drivers/net/mlx5/linux/mlx5_os.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2020 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stddef.h>
   7 #include <unistd.h>
   8 #include <string.h>
   9 #include <stdint.h>
  10 #include <stdlib.h>
  11 #include <errno.h>
  12 #include <net/if.h>
  13 #include <linux/rtnetlink.h>
  14 #include <linux/sockios.h>
  15 #include <linux/ethtool.h>
  16 #include <fcntl.h>
  17
  18 /* Verbs header. */
  19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  20 #ifdef PEDANTIC
  21 #pragma GCC diagnostic ignored "-Wpedantic"
  22 #endif
  23 #include <infiniband/verbs.h>
  24 #ifdef PEDANTIC
  25 #pragma GCC diagnostic error "-Wpedantic"
  26 #endif
  27
  28 #include <rte_malloc.h>
  29 #include <rte_ethdev_driver.h>
  30 #include <rte_ethdev_pci.h>
  31 #include <rte_pci.h>
  32 #include <rte_bus_pci.h>
  33 #include <rte_common.h>
  34 #include <rte_kvargs.h>
  35 #include <rte_rwlock.h>
  36 #include <rte_spinlock.h>
  37 #include <rte_string_fns.h>
  38 #include <rte_alarm.h>
  39 #include <rte_eal_paging.h>
  40
  41 #include <mlx5_glue.h>
  42 #include <mlx5_devx_cmds.h>
  43 #include <mlx5_common.h>
  44 #include <mlx5_common_mp.h>
  45 #include <mlx5_common_mr.h>
  46 #include <mlx5_malloc.h>
  47
  48 #include "mlx5_defs.h"
  49 #include "mlx5.h"
  50 #include "mlx5_common_os.h"
  51 #include "mlx5_utils.h"
  52 #include "mlx5_rxtx.h"
  53 #include "mlx5_autoconf.h"
  54 #include "mlx5_mr.h"
  55 #include "mlx5_flow.h"
  56 #include "rte_pmd_mlx5.h"
  57 #include "mlx5_verbs.h"
  58
  59 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  60
  61 #ifndef HAVE_IBV_MLX5_MOD_MPW
  62 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  63 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  64 #endif
  65
  66 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  67 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  68 #endif
  69
  70 /**
  71  * Get mlx5 device attributes. The glue function query_device_ex() is called
  72  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  73  * device attributes from the glue out parameter.
  74  *
  75  * @param dev
  76  *   Pointer to ibv context.
  77  *
  78  * @param device_attr
  79  *   Pointer to mlx5 device attributes.
  80  *
  81  * @return
  82  *   0 on success, non zero error number otherwise
  83  */
  84 int
  85 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
  86 {
  87         int err;
  88         struct ibv_device_attr_ex attr_ex;
  89         memset(device_attr, 0, sizeof(*device_attr));
  90         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
  91         if (err)
  92                 return err;
  93
  94         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
  95         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
  96         device_attr->max_sge = attr_ex.orig_attr.max_sge;
  97         device_attr->max_cq = attr_ex.orig_attr.max_cq;
  98         device_attr->max_qp = attr_ex.orig_attr.max_qp;
  99         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 100         device_attr->max_rwq_indirection_table_size =
 101                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 102         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 103         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 104
 105         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 106         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 107         if (err)
 108                 return err;
 109
 110         device_attr->flags = dv_attr.flags;
 111         device_attr->comp_mask = dv_attr.comp_mask;
 112 #ifdef HAVE_IBV_MLX5_MOD_SWP
 113         device_attr->sw_parsing_offloads =
 114                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 115 #endif
 116         device_attr->min_single_stride_log_num_of_bytes =
 117                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 118         device_attr->max_single_stride_log_num_of_bytes =
 119                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 120         device_attr->min_single_wqe_log_num_of_strides =
 121                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 122         device_attr->max_single_wqe_log_num_of_strides =
 123                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 124         device_attr->stride_supported_qpts =
 125                 dv_attr.striding_rq_caps.supported_qpts;
 126 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 127         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 128 #endif
 129
 130         return err;
 131 }
 132
 133 /**
 134  * Verbs callback to allocate a memory. This function should allocate the space
 135  * according to the size provided residing inside a huge page.
 136  * Please note that all allocation must respect the alignment from libmlx5
 137  * (i.e. currently rte_mem_page_size()).
 138  *
 139  * @param[in] size
 140  *   The size in bytes of the memory to allocate.
 141  * @param[in] data
 142  *   A pointer to the callback data.
 143  *
 144  * @return
 145  *   Allocated buffer, NULL otherwise and rte_errno is set.
 146  */
 147 static void *
 148 mlx5_alloc_verbs_buf(size_t size, void *data)
 149 {
 150         struct mlx5_priv *priv = data;
 151         void *ret;
 152         unsigned int socket = SOCKET_ID_ANY;
 153         size_t alignment = rte_mem_page_size();
 154         if (alignment == (size_t)-1) {
 155                 DRV_LOG(ERR, "Failed to get mem page size");
 156                 rte_errno = ENOMEM;
 157                 return NULL;
 158         }
 159
 160         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 161                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 162
 163                 socket = ctrl->socket;
 164         } else if (priv->verbs_alloc_ctx.type ==
 165                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 166                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 167
 168                 socket = ctrl->socket;
 169         }
 170         MLX5_ASSERT(data != NULL);
 171         ret = mlx5_malloc(0, size, alignment, socket);
 172         if (!ret && size)
 173                 rte_errno = ENOMEM;
 174         return ret;
 175 }
 176
 177 /**
 178  * Verbs callback to free a memory.
 179  *
 180  * @param[in] ptr
 181  *   A pointer to the memory to free.
 182  * @param[in] data
 183  *   A pointer to the callback data.
 184  */
 185 static void
 186 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 187 {
 188         MLX5_ASSERT(data != NULL);
 189         mlx5_free(ptr);
 190 }
 191
 192 /**
 193  * Initialize DR related data within private structure.
 194  * Routine checks the reference counter and does actual
 195  * resources creation/initialization only if counter is zero.
 196  *
 197  * @param[in] priv
 198  *   Pointer to the private device data structure.
 199  *
 200  * @return
 201  *   Zero on success, positive error code otherwise.
 202  */
 203 static int
 204 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 205 {
 206         struct mlx5_dev_ctx_shared *sh = priv->sh;
 207         char s[MLX5_HLIST_NAMESIZE];
 208         int err = 0;
 209
 210         if (!sh->flow_tbls)
 211                 err = mlx5_alloc_table_hash_list(priv);
 212         else
 213                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 214                         (void *)sh->flow_tbls);
 215         if (err)
 216                 return err;
 217         /* Create tags hash list table. */
 218         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 219         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 220         if (!sh->tag_table) {
 221                 DRV_LOG(ERR, "tags with hash creation failed.");
 222                 err = ENOMEM;
 223                 goto error;
 224         }
 225 #ifdef HAVE_MLX5DV_DR
 226         void *domain;
 227
 228         if (sh->dv_refcnt) {
 229                 /* Shared DV/DR structures is already initialized. */
 230                 sh->dv_refcnt++;
 231                 priv->dr_shared = 1;
 232                 return 0;
 233         }
 234         /* Reference counter is zero, we should initialize structures. */
 235         domain = mlx5_glue->dr_create_domain(sh->ctx,
 236                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 237         if (!domain) {
 238                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 239                 err = errno;
 240                 goto error;
 241         }
 242         sh->rx_domain = domain;
 243         domain = mlx5_glue->dr_create_domain(sh->ctx,
 244                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 245         if (!domain) {
 246                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 247                 err = errno;
 248                 goto error;
 249         }
 250         pthread_mutex_init(&sh->dv_mutex, NULL);
 251         sh->tx_domain = domain;
 252 #ifdef HAVE_MLX5DV_DR_ESWITCH
 253         if (priv->config.dv_esw_en) {
 254                 domain  = mlx5_glue->dr_create_domain
 255                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 256                 if (!domain) {
 257                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 258                         err = errno;
 259                         goto error;
 260                 }
 261                 sh->fdb_domain = domain;
 262                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 263         }
 264 #endif
 265         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 266                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 267                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 268                 if (sh->fdb_domain)
 269                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 270         }
 271         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 272 #endif /* HAVE_MLX5DV_DR */
 273         sh->dv_refcnt++;
 274         priv->dr_shared = 1;
 275         return 0;
 276 error:
 277         /* Rollback the created objects. */
 278         if (sh->rx_domain) {
 279                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 280                 sh->rx_domain = NULL;
 281         }
 282         if (sh->tx_domain) {
 283                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 284                 sh->tx_domain = NULL;
 285         }
 286         if (sh->fdb_domain) {
 287                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 288                 sh->fdb_domain = NULL;
 289         }
 290         if (sh->esw_drop_action) {
 291                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 292                 sh->esw_drop_action = NULL;
 293         }
 294         if (sh->pop_vlan_action) {
 295                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 296                 sh->pop_vlan_action = NULL;
 297         }
 298         if (sh->tag_table) {
 299                 /* tags should be destroyed with flow before. */
 300                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 301                 sh->tag_table = NULL;
 302         }
 303         mlx5_free_table_hash_list(priv);
 304         return err;
 305 }
 306
 307 /**
 308  * Destroy DR related data within private structure.
 309  *
 310  * @param[in] priv
 311  *   Pointer to the private device data structure.
 312  */
 313 void
 314 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 315 {
 316         struct mlx5_dev_ctx_shared *sh;
 317
 318         if (!priv->dr_shared)
 319                 return;
 320         priv->dr_shared = 0;
 321         sh = priv->sh;
 322         MLX5_ASSERT(sh);
 323 #ifdef HAVE_MLX5DV_DR
 324         MLX5_ASSERT(sh->dv_refcnt);
 325         if (sh->dv_refcnt && --sh->dv_refcnt)
 326                 return;
 327         if (sh->rx_domain) {
 328                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 329                 sh->rx_domain = NULL;
 330         }
 331         if (sh->tx_domain) {
 332                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 333                 sh->tx_domain = NULL;
 334         }
 335 #ifdef HAVE_MLX5DV_DR_ESWITCH
 336         if (sh->fdb_domain) {
 337                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 338                 sh->fdb_domain = NULL;
 339         }
 340         if (sh->esw_drop_action) {
 341                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 342                 sh->esw_drop_action = NULL;
 343         }
 344 #endif
 345         if (sh->pop_vlan_action) {
 346                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 347                 sh->pop_vlan_action = NULL;
 348         }
 349         pthread_mutex_destroy(&sh->dv_mutex);
 350 #endif /* HAVE_MLX5DV_DR */
 351         if (sh->tag_table) {
 352                 /* tags should be destroyed with flow before. */
 353                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 354                 sh->tag_table = NULL;
 355         }
 356         mlx5_free_table_hash_list(priv);
 357 }
 358
 359 /**
 360  * Spawn an Ethernet device from Verbs information.
 361  *
 362  * @param dpdk_dev
 363  *   Backing DPDK device.
 364  * @param spawn
 365  *   Verbs device parameters (name, port, switch_info) to spawn.
 366  * @param config
 367  *   Device configuration parameters.
 368  *
 369  * @return
 370  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 371  *   is set. The following errors are defined:
 372  *
 373  *   EBUSY: device is not supposed to be spawned.
 374  *   EEXIST: device is already spawned
 375  */
 376 static struct rte_eth_dev *
 377 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 378                struct mlx5_dev_spawn_data *spawn,
 379                struct mlx5_dev_config config)
 380 {
 381         const struct mlx5_switch_info *switch_info = &spawn->info;
 382         struct mlx5_dev_ctx_shared *sh = NULL;
 383         struct ibv_port_attr port_attr;
 384         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 385         struct rte_eth_dev *eth_dev = NULL;
 386         struct mlx5_priv *priv = NULL;
 387         int err = 0;
 388         unsigned int hw_padding = 0;
 389         unsigned int mps;
 390         unsigned int cqe_comp;
 391         unsigned int cqe_pad = 0;
 392         unsigned int tunnel_en = 0;
 393         unsigned int mpls_en = 0;
 394         unsigned int swp = 0;
 395         unsigned int mprq = 0;
 396         unsigned int mprq_min_stride_size_n = 0;
 397         unsigned int mprq_max_stride_size_n = 0;
 398         unsigned int mprq_min_stride_num_n = 0;
 399         unsigned int mprq_max_stride_num_n = 0;
 400         struct rte_ether_addr mac;
 401         char name[RTE_ETH_NAME_MAX_LEN];
 402         int own_domain_id = 0;
 403         uint16_t port_id;
 404         unsigned int i;
 405 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 406         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 407 #endif
 408
 409         /* Determine if this port representor is supposed to be spawned. */
 410         if (switch_info->representor && dpdk_dev->devargs) {
 411                 struct rte_eth_devargs eth_da;
 412
 413                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 414                 if (err) {
 415                         rte_errno = -err;
 416                         DRV_LOG(ERR, "failed to process device arguments: %s",
 417                                 strerror(rte_errno));
 418                         return NULL;
 419                 }
 420                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 421                         if (eth_da.representor_ports[i] ==
 422                             (uint16_t)switch_info->port_name)
 423                                 break;
 424                 if (i == eth_da.nb_representor_ports) {
 425                         rte_errno = EBUSY;
 426                         return NULL;
 427                 }
 428         }
 429         /* Build device name. */
 430         if (spawn->pf_bond <  0) {
 431                 /* Single device. */
 432                 if (!switch_info->representor)
 433                         strlcpy(name, dpdk_dev->name, sizeof(name));
 434                 else
 435                         snprintf(name, sizeof(name), "%s_representor_%u",
 436                                  dpdk_dev->name, switch_info->port_name);
 437         } else {
 438                 /* Bonding device. */
 439                 if (!switch_info->representor)
 440                         snprintf(name, sizeof(name), "%s_%s",
 441                                  dpdk_dev->name,
 442                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 443                 else
 444                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 445                                  dpdk_dev->name,
 446                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 447                                  switch_info->port_name);
 448         }
 449         /* check if the device is already spawned */
 450         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 451                 rte_errno = EEXIST;
 452                 return NULL;
 453         }
 454         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 455         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 456                 struct mlx5_mp_id mp_id;
 457
 458                 eth_dev = rte_eth_dev_attach_secondary(name);
 459                 if (eth_dev == NULL) {
 460                         DRV_LOG(ERR, "can not attach rte ethdev");
 461                         rte_errno = ENOMEM;
 462                         return NULL;
 463                 }
 464                 eth_dev->device = dpdk_dev;
 465                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 466                 err = mlx5_proc_priv_init(eth_dev);
 467                 if (err)
 468                         return NULL;
 469                 mp_id.port_id = eth_dev->data->port_id;
 470                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 471                 /* Receive command fd from primary process */
 472                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 473                 if (err < 0)
 474                         goto err_secondary;
 475                 /* Remap UAR for Tx queues. */
 476                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 477                 if (err)
 478                         goto err_secondary;
 479                 /*
 480                  * Ethdev pointer is still required as input since
 481                  * the primary device is not accessible from the
 482                  * secondary process.
 483                  */
 484                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 485                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 486                 return eth_dev;
 487 err_secondary:
 488                 mlx5_dev_close(eth_dev);
 489                 return NULL;
 490         }
 491         /*
 492          * Some parameters ("tx_db_nc" in particularly) are needed in
 493          * advance to create dv/verbs device context. We proceed the
 494          * devargs here to get ones, and later proceed devargs again
 495          * to override some hardware settings.
 496          */
 497         err = mlx5_args(&config, dpdk_dev->devargs);
 498         if (err) {
 499                 err = rte_errno;
 500                 DRV_LOG(ERR, "failed to process device arguments: %s",
 501                         strerror(rte_errno));
 502                 goto error;
 503         }
 504         mlx5_malloc_mem_select(config.sys_mem_en);
 505         sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
 506         if (!sh)
 507                 return NULL;
 508         config.devx = sh->devx;
 509 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 510         config.dest_tir = 1;
 511 #endif
 512 #ifdef HAVE_IBV_MLX5_MOD_SWP
 513         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 514 #endif
 515         /*
 516          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 517          * as all ConnectX-5 devices.
 518          */
 519 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 520         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 521 #endif
 522 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 523         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 524 #endif
 525         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 526         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 527                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 528                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 529                         mps = MLX5_MPW_ENHANCED;
 530                 } else {
 531                         DRV_LOG(DEBUG, "MPW is supported");
 532                         mps = MLX5_MPW;
 533                 }
 534         } else {
 535                 DRV_LOG(DEBUG, "MPW isn't supported");
 536                 mps = MLX5_MPW_DISABLED;
 537         }
 538 #ifdef HAVE_IBV_MLX5_MOD_SWP
 539         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 540                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 541         DRV_LOG(DEBUG, "SWP support: %u", swp);
 542 #endif
 543         config.swp = !!swp;
 544 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 545         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 546                 struct mlx5dv_striding_rq_caps mprq_caps =
 547                         dv_attr.striding_rq_caps;
 548
 549                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 550                         mprq_caps.min_single_stride_log_num_of_bytes);
 551                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 552                         mprq_caps.max_single_stride_log_num_of_bytes);
 553                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 554                         mprq_caps.min_single_wqe_log_num_of_strides);
 555                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 556                         mprq_caps.max_single_wqe_log_num_of_strides);
 557                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 558                         mprq_caps.supported_qpts);
 559                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 560                 mprq = 1;
 561                 mprq_min_stride_size_n =
 562                         mprq_caps.min_single_stride_log_num_of_bytes;
 563                 mprq_max_stride_size_n =
 564                         mprq_caps.max_single_stride_log_num_of_bytes;
 565                 mprq_min_stride_num_n =
 566                         mprq_caps.min_single_wqe_log_num_of_strides;
 567                 mprq_max_stride_num_n =
 568                         mprq_caps.max_single_wqe_log_num_of_strides;
 569         }
 570 #endif
 571         if (RTE_CACHE_LINE_SIZE == 128 &&
 572             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 573                 cqe_comp = 0;
 574         else
 575                 cqe_comp = 1;
 576         config.cqe_comp = cqe_comp;
 577 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 578         /* Whether device supports 128B Rx CQE padding. */
 579         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 580                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 581 #endif
 582 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 583         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 584                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 585                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 586                              (dv_attr.tunnel_offloads_caps &
 587                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 588                              (dv_attr.tunnel_offloads_caps &
 589                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 590         }
 591         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 592                 tunnel_en ? "" : "not ");
 593 #else
 594         DRV_LOG(WARNING,
 595                 "tunnel offloading disabled due to old OFED/rdma-core version");
 596 #endif
 597         config.tunnel_en = tunnel_en;
 598 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 599         mpls_en = ((dv_attr.tunnel_offloads_caps &
 600                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 601                    (dv_attr.tunnel_offloads_caps &
 602                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 603         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 604                 mpls_en ? "" : "not ");
 605 #else
 606         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 607                 " old OFED/rdma-core version or firmware configuration");
 608 #endif
 609         config.mpls_en = mpls_en;
 610         /* Check port status. */
 611         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 612         if (err) {
 613                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 614                 goto error;
 615         }
 616         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 617                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 618                 err = EINVAL;
 619                 goto error;
 620         }
 621         if (port_attr.state != IBV_PORT_ACTIVE)
 622                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 623                         mlx5_glue->port_state_str(port_attr.state),
 624                         port_attr.state);
 625         /* Allocate private eth device data. */
 626         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
 627                            sizeof(*priv),
 628                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 629         if (priv == NULL) {
 630                 DRV_LOG(ERR, "priv allocation failure");
 631                 err = ENOMEM;
 632                 goto error;
 633         }
 634         priv->sh = sh;
 635         priv->dev_port = spawn->phys_port;
 636         priv->pci_dev = spawn->pci_dev;
 637         priv->mtu = RTE_ETHER_MTU;
 638         priv->mp_id.port_id = port_id;
 639         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 640         /* Some internal functions rely on Netlink sockets, open them now. */
 641         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 642         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 643         priv->representor = !!switch_info->representor;
 644         priv->master = !!switch_info->master;
 645         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 646         priv->vport_meta_tag = 0;
 647         priv->vport_meta_mask = 0;
 648         priv->pf_bond = spawn->pf_bond;
 649 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 650         /*
 651          * The DevX port query API is implemented. E-Switch may use
 652          * either vport or reg_c[0] metadata register to match on
 653          * vport index. The engaged part of metadata register is
 654          * defined by mask.
 655          */
 656         if (switch_info->representor || switch_info->master) {
 657                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 658                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 659                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 660                                                  &devx_port);
 661                 if (err) {
 662                         DRV_LOG(WARNING,
 663                                 "can't query devx port %d on device %s",
 664                                 spawn->phys_port,
 665                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 666                         devx_port.comp_mask = 0;
 667                 }
 668         }
 669         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 670                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 671                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 672                 if (!priv->vport_meta_mask) {
 673                         DRV_LOG(ERR, "vport zero mask for port %d"
 674                                      " on bonding device %s",
 675                                      spawn->phys_port,
 676                                      mlx5_os_get_dev_device_name
 677                                                         (spawn->phys_dev));
 678                         err = ENOTSUP;
 679                         goto error;
 680                 }
 681                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 682                         DRV_LOG(ERR, "invalid vport tag for port %d"
 683                                      " on bonding device %s",
 684                                      spawn->phys_port,
 685                                      mlx5_os_get_dev_device_name
 686                                                         (spawn->phys_dev));
 687                         err = ENOTSUP;
 688                         goto error;
 689                 }
 690         }
 691         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 692                 priv->vport_id = devx_port.vport_num;
 693         } else if (spawn->pf_bond >= 0) {
 694                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 695                              " on bonding device %s",
 696                              spawn->phys_port,
 697                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 698                 err = ENOTSUP;
 699                 goto error;
 700         } else {
 701                 /* Suppose vport index in compatible way. */
 702                 priv->vport_id = switch_info->representor ?
 703                                  switch_info->port_name + 1 : -1;
 704         }
 705 #else
 706         /*
 707          * Kernel/rdma_core support single E-Switch per PF configurations
 708          * only and vport_id field contains the vport index for
 709          * associated VF, which is deduced from representor port name.
 710          * For example, let's have the IB device port 10, it has
 711          * attached network device eth0, which has port name attribute
 712          * pf0vf2, we can deduce the VF number as 2, and set vport index
 713          * as 3 (2+1). This assigning schema should be changed if the
 714          * multiple E-Switch instances per PF configurations or/and PCI
 715          * subfunctions are added.
 716          */
 717         priv->vport_id = switch_info->representor ?
 718                          switch_info->port_name + 1 : -1;
 719 #endif
 720         /* representor_id field keeps the unmodified VF index. */
 721         priv->representor_id = switch_info->representor ?
 722                                switch_info->port_name : -1;
 723         /*
 724          * Look for sibling devices in order to reuse their switch domain
 725          * if any, otherwise allocate one.
 726          */
 727         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 728                 const struct mlx5_priv *opriv =
 729                         rte_eth_devices[port_id].data->dev_private;
 730
 731                 if (!opriv ||
 732                     opriv->sh != priv->sh ||
 733                         opriv->domain_id ==
 734                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 735                         continue;
 736                 priv->domain_id = opriv->domain_id;
 737                 break;
 738         }
 739         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 740                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 741                 if (err) {
 742                         err = rte_errno;
 743                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 744                                 strerror(rte_errno));
 745                         goto error;
 746                 }
 747                 own_domain_id = 1;
 748         }
 749         /* Override some values set by hardware configuration. */
 750         mlx5_args(&config, dpdk_dev->devargs);
 751         err = mlx5_dev_check_sibling_config(priv, &config);
 752         if (err)
 753                 goto error;
 754         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 755                             IBV_DEVICE_RAW_IP_CSUM);
 756         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 757                 (config.hw_csum ? "" : "not "));
 758 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 759         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 760         DRV_LOG(DEBUG, "counters are not supported");
 761 #endif
 762 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 763         if (config.dv_flow_en) {
 764                 DRV_LOG(WARNING, "DV flow is not supported");
 765                 config.dv_flow_en = 0;
 766         }
 767 #endif
 768         config.ind_table_max_size =
 769                 sh->device_attr.max_rwq_indirection_table_size;
 770         /*
 771          * Remove this check once DPDK supports larger/variable
 772          * indirection tables.
 773          */
 774         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 775                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 776         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 777                 config.ind_table_max_size);
 778         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 779                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 780         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 781                 (config.hw_vlan_strip ? "" : "not "));
 782         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 783                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 784 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 785         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 786 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 787         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 788                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 789 #endif
 790         if (config.hw_padding && !hw_padding) {
 791                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 792                 config.hw_padding = 0;
 793         } else if (config.hw_padding) {
 794                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 795         }
 796         config.tso = (sh->device_attr.max_tso > 0 &&
 797                       (sh->device_attr.tso_supported_qpts &
 798                        (1 << IBV_QPT_RAW_PACKET)));
 799         if (config.tso)
 800                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 801         /*
 802          * MPW is disabled by default, while the Enhanced MPW is enabled
 803          * by default.
 804          */
 805         if (config.mps == MLX5_ARG_UNSET)
 806                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 807                                                           MLX5_MPW_DISABLED;
 808         else
 809                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 810         DRV_LOG(INFO, "%sMPS is %s",
 811                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 812                 config.mps == MLX5_MPW ? "legacy " : "",
 813                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 814         if (config.cqe_comp && !cqe_comp) {
 815                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 816                 config.cqe_comp = 0;
 817         }
 818         if (config.cqe_pad && !cqe_pad) {
 819                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 820                 config.cqe_pad = 0;
 821         } else if (config.cqe_pad) {
 822                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 823         }
 824         if (config.devx) {
 825                 priv->counter_fallback = 0;
 826                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 827                 if (err) {
 828                         err = -err;
 829                         goto error;
 830                 }
 831                 if (!config.hca_attr.flow_counters_dump)
 832                         priv->counter_fallback = 1;
 833 #ifndef HAVE_IBV_DEVX_ASYNC
 834                 priv->counter_fallback = 1;
 835 #endif
 836                 if (priv->counter_fallback)
 837                         DRV_LOG(INFO, "Use fall-back DV counter management");
 838                 /* Check for LRO support. */
 839                 if (config.dest_tir && config.hca_attr.lro_cap &&
 840                     config.dv_flow_en) {
 841                         /* TBD check tunnel lro caps. */
 842                         config.lro.supported = config.hca_attr.lro_cap;
 843                         DRV_LOG(DEBUG, "Device supports LRO");
 844                         /*
 845                          * If LRO timeout is not configured by application,
 846                          * use the minimal supported value.
 847                          */
 848                         if (!config.lro.timeout)
 849                                 config.lro.timeout =
 850                                 config.hca_attr.lro_timer_supported_periods[0];
 851                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 852                                 config.lro.timeout);
 853                 }
 854 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 855                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 856                     config.dv_flow_en) {
 857                         uint8_t reg_c_mask =
 858                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 859                         /*
 860                          * Meter needs two REG_C's for color match and pre-sfx
 861                          * flow match. Here get the REG_C for color match.
 862                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 863                          */
 864                         reg_c_mask &= 0xfc;
 865                         if (__builtin_popcount(reg_c_mask) < 1) {
 866                                 priv->mtr_en = 0;
 867                                 DRV_LOG(WARNING, "No available register for"
 868                                         " meter.");
 869                         } else {
 870                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 871                                                       REG_C_0;
 872                                 priv->mtr_en = 1;
 873                                 priv->mtr_reg_share =
 874                                       config.hca_attr.qos.flow_meter_reg_share;
 875                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 876                                         priv->mtr_color_reg);
 877                         }
 878                 }
 879 #endif
 880         }
 881         if (config.tx_pp) {
 882                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
 883                         config.hca_attr.dev_freq_khz);
 884                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
 885                         config.hca_attr.qos.packet_pacing ? "" : "not ");
 886                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
 887                         config.hca_attr.cross_channel ? "" : "not ");
 888                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
 889                         config.hca_attr.wqe_index_ignore ? "" : "not ");
 890                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
 891                         config.hca_attr.non_wire_sq ? "" : "not ");
 892                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
 893                         config.hca_attr.log_max_static_sq_wq ? "" : "not ",
 894                         config.hca_attr.log_max_static_sq_wq);
 895                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
 896                         config.hca_attr.qos.wqe_rate_pp ? "" : "not ");
 897                 if (!config.devx) {
 898                         DRV_LOG(ERR, "DevX is required for packet pacing");
 899                         err = ENODEV;
 900                         goto error;
 901                 }
 902                 if (!config.hca_attr.qos.packet_pacing) {
 903                         DRV_LOG(ERR, "Packet pacing is not supported");
 904                         err = ENODEV;
 905                         goto error;
 906                 }
 907                 if (!config.hca_attr.cross_channel) {
 908                         DRV_LOG(ERR, "Cross channel operations are"
 909                                      " required for packet pacing");
 910                         err = ENODEV;
 911                         goto error;
 912                 }
 913                 if (!config.hca_attr.wqe_index_ignore) {
 914                         DRV_LOG(ERR, "WQE index ignore feature is"
 915                                      " required for packet pacing");
 916                         err = ENODEV;
 917                         goto error;
 918                 }
 919                 if (!config.hca_attr.non_wire_sq) {
 920                         DRV_LOG(ERR, "Non-wire SQ feature is"
 921                                      " required for packet pacing");
 922                         err = ENODEV;
 923                         goto error;
 924                 }
 925                 if (!config.hca_attr.log_max_static_sq_wq) {
 926                         DRV_LOG(ERR, "Static WQE SQ feature is"
 927                                      " required for packet pacing");
 928                         err = ENODEV;
 929                         goto error;
 930                 }
 931                 if (!config.hca_attr.qos.wqe_rate_pp) {
 932                         DRV_LOG(ERR, "WQE rate mode is required"
 933                                      " for packet pacing");
 934                         err = ENODEV;
 935                         goto error;
 936                 }
 937 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
 938                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
 939                              " can't create queues for packet pacing");
 940                 err = ENODEV;
 941                 goto error;
 942 #endif
 943         }
 944         if (config.devx) {
 945                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
 946
 947                 err = mlx5_devx_cmd_register_read
 948                         (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
 949                         reg, MLX5_ST_SZ_DW(register_mtutc));
 950                 if (!err) {
 951                         uint32_t ts_mode;
 952
 953                         /* MTUTC register is read successfully. */
 954                         ts_mode = MLX5_GET(register_mtutc, reg,
 955                                            time_stamp_mode);
 956                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
 957                                 config.rt_timestamp = 1;
 958                 } else {
 959                         /* Kernel does not support register reading. */
 960                         if (config.hca_attr.dev_freq_khz ==
 961                                                  (NS_PER_S / MS_PER_S))
 962                                 config.rt_timestamp = 1;
 963                 }
 964         }
 965         /*
 966          * If HW has bug working with tunnel packet decapsulation and
 967          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
 968          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
 969          */
 970         if (config.hca_attr.scatter_fcs_w_decap_disable && config.decap_en)
 971                 config.hw_fcs_strip = 0;
 972         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 973                 (config.hw_fcs_strip ? "" : "not "));
 974         if (config.mprq.enabled && mprq) {
 975                 if (config.mprq.stride_num_n &&
 976                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 977                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 978                         config.mprq.stride_num_n =
 979                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 980                                                 mprq_min_stride_num_n),
 981                                         mprq_max_stride_num_n);
 982                         DRV_LOG(WARNING,
 983                                 "the number of strides"
 984                                 " for Multi-Packet RQ is out of range,"
 985                                 " setting default value (%u)",
 986                                 1 << config.mprq.stride_num_n);
 987                 }
 988                 if (config.mprq.stride_size_n &&
 989                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 990                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 991                         config.mprq.stride_size_n =
 992                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 993                                                 mprq_min_stride_size_n),
 994                                         mprq_max_stride_size_n);
 995                         DRV_LOG(WARNING,
 996                                 "the size of a stride"
 997                                 " for Multi-Packet RQ is out of range,"
 998                                 " setting default value (%u)",
 999                                 1 << config.mprq.stride_size_n);
1000                 }
1001                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1002                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1003         } else if (config.mprq.enabled && !mprq) {
1004                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1005                 config.mprq.enabled = 0;
1006         }
1007         if (config.max_dump_files_num == 0)
1008                 config.max_dump_files_num = 128;
1009         eth_dev = rte_eth_dev_allocate(name);
1010         if (eth_dev == NULL) {
1011                 DRV_LOG(ERR, "can not allocate rte ethdev");
1012                 err = ENOMEM;
1013                 goto error;
1014         }
1015         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1016         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1017         if (priv->representor) {
1018                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1019                 eth_dev->data->representor_id = priv->representor_id;
1020         }
1021         /*
1022          * Store associated network device interface index. This index
1023          * is permanent throughout the lifetime of device. So, we may store
1024          * the ifindex here and use the cached value further.
1025          */
1026         MLX5_ASSERT(spawn->ifindex);
1027         priv->if_index = spawn->ifindex;
1028         eth_dev->data->dev_private = priv;
1029         priv->dev_data = eth_dev->data;
1030         eth_dev->data->mac_addrs = priv->mac;
1031         eth_dev->device = dpdk_dev;
1032         /* Configure the first MAC address by default. */
1033         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1034                 DRV_LOG(ERR,
1035                         "port %u cannot get MAC address, is mlx5_en"
1036                         " loaded? (errno: %s)",
1037                         eth_dev->data->port_id, strerror(rte_errno));
1038                 err = ENODEV;
1039                 goto error;
1040         }
1041         DRV_LOG(INFO,
1042                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1043                 eth_dev->data->port_id,
1044                 mac.addr_bytes[0], mac.addr_bytes[1],
1045                 mac.addr_bytes[2], mac.addr_bytes[3],
1046                 mac.addr_bytes[4], mac.addr_bytes[5]);
1047 #ifdef RTE_LIBRTE_MLX5_DEBUG
1048         {
1049                 char ifname[IF_NAMESIZE];
1050
1051                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1052                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1053                                 eth_dev->data->port_id, ifname);
1054                 else
1055                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1056                                 eth_dev->data->port_id);
1057         }
1058 #endif
1059         /* Get actual MTU if possible. */
1060         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1061         if (err) {
1062                 err = rte_errno;
1063                 goto error;
1064         }
1065         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1066                 priv->mtu);
1067         /* Initialize burst functions to prevent crashes before link-up. */
1068         eth_dev->rx_pkt_burst = removed_rx_burst;
1069         eth_dev->tx_pkt_burst = removed_tx_burst;
1070         eth_dev->dev_ops = &mlx5_os_dev_ops;
1071         /* Register MAC address. */
1072         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1073         if (config.vf && config.vf_nl_en)
1074                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1075                                       mlx5_ifindex(eth_dev),
1076                                       eth_dev->data->mac_addrs,
1077                                       MLX5_MAX_MAC_ADDRESSES);
1078         priv->flows = 0;
1079         priv->ctrl_flows = 0;
1080         TAILQ_INIT(&priv->flow_meters);
1081         TAILQ_INIT(&priv->flow_meter_profiles);
1082         /* Hint libmlx5 to use PMD allocator for data plane resources */
1083         mlx5_glue->dv_set_context_attr(sh->ctx,
1084                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1085                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1086                                 .alloc = &mlx5_alloc_verbs_buf,
1087                                 .free = &mlx5_free_verbs_buf,
1088                                 .data = priv,
1089                         }));
1090         /* Bring Ethernet device up. */
1091         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1092                 eth_dev->data->port_id);
1093         mlx5_set_link_up(eth_dev);
1094         /*
1095          * Even though the interrupt handler is not installed yet,
1096          * interrupts will still trigger on the async_fd from
1097          * Verbs context returned by ibv_open_device().
1098          */
1099         mlx5_link_update(eth_dev, 0);
1100 #ifdef HAVE_MLX5DV_DR_ESWITCH
1101         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1102               (switch_info->representor || switch_info->master)))
1103                 config.dv_esw_en = 0;
1104 #else
1105         config.dv_esw_en = 0;
1106 #endif
1107         /* Detect minimal data bytes to inline. */
1108         mlx5_set_min_inline(spawn, &config);
1109         /* Store device configuration on private structure. */
1110         priv->config = config;
1111         /* Create context for virtual machine VLAN workaround. */
1112         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1113         if (config.dv_flow_en) {
1114                 err = mlx5_alloc_shared_dr(priv);
1115                 if (err)
1116                         goto error;
1117                 /*
1118                  * RSS id is shared with meter flow id. Meter flow id can only
1119                  * use the 24 MSB of the register.
1120                  */
1121                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1122                                      MLX5_MTR_COLOR_BITS);
1123                 if (!priv->qrss_id_pool) {
1124                         DRV_LOG(ERR, "can't create flow id pool");
1125                         err = ENOMEM;
1126                         goto error;
1127                 }
1128         }
1129         /* Supported Verbs flow priority number detection. */
1130         err = mlx5_flow_discover_priorities(eth_dev);
1131         if (err < 0) {
1132                 err = -err;
1133                 goto error;
1134         }
1135         priv->config.flow_prio = err;
1136         if (!priv->config.dv_esw_en &&
1137             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1138                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1139                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1140                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1141         }
1142         mlx5_set_metadata_mask(eth_dev);
1143         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1144             !priv->sh->dv_regc0_mask) {
1145                 DRV_LOG(ERR, "metadata mode %u is not supported "
1146                              "(no metadata reg_c[0] is available)",
1147                              priv->config.dv_xmeta_en);
1148                         err = ENOTSUP;
1149                         goto error;
1150         }
1151         /*
1152          * Allocate the buffer for flow creating, just once.
1153          * The allocation must be done before any flow creating.
1154          */
1155         mlx5_flow_alloc_intermediate(eth_dev);
1156         /* Query availability of metadata reg_c's. */
1157         err = mlx5_flow_discover_mreg_c(eth_dev);
1158         if (err < 0) {
1159                 err = -err;
1160                 goto error;
1161         }
1162         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1163                 DRV_LOG(DEBUG,
1164                         "port %u extensive metadata register is not supported",
1165                         eth_dev->data->port_id);
1166                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1167                         DRV_LOG(ERR, "metadata mode %u is not supported "
1168                                      "(no metadata registers available)",
1169                                      priv->config.dv_xmeta_en);
1170                         err = ENOTSUP;
1171                         goto error;
1172                 }
1173         }
1174         if (priv->config.dv_flow_en &&
1175             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1176             mlx5_flow_ext_mreg_supported(eth_dev) &&
1177             priv->sh->dv_regc0_mask) {
1178                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1179                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1180                 if (!priv->mreg_cp_tbl) {
1181                         err = ENOMEM;
1182                         goto error;
1183                 }
1184         }
1185         return eth_dev;
1186 error:
1187         if (priv) {
1188                 if (priv->mreg_cp_tbl)
1189                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1190                 if (priv->sh)
1191                         mlx5_os_free_shared_dr(priv);
1192                 if (priv->nl_socket_route >= 0)
1193                         close(priv->nl_socket_route);
1194                 if (priv->nl_socket_rdma >= 0)
1195                         close(priv->nl_socket_rdma);
1196                 if (priv->vmwa_context)
1197                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1198                 if (priv->qrss_id_pool)
1199                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1200                 if (own_domain_id)
1201                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1202                 mlx5_free(priv);
1203                 if (eth_dev != NULL)
1204                         eth_dev->data->dev_private = NULL;
1205         }
1206         if (eth_dev != NULL) {
1207                 /* mac_addrs must not be freed alone because part of
1208                  * dev_private
1209                  **/
1210                 eth_dev->data->mac_addrs = NULL;
1211                 rte_eth_dev_release_port(eth_dev);
1212         }
1213         if (sh)
1214                 mlx5_free_shared_dev_ctx(sh);
1215         MLX5_ASSERT(err > 0);
1216         rte_errno = err;
1217         return NULL;
1218 }
1219
1220 /**
1221  * Comparison callback to sort device data.
1222  *
1223  * This is meant to be used with qsort().
1224  *
1225  * @param a[in]
1226  *   Pointer to pointer to first data object.
1227  * @param b[in]
1228  *   Pointer to pointer to second data object.
1229  *
1230  * @return
1231  *   0 if both objects are equal, less than 0 if the first argument is less
1232  *   than the second, greater than 0 otherwise.
1233  */
1234 static int
1235 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1236 {
1237         const struct mlx5_switch_info *si_a =
1238                 &((const struct mlx5_dev_spawn_data *)a)->info;
1239         const struct mlx5_switch_info *si_b =
1240                 &((const struct mlx5_dev_spawn_data *)b)->info;
1241         int ret;
1242
1243         /* Master device first. */
1244         ret = si_b->master - si_a->master;
1245         if (ret)
1246                 return ret;
1247         /* Then representor devices. */
1248         ret = si_b->representor - si_a->representor;
1249         if (ret)
1250                 return ret;
1251         /* Unidentified devices come last in no specific order. */
1252         if (!si_a->representor)
1253                 return 0;
1254         /* Order representors by name. */
1255         return si_a->port_name - si_b->port_name;
1256 }
1257
1258 /**
1259  * Match PCI information for possible slaves of bonding device.
1260  *
1261  * @param[in] ibv_dev
1262  *   Pointer to Infiniband device structure.
1263  * @param[in] pci_dev
1264  *   Pointer to PCI device structure to match PCI address.
1265  * @param[in] nl_rdma
1266  *   Netlink RDMA group socket handle.
1267  *
1268  * @return
1269  *   negative value if no bonding device found, otherwise
1270  *   positive index of slave PF in bonding.
1271  */
1272 static int
1273 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1274                            const struct rte_pci_device *pci_dev,
1275                            int nl_rdma)
1276 {
1277         char ifname[IF_NAMESIZE + 1];
1278         unsigned int ifindex;
1279         unsigned int np, i;
1280         FILE *file = NULL;
1281         int pf = -1;
1282
1283         /*
1284          * Try to get master device name. If something goes
1285          * wrong suppose the lack of kernel support and no
1286          * bonding devices.
1287          */
1288         if (nl_rdma < 0)
1289                 return -1;
1290         if (!strstr(ibv_dev->name, "bond"))
1291                 return -1;
1292         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1293         if (!np)
1294                 return -1;
1295         /*
1296          * The Master device might not be on the predefined
1297          * port (not on port index 1, it is not garanted),
1298          * we have to scan all Infiniband device port and
1299          * find master.
1300          */
1301         for (i = 1; i <= np; ++i) {
1302                 /* Check whether Infiniband port is populated. */
1303                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1304                 if (!ifindex)
1305                         continue;
1306                 if (!if_indextoname(ifindex, ifname))
1307                         continue;
1308                 /* Try to read bonding slave names from sysfs. */
1309                 MKSTR(slaves,
1310                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1311                 file = fopen(slaves, "r");
1312                 if (file)
1313                         break;
1314         }
1315         if (!file)
1316                 return -1;
1317         /* Use safe format to check maximal buffer length. */
1318         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1319         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1320                 char tmp_str[IF_NAMESIZE + 32];
1321                 struct rte_pci_addr pci_addr;
1322                 struct mlx5_switch_info info;
1323
1324                 /* Process slave interface names in the loop. */
1325                 snprintf(tmp_str, sizeof(tmp_str),
1326                          "/sys/class/net/%s", ifname);
1327                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1328                         DRV_LOG(WARNING, "can not get PCI address"
1329                                          " for netdev \"%s\"", ifname);
1330                         continue;
1331                 }
1332                 if (pci_dev->addr.domain != pci_addr.domain ||
1333                     pci_dev->addr.bus != pci_addr.bus ||
1334                     pci_dev->addr.devid != pci_addr.devid ||
1335                     pci_dev->addr.function != pci_addr.function)
1336                         continue;
1337                 /* Slave interface PCI address match found. */
1338                 fclose(file);
1339                 snprintf(tmp_str, sizeof(tmp_str),
1340                          "/sys/class/net/%s/phys_port_name", ifname);
1341                 file = fopen(tmp_str, "rb");
1342                 if (!file)
1343                         break;
1344                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1345                 if (fscanf(file, "%32s", tmp_str) == 1)
1346                         mlx5_translate_port_name(tmp_str, &info);
1347                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1348                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1349                         pf = info.port_name;
1350                 break;
1351         }
1352         if (file)
1353                 fclose(file);
1354         return pf;
1355 }
1356
1357 /**
1358  * DPDK callback to register a PCI device.
1359  *
1360  * This function spawns Ethernet devices out of a given PCI device.
1361  *
1362  * @param[in] pci_drv
1363  *   PCI driver structure (mlx5_driver).
1364  * @param[in] pci_dev
1365  *   PCI device information.
1366  *
1367  * @return
1368  *   0 on success, a negative errno value otherwise and rte_errno is set.
1369  */
1370 int
1371 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1372                   struct rte_pci_device *pci_dev)
1373 {
1374         struct ibv_device **ibv_list;
1375         /*
1376          * Number of found IB Devices matching with requested PCI BDF.
1377          * nd != 1 means there are multiple IB devices over the same
1378          * PCI device and we have representors and master.
1379          */
1380         unsigned int nd = 0;
1381         /*
1382          * Number of found IB device Ports. nd = 1 and np = 1..n means
1383          * we have the single multiport IB device, and there may be
1384          * representors attached to some of found ports.
1385          */
1386         unsigned int np = 0;
1387         /*
1388          * Number of DPDK ethernet devices to Spawn - either over
1389          * multiple IB devices or multiple ports of single IB device.
1390          * Actually this is the number of iterations to spawn.
1391          */
1392         unsigned int ns = 0;
1393         /*
1394          * Bonding device
1395          *   < 0 - no bonding device (single one)
1396          *  >= 0 - bonding device (value is slave PF index)
1397          */
1398         int bd = -1;
1399         struct mlx5_dev_spawn_data *list = NULL;
1400         struct mlx5_dev_config dev_config;
1401         int ret;
1402
1403         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1404                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1405                         " driver.");
1406                 return 1;
1407         }
1408         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1409                 mlx5_pmd_socket_init();
1410         ret = mlx5_init_once();
1411         if (ret) {
1412                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1413                         strerror(rte_errno));
1414                 return -rte_errno;
1415         }
1416         MLX5_ASSERT(pci_drv == &mlx5_driver);
1417         errno = 0;
1418         ibv_list = mlx5_glue->get_device_list(&ret);
1419         if (!ibv_list) {
1420                 rte_errno = errno ? errno : ENOSYS;
1421                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1422                 return -rte_errno;
1423         }
1424         /*
1425          * First scan the list of all Infiniband devices to find
1426          * matching ones, gathering into the list.
1427          */
1428         struct ibv_device *ibv_match[ret + 1];
1429         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1430         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1431         unsigned int i;
1432
1433         while (ret-- > 0) {
1434                 struct rte_pci_addr pci_addr;
1435
1436                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1437                 bd = mlx5_device_bond_pci_match
1438                                 (ibv_list[ret], pci_dev, nl_rdma);
1439                 if (bd >= 0) {
1440                         /*
1441                          * Bonding device detected. Only one match is allowed,
1442                          * the bonding is supported over multi-port IB device,
1443                          * there should be no matches on representor PCI
1444                          * functions or non VF LAG bonding devices with
1445                          * specified address.
1446                          */
1447                         if (nd) {
1448                                 DRV_LOG(ERR,
1449                                         "multiple PCI match on bonding device"
1450                                         "\"%s\" found", ibv_list[ret]->name);
1451                                 rte_errno = ENOENT;
1452                                 ret = -rte_errno;
1453                                 goto exit;
1454                         }
1455                         DRV_LOG(INFO, "PCI information matches for"
1456                                       " slave %d bonding device \"%s\"",
1457                                       bd, ibv_list[ret]->name);
1458                         ibv_match[nd++] = ibv_list[ret];
1459                         break;
1460                 }
1461                 if (mlx5_dev_to_pci_addr
1462                         (ibv_list[ret]->ibdev_path, &pci_addr))
1463                         continue;
1464                 if (pci_dev->addr.domain != pci_addr.domain ||
1465                     pci_dev->addr.bus != pci_addr.bus ||
1466                     pci_dev->addr.devid != pci_addr.devid ||
1467                     pci_dev->addr.function != pci_addr.function)
1468                         continue;
1469                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1470                         ibv_list[ret]->name);
1471                 ibv_match[nd++] = ibv_list[ret];
1472         }
1473         ibv_match[nd] = NULL;
1474         if (!nd) {
1475                 /* No device matches, just complain and bail out. */
1476                 DRV_LOG(WARNING,
1477                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1478                         " are kernel drivers loaded?",
1479                         pci_dev->addr.domain, pci_dev->addr.bus,
1480                         pci_dev->addr.devid, pci_dev->addr.function);
1481                 rte_errno = ENOENT;
1482                 ret = -rte_errno;
1483                 goto exit;
1484         }
1485         if (nd == 1) {
1486                 /*
1487                  * Found single matching device may have multiple ports.
1488                  * Each port may be representor, we have to check the port
1489                  * number and check the representors existence.
1490                  */
1491                 if (nl_rdma >= 0)
1492                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1493                 if (!np)
1494                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1495                                          " ports number", ibv_match[0]->name);
1496                 if (bd >= 0 && !np) {
1497                         DRV_LOG(ERR, "can not get ports"
1498                                      " for bonding device");
1499                         rte_errno = ENOENT;
1500                         ret = -rte_errno;
1501                         goto exit;
1502                 }
1503         }
1504 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1505         if (bd >= 0) {
1506                 /*
1507                  * This may happen if there is VF LAG kernel support and
1508                  * application is compiled with older rdma_core library.
1509                  */
1510                 DRV_LOG(ERR,
1511                         "No kernel/verbs support for VF LAG bonding found.");
1512                 rte_errno = ENOTSUP;
1513                 ret = -rte_errno;
1514                 goto exit;
1515         }
1516 #endif
1517         /*
1518          * Now we can determine the maximal
1519          * amount of devices to be spawned.
1520          */
1521         list = mlx5_malloc(MLX5_MEM_ZERO,
1522                            sizeof(struct mlx5_dev_spawn_data) *
1523                            (np ? np : nd),
1524                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1525         if (!list) {
1526                 DRV_LOG(ERR, "spawn data array allocation failure");
1527                 rte_errno = ENOMEM;
1528                 ret = -rte_errno;
1529                 goto exit;
1530         }
1531         if (bd >= 0 || np > 1) {
1532                 /*
1533                  * Single IB device with multiple ports found,
1534                  * it may be E-Switch master device and representors.
1535                  * We have to perform identification through the ports.
1536                  */
1537                 MLX5_ASSERT(nl_rdma >= 0);
1538                 MLX5_ASSERT(ns == 0);
1539                 MLX5_ASSERT(nd == 1);
1540                 MLX5_ASSERT(np);
1541                 for (i = 1; i <= np; ++i) {
1542                         list[ns].max_port = np;
1543                         list[ns].phys_port = i;
1544                         list[ns].phys_dev = ibv_match[0];
1545                         list[ns].eth_dev = NULL;
1546                         list[ns].pci_dev = pci_dev;
1547                         list[ns].pf_bond = bd;
1548                         list[ns].ifindex = mlx5_nl_ifindex
1549                                 (nl_rdma,
1550                                 mlx5_os_get_dev_device_name
1551                                                 (list[ns].phys_dev), i);
1552                         if (!list[ns].ifindex) {
1553                                 /*
1554                                  * No network interface index found for the
1555                                  * specified port, it means there is no
1556                                  * representor on this port. It's OK,
1557                                  * there can be disabled ports, for example
1558                                  * if sriov_numvfs < sriov_totalvfs.
1559                                  */
1560                                 continue;
1561                         }
1562                         ret = -1;
1563                         if (nl_route >= 0)
1564                                 ret = mlx5_nl_switch_info
1565                                                (nl_route,
1566                                                 list[ns].ifindex,
1567                                                 &list[ns].info);
1568                         if (ret || (!list[ns].info.representor &&
1569                                     !list[ns].info.master)) {
1570                                 /*
1571                                  * We failed to recognize representors with
1572                                  * Netlink, let's try to perform the task
1573                                  * with sysfs.
1574                                  */
1575                                 ret =  mlx5_sysfs_switch_info
1576                                                 (list[ns].ifindex,
1577                                                  &list[ns].info);
1578                         }
1579                         if (!ret && bd >= 0) {
1580                                 switch (list[ns].info.name_type) {
1581                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1582                                         if (list[ns].info.port_name == bd)
1583                                                 ns++;
1584                                         break;
1585                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1586                                         /* Fallthrough */
1587                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1588                                         if (list[ns].info.pf_num == bd)
1589                                                 ns++;
1590                                         break;
1591                                 default:
1592                                         break;
1593                                 }
1594                                 continue;
1595                         }
1596                         if (!ret && (list[ns].info.representor ^
1597                                      list[ns].info.master))
1598                                 ns++;
1599                 }
1600                 if (!ns) {
1601                         DRV_LOG(ERR,
1602                                 "unable to recognize master/representors"
1603                                 " on the IB device with multiple ports");
1604                         rte_errno = ENOENT;
1605                         ret = -rte_errno;
1606                         goto exit;
1607                 }
1608         } else {
1609                 /*
1610                  * The existence of several matching entries (nd > 1) means
1611                  * port representors have been instantiated. No existing Verbs
1612                  * call nor sysfs entries can tell them apart, this can only
1613                  * be done through Netlink calls assuming kernel drivers are
1614                  * recent enough to support them.
1615                  *
1616                  * In the event of identification failure through Netlink,
1617                  * try again through sysfs, then:
1618                  *
1619                  * 1. A single IB device matches (nd == 1) with single
1620                  *    port (np=0/1) and is not a representor, assume
1621                  *    no switch support.
1622                  *
1623                  * 2. Otherwise no safe assumptions can be made;
1624                  *    complain louder and bail out.
1625                  */
1626                 for (i = 0; i != nd; ++i) {
1627                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1628                         list[ns].max_port = 1;
1629                         list[ns].phys_port = 1;
1630                         list[ns].phys_dev = ibv_match[i];
1631                         list[ns].eth_dev = NULL;
1632                         list[ns].pci_dev = pci_dev;
1633                         list[ns].pf_bond = -1;
1634                         list[ns].ifindex = 0;
1635                         if (nl_rdma >= 0)
1636                                 list[ns].ifindex = mlx5_nl_ifindex
1637                                 (nl_rdma,
1638                                 mlx5_os_get_dev_device_name
1639                                                 (list[ns].phys_dev), 1);
1640                         if (!list[ns].ifindex) {
1641                                 char ifname[IF_NAMESIZE];
1642
1643                                 /*
1644                                  * Netlink failed, it may happen with old
1645                                  * ib_core kernel driver (before 4.16).
1646                                  * We can assume there is old driver because
1647                                  * here we are processing single ports IB
1648                                  * devices. Let's try sysfs to retrieve
1649                                  * the ifindex. The method works for
1650                                  * master device only.
1651                                  */
1652                                 if (nd > 1) {
1653                                         /*
1654                                          * Multiple devices found, assume
1655                                          * representors, can not distinguish
1656                                          * master/representor and retrieve
1657                                          * ifindex via sysfs.
1658                                          */
1659                                         continue;
1660                                 }
1661                                 ret = mlx5_get_ifname_sysfs
1662                                         (ibv_match[i]->ibdev_path, ifname);
1663                                 if (!ret)
1664                                         list[ns].ifindex =
1665                                                 if_nametoindex(ifname);
1666                                 if (!list[ns].ifindex) {
1667                                         /*
1668                                          * No network interface index found
1669                                          * for the specified device, it means
1670                                          * there it is neither representor
1671                                          * nor master.
1672                                          */
1673                                         continue;
1674                                 }
1675                         }
1676                         ret = -1;
1677                         if (nl_route >= 0)
1678                                 ret = mlx5_nl_switch_info
1679                                                (nl_route,
1680                                                 list[ns].ifindex,
1681                                                 &list[ns].info);
1682                         if (ret || (!list[ns].info.representor &&
1683                                     !list[ns].info.master)) {
1684                                 /*
1685                                  * We failed to recognize representors with
1686                                  * Netlink, let's try to perform the task
1687                                  * with sysfs.
1688                                  */
1689                                 ret =  mlx5_sysfs_switch_info
1690                                                 (list[ns].ifindex,
1691                                                  &list[ns].info);
1692                         }
1693                         if (!ret && (list[ns].info.representor ^
1694                                      list[ns].info.master)) {
1695                                 ns++;
1696                         } else if ((nd == 1) &&
1697                                    !list[ns].info.representor &&
1698                                    !list[ns].info.master) {
1699                                 /*
1700                                  * Single IB device with
1701                                  * one physical port and
1702                                  * attached network device.
1703                                  * May be SRIOV is not enabled
1704                                  * or there is no representors.
1705                                  */
1706                                 DRV_LOG(INFO, "no E-Switch support detected");
1707                                 ns++;
1708                                 break;
1709                         }
1710                 }
1711                 if (!ns) {
1712                         DRV_LOG(ERR,
1713                                 "unable to recognize master/representors"
1714                                 " on the multiple IB devices");
1715                         rte_errno = ENOENT;
1716                         ret = -rte_errno;
1717                         goto exit;
1718                 }
1719         }
1720         MLX5_ASSERT(ns);
1721         /*
1722          * Sort list to probe devices in natural order for users convenience
1723          * (i.e. master first, then representors from lowest to highest ID).
1724          */
1725         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1726         /* Default configuration. */
1727         dev_config = (struct mlx5_dev_config){
1728                 .hw_padding = 0,
1729                 .mps = MLX5_ARG_UNSET,
1730                 .dbnc = MLX5_ARG_UNSET,
1731                 .rx_vec_en = 1,
1732                 .txq_inline_max = MLX5_ARG_UNSET,
1733                 .txq_inline_min = MLX5_ARG_UNSET,
1734                 .txq_inline_mpw = MLX5_ARG_UNSET,
1735                 .txqs_inline = MLX5_ARG_UNSET,
1736                 .vf_nl_en = 1,
1737                 .mr_ext_memseg_en = 1,
1738                 .mprq = {
1739                         .enabled = 0, /* Disabled by default. */
1740                         .stride_num_n = 0,
1741                         .stride_size_n = 0,
1742                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1743                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1744                 },
1745                 .dv_esw_en = 1,
1746                 .dv_flow_en = 1,
1747                 .decap_en = 1,
1748                 .log_hp_size = MLX5_ARG_UNSET,
1749         };
1750         /* Device specific configuration. */
1751         switch (pci_dev->id.device_id) {
1752         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1753         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1754         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1755         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1756         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1757         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1758         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1759                 dev_config.vf = 1;
1760                 break;
1761         default:
1762                 break;
1763         }
1764         for (i = 0; i != ns; ++i) {
1765                 uint32_t restore;
1766
1767                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1768                                                  &list[i],
1769                                                  dev_config);
1770                 if (!list[i].eth_dev) {
1771                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1772                                 break;
1773                         /* Device is disabled or already spawned. Ignore it. */
1774                         continue;
1775                 }
1776                 restore = list[i].eth_dev->data->dev_flags;
1777                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1778                 /* Restore non-PCI flags cleared by the above call. */
1779                 list[i].eth_dev->data->dev_flags |= restore;
1780                 rte_eth_dev_probing_finish(list[i].eth_dev);
1781         }
1782         if (i != ns) {
1783                 DRV_LOG(ERR,
1784                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1785                         " encountering an error: %s",
1786                         pci_dev->addr.domain, pci_dev->addr.bus,
1787                         pci_dev->addr.devid, pci_dev->addr.function,
1788                         strerror(rte_errno));
1789                 ret = -rte_errno;
1790                 /* Roll back. */
1791                 while (i--) {
1792                         if (!list[i].eth_dev)
1793                                 continue;
1794                         mlx5_dev_close(list[i].eth_dev);
1795                         /* mac_addrs must not be freed because in dev_private */
1796                         list[i].eth_dev->data->mac_addrs = NULL;
1797                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1798                 }
1799                 /* Restore original error. */
1800                 rte_errno = -ret;
1801         } else {
1802                 ret = 0;
1803         }
1804 exit:
1805         /*
1806          * Do the routine cleanup:
1807          * - close opened Netlink sockets
1808          * - free allocated spawn data array
1809          * - free the Infiniband device list
1810          */
1811         if (nl_rdma >= 0)
1812                 close(nl_rdma);
1813         if (nl_route >= 0)
1814                 close(nl_route);
1815         if (list)
1816                 mlx5_free(list);
1817         MLX5_ASSERT(ibv_list);
1818         mlx5_glue->free_device_list(ibv_list);
1819         return ret;
1820 }
1821
1822 static int
1823 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1824 {
1825         char *env;
1826         int value;
1827
1828         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1829         /* Get environment variable to store. */
1830         env = getenv(MLX5_SHUT_UP_BF);
1831         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1832         if (config->dbnc == MLX5_ARG_UNSET)
1833                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1834         else
1835                 setenv(MLX5_SHUT_UP_BF,
1836                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1837         return value;
1838 }
1839
1840 static void
1841 mlx5_restore_doorbell_mapping_env(int value)
1842 {
1843         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1844         /* Restore the original environment variable state. */
1845         if (value == MLX5_ARG_UNSET)
1846                 unsetenv(MLX5_SHUT_UP_BF);
1847         else
1848                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1849 }
1850
1851 /**
1852  * Extract pdn of PD object using DV API.
1853  *
1854  * @param[in] pd
1855  *   Pointer to the verbs PD object.
1856  * @param[out] pdn
1857  *   Pointer to the PD object number variable.
1858  *
1859  * @return
1860  *   0 on success, error value otherwise.
1861  */
1862 int
1863 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1864 {
1865 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1866         struct mlx5dv_obj obj;
1867         struct mlx5dv_pd pd_info;
1868         int ret = 0;
1869
1870         obj.pd.in = pd;
1871         obj.pd.out = &pd_info;
1872         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1873         if (ret) {
1874                 DRV_LOG(DEBUG, "Fail to get PD object info");
1875                 return ret;
1876         }
1877         *pdn = pd_info.pdn;
1878         return 0;
1879 #else
1880         (void)pd;
1881         (void)pdn;
1882         return -ENOTSUP;
1883 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1884 }
1885
1886 /**
1887  * Function API to open IB device.
1888  *
1889  * This function calls the Linux glue APIs to open a device.
1890  *
1891  * @param[in] spawn
1892  *   Pointer to the IB device attributes (name, port, etc).
1893  * @param[out] config
1894  *   Pointer to device configuration structure.
1895  * @param[out] sh
1896  *   Pointer to shared context structure.
1897  *
1898  * @return
1899  *   0 on success, a positive error value otherwise.
1900  */
1901 int
1902 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1903                      const struct mlx5_dev_config *config,
1904                      struct mlx5_dev_ctx_shared *sh)
1905 {
1906         int dbmap_env;
1907         int err = 0;
1908
1909         sh->numa_node = spawn->pci_dev->device.numa_node;
1910         pthread_mutex_init(&sh->txpp.mutex, NULL);
1911         /*
1912          * Configure environment variable "MLX5_BF_SHUT_UP"
1913          * before the device creation. The rdma_core library
1914          * checks the variable at device creation and
1915          * stores the result internally.
1916          */
1917         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1918         /* Try to open IB device with DV first, then usual Verbs. */
1919         errno = 0;
1920         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1921         if (sh->ctx) {
1922                 sh->devx = 1;
1923                 DRV_LOG(DEBUG, "DevX is supported");
1924                 /* The device is created, no need for environment. */
1925                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1926         } else {
1927                 /* The environment variable is still configured. */
1928                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1929                 err = errno ? errno : ENODEV;
1930                 /*
1931                  * The environment variable is not needed anymore,
1932                  * all device creation attempts are completed.
1933                  */
1934                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1935                 if (!sh->ctx)
1936                         return err;
1937                 DRV_LOG(DEBUG, "DevX is NOT supported");
1938                 err = 0;
1939         }
1940         return err;
1941 }
1942
1943 /**
1944  * Install shared asynchronous device events handler.
1945  * This function is implemented to support event sharing
1946  * between multiple ports of single IB device.
1947  *
1948  * @param sh
1949  *   Pointer to mlx5_dev_ctx_shared object.
1950  */
1951 void
1952 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1953 {
1954         int ret;
1955         int flags;
1956
1957         sh->intr_handle.fd = -1;
1958         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1959         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1960                     F_SETFL, flags | O_NONBLOCK);
1961         if (ret) {
1962                 DRV_LOG(INFO, "failed to change file descriptor async event"
1963                         " queue");
1964         } else {
1965                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1966                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1967                 if (rte_intr_callback_register(&sh->intr_handle,
1968                                         mlx5_dev_interrupt_handler, sh)) {
1969                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1970                         sh->intr_handle.fd = -1;
1971                 }
1972         }
1973         if (sh->devx) {
1974 #ifdef HAVE_IBV_DEVX_ASYNC
1975                 sh->intr_handle_devx.fd = -1;
1976                 sh->devx_comp =
1977                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1978                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1979                 if (!devx_comp) {
1980                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1981                         return;
1982                 }
1983                 flags = fcntl(devx_comp->fd, F_GETFL);
1984                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1985                 if (ret) {
1986                         DRV_LOG(INFO, "failed to change file descriptor"
1987                                 " devx comp");
1988                         return;
1989                 }
1990                 sh->intr_handle_devx.fd = devx_comp->fd;
1991                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1992                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1993                                         mlx5_dev_interrupt_handler_devx, sh)) {
1994                         DRV_LOG(INFO, "Fail to install the devx shared"
1995                                 " interrupt.");
1996                         sh->intr_handle_devx.fd = -1;
1997                 }
1998 #endif /* HAVE_IBV_DEVX_ASYNC */
1999         }
2000 }
2001
2002 /**
2003  * Uninstall shared asynchronous device events handler.
2004  * This function is implemented to support event sharing
2005  * between multiple ports of single IB device.
2006  *
2007  * @param dev
2008  *   Pointer to mlx5_dev_ctx_shared object.
2009  */
2010 void
2011 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2012 {
2013         if (sh->intr_handle.fd >= 0)
2014                 mlx5_intr_callback_unregister(&sh->intr_handle,
2015                                               mlx5_dev_interrupt_handler, sh);
2016 #ifdef HAVE_IBV_DEVX_ASYNC
2017         if (sh->intr_handle_devx.fd >= 0)
2018                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2019                                   mlx5_dev_interrupt_handler_devx, sh);
2020         if (sh->devx_comp)
2021                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2022 #endif
2023 }
2024
2025 /**
2026  * Read statistics by a named counter.
2027  *
2028  * @param[in] priv
2029  *   Pointer to the private device data structure.
2030  * @param[in] ctr_name
2031  *   Pointer to the name of the statistic counter to read
2032  * @param[out] stat
2033  *   Pointer to read statistic value.
2034  * @return
2035  *   0 on success and stat is valud, 1 if failed to read the value
2036  *   rte_errno is set.
2037  *
2038  */
2039 int
2040 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2041                       uint64_t *stat)
2042 {
2043         int fd;
2044
2045         if (priv->sh) {
2046                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2047                           priv->sh->ibdev_path,
2048                           priv->dev_port,
2049                           ctr_name);
2050                 fd = open(path, O_RDONLY);
2051                 if (fd != -1) {
2052                         char buf[21] = {'\0'};
2053                         ssize_t n = read(fd, buf, sizeof(buf));
2054
2055                         close(fd);
2056                         if (n != -1) {
2057                                 *stat = strtoull(buf, NULL, 10);
2058                                 return 0;
2059                         }
2060                 }
2061         }
2062         *stat = 0;
2063         return 1;
2064 }
2065
2066 /**
2067  * Set the reg_mr and dereg_mr call backs
2068  *
2069  * @param reg_mr_cb[out]
2070  *   Pointer to reg_mr func
2071  * @param dereg_mr_cb[out]
2072  *   Pointer to dereg_mr func
2073  *
2074  */
2075 void
2076 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2077                       mlx5_dereg_mr_t *dereg_mr_cb)
2078 {
2079         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2080         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2081 }
2082
2083 /**
2084  * Remove a MAC address from device
2085  *
2086  * @param dev
2087  *   Pointer to Ethernet device structure.
2088  * @param index
2089  *   MAC address index.
2090  */
2091 void
2092 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2093 {
2094         struct mlx5_priv *priv = dev->data->dev_private;
2095         const int vf = priv->config.vf;
2096
2097         if (vf)
2098                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2099                                         mlx5_ifindex(dev), priv->mac_own,
2100                                         &dev->data->mac_addrs[index], index);
2101 }
2102
2103 /**
2104  * Adds a MAC address to the device
2105  *
2106  * @param dev
2107  *   Pointer to Ethernet device structure.
2108  * @param mac_addr
2109  *   MAC address to register.
2110  * @param index
2111  *   MAC address index.
2112  *
2113  * @return
2114  *   0 on success, a negative errno value otherwise
2115  */
2116 int
2117 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2118                      uint32_t index)
2119 {
2120         struct mlx5_priv *priv = dev->data->dev_private;
2121         const int vf = priv->config.vf;
2122         int ret = 0;
2123
2124         if (vf)
2125                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2126                                            mlx5_ifindex(dev), priv->mac_own,
2127                                            mac, index);
2128         return ret;
2129 }
2130
2131 /**
2132  * Modify a VF MAC address
2133  *
2134  * @param priv
2135  *   Pointer to device private data.
2136  * @param mac_addr
2137  *   MAC address to modify into.
2138  * @param iface_idx
2139  *   Net device interface index
2140  * @param vf_index
2141  *   VF index
2142  *
2143  * @return
2144  *   0 on success, a negative errno value otherwise
2145  */
2146 int
2147 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2148                            unsigned int iface_idx,
2149                            struct rte_ether_addr *mac_addr,
2150                            int vf_index)
2151 {
2152         return mlx5_nl_vf_mac_addr_modify
2153                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2154 }
2155
2156 /**
2157  * Set device promiscuous mode
2158  *
2159  * @param dev
2160  *   Pointer to Ethernet device structure.
2161  * @param enable
2162  *   0 - promiscuous is disabled, otherwise - enabled
2163  *
2164  * @return
2165  *   0 on success, a negative error value otherwise
2166  */
2167 int
2168 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2169 {
2170         struct mlx5_priv *priv = dev->data->dev_private;
2171
2172         return mlx5_nl_promisc(priv->nl_socket_route,
2173                                mlx5_ifindex(dev), !!enable);
2174 }
2175
2176 /**
2177  * Set device promiscuous mode
2178  *
2179  * @param dev
2180  *   Pointer to Ethernet device structure.
2181  * @param enable
2182  *   0 - all multicase is disabled, otherwise - enabled
2183  *
2184  * @return
2185  *   0 on success, a negative error value otherwise
2186  */
2187 int
2188 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2189 {
2190         struct mlx5_priv *priv = dev->data->dev_private;
2191
2192         return mlx5_nl_allmulti(priv->nl_socket_route,
2193                                 mlx5_ifindex(dev), !!enable);
2194 }
2195
2196 const struct eth_dev_ops mlx5_os_dev_ops = {
2197         .dev_configure = mlx5_dev_configure,
2198         .dev_start = mlx5_dev_start,
2199         .dev_stop = mlx5_dev_stop,
2200         .dev_set_link_down = mlx5_set_link_down,
2201         .dev_set_link_up = mlx5_set_link_up,
2202         .dev_close = mlx5_dev_close,
2203         .promiscuous_enable = mlx5_promiscuous_enable,
2204         .promiscuous_disable = mlx5_promiscuous_disable,
2205         .allmulticast_enable = mlx5_allmulticast_enable,
2206         .allmulticast_disable = mlx5_allmulticast_disable,
2207         .link_update = mlx5_link_update,
2208         .stats_get = mlx5_stats_get,
2209         .stats_reset = mlx5_stats_reset,
2210         .xstats_get = mlx5_xstats_get,
2211         .xstats_reset = mlx5_xstats_reset,
2212         .xstats_get_names = mlx5_xstats_get_names,
2213         .fw_version_get = mlx5_fw_version_get,
2214         .dev_infos_get = mlx5_dev_infos_get,
2215         .read_clock = mlx5_txpp_read_clock,
2216         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2217         .vlan_filter_set = mlx5_vlan_filter_set,
2218         .rx_queue_setup = mlx5_rx_queue_setup,
2219         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2220         .tx_queue_setup = mlx5_tx_queue_setup,
2221         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2222         .rx_queue_release = mlx5_rx_queue_release,
2223         .tx_queue_release = mlx5_tx_queue_release,
2224         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2225         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2226         .mac_addr_remove = mlx5_mac_addr_remove,
2227         .mac_addr_add = mlx5_mac_addr_add,
2228         .mac_addr_set = mlx5_mac_addr_set,
2229         .set_mc_addr_list = mlx5_set_mc_addr_list,
2230         .mtu_set = mlx5_dev_set_mtu,
2231         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2232         .vlan_offload_set = mlx5_vlan_offload_set,
2233         .reta_update = mlx5_dev_rss_reta_update,
2234         .reta_query = mlx5_dev_rss_reta_query,
2235         .rss_hash_update = mlx5_rss_hash_update,
2236         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2237         .filter_ctrl = mlx5_dev_filter_ctrl,
2238         .rx_descriptor_status = mlx5_rx_descriptor_status,
2239         .tx_descriptor_status = mlx5_tx_descriptor_status,
2240         .rxq_info_get = mlx5_rxq_info_get,
2241         .txq_info_get = mlx5_txq_info_get,
2242         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2243         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2244         .rx_queue_count = mlx5_rx_queue_count,
2245         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2246         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2247         .is_removed = mlx5_is_removed,
2248         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2249         .get_module_info = mlx5_get_module_info,
2250         .get_module_eeprom = mlx5_get_module_eeprom,
2251         .hairpin_cap_get = mlx5_hairpin_cap_get,
2252         .mtr_ops_get = mlx5_flow_meter_ops_get,
2253 };
2254
2255 /* Available operations from secondary process. */
2256 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2257         .stats_get = mlx5_stats_get,
2258         .stats_reset = mlx5_stats_reset,
2259         .xstats_get = mlx5_xstats_get,
2260         .xstats_reset = mlx5_xstats_reset,
2261         .xstats_get_names = mlx5_xstats_get_names,
2262         .fw_version_get = mlx5_fw_version_get,
2263         .dev_infos_get = mlx5_dev_infos_get,
2264         .read_clock = mlx5_txpp_read_clock,
2265         .rx_descriptor_status = mlx5_rx_descriptor_status,
2266         .tx_descriptor_status = mlx5_tx_descriptor_status,
2267         .rxq_info_get = mlx5_rxq_info_get,
2268         .txq_info_get = mlx5_txq_info_get,
2269         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2270         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2271         .get_module_info = mlx5_get_module_info,
2272         .get_module_eeprom = mlx5_get_module_eeprom,
2273 };
2274
2275 /* Available operations in flow isolated mode. */
2276 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2277         .dev_configure = mlx5_dev_configure,
2278         .dev_start = mlx5_dev_start,
2279         .dev_stop = mlx5_dev_stop,
2280         .dev_set_link_down = mlx5_set_link_down,
2281         .dev_set_link_up = mlx5_set_link_up,
2282         .dev_close = mlx5_dev_close,
2283         .promiscuous_enable = mlx5_promiscuous_enable,
2284         .promiscuous_disable = mlx5_promiscuous_disable,
2285         .allmulticast_enable = mlx5_allmulticast_enable,
2286         .allmulticast_disable = mlx5_allmulticast_disable,
2287         .link_update = mlx5_link_update,
2288         .stats_get = mlx5_stats_get,
2289         .stats_reset = mlx5_stats_reset,
2290         .xstats_get = mlx5_xstats_get,
2291         .xstats_reset = mlx5_xstats_reset,
2292         .xstats_get_names = mlx5_xstats_get_names,
2293         .fw_version_get = mlx5_fw_version_get,
2294         .dev_infos_get = mlx5_dev_infos_get,
2295         .read_clock = mlx5_txpp_read_clock,
2296         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2297         .vlan_filter_set = mlx5_vlan_filter_set,
2298         .rx_queue_setup = mlx5_rx_queue_setup,
2299         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2300         .tx_queue_setup = mlx5_tx_queue_setup,
2301         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2302         .rx_queue_release = mlx5_rx_queue_release,
2303         .tx_queue_release = mlx5_tx_queue_release,
2304         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2305         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2306         .mac_addr_remove = mlx5_mac_addr_remove,
2307         .mac_addr_add = mlx5_mac_addr_add,
2308         .mac_addr_set = mlx5_mac_addr_set,
2309         .set_mc_addr_list = mlx5_set_mc_addr_list,
2310         .mtu_set = mlx5_dev_set_mtu,
2311         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2312         .vlan_offload_set = mlx5_vlan_offload_set,
2313         .filter_ctrl = mlx5_dev_filter_ctrl,
2314         .rx_descriptor_status = mlx5_rx_descriptor_status,
2315         .tx_descriptor_status = mlx5_tx_descriptor_status,
2316         .rxq_info_get = mlx5_rxq_info_get,
2317         .txq_info_get = mlx5_txq_info_get,
2318         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2319         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2320         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2321         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2322         .is_removed = mlx5_is_removed,
2323         .get_module_info = mlx5_get_module_info,
2324         .get_module_eeprom = mlx5_get_module_eeprom,
2325         .hairpin_cap_get = mlx5_hairpin_cap_get,
2326         .mtr_ops_get = mlx5_flow_meter_ops_get,
2327 };