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 <linux/sockios.h>
  16 #include <linux/ethtool.h>
  17 #include <fcntl.h>
  18
  19 /* Verbs header. */
  20 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  21 #ifdef PEDANTIC
  22 #pragma GCC diagnostic ignored "-Wpedantic"
  23 #endif
  24 #include <infiniband/verbs.h>
  25 #ifdef PEDANTIC
  26 #pragma GCC diagnostic error "-Wpedantic"
  27 #endif
  28
  29 #include <rte_malloc.h>
  30 #include <rte_ethdev_driver.h>
  31 #include <rte_ethdev_pci.h>
  32 #include <rte_pci.h>
  33 #include <rte_bus_pci.h>
  34 #include <rte_common.h>
  35 #include <rte_kvargs.h>
  36 #include <rte_rwlock.h>
  37 #include <rte_spinlock.h>
  38 #include <rte_string_fns.h>
  39 #include <rte_alarm.h>
  40
  41 #include <mlx5_glue.h>
  42 #include <mlx5_devx_cmds.h>
  43 #include <mlx5_common.h>
  44 #include <mlx5_common_mp.h>
  45 #include <mlx5_common_mr.h>
  46
  47 #include "mlx5_defs.h"
  48 #include "mlx5.h"
  49 #include "mlx5_common_os.h"
  50 #include "mlx5_utils.h"
  51 #include "mlx5_rxtx.h"
  52 #include "mlx5_autoconf.h"
  53 #include "mlx5_mr.h"
  54 #include "mlx5_flow.h"
  55 #include "rte_pmd_mlx5.h"
  56 #include "mlx5_verbs.h"
  57
  58 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  59
  60 #ifndef HAVE_IBV_MLX5_MOD_MPW
  61 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  62 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  63 #endif
  64
  65 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  66 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  67 #endif
  68
  69 /**
  70  * Get mlx5 device attributes. The glue function query_device_ex() is called
  71  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  72  * device attributes from the glue out parameter.
  73  *
  74  * @param dev
  75  *   Pointer to ibv context.
  76  *
  77  * @param device_attr
  78  *   Pointer to mlx5 device attributes.
  79  *
  80  * @return
  81  *   0 on success, non zero error number otherwise
  82  */
  83 int
  84 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
  85 {
  86         int err;
  87         struct ibv_device_attr_ex attr_ex;
  88         memset(device_attr, 0, sizeof(*device_attr));
  89         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
  90         if (err)
  91                 return err;
  92
  93         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
  94         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
  95         device_attr->max_sge = attr_ex.orig_attr.max_sge;
  96         device_attr->max_cq = attr_ex.orig_attr.max_cq;
  97         device_attr->max_qp = attr_ex.orig_attr.max_qp;
  98         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
  99         device_attr->max_rwq_indirection_table_size =
 100                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 101         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 102         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 103
 104         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 105         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 106         if (err)
 107                 return err;
 108
 109         device_attr->flags = dv_attr.flags;
 110         device_attr->comp_mask = dv_attr.comp_mask;
 111 #ifdef HAVE_IBV_MLX5_MOD_SWP
 112         device_attr->sw_parsing_offloads =
 113                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 114 #endif
 115         device_attr->min_single_stride_log_num_of_bytes =
 116                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 117         device_attr->max_single_stride_log_num_of_bytes =
 118                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 119         device_attr->min_single_wqe_log_num_of_strides =
 120                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 121         device_attr->max_single_wqe_log_num_of_strides =
 122                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 123         device_attr->stride_supported_qpts =
 124                 dv_attr.striding_rq_caps.supported_qpts;
 125 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 126         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 127 #endif
 128
 129         return err;
 130 }
 131
 132 /**
 133  * Verbs callback to allocate a memory. This function should allocate the space
 134  * according to the size provided residing inside a huge page.
 135  * Please note that all allocation must respect the alignment from libmlx5
 136  * (i.e. currently sysconf(_SC_PAGESIZE)).
 137  *
 138  * @param[in] size
 139  *   The size in bytes of the memory to allocate.
 140  * @param[in] data
 141  *   A pointer to the callback data.
 142  *
 143  * @return
 144  *   Allocated buffer, NULL otherwise and rte_errno is set.
 145  */
 146 static void *
 147 mlx5_alloc_verbs_buf(size_t size, void *data)
 148 {
 149         struct mlx5_priv *priv = data;
 150         void *ret;
 151         size_t alignment = sysconf(_SC_PAGESIZE);
 152         unsigned int socket = SOCKET_ID_ANY;
 153
 154         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 155                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 156
 157                 socket = ctrl->socket;
 158         } else if (priv->verbs_alloc_ctx.type ==
 159                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 160                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 161
 162                 socket = ctrl->socket;
 163         }
 164         MLX5_ASSERT(data != NULL);
 165         ret = rte_malloc_socket(__func__, size, alignment, socket);
 166         if (!ret && size)
 167                 rte_errno = ENOMEM;
 168         return ret;
 169 }
 170
 171 /**
 172  * Verbs callback to free a memory.
 173  *
 174  * @param[in] ptr
 175  *   A pointer to the memory to free.
 176  * @param[in] data
 177  *   A pointer to the callback data.
 178  */
 179 static void
 180 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 181 {
 182         MLX5_ASSERT(data != NULL);
 183         rte_free(ptr);
 184 }
 185
 186 /**
 187  * Initialize DR related data within private structure.
 188  * Routine checks the reference counter and does actual
 189  * resources creation/initialization only if counter is zero.
 190  *
 191  * @param[in] priv
 192  *   Pointer to the private device data structure.
 193  *
 194  * @return
 195  *   Zero on success, positive error code otherwise.
 196  */
 197 static int
 198 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 199 {
 200         struct mlx5_dev_ctx_shared *sh = priv->sh;
 201         char s[MLX5_HLIST_NAMESIZE];
 202         int err = 0;
 203
 204         if (!sh->flow_tbls)
 205                 err = mlx5_alloc_table_hash_list(priv);
 206         else
 207                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 208                         (void *)sh->flow_tbls);
 209         if (err)
 210                 return err;
 211         /* Create tags hash list table. */
 212         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 213         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 214         if (!sh->tag_table) {
 215                 DRV_LOG(ERR, "tags with hash creation failed.");
 216                 err = ENOMEM;
 217                 goto error;
 218         }
 219 #ifdef HAVE_MLX5DV_DR
 220         void *domain;
 221
 222         if (sh->dv_refcnt) {
 223                 /* Shared DV/DR structures is already initialized. */
 224                 sh->dv_refcnt++;
 225                 priv->dr_shared = 1;
 226                 return 0;
 227         }
 228         /* Reference counter is zero, we should initialize structures. */
 229         domain = mlx5_glue->dr_create_domain(sh->ctx,
 230                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 231         if (!domain) {
 232                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 233                 err = errno;
 234                 goto error;
 235         }
 236         sh->rx_domain = domain;
 237         domain = mlx5_glue->dr_create_domain(sh->ctx,
 238                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 239         if (!domain) {
 240                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 241                 err = errno;
 242                 goto error;
 243         }
 244         pthread_mutex_init(&sh->dv_mutex, NULL);
 245         sh->tx_domain = domain;
 246 #ifdef HAVE_MLX5DV_DR_ESWITCH
 247         if (priv->config.dv_esw_en) {
 248                 domain  = mlx5_glue->dr_create_domain
 249                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 250                 if (!domain) {
 251                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 252                         err = errno;
 253                         goto error;
 254                 }
 255                 sh->fdb_domain = domain;
 256                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 257         }
 258 #endif
 259         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 260                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 261                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 262                 if (sh->fdb_domain)
 263                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 264         }
 265         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 266 #endif /* HAVE_MLX5DV_DR */
 267         sh->dv_refcnt++;
 268         priv->dr_shared = 1;
 269         return 0;
 270 error:
 271         /* Rollback the created objects. */
 272         if (sh->rx_domain) {
 273                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 274                 sh->rx_domain = NULL;
 275         }
 276         if (sh->tx_domain) {
 277                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 278                 sh->tx_domain = NULL;
 279         }
 280         if (sh->fdb_domain) {
 281                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 282                 sh->fdb_domain = NULL;
 283         }
 284         if (sh->esw_drop_action) {
 285                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 286                 sh->esw_drop_action = NULL;
 287         }
 288         if (sh->pop_vlan_action) {
 289                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 290                 sh->pop_vlan_action = NULL;
 291         }
 292         if (sh->tag_table) {
 293                 /* tags should be destroyed with flow before. */
 294                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 295                 sh->tag_table = NULL;
 296         }
 297         mlx5_free_table_hash_list(priv);
 298         return err;
 299 }
 300
 301 /**
 302  * Destroy DR related data within private structure.
 303  *
 304  * @param[in] priv
 305  *   Pointer to the private device data structure.
 306  */
 307 void
 308 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 309 {
 310         struct mlx5_dev_ctx_shared *sh;
 311
 312         if (!priv->dr_shared)
 313                 return;
 314         priv->dr_shared = 0;
 315         sh = priv->sh;
 316         MLX5_ASSERT(sh);
 317 #ifdef HAVE_MLX5DV_DR
 318         MLX5_ASSERT(sh->dv_refcnt);
 319         if (sh->dv_refcnt && --sh->dv_refcnt)
 320                 return;
 321         if (sh->rx_domain) {
 322                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 323                 sh->rx_domain = NULL;
 324         }
 325         if (sh->tx_domain) {
 326                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 327                 sh->tx_domain = NULL;
 328         }
 329 #ifdef HAVE_MLX5DV_DR_ESWITCH
 330         if (sh->fdb_domain) {
 331                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 332                 sh->fdb_domain = NULL;
 333         }
 334         if (sh->esw_drop_action) {
 335                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 336                 sh->esw_drop_action = NULL;
 337         }
 338 #endif
 339         if (sh->pop_vlan_action) {
 340                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 341                 sh->pop_vlan_action = NULL;
 342         }
 343         pthread_mutex_destroy(&sh->dv_mutex);
 344 #endif /* HAVE_MLX5DV_DR */
 345         if (sh->tag_table) {
 346                 /* tags should be destroyed with flow before. */
 347                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 348                 sh->tag_table = NULL;
 349         }
 350         mlx5_free_table_hash_list(priv);
 351 }
 352
 353 /**
 354  * Spawn an Ethernet device from Verbs information.
 355  *
 356  * @param dpdk_dev
 357  *   Backing DPDK device.
 358  * @param spawn
 359  *   Verbs device parameters (name, port, switch_info) to spawn.
 360  * @param config
 361  *   Device configuration parameters.
 362  *
 363  * @return
 364  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 365  *   is set. The following errors are defined:
 366  *
 367  *   EBUSY: device is not supposed to be spawned.
 368  *   EEXIST: device is already spawned
 369  */
 370 static struct rte_eth_dev *
 371 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 372                struct mlx5_dev_spawn_data *spawn,
 373                struct mlx5_dev_config config)
 374 {
 375         const struct mlx5_switch_info *switch_info = &spawn->info;
 376         struct mlx5_dev_ctx_shared *sh = NULL;
 377         struct ibv_port_attr port_attr;
 378         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 379         struct rte_eth_dev *eth_dev = NULL;
 380         struct mlx5_priv *priv = NULL;
 381         int err = 0;
 382         unsigned int hw_padding = 0;
 383         unsigned int mps;
 384         unsigned int cqe_comp;
 385         unsigned int cqe_pad = 0;
 386         unsigned int tunnel_en = 0;
 387         unsigned int mpls_en = 0;
 388         unsigned int swp = 0;
 389         unsigned int mprq = 0;
 390         unsigned int mprq_min_stride_size_n = 0;
 391         unsigned int mprq_max_stride_size_n = 0;
 392         unsigned int mprq_min_stride_num_n = 0;
 393         unsigned int mprq_max_stride_num_n = 0;
 394         struct rte_ether_addr mac;
 395         char name[RTE_ETH_NAME_MAX_LEN];
 396         int own_domain_id = 0;
 397         uint16_t port_id;
 398         unsigned int i;
 399 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 400         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 401 #endif
 402
 403         /* Determine if this port representor is supposed to be spawned. */
 404         if (switch_info->representor && dpdk_dev->devargs) {
 405                 struct rte_eth_devargs eth_da;
 406
 407                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 408                 if (err) {
 409                         rte_errno = -err;
 410                         DRV_LOG(ERR, "failed to process device arguments: %s",
 411                                 strerror(rte_errno));
 412                         return NULL;
 413                 }
 414                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 415                         if (eth_da.representor_ports[i] ==
 416                             (uint16_t)switch_info->port_name)
 417                                 break;
 418                 if (i == eth_da.nb_representor_ports) {
 419                         rte_errno = EBUSY;
 420                         return NULL;
 421                 }
 422         }
 423         /* Build device name. */
 424         if (spawn->pf_bond <  0) {
 425                 /* Single device. */
 426                 if (!switch_info->representor)
 427                         strlcpy(name, dpdk_dev->name, sizeof(name));
 428                 else
 429                         snprintf(name, sizeof(name), "%s_representor_%u",
 430                                  dpdk_dev->name, switch_info->port_name);
 431         } else {
 432                 /* Bonding device. */
 433                 if (!switch_info->representor)
 434                         snprintf(name, sizeof(name), "%s_%s",
 435                                  dpdk_dev->name,
 436                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 437                 else
 438                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 439                                  dpdk_dev->name,
 440                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 441                                  switch_info->port_name);
 442         }
 443         /* check if the device is already spawned */
 444         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 445                 rte_errno = EEXIST;
 446                 return NULL;
 447         }
 448         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 449         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 450                 struct mlx5_mp_id mp_id;
 451
 452                 eth_dev = rte_eth_dev_attach_secondary(name);
 453                 if (eth_dev == NULL) {
 454                         DRV_LOG(ERR, "can not attach rte ethdev");
 455                         rte_errno = ENOMEM;
 456                         return NULL;
 457                 }
 458                 eth_dev->device = dpdk_dev;
 459                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 460                 err = mlx5_proc_priv_init(eth_dev);
 461                 if (err)
 462                         return NULL;
 463                 mp_id.port_id = eth_dev->data->port_id;
 464                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 465                 /* Receive command fd from primary process */
 466                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 467                 if (err < 0)
 468                         goto err_secondary;
 469                 /* Remap UAR for Tx queues. */
 470                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 471                 if (err)
 472                         goto err_secondary;
 473                 /*
 474                  * Ethdev pointer is still required as input since
 475                  * the primary device is not accessible from the
 476                  * secondary process.
 477                  */
 478                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 479                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 480                 return eth_dev;
 481 err_secondary:
 482                 mlx5_dev_close(eth_dev);
 483                 return NULL;
 484         }
 485         /*
 486          * Some parameters ("tx_db_nc" in particularly) are needed in
 487          * advance to create dv/verbs device context. We proceed the
 488          * devargs here to get ones, and later proceed devargs again
 489          * to override some hardware settings.
 490          */
 491         err = mlx5_args(&config, dpdk_dev->devargs);
 492         if (err) {
 493                 err = rte_errno;
 494                 DRV_LOG(ERR, "failed to process device arguments: %s",
 495                         strerror(rte_errno));
 496                 goto error;
 497         }
 498         sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
 499         if (!sh)
 500                 return NULL;
 501         config.devx = sh->devx;
 502 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 503         config.dest_tir = 1;
 504 #endif
 505 #ifdef HAVE_IBV_MLX5_MOD_SWP
 506         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 507 #endif
 508         /*
 509          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 510          * as all ConnectX-5 devices.
 511          */
 512 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 513         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 514 #endif
 515 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 516         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 517 #endif
 518         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 519         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 520                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 521                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 522                         mps = MLX5_MPW_ENHANCED;
 523                 } else {
 524                         DRV_LOG(DEBUG, "MPW is supported");
 525                         mps = MLX5_MPW;
 526                 }
 527         } else {
 528                 DRV_LOG(DEBUG, "MPW isn't supported");
 529                 mps = MLX5_MPW_DISABLED;
 530         }
 531 #ifdef HAVE_IBV_MLX5_MOD_SWP
 532         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 533                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 534         DRV_LOG(DEBUG, "SWP support: %u", swp);
 535 #endif
 536         config.swp = !!swp;
 537 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 538         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 539                 struct mlx5dv_striding_rq_caps mprq_caps =
 540                         dv_attr.striding_rq_caps;
 541
 542                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 543                         mprq_caps.min_single_stride_log_num_of_bytes);
 544                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 545                         mprq_caps.max_single_stride_log_num_of_bytes);
 546                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 547                         mprq_caps.min_single_wqe_log_num_of_strides);
 548                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 549                         mprq_caps.max_single_wqe_log_num_of_strides);
 550                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 551                         mprq_caps.supported_qpts);
 552                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 553                 mprq = 1;
 554                 mprq_min_stride_size_n =
 555                         mprq_caps.min_single_stride_log_num_of_bytes;
 556                 mprq_max_stride_size_n =
 557                         mprq_caps.max_single_stride_log_num_of_bytes;
 558                 mprq_min_stride_num_n =
 559                         mprq_caps.min_single_wqe_log_num_of_strides;
 560                 mprq_max_stride_num_n =
 561                         mprq_caps.max_single_wqe_log_num_of_strides;
 562         }
 563 #endif
 564         if (RTE_CACHE_LINE_SIZE == 128 &&
 565             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 566                 cqe_comp = 0;
 567         else
 568                 cqe_comp = 1;
 569         config.cqe_comp = cqe_comp;
 570 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 571         /* Whether device supports 128B Rx CQE padding. */
 572         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 573                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 574 #endif
 575 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 576         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 577                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 578                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 579                              (dv_attr.tunnel_offloads_caps &
 580                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 581                              (dv_attr.tunnel_offloads_caps &
 582                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 583         }
 584         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 585                 tunnel_en ? "" : "not ");
 586 #else
 587         DRV_LOG(WARNING,
 588                 "tunnel offloading disabled due to old OFED/rdma-core version");
 589 #endif
 590         config.tunnel_en = tunnel_en;
 591 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 592         mpls_en = ((dv_attr.tunnel_offloads_caps &
 593                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 594                    (dv_attr.tunnel_offloads_caps &
 595                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 596         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 597                 mpls_en ? "" : "not ");
 598 #else
 599         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 600                 " old OFED/rdma-core version or firmware configuration");
 601 #endif
 602         config.mpls_en = mpls_en;
 603         /* Check port status. */
 604         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 605         if (err) {
 606                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 607                 goto error;
 608         }
 609         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 610                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 611                 err = EINVAL;
 612                 goto error;
 613         }
 614         if (port_attr.state != IBV_PORT_ACTIVE)
 615                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 616                         mlx5_glue->port_state_str(port_attr.state),
 617                         port_attr.state);
 618         /* Allocate private eth device data. */
 619         priv = rte_zmalloc("ethdev private structure",
 620                            sizeof(*priv),
 621                            RTE_CACHE_LINE_SIZE);
 622         if (priv == NULL) {
 623                 DRV_LOG(ERR, "priv allocation failure");
 624                 err = ENOMEM;
 625                 goto error;
 626         }
 627         priv->sh = sh;
 628         priv->dev_port = spawn->phys_port;
 629         priv->pci_dev = spawn->pci_dev;
 630         priv->mtu = RTE_ETHER_MTU;
 631         priv->mp_id.port_id = port_id;
 632         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 633         /* Some internal functions rely on Netlink sockets, open them now. */
 634         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 635         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 636         priv->representor = !!switch_info->representor;
 637         priv->master = !!switch_info->master;
 638         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 639         priv->vport_meta_tag = 0;
 640         priv->vport_meta_mask = 0;
 641         priv->pf_bond = spawn->pf_bond;
 642 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 643         /*
 644          * The DevX port query API is implemented. E-Switch may use
 645          * either vport or reg_c[0] metadata register to match on
 646          * vport index. The engaged part of metadata register is
 647          * defined by mask.
 648          */
 649         if (switch_info->representor || switch_info->master) {
 650                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 651                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 652                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 653                                                  &devx_port);
 654                 if (err) {
 655                         DRV_LOG(WARNING,
 656                                 "can't query devx port %d on device %s",
 657                                 spawn->phys_port,
 658                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 659                         devx_port.comp_mask = 0;
 660                 }
 661         }
 662         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 663                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 664                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 665                 if (!priv->vport_meta_mask) {
 666                         DRV_LOG(ERR, "vport zero mask for port %d"
 667                                      " on bonding device %s",
 668                                      spawn->phys_port,
 669                                      mlx5_os_get_dev_device_name
 670                                                         (spawn->phys_dev));
 671                         err = ENOTSUP;
 672                         goto error;
 673                 }
 674                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 675                         DRV_LOG(ERR, "invalid vport tag for port %d"
 676                                      " on bonding device %s",
 677                                      spawn->phys_port,
 678                                      mlx5_os_get_dev_device_name
 679                                                         (spawn->phys_dev));
 680                         err = ENOTSUP;
 681                         goto error;
 682                 }
 683         }
 684         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 685                 priv->vport_id = devx_port.vport_num;
 686         } else if (spawn->pf_bond >= 0) {
 687                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 688                              " on bonding device %s",
 689                              spawn->phys_port,
 690                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 691                 err = ENOTSUP;
 692                 goto error;
 693         } else {
 694                 /* Suppose vport index in compatible way. */
 695                 priv->vport_id = switch_info->representor ?
 696                                  switch_info->port_name + 1 : -1;
 697         }
 698 #else
 699         /*
 700          * Kernel/rdma_core support single E-Switch per PF configurations
 701          * only and vport_id field contains the vport index for
 702          * associated VF, which is deduced from representor port name.
 703          * For example, let's have the IB device port 10, it has
 704          * attached network device eth0, which has port name attribute
 705          * pf0vf2, we can deduce the VF number as 2, and set vport index
 706          * as 3 (2+1). This assigning schema should be changed if the
 707          * multiple E-Switch instances per PF configurations or/and PCI
 708          * subfunctions are added.
 709          */
 710         priv->vport_id = switch_info->representor ?
 711                          switch_info->port_name + 1 : -1;
 712 #endif
 713         /* representor_id field keeps the unmodified VF index. */
 714         priv->representor_id = switch_info->representor ?
 715                                switch_info->port_name : -1;
 716         /*
 717          * Look for sibling devices in order to reuse their switch domain
 718          * if any, otherwise allocate one.
 719          */
 720         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 721                 const struct mlx5_priv *opriv =
 722                         rte_eth_devices[port_id].data->dev_private;
 723
 724                 if (!opriv ||
 725                     opriv->sh != priv->sh ||
 726                         opriv->domain_id ==
 727                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 728                         continue;
 729                 priv->domain_id = opriv->domain_id;
 730                 break;
 731         }
 732         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 733                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 734                 if (err) {
 735                         err = rte_errno;
 736                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 737                                 strerror(rte_errno));
 738                         goto error;
 739                 }
 740                 own_domain_id = 1;
 741         }
 742         /* Override some values set by hardware configuration. */
 743         mlx5_args(&config, dpdk_dev->devargs);
 744         err = mlx5_dev_check_sibling_config(priv, &config);
 745         if (err)
 746                 goto error;
 747         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 748                             IBV_DEVICE_RAW_IP_CSUM);
 749         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 750                 (config.hw_csum ? "" : "not "));
 751 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 752         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 753         DRV_LOG(DEBUG, "counters are not supported");
 754 #endif
 755 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 756         if (config.dv_flow_en) {
 757                 DRV_LOG(WARNING, "DV flow is not supported");
 758                 config.dv_flow_en = 0;
 759         }
 760 #endif
 761         config.ind_table_max_size =
 762                 sh->device_attr.max_rwq_indirection_table_size;
 763         /*
 764          * Remove this check once DPDK supports larger/variable
 765          * indirection tables.
 766          */
 767         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 768                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 769         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 770                 config.ind_table_max_size);
 771         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 772                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 773         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 774                 (config.hw_vlan_strip ? "" : "not "));
 775         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 776                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 777         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 778                 (config.hw_fcs_strip ? "" : "not "));
 779 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 780         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 781 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 782         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 783                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 784 #endif
 785         if (config.hw_padding && !hw_padding) {
 786                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 787                 config.hw_padding = 0;
 788         } else if (config.hw_padding) {
 789                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 790         }
 791         config.tso = (sh->device_attr.max_tso > 0 &&
 792                       (sh->device_attr.tso_supported_qpts &
 793                        (1 << IBV_QPT_RAW_PACKET)));
 794         if (config.tso)
 795                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 796         /*
 797          * MPW is disabled by default, while the Enhanced MPW is enabled
 798          * by default.
 799          */
 800         if (config.mps == MLX5_ARG_UNSET)
 801                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 802                                                           MLX5_MPW_DISABLED;
 803         else
 804                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 805         DRV_LOG(INFO, "%sMPS is %s",
 806                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 807                 config.mps == MLX5_MPW ? "legacy " : "",
 808                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 809         if (config.cqe_comp && !cqe_comp) {
 810                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 811                 config.cqe_comp = 0;
 812         }
 813         if (config.cqe_pad && !cqe_pad) {
 814                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 815                 config.cqe_pad = 0;
 816         } else if (config.cqe_pad) {
 817                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 818         }
 819         if (config.devx) {
 820                 priv->counter_fallback = 0;
 821                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 822                 if (err) {
 823                         err = -err;
 824                         goto error;
 825                 }
 826                 if (!config.hca_attr.flow_counters_dump)
 827                         priv->counter_fallback = 1;
 828 #ifndef HAVE_IBV_DEVX_ASYNC
 829                 priv->counter_fallback = 1;
 830 #endif
 831                 if (priv->counter_fallback)
 832                         DRV_LOG(INFO, "Use fall-back DV counter management");
 833                 /* Check for LRO support. */
 834                 if (config.dest_tir && config.hca_attr.lro_cap &&
 835                     config.dv_flow_en) {
 836                         /* TBD check tunnel lro caps. */
 837                         config.lro.supported = config.hca_attr.lro_cap;
 838                         DRV_LOG(DEBUG, "Device supports LRO");
 839                         /*
 840                          * If LRO timeout is not configured by application,
 841                          * use the minimal supported value.
 842                          */
 843                         if (!config.lro.timeout)
 844                                 config.lro.timeout =
 845                                 config.hca_attr.lro_timer_supported_periods[0];
 846                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 847                                 config.lro.timeout);
 848                 }
 849 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 850                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 851                     config.dv_flow_en) {
 852                         uint8_t reg_c_mask =
 853                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 854                         /*
 855                          * Meter needs two REG_C's for color match and pre-sfx
 856                          * flow match. Here get the REG_C for color match.
 857                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 858                          */
 859                         reg_c_mask &= 0xfc;
 860                         if (__builtin_popcount(reg_c_mask) < 1) {
 861                                 priv->mtr_en = 0;
 862                                 DRV_LOG(WARNING, "No available register for"
 863                                         " meter.");
 864                         } else {
 865                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 866                                                       REG_C_0;
 867                                 priv->mtr_en = 1;
 868                                 priv->mtr_reg_share =
 869                                       config.hca_attr.qos.flow_meter_reg_share;
 870                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 871                                         priv->mtr_color_reg);
 872                         }
 873                 }
 874 #endif
 875         }
 876         if (config.tx_pp) {
 877                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
 878                         config.hca_attr.dev_freq_khz);
 879                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
 880                         config.hca_attr.qos.packet_pacing ? "" : "not ");
 881                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
 882                         config.hca_attr.cross_channel ? "" : "not ");
 883                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
 884                         config.hca_attr.wqe_index_ignore ? "" : "not ");
 885                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
 886                         config.hca_attr.non_wire_sq ? "" : "not ");
 887                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
 888                         config.hca_attr.log_max_static_sq_wq ? "" : "not ",
 889                         config.hca_attr.log_max_static_sq_wq);
 890                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
 891                         config.hca_attr.qos.wqe_rate_pp ? "" : "not ");
 892                 if (!config.devx) {
 893                         DRV_LOG(ERR, "DevX is required for packet pacing");
 894                         err = ENODEV;
 895                         goto error;
 896                 }
 897                 if (!config.hca_attr.qos.packet_pacing) {
 898                         DRV_LOG(ERR, "Packet pacing is not supported");
 899                         err = ENODEV;
 900                         goto error;
 901                 }
 902                 if (!config.hca_attr.cross_channel) {
 903                         DRV_LOG(ERR, "Cross channel operations are"
 904                                      " required for packet pacing");
 905                         err = ENODEV;
 906                         goto error;
 907                 }
 908                 if (!config.hca_attr.wqe_index_ignore) {
 909                         DRV_LOG(ERR, "WQE index ignore feature is"
 910                                      " required for packet pacing");
 911                         err = ENODEV;
 912                         goto error;
 913                 }
 914                 if (!config.hca_attr.non_wire_sq) {
 915                         DRV_LOG(ERR, "Non-wire SQ feature is"
 916                                      " required for packet pacing");
 917                         err = ENODEV;
 918                         goto error;
 919                 }
 920                 if (!config.hca_attr.log_max_static_sq_wq) {
 921                         DRV_LOG(ERR, "Static WQE SQ feature is"
 922                                      " required for packet pacing");
 923                         err = ENODEV;
 924                         goto error;
 925                 }
 926                 if (!config.hca_attr.qos.wqe_rate_pp) {
 927                         DRV_LOG(ERR, "WQE rate mode is required"
 928                                      " for packet pacing");
 929                         err = ENODEV;
 930                         goto error;
 931                 }
 932 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
 933                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
 934                              " can't create queues for packet pacing");
 935                 err = ENODEV;
 936                 goto error;
 937 #endif
 938         }
 939         if (config.devx) {
 940                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
 941
 942                 err = mlx5_devx_cmd_register_read
 943                         (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
 944                         reg, MLX5_ST_SZ_DW(register_mtutc));
 945                 if (!err) {
 946                         uint32_t ts_mode;
 947
 948                         /* MTUTC register is read successfully. */
 949                         ts_mode = MLX5_GET(register_mtutc, reg,
 950                                            time_stamp_mode);
 951                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
 952                                 config.rt_timestamp = 1;
 953                 } else {
 954                         /* Kernel does not support register reading. */
 955                         if (config.hca_attr.dev_freq_khz ==
 956                                                  (NS_PER_S / MS_PER_S))
 957                                 config.rt_timestamp = 1;
 958                 }
 959         }
 960         if (config.mprq.enabled && mprq) {
 961                 if (config.mprq.stride_num_n &&
 962                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 963                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 964                         config.mprq.stride_num_n =
 965                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 966                                                 mprq_min_stride_num_n),
 967                                         mprq_max_stride_num_n);
 968                         DRV_LOG(WARNING,
 969                                 "the number of strides"
 970                                 " for Multi-Packet RQ is out of range,"
 971                                 " setting default value (%u)",
 972                                 1 << config.mprq.stride_num_n);
 973                 }
 974                 if (config.mprq.stride_size_n &&
 975                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 976                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 977                         config.mprq.stride_size_n =
 978                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 979                                                 mprq_min_stride_size_n),
 980                                         mprq_max_stride_size_n);
 981                         DRV_LOG(WARNING,
 982                                 "the size of a stride"
 983                                 " for Multi-Packet RQ is out of range,"
 984                                 " setting default value (%u)",
 985                                 1 << config.mprq.stride_size_n);
 986                 }
 987                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
 988                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
 989         } else if (config.mprq.enabled && !mprq) {
 990                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
 991                 config.mprq.enabled = 0;
 992         }
 993         if (config.max_dump_files_num == 0)
 994                 config.max_dump_files_num = 128;
 995         eth_dev = rte_eth_dev_allocate(name);
 996         if (eth_dev == NULL) {
 997                 DRV_LOG(ERR, "can not allocate rte ethdev");
 998                 err = ENOMEM;
 999                 goto error;
1000         }
1001         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1002         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1003         if (priv->representor) {
1004                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1005                 eth_dev->data->representor_id = priv->representor_id;
1006         }
1007         /*
1008          * Store associated network device interface index. This index
1009          * is permanent throughout the lifetime of device. So, we may store
1010          * the ifindex here and use the cached value further.
1011          */
1012         MLX5_ASSERT(spawn->ifindex);
1013         priv->if_index = spawn->ifindex;
1014         eth_dev->data->dev_private = priv;
1015         priv->dev_data = eth_dev->data;
1016         eth_dev->data->mac_addrs = priv->mac;
1017         eth_dev->device = dpdk_dev;
1018         /* Configure the first MAC address by default. */
1019         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1020                 DRV_LOG(ERR,
1021                         "port %u cannot get MAC address, is mlx5_en"
1022                         " loaded? (errno: %s)",
1023                         eth_dev->data->port_id, strerror(rte_errno));
1024                 err = ENODEV;
1025                 goto error;
1026         }
1027         DRV_LOG(INFO,
1028                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1029                 eth_dev->data->port_id,
1030                 mac.addr_bytes[0], mac.addr_bytes[1],
1031                 mac.addr_bytes[2], mac.addr_bytes[3],
1032                 mac.addr_bytes[4], mac.addr_bytes[5]);
1033 #ifdef RTE_LIBRTE_MLX5_DEBUG
1034         {
1035                 char ifname[IF_NAMESIZE];
1036
1037                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1038                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1039                                 eth_dev->data->port_id, ifname);
1040                 else
1041                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1042                                 eth_dev->data->port_id);
1043         }
1044 #endif
1045         /* Get actual MTU if possible. */
1046         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1047         if (err) {
1048                 err = rte_errno;
1049                 goto error;
1050         }
1051         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1052                 priv->mtu);
1053         /* Initialize burst functions to prevent crashes before link-up. */
1054         eth_dev->rx_pkt_burst = removed_rx_burst;
1055         eth_dev->tx_pkt_burst = removed_tx_burst;
1056         eth_dev->dev_ops = &mlx5_os_dev_ops;
1057         /* Register MAC address. */
1058         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1059         if (config.vf && config.vf_nl_en)
1060                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1061                                       mlx5_ifindex(eth_dev),
1062                                       eth_dev->data->mac_addrs,
1063                                       MLX5_MAX_MAC_ADDRESSES);
1064         priv->flows = 0;
1065         priv->ctrl_flows = 0;
1066         TAILQ_INIT(&priv->flow_meters);
1067         TAILQ_INIT(&priv->flow_meter_profiles);
1068         /* Hint libmlx5 to use PMD allocator for data plane resources */
1069         mlx5_glue->dv_set_context_attr(sh->ctx,
1070                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1071                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1072                                 .alloc = &mlx5_alloc_verbs_buf,
1073                                 .free = &mlx5_free_verbs_buf,
1074                                 .data = priv,
1075                         }));
1076         /* Bring Ethernet device up. */
1077         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1078                 eth_dev->data->port_id);
1079         mlx5_set_link_up(eth_dev);
1080         /*
1081          * Even though the interrupt handler is not installed yet,
1082          * interrupts will still trigger on the async_fd from
1083          * Verbs context returned by ibv_open_device().
1084          */
1085         mlx5_link_update(eth_dev, 0);
1086 #ifdef HAVE_MLX5DV_DR_ESWITCH
1087         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1088               (switch_info->representor || switch_info->master)))
1089                 config.dv_esw_en = 0;
1090 #else
1091         config.dv_esw_en = 0;
1092 #endif
1093         /* Detect minimal data bytes to inline. */
1094         mlx5_set_min_inline(spawn, &config);
1095         /* Store device configuration on private structure. */
1096         priv->config = config;
1097         /* Create context for virtual machine VLAN workaround. */
1098         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1099         if (config.dv_flow_en) {
1100                 err = mlx5_alloc_shared_dr(priv);
1101                 if (err)
1102                         goto error;
1103                 /*
1104                  * RSS id is shared with meter flow id. Meter flow id can only
1105                  * use the 24 MSB of the register.
1106                  */
1107                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1108                                      MLX5_MTR_COLOR_BITS);
1109                 if (!priv->qrss_id_pool) {
1110                         DRV_LOG(ERR, "can't create flow id pool");
1111                         err = ENOMEM;
1112                         goto error;
1113                 }
1114         }
1115         /* Supported Verbs flow priority number detection. */
1116         err = mlx5_flow_discover_priorities(eth_dev);
1117         if (err < 0) {
1118                 err = -err;
1119                 goto error;
1120         }
1121         priv->config.flow_prio = err;
1122         if (!priv->config.dv_esw_en &&
1123             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1124                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1125                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1126                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1127         }
1128         mlx5_set_metadata_mask(eth_dev);
1129         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1130             !priv->sh->dv_regc0_mask) {
1131                 DRV_LOG(ERR, "metadata mode %u is not supported "
1132                              "(no metadata reg_c[0] is available)",
1133                              priv->config.dv_xmeta_en);
1134                         err = ENOTSUP;
1135                         goto error;
1136         }
1137         /*
1138          * Allocate the buffer for flow creating, just once.
1139          * The allocation must be done before any flow creating.
1140          */
1141         mlx5_flow_alloc_intermediate(eth_dev);
1142         /* Query availability of metadata reg_c's. */
1143         err = mlx5_flow_discover_mreg_c(eth_dev);
1144         if (err < 0) {
1145                 err = -err;
1146                 goto error;
1147         }
1148         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1149                 DRV_LOG(DEBUG,
1150                         "port %u extensive metadata register is not supported",
1151                         eth_dev->data->port_id);
1152                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1153                         DRV_LOG(ERR, "metadata mode %u is not supported "
1154                                      "(no metadata registers available)",
1155                                      priv->config.dv_xmeta_en);
1156                         err = ENOTSUP;
1157                         goto error;
1158                 }
1159         }
1160         if (priv->config.dv_flow_en &&
1161             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1162             mlx5_flow_ext_mreg_supported(eth_dev) &&
1163             priv->sh->dv_regc0_mask) {
1164                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1165                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1166                 if (!priv->mreg_cp_tbl) {
1167                         err = ENOMEM;
1168                         goto error;
1169                 }
1170         }
1171         return eth_dev;
1172 error:
1173         if (priv) {
1174                 if (priv->mreg_cp_tbl)
1175                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1176                 if (priv->sh)
1177                         mlx5_os_free_shared_dr(priv);
1178                 if (priv->nl_socket_route >= 0)
1179                         close(priv->nl_socket_route);
1180                 if (priv->nl_socket_rdma >= 0)
1181                         close(priv->nl_socket_rdma);
1182                 if (priv->vmwa_context)
1183                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1184                 if (priv->qrss_id_pool)
1185                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1186                 if (own_domain_id)
1187                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1188                 rte_free(priv);
1189                 if (eth_dev != NULL)
1190                         eth_dev->data->dev_private = NULL;
1191         }
1192         if (eth_dev != NULL) {
1193                 /* mac_addrs must not be freed alone because part of
1194                  * dev_private
1195                  **/
1196                 eth_dev->data->mac_addrs = NULL;
1197                 rte_eth_dev_release_port(eth_dev);
1198         }
1199         if (sh)
1200                 mlx5_free_shared_dev_ctx(sh);
1201         MLX5_ASSERT(err > 0);
1202         rte_errno = err;
1203         return NULL;
1204 }
1205
1206 /**
1207  * Comparison callback to sort device data.
1208  *
1209  * This is meant to be used with qsort().
1210  *
1211  * @param a[in]
1212  *   Pointer to pointer to first data object.
1213  * @param b[in]
1214  *   Pointer to pointer to second data object.
1215  *
1216  * @return
1217  *   0 if both objects are equal, less than 0 if the first argument is less
1218  *   than the second, greater than 0 otherwise.
1219  */
1220 static int
1221 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1222 {
1223         const struct mlx5_switch_info *si_a =
1224                 &((const struct mlx5_dev_spawn_data *)a)->info;
1225         const struct mlx5_switch_info *si_b =
1226                 &((const struct mlx5_dev_spawn_data *)b)->info;
1227         int ret;
1228
1229         /* Master device first. */
1230         ret = si_b->master - si_a->master;
1231         if (ret)
1232                 return ret;
1233         /* Then representor devices. */
1234         ret = si_b->representor - si_a->representor;
1235         if (ret)
1236                 return ret;
1237         /* Unidentified devices come last in no specific order. */
1238         if (!si_a->representor)
1239                 return 0;
1240         /* Order representors by name. */
1241         return si_a->port_name - si_b->port_name;
1242 }
1243
1244 /**
1245  * Match PCI information for possible slaves of bonding device.
1246  *
1247  * @param[in] ibv_dev
1248  *   Pointer to Infiniband device structure.
1249  * @param[in] pci_dev
1250  *   Pointer to PCI device structure to match PCI address.
1251  * @param[in] nl_rdma
1252  *   Netlink RDMA group socket handle.
1253  *
1254  * @return
1255  *   negative value if no bonding device found, otherwise
1256  *   positive index of slave PF in bonding.
1257  */
1258 static int
1259 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1260                            const struct rte_pci_device *pci_dev,
1261                            int nl_rdma)
1262 {
1263         char ifname[IF_NAMESIZE + 1];
1264         unsigned int ifindex;
1265         unsigned int np, i;
1266         FILE *file = NULL;
1267         int pf = -1;
1268
1269         /*
1270          * Try to get master device name. If something goes
1271          * wrong suppose the lack of kernel support and no
1272          * bonding devices.
1273          */
1274         if (nl_rdma < 0)
1275                 return -1;
1276         if (!strstr(ibv_dev->name, "bond"))
1277                 return -1;
1278         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1279         if (!np)
1280                 return -1;
1281         /*
1282          * The Master device might not be on the predefined
1283          * port (not on port index 1, it is not garanted),
1284          * we have to scan all Infiniband device port and
1285          * find master.
1286          */
1287         for (i = 1; i <= np; ++i) {
1288                 /* Check whether Infiniband port is populated. */
1289                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1290                 if (!ifindex)
1291                         continue;
1292                 if (!if_indextoname(ifindex, ifname))
1293                         continue;
1294                 /* Try to read bonding slave names from sysfs. */
1295                 MKSTR(slaves,
1296                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1297                 file = fopen(slaves, "r");
1298                 if (file)
1299                         break;
1300         }
1301         if (!file)
1302                 return -1;
1303         /* Use safe format to check maximal buffer length. */
1304         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1305         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1306                 char tmp_str[IF_NAMESIZE + 32];
1307                 struct rte_pci_addr pci_addr;
1308                 struct mlx5_switch_info info;
1309
1310                 /* Process slave interface names in the loop. */
1311                 snprintf(tmp_str, sizeof(tmp_str),
1312                          "/sys/class/net/%s", ifname);
1313                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1314                         DRV_LOG(WARNING, "can not get PCI address"
1315                                          " for netdev \"%s\"", ifname);
1316                         continue;
1317                 }
1318                 if (pci_dev->addr.domain != pci_addr.domain ||
1319                     pci_dev->addr.bus != pci_addr.bus ||
1320                     pci_dev->addr.devid != pci_addr.devid ||
1321                     pci_dev->addr.function != pci_addr.function)
1322                         continue;
1323                 /* Slave interface PCI address match found. */
1324                 fclose(file);
1325                 snprintf(tmp_str, sizeof(tmp_str),
1326                          "/sys/class/net/%s/phys_port_name", ifname);
1327                 file = fopen(tmp_str, "rb");
1328                 if (!file)
1329                         break;
1330                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1331                 if (fscanf(file, "%32s", tmp_str) == 1)
1332                         mlx5_translate_port_name(tmp_str, &info);
1333                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1334                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1335                         pf = info.port_name;
1336                 break;
1337         }
1338         if (file)
1339                 fclose(file);
1340         return pf;
1341 }
1342
1343 /**
1344  * DPDK callback to register a PCI device.
1345  *
1346  * This function spawns Ethernet devices out of a given PCI device.
1347  *
1348  * @param[in] pci_drv
1349  *   PCI driver structure (mlx5_driver).
1350  * @param[in] pci_dev
1351  *   PCI device information.
1352  *
1353  * @return
1354  *   0 on success, a negative errno value otherwise and rte_errno is set.
1355  */
1356 int
1357 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1358                   struct rte_pci_device *pci_dev)
1359 {
1360         struct ibv_device **ibv_list;
1361         /*
1362          * Number of found IB Devices matching with requested PCI BDF.
1363          * nd != 1 means there are multiple IB devices over the same
1364          * PCI device and we have representors and master.
1365          */
1366         unsigned int nd = 0;
1367         /*
1368          * Number of found IB device Ports. nd = 1 and np = 1..n means
1369          * we have the single multiport IB device, and there may be
1370          * representors attached to some of found ports.
1371          */
1372         unsigned int np = 0;
1373         /*
1374          * Number of DPDK ethernet devices to Spawn - either over
1375          * multiple IB devices or multiple ports of single IB device.
1376          * Actually this is the number of iterations to spawn.
1377          */
1378         unsigned int ns = 0;
1379         /*
1380          * Bonding device
1381          *   < 0 - no bonding device (single one)
1382          *  >= 0 - bonding device (value is slave PF index)
1383          */
1384         int bd = -1;
1385         struct mlx5_dev_spawn_data *list = NULL;
1386         struct mlx5_dev_config dev_config;
1387         int ret;
1388
1389         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1390                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1391                         " driver.");
1392                 return 1;
1393         }
1394         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1395                 mlx5_pmd_socket_init();
1396         ret = mlx5_init_once();
1397         if (ret) {
1398                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1399                         strerror(rte_errno));
1400                 return -rte_errno;
1401         }
1402         MLX5_ASSERT(pci_drv == &mlx5_driver);
1403         errno = 0;
1404         ibv_list = mlx5_glue->get_device_list(&ret);
1405         if (!ibv_list) {
1406                 rte_errno = errno ? errno : ENOSYS;
1407                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1408                 return -rte_errno;
1409         }
1410         /*
1411          * First scan the list of all Infiniband devices to find
1412          * matching ones, gathering into the list.
1413          */
1414         struct ibv_device *ibv_match[ret + 1];
1415         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1416         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1417         unsigned int i;
1418
1419         while (ret-- > 0) {
1420                 struct rte_pci_addr pci_addr;
1421
1422                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1423                 bd = mlx5_device_bond_pci_match
1424                                 (ibv_list[ret], pci_dev, nl_rdma);
1425                 if (bd >= 0) {
1426                         /*
1427                          * Bonding device detected. Only one match is allowed,
1428                          * the bonding is supported over multi-port IB device,
1429                          * there should be no matches on representor PCI
1430                          * functions or non VF LAG bonding devices with
1431                          * specified address.
1432                          */
1433                         if (nd) {
1434                                 DRV_LOG(ERR,
1435                                         "multiple PCI match on bonding device"
1436                                         "\"%s\" found", ibv_list[ret]->name);
1437                                 rte_errno = ENOENT;
1438                                 ret = -rte_errno;
1439                                 goto exit;
1440                         }
1441                         DRV_LOG(INFO, "PCI information matches for"
1442                                       " slave %d bonding device \"%s\"",
1443                                       bd, ibv_list[ret]->name);
1444                         ibv_match[nd++] = ibv_list[ret];
1445                         break;
1446                 }
1447                 if (mlx5_dev_to_pci_addr
1448                         (ibv_list[ret]->ibdev_path, &pci_addr))
1449                         continue;
1450                 if (pci_dev->addr.domain != pci_addr.domain ||
1451                     pci_dev->addr.bus != pci_addr.bus ||
1452                     pci_dev->addr.devid != pci_addr.devid ||
1453                     pci_dev->addr.function != pci_addr.function)
1454                         continue;
1455                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1456                         ibv_list[ret]->name);
1457                 ibv_match[nd++] = ibv_list[ret];
1458         }
1459         ibv_match[nd] = NULL;
1460         if (!nd) {
1461                 /* No device matches, just complain and bail out. */
1462                 DRV_LOG(WARNING,
1463                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1464                         " are kernel drivers loaded?",
1465                         pci_dev->addr.domain, pci_dev->addr.bus,
1466                         pci_dev->addr.devid, pci_dev->addr.function);
1467                 rte_errno = ENOENT;
1468                 ret = -rte_errno;
1469                 goto exit;
1470         }
1471         if (nd == 1) {
1472                 /*
1473                  * Found single matching device may have multiple ports.
1474                  * Each port may be representor, we have to check the port
1475                  * number and check the representors existence.
1476                  */
1477                 if (nl_rdma >= 0)
1478                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1479                 if (!np)
1480                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1481                                          " ports number", ibv_match[0]->name);
1482                 if (bd >= 0 && !np) {
1483                         DRV_LOG(ERR, "can not get ports"
1484                                      " for bonding device");
1485                         rte_errno = ENOENT;
1486                         ret = -rte_errno;
1487                         goto exit;
1488                 }
1489         }
1490 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1491         if (bd >= 0) {
1492                 /*
1493                  * This may happen if there is VF LAG kernel support and
1494                  * application is compiled with older rdma_core library.
1495                  */
1496                 DRV_LOG(ERR,
1497                         "No kernel/verbs support for VF LAG bonding found.");
1498                 rte_errno = ENOTSUP;
1499                 ret = -rte_errno;
1500                 goto exit;
1501         }
1502 #endif
1503         /*
1504          * Now we can determine the maximal
1505          * amount of devices to be spawned.
1506          */
1507         list = rte_zmalloc("device spawn data",
1508                          sizeof(struct mlx5_dev_spawn_data) *
1509                          (np ? np : nd),
1510                          RTE_CACHE_LINE_SIZE);
1511         if (!list) {
1512                 DRV_LOG(ERR, "spawn data array allocation failure");
1513                 rte_errno = ENOMEM;
1514                 ret = -rte_errno;
1515                 goto exit;
1516         }
1517         if (bd >= 0 || np > 1) {
1518                 /*
1519                  * Single IB device with multiple ports found,
1520                  * it may be E-Switch master device and representors.
1521                  * We have to perform identification through the ports.
1522                  */
1523                 MLX5_ASSERT(nl_rdma >= 0);
1524                 MLX5_ASSERT(ns == 0);
1525                 MLX5_ASSERT(nd == 1);
1526                 MLX5_ASSERT(np);
1527                 for (i = 1; i <= np; ++i) {
1528                         list[ns].max_port = np;
1529                         list[ns].phys_port = i;
1530                         list[ns].phys_dev = ibv_match[0];
1531                         list[ns].eth_dev = NULL;
1532                         list[ns].pci_dev = pci_dev;
1533                         list[ns].pf_bond = bd;
1534                         list[ns].ifindex = mlx5_nl_ifindex
1535                                 (nl_rdma,
1536                                 mlx5_os_get_dev_device_name
1537                                                 (list[ns].phys_dev), i);
1538                         if (!list[ns].ifindex) {
1539                                 /*
1540                                  * No network interface index found for the
1541                                  * specified port, it means there is no
1542                                  * representor on this port. It's OK,
1543                                  * there can be disabled ports, for example
1544                                  * if sriov_numvfs < sriov_totalvfs.
1545                                  */
1546                                 continue;
1547                         }
1548                         ret = -1;
1549                         if (nl_route >= 0)
1550                                 ret = mlx5_nl_switch_info
1551                                                (nl_route,
1552                                                 list[ns].ifindex,
1553                                                 &list[ns].info);
1554                         if (ret || (!list[ns].info.representor &&
1555                                     !list[ns].info.master)) {
1556                                 /*
1557                                  * We failed to recognize representors with
1558                                  * Netlink, let's try to perform the task
1559                                  * with sysfs.
1560                                  */
1561                                 ret =  mlx5_sysfs_switch_info
1562                                                 (list[ns].ifindex,
1563                                                  &list[ns].info);
1564                         }
1565                         if (!ret && bd >= 0) {
1566                                 switch (list[ns].info.name_type) {
1567                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1568                                         if (list[ns].info.port_name == bd)
1569                                                 ns++;
1570                                         break;
1571                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1572                                         /* Fallthrough */
1573                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1574                                         if (list[ns].info.pf_num == bd)
1575                                                 ns++;
1576                                         break;
1577                                 default:
1578                                         break;
1579                                 }
1580                                 continue;
1581                         }
1582                         if (!ret && (list[ns].info.representor ^
1583                                      list[ns].info.master))
1584                                 ns++;
1585                 }
1586                 if (!ns) {
1587                         DRV_LOG(ERR,
1588                                 "unable to recognize master/representors"
1589                                 " on the IB device with multiple ports");
1590                         rte_errno = ENOENT;
1591                         ret = -rte_errno;
1592                         goto exit;
1593                 }
1594         } else {
1595                 /*
1596                  * The existence of several matching entries (nd > 1) means
1597                  * port representors have been instantiated. No existing Verbs
1598                  * call nor sysfs entries can tell them apart, this can only
1599                  * be done through Netlink calls assuming kernel drivers are
1600                  * recent enough to support them.
1601                  *
1602                  * In the event of identification failure through Netlink,
1603                  * try again through sysfs, then:
1604                  *
1605                  * 1. A single IB device matches (nd == 1) with single
1606                  *    port (np=0/1) and is not a representor, assume
1607                  *    no switch support.
1608                  *
1609                  * 2. Otherwise no safe assumptions can be made;
1610                  *    complain louder and bail out.
1611                  */
1612                 for (i = 0; i != nd; ++i) {
1613                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1614                         list[ns].max_port = 1;
1615                         list[ns].phys_port = 1;
1616                         list[ns].phys_dev = ibv_match[i];
1617                         list[ns].eth_dev = NULL;
1618                         list[ns].pci_dev = pci_dev;
1619                         list[ns].pf_bond = -1;
1620                         list[ns].ifindex = 0;
1621                         if (nl_rdma >= 0)
1622                                 list[ns].ifindex = mlx5_nl_ifindex
1623                                 (nl_rdma,
1624                                 mlx5_os_get_dev_device_name
1625                                                 (list[ns].phys_dev), 1);
1626                         if (!list[ns].ifindex) {
1627                                 char ifname[IF_NAMESIZE];
1628
1629                                 /*
1630                                  * Netlink failed, it may happen with old
1631                                  * ib_core kernel driver (before 4.16).
1632                                  * We can assume there is old driver because
1633                                  * here we are processing single ports IB
1634                                  * devices. Let's try sysfs to retrieve
1635                                  * the ifindex. The method works for
1636                                  * master device only.
1637                                  */
1638                                 if (nd > 1) {
1639                                         /*
1640                                          * Multiple devices found, assume
1641                                          * representors, can not distinguish
1642                                          * master/representor and retrieve
1643                                          * ifindex via sysfs.
1644                                          */
1645                                         continue;
1646                                 }
1647                                 ret = mlx5_get_ifname_sysfs
1648                                         (ibv_match[i]->ibdev_path, ifname);
1649                                 if (!ret)
1650                                         list[ns].ifindex =
1651                                                 if_nametoindex(ifname);
1652                                 if (!list[ns].ifindex) {
1653                                         /*
1654                                          * No network interface index found
1655                                          * for the specified device, it means
1656                                          * there it is neither representor
1657                                          * nor master.
1658                                          */
1659                                         continue;
1660                                 }
1661                         }
1662                         ret = -1;
1663                         if (nl_route >= 0)
1664                                 ret = mlx5_nl_switch_info
1665                                                (nl_route,
1666                                                 list[ns].ifindex,
1667                                                 &list[ns].info);
1668                         if (ret || (!list[ns].info.representor &&
1669                                     !list[ns].info.master)) {
1670                                 /*
1671                                  * We failed to recognize representors with
1672                                  * Netlink, let's try to perform the task
1673                                  * with sysfs.
1674                                  */
1675                                 ret =  mlx5_sysfs_switch_info
1676                                                 (list[ns].ifindex,
1677                                                  &list[ns].info);
1678                         }
1679                         if (!ret && (list[ns].info.representor ^
1680                                      list[ns].info.master)) {
1681                                 ns++;
1682                         } else if ((nd == 1) &&
1683                                    !list[ns].info.representor &&
1684                                    !list[ns].info.master) {
1685                                 /*
1686                                  * Single IB device with
1687                                  * one physical port and
1688                                  * attached network device.
1689                                  * May be SRIOV is not enabled
1690                                  * or there is no representors.
1691                                  */
1692                                 DRV_LOG(INFO, "no E-Switch support detected");
1693                                 ns++;
1694                                 break;
1695                         }
1696                 }
1697                 if (!ns) {
1698                         DRV_LOG(ERR,
1699                                 "unable to recognize master/representors"
1700                                 " on the multiple IB devices");
1701                         rte_errno = ENOENT;
1702                         ret = -rte_errno;
1703                         goto exit;
1704                 }
1705         }
1706         MLX5_ASSERT(ns);
1707         /*
1708          * Sort list to probe devices in natural order for users convenience
1709          * (i.e. master first, then representors from lowest to highest ID).
1710          */
1711         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1712         /* Default configuration. */
1713         dev_config = (struct mlx5_dev_config){
1714                 .hw_padding = 0,
1715                 .mps = MLX5_ARG_UNSET,
1716                 .dbnc = MLX5_ARG_UNSET,
1717                 .rx_vec_en = 1,
1718                 .txq_inline_max = MLX5_ARG_UNSET,
1719                 .txq_inline_min = MLX5_ARG_UNSET,
1720                 .txq_inline_mpw = MLX5_ARG_UNSET,
1721                 .txqs_inline = MLX5_ARG_UNSET,
1722                 .vf_nl_en = 1,
1723                 .mr_ext_memseg_en = 1,
1724                 .mprq = {
1725                         .enabled = 0, /* Disabled by default. */
1726                         .stride_num_n = 0,
1727                         .stride_size_n = 0,
1728                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1729                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1730                 },
1731                 .dv_esw_en = 1,
1732                 .dv_flow_en = 1,
1733                 .log_hp_size = MLX5_ARG_UNSET,
1734         };
1735         /* Device specific configuration. */
1736         switch (pci_dev->id.device_id) {
1737         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1738         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1739         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1740         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1741         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1742         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1743         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1744                 dev_config.vf = 1;
1745                 break;
1746         default:
1747                 break;
1748         }
1749         for (i = 0; i != ns; ++i) {
1750                 uint32_t restore;
1751
1752                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1753                                                  &list[i],
1754                                                  dev_config);
1755                 if (!list[i].eth_dev) {
1756                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1757                                 break;
1758                         /* Device is disabled or already spawned. Ignore it. */
1759                         continue;
1760                 }
1761                 restore = list[i].eth_dev->data->dev_flags;
1762                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1763                 /* Restore non-PCI flags cleared by the above call. */
1764                 list[i].eth_dev->data->dev_flags |= restore;
1765                 rte_eth_dev_probing_finish(list[i].eth_dev);
1766         }
1767         if (i != ns) {
1768                 DRV_LOG(ERR,
1769                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1770                         " encountering an error: %s",
1771                         pci_dev->addr.domain, pci_dev->addr.bus,
1772                         pci_dev->addr.devid, pci_dev->addr.function,
1773                         strerror(rte_errno));
1774                 ret = -rte_errno;
1775                 /* Roll back. */
1776                 while (i--) {
1777                         if (!list[i].eth_dev)
1778                                 continue;
1779                         mlx5_dev_close(list[i].eth_dev);
1780                         /* mac_addrs must not be freed because in dev_private */
1781                         list[i].eth_dev->data->mac_addrs = NULL;
1782                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1783                 }
1784                 /* Restore original error. */
1785                 rte_errno = -ret;
1786         } else {
1787                 ret = 0;
1788         }
1789 exit:
1790         /*
1791          * Do the routine cleanup:
1792          * - close opened Netlink sockets
1793          * - free allocated spawn data array
1794          * - free the Infiniband device list
1795          */
1796         if (nl_rdma >= 0)
1797                 close(nl_rdma);
1798         if (nl_route >= 0)
1799                 close(nl_route);
1800         if (list)
1801                 rte_free(list);
1802         MLX5_ASSERT(ibv_list);
1803         mlx5_glue->free_device_list(ibv_list);
1804         return ret;
1805 }
1806
1807 static int
1808 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1809 {
1810         char *env;
1811         int value;
1812
1813         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1814         /* Get environment variable to store. */
1815         env = getenv(MLX5_SHUT_UP_BF);
1816         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1817         if (config->dbnc == MLX5_ARG_UNSET)
1818                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1819         else
1820                 setenv(MLX5_SHUT_UP_BF,
1821                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1822         return value;
1823 }
1824
1825 static void
1826 mlx5_restore_doorbell_mapping_env(int value)
1827 {
1828         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1829         /* Restore the original environment variable state. */
1830         if (value == MLX5_ARG_UNSET)
1831                 unsetenv(MLX5_SHUT_UP_BF);
1832         else
1833                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1834 }
1835
1836 /**
1837  * Extract pdn of PD object using DV API.
1838  *
1839  * @param[in] pd
1840  *   Pointer to the verbs PD object.
1841  * @param[out] pdn
1842  *   Pointer to the PD object number variable.
1843  *
1844  * @return
1845  *   0 on success, error value otherwise.
1846  */
1847 int
1848 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1849 {
1850 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1851         struct mlx5dv_obj obj;
1852         struct mlx5dv_pd pd_info;
1853         int ret = 0;
1854
1855         obj.pd.in = pd;
1856         obj.pd.out = &pd_info;
1857         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1858         if (ret) {
1859                 DRV_LOG(DEBUG, "Fail to get PD object info");
1860                 return ret;
1861         }
1862         *pdn = pd_info.pdn;
1863         return 0;
1864 #else
1865         (void)pd;
1866         (void)pdn;
1867         return -ENOTSUP;
1868 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1869 }
1870
1871 /**
1872  * Function API to open IB device.
1873  *
1874  * This function calls the Linux glue APIs to open a device.
1875  *
1876  * @param[in] spawn
1877  *   Pointer to the IB device attributes (name, port, etc).
1878  * @param[out] config
1879  *   Pointer to device configuration structure.
1880  * @param[out] sh
1881  *   Pointer to shared context structure.
1882  *
1883  * @return
1884  *   0 on success, a positive error value otherwise.
1885  */
1886 int
1887 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1888                      const struct mlx5_dev_config *config,
1889                      struct mlx5_dev_ctx_shared *sh)
1890 {
1891         int dbmap_env;
1892         int err = 0;
1893
1894         sh->numa_node = spawn->pci_dev->device.numa_node;
1895         pthread_mutex_init(&sh->txpp.mutex, NULL);
1896         /*
1897          * Configure environment variable "MLX5_BF_SHUT_UP"
1898          * before the device creation. The rdma_core library
1899          * checks the variable at device creation and
1900          * stores the result internally.
1901          */
1902         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1903         /* Try to open IB device with DV first, then usual Verbs. */
1904         errno = 0;
1905         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1906         if (sh->ctx) {
1907                 sh->devx = 1;
1908                 DRV_LOG(DEBUG, "DevX is supported");
1909                 /* The device is created, no need for environment. */
1910                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1911         } else {
1912                 /* The environment variable is still configured. */
1913                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1914                 err = errno ? errno : ENODEV;
1915                 /*
1916                  * The environment variable is not needed anymore,
1917                  * all device creation attempts are completed.
1918                  */
1919                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1920                 if (!sh->ctx)
1921                         return err;
1922                 DRV_LOG(DEBUG, "DevX is NOT supported");
1923                 err = 0;
1924         }
1925         return err;
1926 }
1927
1928 /**
1929  * Install shared asynchronous device events handler.
1930  * This function is implemented to support event sharing
1931  * between multiple ports of single IB device.
1932  *
1933  * @param sh
1934  *   Pointer to mlx5_dev_ctx_shared object.
1935  */
1936 void
1937 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1938 {
1939         int ret;
1940         int flags;
1941
1942         sh->intr_handle.fd = -1;
1943         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1944         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1945                     F_SETFL, flags | O_NONBLOCK);
1946         if (ret) {
1947                 DRV_LOG(INFO, "failed to change file descriptor async event"
1948                         " queue");
1949         } else {
1950                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1951                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1952                 if (rte_intr_callback_register(&sh->intr_handle,
1953                                         mlx5_dev_interrupt_handler, sh)) {
1954                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1955                         sh->intr_handle.fd = -1;
1956                 }
1957         }
1958         if (sh->devx) {
1959 #ifdef HAVE_IBV_DEVX_ASYNC
1960                 sh->intr_handle_devx.fd = -1;
1961                 sh->devx_comp =
1962                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1963                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1964                 if (!devx_comp) {
1965                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1966                         return;
1967                 }
1968                 flags = fcntl(devx_comp->fd, F_GETFL);
1969                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1970                 if (ret) {
1971                         DRV_LOG(INFO, "failed to change file descriptor"
1972                                 " devx comp");
1973                         return;
1974                 }
1975                 sh->intr_handle_devx.fd = devx_comp->fd;
1976                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1977                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1978                                         mlx5_dev_interrupt_handler_devx, sh)) {
1979                         DRV_LOG(INFO, "Fail to install the devx shared"
1980                                 " interrupt.");
1981                         sh->intr_handle_devx.fd = -1;
1982                 }
1983 #endif /* HAVE_IBV_DEVX_ASYNC */
1984         }
1985 }
1986
1987 /**
1988  * Uninstall shared asynchronous device events handler.
1989  * This function is implemented to support event sharing
1990  * between multiple ports of single IB device.
1991  *
1992  * @param dev
1993  *   Pointer to mlx5_dev_ctx_shared object.
1994  */
1995 void
1996 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
1997 {
1998         if (sh->intr_handle.fd >= 0)
1999                 mlx5_intr_callback_unregister(&sh->intr_handle,
2000                                               mlx5_dev_interrupt_handler, sh);
2001 #ifdef HAVE_IBV_DEVX_ASYNC
2002         if (sh->intr_handle_devx.fd >= 0)
2003                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2004                                   mlx5_dev_interrupt_handler_devx, sh);
2005         if (sh->devx_comp)
2006                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2007 #endif
2008 }
2009
2010 /**
2011  * Read statistics by a named counter.
2012  *
2013  * @param[in] priv
2014  *   Pointer to the private device data structure.
2015  * @param[in] ctr_name
2016  *   Pointer to the name of the statistic counter to read
2017  * @param[out] stat
2018  *   Pointer to read statistic value.
2019  * @return
2020  *   0 on success and stat is valud, 1 if failed to read the value
2021  *   rte_errno is set.
2022  *
2023  */
2024 int
2025 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2026                       uint64_t *stat)
2027 {
2028         int fd;
2029
2030         if (priv->sh) {
2031                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2032                           priv->sh->ibdev_path,
2033                           priv->dev_port,
2034                           ctr_name);
2035                 fd = open(path, O_RDONLY);
2036                 if (fd != -1) {
2037                         char buf[21] = {'\0'};
2038                         ssize_t n = read(fd, buf, sizeof(buf));
2039
2040                         close(fd);
2041                         if (n != -1) {
2042                                 *stat = strtoull(buf, NULL, 10);
2043                                 return 0;
2044                         }
2045                 }
2046         }
2047         *stat = 0;
2048         return 1;
2049 }
2050
2051 /**
2052  * Read device counters table.
2053  *
2054  * @param dev
2055  *   Pointer to Ethernet device.
2056  * @param[out] stats
2057  *   Counters table output buffer.
2058  *
2059  * @return
2060  *   0 on success and stats is filled, negative errno value otherwise and
2061  *   rte_errno is set.
2062  */
2063 int
2064 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
2065 {
2066         struct mlx5_priv *priv = dev->data->dev_private;
2067         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2068         unsigned int i;
2069         struct ifreq ifr;
2070         unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
2071         unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
2072         struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
2073         int ret;
2074
2075         et_stats->cmd = ETHTOOL_GSTATS;
2076         et_stats->n_stats = xstats_ctrl->stats_n;
2077         ifr.ifr_data = (caddr_t)et_stats;
2078         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2079         if (ret) {
2080                 DRV_LOG(WARNING,
2081                         "port %u unable to read statistic values from device",
2082                         dev->data->port_id);
2083                 return ret;
2084         }
2085         for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2086                 if (xstats_ctrl->info[i].dev) {
2087                         ret = mlx5_os_read_dev_stat(priv,
2088                                             xstats_ctrl->info[i].ctr_name,
2089                                             &stats[i]);
2090                         /* return last xstats counter if fail to read. */
2091                         if (ret == 0)
2092                                 xstats_ctrl->xstats[i] = stats[i];
2093                         else
2094                                 stats[i] = xstats_ctrl->xstats[i];
2095                 } else {
2096                         stats[i] = (uint64_t)
2097                                 et_stats->data[xstats_ctrl->dev_table_idx[i]];
2098                 }
2099         }
2100         return 0;
2101 }
2102
2103 /**
2104  * Query the number of statistics provided by ETHTOOL.
2105  *
2106  * @param dev
2107  *   Pointer to Ethernet device.
2108  *
2109  * @return
2110  *   Number of statistics on success, negative errno value otherwise and
2111  *   rte_errno is set.
2112  */
2113 int
2114 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2115 {
2116         struct ethtool_drvinfo drvinfo;
2117         struct ifreq ifr;
2118         int ret;
2119
2120         drvinfo.cmd = ETHTOOL_GDRVINFO;
2121         ifr.ifr_data = (caddr_t)&drvinfo;
2122         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2123         if (ret) {
2124                 DRV_LOG(WARNING, "port %u unable to query number of statistics",
2125                         dev->data->port_id);
2126                 return ret;
2127         }
2128         return drvinfo.n_stats;
2129 }
2130
2131 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2132         {
2133                 .dpdk_name = "rx_port_unicast_bytes",
2134                 .ctr_name = "rx_vport_unicast_bytes",
2135         },
2136         {
2137                 .dpdk_name = "rx_port_multicast_bytes",
2138                 .ctr_name = "rx_vport_multicast_bytes",
2139         },
2140         {
2141                 .dpdk_name = "rx_port_broadcast_bytes",
2142                 .ctr_name = "rx_vport_broadcast_bytes",
2143         },
2144         {
2145                 .dpdk_name = "rx_port_unicast_packets",
2146                 .ctr_name = "rx_vport_unicast_packets",
2147         },
2148         {
2149                 .dpdk_name = "rx_port_multicast_packets",
2150                 .ctr_name = "rx_vport_multicast_packets",
2151         },
2152         {
2153                 .dpdk_name = "rx_port_broadcast_packets",
2154                 .ctr_name = "rx_vport_broadcast_packets",
2155         },
2156         {
2157                 .dpdk_name = "tx_port_unicast_bytes",
2158                 .ctr_name = "tx_vport_unicast_bytes",
2159         },
2160         {
2161                 .dpdk_name = "tx_port_multicast_bytes",
2162                 .ctr_name = "tx_vport_multicast_bytes",
2163         },
2164         {
2165                 .dpdk_name = "tx_port_broadcast_bytes",
2166                 .ctr_name = "tx_vport_broadcast_bytes",
2167         },
2168         {
2169                 .dpdk_name = "tx_port_unicast_packets",
2170                 .ctr_name = "tx_vport_unicast_packets",
2171         },
2172         {
2173                 .dpdk_name = "tx_port_multicast_packets",
2174                 .ctr_name = "tx_vport_multicast_packets",
2175         },
2176         {
2177                 .dpdk_name = "tx_port_broadcast_packets",
2178                 .ctr_name = "tx_vport_broadcast_packets",
2179         },
2180         {
2181                 .dpdk_name = "rx_wqe_err",
2182                 .ctr_name = "rx_wqe_err",
2183         },
2184         {
2185                 .dpdk_name = "rx_crc_errors_phy",
2186                 .ctr_name = "rx_crc_errors_phy",
2187         },
2188         {
2189                 .dpdk_name = "rx_in_range_len_errors_phy",
2190                 .ctr_name = "rx_in_range_len_errors_phy",
2191         },
2192         {
2193                 .dpdk_name = "rx_symbol_err_phy",
2194                 .ctr_name = "rx_symbol_err_phy",
2195         },
2196         {
2197                 .dpdk_name = "tx_errors_phy",
2198                 .ctr_name = "tx_errors_phy",
2199         },
2200         {
2201                 .dpdk_name = "rx_out_of_buffer",
2202                 .ctr_name = "out_of_buffer",
2203                 .dev = 1,
2204         },
2205         {
2206                 .dpdk_name = "tx_packets_phy",
2207                 .ctr_name = "tx_packets_phy",
2208         },
2209         {
2210                 .dpdk_name = "rx_packets_phy",
2211                 .ctr_name = "rx_packets_phy",
2212         },
2213         {
2214                 .dpdk_name = "tx_discards_phy",
2215                 .ctr_name = "tx_discards_phy",
2216         },
2217         {
2218                 .dpdk_name = "rx_discards_phy",
2219                 .ctr_name = "rx_discards_phy",
2220         },
2221         {
2222                 .dpdk_name = "tx_bytes_phy",
2223                 .ctr_name = "tx_bytes_phy",
2224         },
2225         {
2226                 .dpdk_name = "rx_bytes_phy",
2227                 .ctr_name = "rx_bytes_phy",
2228         },
2229         /* Representor only */
2230         {
2231                 .dpdk_name = "rx_packets",
2232                 .ctr_name = "vport_rx_packets",
2233         },
2234         {
2235                 .dpdk_name = "rx_bytes",
2236                 .ctr_name = "vport_rx_bytes",
2237         },
2238         {
2239                 .dpdk_name = "tx_packets",
2240                 .ctr_name = "vport_tx_packets",
2241         },
2242         {
2243                 .dpdk_name = "tx_bytes",
2244                 .ctr_name = "vport_tx_bytes",
2245         },
2246 };
2247
2248 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2249
2250 /**
2251  * Init the structures to read device counters.
2252  *
2253  * @param dev
2254  *   Pointer to Ethernet device.
2255  */
2256 void
2257 mlx5_os_stats_init(struct rte_eth_dev *dev)
2258 {
2259         struct mlx5_priv *priv = dev->data->dev_private;
2260         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2261         struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2262         unsigned int i;
2263         unsigned int j;
2264         struct ifreq ifr;
2265         struct ethtool_gstrings *strings = NULL;
2266         unsigned int dev_stats_n;
2267         unsigned int str_sz;
2268         int ret;
2269
2270         /* So that it won't aggregate for each init. */
2271         xstats_ctrl->mlx5_stats_n = 0;
2272         ret = mlx5_os_get_stats_n(dev);
2273         if (ret < 0) {
2274                 DRV_LOG(WARNING, "port %u no extended statistics available",
2275                         dev->data->port_id);
2276                 return;
2277         }
2278         dev_stats_n = ret;
2279         /* Allocate memory to grab stat names and values. */
2280         str_sz = dev_stats_n * ETH_GSTRING_LEN;
2281         strings = (struct ethtool_gstrings *)
2282                   rte_malloc("xstats_strings",
2283                              str_sz + sizeof(struct ethtool_gstrings), 0);
2284         if (!strings) {
2285                 DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2286                      dev->data->port_id);
2287                 return;
2288         }
2289         strings->cmd = ETHTOOL_GSTRINGS;
2290         strings->string_set = ETH_SS_STATS;
2291         strings->len = dev_stats_n;
2292         ifr.ifr_data = (caddr_t)strings;
2293         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2294         if (ret) {
2295                 DRV_LOG(WARNING, "port %u unable to get statistic names",
2296                         dev->data->port_id);
2297                 goto free;
2298         }
2299         for (i = 0; i != dev_stats_n; ++i) {
2300                 const char *curr_string = (const char *)
2301                         &strings->data[i * ETH_GSTRING_LEN];
2302
2303                 for (j = 0; j != xstats_n; ++j) {
2304                         if (!strcmp(mlx5_counters_init[j].ctr_name,
2305                                     curr_string)) {
2306                                 unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2307
2308                                 xstats_ctrl->dev_table_idx[idx] = i;
2309                                 xstats_ctrl->info[idx] = mlx5_counters_init[j];
2310                                 break;
2311                         }
2312                 }
2313         }
2314         /* Add dev counters. */
2315         for (i = 0; i != xstats_n; ++i) {
2316                 if (mlx5_counters_init[i].dev) {
2317                         unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2318
2319                         xstats_ctrl->info[idx] = mlx5_counters_init[i];
2320                         xstats_ctrl->hw_stats[idx] = 0;
2321                 }
2322         }
2323         MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2324         xstats_ctrl->stats_n = dev_stats_n;
2325         /* Copy to base at first time. */
2326         ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2327         if (ret)
2328                 DRV_LOG(ERR, "port %u cannot read device counters: %s",
2329                         dev->data->port_id, strerror(rte_errno));
2330         mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2331         stats_ctrl->imissed = 0;
2332 free:
2333         rte_free(strings);
2334 }
2335
2336 /**
2337  * Set the reg_mr and dereg_mr call backs
2338  *
2339  * @param reg_mr_cb[out]
2340  *   Pointer to reg_mr func
2341  * @param dereg_mr_cb[out]
2342  *   Pointer to dereg_mr func
2343  *
2344  */
2345 void
2346 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2347                       mlx5_dereg_mr_t *dereg_mr_cb)
2348 {
2349         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2350         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2351 }
2352
2353 const struct eth_dev_ops mlx5_os_dev_ops = {
2354         .dev_configure = mlx5_dev_configure,
2355         .dev_start = mlx5_dev_start,
2356         .dev_stop = mlx5_dev_stop,
2357         .dev_set_link_down = mlx5_set_link_down,
2358         .dev_set_link_up = mlx5_set_link_up,
2359         .dev_close = mlx5_dev_close,
2360         .promiscuous_enable = mlx5_promiscuous_enable,
2361         .promiscuous_disable = mlx5_promiscuous_disable,
2362         .allmulticast_enable = mlx5_allmulticast_enable,
2363         .allmulticast_disable = mlx5_allmulticast_disable,
2364         .link_update = mlx5_link_update,
2365         .stats_get = mlx5_stats_get,
2366         .stats_reset = mlx5_stats_reset,
2367         .xstats_get = mlx5_xstats_get,
2368         .xstats_reset = mlx5_xstats_reset,
2369         .xstats_get_names = mlx5_xstats_get_names,
2370         .fw_version_get = mlx5_fw_version_get,
2371         .dev_infos_get = mlx5_dev_infos_get,
2372         .read_clock = mlx5_txpp_read_clock,
2373         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2374         .vlan_filter_set = mlx5_vlan_filter_set,
2375         .rx_queue_setup = mlx5_rx_queue_setup,
2376         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2377         .tx_queue_setup = mlx5_tx_queue_setup,
2378         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2379         .rx_queue_release = mlx5_rx_queue_release,
2380         .tx_queue_release = mlx5_tx_queue_release,
2381         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2382         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2383         .mac_addr_remove = mlx5_mac_addr_remove,
2384         .mac_addr_add = mlx5_mac_addr_add,
2385         .mac_addr_set = mlx5_mac_addr_set,
2386         .set_mc_addr_list = mlx5_set_mc_addr_list,
2387         .mtu_set = mlx5_dev_set_mtu,
2388         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2389         .vlan_offload_set = mlx5_vlan_offload_set,
2390         .reta_update = mlx5_dev_rss_reta_update,
2391         .reta_query = mlx5_dev_rss_reta_query,
2392         .rss_hash_update = mlx5_rss_hash_update,
2393         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2394         .filter_ctrl = mlx5_dev_filter_ctrl,
2395         .rx_descriptor_status = mlx5_rx_descriptor_status,
2396         .tx_descriptor_status = mlx5_tx_descriptor_status,
2397         .rxq_info_get = mlx5_rxq_info_get,
2398         .txq_info_get = mlx5_txq_info_get,
2399         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2400         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2401         .rx_queue_count = mlx5_rx_queue_count,
2402         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2403         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2404         .is_removed = mlx5_is_removed,
2405         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2406         .get_module_info = mlx5_get_module_info,
2407         .get_module_eeprom = mlx5_get_module_eeprom,
2408         .hairpin_cap_get = mlx5_hairpin_cap_get,
2409         .mtr_ops_get = mlx5_flow_meter_ops_get,
2410 };
2411
2412 /* Available operations from secondary process. */
2413 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2414         .stats_get = mlx5_stats_get,
2415         .stats_reset = mlx5_stats_reset,
2416         .xstats_get = mlx5_xstats_get,
2417         .xstats_reset = mlx5_xstats_reset,
2418         .xstats_get_names = mlx5_xstats_get_names,
2419         .fw_version_get = mlx5_fw_version_get,
2420         .dev_infos_get = mlx5_dev_infos_get,
2421         .read_clock = mlx5_txpp_read_clock,
2422         .rx_descriptor_status = mlx5_rx_descriptor_status,
2423         .tx_descriptor_status = mlx5_tx_descriptor_status,
2424         .rxq_info_get = mlx5_rxq_info_get,
2425         .txq_info_get = mlx5_txq_info_get,
2426         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2427         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2428         .get_module_info = mlx5_get_module_info,
2429         .get_module_eeprom = mlx5_get_module_eeprom,
2430 };
2431
2432 /* Available operations in flow isolated mode. */
2433 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2434         .dev_configure = mlx5_dev_configure,
2435         .dev_start = mlx5_dev_start,
2436         .dev_stop = mlx5_dev_stop,
2437         .dev_set_link_down = mlx5_set_link_down,
2438         .dev_set_link_up = mlx5_set_link_up,
2439         .dev_close = mlx5_dev_close,
2440         .promiscuous_enable = mlx5_promiscuous_enable,
2441         .promiscuous_disable = mlx5_promiscuous_disable,
2442         .allmulticast_enable = mlx5_allmulticast_enable,
2443         .allmulticast_disable = mlx5_allmulticast_disable,
2444         .link_update = mlx5_link_update,
2445         .stats_get = mlx5_stats_get,
2446         .stats_reset = mlx5_stats_reset,
2447         .xstats_get = mlx5_xstats_get,
2448         .xstats_reset = mlx5_xstats_reset,
2449         .xstats_get_names = mlx5_xstats_get_names,
2450         .fw_version_get = mlx5_fw_version_get,
2451         .dev_infos_get = mlx5_dev_infos_get,
2452         .read_clock = mlx5_txpp_read_clock,
2453         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2454         .vlan_filter_set = mlx5_vlan_filter_set,
2455         .rx_queue_setup = mlx5_rx_queue_setup,
2456         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2457         .tx_queue_setup = mlx5_tx_queue_setup,
2458         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2459         .rx_queue_release = mlx5_rx_queue_release,
2460         .tx_queue_release = mlx5_tx_queue_release,
2461         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2462         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2463         .mac_addr_remove = mlx5_mac_addr_remove,
2464         .mac_addr_add = mlx5_mac_addr_add,
2465         .mac_addr_set = mlx5_mac_addr_set,
2466         .set_mc_addr_list = mlx5_set_mc_addr_list,
2467         .mtu_set = mlx5_dev_set_mtu,
2468         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2469         .vlan_offload_set = mlx5_vlan_offload_set,
2470         .filter_ctrl = mlx5_dev_filter_ctrl,
2471         .rx_descriptor_status = mlx5_rx_descriptor_status,
2472         .tx_descriptor_status = mlx5_tx_descriptor_status,
2473         .rxq_info_get = mlx5_rxq_info_get,
2474         .txq_info_get = mlx5_txq_info_get,
2475         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2476         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2477         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2478         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2479         .is_removed = mlx5_is_removed,
2480         .get_module_info = mlx5_get_module_info,
2481         .get_module_eeprom = mlx5_get_module_eeprom,
2482         .hairpin_cap_get = mlx5_hairpin_cap_get,
2483         .mtr_ops_get = mlx5_flow_meter_ops_get,
2484 };