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 <ethdev_driver.h>
  20 #include <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_rx.h"
  44 #include "mlx5_tx.h"
  45 #include "mlx5_autoconf.h"
  46 #include "mlx5_mr.h"
  47 #include "mlx5_flow.h"
  48 #include "rte_pmd_mlx5.h"
  49 #include "mlx5_verbs.h"
  50 #include "mlx5_nl.h"
  51 #include "mlx5_devx.h"
  52
  53 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  54
  55 #ifndef HAVE_IBV_MLX5_MOD_MPW
  56 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  57 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  58 #endif
  59
  60 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  61 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  62 #endif
  63
  64 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
  65
  66 /* Spinlock for mlx5_shared_data allocation. */
  67 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
  68
  69 /* Process local data for secondary processes. */
  70 static struct mlx5_local_data mlx5_local_data;
  71
  72 /**
  73  * Set the completion channel file descriptor interrupt as non-blocking.
  74  *
  75  * @param[in] rxq_obj
  76  *   Pointer to RQ channel object, which includes the channel fd
  77  *
  78  * @param[out] fd
  79  *   The file descriptor (representing the intetrrupt) used in this channel.
  80  *
  81  * @return
  82  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
  83  */
  84 int
  85 mlx5_os_set_nonblock_channel_fd(int fd)
  86 {
  87         int flags;
  88
  89         flags = fcntl(fd, F_GETFL);
  90         return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
  91 }
  92
  93 /**
  94  * Get mlx5 device attributes. The glue function query_device_ex() is called
  95  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  96  * device attributes from the glue out parameter.
  97  *
  98  * @param dev
  99  *   Pointer to ibv context.
 100  *
 101  * @param device_attr
 102  *   Pointer to mlx5 device attributes.
 103  *
 104  * @return
 105  *   0 on success, non zero error number otherwise
 106  */
 107 int
 108 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
 109 {
 110         int err;
 111         struct ibv_device_attr_ex attr_ex;
 112         memset(device_attr, 0, sizeof(*device_attr));
 113         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
 114         if (err)
 115                 return err;
 116
 117         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
 118         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
 119         device_attr->max_sge = attr_ex.orig_attr.max_sge;
 120         device_attr->max_cq = attr_ex.orig_attr.max_cq;
 121         device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
 122         device_attr->max_mr = attr_ex.orig_attr.max_mr;
 123         device_attr->max_pd = attr_ex.orig_attr.max_pd;
 124         device_attr->max_qp = attr_ex.orig_attr.max_qp;
 125         device_attr->max_srq = attr_ex.orig_attr.max_srq;
 126         device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
 127         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 128         device_attr->max_rwq_indirection_table_size =
 129                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 130         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 131         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 132
 133         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 134         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 135         if (err)
 136                 return err;
 137
 138         device_attr->flags = dv_attr.flags;
 139         device_attr->comp_mask = dv_attr.comp_mask;
 140 #ifdef HAVE_IBV_MLX5_MOD_SWP
 141         device_attr->sw_parsing_offloads =
 142                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 143 #endif
 144         device_attr->min_single_stride_log_num_of_bytes =
 145                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 146         device_attr->max_single_stride_log_num_of_bytes =
 147                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 148         device_attr->min_single_wqe_log_num_of_strides =
 149                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 150         device_attr->max_single_wqe_log_num_of_strides =
 151                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 152         device_attr->stride_supported_qpts =
 153                 dv_attr.striding_rq_caps.supported_qpts;
 154 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 155         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 156 #endif
 157
 158         return err;
 159 }
 160
 161 /**
 162  * Verbs callback to allocate a memory. This function should allocate the space
 163  * according to the size provided residing inside a huge page.
 164  * Please note that all allocation must respect the alignment from libmlx5
 165  * (i.e. currently rte_mem_page_size()).
 166  *
 167  * @param[in] size
 168  *   The size in bytes of the memory to allocate.
 169  * @param[in] data
 170  *   A pointer to the callback data.
 171  *
 172  * @return
 173  *   Allocated buffer, NULL otherwise and rte_errno is set.
 174  */
 175 static void *
 176 mlx5_alloc_verbs_buf(size_t size, void *data)
 177 {
 178         struct mlx5_dev_ctx_shared *sh = data;
 179         void *ret;
 180         size_t alignment = rte_mem_page_size();
 181         if (alignment == (size_t)-1) {
 182                 DRV_LOG(ERR, "Failed to get mem page size");
 183                 rte_errno = ENOMEM;
 184                 return NULL;
 185         }
 186
 187         MLX5_ASSERT(data != NULL);
 188         ret = mlx5_malloc(0, size, alignment, sh->numa_node);
 189         if (!ret && size)
 190                 rte_errno = ENOMEM;
 191         return ret;
 192 }
 193
 194 /**
 195  * Verbs callback to free a memory.
 196  *
 197  * @param[in] ptr
 198  *   A pointer to the memory to free.
 199  * @param[in] data
 200  *   A pointer to the callback data.
 201  */
 202 static void
 203 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 204 {
 205         MLX5_ASSERT(data != NULL);
 206         mlx5_free(ptr);
 207 }
 208
 209 /**
 210  * Initialize DR related data within private structure.
 211  * Routine checks the reference counter and does actual
 212  * resources creation/initialization only if counter is zero.
 213  *
 214  * @param[in] priv
 215  *   Pointer to the private device data structure.
 216  *
 217  * @return
 218  *   Zero on success, positive error code otherwise.
 219  */
 220 static int
 221 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 222 {
 223         struct mlx5_dev_ctx_shared *sh = priv->sh;
 224         char s[MLX5_HLIST_NAMESIZE] __rte_unused;
 225         int err;
 226
 227         MLX5_ASSERT(sh && sh->refcnt);
 228         if (sh->refcnt > 1)
 229                 return 0;
 230         err = mlx5_alloc_table_hash_list(priv);
 231         if (err)
 232                 goto error;
 233         /* The resources below are only valid with DV support. */
 234 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 235         /* Init port id action cache list. */
 236         snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
 237         mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
 238                              flow_dv_port_id_create_cb,
 239                              flow_dv_port_id_match_cb,
 240                              flow_dv_port_id_remove_cb);
 241         /* Init push vlan action cache list. */
 242         snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
 243         mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
 244                              flow_dv_push_vlan_create_cb,
 245                              flow_dv_push_vlan_match_cb,
 246                              flow_dv_push_vlan_remove_cb);
 247         /* Init sample action cache list. */
 248         snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
 249         mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
 250                              flow_dv_sample_create_cb,
 251                              flow_dv_sample_match_cb,
 252                              flow_dv_sample_remove_cb);
 253         /* Init dest array action cache list. */
 254         snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
 255         mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
 256                              flow_dv_dest_array_create_cb,
 257                              flow_dv_dest_array_match_cb,
 258                              flow_dv_dest_array_remove_cb);
 259         /* Create tags hash list table. */
 260         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 261         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
 262                                           MLX5_HLIST_WRITE_MOST,
 263                                           flow_dv_tag_create_cb,
 264                                           flow_dv_tag_match_cb,
 265                                           flow_dv_tag_remove_cb);
 266         if (!sh->tag_table) {
 267                 DRV_LOG(ERR, "tags with hash creation failed.");
 268                 err = ENOMEM;
 269                 goto error;
 270         }
 271         sh->tag_table->ctx = sh;
 272         snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
 273         sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
 274                                             0, MLX5_HLIST_WRITE_MOST |
 275                                             MLX5_HLIST_DIRECT_KEY,
 276                                             flow_dv_modify_create_cb,
 277                                             flow_dv_modify_match_cb,
 278                                             flow_dv_modify_remove_cb);
 279         if (!sh->modify_cmds) {
 280                 DRV_LOG(ERR, "hdr modify hash creation failed");
 281                 err = ENOMEM;
 282                 goto error;
 283         }
 284         sh->modify_cmds->ctx = sh;
 285         snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
 286         sh->encaps_decaps = mlx5_hlist_create(s,
 287                                               MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
 288                                               0, MLX5_HLIST_DIRECT_KEY |
 289                                               MLX5_HLIST_WRITE_MOST,
 290                                               flow_dv_encap_decap_create_cb,
 291                                               flow_dv_encap_decap_match_cb,
 292                                               flow_dv_encap_decap_remove_cb);
 293         if (!sh->encaps_decaps) {
 294                 DRV_LOG(ERR, "encap decap hash creation failed");
 295                 err = ENOMEM;
 296                 goto error;
 297         }
 298         sh->encaps_decaps->ctx = sh;
 299 #endif
 300 #ifdef HAVE_MLX5DV_DR
 301         void *domain;
 302
 303         /* Reference counter is zero, we should initialize structures. */
 304         domain = mlx5_glue->dr_create_domain(sh->ctx,
 305                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 306         if (!domain) {
 307                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 308                 err = errno;
 309                 goto error;
 310         }
 311         sh->rx_domain = domain;
 312         domain = mlx5_glue->dr_create_domain(sh->ctx,
 313                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 314         if (!domain) {
 315                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 316                 err = errno;
 317                 goto error;
 318         }
 319         sh->tx_domain = domain;
 320 #ifdef HAVE_MLX5DV_DR_ESWITCH
 321         if (priv->config.dv_esw_en) {
 322                 domain  = mlx5_glue->dr_create_domain
 323                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 324                 if (!domain) {
 325                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 326                         err = errno;
 327                         goto error;
 328                 }
 329                 sh->fdb_domain = domain;
 330         }
 331         /*
 332          * The drop action is just some dummy placeholder in rdma-core. It
 333          * does not belong to domains and has no any attributes, and, can be
 334          * shared by the entire device.
 335          */
 336         sh->dr_drop_action = mlx5_glue->dr_create_flow_action_drop();
 337         if (!sh->dr_drop_action) {
 338                 DRV_LOG(ERR, "FDB mlx5dv_dr_create_flow_action_drop");
 339                 err = errno;
 340                 goto error;
 341         }
 342 #endif
 343         if (!sh->tunnel_hub)
 344                 err = mlx5_alloc_tunnel_hub(sh);
 345         if (err) {
 346                 DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
 347                 goto error;
 348         }
 349         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 350                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 351                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 352                 if (sh->fdb_domain)
 353                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 354         }
 355         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 356 #endif /* HAVE_MLX5DV_DR */
 357         sh->default_miss_action =
 358                         mlx5_glue->dr_create_flow_action_default_miss();
 359         if (!sh->default_miss_action)
 360                 DRV_LOG(WARNING, "Default miss action is not supported.");
 361         return 0;
 362 error:
 363         /* Rollback the created objects. */
 364         if (sh->rx_domain) {
 365                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 366                 sh->rx_domain = NULL;
 367         }
 368         if (sh->tx_domain) {
 369                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 370                 sh->tx_domain = NULL;
 371         }
 372         if (sh->fdb_domain) {
 373                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 374                 sh->fdb_domain = NULL;
 375         }
 376         if (sh->dr_drop_action) {
 377                 mlx5_glue->destroy_flow_action(sh->dr_drop_action);
 378                 sh->dr_drop_action = NULL;
 379         }
 380         if (sh->pop_vlan_action) {
 381                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 382                 sh->pop_vlan_action = NULL;
 383         }
 384         if (sh->encaps_decaps) {
 385                 mlx5_hlist_destroy(sh->encaps_decaps);
 386                 sh->encaps_decaps = NULL;
 387         }
 388         if (sh->modify_cmds) {
 389                 mlx5_hlist_destroy(sh->modify_cmds);
 390                 sh->modify_cmds = NULL;
 391         }
 392         if (sh->tag_table) {
 393                 /* tags should be destroyed with flow before. */
 394                 mlx5_hlist_destroy(sh->tag_table);
 395                 sh->tag_table = NULL;
 396         }
 397         if (sh->tunnel_hub) {
 398                 mlx5_release_tunnel_hub(sh, priv->dev_port);
 399                 sh->tunnel_hub = NULL;
 400         }
 401         mlx5_free_table_hash_list(priv);
 402         return err;
 403 }
 404
 405 /**
 406  * Destroy DR related data within private structure.
 407  *
 408  * @param[in] priv
 409  *   Pointer to the private device data structure.
 410  */
 411 void
 412 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 413 {
 414         struct mlx5_dev_ctx_shared *sh = priv->sh;
 415
 416         MLX5_ASSERT(sh && sh->refcnt);
 417         if (sh->refcnt > 1)
 418                 return;
 419 #ifdef HAVE_MLX5DV_DR
 420         if (sh->rx_domain) {
 421                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 422                 sh->rx_domain = NULL;
 423         }
 424         if (sh->tx_domain) {
 425                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 426                 sh->tx_domain = NULL;
 427         }
 428 #ifdef HAVE_MLX5DV_DR_ESWITCH
 429         if (sh->fdb_domain) {
 430                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 431                 sh->fdb_domain = NULL;
 432         }
 433         if (sh->dr_drop_action) {
 434                 mlx5_glue->destroy_flow_action(sh->dr_drop_action);
 435                 sh->dr_drop_action = NULL;
 436         }
 437 #endif
 438         if (sh->pop_vlan_action) {
 439                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 440                 sh->pop_vlan_action = NULL;
 441         }
 442 #endif /* HAVE_MLX5DV_DR */
 443         if (sh->default_miss_action)
 444                 mlx5_glue->destroy_flow_action
 445                                 (sh->default_miss_action);
 446         if (sh->encaps_decaps) {
 447                 mlx5_hlist_destroy(sh->encaps_decaps);
 448                 sh->encaps_decaps = NULL;
 449         }
 450         if (sh->modify_cmds) {
 451                 mlx5_hlist_destroy(sh->modify_cmds);
 452                 sh->modify_cmds = NULL;
 453         }
 454         if (sh->tag_table) {
 455                 /* tags should be destroyed with flow before. */
 456                 mlx5_hlist_destroy(sh->tag_table);
 457                 sh->tag_table = NULL;
 458         }
 459         if (sh->tunnel_hub) {
 460                 mlx5_release_tunnel_hub(sh, priv->dev_port);
 461                 sh->tunnel_hub = NULL;
 462         }
 463         mlx5_cache_list_destroy(&sh->port_id_action_list);
 464         mlx5_cache_list_destroy(&sh->push_vlan_action_list);
 465         mlx5_free_table_hash_list(priv);
 466 }
 467
 468 /**
 469  * Initialize shared data between primary and secondary process.
 470  *
 471  * A memzone is reserved by primary process and secondary processes attach to
 472  * the memzone.
 473  *
 474  * @return
 475  *   0 on success, a negative errno value otherwise and rte_errno is set.
 476  */
 477 static int
 478 mlx5_init_shared_data(void)
 479 {
 480         const struct rte_memzone *mz;
 481         int ret = 0;
 482
 483         rte_spinlock_lock(&mlx5_shared_data_lock);
 484         if (mlx5_shared_data == NULL) {
 485                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 486                         /* Allocate shared memory. */
 487                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 488                                                  sizeof(*mlx5_shared_data),
 489                                                  SOCKET_ID_ANY, 0);
 490                         if (mz == NULL) {
 491                                 DRV_LOG(ERR,
 492                                         "Cannot allocate mlx5 shared data");
 493                                 ret = -rte_errno;
 494                                 goto error;
 495                         }
 496                         mlx5_shared_data = mz->addr;
 497                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 498                         rte_spinlock_init(&mlx5_shared_data->lock);
 499                 } else {
 500                         /* Lookup allocated shared memory. */
 501                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 502                         if (mz == NULL) {
 503                                 DRV_LOG(ERR,
 504                                         "Cannot attach mlx5 shared data");
 505                                 ret = -rte_errno;
 506                                 goto error;
 507                         }
 508                         mlx5_shared_data = mz->addr;
 509                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 510                 }
 511         }
 512 error:
 513         rte_spinlock_unlock(&mlx5_shared_data_lock);
 514         return ret;
 515 }
 516
 517 /**
 518  * PMD global initialization.
 519  *
 520  * Independent from individual device, this function initializes global
 521  * per-PMD data structures distinguishing primary and secondary processes.
 522  * Hence, each initialization is called once per a process.
 523  *
 524  * @return
 525  *   0 on success, a negative errno value otherwise and rte_errno is set.
 526  */
 527 static int
 528 mlx5_init_once(void)
 529 {
 530         struct mlx5_shared_data *sd;
 531         struct mlx5_local_data *ld = &mlx5_local_data;
 532         int ret = 0;
 533
 534         if (mlx5_init_shared_data())
 535                 return -rte_errno;
 536         sd = mlx5_shared_data;
 537         MLX5_ASSERT(sd);
 538         rte_spinlock_lock(&sd->lock);
 539         switch (rte_eal_process_type()) {
 540         case RTE_PROC_PRIMARY:
 541                 if (sd->init_done)
 542                         break;
 543                 LIST_INIT(&sd->mem_event_cb_list);
 544                 rte_rwlock_init(&sd->mem_event_rwlock);
 545                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
 546                                                 mlx5_mr_mem_event_cb, NULL);
 547                 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
 548                                            mlx5_mp_os_primary_handle);
 549                 if (ret)
 550                         goto out;
 551                 sd->init_done = true;
 552                 break;
 553         case RTE_PROC_SECONDARY:
 554                 if (ld->init_done)
 555                         break;
 556                 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
 557                                              mlx5_mp_os_secondary_handle);
 558                 if (ret)
 559                         goto out;
 560                 ++sd->secondary_cnt;
 561                 ld->init_done = true;
 562                 break;
 563         default:
 564                 break;
 565         }
 566 out:
 567         rte_spinlock_unlock(&sd->lock);
 568         return ret;
 569 }
 570
 571 /**
 572  * Create the Tx queue DevX/Verbs object.
 573  *
 574  * @param dev
 575  *   Pointer to Ethernet device.
 576  * @param idx
 577  *   Queue index in DPDK Tx queue array.
 578  *
 579  * @return
 580  *   0 on success, a negative errno value otherwise and rte_errno is set.
 581  */
 582 static int
 583 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
 584 {
 585         struct mlx5_priv *priv = dev->data->dev_private;
 586         struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
 587         struct mlx5_txq_ctrl *txq_ctrl =
 588                         container_of(txq_data, struct mlx5_txq_ctrl, txq);
 589
 590         if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
 591                 return mlx5_txq_devx_obj_new(dev, idx);
 592 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
 593         if (!priv->config.dv_esw_en)
 594                 return mlx5_txq_devx_obj_new(dev, idx);
 595 #endif
 596         return mlx5_txq_ibv_obj_new(dev, idx);
 597 }
 598
 599 /**
 600  * Release an Tx DevX/verbs queue object.
 601  *
 602  * @param txq_obj
 603  *   DevX/Verbs Tx queue object.
 604  */
 605 static void
 606 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
 607 {
 608         if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
 609                 mlx5_txq_devx_obj_release(txq_obj);
 610                 return;
 611         }
 612 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
 613         if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
 614                 mlx5_txq_devx_obj_release(txq_obj);
 615                 return;
 616         }
 617 #endif
 618         mlx5_txq_ibv_obj_release(txq_obj);
 619 }
 620
 621 /**
 622  * DV flow counter mode detect and config.
 623  *
 624  * @param dev
 625  *   Pointer to rte_eth_dev structure.
 626  *
 627  */
 628 static void
 629 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
 630 {
 631 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 632         struct mlx5_priv *priv = dev->data->dev_private;
 633         struct mlx5_dev_ctx_shared *sh = priv->sh;
 634         bool fallback;
 635
 636 #ifndef HAVE_IBV_DEVX_ASYNC
 637         fallback = true;
 638 #else
 639         fallback = false;
 640         if (!priv->config.devx || !priv->config.dv_flow_en ||
 641             !priv->config.hca_attr.flow_counters_dump ||
 642             !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
 643             (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
 644                 fallback = true;
 645 #endif
 646         if (fallback)
 647                 DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
 648                         "counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
 649                         priv->config.hca_attr.flow_counters_dump,
 650                         priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
 651         /* Initialize fallback mode only on the port initializes sh. */
 652         if (sh->refcnt == 1)
 653                 sh->cmng.counter_fallback = fallback;
 654         else if (fallback != sh->cmng.counter_fallback)
 655                 DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
 656                         "with others:%d.", PORT_ID(priv), fallback);
 657 #endif
 658 }
 659
 660 static void
 661 mlx5_queue_counter_id_prepare(struct rte_eth_dev *dev)
 662 {
 663         struct mlx5_priv *priv = dev->data->dev_private;
 664         void *ctx = priv->sh->ctx;
 665
 666         priv->q_counters = mlx5_devx_cmd_queue_counter_alloc(ctx);
 667         if (!priv->q_counters) {
 668                 struct ibv_cq *cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
 669                 struct ibv_wq *wq;
 670
 671                 DRV_LOG(DEBUG, "Port %d queue counter object cannot be created "
 672                         "by DevX - fall-back to use the kernel driver global "
 673                         "queue counter.", dev->data->port_id);
 674                 /* Create WQ by kernel and query its queue counter ID. */
 675                 if (cq) {
 676                         wq = mlx5_glue->create_wq(ctx,
 677                                                   &(struct ibv_wq_init_attr){
 678                                                     .wq_type = IBV_WQT_RQ,
 679                                                     .max_wr = 1,
 680                                                     .max_sge = 1,
 681                                                     .pd = priv->sh->pd,
 682                                                     .cq = cq,
 683                                                 });
 684                         if (wq) {
 685                                 /* Counter is assigned only on RDY state. */
 686                                 int ret = mlx5_glue->modify_wq(wq,
 687                                                  &(struct ibv_wq_attr){
 688                                                  .attr_mask = IBV_WQ_ATTR_STATE,
 689                                                  .wq_state = IBV_WQS_RDY,
 690                                                 });
 691
 692                                 if (ret == 0)
 693                                         mlx5_devx_cmd_wq_query(wq,
 694                                                          &priv->counter_set_id);
 695                                 claim_zero(mlx5_glue->destroy_wq(wq));
 696                         }
 697                         claim_zero(mlx5_glue->destroy_cq(cq));
 698                 }
 699         } else {
 700                 priv->counter_set_id = priv->q_counters->id;
 701         }
 702         if (priv->counter_set_id == 0)
 703                 DRV_LOG(INFO, "Part of the port %d statistics will not be "
 704                         "available.", dev->data->port_id);
 705 }
 706
 707 /**
 708  * Check if representor spawn info match devargs.
 709  *
 710  * @param spawn
 711  *   Verbs device parameters (name, port, switch_info) to spawn.
 712  * @param eth_da
 713  *   Device devargs to probe.
 714  *
 715  * @return
 716  *   Match result.
 717  */
 718 static bool
 719 mlx5_representor_match(struct mlx5_dev_spawn_data *spawn,
 720                        struct rte_eth_devargs *eth_da)
 721 {
 722         struct mlx5_switch_info *switch_info = &spawn->info;
 723         unsigned int p, f;
 724         uint16_t id;
 725         uint16_t repr_id = mlx5_representor_id_encode(switch_info,
 726                                                       eth_da->type);
 727
 728         switch (eth_da->type) {
 729         case RTE_ETH_REPRESENTOR_SF:
 730                 if (!(spawn->info.port_name == -1 &&
 731                       switch_info->name_type ==
 732                                 MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
 733                     switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFSF) {
 734                         rte_errno = EBUSY;
 735                         return false;
 736                 }
 737                 break;
 738         case RTE_ETH_REPRESENTOR_VF:
 739                 /* Allows HPF representor index -1 as exception. */
 740                 if (!(spawn->info.port_name == -1 &&
 741                       switch_info->name_type ==
 742                                 MLX5_PHYS_PORT_NAME_TYPE_PFHPF) &&
 743                     switch_info->name_type != MLX5_PHYS_PORT_NAME_TYPE_PFVF) {
 744                         rte_errno = EBUSY;
 745                         return false;
 746                 }
 747                 break;
 748         case RTE_ETH_REPRESENTOR_NONE:
 749                 rte_errno = EBUSY;
 750                 return false;
 751         default:
 752                 rte_errno = ENOTSUP;
 753                 DRV_LOG(ERR, "unsupported representor type");
 754                 return false;
 755         }
 756         /* Check representor ID: */
 757         for (p = 0; p < eth_da->nb_ports; ++p) {
 758                 if (spawn->pf_bond < 0) {
 759                         /* For non-LAG mode, allow and ignore pf. */
 760                         switch_info->pf_num = eth_da->ports[p];
 761                         repr_id = mlx5_representor_id_encode(switch_info,
 762                                                              eth_da->type);
 763                 }
 764                 for (f = 0; f < eth_da->nb_representor_ports; ++f) {
 765                         id = MLX5_REPRESENTOR_ID
 766                                 (eth_da->ports[p], eth_da->type,
 767                                  eth_da->representor_ports[f]);
 768                         if (repr_id == id)
 769                                 return true;
 770                 }
 771         }
 772         rte_errno = EBUSY;
 773         return false;
 774 }
 775
 776
 777 /**
 778  * Spawn an Ethernet device from Verbs information.
 779  *
 780  * @param dpdk_dev
 781  *   Backing DPDK device.
 782  * @param spawn
 783  *   Verbs device parameters (name, port, switch_info) to spawn.
 784  * @param config
 785  *   Device configuration parameters.
 786  * @param config
 787  *   Device arguments.
 788  *
 789  * @return
 790  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 791  *   is set. The following errors are defined:
 792  *
 793  *   EBUSY: device is not supposed to be spawned.
 794  *   EEXIST: device is already spawned
 795  */
 796 static struct rte_eth_dev *
 797 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 798                struct mlx5_dev_spawn_data *spawn,
 799                struct mlx5_dev_config *config,
 800                struct rte_eth_devargs *eth_da)
 801 {
 802         const struct mlx5_switch_info *switch_info = &spawn->info;
 803         struct mlx5_dev_ctx_shared *sh = NULL;
 804         struct ibv_port_attr port_attr;
 805         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 806         struct rte_eth_dev *eth_dev = NULL;
 807         struct mlx5_priv *priv = NULL;
 808         int err = 0;
 809         unsigned int hw_padding = 0;
 810         unsigned int mps;
 811         unsigned int tunnel_en = 0;
 812         unsigned int mpls_en = 0;
 813         unsigned int swp = 0;
 814         unsigned int mprq = 0;
 815         unsigned int mprq_min_stride_size_n = 0;
 816         unsigned int mprq_max_stride_size_n = 0;
 817         unsigned int mprq_min_stride_num_n = 0;
 818         unsigned int mprq_max_stride_num_n = 0;
 819         struct rte_ether_addr mac;
 820         char name[RTE_ETH_NAME_MAX_LEN];
 821         int own_domain_id = 0;
 822         uint16_t port_id;
 823 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 824         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 825 #endif
 826
 827         /* Determine if this port representor is supposed to be spawned. */
 828         if (switch_info->representor && dpdk_dev->devargs &&
 829             !mlx5_representor_match(spawn, eth_da))
 830                 return NULL;
 831         /* Build device name. */
 832         if (spawn->pf_bond < 0) {
 833                 /* Single device. */
 834                 if (!switch_info->representor)
 835                         strlcpy(name, dpdk_dev->name, sizeof(name));
 836                 else
 837                         err = snprintf(name, sizeof(name), "%s_representor_%s%u",
 838                                  dpdk_dev->name,
 839                                  switch_info->name_type ==
 840                                  MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
 841                                  switch_info->port_name);
 842         } else {
 843                 /* Bonding device. */
 844                 if (!switch_info->representor) {
 845                         err = snprintf(name, sizeof(name), "%s_%s",
 846                                  dpdk_dev->name,
 847                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 848                 } else {
 849                         err = snprintf(name, sizeof(name), "%s_%s_representor_c%dpf%d%s%u",
 850                                 dpdk_dev->name,
 851                                 mlx5_os_get_dev_device_name(spawn->phys_dev),
 852                                 switch_info->ctrl_num,
 853                                 switch_info->pf_num,
 854                                 switch_info->name_type ==
 855                                 MLX5_PHYS_PORT_NAME_TYPE_PFSF ? "sf" : "vf",
 856                                 switch_info->port_name);
 857                 }
 858         }
 859         if (err >= (int)sizeof(name))
 860                 DRV_LOG(WARNING, "device name overflow %s", name);
 861         /* check if the device is already spawned */
 862         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 863                 rte_errno = EEXIST;
 864                 return NULL;
 865         }
 866         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 867         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 868                 struct mlx5_mp_id mp_id;
 869
 870                 eth_dev = rte_eth_dev_attach_secondary(name);
 871                 if (eth_dev == NULL) {
 872                         DRV_LOG(ERR, "can not attach rte ethdev");
 873                         rte_errno = ENOMEM;
 874                         return NULL;
 875                 }
 876                 eth_dev->device = dpdk_dev;
 877                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
 878                 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
 879                 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
 880                 err = mlx5_proc_priv_init(eth_dev);
 881                 if (err)
 882                         return NULL;
 883                 mp_id.port_id = eth_dev->data->port_id;
 884                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 885                 /* Receive command fd from primary process */
 886                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 887                 if (err < 0)
 888                         goto err_secondary;
 889                 /* Remap UAR for Tx queues. */
 890                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 891                 if (err)
 892                         goto err_secondary;
 893                 /*
 894                  * Ethdev pointer is still required as input since
 895                  * the primary device is not accessible from the
 896                  * secondary process.
 897                  */
 898                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 899                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 900                 return eth_dev;
 901 err_secondary:
 902                 mlx5_dev_close(eth_dev);
 903                 return NULL;
 904         }
 905         /*
 906          * Some parameters ("tx_db_nc" in particularly) are needed in
 907          * advance to create dv/verbs device context. We proceed the
 908          * devargs here to get ones, and later proceed devargs again
 909          * to override some hardware settings.
 910          */
 911         err = mlx5_args(config, dpdk_dev->devargs);
 912         if (err) {
 913                 err = rte_errno;
 914                 DRV_LOG(ERR, "failed to process device arguments: %s",
 915                         strerror(rte_errno));
 916                 goto error;
 917         }
 918         if (config->dv_miss_info) {
 919                 if (switch_info->master || switch_info->representor)
 920                         config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
 921         }
 922         mlx5_malloc_mem_select(config->sys_mem_en);
 923         sh = mlx5_alloc_shared_dev_ctx(spawn, config);
 924         if (!sh)
 925                 return NULL;
 926         config->devx = sh->devx;
 927 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 928         config->dest_tir = 1;
 929 #endif
 930 #ifdef HAVE_IBV_MLX5_MOD_SWP
 931         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 932 #endif
 933         /*
 934          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 935          * as all ConnectX-5 devices.
 936          */
 937 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 938         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 939 #endif
 940 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 941         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 942 #endif
 943         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 944         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 945                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 946                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 947                         mps = MLX5_MPW_ENHANCED;
 948                 } else {
 949                         DRV_LOG(DEBUG, "MPW is supported");
 950                         mps = MLX5_MPW;
 951                 }
 952         } else {
 953                 DRV_LOG(DEBUG, "MPW isn't supported");
 954                 mps = MLX5_MPW_DISABLED;
 955         }
 956 #ifdef HAVE_IBV_MLX5_MOD_SWP
 957         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 958                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 959         DRV_LOG(DEBUG, "SWP support: %u", swp);
 960 #endif
 961         config->swp = !!swp;
 962 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 963         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 964                 struct mlx5dv_striding_rq_caps mprq_caps =
 965                         dv_attr.striding_rq_caps;
 966
 967                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 968                         mprq_caps.min_single_stride_log_num_of_bytes);
 969                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 970                         mprq_caps.max_single_stride_log_num_of_bytes);
 971                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 972                         mprq_caps.min_single_wqe_log_num_of_strides);
 973                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 974                         mprq_caps.max_single_wqe_log_num_of_strides);
 975                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 976                         mprq_caps.supported_qpts);
 977                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 978                 mprq = 1;
 979                 mprq_min_stride_size_n =
 980                         mprq_caps.min_single_stride_log_num_of_bytes;
 981                 mprq_max_stride_size_n =
 982                         mprq_caps.max_single_stride_log_num_of_bytes;
 983                 mprq_min_stride_num_n =
 984                         mprq_caps.min_single_wqe_log_num_of_strides;
 985                 mprq_max_stride_num_n =
 986                         mprq_caps.max_single_wqe_log_num_of_strides;
 987         }
 988 #endif
 989         /* Rx CQE compression is enabled by default. */
 990         config->cqe_comp = 1;
 991 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 992         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 993                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 994                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 995                              (dv_attr.tunnel_offloads_caps &
 996                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 997                              (dv_attr.tunnel_offloads_caps &
 998                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 999         }
1000         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1001                 tunnel_en ? "" : "not ");
1002 #else
1003         DRV_LOG(WARNING,
1004                 "tunnel offloading disabled due to old OFED/rdma-core version");
1005 #endif
1006         config->tunnel_en = tunnel_en;
1007 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1008         mpls_en = ((dv_attr.tunnel_offloads_caps &
1009                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1010                    (dv_attr.tunnel_offloads_caps &
1011                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1012         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1013                 mpls_en ? "" : "not ");
1014 #else
1015         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1016                 " old OFED/rdma-core version or firmware configuration");
1017 #endif
1018         config->mpls_en = mpls_en;
1019         /* Check port status. */
1020         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
1021         if (err) {
1022                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1023                 goto error;
1024         }
1025         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1026                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1027                 err = EINVAL;
1028                 goto error;
1029         }
1030         if (port_attr.state != IBV_PORT_ACTIVE)
1031                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1032                         mlx5_glue->port_state_str(port_attr.state),
1033                         port_attr.state);
1034         /* Allocate private eth device data. */
1035         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
1036                            sizeof(*priv),
1037                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1038         if (priv == NULL) {
1039                 DRV_LOG(ERR, "priv allocation failure");
1040                 err = ENOMEM;
1041                 goto error;
1042         }
1043         priv->sh = sh;
1044         priv->dev_port = spawn->phys_port;
1045         priv->pci_dev = spawn->pci_dev;
1046         priv->mtu = RTE_ETHER_MTU;
1047         /* Some internal functions rely on Netlink sockets, open them now. */
1048         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1049         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1050         priv->representor = !!switch_info->representor;
1051         priv->master = !!switch_info->master;
1052         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1053         priv->vport_meta_tag = 0;
1054         priv->vport_meta_mask = 0;
1055         priv->pf_bond = spawn->pf_bond;
1056 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1057         /*
1058          * The DevX port query API is implemented. E-Switch may use
1059          * either vport or reg_c[0] metadata register to match on
1060          * vport index. The engaged part of metadata register is
1061          * defined by mask.
1062          */
1063         if (switch_info->representor || switch_info->master) {
1064                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
1065                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
1066                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
1067                                                  &devx_port);
1068                 if (err) {
1069                         DRV_LOG(WARNING,
1070                                 "can't query devx port %d on device %s",
1071                                 spawn->phys_port,
1072                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
1073                         devx_port.comp_mask = 0;
1074                 }
1075         }
1076         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
1077                 priv->vport_meta_tag = devx_port.reg_c_0.value;
1078                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
1079                 if (!priv->vport_meta_mask) {
1080                         DRV_LOG(ERR, "vport zero mask for port %d"
1081                                      " on bonding device %s",
1082                                      spawn->phys_port,
1083                                      mlx5_os_get_dev_device_name
1084                                                         (spawn->phys_dev));
1085                         err = ENOTSUP;
1086                         goto error;
1087                 }
1088                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
1089                         DRV_LOG(ERR, "invalid vport tag for port %d"
1090                                      " on bonding device %s",
1091                                      spawn->phys_port,
1092                                      mlx5_os_get_dev_device_name
1093                                                         (spawn->phys_dev));
1094                         err = ENOTSUP;
1095                         goto error;
1096                 }
1097         }
1098         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
1099                 priv->vport_id = devx_port.vport_num;
1100         } else if (spawn->pf_bond >= 0) {
1101                 DRV_LOG(ERR, "can't deduce vport index for port %d"
1102                              " on bonding device %s",
1103                              spawn->phys_port,
1104                              mlx5_os_get_dev_device_name(spawn->phys_dev));
1105                 err = ENOTSUP;
1106                 goto error;
1107         } else {
1108                 /* Suppose vport index in compatible way. */
1109                 priv->vport_id = switch_info->representor ?
1110                                  switch_info->port_name + 1 : -1;
1111         }
1112 #else
1113         /*
1114          * Kernel/rdma_core support single E-Switch per PF configurations
1115          * only and vport_id field contains the vport index for
1116          * associated VF, which is deduced from representor port name.
1117          * For example, let's have the IB device port 10, it has
1118          * attached network device eth0, which has port name attribute
1119          * pf0vf2, we can deduce the VF number as 2, and set vport index
1120          * as 3 (2+1). This assigning schema should be changed if the
1121          * multiple E-Switch instances per PF configurations or/and PCI
1122          * subfunctions are added.
1123          */
1124         priv->vport_id = switch_info->representor ?
1125                          switch_info->port_name + 1 : -1;
1126 #endif
1127         priv->representor_id = mlx5_representor_id_encode(switch_info,
1128                                                           eth_da->type);
1129         /*
1130          * Look for sibling devices in order to reuse their switch domain
1131          * if any, otherwise allocate one.
1132          */
1133         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1134                 const struct mlx5_priv *opriv =
1135                         rte_eth_devices[port_id].data->dev_private;
1136
1137                 if (!opriv ||
1138                     opriv->sh != priv->sh ||
1139                         opriv->domain_id ==
1140                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1141                         continue;
1142                 priv->domain_id = opriv->domain_id;
1143                 break;
1144         }
1145         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1146                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1147                 if (err) {
1148                         err = rte_errno;
1149                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1150                                 strerror(rte_errno));
1151                         goto error;
1152                 }
1153                 own_domain_id = 1;
1154         }
1155         /* Override some values set by hardware configuration. */
1156         mlx5_args(config, dpdk_dev->devargs);
1157         err = mlx5_dev_check_sibling_config(priv, config);
1158         if (err)
1159                 goto error;
1160         config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1161                             IBV_DEVICE_RAW_IP_CSUM);
1162         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1163                 (config->hw_csum ? "" : "not "));
1164 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1165         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1166         DRV_LOG(DEBUG, "counters are not supported");
1167 #endif
1168 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1169         if (config->dv_flow_en) {
1170                 DRV_LOG(WARNING, "DV flow is not supported");
1171                 config->dv_flow_en = 0;
1172         }
1173 #endif
1174         config->ind_table_max_size =
1175                 sh->device_attr.max_rwq_indirection_table_size;
1176         /*
1177          * Remove this check once DPDK supports larger/variable
1178          * indirection tables.
1179          */
1180         if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1181                 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1182         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1183                 config->ind_table_max_size);
1184         config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1185                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1186         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1187                 (config->hw_vlan_strip ? "" : "not "));
1188         config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1189                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1190 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1191         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1192 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1193         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1194                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1195 #endif
1196         if (config->hw_padding && !hw_padding) {
1197                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1198                 config->hw_padding = 0;
1199         } else if (config->hw_padding) {
1200                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1201         }
1202         config->tso = (sh->device_attr.max_tso > 0 &&
1203                       (sh->device_attr.tso_supported_qpts &
1204                        (1 << IBV_QPT_RAW_PACKET)));
1205         if (config->tso)
1206                 config->tso_max_payload_sz = sh->device_attr.max_tso;
1207         /*
1208          * MPW is disabled by default, while the Enhanced MPW is enabled
1209          * by default.
1210          */
1211         if (config->mps == MLX5_ARG_UNSET)
1212                 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1213                                                           MLX5_MPW_DISABLED;
1214         else
1215                 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1216         DRV_LOG(INFO, "%sMPS is %s",
1217                 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1218                 config->mps == MLX5_MPW ? "legacy " : "",
1219                 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1220         if (config->devx) {
1221                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1222                 if (err) {
1223                         err = -err;
1224                         goto error;
1225                 }
1226                 /* Check relax ordering support. */
1227                 if (!haswell_broadwell_cpu) {
1228                         sh->cmng.relaxed_ordering_write =
1229                                 config->hca_attr.relaxed_ordering_write;
1230                         sh->cmng.relaxed_ordering_read =
1231                                 config->hca_attr.relaxed_ordering_read;
1232                 } else {
1233                         sh->cmng.relaxed_ordering_read = 0;
1234                         sh->cmng.relaxed_ordering_write = 0;
1235                 }
1236                 sh->rq_ts_format = config->hca_attr.rq_ts_format;
1237                 sh->sq_ts_format = config->hca_attr.sq_ts_format;
1238                 sh->qp_ts_format = config->hca_attr.qp_ts_format;
1239                 /* Check for LRO support. */
1240                 if (config->dest_tir && config->hca_attr.lro_cap &&
1241                     config->dv_flow_en) {
1242                         /* TBD check tunnel lro caps. */
1243                         config->lro.supported = config->hca_attr.lro_cap;
1244                         DRV_LOG(DEBUG, "Device supports LRO");
1245                         /*
1246                          * If LRO timeout is not configured by application,
1247                          * use the minimal supported value.
1248                          */
1249                         if (!config->lro.timeout)
1250                                 config->lro.timeout =
1251                                 config->hca_attr.lro_timer_supported_periods[0];
1252                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1253                                 config->lro.timeout);
1254                         DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1255                                 "required for coalescing is %d bytes",
1256                                 config->hca_attr.lro_min_mss_size);
1257                 }
1258 #if defined(HAVE_MLX5DV_DR) && \
1259         (defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER) || \
1260          defined(HAVE_MLX5_DR_CREATE_ACTION_ASO))
1261                 if (config->hca_attr.qos.sup &&
1262                     config->hca_attr.qos.flow_meter_old &&
1263                     config->dv_flow_en) {
1264                         uint8_t reg_c_mask =
1265                                 config->hca_attr.qos.flow_meter_reg_c_ids;
1266                         /*
1267                          * Meter needs two REG_C's for color match and pre-sfx
1268                          * flow match. Here get the REG_C for color match.
1269                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
1270                          */
1271                         reg_c_mask &= 0xfc;
1272                         if (__builtin_popcount(reg_c_mask) < 1) {
1273                                 priv->mtr_en = 0;
1274                                 DRV_LOG(WARNING, "No available register for"
1275                                         " meter.");
1276                         } else {
1277                                 /*
1278                                  * The meter color register is used by the
1279                                  * flow-hit feature as well.
1280                                  * The flow-hit feature must use REG_C_3
1281                                  * Prefer REG_C_3 if it is available.
1282                                  */
1283                                 if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1284                                         priv->mtr_color_reg = REG_C_3;
1285                                 else
1286                                         priv->mtr_color_reg = ffs(reg_c_mask)
1287                                                               - 1 + REG_C_0;
1288                                 priv->mtr_en = 1;
1289                                 priv->mtr_reg_share =
1290                                       config->hca_attr.qos.flow_meter;
1291                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1292                                         priv->mtr_color_reg);
1293                         }
1294                 }
1295                 if (config->hca_attr.qos.sup &&
1296                         config->hca_attr.qos.flow_meter_aso_sup) {
1297                         uint32_t log_obj_size =
1298                                 rte_log2_u32(MLX5_ASO_MTRS_PER_POOL >> 1);
1299                         if (log_obj_size >=
1300                         config->hca_attr.qos.log_meter_aso_granularity &&
1301                         log_obj_size <=
1302                         config->hca_attr.qos.log_meter_aso_max_alloc) {
1303                                 sh->meter_aso_en = 1;
1304                                 err = mlx5_aso_flow_mtrs_mng_init(priv);
1305                                 if (err) {
1306                                         err = -err;
1307                                         goto error;
1308                                 }
1309                         }
1310                 }
1311 #endif
1312 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1313                 if (config->hca_attr.flow_hit_aso &&
1314                     priv->mtr_color_reg == REG_C_3) {
1315                         sh->flow_hit_aso_en = 1;
1316                         err = mlx5_flow_aso_age_mng_init(sh);
1317                         if (err) {
1318                                 err = -err;
1319                                 goto error;
1320                         }
1321                         DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1322                 }
1323 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1324 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1325                 if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1326                     config->dv_flow_en) {
1327                         priv->sampler_en = 1;
1328                         DRV_LOG(DEBUG, "Sampler enabled!");
1329                 } else {
1330                         priv->sampler_en = 0;
1331                         if (!config->hca_attr.log_max_ft_sampler_num)
1332                                 DRV_LOG(WARNING,
1333                                         "No available register for sampler.");
1334                         else
1335                                 DRV_LOG(DEBUG, "DV flow is not supported!");
1336                 }
1337 #endif
1338         }
1339         if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1340             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1341                 DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1342                 config->cqe_comp = 0;
1343         }
1344         if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1345             (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1346                 DRV_LOG(WARNING, "Flow Tag CQE compression"
1347                                  " format isn't supported.");
1348                 config->cqe_comp = 0;
1349         }
1350         if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1351             (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1352                 DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1353                                  " format isn't supported.");
1354                 config->cqe_comp = 0;
1355         }
1356         DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1357                         config->cqe_comp ? "" : "not ");
1358         if (config->tx_pp) {
1359                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1360                         config->hca_attr.dev_freq_khz);
1361                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1362                         config->hca_attr.qos.packet_pacing ? "" : "not ");
1363                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1364                         config->hca_attr.cross_channel ? "" : "not ");
1365                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1366                         config->hca_attr.wqe_index_ignore ? "" : "not ");
1367                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1368                         config->hca_attr.non_wire_sq ? "" : "not ");
1369                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1370                         config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1371                         config->hca_attr.log_max_static_sq_wq);
1372                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1373                         config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1374                 if (!config->devx) {
1375                         DRV_LOG(ERR, "DevX is required for packet pacing");
1376                         err = ENODEV;
1377                         goto error;
1378                 }
1379                 if (!config->hca_attr.qos.packet_pacing) {
1380                         DRV_LOG(ERR, "Packet pacing is not supported");
1381                         err = ENODEV;
1382                         goto error;
1383                 }
1384                 if (!config->hca_attr.cross_channel) {
1385                         DRV_LOG(ERR, "Cross channel operations are"
1386                                      " required for packet pacing");
1387                         err = ENODEV;
1388                         goto error;
1389                 }
1390                 if (!config->hca_attr.wqe_index_ignore) {
1391                         DRV_LOG(ERR, "WQE index ignore feature is"
1392                                      " required for packet pacing");
1393                         err = ENODEV;
1394                         goto error;
1395                 }
1396                 if (!config->hca_attr.non_wire_sq) {
1397                         DRV_LOG(ERR, "Non-wire SQ feature is"
1398                                      " required for packet pacing");
1399                         err = ENODEV;
1400                         goto error;
1401                 }
1402                 if (!config->hca_attr.log_max_static_sq_wq) {
1403                         DRV_LOG(ERR, "Static WQE SQ feature is"
1404                                      " required for packet pacing");
1405                         err = ENODEV;
1406                         goto error;
1407                 }
1408                 if (!config->hca_attr.qos.wqe_rate_pp) {
1409                         DRV_LOG(ERR, "WQE rate mode is required"
1410                                      " for packet pacing");
1411                         err = ENODEV;
1412                         goto error;
1413                 }
1414 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1415                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1416                              " can't create queues for packet pacing");
1417                 err = ENODEV;
1418                 goto error;
1419 #endif
1420         }
1421         if (config->devx) {
1422                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1423
1424                 err = config->hca_attr.access_register_user ?
1425                         mlx5_devx_cmd_register_read
1426                                 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1427                                 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1428                 if (!err) {
1429                         uint32_t ts_mode;
1430
1431                         /* MTUTC register is read successfully. */
1432                         ts_mode = MLX5_GET(register_mtutc, reg,
1433                                            time_stamp_mode);
1434                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1435                                 config->rt_timestamp = 1;
1436                 } else {
1437                         /* Kernel does not support register reading. */
1438                         if (config->hca_attr.dev_freq_khz ==
1439                                                  (NS_PER_S / MS_PER_S))
1440                                 config->rt_timestamp = 1;
1441                 }
1442         }
1443         /*
1444          * If HW has bug working with tunnel packet decapsulation and
1445          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1446          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1447          */
1448         if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1449                 config->hw_fcs_strip = 0;
1450         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1451                 (config->hw_fcs_strip ? "" : "not "));
1452         if (config->mprq.enabled && mprq) {
1453                 if (config->mprq.stride_num_n &&
1454                     (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1455                      config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1456                         config->mprq.stride_num_n =
1457                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1458                                                 mprq_min_stride_num_n),
1459                                         mprq_max_stride_num_n);
1460                         DRV_LOG(WARNING,
1461                                 "the number of strides"
1462                                 " for Multi-Packet RQ is out of range,"
1463                                 " setting default value (%u)",
1464                                 1 << config->mprq.stride_num_n);
1465                 }
1466                 if (config->mprq.stride_size_n &&
1467                     (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1468                      config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1469                         config->mprq.stride_size_n =
1470                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1471                                                 mprq_min_stride_size_n),
1472                                         mprq_max_stride_size_n);
1473                         DRV_LOG(WARNING,
1474                                 "the size of a stride"
1475                                 " for Multi-Packet RQ is out of range,"
1476                                 " setting default value (%u)",
1477                                 1 << config->mprq.stride_size_n);
1478                 }
1479                 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1480                 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1481         } else if (config->mprq.enabled && !mprq) {
1482                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1483                 config->mprq.enabled = 0;
1484         }
1485         if (config->max_dump_files_num == 0)
1486                 config->max_dump_files_num = 128;
1487         eth_dev = rte_eth_dev_allocate(name);
1488         if (eth_dev == NULL) {
1489                 DRV_LOG(ERR, "can not allocate rte ethdev");
1490                 err = ENOMEM;
1491                 goto error;
1492         }
1493         if (priv->representor) {
1494                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1495                 eth_dev->data->representor_id = priv->representor_id;
1496         }
1497         priv->mp_id.port_id = eth_dev->data->port_id;
1498         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1499         /*
1500          * Store associated network device interface index. This index
1501          * is permanent throughout the lifetime of device. So, we may store
1502          * the ifindex here and use the cached value further.
1503          */
1504         MLX5_ASSERT(spawn->ifindex);
1505         priv->if_index = spawn->ifindex;
1506         eth_dev->data->dev_private = priv;
1507         priv->dev_data = eth_dev->data;
1508         eth_dev->data->mac_addrs = priv->mac;
1509         eth_dev->device = dpdk_dev;
1510         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1511         /* Configure the first MAC address by default. */
1512         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1513                 DRV_LOG(ERR,
1514                         "port %u cannot get MAC address, is mlx5_en"
1515                         " loaded? (errno: %s)",
1516                         eth_dev->data->port_id, strerror(rte_errno));
1517                 err = ENODEV;
1518                 goto error;
1519         }
1520         DRV_LOG(INFO,
1521                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1522                 eth_dev->data->port_id,
1523                 mac.addr_bytes[0], mac.addr_bytes[1],
1524                 mac.addr_bytes[2], mac.addr_bytes[3],
1525                 mac.addr_bytes[4], mac.addr_bytes[5]);
1526 #ifdef RTE_LIBRTE_MLX5_DEBUG
1527         {
1528                 char ifname[MLX5_NAMESIZE];
1529
1530                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1531                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1532                                 eth_dev->data->port_id, ifname);
1533                 else
1534                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1535                                 eth_dev->data->port_id);
1536         }
1537 #endif
1538         /* Get actual MTU if possible. */
1539         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1540         if (err) {
1541                 err = rte_errno;
1542                 goto error;
1543         }
1544         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1545                 priv->mtu);
1546         /* Initialize burst functions to prevent crashes before link-up. */
1547         eth_dev->rx_pkt_burst = removed_rx_burst;
1548         eth_dev->tx_pkt_burst = removed_tx_burst;
1549         eth_dev->dev_ops = &mlx5_dev_ops;
1550         eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1551         eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1552         eth_dev->rx_queue_count = mlx5_rx_queue_count;
1553         /* Register MAC address. */
1554         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1555         if (config->vf && config->vf_nl_en)
1556                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1557                                       mlx5_ifindex(eth_dev),
1558                                       eth_dev->data->mac_addrs,
1559                                       MLX5_MAX_MAC_ADDRESSES);
1560         priv->flows = 0;
1561         priv->ctrl_flows = 0;
1562         rte_spinlock_init(&priv->flow_list_lock);
1563         TAILQ_INIT(&priv->flow_meters);
1564         TAILQ_INIT(&priv->flow_meter_profiles);
1565         /* Hint libmlx5 to use PMD allocator for data plane resources */
1566         mlx5_glue->dv_set_context_attr(sh->ctx,
1567                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1568                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1569                                 .alloc = &mlx5_alloc_verbs_buf,
1570                                 .free = &mlx5_free_verbs_buf,
1571                                 .data = sh,
1572                         }));
1573         /* Bring Ethernet device up. */
1574         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1575                 eth_dev->data->port_id);
1576         mlx5_set_link_up(eth_dev);
1577         /*
1578          * Even though the interrupt handler is not installed yet,
1579          * interrupts will still trigger on the async_fd from
1580          * Verbs context returned by ibv_open_device().
1581          */
1582         mlx5_link_update(eth_dev, 0);
1583 #ifdef HAVE_MLX5DV_DR_ESWITCH
1584         if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1585               (switch_info->representor || switch_info->master)))
1586                 config->dv_esw_en = 0;
1587 #else
1588         config->dv_esw_en = 0;
1589 #endif
1590         /* Detect minimal data bytes to inline. */
1591         mlx5_set_min_inline(spawn, config);
1592         /* Store device configuration on private structure. */
1593         priv->config = *config;
1594         /* Create context for virtual machine VLAN workaround. */
1595         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1596         if (config->dv_flow_en) {
1597                 err = mlx5_alloc_shared_dr(priv);
1598                 if (err)
1599                         goto error;
1600         }
1601         if (config->devx && config->dv_flow_en && config->dest_tir) {
1602                 priv->obj_ops = devx_obj_ops;
1603                 priv->obj_ops.drop_action_create =
1604                                                 ibv_obj_ops.drop_action_create;
1605                 priv->obj_ops.drop_action_destroy =
1606                                                 ibv_obj_ops.drop_action_destroy;
1607 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1608                 priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1609 #else
1610                 if (config->dv_esw_en)
1611                         priv->obj_ops.txq_obj_modify =
1612                                                 ibv_obj_ops.txq_obj_modify;
1613 #endif
1614                 /* Use specific wrappers for Tx object. */
1615                 priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1616                 priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1617                 mlx5_queue_counter_id_prepare(eth_dev);
1618
1619         } else {
1620                 priv->obj_ops = ibv_obj_ops;
1621         }
1622         priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1623         if (!priv->drop_queue.hrxq)
1624                 goto error;
1625         /* Supported Verbs flow priority number detection. */
1626         err = mlx5_flow_discover_priorities(eth_dev);
1627         if (err < 0) {
1628                 err = -err;
1629                 goto error;
1630         }
1631         priv->config.flow_prio = err;
1632         if (!priv->config.dv_esw_en &&
1633             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1634                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1635                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1636                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1637         }
1638         mlx5_set_metadata_mask(eth_dev);
1639         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1640             !priv->sh->dv_regc0_mask) {
1641                 DRV_LOG(ERR, "metadata mode %u is not supported "
1642                              "(no metadata reg_c[0] is available)",
1643                              priv->config.dv_xmeta_en);
1644                         err = ENOTSUP;
1645                         goto error;
1646         }
1647         mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1648                              mlx5_hrxq_create_cb,
1649                              mlx5_hrxq_match_cb,
1650                              mlx5_hrxq_remove_cb);
1651         /* Query availability of metadata reg_c's. */
1652         err = mlx5_flow_discover_mreg_c(eth_dev);
1653         if (err < 0) {
1654                 err = -err;
1655                 goto error;
1656         }
1657         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1658                 DRV_LOG(DEBUG,
1659                         "port %u extensive metadata register is not supported",
1660                         eth_dev->data->port_id);
1661                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1662                         DRV_LOG(ERR, "metadata mode %u is not supported "
1663                                      "(no metadata registers available)",
1664                                      priv->config.dv_xmeta_en);
1665                         err = ENOTSUP;
1666                         goto error;
1667                 }
1668         }
1669         if (priv->config.dv_flow_en &&
1670             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1671             mlx5_flow_ext_mreg_supported(eth_dev) &&
1672             priv->sh->dv_regc0_mask) {
1673                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1674                                                       MLX5_FLOW_MREG_HTABLE_SZ,
1675                                                       0, 0,
1676                                                       flow_dv_mreg_create_cb,
1677                                                       flow_dv_mreg_match_cb,
1678                                                       flow_dv_mreg_remove_cb);
1679                 if (!priv->mreg_cp_tbl) {
1680                         err = ENOMEM;
1681                         goto error;
1682                 }
1683                 priv->mreg_cp_tbl->ctx = eth_dev;
1684         }
1685         rte_spinlock_init(&priv->shared_act_sl);
1686         mlx5_flow_counter_mode_config(eth_dev);
1687         if (priv->config.dv_flow_en)
1688                 eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1689         return eth_dev;
1690 error:
1691         if (priv) {
1692                 if (priv->mreg_cp_tbl)
1693                         mlx5_hlist_destroy(priv->mreg_cp_tbl);
1694                 if (priv->sh)
1695                         mlx5_os_free_shared_dr(priv);
1696                 if (priv->nl_socket_route >= 0)
1697                         close(priv->nl_socket_route);
1698                 if (priv->nl_socket_rdma >= 0)
1699                         close(priv->nl_socket_rdma);
1700                 if (priv->vmwa_context)
1701                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1702                 if (eth_dev && priv->drop_queue.hrxq)
1703                         mlx5_drop_action_destroy(eth_dev);
1704                 if (own_domain_id)
1705                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1706                 mlx5_cache_list_destroy(&priv->hrxqs);
1707                 mlx5_free(priv);
1708                 if (eth_dev != NULL)
1709                         eth_dev->data->dev_private = NULL;
1710         }
1711         if (eth_dev != NULL) {
1712                 /* mac_addrs must not be freed alone because part of
1713                  * dev_private
1714                  **/
1715                 eth_dev->data->mac_addrs = NULL;
1716                 rte_eth_dev_release_port(eth_dev);
1717         }
1718         if (sh)
1719                 mlx5_free_shared_dev_ctx(sh);
1720         MLX5_ASSERT(err > 0);
1721         rte_errno = err;
1722         return NULL;
1723 }
1724
1725 /**
1726  * Comparison callback to sort device data.
1727  *
1728  * This is meant to be used with qsort().
1729  *
1730  * @param a[in]
1731  *   Pointer to pointer to first data object.
1732  * @param b[in]
1733  *   Pointer to pointer to second data object.
1734  *
1735  * @return
1736  *   0 if both objects are equal, less than 0 if the first argument is less
1737  *   than the second, greater than 0 otherwise.
1738  */
1739 static int
1740 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1741 {
1742         const struct mlx5_switch_info *si_a =
1743                 &((const struct mlx5_dev_spawn_data *)a)->info;
1744         const struct mlx5_switch_info *si_b =
1745                 &((const struct mlx5_dev_spawn_data *)b)->info;
1746         int ret;
1747
1748         /* Master device first. */
1749         ret = si_b->master - si_a->master;
1750         if (ret)
1751                 return ret;
1752         /* Then representor devices. */
1753         ret = si_b->representor - si_a->representor;
1754         if (ret)
1755                 return ret;
1756         /* Unidentified devices come last in no specific order. */
1757         if (!si_a->representor)
1758                 return 0;
1759         /* Order representors by name. */
1760         return si_a->port_name - si_b->port_name;
1761 }
1762
1763 /**
1764  * Match PCI information for possible slaves of bonding device.
1765  *
1766  * @param[in] ibv_dev
1767  *   Pointer to Infiniband device structure.
1768  * @param[in] pci_dev
1769  *   Pointer to primary PCI address structure to match.
1770  * @param[in] nl_rdma
1771  *   Netlink RDMA group socket handle.
1772  * @param[in] owner
1773  *   Rerepsentor owner PF index.
1774  * @param[out] bond_info
1775  *   Pointer to bonding information.
1776  *
1777  * @return
1778  *   negative value if no bonding device found, otherwise
1779  *   positive index of slave PF in bonding.
1780  */
1781 static int
1782 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1783                            const struct rte_pci_addr *pci_dev,
1784                            int nl_rdma, uint16_t owner,
1785                            struct mlx5_bond_info *bond_info)
1786 {
1787         char ifname[IF_NAMESIZE + 1];
1788         unsigned int ifindex;
1789         unsigned int np, i;
1790         FILE *bond_file = NULL, *file;
1791         int pf = -1;
1792         int ret;
1793
1794         /*
1795          * Try to get master device name. If something goes
1796          * wrong suppose the lack of kernel support and no
1797          * bonding devices.
1798          */
1799         memset(bond_info, 0, sizeof(*bond_info));
1800         if (nl_rdma < 0)
1801                 return -1;
1802         if (!strstr(ibv_dev->name, "bond"))
1803                 return -1;
1804         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1805         if (!np)
1806                 return -1;
1807         /*
1808          * The Master device might not be on the predefined
1809          * port (not on port index 1, it is not garanted),
1810          * we have to scan all Infiniband device port and
1811          * find master.
1812          */
1813         for (i = 1; i <= np; ++i) {
1814                 /* Check whether Infiniband port is populated. */
1815                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1816                 if (!ifindex)
1817                         continue;
1818                 if (!if_indextoname(ifindex, ifname))
1819                         continue;
1820                 /* Try to read bonding slave names from sysfs. */
1821                 MKSTR(slaves,
1822                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1823                 bond_file = fopen(slaves, "r");
1824                 if (bond_file)
1825                         break;
1826         }
1827         if (!bond_file)
1828                 return -1;
1829         /* Use safe format to check maximal buffer length. */
1830         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1831         while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1832                 char tmp_str[IF_NAMESIZE + 32];
1833                 struct rte_pci_addr pci_addr;
1834                 struct mlx5_switch_info info;
1835
1836                 /* Process slave interface names in the loop. */
1837                 snprintf(tmp_str, sizeof(tmp_str),
1838                          "/sys/class/net/%s", ifname);
1839                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1840                         DRV_LOG(WARNING, "can not get PCI address"
1841                                          " for netdev \"%s\"", ifname);
1842                         continue;
1843                 }
1844                 /* Slave interface PCI address match found. */
1845                 snprintf(tmp_str, sizeof(tmp_str),
1846                          "/sys/class/net/%s/phys_port_name", ifname);
1847                 file = fopen(tmp_str, "rb");
1848                 if (!file)
1849                         break;
1850                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1851                 if (fscanf(file, "%32s", tmp_str) == 1)
1852                         mlx5_translate_port_name(tmp_str, &info);
1853                 fclose(file);
1854                 /* Only process PF ports. */
1855                 if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1856                     info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1857                         continue;
1858                 /* Check max bonding member. */
1859                 if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1860                         DRV_LOG(WARNING, "bonding index out of range, "
1861                                 "please increase MLX5_BOND_MAX_PORTS: %s",
1862                                 tmp_str);
1863                         break;
1864                 }
1865                 /* Match PCI address. */
1866                 if (pci_dev->domain == pci_addr.domain &&
1867                     pci_dev->bus == pci_addr.bus &&
1868                     pci_dev->devid == pci_addr.devid &&
1869                     pci_dev->function + owner == pci_addr.function)
1870                         pf = info.port_name;
1871                 /* Get ifindex. */
1872                 snprintf(tmp_str, sizeof(tmp_str),
1873                          "/sys/class/net/%s/ifindex", ifname);
1874                 file = fopen(tmp_str, "rb");
1875                 if (!file)
1876                         break;
1877                 ret = fscanf(file, "%u", &ifindex);
1878                 fclose(file);
1879                 if (ret != 1)
1880                         break;
1881                 /* Save bonding info. */
1882                 strncpy(bond_info->ports[info.port_name].ifname, ifname,
1883                         sizeof(bond_info->ports[0].ifname));
1884                 bond_info->ports[info.port_name].pci_addr = pci_addr;
1885                 bond_info->ports[info.port_name].ifindex = ifindex;
1886                 bond_info->n_port++;
1887         }
1888         if (pf >= 0) {
1889                 /* Get bond interface info */
1890                 ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1891                                            bond_info->ifname);
1892                 if (ret)
1893                         DRV_LOG(ERR, "unable to get bond info: %s",
1894                                 strerror(rte_errno));
1895                 else
1896                         DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1897                                 ifindex, bond_info->ifindex, bond_info->ifname);
1898         }
1899         return pf;
1900 }
1901
1902 /**
1903  * Register a PCI device within bonding.
1904  *
1905  * This function spawns Ethernet devices out of a given PCI device and
1906  * bonding owner PF index.
1907  *
1908  * @param[in] pci_dev
1909  *   PCI device information.
1910  * @param[in] req_eth_da
1911  *   Requested ethdev device argument.
1912  * @param[in] owner_id
1913  *   Requested owner PF port ID within bonding device, default to 0.
1914  *
1915  * @return
1916  *   0 on success, a negative errno value otherwise and rte_errno is set.
1917  */
1918 static int
1919 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1920                      struct rte_eth_devargs *req_eth_da,
1921                      uint16_t owner_id)
1922 {
1923         struct ibv_device **ibv_list;
1924         /*
1925          * Number of found IB Devices matching with requested PCI BDF.
1926          * nd != 1 means there are multiple IB devices over the same
1927          * PCI device and we have representors and master.
1928          */
1929         unsigned int nd = 0;
1930         /*
1931          * Number of found IB device Ports. nd = 1 and np = 1..n means
1932          * we have the single multiport IB device, and there may be
1933          * representors attached to some of found ports.
1934          */
1935         unsigned int np = 0;
1936         /*
1937          * Number of DPDK ethernet devices to Spawn - either over
1938          * multiple IB devices or multiple ports of single IB device.
1939          * Actually this is the number of iterations to spawn.
1940          */
1941         unsigned int ns = 0;
1942         /*
1943          * Bonding device
1944          *   < 0 - no bonding device (single one)
1945          *  >= 0 - bonding device (value is slave PF index)
1946          */
1947         int bd = -1;
1948         struct mlx5_dev_spawn_data *list = NULL;
1949         struct mlx5_dev_config dev_config;
1950         unsigned int dev_config_vf;
1951         struct rte_eth_devargs eth_da = *req_eth_da;
1952         struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1953         struct mlx5_bond_info bond_info;
1954         int ret = -1;
1955
1956         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1957                 mlx5_pmd_socket_init();
1958         ret = mlx5_init_once();
1959         if (ret) {
1960                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1961                         strerror(rte_errno));
1962                 return -rte_errno;
1963         }
1964         errno = 0;
1965         ibv_list = mlx5_glue->get_device_list(&ret);
1966         if (!ibv_list) {
1967                 rte_errno = errno ? errno : ENOSYS;
1968                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1969                 return -rte_errno;
1970         }
1971         /*
1972          * First scan the list of all Infiniband devices to find
1973          * matching ones, gathering into the list.
1974          */
1975         struct ibv_device *ibv_match[ret + 1];
1976         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1977         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1978         unsigned int i;
1979
1980         while (ret-- > 0) {
1981                 struct rte_pci_addr pci_addr;
1982
1983                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1984                 bd = mlx5_device_bond_pci_match
1985                                 (ibv_list[ret], &owner_pci, nl_rdma, owner_id,
1986                                  &bond_info);
1987                 if (bd >= 0) {
1988                         /*
1989                          * Bonding device detected. Only one match is allowed,
1990                          * the bonding is supported over multi-port IB device,
1991                          * there should be no matches on representor PCI
1992                          * functions or non VF LAG bonding devices with
1993                          * specified address.
1994                          */
1995                         if (nd) {
1996                                 DRV_LOG(ERR,
1997                                         "multiple PCI match on bonding device"
1998                                         "\"%s\" found", ibv_list[ret]->name);
1999                                 rte_errno = ENOENT;
2000                                 ret = -rte_errno;
2001                                 goto exit;
2002                         }
2003                         /* Amend owner pci address if owner PF ID specified. */
2004                         if (eth_da.nb_representor_ports)
2005                                 owner_pci.function += owner_id;
2006                         DRV_LOG(INFO, "PCI information matches for"
2007                                       " slave %d bonding device \"%s\"",
2008                                       bd, ibv_list[ret]->name);
2009                         ibv_match[nd++] = ibv_list[ret];
2010                         break;
2011                 } else {
2012                         /* Bonding device not found. */
2013                         if (mlx5_dev_to_pci_addr
2014                                 (ibv_list[ret]->ibdev_path, &pci_addr))
2015                                 continue;
2016                         if (owner_pci.domain != pci_addr.domain ||
2017                             owner_pci.bus != pci_addr.bus ||
2018                             owner_pci.devid != pci_addr.devid ||
2019                             owner_pci.function != pci_addr.function)
2020                                 continue;
2021                         DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2022                                 ibv_list[ret]->name);
2023                         ibv_match[nd++] = ibv_list[ret];
2024                 }
2025         }
2026         ibv_match[nd] = NULL;
2027         if (!nd) {
2028                 /* No device matches, just complain and bail out. */
2029                 DRV_LOG(WARNING,
2030                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
2031                         " are kernel drivers loaded?",
2032                         owner_pci.domain, owner_pci.bus,
2033                         owner_pci.devid, owner_pci.function);
2034                 rte_errno = ENOENT;
2035                 ret = -rte_errno;
2036                 goto exit;
2037         }
2038         if (nd == 1) {
2039                 /*
2040                  * Found single matching device may have multiple ports.
2041                  * Each port may be representor, we have to check the port
2042                  * number and check the representors existence.
2043                  */
2044                 if (nl_rdma >= 0)
2045                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2046                 if (!np)
2047                         DRV_LOG(WARNING, "can not get IB device \"%s\""
2048                                          " ports number", ibv_match[0]->name);
2049                 if (bd >= 0 && !np) {
2050                         DRV_LOG(ERR, "can not get ports"
2051                                      " for bonding device");
2052                         rte_errno = ENOENT;
2053                         ret = -rte_errno;
2054                         goto exit;
2055                 }
2056         }
2057 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2058         if (bd >= 0) {
2059                 /*
2060                  * This may happen if there is VF LAG kernel support and
2061                  * application is compiled with older rdma_core library.
2062                  */
2063                 DRV_LOG(ERR,
2064                         "No kernel/verbs support for VF LAG bonding found.");
2065                 rte_errno = ENOTSUP;
2066                 ret = -rte_errno;
2067                 goto exit;
2068         }
2069 #endif
2070         /*
2071          * Now we can determine the maximal
2072          * amount of devices to be spawned.
2073          */
2074         list = mlx5_malloc(MLX5_MEM_ZERO,
2075                            sizeof(struct mlx5_dev_spawn_data) *
2076                            (np ? np : nd),
2077                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2078         if (!list) {
2079                 DRV_LOG(ERR, "spawn data array allocation failure");
2080                 rte_errno = ENOMEM;
2081                 ret = -rte_errno;
2082                 goto exit;
2083         }
2084         if (bd >= 0 || np > 1) {
2085                 /*
2086                  * Single IB device with multiple ports found,
2087                  * it may be E-Switch master device and representors.
2088                  * We have to perform identification through the ports.
2089                  */
2090                 MLX5_ASSERT(nl_rdma >= 0);
2091                 MLX5_ASSERT(ns == 0);
2092                 MLX5_ASSERT(nd == 1);
2093                 MLX5_ASSERT(np);
2094                 for (i = 1; i <= np; ++i) {
2095                         list[ns].bond_info = &bond_info;
2096                         list[ns].max_port = np;
2097                         list[ns].phys_port = i;
2098                         list[ns].phys_dev = ibv_match[0];
2099                         list[ns].eth_dev = NULL;
2100                         list[ns].pci_dev = pci_dev;
2101                         list[ns].pf_bond = bd;
2102                         list[ns].ifindex = mlx5_nl_ifindex
2103                                 (nl_rdma,
2104                                 mlx5_os_get_dev_device_name
2105                                                 (list[ns].phys_dev), i);
2106                         if (!list[ns].ifindex) {
2107                                 /*
2108                                  * No network interface index found for the
2109                                  * specified port, it means there is no
2110                                  * representor on this port. It's OK,
2111                                  * there can be disabled ports, for example
2112                                  * if sriov_numvfs < sriov_totalvfs.
2113                                  */
2114                                 continue;
2115                         }
2116                         ret = -1;
2117                         if (nl_route >= 0)
2118                                 ret = mlx5_nl_switch_info
2119                                                (nl_route,
2120                                                 list[ns].ifindex,
2121                                                 &list[ns].info);
2122                         if (ret || (!list[ns].info.representor &&
2123                                     !list[ns].info.master)) {
2124                                 /*
2125                                  * We failed to recognize representors with
2126                                  * Netlink, let's try to perform the task
2127                                  * with sysfs.
2128                                  */
2129                                 ret =  mlx5_sysfs_switch_info
2130                                                 (list[ns].ifindex,
2131                                                  &list[ns].info);
2132                         }
2133 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2134                         if (!ret && bd >= 0) {
2135                                 switch (list[ns].info.name_type) {
2136                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2137                                         if (list[ns].info.port_name == bd)
2138                                                 ns++;
2139                                         break;
2140                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2141                                         /* Fallthrough */
2142                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2143                                         /* Fallthrough */
2144                                 case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2145                                         if (list[ns].info.pf_num == bd)
2146                                                 ns++;
2147                                         break;
2148                                 default:
2149                                         break;
2150                                 }
2151                                 continue;
2152                         }
2153 #endif
2154                         if (!ret && (list[ns].info.representor ^
2155                                      list[ns].info.master))
2156                                 ns++;
2157                 }
2158                 if (!ns) {
2159                         DRV_LOG(ERR,
2160                                 "unable to recognize master/representors"
2161                                 " on the IB device with multiple ports");
2162                         rte_errno = ENOENT;
2163                         ret = -rte_errno;
2164                         goto exit;
2165                 }
2166         } else {
2167                 /*
2168                  * The existence of several matching entries (nd > 1) means
2169                  * port representors have been instantiated. No existing Verbs
2170                  * call nor sysfs entries can tell them apart, this can only
2171                  * be done through Netlink calls assuming kernel drivers are
2172                  * recent enough to support them.
2173                  *
2174                  * In the event of identification failure through Netlink,
2175                  * try again through sysfs, then:
2176                  *
2177                  * 1. A single IB device matches (nd == 1) with single
2178                  *    port (np=0/1) and is not a representor, assume
2179                  *    no switch support.
2180                  *
2181                  * 2. Otherwise no safe assumptions can be made;
2182                  *    complain louder and bail out.
2183                  */
2184                 for (i = 0; i != nd; ++i) {
2185                         memset(&list[ns].info, 0, sizeof(list[ns].info));
2186                         list[ns].bond_info = NULL;
2187                         list[ns].max_port = 1;
2188                         list[ns].phys_port = 1;
2189                         list[ns].phys_dev = ibv_match[i];
2190                         list[ns].eth_dev = NULL;
2191                         list[ns].pci_dev = pci_dev;
2192                         list[ns].pf_bond = -1;
2193                         list[ns].ifindex = 0;
2194                         if (nl_rdma >= 0)
2195                                 list[ns].ifindex = mlx5_nl_ifindex
2196                                 (nl_rdma,
2197                                 mlx5_os_get_dev_device_name
2198                                                 (list[ns].phys_dev), 1);
2199                         if (!list[ns].ifindex) {
2200                                 char ifname[IF_NAMESIZE];
2201
2202                                 /*
2203                                  * Netlink failed, it may happen with old
2204                                  * ib_core kernel driver (before 4.16).
2205                                  * We can assume there is old driver because
2206                                  * here we are processing single ports IB
2207                                  * devices. Let's try sysfs to retrieve
2208                                  * the ifindex. The method works for
2209                                  * master device only.
2210                                  */
2211                                 if (nd > 1) {
2212                                         /*
2213                                          * Multiple devices found, assume
2214                                          * representors, can not distinguish
2215                                          * master/representor and retrieve
2216                                          * ifindex via sysfs.
2217                                          */
2218                                         continue;
2219                                 }
2220                                 ret = mlx5_get_ifname_sysfs
2221                                         (ibv_match[i]->ibdev_path, ifname);
2222                                 if (!ret)
2223                                         list[ns].ifindex =
2224                                                 if_nametoindex(ifname);
2225                                 if (!list[ns].ifindex) {
2226                                         /*
2227                                          * No network interface index found
2228                                          * for the specified device, it means
2229                                          * there it is neither representor
2230                                          * nor master.
2231                                          */
2232                                         continue;
2233                                 }
2234                         }
2235                         ret = -1;
2236                         if (nl_route >= 0)
2237                                 ret = mlx5_nl_switch_info
2238                                                (nl_route,
2239                                                 list[ns].ifindex,
2240                                                 &list[ns].info);
2241                         if (ret || (!list[ns].info.representor &&
2242                                     !list[ns].info.master)) {
2243                                 /*
2244                                  * We failed to recognize representors with
2245                                  * Netlink, let's try to perform the task
2246                                  * with sysfs.
2247                                  */
2248                                 ret =  mlx5_sysfs_switch_info
2249                                                 (list[ns].ifindex,
2250                                                  &list[ns].info);
2251                         }
2252                         if (!ret && (list[ns].info.representor ^
2253                                      list[ns].info.master)) {
2254                                 ns++;
2255                         } else if ((nd == 1) &&
2256                                    !list[ns].info.representor &&
2257                                    !list[ns].info.master) {
2258                                 /*
2259                                  * Single IB device with
2260                                  * one physical port and
2261                                  * attached network device.
2262                                  * May be SRIOV is not enabled
2263                                  * or there is no representors.
2264                                  */
2265                                 DRV_LOG(INFO, "no E-Switch support detected");
2266                                 ns++;
2267                                 break;
2268                         }
2269                 }
2270                 if (!ns) {
2271                         DRV_LOG(ERR,
2272                                 "unable to recognize master/representors"
2273                                 " on the multiple IB devices");
2274                         rte_errno = ENOENT;
2275                         ret = -rte_errno;
2276                         goto exit;
2277                 }
2278         }
2279         MLX5_ASSERT(ns);
2280         /*
2281          * Sort list to probe devices in natural order for users convenience
2282          * (i.e. master first, then representors from lowest to highest ID).
2283          */
2284         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2285         /* Device specific configuration. */
2286         switch (pci_dev->id.device_id) {
2287         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2288         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2289         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2290         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2291         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2292         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2293         case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2294                 dev_config_vf = 1;
2295                 break;
2296         default:
2297                 dev_config_vf = 0;
2298                 break;
2299         }
2300         if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2301                 /* Set devargs default values. */
2302                 if (eth_da.nb_mh_controllers == 0) {
2303                         eth_da.nb_mh_controllers = 1;
2304                         eth_da.mh_controllers[0] = 0;
2305                 }
2306                 if (eth_da.nb_ports == 0 && ns > 0) {
2307                         if (list[0].pf_bond >= 0 && list[0].info.representor)
2308                                 DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2309                                         pci_dev->device.devargs->args);
2310                         eth_da.nb_ports = 1;
2311                         eth_da.ports[0] = list[0].info.pf_num;
2312                 }
2313                 if (eth_da.nb_representor_ports == 0) {
2314                         eth_da.nb_representor_ports = 1;
2315                         eth_da.representor_ports[0] = 0;
2316                 }
2317         }
2318         for (i = 0; i != ns; ++i) {
2319                 uint32_t restore;
2320
2321                 /* Default configuration. */
2322                 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2323                 dev_config.vf = dev_config_vf;
2324                 dev_config.mps = MLX5_ARG_UNSET;
2325                 dev_config.dbnc = MLX5_ARG_UNSET;
2326                 dev_config.rx_vec_en = 1;
2327                 dev_config.txq_inline_max = MLX5_ARG_UNSET;
2328                 dev_config.txq_inline_min = MLX5_ARG_UNSET;
2329                 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2330                 dev_config.txqs_inline = MLX5_ARG_UNSET;
2331                 dev_config.vf_nl_en = 1;
2332                 dev_config.mr_ext_memseg_en = 1;
2333                 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2334                 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2335                 dev_config.dv_esw_en = 1;
2336                 dev_config.dv_flow_en = 1;
2337                 dev_config.decap_en = 1;
2338                 dev_config.log_hp_size = MLX5_ARG_UNSET;
2339                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2340                                                  &list[i],
2341                                                  &dev_config,
2342                                                  &eth_da);
2343                 if (!list[i].eth_dev) {
2344                         if (rte_errno != EBUSY && rte_errno != EEXIST)
2345                                 break;
2346                         /* Device is disabled or already spawned. Ignore it. */
2347                         continue;
2348                 }
2349                 restore = list[i].eth_dev->data->dev_flags;
2350                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2351                 /* Restore non-PCI flags cleared by the above call. */
2352                 list[i].eth_dev->data->dev_flags |= restore;
2353                 rte_eth_dev_probing_finish(list[i].eth_dev);
2354         }
2355         if (i != ns) {
2356                 DRV_LOG(ERR,
2357                         "probe of PCI device " PCI_PRI_FMT " aborted after"
2358                         " encountering an error: %s",
2359                         owner_pci.domain, owner_pci.bus,
2360                         owner_pci.devid, owner_pci.function,
2361                         strerror(rte_errno));
2362                 ret = -rte_errno;
2363                 /* Roll back. */
2364                 while (i--) {
2365                         if (!list[i].eth_dev)
2366                                 continue;
2367                         mlx5_dev_close(list[i].eth_dev);
2368                         /* mac_addrs must not be freed because in dev_private */
2369                         list[i].eth_dev->data->mac_addrs = NULL;
2370                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2371                 }
2372                 /* Restore original error. */
2373                 rte_errno = -ret;
2374         } else {
2375                 ret = 0;
2376         }
2377 exit:
2378         /*
2379          * Do the routine cleanup:
2380          * - close opened Netlink sockets
2381          * - free allocated spawn data array
2382          * - free the Infiniband device list
2383          */
2384         if (nl_rdma >= 0)
2385                 close(nl_rdma);
2386         if (nl_route >= 0)
2387                 close(nl_route);
2388         if (list)
2389                 mlx5_free(list);
2390         MLX5_ASSERT(ibv_list);
2391         mlx5_glue->free_device_list(ibv_list);
2392         return ret;
2393 }
2394
2395 /**
2396  * DPDK callback to register a PCI device.
2397  *
2398  * This function spawns Ethernet devices out of a given PCI device.
2399  *
2400  * @param[in] pci_drv
2401  *   PCI driver structure (mlx5_driver).
2402  * @param[in] pci_dev
2403  *   PCI device information.
2404  *
2405  * @return
2406  *   0 on success, a negative errno value otherwise and rte_errno is set.
2407  */
2408 int
2409 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2410                   struct rte_pci_device *pci_dev)
2411 {
2412         struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2413         int ret = 0;
2414         uint16_t p;
2415
2416         if (pci_dev->device.devargs) {
2417                 /* Parse representor information from device argument. */
2418                 if (pci_dev->device.devargs->cls_str)
2419                         ret = rte_eth_devargs_parse
2420                                 (pci_dev->device.devargs->cls_str, &eth_da);
2421                 if (ret) {
2422                         DRV_LOG(ERR, "failed to parse device arguments: %s",
2423                                 pci_dev->device.devargs->cls_str);
2424                         return -rte_errno;
2425                 }
2426                 if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2427                         /* Support legacy device argument */
2428                         ret = rte_eth_devargs_parse
2429                                 (pci_dev->device.devargs->args, &eth_da);
2430                         if (ret) {
2431                                 DRV_LOG(ERR, "failed to parse device arguments: %s",
2432                                         pci_dev->device.devargs->args);
2433                                 return -rte_errno;
2434                         }
2435                 }
2436         }
2437
2438         if (eth_da.nb_ports > 0) {
2439                 /* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2440                 for (p = 0; p < eth_da.nb_ports; p++)
2441                         ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2442                                                    eth_da.ports[p]);
2443         } else {
2444                 ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2445         }
2446         return ret;
2447 }
2448
2449 static int
2450 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2451 {
2452         char *env;
2453         int value;
2454
2455         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2456         /* Get environment variable to store. */
2457         env = getenv(MLX5_SHUT_UP_BF);
2458         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2459         if (config->dbnc == MLX5_ARG_UNSET)
2460                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2461         else
2462                 setenv(MLX5_SHUT_UP_BF,
2463                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2464         return value;
2465 }
2466
2467 static void
2468 mlx5_restore_doorbell_mapping_env(int value)
2469 {
2470         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2471         /* Restore the original environment variable state. */
2472         if (value == MLX5_ARG_UNSET)
2473                 unsetenv(MLX5_SHUT_UP_BF);
2474         else
2475                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2476 }
2477
2478 /**
2479  * Extract pdn of PD object using DV API.
2480  *
2481  * @param[in] pd
2482  *   Pointer to the verbs PD object.
2483  * @param[out] pdn
2484  *   Pointer to the PD object number variable.
2485  *
2486  * @return
2487  *   0 on success, error value otherwise.
2488  */
2489 int
2490 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2491 {
2492 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2493         struct mlx5dv_obj obj;
2494         struct mlx5dv_pd pd_info;
2495         int ret = 0;
2496
2497         obj.pd.in = pd;
2498         obj.pd.out = &pd_info;
2499         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2500         if (ret) {
2501                 DRV_LOG(DEBUG, "Fail to get PD object info");
2502                 return ret;
2503         }
2504         *pdn = pd_info.pdn;
2505         return 0;
2506 #else
2507         (void)pd;
2508         (void)pdn;
2509         return -ENOTSUP;
2510 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2511 }
2512
2513 /**
2514  * Function API to open IB device.
2515  *
2516  * This function calls the Linux glue APIs to open a device.
2517  *
2518  * @param[in] spawn
2519  *   Pointer to the IB device attributes (name, port, etc).
2520  * @param[out] config
2521  *   Pointer to device configuration structure.
2522  * @param[out] sh
2523  *   Pointer to shared context structure.
2524  *
2525  * @return
2526  *   0 on success, a positive error value otherwise.
2527  */
2528 int
2529 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2530                      const struct mlx5_dev_config *config,
2531                      struct mlx5_dev_ctx_shared *sh)
2532 {
2533         int dbmap_env;
2534         int err = 0;
2535
2536         sh->numa_node = spawn->pci_dev->device.numa_node;
2537         pthread_mutex_init(&sh->txpp.mutex, NULL);
2538         /*
2539          * Configure environment variable "MLX5_BF_SHUT_UP"
2540          * before the device creation. The rdma_core library
2541          * checks the variable at device creation and
2542          * stores the result internally.
2543          */
2544         dbmap_env = mlx5_config_doorbell_mapping_env(config);
2545         /* Try to open IB device with DV first, then usual Verbs. */
2546         errno = 0;
2547         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2548         if (sh->ctx) {
2549                 sh->devx = 1;
2550                 DRV_LOG(DEBUG, "DevX is supported");
2551                 /* The device is created, no need for environment. */
2552                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2553         } else {
2554                 /* The environment variable is still configured. */
2555                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2556                 err = errno ? errno : ENODEV;
2557                 /*
2558                  * The environment variable is not needed anymore,
2559                  * all device creation attempts are completed.
2560                  */
2561                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2562                 if (!sh->ctx)
2563                         return err;
2564                 DRV_LOG(DEBUG, "DevX is NOT supported");
2565                 err = 0;
2566         }
2567         if (!err && sh->ctx) {
2568                 /* Hint libmlx5 to use PMD allocator for data plane resources */
2569                 mlx5_glue->dv_set_context_attr(sh->ctx,
2570                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2571                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2572                                 .alloc = &mlx5_alloc_verbs_buf,
2573                                 .free = &mlx5_free_verbs_buf,
2574                                 .data = sh,
2575                         }));
2576         }
2577         return err;
2578 }
2579
2580 /**
2581  * Install shared asynchronous device events handler.
2582  * This function is implemented to support event sharing
2583  * between multiple ports of single IB device.
2584  *
2585  * @param sh
2586  *   Pointer to mlx5_dev_ctx_shared object.
2587  */
2588 void
2589 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2590 {
2591         int ret;
2592         int flags;
2593
2594         sh->intr_handle.fd = -1;
2595         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2596         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2597                     F_SETFL, flags | O_NONBLOCK);
2598         if (ret) {
2599                 DRV_LOG(INFO, "failed to change file descriptor async event"
2600                         " queue");
2601         } else {
2602                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2603                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2604                 if (rte_intr_callback_register(&sh->intr_handle,
2605                                         mlx5_dev_interrupt_handler, sh)) {
2606                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
2607                         sh->intr_handle.fd = -1;
2608                 }
2609         }
2610         if (sh->devx) {
2611 #ifdef HAVE_IBV_DEVX_ASYNC
2612                 sh->intr_handle_devx.fd = -1;
2613                 sh->devx_comp =
2614                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2615                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2616                 if (!devx_comp) {
2617                         DRV_LOG(INFO, "failed to allocate devx_comp.");
2618                         return;
2619                 }
2620                 flags = fcntl(devx_comp->fd, F_GETFL);
2621                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2622                 if (ret) {
2623                         DRV_LOG(INFO, "failed to change file descriptor"
2624                                 " devx comp");
2625                         return;
2626                 }
2627                 sh->intr_handle_devx.fd = devx_comp->fd;
2628                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2629                 if (rte_intr_callback_register(&sh->intr_handle_devx,
2630                                         mlx5_dev_interrupt_handler_devx, sh)) {
2631                         DRV_LOG(INFO, "Fail to install the devx shared"
2632                                 " interrupt.");
2633                         sh->intr_handle_devx.fd = -1;
2634                 }
2635 #endif /* HAVE_IBV_DEVX_ASYNC */
2636         }
2637 }
2638
2639 /**
2640  * Uninstall shared asynchronous device events handler.
2641  * This function is implemented to support event sharing
2642  * between multiple ports of single IB device.
2643  *
2644  * @param dev
2645  *   Pointer to mlx5_dev_ctx_shared object.
2646  */
2647 void
2648 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2649 {
2650         if (sh->intr_handle.fd >= 0)
2651                 mlx5_intr_callback_unregister(&sh->intr_handle,
2652                                               mlx5_dev_interrupt_handler, sh);
2653 #ifdef HAVE_IBV_DEVX_ASYNC
2654         if (sh->intr_handle_devx.fd >= 0)
2655                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2656                                   mlx5_dev_interrupt_handler_devx, sh);
2657         if (sh->devx_comp)
2658                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2659 #endif
2660 }
2661
2662 /**
2663  * Read statistics by a named counter.
2664  *
2665  * @param[in] priv
2666  *   Pointer to the private device data structure.
2667  * @param[in] ctr_name
2668  *   Pointer to the name of the statistic counter to read
2669  * @param[out] stat
2670  *   Pointer to read statistic value.
2671  * @return
2672  *   0 on success and stat is valud, 1 if failed to read the value
2673  *   rte_errno is set.
2674  *
2675  */
2676 int
2677 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2678                       uint64_t *stat)
2679 {
2680         int fd;
2681
2682         if (priv->sh) {
2683                 if (priv->q_counters != NULL &&
2684                     strcmp(ctr_name, "out_of_buffer") == 0)
2685                         return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2686                                                            0, (uint32_t *)stat);
2687                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2688                       priv->sh->ibdev_path,
2689                       priv->dev_port,
2690                       ctr_name);
2691                 fd = open(path, O_RDONLY);
2692                 /*
2693                  * in switchdev the file location is not per port
2694                  * but rather in <ibdev_path>/hw_counters/<file_name>.
2695                  */
2696                 if (fd == -1) {
2697                         MKSTR(path1, "%s/hw_counters/%s",
2698                               priv->sh->ibdev_path,
2699                               ctr_name);
2700                         fd = open(path1, O_RDONLY);
2701                 }
2702                 if (fd != -1) {
2703                         char buf[21] = {'\0'};
2704                         ssize_t n = read(fd, buf, sizeof(buf));
2705
2706                         close(fd);
2707                         if (n != -1) {
2708                                 *stat = strtoull(buf, NULL, 10);
2709                                 return 0;
2710                         }
2711                 }
2712         }
2713         *stat = 0;
2714         return 1;
2715 }
2716
2717 /**
2718  * Set the reg_mr and dereg_mr call backs
2719  *
2720  * @param reg_mr_cb[out]
2721  *   Pointer to reg_mr func
2722  * @param dereg_mr_cb[out]
2723  *   Pointer to dereg_mr func
2724  *
2725  */
2726 void
2727 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2728                       mlx5_dereg_mr_t *dereg_mr_cb)
2729 {
2730         *reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2731         *dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2732 }
2733
2734 /**
2735  * Remove a MAC address from device
2736  *
2737  * @param dev
2738  *   Pointer to Ethernet device structure.
2739  * @param index
2740  *   MAC address index.
2741  */
2742 void
2743 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2744 {
2745         struct mlx5_priv *priv = dev->data->dev_private;
2746         const int vf = priv->config.vf;
2747
2748         if (vf)
2749                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2750                                         mlx5_ifindex(dev), priv->mac_own,
2751                                         &dev->data->mac_addrs[index], index);
2752 }
2753
2754 /**
2755  * Adds a MAC address to the device
2756  *
2757  * @param dev
2758  *   Pointer to Ethernet device structure.
2759  * @param mac_addr
2760  *   MAC address to register.
2761  * @param index
2762  *   MAC address index.
2763  *
2764  * @return
2765  *   0 on success, a negative errno value otherwise
2766  */
2767 int
2768 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2769                      uint32_t index)
2770 {
2771         struct mlx5_priv *priv = dev->data->dev_private;
2772         const int vf = priv->config.vf;
2773         int ret = 0;
2774
2775         if (vf)
2776                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2777                                            mlx5_ifindex(dev), priv->mac_own,
2778                                            mac, index);
2779         return ret;
2780 }
2781
2782 /**
2783  * Modify a VF MAC address
2784  *
2785  * @param priv
2786  *   Pointer to device private data.
2787  * @param mac_addr
2788  *   MAC address to modify into.
2789  * @param iface_idx
2790  *   Net device interface index
2791  * @param vf_index
2792  *   VF index
2793  *
2794  * @return
2795  *   0 on success, a negative errno value otherwise
2796  */
2797 int
2798 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2799                            unsigned int iface_idx,
2800                            struct rte_ether_addr *mac_addr,
2801                            int vf_index)
2802 {
2803         return mlx5_nl_vf_mac_addr_modify
2804                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2805 }
2806
2807 /**
2808  * Set device promiscuous mode
2809  *
2810  * @param dev
2811  *   Pointer to Ethernet device structure.
2812  * @param enable
2813  *   0 - promiscuous is disabled, otherwise - enabled
2814  *
2815  * @return
2816  *   0 on success, a negative error value otherwise
2817  */
2818 int
2819 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2820 {
2821         struct mlx5_priv *priv = dev->data->dev_private;
2822
2823         return mlx5_nl_promisc(priv->nl_socket_route,
2824                                mlx5_ifindex(dev), !!enable);
2825 }
2826
2827 /**
2828  * Set device promiscuous mode
2829  *
2830  * @param dev
2831  *   Pointer to Ethernet device structure.
2832  * @param enable
2833  *   0 - all multicase is disabled, otherwise - enabled
2834  *
2835  * @return
2836  *   0 on success, a negative error value otherwise
2837  */
2838 int
2839 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2840 {
2841         struct mlx5_priv *priv = dev->data->dev_private;
2842
2843         return mlx5_nl_allmulti(priv->nl_socket_route,
2844                                 mlx5_ifindex(dev), !!enable);
2845 }
2846
2847 /**
2848  * Flush device MAC addresses
2849  *
2850  * @param dev
2851  *   Pointer to Ethernet device structure.
2852  *
2853  */
2854 void
2855 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2856 {
2857         struct mlx5_priv *priv = dev->data->dev_private;
2858
2859         mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2860                                dev->data->mac_addrs,
2861                                MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2862 }