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 #include <mlx5_malloc.h>
  47
  48 #include "mlx5_defs.h"
  49 #include "mlx5.h"
  50 #include "mlx5_common_os.h"
  51 #include "mlx5_utils.h"
  52 #include "mlx5_rxtx.h"
  53 #include "mlx5_autoconf.h"
  54 #include "mlx5_mr.h"
  55 #include "mlx5_flow.h"
  56 #include "rte_pmd_mlx5.h"
  57 #include "mlx5_verbs.h"
  58
  59 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  60
  61 #ifndef HAVE_IBV_MLX5_MOD_MPW
  62 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  63 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  64 #endif
  65
  66 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  67 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  68 #endif
  69
  70 /**
  71  * Get mlx5 device attributes. The glue function query_device_ex() is called
  72  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  73  * device attributes from the glue out parameter.
  74  *
  75  * @param dev
  76  *   Pointer to ibv context.
  77  *
  78  * @param device_attr
  79  *   Pointer to mlx5 device attributes.
  80  *
  81  * @return
  82  *   0 on success, non zero error number otherwise
  83  */
  84 int
  85 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
  86 {
  87         int err;
  88         struct ibv_device_attr_ex attr_ex;
  89         memset(device_attr, 0, sizeof(*device_attr));
  90         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
  91         if (err)
  92                 return err;
  93
  94         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
  95         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
  96         device_attr->max_sge = attr_ex.orig_attr.max_sge;
  97         device_attr->max_cq = attr_ex.orig_attr.max_cq;
  98         device_attr->max_qp = attr_ex.orig_attr.max_qp;
  99         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 100         device_attr->max_rwq_indirection_table_size =
 101                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 102         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 103         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 104
 105         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 106         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 107         if (err)
 108                 return err;
 109
 110         device_attr->flags = dv_attr.flags;
 111         device_attr->comp_mask = dv_attr.comp_mask;
 112 #ifdef HAVE_IBV_MLX5_MOD_SWP
 113         device_attr->sw_parsing_offloads =
 114                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 115 #endif
 116         device_attr->min_single_stride_log_num_of_bytes =
 117                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 118         device_attr->max_single_stride_log_num_of_bytes =
 119                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 120         device_attr->min_single_wqe_log_num_of_strides =
 121                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 122         device_attr->max_single_wqe_log_num_of_strides =
 123                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 124         device_attr->stride_supported_qpts =
 125                 dv_attr.striding_rq_caps.supported_qpts;
 126 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 127         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 128 #endif
 129
 130         return err;
 131 }
 132
 133 /**
 134  * Verbs callback to allocate a memory. This function should allocate the space
 135  * according to the size provided residing inside a huge page.
 136  * Please note that all allocation must respect the alignment from libmlx5
 137  * (i.e. currently sysconf(_SC_PAGESIZE)).
 138  *
 139  * @param[in] size
 140  *   The size in bytes of the memory to allocate.
 141  * @param[in] data
 142  *   A pointer to the callback data.
 143  *
 144  * @return
 145  *   Allocated buffer, NULL otherwise and rte_errno is set.
 146  */
 147 static void *
 148 mlx5_alloc_verbs_buf(size_t size, void *data)
 149 {
 150         struct mlx5_priv *priv = data;
 151         void *ret;
 152         size_t alignment = sysconf(_SC_PAGESIZE);
 153         unsigned int socket = SOCKET_ID_ANY;
 154
 155         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 156                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 157
 158                 socket = ctrl->socket;
 159         } else if (priv->verbs_alloc_ctx.type ==
 160                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 161                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 162
 163                 socket = ctrl->socket;
 164         }
 165         MLX5_ASSERT(data != NULL);
 166         ret = mlx5_malloc(0, size, alignment, socket);
 167         if (!ret && size)
 168                 rte_errno = ENOMEM;
 169         return ret;
 170 }
 171
 172 /**
 173  * Verbs callback to free a memory.
 174  *
 175  * @param[in] ptr
 176  *   A pointer to the memory to free.
 177  * @param[in] data
 178  *   A pointer to the callback data.
 179  */
 180 static void
 181 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 182 {
 183         MLX5_ASSERT(data != NULL);
 184         mlx5_free(ptr);
 185 }
 186
 187 /**
 188  * Initialize DR related data within private structure.
 189  * Routine checks the reference counter and does actual
 190  * resources creation/initialization only if counter is zero.
 191  *
 192  * @param[in] priv
 193  *   Pointer to the private device data structure.
 194  *
 195  * @return
 196  *   Zero on success, positive error code otherwise.
 197  */
 198 static int
 199 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 200 {
 201         struct mlx5_dev_ctx_shared *sh = priv->sh;
 202         char s[MLX5_HLIST_NAMESIZE];
 203         int err = 0;
 204
 205         if (!sh->flow_tbls)
 206                 err = mlx5_alloc_table_hash_list(priv);
 207         else
 208                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 209                         (void *)sh->flow_tbls);
 210         if (err)
 211                 return err;
 212         /* Create tags hash list table. */
 213         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 214         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 215         if (!sh->tag_table) {
 216                 DRV_LOG(ERR, "tags with hash creation failed.");
 217                 err = ENOMEM;
 218                 goto error;
 219         }
 220 #ifdef HAVE_MLX5DV_DR
 221         void *domain;
 222
 223         if (sh->dv_refcnt) {
 224                 /* Shared DV/DR structures is already initialized. */
 225                 sh->dv_refcnt++;
 226                 priv->dr_shared = 1;
 227                 return 0;
 228         }
 229         /* Reference counter is zero, we should initialize structures. */
 230         domain = mlx5_glue->dr_create_domain(sh->ctx,
 231                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 232         if (!domain) {
 233                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 234                 err = errno;
 235                 goto error;
 236         }
 237         sh->rx_domain = domain;
 238         domain = mlx5_glue->dr_create_domain(sh->ctx,
 239                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 240         if (!domain) {
 241                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 242                 err = errno;
 243                 goto error;
 244         }
 245         pthread_mutex_init(&sh->dv_mutex, NULL);
 246         sh->tx_domain = domain;
 247 #ifdef HAVE_MLX5DV_DR_ESWITCH
 248         if (priv->config.dv_esw_en) {
 249                 domain  = mlx5_glue->dr_create_domain
 250                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 251                 if (!domain) {
 252                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 253                         err = errno;
 254                         goto error;
 255                 }
 256                 sh->fdb_domain = domain;
 257                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 258         }
 259 #endif
 260         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 261                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 262                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 263                 if (sh->fdb_domain)
 264                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 265         }
 266         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 267 #endif /* HAVE_MLX5DV_DR */
 268         sh->dv_refcnt++;
 269         priv->dr_shared = 1;
 270         return 0;
 271 error:
 272         /* Rollback the created objects. */
 273         if (sh->rx_domain) {
 274                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 275                 sh->rx_domain = NULL;
 276         }
 277         if (sh->tx_domain) {
 278                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 279                 sh->tx_domain = NULL;
 280         }
 281         if (sh->fdb_domain) {
 282                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 283                 sh->fdb_domain = NULL;
 284         }
 285         if (sh->esw_drop_action) {
 286                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 287                 sh->esw_drop_action = NULL;
 288         }
 289         if (sh->pop_vlan_action) {
 290                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 291                 sh->pop_vlan_action = NULL;
 292         }
 293         if (sh->tag_table) {
 294                 /* tags should be destroyed with flow before. */
 295                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 296                 sh->tag_table = NULL;
 297         }
 298         mlx5_free_table_hash_list(priv);
 299         return err;
 300 }
 301
 302 /**
 303  * Destroy DR related data within private structure.
 304  *
 305  * @param[in] priv
 306  *   Pointer to the private device data structure.
 307  */
 308 void
 309 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 310 {
 311         struct mlx5_dev_ctx_shared *sh;
 312
 313         if (!priv->dr_shared)
 314                 return;
 315         priv->dr_shared = 0;
 316         sh = priv->sh;
 317         MLX5_ASSERT(sh);
 318 #ifdef HAVE_MLX5DV_DR
 319         MLX5_ASSERT(sh->dv_refcnt);
 320         if (sh->dv_refcnt && --sh->dv_refcnt)
 321                 return;
 322         if (sh->rx_domain) {
 323                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 324                 sh->rx_domain = NULL;
 325         }
 326         if (sh->tx_domain) {
 327                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 328                 sh->tx_domain = NULL;
 329         }
 330 #ifdef HAVE_MLX5DV_DR_ESWITCH
 331         if (sh->fdb_domain) {
 332                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 333                 sh->fdb_domain = NULL;
 334         }
 335         if (sh->esw_drop_action) {
 336                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 337                 sh->esw_drop_action = NULL;
 338         }
 339 #endif
 340         if (sh->pop_vlan_action) {
 341                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 342                 sh->pop_vlan_action = NULL;
 343         }
 344         pthread_mutex_destroy(&sh->dv_mutex);
 345 #endif /* HAVE_MLX5DV_DR */
 346         if (sh->tag_table) {
 347                 /* tags should be destroyed with flow before. */
 348                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 349                 sh->tag_table = NULL;
 350         }
 351         mlx5_free_table_hash_list(priv);
 352 }
 353
 354 /**
 355  * Spawn an Ethernet device from Verbs information.
 356  *
 357  * @param dpdk_dev
 358  *   Backing DPDK device.
 359  * @param spawn
 360  *   Verbs device parameters (name, port, switch_info) to spawn.
 361  * @param config
 362  *   Device configuration parameters.
 363  *
 364  * @return
 365  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 366  *   is set. The following errors are defined:
 367  *
 368  *   EBUSY: device is not supposed to be spawned.
 369  *   EEXIST: device is already spawned
 370  */
 371 static struct rte_eth_dev *
 372 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 373                struct mlx5_dev_spawn_data *spawn,
 374                struct mlx5_dev_config config)
 375 {
 376         const struct mlx5_switch_info *switch_info = &spawn->info;
 377         struct mlx5_dev_ctx_shared *sh = NULL;
 378         struct ibv_port_attr port_attr;
 379         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 380         struct rte_eth_dev *eth_dev = NULL;
 381         struct mlx5_priv *priv = NULL;
 382         int err = 0;
 383         unsigned int hw_padding = 0;
 384         unsigned int mps;
 385         unsigned int cqe_comp;
 386         unsigned int cqe_pad = 0;
 387         unsigned int tunnel_en = 0;
 388         unsigned int mpls_en = 0;
 389         unsigned int swp = 0;
 390         unsigned int mprq = 0;
 391         unsigned int mprq_min_stride_size_n = 0;
 392         unsigned int mprq_max_stride_size_n = 0;
 393         unsigned int mprq_min_stride_num_n = 0;
 394         unsigned int mprq_max_stride_num_n = 0;
 395         struct rte_ether_addr mac;
 396         char name[RTE_ETH_NAME_MAX_LEN];
 397         int own_domain_id = 0;
 398         uint16_t port_id;
 399         unsigned int i;
 400 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 401         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 402 #endif
 403
 404         /* Determine if this port representor is supposed to be spawned. */
 405         if (switch_info->representor && dpdk_dev->devargs) {
 406                 struct rte_eth_devargs eth_da;
 407
 408                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 409                 if (err) {
 410                         rte_errno = -err;
 411                         DRV_LOG(ERR, "failed to process device arguments: %s",
 412                                 strerror(rte_errno));
 413                         return NULL;
 414                 }
 415                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 416                         if (eth_da.representor_ports[i] ==
 417                             (uint16_t)switch_info->port_name)
 418                                 break;
 419                 if (i == eth_da.nb_representor_ports) {
 420                         rte_errno = EBUSY;
 421                         return NULL;
 422                 }
 423         }
 424         /* Build device name. */
 425         if (spawn->pf_bond <  0) {
 426                 /* Single device. */
 427                 if (!switch_info->representor)
 428                         strlcpy(name, dpdk_dev->name, sizeof(name));
 429                 else
 430                         snprintf(name, sizeof(name), "%s_representor_%u",
 431                                  dpdk_dev->name, switch_info->port_name);
 432         } else {
 433                 /* Bonding device. */
 434                 if (!switch_info->representor)
 435                         snprintf(name, sizeof(name), "%s_%s",
 436                                  dpdk_dev->name,
 437                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 438                 else
 439                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 440                                  dpdk_dev->name,
 441                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 442                                  switch_info->port_name);
 443         }
 444         /* check if the device is already spawned */
 445         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 446                 rte_errno = EEXIST;
 447                 return NULL;
 448         }
 449         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 450         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 451                 struct mlx5_mp_id mp_id;
 452
 453                 eth_dev = rte_eth_dev_attach_secondary(name);
 454                 if (eth_dev == NULL) {
 455                         DRV_LOG(ERR, "can not attach rte ethdev");
 456                         rte_errno = ENOMEM;
 457                         return NULL;
 458                 }
 459                 eth_dev->device = dpdk_dev;
 460                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 461                 err = mlx5_proc_priv_init(eth_dev);
 462                 if (err)
 463                         return NULL;
 464                 mp_id.port_id = eth_dev->data->port_id;
 465                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 466                 /* Receive command fd from primary process */
 467                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 468                 if (err < 0)
 469                         goto err_secondary;
 470                 /* Remap UAR for Tx queues. */
 471                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 472                 if (err)
 473                         goto err_secondary;
 474                 /*
 475                  * Ethdev pointer is still required as input since
 476                  * the primary device is not accessible from the
 477                  * secondary process.
 478                  */
 479                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 480                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 481                 return eth_dev;
 482 err_secondary:
 483                 mlx5_dev_close(eth_dev);
 484                 return NULL;
 485         }
 486         /*
 487          * Some parameters ("tx_db_nc" in particularly) are needed in
 488          * advance to create dv/verbs device context. We proceed the
 489          * devargs here to get ones, and later proceed devargs again
 490          * to override some hardware settings.
 491          */
 492         err = mlx5_args(&config, dpdk_dev->devargs);
 493         if (err) {
 494                 err = rte_errno;
 495                 DRV_LOG(ERR, "failed to process device arguments: %s",
 496                         strerror(rte_errno));
 497                 goto error;
 498         }
 499         mlx5_malloc_mem_select(config.sys_mem_en);
 500         sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
 501         if (!sh)
 502                 return NULL;
 503         config.devx = sh->devx;
 504 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 505         config.dest_tir = 1;
 506 #endif
 507 #ifdef HAVE_IBV_MLX5_MOD_SWP
 508         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 509 #endif
 510         /*
 511          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 512          * as all ConnectX-5 devices.
 513          */
 514 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 515         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 516 #endif
 517 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 518         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 519 #endif
 520         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 521         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 522                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 523                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 524                         mps = MLX5_MPW_ENHANCED;
 525                 } else {
 526                         DRV_LOG(DEBUG, "MPW is supported");
 527                         mps = MLX5_MPW;
 528                 }
 529         } else {
 530                 DRV_LOG(DEBUG, "MPW isn't supported");
 531                 mps = MLX5_MPW_DISABLED;
 532         }
 533 #ifdef HAVE_IBV_MLX5_MOD_SWP
 534         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 535                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 536         DRV_LOG(DEBUG, "SWP support: %u", swp);
 537 #endif
 538         config.swp = !!swp;
 539 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 540         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 541                 struct mlx5dv_striding_rq_caps mprq_caps =
 542                         dv_attr.striding_rq_caps;
 543
 544                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 545                         mprq_caps.min_single_stride_log_num_of_bytes);
 546                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 547                         mprq_caps.max_single_stride_log_num_of_bytes);
 548                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 549                         mprq_caps.min_single_wqe_log_num_of_strides);
 550                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 551                         mprq_caps.max_single_wqe_log_num_of_strides);
 552                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 553                         mprq_caps.supported_qpts);
 554                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 555                 mprq = 1;
 556                 mprq_min_stride_size_n =
 557                         mprq_caps.min_single_stride_log_num_of_bytes;
 558                 mprq_max_stride_size_n =
 559                         mprq_caps.max_single_stride_log_num_of_bytes;
 560                 mprq_min_stride_num_n =
 561                         mprq_caps.min_single_wqe_log_num_of_strides;
 562                 mprq_max_stride_num_n =
 563                         mprq_caps.max_single_wqe_log_num_of_strides;
 564         }
 565 #endif
 566         if (RTE_CACHE_LINE_SIZE == 128 &&
 567             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 568                 cqe_comp = 0;
 569         else
 570                 cqe_comp = 1;
 571         config.cqe_comp = cqe_comp;
 572 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 573         /* Whether device supports 128B Rx CQE padding. */
 574         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 575                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 576 #endif
 577 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 578         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 579                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 580                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 581                              (dv_attr.tunnel_offloads_caps &
 582                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 583                              (dv_attr.tunnel_offloads_caps &
 584                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 585         }
 586         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 587                 tunnel_en ? "" : "not ");
 588 #else
 589         DRV_LOG(WARNING,
 590                 "tunnel offloading disabled due to old OFED/rdma-core version");
 591 #endif
 592         config.tunnel_en = tunnel_en;
 593 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 594         mpls_en = ((dv_attr.tunnel_offloads_caps &
 595                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 596                    (dv_attr.tunnel_offloads_caps &
 597                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 598         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 599                 mpls_en ? "" : "not ");
 600 #else
 601         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 602                 " old OFED/rdma-core version or firmware configuration");
 603 #endif
 604         config.mpls_en = mpls_en;
 605         /* Check port status. */
 606         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 607         if (err) {
 608                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 609                 goto error;
 610         }
 611         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 612                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 613                 err = EINVAL;
 614                 goto error;
 615         }
 616         if (port_attr.state != IBV_PORT_ACTIVE)
 617                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 618                         mlx5_glue->port_state_str(port_attr.state),
 619                         port_attr.state);
 620         /* Allocate private eth device data. */
 621         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
 622                            sizeof(*priv),
 623                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 624         if (priv == NULL) {
 625                 DRV_LOG(ERR, "priv allocation failure");
 626                 err = ENOMEM;
 627                 goto error;
 628         }
 629         priv->sh = sh;
 630         priv->dev_port = spawn->phys_port;
 631         priv->pci_dev = spawn->pci_dev;
 632         priv->mtu = RTE_ETHER_MTU;
 633         priv->mp_id.port_id = port_id;
 634         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 635         /* Some internal functions rely on Netlink sockets, open them now. */
 636         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 637         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 638         priv->representor = !!switch_info->representor;
 639         priv->master = !!switch_info->master;
 640         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 641         priv->vport_meta_tag = 0;
 642         priv->vport_meta_mask = 0;
 643         priv->pf_bond = spawn->pf_bond;
 644 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 645         /*
 646          * The DevX port query API is implemented. E-Switch may use
 647          * either vport or reg_c[0] metadata register to match on
 648          * vport index. The engaged part of metadata register is
 649          * defined by mask.
 650          */
 651         if (switch_info->representor || switch_info->master) {
 652                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 653                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 654                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 655                                                  &devx_port);
 656                 if (err) {
 657                         DRV_LOG(WARNING,
 658                                 "can't query devx port %d on device %s",
 659                                 spawn->phys_port,
 660                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 661                         devx_port.comp_mask = 0;
 662                 }
 663         }
 664         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 665                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 666                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 667                 if (!priv->vport_meta_mask) {
 668                         DRV_LOG(ERR, "vport zero mask for port %d"
 669                                      " on bonding device %s",
 670                                      spawn->phys_port,
 671                                      mlx5_os_get_dev_device_name
 672                                                         (spawn->phys_dev));
 673                         err = ENOTSUP;
 674                         goto error;
 675                 }
 676                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 677                         DRV_LOG(ERR, "invalid vport tag for port %d"
 678                                      " on bonding device %s",
 679                                      spawn->phys_port,
 680                                      mlx5_os_get_dev_device_name
 681                                                         (spawn->phys_dev));
 682                         err = ENOTSUP;
 683                         goto error;
 684                 }
 685         }
 686         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 687                 priv->vport_id = devx_port.vport_num;
 688         } else if (spawn->pf_bond >= 0) {
 689                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 690                              " on bonding device %s",
 691                              spawn->phys_port,
 692                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 693                 err = ENOTSUP;
 694                 goto error;
 695         } else {
 696                 /* Suppose vport index in compatible way. */
 697                 priv->vport_id = switch_info->representor ?
 698                                  switch_info->port_name + 1 : -1;
 699         }
 700 #else
 701         /*
 702          * Kernel/rdma_core support single E-Switch per PF configurations
 703          * only and vport_id field contains the vport index for
 704          * associated VF, which is deduced from representor port name.
 705          * For example, let's have the IB device port 10, it has
 706          * attached network device eth0, which has port name attribute
 707          * pf0vf2, we can deduce the VF number as 2, and set vport index
 708          * as 3 (2+1). This assigning schema should be changed if the
 709          * multiple E-Switch instances per PF configurations or/and PCI
 710          * subfunctions are added.
 711          */
 712         priv->vport_id = switch_info->representor ?
 713                          switch_info->port_name + 1 : -1;
 714 #endif
 715         /* representor_id field keeps the unmodified VF index. */
 716         priv->representor_id = switch_info->representor ?
 717                                switch_info->port_name : -1;
 718         /*
 719          * Look for sibling devices in order to reuse their switch domain
 720          * if any, otherwise allocate one.
 721          */
 722         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 723                 const struct mlx5_priv *opriv =
 724                         rte_eth_devices[port_id].data->dev_private;
 725
 726                 if (!opriv ||
 727                     opriv->sh != priv->sh ||
 728                         opriv->domain_id ==
 729                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 730                         continue;
 731                 priv->domain_id = opriv->domain_id;
 732                 break;
 733         }
 734         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 735                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 736                 if (err) {
 737                         err = rte_errno;
 738                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 739                                 strerror(rte_errno));
 740                         goto error;
 741                 }
 742                 own_domain_id = 1;
 743         }
 744         /* Override some values set by hardware configuration. */
 745         mlx5_args(&config, dpdk_dev->devargs);
 746         err = mlx5_dev_check_sibling_config(priv, &config);
 747         if (err)
 748                 goto error;
 749         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 750                             IBV_DEVICE_RAW_IP_CSUM);
 751         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 752                 (config.hw_csum ? "" : "not "));
 753 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 754         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 755         DRV_LOG(DEBUG, "counters are not supported");
 756 #endif
 757 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 758         if (config.dv_flow_en) {
 759                 DRV_LOG(WARNING, "DV flow is not supported");
 760                 config.dv_flow_en = 0;
 761         }
 762 #endif
 763         config.ind_table_max_size =
 764                 sh->device_attr.max_rwq_indirection_table_size;
 765         /*
 766          * Remove this check once DPDK supports larger/variable
 767          * indirection tables.
 768          */
 769         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 770                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 771         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 772                 config.ind_table_max_size);
 773         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 774                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 775         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 776                 (config.hw_vlan_strip ? "" : "not "));
 777         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 778                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 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         /*
 961          * If HW has bug working with tunnel packet decapsulation and
 962          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
 963          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
 964          */
 965         if (config.hca_attr.scatter_fcs_w_decap_disable && config.decap_en)
 966                 config.hw_fcs_strip = 0;
 967         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 968                 (config.hw_fcs_strip ? "" : "not "));
 969         if (config.mprq.enabled && mprq) {
 970                 if (config.mprq.stride_num_n &&
 971                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 972                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 973                         config.mprq.stride_num_n =
 974                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 975                                                 mprq_min_stride_num_n),
 976                                         mprq_max_stride_num_n);
 977                         DRV_LOG(WARNING,
 978                                 "the number of strides"
 979                                 " for Multi-Packet RQ is out of range,"
 980                                 " setting default value (%u)",
 981                                 1 << config.mprq.stride_num_n);
 982                 }
 983                 if (config.mprq.stride_size_n &&
 984                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 985                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 986                         config.mprq.stride_size_n =
 987                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 988                                                 mprq_min_stride_size_n),
 989                                         mprq_max_stride_size_n);
 990                         DRV_LOG(WARNING,
 991                                 "the size of a stride"
 992                                 " for Multi-Packet RQ is out of range,"
 993                                 " setting default value (%u)",
 994                                 1 << config.mprq.stride_size_n);
 995                 }
 996                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
 997                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
 998         } else if (config.mprq.enabled && !mprq) {
 999                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1000                 config.mprq.enabled = 0;
1001         }
1002         if (config.max_dump_files_num == 0)
1003                 config.max_dump_files_num = 128;
1004         eth_dev = rte_eth_dev_allocate(name);
1005         if (eth_dev == NULL) {
1006                 DRV_LOG(ERR, "can not allocate rte ethdev");
1007                 err = ENOMEM;
1008                 goto error;
1009         }
1010         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1011         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1012         if (priv->representor) {
1013                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1014                 eth_dev->data->representor_id = priv->representor_id;
1015         }
1016         /*
1017          * Store associated network device interface index. This index
1018          * is permanent throughout the lifetime of device. So, we may store
1019          * the ifindex here and use the cached value further.
1020          */
1021         MLX5_ASSERT(spawn->ifindex);
1022         priv->if_index = spawn->ifindex;
1023         eth_dev->data->dev_private = priv;
1024         priv->dev_data = eth_dev->data;
1025         eth_dev->data->mac_addrs = priv->mac;
1026         eth_dev->device = dpdk_dev;
1027         /* Configure the first MAC address by default. */
1028         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1029                 DRV_LOG(ERR,
1030                         "port %u cannot get MAC address, is mlx5_en"
1031                         " loaded? (errno: %s)",
1032                         eth_dev->data->port_id, strerror(rte_errno));
1033                 err = ENODEV;
1034                 goto error;
1035         }
1036         DRV_LOG(INFO,
1037                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1038                 eth_dev->data->port_id,
1039                 mac.addr_bytes[0], mac.addr_bytes[1],
1040                 mac.addr_bytes[2], mac.addr_bytes[3],
1041                 mac.addr_bytes[4], mac.addr_bytes[5]);
1042 #ifdef RTE_LIBRTE_MLX5_DEBUG
1043         {
1044                 char ifname[IF_NAMESIZE];
1045
1046                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1047                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1048                                 eth_dev->data->port_id, ifname);
1049                 else
1050                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1051                                 eth_dev->data->port_id);
1052         }
1053 #endif
1054         /* Get actual MTU if possible. */
1055         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1056         if (err) {
1057                 err = rte_errno;
1058                 goto error;
1059         }
1060         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1061                 priv->mtu);
1062         /* Initialize burst functions to prevent crashes before link-up. */
1063         eth_dev->rx_pkt_burst = removed_rx_burst;
1064         eth_dev->tx_pkt_burst = removed_tx_burst;
1065         eth_dev->dev_ops = &mlx5_os_dev_ops;
1066         /* Register MAC address. */
1067         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1068         if (config.vf && config.vf_nl_en)
1069                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1070                                       mlx5_ifindex(eth_dev),
1071                                       eth_dev->data->mac_addrs,
1072                                       MLX5_MAX_MAC_ADDRESSES);
1073         priv->flows = 0;
1074         priv->ctrl_flows = 0;
1075         TAILQ_INIT(&priv->flow_meters);
1076         TAILQ_INIT(&priv->flow_meter_profiles);
1077         /* Hint libmlx5 to use PMD allocator for data plane resources */
1078         mlx5_glue->dv_set_context_attr(sh->ctx,
1079                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1080                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1081                                 .alloc = &mlx5_alloc_verbs_buf,
1082                                 .free = &mlx5_free_verbs_buf,
1083                                 .data = priv,
1084                         }));
1085         /* Bring Ethernet device up. */
1086         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1087                 eth_dev->data->port_id);
1088         mlx5_set_link_up(eth_dev);
1089         /*
1090          * Even though the interrupt handler is not installed yet,
1091          * interrupts will still trigger on the async_fd from
1092          * Verbs context returned by ibv_open_device().
1093          */
1094         mlx5_link_update(eth_dev, 0);
1095 #ifdef HAVE_MLX5DV_DR_ESWITCH
1096         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1097               (switch_info->representor || switch_info->master)))
1098                 config.dv_esw_en = 0;
1099 #else
1100         config.dv_esw_en = 0;
1101 #endif
1102         /* Detect minimal data bytes to inline. */
1103         mlx5_set_min_inline(spawn, &config);
1104         /* Store device configuration on private structure. */
1105         priv->config = config;
1106         /* Create context for virtual machine VLAN workaround. */
1107         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1108         if (config.dv_flow_en) {
1109                 err = mlx5_alloc_shared_dr(priv);
1110                 if (err)
1111                         goto error;
1112                 /*
1113                  * RSS id is shared with meter flow id. Meter flow id can only
1114                  * use the 24 MSB of the register.
1115                  */
1116                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1117                                      MLX5_MTR_COLOR_BITS);
1118                 if (!priv->qrss_id_pool) {
1119                         DRV_LOG(ERR, "can't create flow id pool");
1120                         err = ENOMEM;
1121                         goto error;
1122                 }
1123         }
1124         /* Supported Verbs flow priority number detection. */
1125         err = mlx5_flow_discover_priorities(eth_dev);
1126         if (err < 0) {
1127                 err = -err;
1128                 goto error;
1129         }
1130         priv->config.flow_prio = err;
1131         if (!priv->config.dv_esw_en &&
1132             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1133                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1134                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1135                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1136         }
1137         mlx5_set_metadata_mask(eth_dev);
1138         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1139             !priv->sh->dv_regc0_mask) {
1140                 DRV_LOG(ERR, "metadata mode %u is not supported "
1141                              "(no metadata reg_c[0] is available)",
1142                              priv->config.dv_xmeta_en);
1143                         err = ENOTSUP;
1144                         goto error;
1145         }
1146         /*
1147          * Allocate the buffer for flow creating, just once.
1148          * The allocation must be done before any flow creating.
1149          */
1150         mlx5_flow_alloc_intermediate(eth_dev);
1151         /* Query availability of metadata reg_c's. */
1152         err = mlx5_flow_discover_mreg_c(eth_dev);
1153         if (err < 0) {
1154                 err = -err;
1155                 goto error;
1156         }
1157         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1158                 DRV_LOG(DEBUG,
1159                         "port %u extensive metadata register is not supported",
1160                         eth_dev->data->port_id);
1161                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1162                         DRV_LOG(ERR, "metadata mode %u is not supported "
1163                                      "(no metadata registers available)",
1164                                      priv->config.dv_xmeta_en);
1165                         err = ENOTSUP;
1166                         goto error;
1167                 }
1168         }
1169         if (priv->config.dv_flow_en &&
1170             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1171             mlx5_flow_ext_mreg_supported(eth_dev) &&
1172             priv->sh->dv_regc0_mask) {
1173                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1174                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1175                 if (!priv->mreg_cp_tbl) {
1176                         err = ENOMEM;
1177                         goto error;
1178                 }
1179         }
1180         return eth_dev;
1181 error:
1182         if (priv) {
1183                 if (priv->mreg_cp_tbl)
1184                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1185                 if (priv->sh)
1186                         mlx5_os_free_shared_dr(priv);
1187                 if (priv->nl_socket_route >= 0)
1188                         close(priv->nl_socket_route);
1189                 if (priv->nl_socket_rdma >= 0)
1190                         close(priv->nl_socket_rdma);
1191                 if (priv->vmwa_context)
1192                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1193                 if (priv->qrss_id_pool)
1194                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1195                 if (own_domain_id)
1196                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1197                 mlx5_free(priv);
1198                 if (eth_dev != NULL)
1199                         eth_dev->data->dev_private = NULL;
1200         }
1201         if (eth_dev != NULL) {
1202                 /* mac_addrs must not be freed alone because part of
1203                  * dev_private
1204                  **/
1205                 eth_dev->data->mac_addrs = NULL;
1206                 rte_eth_dev_release_port(eth_dev);
1207         }
1208         if (sh)
1209                 mlx5_free_shared_dev_ctx(sh);
1210         MLX5_ASSERT(err > 0);
1211         rte_errno = err;
1212         return NULL;
1213 }
1214
1215 /**
1216  * Comparison callback to sort device data.
1217  *
1218  * This is meant to be used with qsort().
1219  *
1220  * @param a[in]
1221  *   Pointer to pointer to first data object.
1222  * @param b[in]
1223  *   Pointer to pointer to second data object.
1224  *
1225  * @return
1226  *   0 if both objects are equal, less than 0 if the first argument is less
1227  *   than the second, greater than 0 otherwise.
1228  */
1229 static int
1230 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1231 {
1232         const struct mlx5_switch_info *si_a =
1233                 &((const struct mlx5_dev_spawn_data *)a)->info;
1234         const struct mlx5_switch_info *si_b =
1235                 &((const struct mlx5_dev_spawn_data *)b)->info;
1236         int ret;
1237
1238         /* Master device first. */
1239         ret = si_b->master - si_a->master;
1240         if (ret)
1241                 return ret;
1242         /* Then representor devices. */
1243         ret = si_b->representor - si_a->representor;
1244         if (ret)
1245                 return ret;
1246         /* Unidentified devices come last in no specific order. */
1247         if (!si_a->representor)
1248                 return 0;
1249         /* Order representors by name. */
1250         return si_a->port_name - si_b->port_name;
1251 }
1252
1253 /**
1254  * Match PCI information for possible slaves of bonding device.
1255  *
1256  * @param[in] ibv_dev
1257  *   Pointer to Infiniband device structure.
1258  * @param[in] pci_dev
1259  *   Pointer to PCI device structure to match PCI address.
1260  * @param[in] nl_rdma
1261  *   Netlink RDMA group socket handle.
1262  *
1263  * @return
1264  *   negative value if no bonding device found, otherwise
1265  *   positive index of slave PF in bonding.
1266  */
1267 static int
1268 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1269                            const struct rte_pci_device *pci_dev,
1270                            int nl_rdma)
1271 {
1272         char ifname[IF_NAMESIZE + 1];
1273         unsigned int ifindex;
1274         unsigned int np, i;
1275         FILE *file = NULL;
1276         int pf = -1;
1277
1278         /*
1279          * Try to get master device name. If something goes
1280          * wrong suppose the lack of kernel support and no
1281          * bonding devices.
1282          */
1283         if (nl_rdma < 0)
1284                 return -1;
1285         if (!strstr(ibv_dev->name, "bond"))
1286                 return -1;
1287         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1288         if (!np)
1289                 return -1;
1290         /*
1291          * The Master device might not be on the predefined
1292          * port (not on port index 1, it is not garanted),
1293          * we have to scan all Infiniband device port and
1294          * find master.
1295          */
1296         for (i = 1; i <= np; ++i) {
1297                 /* Check whether Infiniband port is populated. */
1298                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1299                 if (!ifindex)
1300                         continue;
1301                 if (!if_indextoname(ifindex, ifname))
1302                         continue;
1303                 /* Try to read bonding slave names from sysfs. */
1304                 MKSTR(slaves,
1305                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1306                 file = fopen(slaves, "r");
1307                 if (file)
1308                         break;
1309         }
1310         if (!file)
1311                 return -1;
1312         /* Use safe format to check maximal buffer length. */
1313         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1314         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1315                 char tmp_str[IF_NAMESIZE + 32];
1316                 struct rte_pci_addr pci_addr;
1317                 struct mlx5_switch_info info;
1318
1319                 /* Process slave interface names in the loop. */
1320                 snprintf(tmp_str, sizeof(tmp_str),
1321                          "/sys/class/net/%s", ifname);
1322                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1323                         DRV_LOG(WARNING, "can not get PCI address"
1324                                          " for netdev \"%s\"", ifname);
1325                         continue;
1326                 }
1327                 if (pci_dev->addr.domain != pci_addr.domain ||
1328                     pci_dev->addr.bus != pci_addr.bus ||
1329                     pci_dev->addr.devid != pci_addr.devid ||
1330                     pci_dev->addr.function != pci_addr.function)
1331                         continue;
1332                 /* Slave interface PCI address match found. */
1333                 fclose(file);
1334                 snprintf(tmp_str, sizeof(tmp_str),
1335                          "/sys/class/net/%s/phys_port_name", ifname);
1336                 file = fopen(tmp_str, "rb");
1337                 if (!file)
1338                         break;
1339                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1340                 if (fscanf(file, "%32s", tmp_str) == 1)
1341                         mlx5_translate_port_name(tmp_str, &info);
1342                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1343                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1344                         pf = info.port_name;
1345                 break;
1346         }
1347         if (file)
1348                 fclose(file);
1349         return pf;
1350 }
1351
1352 /**
1353  * DPDK callback to register a PCI device.
1354  *
1355  * This function spawns Ethernet devices out of a given PCI device.
1356  *
1357  * @param[in] pci_drv
1358  *   PCI driver structure (mlx5_driver).
1359  * @param[in] pci_dev
1360  *   PCI device information.
1361  *
1362  * @return
1363  *   0 on success, a negative errno value otherwise and rte_errno is set.
1364  */
1365 int
1366 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1367                   struct rte_pci_device *pci_dev)
1368 {
1369         struct ibv_device **ibv_list;
1370         /*
1371          * Number of found IB Devices matching with requested PCI BDF.
1372          * nd != 1 means there are multiple IB devices over the same
1373          * PCI device and we have representors and master.
1374          */
1375         unsigned int nd = 0;
1376         /*
1377          * Number of found IB device Ports. nd = 1 and np = 1..n means
1378          * we have the single multiport IB device, and there may be
1379          * representors attached to some of found ports.
1380          */
1381         unsigned int np = 0;
1382         /*
1383          * Number of DPDK ethernet devices to Spawn - either over
1384          * multiple IB devices or multiple ports of single IB device.
1385          * Actually this is the number of iterations to spawn.
1386          */
1387         unsigned int ns = 0;
1388         /*
1389          * Bonding device
1390          *   < 0 - no bonding device (single one)
1391          *  >= 0 - bonding device (value is slave PF index)
1392          */
1393         int bd = -1;
1394         struct mlx5_dev_spawn_data *list = NULL;
1395         struct mlx5_dev_config dev_config;
1396         int ret;
1397
1398         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1399                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1400                         " driver.");
1401                 return 1;
1402         }
1403         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1404                 mlx5_pmd_socket_init();
1405         ret = mlx5_init_once();
1406         if (ret) {
1407                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1408                         strerror(rte_errno));
1409                 return -rte_errno;
1410         }
1411         MLX5_ASSERT(pci_drv == &mlx5_driver);
1412         errno = 0;
1413         ibv_list = mlx5_glue->get_device_list(&ret);
1414         if (!ibv_list) {
1415                 rte_errno = errno ? errno : ENOSYS;
1416                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1417                 return -rte_errno;
1418         }
1419         /*
1420          * First scan the list of all Infiniband devices to find
1421          * matching ones, gathering into the list.
1422          */
1423         struct ibv_device *ibv_match[ret + 1];
1424         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1425         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1426         unsigned int i;
1427
1428         while (ret-- > 0) {
1429                 struct rte_pci_addr pci_addr;
1430
1431                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1432                 bd = mlx5_device_bond_pci_match
1433                                 (ibv_list[ret], pci_dev, nl_rdma);
1434                 if (bd >= 0) {
1435                         /*
1436                          * Bonding device detected. Only one match is allowed,
1437                          * the bonding is supported over multi-port IB device,
1438                          * there should be no matches on representor PCI
1439                          * functions or non VF LAG bonding devices with
1440                          * specified address.
1441                          */
1442                         if (nd) {
1443                                 DRV_LOG(ERR,
1444                                         "multiple PCI match on bonding device"
1445                                         "\"%s\" found", ibv_list[ret]->name);
1446                                 rte_errno = ENOENT;
1447                                 ret = -rte_errno;
1448                                 goto exit;
1449                         }
1450                         DRV_LOG(INFO, "PCI information matches for"
1451                                       " slave %d bonding device \"%s\"",
1452                                       bd, ibv_list[ret]->name);
1453                         ibv_match[nd++] = ibv_list[ret];
1454                         break;
1455                 }
1456                 if (mlx5_dev_to_pci_addr
1457                         (ibv_list[ret]->ibdev_path, &pci_addr))
1458                         continue;
1459                 if (pci_dev->addr.domain != pci_addr.domain ||
1460                     pci_dev->addr.bus != pci_addr.bus ||
1461                     pci_dev->addr.devid != pci_addr.devid ||
1462                     pci_dev->addr.function != pci_addr.function)
1463                         continue;
1464                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1465                         ibv_list[ret]->name);
1466                 ibv_match[nd++] = ibv_list[ret];
1467         }
1468         ibv_match[nd] = NULL;
1469         if (!nd) {
1470                 /* No device matches, just complain and bail out. */
1471                 DRV_LOG(WARNING,
1472                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1473                         " are kernel drivers loaded?",
1474                         pci_dev->addr.domain, pci_dev->addr.bus,
1475                         pci_dev->addr.devid, pci_dev->addr.function);
1476                 rte_errno = ENOENT;
1477                 ret = -rte_errno;
1478                 goto exit;
1479         }
1480         if (nd == 1) {
1481                 /*
1482                  * Found single matching device may have multiple ports.
1483                  * Each port may be representor, we have to check the port
1484                  * number and check the representors existence.
1485                  */
1486                 if (nl_rdma >= 0)
1487                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1488                 if (!np)
1489                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1490                                          " ports number", ibv_match[0]->name);
1491                 if (bd >= 0 && !np) {
1492                         DRV_LOG(ERR, "can not get ports"
1493                                      " for bonding device");
1494                         rte_errno = ENOENT;
1495                         ret = -rte_errno;
1496                         goto exit;
1497                 }
1498         }
1499 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1500         if (bd >= 0) {
1501                 /*
1502                  * This may happen if there is VF LAG kernel support and
1503                  * application is compiled with older rdma_core library.
1504                  */
1505                 DRV_LOG(ERR,
1506                         "No kernel/verbs support for VF LAG bonding found.");
1507                 rte_errno = ENOTSUP;
1508                 ret = -rte_errno;
1509                 goto exit;
1510         }
1511 #endif
1512         /*
1513          * Now we can determine the maximal
1514          * amount of devices to be spawned.
1515          */
1516         list = mlx5_malloc(MLX5_MEM_ZERO,
1517                            sizeof(struct mlx5_dev_spawn_data) *
1518                            (np ? np : nd),
1519                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1520         if (!list) {
1521                 DRV_LOG(ERR, "spawn data array allocation failure");
1522                 rte_errno = ENOMEM;
1523                 ret = -rte_errno;
1524                 goto exit;
1525         }
1526         if (bd >= 0 || np > 1) {
1527                 /*
1528                  * Single IB device with multiple ports found,
1529                  * it may be E-Switch master device and representors.
1530                  * We have to perform identification through the ports.
1531                  */
1532                 MLX5_ASSERT(nl_rdma >= 0);
1533                 MLX5_ASSERT(ns == 0);
1534                 MLX5_ASSERT(nd == 1);
1535                 MLX5_ASSERT(np);
1536                 for (i = 1; i <= np; ++i) {
1537                         list[ns].max_port = np;
1538                         list[ns].phys_port = i;
1539                         list[ns].phys_dev = ibv_match[0];
1540                         list[ns].eth_dev = NULL;
1541                         list[ns].pci_dev = pci_dev;
1542                         list[ns].pf_bond = bd;
1543                         list[ns].ifindex = mlx5_nl_ifindex
1544                                 (nl_rdma,
1545                                 mlx5_os_get_dev_device_name
1546                                                 (list[ns].phys_dev), i);
1547                         if (!list[ns].ifindex) {
1548                                 /*
1549                                  * No network interface index found for the
1550                                  * specified port, it means there is no
1551                                  * representor on this port. It's OK,
1552                                  * there can be disabled ports, for example
1553                                  * if sriov_numvfs < sriov_totalvfs.
1554                                  */
1555                                 continue;
1556                         }
1557                         ret = -1;
1558                         if (nl_route >= 0)
1559                                 ret = mlx5_nl_switch_info
1560                                                (nl_route,
1561                                                 list[ns].ifindex,
1562                                                 &list[ns].info);
1563                         if (ret || (!list[ns].info.representor &&
1564                                     !list[ns].info.master)) {
1565                                 /*
1566                                  * We failed to recognize representors with
1567                                  * Netlink, let's try to perform the task
1568                                  * with sysfs.
1569                                  */
1570                                 ret =  mlx5_sysfs_switch_info
1571                                                 (list[ns].ifindex,
1572                                                  &list[ns].info);
1573                         }
1574                         if (!ret && bd >= 0) {
1575                                 switch (list[ns].info.name_type) {
1576                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1577                                         if (list[ns].info.port_name == bd)
1578                                                 ns++;
1579                                         break;
1580                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1581                                         /* Fallthrough */
1582                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1583                                         if (list[ns].info.pf_num == bd)
1584                                                 ns++;
1585                                         break;
1586                                 default:
1587                                         break;
1588                                 }
1589                                 continue;
1590                         }
1591                         if (!ret && (list[ns].info.representor ^
1592                                      list[ns].info.master))
1593                                 ns++;
1594                 }
1595                 if (!ns) {
1596                         DRV_LOG(ERR,
1597                                 "unable to recognize master/representors"
1598                                 " on the IB device with multiple ports");
1599                         rte_errno = ENOENT;
1600                         ret = -rte_errno;
1601                         goto exit;
1602                 }
1603         } else {
1604                 /*
1605                  * The existence of several matching entries (nd > 1) means
1606                  * port representors have been instantiated. No existing Verbs
1607                  * call nor sysfs entries can tell them apart, this can only
1608                  * be done through Netlink calls assuming kernel drivers are
1609                  * recent enough to support them.
1610                  *
1611                  * In the event of identification failure through Netlink,
1612                  * try again through sysfs, then:
1613                  *
1614                  * 1. A single IB device matches (nd == 1) with single
1615                  *    port (np=0/1) and is not a representor, assume
1616                  *    no switch support.
1617                  *
1618                  * 2. Otherwise no safe assumptions can be made;
1619                  *    complain louder and bail out.
1620                  */
1621                 for (i = 0; i != nd; ++i) {
1622                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1623                         list[ns].max_port = 1;
1624                         list[ns].phys_port = 1;
1625                         list[ns].phys_dev = ibv_match[i];
1626                         list[ns].eth_dev = NULL;
1627                         list[ns].pci_dev = pci_dev;
1628                         list[ns].pf_bond = -1;
1629                         list[ns].ifindex = 0;
1630                         if (nl_rdma >= 0)
1631                                 list[ns].ifindex = mlx5_nl_ifindex
1632                                 (nl_rdma,
1633                                 mlx5_os_get_dev_device_name
1634                                                 (list[ns].phys_dev), 1);
1635                         if (!list[ns].ifindex) {
1636                                 char ifname[IF_NAMESIZE];
1637
1638                                 /*
1639                                  * Netlink failed, it may happen with old
1640                                  * ib_core kernel driver (before 4.16).
1641                                  * We can assume there is old driver because
1642                                  * here we are processing single ports IB
1643                                  * devices. Let's try sysfs to retrieve
1644                                  * the ifindex. The method works for
1645                                  * master device only.
1646                                  */
1647                                 if (nd > 1) {
1648                                         /*
1649                                          * Multiple devices found, assume
1650                                          * representors, can not distinguish
1651                                          * master/representor and retrieve
1652                                          * ifindex via sysfs.
1653                                          */
1654                                         continue;
1655                                 }
1656                                 ret = mlx5_get_ifname_sysfs
1657                                         (ibv_match[i]->ibdev_path, ifname);
1658                                 if (!ret)
1659                                         list[ns].ifindex =
1660                                                 if_nametoindex(ifname);
1661                                 if (!list[ns].ifindex) {
1662                                         /*
1663                                          * No network interface index found
1664                                          * for the specified device, it means
1665                                          * there it is neither representor
1666                                          * nor master.
1667                                          */
1668                                         continue;
1669                                 }
1670                         }
1671                         ret = -1;
1672                         if (nl_route >= 0)
1673                                 ret = mlx5_nl_switch_info
1674                                                (nl_route,
1675                                                 list[ns].ifindex,
1676                                                 &list[ns].info);
1677                         if (ret || (!list[ns].info.representor &&
1678                                     !list[ns].info.master)) {
1679                                 /*
1680                                  * We failed to recognize representors with
1681                                  * Netlink, let's try to perform the task
1682                                  * with sysfs.
1683                                  */
1684                                 ret =  mlx5_sysfs_switch_info
1685                                                 (list[ns].ifindex,
1686                                                  &list[ns].info);
1687                         }
1688                         if (!ret && (list[ns].info.representor ^
1689                                      list[ns].info.master)) {
1690                                 ns++;
1691                         } else if ((nd == 1) &&
1692                                    !list[ns].info.representor &&
1693                                    !list[ns].info.master) {
1694                                 /*
1695                                  * Single IB device with
1696                                  * one physical port and
1697                                  * attached network device.
1698                                  * May be SRIOV is not enabled
1699                                  * or there is no representors.
1700                                  */
1701                                 DRV_LOG(INFO, "no E-Switch support detected");
1702                                 ns++;
1703                                 break;
1704                         }
1705                 }
1706                 if (!ns) {
1707                         DRV_LOG(ERR,
1708                                 "unable to recognize master/representors"
1709                                 " on the multiple IB devices");
1710                         rte_errno = ENOENT;
1711                         ret = -rte_errno;
1712                         goto exit;
1713                 }
1714         }
1715         MLX5_ASSERT(ns);
1716         /*
1717          * Sort list to probe devices in natural order for users convenience
1718          * (i.e. master first, then representors from lowest to highest ID).
1719          */
1720         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1721         /* Default configuration. */
1722         dev_config = (struct mlx5_dev_config){
1723                 .hw_padding = 0,
1724                 .mps = MLX5_ARG_UNSET,
1725                 .dbnc = MLX5_ARG_UNSET,
1726                 .rx_vec_en = 1,
1727                 .txq_inline_max = MLX5_ARG_UNSET,
1728                 .txq_inline_min = MLX5_ARG_UNSET,
1729                 .txq_inline_mpw = MLX5_ARG_UNSET,
1730                 .txqs_inline = MLX5_ARG_UNSET,
1731                 .vf_nl_en = 1,
1732                 .mr_ext_memseg_en = 1,
1733                 .mprq = {
1734                         .enabled = 0, /* Disabled by default. */
1735                         .stride_num_n = 0,
1736                         .stride_size_n = 0,
1737                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1738                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1739                 },
1740                 .dv_esw_en = 1,
1741                 .dv_flow_en = 1,
1742                 .decap_en = 1,
1743                 .log_hp_size = MLX5_ARG_UNSET,
1744         };
1745         /* Device specific configuration. */
1746         switch (pci_dev->id.device_id) {
1747         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1748         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1749         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1750         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1751         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1752         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1753         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1754                 dev_config.vf = 1;
1755                 break;
1756         default:
1757                 break;
1758         }
1759         for (i = 0; i != ns; ++i) {
1760                 uint32_t restore;
1761
1762                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1763                                                  &list[i],
1764                                                  dev_config);
1765                 if (!list[i].eth_dev) {
1766                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1767                                 break;
1768                         /* Device is disabled or already spawned. Ignore it. */
1769                         continue;
1770                 }
1771                 restore = list[i].eth_dev->data->dev_flags;
1772                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1773                 /* Restore non-PCI flags cleared by the above call. */
1774                 list[i].eth_dev->data->dev_flags |= restore;
1775                 rte_eth_dev_probing_finish(list[i].eth_dev);
1776         }
1777         if (i != ns) {
1778                 DRV_LOG(ERR,
1779                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1780                         " encountering an error: %s",
1781                         pci_dev->addr.domain, pci_dev->addr.bus,
1782                         pci_dev->addr.devid, pci_dev->addr.function,
1783                         strerror(rte_errno));
1784                 ret = -rte_errno;
1785                 /* Roll back. */
1786                 while (i--) {
1787                         if (!list[i].eth_dev)
1788                                 continue;
1789                         mlx5_dev_close(list[i].eth_dev);
1790                         /* mac_addrs must not be freed because in dev_private */
1791                         list[i].eth_dev->data->mac_addrs = NULL;
1792                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1793                 }
1794                 /* Restore original error. */
1795                 rte_errno = -ret;
1796         } else {
1797                 ret = 0;
1798         }
1799 exit:
1800         /*
1801          * Do the routine cleanup:
1802          * - close opened Netlink sockets
1803          * - free allocated spawn data array
1804          * - free the Infiniband device list
1805          */
1806         if (nl_rdma >= 0)
1807                 close(nl_rdma);
1808         if (nl_route >= 0)
1809                 close(nl_route);
1810         if (list)
1811                 mlx5_free(list);
1812         MLX5_ASSERT(ibv_list);
1813         mlx5_glue->free_device_list(ibv_list);
1814         return ret;
1815 }
1816
1817 static int
1818 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1819 {
1820         char *env;
1821         int value;
1822
1823         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1824         /* Get environment variable to store. */
1825         env = getenv(MLX5_SHUT_UP_BF);
1826         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1827         if (config->dbnc == MLX5_ARG_UNSET)
1828                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1829         else
1830                 setenv(MLX5_SHUT_UP_BF,
1831                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1832         return value;
1833 }
1834
1835 static void
1836 mlx5_restore_doorbell_mapping_env(int value)
1837 {
1838         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1839         /* Restore the original environment variable state. */
1840         if (value == MLX5_ARG_UNSET)
1841                 unsetenv(MLX5_SHUT_UP_BF);
1842         else
1843                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1844 }
1845
1846 /**
1847  * Extract pdn of PD object using DV API.
1848  *
1849  * @param[in] pd
1850  *   Pointer to the verbs PD object.
1851  * @param[out] pdn
1852  *   Pointer to the PD object number variable.
1853  *
1854  * @return
1855  *   0 on success, error value otherwise.
1856  */
1857 int
1858 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1859 {
1860 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1861         struct mlx5dv_obj obj;
1862         struct mlx5dv_pd pd_info;
1863         int ret = 0;
1864
1865         obj.pd.in = pd;
1866         obj.pd.out = &pd_info;
1867         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1868         if (ret) {
1869                 DRV_LOG(DEBUG, "Fail to get PD object info");
1870                 return ret;
1871         }
1872         *pdn = pd_info.pdn;
1873         return 0;
1874 #else
1875         (void)pd;
1876         (void)pdn;
1877         return -ENOTSUP;
1878 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1879 }
1880
1881 /**
1882  * Function API to open IB device.
1883  *
1884  * This function calls the Linux glue APIs to open a device.
1885  *
1886  * @param[in] spawn
1887  *   Pointer to the IB device attributes (name, port, etc).
1888  * @param[out] config
1889  *   Pointer to device configuration structure.
1890  * @param[out] sh
1891  *   Pointer to shared context structure.
1892  *
1893  * @return
1894  *   0 on success, a positive error value otherwise.
1895  */
1896 int
1897 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1898                      const struct mlx5_dev_config *config,
1899                      struct mlx5_dev_ctx_shared *sh)
1900 {
1901         int dbmap_env;
1902         int err = 0;
1903
1904         sh->numa_node = spawn->pci_dev->device.numa_node;
1905         pthread_mutex_init(&sh->txpp.mutex, NULL);
1906         /*
1907          * Configure environment variable "MLX5_BF_SHUT_UP"
1908          * before the device creation. The rdma_core library
1909          * checks the variable at device creation and
1910          * stores the result internally.
1911          */
1912         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1913         /* Try to open IB device with DV first, then usual Verbs. */
1914         errno = 0;
1915         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1916         if (sh->ctx) {
1917                 sh->devx = 1;
1918                 DRV_LOG(DEBUG, "DevX is supported");
1919                 /* The device is created, no need for environment. */
1920                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1921         } else {
1922                 /* The environment variable is still configured. */
1923                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1924                 err = errno ? errno : ENODEV;
1925                 /*
1926                  * The environment variable is not needed anymore,
1927                  * all device creation attempts are completed.
1928                  */
1929                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1930                 if (!sh->ctx)
1931                         return err;
1932                 DRV_LOG(DEBUG, "DevX is NOT supported");
1933                 err = 0;
1934         }
1935         return err;
1936 }
1937
1938 /**
1939  * Install shared asynchronous device events handler.
1940  * This function is implemented to support event sharing
1941  * between multiple ports of single IB device.
1942  *
1943  * @param sh
1944  *   Pointer to mlx5_dev_ctx_shared object.
1945  */
1946 void
1947 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1948 {
1949         int ret;
1950         int flags;
1951
1952         sh->intr_handle.fd = -1;
1953         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1954         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1955                     F_SETFL, flags | O_NONBLOCK);
1956         if (ret) {
1957                 DRV_LOG(INFO, "failed to change file descriptor async event"
1958                         " queue");
1959         } else {
1960                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1961                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1962                 if (rte_intr_callback_register(&sh->intr_handle,
1963                                         mlx5_dev_interrupt_handler, sh)) {
1964                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1965                         sh->intr_handle.fd = -1;
1966                 }
1967         }
1968         if (sh->devx) {
1969 #ifdef HAVE_IBV_DEVX_ASYNC
1970                 sh->intr_handle_devx.fd = -1;
1971                 sh->devx_comp =
1972                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1973                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1974                 if (!devx_comp) {
1975                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1976                         return;
1977                 }
1978                 flags = fcntl(devx_comp->fd, F_GETFL);
1979                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1980                 if (ret) {
1981                         DRV_LOG(INFO, "failed to change file descriptor"
1982                                 " devx comp");
1983                         return;
1984                 }
1985                 sh->intr_handle_devx.fd = devx_comp->fd;
1986                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1987                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1988                                         mlx5_dev_interrupt_handler_devx, sh)) {
1989                         DRV_LOG(INFO, "Fail to install the devx shared"
1990                                 " interrupt.");
1991                         sh->intr_handle_devx.fd = -1;
1992                 }
1993 #endif /* HAVE_IBV_DEVX_ASYNC */
1994         }
1995 }
1996
1997 /**
1998  * Uninstall shared asynchronous device events handler.
1999  * This function is implemented to support event sharing
2000  * between multiple ports of single IB device.
2001  *
2002  * @param dev
2003  *   Pointer to mlx5_dev_ctx_shared object.
2004  */
2005 void
2006 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2007 {
2008         if (sh->intr_handle.fd >= 0)
2009                 mlx5_intr_callback_unregister(&sh->intr_handle,
2010                                               mlx5_dev_interrupt_handler, sh);
2011 #ifdef HAVE_IBV_DEVX_ASYNC
2012         if (sh->intr_handle_devx.fd >= 0)
2013                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2014                                   mlx5_dev_interrupt_handler_devx, sh);
2015         if (sh->devx_comp)
2016                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2017 #endif
2018 }
2019
2020 /**
2021  * Read statistics by a named counter.
2022  *
2023  * @param[in] priv
2024  *   Pointer to the private device data structure.
2025  * @param[in] ctr_name
2026  *   Pointer to the name of the statistic counter to read
2027  * @param[out] stat
2028  *   Pointer to read statistic value.
2029  * @return
2030  *   0 on success and stat is valud, 1 if failed to read the value
2031  *   rte_errno is set.
2032  *
2033  */
2034 int
2035 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2036                       uint64_t *stat)
2037 {
2038         int fd;
2039
2040         if (priv->sh) {
2041                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2042                           priv->sh->ibdev_path,
2043                           priv->dev_port,
2044                           ctr_name);
2045                 fd = open(path, O_RDONLY);
2046                 if (fd != -1) {
2047                         char buf[21] = {'\0'};
2048                         ssize_t n = read(fd, buf, sizeof(buf));
2049
2050                         close(fd);
2051                         if (n != -1) {
2052                                 *stat = strtoull(buf, NULL, 10);
2053                                 return 0;
2054                         }
2055                 }
2056         }
2057         *stat = 0;
2058         return 1;
2059 }
2060
2061 /**
2062  * Read device counters table.
2063  *
2064  * @param dev
2065  *   Pointer to Ethernet device.
2066  * @param[out] stats
2067  *   Counters table output buffer.
2068  *
2069  * @return
2070  *   0 on success and stats is filled, negative errno value otherwise and
2071  *   rte_errno is set.
2072  */
2073 int
2074 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
2075 {
2076         struct mlx5_priv *priv = dev->data->dev_private;
2077         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2078         unsigned int i;
2079         struct ifreq ifr;
2080         unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
2081         unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
2082         struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
2083         int ret;
2084
2085         et_stats->cmd = ETHTOOL_GSTATS;
2086         et_stats->n_stats = xstats_ctrl->stats_n;
2087         ifr.ifr_data = (caddr_t)et_stats;
2088         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2089         if (ret) {
2090                 DRV_LOG(WARNING,
2091                         "port %u unable to read statistic values from device",
2092                         dev->data->port_id);
2093                 return ret;
2094         }
2095         for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2096                 if (xstats_ctrl->info[i].dev) {
2097                         ret = mlx5_os_read_dev_stat(priv,
2098                                             xstats_ctrl->info[i].ctr_name,
2099                                             &stats[i]);
2100                         /* return last xstats counter if fail to read. */
2101                         if (ret == 0)
2102                                 xstats_ctrl->xstats[i] = stats[i];
2103                         else
2104                                 stats[i] = xstats_ctrl->xstats[i];
2105                 } else {
2106                         stats[i] = (uint64_t)
2107                                 et_stats->data[xstats_ctrl->dev_table_idx[i]];
2108                 }
2109         }
2110         return 0;
2111 }
2112
2113 /**
2114  * Query the number of statistics provided by ETHTOOL.
2115  *
2116  * @param dev
2117  *   Pointer to Ethernet device.
2118  *
2119  * @return
2120  *   Number of statistics on success, negative errno value otherwise and
2121  *   rte_errno is set.
2122  */
2123 int
2124 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2125 {
2126         struct ethtool_drvinfo drvinfo;
2127         struct ifreq ifr;
2128         int ret;
2129
2130         drvinfo.cmd = ETHTOOL_GDRVINFO;
2131         ifr.ifr_data = (caddr_t)&drvinfo;
2132         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2133         if (ret) {
2134                 DRV_LOG(WARNING, "port %u unable to query number of statistics",
2135                         dev->data->port_id);
2136                 return ret;
2137         }
2138         return drvinfo.n_stats;
2139 }
2140
2141 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2142         {
2143                 .dpdk_name = "rx_port_unicast_bytes",
2144                 .ctr_name = "rx_vport_unicast_bytes",
2145         },
2146         {
2147                 .dpdk_name = "rx_port_multicast_bytes",
2148                 .ctr_name = "rx_vport_multicast_bytes",
2149         },
2150         {
2151                 .dpdk_name = "rx_port_broadcast_bytes",
2152                 .ctr_name = "rx_vport_broadcast_bytes",
2153         },
2154         {
2155                 .dpdk_name = "rx_port_unicast_packets",
2156                 .ctr_name = "rx_vport_unicast_packets",
2157         },
2158         {
2159                 .dpdk_name = "rx_port_multicast_packets",
2160                 .ctr_name = "rx_vport_multicast_packets",
2161         },
2162         {
2163                 .dpdk_name = "rx_port_broadcast_packets",
2164                 .ctr_name = "rx_vport_broadcast_packets",
2165         },
2166         {
2167                 .dpdk_name = "tx_port_unicast_bytes",
2168                 .ctr_name = "tx_vport_unicast_bytes",
2169         },
2170         {
2171                 .dpdk_name = "tx_port_multicast_bytes",
2172                 .ctr_name = "tx_vport_multicast_bytes",
2173         },
2174         {
2175                 .dpdk_name = "tx_port_broadcast_bytes",
2176                 .ctr_name = "tx_vport_broadcast_bytes",
2177         },
2178         {
2179                 .dpdk_name = "tx_port_unicast_packets",
2180                 .ctr_name = "tx_vport_unicast_packets",
2181         },
2182         {
2183                 .dpdk_name = "tx_port_multicast_packets",
2184                 .ctr_name = "tx_vport_multicast_packets",
2185         },
2186         {
2187                 .dpdk_name = "tx_port_broadcast_packets",
2188                 .ctr_name = "tx_vport_broadcast_packets",
2189         },
2190         {
2191                 .dpdk_name = "rx_wqe_err",
2192                 .ctr_name = "rx_wqe_err",
2193         },
2194         {
2195                 .dpdk_name = "rx_crc_errors_phy",
2196                 .ctr_name = "rx_crc_errors_phy",
2197         },
2198         {
2199                 .dpdk_name = "rx_in_range_len_errors_phy",
2200                 .ctr_name = "rx_in_range_len_errors_phy",
2201         },
2202         {
2203                 .dpdk_name = "rx_symbol_err_phy",
2204                 .ctr_name = "rx_symbol_err_phy",
2205         },
2206         {
2207                 .dpdk_name = "tx_errors_phy",
2208                 .ctr_name = "tx_errors_phy",
2209         },
2210         {
2211                 .dpdk_name = "rx_out_of_buffer",
2212                 .ctr_name = "out_of_buffer",
2213                 .dev = 1,
2214         },
2215         {
2216                 .dpdk_name = "tx_packets_phy",
2217                 .ctr_name = "tx_packets_phy",
2218         },
2219         {
2220                 .dpdk_name = "rx_packets_phy",
2221                 .ctr_name = "rx_packets_phy",
2222         },
2223         {
2224                 .dpdk_name = "tx_discards_phy",
2225                 .ctr_name = "tx_discards_phy",
2226         },
2227         {
2228                 .dpdk_name = "rx_discards_phy",
2229                 .ctr_name = "rx_discards_phy",
2230         },
2231         {
2232                 .dpdk_name = "tx_bytes_phy",
2233                 .ctr_name = "tx_bytes_phy",
2234         },
2235         {
2236                 .dpdk_name = "rx_bytes_phy",
2237                 .ctr_name = "rx_bytes_phy",
2238         },
2239         /* Representor only */
2240         {
2241                 .dpdk_name = "rx_packets",
2242                 .ctr_name = "vport_rx_packets",
2243         },
2244         {
2245                 .dpdk_name = "rx_bytes",
2246                 .ctr_name = "vport_rx_bytes",
2247         },
2248         {
2249                 .dpdk_name = "tx_packets",
2250                 .ctr_name = "vport_tx_packets",
2251         },
2252         {
2253                 .dpdk_name = "tx_bytes",
2254                 .ctr_name = "vport_tx_bytes",
2255         },
2256 };
2257
2258 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2259
2260 /**
2261  * Init the structures to read device counters.
2262  *
2263  * @param dev
2264  *   Pointer to Ethernet device.
2265  */
2266 void
2267 mlx5_os_stats_init(struct rte_eth_dev *dev)
2268 {
2269         struct mlx5_priv *priv = dev->data->dev_private;
2270         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2271         struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2272         unsigned int i;
2273         unsigned int j;
2274         struct ifreq ifr;
2275         struct ethtool_gstrings *strings = NULL;
2276         unsigned int dev_stats_n;
2277         unsigned int str_sz;
2278         int ret;
2279
2280         /* So that it won't aggregate for each init. */
2281         xstats_ctrl->mlx5_stats_n = 0;
2282         ret = mlx5_os_get_stats_n(dev);
2283         if (ret < 0) {
2284                 DRV_LOG(WARNING, "port %u no extended statistics available",
2285                         dev->data->port_id);
2286                 return;
2287         }
2288         dev_stats_n = ret;
2289         /* Allocate memory to grab stat names and values. */
2290         str_sz = dev_stats_n * ETH_GSTRING_LEN;
2291         strings = (struct ethtool_gstrings *)
2292                   mlx5_malloc(0, str_sz + sizeof(struct ethtool_gstrings), 0,
2293                               SOCKET_ID_ANY);
2294         if (!strings) {
2295                 DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2296                      dev->data->port_id);
2297                 return;
2298         }
2299         strings->cmd = ETHTOOL_GSTRINGS;
2300         strings->string_set = ETH_SS_STATS;
2301         strings->len = dev_stats_n;
2302         ifr.ifr_data = (caddr_t)strings;
2303         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2304         if (ret) {
2305                 DRV_LOG(WARNING, "port %u unable to get statistic names",
2306                         dev->data->port_id);
2307                 goto free;
2308         }
2309         for (i = 0; i != dev_stats_n; ++i) {
2310                 const char *curr_string = (const char *)
2311                         &strings->data[i * ETH_GSTRING_LEN];
2312
2313                 for (j = 0; j != xstats_n; ++j) {
2314                         if (!strcmp(mlx5_counters_init[j].ctr_name,
2315                                     curr_string)) {
2316                                 unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2317
2318                                 xstats_ctrl->dev_table_idx[idx] = i;
2319                                 xstats_ctrl->info[idx] = mlx5_counters_init[j];
2320                                 break;
2321                         }
2322                 }
2323         }
2324         /* Add dev counters. */
2325         for (i = 0; i != xstats_n; ++i) {
2326                 if (mlx5_counters_init[i].dev) {
2327                         unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2328
2329                         xstats_ctrl->info[idx] = mlx5_counters_init[i];
2330                         xstats_ctrl->hw_stats[idx] = 0;
2331                 }
2332         }
2333         MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2334         xstats_ctrl->stats_n = dev_stats_n;
2335         /* Copy to base at first time. */
2336         ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2337         if (ret)
2338                 DRV_LOG(ERR, "port %u cannot read device counters: %s",
2339                         dev->data->port_id, strerror(rte_errno));
2340         mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2341         stats_ctrl->imissed = 0;
2342 free:
2343         mlx5_free(strings);
2344 }
2345
2346 /**
2347  * Set the reg_mr and dereg_mr call backs
2348  *
2349  * @param reg_mr_cb[out]
2350  *   Pointer to reg_mr func
2351  * @param dereg_mr_cb[out]
2352  *   Pointer to dereg_mr func
2353  *
2354  */
2355 void
2356 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2357                       mlx5_dereg_mr_t *dereg_mr_cb)
2358 {
2359         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2360         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2361 }
2362
2363 const struct eth_dev_ops mlx5_os_dev_ops = {
2364         .dev_configure = mlx5_dev_configure,
2365         .dev_start = mlx5_dev_start,
2366         .dev_stop = mlx5_dev_stop,
2367         .dev_set_link_down = mlx5_set_link_down,
2368         .dev_set_link_up = mlx5_set_link_up,
2369         .dev_close = mlx5_dev_close,
2370         .promiscuous_enable = mlx5_promiscuous_enable,
2371         .promiscuous_disable = mlx5_promiscuous_disable,
2372         .allmulticast_enable = mlx5_allmulticast_enable,
2373         .allmulticast_disable = mlx5_allmulticast_disable,
2374         .link_update = mlx5_link_update,
2375         .stats_get = mlx5_stats_get,
2376         .stats_reset = mlx5_stats_reset,
2377         .xstats_get = mlx5_xstats_get,
2378         .xstats_reset = mlx5_xstats_reset,
2379         .xstats_get_names = mlx5_xstats_get_names,
2380         .fw_version_get = mlx5_fw_version_get,
2381         .dev_infos_get = mlx5_dev_infos_get,
2382         .read_clock = mlx5_txpp_read_clock,
2383         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2384         .vlan_filter_set = mlx5_vlan_filter_set,
2385         .rx_queue_setup = mlx5_rx_queue_setup,
2386         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2387         .tx_queue_setup = mlx5_tx_queue_setup,
2388         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2389         .rx_queue_release = mlx5_rx_queue_release,
2390         .tx_queue_release = mlx5_tx_queue_release,
2391         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2392         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2393         .mac_addr_remove = mlx5_mac_addr_remove,
2394         .mac_addr_add = mlx5_mac_addr_add,
2395         .mac_addr_set = mlx5_mac_addr_set,
2396         .set_mc_addr_list = mlx5_set_mc_addr_list,
2397         .mtu_set = mlx5_dev_set_mtu,
2398         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2399         .vlan_offload_set = mlx5_vlan_offload_set,
2400         .reta_update = mlx5_dev_rss_reta_update,
2401         .reta_query = mlx5_dev_rss_reta_query,
2402         .rss_hash_update = mlx5_rss_hash_update,
2403         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2404         .filter_ctrl = mlx5_dev_filter_ctrl,
2405         .rx_descriptor_status = mlx5_rx_descriptor_status,
2406         .tx_descriptor_status = mlx5_tx_descriptor_status,
2407         .rxq_info_get = mlx5_rxq_info_get,
2408         .txq_info_get = mlx5_txq_info_get,
2409         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2410         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2411         .rx_queue_count = mlx5_rx_queue_count,
2412         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2413         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2414         .is_removed = mlx5_is_removed,
2415         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2416         .get_module_info = mlx5_get_module_info,
2417         .get_module_eeprom = mlx5_get_module_eeprom,
2418         .hairpin_cap_get = mlx5_hairpin_cap_get,
2419         .mtr_ops_get = mlx5_flow_meter_ops_get,
2420 };
2421
2422 /* Available operations from secondary process. */
2423 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2424         .stats_get = mlx5_stats_get,
2425         .stats_reset = mlx5_stats_reset,
2426         .xstats_get = mlx5_xstats_get,
2427         .xstats_reset = mlx5_xstats_reset,
2428         .xstats_get_names = mlx5_xstats_get_names,
2429         .fw_version_get = mlx5_fw_version_get,
2430         .dev_infos_get = mlx5_dev_infos_get,
2431         .read_clock = mlx5_txpp_read_clock,
2432         .rx_descriptor_status = mlx5_rx_descriptor_status,
2433         .tx_descriptor_status = mlx5_tx_descriptor_status,
2434         .rxq_info_get = mlx5_rxq_info_get,
2435         .txq_info_get = mlx5_txq_info_get,
2436         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2437         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2438         .get_module_info = mlx5_get_module_info,
2439         .get_module_eeprom = mlx5_get_module_eeprom,
2440 };
2441
2442 /* Available operations in flow isolated mode. */
2443 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2444         .dev_configure = mlx5_dev_configure,
2445         .dev_start = mlx5_dev_start,
2446         .dev_stop = mlx5_dev_stop,
2447         .dev_set_link_down = mlx5_set_link_down,
2448         .dev_set_link_up = mlx5_set_link_up,
2449         .dev_close = mlx5_dev_close,
2450         .promiscuous_enable = mlx5_promiscuous_enable,
2451         .promiscuous_disable = mlx5_promiscuous_disable,
2452         .allmulticast_enable = mlx5_allmulticast_enable,
2453         .allmulticast_disable = mlx5_allmulticast_disable,
2454         .link_update = mlx5_link_update,
2455         .stats_get = mlx5_stats_get,
2456         .stats_reset = mlx5_stats_reset,
2457         .xstats_get = mlx5_xstats_get,
2458         .xstats_reset = mlx5_xstats_reset,
2459         .xstats_get_names = mlx5_xstats_get_names,
2460         .fw_version_get = mlx5_fw_version_get,
2461         .dev_infos_get = mlx5_dev_infos_get,
2462         .read_clock = mlx5_txpp_read_clock,
2463         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2464         .vlan_filter_set = mlx5_vlan_filter_set,
2465         .rx_queue_setup = mlx5_rx_queue_setup,
2466         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2467         .tx_queue_setup = mlx5_tx_queue_setup,
2468         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2469         .rx_queue_release = mlx5_rx_queue_release,
2470         .tx_queue_release = mlx5_tx_queue_release,
2471         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2472         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2473         .mac_addr_remove = mlx5_mac_addr_remove,
2474         .mac_addr_add = mlx5_mac_addr_add,
2475         .mac_addr_set = mlx5_mac_addr_set,
2476         .set_mc_addr_list = mlx5_set_mc_addr_list,
2477         .mtu_set = mlx5_dev_set_mtu,
2478         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2479         .vlan_offload_set = mlx5_vlan_offload_set,
2480         .filter_ctrl = mlx5_dev_filter_ctrl,
2481         .rx_descriptor_status = mlx5_rx_descriptor_status,
2482         .tx_descriptor_status = mlx5_tx_descriptor_status,
2483         .rxq_info_get = mlx5_rxq_info_get,
2484         .txq_info_get = mlx5_txq_info_get,
2485         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2486         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2487         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2488         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2489         .is_removed = mlx5_is_removed,
2490         .get_module_info = mlx5_get_module_info,
2491         .get_module_eeprom = mlx5_get_module_eeprom,
2492         .hairpin_cap_get = mlx5_hairpin_cap_get,
2493         .mtr_ops_get = mlx5_flow_meter_ops_get,
2494 };