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 #include <rte_malloc.h>
  19 #include <rte_ethdev_driver.h>
  20 #include <rte_ethdev_pci.h>
  21 #include <rte_pci.h>
  22 #include <rte_bus_pci.h>
  23 #include <rte_common.h>
  24 #include <rte_kvargs.h>
  25 #include <rte_rwlock.h>
  26 #include <rte_spinlock.h>
  27 #include <rte_string_fns.h>
  28 #include <rte_alarm.h>
  29 #include <rte_eal_paging.h>
  30
  31 #include <mlx5_glue.h>
  32 #include <mlx5_devx_cmds.h>
  33 #include <mlx5_common.h>
  34 #include <mlx5_common_mp.h>
  35 #include <mlx5_common_mr.h>
  36 #include <mlx5_malloc.h>
  37
  38 #include "mlx5_defs.h"
  39 #include "mlx5.h"
  40 #include "mlx5_common_os.h"
  41 #include "mlx5_utils.h"
  42 #include "mlx5_rxtx.h"
  43 #include "mlx5_autoconf.h"
  44 #include "mlx5_mr.h"
  45 #include "mlx5_flow.h"
  46 #include "rte_pmd_mlx5.h"
  47 #include "mlx5_verbs.h"
  48
  49 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  50
  51 #ifndef HAVE_IBV_MLX5_MOD_MPW
  52 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  53 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  54 #endif
  55
  56 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  57 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  58 #endif
  59
  60 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
  61
  62 /* Spinlock for mlx5_shared_data allocation. */
  63 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
  64
  65 /* Process local data for secondary processes. */
  66 static struct mlx5_local_data mlx5_local_data;
  67
  68 /**
  69  * Set the completion channel file descriptor interrupt as non-blocking.
  70  *
  71  * @param[in] rxq_obj
  72  *   Pointer to RQ channel object, which includes the channel fd
  73  *
  74  * @param[out] fd
  75  *   The file descriptor (representing the intetrrupt) used in this channel.
  76  *
  77  * @return
  78  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
  79  */
  80 int
  81 mlx5_os_set_nonblock_channel_fd(int fd)
  82 {
  83         int flags;
  84
  85         flags = fcntl(fd, F_GETFL);
  86         return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
  87 }
  88
  89 /**
  90  * Get mlx5 device attributes. The glue function query_device_ex() is called
  91  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  92  * device attributes from the glue out parameter.
  93  *
  94  * @param dev
  95  *   Pointer to ibv context.
  96  *
  97  * @param device_attr
  98  *   Pointer to mlx5 device attributes.
  99  *
 100  * @return
 101  *   0 on success, non zero error number otherwise
 102  */
 103 int
 104 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
 105 {
 106         int err;
 107         struct ibv_device_attr_ex attr_ex;
 108         memset(device_attr, 0, sizeof(*device_attr));
 109         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
 110         if (err)
 111                 return err;
 112
 113         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
 114         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
 115         device_attr->max_sge = attr_ex.orig_attr.max_sge;
 116         device_attr->max_cq = attr_ex.orig_attr.max_cq;
 117         device_attr->max_qp = attr_ex.orig_attr.max_qp;
 118         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 119         device_attr->max_rwq_indirection_table_size =
 120                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 121         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 122         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 123
 124         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 125         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 126         if (err)
 127                 return err;
 128
 129         device_attr->flags = dv_attr.flags;
 130         device_attr->comp_mask = dv_attr.comp_mask;
 131 #ifdef HAVE_IBV_MLX5_MOD_SWP
 132         device_attr->sw_parsing_offloads =
 133                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 134 #endif
 135         device_attr->min_single_stride_log_num_of_bytes =
 136                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 137         device_attr->max_single_stride_log_num_of_bytes =
 138                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 139         device_attr->min_single_wqe_log_num_of_strides =
 140                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 141         device_attr->max_single_wqe_log_num_of_strides =
 142                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 143         device_attr->stride_supported_qpts =
 144                 dv_attr.striding_rq_caps.supported_qpts;
 145 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 146         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 147 #endif
 148
 149         return err;
 150 }
 151
 152 /**
 153  * Verbs callback to allocate a memory. This function should allocate the space
 154  * according to the size provided residing inside a huge page.
 155  * Please note that all allocation must respect the alignment from libmlx5
 156  * (i.e. currently rte_mem_page_size()).
 157  *
 158  * @param[in] size
 159  *   The size in bytes of the memory to allocate.
 160  * @param[in] data
 161  *   A pointer to the callback data.
 162  *
 163  * @return
 164  *   Allocated buffer, NULL otherwise and rte_errno is set.
 165  */
 166 static void *
 167 mlx5_alloc_verbs_buf(size_t size, void *data)
 168 {
 169         struct mlx5_priv *priv = data;
 170         void *ret;
 171         unsigned int socket = SOCKET_ID_ANY;
 172         size_t alignment = rte_mem_page_size();
 173         if (alignment == (size_t)-1) {
 174                 DRV_LOG(ERR, "Failed to get mem page size");
 175                 rte_errno = ENOMEM;
 176                 return NULL;
 177         }
 178
 179         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 180                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 181
 182                 socket = ctrl->socket;
 183         } else if (priv->verbs_alloc_ctx.type ==
 184                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 185                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 186
 187                 socket = ctrl->socket;
 188         }
 189         MLX5_ASSERT(data != NULL);
 190         ret = mlx5_malloc(0, size, alignment, socket);
 191         if (!ret && size)
 192                 rte_errno = ENOMEM;
 193         return ret;
 194 }
 195
 196 /**
 197  * Verbs callback to free a memory.
 198  *
 199  * @param[in] ptr
 200  *   A pointer to the memory to free.
 201  * @param[in] data
 202  *   A pointer to the callback data.
 203  */
 204 static void
 205 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 206 {
 207         MLX5_ASSERT(data != NULL);
 208         mlx5_free(ptr);
 209 }
 210
 211 /**
 212  * Initialize DR related data within private structure.
 213  * Routine checks the reference counter and does actual
 214  * resources creation/initialization only if counter is zero.
 215  *
 216  * @param[in] priv
 217  *   Pointer to the private device data structure.
 218  *
 219  * @return
 220  *   Zero on success, positive error code otherwise.
 221  */
 222 static int
 223 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 224 {
 225         struct mlx5_dev_ctx_shared *sh = priv->sh;
 226         char s[MLX5_HLIST_NAMESIZE];
 227         int err = 0;
 228
 229         if (!sh->flow_tbls)
 230                 err = mlx5_alloc_table_hash_list(priv);
 231         else
 232                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 233                         (void *)sh->flow_tbls);
 234         if (err)
 235                 return err;
 236         /* Create tags hash list table. */
 237         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 238         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 239         if (!sh->tag_table) {
 240                 DRV_LOG(ERR, "tags with hash creation failed.");
 241                 err = ENOMEM;
 242                 goto error;
 243         }
 244 #ifdef HAVE_MLX5DV_DR
 245         void *domain;
 246
 247         if (sh->dv_refcnt) {
 248                 /* Shared DV/DR structures is already initialized. */
 249                 sh->dv_refcnt++;
 250                 priv->dr_shared = 1;
 251                 return 0;
 252         }
 253         /* Reference counter is zero, we should initialize structures. */
 254         domain = mlx5_glue->dr_create_domain(sh->ctx,
 255                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 256         if (!domain) {
 257                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 258                 err = errno;
 259                 goto error;
 260         }
 261         sh->rx_domain = domain;
 262         domain = mlx5_glue->dr_create_domain(sh->ctx,
 263                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 264         if (!domain) {
 265                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 266                 err = errno;
 267                 goto error;
 268         }
 269         pthread_mutex_init(&sh->dv_mutex, NULL);
 270         sh->tx_domain = domain;
 271 #ifdef HAVE_MLX5DV_DR_ESWITCH
 272         if (priv->config.dv_esw_en) {
 273                 domain  = mlx5_glue->dr_create_domain
 274                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 275                 if (!domain) {
 276                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 277                         err = errno;
 278                         goto error;
 279                 }
 280                 sh->fdb_domain = domain;
 281                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 282         }
 283 #endif
 284         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 285                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 286                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 287                 if (sh->fdb_domain)
 288                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 289         }
 290         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 291 #endif /* HAVE_MLX5DV_DR */
 292         sh->dv_refcnt++;
 293         priv->dr_shared = 1;
 294         return 0;
 295 error:
 296         /* Rollback the created objects. */
 297         if (sh->rx_domain) {
 298                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 299                 sh->rx_domain = NULL;
 300         }
 301         if (sh->tx_domain) {
 302                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 303                 sh->tx_domain = NULL;
 304         }
 305         if (sh->fdb_domain) {
 306                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 307                 sh->fdb_domain = NULL;
 308         }
 309         if (sh->esw_drop_action) {
 310                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 311                 sh->esw_drop_action = NULL;
 312         }
 313         if (sh->pop_vlan_action) {
 314                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 315                 sh->pop_vlan_action = NULL;
 316         }
 317         if (sh->tag_table) {
 318                 /* tags should be destroyed with flow before. */
 319                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 320                 sh->tag_table = NULL;
 321         }
 322         mlx5_free_table_hash_list(priv);
 323         return err;
 324 }
 325
 326 /**
 327  * Destroy DR related data within private structure.
 328  *
 329  * @param[in] priv
 330  *   Pointer to the private device data structure.
 331  */
 332 void
 333 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 334 {
 335         struct mlx5_dev_ctx_shared *sh;
 336
 337         if (!priv->dr_shared)
 338                 return;
 339         priv->dr_shared = 0;
 340         sh = priv->sh;
 341         MLX5_ASSERT(sh);
 342 #ifdef HAVE_MLX5DV_DR
 343         MLX5_ASSERT(sh->dv_refcnt);
 344         if (sh->dv_refcnt && --sh->dv_refcnt)
 345                 return;
 346         if (sh->rx_domain) {
 347                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 348                 sh->rx_domain = NULL;
 349         }
 350         if (sh->tx_domain) {
 351                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 352                 sh->tx_domain = NULL;
 353         }
 354 #ifdef HAVE_MLX5DV_DR_ESWITCH
 355         if (sh->fdb_domain) {
 356                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 357                 sh->fdb_domain = NULL;
 358         }
 359         if (sh->esw_drop_action) {
 360                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 361                 sh->esw_drop_action = NULL;
 362         }
 363 #endif
 364         if (sh->pop_vlan_action) {
 365                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 366                 sh->pop_vlan_action = NULL;
 367         }
 368         pthread_mutex_destroy(&sh->dv_mutex);
 369 #endif /* HAVE_MLX5DV_DR */
 370         if (sh->tag_table) {
 371                 /* tags should be destroyed with flow before. */
 372                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 373                 sh->tag_table = NULL;
 374         }
 375         mlx5_free_table_hash_list(priv);
 376 }
 377
 378 /**
 379  * Initialize shared data between primary and secondary process.
 380  *
 381  * A memzone is reserved by primary process and secondary processes attach to
 382  * the memzone.
 383  *
 384  * @return
 385  *   0 on success, a negative errno value otherwise and rte_errno is set.
 386  */
 387 static int
 388 mlx5_init_shared_data(void)
 389 {
 390         const struct rte_memzone *mz;
 391         int ret = 0;
 392
 393         rte_spinlock_lock(&mlx5_shared_data_lock);
 394         if (mlx5_shared_data == NULL) {
 395                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 396                         /* Allocate shared memory. */
 397                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 398                                                  sizeof(*mlx5_shared_data),
 399                                                  SOCKET_ID_ANY, 0);
 400                         if (mz == NULL) {
 401                                 DRV_LOG(ERR,
 402                                         "Cannot allocate mlx5 shared data");
 403                                 ret = -rte_errno;
 404                                 goto error;
 405                         }
 406                         mlx5_shared_data = mz->addr;
 407                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 408                         rte_spinlock_init(&mlx5_shared_data->lock);
 409                 } else {
 410                         /* Lookup allocated shared memory. */
 411                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 412                         if (mz == NULL) {
 413                                 DRV_LOG(ERR,
 414                                         "Cannot attach mlx5 shared data");
 415                                 ret = -rte_errno;
 416                                 goto error;
 417                         }
 418                         mlx5_shared_data = mz->addr;
 419                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 420                 }
 421         }
 422 error:
 423         rte_spinlock_unlock(&mlx5_shared_data_lock);
 424         return ret;
 425 }
 426
 427 /**
 428  * PMD global initialization.
 429  *
 430  * Independent from individual device, this function initializes global
 431  * per-PMD data structures distinguishing primary and secondary processes.
 432  * Hence, each initialization is called once per a process.
 433  *
 434  * @return
 435  *   0 on success, a negative errno value otherwise and rte_errno is set.
 436  */
 437 static int
 438 mlx5_init_once(void)
 439 {
 440         struct mlx5_shared_data *sd;
 441         struct mlx5_local_data *ld = &mlx5_local_data;
 442         int ret = 0;
 443
 444         if (mlx5_init_shared_data())
 445                 return -rte_errno;
 446         sd = mlx5_shared_data;
 447         MLX5_ASSERT(sd);
 448         rte_spinlock_lock(&sd->lock);
 449         switch (rte_eal_process_type()) {
 450         case RTE_PROC_PRIMARY:
 451                 if (sd->init_done)
 452                         break;
 453                 LIST_INIT(&sd->mem_event_cb_list);
 454                 rte_rwlock_init(&sd->mem_event_rwlock);
 455                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
 456                                                 mlx5_mr_mem_event_cb, NULL);
 457                 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
 458                                            mlx5_mp_os_primary_handle);
 459                 if (ret)
 460                         goto out;
 461                 sd->init_done = true;
 462                 break;
 463         case RTE_PROC_SECONDARY:
 464                 if (ld->init_done)
 465                         break;
 466                 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
 467                                              mlx5_mp_os_secondary_handle);
 468                 if (ret)
 469                         goto out;
 470                 ++sd->secondary_cnt;
 471                 ld->init_done = true;
 472                 break;
 473         default:
 474                 break;
 475         }
 476 out:
 477         rte_spinlock_unlock(&sd->lock);
 478         return ret;
 479 }
 480
 481 /**
 482  * Spawn an Ethernet device from Verbs information.
 483  *
 484  * @param dpdk_dev
 485  *   Backing DPDK device.
 486  * @param spawn
 487  *   Verbs device parameters (name, port, switch_info) to spawn.
 488  * @param config
 489  *   Device configuration parameters.
 490  *
 491  * @return
 492  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 493  *   is set. The following errors are defined:
 494  *
 495  *   EBUSY: device is not supposed to be spawned.
 496  *   EEXIST: device is already spawned
 497  */
 498 static struct rte_eth_dev *
 499 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 500                struct mlx5_dev_spawn_data *spawn,
 501                struct mlx5_dev_config *config)
 502 {
 503         const struct mlx5_switch_info *switch_info = &spawn->info;
 504         struct mlx5_dev_ctx_shared *sh = NULL;
 505         struct ibv_port_attr port_attr;
 506         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 507         struct rte_eth_dev *eth_dev = NULL;
 508         struct mlx5_priv *priv = NULL;
 509         int err = 0;
 510         unsigned int hw_padding = 0;
 511         unsigned int mps;
 512         unsigned int cqe_comp;
 513         unsigned int cqe_pad = 0;
 514         unsigned int tunnel_en = 0;
 515         unsigned int mpls_en = 0;
 516         unsigned int swp = 0;
 517         unsigned int mprq = 0;
 518         unsigned int mprq_min_stride_size_n = 0;
 519         unsigned int mprq_max_stride_size_n = 0;
 520         unsigned int mprq_min_stride_num_n = 0;
 521         unsigned int mprq_max_stride_num_n = 0;
 522         struct rte_ether_addr mac;
 523         char name[RTE_ETH_NAME_MAX_LEN];
 524         int own_domain_id = 0;
 525         uint16_t port_id;
 526         unsigned int i;
 527 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 528         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 529 #endif
 530
 531         /* Determine if this port representor is supposed to be spawned. */
 532         if (switch_info->representor && dpdk_dev->devargs) {
 533                 struct rte_eth_devargs eth_da;
 534
 535                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 536                 if (err) {
 537                         rte_errno = -err;
 538                         DRV_LOG(ERR, "failed to process device arguments: %s",
 539                                 strerror(rte_errno));
 540                         return NULL;
 541                 }
 542                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 543                         if (eth_da.representor_ports[i] ==
 544                             (uint16_t)switch_info->port_name)
 545                                 break;
 546                 if (i == eth_da.nb_representor_ports) {
 547                         rte_errno = EBUSY;
 548                         return NULL;
 549                 }
 550         }
 551         /* Build device name. */
 552         if (spawn->pf_bond <  0) {
 553                 /* Single device. */
 554                 if (!switch_info->representor)
 555                         strlcpy(name, dpdk_dev->name, sizeof(name));
 556                 else
 557                         snprintf(name, sizeof(name), "%s_representor_%u",
 558                                  dpdk_dev->name, switch_info->port_name);
 559         } else {
 560                 /* Bonding device. */
 561                 if (!switch_info->representor)
 562                         snprintf(name, sizeof(name), "%s_%s",
 563                                  dpdk_dev->name,
 564                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 565                 else
 566                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 567                                  dpdk_dev->name,
 568                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 569                                  switch_info->port_name);
 570         }
 571         /* check if the device is already spawned */
 572         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 573                 rte_errno = EEXIST;
 574                 return NULL;
 575         }
 576         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 577         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 578                 struct mlx5_mp_id mp_id;
 579
 580                 eth_dev = rte_eth_dev_attach_secondary(name);
 581                 if (eth_dev == NULL) {
 582                         DRV_LOG(ERR, "can not attach rte ethdev");
 583                         rte_errno = ENOMEM;
 584                         return NULL;
 585                 }
 586                 eth_dev->device = dpdk_dev;
 587                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 588                 err = mlx5_proc_priv_init(eth_dev);
 589                 if (err)
 590                         return NULL;
 591                 mp_id.port_id = eth_dev->data->port_id;
 592                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 593                 /* Receive command fd from primary process */
 594                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 595                 if (err < 0)
 596                         goto err_secondary;
 597                 /* Remap UAR for Tx queues. */
 598                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 599                 if (err)
 600                         goto err_secondary;
 601                 /*
 602                  * Ethdev pointer is still required as input since
 603                  * the primary device is not accessible from the
 604                  * secondary process.
 605                  */
 606                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 607                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 608                 return eth_dev;
 609 err_secondary:
 610                 mlx5_dev_close(eth_dev);
 611                 return NULL;
 612         }
 613         /*
 614          * Some parameters ("tx_db_nc" in particularly) are needed in
 615          * advance to create dv/verbs device context. We proceed the
 616          * devargs here to get ones, and later proceed devargs again
 617          * to override some hardware settings.
 618          */
 619         err = mlx5_args(config, dpdk_dev->devargs);
 620         if (err) {
 621                 err = rte_errno;
 622                 DRV_LOG(ERR, "failed to process device arguments: %s",
 623                         strerror(rte_errno));
 624                 goto error;
 625         }
 626         mlx5_malloc_mem_select(config->sys_mem_en);
 627         sh = mlx5_alloc_shared_dev_ctx(spawn, config);
 628         if (!sh)
 629                 return NULL;
 630         config->devx = sh->devx;
 631 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 632         config->dest_tir = 1;
 633 #endif
 634 #ifdef HAVE_IBV_MLX5_MOD_SWP
 635         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 636 #endif
 637         /*
 638          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 639          * as all ConnectX-5 devices.
 640          */
 641 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 642         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 643 #endif
 644 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 645         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 646 #endif
 647         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 648         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 649                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 650                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 651                         mps = MLX5_MPW_ENHANCED;
 652                 } else {
 653                         DRV_LOG(DEBUG, "MPW is supported");
 654                         mps = MLX5_MPW;
 655                 }
 656         } else {
 657                 DRV_LOG(DEBUG, "MPW isn't supported");
 658                 mps = MLX5_MPW_DISABLED;
 659         }
 660 #ifdef HAVE_IBV_MLX5_MOD_SWP
 661         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 662                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 663         DRV_LOG(DEBUG, "SWP support: %u", swp);
 664 #endif
 665         config->swp = !!swp;
 666 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 667         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 668                 struct mlx5dv_striding_rq_caps mprq_caps =
 669                         dv_attr.striding_rq_caps;
 670
 671                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 672                         mprq_caps.min_single_stride_log_num_of_bytes);
 673                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 674                         mprq_caps.max_single_stride_log_num_of_bytes);
 675                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 676                         mprq_caps.min_single_wqe_log_num_of_strides);
 677                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 678                         mprq_caps.max_single_wqe_log_num_of_strides);
 679                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 680                         mprq_caps.supported_qpts);
 681                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 682                 mprq = 1;
 683                 mprq_min_stride_size_n =
 684                         mprq_caps.min_single_stride_log_num_of_bytes;
 685                 mprq_max_stride_size_n =
 686                         mprq_caps.max_single_stride_log_num_of_bytes;
 687                 mprq_min_stride_num_n =
 688                         mprq_caps.min_single_wqe_log_num_of_strides;
 689                 mprq_max_stride_num_n =
 690                         mprq_caps.max_single_wqe_log_num_of_strides;
 691         }
 692 #endif
 693         if (RTE_CACHE_LINE_SIZE == 128 &&
 694             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 695                 cqe_comp = 0;
 696         else
 697                 cqe_comp = 1;
 698         config->cqe_comp = cqe_comp;
 699 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 700         /* Whether device supports 128B Rx CQE padding. */
 701         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 702                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 703 #endif
 704 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 705         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 706                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 707                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 708                              (dv_attr.tunnel_offloads_caps &
 709                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 710                              (dv_attr.tunnel_offloads_caps &
 711                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 712         }
 713         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 714                 tunnel_en ? "" : "not ");
 715 #else
 716         DRV_LOG(WARNING,
 717                 "tunnel offloading disabled due to old OFED/rdma-core version");
 718 #endif
 719         config->tunnel_en = tunnel_en;
 720 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 721         mpls_en = ((dv_attr.tunnel_offloads_caps &
 722                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 723                    (dv_attr.tunnel_offloads_caps &
 724                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 725         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 726                 mpls_en ? "" : "not ");
 727 #else
 728         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 729                 " old OFED/rdma-core version or firmware configuration");
 730 #endif
 731         config->mpls_en = mpls_en;
 732         /* Check port status. */
 733         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 734         if (err) {
 735                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 736                 goto error;
 737         }
 738         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 739                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 740                 err = EINVAL;
 741                 goto error;
 742         }
 743         if (port_attr.state != IBV_PORT_ACTIVE)
 744                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 745                         mlx5_glue->port_state_str(port_attr.state),
 746                         port_attr.state);
 747         /* Allocate private eth device data. */
 748         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
 749                            sizeof(*priv),
 750                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 751         if (priv == NULL) {
 752                 DRV_LOG(ERR, "priv allocation failure");
 753                 err = ENOMEM;
 754                 goto error;
 755         }
 756         priv->sh = sh;
 757         priv->dev_port = spawn->phys_port;
 758         priv->pci_dev = spawn->pci_dev;
 759         priv->mtu = RTE_ETHER_MTU;
 760         priv->mp_id.port_id = port_id;
 761         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 762         /* Some internal functions rely on Netlink sockets, open them now. */
 763         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 764         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 765         priv->representor = !!switch_info->representor;
 766         priv->master = !!switch_info->master;
 767         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 768         priv->vport_meta_tag = 0;
 769         priv->vport_meta_mask = 0;
 770         priv->pf_bond = spawn->pf_bond;
 771 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 772         /*
 773          * The DevX port query API is implemented. E-Switch may use
 774          * either vport or reg_c[0] metadata register to match on
 775          * vport index. The engaged part of metadata register is
 776          * defined by mask.
 777          */
 778         if (switch_info->representor || switch_info->master) {
 779                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 780                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 781                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 782                                                  &devx_port);
 783                 if (err) {
 784                         DRV_LOG(WARNING,
 785                                 "can't query devx port %d on device %s",
 786                                 spawn->phys_port,
 787                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 788                         devx_port.comp_mask = 0;
 789                 }
 790         }
 791         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 792                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 793                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 794                 if (!priv->vport_meta_mask) {
 795                         DRV_LOG(ERR, "vport zero mask for port %d"
 796                                      " on bonding device %s",
 797                                      spawn->phys_port,
 798                                      mlx5_os_get_dev_device_name
 799                                                         (spawn->phys_dev));
 800                         err = ENOTSUP;
 801                         goto error;
 802                 }
 803                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 804                         DRV_LOG(ERR, "invalid vport tag for port %d"
 805                                      " on bonding device %s",
 806                                      spawn->phys_port,
 807                                      mlx5_os_get_dev_device_name
 808                                                         (spawn->phys_dev));
 809                         err = ENOTSUP;
 810                         goto error;
 811                 }
 812         }
 813         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 814                 priv->vport_id = devx_port.vport_num;
 815         } else if (spawn->pf_bond >= 0) {
 816                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 817                              " on bonding device %s",
 818                              spawn->phys_port,
 819                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 820                 err = ENOTSUP;
 821                 goto error;
 822         } else {
 823                 /* Suppose vport index in compatible way. */
 824                 priv->vport_id = switch_info->representor ?
 825                                  switch_info->port_name + 1 : -1;
 826         }
 827 #else
 828         /*
 829          * Kernel/rdma_core support single E-Switch per PF configurations
 830          * only and vport_id field contains the vport index for
 831          * associated VF, which is deduced from representor port name.
 832          * For example, let's have the IB device port 10, it has
 833          * attached network device eth0, which has port name attribute
 834          * pf0vf2, we can deduce the VF number as 2, and set vport index
 835          * as 3 (2+1). This assigning schema should be changed if the
 836          * multiple E-Switch instances per PF configurations or/and PCI
 837          * subfunctions are added.
 838          */
 839         priv->vport_id = switch_info->representor ?
 840                          switch_info->port_name + 1 : -1;
 841 #endif
 842         /* representor_id field keeps the unmodified VF index. */
 843         priv->representor_id = switch_info->representor ?
 844                                switch_info->port_name : -1;
 845         /*
 846          * Look for sibling devices in order to reuse their switch domain
 847          * if any, otherwise allocate one.
 848          */
 849         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 850                 const struct mlx5_priv *opriv =
 851                         rte_eth_devices[port_id].data->dev_private;
 852
 853                 if (!opriv ||
 854                     opriv->sh != priv->sh ||
 855                         opriv->domain_id ==
 856                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 857                         continue;
 858                 priv->domain_id = opriv->domain_id;
 859                 break;
 860         }
 861         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 862                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 863                 if (err) {
 864                         err = rte_errno;
 865                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 866                                 strerror(rte_errno));
 867                         goto error;
 868                 }
 869                 own_domain_id = 1;
 870         }
 871         /* Override some values set by hardware configuration. */
 872         mlx5_args(config, dpdk_dev->devargs);
 873         err = mlx5_dev_check_sibling_config(priv, config);
 874         if (err)
 875                 goto error;
 876         config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 877                             IBV_DEVICE_RAW_IP_CSUM);
 878         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 879                 (config->hw_csum ? "" : "not "));
 880 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 881         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 882         DRV_LOG(DEBUG, "counters are not supported");
 883 #endif
 884 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 885         if (config->dv_flow_en) {
 886                 DRV_LOG(WARNING, "DV flow is not supported");
 887                 config->dv_flow_en = 0;
 888         }
 889 #endif
 890         config->ind_table_max_size =
 891                 sh->device_attr.max_rwq_indirection_table_size;
 892         /*
 893          * Remove this check once DPDK supports larger/variable
 894          * indirection tables.
 895          */
 896         if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 897                 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 898         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 899                 config->ind_table_max_size);
 900         config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 901                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 902         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 903                 (config->hw_vlan_strip ? "" : "not "));
 904         config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 905                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 906 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 907         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 908 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 909         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 910                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 911 #endif
 912         if (config->hw_padding && !hw_padding) {
 913                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 914                 config->hw_padding = 0;
 915         } else if (config->hw_padding) {
 916                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 917         }
 918         config->tso = (sh->device_attr.max_tso > 0 &&
 919                       (sh->device_attr.tso_supported_qpts &
 920                        (1 << IBV_QPT_RAW_PACKET)));
 921         if (config->tso)
 922                 config->tso_max_payload_sz = sh->device_attr.max_tso;
 923         /*
 924          * MPW is disabled by default, while the Enhanced MPW is enabled
 925          * by default.
 926          */
 927         if (config->mps == MLX5_ARG_UNSET)
 928                 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 929                                                           MLX5_MPW_DISABLED;
 930         else
 931                 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
 932         DRV_LOG(INFO, "%sMPS is %s",
 933                 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
 934                 config->mps == MLX5_MPW ? "legacy " : "",
 935                 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 936         if (config->cqe_comp && !cqe_comp) {
 937                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 938                 config->cqe_comp = 0;
 939         }
 940         if (config->cqe_pad && !cqe_pad) {
 941                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 942                 config->cqe_pad = 0;
 943         } else if (config->cqe_pad) {
 944                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 945         }
 946         if (config->devx) {
 947                 priv->counter_fallback = 0;
 948                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
 949                 if (err) {
 950                         err = -err;
 951                         goto error;
 952                 }
 953                 if (!config->hca_attr.flow_counters_dump)
 954                         priv->counter_fallback = 1;
 955 #ifndef HAVE_IBV_DEVX_ASYNC
 956                 priv->counter_fallback = 1;
 957 #endif
 958                 if (priv->counter_fallback)
 959                         DRV_LOG(INFO, "Use fall-back DV counter management");
 960                 /* Check for LRO support. */
 961                 if (config->dest_tir && config->hca_attr.lro_cap &&
 962                     config->dv_flow_en) {
 963                         /* TBD check tunnel lro caps. */
 964                         config->lro.supported = config->hca_attr.lro_cap;
 965                         DRV_LOG(DEBUG, "Device supports LRO");
 966                         /*
 967                          * If LRO timeout is not configured by application,
 968                          * use the minimal supported value.
 969                          */
 970                         if (!config->lro.timeout)
 971                                 config->lro.timeout =
 972                                 config->hca_attr.lro_timer_supported_periods[0];
 973                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 974                                 config->lro.timeout);
 975                 }
 976 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 977                 if (config->hca_attr.qos.sup &&
 978                     config->hca_attr.qos.srtcm_sup &&
 979                     config->dv_flow_en) {
 980                         uint8_t reg_c_mask =
 981                                 config->hca_attr.qos.flow_meter_reg_c_ids;
 982                         /*
 983                          * Meter needs two REG_C's for color match and pre-sfx
 984                          * flow match. Here get the REG_C for color match.
 985                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 986                          */
 987                         reg_c_mask &= 0xfc;
 988                         if (__builtin_popcount(reg_c_mask) < 1) {
 989                                 priv->mtr_en = 0;
 990                                 DRV_LOG(WARNING, "No available register for"
 991                                         " meter.");
 992                         } else {
 993                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 994                                                       REG_C_0;
 995                                 priv->mtr_en = 1;
 996                                 priv->mtr_reg_share =
 997                                       config->hca_attr.qos.flow_meter_reg_share;
 998                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 999                                         priv->mtr_color_reg);
1000                         }
1001                 }
1002 #endif
1003         }
1004         if (config->tx_pp) {
1005                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1006                         config->hca_attr.dev_freq_khz);
1007                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1008                         config->hca_attr.qos.packet_pacing ? "" : "not ");
1009                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1010                         config->hca_attr.cross_channel ? "" : "not ");
1011                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1012                         config->hca_attr.wqe_index_ignore ? "" : "not ");
1013                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1014                         config->hca_attr.non_wire_sq ? "" : "not ");
1015                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1016                         config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1017                         config->hca_attr.log_max_static_sq_wq);
1018                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1019                         config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1020                 if (!config->devx) {
1021                         DRV_LOG(ERR, "DevX is required for packet pacing");
1022                         err = ENODEV;
1023                         goto error;
1024                 }
1025                 if (!config->hca_attr.qos.packet_pacing) {
1026                         DRV_LOG(ERR, "Packet pacing is not supported");
1027                         err = ENODEV;
1028                         goto error;
1029                 }
1030                 if (!config->hca_attr.cross_channel) {
1031                         DRV_LOG(ERR, "Cross channel operations are"
1032                                      " required for packet pacing");
1033                         err = ENODEV;
1034                         goto error;
1035                 }
1036                 if (!config->hca_attr.wqe_index_ignore) {
1037                         DRV_LOG(ERR, "WQE index ignore feature is"
1038                                      " required for packet pacing");
1039                         err = ENODEV;
1040                         goto error;
1041                 }
1042                 if (!config->hca_attr.non_wire_sq) {
1043                         DRV_LOG(ERR, "Non-wire SQ feature is"
1044                                      " required for packet pacing");
1045                         err = ENODEV;
1046                         goto error;
1047                 }
1048                 if (!config->hca_attr.log_max_static_sq_wq) {
1049                         DRV_LOG(ERR, "Static WQE SQ feature is"
1050                                      " required for packet pacing");
1051                         err = ENODEV;
1052                         goto error;
1053                 }
1054                 if (!config->hca_attr.qos.wqe_rate_pp) {
1055                         DRV_LOG(ERR, "WQE rate mode is required"
1056                                      " for packet pacing");
1057                         err = ENODEV;
1058                         goto error;
1059                 }
1060 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1061                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1062                              " can't create queues for packet pacing");
1063                 err = ENODEV;
1064                 goto error;
1065 #endif
1066         }
1067         if (config->devx) {
1068                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1069
1070                 err = config->hca_attr.access_register_user ?
1071                         mlx5_devx_cmd_register_read
1072                                 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1073                                 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1074                 if (!err) {
1075                         uint32_t ts_mode;
1076
1077                         /* MTUTC register is read successfully. */
1078                         ts_mode = MLX5_GET(register_mtutc, reg,
1079                                            time_stamp_mode);
1080                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1081                                 config->rt_timestamp = 1;
1082                 } else {
1083                         /* Kernel does not support register reading. */
1084                         if (config->hca_attr.dev_freq_khz ==
1085                                                  (NS_PER_S / MS_PER_S))
1086                                 config->rt_timestamp = 1;
1087                 }
1088         }
1089         /*
1090          * If HW has bug working with tunnel packet decapsulation and
1091          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1092          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1093          */
1094         if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1095                 config->hw_fcs_strip = 0;
1096         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1097                 (config->hw_fcs_strip ? "" : "not "));
1098         if (config->mprq.enabled && mprq) {
1099                 if (config->mprq.stride_num_n &&
1100                     (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1101                      config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1102                         config->mprq.stride_num_n =
1103                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1104                                                 mprq_min_stride_num_n),
1105                                         mprq_max_stride_num_n);
1106                         DRV_LOG(WARNING,
1107                                 "the number of strides"
1108                                 " for Multi-Packet RQ is out of range,"
1109                                 " setting default value (%u)",
1110                                 1 << config->mprq.stride_num_n);
1111                 }
1112                 if (config->mprq.stride_size_n &&
1113                     (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1114                      config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1115                         config->mprq.stride_size_n =
1116                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1117                                                 mprq_min_stride_size_n),
1118                                         mprq_max_stride_size_n);
1119                         DRV_LOG(WARNING,
1120                                 "the size of a stride"
1121                                 " for Multi-Packet RQ is out of range,"
1122                                 " setting default value (%u)",
1123                                 1 << config->mprq.stride_size_n);
1124                 }
1125                 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1126                 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1127         } else if (config->mprq.enabled && !mprq) {
1128                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1129                 config->mprq.enabled = 0;
1130         }
1131         if (config->max_dump_files_num == 0)
1132                 config->max_dump_files_num = 128;
1133         eth_dev = rte_eth_dev_allocate(name);
1134         if (eth_dev == NULL) {
1135                 DRV_LOG(ERR, "can not allocate rte ethdev");
1136                 err = ENOMEM;
1137                 goto error;
1138         }
1139         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1140         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1141         if (priv->representor) {
1142                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1143                 eth_dev->data->representor_id = priv->representor_id;
1144         }
1145         /*
1146          * Store associated network device interface index. This index
1147          * is permanent throughout the lifetime of device. So, we may store
1148          * the ifindex here and use the cached value further.
1149          */
1150         MLX5_ASSERT(spawn->ifindex);
1151         priv->if_index = spawn->ifindex;
1152         eth_dev->data->dev_private = priv;
1153         priv->dev_data = eth_dev->data;
1154         eth_dev->data->mac_addrs = priv->mac;
1155         eth_dev->device = dpdk_dev;
1156         /* Configure the first MAC address by default. */
1157         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1158                 DRV_LOG(ERR,
1159                         "port %u cannot get MAC address, is mlx5_en"
1160                         " loaded? (errno: %s)",
1161                         eth_dev->data->port_id, strerror(rte_errno));
1162                 err = ENODEV;
1163                 goto error;
1164         }
1165         DRV_LOG(INFO,
1166                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1167                 eth_dev->data->port_id,
1168                 mac.addr_bytes[0], mac.addr_bytes[1],
1169                 mac.addr_bytes[2], mac.addr_bytes[3],
1170                 mac.addr_bytes[4], mac.addr_bytes[5]);
1171 #ifdef RTE_LIBRTE_MLX5_DEBUG
1172         {
1173                 char ifname[IF_NAMESIZE];
1174
1175                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1176                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1177                                 eth_dev->data->port_id, ifname);
1178                 else
1179                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1180                                 eth_dev->data->port_id);
1181         }
1182 #endif
1183         /* Get actual MTU if possible. */
1184         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1185         if (err) {
1186                 err = rte_errno;
1187                 goto error;
1188         }
1189         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1190                 priv->mtu);
1191         /* Initialize burst functions to prevent crashes before link-up. */
1192         eth_dev->rx_pkt_burst = removed_rx_burst;
1193         eth_dev->tx_pkt_burst = removed_tx_burst;
1194         eth_dev->dev_ops = &mlx5_os_dev_ops;
1195         /* Register MAC address. */
1196         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1197         if (config->vf && config->vf_nl_en)
1198                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1199                                       mlx5_ifindex(eth_dev),
1200                                       eth_dev->data->mac_addrs,
1201                                       MLX5_MAX_MAC_ADDRESSES);
1202         priv->flows = 0;
1203         priv->ctrl_flows = 0;
1204         TAILQ_INIT(&priv->flow_meters);
1205         TAILQ_INIT(&priv->flow_meter_profiles);
1206         /* Hint libmlx5 to use PMD allocator for data plane resources */
1207         mlx5_glue->dv_set_context_attr(sh->ctx,
1208                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1209                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1210                                 .alloc = &mlx5_alloc_verbs_buf,
1211                                 .free = &mlx5_free_verbs_buf,
1212                                 .data = priv,
1213                         }));
1214         /* Bring Ethernet device up. */
1215         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1216                 eth_dev->data->port_id);
1217         mlx5_set_link_up(eth_dev);
1218         /*
1219          * Even though the interrupt handler is not installed yet,
1220          * interrupts will still trigger on the async_fd from
1221          * Verbs context returned by ibv_open_device().
1222          */
1223         mlx5_link_update(eth_dev, 0);
1224 #ifdef HAVE_MLX5DV_DR_ESWITCH
1225         if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1226               (switch_info->representor || switch_info->master)))
1227                 config->dv_esw_en = 0;
1228 #else
1229         config->dv_esw_en = 0;
1230 #endif
1231         /* Detect minimal data bytes to inline. */
1232         mlx5_set_min_inline(spawn, config);
1233         /* Store device configuration on private structure. */
1234         priv->config = *config;
1235         /* Create context for virtual machine VLAN workaround. */
1236         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1237         if (config->dv_flow_en) {
1238                 err = mlx5_alloc_shared_dr(priv);
1239                 if (err)
1240                         goto error;
1241                 /*
1242                  * RSS id is shared with meter flow id. Meter flow id can only
1243                  * use the 24 MSB of the register.
1244                  */
1245                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1246                                      MLX5_MTR_COLOR_BITS);
1247                 if (!priv->qrss_id_pool) {
1248                         DRV_LOG(ERR, "can't create flow id pool");
1249                         err = ENOMEM;
1250                         goto error;
1251                 }
1252         }
1253         /* Supported Verbs flow priority number detection. */
1254         err = mlx5_flow_discover_priorities(eth_dev);
1255         if (err < 0) {
1256                 err = -err;
1257                 goto error;
1258         }
1259         priv->config.flow_prio = err;
1260         if (!priv->config.dv_esw_en &&
1261             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1262                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1263                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1264                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1265         }
1266         mlx5_set_metadata_mask(eth_dev);
1267         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1268             !priv->sh->dv_regc0_mask) {
1269                 DRV_LOG(ERR, "metadata mode %u is not supported "
1270                              "(no metadata reg_c[0] is available)",
1271                              priv->config.dv_xmeta_en);
1272                         err = ENOTSUP;
1273                         goto error;
1274         }
1275         /*
1276          * Allocate the buffer for flow creating, just once.
1277          * The allocation must be done before any flow creating.
1278          */
1279         mlx5_flow_alloc_intermediate(eth_dev);
1280         /* Query availability of metadata reg_c's. */
1281         err = mlx5_flow_discover_mreg_c(eth_dev);
1282         if (err < 0) {
1283                 err = -err;
1284                 goto error;
1285         }
1286         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1287                 DRV_LOG(DEBUG,
1288                         "port %u extensive metadata register is not supported",
1289                         eth_dev->data->port_id);
1290                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1291                         DRV_LOG(ERR, "metadata mode %u is not supported "
1292                                      "(no metadata registers available)",
1293                                      priv->config.dv_xmeta_en);
1294                         err = ENOTSUP;
1295                         goto error;
1296                 }
1297         }
1298         if (priv->config.dv_flow_en &&
1299             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1300             mlx5_flow_ext_mreg_supported(eth_dev) &&
1301             priv->sh->dv_regc0_mask) {
1302                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1303                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1304                 if (!priv->mreg_cp_tbl) {
1305                         err = ENOMEM;
1306                         goto error;
1307                 }
1308         }
1309         return eth_dev;
1310 error:
1311         if (priv) {
1312                 if (priv->mreg_cp_tbl)
1313                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1314                 if (priv->sh)
1315                         mlx5_os_free_shared_dr(priv);
1316                 if (priv->nl_socket_route >= 0)
1317                         close(priv->nl_socket_route);
1318                 if (priv->nl_socket_rdma >= 0)
1319                         close(priv->nl_socket_rdma);
1320                 if (priv->vmwa_context)
1321                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1322                 if (priv->qrss_id_pool)
1323                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1324                 if (own_domain_id)
1325                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1326                 mlx5_free(priv);
1327                 if (eth_dev != NULL)
1328                         eth_dev->data->dev_private = NULL;
1329         }
1330         if (eth_dev != NULL) {
1331                 /* mac_addrs must not be freed alone because part of
1332                  * dev_private
1333                  **/
1334                 eth_dev->data->mac_addrs = NULL;
1335                 rte_eth_dev_release_port(eth_dev);
1336         }
1337         if (sh)
1338                 mlx5_free_shared_dev_ctx(sh);
1339         MLX5_ASSERT(err > 0);
1340         rte_errno = err;
1341         return NULL;
1342 }
1343
1344 /**
1345  * Comparison callback to sort device data.
1346  *
1347  * This is meant to be used with qsort().
1348  *
1349  * @param a[in]
1350  *   Pointer to pointer to first data object.
1351  * @param b[in]
1352  *   Pointer to pointer to second data object.
1353  *
1354  * @return
1355  *   0 if both objects are equal, less than 0 if the first argument is less
1356  *   than the second, greater than 0 otherwise.
1357  */
1358 static int
1359 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1360 {
1361         const struct mlx5_switch_info *si_a =
1362                 &((const struct mlx5_dev_spawn_data *)a)->info;
1363         const struct mlx5_switch_info *si_b =
1364                 &((const struct mlx5_dev_spawn_data *)b)->info;
1365         int ret;
1366
1367         /* Master device first. */
1368         ret = si_b->master - si_a->master;
1369         if (ret)
1370                 return ret;
1371         /* Then representor devices. */
1372         ret = si_b->representor - si_a->representor;
1373         if (ret)
1374                 return ret;
1375         /* Unidentified devices come last in no specific order. */
1376         if (!si_a->representor)
1377                 return 0;
1378         /* Order representors by name. */
1379         return si_a->port_name - si_b->port_name;
1380 }
1381
1382 /**
1383  * Match PCI information for possible slaves of bonding device.
1384  *
1385  * @param[in] ibv_dev
1386  *   Pointer to Infiniband device structure.
1387  * @param[in] pci_dev
1388  *   Pointer to PCI device structure to match PCI address.
1389  * @param[in] nl_rdma
1390  *   Netlink RDMA group socket handle.
1391  *
1392  * @return
1393  *   negative value if no bonding device found, otherwise
1394  *   positive index of slave PF in bonding.
1395  */
1396 static int
1397 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1398                            const struct rte_pci_device *pci_dev,
1399                            int nl_rdma)
1400 {
1401         char ifname[IF_NAMESIZE + 1];
1402         unsigned int ifindex;
1403         unsigned int np, i;
1404         FILE *file = NULL;
1405         int pf = -1;
1406
1407         /*
1408          * Try to get master device name. If something goes
1409          * wrong suppose the lack of kernel support and no
1410          * bonding devices.
1411          */
1412         if (nl_rdma < 0)
1413                 return -1;
1414         if (!strstr(ibv_dev->name, "bond"))
1415                 return -1;
1416         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1417         if (!np)
1418                 return -1;
1419         /*
1420          * The Master device might not be on the predefined
1421          * port (not on port index 1, it is not garanted),
1422          * we have to scan all Infiniband device port and
1423          * find master.
1424          */
1425         for (i = 1; i <= np; ++i) {
1426                 /* Check whether Infiniband port is populated. */
1427                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1428                 if (!ifindex)
1429                         continue;
1430                 if (!if_indextoname(ifindex, ifname))
1431                         continue;
1432                 /* Try to read bonding slave names from sysfs. */
1433                 MKSTR(slaves,
1434                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1435                 file = fopen(slaves, "r");
1436                 if (file)
1437                         break;
1438         }
1439         if (!file)
1440                 return -1;
1441         /* Use safe format to check maximal buffer length. */
1442         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1443         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1444                 char tmp_str[IF_NAMESIZE + 32];
1445                 struct rte_pci_addr pci_addr;
1446                 struct mlx5_switch_info info;
1447
1448                 /* Process slave interface names in the loop. */
1449                 snprintf(tmp_str, sizeof(tmp_str),
1450                          "/sys/class/net/%s", ifname);
1451                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1452                         DRV_LOG(WARNING, "can not get PCI address"
1453                                          " for netdev \"%s\"", ifname);
1454                         continue;
1455                 }
1456                 if (pci_dev->addr.domain != pci_addr.domain ||
1457                     pci_dev->addr.bus != pci_addr.bus ||
1458                     pci_dev->addr.devid != pci_addr.devid ||
1459                     pci_dev->addr.function != pci_addr.function)
1460                         continue;
1461                 /* Slave interface PCI address match found. */
1462                 fclose(file);
1463                 snprintf(tmp_str, sizeof(tmp_str),
1464                          "/sys/class/net/%s/phys_port_name", ifname);
1465                 file = fopen(tmp_str, "rb");
1466                 if (!file)
1467                         break;
1468                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1469                 if (fscanf(file, "%32s", tmp_str) == 1)
1470                         mlx5_translate_port_name(tmp_str, &info);
1471                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1472                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1473                         pf = info.port_name;
1474                 break;
1475         }
1476         if (file)
1477                 fclose(file);
1478         return pf;
1479 }
1480
1481 /**
1482  * DPDK callback to register a PCI device.
1483  *
1484  * This function spawns Ethernet devices out of a given PCI device.
1485  *
1486  * @param[in] pci_drv
1487  *   PCI driver structure (mlx5_driver).
1488  * @param[in] pci_dev
1489  *   PCI device information.
1490  *
1491  * @return
1492  *   0 on success, a negative errno value otherwise and rte_errno is set.
1493  */
1494 int
1495 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1496                   struct rte_pci_device *pci_dev)
1497 {
1498         struct ibv_device **ibv_list;
1499         /*
1500          * Number of found IB Devices matching with requested PCI BDF.
1501          * nd != 1 means there are multiple IB devices over the same
1502          * PCI device and we have representors and master.
1503          */
1504         unsigned int nd = 0;
1505         /*
1506          * Number of found IB device Ports. nd = 1 and np = 1..n means
1507          * we have the single multiport IB device, and there may be
1508          * representors attached to some of found ports.
1509          */
1510         unsigned int np = 0;
1511         /*
1512          * Number of DPDK ethernet devices to Spawn - either over
1513          * multiple IB devices or multiple ports of single IB device.
1514          * Actually this is the number of iterations to spawn.
1515          */
1516         unsigned int ns = 0;
1517         /*
1518          * Bonding device
1519          *   < 0 - no bonding device (single one)
1520          *  >= 0 - bonding device (value is slave PF index)
1521          */
1522         int bd = -1;
1523         struct mlx5_dev_spawn_data *list = NULL;
1524         struct mlx5_dev_config dev_config;
1525         unsigned int dev_config_vf;
1526         int ret;
1527
1528         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1529                 mlx5_pmd_socket_init();
1530         ret = mlx5_init_once();
1531         if (ret) {
1532                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1533                         strerror(rte_errno));
1534                 return -rte_errno;
1535         }
1536         errno = 0;
1537         ibv_list = mlx5_glue->get_device_list(&ret);
1538         if (!ibv_list) {
1539                 rte_errno = errno ? errno : ENOSYS;
1540                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1541                 return -rte_errno;
1542         }
1543         /*
1544          * First scan the list of all Infiniband devices to find
1545          * matching ones, gathering into the list.
1546          */
1547         struct ibv_device *ibv_match[ret + 1];
1548         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1549         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1550         unsigned int i;
1551
1552         while (ret-- > 0) {
1553                 struct rte_pci_addr pci_addr;
1554
1555                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1556                 bd = mlx5_device_bond_pci_match
1557                                 (ibv_list[ret], pci_dev, nl_rdma);
1558                 if (bd >= 0) {
1559                         /*
1560                          * Bonding device detected. Only one match is allowed,
1561                          * the bonding is supported over multi-port IB device,
1562                          * there should be no matches on representor PCI
1563                          * functions or non VF LAG bonding devices with
1564                          * specified address.
1565                          */
1566                         if (nd) {
1567                                 DRV_LOG(ERR,
1568                                         "multiple PCI match on bonding device"
1569                                         "\"%s\" found", ibv_list[ret]->name);
1570                                 rte_errno = ENOENT;
1571                                 ret = -rte_errno;
1572                                 goto exit;
1573                         }
1574                         DRV_LOG(INFO, "PCI information matches for"
1575                                       " slave %d bonding device \"%s\"",
1576                                       bd, ibv_list[ret]->name);
1577                         ibv_match[nd++] = ibv_list[ret];
1578                         break;
1579                 }
1580                 if (mlx5_dev_to_pci_addr
1581                         (ibv_list[ret]->ibdev_path, &pci_addr))
1582                         continue;
1583                 if (pci_dev->addr.domain != pci_addr.domain ||
1584                     pci_dev->addr.bus != pci_addr.bus ||
1585                     pci_dev->addr.devid != pci_addr.devid ||
1586                     pci_dev->addr.function != pci_addr.function)
1587                         continue;
1588                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1589                         ibv_list[ret]->name);
1590                 ibv_match[nd++] = ibv_list[ret];
1591         }
1592         ibv_match[nd] = NULL;
1593         if (!nd) {
1594                 /* No device matches, just complain and bail out. */
1595                 DRV_LOG(WARNING,
1596                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1597                         " are kernel drivers loaded?",
1598                         pci_dev->addr.domain, pci_dev->addr.bus,
1599                         pci_dev->addr.devid, pci_dev->addr.function);
1600                 rte_errno = ENOENT;
1601                 ret = -rte_errno;
1602                 goto exit;
1603         }
1604         if (nd == 1) {
1605                 /*
1606                  * Found single matching device may have multiple ports.
1607                  * Each port may be representor, we have to check the port
1608                  * number and check the representors existence.
1609                  */
1610                 if (nl_rdma >= 0)
1611                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1612                 if (!np)
1613                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1614                                          " ports number", ibv_match[0]->name);
1615                 if (bd >= 0 && !np) {
1616                         DRV_LOG(ERR, "can not get ports"
1617                                      " for bonding device");
1618                         rte_errno = ENOENT;
1619                         ret = -rte_errno;
1620                         goto exit;
1621                 }
1622         }
1623 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1624         if (bd >= 0) {
1625                 /*
1626                  * This may happen if there is VF LAG kernel support and
1627                  * application is compiled with older rdma_core library.
1628                  */
1629                 DRV_LOG(ERR,
1630                         "No kernel/verbs support for VF LAG bonding found.");
1631                 rte_errno = ENOTSUP;
1632                 ret = -rte_errno;
1633                 goto exit;
1634         }
1635 #endif
1636         /*
1637          * Now we can determine the maximal
1638          * amount of devices to be spawned.
1639          */
1640         list = mlx5_malloc(MLX5_MEM_ZERO,
1641                            sizeof(struct mlx5_dev_spawn_data) *
1642                            (np ? np : nd),
1643                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1644         if (!list) {
1645                 DRV_LOG(ERR, "spawn data array allocation failure");
1646                 rte_errno = ENOMEM;
1647                 ret = -rte_errno;
1648                 goto exit;
1649         }
1650         if (bd >= 0 || np > 1) {
1651                 /*
1652                  * Single IB device with multiple ports found,
1653                  * it may be E-Switch master device and representors.
1654                  * We have to perform identification through the ports.
1655                  */
1656                 MLX5_ASSERT(nl_rdma >= 0);
1657                 MLX5_ASSERT(ns == 0);
1658                 MLX5_ASSERT(nd == 1);
1659                 MLX5_ASSERT(np);
1660                 for (i = 1; i <= np; ++i) {
1661                         list[ns].max_port = np;
1662                         list[ns].phys_port = i;
1663                         list[ns].phys_dev = ibv_match[0];
1664                         list[ns].eth_dev = NULL;
1665                         list[ns].pci_dev = pci_dev;
1666                         list[ns].pf_bond = bd;
1667                         list[ns].ifindex = mlx5_nl_ifindex
1668                                 (nl_rdma,
1669                                 mlx5_os_get_dev_device_name
1670                                                 (list[ns].phys_dev), i);
1671                         if (!list[ns].ifindex) {
1672                                 /*
1673                                  * No network interface index found for the
1674                                  * specified port, it means there is no
1675                                  * representor on this port. It's OK,
1676                                  * there can be disabled ports, for example
1677                                  * if sriov_numvfs < sriov_totalvfs.
1678                                  */
1679                                 continue;
1680                         }
1681                         ret = -1;
1682                         if (nl_route >= 0)
1683                                 ret = mlx5_nl_switch_info
1684                                                (nl_route,
1685                                                 list[ns].ifindex,
1686                                                 &list[ns].info);
1687                         if (ret || (!list[ns].info.representor &&
1688                                     !list[ns].info.master)) {
1689                                 /*
1690                                  * We failed to recognize representors with
1691                                  * Netlink, let's try to perform the task
1692                                  * with sysfs.
1693                                  */
1694                                 ret =  mlx5_sysfs_switch_info
1695                                                 (list[ns].ifindex,
1696                                                  &list[ns].info);
1697                         }
1698                         if (!ret && bd >= 0) {
1699                                 switch (list[ns].info.name_type) {
1700                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1701                                         if (list[ns].info.port_name == bd)
1702                                                 ns++;
1703                                         break;
1704                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1705                                         /* Fallthrough */
1706                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1707                                         if (list[ns].info.pf_num == bd)
1708                                                 ns++;
1709                                         break;
1710                                 default:
1711                                         break;
1712                                 }
1713                                 continue;
1714                         }
1715                         if (!ret && (list[ns].info.representor ^
1716                                      list[ns].info.master))
1717                                 ns++;
1718                 }
1719                 if (!ns) {
1720                         DRV_LOG(ERR,
1721                                 "unable to recognize master/representors"
1722                                 " on the IB device with multiple ports");
1723                         rte_errno = ENOENT;
1724                         ret = -rte_errno;
1725                         goto exit;
1726                 }
1727         } else {
1728                 /*
1729                  * The existence of several matching entries (nd > 1) means
1730                  * port representors have been instantiated. No existing Verbs
1731                  * call nor sysfs entries can tell them apart, this can only
1732                  * be done through Netlink calls assuming kernel drivers are
1733                  * recent enough to support them.
1734                  *
1735                  * In the event of identification failure through Netlink,
1736                  * try again through sysfs, then:
1737                  *
1738                  * 1. A single IB device matches (nd == 1) with single
1739                  *    port (np=0/1) and is not a representor, assume
1740                  *    no switch support.
1741                  *
1742                  * 2. Otherwise no safe assumptions can be made;
1743                  *    complain louder and bail out.
1744                  */
1745                 for (i = 0; i != nd; ++i) {
1746                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1747                         list[ns].max_port = 1;
1748                         list[ns].phys_port = 1;
1749                         list[ns].phys_dev = ibv_match[i];
1750                         list[ns].eth_dev = NULL;
1751                         list[ns].pci_dev = pci_dev;
1752                         list[ns].pf_bond = -1;
1753                         list[ns].ifindex = 0;
1754                         if (nl_rdma >= 0)
1755                                 list[ns].ifindex = mlx5_nl_ifindex
1756                                 (nl_rdma,
1757                                 mlx5_os_get_dev_device_name
1758                                                 (list[ns].phys_dev), 1);
1759                         if (!list[ns].ifindex) {
1760                                 char ifname[IF_NAMESIZE];
1761
1762                                 /*
1763                                  * Netlink failed, it may happen with old
1764                                  * ib_core kernel driver (before 4.16).
1765                                  * We can assume there is old driver because
1766                                  * here we are processing single ports IB
1767                                  * devices. Let's try sysfs to retrieve
1768                                  * the ifindex. The method works for
1769                                  * master device only.
1770                                  */
1771                                 if (nd > 1) {
1772                                         /*
1773                                          * Multiple devices found, assume
1774                                          * representors, can not distinguish
1775                                          * master/representor and retrieve
1776                                          * ifindex via sysfs.
1777                                          */
1778                                         continue;
1779                                 }
1780                                 ret = mlx5_get_ifname_sysfs
1781                                         (ibv_match[i]->ibdev_path, ifname);
1782                                 if (!ret)
1783                                         list[ns].ifindex =
1784                                                 if_nametoindex(ifname);
1785                                 if (!list[ns].ifindex) {
1786                                         /*
1787                                          * No network interface index found
1788                                          * for the specified device, it means
1789                                          * there it is neither representor
1790                                          * nor master.
1791                                          */
1792                                         continue;
1793                                 }
1794                         }
1795                         ret = -1;
1796                         if (nl_route >= 0)
1797                                 ret = mlx5_nl_switch_info
1798                                                (nl_route,
1799                                                 list[ns].ifindex,
1800                                                 &list[ns].info);
1801                         if (ret || (!list[ns].info.representor &&
1802                                     !list[ns].info.master)) {
1803                                 /*
1804                                  * We failed to recognize representors with
1805                                  * Netlink, let's try to perform the task
1806                                  * with sysfs.
1807                                  */
1808                                 ret =  mlx5_sysfs_switch_info
1809                                                 (list[ns].ifindex,
1810                                                  &list[ns].info);
1811                         }
1812                         if (!ret && (list[ns].info.representor ^
1813                                      list[ns].info.master)) {
1814                                 ns++;
1815                         } else if ((nd == 1) &&
1816                                    !list[ns].info.representor &&
1817                                    !list[ns].info.master) {
1818                                 /*
1819                                  * Single IB device with
1820                                  * one physical port and
1821                                  * attached network device.
1822                                  * May be SRIOV is not enabled
1823                                  * or there is no representors.
1824                                  */
1825                                 DRV_LOG(INFO, "no E-Switch support detected");
1826                                 ns++;
1827                                 break;
1828                         }
1829                 }
1830                 if (!ns) {
1831                         DRV_LOG(ERR,
1832                                 "unable to recognize master/representors"
1833                                 " on the multiple IB devices");
1834                         rte_errno = ENOENT;
1835                         ret = -rte_errno;
1836                         goto exit;
1837                 }
1838         }
1839         MLX5_ASSERT(ns);
1840         /*
1841          * Sort list to probe devices in natural order for users convenience
1842          * (i.e. master first, then representors from lowest to highest ID).
1843          */
1844         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1845         /* Device specific configuration. */
1846         switch (pci_dev->id.device_id) {
1847         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1848         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1849         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1850         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1851         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1852         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1853         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1854                 dev_config_vf = 1;
1855                 break;
1856         default:
1857                 dev_config_vf = 0;
1858                 break;
1859         }
1860         for (i = 0; i != ns; ++i) {
1861                 uint32_t restore;
1862
1863                 /* Default configuration. */
1864                 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1865                 dev_config.vf = dev_config_vf;
1866                 dev_config.mps = MLX5_ARG_UNSET;
1867                 dev_config.dbnc = MLX5_ARG_UNSET;
1868                 dev_config.rx_vec_en = 1;
1869                 dev_config.txq_inline_max = MLX5_ARG_UNSET;
1870                 dev_config.txq_inline_min = MLX5_ARG_UNSET;
1871                 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
1872                 dev_config.txqs_inline = MLX5_ARG_UNSET;
1873                 dev_config.vf_nl_en = 1;
1874                 dev_config.mr_ext_memseg_en = 1;
1875                 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1876                 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1877                 dev_config.dv_esw_en = 1;
1878                 dev_config.dv_flow_en = 1;
1879                 dev_config.decap_en = 1;
1880                 dev_config.log_hp_size = MLX5_ARG_UNSET;
1881                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1882                                                  &list[i],
1883                                                  &dev_config);
1884                 if (!list[i].eth_dev) {
1885                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1886                                 break;
1887                         /* Device is disabled or already spawned. Ignore it. */
1888                         continue;
1889                 }
1890                 restore = list[i].eth_dev->data->dev_flags;
1891                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1892                 /* Restore non-PCI flags cleared by the above call. */
1893                 list[i].eth_dev->data->dev_flags |= restore;
1894                 rte_eth_dev_probing_finish(list[i].eth_dev);
1895         }
1896         if (i != ns) {
1897                 DRV_LOG(ERR,
1898                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1899                         " encountering an error: %s",
1900                         pci_dev->addr.domain, pci_dev->addr.bus,
1901                         pci_dev->addr.devid, pci_dev->addr.function,
1902                         strerror(rte_errno));
1903                 ret = -rte_errno;
1904                 /* Roll back. */
1905                 while (i--) {
1906                         if (!list[i].eth_dev)
1907                                 continue;
1908                         mlx5_dev_close(list[i].eth_dev);
1909                         /* mac_addrs must not be freed because in dev_private */
1910                         list[i].eth_dev->data->mac_addrs = NULL;
1911                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1912                 }
1913                 /* Restore original error. */
1914                 rte_errno = -ret;
1915         } else {
1916                 ret = 0;
1917         }
1918 exit:
1919         /*
1920          * Do the routine cleanup:
1921          * - close opened Netlink sockets
1922          * - free allocated spawn data array
1923          * - free the Infiniband device list
1924          */
1925         if (nl_rdma >= 0)
1926                 close(nl_rdma);
1927         if (nl_route >= 0)
1928                 close(nl_route);
1929         if (list)
1930                 mlx5_free(list);
1931         MLX5_ASSERT(ibv_list);
1932         mlx5_glue->free_device_list(ibv_list);
1933         return ret;
1934 }
1935
1936 static int
1937 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1938 {
1939         char *env;
1940         int value;
1941
1942         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1943         /* Get environment variable to store. */
1944         env = getenv(MLX5_SHUT_UP_BF);
1945         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1946         if (config->dbnc == MLX5_ARG_UNSET)
1947                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1948         else
1949                 setenv(MLX5_SHUT_UP_BF,
1950                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1951         return value;
1952 }
1953
1954 static void
1955 mlx5_restore_doorbell_mapping_env(int value)
1956 {
1957         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1958         /* Restore the original environment variable state. */
1959         if (value == MLX5_ARG_UNSET)
1960                 unsetenv(MLX5_SHUT_UP_BF);
1961         else
1962                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1963 }
1964
1965 /**
1966  * Extract pdn of PD object using DV API.
1967  *
1968  * @param[in] pd
1969  *   Pointer to the verbs PD object.
1970  * @param[out] pdn
1971  *   Pointer to the PD object number variable.
1972  *
1973  * @return
1974  *   0 on success, error value otherwise.
1975  */
1976 int
1977 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1978 {
1979 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1980         struct mlx5dv_obj obj;
1981         struct mlx5dv_pd pd_info;
1982         int ret = 0;
1983
1984         obj.pd.in = pd;
1985         obj.pd.out = &pd_info;
1986         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1987         if (ret) {
1988                 DRV_LOG(DEBUG, "Fail to get PD object info");
1989                 return ret;
1990         }
1991         *pdn = pd_info.pdn;
1992         return 0;
1993 #else
1994         (void)pd;
1995         (void)pdn;
1996         return -ENOTSUP;
1997 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1998 }
1999
2000 /**
2001  * Function API to open IB device.
2002  *
2003  * This function calls the Linux glue APIs to open a device.
2004  *
2005  * @param[in] spawn
2006  *   Pointer to the IB device attributes (name, port, etc).
2007  * @param[out] config
2008  *   Pointer to device configuration structure.
2009  * @param[out] sh
2010  *   Pointer to shared context structure.
2011  *
2012  * @return
2013  *   0 on success, a positive error value otherwise.
2014  */
2015 int
2016 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2017                      const struct mlx5_dev_config *config,
2018                      struct mlx5_dev_ctx_shared *sh)
2019 {
2020         int dbmap_env;
2021         int err = 0;
2022
2023         sh->numa_node = spawn->pci_dev->device.numa_node;
2024         pthread_mutex_init(&sh->txpp.mutex, NULL);
2025         /*
2026          * Configure environment variable "MLX5_BF_SHUT_UP"
2027          * before the device creation. The rdma_core library
2028          * checks the variable at device creation and
2029          * stores the result internally.
2030          */
2031         dbmap_env = mlx5_config_doorbell_mapping_env(config);
2032         /* Try to open IB device with DV first, then usual Verbs. */
2033         errno = 0;
2034         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2035         if (sh->ctx) {
2036                 sh->devx = 1;
2037                 DRV_LOG(DEBUG, "DevX is supported");
2038                 /* The device is created, no need for environment. */
2039                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2040         } else {
2041                 /* The environment variable is still configured. */
2042                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2043                 err = errno ? errno : ENODEV;
2044                 /*
2045                  * The environment variable is not needed anymore,
2046                  * all device creation attempts are completed.
2047                  */
2048                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2049                 if (!sh->ctx)
2050                         return err;
2051                 DRV_LOG(DEBUG, "DevX is NOT supported");
2052                 err = 0;
2053         }
2054         return err;
2055 }
2056
2057 /**
2058  * Install shared asynchronous device events handler.
2059  * This function is implemented to support event sharing
2060  * between multiple ports of single IB device.
2061  *
2062  * @param sh
2063  *   Pointer to mlx5_dev_ctx_shared object.
2064  */
2065 void
2066 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2067 {
2068         int ret;
2069         int flags;
2070
2071         sh->intr_handle.fd = -1;
2072         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2073         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2074                     F_SETFL, flags | O_NONBLOCK);
2075         if (ret) {
2076                 DRV_LOG(INFO, "failed to change file descriptor async event"
2077                         " queue");
2078         } else {
2079                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2080                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2081                 if (rte_intr_callback_register(&sh->intr_handle,
2082                                         mlx5_dev_interrupt_handler, sh)) {
2083                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
2084                         sh->intr_handle.fd = -1;
2085                 }
2086         }
2087         if (sh->devx) {
2088 #ifdef HAVE_IBV_DEVX_ASYNC
2089                 sh->intr_handle_devx.fd = -1;
2090                 sh->devx_comp =
2091                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2092                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2093                 if (!devx_comp) {
2094                         DRV_LOG(INFO, "failed to allocate devx_comp.");
2095                         return;
2096                 }
2097                 flags = fcntl(devx_comp->fd, F_GETFL);
2098                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2099                 if (ret) {
2100                         DRV_LOG(INFO, "failed to change file descriptor"
2101                                 " devx comp");
2102                         return;
2103                 }
2104                 sh->intr_handle_devx.fd = devx_comp->fd;
2105                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2106                 if (rte_intr_callback_register(&sh->intr_handle_devx,
2107                                         mlx5_dev_interrupt_handler_devx, sh)) {
2108                         DRV_LOG(INFO, "Fail to install the devx shared"
2109                                 " interrupt.");
2110                         sh->intr_handle_devx.fd = -1;
2111                 }
2112 #endif /* HAVE_IBV_DEVX_ASYNC */
2113         }
2114 }
2115
2116 /**
2117  * Uninstall shared asynchronous device events handler.
2118  * This function is implemented to support event sharing
2119  * between multiple ports of single IB device.
2120  *
2121  * @param dev
2122  *   Pointer to mlx5_dev_ctx_shared object.
2123  */
2124 void
2125 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2126 {
2127         if (sh->intr_handle.fd >= 0)
2128                 mlx5_intr_callback_unregister(&sh->intr_handle,
2129                                               mlx5_dev_interrupt_handler, sh);
2130 #ifdef HAVE_IBV_DEVX_ASYNC
2131         if (sh->intr_handle_devx.fd >= 0)
2132                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2133                                   mlx5_dev_interrupt_handler_devx, sh);
2134         if (sh->devx_comp)
2135                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2136 #endif
2137 }
2138
2139 /**
2140  * Read statistics by a named counter.
2141  *
2142  * @param[in] priv
2143  *   Pointer to the private device data structure.
2144  * @param[in] ctr_name
2145  *   Pointer to the name of the statistic counter to read
2146  * @param[out] stat
2147  *   Pointer to read statistic value.
2148  * @return
2149  *   0 on success and stat is valud, 1 if failed to read the value
2150  *   rte_errno is set.
2151  *
2152  */
2153 int
2154 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2155                       uint64_t *stat)
2156 {
2157         int fd;
2158
2159         if (priv->sh) {
2160                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2161                       priv->sh->ibdev_path,
2162                       priv->dev_port,
2163                       ctr_name);
2164                 fd = open(path, O_RDONLY);
2165                 /*
2166                  * in switchdev the file location is not per port
2167                  * but rather in <ibdev_path>/hw_counters/<file_name>.
2168                  */
2169                 if (fd == -1) {
2170                         MKSTR(path1, "%s/hw_counters/%s",
2171                               priv->sh->ibdev_path,
2172                               ctr_name);
2173                         fd = open(path1, O_RDONLY);
2174                 }
2175                 if (fd != -1) {
2176                         char buf[21] = {'\0'};
2177                         ssize_t n = read(fd, buf, sizeof(buf));
2178
2179                         close(fd);
2180                         if (n != -1) {
2181                                 *stat = strtoull(buf, NULL, 10);
2182                                 return 0;
2183                         }
2184                 }
2185         }
2186         *stat = 0;
2187         return 1;
2188 }
2189
2190 /**
2191  * Set the reg_mr and dereg_mr call backs
2192  *
2193  * @param reg_mr_cb[out]
2194  *   Pointer to reg_mr func
2195  * @param dereg_mr_cb[out]
2196  *   Pointer to dereg_mr func
2197  *
2198  */
2199 void
2200 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2201                       mlx5_dereg_mr_t *dereg_mr_cb)
2202 {
2203         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2204         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2205 }
2206
2207 /**
2208  * Remove a MAC address from device
2209  *
2210  * @param dev
2211  *   Pointer to Ethernet device structure.
2212  * @param index
2213  *   MAC address index.
2214  */
2215 void
2216 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2217 {
2218         struct mlx5_priv *priv = dev->data->dev_private;
2219         const int vf = priv->config.vf;
2220
2221         if (vf)
2222                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2223                                         mlx5_ifindex(dev), priv->mac_own,
2224                                         &dev->data->mac_addrs[index], index);
2225 }
2226
2227 /**
2228  * Adds a MAC address to the device
2229  *
2230  * @param dev
2231  *   Pointer to Ethernet device structure.
2232  * @param mac_addr
2233  *   MAC address to register.
2234  * @param index
2235  *   MAC address index.
2236  *
2237  * @return
2238  *   0 on success, a negative errno value otherwise
2239  */
2240 int
2241 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2242                      uint32_t index)
2243 {
2244         struct mlx5_priv *priv = dev->data->dev_private;
2245         const int vf = priv->config.vf;
2246         int ret = 0;
2247
2248         if (vf)
2249                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2250                                            mlx5_ifindex(dev), priv->mac_own,
2251                                            mac, index);
2252         return ret;
2253 }
2254
2255 /**
2256  * Modify a VF MAC address
2257  *
2258  * @param priv
2259  *   Pointer to device private data.
2260  * @param mac_addr
2261  *   MAC address to modify into.
2262  * @param iface_idx
2263  *   Net device interface index
2264  * @param vf_index
2265  *   VF index
2266  *
2267  * @return
2268  *   0 on success, a negative errno value otherwise
2269  */
2270 int
2271 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2272                            unsigned int iface_idx,
2273                            struct rte_ether_addr *mac_addr,
2274                            int vf_index)
2275 {
2276         return mlx5_nl_vf_mac_addr_modify
2277                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2278 }
2279
2280 /**
2281  * Set device promiscuous mode
2282  *
2283  * @param dev
2284  *   Pointer to Ethernet device structure.
2285  * @param enable
2286  *   0 - promiscuous is disabled, otherwise - enabled
2287  *
2288  * @return
2289  *   0 on success, a negative error value otherwise
2290  */
2291 int
2292 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2293 {
2294         struct mlx5_priv *priv = dev->data->dev_private;
2295
2296         return mlx5_nl_promisc(priv->nl_socket_route,
2297                                mlx5_ifindex(dev), !!enable);
2298 }
2299
2300 /**
2301  * Set device promiscuous mode
2302  *
2303  * @param dev
2304  *   Pointer to Ethernet device structure.
2305  * @param enable
2306  *   0 - all multicase is disabled, otherwise - enabled
2307  *
2308  * @return
2309  *   0 on success, a negative error value otherwise
2310  */
2311 int
2312 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2313 {
2314         struct mlx5_priv *priv = dev->data->dev_private;
2315
2316         return mlx5_nl_allmulti(priv->nl_socket_route,
2317                                 mlx5_ifindex(dev), !!enable);
2318 }
2319
2320 const struct eth_dev_ops mlx5_os_dev_ops = {
2321         .dev_configure = mlx5_dev_configure,
2322         .dev_start = mlx5_dev_start,
2323         .dev_stop = mlx5_dev_stop,
2324         .dev_set_link_down = mlx5_set_link_down,
2325         .dev_set_link_up = mlx5_set_link_up,
2326         .dev_close = mlx5_dev_close,
2327         .promiscuous_enable = mlx5_promiscuous_enable,
2328         .promiscuous_disable = mlx5_promiscuous_disable,
2329         .allmulticast_enable = mlx5_allmulticast_enable,
2330         .allmulticast_disable = mlx5_allmulticast_disable,
2331         .link_update = mlx5_link_update,
2332         .stats_get = mlx5_stats_get,
2333         .stats_reset = mlx5_stats_reset,
2334         .xstats_get = mlx5_xstats_get,
2335         .xstats_reset = mlx5_xstats_reset,
2336         .xstats_get_names = mlx5_xstats_get_names,
2337         .fw_version_get = mlx5_fw_version_get,
2338         .dev_infos_get = mlx5_dev_infos_get,
2339         .read_clock = mlx5_txpp_read_clock,
2340         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2341         .vlan_filter_set = mlx5_vlan_filter_set,
2342         .rx_queue_setup = mlx5_rx_queue_setup,
2343         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2344         .tx_queue_setup = mlx5_tx_queue_setup,
2345         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2346         .rx_queue_release = mlx5_rx_queue_release,
2347         .tx_queue_release = mlx5_tx_queue_release,
2348         .rx_queue_start = mlx5_rx_queue_start,
2349         .rx_queue_stop = mlx5_rx_queue_stop,
2350         .tx_queue_start = mlx5_tx_queue_start,
2351         .tx_queue_stop = mlx5_tx_queue_stop,
2352         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2353         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2354         .mac_addr_remove = mlx5_mac_addr_remove,
2355         .mac_addr_add = mlx5_mac_addr_add,
2356         .mac_addr_set = mlx5_mac_addr_set,
2357         .set_mc_addr_list = mlx5_set_mc_addr_list,
2358         .mtu_set = mlx5_dev_set_mtu,
2359         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2360         .vlan_offload_set = mlx5_vlan_offload_set,
2361         .reta_update = mlx5_dev_rss_reta_update,
2362         .reta_query = mlx5_dev_rss_reta_query,
2363         .rss_hash_update = mlx5_rss_hash_update,
2364         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2365         .filter_ctrl = mlx5_dev_filter_ctrl,
2366         .rx_descriptor_status = mlx5_rx_descriptor_status,
2367         .tx_descriptor_status = mlx5_tx_descriptor_status,
2368         .rxq_info_get = mlx5_rxq_info_get,
2369         .txq_info_get = mlx5_txq_info_get,
2370         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2371         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2372         .rx_queue_count = mlx5_rx_queue_count,
2373         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2374         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2375         .is_removed = mlx5_is_removed,
2376         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2377         .get_module_info = mlx5_get_module_info,
2378         .get_module_eeprom = mlx5_get_module_eeprom,
2379         .hairpin_cap_get = mlx5_hairpin_cap_get,
2380         .mtr_ops_get = mlx5_flow_meter_ops_get,
2381 };
2382
2383 /* Available operations from secondary process. */
2384 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2385         .stats_get = mlx5_stats_get,
2386         .stats_reset = mlx5_stats_reset,
2387         .xstats_get = mlx5_xstats_get,
2388         .xstats_reset = mlx5_xstats_reset,
2389         .xstats_get_names = mlx5_xstats_get_names,
2390         .fw_version_get = mlx5_fw_version_get,
2391         .dev_infos_get = mlx5_dev_infos_get,
2392         .read_clock = mlx5_txpp_read_clock,
2393         .rx_queue_start = mlx5_rx_queue_start,
2394         .rx_queue_stop = mlx5_rx_queue_stop,
2395         .tx_queue_start = mlx5_tx_queue_start,
2396         .tx_queue_stop = mlx5_tx_queue_stop,
2397         .rx_descriptor_status = mlx5_rx_descriptor_status,
2398         .tx_descriptor_status = mlx5_tx_descriptor_status,
2399         .rxq_info_get = mlx5_rxq_info_get,
2400         .txq_info_get = mlx5_txq_info_get,
2401         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2402         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2403         .get_module_info = mlx5_get_module_info,
2404         .get_module_eeprom = mlx5_get_module_eeprom,
2405 };
2406
2407 /* Available operations in flow isolated mode. */
2408 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2409         .dev_configure = mlx5_dev_configure,
2410         .dev_start = mlx5_dev_start,
2411         .dev_stop = mlx5_dev_stop,
2412         .dev_set_link_down = mlx5_set_link_down,
2413         .dev_set_link_up = mlx5_set_link_up,
2414         .dev_close = mlx5_dev_close,
2415         .promiscuous_enable = mlx5_promiscuous_enable,
2416         .promiscuous_disable = mlx5_promiscuous_disable,
2417         .allmulticast_enable = mlx5_allmulticast_enable,
2418         .allmulticast_disable = mlx5_allmulticast_disable,
2419         .link_update = mlx5_link_update,
2420         .stats_get = mlx5_stats_get,
2421         .stats_reset = mlx5_stats_reset,
2422         .xstats_get = mlx5_xstats_get,
2423         .xstats_reset = mlx5_xstats_reset,
2424         .xstats_get_names = mlx5_xstats_get_names,
2425         .fw_version_get = mlx5_fw_version_get,
2426         .dev_infos_get = mlx5_dev_infos_get,
2427         .read_clock = mlx5_txpp_read_clock,
2428         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2429         .vlan_filter_set = mlx5_vlan_filter_set,
2430         .rx_queue_setup = mlx5_rx_queue_setup,
2431         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2432         .tx_queue_setup = mlx5_tx_queue_setup,
2433         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2434         .rx_queue_release = mlx5_rx_queue_release,
2435         .tx_queue_release = mlx5_tx_queue_release,
2436         .rx_queue_start = mlx5_rx_queue_start,
2437         .rx_queue_stop = mlx5_rx_queue_stop,
2438         .tx_queue_start = mlx5_tx_queue_start,
2439         .tx_queue_stop = mlx5_tx_queue_stop,
2440         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2441         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2442         .mac_addr_remove = mlx5_mac_addr_remove,
2443         .mac_addr_add = mlx5_mac_addr_add,
2444         .mac_addr_set = mlx5_mac_addr_set,
2445         .set_mc_addr_list = mlx5_set_mc_addr_list,
2446         .mtu_set = mlx5_dev_set_mtu,
2447         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2448         .vlan_offload_set = mlx5_vlan_offload_set,
2449         .filter_ctrl = mlx5_dev_filter_ctrl,
2450         .rx_descriptor_status = mlx5_rx_descriptor_status,
2451         .tx_descriptor_status = mlx5_tx_descriptor_status,
2452         .rxq_info_get = mlx5_rxq_info_get,
2453         .txq_info_get = mlx5_txq_info_get,
2454         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2455         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2456         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2457         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2458         .is_removed = mlx5_is_removed,
2459         .get_module_info = mlx5_get_module_info,
2460         .get_module_eeprom = mlx5_get_module_eeprom,
2461         .hairpin_cap_get = mlx5_hairpin_cap_get,
2462         .mtr_ops_get = mlx5_flow_meter_ops_get,
2463 };