;
3304         int ix;
3305
3306         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
3307         in = kvzalloc(inlen, GFP_KERNEL);
3308         if (!in)
3309                 return -ENOMEM;
3310
3311         for (ix = 0; ix < n; ix++) {
3312                 memset(in, 0, inlen);
3313                 tir = &tirs[ix];
3314                 tirc = MLX5_ADDR_OF(create_tir_in, in, ctx);
3315                 mlx5e_build_direct_tir_ctx(priv, tir->rqt.rqtn, tirc);
3316                 err = mlx5e_create_tir(priv->mdev, tir, in);
3317                 if (unlikely(err))
3318                         goto err_destroy_ch_tirs;
3319         }
3320
3321         goto out;
3322
3323 err_destroy_ch_tirs:
3324         mlx5_core_warn(priv->mdev, "create tirs failed, %d\n", err);
3325         for (ix--; ix >= 0; ix--)
3326                 mlx5e_destroy_tir(priv->mdev, &tirs[ix]);
3327
3328 out:
3329         kvfree(in);
3330
3331         return err;
3332 }
3333
3334 void mlx5e_destroy_indirect_tirs(struct mlx5e_priv *priv)
3335 {
3336         int i;
3337
3338         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3339                 mlx5e_destroy_tir(priv->mdev, &priv->indir_tir[i]);
3340
3341         /* Verify inner tirs resources allocated */
3342         if (!priv->inner_indir_tir[0].tirn)
3343                 return;
3344
3345         for (i = 0; i < MLX5E_NUM_INDIR_TIRS; i++)
3346                 mlx5e_destroy_tir(priv->mdev, &priv->inner_indir_tir[i]);
3347 }
3348
3349 void mlx5e_destroy_direct_tirs(struct mlx5e_priv *priv, struct mlx5e_tir *tirs, int n)
3350 {
3351         int i;
3352
3353         for (i = 0; i < n; i++)
3354                 mlx5e_destroy_tir(priv->mdev, &tirs[i]);
3355 }
3356
3357 static int mlx5e_modify_channels_scatter_fcs(struct mlx5e_channels *chs, bool enable)
3358 {
3359         int err = 0;
3360         int i;
3361
3362         for (i = 0; i < chs->num; i++) {
3363                 err = mlx5e_modify_rq_scatter_fcs(&chs->c[i]->rq, enable);
3364                 if (err)
3365                         return err;
3366         }
3367
3368         return 0;
3369 }
3370
3371 static int mlx5e_modify_channels_vsd(struct mlx5e_channels *chs, bool vsd)
3372 {
3373         int err = 0;
3374         int i;
3375
3376         for (i = 0; i < chs->num; i++) {
3377                 err = mlx5e_modify_rq_vsd(&chs->c[i]->rq, vsd);
3378                 if (err)
3379                         return err;
3380         }
3381
3382         return 0;
3383 }
3384
3385 static int mlx5e_setup_tc_mqprio(struct mlx5e_priv *priv,
3386                                  struct tc_mqprio_qopt *mqprio)
3387 {
3388         struct mlx5e_channels new_channels = {};
3389         u8 tc = mqprio->num_tc;
3390         int err = 0;
3391
3392         mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
3393
3394         if (tc && tc != MLX5E_MAX_NUM_TC)
3395                 return -EINVAL;
3396
3397         mutex_lock(&priv->state_lock);
3398
3399         /* MQPRIO is another toplevel qdisc that can't be attached
3400          * simultaneously with the offloaded HTB.
3401          */
3402         if (WARN_ON(priv->htb.maj_id)) {
3403                 err = -EINVAL;
3404                 goto out;
3405         }
3406
3407         new_channels.params = priv->channels.params;
3408         new_channels.params.num_tc = tc ? tc : 1;
3409
3410         if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
3411                 struct mlx5e_params old_params;
3412
3413                 old_params = priv->channels.params;
3414                 priv->channels.params = new_channels.params;
3415                 err = mlx5e_num_channels_changed(priv);
3416                 if (err)
3417                         priv->channels.params = old_params;
3418
3419                 goto out;
3420         }
3421
3422         err = mlx5e_safe_switch_channels(priv, &new_channels,
3423                                          mlx5e_num_channels_changed_ctx, NULL);
3424
3425 out:
3426         priv->max_opened_tc = max_t(u8, priv->max_opened_tc,
3427                                     priv->channels.params.num_tc);
3428         mutex_unlock(&priv->state_lock);
3429         return err;
3430 }
3431
3432 static int mlx5e_setup_tc_htb(struct mlx5e_priv *priv, struct tc_htb_qopt_offload *htb)
3433 {
3434         int res;
3435
3436         switch (htb->command) {
3437         case TC_HTB_CREATE:
3438                 return mlx5e_htb_root_add(priv, htb->parent_classid, htb->classid,
3439                                           htb->extack);
3440         case TC_HTB_DESTROY:
3441                 return mlx5e_htb_root_del(priv);
3442         case TC_HTB_LEAF_ALLOC_QUEUE:
3443                 res = mlx5e_htb_leaf_alloc_queue(priv, htb->classid, htb->parent_classid,
3444                                                  htb->rate, htb->ceil, htb->extack);
3445                 if (res < 0)
3446                         return res;
3447                 htb->qid = res;
3448                 return 0;
3449         case TC_HTB_LEAF_TO_INNER:
3450                 return mlx5e_htb_leaf_to_inner(priv, htb->parent_classid, htb->classid,
3451                                                htb->rate, htb->ceil, htb->extack);
3452         case TC_HTB_LEAF_DEL:
3453                 return mlx5e_htb_leaf_del(priv, htb->classid, &htb->moved_qid, &htb->qid,
3454                                           htb->extack);
3455         case TC_HTB_LEAF_DEL_LAST:
3456         case TC_HTB_LEAF_DEL_LAST_FORCE:
3457                 return mlx5e_htb_leaf_del_last(priv, htb->classid,
3458                                                htb->command == TC_HTB_LEAF_DEL_LAST_FORCE,
3459                                                htb->extack);
3460         case TC_HTB_NODE_MODIFY:
3461                 return mlx5e_htb_node_modify(priv, htb->classid, htb->rate, htb->ceil,
3462                                              htb->extack);
3463         case TC_HTB_LEAF_QUERY_QUEUE:
3464                 res = mlx5e_get_txq_by_classid(priv, htb->classid);
3465                 if (res < 0)
3466                         return res;
3467                 htb->qid = res;
3468                 return 0;
3469         default:
3470                 return -EOPNOTSUPP;
3471         }
3472 }
3473
3474 static LIST_HEAD(mlx5e_block_cb_list);
3475
3476 static int mlx5e_setup_tc(struct net_device *dev, enum tc_setup_type type,
3477                           void *type_data)
3478 {
3479         struct mlx5e_priv *priv = netdev_priv(dev);
3480         bool tc_unbind = false;
3481         int err;
3482
3483         if (type == TC_SETUP_BLOCK &&
3484             ((struct flow_block_offload *)type_data)->command == FLOW_BLOCK_UNBIND)
3485                 tc_unbind = true;
3486
3487         if (!netif_device_present(dev) && !tc_unbind)
3488                 return -ENODEV;
3489
3490         switch (type) {
3491         case TC_SETUP_BLOCK: {
3492                 struct flow_block_offload *f = type_data;
3493
3494                 f->unlocked_driver_cb = true;
3495                 return flow_block_cb_setup_simple(type_data,
3496                                                   &mlx5e_block_cb_list,
3497                                                   mlx5e_setup_tc_block_cb,
3498                                                   priv, priv, true);
3499         }
3500         case TC_SETUP_QDISC_MQPRIO:
3501                 return mlx5e_setup_tc_mqprio(priv, type_data);
3502         case TC_SETUP_QDISC_HTB:
3503                 mutex_lock(&priv->state_lock);
3504                 err = mlx5e_setup_tc_htb(priv, type_data);
3505                 mutex_unlock(&priv->state_lock);
3506                 return err;
3507         default:
3508                 return -EOPNOTSUPP;
3509         }
3510 }
3511
3512 void mlx5e_fold_sw_stats64(struct mlx5e_priv *priv, struct rtnl_link_stats64 *s)
3513 {
3514         int i;
3515
3516         for (i = 0; i < priv->max_nch; i++) {
3517                 struct mlx5e_channel_stats *channel_stats = &priv->channel_stats[i];
3518                 struct mlx5e_rq_stats *xskrq_stats = &channel_stats->xskrq;
3519                 struct mlx5e_rq_stats *rq_stats = &channel_stats->rq;
3520                 int j;
3521
3522                 s->rx_packets   += rq_stats->packets + xskrq_stats->packets;
3523                 s->rx_bytes     += rq_stats->bytes + xskrq_stats->bytes;
3524                 s->multicast    += rq_stats->mcast_packets + xskrq_stats->mcast_packets;
3525
3526                 for (j = 0; j < priv->max_opened_tc; j++) {
3527                         struct mlx5e_sq_stats *sq_stats = &channel_stats->sq[j];
3528
3529                         s->tx_packets    += sq_stats->packets;
3530                         s->tx_bytes      += sq_stats->bytes;
3531                         s->tx_dropped    += sq_stats->dropped;
3532                 }
3533         }
3534         if (priv->tx_ptp_opened) {
3535                 for (i = 0; i < priv->max_opened_tc; i++) {
3536                         struct mlx5e_sq_stats *sq_stats = &priv->ptp_stats.sq[i];
3537
3538                         s->tx_packets    += sq_stats->packets;
3539                         s->tx_bytes      += sq_stats->bytes;
3540                         s->tx_dropped    += sq_stats->dropped;
3541                 }
3542         }
3543         if (priv->rx_ptp_opened) {
3544                 struct mlx5e_rq_stats *rq_stats = &priv->ptp_stats.rq;
3545
3546                 s->rx_packets   += rq_stats->packets;
3547                 s->rx_bytes     += rq_stats->bytes;
3548                 s->multicast    += rq_stats->mcast_packets;
3549         }
3550 }
3551
3552 void
3553 mlx5e_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
3554 {
3555         struct mlx5e_priv *priv = netdev_priv(dev);
3556         struct mlx5e_pport_stats *pstats = &priv->stats.pport;
3557
3558         if (!netif_device_present(dev))
3559                 return;
3560
3561         /* In switchdev mode, monitor counters doesn't monitor
3562          * rx/tx stats of 802_3. The update stats mechanism
3563          * should keep the 802_3 layout counters updated
3564          */
3565         if (!mlx5e_monitor_counter_supported(priv) ||
3566             mlx5e_is_uplink_rep(priv)) {
3567                 /* update HW stats in background for next time */
3568                 mlx5e_queue_update_stats(priv);
3569         }
3570
3571         if (mlx5e_is_uplink_rep(priv)) {
3572                 struct mlx5e_vport_stats *vstats = &priv->stats.vport;
3573
3574                 stats->rx_packets = PPORT_802_3_GET(pstats, a_frames_received_ok);
3575                 stats->rx_bytes   = PPORT_802_3_GET(pstats, a_octets_received_ok);
3576                 stats->tx_packets = PPORT_802_3_GET(pstats, a_frames_transmitted_ok);
3577                 stats->tx_bytes   = PPORT_802_3_GET(pstats, a_octets_transmitted_ok);
3578
3579                 /* vport multicast also counts packets that are dropped due to steering
3580                  * or rx out of buffer
3581                  */
3582                 stats->multicast = VPORT_COUNTER_GET(vstats, received_eth_multicast.packets);
3583         } else {
3584                 mlx5e_fold_sw_stats64(priv, stats);
3585         }
3586
3587         stats->rx_dropped = priv->stats.qcnt.rx_out_of_buffer;
3588
3589         stats->rx_length_errors =
3590                 PPORT_802_3_GET(pstats, a_in_range_length_errors) +
3591                 PPORT_802_3_GET(pstats, a_out_of_range_length_field) +
3592                 PPORT_802_3_GET(pstats, a_frame_too_long_errors);
3593         stats->rx_crc_errors =
3594                 PPORT_802_3_GET(pstats, a_frame_check_sequence_errors);
3595         stats->rx_frame_errors = PPORT_802_3_GET(pstats, a_alignment_errors);
3596         stats->tx_aborted_errors = PPORT_2863_GET(pstats, if_out_discards);
3597         stats->rx_errors = stats->rx_length_errors + stats->rx_crc_errors +
3598                            stats->rx_frame_errors;
3599         stats->tx_errors = stats->tx_aborted_errors + stats->tx_carrier_errors;
3600 }
3601
3602 static void mlx5e_nic_set_rx_mode(struct mlx5e_priv *priv)
3603 {
3604         if (mlx5e_is_uplink_rep(priv))
3605                 return; /* no rx mode for uplink rep */
3606
3607         queue_work(priv->wq, &priv->set_rx_mode_work);
3608 }
3609
3610 static void mlx5e_set_rx_mode(struct net_device *dev)
3611 {
3612         struct mlx5e_priv *priv = netdev_priv(dev);
3613
3614         mlx5e_nic_set_rx_mode(priv);
3615 }
3616
3617 static int mlx5e_set_mac(struct net_device *netdev, void *addr)
3618 {
3619         struct mlx5e_priv *priv = netdev_priv(netdev);
3620         struct sockaddr *saddr = addr;
3621
3622         if (!is_valid_ether_addr(saddr->sa_data))
3623                 return -EADDRNOTAVAIL;
3624
3625         netif_addr_lock_bh(netdev);
3626         ether_addr_copy(netdev->dev_addr, saddr->sa_data);
3627         netif_addr_unlock_bh(netdev);
3628
3629         mlx5e_nic_set_rx_mode(priv);
3630
3631         return 0;
3632 }
3633
3634 #define MLX5E_SET_FEATURE(features, feature, enable)    \
3635         do {                                            \
3636                 if (enable)                             \
3637                         *features |= feature;           \
3638                 else                                    \
3639                         *features &= ~feature;          \
3640         } while (0)
3641
3642 typedef int (*mlx5e_feature_handler)(struct net_device *netdev, bool enable);
3643
3644 static int set_feature_lro(struct net_device *netdev, bool enable)
3645 {
3646         struct mlx5e_priv *priv = netdev_priv(netdev);
3647         struct mlx5_core_dev *mdev = priv->mdev;
3648         struct mlx5e_channels new_channels = {};
3649         struct mlx5e_params *cur_params;
3650         int err = 0;
3651         bool reset;
3652
3653         mutex_lock(&priv->state_lock);
3654
3655         if (enable && priv->xsk.refcnt) {
3656                 netdev_warn(netdev, "LRO is incompatible with AF_XDP (%u XSKs are active)\n",
3657                             priv->xsk.refcnt);
3658                 err = -EINVAL;
3659                 goto out;
3660         }
3661
3662         cur_params = &priv->channels.params;
3663         if (enable && !MLX5E_GET_PFLAG(cur_params, MLX5E_PFLAG_RX_STRIDING_RQ)) {
3664                 netdev_warn(netdev, "can't set LRO with legacy RQ\n");
3665                 err = -EINVAL;
3666                 goto out;
3667         }
3668
3669         reset = test_bit(MLX5E_STATE_OPENED, &priv->state);
3670
3671         new_channels.params = *cur_params;
3672         new_channels.params.lro_en = enable;
3673
3674         if (cur_params->rq_wq_type != MLX5_WQ_TYPE_CYCLIC) {
3675                 if (mlx5e_rx_mpwqe_is_linear_skb(mdev, cur_params, NULL) ==
3676                     mlx5e_rx_mpwqe_is_linear_skb(mdev, &new_channels.params, NULL))
3677                         reset = false;
3678         }
3679
3680         if (!reset) {
3681                 struct mlx5e_params old_params;
3682
3683                 old_params = *cur_params;
3684                 *cur_params = new_channels.params;
3685                 err = mlx5e_modify_tirs_lro(priv);
3686                 if (err)
3687                         *cur_params = old_params;
3688                 goto out;
3689         }
3690
3691         err = mlx5e_safe_switch_channels(priv, &new_channels,
3692                                          mlx5e_modify_tirs_lro_ctx, NULL);
3693 out:
3694         mutex_unlock(&priv->state_lock);
3695         return err;
3696 }
3697
3698 static int set_feature_cvlan_filter(struct net_device *netdev, bool enable)
3699 {
3700         struct mlx5e_priv *priv = netdev_priv(netdev);
3701
3702         if (enable)
3703                 mlx5e_enable_cvlan_filter(priv);
3704         else
3705                 mlx5e_disable_cvlan_filter(priv);
3706
3707         return 0;
3708 }
3709
3710 static int set_feature_hw_tc(struct net_device *netdev, bool enable)
3711 {
3712         struct mlx5e_priv *priv = netdev_priv(netdev);
3713
3714 #if IS_ENABLED(CONFIG_MLX5_CLS_ACT)
3715         if (!enable && mlx5e_tc_num_filters(priv, MLX5_TC_FLAG(NIC_OFFLOAD))) {
3716                 netdev_err(netdev,
3717                            "Active offloaded tc filters, can't turn hw_tc_offload off\n");
3718                 return -EINVAL;
3719         }
3720 #endif
3721
3722         if (!enable && priv->htb.maj_id) {
3723                 netdev_err(netdev, "Active HTB offload, can't turn hw_tc_offload off\n");
3724                 return -EINVAL;
3725         }
3726
3727         return 0;
3728 }
3729
3730 static int set_feature_rx_all(struct net_device *netdev, bool enable)
3731 {
3732         struct mlx5e_priv *priv = netdev_priv(netdev);
3733         struct mlx5_core_dev *mdev = priv->mdev;
3734
3735         return mlx5_set_port_fcs(mdev, !enable);
3736 }
3737
3738 static int set_feature_rx_fcs(struct net_device *netdev, bool enable)
3739 {
3740         struct mlx5e_priv *priv = netdev_priv(netdev);
3741         int err;
3742
3743         mutex_lock(&priv->state_lock);
3744
3745         priv->channels.params.scatter_fcs_en = enable;
3746         err = mlx5e_modify_channels_scatter_fcs(&priv->channels, enable);
3747         if (err)
3748                 priv->channels.params.scatter_fcs_en = !enable;
3749
3750         mutex_unlock(&priv->state_lock);
3751
3752         return err;
3753 }
3754
3755 static int set_feature_rx_vlan(struct net_device *netdev, bool enable)
3756 {
3757         struct mlx5e_priv *priv = netdev_priv(netdev);
3758         int err = 0;
3759
3760         mutex_lock(&priv->state_lock);
3761
3762         priv->channels.params.vlan_strip_disable = !enable;
3763         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
3764                 goto unlock;
3765
3766         err = mlx5e_modify_channels_vsd(&priv->channels, !enable);
3767         if (err)
3768                 priv->channels.params.vlan_strip_disable = enable;
3769
3770 unlock:
3771         mutex_unlock(&priv->state_lock);
3772
3773         return err;
3774 }
3775
3776 #ifdef CONFIG_MLX5_EN_ARFS
3777 static int set_feature_arfs(struct net_device *netdev, bool enable)
3778 {
3779         struct mlx5e_priv *priv = netdev_priv(netdev);
3780         int err;
3781
3782         if (enable)
3783                 err = mlx5e_arfs_enable(priv);
3784         else
3785                 err = mlx5e_arfs_disable(priv);
3786
3787         return err;
3788 }
3789 #endif
3790
3791 static int mlx5e_handle_feature(struct net_device *netdev,
3792                                 netdev_features_t *features,
3793                                 netdev_features_t wanted_features,
3794                                 netdev_features_t feature,
3795                                 mlx5e_feature_handler feature_handler)
3796 {
3797         netdev_features_t changes = wanted_features ^ netdev->features;
3798         bool enable = !!(wanted_features & feature);
3799         int err;
3800
3801         if (!(changes & feature))
3802                 return 0;
3803
3804         err = feature_handler(netdev, enable);
3805         if (err) {
3806                 netdev_err(netdev, "%s feature %pNF failed, err %d\n",
3807                            enable ? "Enable" : "Disable", &feature, err);
3808                 return err;
3809         }
3810
3811         MLX5E_SET_FEATURE(features, feature, enable);
3812         return 0;
3813 }
3814
3815 int mlx5e_set_features(struct net_device *netdev, netdev_features_t features)
3816 {
3817         netdev_features_t oper_features = netdev->features;
3818         int err = 0;
3819
3820 #define MLX5E_HANDLE_FEATURE(feature, handler) \
3821         mlx5e_handle_feature(netdev, &oper_features, features, feature, handler)
3822
3823         err |= MLX5E_HANDLE_FEATURE(NETIF_F_LRO, set_feature_lro);
3824         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_VLAN_CTAG_FILTER,
3825                                     set_feature_cvlan_filter);
3826         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_TC, set_feature_hw_tc);
3827         err |= MLX5E_HANDLE_FEATURE(NETIF_F_RXALL, set_feature_rx_all);
3828         err |= MLX5E_HANDLE_FEATURE(NETIF_F_RXFCS, set_feature_rx_fcs);
3829         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_VLAN_CTAG_RX, set_feature_rx_vlan);
3830 #ifdef CONFIG_MLX5_EN_ARFS
3831         err |= MLX5E_HANDLE_FEATURE(NETIF_F_NTUPLE, set_feature_arfs);
3832 #endif
3833         err |= MLX5E_HANDLE_FEATURE(NETIF_F_HW_TLS_RX, mlx5e_ktls_set_feature_rx);
3834
3835         if (err) {
3836                 netdev->features = oper_features;
3837                 return -EINVAL;
3838         }
3839
3840         return 0;
3841 }
3842
3843 static netdev_features_t mlx5e_fix_features(struct net_device *netdev,
3844                                             netdev_features_t features)
3845 {
3846         struct mlx5e_priv *priv = netdev_priv(netdev);
3847         struct mlx5e_params *params;
3848
3849         mutex_lock(&priv->state_lock);
3850         params = &priv->channels.params;
3851         if (!priv->fs.vlan ||
3852             !bitmap_empty(mlx5e_vlan_get_active_svlans(priv->fs.vlan), VLAN_N_VID)) {
3853                 /* HW strips the outer C-tag header, this is a problem
3854                  * for S-tag traffic.
3855                  */
3856                 features &= ~NETIF_F_HW_VLAN_CTAG_RX;
3857                 if (!params->vlan_strip_disable)
3858                         netdev_warn(netdev, "Dropping C-tag vlan stripping offload due to S-tag vlan\n");
3859         }
3860
3861         if (!MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ)) {
3862                 if (features & NETIF_F_LRO) {
3863                         netdev_warn(netdev, "Disabling LRO, not supported in legacy RQ\n");
3864                         features &= ~NETIF_F_LRO;
3865                 }
3866         }
3867
3868         if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
3869                 features &= ~NETIF_F_RXHASH;
3870                 if (netdev->features & NETIF_F_RXHASH)
3871                         netdev_warn(netdev, "Disabling rxhash, not supported when CQE compress is active\n");
3872         }
3873
3874         mutex_unlock(&priv->state_lock);
3875
3876         return features;
3877 }
3878
3879 static bool mlx5e_xsk_validate_mtu(struct net_device *netdev,
3880                                    struct mlx5e_channels *chs,
3881                                    struct mlx5e_params *new_params,
3882                                    struct mlx5_core_dev *mdev)
3883 {
3884         u16 ix;
3885
3886         for (ix = 0; ix < chs->params.num_channels; ix++) {
3887                 struct xsk_buff_pool *xsk_pool =
3888                         mlx5e_xsk_get_pool(&chs->params, chs->params.xsk, ix);
3889                 struct mlx5e_xsk_param xsk;
3890
3891                 if (!xsk_pool)
3892                         continue;
3893
3894                 mlx5e_build_xsk_param(xsk_pool, &xsk);
3895
3896                 if (!mlx5e_validate_xsk_param(new_params, &xsk, mdev)) {
3897                         u32 hr = mlx5e_get_linear_rq_headroom(new_params, &xsk);
3898                         int max_mtu_frame, max_mtu_page, max_mtu;
3899
3900                         /* Two criteria must be met:
3901                          * 1. HW MTU + all headrooms <= XSK frame size.
3902                          * 2. Size of SKBs allocated on XDP_PASS <= PAGE_SIZE.
3903                          */
3904                         max_mtu_frame = MLX5E_HW2SW_MTU(new_params, xsk.chunk_size - hr);
3905                         max_mtu_page = mlx5e_xdp_max_mtu(new_params, &xsk);
3906                         max_mtu = min(max_mtu_frame, max_mtu_page);
3907
3908                         netdev_err(netdev, "MTU %d is too big for an XSK running on channel %u. Try MTU <= %d\n",
3909                                    new_params->sw_mtu, ix, max_mtu);
3910                         return false;
3911                 }
3912         }
3913
3914         return true;
3915 }
3916
3917 int mlx5e_change_mtu(struct net_device *netdev, int new_mtu,
3918                      mlx5e_fp_preactivate preactivate)
3919 {
3920         struct mlx5e_priv *priv = netdev_priv(netdev);
3921         struct mlx5e_channels new_channels = {};
3922         struct mlx5e_params *params;
3923         int err = 0;
3924         bool reset;
3925
3926         mutex_lock(&priv->state_lock);
3927
3928         params = &priv->channels.params;
3929
3930         reset = !params->lro_en;
3931         reset = reset && test_bit(MLX5E_STATE_OPENED, &priv->state);
3932
3933         new_channels.params = *params;
3934         new_channels.params.sw_mtu = new_mtu;
3935         err = mlx5e_validate_params(priv->mdev, &new_channels.params);
3936         if (err)
3937                 goto out;
3938
3939         if (params->xdp_prog &&
3940             !mlx5e_rx_is_linear_skb(&new_channels.params, NULL)) {
3941                 netdev_err(netdev, "MTU(%d) > %d is not allowed while XDP enabled\n",
3942                            new_mtu, mlx5e_xdp_max_mtu(params, NULL));
3943                 err = -EINVAL;
3944                 goto out;
3945         }
3946
3947         if (priv->xsk.refcnt &&
3948             !mlx5e_xsk_validate_mtu(netdev, &priv->channels,
3949                                     &new_channels.params, priv->mdev)) {
3950                 err = -EINVAL;
3951                 goto out;
3952         }
3953
3954         if (params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
3955                 bool is_linear = mlx5e_rx_mpwqe_is_linear_skb(priv->mdev,
3956                                                               &new_channels.params,
3957                                                               NULL);
3958                 u8 ppw_old = mlx5e_mpwqe_log_pkts_per_wqe(params, NULL);
3959                 u8 ppw_new = mlx5e_mpwqe_log_pkts_per_wqe(&new_channels.params, NULL);
3960
3961                 /* If XSK is active, XSK RQs are linear. */
3962                 is_linear |= priv->xsk.refcnt;
3963
3964                 /* Always reset in linear mode - hw_mtu is used in data path. */
3965                 reset = reset && (is_linear || (ppw_old != ppw_new));
3966         }
3967
3968         if (!reset) {
3969                 unsigned int old_mtu = params->sw_mtu;
3970
3971                 params->sw_mtu = new_mtu;
3972                 if (preactivate) {
3973                         err = preactivate(priv, NULL);
3974                         if (err) {
3975                                 params->sw_mtu = old_mtu;
3976                                 goto out;
3977                         }
3978                 }
3979                 netdev->mtu = params->sw_mtu;
3980                 goto out;
3981         }
3982
3983         err = mlx5e_safe_switch_channels(priv, &new_channels, preactivate, NULL);
3984         if (err)
3985                 goto out;
3986
3987         netdev->mtu = new_channels.params.sw_mtu;
3988
3989 out:
3990         mutex_unlock(&priv->state_lock);
3991         return err;
3992 }
3993
3994 static int mlx5e_change_nic_mtu(struct net_device *netdev, int new_mtu)
3995 {
3996         return mlx5e_change_mtu(netdev, new_mtu, mlx5e_set_dev_port_mtu_ctx);
3997 }
3998
3999 int mlx5e_ptp_rx_manage_fs_ctx(struct mlx5e_priv *priv, void *ctx)
4000 {
4001         bool set  = *(bool *)ctx;
4002
4003         return mlx5e_ptp_rx_manage_fs(priv, set);
4004 }
4005
4006 int mlx5e_hwstamp_set(struct mlx5e_priv *priv, struct ifreq *ifr)
4007 {
4008         struct mlx5e_channels new_channels = {};
4009         struct hwtstamp_config config;
4010         bool rx_cqe_compress_def;
4011         int err;
4012
4013         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz) ||
4014             (mlx5_clock_get_ptp_index(priv->mdev) == -1))
4015                 return -EOPNOTSUPP;
4016
4017         if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
4018                 return -EFAULT;
4019
4020         /* TX HW timestamp */
4021         switch (config.tx_type) {
4022         case HWTSTAMP_TX_OFF:
4023         case HWTSTAMP_TX_ON:
4024                 break;
4025         default:
4026                 return -ERANGE;
4027         }
4028
4029         mutex_lock(&priv->state_lock);
4030         new_channels.params = priv->channels.params;
4031         rx_cqe_compress_def = priv->channels.params.rx_cqe_compress_def;
4032
4033         /* RX HW timestamp */
4034         switch (config.rx_filter) {
4035         case HWTSTAMP_FILTER_NONE:
4036                 new_channels.params.ptp_rx = false;
4037                 break;
4038         case HWTSTAMP_FILTER_ALL:
4039         case HWTSTAMP_FILTER_SOME:
4040         case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
4041         case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
4042         case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
4043         case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
4044         case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
4045         case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
4046         case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
4047         case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
4048         case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
4049         case HWTSTAMP_FILTER_PTP_V2_EVENT:
4050         case HWTSTAMP_FILTER_PTP_V2_SYNC:
4051         case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
4052         case HWTSTAMP_FILTER_NTP_ALL:
4053                 new_channels.params.ptp_rx = rx_cqe_compress_def;
4054                 config.rx_filter = HWTSTAMP_FILTER_ALL;
4055                 break;
4056         default:
4057                 mutex_unlock(&priv->state_lock);
4058                 return -ERANGE;
4059         }
4060
4061         if (new_channels.params.ptp_rx == priv->channels.params.ptp_rx)
4062                 goto out;
4063
4064         if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
4065                 priv->channels.params = new_channels.params;
4066                 goto out;
4067         }
4068         err = mlx5e_safe_switch_channels(priv, &new_channels, mlx5e_ptp_rx_manage_fs_ctx,
4069                                          &new_channels.params.ptp_rx);
4070         if (err) {
4071                 mutex_unlock(&priv->state_lock);
4072                 return err;
4073         }
4074 out:
4075         memcpy(&priv->tstamp, &config, sizeof(config));
4076         mutex_unlock(&priv->state_lock);
4077
4078         /* might need to fix some features */
4079         netdev_update_features(priv->netdev);
4080
4081         return copy_to_user(ifr->ifr_data, &config,
4082                             sizeof(config)) ? -EFAULT : 0;
4083 }
4084
4085 int mlx5e_hwstamp_get(struct mlx5e_priv *priv, struct ifreq *ifr)
4086 {
4087         struct hwtstamp_config *cfg = &priv->tstamp;
4088
4089         if (!MLX5_CAP_GEN(priv->mdev, device_frequency_khz))
4090                 return -EOPNOTSUPP;
4091
4092         return copy_to_user(ifr->ifr_data, cfg, sizeof(*cfg)) ? -EFAULT : 0;
4093 }
4094
4095 static int mlx5e_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
4096 {
4097         struct mlx5e_priv *priv = netdev_priv(dev);
4098
4099         switch (cmd) {
4100         case SIOCSHWTSTAMP:
4101                 return mlx5e_hwstamp_set(priv, ifr);
4102         case SIOCGHWTSTAMP:
4103                 return mlx5e_hwstamp_get(priv, ifr);
4104         default:
4105                 return -EOPNOTSUPP;
4106         }
4107 }
4108
4109 #ifdef CONFIG_MLX5_ESWITCH
4110 int mlx5e_set_vf_mac(struct net_device *dev, int vf, u8 *mac)
4111 {
4112         struct mlx5e_priv *priv = netdev_priv(dev);
4113         struct mlx5_core_dev *mdev = priv->mdev;
4114
4115         return mlx5_eswitch_set_vport_mac(mdev->priv.eswitch, vf + 1, mac);
4116 }
4117
4118 static int mlx5e_set_vf_vlan(struct net_device *dev, int vf, u16 vlan, u8 qos,
4119                              __be16 vlan_proto)
4120 {
4121         struct mlx5e_priv *priv = netdev_priv(dev);
4122         struct mlx5_core_dev *mdev = priv->mdev;
4123
4124         if (vlan_proto != htons(ETH_P_8021Q))
4125                 return -EPROTONOSUPPORT;
4126
4127         return mlx5_eswitch_set_vport_vlan(mdev->priv.eswitch, vf + 1,
4128                                            vlan, qos);
4129 }
4130
4131 static int mlx5e_set_vf_spoofchk(struct net_device *dev, int vf, bool setting)
4132 {
4133         struct mlx5e_priv *priv = netdev_priv(dev);
4134         struct mlx5_core_dev *mdev = priv->mdev;
4135
4136         return mlx5_eswitch_set_vport_spoofchk(mdev->priv.eswitch, vf + 1, setting);
4137 }
4138
4139 static int mlx5e_set_vf_trust(struct net_device *dev, int vf, bool setting)
4140 {
4141         struct mlx5e_priv *priv = netdev_priv(dev);
4142         struct mlx5_core_dev *mdev = priv->mdev;
4143
4144         return mlx5_eswitch_set_vport_trust(mdev->priv.eswitch, vf + 1, setting);
4145 }
4146
4147 int mlx5e_set_vf_rate(struct net_device *dev, int vf, int min_tx_rate,
4148                       int max_tx_rate)
4149 {
4150         struct mlx5e_priv *priv = netdev_priv(dev);
4151         struct mlx5_core_dev *mdev = priv->mdev;
4152
4153         return mlx5_eswitch_set_vport_rate(mdev->priv.eswitch, vf + 1,
4154                                            max_tx_rate, min_tx_rate);
4155 }
4156
4157 static int mlx5_vport_link2ifla(u8 esw_link)
4158 {
4159         switch (esw_link) {
4160         case MLX5_VPORT_ADMIN_STATE_DOWN:
4161                 return IFLA_VF_LINK_STATE_DISABLE;
4162         case MLX5_VPORT_ADMIN_STATE_UP:
4163                 return IFLA_VF_LINK_STATE_ENABLE;
4164         }
4165         return IFLA_VF_LINK_STATE_AUTO;
4166 }
4167
4168 static int mlx5_ifla_link2vport(u8 ifla_link)
4169 {
4170         switch (ifla_link) {
4171         case IFLA_VF_LINK_STATE_DISABLE:
4172                 return MLX5_VPORT_ADMIN_STATE_DOWN;
4173         case IFLA_VF_LINK_STATE_ENABLE:
4174                 return MLX5_VPORT_ADMIN_STATE_UP;
4175         }
4176         return MLX5_VPORT_ADMIN_STATE_AUTO;
4177 }
4178
4179 static int mlx5e_set_vf_link_state(struct net_device *dev, int vf,
4180                                    int link_state)
4181 {
4182         struct mlx5e_priv *priv = netdev_priv(dev);
4183         struct mlx5_core_dev *mdev = priv->mdev;
4184
4185         if (mlx5e_is_uplink_rep(priv))
4186                 return -EOPNOTSUPP;
4187
4188         return mlx5_eswitch_set_vport_state(mdev->priv.eswitch, vf + 1,
4189                                             mlx5_ifla_link2vport(link_state));
4190 }
4191
4192 int mlx5e_get_vf_config(struct net_device *dev,
4193                         int vf, struct ifla_vf_info *ivi)
4194 {
4195         struct mlx5e_priv *priv = netdev_priv(dev);
4196         struct mlx5_core_dev *mdev = priv->mdev;
4197         int err;
4198
4199         if (!netif_device_present(dev))
4200                 return -EOPNOTSUPP;
4201
4202         err = mlx5_eswitch_get_vport_config(mdev->priv.eswitch, vf + 1, ivi);
4203         if (err)
4204                 return err;
4205         ivi->linkstate = mlx5_vport_link2ifla(ivi->linkstate);
4206         return 0;
4207 }
4208
4209 int mlx5e_get_vf_stats(struct net_device *dev,
4210                        int vf, struct ifla_vf_stats *vf_stats)
4211 {
4212         struct mlx5e_priv *priv = netdev_priv(dev);
4213         struct mlx5_core_dev *mdev = priv->mdev;
4214
4215         return mlx5_eswitch_get_vport_stats(mdev->priv.eswitch, vf + 1,
4216                                             vf_stats);
4217 }
4218
4219 static bool
4220 mlx5e_has_offload_stats(const struct net_device *dev, int attr_id)
4221 {
4222         struct mlx5e_priv *priv = netdev_priv(dev);
4223
4224         if (!netif_device_present(dev))
4225                 return false;
4226
4227         if (!mlx5e_is_uplink_rep(priv))
4228                 return false;
4229
4230         return mlx5e_rep_has_offload_stats(dev, attr_id);
4231 }
4232
4233 static int
4234 mlx5e_get_offload_stats(int attr_id, const struct net_device *dev,
4235                         void *sp)
4236 {
4237         struct mlx5e_priv *priv = netdev_priv(dev);
4238
4239         if (!mlx5e_is_uplink_rep(priv))
4240                 return -EOPNOTSUPP;
4241
4242         return mlx5e_rep_get_offload_stats(attr_id, dev, sp);
4243 }
4244 #endif
4245
4246 static bool mlx5e_tunnel_proto_supported_tx(struct mlx5_core_dev *mdev, u8 proto_type)
4247 {
4248         switch (proto_type) {
4249         case IPPROTO_GRE:
4250                 return MLX5_CAP_ETH(mdev, tunnel_stateless_gre);
4251         case IPPROTO_IPIP:
4252         case IPPROTO_IPV6:
4253                 return (MLX5_CAP_ETH(mdev, tunnel_stateless_ip_over_ip) ||
4254                         MLX5_CAP_ETH(mdev, tunnel_stateless_ip_over_ip_tx));
4255         default:
4256                 return false;
4257         }
4258 }
4259
4260 static bool mlx5e_gre_tunnel_inner_proto_offload_supported(struct mlx5_core_dev *mdev,
4261                                                            struct sk_buff *skb)
4262 {
4263         switch (skb->inner_protocol) {
4264         case htons(ETH_P_IP):
4265         case htons(ETH_P_IPV6):
4266         case htons(ETH_P_TEB):
4267                 return true;
4268         case htons(ETH_P_MPLS_UC):
4269         case htons(ETH_P_MPLS_MC):
4270                 return MLX5_CAP_ETH(mdev, tunnel_stateless_mpls_over_gre);
4271         }
4272         return false;
4273 }
4274
4275 static netdev_features_t mlx5e_tunnel_features_check(struct mlx5e_priv *priv,
4276                                                      struct sk_buff *skb,
4277                                                      netdev_features_t features)
4278 {
4279         unsigned int offset = 0;
4280         struct udphdr *udph;
4281         u8 proto;
4282         u16 port;
4283
4284         switch (vlan_get_protocol(skb)) {
4285         case htons(ETH_P_IP):
4286                 proto = ip_hdr(skb)->protocol;
4287                 break;
4288         case htons(ETH_P_IPV6):
4289                 proto = ipv6_find_hdr(skb, &offset, -1, NULL, NULL);
4290                 break;
4291         default:
4292                 goto out;
4293         }
4294
4295         switch (proto) {
4296         case IPPROTO_GRE:
4297                 if (mlx5e_gre_tunnel_inner_proto_offload_supported(priv->mdev, skb))
4298                         return features;
4299                 break;
4300         case IPPROTO_IPIP:
4301         case IPPROTO_IPV6:
4302                 if (mlx5e_tunnel_proto_supported_tx(priv->mdev, IPPROTO_IPIP))
4303                         return features;
4304                 break;
4305         case IPPROTO_UDP:
4306                 udph = udp_hdr(skb);
4307                 port = be16_to_cpu(udph->dest);
4308
4309                 /* Verify if UDP port is being offloaded by HW */
4310                 if (mlx5_vxlan_lookup_port(priv->mdev->vxlan, port))
4311                         return features;
4312
4313 #if IS_ENABLED(CONFIG_GENEVE)
4314                 /* Support Geneve offload for default UDP port */
4315                 if (port == GENEVE_UDP_PORT && mlx5_geneve_tx_allowed(priv->mdev))
4316                         return features;
4317 #endif
4318         }
4319
4320 out:
4321         /* Disable CSUM and GSO if the udp dport is not offloaded by HW */
4322         return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
4323 }
4324
4325 netdev_features_t mlx5e_features_check(struct sk_buff *skb,
4326                                        struct net_device *netdev,
4327                                        netdev_features_t features)
4328 {
4329         struct mlx5e_priv *priv = netdev_priv(netdev);
4330
4331         features = vlan_features_check(skb, features);
4332         features = vxlan_features_check(skb, features);
4333
4334         if (mlx5e_ipsec_feature_check(skb, netdev, features))
4335                 return features;
4336
4337         /* Validate if the tunneled packet is being offloaded by HW */
4338         if (skb->encapsulation &&
4339             (features & NETIF_F_CSUM_MASK || features & NETIF_F_GSO_MASK))
4340                 return mlx5e_tunnel_features_check(priv, skb, features);
4341
4342         return features;
4343 }
4344
4345 static void mlx5e_tx_timeout_work(struct work_struct *work)
4346 {
4347         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
4348                                                tx_timeout_work);
4349         struct net_device *netdev = priv->netdev;
4350         int i;
4351
4352         rtnl_lock();
4353         mutex_lock(&priv->state_lock);
4354
4355         if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
4356                 goto unlock;
4357
4358         for (i = 0; i < netdev->real_num_tx_queues; i++) {
4359                 struct netdev_queue *dev_queue =
4360                         netdev_get_tx_queue(netdev, i);
4361                 struct mlx5e_txqsq *sq = priv->txq2sq[i];
4362
4363                 if (!netif_xmit_stopped(dev_queue))
4364                         continue;
4365
4366                 if (mlx5e_reporter_tx_timeout(sq))
4367                 /* break if tried to reopened channels */
4368                         break;
4369         }
4370
4371 unlock:
4372         mutex_unlock(&priv->state_lock);
4373         rtnl_unlock();
4374 }
4375
4376 static void mlx5e_tx_timeout(struct net_device *dev, unsigned int txqueue)
4377 {
4378         struct mlx5e_priv *priv = netdev_priv(dev);
4379
4380         netdev_err(dev, "TX timeout detected\n");
4381         queue_work(priv->wq, &priv->tx_timeout_work);
4382 }
4383
4384 static int mlx5e_xdp_allowed(struct mlx5e_priv *priv, struct bpf_prog *prog)
4385 {
4386         struct net_device *netdev = priv->netdev;
4387         struct mlx5e_channels new_channels = {};
4388
4389         if (priv->channels.params.lro_en) {
4390                 netdev_warn(netdev, "can't set XDP while LRO is on, disable LRO first\n");
4391                 return -EINVAL;
4392         }
4393
4394         if (mlx5_fpga_is_ipsec_device(priv->mdev)) {
4395                 netdev_warn(netdev,
4396                             "XDP is not available on Innova cards with IPsec support\n");
4397                 return -EINVAL;
4398         }
4399
4400         new_channels.params = priv->channels.params;
4401         new_channels.params.xdp_prog = prog;
4402
4403         /* No XSK params: AF_XDP can't be enabled yet at the point of setting
4404          * the XDP program.
4405          */
4406         if (!mlx5e_rx_is_linear_skb(&new_channels.params, NULL)) {
4407                 netdev_warn(netdev, "XDP is not allowed with MTU(%d) > %d\n",
4408                             new_channels.params.sw_mtu,
4409                             mlx5e_xdp_max_mtu(&new_channels.params, NULL));
4410                 return -EINVAL;
4411         }
4412
4413         return 0;
4414 }
4415
4416 static void mlx5e_rq_replace_xdp_prog(struct mlx5e_rq *rq, struct bpf_prog *prog)
4417 {
4418         struct bpf_prog *old_prog;
4419
4420         old_prog = rcu_replace_pointer(rq->xdp_prog, prog,
4421                                        lockdep_is_held(&rq->priv->state_lock));
4422         if (old_prog)
4423                 bpf_prog_put(old_prog);
4424 }
4425
4426 static int mlx5e_xdp_set(struct net_device *netdev, struct bpf_prog *prog)
4427 {
4428         struct mlx5e_priv *priv = netdev_priv(netdev);
4429         struct bpf_prog *old_prog;
4430         bool reset, was_opened;
4431         int err = 0;
4432         int i;
4433
4434         mutex_lock(&priv->state_lock);
4435
4436         if (prog) {
4437                 err = mlx5e_xdp_allowed(priv, prog);
4438                 if (err)
4439                         goto unlock;
4440         }
4441
4442         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
4443         /* no need for full reset when exchanging programs */
4444         reset = (!priv->channels.params.xdp_prog || !prog);
4445
4446         if (was_opened && !reset)
4447                 /* num_channels is invariant here, so we can take the
4448                  * batched reference right upfront.
4449                  */
4450                 bpf_prog_add(prog, priv->channels.num);
4451
4452         if (was_opened && reset) {
4453                 struct mlx5e_channels new_channels = {};
4454
4455                 new_channels.params = priv->channels.params;
4456                 new_channels.params.xdp_prog = prog;
4457                 mlx5e_set_rq_type(priv->mdev, &new_channels.params);
4458                 old_prog = priv->channels.params.xdp_prog;
4459
4460                 err = mlx5e_safe_switch_channels(priv, &new_channels, NULL, NULL);
4461                 if (err)
4462                         goto unlock;
4463         } else {
4464                 /* exchange programs, extra prog reference we got from caller
4465                  * as long as we don't fail from this point onwards.
4466                  */
4467                 old_prog = xchg(&priv->channels.params.xdp_prog, prog);
4468         }
4469
4470         if (old_prog)
4471                 bpf_prog_put(old_prog);
4472
4473         if (!was_opened && reset) /* change RQ type according to priv->xdp_prog */
4474                 mlx5e_set_rq_type(priv->mdev, &priv->channels.params);
4475
4476         if (!was_opened || reset)
4477                 goto unlock;
4478
4479         /* exchanging programs w/o reset, we update ref counts on behalf
4480          * of the channels RQs here.
4481          */
4482         for (i = 0; i < priv->channels.num; i++) {
4483                 struct mlx5e_channel *c = priv->channels.c[i];
4484
4485                 mlx5e_rq_replace_xdp_prog(&c->rq, prog);
4486                 if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state)) {
4487                         bpf_prog_inc(prog);
4488                         mlx5e_rq_replace_xdp_prog(&c->xskrq, prog);
4489                 }
4490         }
4491
4492 unlock:
4493         mutex_unlock(&priv->state_lock);
4494         return err;
4495 }
4496
4497 static int mlx5e_xdp(struct net_device *dev, struct netdev_bpf *xdp)
4498 {
4499         switch (xdp->command) {
4500         case XDP_SETUP_PROG:
4501                 return mlx5e_xdp_set(dev, xdp->prog);
4502         case XDP_SETUP_XSK_POOL:
4503                 return mlx5e_xsk_setup_pool(dev, xdp->xsk.pool,
4504                                             xdp->xsk.queue_id);
4505         default:
4506                 return -EINVAL;
4507         }
4508 }
4509
4510 #ifdef CONFIG_MLX5_ESWITCH
4511 static int mlx5e_bridge_getlink(struct sk_buff *skb, u32 pid, u32 seq,
4512                                 struct net_device *dev, u32 filter_mask,
4513                                 int nlflags)
4514 {
4515         struct mlx5e_priv *priv = netdev_priv(dev);
4516         struct mlx5_core_dev *mdev = priv->mdev;
4517         u8 mode, setting;
4518         int err;
4519
4520         err = mlx5_eswitch_get_vepa(mdev->priv.eswitch, &setting);
4521         if (err)
4522                 return err;
4523         mode = setting ? BRIDGE_MODE_VEPA : BRIDGE_MODE_VEB;
4524         return ndo_dflt_bridge_getlink(skb, pid, seq, dev,
4525                                        mode,
4526                                        0, 0, nlflags, filter_mask, NULL);
4527 }
4528
4529 static int mlx5e_bridge_setlink(struct net_device *dev, struct nlmsghdr *nlh,
4530                                 u16 flags, struct netlink_ext_ack *extack)
4531 {
4532         struct mlx5e_priv *priv = netdev_priv(dev);
4533         struct mlx5_core_dev *mdev = priv->mdev;
4534         struct nlattr *attr, *br_spec;
4535         u16 mode = BRIDGE_MODE_UNDEF;
4536         u8 setting;
4537         int rem;
4538
4539         br_spec = nlmsg_find_attr(nlh, sizeof(struct ifinfomsg), IFLA_AF_SPEC);
4540         if (!br_spec)
4541                 return -EINVAL;
4542
4543         nla_for_each_nested(attr, br_spec, rem) {
4544                 if (nla_type(attr) != IFLA_BRIDGE_MODE)
4545                         continue;
4546
4547                 if (nla_len(attr) < sizeof(mode))
4548                         return -EINVAL;
4549
4550                 mode = nla_get_u16(attr);
4551                 if (mode > BRIDGE_MODE_VEPA)
4552                         return -EINVAL;
4553
4554                 break;
4555         }
4556
4557         if (mode == BRIDGE_MODE_UNDEF)
4558                 return -EINVAL;
4559
4560         setting = (mode == BRIDGE_MODE_VEPA) ?  1 : 0;
4561         return mlx5_eswitch_set_vepa(mdev->priv.eswitch, setting);
4562 }
4563 #endif
4564
4565 const struct net_device_ops mlx5e_netdev_ops = {
4566         .ndo_open                = mlx5e_open,
4567         .ndo_stop                = mlx5e_close,
4568         .ndo_start_xmit          = mlx5e_xmit,
4569         .ndo_setup_tc            = mlx5e_setup_tc,
4570         .ndo_select_queue        = mlx5e_select_queue,
4571         .ndo_get_stats64         = mlx5e_get_stats,
4572         .ndo_set_rx_mode         = mlx5e_set_rx_mode,
4573         .ndo_set_mac_address     = mlx5e_set_mac,
4574         .ndo_vlan_rx_add_vid     = mlx5e_vlan_rx_add_vid,
4575         .ndo_vlan_rx_kill_vid    = mlx5e_vlan_rx_kill_vid,
4576         .ndo_set_features        = mlx5e_set_features,
4577         .ndo_fix_features        = mlx5e_fix_features,
4578         .ndo_change_mtu          = mlx5e_change_nic_mtu,
4579         .ndo_do_ioctl            = mlx5e_ioctl,
4580         .ndo_set_tx_maxrate      = mlx5e_set_tx_maxrate,
4581         .ndo_features_check      = mlx5e_features_check,
4582         .ndo_tx_timeout          = mlx5e_tx_timeout,
4583         .ndo_bpf                 = mlx5e_xdp,
4584         .ndo_xdp_xmit            = mlx5e_xdp_xmit,
4585         .ndo_xsk_wakeup          = mlx5e_xsk_wakeup,
4586 #ifdef CONFIG_MLX5_EN_ARFS
4587         .ndo_rx_flow_steer       = mlx5e_rx_flow_steer,
4588 #endif
4589 #ifdef CONFIG_MLX5_ESWITCH
4590         .ndo_bridge_setlink      = mlx5e_bridge_setlink,
4591         .ndo_bridge_getlink      = mlx5e_bridge_getlink,
4592
4593         /* SRIOV E-Switch NDOs */
4594         .ndo_set_vf_mac          = mlx5e_set_vf_mac,
4595         .ndo_set_vf_vlan         = mlx5e_set_vf_vlan,
4596         .ndo_set_vf_spoofchk     = mlx5e_set_vf_spoofchk,
4597         .ndo_set_vf_trust        = mlx5e_set_vf_trust,
4598         .ndo_set_vf_rate         = mlx5e_set_vf_rate,
4599         .ndo_get_vf_config       = mlx5e_get_vf_config,
4600         .ndo_set_vf_link_state   = mlx5e_set_vf_link_state,
4601         .ndo_get_vf_stats        = mlx5e_get_vf_stats,
4602         .ndo_has_offload_stats   = mlx5e_has_offload_stats,
4603         .ndo_get_offload_stats   = mlx5e_get_offload_stats,
4604 #endif
4605         .ndo_get_devlink_port    = mlx5e_get_devlink_port,
4606 };
4607
4608 void mlx5e_build_default_indir_rqt(u32 *indirection_rqt, int len,
4609                                    int num_channels)
4610 {
4611         int i;
4612
4613         for (i = 0; i < len; i++)
4614                 indirection_rqt[i] = i % num_channels;
4615 }
4616
4617 static u32 mlx5e_choose_lro_timeout(struct mlx5_core_dev *mdev, u32 wanted_timeout)
4618 {
4619         int i;
4620
4621         /* The supported periods are organized in ascending order */
4622         for (i = 0; i < MLX5E_LRO_TIMEOUT_ARR_SIZE - 1; i++)
4623                 if (MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]) >= wanted_timeout)
4624                         break;
4625
4626         return MLX5_CAP_ETH(mdev, lro_timer_supported_periods[i]);
4627 }
4628
4629 void mlx5e_build_rss_params(struct mlx5e_rss_params *rss_params,
4630                             u16 num_channels)
4631 {
4632         enum mlx5e_traffic_types tt;
4633
4634         rss_params->hfunc = ETH_RSS_HASH_TOP;
4635         netdev_rss_key_fill(rss_params->toeplitz_hash_key,
4636                             sizeof(rss_params->toeplitz_hash_key));
4637         mlx5e_build_default_indir_rqt(rss_params->indirection_rqt,
4638                                       MLX5E_INDIR_RQT_SIZE, num_channels);
4639         for (tt = 0; tt < MLX5E_NUM_INDIR_TIRS; tt++)
4640                 rss_params->rx_hash_fields[tt] =
4641                         tirc_default_config[tt].rx_hash_fields;
4642 }
4643
4644 void mlx5e_build_nic_params(struct mlx5e_priv *priv, struct mlx5e_xsk *xsk, u16 mtu)
4645 {
4646         struct mlx5e_rss_params *rss_params = &priv->rss_params;
4647         struct mlx5e_params *params = &priv->channels.params;
4648         struct mlx5_core_dev *mdev = priv->mdev;
4649         u8 rx_cq_period_mode;
4650
4651         priv->max_nch = mlx5e_calc_max_nch(priv, priv->profile);
4652
4653         params->sw_mtu = mtu;
4654         params->hard_mtu = MLX5E_ETH_HARD_MTU;
4655         params->num_channels = min_t(unsigned int, MLX5E_MAX_NUM_CHANNELS / 2,
4656                                      priv->max_nch);
4657         params->num_tc       = 1;
4658
4659         /* Set an initial non-zero value, so that mlx5e_select_queue won't
4660          * divide by zero if called before first activating channels.
4661          */
4662         priv->num_tc_x_num_ch = params->num_channels * params->num_tc;
4663
4664         /* SQ */
4665         params->log_sq_size = is_kdump_kernel() ?
4666                 MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE :
4667                 MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
4668         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_SKB_TX_MPWQE,
4669                         MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe));
4670
4671         /* XDP SQ */
4672         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_XDP_TX_MPWQE,
4673                         MLX5_CAP_ETH(mdev, enhanced_multi_pkt_send_wqe));
4674
4675         /* set CQE compression */
4676         params->rx_cqe_compress_def = false;
4677         if (MLX5_CAP_GEN(mdev, cqe_compression) &&
4678             MLX5_CAP_GEN(mdev, vport_group_manager))
4679                 params->rx_cqe_compress_def = slow_pci_heuristic(mdev);
4680
4681         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS, params->rx_cqe_compress_def);
4682         MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_NO_CSUM_COMPLETE, false);
4683
4684         /* RQ */
4685         mlx5e_build_rq_params(mdev, params);
4686
4687         /* HW LRO */
4688         if (MLX5_CAP_ETH(mdev, lro_cap) &&
4689             params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ) {
4690                 /* No XSK params: checking the availability of striding RQ in general. */
4691                 if (!mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL))
4692                         params->lro_en = !slow_pci_heuristic(mdev);
4693         }
4694         params->lro_timeout = mlx5e_choose_lro_timeout(mdev, MLX5E_DEFAULT_LRO_TIMEOUT);
4695
4696         /* CQ moderation params */
4697         rx_cq_period_mode = MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
4698                         MLX5_CQ_PERIOD_MODE_START_FROM_CQE :
4699                         MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
4700         params->rx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
4701         params->tx_dim_enabled = MLX5_CAP_GEN(mdev, cq_moderation);
4702         mlx5e_set_rx_cq_mode_params(params, rx_cq_period_mode);
4703         mlx5e_set_tx_cq_mode_params(params, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
4704
4705         /* TX inline */
4706         mlx5_query_min_inline(mdev, &params->tx_min_inline_mode);
4707
4708         /* RSS */
4709         mlx5e_build_rss_params(rss_params, params->num_channels);
4710         params->tunneled_offload_en =
4711                 mlx5e_tunnel_inner_ft_supported(mdev);
4712
4713         /* AF_XDP */
4714         params->xsk = xsk;
4715
4716         /* Do not update netdev->features directly in here
4717          * on mlx5e_attach_netdev() we will call mlx5e_update_features()
4718          * To update netdev->features please modify mlx5e_fix_features()
4719          */
4720 }
4721
4722 static void mlx5e_set_netdev_dev_addr(struct net_device *netdev)
4723 {
4724         struct mlx5e_priv *priv = netdev_priv(netdev);
4725
4726         mlx5_query_mac_address(priv->mdev, netdev->dev_addr);
4727         if (is_zero_ether_addr(netdev->dev_addr) &&
4728             !MLX5_CAP_GEN(priv->mdev, vport_group_manager)) {
4729                 eth_hw_addr_random(netdev);
4730                 mlx5_core_info(priv->mdev, "Assigned random MAC address %pM\n", netdev->dev_addr);
4731         }
4732 }
4733
4734 static int mlx5e_vxlan_set_port(struct net_device *netdev, unsigned int table,
4735                                 unsigned int entry, struct udp_tunnel_info *ti)
4736 {
4737         struct mlx5e_priv *priv = netdev_priv(netdev);
4738
4739         return mlx5_vxlan_add_port(priv->mdev->vxlan, ntohs(ti->port));
4740 }
4741
4742 static int mlx5e_vxlan_unset_port(struct net_device *netdev, unsigned int table,
4743                                   unsigned int entry, struct udp_tunnel_info *ti)
4744 {
4745         struct mlx5e_priv *priv = netdev_priv(netdev);
4746
4747         return mlx5_vxlan_del_port(priv->mdev->vxlan, ntohs(ti->port));
4748 }
4749
4750 void mlx5e_vxlan_set_netdev_info(struct mlx5e_priv *priv)
4751 {
4752         if (!mlx5_vxlan_allowed(priv->mdev->vxlan))
4753                 return;
4754
4755         priv->nic_info.set_port = mlx5e_vxlan_set_port;
4756         priv->nic_info.unset_port = mlx5e_vxlan_unset_port;
4757         priv->nic_info.flags = UDP_TUNNEL_NIC_INFO_MAY_SLEEP |
4758                                 UDP_TUNNEL_NIC_INFO_STATIC_IANA_VXLAN;
4759         priv->nic_info.tables[0].tunnel_types = UDP_TUNNEL_TYPE_VXLAN;
4760         /* Don't count the space hard-coded to the IANA port */
4761         priv->nic_info.tables[0].n_entries =
4762                 mlx5_vxlan_max_udp_ports(priv->mdev) - 1;
4763
4764         priv->netdev->udp_tunnel_nic_info = &priv->nic_info;
4765 }
4766
4767 static bool mlx5e_tunnel_any_tx_proto_supported(struct mlx5_core_dev *mdev)
4768 {
4769         int tt;
4770
4771         for (tt = 0; tt < MLX5E_NUM_TUNNEL_TT; tt++) {
4772                 if (mlx5e_tunnel_proto_supported_tx(mdev, mlx5e_get_proto_by_tunnel_type(tt)))
4773                         return true;
4774         }
4775         return (mlx5_vxlan_allowed(mdev->vxlan) || mlx5_geneve_tx_allowed(mdev));
4776 }
4777
4778 static void mlx5e_build_nic_netdev(struct net_device *netdev)
4779 {
4780         struct mlx5e_priv *priv = netdev_priv(netdev);
4781         struct mlx5_core_dev *mdev = priv->mdev;
4782         bool fcs_supported;
4783         bool fcs_enabled;
4784
4785         SET_NETDEV_DEV(netdev, mdev->device);
4786
4787         netdev->netdev_ops = &mlx5e_netdev_ops;
4788
4789         mlx5e_dcbnl_build_netdev(netdev);
4790
4791         netdev->watchdog_timeo    = 15 * HZ;
4792
4793         netdev->ethtool_ops       = &mlx5e_ethtool_ops;
4794
4795         netdev->vlan_features    |= NETIF_F_SG;
4796         netdev->vlan_features    |= NETIF_F_HW_CSUM;
4797         netdev->vlan_features    |= NETIF_F_GRO;
4798         netdev->vlan_features    |= NETIF_F_TSO;
4799         netdev->vlan_features    |= NETIF_F_TSO6;
4800         netdev->vlan_features    |= NETIF_F_RXCSUM;
4801         netdev->vlan_features    |= NETIF_F_RXHASH;
4802
4803         netdev->mpls_features    |= NETIF_F_SG;
4804         netdev->mpls_features    |= NETIF_F_HW_CSUM;
4805         netdev->mpls_features    |= NETIF_F_TSO;
4806         netdev->mpls_features    |= NETIF_F_TSO6;
4807
4808         netdev->hw_enc_features  |= NETIF_F_HW_VLAN_CTAG_TX;
4809         netdev->hw_enc_features  |= NETIF_F_HW_VLAN_CTAG_RX;
4810
4811         if (!!MLX5_CAP_ETH(mdev, lro_cap) &&
4812             mlx5e_check_fragmented_striding_rq_cap(mdev))
4813                 netdev->vlan_features    |= NETIF_F_LRO;
4814
4815         netdev->hw_features       = netdev->vlan_features;
4816         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_TX;
4817         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_RX;
4818         netdev->hw_features      |= NETIF_F_HW_VLAN_CTAG_FILTER;
4819         netdev->hw_features      |= NETIF_F_HW_VLAN_STAG_TX;
4820
4821         if (mlx5e_tunnel_any_tx_proto_supported(mdev)) {
4822                 netdev->hw_enc_features |= NETIF_F_HW_CSUM;
4823                 netdev->hw_enc_features |= NETIF_F_TSO;
4824                 netdev->hw_enc_features |= NETIF_F_TSO6;
4825                 netdev->hw_enc_features |= NETIF_F_GSO_PARTIAL;
4826         }
4827
4828         if (mlx5_vxlan_allowed(mdev->vxlan) || mlx5_geneve_tx_allowed(mdev)) {
4829                 netdev->hw_features     |= NETIF_F_GSO_UDP_TUNNEL |
4830                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4831                 netdev->hw_enc_features |= NETIF_F_GSO_UDP_TUNNEL |
4832                                            NETIF_F_GSO_UDP_TUNNEL_CSUM;
4833                 netdev->gso_partial_features = NETIF_F_GSO_UDP_TUNNEL_CSUM;
4834                 netdev->vlan_features |= NETIF_F_GSO_UDP_TUNNEL |
4835                                          NETIF_F_GSO_UDP_TUNNEL_CSUM;
4836         }
4837
4838         if (mlx5e_tunnel_proto_supported_tx(mdev, IPPROTO_GRE)) {
4839                 netdev->hw_features     |= NETIF_F_GSO_GRE |
4840                                            NETIF_F_GSO_GRE_CSUM;
4841                 netdev->hw_enc_features |= NETIF_F_GSO_GRE |
4842                                            NETIF_F_GSO_GRE_CSUM;
4843                 netdev->gso_partial_features |= NETIF_F_GSO_GRE |
4844                                                 NETIF_F_GSO_GRE_CSUM;
4845         }
4846
4847         if (mlx5e_tunnel_proto_supported_tx(mdev, IPPROTO_IPIP)) {
4848                 netdev->hw_features |= NETIF_F_GSO_IPXIP4 |
4849                                        NETIF_F_GSO_IPXIP6;
4850                 netdev->hw_enc_features |= NETIF_F_GSO_IPXIP4 |
4851                                            NETIF_F_GSO_IPXIP6;
4852                 netdev->gso_partial_features |= NETIF_F_GSO_IPXIP4 |
4853                                                 NETIF_F_GSO_IPXIP6;
4854         }
4855
4856         netdev->hw_features                      |= NETIF_F_GSO_PARTIAL;
4857         netdev->gso_partial_features             |= NETIF_F_GSO_UDP_L4;
4858         netdev->hw_features                      |= NETIF_F_GSO_UDP_L4;
4859         netdev->features                         |= NETIF_F_GSO_UDP_L4;
4860
4861         mlx5_query_port_fcs(mdev, &fcs_supported, &fcs_enabled);
4862
4863         if (fcs_supported)
4864                 netdev->hw_features |= NETIF_F_RXALL;
4865
4866         if (MLX5_CAP_ETH(mdev, scatter_fcs))
4867                 netdev->hw_features |= NETIF_F_RXFCS;
4868
4869         netdev->features          = netdev->hw_features;
4870
4871         /* Defaults */
4872         if (fcs_enabled)
4873                 netdev->features  &= ~NETIF_F_RXALL;
4874         netdev->features  &= ~NETIF_F_LRO;
4875         netdev->features  &= ~NETIF_F_RXFCS;
4876
4877 #define FT_CAP(f) MLX5_CAP_FLOWTABLE(mdev, flow_table_properties_nic_receive.f)
4878         if (FT_CAP(flow_modify_en) &&
4879             FT_CAP(modify_root) &&
4880             FT_CAP(identified_miss_table_mode) &&
4881             FT_CAP(flow_table_modify)) {
4882 #if IS_ENABLED(CONFIG_MLX5_CLS_ACT)
4883                 netdev->hw_features      |= NETIF_F_HW_TC;
4884 #endif
4885 #ifdef CONFIG_MLX5_EN_ARFS
4886                 netdev->hw_features      |= NETIF_F_NTUPLE;
4887 #endif
4888         }
4889         if (mlx5_qos_is_supported(mdev))
4890                 netdev->features |= NETIF_F_HW_TC;
4891
4892         netdev->features         |= NETIF_F_HIGHDMA;
4893         netdev->features         |= NETIF_F_HW_VLAN_STAG_FILTER;
4894
4895         netdev->priv_flags       |= IFF_UNICAST_FLT;
4896
4897         mlx5e_set_netdev_dev_addr(netdev);
4898         mlx5e_ipsec_build_netdev(priv);
4899         mlx5e_tls_build_netdev(priv);
4900 }
4901
4902 void mlx5e_create_q_counters(struct mlx5e_priv *priv)
4903 {
4904         u32 out[MLX5_ST_SZ_DW(alloc_q_counter_out)] = {};
4905         u32 in[MLX5_ST_SZ_DW(alloc_q_counter_in)] = {};
4906         struct mlx5_core_dev *mdev = priv->mdev;
4907         int err;
4908
4909         MLX5_SET(alloc_q_counter_in, in, opcode, MLX5_CMD_OP_ALLOC_Q_COUNTER);
4910         err = mlx5_cmd_exec_inout(mdev, alloc_q_counter, in, out);
4911         if (!err)
4912                 priv->q_counter =
4913                         MLX5_GET(alloc_q_counter_out, out, counter_set_id);
4914
4915         err = mlx5_cmd_exec_inout(mdev, alloc_q_counter, in, out);
4916         if (!err)
4917                 priv->drop_rq_q_counter =
4918                         MLX5_GET(alloc_q_counter_out, out, counter_set_id);
4919 }
4920
4921 void mlx5e_destroy_q_counters(struct mlx5e_priv *priv)
4922 {
4923         u32 in[MLX5_ST_SZ_DW(dealloc_q_counter_in)] = {};
4924
4925         MLX5_SET(dealloc_q_counter_in, in, opcode,
4926                  MLX5_CMD_OP_DEALLOC_Q_COUNTER);
4927         if (priv->q_counter) {
4928                 MLX5_SET(dealloc_q_counter_in, in, counter_set_id,
4929                          priv->q_counter);
4930                 mlx5_cmd_exec_in(priv->mdev, dealloc_q_counter, in);
4931         }
4932
4933         if (priv->drop_rq_q_counter) {
4934                 MLX5_SET(dealloc_q_counter_in, in, counter_set_id,
4935                          priv->drop_rq_q_counter);
4936                 mlx5_cmd_exec_in(priv->mdev, dealloc_q_counter, in);
4937         }
4938 }
4939
4940 static int mlx5e_nic_init(struct mlx5_core_dev *mdev,
4941                           struct net_device *netdev)
4942 {
4943         struct mlx5e_priv *priv = netdev_priv(netdev);
4944         int err;
4945
4946         mlx5e_build_nic_params(priv, &priv->xsk, netdev->mtu);
4947         mlx5e_vxlan_set_netdev_info(priv);
4948
4949         mlx5e_timestamp_init(priv);
4950
4951         err = mlx5e_ipsec_init(priv);
4952         if (err)
4953                 mlx5_core_err(mdev, "IPSec initialization failed, %d\n", err);
4954
4955         err = mlx5e_tls_init(priv);
4956         if (err)
4957                 mlx5_core_err(mdev, "TLS initialization failed, %d\n", err);
4958
4959         mlx5e_health_create_reporters(priv);
4960
4961         return 0;
4962 }
4963
4964 static void mlx5e_nic_cleanup(struct mlx5e_priv *priv)
4965 {
4966         mlx5e_health_destroy_reporters(priv);
4967         mlx5e_tls_cleanup(priv);
4968         mlx5e_ipsec_cleanup(priv);
4969 }
4970
4971 static int mlx5e_init_nic_rx(struct mlx5e_priv *priv)
4972 {
4973         struct mlx5_core_dev *mdev = priv->mdev;
4974         u16 max_nch = priv->max_nch;
4975         int err;
4976
4977         mlx5e_create_q_counters(priv);
4978
4979         err = mlx5e_open_drop_rq(priv, &priv->drop_rq);
4980         if (err) {
4981                 mlx5_core_err(mdev, "open drop rq failed, %d\n", err);
4982                 goto err_destroy_q_counters;
4983         }
4984
4985         err = mlx5e_create_indirect_rqt(priv);
4986         if (err)
4987                 goto err_close_drop_rq;
4988
4989         err = mlx5e_create_direct_rqts(priv, priv->direct_tir, max_nch);
4990         if (err)
4991                 goto err_destroy_indirect_rqts;
4992
4993         err = mlx5e_create_indirect_tirs(priv, true);
4994         if (err)
4995                 goto err_destroy_direct_rqts;
4996
4997         err = mlx5e_create_direct_tirs(priv, priv->direct_tir, max_nch);
4998         if (err)
4999                 goto err_destroy_indirect_tirs;
5000
5001         err = mlx5e_create_direct_rqts(priv, priv->xsk_tir, max_nch);
5002         if (unlikely(err))
5003                 goto err_destroy_direct_tirs;
5004
5005         err = mlx5e_create_direct_tirs(priv, priv->xsk_tir, max_nch);
5006         if (unlikely(err))
5007                 goto err_destroy_xsk_rqts;
5008
5009         err = mlx5e_create_direct_rqts(priv, &priv->ptp_tir, 1);
5010         if (err)
5011                 goto err_destroy_xsk_tirs;
5012
5013         err = mlx5e_create_direct_tirs(priv, &priv->ptp_tir, 1);
5014         if (err)
5015                 goto err_destroy_ptp_rqt;
5016
5017         err = mlx5e_create_flow_steering(priv);
5018         if (err) {
5019                 mlx5_core_warn(mdev, "create flow steering failed, %d\n", err);
5020                 goto err_destroy_ptp_direct_tir;
5021         }
5022
5023         err = mlx5e_tc_nic_init(priv);
5024         if (err)
5025                 goto err_destroy_flow_steering;
5026
5027         err = mlx5e_accel_init_rx(priv);
5028         if (err)
5029                 goto err_tc_nic_cleanup;
5030
5031 #ifdef CONFIG_MLX5_EN_ARFS
5032         priv->netdev->rx_cpu_rmap =  mlx5_eq_table_get_rmap(priv->mdev);
5033 #endif
5034
5035         return 0;
5036
5037 err_tc_nic_cleanup:
5038         mlx5e_tc_nic_cleanup(priv);
5039 err_destroy_flow_steering:
5040         mlx5e_destroy_flow_steering(priv);
5041 err_destroy_ptp_direct_tir:
5042         mlx5e_destroy_direct_tirs(priv, &priv->ptp_tir, 1);
5043 err_destroy_ptp_rqt:
5044         mlx5e_destroy_direct_rqts(priv, &priv->ptp_tir, 1);
5045 err_destroy_xsk_tirs:
5046         mlx5e_destroy_direct_tirs(priv, priv->xsk_tir, max_nch);
5047 err_destroy_xsk_rqts:
5048         mlx5e_destroy_direct_rqts(priv, priv->xsk_tir, max_nch);
5049 err_destroy_direct_tirs:
5050         mlx5e_destroy_direct_tirs(priv, priv->direct_tir, max_nch);
5051 err_destroy_indirect_tirs:
5052         mlx5e_destroy_indirect_tirs(priv);
5053 err_destroy_direct_rqts:
5054         mlx5e_destroy_direct_rqts(priv, priv->direct_tir, max_nch);
5055 err_destroy_indirect_rqts:
5056         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
5057 err_close_drop_rq:
5058         mlx5e_close_drop_rq(&priv->drop_rq);
5059 err_destroy_q_counters:
5060         mlx5e_destroy_q_counters(priv);
5061         return err;
5062 }
5063
5064 static void mlx5e_cleanup_nic_rx(struct mlx5e_priv *priv)
5065 {
5066         u16 max_nch = priv->max_nch;
5067
5068         mlx5e_accel_cleanup_rx(priv);
5069         mlx5e_tc_nic_cleanup(priv);
5070         mlx5e_destroy_flow_steering(priv);
5071         mlx5e_destroy_direct_tirs(priv, &priv->ptp_tir, 1);
5072         mlx5e_destroy_direct_rqts(priv, &priv->ptp_tir, 1);
5073         mlx5e_destroy_direct_tirs(priv, priv->xsk_tir, max_nch);
5074         mlx5e_destroy_direct_rqts(priv, priv->xsk_tir, max_nch);
5075         mlx5e_destroy_direct_tirs(priv, priv->direct_tir, max_nch);
5076         mlx5e_destroy_indirect_tirs(priv);
5077         mlx5e_destroy_direct_rqts(priv, priv->direct_tir, max_nch);
5078         mlx5e_destroy_rqt(priv, &priv->indir_rqt);
5079         mlx5e_close_drop_rq(&priv->drop_rq);
5080         mlx5e_destroy_q_counters(priv);
5081 }
5082
5083 static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
5084 {
5085         int err;
5086
5087         err = mlx5e_create_tises(priv);
5088         if (err) {
5089                 mlx5_core_warn(priv->mdev, "create tises failed, %d\n", err);
5090                 return err;
5091         }
5092
5093         mlx5e_dcbnl_initialize(priv);
5094         return 0;
5095 }
5096
5097 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
5098 {
5099         struct net_device *netdev = priv->netdev;
5100         struct mlx5_core_dev *mdev = priv->mdev;
5101
5102         mlx5e_init_l2_addr(priv);
5103
5104         /* Marking the link as currently not needed by the Driver */
5105         if (!netif_running(netdev))
5106                 mlx5e_modify_admin_state(mdev, MLX5_PORT_DOWN);
5107
5108         mlx5e_set_netdev_mtu_boundaries(priv);
5109         mlx5e_set_dev_port_mtu(priv);
5110
5111         mlx5_lag_add(mdev, netdev);
5112
5113         mlx5e_enable_async_events(priv);
5114         mlx5e_enable_blocking_events(priv);
5115         if (mlx5e_monitor_counter_supported(priv))
5116                 mlx5e_monitor_counter_init(priv);
5117
5118         mlx5e_hv_vhca_stats_create(priv);
5119         if (netdev->reg_state != NETREG_REGISTERED)
5120                 return;
5121         mlx5e_dcbnl_init_app(priv);
5122
5123         mlx5e_nic_set_rx_mode(priv);
5124
5125         rtnl_lock();
5126         if (netif_running(netdev))
5127                 mlx5e_open(netdev);
5128         udp_tunnel_nic_reset_ntf(priv->netdev);
5129         netif_device_attach(netdev);
5130         rtnl_unlock();
5131 }
5132
5133 static void mlx5e_nic_disable(struct mlx5e_priv *priv)
5134 {
5135         struct mlx5_core_dev *mdev = priv->mdev;
5136
5137         if (priv->netdev->reg_state == NETREG_REGISTERED)
5138                 mlx5e_dcbnl_delete_app(priv);
5139
5140         rtnl_lock();
5141         if (netif_running(priv->netdev))
5142                 mlx5e_close(priv->netdev);
5143         netif_device_detach(priv->netdev);
5144         rtnl_unlock();
5145
5146         mlx5e_nic_set_rx_mode(priv);
5147
5148         mlx5e_hv_vhca_stats_destroy(priv);
5149         if (mlx5e_monitor_counter_supported(priv))
5150                 mlx5e_monitor_counter_cleanup(priv);
5151
5152         mlx5e_disable_blocking_events(priv);
5153         if (priv->en_trap) {
5154                 mlx5e_deactivate_trap(priv);
5155                 mlx5e_close_trap(priv->en_trap);
5156                 priv->en_trap = NULL;
5157         }
5158         mlx5e_disable_async_events(priv);
5159         mlx5_lag_remove(mdev);
5160         mlx5_vxlan_reset_to_default(mdev->vxlan);
5161 }
5162
5163 int mlx5e_update_nic_rx(struct mlx5e_priv *priv)
5164 {
5165         return mlx5e_refresh_tirs(priv, false, false);
5166 }
5167
5168 static const struct mlx5e_profile mlx5e_nic_profile = {
5169         .init              = mlx5e_nic_init,
5170         .cleanup           = mlx5e_nic_cleanup,
5171         .init_rx           = mlx5e_init_nic_rx,
5172         .cleanup_rx        = mlx5e_cleanup_nic_rx,
5173         .init_tx           = mlx5e_init_nic_tx,
5174         .cleanup_tx        = mlx5e_cleanup_nic_tx,
5175         .enable            = mlx5e_nic_enable,
5176         .disable           = mlx5e_nic_disable,
5177         .update_rx         = mlx5e_update_nic_rx,
5178         .update_stats      = mlx5e_stats_update_ndo_stats,
5179         .update_carrier    = mlx5e_update_carrier,
5180         .rx_handlers       = &mlx5e_rx_handlers_nic,
5181         .max_tc            = MLX5E_MAX_NUM_TC,
5182         .rq_groups         = MLX5E_NUM_RQ_GROUPS(XSK),
5183         .stats_grps        = mlx5e_nic_stats_grps,
5184         .stats_grps_num    = mlx5e_nic_stats_grps_num,
5185         .rx_ptp_support    = true,
5186 };
5187
5188 /* mlx5e generic netdev management API (move to en_common.c) */
5189 int mlx5e_priv_init(struct mlx5e_priv *priv,
5190                     struct net_device *netdev,
5191                     struct mlx5_core_dev *mdev)
5192 {
5193         /* priv init */
5194         priv->mdev        = mdev;
5195         priv->netdev      = netdev;
5196         priv->msglevel    = MLX5E_MSG_LEVEL;
5197         priv->max_opened_tc = 1;
5198
5199         if (!alloc_cpumask_var(&priv->scratchpad.cpumask, GFP_KERNEL))
5200                 return -ENOMEM;
5201
5202         mutex_init(&priv->state_lock);
5203         hash_init(priv->htb.qos_tc2node);
5204         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
5205         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
5206         INIT_WORK(&priv->tx_timeout_work, mlx5e_tx_timeout_work);
5207         INIT_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
5208
5209         priv->wq = create_singlethread_workqueue("mlx5e");
5210         if (!priv->wq)
5211                 goto err_free_cpumask;
5212
5213         return 0;
5214
5215 err_free_cpumask:
5216         free_cpumask_var(priv->scratchpad.cpumask);
5217
5218         return -ENOMEM;
5219 }
5220
5221 void mlx5e_priv_cleanup(struct mlx5e_priv *priv)
5222 {
5223         int i;
5224
5225         /* bail if change profile failed and also rollback failed */
5226         if (!priv->mdev)
5227                 return;
5228
5229         destroy_workqueue(priv->wq);
5230         free_cpumask_var(priv->scratchpad.cpumask);
5231
5232         for (i = 0; i < priv->htb.max_qos_sqs; i++)
5233                 kfree(priv->htb.qos_sq_stats[i]);
5234         kvfree(priv->htb.qos_sq_stats);
5235
5236         memset(priv, 0, sizeof(*priv));
5237 }
5238
5239 struct net_device *
5240 mlx5e_create_netdev(struct mlx5_core_dev *mdev, unsigned int txqs, unsigned int rxqs)
5241 {
5242         struct net_device *netdev;
5243         int err;
5244
5245         netdev = alloc_etherdev_mqs(sizeof(struct mlx5e_priv), txqs, rxqs);
5246         if (!netdev) {
5247                 mlx5_core_err(mdev, "alloc_etherdev_mqs() failed\n");
5248                 return NULL;
5249         }
5250
5251         err = mlx5e_priv_init(netdev_priv(netdev), netdev, mdev);
5252         if (err) {
5253                 mlx5_core_err(mdev, "mlx5e_priv_init failed, err=%d\n", err);
5254                 goto err_free_netdev;
5255         }
5256
5257         netif_carrier_off(netdev);
5258         dev_net_set(netdev, mlx5_core_net(mdev));
5259
5260         return netdev;
5261
5262 err_free_netdev:
5263         free_netdev(netdev);
5264
5265         return NULL;
5266 }
5267
5268 static void mlx5e_update_features(struct net_device *netdev)
5269 {
5270         if (netdev->reg_state != NETREG_REGISTERED)
5271                 return; /* features will be updated on netdev registration */
5272
5273         rtnl_lock();
5274         netdev_update_features(netdev);
5275         rtnl_unlock();
5276 }
5277
5278 int mlx5e_attach_netdev(struct mlx5e_priv *priv)
5279 {
5280         const bool take_rtnl = priv->netdev->reg_state == NETREG_REGISTERED;
5281         const struct mlx5e_profile *profile = priv->profile;
5282         int max_nch;
5283         int err;
5284
5285         clear_bit(MLX5E_STATE_DESTROYING, &priv->state);
5286
5287         /* max number of channels may have changed */
5288         max_nch = mlx5e_get_max_num_channels(priv->mdev);
5289         if (priv->channels.params.num_channels > max_nch) {
5290                 mlx5_core_warn(priv->mdev, "MLX5E: Reducing number of channels to %d\n", max_nch);
5291                 /* Reducing the number of channels - RXFH has to be reset, and
5292                  * mlx5e_num_channels_changed below will build the RQT.
5293                  */
5294                 priv->netdev->priv_flags &= ~IFF_RXFH_CONFIGURED;
5295                 priv->channels.params.num_channels = max_nch;
5296         }
5297         /* 1. Set the real number of queues in the kernel the first time.
5298          * 2. Set our default XPS cpumask.
5299          * 3. Build the RQT.
5300          *
5301          * rtnl_lock is required by netif_set_real_num_*_queues in case the
5302          * netdev has been registered by this point (if this function was called
5303          * in the reload or resume flow).
5304          */
5305         if (take_rtnl)
5306                 rtnl_lock();
5307         err = mlx5e_num_channels_changed(priv);
5308         if (take_rtnl)
5309                 rtnl_unlock();
5310         if (err)
5311                 goto out;
5312
5313         err = profile->init_tx(priv);
5314         if (err)
5315                 goto out;
5316
5317         err = profile->init_rx(priv);
5318         if (err)
5319                 goto err_cleanup_tx;
5320
5321         if (profile->enable)
5322                 profile->enable(priv);
5323
5324         mlx5e_update_features(priv->netdev);
5325
5326         return 0;
5327
5328 err_cleanup_tx:
5329         profile->cleanup_tx(priv);
5330
5331 out:
5332         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
5333         cancel_work_sync(&priv->update_stats_work);
5334         return err;
5335 }
5336
5337 void mlx5e_detach_netdev(struct mlx5e_priv *priv)
5338 {
5339         const struct mlx5e_profile *profile = priv->profile;
5340
5341         set_bit(MLX5E_STATE_DESTROYING, &priv->state);
5342
5343         if (profile->disable)
5344                 profile->disable(priv);
5345         flush_workqueue(priv->wq);
5346
5347         profile->cleanup_rx(priv);
5348         profile->cleanup_tx(priv);
5349         cancel_work_sync(&priv->update_stats_work);
5350 }
5351
5352 static int
5353 mlx5e_netdev_attach_profile(struct net_device *netdev, struct mlx5_core_dev *mdev,
5354                             const struct mlx5e_profile *new_profile, void *new_ppriv)
5355 {
5356         struct mlx5e_priv *priv = netdev_priv(netdev);
5357         int err;
5358
5359         err = mlx5e_priv_init(priv, netdev, mdev);
5360         if (err) {
5361                 mlx5_core_err(mdev, "mlx5e_priv_init failed, err=%d\n", err);
5362                 return err;
5363         }
5364         netif_carrier_off(netdev);
5365         priv->profile = new_profile;
5366         priv->ppriv = new_ppriv;
5367         err = new_profile->init(priv->mdev, priv->netdev);
5368         if (err)
5369                 goto priv_cleanup;
5370         err = mlx5e_attach_netdev(priv);
5371         if (err)
5372                 goto profile_cleanup;
5373         return err;
5374
5375 profile_cleanup:
5376         new_profile->cleanup(priv);
5377 priv_cleanup:
5378         mlx5e_priv_cleanup(priv);
5379         return err;
5380 }
5381
5382 int mlx5e_netdev_change_profile(struct mlx5e_priv *priv,
5383                                 const struct mlx5e_profile *new_profile, void *new_ppriv)
5384 {
5385         unsigned int new_max_nch = mlx5e_calc_max_nch(priv, new_profile);
5386         const struct mlx5e_profile *orig_profile = priv->profile;
5387         struct net_device *netdev = priv->netdev;
5388         struct mlx5_core_dev *mdev = priv->mdev;
5389         void *orig_ppriv = priv->ppriv;
5390         int err, rollback_err;
5391
5392         /* sanity */
5393         if (new_max_nch != priv->max_nch) {
5394                 netdev_warn(netdev, "%s: Replacing profile with different max channels\n",
5395                             __func__);
5396                 return -EINVAL;
5397         }
5398
5399         /* cleanup old profile */
5400         mlx5e_detach_netdev(priv);
5401         priv->profile->cleanup(priv);
5402         mlx5e_priv_cleanup(priv);
5403
5404         err = mlx5e_netdev_attach_profile(netdev, mdev, new_profile, new_ppriv);
5405         if (err) { /* roll back to original profile */
5406                 netdev_warn(netdev, "%s: new profile init failed, %d\n", __func__, err);
5407                 goto rollback;
5408         }
5409
5410         return 0;
5411
5412 rollback:
5413         rollback_err = mlx5e_netdev_attach_profile(netdev, mdev, orig_profile, orig_ppriv);
5414         if (rollback_err)
5415                 netdev_err(netdev, "%s: failed to rollback to orig profile, %d\n",
5416                            __func__, rollback_err);
5417         return err;
5418 }
5419
5420 void mlx5e_netdev_attach_nic_profile(struct mlx5e_priv *priv)
5421 {
5422         mlx5e_netdev_change_profile(priv, &mlx5e_nic_profile, NULL);
5423 }
5424
5425 void mlx5e_destroy_netdev(struct mlx5e_priv *priv)
5426 {
5427         struct net_device *netdev = priv->netdev;
5428
5429         mlx5e_priv_cleanup(priv);
5430         free_netdev(netdev);
5431 }
5432
5433 static int mlx5e_resume(struct auxiliary_device *adev)
5434 {
5435         struct mlx5_adev *edev = container_of(adev, struct mlx5_adev, adev);
5436         struct mlx5e_priv *priv = dev_get_drvdata(&adev->dev);
5437         struct net_device *netdev = priv->netdev;
5438         struct mlx5_core_dev *mdev = edev->mdev;
5439         int err;
5440
5441         if (netif_device_present(netdev))
5442                 return 0;
5443
5444         err = mlx5e_create_mdev_resources(mdev);
5445         if (err)
5446                 return err;
5447
5448         err = mlx5e_attach_netdev(priv);
5449         if (err) {
5450                 mlx5e_destroy_mdev_resources(mdev);
5451                 return err;
5452         }
5453
5454         return 0;
5455 }
5456
5457 static int mlx5e_suspend(struct auxiliary_device *adev, pm_message_t state)
5458 {
5459         struct mlx5e_priv *priv = dev_get_drvdata(&adev->dev);
5460         struct net_device *netdev = priv->netdev;
5461         struct mlx5_core_dev *mdev = priv->mdev;
5462
5463         if (!netif_device_present(netdev))
5464                 return -ENODEV;
5465
5466         mlx5e_detach_netdev(priv);
5467         mlx5e_destroy_mdev_resources(mdev);
5468         return 0;
5469 }
5470
5471 static int mlx5e_probe(struct auxiliary_device *adev,
5472                        const struct auxiliary_device_id *id)
5473 {
5474         struct mlx5_adev *edev = container_of(adev, struct mlx5_adev, adev);
5475         const struct mlx5e_profile *profile = &mlx5e_nic_profile;
5476         struct mlx5_core_dev *mdev = edev->mdev;
5477         struct net_device *netdev;
5478         pm_message_t state = {};
5479         unsigned int txqs, rxqs, ptp_txqs = 0;
5480         struct mlx5e_priv *priv;
5481         int qos_sqs = 0;
5482         int err;
5483         int nch;
5484
5485         if (MLX5_CAP_GEN(mdev, ts_cqe_to_dest_cqn))
5486                 ptp_txqs = profile->max_tc;
5487
5488         if (mlx5_qos_is_supported(mdev))
5489                 qos_sqs = mlx5e_qos_max_leaf_nodes(mdev);
5490
5491         nch = mlx5e_get_max_num_channels(mdev);
5492         txqs = nch * profile->max_tc + ptp_txqs + qos_sqs;
5493         rxqs = nch * profile->rq_groups;
5494         netdev = mlx5e_create_netdev(mdev, txqs, rxqs);
5495         if (!netdev) {
5496                 mlx5_core_err(mdev, "mlx5e_create_netdev failed\n");
5497                 return -ENOMEM;
5498         }
5499
5500         mlx5e_build_nic_netdev(netdev);
5501
5502         priv = netdev_priv(netdev);
5503         dev_set_drvdata(&adev->dev, priv);
5504
5505         priv->profile = profile;
5506         priv->ppriv = NULL;
5507
5508         err = mlx5e_devlink_port_register(priv);
5509         if (err) {
5510                 mlx5_core_err(mdev, "mlx5e_devlink_port_register failed, %d\n", err);
5511                 goto err_destroy_netdev;
5512         }
5513
5514         err = profile->init(mdev, netdev);
5515         if (err) {
5516                 mlx5_core_err(mdev, "mlx5e_nic_profile init failed, %d\n", err);
5517                 goto err_devlink_cleanup;
5518         }
5519
5520         err = mlx5e_resume(adev);
5521         if (err) {
5522                 mlx5_core_err(mdev, "mlx5e_resume failed, %d\n", err);
5523                 goto err_profile_cleanup;
5524         }
5525
5526         err = register_netdev(netdev);
5527         if (err) {
5528                 mlx5_core_err(mdev, "register_netdev failed, %d\n", err);
5529                 goto err_resume;
5530         }
5531
5532         mlx5e_devlink_port_type_eth_set(priv);
5533
5534         mlx5e_dcbnl_init_app(priv);
5535         mlx5_uplink_netdev_set(mdev, netdev);
5536         return 0;
5537
5538 err_resume:
5539         mlx5e_suspend(adev, state);
5540 err_profile_cleanup:
5541         profile->cleanup(priv);
5542 err_devlink_cleanup:
5543         mlx5e_devlink_port_unregister(priv);
5544 err_destroy_netdev:
5545         mlx5e_destroy_netdev(priv);
5546         return err;
5547 }
5548
5549 static void mlx5e_remove(struct auxiliary_device *adev)
5550 {
5551         struct mlx5e_priv *priv = dev_get_drvdata(&adev->dev);
5552         pm_message_t state = {};
5553
5554         mlx5e_dcbnl_delete_app(priv);
5555         unregister_netdev(priv->netdev);
5556         mlx5e_suspend(adev, state);
5557         priv->profile->cleanup(priv);
5558         mlx5e_devlink_port_unregister(priv);
5559         mlx5e_destroy_netdev(priv);
5560 }
5561
5562 static const struct auxiliary_device_id mlx5e_id_table[] = {
5563         { .name = MLX5_ADEV_NAME ".eth", },
5564         {},
5565 };
5566
5567 MODULE_DEVICE_TABLE(auxiliary, mlx5e_id_table);
5568
5569 static struct auxiliary_driver mlx5e_driver = {
5570         .name = "eth",
5571         .probe = mlx5e_probe,
5572         .remove = mlx5e_remove,
5573         .suspend = mlx5e_suspend,
5574         .resume = mlx5e_resume,
5575         .id_table = mlx5e_id_table,
5576 };
5577
5578 int mlx5e_init(void)
5579 {
5580         int ret;
5581
5582         mlx5e_ipsec_build_inverse_table();
5583         mlx5e_build_ptys2ethtool_map();
5584         ret = auxiliary_driver_register(&mlx5e_driver);
5585         if (ret)
5586                 return ret;
5587
5588         ret = mlx5e_rep_init();
5589         if (ret)
5590                 auxiliary_driver_unregister(&mlx5e_driver);
5591         return ret;
5592 }
5593
5594 void mlx5e_cleanup(void)
5595 {
5596         mlx5e_rep_cleanup();
5597         auxiliary_driver_unregister(&mlx5e_driver);
5598 }