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 <sys/mman.h>
  14 #include <linux/rtnetlink.h>
  15 #include <fcntl.h>
  16
  17 /* Verbs header. */
  18 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  19 #ifdef PEDANTIC
  20 #pragma GCC diagnostic ignored "-Wpedantic"
  21 #endif
  22 #include <infiniband/verbs.h>
  23 #ifdef PEDANTIC
  24 #pragma GCC diagnostic error "-Wpedantic"
  25 #endif
  26
  27 #include <rte_malloc.h>
  28 #include <rte_ethdev_driver.h>
  29 #include <rte_ethdev_pci.h>
  30 #include <rte_pci.h>
  31 #include <rte_bus_pci.h>
  32 #include <rte_common.h>
  33 #include <rte_kvargs.h>
  34 #include <rte_rwlock.h>
  35 #include <rte_spinlock.h>
  36 #include <rte_string_fns.h>
  37 #include <rte_alarm.h>
  38
  39 #include <mlx5_glue.h>
  40 #include <mlx5_devx_cmds.h>
  41 #include <mlx5_common.h>
  42 #include <mlx5_common_mp.h>
  43
  44 #include "mlx5_defs.h"
  45 #include "mlx5.h"
  46 #include "mlx5_utils.h"
  47 #include "mlx5_rxtx.h"
  48 #include "mlx5_autoconf.h"
  49 #include "mlx5_mr.h"
  50 #include "mlx5_flow.h"
  51 #include "rte_pmd_mlx5.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 /**
  65  * Get device name. Given an ibv_device pointer - return a
  66  * pointer to the corresponding device name.
  67  *
  68  * @param[in] dev
  69  *   Pointer to ibv device.
  70  *
  71  * @return
  72  *   Pointer to device name if dev is valid, NULL otherwise.
  73  */
  74 const char *
  75 mlx5_os_get_dev_device_name(void *dev)
  76 {
  77         if (!dev)
  78                 return NULL;
  79         return ((struct ibv_device *)dev)->name;
  80 }
  81
  82 /**
  83  * Get ibv device name. Given an ibv_context pointer - return a
  84  * pointer to the corresponding device name.
  85  *
  86  * @param[in] ctx
  87  *   Pointer to ibv context.
  88  *
  89  * @return
  90  *   Pointer to device name if ctx is valid, NULL otherwise.
  91  */
  92 const char *
  93 mlx5_os_get_ctx_device_name(void *ctx)
  94 {
  95         if (!ctx)
  96                 return NULL;
  97         return ((struct ibv_context *)ctx)->device->name;
  98 }
  99
 100 /**
 101  * Get ibv device path name. Given an ibv_context pointer - return a
 102  * pointer to the corresponding device path name.
 103  *
 104  * @param[in] ctx
 105  *   Pointer to ibv context.
 106  *
 107  * @return
 108  *   Pointer to device path name if ctx is valid, NULL otherwise.
 109  */
 110 const char *
 111 mlx5_os_get_ctx_device_path(void *ctx)
 112 {
 113         if (!ctx)
 114                 return NULL;
 115
 116         return ((struct ibv_context *)ctx)->device->ibdev_path;
 117 }
 118
 119 /**
 120  * Get umem id. Given a pointer to umem object of type
 121  * 'struct mlx5dv_devx_umem *' - return its id.
 122  *
 123  * @param[in] umem
 124  *   Pointer to umem object.
 125  *
 126  * @return
 127  *   The umem id if umem is valid, 0 otherwise.
 128  */
 129 uint32_t
 130 mlx5_os_get_umem_id(void *umem)
 131 {
 132         if (!umem)
 133                 return 0;
 134         return ((struct mlx5dv_devx_umem *)umem)->umem_id;
 135 }
 136
 137 /**
 138  * Get mlx5 device attributes. The glue function query_device_ex() is called
 139  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
 140  * device attributes from the glue out parameter.
 141  *
 142  * @param dev
 143  *   Pointer to ibv context.
 144  *
 145  * @param device_attr
 146  *   Pointer to mlx5 device attributes.
 147  *
 148  * @return
 149  *   0 on success, non zero error number otherwise
 150  */
 151 int
 152 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
 153 {
 154         int err;
 155         struct ibv_device_attr_ex attr_ex;
 156         memset(device_attr, 0, sizeof(*device_attr));
 157         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
 158         if (err)
 159                 return err;
 160
 161         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
 162         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
 163         device_attr->max_sge = attr_ex.orig_attr.max_sge;
 164         device_attr->max_cq = attr_ex.orig_attr.max_cq;
 165         device_attr->max_qp = attr_ex.orig_attr.max_qp;
 166         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 167         device_attr->max_rwq_indirection_table_size =
 168                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 169         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 170         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 171
 172         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 173         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 174         if (err)
 175                 return err;
 176
 177         device_attr->flags = dv_attr.flags;
 178         device_attr->comp_mask = dv_attr.comp_mask;
 179 #ifdef HAVE_IBV_MLX5_MOD_SWP
 180         device_attr->sw_parsing_offloads =
 181                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 182 #endif
 183         device_attr->min_single_stride_log_num_of_bytes =
 184                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 185         device_attr->max_single_stride_log_num_of_bytes =
 186                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 187         device_attr->min_single_wqe_log_num_of_strides =
 188                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 189         device_attr->max_single_wqe_log_num_of_strides =
 190                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 191         device_attr->stride_supported_qpts =
 192                 dv_attr.striding_rq_caps.supported_qpts;
 193 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 194         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 195 #endif
 196
 197         return err;
 198 }
 199
 200 /**
 201  * Verbs callback to allocate a memory. This function should allocate the space
 202  * according to the size provided residing inside a huge page.
 203  * Please note that all allocation must respect the alignment from libmlx5
 204  * (i.e. currently sysconf(_SC_PAGESIZE)).
 205  *
 206  * @param[in] size
 207  *   The size in bytes of the memory to allocate.
 208  * @param[in] data
 209  *   A pointer to the callback data.
 210  *
 211  * @return
 212  *   Allocated buffer, NULL otherwise and rte_errno is set.
 213  */
 214 static void *
 215 mlx5_alloc_verbs_buf(size_t size, void *data)
 216 {
 217         struct mlx5_priv *priv = data;
 218         void *ret;
 219         size_t alignment = sysconf(_SC_PAGESIZE);
 220         unsigned int socket = SOCKET_ID_ANY;
 221
 222         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 223                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 224
 225                 socket = ctrl->socket;
 226         } else if (priv->verbs_alloc_ctx.type ==
 227                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 228                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 229
 230                 socket = ctrl->socket;
 231         }
 232         MLX5_ASSERT(data != NULL);
 233         ret = rte_malloc_socket(__func__, size, alignment, socket);
 234         if (!ret && size)
 235                 rte_errno = ENOMEM;
 236         return ret;
 237 }
 238
 239 /**
 240  * Verbs callback to free a memory.
 241  *
 242  * @param[in] ptr
 243  *   A pointer to the memory to free.
 244  * @param[in] data
 245  *   A pointer to the callback data.
 246  */
 247 static void
 248 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 249 {
 250         MLX5_ASSERT(data != NULL);
 251         rte_free(ptr);
 252 }
 253
 254 /**
 255  * Initialize DR related data within private structure.
 256  * Routine checks the reference counter and does actual
 257  * resources creation/initialization only if counter is zero.
 258  *
 259  * @param[in] priv
 260  *   Pointer to the private device data structure.
 261  *
 262  * @return
 263  *   Zero on success, positive error code otherwise.
 264  */
 265 static int
 266 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 267 {
 268         struct mlx5_dev_ctx_shared *sh = priv->sh;
 269         char s[MLX5_HLIST_NAMESIZE];
 270         int err = 0;
 271
 272         if (!sh->flow_tbls)
 273                 err = mlx5_alloc_table_hash_list(priv);
 274         else
 275                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 276                         (void *)sh->flow_tbls);
 277         if (err)
 278                 return err;
 279         /* Create tags hash list table. */
 280         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 281         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 282         if (!sh->tag_table) {
 283                 DRV_LOG(ERR, "tags with hash creation failed.\n");
 284                 err = ENOMEM;
 285                 goto error;
 286         }
 287 #ifdef HAVE_MLX5DV_DR
 288         void *domain;
 289
 290         if (sh->dv_refcnt) {
 291                 /* Shared DV/DR structures is already initialized. */
 292                 sh->dv_refcnt++;
 293                 priv->dr_shared = 1;
 294                 return 0;
 295         }
 296         /* Reference counter is zero, we should initialize structures. */
 297         domain = mlx5_glue->dr_create_domain(sh->ctx,
 298                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 299         if (!domain) {
 300                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 301                 err = errno;
 302                 goto error;
 303         }
 304         sh->rx_domain = domain;
 305         domain = mlx5_glue->dr_create_domain(sh->ctx,
 306                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 307         if (!domain) {
 308                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 309                 err = errno;
 310                 goto error;
 311         }
 312         pthread_mutex_init(&sh->dv_mutex, NULL);
 313         sh->tx_domain = domain;
 314 #ifdef HAVE_MLX5DV_DR_ESWITCH
 315         if (priv->config.dv_esw_en) {
 316                 domain  = mlx5_glue->dr_create_domain
 317                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 318                 if (!domain) {
 319                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 320                         err = errno;
 321                         goto error;
 322                 }
 323                 sh->fdb_domain = domain;
 324                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 325         }
 326 #endif
 327         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 328                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 329                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 330                 if (sh->fdb_domain)
 331                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 332         }
 333         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 334 #endif /* HAVE_MLX5DV_DR */
 335         sh->dv_refcnt++;
 336         priv->dr_shared = 1;
 337         return 0;
 338 error:
 339         /* Rollback the created objects. */
 340         if (sh->rx_domain) {
 341                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 342                 sh->rx_domain = NULL;
 343         }
 344         if (sh->tx_domain) {
 345                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 346                 sh->tx_domain = NULL;
 347         }
 348         if (sh->fdb_domain) {
 349                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 350                 sh->fdb_domain = NULL;
 351         }
 352         if (sh->esw_drop_action) {
 353                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 354                 sh->esw_drop_action = NULL;
 355         }
 356         if (sh->pop_vlan_action) {
 357                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 358                 sh->pop_vlan_action = NULL;
 359         }
 360         if (sh->tag_table) {
 361                 /* tags should be destroyed with flow before. */
 362                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 363                 sh->tag_table = NULL;
 364         }
 365         mlx5_free_table_hash_list(priv);
 366         return err;
 367 }
 368
 369 /**
 370  * Destroy DR related data within private structure.
 371  *
 372  * @param[in] priv
 373  *   Pointer to the private device data structure.
 374  */
 375 void
 376 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 377 {
 378         struct mlx5_dev_ctx_shared *sh;
 379
 380         if (!priv->dr_shared)
 381                 return;
 382         priv->dr_shared = 0;
 383         sh = priv->sh;
 384         MLX5_ASSERT(sh);
 385 #ifdef HAVE_MLX5DV_DR
 386         MLX5_ASSERT(sh->dv_refcnt);
 387         if (sh->dv_refcnt && --sh->dv_refcnt)
 388                 return;
 389         if (sh->rx_domain) {
 390                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 391                 sh->rx_domain = NULL;
 392         }
 393         if (sh->tx_domain) {
 394                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 395                 sh->tx_domain = NULL;
 396         }
 397 #ifdef HAVE_MLX5DV_DR_ESWITCH
 398         if (sh->fdb_domain) {
 399                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 400                 sh->fdb_domain = NULL;
 401         }
 402         if (sh->esw_drop_action) {
 403                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 404                 sh->esw_drop_action = NULL;
 405         }
 406 #endif
 407         if (sh->pop_vlan_action) {
 408                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 409                 sh->pop_vlan_action = NULL;
 410         }
 411         pthread_mutex_destroy(&sh->dv_mutex);
 412 #endif /* HAVE_MLX5DV_DR */
 413         if (sh->tag_table) {
 414                 /* tags should be destroyed with flow before. */
 415                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 416                 sh->tag_table = NULL;
 417         }
 418         mlx5_free_table_hash_list(priv);
 419 }
 420
 421 /**
 422  * Spawn an Ethernet device from Verbs information.
 423  *
 424  * @param dpdk_dev
 425  *   Backing DPDK device.
 426  * @param spawn
 427  *   Verbs device parameters (name, port, switch_info) to spawn.
 428  * @param config
 429  *   Device configuration parameters.
 430  *
 431  * @return
 432  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 433  *   is set. The following errors are defined:
 434  *
 435  *   EBUSY: device is not supposed to be spawned.
 436  *   EEXIST: device is already spawned
 437  */
 438 static struct rte_eth_dev *
 439 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 440                struct mlx5_dev_spawn_data *spawn,
 441                struct mlx5_dev_config config)
 442 {
 443         const struct mlx5_switch_info *switch_info = &spawn->info;
 444         struct mlx5_dev_ctx_shared *sh = NULL;
 445         struct ibv_port_attr port_attr;
 446         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 447         struct rte_eth_dev *eth_dev = NULL;
 448         struct mlx5_priv *priv = NULL;
 449         int err = 0;
 450         unsigned int hw_padding = 0;
 451         unsigned int mps;
 452         unsigned int cqe_comp;
 453         unsigned int cqe_pad = 0;
 454         unsigned int tunnel_en = 0;
 455         unsigned int mpls_en = 0;
 456         unsigned int swp = 0;
 457         unsigned int mprq = 0;
 458         unsigned int mprq_min_stride_size_n = 0;
 459         unsigned int mprq_max_stride_size_n = 0;
 460         unsigned int mprq_min_stride_num_n = 0;
 461         unsigned int mprq_max_stride_num_n = 0;
 462         struct rte_ether_addr mac;
 463         char name[RTE_ETH_NAME_MAX_LEN];
 464         int own_domain_id = 0;
 465         uint16_t port_id;
 466         unsigned int i;
 467 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 468         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 469 #endif
 470
 471         /* Determine if this port representor is supposed to be spawned. */
 472         if (switch_info->representor && dpdk_dev->devargs) {
 473                 struct rte_eth_devargs eth_da;
 474
 475                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 476                 if (err) {
 477                         rte_errno = -err;
 478                         DRV_LOG(ERR, "failed to process device arguments: %s",
 479                                 strerror(rte_errno));
 480                         return NULL;
 481                 }
 482                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 483                         if (eth_da.representor_ports[i] ==
 484                             (uint16_t)switch_info->port_name)
 485                                 break;
 486                 if (i == eth_da.nb_representor_ports) {
 487                         rte_errno = EBUSY;
 488                         return NULL;
 489                 }
 490         }
 491         /* Build device name. */
 492         if (spawn->pf_bond <  0) {
 493                 /* Single device. */
 494                 if (!switch_info->representor)
 495                         strlcpy(name, dpdk_dev->name, sizeof(name));
 496                 else
 497                         snprintf(name, sizeof(name), "%s_representor_%u",
 498                                  dpdk_dev->name, switch_info->port_name);
 499         } else {
 500                 /* Bonding device. */
 501                 if (!switch_info->representor)
 502                         snprintf(name, sizeof(name), "%s_%s",
 503                                  dpdk_dev->name,
 504                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 505                 else
 506                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 507                                  dpdk_dev->name,
 508                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 509                                  switch_info->port_name);
 510         }
 511         /* check if the device is already spawned */
 512         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 513                 rte_errno = EEXIST;
 514                 return NULL;
 515         }
 516         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 517         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 518                 struct mlx5_mp_id mp_id;
 519
 520                 eth_dev = rte_eth_dev_attach_secondary(name);
 521                 if (eth_dev == NULL) {
 522                         DRV_LOG(ERR, "can not attach rte ethdev");
 523                         rte_errno = ENOMEM;
 524                         return NULL;
 525                 }
 526                 eth_dev->device = dpdk_dev;
 527                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
 528                 err = mlx5_proc_priv_init(eth_dev);
 529                 if (err)
 530                         return NULL;
 531                 mp_id.port_id = eth_dev->data->port_id;
 532                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 533                 /* Receive command fd from primary process */
 534                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 535                 if (err < 0)
 536                         goto err_secondary;
 537                 /* Remap UAR for Tx queues. */
 538                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 539                 if (err)
 540                         goto err_secondary;
 541                 /*
 542                  * Ethdev pointer is still required as input since
 543                  * the primary device is not accessible from the
 544                  * secondary process.
 545                  */
 546                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 547                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 548                 return eth_dev;
 549 err_secondary:
 550                 mlx5_dev_close(eth_dev);
 551                 return NULL;
 552         }
 553         /*
 554          * Some parameters ("tx_db_nc" in particularly) are needed in
 555          * advance to create dv/verbs device context. We proceed the
 556          * devargs here to get ones, and later proceed devargs again
 557          * to override some hardware settings.
 558          */
 559         err = mlx5_args(&config, dpdk_dev->devargs);
 560         if (err) {
 561                 err = rte_errno;
 562                 DRV_LOG(ERR, "failed to process device arguments: %s",
 563                         strerror(rte_errno));
 564                 goto error;
 565         }
 566         sh = mlx5_alloc_shared_ibctx(spawn, &config);
 567         if (!sh)
 568                 return NULL;
 569         config.devx = sh->devx;
 570 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 571         config.dest_tir = 1;
 572 #endif
 573 #ifdef HAVE_IBV_MLX5_MOD_SWP
 574         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 575 #endif
 576         /*
 577          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 578          * as all ConnectX-5 devices.
 579          */
 580 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 581         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 582 #endif
 583 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 584         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 585 #endif
 586         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 587         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 588                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 589                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 590                         mps = MLX5_MPW_ENHANCED;
 591                 } else {
 592                         DRV_LOG(DEBUG, "MPW is supported");
 593                         mps = MLX5_MPW;
 594                 }
 595         } else {
 596                 DRV_LOG(DEBUG, "MPW isn't supported");
 597                 mps = MLX5_MPW_DISABLED;
 598         }
 599 #ifdef HAVE_IBV_MLX5_MOD_SWP
 600         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 601                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 602         DRV_LOG(DEBUG, "SWP support: %u", swp);
 603 #endif
 604         config.swp = !!swp;
 605 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 606         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 607                 struct mlx5dv_striding_rq_caps mprq_caps =
 608                         dv_attr.striding_rq_caps;
 609
 610                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 611                         mprq_caps.min_single_stride_log_num_of_bytes);
 612                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 613                         mprq_caps.max_single_stride_log_num_of_bytes);
 614                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 615                         mprq_caps.min_single_wqe_log_num_of_strides);
 616                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 617                         mprq_caps.max_single_wqe_log_num_of_strides);
 618                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 619                         mprq_caps.supported_qpts);
 620                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 621                 mprq = 1;
 622                 mprq_min_stride_size_n =
 623                         mprq_caps.min_single_stride_log_num_of_bytes;
 624                 mprq_max_stride_size_n =
 625                         mprq_caps.max_single_stride_log_num_of_bytes;
 626                 mprq_min_stride_num_n =
 627                         mprq_caps.min_single_wqe_log_num_of_strides;
 628                 mprq_max_stride_num_n =
 629                         mprq_caps.max_single_wqe_log_num_of_strides;
 630         }
 631 #endif
 632         if (RTE_CACHE_LINE_SIZE == 128 &&
 633             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 634                 cqe_comp = 0;
 635         else
 636                 cqe_comp = 1;
 637         config.cqe_comp = cqe_comp;
 638 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 639         /* Whether device supports 128B Rx CQE padding. */
 640         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 641                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 642 #endif
 643 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 644         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 645                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 646                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 647                              (dv_attr.tunnel_offloads_caps &
 648                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 649                              (dv_attr.tunnel_offloads_caps &
 650                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 651         }
 652         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 653                 tunnel_en ? "" : "not ");
 654 #else
 655         DRV_LOG(WARNING,
 656                 "tunnel offloading disabled due to old OFED/rdma-core version");
 657 #endif
 658         config.tunnel_en = tunnel_en;
 659 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 660         mpls_en = ((dv_attr.tunnel_offloads_caps &
 661                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 662                    (dv_attr.tunnel_offloads_caps &
 663                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 664         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 665                 mpls_en ? "" : "not ");
 666 #else
 667         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 668                 " old OFED/rdma-core version or firmware configuration");
 669 #endif
 670         config.mpls_en = mpls_en;
 671         /* Check port status. */
 672         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 673         if (err) {
 674                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 675                 goto error;
 676         }
 677         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 678                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 679                 err = EINVAL;
 680                 goto error;
 681         }
 682         if (port_attr.state != IBV_PORT_ACTIVE)
 683                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 684                         mlx5_glue->port_state_str(port_attr.state),
 685                         port_attr.state);
 686         /* Allocate private eth device data. */
 687         priv = rte_zmalloc("ethdev private structure",
 688                            sizeof(*priv),
 689                            RTE_CACHE_LINE_SIZE);
 690         if (priv == NULL) {
 691                 DRV_LOG(ERR, "priv allocation failure");
 692                 err = ENOMEM;
 693                 goto error;
 694         }
 695         priv->sh = sh;
 696         priv->ibv_port = spawn->phys_port;
 697         priv->pci_dev = spawn->pci_dev;
 698         priv->mtu = RTE_ETHER_MTU;
 699         priv->mp_id.port_id = port_id;
 700         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 701 #ifndef RTE_ARCH_64
 702         /* Initialize UAR access locks for 32bit implementations. */
 703         rte_spinlock_init(&priv->uar_lock_cq);
 704         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
 705                 rte_spinlock_init(&priv->uar_lock[i]);
 706 #endif
 707         /* Some internal functions rely on Netlink sockets, open them now. */
 708         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 709         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 710         priv->representor = !!switch_info->representor;
 711         priv->master = !!switch_info->master;
 712         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 713         priv->vport_meta_tag = 0;
 714         priv->vport_meta_mask = 0;
 715         priv->pf_bond = spawn->pf_bond;
 716 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 717         /*
 718          * The DevX port query API is implemented. E-Switch may use
 719          * either vport or reg_c[0] metadata register to match on
 720          * vport index. The engaged part of metadata register is
 721          * defined by mask.
 722          */
 723         if (switch_info->representor || switch_info->master) {
 724                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 725                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 726                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 727                                                  &devx_port);
 728                 if (err) {
 729                         DRV_LOG(WARNING,
 730                                 "can't query devx port %d on device %s",
 731                                 spawn->phys_port,
 732                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 733                         devx_port.comp_mask = 0;
 734                 }
 735         }
 736         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 737                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 738                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 739                 if (!priv->vport_meta_mask) {
 740                         DRV_LOG(ERR, "vport zero mask for port %d"
 741                                      " on bonding device %s",
 742                                      spawn->phys_port,
 743                                      mlx5_os_get_dev_device_name
 744                                                         (spawn->phys_dev));
 745                         err = ENOTSUP;
 746                         goto error;
 747                 }
 748                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 749                         DRV_LOG(ERR, "invalid vport tag for port %d"
 750                                      " on bonding device %s",
 751                                      spawn->phys_port,
 752                                      mlx5_os_get_dev_device_name
 753                                                         (spawn->phys_dev));
 754                         err = ENOTSUP;
 755                         goto error;
 756                 }
 757         }
 758         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 759                 priv->vport_id = devx_port.vport_num;
 760         } else if (spawn->pf_bond >= 0) {
 761                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 762                              " on bonding device %s",
 763                              spawn->phys_port,
 764                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 765                 err = ENOTSUP;
 766                 goto error;
 767         } else {
 768                 /* Suppose vport index in compatible way. */
 769                 priv->vport_id = switch_info->representor ?
 770                                  switch_info->port_name + 1 : -1;
 771         }
 772 #else
 773         /*
 774          * Kernel/rdma_core support single E-Switch per PF configurations
 775          * only and vport_id field contains the vport index for
 776          * associated VF, which is deduced from representor port name.
 777          * For example, let's have the IB device port 10, it has
 778          * attached network device eth0, which has port name attribute
 779          * pf0vf2, we can deduce the VF number as 2, and set vport index
 780          * as 3 (2+1). This assigning schema should be changed if the
 781          * multiple E-Switch instances per PF configurations or/and PCI
 782          * subfunctions are added.
 783          */
 784         priv->vport_id = switch_info->representor ?
 785                          switch_info->port_name + 1 : -1;
 786 #endif
 787         /* representor_id field keeps the unmodified VF index. */
 788         priv->representor_id = switch_info->representor ?
 789                                switch_info->port_name : -1;
 790         /*
 791          * Look for sibling devices in order to reuse their switch domain
 792          * if any, otherwise allocate one.
 793          */
 794         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 795                 const struct mlx5_priv *opriv =
 796                         rte_eth_devices[port_id].data->dev_private;
 797
 798                 if (!opriv ||
 799                     opriv->sh != priv->sh ||
 800                         opriv->domain_id ==
 801                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 802                         continue;
 803                 priv->domain_id = opriv->domain_id;
 804                 break;
 805         }
 806         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 807                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 808                 if (err) {
 809                         err = rte_errno;
 810                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 811                                 strerror(rte_errno));
 812                         goto error;
 813                 }
 814                 own_domain_id = 1;
 815         }
 816         /* Override some values set by hardware configuration. */
 817         mlx5_args(&config, dpdk_dev->devargs);
 818         err = mlx5_dev_check_sibling_config(priv, &config);
 819         if (err)
 820                 goto error;
 821         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 822                             IBV_DEVICE_RAW_IP_CSUM);
 823         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 824                 (config.hw_csum ? "" : "not "));
 825 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 826         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 827         DRV_LOG(DEBUG, "counters are not supported");
 828 #endif
 829 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 830         if (config.dv_flow_en) {
 831                 DRV_LOG(WARNING, "DV flow is not supported");
 832                 config.dv_flow_en = 0;
 833         }
 834 #endif
 835         config.ind_table_max_size =
 836                 sh->device_attr.max_rwq_indirection_table_size;
 837         /*
 838          * Remove this check once DPDK supports larger/variable
 839          * indirection tables.
 840          */
 841         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 842                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 843         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 844                 config.ind_table_max_size);
 845         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 846                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 847         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 848                 (config.hw_vlan_strip ? "" : "not "));
 849         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 850                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 851         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 852                 (config.hw_fcs_strip ? "" : "not "));
 853 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 854         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 855 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 856         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 857                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 858 #endif
 859         if (config.hw_padding && !hw_padding) {
 860                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 861                 config.hw_padding = 0;
 862         } else if (config.hw_padding) {
 863                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 864         }
 865         config.tso = (sh->device_attr.max_tso > 0 &&
 866                       (sh->device_attr.tso_supported_qpts &
 867                        (1 << IBV_QPT_RAW_PACKET)));
 868         if (config.tso)
 869                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 870         /*
 871          * MPW is disabled by default, while the Enhanced MPW is enabled
 872          * by default.
 873          */
 874         if (config.mps == MLX5_ARG_UNSET)
 875                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 876                                                           MLX5_MPW_DISABLED;
 877         else
 878                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 879         DRV_LOG(INFO, "%sMPS is %s",
 880                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 881                 config.mps == MLX5_MPW ? "legacy " : "",
 882                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 883         if (config.cqe_comp && !cqe_comp) {
 884                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 885                 config.cqe_comp = 0;
 886         }
 887         if (config.cqe_pad && !cqe_pad) {
 888                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 889                 config.cqe_pad = 0;
 890         } else if (config.cqe_pad) {
 891                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 892         }
 893         if (config.devx) {
 894                 priv->counter_fallback = 0;
 895                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 896                 if (err) {
 897                         err = -err;
 898                         goto error;
 899                 }
 900                 if (!config.hca_attr.flow_counters_dump)
 901                         priv->counter_fallback = 1;
 902 #ifndef HAVE_IBV_DEVX_ASYNC
 903                 priv->counter_fallback = 1;
 904 #endif
 905                 if (priv->counter_fallback)
 906                         DRV_LOG(INFO, "Use fall-back DV counter management");
 907                 /* Check for LRO support. */
 908                 if (config.dest_tir && config.hca_attr.lro_cap &&
 909                     config.dv_flow_en) {
 910                         /* TBD check tunnel lro caps. */
 911                         config.lro.supported = config.hca_attr.lro_cap;
 912                         DRV_LOG(DEBUG, "Device supports LRO");
 913                         /*
 914                          * If LRO timeout is not configured by application,
 915                          * use the minimal supported value.
 916                          */
 917                         if (!config.lro.timeout)
 918                                 config.lro.timeout =
 919                                 config.hca_attr.lro_timer_supported_periods[0];
 920                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 921                                 config.lro.timeout);
 922                 }
 923 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 924                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 925                     config.dv_flow_en) {
 926                         uint8_t reg_c_mask =
 927                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 928                         /*
 929                          * Meter needs two REG_C's for color match and pre-sfx
 930                          * flow match. Here get the REG_C for color match.
 931                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 932                          */
 933                         reg_c_mask &= 0xfc;
 934                         if (__builtin_popcount(reg_c_mask) < 1) {
 935                                 priv->mtr_en = 0;
 936                                 DRV_LOG(WARNING, "No available register for"
 937                                         " meter.");
 938                         } else {
 939                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 940                                                       REG_C_0;
 941                                 priv->mtr_en = 1;
 942                                 priv->mtr_reg_share =
 943                                       config.hca_attr.qos.flow_meter_reg_share;
 944                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 945                                         priv->mtr_color_reg);
 946                         }
 947                 }
 948 #endif
 949         }
 950         if (config.mprq.enabled && mprq) {
 951                 if (config.mprq.stride_num_n &&
 952                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 953                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 954                         config.mprq.stride_num_n =
 955                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 956                                                 mprq_min_stride_num_n),
 957                                         mprq_max_stride_num_n);
 958                         DRV_LOG(WARNING,
 959                                 "the number of strides"
 960                                 " for Multi-Packet RQ is out of range,"
 961                                 " setting default value (%u)",
 962                                 1 << config.mprq.stride_num_n);
 963                 }
 964                 if (config.mprq.stride_size_n &&
 965                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 966                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 967                         config.mprq.stride_size_n =
 968                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 969                                                 mprq_min_stride_size_n),
 970                                         mprq_max_stride_size_n);
 971                         DRV_LOG(WARNING,
 972                                 "the size of a stride"
 973                                 " for Multi-Packet RQ is out of range,"
 974                                 " setting default value (%u)",
 975                                 1 << config.mprq.stride_size_n);
 976                 }
 977                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
 978                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
 979         } else if (config.mprq.enabled && !mprq) {
 980                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
 981                 config.mprq.enabled = 0;
 982         }
 983         if (config.max_dump_files_num == 0)
 984                 config.max_dump_files_num = 128;
 985         eth_dev = rte_eth_dev_allocate(name);
 986         if (eth_dev == NULL) {
 987                 DRV_LOG(ERR, "can not allocate rte ethdev");
 988                 err = ENOMEM;
 989                 goto error;
 990         }
 991         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
 992         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 993         if (priv->representor) {
 994                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
 995                 eth_dev->data->representor_id = priv->representor_id;
 996         }
 997         /*
 998          * Store associated network device interface index. This index
 999          * is permanent throughout the lifetime of device. So, we may store
1000          * the ifindex here and use the cached value further.
1001          */
1002         MLX5_ASSERT(spawn->ifindex);
1003         priv->if_index = spawn->ifindex;
1004         eth_dev->data->dev_private = priv;
1005         priv->dev_data = eth_dev->data;
1006         eth_dev->data->mac_addrs = priv->mac;
1007         eth_dev->device = dpdk_dev;
1008         /* Configure the first MAC address by default. */
1009         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1010                 DRV_LOG(ERR,
1011                         "port %u cannot get MAC address, is mlx5_en"
1012                         " loaded? (errno: %s)",
1013                         eth_dev->data->port_id, strerror(rte_errno));
1014                 err = ENODEV;
1015                 goto error;
1016         }
1017         DRV_LOG(INFO,
1018                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1019                 eth_dev->data->port_id,
1020                 mac.addr_bytes[0], mac.addr_bytes[1],
1021                 mac.addr_bytes[2], mac.addr_bytes[3],
1022                 mac.addr_bytes[4], mac.addr_bytes[5]);
1023 #ifdef RTE_LIBRTE_MLX5_DEBUG
1024         {
1025                 char ifname[IF_NAMESIZE];
1026
1027                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1028                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1029                                 eth_dev->data->port_id, ifname);
1030                 else
1031                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1032                                 eth_dev->data->port_id);
1033         }
1034 #endif
1035         /* Get actual MTU if possible. */
1036         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1037         if (err) {
1038                 err = rte_errno;
1039                 goto error;
1040         }
1041         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1042                 priv->mtu);
1043         /* Initialize burst functions to prevent crashes before link-up. */
1044         eth_dev->rx_pkt_burst = removed_rx_burst;
1045         eth_dev->tx_pkt_burst = removed_tx_burst;
1046         eth_dev->dev_ops = &mlx5_dev_ops;
1047         /* Register MAC address. */
1048         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1049         if (config.vf && config.vf_nl_en)
1050                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1051                                       mlx5_ifindex(eth_dev),
1052                                       eth_dev->data->mac_addrs,
1053                                       MLX5_MAX_MAC_ADDRESSES);
1054         priv->flows = 0;
1055         priv->ctrl_flows = 0;
1056         TAILQ_INIT(&priv->flow_meters);
1057         TAILQ_INIT(&priv->flow_meter_profiles);
1058         /* Hint libmlx5 to use PMD allocator for data plane resources */
1059         struct mlx5dv_ctx_allocators alctr = {
1060                 .alloc = &mlx5_alloc_verbs_buf,
1061                 .free = &mlx5_free_verbs_buf,
1062                 .data = priv,
1063         };
1064         mlx5_glue->dv_set_context_attr(sh->ctx,
1065                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1066                                        (void *)((uintptr_t)&alctr));
1067         /* Bring Ethernet device up. */
1068         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1069                 eth_dev->data->port_id);
1070         mlx5_set_link_up(eth_dev);
1071         /*
1072          * Even though the interrupt handler is not installed yet,
1073          * interrupts will still trigger on the async_fd from
1074          * Verbs context returned by ibv_open_device().
1075          */
1076         mlx5_link_update(eth_dev, 0);
1077 #ifdef HAVE_MLX5DV_DR_ESWITCH
1078         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1079               (switch_info->representor || switch_info->master)))
1080                 config.dv_esw_en = 0;
1081 #else
1082         config.dv_esw_en = 0;
1083 #endif
1084         /* Detect minimal data bytes to inline. */
1085         mlx5_set_min_inline(spawn, &config);
1086         /* Store device configuration on private structure. */
1087         priv->config = config;
1088         /* Create context for virtual machine VLAN workaround. */
1089         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1090         if (config.dv_flow_en) {
1091                 err = mlx5_alloc_shared_dr(priv);
1092                 if (err)
1093                         goto error;
1094                 /*
1095                  * RSS id is shared with meter flow id. Meter flow id can only
1096                  * use the 24 MSB of the register.
1097                  */
1098                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1099                                      MLX5_MTR_COLOR_BITS);
1100                 if (!priv->qrss_id_pool) {
1101                         DRV_LOG(ERR, "can't create flow id pool");
1102                         err = ENOMEM;
1103                         goto error;
1104                 }
1105         }
1106         /* Supported Verbs flow priority number detection. */
1107         err = mlx5_flow_discover_priorities(eth_dev);
1108         if (err < 0) {
1109                 err = -err;
1110                 goto error;
1111         }
1112         priv->config.flow_prio = err;
1113         if (!priv->config.dv_esw_en &&
1114             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1115                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1116                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1117                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1118         }
1119         mlx5_set_metadata_mask(eth_dev);
1120         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1121             !priv->sh->dv_regc0_mask) {
1122                 DRV_LOG(ERR, "metadata mode %u is not supported "
1123                              "(no metadata reg_c[0] is available)",
1124                              priv->config.dv_xmeta_en);
1125                         err = ENOTSUP;
1126                         goto error;
1127         }
1128         /*
1129          * Allocate the buffer for flow creating, just once.
1130          * The allocation must be done before any flow creating.
1131          */
1132         mlx5_flow_alloc_intermediate(eth_dev);
1133         /* Query availability of metadata reg_c's. */
1134         err = mlx5_flow_discover_mreg_c(eth_dev);
1135         if (err < 0) {
1136                 err = -err;
1137                 goto error;
1138         }
1139         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1140                 DRV_LOG(DEBUG,
1141                         "port %u extensive metadata register is not supported",
1142                         eth_dev->data->port_id);
1143                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1144                         DRV_LOG(ERR, "metadata mode %u is not supported "
1145                                      "(no metadata registers available)",
1146                                      priv->config.dv_xmeta_en);
1147                         err = ENOTSUP;
1148                         goto error;
1149                 }
1150         }
1151         if (priv->config.dv_flow_en &&
1152             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1153             mlx5_flow_ext_mreg_supported(eth_dev) &&
1154             priv->sh->dv_regc0_mask) {
1155                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1156                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1157                 if (!priv->mreg_cp_tbl) {
1158                         err = ENOMEM;
1159                         goto error;
1160                 }
1161         }
1162         return eth_dev;
1163 error:
1164         if (priv) {
1165                 if (priv->mreg_cp_tbl)
1166                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1167                 if (priv->sh)
1168                         mlx5_os_free_shared_dr(priv);
1169                 if (priv->nl_socket_route >= 0)
1170                         close(priv->nl_socket_route);
1171                 if (priv->nl_socket_rdma >= 0)
1172                         close(priv->nl_socket_rdma);
1173                 if (priv->vmwa_context)
1174                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1175                 if (priv->qrss_id_pool)
1176                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1177                 if (own_domain_id)
1178                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1179                 rte_free(priv);
1180                 if (eth_dev != NULL)
1181                         eth_dev->data->dev_private = NULL;
1182         }
1183         if (eth_dev != NULL) {
1184                 /* mac_addrs must not be freed alone because part of
1185                  * dev_private
1186                  **/
1187                 eth_dev->data->mac_addrs = NULL;
1188                 rte_eth_dev_release_port(eth_dev);
1189         }
1190         if (sh)
1191                 mlx5_free_shared_ibctx(sh);
1192         MLX5_ASSERT(err > 0);
1193         rte_errno = err;
1194         return NULL;
1195 }
1196
1197 /**
1198  * Comparison callback to sort device data.
1199  *
1200  * This is meant to be used with qsort().
1201  *
1202  * @param a[in]
1203  *   Pointer to pointer to first data object.
1204  * @param b[in]
1205  *   Pointer to pointer to second data object.
1206  *
1207  * @return
1208  *   0 if both objects are equal, less than 0 if the first argument is less
1209  *   than the second, greater than 0 otherwise.
1210  */
1211 static int
1212 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1213 {
1214         const struct mlx5_switch_info *si_a =
1215                 &((const struct mlx5_dev_spawn_data *)a)->info;
1216         const struct mlx5_switch_info *si_b =
1217                 &((const struct mlx5_dev_spawn_data *)b)->info;
1218         int ret;
1219
1220         /* Master device first. */
1221         ret = si_b->master - si_a->master;
1222         if (ret)
1223                 return ret;
1224         /* Then representor devices. */
1225         ret = si_b->representor - si_a->representor;
1226         if (ret)
1227                 return ret;
1228         /* Unidentified devices come last in no specific order. */
1229         if (!si_a->representor)
1230                 return 0;
1231         /* Order representors by name. */
1232         return si_a->port_name - si_b->port_name;
1233 }
1234
1235 /**
1236  * Match PCI information for possible slaves of bonding device.
1237  *
1238  * @param[in] ibv_dev
1239  *   Pointer to Infiniband device structure.
1240  * @param[in] pci_dev
1241  *   Pointer to PCI device structure to match PCI address.
1242  * @param[in] nl_rdma
1243  *   Netlink RDMA group socket handle.
1244  *
1245  * @return
1246  *   negative value if no bonding device found, otherwise
1247  *   positive index of slave PF in bonding.
1248  */
1249 static int
1250 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1251                            const struct rte_pci_device *pci_dev,
1252                            int nl_rdma)
1253 {
1254         char ifname[IF_NAMESIZE + 1];
1255         unsigned int ifindex;
1256         unsigned int np, i;
1257         FILE *file = NULL;
1258         int pf = -1;
1259
1260         /*
1261          * Try to get master device name. If something goes
1262          * wrong suppose the lack of kernel support and no
1263          * bonding devices.
1264          */
1265         if (nl_rdma < 0)
1266                 return -1;
1267         if (!strstr(ibv_dev->name, "bond"))
1268                 return -1;
1269         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1270         if (!np)
1271                 return -1;
1272         /*
1273          * The Master device might not be on the predefined
1274          * port (not on port index 1, it is not garanted),
1275          * we have to scan all Infiniband device port and
1276          * find master.
1277          */
1278         for (i = 1; i <= np; ++i) {
1279                 /* Check whether Infiniband port is populated. */
1280                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1281                 if (!ifindex)
1282                         continue;
1283                 if (!if_indextoname(ifindex, ifname))
1284                         continue;
1285                 /* Try to read bonding slave names from sysfs. */
1286                 MKSTR(slaves,
1287                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1288                 file = fopen(slaves, "r");
1289                 if (file)
1290                         break;
1291         }
1292         if (!file)
1293                 return -1;
1294         /* Use safe format to check maximal buffer length. */
1295         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1296         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1297                 char tmp_str[IF_NAMESIZE + 32];
1298                 struct rte_pci_addr pci_addr;
1299                 struct mlx5_switch_info info;
1300
1301                 /* Process slave interface names in the loop. */
1302                 snprintf(tmp_str, sizeof(tmp_str),
1303                          "/sys/class/net/%s", ifname);
1304                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1305                         DRV_LOG(WARNING, "can not get PCI address"
1306                                          " for netdev \"%s\"", ifname);
1307                         continue;
1308                 }
1309                 if (pci_dev->addr.domain != pci_addr.domain ||
1310                     pci_dev->addr.bus != pci_addr.bus ||
1311                     pci_dev->addr.devid != pci_addr.devid ||
1312                     pci_dev->addr.function != pci_addr.function)
1313                         continue;
1314                 /* Slave interface PCI address match found. */
1315                 fclose(file);
1316                 snprintf(tmp_str, sizeof(tmp_str),
1317                          "/sys/class/net/%s/phys_port_name", ifname);
1318                 file = fopen(tmp_str, "rb");
1319                 if (!file)
1320                         break;
1321                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1322                 if (fscanf(file, "%32s", tmp_str) == 1)
1323                         mlx5_translate_port_name(tmp_str, &info);
1324                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1325                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1326                         pf = info.port_name;
1327                 break;
1328         }
1329         if (file)
1330                 fclose(file);
1331         return pf;
1332 }
1333
1334 /**
1335  * DPDK callback to register a PCI device.
1336  *
1337  * This function spawns Ethernet devices out of a given PCI device.
1338  *
1339  * @param[in] pci_drv
1340  *   PCI driver structure (mlx5_driver).
1341  * @param[in] pci_dev
1342  *   PCI device information.
1343  *
1344  * @return
1345  *   0 on success, a negative errno value otherwise and rte_errno is set.
1346  */
1347 int
1348 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1349                   struct rte_pci_device *pci_dev)
1350 {
1351         struct ibv_device **ibv_list;
1352         /*
1353          * Number of found IB Devices matching with requested PCI BDF.
1354          * nd != 1 means there are multiple IB devices over the same
1355          * PCI device and we have representors and master.
1356          */
1357         unsigned int nd = 0;
1358         /*
1359          * Number of found IB device Ports. nd = 1 and np = 1..n means
1360          * we have the single multiport IB device, and there may be
1361          * representors attached to some of found ports.
1362          */
1363         unsigned int np = 0;
1364         /*
1365          * Number of DPDK ethernet devices to Spawn - either over
1366          * multiple IB devices or multiple ports of single IB device.
1367          * Actually this is the number of iterations to spawn.
1368          */
1369         unsigned int ns = 0;
1370         /*
1371          * Bonding device
1372          *   < 0 - no bonding device (single one)
1373          *  >= 0 - bonding device (value is slave PF index)
1374          */
1375         int bd = -1;
1376         struct mlx5_dev_spawn_data *list = NULL;
1377         struct mlx5_dev_config dev_config;
1378         int ret;
1379
1380         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1381                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1382                         " driver.");
1383                 return 1;
1384         }
1385         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1386                 mlx5_pmd_socket_init();
1387         ret = mlx5_init_once();
1388         if (ret) {
1389                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1390                         strerror(rte_errno));
1391                 return -rte_errno;
1392         }
1393         MLX5_ASSERT(pci_drv == &mlx5_driver);
1394         errno = 0;
1395         ibv_list = mlx5_glue->get_device_list(&ret);
1396         if (!ibv_list) {
1397                 rte_errno = errno ? errno : ENOSYS;
1398                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1399                 return -rte_errno;
1400         }
1401         /*
1402          * First scan the list of all Infiniband devices to find
1403          * matching ones, gathering into the list.
1404          */
1405         struct ibv_device *ibv_match[ret + 1];
1406         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1407         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1408         unsigned int i;
1409
1410         while (ret-- > 0) {
1411                 struct rte_pci_addr pci_addr;
1412
1413                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1414                 bd = mlx5_device_bond_pci_match
1415                                 (ibv_list[ret], pci_dev, nl_rdma);
1416                 if (bd >= 0) {
1417                         /*
1418                          * Bonding device detected. Only one match is allowed,
1419                          * the bonding is supported over multi-port IB device,
1420                          * there should be no matches on representor PCI
1421                          * functions or non VF LAG bonding devices with
1422                          * specified address.
1423                          */
1424                         if (nd) {
1425                                 DRV_LOG(ERR,
1426                                         "multiple PCI match on bonding device"
1427                                         "\"%s\" found", ibv_list[ret]->name);
1428                                 rte_errno = ENOENT;
1429                                 ret = -rte_errno;
1430                                 goto exit;
1431                         }
1432                         DRV_LOG(INFO, "PCI information matches for"
1433                                       " slave %d bonding device \"%s\"",
1434                                       bd, ibv_list[ret]->name);
1435                         ibv_match[nd++] = ibv_list[ret];
1436                         break;
1437                 }
1438                 if (mlx5_dev_to_pci_addr
1439                         (ibv_list[ret]->ibdev_path, &pci_addr))
1440                         continue;
1441                 if (pci_dev->addr.domain != pci_addr.domain ||
1442                     pci_dev->addr.bus != pci_addr.bus ||
1443                     pci_dev->addr.devid != pci_addr.devid ||
1444                     pci_dev->addr.function != pci_addr.function)
1445                         continue;
1446                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1447                         ibv_list[ret]->name);
1448                 ibv_match[nd++] = ibv_list[ret];
1449         }
1450         ibv_match[nd] = NULL;
1451         if (!nd) {
1452                 /* No device matches, just complain and bail out. */
1453                 DRV_LOG(WARNING,
1454                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1455                         " are kernel drivers loaded?",
1456                         pci_dev->addr.domain, pci_dev->addr.bus,
1457                         pci_dev->addr.devid, pci_dev->addr.function);
1458                 rte_errno = ENOENT;
1459                 ret = -rte_errno;
1460                 goto exit;
1461         }
1462         if (nd == 1) {
1463                 /*
1464                  * Found single matching device may have multiple ports.
1465                  * Each port may be representor, we have to check the port
1466                  * number and check the representors existence.
1467                  */
1468                 if (nl_rdma >= 0)
1469                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1470                 if (!np)
1471                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1472                                          " ports number", ibv_match[0]->name);
1473                 if (bd >= 0 && !np) {
1474                         DRV_LOG(ERR, "can not get ports"
1475                                      " for bonding device");
1476                         rte_errno = ENOENT;
1477                         ret = -rte_errno;
1478                         goto exit;
1479                 }
1480         }
1481 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1482         if (bd >= 0) {
1483                 /*
1484                  * This may happen if there is VF LAG kernel support and
1485                  * application is compiled with older rdma_core library.
1486                  */
1487                 DRV_LOG(ERR,
1488                         "No kernel/verbs support for VF LAG bonding found.");
1489                 rte_errno = ENOTSUP;
1490                 ret = -rte_errno;
1491                 goto exit;
1492         }
1493 #endif
1494         /*
1495          * Now we can determine the maximal
1496          * amount of devices to be spawned.
1497          */
1498         list = rte_zmalloc("device spawn data",
1499                          sizeof(struct mlx5_dev_spawn_data) *
1500                          (np ? np : nd),
1501                          RTE_CACHE_LINE_SIZE);
1502         if (!list) {
1503                 DRV_LOG(ERR, "spawn data array allocation failure");
1504                 rte_errno = ENOMEM;
1505                 ret = -rte_errno;
1506                 goto exit;
1507         }
1508         if (bd >= 0 || np > 1) {
1509                 /*
1510                  * Single IB device with multiple ports found,
1511                  * it may be E-Switch master device and representors.
1512                  * We have to perform identification through the ports.
1513                  */
1514                 MLX5_ASSERT(nl_rdma >= 0);
1515                 MLX5_ASSERT(ns == 0);
1516                 MLX5_ASSERT(nd == 1);
1517                 MLX5_ASSERT(np);
1518                 for (i = 1; i <= np; ++i) {
1519                         list[ns].max_port = np;
1520                         list[ns].phys_port = i;
1521                         list[ns].phys_dev = ibv_match[0];
1522                         list[ns].eth_dev = NULL;
1523                         list[ns].pci_dev = pci_dev;
1524                         list[ns].pf_bond = bd;
1525                         list[ns].ifindex = mlx5_nl_ifindex
1526                                 (nl_rdma,
1527                                 mlx5_os_get_dev_device_name
1528                                                 (list[ns].phys_dev), i);
1529                         if (!list[ns].ifindex) {
1530                                 /*
1531                                  * No network interface index found for the
1532                                  * specified port, it means there is no
1533                                  * representor on this port. It's OK,
1534                                  * there can be disabled ports, for example
1535                                  * if sriov_numvfs < sriov_totalvfs.
1536                                  */
1537                                 continue;
1538                         }
1539                         ret = -1;
1540                         if (nl_route >= 0)
1541                                 ret = mlx5_nl_switch_info
1542                                                (nl_route,
1543                                                 list[ns].ifindex,
1544                                                 &list[ns].info);
1545                         if (ret || (!list[ns].info.representor &&
1546                                     !list[ns].info.master)) {
1547                                 /*
1548                                  * We failed to recognize representors with
1549                                  * Netlink, let's try to perform the task
1550                                  * with sysfs.
1551                                  */
1552                                 ret =  mlx5_sysfs_switch_info
1553                                                 (list[ns].ifindex,
1554                                                  &list[ns].info);
1555                         }
1556                         if (!ret && bd >= 0) {
1557                                 switch (list[ns].info.name_type) {
1558                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1559                                         if (list[ns].info.port_name == bd)
1560                                                 ns++;
1561                                         break;
1562                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1563                                         if (list[ns].info.pf_num == bd)
1564                                                 ns++;
1565                                         break;
1566                                 default:
1567                                         break;
1568                                 }
1569                                 continue;
1570                         }
1571                         if (!ret && (list[ns].info.representor ^
1572                                      list[ns].info.master))
1573                                 ns++;
1574                 }
1575                 if (!ns) {
1576                         DRV_LOG(ERR,
1577                                 "unable to recognize master/representors"
1578                                 " on the IB device with multiple ports");
1579                         rte_errno = ENOENT;
1580                         ret = -rte_errno;
1581                         goto exit;
1582                 }
1583         } else {
1584                 /*
1585                  * The existence of several matching entries (nd > 1) means
1586                  * port representors have been instantiated. No existing Verbs
1587                  * call nor sysfs entries can tell them apart, this can only
1588                  * be done through Netlink calls assuming kernel drivers are
1589                  * recent enough to support them.
1590                  *
1591                  * In the event of identification failure through Netlink,
1592                  * try again through sysfs, then:
1593                  *
1594                  * 1. A single IB device matches (nd == 1) with single
1595                  *    port (np=0/1) and is not a representor, assume
1596                  *    no switch support.
1597                  *
1598                  * 2. Otherwise no safe assumptions can be made;
1599                  *    complain louder and bail out.
1600                  */
1601                 for (i = 0; i != nd; ++i) {
1602                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1603                         list[ns].max_port = 1;
1604                         list[ns].phys_port = 1;
1605                         list[ns].phys_dev = ibv_match[i];
1606                         list[ns].eth_dev = NULL;
1607                         list[ns].pci_dev = pci_dev;
1608                         list[ns].pf_bond = -1;
1609                         list[ns].ifindex = 0;
1610                         if (nl_rdma >= 0)
1611                                 list[ns].ifindex = mlx5_nl_ifindex
1612                                 (nl_rdma,
1613                                 mlx5_os_get_dev_device_name
1614                                                 (list[ns].phys_dev), 1);
1615                         if (!list[ns].ifindex) {
1616                                 char ifname[IF_NAMESIZE];
1617
1618                                 /*
1619                                  * Netlink failed, it may happen with old
1620                                  * ib_core kernel driver (before 4.16).
1621                                  * We can assume there is old driver because
1622                                  * here we are processing single ports IB
1623                                  * devices. Let's try sysfs to retrieve
1624                                  * the ifindex. The method works for
1625                                  * master device only.
1626                                  */
1627                                 if (nd > 1) {
1628                                         /*
1629                                          * Multiple devices found, assume
1630                                          * representors, can not distinguish
1631                                          * master/representor and retrieve
1632                                          * ifindex via sysfs.
1633                                          */
1634                                         continue;
1635                                 }
1636                                 ret = mlx5_get_master_ifname
1637                                         (ibv_match[i]->ibdev_path, &ifname);
1638                                 if (!ret)
1639                                         list[ns].ifindex =
1640                                                 if_nametoindex(ifname);
1641                                 if (!list[ns].ifindex) {
1642                                         /*
1643                                          * No network interface index found
1644                                          * for the specified device, it means
1645                                          * there it is neither representor
1646                                          * nor master.
1647                                          */
1648                                         continue;
1649                                 }
1650                         }
1651                         ret = -1;
1652                         if (nl_route >= 0)
1653                                 ret = mlx5_nl_switch_info
1654                                                (nl_route,
1655                                                 list[ns].ifindex,
1656                                                 &list[ns].info);
1657                         if (ret || (!list[ns].info.representor &&
1658                                     !list[ns].info.master)) {
1659                                 /*
1660                                  * We failed to recognize representors with
1661                                  * Netlink, let's try to perform the task
1662                                  * with sysfs.
1663                                  */
1664                                 ret =  mlx5_sysfs_switch_info
1665                                                 (list[ns].ifindex,
1666                                                  &list[ns].info);
1667                         }
1668                         if (!ret && (list[ns].info.representor ^
1669                                      list[ns].info.master)) {
1670                                 ns++;
1671                         } else if ((nd == 1) &&
1672                                    !list[ns].info.representor &&
1673                                    !list[ns].info.master) {
1674                                 /*
1675                                  * Single IB device with
1676                                  * one physical port and
1677                                  * attached network device.
1678                                  * May be SRIOV is not enabled
1679                                  * or there is no representors.
1680                                  */
1681                                 DRV_LOG(INFO, "no E-Switch support detected");
1682                                 ns++;
1683                                 break;
1684                         }
1685                 }
1686                 if (!ns) {
1687                         DRV_LOG(ERR,
1688                                 "unable to recognize master/representors"
1689                                 " on the multiple IB devices");
1690                         rte_errno = ENOENT;
1691                         ret = -rte_errno;
1692                         goto exit;
1693                 }
1694         }
1695         MLX5_ASSERT(ns);
1696         /*
1697          * Sort list to probe devices in natural order for users convenience
1698          * (i.e. master first, then representors from lowest to highest ID).
1699          */
1700         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1701         /* Default configuration. */
1702         dev_config = (struct mlx5_dev_config){
1703                 .hw_padding = 0,
1704                 .mps = MLX5_ARG_UNSET,
1705                 .dbnc = MLX5_ARG_UNSET,
1706                 .rx_vec_en = 1,
1707                 .txq_inline_max = MLX5_ARG_UNSET,
1708                 .txq_inline_min = MLX5_ARG_UNSET,
1709                 .txq_inline_mpw = MLX5_ARG_UNSET,
1710                 .txqs_inline = MLX5_ARG_UNSET,
1711                 .vf_nl_en = 1,
1712                 .mr_ext_memseg_en = 1,
1713                 .mprq = {
1714                         .enabled = 0, /* Disabled by default. */
1715                         .stride_num_n = 0,
1716                         .stride_size_n = 0,
1717                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1718                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1719                 },
1720                 .dv_esw_en = 1,
1721                 .dv_flow_en = 1,
1722                 .log_hp_size = MLX5_ARG_UNSET,
1723         };
1724         /* Device specific configuration. */
1725         switch (pci_dev->id.device_id) {
1726         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1727         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1728         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1729         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1730         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1731         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1732         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1733                 dev_config.vf = 1;
1734                 break;
1735         default:
1736                 break;
1737         }
1738         for (i = 0; i != ns; ++i) {
1739                 uint32_t restore;
1740
1741                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1742                                                  &list[i],
1743                                                  dev_config);
1744                 if (!list[i].eth_dev) {
1745                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1746                                 break;
1747                         /* Device is disabled or already spawned. Ignore it. */
1748                         continue;
1749                 }
1750                 restore = list[i].eth_dev->data->dev_flags;
1751                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1752                 /* Restore non-PCI flags cleared by the above call. */
1753                 list[i].eth_dev->data->dev_flags |= restore;
1754                 rte_eth_dev_probing_finish(list[i].eth_dev);
1755         }
1756         if (i != ns) {
1757                 DRV_LOG(ERR,
1758                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1759                         " encountering an error: %s",
1760                         pci_dev->addr.domain, pci_dev->addr.bus,
1761                         pci_dev->addr.devid, pci_dev->addr.function,
1762                         strerror(rte_errno));
1763                 ret = -rte_errno;
1764                 /* Roll back. */
1765                 while (i--) {
1766                         if (!list[i].eth_dev)
1767                                 continue;
1768                         mlx5_dev_close(list[i].eth_dev);
1769                         /* mac_addrs must not be freed because in dev_private */
1770                         list[i].eth_dev->data->mac_addrs = NULL;
1771                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1772                 }
1773                 /* Restore original error. */
1774                 rte_errno = -ret;
1775         } else {
1776                 ret = 0;
1777         }
1778 exit:
1779         /*
1780          * Do the routine cleanup:
1781          * - close opened Netlink sockets
1782          * - free allocated spawn data array
1783          * - free the Infiniband device list
1784          */
1785         if (nl_rdma >= 0)
1786                 close(nl_rdma);
1787         if (nl_route >= 0)
1788                 close(nl_route);
1789         if (list)
1790                 rte_free(list);
1791         MLX5_ASSERT(ibv_list);
1792         mlx5_glue->free_device_list(ibv_list);
1793         return ret;
1794 }
1795
1796 static int
1797 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1798 {
1799         char *env;
1800         int value;
1801
1802         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1803         /* Get environment variable to store. */
1804         env = getenv(MLX5_SHUT_UP_BF);
1805         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1806         if (config->dbnc == MLX5_ARG_UNSET)
1807                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1808         else
1809                 setenv(MLX5_SHUT_UP_BF,
1810                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1811         return value;
1812 }
1813
1814 static void
1815 mlx5_restore_doorbell_mapping_env(int value)
1816 {
1817         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1818         /* Restore the original environment variable state. */
1819         if (value == MLX5_ARG_UNSET)
1820                 unsetenv(MLX5_SHUT_UP_BF);
1821         else
1822                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1823 }
1824
1825 /**
1826  * Extract pdn of PD object using DV API.
1827  *
1828  * @param[in] pd
1829  *   Pointer to the verbs PD object.
1830  * @param[out] pdn
1831  *   Pointer to the PD object number variable.
1832  *
1833  * @return
1834  *   0 on success, error value otherwise.
1835  */
1836 int
1837 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1838 {
1839 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1840         struct mlx5dv_obj obj;
1841         struct mlx5dv_pd pd_info;
1842         int ret = 0;
1843
1844         obj.pd.in = pd;
1845         obj.pd.out = &pd_info;
1846         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1847         if (ret) {
1848                 DRV_LOG(DEBUG, "Fail to get PD object info");
1849                 return ret;
1850         }
1851         *pdn = pd_info.pdn;
1852         return 0;
1853 #else
1854         (void)pd;
1855         (void)pdn;
1856         return -ENOTSUP;
1857 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1858 }
1859
1860 /**
1861  * Function API to open IB device.
1862  *
1863  * This function calls the Linux glue APIs to open a device.
1864  *
1865  * @param[in] spawn
1866  *   Pointer to the IB device attributes (name, port, etc).
1867  * @param[out] config
1868  *   Pointer to device configuration structure.
1869  * @param[out] sh
1870  *   Pointer to shared context structure.
1871  *
1872  * @return
1873  *   0 on success, a positive error value otherwise.
1874  */
1875 int
1876 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1877                      const struct mlx5_dev_config *config,
1878                      struct mlx5_dev_ctx_shared *sh)
1879 {
1880         int dbmap_env;
1881         int err = 0;
1882         /*
1883          * Configure environment variable "MLX5_BF_SHUT_UP"
1884          * before the device creation. The rdma_core library
1885          * checks the variable at device creation and
1886          * stores the result internally.
1887          */
1888         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1889         /* Try to open IB device with DV first, then usual Verbs. */
1890         errno = 0;
1891         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1892         if (sh->ctx) {
1893                 sh->devx = 1;
1894                 DRV_LOG(DEBUG, "DevX is supported");
1895                 /* The device is created, no need for environment. */
1896                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1897         } else {
1898                 /* The environment variable is still configured. */
1899                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1900                 err = errno ? errno : ENODEV;
1901                 /*
1902                  * The environment variable is not needed anymore,
1903                  * all device creation attempts are completed.
1904                  */
1905                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1906                 if (!sh->ctx)
1907                         return err;
1908                 DRV_LOG(DEBUG, "DevX is NOT supported");
1909                 err = 0;
1910         }
1911         return err;
1912 }
1913
1914 /**
1915  * Install shared asynchronous device events handler.
1916  * This function is implemented to support event sharing
1917  * between multiple ports of single IB device.
1918  *
1919  * @param sh
1920  *   Pointer to mlx5_dev_ctx_shared object.
1921  */
1922 void
1923 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1924 {
1925         int ret;
1926         int flags;
1927
1928         sh->intr_handle.fd = -1;
1929         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1930         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1931                     F_SETFL, flags | O_NONBLOCK);
1932         if (ret) {
1933                 DRV_LOG(INFO, "failed to change file descriptor async event"
1934                         " queue");
1935         } else {
1936                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1937                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1938                 if (rte_intr_callback_register(&sh->intr_handle,
1939                                         mlx5_dev_interrupt_handler, sh)) {
1940                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1941                         sh->intr_handle.fd = -1;
1942                 }
1943         }
1944         if (sh->devx) {
1945 #ifdef HAVE_IBV_DEVX_ASYNC
1946                 sh->intr_handle_devx.fd = -1;
1947                 sh->devx_comp = mlx5_glue->devx_create_cmd_comp(sh->ctx);
1948                 if (!sh->devx_comp) {
1949                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1950                         return;
1951                 }
1952                 flags = fcntl(sh->devx_comp->fd, F_GETFL);
1953                 ret = fcntl(sh->devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1954                 if (ret) {
1955                         DRV_LOG(INFO, "failed to change file descriptor"
1956                                 " devx comp");
1957                         return;
1958                 }
1959                 sh->intr_handle_devx.fd = sh->devx_comp->fd;
1960                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1961                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1962                                         mlx5_dev_interrupt_handler_devx, sh)) {
1963                         DRV_LOG(INFO, "Fail to install the devx shared"
1964                                 " interrupt.");
1965                         sh->intr_handle_devx.fd = -1;
1966                 }
1967 #endif /* HAVE_IBV_DEVX_ASYNC */
1968         }
1969 }
1970
1971 /**
1972  * Uninstall shared asynchronous device events handler.
1973  * This function is implemented to support event sharing
1974  * between multiple ports of single IB device.
1975  *
1976  * @param dev
1977  *   Pointer to mlx5_dev_ctx_shared object.
1978  */
1979 void
1980 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
1981 {
1982         if (sh->intr_handle.fd >= 0)
1983                 mlx5_intr_callback_unregister(&sh->intr_handle,
1984                                               mlx5_dev_interrupt_handler, sh);
1985 #ifdef HAVE_IBV_DEVX_ASYNC
1986         if (sh->intr_handle_devx.fd >= 0)
1987                 rte_intr_callback_unregister(&sh->intr_handle_devx,
1988                                   mlx5_dev_interrupt_handler_devx, sh);
1989         if (sh->devx_comp)
1990                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1991 #endif
1992 }