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 = rte_malloc_socket(__func__, 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         rte_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 = rte_zmalloc("ethdev private structure",
 622                            sizeof(*priv),
 623                            RTE_CACHE_LINE_SIZE);
 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         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 780                 (config.hw_fcs_strip ? "" : "not "));
 781 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 782         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 783 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 784         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 785                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 786 #endif
 787         if (config.hw_padding && !hw_padding) {
 788                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 789                 config.hw_padding = 0;
 790         } else if (config.hw_padding) {
 791                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 792         }
 793         config.tso = (sh->device_attr.max_tso > 0 &&
 794                       (sh->device_attr.tso_supported_qpts &
 795                        (1 << IBV_QPT_RAW_PACKET)));
 796         if (config.tso)
 797                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 798         /*
 799          * MPW is disabled by default, while the Enhanced MPW is enabled
 800          * by default.
 801          */
 802         if (config.mps == MLX5_ARG_UNSET)
 803                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 804                                                           MLX5_MPW_DISABLED;
 805         else
 806                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 807         DRV_LOG(INFO, "%sMPS is %s",
 808                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 809                 config.mps == MLX5_MPW ? "legacy " : "",
 810                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 811         if (config.cqe_comp && !cqe_comp) {
 812                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 813                 config.cqe_comp = 0;
 814         }
 815         if (config.cqe_pad && !cqe_pad) {
 816                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 817                 config.cqe_pad = 0;
 818         } else if (config.cqe_pad) {
 819                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 820         }
 821         if (config.devx) {
 822                 priv->counter_fallback = 0;
 823                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 824                 if (err) {
 825                         err = -err;
 826                         goto error;
 827                 }
 828                 if (!config.hca_attr.flow_counters_dump)
 829                         priv->counter_fallback = 1;
 830 #ifndef HAVE_IBV_DEVX_ASYNC
 831                 priv->counter_fallback = 1;
 832 #endif
 833                 if (priv->counter_fallback)
 834                         DRV_LOG(INFO, "Use fall-back DV counter management");
 835                 /* Check for LRO support. */
 836                 if (config.dest_tir && config.hca_attr.lro_cap &&
 837                     config.dv_flow_en) {
 838                         /* TBD check tunnel lro caps. */
 839                         config.lro.supported = config.hca_attr.lro_cap;
 840                         DRV_LOG(DEBUG, "Device supports LRO");
 841                         /*
 842                          * If LRO timeout is not configured by application,
 843                          * use the minimal supported value.
 844                          */
 845                         if (!config.lro.timeout)
 846                                 config.lro.timeout =
 847                                 config.hca_attr.lro_timer_supported_periods[0];
 848                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 849                                 config.lro.timeout);
 850                 }
 851 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 852                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 853                     config.dv_flow_en) {
 854                         uint8_t reg_c_mask =
 855                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 856                         /*
 857                          * Meter needs two REG_C's for color match and pre-sfx
 858                          * flow match. Here get the REG_C for color match.
 859                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 860                          */
 861                         reg_c_mask &= 0xfc;
 862                         if (__builtin_popcount(reg_c_mask) < 1) {
 863                                 priv->mtr_en = 0;
 864                                 DRV_LOG(WARNING, "No available register for"
 865                                         " meter.");
 866                         } else {
 867                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 868                                                       REG_C_0;
 869                                 priv->mtr_en = 1;
 870                                 priv->mtr_reg_share =
 871                                       config.hca_attr.qos.flow_meter_reg_share;
 872                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 873                                         priv->mtr_color_reg);
 874                         }
 875                 }
 876 #endif
 877         }
 878         if (config.tx_pp) {
 879                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
 880                         config.hca_attr.dev_freq_khz);
 881                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
 882                         config.hca_attr.qos.packet_pacing ? "" : "not ");
 883                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
 884                         config.hca_attr.cross_channel ? "" : "not ");
 885                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
 886                         config.hca_attr.wqe_index_ignore ? "" : "not ");
 887                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
 888                         config.hca_attr.non_wire_sq ? "" : "not ");
 889                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
 890                         config.hca_attr.log_max_static_sq_wq ? "" : "not ",
 891                         config.hca_attr.log_max_static_sq_wq);
 892                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
 893                         config.hca_attr.qos.wqe_rate_pp ? "" : "not ");
 894                 if (!config.devx) {
 895                         DRV_LOG(ERR, "DevX is required for packet pacing");
 896                         err = ENODEV;
 897                         goto error;
 898                 }
 899                 if (!config.hca_attr.qos.packet_pacing) {
 900                         DRV_LOG(ERR, "Packet pacing is not supported");
 901                         err = ENODEV;
 902                         goto error;
 903                 }
 904                 if (!config.hca_attr.cross_channel) {
 905                         DRV_LOG(ERR, "Cross channel operations are"
 906                                      " required for packet pacing");
 907                         err = ENODEV;
 908                         goto error;
 909                 }
 910                 if (!config.hca_attr.wqe_index_ignore) {
 911                         DRV_LOG(ERR, "WQE index ignore feature is"
 912                                      " required for packet pacing");
 913                         err = ENODEV;
 914                         goto error;
 915                 }
 916                 if (!config.hca_attr.non_wire_sq) {
 917                         DRV_LOG(ERR, "Non-wire SQ feature is"
 918                                      " required for packet pacing");
 919                         err = ENODEV;
 920                         goto error;
 921                 }
 922                 if (!config.hca_attr.log_max_static_sq_wq) {
 923                         DRV_LOG(ERR, "Static WQE SQ feature is"
 924                                      " required for packet pacing");
 925                         err = ENODEV;
 926                         goto error;
 927                 }
 928                 if (!config.hca_attr.qos.wqe_rate_pp) {
 929                         DRV_LOG(ERR, "WQE rate mode is required"
 930                                      " for packet pacing");
 931                         err = ENODEV;
 932                         goto error;
 933                 }
 934 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
 935                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
 936                              " can't create queues for packet pacing");
 937                 err = ENODEV;
 938                 goto error;
 939 #endif
 940         }
 941         if (config.devx) {
 942                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
 943
 944                 err = mlx5_devx_cmd_register_read
 945                         (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
 946                         reg, MLX5_ST_SZ_DW(register_mtutc));
 947                 if (!err) {
 948                         uint32_t ts_mode;
 949
 950                         /* MTUTC register is read successfully. */
 951                         ts_mode = MLX5_GET(register_mtutc, reg,
 952                                            time_stamp_mode);
 953                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
 954                                 config.rt_timestamp = 1;
 955                 } else {
 956                         /* Kernel does not support register reading. */
 957                         if (config.hca_attr.dev_freq_khz ==
 958                                                  (NS_PER_S / MS_PER_S))
 959                                 config.rt_timestamp = 1;
 960                 }
 961         }
 962         if (config.mprq.enabled && mprq) {
 963                 if (config.mprq.stride_num_n &&
 964                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 965                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 966                         config.mprq.stride_num_n =
 967                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 968                                                 mprq_min_stride_num_n),
 969                                         mprq_max_stride_num_n);
 970                         DRV_LOG(WARNING,
 971                                 "the number of strides"
 972                                 " for Multi-Packet RQ is out of range,"
 973                                 " setting default value (%u)",
 974                                 1 << config.mprq.stride_num_n);
 975                 }
 976                 if (config.mprq.stride_size_n &&
 977                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 978                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 979                         config.mprq.stride_size_n =
 980                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 981                                                 mprq_min_stride_size_n),
 982                                         mprq_max_stride_size_n);
 983                         DRV_LOG(WARNING,
 984                                 "the size of a stride"
 985                                 " for Multi-Packet RQ is out of range,"
 986                                 " setting default value (%u)",
 987                                 1 << config.mprq.stride_size_n);
 988                 }
 989                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
 990                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
 991         } else if (config.mprq.enabled && !mprq) {
 992                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
 993                 config.mprq.enabled = 0;
 994         }
 995         if (config.max_dump_files_num == 0)
 996                 config.max_dump_files_num = 128;
 997         eth_dev = rte_eth_dev_allocate(name);
 998         if (eth_dev == NULL) {
 999                 DRV_LOG(ERR, "can not allocate rte ethdev");
1000                 err = ENOMEM;
1001                 goto error;
1002         }
1003         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1004         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1005         if (priv->representor) {
1006                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1007                 eth_dev->data->representor_id = priv->representor_id;
1008         }
1009         /*
1010          * Store associated network device interface index. This index
1011          * is permanent throughout the lifetime of device. So, we may store
1012          * the ifindex here and use the cached value further.
1013          */
1014         MLX5_ASSERT(spawn->ifindex);
1015         priv->if_index = spawn->ifindex;
1016         eth_dev->data->dev_private = priv;
1017         priv->dev_data = eth_dev->data;
1018         eth_dev->data->mac_addrs = priv->mac;
1019         eth_dev->device = dpdk_dev;
1020         /* Configure the first MAC address by default. */
1021         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1022                 DRV_LOG(ERR,
1023                         "port %u cannot get MAC address, is mlx5_en"
1024                         " loaded? (errno: %s)",
1025                         eth_dev->data->port_id, strerror(rte_errno));
1026                 err = ENODEV;
1027                 goto error;
1028         }
1029         DRV_LOG(INFO,
1030                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1031                 eth_dev->data->port_id,
1032                 mac.addr_bytes[0], mac.addr_bytes[1],
1033                 mac.addr_bytes[2], mac.addr_bytes[3],
1034                 mac.addr_bytes[4], mac.addr_bytes[5]);
1035 #ifdef RTE_LIBRTE_MLX5_DEBUG
1036         {
1037                 char ifname[IF_NAMESIZE];
1038
1039                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1040                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1041                                 eth_dev->data->port_id, ifname);
1042                 else
1043                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1044                                 eth_dev->data->port_id);
1045         }
1046 #endif
1047         /* Get actual MTU if possible. */
1048         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1049         if (err) {
1050                 err = rte_errno;
1051                 goto error;
1052         }
1053         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1054                 priv->mtu);
1055         /* Initialize burst functions to prevent crashes before link-up. */
1056         eth_dev->rx_pkt_burst = removed_rx_burst;
1057         eth_dev->tx_pkt_burst = removed_tx_burst;
1058         eth_dev->dev_ops = &mlx5_os_dev_ops;
1059         /* Register MAC address. */
1060         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1061         if (config.vf && config.vf_nl_en)
1062                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1063                                       mlx5_ifindex(eth_dev),
1064                                       eth_dev->data->mac_addrs,
1065                                       MLX5_MAX_MAC_ADDRESSES);
1066         priv->flows = 0;
1067         priv->ctrl_flows = 0;
1068         TAILQ_INIT(&priv->flow_meters);
1069         TAILQ_INIT(&priv->flow_meter_profiles);
1070         /* Hint libmlx5 to use PMD allocator for data plane resources */
1071         mlx5_glue->dv_set_context_attr(sh->ctx,
1072                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1073                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1074                                 .alloc = &mlx5_alloc_verbs_buf,
1075                                 .free = &mlx5_free_verbs_buf,
1076                                 .data = priv,
1077                         }));
1078         /* Bring Ethernet device up. */
1079         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1080                 eth_dev->data->port_id);
1081         mlx5_set_link_up(eth_dev);
1082         /*
1083          * Even though the interrupt handler is not installed yet,
1084          * interrupts will still trigger on the async_fd from
1085          * Verbs context returned by ibv_open_device().
1086          */
1087         mlx5_link_update(eth_dev, 0);
1088 #ifdef HAVE_MLX5DV_DR_ESWITCH
1089         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1090               (switch_info->representor || switch_info->master)))
1091                 config.dv_esw_en = 0;
1092 #else
1093         config.dv_esw_en = 0;
1094 #endif
1095         /* Detect minimal data bytes to inline. */
1096         mlx5_set_min_inline(spawn, &config);
1097         /* Store device configuration on private structure. */
1098         priv->config = config;
1099         /* Create context for virtual machine VLAN workaround. */
1100         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1101         if (config.dv_flow_en) {
1102                 err = mlx5_alloc_shared_dr(priv);
1103                 if (err)
1104                         goto error;
1105                 /*
1106                  * RSS id is shared with meter flow id. Meter flow id can only
1107                  * use the 24 MSB of the register.
1108                  */
1109                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1110                                      MLX5_MTR_COLOR_BITS);
1111                 if (!priv->qrss_id_pool) {
1112                         DRV_LOG(ERR, "can't create flow id pool");
1113                         err = ENOMEM;
1114                         goto error;
1115                 }
1116         }
1117         /* Supported Verbs flow priority number detection. */
1118         err = mlx5_flow_discover_priorities(eth_dev);
1119         if (err < 0) {
1120                 err = -err;
1121                 goto error;
1122         }
1123         priv->config.flow_prio = err;
1124         if (!priv->config.dv_esw_en &&
1125             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1126                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1127                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1128                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1129         }
1130         mlx5_set_metadata_mask(eth_dev);
1131         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1132             !priv->sh->dv_regc0_mask) {
1133                 DRV_LOG(ERR, "metadata mode %u is not supported "
1134                              "(no metadata reg_c[0] is available)",
1135                              priv->config.dv_xmeta_en);
1136                         err = ENOTSUP;
1137                         goto error;
1138         }
1139         /*
1140          * Allocate the buffer for flow creating, just once.
1141          * The allocation must be done before any flow creating.
1142          */
1143         mlx5_flow_alloc_intermediate(eth_dev);
1144         /* Query availability of metadata reg_c's. */
1145         err = mlx5_flow_discover_mreg_c(eth_dev);
1146         if (err < 0) {
1147                 err = -err;
1148                 goto error;
1149         }
1150         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1151                 DRV_LOG(DEBUG,
1152                         "port %u extensive metadata register is not supported",
1153                         eth_dev->data->port_id);
1154                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1155                         DRV_LOG(ERR, "metadata mode %u is not supported "
1156                                      "(no metadata registers available)",
1157                                      priv->config.dv_xmeta_en);
1158                         err = ENOTSUP;
1159                         goto error;
1160                 }
1161         }
1162         if (priv->config.dv_flow_en &&
1163             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1164             mlx5_flow_ext_mreg_supported(eth_dev) &&
1165             priv->sh->dv_regc0_mask) {
1166                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1167                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1168                 if (!priv->mreg_cp_tbl) {
1169                         err = ENOMEM;
1170                         goto error;
1171                 }
1172         }
1173         return eth_dev;
1174 error:
1175         if (priv) {
1176                 if (priv->mreg_cp_tbl)
1177                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1178                 if (priv->sh)
1179                         mlx5_os_free_shared_dr(priv);
1180                 if (priv->nl_socket_route >= 0)
1181                         close(priv->nl_socket_route);
1182                 if (priv->nl_socket_rdma >= 0)
1183                         close(priv->nl_socket_rdma);
1184                 if (priv->vmwa_context)
1185                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1186                 if (priv->qrss_id_pool)
1187                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1188                 if (own_domain_id)
1189                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1190                 rte_free(priv);
1191                 if (eth_dev != NULL)
1192                         eth_dev->data->dev_private = NULL;
1193         }
1194         if (eth_dev != NULL) {
1195                 /* mac_addrs must not be freed alone because part of
1196                  * dev_private
1197                  **/
1198                 eth_dev->data->mac_addrs = NULL;
1199                 rte_eth_dev_release_port(eth_dev);
1200         }
1201         if (sh)
1202                 mlx5_free_shared_dev_ctx(sh);
1203         MLX5_ASSERT(err > 0);
1204         rte_errno = err;
1205         return NULL;
1206 }
1207
1208 /**
1209  * Comparison callback to sort device data.
1210  *
1211  * This is meant to be used with qsort().
1212  *
1213  * @param a[in]
1214  *   Pointer to pointer to first data object.
1215  * @param b[in]
1216  *   Pointer to pointer to second data object.
1217  *
1218  * @return
1219  *   0 if both objects are equal, less than 0 if the first argument is less
1220  *   than the second, greater than 0 otherwise.
1221  */
1222 static int
1223 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1224 {
1225         const struct mlx5_switch_info *si_a =
1226                 &((const struct mlx5_dev_spawn_data *)a)->info;
1227         const struct mlx5_switch_info *si_b =
1228                 &((const struct mlx5_dev_spawn_data *)b)->info;
1229         int ret;
1230
1231         /* Master device first. */
1232         ret = si_b->master - si_a->master;
1233         if (ret)
1234                 return ret;
1235         /* Then representor devices. */
1236         ret = si_b->representor - si_a->representor;
1237         if (ret)
1238                 return ret;
1239         /* Unidentified devices come last in no specific order. */
1240         if (!si_a->representor)
1241                 return 0;
1242         /* Order representors by name. */
1243         return si_a->port_name - si_b->port_name;
1244 }
1245
1246 /**
1247  * Match PCI information for possible slaves of bonding device.
1248  *
1249  * @param[in] ibv_dev
1250  *   Pointer to Infiniband device structure.
1251  * @param[in] pci_dev
1252  *   Pointer to PCI device structure to match PCI address.
1253  * @param[in] nl_rdma
1254  *   Netlink RDMA group socket handle.
1255  *
1256  * @return
1257  *   negative value if no bonding device found, otherwise
1258  *   positive index of slave PF in bonding.
1259  */
1260 static int
1261 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1262                            const struct rte_pci_device *pci_dev,
1263                            int nl_rdma)
1264 {
1265         char ifname[IF_NAMESIZE + 1];
1266         unsigned int ifindex;
1267         unsigned int np, i;
1268         FILE *file = NULL;
1269         int pf = -1;
1270
1271         /*
1272          * Try to get master device name. If something goes
1273          * wrong suppose the lack of kernel support and no
1274          * bonding devices.
1275          */
1276         if (nl_rdma < 0)
1277                 return -1;
1278         if (!strstr(ibv_dev->name, "bond"))
1279                 return -1;
1280         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1281         if (!np)
1282                 return -1;
1283         /*
1284          * The Master device might not be on the predefined
1285          * port (not on port index 1, it is not garanted),
1286          * we have to scan all Infiniband device port and
1287          * find master.
1288          */
1289         for (i = 1; i <= np; ++i) {
1290                 /* Check whether Infiniband port is populated. */
1291                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1292                 if (!ifindex)
1293                         continue;
1294                 if (!if_indextoname(ifindex, ifname))
1295                         continue;
1296                 /* Try to read bonding slave names from sysfs. */
1297                 MKSTR(slaves,
1298                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1299                 file = fopen(slaves, "r");
1300                 if (file)
1301                         break;
1302         }
1303         if (!file)
1304                 return -1;
1305         /* Use safe format to check maximal buffer length. */
1306         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1307         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1308                 char tmp_str[IF_NAMESIZE + 32];
1309                 struct rte_pci_addr pci_addr;
1310                 struct mlx5_switch_info info;
1311
1312                 /* Process slave interface names in the loop. */
1313                 snprintf(tmp_str, sizeof(tmp_str),
1314                          "/sys/class/net/%s", ifname);
1315                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1316                         DRV_LOG(WARNING, "can not get PCI address"
1317                                          " for netdev \"%s\"", ifname);
1318                         continue;
1319                 }
1320                 if (pci_dev->addr.domain != pci_addr.domain ||
1321                     pci_dev->addr.bus != pci_addr.bus ||
1322                     pci_dev->addr.devid != pci_addr.devid ||
1323                     pci_dev->addr.function != pci_addr.function)
1324                         continue;
1325                 /* Slave interface PCI address match found. */
1326                 fclose(file);
1327                 snprintf(tmp_str, sizeof(tmp_str),
1328                          "/sys/class/net/%s/phys_port_name", ifname);
1329                 file = fopen(tmp_str, "rb");
1330                 if (!file)
1331                         break;
1332                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1333                 if (fscanf(file, "%32s", tmp_str) == 1)
1334                         mlx5_translate_port_name(tmp_str, &info);
1335                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1336                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1337                         pf = info.port_name;
1338                 break;
1339         }
1340         if (file)
1341                 fclose(file);
1342         return pf;
1343 }
1344
1345 /**
1346  * DPDK callback to register a PCI device.
1347  *
1348  * This function spawns Ethernet devices out of a given PCI device.
1349  *
1350  * @param[in] pci_drv
1351  *   PCI driver structure (mlx5_driver).
1352  * @param[in] pci_dev
1353  *   PCI device information.
1354  *
1355  * @return
1356  *   0 on success, a negative errno value otherwise and rte_errno is set.
1357  */
1358 int
1359 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1360                   struct rte_pci_device *pci_dev)
1361 {
1362         struct ibv_device **ibv_list;
1363         /*
1364          * Number of found IB Devices matching with requested PCI BDF.
1365          * nd != 1 means there are multiple IB devices over the same
1366          * PCI device and we have representors and master.
1367          */
1368         unsigned int nd = 0;
1369         /*
1370          * Number of found IB device Ports. nd = 1 and np = 1..n means
1371          * we have the single multiport IB device, and there may be
1372          * representors attached to some of found ports.
1373          */
1374         unsigned int np = 0;
1375         /*
1376          * Number of DPDK ethernet devices to Spawn - either over
1377          * multiple IB devices or multiple ports of single IB device.
1378          * Actually this is the number of iterations to spawn.
1379          */
1380         unsigned int ns = 0;
1381         /*
1382          * Bonding device
1383          *   < 0 - no bonding device (single one)
1384          *  >= 0 - bonding device (value is slave PF index)
1385          */
1386         int bd = -1;
1387         struct mlx5_dev_spawn_data *list = NULL;
1388         struct mlx5_dev_config dev_config;
1389         int ret;
1390
1391         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1392                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1393                         " driver.");
1394                 return 1;
1395         }
1396         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1397                 mlx5_pmd_socket_init();
1398         ret = mlx5_init_once();
1399         if (ret) {
1400                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1401                         strerror(rte_errno));
1402                 return -rte_errno;
1403         }
1404         MLX5_ASSERT(pci_drv == &mlx5_driver);
1405         errno = 0;
1406         ibv_list = mlx5_glue->get_device_list(&ret);
1407         if (!ibv_list) {
1408                 rte_errno = errno ? errno : ENOSYS;
1409                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1410                 return -rte_errno;
1411         }
1412         /*
1413          * First scan the list of all Infiniband devices to find
1414          * matching ones, gathering into the list.
1415          */
1416         struct ibv_device *ibv_match[ret + 1];
1417         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1418         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1419         unsigned int i;
1420
1421         while (ret-- > 0) {
1422                 struct rte_pci_addr pci_addr;
1423
1424                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1425                 bd = mlx5_device_bond_pci_match
1426                                 (ibv_list[ret], pci_dev, nl_rdma);
1427                 if (bd >= 0) {
1428                         /*
1429                          * Bonding device detected. Only one match is allowed,
1430                          * the bonding is supported over multi-port IB device,
1431                          * there should be no matches on representor PCI
1432                          * functions or non VF LAG bonding devices with
1433                          * specified address.
1434                          */
1435                         if (nd) {
1436                                 DRV_LOG(ERR,
1437                                         "multiple PCI match on bonding device"
1438                                         "\"%s\" found", ibv_list[ret]->name);
1439                                 rte_errno = ENOENT;
1440                                 ret = -rte_errno;
1441                                 goto exit;
1442                         }
1443                         DRV_LOG(INFO, "PCI information matches for"
1444                                       " slave %d bonding device \"%s\"",
1445                                       bd, ibv_list[ret]->name);
1446                         ibv_match[nd++] = ibv_list[ret];
1447                         break;
1448                 }
1449                 if (mlx5_dev_to_pci_addr
1450                         (ibv_list[ret]->ibdev_path, &pci_addr))
1451                         continue;
1452                 if (pci_dev->addr.domain != pci_addr.domain ||
1453                     pci_dev->addr.bus != pci_addr.bus ||
1454                     pci_dev->addr.devid != pci_addr.devid ||
1455                     pci_dev->addr.function != pci_addr.function)
1456                         continue;
1457                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1458                         ibv_list[ret]->name);
1459                 ibv_match[nd++] = ibv_list[ret];
1460         }
1461         ibv_match[nd] = NULL;
1462         if (!nd) {
1463                 /* No device matches, just complain and bail out. */
1464                 DRV_LOG(WARNING,
1465                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1466                         " are kernel drivers loaded?",
1467                         pci_dev->addr.domain, pci_dev->addr.bus,
1468                         pci_dev->addr.devid, pci_dev->addr.function);
1469                 rte_errno = ENOENT;
1470                 ret = -rte_errno;
1471                 goto exit;
1472         }
1473         if (nd == 1) {
1474                 /*
1475                  * Found single matching device may have multiple ports.
1476                  * Each port may be representor, we have to check the port
1477                  * number and check the representors existence.
1478                  */
1479                 if (nl_rdma >= 0)
1480                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1481                 if (!np)
1482                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1483                                          " ports number", ibv_match[0]->name);
1484                 if (bd >= 0 && !np) {
1485                         DRV_LOG(ERR, "can not get ports"
1486                                      " for bonding device");
1487                         rte_errno = ENOENT;
1488                         ret = -rte_errno;
1489                         goto exit;
1490                 }
1491         }
1492 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1493         if (bd >= 0) {
1494                 /*
1495                  * This may happen if there is VF LAG kernel support and
1496                  * application is compiled with older rdma_core library.
1497                  */
1498                 DRV_LOG(ERR,
1499                         "No kernel/verbs support for VF LAG bonding found.");
1500                 rte_errno = ENOTSUP;
1501                 ret = -rte_errno;
1502                 goto exit;
1503         }
1504 #endif
1505         /*
1506          * Now we can determine the maximal
1507          * amount of devices to be spawned.
1508          */
1509         list = rte_zmalloc("device spawn data",
1510                          sizeof(struct mlx5_dev_spawn_data) *
1511                          (np ? np : nd),
1512                          RTE_CACHE_LINE_SIZE);
1513         if (!list) {
1514                 DRV_LOG(ERR, "spawn data array allocation failure");
1515                 rte_errno = ENOMEM;
1516                 ret = -rte_errno;
1517                 goto exit;
1518         }
1519         if (bd >= 0 || np > 1) {
1520                 /*
1521                  * Single IB device with multiple ports found,
1522                  * it may be E-Switch master device and representors.
1523                  * We have to perform identification through the ports.
1524                  */
1525                 MLX5_ASSERT(nl_rdma >= 0);
1526                 MLX5_ASSERT(ns == 0);
1527                 MLX5_ASSERT(nd == 1);
1528                 MLX5_ASSERT(np);
1529                 for (i = 1; i <= np; ++i) {
1530                         list[ns].max_port = np;
1531                         list[ns].phys_port = i;
1532                         list[ns].phys_dev = ibv_match[0];
1533                         list[ns].eth_dev = NULL;
1534                         list[ns].pci_dev = pci_dev;
1535                         list[ns].pf_bond = bd;
1536                         list[ns].ifindex = mlx5_nl_ifindex
1537                                 (nl_rdma,
1538                                 mlx5_os_get_dev_device_name
1539                                                 (list[ns].phys_dev), i);
1540                         if (!list[ns].ifindex) {
1541                                 /*
1542                                  * No network interface index found for the
1543                                  * specified port, it means there is no
1544                                  * representor on this port. It's OK,
1545                                  * there can be disabled ports, for example
1546                                  * if sriov_numvfs < sriov_totalvfs.
1547                                  */
1548                                 continue;
1549                         }
1550                         ret = -1;
1551                         if (nl_route >= 0)
1552                                 ret = mlx5_nl_switch_info
1553                                                (nl_route,
1554                                                 list[ns].ifindex,
1555                                                 &list[ns].info);
1556                         if (ret || (!list[ns].info.representor &&
1557                                     !list[ns].info.master)) {
1558                                 /*
1559                                  * We failed to recognize representors with
1560                                  * Netlink, let's try to perform the task
1561                                  * with sysfs.
1562                                  */
1563                                 ret =  mlx5_sysfs_switch_info
1564                                                 (list[ns].ifindex,
1565                                                  &list[ns].info);
1566                         }
1567                         if (!ret && bd >= 0) {
1568                                 switch (list[ns].info.name_type) {
1569                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1570                                         if (list[ns].info.port_name == bd)
1571                                                 ns++;
1572                                         break;
1573                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1574                                         /* Fallthrough */
1575                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1576                                         if (list[ns].info.pf_num == bd)
1577                                                 ns++;
1578                                         break;
1579                                 default:
1580                                         break;
1581                                 }
1582                                 continue;
1583                         }
1584                         if (!ret && (list[ns].info.representor ^
1585                                      list[ns].info.master))
1586                                 ns++;
1587                 }
1588                 if (!ns) {
1589                         DRV_LOG(ERR,
1590                                 "unable to recognize master/representors"
1591                                 " on the IB device with multiple ports");
1592                         rte_errno = ENOENT;
1593                         ret = -rte_errno;
1594                         goto exit;
1595                 }
1596         } else {
1597                 /*
1598                  * The existence of several matching entries (nd > 1) means
1599                  * port representors have been instantiated. No existing Verbs
1600                  * call nor sysfs entries can tell them apart, this can only
1601                  * be done through Netlink calls assuming kernel drivers are
1602                  * recent enough to support them.
1603                  *
1604                  * In the event of identification failure through Netlink,
1605                  * try again through sysfs, then:
1606                  *
1607                  * 1. A single IB device matches (nd == 1) with single
1608                  *    port (np=0/1) and is not a representor, assume
1609                  *    no switch support.
1610                  *
1611                  * 2. Otherwise no safe assumptions can be made;
1612                  *    complain louder and bail out.
1613                  */
1614                 for (i = 0; i != nd; ++i) {
1615                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1616                         list[ns].max_port = 1;
1617                         list[ns].phys_port = 1;
1618                         list[ns].phys_dev = ibv_match[i];
1619                         list[ns].eth_dev = NULL;
1620                         list[ns].pci_dev = pci_dev;
1621                         list[ns].pf_bond = -1;
1622                         list[ns].ifindex = 0;
1623                         if (nl_rdma >= 0)
1624                                 list[ns].ifindex = mlx5_nl_ifindex
1625                                 (nl_rdma,
1626                                 mlx5_os_get_dev_device_name
1627                                                 (list[ns].phys_dev), 1);
1628                         if (!list[ns].ifindex) {
1629                                 char ifname[IF_NAMESIZE];
1630
1631                                 /*
1632                                  * Netlink failed, it may happen with old
1633                                  * ib_core kernel driver (before 4.16).
1634                                  * We can assume there is old driver because
1635                                  * here we are processing single ports IB
1636                                  * devices. Let's try sysfs to retrieve
1637                                  * the ifindex. The method works for
1638                                  * master device only.
1639                                  */
1640                                 if (nd > 1) {
1641                                         /*
1642                                          * Multiple devices found, assume
1643                                          * representors, can not distinguish
1644                                          * master/representor and retrieve
1645                                          * ifindex via sysfs.
1646                                          */
1647                                         continue;
1648                                 }
1649                                 ret = mlx5_get_ifname_sysfs
1650                                         (ibv_match[i]->ibdev_path, ifname);
1651                                 if (!ret)
1652                                         list[ns].ifindex =
1653                                                 if_nametoindex(ifname);
1654                                 if (!list[ns].ifindex) {
1655                                         /*
1656                                          * No network interface index found
1657                                          * for the specified device, it means
1658                                          * there it is neither representor
1659                                          * nor master.
1660                                          */
1661                                         continue;
1662                                 }
1663                         }
1664                         ret = -1;
1665                         if (nl_route >= 0)
1666                                 ret = mlx5_nl_switch_info
1667                                                (nl_route,
1668                                                 list[ns].ifindex,
1669                                                 &list[ns].info);
1670                         if (ret || (!list[ns].info.representor &&
1671                                     !list[ns].info.master)) {
1672                                 /*
1673                                  * We failed to recognize representors with
1674                                  * Netlink, let's try to perform the task
1675                                  * with sysfs.
1676                                  */
1677                                 ret =  mlx5_sysfs_switch_info
1678                                                 (list[ns].ifindex,
1679                                                  &list[ns].info);
1680                         }
1681                         if (!ret && (list[ns].info.representor ^
1682                                      list[ns].info.master)) {
1683                                 ns++;
1684                         } else if ((nd == 1) &&
1685                                    !list[ns].info.representor &&
1686                                    !list[ns].info.master) {
1687                                 /*
1688                                  * Single IB device with
1689                                  * one physical port and
1690                                  * attached network device.
1691                                  * May be SRIOV is not enabled
1692                                  * or there is no representors.
1693                                  */
1694                                 DRV_LOG(INFO, "no E-Switch support detected");
1695                                 ns++;
1696                                 break;
1697                         }
1698                 }
1699                 if (!ns) {
1700                         DRV_LOG(ERR,
1701                                 "unable to recognize master/representors"
1702                                 " on the multiple IB devices");
1703                         rte_errno = ENOENT;
1704                         ret = -rte_errno;
1705                         goto exit;
1706                 }
1707         }
1708         MLX5_ASSERT(ns);
1709         /*
1710          * Sort list to probe devices in natural order for users convenience
1711          * (i.e. master first, then representors from lowest to highest ID).
1712          */
1713         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1714         /* Default configuration. */
1715         dev_config = (struct mlx5_dev_config){
1716                 .hw_padding = 0,
1717                 .mps = MLX5_ARG_UNSET,
1718                 .dbnc = MLX5_ARG_UNSET,
1719                 .rx_vec_en = 1,
1720                 .txq_inline_max = MLX5_ARG_UNSET,
1721                 .txq_inline_min = MLX5_ARG_UNSET,
1722                 .txq_inline_mpw = MLX5_ARG_UNSET,
1723                 .txqs_inline = MLX5_ARG_UNSET,
1724                 .vf_nl_en = 1,
1725                 .mr_ext_memseg_en = 1,
1726                 .mprq = {
1727                         .enabled = 0, /* Disabled by default. */
1728                         .stride_num_n = 0,
1729                         .stride_size_n = 0,
1730                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1731                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1732                 },
1733                 .dv_esw_en = 1,
1734                 .dv_flow_en = 1,
1735                 .log_hp_size = MLX5_ARG_UNSET,
1736         };
1737         /* Device specific configuration. */
1738         switch (pci_dev->id.device_id) {
1739         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1740         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1741         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1742         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1743         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1744         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1745         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1746                 dev_config.vf = 1;
1747                 break;
1748         default:
1749                 break;
1750         }
1751         for (i = 0; i != ns; ++i) {
1752                 uint32_t restore;
1753
1754                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1755                                                  &list[i],
1756                                                  dev_config);
1757                 if (!list[i].eth_dev) {
1758                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1759                                 break;
1760                         /* Device is disabled or already spawned. Ignore it. */
1761                         continue;
1762                 }
1763                 restore = list[i].eth_dev->data->dev_flags;
1764                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1765                 /* Restore non-PCI flags cleared by the above call. */
1766                 list[i].eth_dev->data->dev_flags |= restore;
1767                 rte_eth_dev_probing_finish(list[i].eth_dev);
1768         }
1769         if (i != ns) {
1770                 DRV_LOG(ERR,
1771                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1772                         " encountering an error: %s",
1773                         pci_dev->addr.domain, pci_dev->addr.bus,
1774                         pci_dev->addr.devid, pci_dev->addr.function,
1775                         strerror(rte_errno));
1776                 ret = -rte_errno;
1777                 /* Roll back. */
1778                 while (i--) {
1779                         if (!list[i].eth_dev)
1780                                 continue;
1781                         mlx5_dev_close(list[i].eth_dev);
1782                         /* mac_addrs must not be freed because in dev_private */
1783                         list[i].eth_dev->data->mac_addrs = NULL;
1784                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1785                 }
1786                 /* Restore original error. */
1787                 rte_errno = -ret;
1788         } else {
1789                 ret = 0;
1790         }
1791 exit:
1792         /*
1793          * Do the routine cleanup:
1794          * - close opened Netlink sockets
1795          * - free allocated spawn data array
1796          * - free the Infiniband device list
1797          */
1798         if (nl_rdma >= 0)
1799                 close(nl_rdma);
1800         if (nl_route >= 0)
1801                 close(nl_route);
1802         if (list)
1803                 rte_free(list);
1804         MLX5_ASSERT(ibv_list);
1805         mlx5_glue->free_device_list(ibv_list);
1806         return ret;
1807 }
1808
1809 static int
1810 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1811 {
1812         char *env;
1813         int value;
1814
1815         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1816         /* Get environment variable to store. */
1817         env = getenv(MLX5_SHUT_UP_BF);
1818         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1819         if (config->dbnc == MLX5_ARG_UNSET)
1820                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1821         else
1822                 setenv(MLX5_SHUT_UP_BF,
1823                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1824         return value;
1825 }
1826
1827 static void
1828 mlx5_restore_doorbell_mapping_env(int value)
1829 {
1830         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1831         /* Restore the original environment variable state. */
1832         if (value == MLX5_ARG_UNSET)
1833                 unsetenv(MLX5_SHUT_UP_BF);
1834         else
1835                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1836 }
1837
1838 /**
1839  * Extract pdn of PD object using DV API.
1840  *
1841  * @param[in] pd
1842  *   Pointer to the verbs PD object.
1843  * @param[out] pdn
1844  *   Pointer to the PD object number variable.
1845  *
1846  * @return
1847  *   0 on success, error value otherwise.
1848  */
1849 int
1850 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1851 {
1852 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1853         struct mlx5dv_obj obj;
1854         struct mlx5dv_pd pd_info;
1855         int ret = 0;
1856
1857         obj.pd.in = pd;
1858         obj.pd.out = &pd_info;
1859         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1860         if (ret) {
1861                 DRV_LOG(DEBUG, "Fail to get PD object info");
1862                 return ret;
1863         }
1864         *pdn = pd_info.pdn;
1865         return 0;
1866 #else
1867         (void)pd;
1868         (void)pdn;
1869         return -ENOTSUP;
1870 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1871 }
1872
1873 /**
1874  * Function API to open IB device.
1875  *
1876  * This function calls the Linux glue APIs to open a device.
1877  *
1878  * @param[in] spawn
1879  *   Pointer to the IB device attributes (name, port, etc).
1880  * @param[out] config
1881  *   Pointer to device configuration structure.
1882  * @param[out] sh
1883  *   Pointer to shared context structure.
1884  *
1885  * @return
1886  *   0 on success, a positive error value otherwise.
1887  */
1888 int
1889 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1890                      const struct mlx5_dev_config *config,
1891                      struct mlx5_dev_ctx_shared *sh)
1892 {
1893         int dbmap_env;
1894         int err = 0;
1895
1896         sh->numa_node = spawn->pci_dev->device.numa_node;
1897         pthread_mutex_init(&sh->txpp.mutex, NULL);
1898         /*
1899          * Configure environment variable "MLX5_BF_SHUT_UP"
1900          * before the device creation. The rdma_core library
1901          * checks the variable at device creation and
1902          * stores the result internally.
1903          */
1904         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1905         /* Try to open IB device with DV first, then usual Verbs. */
1906         errno = 0;
1907         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1908         if (sh->ctx) {
1909                 sh->devx = 1;
1910                 DRV_LOG(DEBUG, "DevX is supported");
1911                 /* The device is created, no need for environment. */
1912                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1913         } else {
1914                 /* The environment variable is still configured. */
1915                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1916                 err = errno ? errno : ENODEV;
1917                 /*
1918                  * The environment variable is not needed anymore,
1919                  * all device creation attempts are completed.
1920                  */
1921                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1922                 if (!sh->ctx)
1923                         return err;
1924                 DRV_LOG(DEBUG, "DevX is NOT supported");
1925                 err = 0;
1926         }
1927         return err;
1928 }
1929
1930 /**
1931  * Install shared asynchronous device events handler.
1932  * This function is implemented to support event sharing
1933  * between multiple ports of single IB device.
1934  *
1935  * @param sh
1936  *   Pointer to mlx5_dev_ctx_shared object.
1937  */
1938 void
1939 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1940 {
1941         int ret;
1942         int flags;
1943
1944         sh->intr_handle.fd = -1;
1945         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1946         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1947                     F_SETFL, flags | O_NONBLOCK);
1948         if (ret) {
1949                 DRV_LOG(INFO, "failed to change file descriptor async event"
1950                         " queue");
1951         } else {
1952                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1953                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1954                 if (rte_intr_callback_register(&sh->intr_handle,
1955                                         mlx5_dev_interrupt_handler, sh)) {
1956                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1957                         sh->intr_handle.fd = -1;
1958                 }
1959         }
1960         if (sh->devx) {
1961 #ifdef HAVE_IBV_DEVX_ASYNC
1962                 sh->intr_handle_devx.fd = -1;
1963                 sh->devx_comp =
1964                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1965                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1966                 if (!devx_comp) {
1967                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1968                         return;
1969                 }
1970                 flags = fcntl(devx_comp->fd, F_GETFL);
1971                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1972                 if (ret) {
1973                         DRV_LOG(INFO, "failed to change file descriptor"
1974                                 " devx comp");
1975                         return;
1976                 }
1977                 sh->intr_handle_devx.fd = devx_comp->fd;
1978                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1979                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1980                                         mlx5_dev_interrupt_handler_devx, sh)) {
1981                         DRV_LOG(INFO, "Fail to install the devx shared"
1982                                 " interrupt.");
1983                         sh->intr_handle_devx.fd = -1;
1984                 }
1985 #endif /* HAVE_IBV_DEVX_ASYNC */
1986         }
1987 }
1988
1989 /**
1990  * Uninstall shared asynchronous device events handler.
1991  * This function is implemented to support event sharing
1992  * between multiple ports of single IB device.
1993  *
1994  * @param dev
1995  *   Pointer to mlx5_dev_ctx_shared object.
1996  */
1997 void
1998 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
1999 {
2000         if (sh->intr_handle.fd >= 0)
2001                 mlx5_intr_callback_unregister(&sh->intr_handle,
2002                                               mlx5_dev_interrupt_handler, sh);
2003 #ifdef HAVE_IBV_DEVX_ASYNC
2004         if (sh->intr_handle_devx.fd >= 0)
2005                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2006                                   mlx5_dev_interrupt_handler_devx, sh);
2007         if (sh->devx_comp)
2008                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2009 #endif
2010 }
2011
2012 /**
2013  * Read statistics by a named counter.
2014  *
2015  * @param[in] priv
2016  *   Pointer to the private device data structure.
2017  * @param[in] ctr_name
2018  *   Pointer to the name of the statistic counter to read
2019  * @param[out] stat
2020  *   Pointer to read statistic value.
2021  * @return
2022  *   0 on success and stat is valud, 1 if failed to read the value
2023  *   rte_errno is set.
2024  *
2025  */
2026 int
2027 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2028                       uint64_t *stat)
2029 {
2030         int fd;
2031
2032         if (priv->sh) {
2033                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2034                           priv->sh->ibdev_path,
2035                           priv->dev_port,
2036                           ctr_name);
2037                 fd = open(path, O_RDONLY);
2038                 if (fd != -1) {
2039                         char buf[21] = {'\0'};
2040                         ssize_t n = read(fd, buf, sizeof(buf));
2041
2042                         close(fd);
2043                         if (n != -1) {
2044                                 *stat = strtoull(buf, NULL, 10);
2045                                 return 0;
2046                         }
2047                 }
2048         }
2049         *stat = 0;
2050         return 1;
2051 }
2052
2053 /**
2054  * Read device counters table.
2055  *
2056  * @param dev
2057  *   Pointer to Ethernet device.
2058  * @param[out] stats
2059  *   Counters table output buffer.
2060  *
2061  * @return
2062  *   0 on success and stats is filled, negative errno value otherwise and
2063  *   rte_errno is set.
2064  */
2065 int
2066 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
2067 {
2068         struct mlx5_priv *priv = dev->data->dev_private;
2069         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2070         unsigned int i;
2071         struct ifreq ifr;
2072         unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
2073         unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
2074         struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
2075         int ret;
2076
2077         et_stats->cmd = ETHTOOL_GSTATS;
2078         et_stats->n_stats = xstats_ctrl->stats_n;
2079         ifr.ifr_data = (caddr_t)et_stats;
2080         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2081         if (ret) {
2082                 DRV_LOG(WARNING,
2083                         "port %u unable to read statistic values from device",
2084                         dev->data->port_id);
2085                 return ret;
2086         }
2087         for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2088                 if (xstats_ctrl->info[i].dev) {
2089                         ret = mlx5_os_read_dev_stat(priv,
2090                                             xstats_ctrl->info[i].ctr_name,
2091                                             &stats[i]);
2092                         /* return last xstats counter if fail to read. */
2093                         if (ret == 0)
2094                                 xstats_ctrl->xstats[i] = stats[i];
2095                         else
2096                                 stats[i] = xstats_ctrl->xstats[i];
2097                 } else {
2098                         stats[i] = (uint64_t)
2099                                 et_stats->data[xstats_ctrl->dev_table_idx[i]];
2100                 }
2101         }
2102         return 0;
2103 }
2104
2105 /**
2106  * Query the number of statistics provided by ETHTOOL.
2107  *
2108  * @param dev
2109  *   Pointer to Ethernet device.
2110  *
2111  * @return
2112  *   Number of statistics on success, negative errno value otherwise and
2113  *   rte_errno is set.
2114  */
2115 int
2116 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2117 {
2118         struct ethtool_drvinfo drvinfo;
2119         struct ifreq ifr;
2120         int ret;
2121
2122         drvinfo.cmd = ETHTOOL_GDRVINFO;
2123         ifr.ifr_data = (caddr_t)&drvinfo;
2124         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2125         if (ret) {
2126                 DRV_LOG(WARNING, "port %u unable to query number of statistics",
2127                         dev->data->port_id);
2128                 return ret;
2129         }
2130         return drvinfo.n_stats;
2131 }
2132
2133 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2134         {
2135                 .dpdk_name = "rx_port_unicast_bytes",
2136                 .ctr_name = "rx_vport_unicast_bytes",
2137         },
2138         {
2139                 .dpdk_name = "rx_port_multicast_bytes",
2140                 .ctr_name = "rx_vport_multicast_bytes",
2141         },
2142         {
2143                 .dpdk_name = "rx_port_broadcast_bytes",
2144                 .ctr_name = "rx_vport_broadcast_bytes",
2145         },
2146         {
2147                 .dpdk_name = "rx_port_unicast_packets",
2148                 .ctr_name = "rx_vport_unicast_packets",
2149         },
2150         {
2151                 .dpdk_name = "rx_port_multicast_packets",
2152                 .ctr_name = "rx_vport_multicast_packets",
2153         },
2154         {
2155                 .dpdk_name = "rx_port_broadcast_packets",
2156                 .ctr_name = "rx_vport_broadcast_packets",
2157         },
2158         {
2159                 .dpdk_name = "tx_port_unicast_bytes",
2160                 .ctr_name = "tx_vport_unicast_bytes",
2161         },
2162         {
2163                 .dpdk_name = "tx_port_multicast_bytes",
2164                 .ctr_name = "tx_vport_multicast_bytes",
2165         },
2166         {
2167                 .dpdk_name = "tx_port_broadcast_bytes",
2168                 .ctr_name = "tx_vport_broadcast_bytes",
2169         },
2170         {
2171                 .dpdk_name = "tx_port_unicast_packets",
2172                 .ctr_name = "tx_vport_unicast_packets",
2173         },
2174         {
2175                 .dpdk_name = "tx_port_multicast_packets",
2176                 .ctr_name = "tx_vport_multicast_packets",
2177         },
2178         {
2179                 .dpdk_name = "tx_port_broadcast_packets",
2180                 .ctr_name = "tx_vport_broadcast_packets",
2181         },
2182         {
2183                 .dpdk_name = "rx_wqe_err",
2184                 .ctr_name = "rx_wqe_err",
2185         },
2186         {
2187                 .dpdk_name = "rx_crc_errors_phy",
2188                 .ctr_name = "rx_crc_errors_phy",
2189         },
2190         {
2191                 .dpdk_name = "rx_in_range_len_errors_phy",
2192                 .ctr_name = "rx_in_range_len_errors_phy",
2193         },
2194         {
2195                 .dpdk_name = "rx_symbol_err_phy",
2196                 .ctr_name = "rx_symbol_err_phy",
2197         },
2198         {
2199                 .dpdk_name = "tx_errors_phy",
2200                 .ctr_name = "tx_errors_phy",
2201         },
2202         {
2203                 .dpdk_name = "rx_out_of_buffer",
2204                 .ctr_name = "out_of_buffer",
2205                 .dev = 1,
2206         },
2207         {
2208                 .dpdk_name = "tx_packets_phy",
2209                 .ctr_name = "tx_packets_phy",
2210         },
2211         {
2212                 .dpdk_name = "rx_packets_phy",
2213                 .ctr_name = "rx_packets_phy",
2214         },
2215         {
2216                 .dpdk_name = "tx_discards_phy",
2217                 .ctr_name = "tx_discards_phy",
2218         },
2219         {
2220                 .dpdk_name = "rx_discards_phy",
2221                 .ctr_name = "rx_discards_phy",
2222         },
2223         {
2224                 .dpdk_name = "tx_bytes_phy",
2225                 .ctr_name = "tx_bytes_phy",
2226         },
2227         {
2228                 .dpdk_name = "rx_bytes_phy",
2229                 .ctr_name = "rx_bytes_phy",
2230         },
2231         /* Representor only */
2232         {
2233                 .dpdk_name = "rx_packets",
2234                 .ctr_name = "vport_rx_packets",
2235         },
2236         {
2237                 .dpdk_name = "rx_bytes",
2238                 .ctr_name = "vport_rx_bytes",
2239         },
2240         {
2241                 .dpdk_name = "tx_packets",
2242                 .ctr_name = "vport_tx_packets",
2243         },
2244         {
2245                 .dpdk_name = "tx_bytes",
2246                 .ctr_name = "vport_tx_bytes",
2247         },
2248 };
2249
2250 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2251
2252 /**
2253  * Init the structures to read device counters.
2254  *
2255  * @param dev
2256  *   Pointer to Ethernet device.
2257  */
2258 void
2259 mlx5_os_stats_init(struct rte_eth_dev *dev)
2260 {
2261         struct mlx5_priv *priv = dev->data->dev_private;
2262         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2263         struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2264         unsigned int i;
2265         unsigned int j;
2266         struct ifreq ifr;
2267         struct ethtool_gstrings *strings = NULL;
2268         unsigned int dev_stats_n;
2269         unsigned int str_sz;
2270         int ret;
2271
2272         /* So that it won't aggregate for each init. */
2273         xstats_ctrl->mlx5_stats_n = 0;
2274         ret = mlx5_os_get_stats_n(dev);
2275         if (ret < 0) {
2276                 DRV_LOG(WARNING, "port %u no extended statistics available",
2277                         dev->data->port_id);
2278                 return;
2279         }
2280         dev_stats_n = ret;
2281         /* Allocate memory to grab stat names and values. */
2282         str_sz = dev_stats_n * ETH_GSTRING_LEN;
2283         strings = (struct ethtool_gstrings *)
2284                   rte_malloc("xstats_strings",
2285                              str_sz + sizeof(struct ethtool_gstrings), 0);
2286         if (!strings) {
2287                 DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2288                      dev->data->port_id);
2289                 return;
2290         }
2291         strings->cmd = ETHTOOL_GSTRINGS;
2292         strings->string_set = ETH_SS_STATS;
2293         strings->len = dev_stats_n;
2294         ifr.ifr_data = (caddr_t)strings;
2295         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2296         if (ret) {
2297                 DRV_LOG(WARNING, "port %u unable to get statistic names",
2298                         dev->data->port_id);
2299                 goto free;
2300         }
2301         for (i = 0; i != dev_stats_n; ++i) {
2302                 const char *curr_string = (const char *)
2303                         &strings->data[i * ETH_GSTRING_LEN];
2304
2305                 for (j = 0; j != xstats_n; ++j) {
2306                         if (!strcmp(mlx5_counters_init[j].ctr_name,
2307                                     curr_string)) {
2308                                 unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2309
2310                                 xstats_ctrl->dev_table_idx[idx] = i;
2311                                 xstats_ctrl->info[idx] = mlx5_counters_init[j];
2312                                 break;
2313                         }
2314                 }
2315         }
2316         /* Add dev counters. */
2317         for (i = 0; i != xstats_n; ++i) {
2318                 if (mlx5_counters_init[i].dev) {
2319                         unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2320
2321                         xstats_ctrl->info[idx] = mlx5_counters_init[i];
2322                         xstats_ctrl->hw_stats[idx] = 0;
2323                 }
2324         }
2325         MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2326         xstats_ctrl->stats_n = dev_stats_n;
2327         /* Copy to base at first time. */
2328         ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2329         if (ret)
2330                 DRV_LOG(ERR, "port %u cannot read device counters: %s",
2331                         dev->data->port_id, strerror(rte_errno));
2332         mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2333         stats_ctrl->imissed = 0;
2334 free:
2335         rte_free(strings);
2336 }
2337
2338 /**
2339  * Set the reg_mr and dereg_mr call backs
2340  *
2341  * @param reg_mr_cb[out]
2342  *   Pointer to reg_mr func
2343  * @param dereg_mr_cb[out]
2344  *   Pointer to dereg_mr func
2345  *
2346  */
2347 void
2348 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2349                       mlx5_dereg_mr_t *dereg_mr_cb)
2350 {
2351         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2352         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2353 }
2354
2355 const struct eth_dev_ops mlx5_os_dev_ops = {
2356         .dev_configure = mlx5_dev_configure,
2357         .dev_start = mlx5_dev_start,
2358         .dev_stop = mlx5_dev_stop,
2359         .dev_set_link_down = mlx5_set_link_down,
2360         .dev_set_link_up = mlx5_set_link_up,
2361         .dev_close = mlx5_dev_close,
2362         .promiscuous_enable = mlx5_promiscuous_enable,
2363         .promiscuous_disable = mlx5_promiscuous_disable,
2364         .allmulticast_enable = mlx5_allmulticast_enable,
2365         .allmulticast_disable = mlx5_allmulticast_disable,
2366         .link_update = mlx5_link_update,
2367         .stats_get = mlx5_stats_get,
2368         .stats_reset = mlx5_stats_reset,
2369         .xstats_get = mlx5_xstats_get,
2370         .xstats_reset = mlx5_xstats_reset,
2371         .xstats_get_names = mlx5_xstats_get_names,
2372         .fw_version_get = mlx5_fw_version_get,
2373         .dev_infos_get = mlx5_dev_infos_get,
2374         .read_clock = mlx5_txpp_read_clock,
2375         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2376         .vlan_filter_set = mlx5_vlan_filter_set,
2377         .rx_queue_setup = mlx5_rx_queue_setup,
2378         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2379         .tx_queue_setup = mlx5_tx_queue_setup,
2380         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2381         .rx_queue_release = mlx5_rx_queue_release,
2382         .tx_queue_release = mlx5_tx_queue_release,
2383         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2384         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2385         .mac_addr_remove = mlx5_mac_addr_remove,
2386         .mac_addr_add = mlx5_mac_addr_add,
2387         .mac_addr_set = mlx5_mac_addr_set,
2388         .set_mc_addr_list = mlx5_set_mc_addr_list,
2389         .mtu_set = mlx5_dev_set_mtu,
2390         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2391         .vlan_offload_set = mlx5_vlan_offload_set,
2392         .reta_update = mlx5_dev_rss_reta_update,
2393         .reta_query = mlx5_dev_rss_reta_query,
2394         .rss_hash_update = mlx5_rss_hash_update,
2395         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2396         .filter_ctrl = mlx5_dev_filter_ctrl,
2397         .rx_descriptor_status = mlx5_rx_descriptor_status,
2398         .tx_descriptor_status = mlx5_tx_descriptor_status,
2399         .rxq_info_get = mlx5_rxq_info_get,
2400         .txq_info_get = mlx5_txq_info_get,
2401         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2402         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2403         .rx_queue_count = mlx5_rx_queue_count,
2404         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2405         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2406         .is_removed = mlx5_is_removed,
2407         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2408         .get_module_info = mlx5_get_module_info,
2409         .get_module_eeprom = mlx5_get_module_eeprom,
2410         .hairpin_cap_get = mlx5_hairpin_cap_get,
2411         .mtr_ops_get = mlx5_flow_meter_ops_get,
2412 };
2413
2414 /* Available operations from secondary process. */
2415 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2416         .stats_get = mlx5_stats_get,
2417         .stats_reset = mlx5_stats_reset,
2418         .xstats_get = mlx5_xstats_get,
2419         .xstats_reset = mlx5_xstats_reset,
2420         .xstats_get_names = mlx5_xstats_get_names,
2421         .fw_version_get = mlx5_fw_version_get,
2422         .dev_infos_get = mlx5_dev_infos_get,
2423         .read_clock = mlx5_txpp_read_clock,
2424         .rx_descriptor_status = mlx5_rx_descriptor_status,
2425         .tx_descriptor_status = mlx5_tx_descriptor_status,
2426         .rxq_info_get = mlx5_rxq_info_get,
2427         .txq_info_get = mlx5_txq_info_get,
2428         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2429         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2430         .get_module_info = mlx5_get_module_info,
2431         .get_module_eeprom = mlx5_get_module_eeprom,
2432 };
2433
2434 /* Available operations in flow isolated mode. */
2435 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2436         .dev_configure = mlx5_dev_configure,
2437         .dev_start = mlx5_dev_start,
2438         .dev_stop = mlx5_dev_stop,
2439         .dev_set_link_down = mlx5_set_link_down,
2440         .dev_set_link_up = mlx5_set_link_up,
2441         .dev_close = mlx5_dev_close,
2442         .promiscuous_enable = mlx5_promiscuous_enable,
2443         .promiscuous_disable = mlx5_promiscuous_disable,
2444         .allmulticast_enable = mlx5_allmulticast_enable,
2445         .allmulticast_disable = mlx5_allmulticast_disable,
2446         .link_update = mlx5_link_update,
2447         .stats_get = mlx5_stats_get,
2448         .stats_reset = mlx5_stats_reset,
2449         .xstats_get = mlx5_xstats_get,
2450         .xstats_reset = mlx5_xstats_reset,
2451         .xstats_get_names = mlx5_xstats_get_names,
2452         .fw_version_get = mlx5_fw_version_get,
2453         .dev_infos_get = mlx5_dev_infos_get,
2454         .read_clock = mlx5_txpp_read_clock,
2455         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2456         .vlan_filter_set = mlx5_vlan_filter_set,
2457         .rx_queue_setup = mlx5_rx_queue_setup,
2458         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2459         .tx_queue_setup = mlx5_tx_queue_setup,
2460         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2461         .rx_queue_release = mlx5_rx_queue_release,
2462         .tx_queue_release = mlx5_tx_queue_release,
2463         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2464         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2465         .mac_addr_remove = mlx5_mac_addr_remove,
2466         .mac_addr_add = mlx5_mac_addr_add,
2467         .mac_addr_set = mlx5_mac_addr_set,
2468         .set_mc_addr_list = mlx5_set_mc_addr_list,
2469         .mtu_set = mlx5_dev_set_mtu,
2470         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2471         .vlan_offload_set = mlx5_vlan_offload_set,
2472         .filter_ctrl = mlx5_dev_filter_ctrl,
2473         .rx_descriptor_status = mlx5_rx_descriptor_status,
2474         .tx_descriptor_status = mlx5_tx_descriptor_status,
2475         .rxq_info_get = mlx5_rxq_info_get,
2476         .txq_info_get = mlx5_txq_info_get,
2477         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2478         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2479         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2480         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2481         .is_removed = mlx5_is_removed,
2482         .get_module_info = mlx5_get_module_info,
2483         .get_module_eeprom = mlx5_get_module_eeprom,
2484         .hairpin_cap_get = mlx5_hairpin_cap_get,
2485         .mtr_ops_get = mlx5_flow_meter_ops_get,
2486 };