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                 }
1305                 if (priv->mtr_en) {
1306                         err = mlx5_aso_flow_mtrs_mng_init(priv->sh);
1307                         if (err) {
1308                                 err = -err;
1309                                 goto error;
1310                         }
1311                 }
1312 #endif
1313 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1314                 if (config->hca_attr.flow_hit_aso &&
1315                     priv->mtr_color_reg == REG_C_3) {
1316                         sh->flow_hit_aso_en = 1;
1317                         err = mlx5_flow_aso_age_mng_init(sh);
1318                         if (err) {
1319                                 err = -err;
1320                                 goto error;
1321                         }
1322                         DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1323                 }
1324 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1325 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1326                 if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1327                     config->dv_flow_en) {
1328                         priv->sampler_en = 1;
1329                         DRV_LOG(DEBUG, "Sampler enabled!");
1330                 } else {
1331                         priv->sampler_en = 0;
1332                         if (!config->hca_attr.log_max_ft_sampler_num)
1333                                 DRV_LOG(WARNING,
1334                                         "No available register for sampler.");
1335                         else
1336                                 DRV_LOG(DEBUG, "DV flow is not supported!");
1337                 }
1338 #endif
1339         }
1340         if (config->cqe_comp && RTE_CACHE_LINE_SIZE == 128 &&
1341             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP)) {
1342                 DRV_LOG(WARNING, "Rx CQE 128B compression is not supported");
1343                 config->cqe_comp = 0;
1344         }
1345         if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX &&
1346             (!config->devx || !config->hca_attr.mini_cqe_resp_flow_tag)) {
1347                 DRV_LOG(WARNING, "Flow Tag CQE compression"
1348                                  " format isn't supported.");
1349                 config->cqe_comp = 0;
1350         }
1351         if (config->cqe_comp_fmt == MLX5_CQE_RESP_FORMAT_L34H_STRIDX &&
1352             (!config->devx || !config->hca_attr.mini_cqe_resp_l3_l4_tag)) {
1353                 DRV_LOG(WARNING, "L3/L4 Header CQE compression"
1354                                  " format isn't supported.");
1355                 config->cqe_comp = 0;
1356         }
1357         DRV_LOG(DEBUG, "Rx CQE compression is %ssupported",
1358                         config->cqe_comp ? "" : "not ");
1359         if (config->tx_pp) {
1360                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1361                         config->hca_attr.dev_freq_khz);
1362                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1363                         config->hca_attr.qos.packet_pacing ? "" : "not ");
1364                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1365                         config->hca_attr.cross_channel ? "" : "not ");
1366                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1367                         config->hca_attr.wqe_index_ignore ? "" : "not ");
1368                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1369                         config->hca_attr.non_wire_sq ? "" : "not ");
1370                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1371                         config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1372                         config->hca_attr.log_max_static_sq_wq);
1373                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1374                         config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1375                 if (!config->devx) {
1376                         DRV_LOG(ERR, "DevX is required for packet pacing");
1377                         err = ENODEV;
1378                         goto error;
1379                 }
1380                 if (!config->hca_attr.qos.packet_pacing) {
1381                         DRV_LOG(ERR, "Packet pacing is not supported");
1382                         err = ENODEV;
1383                         goto error;
1384                 }
1385                 if (!config->hca_attr.cross_channel) {
1386                         DRV_LOG(ERR, "Cross channel operations are"
1387                                      " required for packet pacing");
1388                         err = ENODEV;
1389                         goto error;
1390                 }
1391                 if (!config->hca_attr.wqe_index_ignore) {
1392                         DRV_LOG(ERR, "WQE index ignore feature is"
1393                                      " required for packet pacing");
1394                         err = ENODEV;
1395                         goto error;
1396                 }
1397                 if (!config->hca_attr.non_wire_sq) {
1398                         DRV_LOG(ERR, "Non-wire SQ feature is"
1399                                      " required for packet pacing");
1400                         err = ENODEV;
1401                         goto error;
1402                 }
1403                 if (!config->hca_attr.log_max_static_sq_wq) {
1404                         DRV_LOG(ERR, "Static WQE SQ feature is"
1405                                      " required for packet pacing");
1406                         err = ENODEV;
1407                         goto error;
1408                 }
1409                 if (!config->hca_attr.qos.wqe_rate_pp) {
1410                         DRV_LOG(ERR, "WQE rate mode is required"
1411                                      " for packet pacing");
1412                         err = ENODEV;
1413                         goto error;
1414                 }
1415 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1416                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1417                              " can't create queues for packet pacing");
1418                 err = ENODEV;
1419                 goto error;
1420 #endif
1421         }
1422         if (config->devx) {
1423                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1424
1425                 err = config->hca_attr.access_register_user ?
1426                         mlx5_devx_cmd_register_read
1427                                 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1428                                 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1429                 if (!err) {
1430                         uint32_t ts_mode;
1431
1432                         /* MTUTC register is read successfully. */
1433                         ts_mode = MLX5_GET(register_mtutc, reg,
1434                                            time_stamp_mode);
1435                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1436                                 config->rt_timestamp = 1;
1437                 } else {
1438                         /* Kernel does not support register reading. */
1439                         if (config->hca_attr.dev_freq_khz ==
1440                                                  (NS_PER_S / MS_PER_S))
1441                                 config->rt_timestamp = 1;
1442                 }
1443         }
1444         /*
1445          * If HW has bug working with tunnel packet decapsulation and
1446          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1447          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1448          */
1449         if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1450                 config->hw_fcs_strip = 0;
1451         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1452                 (config->hw_fcs_strip ? "" : "not "));
1453         if (config->mprq.enabled && mprq) {
1454                 if (config->mprq.stride_num_n &&
1455                     (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1456                      config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1457                         config->mprq.stride_num_n =
1458                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1459                                                 mprq_min_stride_num_n),
1460                                         mprq_max_stride_num_n);
1461                         DRV_LOG(WARNING,
1462                                 "the number of strides"
1463                                 " for Multi-Packet RQ is out of range,"
1464                                 " setting default value (%u)",
1465                                 1 << config->mprq.stride_num_n);
1466                 }
1467                 if (config->mprq.stride_size_n &&
1468                     (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1469                      config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1470                         config->mprq.stride_size_n =
1471                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1472                                                 mprq_min_stride_size_n),
1473                                         mprq_max_stride_size_n);
1474                         DRV_LOG(WARNING,
1475                                 "the size of a stride"
1476                                 " for Multi-Packet RQ is out of range,"
1477                                 " setting default value (%u)",
1478                                 1 << config->mprq.stride_size_n);
1479                 }
1480                 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1481                 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1482         } else if (config->mprq.enabled && !mprq) {
1483                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1484                 config->mprq.enabled = 0;
1485         }
1486         if (config->max_dump_files_num == 0)
1487                 config->max_dump_files_num = 128;
1488         eth_dev = rte_eth_dev_allocate(name);
1489         if (eth_dev == NULL) {
1490                 DRV_LOG(ERR, "can not allocate rte ethdev");
1491                 err = ENOMEM;
1492                 goto error;
1493         }
1494         if (priv->representor) {
1495                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1496                 eth_dev->data->representor_id = priv->representor_id;
1497         }
1498         priv->mp_id.port_id = eth_dev->data->port_id;
1499         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
1500         /*
1501          * Store associated network device interface index. This index
1502          * is permanent throughout the lifetime of device. So, we may store
1503          * the ifindex here and use the cached value further.
1504          */
1505         MLX5_ASSERT(spawn->ifindex);
1506         priv->if_index = spawn->ifindex;
1507         eth_dev->data->dev_private = priv;
1508         priv->dev_data = eth_dev->data;
1509         eth_dev->data->mac_addrs = priv->mac;
1510         eth_dev->device = dpdk_dev;
1511         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1512         /* Configure the first MAC address by default. */
1513         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1514                 DRV_LOG(ERR,
1515                         "port %u cannot get MAC address, is mlx5_en"
1516                         " loaded? (errno: %s)",
1517                         eth_dev->data->port_id, strerror(rte_errno));
1518                 err = ENODEV;
1519                 goto error;
1520         }
1521         DRV_LOG(INFO,
1522                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1523                 eth_dev->data->port_id,
1524                 mac.addr_bytes[0], mac.addr_bytes[1],
1525                 mac.addr_bytes[2], mac.addr_bytes[3],
1526                 mac.addr_bytes[4], mac.addr_bytes[5]);
1527 #ifdef RTE_LIBRTE_MLX5_DEBUG
1528         {
1529                 char ifname[MLX5_NAMESIZE];
1530
1531                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1532                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1533                                 eth_dev->data->port_id, ifname);
1534                 else
1535                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1536                                 eth_dev->data->port_id);
1537         }
1538 #endif
1539         /* Get actual MTU if possible. */
1540         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1541         if (err) {
1542                 err = rte_errno;
1543                 goto error;
1544         }
1545         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1546                 priv->mtu);
1547         /* Initialize burst functions to prevent crashes before link-up. */
1548         eth_dev->rx_pkt_burst = removed_rx_burst;
1549         eth_dev->tx_pkt_burst = removed_tx_burst;
1550         eth_dev->dev_ops = &mlx5_dev_ops;
1551         eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1552         eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1553         eth_dev->rx_queue_count = mlx5_rx_queue_count;
1554         /* Register MAC address. */
1555         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1556         if (config->vf && config->vf_nl_en)
1557                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1558                                       mlx5_ifindex(eth_dev),
1559                                       eth_dev->data->mac_addrs,
1560                                       MLX5_MAX_MAC_ADDRESSES);
1561         priv->flows = 0;
1562         priv->ctrl_flows = 0;
1563         rte_spinlock_init(&priv->flow_list_lock);
1564         TAILQ_INIT(&priv->flow_meters);
1565         TAILQ_INIT(&priv->flow_meter_profiles);
1566         /* Hint libmlx5 to use PMD allocator for data plane resources */
1567         mlx5_glue->dv_set_context_attr(sh->ctx,
1568                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1569                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1570                                 .alloc = &mlx5_alloc_verbs_buf,
1571                                 .free = &mlx5_free_verbs_buf,
1572                                 .data = sh,
1573                         }));
1574         /* Bring Ethernet device up. */
1575         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1576                 eth_dev->data->port_id);
1577         mlx5_set_link_up(eth_dev);
1578         /*
1579          * Even though the interrupt handler is not installed yet,
1580          * interrupts will still trigger on the async_fd from
1581          * Verbs context returned by ibv_open_device().
1582          */
1583         mlx5_link_update(eth_dev, 0);
1584 #ifdef HAVE_MLX5DV_DR_ESWITCH
1585         if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1586               (switch_info->representor || switch_info->master)))
1587                 config->dv_esw_en = 0;
1588 #else
1589         config->dv_esw_en = 0;
1590 #endif
1591         /* Detect minimal data bytes to inline. */
1592         mlx5_set_min_inline(spawn, config);
1593         /* Store device configuration on private structure. */
1594         priv->config = *config;
1595         /* Create context for virtual machine VLAN workaround. */
1596         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1597         if (config->dv_flow_en) {
1598                 err = mlx5_alloc_shared_dr(priv);
1599                 if (err)
1600                         goto error;
1601         }
1602         if (config->devx && config->dv_flow_en && config->dest_tir) {
1603                 priv->obj_ops = devx_obj_ops;
1604                 priv->obj_ops.drop_action_create =
1605                                                 ibv_obj_ops.drop_action_create;
1606                 priv->obj_ops.drop_action_destroy =
1607                                                 ibv_obj_ops.drop_action_destroy;
1608 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1609                 priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1610 #else
1611                 if (config->dv_esw_en)
1612                         priv->obj_ops.txq_obj_modify =
1613                                                 ibv_obj_ops.txq_obj_modify;
1614 #endif
1615                 /* Use specific wrappers for Tx object. */
1616                 priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1617                 priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1618                 mlx5_queue_counter_id_prepare(eth_dev);
1619
1620         } else {
1621                 priv->obj_ops = ibv_obj_ops;
1622         }
1623         priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1624         if (!priv->drop_queue.hrxq)
1625                 goto error;
1626         /* Supported Verbs flow priority number detection. */
1627         err = mlx5_flow_discover_priorities(eth_dev);
1628         if (err < 0) {
1629                 err = -err;
1630                 goto error;
1631         }
1632         priv->config.flow_prio = err;
1633         if (!priv->config.dv_esw_en &&
1634             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1635                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1636                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1637                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1638         }
1639         mlx5_set_metadata_mask(eth_dev);
1640         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1641             !priv->sh->dv_regc0_mask) {
1642                 DRV_LOG(ERR, "metadata mode %u is not supported "
1643                              "(no metadata reg_c[0] is available)",
1644                              priv->config.dv_xmeta_en);
1645                         err = ENOTSUP;
1646                         goto error;
1647         }
1648         mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1649                              mlx5_hrxq_create_cb,
1650                              mlx5_hrxq_match_cb,
1651                              mlx5_hrxq_remove_cb);
1652         /* Query availability of metadata reg_c's. */
1653         err = mlx5_flow_discover_mreg_c(eth_dev);
1654         if (err < 0) {
1655                 err = -err;
1656                 goto error;
1657         }
1658         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1659                 DRV_LOG(DEBUG,
1660                         "port %u extensive metadata register is not supported",
1661                         eth_dev->data->port_id);
1662                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1663                         DRV_LOG(ERR, "metadata mode %u is not supported "
1664                                      "(no metadata registers available)",
1665                                      priv->config.dv_xmeta_en);
1666                         err = ENOTSUP;
1667                         goto error;
1668                 }
1669         }
1670         if (priv->config.dv_flow_en &&
1671             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1672             mlx5_flow_ext_mreg_supported(eth_dev) &&
1673             priv->sh->dv_regc0_mask) {
1674                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1675                                                       MLX5_FLOW_MREG_HTABLE_SZ,
1676                                                       0, 0,
1677                                                       flow_dv_mreg_create_cb,
1678                                                       flow_dv_mreg_match_cb,
1679                                                       flow_dv_mreg_remove_cb);
1680                 if (!priv->mreg_cp_tbl) {
1681                         err = ENOMEM;
1682                         goto error;
1683                 }
1684                 priv->mreg_cp_tbl->ctx = eth_dev;
1685         }
1686         rte_spinlock_init(&priv->shared_act_sl);
1687         mlx5_flow_counter_mode_config(eth_dev);
1688         if (priv->config.dv_flow_en)
1689                 eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1690         return eth_dev;
1691 error:
1692         if (priv) {
1693                 if (priv->mreg_cp_tbl)
1694                         mlx5_hlist_destroy(priv->mreg_cp_tbl);
1695                 if (priv->sh)
1696                         mlx5_os_free_shared_dr(priv);
1697                 if (priv->nl_socket_route >= 0)
1698                         close(priv->nl_socket_route);
1699                 if (priv->nl_socket_rdma >= 0)
1700                         close(priv->nl_socket_rdma);
1701                 if (priv->vmwa_context)
1702                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1703                 if (eth_dev && priv->drop_queue.hrxq)
1704                         mlx5_drop_action_destroy(eth_dev);
1705                 if (own_domain_id)
1706                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1707                 mlx5_cache_list_destroy(&priv->hrxqs);
1708                 mlx5_free(priv);
1709                 if (eth_dev != NULL)
1710                         eth_dev->data->dev_private = NULL;
1711         }
1712         if (eth_dev != NULL) {
1713                 /* mac_addrs must not be freed alone because part of
1714                  * dev_private
1715                  **/
1716                 eth_dev->data->mac_addrs = NULL;
1717                 rte_eth_dev_release_port(eth_dev);
1718         }
1719         if (sh)
1720                 mlx5_free_shared_dev_ctx(sh);
1721         MLX5_ASSERT(err > 0);
1722         rte_errno = err;
1723         return NULL;
1724 }
1725
1726 /**
1727  * Comparison callback to sort device data.
1728  *
1729  * This is meant to be used with qsort().
1730  *
1731  * @param a[in]
1732  *   Pointer to pointer to first data object.
1733  * @param b[in]
1734  *   Pointer to pointer to second data object.
1735  *
1736  * @return
1737  *   0 if both objects are equal, less than 0 if the first argument is less
1738  *   than the second, greater than 0 otherwise.
1739  */
1740 static int
1741 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1742 {
1743         const struct mlx5_switch_info *si_a =
1744                 &((const struct mlx5_dev_spawn_data *)a)->info;
1745         const struct mlx5_switch_info *si_b =
1746                 &((const struct mlx5_dev_spawn_data *)b)->info;
1747         int ret;
1748
1749         /* Master device first. */
1750         ret = si_b->master - si_a->master;
1751         if (ret)
1752                 return ret;
1753         /* Then representor devices. */
1754         ret = si_b->representor - si_a->representor;
1755         if (ret)
1756                 return ret;
1757         /* Unidentified devices come last in no specific order. */
1758         if (!si_a->representor)
1759                 return 0;
1760         /* Order representors by name. */
1761         return si_a->port_name - si_b->port_name;
1762 }
1763
1764 /**
1765  * Match PCI information for possible slaves of bonding device.
1766  *
1767  * @param[in] ibv_dev
1768  *   Pointer to Infiniband device structure.
1769  * @param[in] pci_dev
1770  *   Pointer to primary PCI address structure to match.
1771  * @param[in] nl_rdma
1772  *   Netlink RDMA group socket handle.
1773  * @param[in] owner
1774  *   Rerepsentor owner PF index.
1775  * @param[out] bond_info
1776  *   Pointer to bonding information.
1777  *
1778  * @return
1779  *   negative value if no bonding device found, otherwise
1780  *   positive index of slave PF in bonding.
1781  */
1782 static int
1783 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1784                            const struct rte_pci_addr *pci_dev,
1785                            int nl_rdma, uint16_t owner,
1786                            struct mlx5_bond_info *bond_info)
1787 {
1788         char ifname[IF_NAMESIZE + 1];
1789         unsigned int ifindex;
1790         unsigned int np, i;
1791         FILE *bond_file = NULL, *file;
1792         int pf = -1;
1793         int ret;
1794
1795         /*
1796          * Try to get master device name. If something goes
1797          * wrong suppose the lack of kernel support and no
1798          * bonding devices.
1799          */
1800         memset(bond_info, 0, sizeof(*bond_info));
1801         if (nl_rdma < 0)
1802                 return -1;
1803         if (!strstr(ibv_dev->name, "bond"))
1804                 return -1;
1805         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1806         if (!np)
1807                 return -1;
1808         /*
1809          * The Master device might not be on the predefined
1810          * port (not on port index 1, it is not garanted),
1811          * we have to scan all Infiniband device port and
1812          * find master.
1813          */
1814         for (i = 1; i <= np; ++i) {
1815                 /* Check whether Infiniband port is populated. */
1816                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1817                 if (!ifindex)
1818                         continue;
1819                 if (!if_indextoname(ifindex, ifname))
1820                         continue;
1821                 /* Try to read bonding slave names from sysfs. */
1822                 MKSTR(slaves,
1823                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1824                 bond_file = fopen(slaves, "r");
1825                 if (bond_file)
1826                         break;
1827         }
1828         if (!bond_file)
1829                 return -1;
1830         /* Use safe format to check maximal buffer length. */
1831         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1832         while (fscanf(bond_file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1833                 char tmp_str[IF_NAMESIZE + 32];
1834                 struct rte_pci_addr pci_addr;
1835                 struct mlx5_switch_info info;
1836
1837                 /* Process slave interface names in the loop. */
1838                 snprintf(tmp_str, sizeof(tmp_str),
1839                          "/sys/class/net/%s", ifname);
1840                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1841                         DRV_LOG(WARNING, "can not get PCI address"
1842                                          " for netdev \"%s\"", ifname);
1843                         continue;
1844                 }
1845                 /* Slave interface PCI address match found. */
1846                 snprintf(tmp_str, sizeof(tmp_str),
1847                          "/sys/class/net/%s/phys_port_name", ifname);
1848                 file = fopen(tmp_str, "rb");
1849                 if (!file)
1850                         break;
1851                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1852                 if (fscanf(file, "%32s", tmp_str) == 1)
1853                         mlx5_translate_port_name(tmp_str, &info);
1854                 fclose(file);
1855                 /* Only process PF ports. */
1856                 if (info.name_type != MLX5_PHYS_PORT_NAME_TYPE_LEGACY &&
1857                     info.name_type != MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1858                         continue;
1859                 /* Check max bonding member. */
1860                 if (info.port_name >= MLX5_BOND_MAX_PORTS) {
1861                         DRV_LOG(WARNING, "bonding index out of range, "
1862                                 "please increase MLX5_BOND_MAX_PORTS: %s",
1863                                 tmp_str);
1864                         break;
1865                 }
1866                 /* Match PCI address. */
1867                 if (pci_dev->domain == pci_addr.domain &&
1868                     pci_dev->bus == pci_addr.bus &&
1869                     pci_dev->devid == pci_addr.devid &&
1870                     pci_dev->function + owner == pci_addr.function)
1871                         pf = info.port_name;
1872                 /* Get ifindex. */
1873                 snprintf(tmp_str, sizeof(tmp_str),
1874                          "/sys/class/net/%s/ifindex", ifname);
1875                 file = fopen(tmp_str, "rb");
1876                 if (!file)
1877                         break;
1878                 ret = fscanf(file, "%u", &ifindex);
1879                 fclose(file);
1880                 if (ret != 1)
1881                         break;
1882                 /* Save bonding info. */
1883                 strncpy(bond_info->ports[info.port_name].ifname, ifname,
1884                         sizeof(bond_info->ports[0].ifname));
1885                 bond_info->ports[info.port_name].pci_addr = pci_addr;
1886                 bond_info->ports[info.port_name].ifindex = ifindex;
1887                 bond_info->n_port++;
1888         }
1889         if (pf >= 0) {
1890                 /* Get bond interface info */
1891                 ret = mlx5_sysfs_bond_info(ifindex, &bond_info->ifindex,
1892                                            bond_info->ifname);
1893                 if (ret)
1894                         DRV_LOG(ERR, "unable to get bond info: %s",
1895                                 strerror(rte_errno));
1896                 else
1897                         DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1898                                 ifindex, bond_info->ifindex, bond_info->ifname);
1899         }
1900         return pf;
1901 }
1902
1903 /**
1904  * Register a PCI device within bonding.
1905  *
1906  * This function spawns Ethernet devices out of a given PCI device and
1907  * bonding owner PF index.
1908  *
1909  * @param[in] pci_dev
1910  *   PCI device information.
1911  * @param[in] req_eth_da
1912  *   Requested ethdev device argument.
1913  * @param[in] owner_id
1914  *   Requested owner PF port ID within bonding device, default to 0.
1915  *
1916  * @return
1917  *   0 on success, a negative errno value otherwise and rte_errno is set.
1918  */
1919 static int
1920 mlx5_os_pci_probe_pf(struct rte_pci_device *pci_dev,
1921                      struct rte_eth_devargs *req_eth_da,
1922                      uint16_t owner_id)
1923 {
1924         struct ibv_device **ibv_list;
1925         /*
1926          * Number of found IB Devices matching with requested PCI BDF.
1927          * nd != 1 means there are multiple IB devices over the same
1928          * PCI device and we have representors and master.
1929          */
1930         unsigned int nd = 0;
1931         /*
1932          * Number of found IB device Ports. nd = 1 and np = 1..n means
1933          * we have the single multiport IB device, and there may be
1934          * representors attached to some of found ports.
1935          */
1936         unsigned int np = 0;
1937         /*
1938          * Number of DPDK ethernet devices to Spawn - either over
1939          * multiple IB devices or multiple ports of single IB device.
1940          * Actually this is the number of iterations to spawn.
1941          */
1942         unsigned int ns = 0;
1943         /*
1944          * Bonding device
1945          *   < 0 - no bonding device (single one)
1946          *  >= 0 - bonding device (value is slave PF index)
1947          */
1948         int bd = -1;
1949         struct mlx5_dev_spawn_data *list = NULL;
1950         struct mlx5_dev_config dev_config;
1951         unsigned int dev_config_vf;
1952         struct rte_eth_devargs eth_da = *req_eth_da;
1953         struct rte_pci_addr owner_pci = pci_dev->addr; /* Owner PF. */
1954         struct mlx5_bond_info bond_info;
1955         int ret = -1;
1956
1957         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1958                 mlx5_pmd_socket_init();
1959         ret = mlx5_init_once();
1960         if (ret) {
1961                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1962                         strerror(rte_errno));
1963                 return -rte_errno;
1964         }
1965         errno = 0;
1966         ibv_list = mlx5_glue->get_device_list(&ret);
1967         if (!ibv_list) {
1968                 rte_errno = errno ? errno : ENOSYS;
1969                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1970                 return -rte_errno;
1971         }
1972         /*
1973          * First scan the list of all Infiniband devices to find
1974          * matching ones, gathering into the list.
1975          */
1976         struct ibv_device *ibv_match[ret + 1];
1977         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1978         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1979         unsigned int i;
1980
1981         while (ret-- > 0) {
1982                 struct rte_pci_addr pci_addr;
1983
1984                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1985                 bd = mlx5_device_bond_pci_match
1986                                 (ibv_list[ret], &owner_pci, nl_rdma, owner_id,
1987                                  &bond_info);
1988                 if (bd >= 0) {
1989                         /*
1990                          * Bonding device detected. Only one match is allowed,
1991                          * the bonding is supported over multi-port IB device,
1992                          * there should be no matches on representor PCI
1993                          * functions or non VF LAG bonding devices with
1994                          * specified address.
1995                          */
1996                         if (nd) {
1997                                 DRV_LOG(ERR,
1998                                         "multiple PCI match on bonding device"
1999                                         "\"%s\" found", ibv_list[ret]->name);
2000                                 rte_errno = ENOENT;
2001                                 ret = -rte_errno;
2002                                 goto exit;
2003                         }
2004                         /* Amend owner pci address if owner PF ID specified. */
2005                         if (eth_da.nb_representor_ports)
2006                                 owner_pci.function += owner_id;
2007                         DRV_LOG(INFO, "PCI information matches for"
2008                                       " slave %d bonding device \"%s\"",
2009                                       bd, ibv_list[ret]->name);
2010                         ibv_match[nd++] = ibv_list[ret];
2011                         break;
2012                 } else {
2013                         /* Bonding device not found. */
2014                         if (mlx5_dev_to_pci_addr
2015                                 (ibv_list[ret]->ibdev_path, &pci_addr))
2016                                 continue;
2017                         if (owner_pci.domain != pci_addr.domain ||
2018                             owner_pci.bus != pci_addr.bus ||
2019                             owner_pci.devid != pci_addr.devid ||
2020                             owner_pci.function != pci_addr.function)
2021                                 continue;
2022                         DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2023                                 ibv_list[ret]->name);
2024                         ibv_match[nd++] = ibv_list[ret];
2025                 }
2026         }
2027         ibv_match[nd] = NULL;
2028         if (!nd) {
2029                 /* No device matches, just complain and bail out. */
2030                 DRV_LOG(WARNING,
2031                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
2032                         " are kernel drivers loaded?",
2033                         owner_pci.domain, owner_pci.bus,
2034                         owner_pci.devid, owner_pci.function);
2035                 rte_errno = ENOENT;
2036                 ret = -rte_errno;
2037                 goto exit;
2038         }
2039         if (nd == 1) {
2040                 /*
2041                  * Found single matching device may have multiple ports.
2042                  * Each port may be representor, we have to check the port
2043                  * number and check the representors existence.
2044                  */
2045                 if (nl_rdma >= 0)
2046                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2047                 if (!np)
2048                         DRV_LOG(WARNING, "can not get IB device \"%s\""
2049                                          " ports number", ibv_match[0]->name);
2050                 if (bd >= 0 && !np) {
2051                         DRV_LOG(ERR, "can not get ports"
2052                                      " for bonding device");
2053                         rte_errno = ENOENT;
2054                         ret = -rte_errno;
2055                         goto exit;
2056                 }
2057         }
2058 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2059         if (bd >= 0) {
2060                 /*
2061                  * This may happen if there is VF LAG kernel support and
2062                  * application is compiled with older rdma_core library.
2063                  */
2064                 DRV_LOG(ERR,
2065                         "No kernel/verbs support for VF LAG bonding found.");
2066                 rte_errno = ENOTSUP;
2067                 ret = -rte_errno;
2068                 goto exit;
2069         }
2070 #endif
2071         /*
2072          * Now we can determine the maximal
2073          * amount of devices to be spawned.
2074          */
2075         list = mlx5_malloc(MLX5_MEM_ZERO,
2076                            sizeof(struct mlx5_dev_spawn_data) *
2077                            (np ? np : nd),
2078                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
2079         if (!list) {
2080                 DRV_LOG(ERR, "spawn data array allocation failure");
2081                 rte_errno = ENOMEM;
2082                 ret = -rte_errno;
2083                 goto exit;
2084         }
2085         if (bd >= 0 || np > 1) {
2086                 /*
2087                  * Single IB device with multiple ports found,
2088                  * it may be E-Switch master device and representors.
2089                  * We have to perform identification through the ports.
2090                  */
2091                 MLX5_ASSERT(nl_rdma >= 0);
2092                 MLX5_ASSERT(ns == 0);
2093                 MLX5_ASSERT(nd == 1);
2094                 MLX5_ASSERT(np);
2095                 for (i = 1; i <= np; ++i) {
2096                         list[ns].bond_info = &bond_info;
2097                         list[ns].max_port = np;
2098                         list[ns].phys_port = i;
2099                         list[ns].phys_dev = ibv_match[0];
2100                         list[ns].eth_dev = NULL;
2101                         list[ns].pci_dev = pci_dev;
2102                         list[ns].pf_bond = bd;
2103                         list[ns].ifindex = mlx5_nl_ifindex
2104                                 (nl_rdma,
2105                                 mlx5_os_get_dev_device_name
2106                                                 (list[ns].phys_dev), i);
2107                         if (!list[ns].ifindex) {
2108                                 /*
2109                                  * No network interface index found for the
2110                                  * specified port, it means there is no
2111                                  * representor on this port. It's OK,
2112                                  * there can be disabled ports, for example
2113                                  * if sriov_numvfs < sriov_totalvfs.
2114                                  */
2115                                 continue;
2116                         }
2117                         ret = -1;
2118                         if (nl_route >= 0)
2119                                 ret = mlx5_nl_switch_info
2120                                                (nl_route,
2121                                                 list[ns].ifindex,
2122                                                 &list[ns].info);
2123                         if (ret || (!list[ns].info.representor &&
2124                                     !list[ns].info.master)) {
2125                                 /*
2126                                  * We failed to recognize representors with
2127                                  * Netlink, let's try to perform the task
2128                                  * with sysfs.
2129                                  */
2130                                 ret =  mlx5_sysfs_switch_info
2131                                                 (list[ns].ifindex,
2132                                                  &list[ns].info);
2133                         }
2134 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
2135                         if (!ret && bd >= 0) {
2136                                 switch (list[ns].info.name_type) {
2137                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2138                                         if (list[ns].info.port_name == bd)
2139                                                 ns++;
2140                                         break;
2141                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
2142                                         /* Fallthrough */
2143                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2144                                         /* Fallthrough */
2145                                 case MLX5_PHYS_PORT_NAME_TYPE_PFSF:
2146                                         if (list[ns].info.pf_num == bd)
2147                                                 ns++;
2148                                         break;
2149                                 default:
2150                                         break;
2151                                 }
2152                                 continue;
2153                         }
2154 #endif
2155                         if (!ret && (list[ns].info.representor ^
2156                                      list[ns].info.master))
2157                                 ns++;
2158                 }
2159                 if (!ns) {
2160                         DRV_LOG(ERR,
2161                                 "unable to recognize master/representors"
2162                                 " on the IB device with multiple ports");
2163                         rte_errno = ENOENT;
2164                         ret = -rte_errno;
2165                         goto exit;
2166                 }
2167         } else {
2168                 /*
2169                  * The existence of several matching entries (nd > 1) means
2170                  * port representors have been instantiated. No existing Verbs
2171                  * call nor sysfs entries can tell them apart, this can only
2172                  * be done through Netlink calls assuming kernel drivers are
2173                  * recent enough to support them.
2174                  *
2175                  * In the event of identification failure through Netlink,
2176                  * try again through sysfs, then:
2177                  *
2178                  * 1. A single IB device matches (nd == 1) with single
2179                  *    port (np=0/1) and is not a representor, assume
2180                  *    no switch support.
2181                  *
2182                  * 2. Otherwise no safe assumptions can be made;
2183                  *    complain louder and bail out.
2184                  */
2185                 for (i = 0; i != nd; ++i) {
2186                         memset(&list[ns].info, 0, sizeof(list[ns].info));
2187                         list[ns].bond_info = NULL;
2188                         list[ns].max_port = 1;
2189                         list[ns].phys_port = 1;
2190                         list[ns].phys_dev = ibv_match[i];
2191                         list[ns].eth_dev = NULL;
2192                         list[ns].pci_dev = pci_dev;
2193                         list[ns].pf_bond = -1;
2194                         list[ns].ifindex = 0;
2195                         if (nl_rdma >= 0)
2196                                 list[ns].ifindex = mlx5_nl_ifindex
2197                                 (nl_rdma,
2198                                 mlx5_os_get_dev_device_name
2199                                                 (list[ns].phys_dev), 1);
2200                         if (!list[ns].ifindex) {
2201                                 char ifname[IF_NAMESIZE];
2202
2203                                 /*
2204                                  * Netlink failed, it may happen with old
2205                                  * ib_core kernel driver (before 4.16).
2206                                  * We can assume there is old driver because
2207                                  * here we are processing single ports IB
2208                                  * devices. Let's try sysfs to retrieve
2209                                  * the ifindex. The method works for
2210                                  * master device only.
2211                                  */
2212                                 if (nd > 1) {
2213                                         /*
2214                                          * Multiple devices found, assume
2215                                          * representors, can not distinguish
2216                                          * master/representor and retrieve
2217                                          * ifindex via sysfs.
2218                                          */
2219                                         continue;
2220                                 }
2221                                 ret = mlx5_get_ifname_sysfs
2222                                         (ibv_match[i]->ibdev_path, ifname);
2223                                 if (!ret)
2224                                         list[ns].ifindex =
2225                                                 if_nametoindex(ifname);
2226                                 if (!list[ns].ifindex) {
2227                                         /*
2228                                          * No network interface index found
2229                                          * for the specified device, it means
2230                                          * there it is neither representor
2231                                          * nor master.
2232                                          */
2233                                         continue;
2234                                 }
2235                         }
2236                         ret = -1;
2237                         if (nl_route >= 0)
2238                                 ret = mlx5_nl_switch_info
2239                                                (nl_route,
2240                                                 list[ns].ifindex,
2241                                                 &list[ns].info);
2242                         if (ret || (!list[ns].info.representor &&
2243                                     !list[ns].info.master)) {
2244                                 /*
2245                                  * We failed to recognize representors with
2246                                  * Netlink, let's try to perform the task
2247                                  * with sysfs.
2248                                  */
2249                                 ret =  mlx5_sysfs_switch_info
2250                                                 (list[ns].ifindex,
2251                                                  &list[ns].info);
2252                         }
2253                         if (!ret && (list[ns].info.representor ^
2254                                      list[ns].info.master)) {
2255                                 ns++;
2256                         } else if ((nd == 1) &&
2257                                    !list[ns].info.representor &&
2258                                    !list[ns].info.master) {
2259                                 /*
2260                                  * Single IB device with
2261                                  * one physical port and
2262                                  * attached network device.
2263                                  * May be SRIOV is not enabled
2264                                  * or there is no representors.
2265                                  */
2266                                 DRV_LOG(INFO, "no E-Switch support detected");
2267                                 ns++;
2268                                 break;
2269                         }
2270                 }
2271                 if (!ns) {
2272                         DRV_LOG(ERR,
2273                                 "unable to recognize master/representors"
2274                                 " on the multiple IB devices");
2275                         rte_errno = ENOENT;
2276                         ret = -rte_errno;
2277                         goto exit;
2278                 }
2279         }
2280         MLX5_ASSERT(ns);
2281         /*
2282          * Sort list to probe devices in natural order for users convenience
2283          * (i.e. master first, then representors from lowest to highest ID).
2284          */
2285         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2286         /* Device specific configuration. */
2287         switch (pci_dev->id.device_id) {
2288         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2289         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2290         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2291         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2292         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2293         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2294         case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2295                 dev_config_vf = 1;
2296                 break;
2297         default:
2298                 dev_config_vf = 0;
2299                 break;
2300         }
2301         if (eth_da.type != RTE_ETH_REPRESENTOR_NONE) {
2302                 /* Set devargs default values. */
2303                 if (eth_da.nb_mh_controllers == 0) {
2304                         eth_da.nb_mh_controllers = 1;
2305                         eth_da.mh_controllers[0] = 0;
2306                 }
2307                 if (eth_da.nb_ports == 0 && ns > 0) {
2308                         if (list[0].pf_bond >= 0 && list[0].info.representor)
2309                                 DRV_LOG(WARNING, "Representor on Bonding device should use pf#vf# syntax: %s",
2310                                         pci_dev->device.devargs->args);
2311                         eth_da.nb_ports = 1;
2312                         eth_da.ports[0] = list[0].info.pf_num;
2313                 }
2314                 if (eth_da.nb_representor_ports == 0) {
2315                         eth_da.nb_representor_ports = 1;
2316                         eth_da.representor_ports[0] = 0;
2317                 }
2318         }
2319         for (i = 0; i != ns; ++i) {
2320                 uint32_t restore;
2321
2322                 /* Default configuration. */
2323                 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2324                 dev_config.vf = dev_config_vf;
2325                 dev_config.mps = MLX5_ARG_UNSET;
2326                 dev_config.dbnc = MLX5_ARG_UNSET;
2327                 dev_config.rx_vec_en = 1;
2328                 dev_config.txq_inline_max = MLX5_ARG_UNSET;
2329                 dev_config.txq_inline_min = MLX5_ARG_UNSET;
2330                 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2331                 dev_config.txqs_inline = MLX5_ARG_UNSET;
2332                 dev_config.vf_nl_en = 1;
2333                 dev_config.mr_ext_memseg_en = 1;
2334                 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2335                 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2336                 dev_config.dv_esw_en = 1;
2337                 dev_config.dv_flow_en = 1;
2338                 dev_config.decap_en = 1;
2339                 dev_config.log_hp_size = MLX5_ARG_UNSET;
2340                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2341                                                  &list[i],
2342                                                  &dev_config,
2343                                                  &eth_da);
2344                 if (!list[i].eth_dev) {
2345                         if (rte_errno != EBUSY && rte_errno != EEXIST)
2346                                 break;
2347                         /* Device is disabled or already spawned. Ignore it. */
2348                         continue;
2349                 }
2350                 restore = list[i].eth_dev->data->dev_flags;
2351                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2352                 /* Restore non-PCI flags cleared by the above call. */
2353                 list[i].eth_dev->data->dev_flags |= restore;
2354                 rte_eth_dev_probing_finish(list[i].eth_dev);
2355         }
2356         if (i != ns) {
2357                 DRV_LOG(ERR,
2358                         "probe of PCI device " PCI_PRI_FMT " aborted after"
2359                         " encountering an error: %s",
2360                         owner_pci.domain, owner_pci.bus,
2361                         owner_pci.devid, owner_pci.function,
2362                         strerror(rte_errno));
2363                 ret = -rte_errno;
2364                 /* Roll back. */
2365                 while (i--) {
2366                         if (!list[i].eth_dev)
2367                                 continue;
2368                         mlx5_dev_close(list[i].eth_dev);
2369                         /* mac_addrs must not be freed because in dev_private */
2370                         list[i].eth_dev->data->mac_addrs = NULL;
2371                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2372                 }
2373                 /* Restore original error. */
2374                 rte_errno = -ret;
2375         } else {
2376                 ret = 0;
2377         }
2378 exit:
2379         /*
2380          * Do the routine cleanup:
2381          * - close opened Netlink sockets
2382          * - free allocated spawn data array
2383          * - free the Infiniband device list
2384          */
2385         if (nl_rdma >= 0)
2386                 close(nl_rdma);
2387         if (nl_route >= 0)
2388                 close(nl_route);
2389         if (list)
2390                 mlx5_free(list);
2391         MLX5_ASSERT(ibv_list);
2392         mlx5_glue->free_device_list(ibv_list);
2393         return ret;
2394 }
2395
2396 /**
2397  * DPDK callback to register a PCI device.
2398  *
2399  * This function spawns Ethernet devices out of a given PCI device.
2400  *
2401  * @param[in] pci_drv
2402  *   PCI driver structure (mlx5_driver).
2403  * @param[in] pci_dev
2404  *   PCI device information.
2405  *
2406  * @return
2407  *   0 on success, a negative errno value otherwise and rte_errno is set.
2408  */
2409 int
2410 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2411                   struct rte_pci_device *pci_dev)
2412 {
2413         struct rte_eth_devargs eth_da = { .type = RTE_ETH_REPRESENTOR_NONE };
2414         int ret = 0;
2415         uint16_t p;
2416
2417         if (pci_dev->device.devargs) {
2418                 /* Parse representor information from device argument. */
2419                 if (pci_dev->device.devargs->cls_str)
2420                         ret = rte_eth_devargs_parse
2421                                 (pci_dev->device.devargs->cls_str, &eth_da);
2422                 if (ret) {
2423                         DRV_LOG(ERR, "failed to parse device arguments: %s",
2424                                 pci_dev->device.devargs->cls_str);
2425                         return -rte_errno;
2426                 }
2427                 if (eth_da.type == RTE_ETH_REPRESENTOR_NONE) {
2428                         /* Support legacy device argument */
2429                         ret = rte_eth_devargs_parse
2430                                 (pci_dev->device.devargs->args, &eth_da);
2431                         if (ret) {
2432                                 DRV_LOG(ERR, "failed to parse device arguments: %s",
2433                                         pci_dev->device.devargs->args);
2434                                 return -rte_errno;
2435                         }
2436                 }
2437         }
2438
2439         if (eth_da.nb_ports > 0) {
2440                 /* Iterate all port if devargs pf is range: "pf[0-1]vf[...]". */
2441                 for (p = 0; p < eth_da.nb_ports; p++)
2442                         ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da,
2443                                                    eth_da.ports[p]);
2444         } else {
2445                 ret = mlx5_os_pci_probe_pf(pci_dev, &eth_da, 0);
2446         }
2447         return ret;
2448 }
2449
2450 static int
2451 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2452 {
2453         char *env;
2454         int value;
2455
2456         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2457         /* Get environment variable to store. */
2458         env = getenv(MLX5_SHUT_UP_BF);
2459         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2460         if (config->dbnc == MLX5_ARG_UNSET)
2461                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2462         else
2463                 setenv(MLX5_SHUT_UP_BF,
2464                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2465         return value;
2466 }
2467
2468 static void
2469 mlx5_restore_doorbell_mapping_env(int value)
2470 {
2471         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2472         /* Restore the original environment variable state. */
2473         if (value == MLX5_ARG_UNSET)
2474                 unsetenv(MLX5_SHUT_UP_BF);
2475         else
2476                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2477 }
2478
2479 /**
2480  * Extract pdn of PD object using DV API.
2481  *
2482  * @param[in] pd
2483  *   Pointer to the verbs PD object.
2484  * @param[out] pdn
2485  *   Pointer to the PD object number variable.
2486  *
2487  * @return
2488  *   0 on success, error value otherwise.
2489  */
2490 int
2491 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2492 {
2493 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2494         struct mlx5dv_obj obj;
2495         struct mlx5dv_pd pd_info;
2496         int ret = 0;
2497
2498         obj.pd.in = pd;
2499         obj.pd.out = &pd_info;
2500         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2501         if (ret) {
2502                 DRV_LOG(DEBUG, "Fail to get PD object info");
2503                 return ret;
2504         }
2505         *pdn = pd_info.pdn;
2506         return 0;
2507 #else
2508         (void)pd;
2509         (void)pdn;
2510         return -ENOTSUP;
2511 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2512 }
2513
2514 /**
2515  * Function API to open IB device.
2516  *
2517  * This function calls the Linux glue APIs to open a device.
2518  *
2519  * @param[in] spawn
2520  *   Pointer to the IB device attributes (name, port, etc).
2521  * @param[out] config
2522  *   Pointer to device configuration structure.
2523  * @param[out] sh
2524  *   Pointer to shared context structure.
2525  *
2526  * @return
2527  *   0 on success, a positive error value otherwise.
2528  */
2529 int
2530 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2531                      const struct mlx5_dev_config *config,
2532                      struct mlx5_dev_ctx_shared *sh)
2533 {
2534         int dbmap_env;
2535         int err = 0;
2536
2537         sh->numa_node = spawn->pci_dev->device.numa_node;
2538         pthread_mutex_init(&sh->txpp.mutex, NULL);
2539         /*
2540          * Configure environment variable "MLX5_BF_SHUT_UP"
2541          * before the device creation. The rdma_core library
2542          * checks the variable at device creation and
2543          * stores the result internally.
2544          */
2545         dbmap_env = mlx5_config_doorbell_mapping_env(config);
2546         /* Try to open IB device with DV first, then usual Verbs. */
2547         errno = 0;
2548         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2549         if (sh->ctx) {
2550                 sh->devx = 1;
2551                 DRV_LOG(DEBUG, "DevX is supported");
2552                 /* The device is created, no need for environment. */
2553                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2554         } else {
2555                 /* The environment variable is still configured. */
2556                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2557                 err = errno ? errno : ENODEV;
2558                 /*
2559                  * The environment variable is not needed anymore,
2560                  * all device creation attempts are completed.
2561                  */
2562                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2563                 if (!sh->ctx)
2564                         return err;
2565                 DRV_LOG(DEBUG, "DevX is NOT supported");
2566                 err = 0;
2567         }
2568         if (!err && sh->ctx) {
2569                 /* Hint libmlx5 to use PMD allocator for data plane resources */
2570                 mlx5_glue->dv_set_context_attr(sh->ctx,
2571                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2572                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2573                                 .alloc = &mlx5_alloc_verbs_buf,
2574                                 .free = &mlx5_free_verbs_buf,
2575                                 .data = sh,
2576                         }));
2577         }
2578         return err;
2579 }
2580
2581 /**
2582  * Install shared asynchronous device events handler.
2583  * This function is implemented to support event sharing
2584  * between multiple ports of single IB device.
2585  *
2586  * @param sh
2587  *   Pointer to mlx5_dev_ctx_shared object.
2588  */
2589 void
2590 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2591 {
2592         int ret;
2593         int flags;
2594
2595         sh->intr_handle.fd = -1;
2596         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2597         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2598                     F_SETFL, flags | O_NONBLOCK);
2599         if (ret) {
2600                 DRV_LOG(INFO, "failed to change file descriptor async event"
2601                         " queue");
2602         } else {
2603                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2604                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2605                 if (rte_intr_callback_register(&sh->intr_handle,
2606                                         mlx5_dev_interrupt_handler, sh)) {
2607                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
2608                         sh->intr_handle.fd = -1;
2609                 }
2610         }
2611         if (sh->devx) {
2612 #ifdef HAVE_IBV_DEVX_ASYNC
2613                 sh->intr_handle_devx.fd = -1;
2614                 sh->devx_comp =
2615                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2616                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2617                 if (!devx_comp) {
2618                         DRV_LOG(INFO, "failed to allocate devx_comp.");
2619                         return;
2620                 }
2621                 flags = fcntl(devx_comp->fd, F_GETFL);
2622                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2623                 if (ret) {
2624                         DRV_LOG(INFO, "failed to change file descriptor"
2625                                 " devx comp");
2626                         return;
2627                 }
2628                 sh->intr_handle_devx.fd = devx_comp->fd;
2629                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2630                 if (rte_intr_callback_register(&sh->intr_handle_devx,
2631                                         mlx5_dev_interrupt_handler_devx, sh)) {
2632                         DRV_LOG(INFO, "Fail to install the devx shared"
2633                                 " interrupt.");
2634                         sh->intr_handle_devx.fd = -1;
2635                 }
2636 #endif /* HAVE_IBV_DEVX_ASYNC */
2637         }
2638 }
2639
2640 /**
2641  * Uninstall shared asynchronous device events handler.
2642  * This function is implemented to support event sharing
2643  * between multiple ports of single IB device.
2644  *
2645  * @param dev
2646  *   Pointer to mlx5_dev_ctx_shared object.
2647  */
2648 void
2649 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2650 {
2651         if (sh->intr_handle.fd >= 0)
2652                 mlx5_intr_callback_unregister(&sh->intr_handle,
2653                                               mlx5_dev_interrupt_handler, sh);
2654 #ifdef HAVE_IBV_DEVX_ASYNC
2655         if (sh->intr_handle_devx.fd >= 0)
2656                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2657                                   mlx5_dev_interrupt_handler_devx, sh);
2658         if (sh->devx_comp)
2659                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2660 #endif
2661 }
2662
2663 /**
2664  * Read statistics by a named counter.
2665  *
2666  * @param[in] priv
2667  *   Pointer to the private device data structure.
2668  * @param[in] ctr_name
2669  *   Pointer to the name of the statistic counter to read
2670  * @param[out] stat
2671  *   Pointer to read statistic value.
2672  * @return
2673  *   0 on success and stat is valud, 1 if failed to read the value
2674  *   rte_errno is set.
2675  *
2676  */
2677 int
2678 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2679                       uint64_t *stat)
2680 {
2681         int fd;
2682
2683         if (priv->sh) {
2684                 if (priv->q_counters != NULL &&
2685                     strcmp(ctr_name, "out_of_buffer") == 0)
2686                         return mlx5_devx_cmd_queue_counter_query(priv->sh->ctx,
2687                                                            0, (uint32_t *)stat);
2688                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2689                       priv->sh->ibdev_path,
2690                       priv->dev_port,
2691                       ctr_name);
2692                 fd = open(path, O_RDONLY);
2693                 /*
2694                  * in switchdev the file location is not per port
2695                  * but rather in <ibdev_path>/hw_counters/<file_name>.
2696                  */
2697                 if (fd == -1) {
2698                         MKSTR(path1, "%s/hw_counters/%s",
2699                               priv->sh->ibdev_path,
2700                               ctr_name);
2701                         fd = open(path1, O_RDONLY);
2702                 }
2703                 if (fd != -1) {
2704                         char buf[21] = {'\0'};
2705                         ssize_t n = read(fd, buf, sizeof(buf));
2706
2707                         close(fd);
2708                         if (n != -1) {
2709                                 *stat = strtoull(buf, NULL, 10);
2710                                 return 0;
2711                         }
2712                 }
2713         }
2714         *stat = 0;
2715         return 1;
2716 }
2717
2718 /**
2719  * Set the reg_mr and dereg_mr call backs
2720  *
2721  * @param reg_mr_cb[out]
2722  *   Pointer to reg_mr func
2723  * @param dereg_mr_cb[out]
2724  *   Pointer to dereg_mr func
2725  *
2726  */
2727 void
2728 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2729                       mlx5_dereg_mr_t *dereg_mr_cb)
2730 {
2731         *reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2732         *dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2733 }
2734
2735 /**
2736  * Remove a MAC address from device
2737  *
2738  * @param dev
2739  *   Pointer to Ethernet device structure.
2740  * @param index
2741  *   MAC address index.
2742  */
2743 void
2744 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2745 {
2746         struct mlx5_priv *priv = dev->data->dev_private;
2747         const int vf = priv->config.vf;
2748
2749         if (vf)
2750                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2751                                         mlx5_ifindex(dev), priv->mac_own,
2752                                         &dev->data->mac_addrs[index], index);
2753 }
2754
2755 /**
2756  * Adds a MAC address to the device
2757  *
2758  * @param dev
2759  *   Pointer to Ethernet device structure.
2760  * @param mac_addr
2761  *   MAC address to register.
2762  * @param index
2763  *   MAC address index.
2764  *
2765  * @return
2766  *   0 on success, a negative errno value otherwise
2767  */
2768 int
2769 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2770                      uint32_t index)
2771 {
2772         struct mlx5_priv *priv = dev->data->dev_private;
2773         const int vf = priv->config.vf;
2774         int ret = 0;
2775
2776         if (vf)
2777                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2778                                            mlx5_ifindex(dev), priv->mac_own,
2779                                            mac, index);
2780         return ret;
2781 }
2782
2783 /**
2784  * Modify a VF MAC address
2785  *
2786  * @param priv
2787  *   Pointer to device private data.
2788  * @param mac_addr
2789  *   MAC address to modify into.
2790  * @param iface_idx
2791  *   Net device interface index
2792  * @param vf_index
2793  *   VF index
2794  *
2795  * @return
2796  *   0 on success, a negative errno value otherwise
2797  */
2798 int
2799 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2800                            unsigned int iface_idx,
2801                            struct rte_ether_addr *mac_addr,
2802                            int vf_index)
2803 {
2804         return mlx5_nl_vf_mac_addr_modify
2805                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2806 }
2807
2808 /**
2809  * Set device promiscuous mode
2810  *
2811  * @param dev
2812  *   Pointer to Ethernet device structure.
2813  * @param enable
2814  *   0 - promiscuous is disabled, otherwise - enabled
2815  *
2816  * @return
2817  *   0 on success, a negative error value otherwise
2818  */
2819 int
2820 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2821 {
2822         struct mlx5_priv *priv = dev->data->dev_private;
2823
2824         return mlx5_nl_promisc(priv->nl_socket_route,
2825                                mlx5_ifindex(dev), !!enable);
2826 }
2827
2828 /**
2829  * Set device promiscuous mode
2830  *
2831  * @param dev
2832  *   Pointer to Ethernet device structure.
2833  * @param enable
2834  *   0 - all multicase is disabled, otherwise - enabled
2835  *
2836  * @return
2837  *   0 on success, a negative error value otherwise
2838  */
2839 int
2840 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2841 {
2842         struct mlx5_priv *priv = dev->data->dev_private;
2843
2844         return mlx5_nl_allmulti(priv->nl_socket_route,
2845                                 mlx5_ifindex(dev), !!enable);
2846 }
2847
2848 /**
2849  * Flush device MAC addresses
2850  *
2851  * @param dev
2852  *   Pointer to Ethernet device structure.
2853  *
2854  */
2855 void
2856 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2857 {
2858         struct mlx5_priv *priv = dev->data->dev_private;
2859
2860         mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2861                                dev->data->mac_addrs,
2862                                MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2863 }