drivers/net/mlx5/linux/mlx5_os.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2020 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stddef.h>
   7 #include <unistd.h>
   8 #include <string.h>
   9 #include <stdint.h>
  10 #include <stdlib.h>
  11 #include <errno.h>
  12 #include <net/if.h>
  13 #include <sys/mman.h>
  14 #include <linux/rtnetlink.h>
  15 #include <linux/sockios.h>
  16 #include <linux/ethtool.h>
  17 #include <fcntl.h>
  18
  19 /* Verbs header. */
  20 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  21 #ifdef PEDANTIC
  22 #pragma GCC diagnostic ignored "-Wpedantic"
  23 #endif
  24 #include <infiniband/verbs.h>
  25 #ifdef PEDANTIC
  26 #pragma GCC diagnostic error "-Wpedantic"
  27 #endif
  28
  29 #include <rte_malloc.h>
  30 #include <rte_ethdev_driver.h>
  31 #include <rte_ethdev_pci.h>
  32 #include <rte_pci.h>
  33 #include <rte_bus_pci.h>
  34 #include <rte_common.h>
  35 #include <rte_kvargs.h>
  36 #include <rte_rwlock.h>
  37 #include <rte_spinlock.h>
  38 #include <rte_string_fns.h>
  39 #include <rte_alarm.h>
  40
  41 #include <mlx5_glue.h>
  42 #include <mlx5_devx_cmds.h>
  43 #include <mlx5_common.h>
  44 #include <mlx5_common_mp.h>
  45 #include <mlx5_common_mr.h>
  46
  47 #include "mlx5_defs.h"
  48 #include "mlx5.h"
  49 #include "mlx5_common_os.h"
  50 #include "mlx5_utils.h"
  51 #include "mlx5_rxtx.h"
  52 #include "mlx5_autoconf.h"
  53 #include "mlx5_mr.h"
  54 #include "mlx5_flow.h"
  55 #include "rte_pmd_mlx5.h"
  56 #include "mlx5_verbs.h"
  57
  58 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  59
  60 #ifndef HAVE_IBV_MLX5_MOD_MPW
  61 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  62 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  63 #endif
  64
  65 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  66 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  67 #endif
  68
  69 /**
  70  * Get mlx5 device attributes. The glue function query_device_ex() is called
  71  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  72  * device attributes from the glue out parameter.
  73  *
  74  * @param dev
  75  *   Pointer to ibv context.
  76  *
  77  * @param device_attr
  78  *   Pointer to mlx5 device attributes.
  79  *
  80  * @return
  81  *   0 on success, non zero error number otherwise
  82  */
  83 int
  84 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
  85 {
  86         int err;
  87         struct ibv_device_attr_ex attr_ex;
  88         memset(device_attr, 0, sizeof(*device_attr));
  89         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
  90         if (err)
  91                 return err;
  92
  93         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
  94         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
  95         device_attr->max_sge = attr_ex.orig_attr.max_sge;
  96         device_attr->max_cq = attr_ex.orig_attr.max_cq;
  97         device_attr->max_qp = attr_ex.orig_attr.max_qp;
  98         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
  99         device_attr->max_rwq_indirection_table_size =
 100                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 101         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 102         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 103
 104         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 105         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 106         if (err)
 107                 return err;
 108
 109         device_attr->flags = dv_attr.flags;
 110         device_attr->comp_mask = dv_attr.comp_mask;
 111 #ifdef HAVE_IBV_MLX5_MOD_SWP
 112         device_attr->sw_parsing_offloads =
 113                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 114 #endif
 115         device_attr->min_single_stride_log_num_of_bytes =
 116                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 117         device_attr->max_single_stride_log_num_of_bytes =
 118                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 119         device_attr->min_single_wqe_log_num_of_strides =
 120                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 121         device_attr->max_single_wqe_log_num_of_strides =
 122                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 123         device_attr->stride_supported_qpts =
 124                 dv_attr.striding_rq_caps.supported_qpts;
 125 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 126         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 127 #endif
 128
 129         return err;
 130 }
 131
 132 /**
 133  * Verbs callback to allocate a memory. This function should allocate the space
 134  * according to the size provided residing inside a huge page.
 135  * Please note that all allocation must respect the alignment from libmlx5
 136  * (i.e. currently sysconf(_SC_PAGESIZE)).
 137  *
 138  * @param[in] size
 139  *   The size in bytes of the memory to allocate.
 140  * @param[in] data
 141  *   A pointer to the callback data.
 142  *
 143  * @return
 144  *   Allocated buffer, NULL otherwise and rte_errno is set.
 145  */
 146 static void *
 147 mlx5_alloc_verbs_buf(size_t size, void *data)
 148 {
 149         struct mlx5_priv *priv = data;
 150         void *ret;
 151         size_t alignment = sysconf(_SC_PAGESIZE);
 152         unsigned int socket = SOCKET_ID_ANY;
 153
 154         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 155                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 156
 157                 socket = ctrl->socket;
 158         } else if (priv->verbs_alloc_ctx.type ==
 159                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 160                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 161
 162                 socket = ctrl->socket;
 163         }
 164         MLX5_ASSERT(data != NULL);
 165         ret = rte_malloc_socket(__func__, size, alignment, socket);
 166         if (!ret && size)
 167                 rte_errno = ENOMEM;
 168         return ret;
 169 }
 170
 171 /**
 172  * Verbs callback to free a memory.
 173  *
 174  * @param[in] ptr
 175  *   A pointer to the memory to free.
 176  * @param[in] data
 177  *   A pointer to the callback data.
 178  */
 179 static void
 180 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 181 {
 182         MLX5_ASSERT(data != NULL);
 183         rte_free(ptr);
 184 }
 185
 186 /**
 187  * Initialize DR related data within private structure.
 188  * Routine checks the reference counter and does actual
 189  * resources creation/initialization only if counter is zero.
 190  *
 191  * @param[in] priv
 192  *   Pointer to the private device data structure.
 193  *
 194  * @return
 195  *   Zero on success, positive error code otherwise.
 196  */
 197 static int
 198 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 199 {
 200         struct mlx5_dev_ctx_shared *sh = priv->sh;
 201         char s[MLX5_HLIST_NAMESIZE];
 202         int err = 0;
 203
 204         if (!sh->flow_tbls)
 205                 err = mlx5_alloc_table_hash_list(priv);
 206         else
 207                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 208                         (void *)sh->flow_tbls);
 209         if (err)
 210                 return err;
 211         /* Create tags hash list table. */
 212         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 213         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 214         if (!sh->tag_table) {
 215                 DRV_LOG(ERR, "tags with hash creation failed.");
 216                 err = ENOMEM;
 217                 goto error;
 218         }
 219 #ifdef HAVE_MLX5DV_DR
 220         void *domain;
 221
 222         if (sh->dv_refcnt) {
 223                 /* Shared DV/DR structures is already initialized. */
 224                 sh->dv_refcnt++;
 225                 priv->dr_shared = 1;
 226                 return 0;
 227         }
 228         /* Reference counter is zero, we should initialize structures. */
 229         domain = mlx5_glue->dr_create_domain(sh->ctx,
 230                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 231         if (!domain) {
 232                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 233                 err = errno;
 234                 goto error;
 235         }
 236         sh->rx_domain = domain;
 237         domain = mlx5_glue->dr_create_domain(sh->ctx,
 238                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 239         if (!domain) {
 240                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 241                 err = errno;
 242                 goto error;
 243         }
 244         pthread_mutex_init(&sh->dv_mutex, NULL);
 245         sh->tx_domain = domain;
 246 #ifdef HAVE_MLX5DV_DR_ESWITCH
 247         if (priv->config.dv_esw_en) {
 248                 domain  = mlx5_glue->dr_create_domain
 249                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 250                 if (!domain) {
 251                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 252                         err = errno;
 253                         goto error;
 254                 }
 255                 sh->fdb_domain = domain;
 256                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 257         }
 258 #endif
 259         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 260                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 261                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 262                 if (sh->fdb_domain)
 263                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 264         }
 265         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 266 #endif /* HAVE_MLX5DV_DR */
 267         sh->dv_refcnt++;
 268         priv->dr_shared = 1;
 269         return 0;
 270 error:
 271         /* Rollback the created objects. */
 272         if (sh->rx_domain) {
 273                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 274                 sh->rx_domain = NULL;
 275         }
 276         if (sh->tx_domain) {
 277                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 278                 sh->tx_domain = NULL;
 279         }
 280         if (sh->fdb_domain) {
 281                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 282                 sh->fdb_domain = NULL;
 283         }
 284         if (sh->esw_drop_action) {
 285                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 286                 sh->esw_drop_action = NULL;
 287         }
 288         if (sh->pop_vlan_action) {
 289                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 290                 sh->pop_vlan_action = NULL;
 291         }
 292         if (sh->tag_table) {
 293                 /* tags should be destroyed with flow before. */
 294                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 295                 sh->tag_table = NULL;
 296         }
 297         mlx5_free_table_hash_list(priv);
 298         return err;
 299 }
 300
 301 /**
 302  * Destroy DR related data within private structure.
 303  *
 304  * @param[in] priv
 305  *   Pointer to the private device data structure.
 306  */
 307 void
 308 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 309 {
 310         struct mlx5_dev_ctx_shared *sh;
 311
 312         if (!priv->dr_shared)
 313                 return;
 314         priv->dr_shared = 0;
 315         sh = priv->sh;
 316         MLX5_ASSERT(sh);
 317 #ifdef HAVE_MLX5DV_DR
 318         MLX5_ASSERT(sh->dv_refcnt);
 319         if (sh->dv_refcnt && --sh->dv_refcnt)
 320                 return;
 321         if (sh->rx_domain) {
 322                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 323                 sh->rx_domain = NULL;
 324         }
 325         if (sh->tx_domain) {
 326                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 327                 sh->tx_domain = NULL;
 328         }
 329 #ifdef HAVE_MLX5DV_DR_ESWITCH
 330         if (sh->fdb_domain) {
 331                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 332                 sh->fdb_domain = NULL;
 333         }
 334         if (sh->esw_drop_action) {
 335                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 336                 sh->esw_drop_action = NULL;
 337         }
 338 #endif
 339         if (sh->pop_vlan_action) {
 340                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 341                 sh->pop_vlan_action = NULL;
 342         }
 343         pthread_mutex_destroy(&sh->dv_mutex);
 344 #endif /* HAVE_MLX5DV_DR */
 345         if (sh->tag_table) {
 346                 /* tags should be destroyed with flow before. */
 347                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 348                 sh->tag_table = NULL;
 349         }
 350         mlx5_free_table_hash_list(priv);
 351 }
 352
 353 /**
 354  * Spawn an Ethernet device from Verbs information.
 355  *
 356  * @param dpdk_dev
 357  *   Backing DPDK device.
 358  * @param spawn
 359  *   Verbs device parameters (name, port, switch_info) to spawn.
 360  * @param config
 361  *   Device configuration parameters.
 362  *
 363  * @return
 364  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 365  *   is set. The following errors are defined:
 366  *
 367  *   EBUSY: device is not supposed to be spawned.
 368  *   EEXIST: device is already spawned
 369  */
 370 static struct rte_eth_dev *
 371 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 372                struct mlx5_dev_spawn_data *spawn,
 373                struct mlx5_dev_config config)
 374 {
 375         const struct mlx5_switch_info *switch_info = &spawn->info;
 376         struct mlx5_dev_ctx_shared *sh = NULL;
 377         struct ibv_port_attr port_attr;
 378         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 379         struct rte_eth_dev *eth_dev = NULL;
 380         struct mlx5_priv *priv = NULL;
 381         int err = 0;
 382         unsigned int hw_padding = 0;
 383         unsigned int mps;
 384         unsigned int cqe_comp;
 385         unsigned int cqe_pad = 0;
 386         unsigned int tunnel_en = 0;
 387         unsigned int mpls_en = 0;
 388         unsigned int swp = 0;
 389         unsigned int mprq = 0;
 390         unsigned int mprq_min_stride_size_n = 0;
 391         unsigned int mprq_max_stride_size_n = 0;
 392         unsigned int mprq_min_stride_num_n = 0;
 393         unsigned int mprq_max_stride_num_n = 0;
 394         struct rte_ether_addr mac;
 395         char name[RTE_ETH_NAME_MAX_LEN];
 396         int own_domain_id = 0;
 397         uint16_t port_id;
 398         unsigned int i;
 399 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 400         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 401 #endif
 402
 403         /* Determine if this port representor is supposed to be spawned. */
 404         if (switch_info->representor && dpdk_dev->devargs) {
 405                 struct rte_eth_devargs eth_da;
 406
 407                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 408                 if (err) {
 409                         rte_errno = -err;
 410                         DRV_LOG(ERR, "failed to process device arguments: %s",
 411                                 strerror(rte_errno));
 412                         return NULL;
 413                 }
 414                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 415                         if (eth_da.representor_ports[i] ==
 416                             (uint16_t)switch_info->port_name)
 417                                 break;
 418                 if (i == eth_da.nb_representor_ports) {
 419                         rte_errno = EBUSY;
 420                         return NULL;
 421                 }
 422         }
 423         /* Build device name. */
 424         if (spawn->pf_bond <  0) {
 425                 /* Single device. */
 426                 if (!switch_info->representor)
 427                         strlcpy(name, dpdk_dev->name, sizeof(name));
 428                 else
 429                         snprintf(name, sizeof(name), "%s_representor_%u",
 430                                  dpdk_dev->name, switch_info->port_name);
 431         } else {
 432                 /* Bonding device. */
 433                 if (!switch_info->representor)
 434                         snprintf(name, sizeof(name), "%s_%s",
 435                                  dpdk_dev->name,
 436                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 437                 else
 438                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 439                                  dpdk_dev->name,
 440                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 441                                  switch_info->port_name);
 442         }
 443         /* check if the device is already spawned */
 444         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 445                 rte_errno = EEXIST;
 446                 return NULL;
 447         }
 448         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 449         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 450                 struct mlx5_mp_id mp_id;
 451
 452                 eth_dev = rte_eth_dev_attach_secondary(name);
 453                 if (eth_dev == NULL) {
 454                         DRV_LOG(ERR, "can not attach rte ethdev");
 455                         rte_errno = ENOMEM;
 456                         return NULL;
 457                 }
 458                 eth_dev->device = dpdk_dev;
 459                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 460                 err = mlx5_proc_priv_init(eth_dev);
 461                 if (err)
 462                         return NULL;
 463                 mp_id.port_id = eth_dev->data->port_id;
 464                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 465                 /* Receive command fd from primary process */
 466                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 467                 if (err < 0)
 468                         goto err_secondary;
 469                 /* Remap UAR for Tx queues. */
 470                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 471                 if (err)
 472                         goto err_secondary;
 473                 /*
 474                  * Ethdev pointer is still required as input since
 475                  * the primary device is not accessible from the
 476                  * secondary process.
 477                  */
 478                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 479                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 480                 return eth_dev;
 481 err_secondary:
 482                 mlx5_dev_close(eth_dev);
 483                 return NULL;
 484         }
 485         /*
 486          * Some parameters ("tx_db_nc" in particularly) are needed in
 487          * advance to create dv/verbs device context. We proceed the
 488          * devargs here to get ones, and later proceed devargs again
 489          * to override some hardware settings.
 490          */
 491         err = mlx5_args(&config, dpdk_dev->devargs);
 492         if (err) {
 493                 err = rte_errno;
 494                 DRV_LOG(ERR, "failed to process device arguments: %s",
 495                         strerror(rte_errno));
 496                 goto error;
 497         }
 498         sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
 499         if (!sh)
 500                 return NULL;
 501         config.devx = sh->devx;
 502 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 503         config.dest_tir = 1;
 504 #endif
 505 #ifdef HAVE_IBV_MLX5_MOD_SWP
 506         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 507 #endif
 508         /*
 509          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 510          * as all ConnectX-5 devices.
 511          */
 512 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 513         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 514 #endif
 515 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 516         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 517 #endif
 518         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 519         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 520                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 521                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 522                         mps = MLX5_MPW_ENHANCED;
 523                 } else {
 524                         DRV_LOG(DEBUG, "MPW is supported");
 525                         mps = MLX5_MPW;
 526                 }
 527         } else {
 528                 DRV_LOG(DEBUG, "MPW isn't supported");
 529                 mps = MLX5_MPW_DISABLED;
 530         }
 531 #ifdef HAVE_IBV_MLX5_MOD_SWP
 532         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 533                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 534         DRV_LOG(DEBUG, "SWP support: %u", swp);
 535 #endif
 536         config.swp = !!swp;
 537 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 538         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 539                 struct mlx5dv_striding_rq_caps mprq_caps =
 540                         dv_attr.striding_rq_caps;
 541
 542                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 543                         mprq_caps.min_single_stride_log_num_of_bytes);
 544                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 545                         mprq_caps.max_single_stride_log_num_of_bytes);
 546                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 547                         mprq_caps.min_single_wqe_log_num_of_strides);
 548                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 549                         mprq_caps.max_single_wqe_log_num_of_strides);
 550                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 551                         mprq_caps.supported_qpts);
 552                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 553                 mprq = 1;
 554                 mprq_min_stride_size_n =
 555                         mprq_caps.min_single_stride_log_num_of_bytes;
 556                 mprq_max_stride_size_n =
 557                         mprq_caps.max_single_stride_log_num_of_bytes;
 558                 mprq_min_stride_num_n =
 559                         mprq_caps.min_single_wqe_log_num_of_strides;
 560                 mprq_max_stride_num_n =
 561                         mprq_caps.max_single_wqe_log_num_of_strides;
 562         }
 563 #endif
 564         if (RTE_CACHE_LINE_SIZE == 128 &&
 565             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 566                 cqe_comp = 0;
 567         else
 568                 cqe_comp = 1;
 569         config.cqe_comp = cqe_comp;
 570 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 571         /* Whether device supports 128B Rx CQE padding. */
 572         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 573                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 574 #endif
 575 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 576         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 577                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 578                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 579                              (dv_attr.tunnel_offloads_caps &
 580                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 581                              (dv_attr.tunnel_offloads_caps &
 582                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 583         }
 584         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 585                 tunnel_en ? "" : "not ");
 586 #else
 587         DRV_LOG(WARNING,
 588                 "tunnel offloading disabled due to old OFED/rdma-core version");
 589 #endif
 590         config.tunnel_en = tunnel_en;
 591 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 592         mpls_en = ((dv_attr.tunnel_offloads_caps &
 593                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 594                    (dv_attr.tunnel_offloads_caps &
 595                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 596         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 597                 mpls_en ? "" : "not ");
 598 #else
 599         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 600                 " old OFED/rdma-core version or firmware configuration");
 601 #endif
 602         config.mpls_en = mpls_en;
 603         /* Check port status. */
 604         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 605         if (err) {
 606                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 607                 goto error;
 608         }
 609         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 610                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 611                 err = EINVAL;
 612                 goto error;
 613         }
 614         if (port_attr.state != IBV_PORT_ACTIVE)
 615                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 616                         mlx5_glue->port_state_str(port_attr.state),
 617                         port_attr.state);
 618         /* Allocate private eth device data. */
 619         priv = rte_zmalloc("ethdev private structure",
 620                            sizeof(*priv),
 621                            RTE_CACHE_LINE_SIZE);
 622         if (priv == NULL) {
 623                 DRV_LOG(ERR, "priv allocation failure");
 624                 err = ENOMEM;
 625                 goto error;
 626         }
 627         priv->sh = sh;
 628         priv->dev_port = spawn->phys_port;
 629         priv->pci_dev = spawn->pci_dev;
 630         priv->mtu = RTE_ETHER_MTU;
 631         priv->mp_id.port_id = port_id;
 632         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 633         /* Some internal functions rely on Netlink sockets, open them now. */
 634         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 635         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 636         priv->representor = !!switch_info->representor;
 637         priv->master = !!switch_info->master;
 638         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 639         priv->vport_meta_tag = 0;
 640         priv->vport_meta_mask = 0;
 641         priv->pf_bond = spawn->pf_bond;
 642 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 643         /*
 644          * The DevX port query API is implemented. E-Switch may use
 645          * either vport or reg_c[0] metadata register to match on
 646          * vport index. The engaged part of metadata register is
 647          * defined by mask.
 648          */
 649         if (switch_info->representor || switch_info->master) {
 650                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 651                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 652                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 653                                                  &devx_port);
 654                 if (err) {
 655                         DRV_LOG(WARNING,
 656                                 "can't query devx port %d on device %s",
 657                                 spawn->phys_port,
 658                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 659                         devx_port.comp_mask = 0;
 660                 }
 661         }
 662         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 663                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 664                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 665                 if (!priv->vport_meta_mask) {
 666                         DRV_LOG(ERR, "vport zero mask for port %d"
 667                                      " on bonding device %s",
 668                                      spawn->phys_port,
 669                                      mlx5_os_get_dev_device_name
 670                                                         (spawn->phys_dev));
 671                         err = ENOTSUP;
 672                         goto error;
 673                 }
 674                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 675                         DRV_LOG(ERR, "invalid vport tag for port %d"
 676                                      " on bonding device %s",
 677                                      spawn->phys_port,
 678                                      mlx5_os_get_dev_device_name
 679                                                         (spawn->phys_dev));
 680                         err = ENOTSUP;
 681                         goto error;
 682                 }
 683         }
 684         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 685                 priv->vport_id = devx_port.vport_num;
 686         } else if (spawn->pf_bond >= 0) {
 687                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 688                              " on bonding device %s",
 689                              spawn->phys_port,
 690                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 691                 err = ENOTSUP;
 692                 goto error;
 693         } else {
 694                 /* Suppose vport index in compatible way. */
 695                 priv->vport_id = switch_info->representor ?
 696                                  switch_info->port_name + 1 : -1;
 697         }
 698 #else
 699         /*
 700          * Kernel/rdma_core support single E-Switch per PF configurations
 701          * only and vport_id field contains the vport index for
 702          * associated VF, which is deduced from representor port name.
 703          * For example, let's have the IB device port 10, it has
 704          * attached network device eth0, which has port name attribute
 705          * pf0vf2, we can deduce the VF number as 2, and set vport index
 706          * as 3 (2+1). This assigning schema should be changed if the
 707          * multiple E-Switch instances per PF configurations or/and PCI
 708          * subfunctions are added.
 709          */
 710         priv->vport_id = switch_info->representor ?
 711                          switch_info->port_name + 1 : -1;
 712 #endif
 713         /* representor_id field keeps the unmodified VF index. */
 714         priv->representor_id = switch_info->representor ?
 715                                switch_info->port_name : -1;
 716         /*
 717          * Look for sibling devices in order to reuse their switch domain
 718          * if any, otherwise allocate one.
 719          */
 720         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 721                 const struct mlx5_priv *opriv =
 722                         rte_eth_devices[port_id].data->dev_private;
 723
 724                 if (!opriv ||
 725                     opriv->sh != priv->sh ||
 726                         opriv->domain_id ==
 727                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 728                         continue;
 729                 priv->domain_id = opriv->domain_id;
 730                 break;
 731         }
 732         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 733                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 734                 if (err) {
 735                         err = rte_errno;
 736                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 737                                 strerror(rte_errno));
 738                         goto error;
 739                 }
 740                 own_domain_id = 1;
 741         }
 742         /* Override some values set by hardware configuration. */
 743         mlx5_args(&config, dpdk_dev->devargs);
 744         err = mlx5_dev_check_sibling_config(priv, &config);
 745         if (err)
 746                 goto error;
 747         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 748                             IBV_DEVICE_RAW_IP_CSUM);
 749         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 750                 (config.hw_csum ? "" : "not "));
 751 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 752         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 753         DRV_LOG(DEBUG, "counters are not supported");
 754 #endif
 755 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 756         if (config.dv_flow_en) {
 757                 DRV_LOG(WARNING, "DV flow is not supported");
 758                 config.dv_flow_en = 0;
 759         }
 760 #endif
 761         config.ind_table_max_size =
 762                 sh->device_attr.max_rwq_indirection_table_size;
 763         /*
 764          * Remove this check once DPDK supports larger/variable
 765          * indirection tables.
 766          */
 767         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 768                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 769         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 770                 config.ind_table_max_size);
 771         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 772                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 773         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 774                 (config.hw_vlan_strip ? "" : "not "));
 775         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 776                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 777         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 778                 (config.hw_fcs_strip ? "" : "not "));
 779 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 780         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 781 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 782         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 783                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 784 #endif
 785         if (config.hw_padding && !hw_padding) {
 786                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 787                 config.hw_padding = 0;
 788         } else if (config.hw_padding) {
 789                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 790         }
 791         config.tso = (sh->device_attr.max_tso > 0 &&
 792                       (sh->device_attr.tso_supported_qpts &
 793                        (1 << IBV_QPT_RAW_PACKET)));
 794         if (config.tso)
 795                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 796         /*
 797          * MPW is disabled by default, while the Enhanced MPW is enabled
 798          * by default.
 799          */
 800         if (config.mps == MLX5_ARG_UNSET)
 801                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 802                                                           MLX5_MPW_DISABLED;
 803         else
 804                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 805         DRV_LOG(INFO, "%sMPS is %s",
 806                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 807                 config.mps == MLX5_MPW ? "legacy " : "",
 808                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 809         if (config.cqe_comp && !cqe_comp) {
 810                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 811                 config.cqe_comp = 0;
 812         }
 813         if (config.cqe_pad && !cqe_pad) {
 814                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 815                 config.cqe_pad = 0;
 816         } else if (config.cqe_pad) {
 817                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 818         }
 819         if (config.devx) {
 820                 priv->counter_fallback = 0;
 821                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 822                 if (err) {
 823                         err = -err;
 824                         goto error;
 825                 }
 826                 if (!config.hca_attr.flow_counters_dump)
 827                         priv->counter_fallback = 1;
 828 #ifndef HAVE_IBV_DEVX_ASYNC
 829                 priv->counter_fallback = 1;
 830 #endif
 831                 if (priv->counter_fallback)
 832                         DRV_LOG(INFO, "Use fall-back DV counter management");
 833                 /* Check for LRO support. */
 834                 if (config.dest_tir && config.hca_attr.lro_cap &&
 835                     config.dv_flow_en) {
 836                         /* TBD check tunnel lro caps. */
 837                         config.lro.supported = config.hca_attr.lro_cap;
 838                         DRV_LOG(DEBUG, "Device supports LRO");
 839                         /*
 840                          * If LRO timeout is not configured by application,
 841                          * use the minimal supported value.
 842                          */
 843                         if (!config.lro.timeout)
 844                                 config.lro.timeout =
 845                                 config.hca_attr.lro_timer_supported_periods[0];
 846                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 847                                 config.lro.timeout);
 848                 }
 849 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 850                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 851                     config.dv_flow_en) {
 852                         uint8_t reg_c_mask =
 853                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 854                         /*
 855                          * Meter needs two REG_C's for color match and pre-sfx
 856                          * flow match. Here get the REG_C for color match.
 857                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 858                          */
 859                         reg_c_mask &= 0xfc;
 860                         if (__builtin_popcount(reg_c_mask) < 1) {
 861                                 priv->mtr_en = 0;
 862                                 DRV_LOG(WARNING, "No available register for"
 863                                         " meter.");
 864                         } else {
 865                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 866                                                       REG_C_0;
 867                                 priv->mtr_en = 1;
 868                                 priv->mtr_reg_share =
 869                                       config.hca_attr.qos.flow_meter_reg_share;
 870                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 871                                         priv->mtr_color_reg);
 872                         }
 873                 }
 874 #endif
 875         }
 876         if (config.tx_pp) {
 877                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
 878                         config.hca_attr.dev_freq_khz);
 879                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
 880                         config.hca_attr.qos.packet_pacing ? "" : "not ");
 881                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
 882                         config.hca_attr.cross_channel ? "" : "not ");
 883                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
 884                         config.hca_attr.wqe_index_ignore ? "" : "not ");
 885                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
 886                         config.hca_attr.non_wire_sq ? "" : "not ");
 887                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
 888                         config.hca_attr.log_max_static_sq_wq ? "" : "not ",
 889                         config.hca_attr.log_max_static_sq_wq);
 890                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
 891                         config.hca_attr.qos.wqe_rate_pp ? "" : "not ");
 892                 if (!config.devx) {
 893                         DRV_LOG(ERR, "DevX is required for packet pacing");
 894                         err = ENODEV;
 895                         goto error;
 896                 }
 897                 if (!config.hca_attr.qos.packet_pacing) {
 898                         DRV_LOG(ERR, "Packet pacing is not supported");
 899                         err = ENODEV;
 900                         goto error;
 901                 }
 902                 if (!config.hca_attr.cross_channel) {
 903                         DRV_LOG(ERR, "Cross channel operations are"
 904                                      " required for packet pacing");
 905                         err = ENODEV;
 906                         goto error;
 907                 }
 908                 if (!config.hca_attr.wqe_index_ignore) {
 909                         DRV_LOG(ERR, "WQE index ignore feature is"
 910                                      " required for packet pacing");
 911                         err = ENODEV;
 912                         goto error;
 913                 }
 914                 if (!config.hca_attr.non_wire_sq) {
 915                         DRV_LOG(ERR, "Non-wire SQ feature is"
 916                                      " required for packet pacing");
 917                         err = ENODEV;
 918                         goto error;
 919                 }
 920                 if (!config.hca_attr.log_max_static_sq_wq) {
 921                         DRV_LOG(ERR, "Static WQE SQ feature is"
 922                                      " required for packet pacing");
 923                         err = ENODEV;
 924                         goto error;
 925                 }
 926                 if (!config.hca_attr.qos.wqe_rate_pp) {
 927                         DRV_LOG(ERR, "WQE rate mode is required"
 928                                      " for packet pacing");
 929                         err = ENODEV;
 930                         goto error;
 931                 }
 932 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
 933                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
 934                              " can't create queues for packet pacing");
 935                 err = ENODEV;
 936                 goto error;
 937 #endif
 938         }
 939         if (config.mprq.enabled && mprq) {
 940                 if (config.mprq.stride_num_n &&
 941                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 942                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 943                         config.mprq.stride_num_n =
 944                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 945                                                 mprq_min_stride_num_n),
 946                                         mprq_max_stride_num_n);
 947                         DRV_LOG(WARNING,
 948                                 "the number of strides"
 949                                 " for Multi-Packet RQ is out of range,"
 950                                 " setting default value (%u)",
 951                                 1 << config.mprq.stride_num_n);
 952                 }
 953                 if (config.mprq.stride_size_n &&
 954                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 955                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 956                         config.mprq.stride_size_n =
 957                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 958                                                 mprq_min_stride_size_n),
 959                                         mprq_max_stride_size_n);
 960                         DRV_LOG(WARNING,
 961                                 "the size of a stride"
 962                                 " for Multi-Packet RQ is out of range,"
 963                                 " setting default value (%u)",
 964                                 1 << config.mprq.stride_size_n);
 965                 }
 966                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
 967                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
 968         } else if (config.mprq.enabled && !mprq) {
 969                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
 970                 config.mprq.enabled = 0;
 971         }
 972         if (config.max_dump_files_num == 0)
 973                 config.max_dump_files_num = 128;
 974         eth_dev = rte_eth_dev_allocate(name);
 975         if (eth_dev == NULL) {
 976                 DRV_LOG(ERR, "can not allocate rte ethdev");
 977                 err = ENOMEM;
 978                 goto error;
 979         }
 980         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
 981         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 982         if (priv->representor) {
 983                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
 984                 eth_dev->data->representor_id = priv->representor_id;
 985         }
 986         /*
 987          * Store associated network device interface index. This index
 988          * is permanent throughout the lifetime of device. So, we may store
 989          * the ifindex here and use the cached value further.
 990          */
 991         MLX5_ASSERT(spawn->ifindex);
 992         priv->if_index = spawn->ifindex;
 993         eth_dev->data->dev_private = priv;
 994         priv->dev_data = eth_dev->data;
 995         eth_dev->data->mac_addrs = priv->mac;
 996         eth_dev->device = dpdk_dev;
 997         /* Configure the first MAC address by default. */
 998         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
 999                 DRV_LOG(ERR,
1000                         "port %u cannot get MAC address, is mlx5_en"
1001                         " loaded? (errno: %s)",
1002                         eth_dev->data->port_id, strerror(rte_errno));
1003                 err = ENODEV;
1004                 goto error;
1005         }
1006         DRV_LOG(INFO,
1007                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1008                 eth_dev->data->port_id,
1009                 mac.addr_bytes[0], mac.addr_bytes[1],
1010                 mac.addr_bytes[2], mac.addr_bytes[3],
1011                 mac.addr_bytes[4], mac.addr_bytes[5]);
1012 #ifdef RTE_LIBRTE_MLX5_DEBUG
1013         {
1014                 char ifname[IF_NAMESIZE];
1015
1016                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1017                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1018                                 eth_dev->data->port_id, ifname);
1019                 else
1020                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1021                                 eth_dev->data->port_id);
1022         }
1023 #endif
1024         /* Get actual MTU if possible. */
1025         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1026         if (err) {
1027                 err = rte_errno;
1028                 goto error;
1029         }
1030         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1031                 priv->mtu);
1032         /* Initialize burst functions to prevent crashes before link-up. */
1033         eth_dev->rx_pkt_burst = removed_rx_burst;
1034         eth_dev->tx_pkt_burst = removed_tx_burst;
1035         eth_dev->dev_ops = &mlx5_os_dev_ops;
1036         /* Register MAC address. */
1037         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1038         if (config.vf && config.vf_nl_en)
1039                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1040                                       mlx5_ifindex(eth_dev),
1041                                       eth_dev->data->mac_addrs,
1042                                       MLX5_MAX_MAC_ADDRESSES);
1043         priv->flows = 0;
1044         priv->ctrl_flows = 0;
1045         TAILQ_INIT(&priv->flow_meters);
1046         TAILQ_INIT(&priv->flow_meter_profiles);
1047         /* Hint libmlx5 to use PMD allocator for data plane resources */
1048         mlx5_glue->dv_set_context_attr(sh->ctx,
1049                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1050                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1051                                 .alloc = &mlx5_alloc_verbs_buf,
1052                                 .free = &mlx5_free_verbs_buf,
1053                                 .data = priv,
1054                         }));
1055         /* Bring Ethernet device up. */
1056         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1057                 eth_dev->data->port_id);
1058         mlx5_set_link_up(eth_dev);
1059         /*
1060          * Even though the interrupt handler is not installed yet,
1061          * interrupts will still trigger on the async_fd from
1062          * Verbs context returned by ibv_open_device().
1063          */
1064         mlx5_link_update(eth_dev, 0);
1065 #ifdef HAVE_MLX5DV_DR_ESWITCH
1066         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1067               (switch_info->representor || switch_info->master)))
1068                 config.dv_esw_en = 0;
1069 #else
1070         config.dv_esw_en = 0;
1071 #endif
1072         /* Detect minimal data bytes to inline. */
1073         mlx5_set_min_inline(spawn, &config);
1074         /* Store device configuration on private structure. */
1075         priv->config = config;
1076         /* Create context for virtual machine VLAN workaround. */
1077         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1078         if (config.dv_flow_en) {
1079                 err = mlx5_alloc_shared_dr(priv);
1080                 if (err)
1081                         goto error;
1082                 /*
1083                  * RSS id is shared with meter flow id. Meter flow id can only
1084                  * use the 24 MSB of the register.
1085                  */
1086                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1087                                      MLX5_MTR_COLOR_BITS);
1088                 if (!priv->qrss_id_pool) {
1089                         DRV_LOG(ERR, "can't create flow id pool");
1090                         err = ENOMEM;
1091                         goto error;
1092                 }
1093         }
1094         /* Supported Verbs flow priority number detection. */
1095         err = mlx5_flow_discover_priorities(eth_dev);
1096         if (err < 0) {
1097                 err = -err;
1098                 goto error;
1099         }
1100         priv->config.flow_prio = err;
1101         if (!priv->config.dv_esw_en &&
1102             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1103                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1104                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1105                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1106         }
1107         mlx5_set_metadata_mask(eth_dev);
1108         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1109             !priv->sh->dv_regc0_mask) {
1110                 DRV_LOG(ERR, "metadata mode %u is not supported "
1111                              "(no metadata reg_c[0] is available)",
1112                              priv->config.dv_xmeta_en);
1113                         err = ENOTSUP;
1114                         goto error;
1115         }
1116         /*
1117          * Allocate the buffer for flow creating, just once.
1118          * The allocation must be done before any flow creating.
1119          */
1120         mlx5_flow_alloc_intermediate(eth_dev);
1121         /* Query availability of metadata reg_c's. */
1122         err = mlx5_flow_discover_mreg_c(eth_dev);
1123         if (err < 0) {
1124                 err = -err;
1125                 goto error;
1126         }
1127         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1128                 DRV_LOG(DEBUG,
1129                         "port %u extensive metadata register is not supported",
1130                         eth_dev->data->port_id);
1131                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1132                         DRV_LOG(ERR, "metadata mode %u is not supported "
1133                                      "(no metadata registers available)",
1134                                      priv->config.dv_xmeta_en);
1135                         err = ENOTSUP;
1136                         goto error;
1137                 }
1138         }
1139         if (priv->config.dv_flow_en &&
1140             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1141             mlx5_flow_ext_mreg_supported(eth_dev) &&
1142             priv->sh->dv_regc0_mask) {
1143                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1144                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1145                 if (!priv->mreg_cp_tbl) {
1146                         err = ENOMEM;
1147                         goto error;
1148                 }
1149         }
1150         return eth_dev;
1151 error:
1152         if (priv) {
1153                 if (priv->mreg_cp_tbl)
1154                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1155                 if (priv->sh)
1156                         mlx5_os_free_shared_dr(priv);
1157                 if (priv->nl_socket_route >= 0)
1158                         close(priv->nl_socket_route);
1159                 if (priv->nl_socket_rdma >= 0)
1160                         close(priv->nl_socket_rdma);
1161                 if (priv->vmwa_context)
1162                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1163                 if (priv->qrss_id_pool)
1164                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1165                 if (own_domain_id)
1166                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1167                 rte_free(priv);
1168                 if (eth_dev != NULL)
1169                         eth_dev->data->dev_private = NULL;
1170         }
1171         if (eth_dev != NULL) {
1172                 /* mac_addrs must not be freed alone because part of
1173                  * dev_private
1174                  **/
1175                 eth_dev->data->mac_addrs = NULL;
1176                 rte_eth_dev_release_port(eth_dev);
1177         }
1178         if (sh)
1179                 mlx5_free_shared_dev_ctx(sh);
1180         MLX5_ASSERT(err > 0);
1181         rte_errno = err;
1182         return NULL;
1183 }
1184
1185 /**
1186  * Comparison callback to sort device data.
1187  *
1188  * This is meant to be used with qsort().
1189  *
1190  * @param a[in]
1191  *   Pointer to pointer to first data object.
1192  * @param b[in]
1193  *   Pointer to pointer to second data object.
1194  *
1195  * @return
1196  *   0 if both objects are equal, less than 0 if the first argument is less
1197  *   than the second, greater than 0 otherwise.
1198  */
1199 static int
1200 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1201 {
1202         const struct mlx5_switch_info *si_a =
1203                 &((const struct mlx5_dev_spawn_data *)a)->info;
1204         const struct mlx5_switch_info *si_b =
1205                 &((const struct mlx5_dev_spawn_data *)b)->info;
1206         int ret;
1207
1208         /* Master device first. */
1209         ret = si_b->master - si_a->master;
1210         if (ret)
1211                 return ret;
1212         /* Then representor devices. */
1213         ret = si_b->representor - si_a->representor;
1214         if (ret)
1215                 return ret;
1216         /* Unidentified devices come last in no specific order. */
1217         if (!si_a->representor)
1218                 return 0;
1219         /* Order representors by name. */
1220         return si_a->port_name - si_b->port_name;
1221 }
1222
1223 /**
1224  * Match PCI information for possible slaves of bonding device.
1225  *
1226  * @param[in] ibv_dev
1227  *   Pointer to Infiniband device structure.
1228  * @param[in] pci_dev
1229  *   Pointer to PCI device structure to match PCI address.
1230  * @param[in] nl_rdma
1231  *   Netlink RDMA group socket handle.
1232  *
1233  * @return
1234  *   negative value if no bonding device found, otherwise
1235  *   positive index of slave PF in bonding.
1236  */
1237 static int
1238 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1239                            const struct rte_pci_device *pci_dev,
1240                            int nl_rdma)
1241 {
1242         char ifname[IF_NAMESIZE + 1];
1243         unsigned int ifindex;
1244         unsigned int np, i;
1245         FILE *file = NULL;
1246         int pf = -1;
1247
1248         /*
1249          * Try to get master device name. If something goes
1250          * wrong suppose the lack of kernel support and no
1251          * bonding devices.
1252          */
1253         if (nl_rdma < 0)
1254                 return -1;
1255         if (!strstr(ibv_dev->name, "bond"))
1256                 return -1;
1257         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1258         if (!np)
1259                 return -1;
1260         /*
1261          * The Master device might not be on the predefined
1262          * port (not on port index 1, it is not garanted),
1263          * we have to scan all Infiniband device port and
1264          * find master.
1265          */
1266         for (i = 1; i <= np; ++i) {
1267                 /* Check whether Infiniband port is populated. */
1268                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1269                 if (!ifindex)
1270                         continue;
1271                 if (!if_indextoname(ifindex, ifname))
1272                         continue;
1273                 /* Try to read bonding slave names from sysfs. */
1274                 MKSTR(slaves,
1275                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1276                 file = fopen(slaves, "r");
1277                 if (file)
1278                         break;
1279         }
1280         if (!file)
1281                 return -1;
1282         /* Use safe format to check maximal buffer length. */
1283         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1284         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1285                 char tmp_str[IF_NAMESIZE + 32];
1286                 struct rte_pci_addr pci_addr;
1287                 struct mlx5_switch_info info;
1288
1289                 /* Process slave interface names in the loop. */
1290                 snprintf(tmp_str, sizeof(tmp_str),
1291                          "/sys/class/net/%s", ifname);
1292                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1293                         DRV_LOG(WARNING, "can not get PCI address"
1294                                          " for netdev \"%s\"", ifname);
1295                         continue;
1296                 }
1297                 if (pci_dev->addr.domain != pci_addr.domain ||
1298                     pci_dev->addr.bus != pci_addr.bus ||
1299                     pci_dev->addr.devid != pci_addr.devid ||
1300                     pci_dev->addr.function != pci_addr.function)
1301                         continue;
1302                 /* Slave interface PCI address match found. */
1303                 fclose(file);
1304                 snprintf(tmp_str, sizeof(tmp_str),
1305                          "/sys/class/net/%s/phys_port_name", ifname);
1306                 file = fopen(tmp_str, "rb");
1307                 if (!file)
1308                         break;
1309                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1310                 if (fscanf(file, "%32s", tmp_str) == 1)
1311                         mlx5_translate_port_name(tmp_str, &info);
1312                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1313                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1314                         pf = info.port_name;
1315                 break;
1316         }
1317         if (file)
1318                 fclose(file);
1319         return pf;
1320 }
1321
1322 /**
1323  * DPDK callback to register a PCI device.
1324  *
1325  * This function spawns Ethernet devices out of a given PCI device.
1326  *
1327  * @param[in] pci_drv
1328  *   PCI driver structure (mlx5_driver).
1329  * @param[in] pci_dev
1330  *   PCI device information.
1331  *
1332  * @return
1333  *   0 on success, a negative errno value otherwise and rte_errno is set.
1334  */
1335 int
1336 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1337                   struct rte_pci_device *pci_dev)
1338 {
1339         struct ibv_device **ibv_list;
1340         /*
1341          * Number of found IB Devices matching with requested PCI BDF.
1342          * nd != 1 means there are multiple IB devices over the same
1343          * PCI device and we have representors and master.
1344          */
1345         unsigned int nd = 0;
1346         /*
1347          * Number of found IB device Ports. nd = 1 and np = 1..n means
1348          * we have the single multiport IB device, and there may be
1349          * representors attached to some of found ports.
1350          */
1351         unsigned int np = 0;
1352         /*
1353          * Number of DPDK ethernet devices to Spawn - either over
1354          * multiple IB devices or multiple ports of single IB device.
1355          * Actually this is the number of iterations to spawn.
1356          */
1357         unsigned int ns = 0;
1358         /*
1359          * Bonding device
1360          *   < 0 - no bonding device (single one)
1361          *  >= 0 - bonding device (value is slave PF index)
1362          */
1363         int bd = -1;
1364         struct mlx5_dev_spawn_data *list = NULL;
1365         struct mlx5_dev_config dev_config;
1366         int ret;
1367
1368         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1369                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1370                         " driver.");
1371                 return 1;
1372         }
1373         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1374                 mlx5_pmd_socket_init();
1375         ret = mlx5_init_once();
1376         if (ret) {
1377                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1378                         strerror(rte_errno));
1379                 return -rte_errno;
1380         }
1381         MLX5_ASSERT(pci_drv == &mlx5_driver);
1382         errno = 0;
1383         ibv_list = mlx5_glue->get_device_list(&ret);
1384         if (!ibv_list) {
1385                 rte_errno = errno ? errno : ENOSYS;
1386                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1387                 return -rte_errno;
1388         }
1389         /*
1390          * First scan the list of all Infiniband devices to find
1391          * matching ones, gathering into the list.
1392          */
1393         struct ibv_device *ibv_match[ret + 1];
1394         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1395         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1396         unsigned int i;
1397
1398         while (ret-- > 0) {
1399                 struct rte_pci_addr pci_addr;
1400
1401                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1402                 bd = mlx5_device_bond_pci_match
1403                                 (ibv_list[ret], pci_dev, nl_rdma);
1404                 if (bd >= 0) {
1405                         /*
1406                          * Bonding device detected. Only one match is allowed,
1407                          * the bonding is supported over multi-port IB device,
1408                          * there should be no matches on representor PCI
1409                          * functions or non VF LAG bonding devices with
1410                          * specified address.
1411                          */
1412                         if (nd) {
1413                                 DRV_LOG(ERR,
1414                                         "multiple PCI match on bonding device"
1415                                         "\"%s\" found", ibv_list[ret]->name);
1416                                 rte_errno = ENOENT;
1417                                 ret = -rte_errno;
1418                                 goto exit;
1419                         }
1420                         DRV_LOG(INFO, "PCI information matches for"
1421                                       " slave %d bonding device \"%s\"",
1422                                       bd, ibv_list[ret]->name);
1423                         ibv_match[nd++] = ibv_list[ret];
1424                         break;
1425                 }
1426                 if (mlx5_dev_to_pci_addr
1427                         (ibv_list[ret]->ibdev_path, &pci_addr))
1428                         continue;
1429                 if (pci_dev->addr.domain != pci_addr.domain ||
1430                     pci_dev->addr.bus != pci_addr.bus ||
1431                     pci_dev->addr.devid != pci_addr.devid ||
1432                     pci_dev->addr.function != pci_addr.function)
1433                         continue;
1434                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1435                         ibv_list[ret]->name);
1436                 ibv_match[nd++] = ibv_list[ret];
1437         }
1438         ibv_match[nd] = NULL;
1439         if (!nd) {
1440                 /* No device matches, just complain and bail out. */
1441                 DRV_LOG(WARNING,
1442                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1443                         " are kernel drivers loaded?",
1444                         pci_dev->addr.domain, pci_dev->addr.bus,
1445                         pci_dev->addr.devid, pci_dev->addr.function);
1446                 rte_errno = ENOENT;
1447                 ret = -rte_errno;
1448                 goto exit;
1449         }
1450         if (nd == 1) {
1451                 /*
1452                  * Found single matching device may have multiple ports.
1453                  * Each port may be representor, we have to check the port
1454                  * number and check the representors existence.
1455                  */
1456                 if (nl_rdma >= 0)
1457                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1458                 if (!np)
1459                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1460                                          " ports number", ibv_match[0]->name);
1461                 if (bd >= 0 && !np) {
1462                         DRV_LOG(ERR, "can not get ports"
1463                                      " for bonding device");
1464                         rte_errno = ENOENT;
1465                         ret = -rte_errno;
1466                         goto exit;
1467                 }
1468         }
1469 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1470         if (bd >= 0) {
1471                 /*
1472                  * This may happen if there is VF LAG kernel support and
1473                  * application is compiled with older rdma_core library.
1474                  */
1475                 DRV_LOG(ERR,
1476                         "No kernel/verbs support for VF LAG bonding found.");
1477                 rte_errno = ENOTSUP;
1478                 ret = -rte_errno;
1479                 goto exit;
1480         }
1481 #endif
1482         /*
1483          * Now we can determine the maximal
1484          * amount of devices to be spawned.
1485          */
1486         list = rte_zmalloc("device spawn data",
1487                          sizeof(struct mlx5_dev_spawn_data) *
1488                          (np ? np : nd),
1489                          RTE_CACHE_LINE_SIZE);
1490         if (!list) {
1491                 DRV_LOG(ERR, "spawn data array allocation failure");
1492                 rte_errno = ENOMEM;
1493                 ret = -rte_errno;
1494                 goto exit;
1495         }
1496         if (bd >= 0 || np > 1) {
1497                 /*
1498                  * Single IB device with multiple ports found,
1499                  * it may be E-Switch master device and representors.
1500                  * We have to perform identification through the ports.
1501                  */
1502                 MLX5_ASSERT(nl_rdma >= 0);
1503                 MLX5_ASSERT(ns == 0);
1504                 MLX5_ASSERT(nd == 1);
1505                 MLX5_ASSERT(np);
1506                 for (i = 1; i <= np; ++i) {
1507                         list[ns].max_port = np;
1508                         list[ns].phys_port = i;
1509                         list[ns].phys_dev = ibv_match[0];
1510                         list[ns].eth_dev = NULL;
1511                         list[ns].pci_dev = pci_dev;
1512                         list[ns].pf_bond = bd;
1513                         list[ns].ifindex = mlx5_nl_ifindex
1514                                 (nl_rdma,
1515                                 mlx5_os_get_dev_device_name
1516                                                 (list[ns].phys_dev), i);
1517                         if (!list[ns].ifindex) {
1518                                 /*
1519                                  * No network interface index found for the
1520                                  * specified port, it means there is no
1521                                  * representor on this port. It's OK,
1522                                  * there can be disabled ports, for example
1523                                  * if sriov_numvfs < sriov_totalvfs.
1524                                  */
1525                                 continue;
1526                         }
1527                         ret = -1;
1528                         if (nl_route >= 0)
1529                                 ret = mlx5_nl_switch_info
1530                                                (nl_route,
1531                                                 list[ns].ifindex,
1532                                                 &list[ns].info);
1533                         if (ret || (!list[ns].info.representor &&
1534                                     !list[ns].info.master)) {
1535                                 /*
1536                                  * We failed to recognize representors with
1537                                  * Netlink, let's try to perform the task
1538                                  * with sysfs.
1539                                  */
1540                                 ret =  mlx5_sysfs_switch_info
1541                                                 (list[ns].ifindex,
1542                                                  &list[ns].info);
1543                         }
1544                         if (!ret && bd >= 0) {
1545                                 switch (list[ns].info.name_type) {
1546                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1547                                         if (list[ns].info.port_name == bd)
1548                                                 ns++;
1549                                         break;
1550                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1551                                         /* Fallthrough */
1552                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1553                                         if (list[ns].info.pf_num == bd)
1554                                                 ns++;
1555                                         break;
1556                                 default:
1557                                         break;
1558                                 }
1559                                 continue;
1560                         }
1561                         if (!ret && (list[ns].info.representor ^
1562                                      list[ns].info.master))
1563                                 ns++;
1564                 }
1565                 if (!ns) {
1566                         DRV_LOG(ERR,
1567                                 "unable to recognize master/representors"
1568                                 " on the IB device with multiple ports");
1569                         rte_errno = ENOENT;
1570                         ret = -rte_errno;
1571                         goto exit;
1572                 }
1573         } else {
1574                 /*
1575                  * The existence of several matching entries (nd > 1) means
1576                  * port representors have been instantiated. No existing Verbs
1577                  * call nor sysfs entries can tell them apart, this can only
1578                  * be done through Netlink calls assuming kernel drivers are
1579                  * recent enough to support them.
1580                  *
1581                  * In the event of identification failure through Netlink,
1582                  * try again through sysfs, then:
1583                  *
1584                  * 1. A single IB device matches (nd == 1) with single
1585                  *    port (np=0/1) and is not a representor, assume
1586                  *    no switch support.
1587                  *
1588                  * 2. Otherwise no safe assumptions can be made;
1589                  *    complain louder and bail out.
1590                  */
1591                 for (i = 0; i != nd; ++i) {
1592                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1593                         list[ns].max_port = 1;
1594                         list[ns].phys_port = 1;
1595                         list[ns].phys_dev = ibv_match[i];
1596                         list[ns].eth_dev = NULL;
1597                         list[ns].pci_dev = pci_dev;
1598                         list[ns].pf_bond = -1;
1599                         list[ns].ifindex = 0;
1600                         if (nl_rdma >= 0)
1601                                 list[ns].ifindex = mlx5_nl_ifindex
1602                                 (nl_rdma,
1603                                 mlx5_os_get_dev_device_name
1604                                                 (list[ns].phys_dev), 1);
1605                         if (!list[ns].ifindex) {
1606                                 char ifname[IF_NAMESIZE];
1607
1608                                 /*
1609                                  * Netlink failed, it may happen with old
1610                                  * ib_core kernel driver (before 4.16).
1611                                  * We can assume there is old driver because
1612                                  * here we are processing single ports IB
1613                                  * devices. Let's try sysfs to retrieve
1614                                  * the ifindex. The method works for
1615                                  * master device only.
1616                                  */
1617                                 if (nd > 1) {
1618                                         /*
1619                                          * Multiple devices found, assume
1620                                          * representors, can not distinguish
1621                                          * master/representor and retrieve
1622                                          * ifindex via sysfs.
1623                                          */
1624                                         continue;
1625                                 }
1626                                 ret = mlx5_get_ifname_sysfs
1627                                         (ibv_match[i]->ibdev_path, ifname);
1628                                 if (!ret)
1629                                         list[ns].ifindex =
1630                                                 if_nametoindex(ifname);
1631                                 if (!list[ns].ifindex) {
1632                                         /*
1633                                          * No network interface index found
1634                                          * for the specified device, it means
1635                                          * there it is neither representor
1636                                          * nor master.
1637                                          */
1638                                         continue;
1639                                 }
1640                         }
1641                         ret = -1;
1642                         if (nl_route >= 0)
1643                                 ret = mlx5_nl_switch_info
1644                                                (nl_route,
1645                                                 list[ns].ifindex,
1646                                                 &list[ns].info);
1647                         if (ret || (!list[ns].info.representor &&
1648                                     !list[ns].info.master)) {
1649                                 /*
1650                                  * We failed to recognize representors with
1651                                  * Netlink, let's try to perform the task
1652                                  * with sysfs.
1653                                  */
1654                                 ret =  mlx5_sysfs_switch_info
1655                                                 (list[ns].ifindex,
1656                                                  &list[ns].info);
1657                         }
1658                         if (!ret && (list[ns].info.representor ^
1659                                      list[ns].info.master)) {
1660                                 ns++;
1661                         } else if ((nd == 1) &&
1662                                    !list[ns].info.representor &&
1663                                    !list[ns].info.master) {
1664                                 /*
1665                                  * Single IB device with
1666                                  * one physical port and
1667                                  * attached network device.
1668                                  * May be SRIOV is not enabled
1669                                  * or there is no representors.
1670                                  */
1671                                 DRV_LOG(INFO, "no E-Switch support detected");
1672                                 ns++;
1673                                 break;
1674                         }
1675                 }
1676                 if (!ns) {
1677                         DRV_LOG(ERR,
1678                                 "unable to recognize master/representors"
1679                                 " on the multiple IB devices");
1680                         rte_errno = ENOENT;
1681                         ret = -rte_errno;
1682                         goto exit;
1683                 }
1684         }
1685         MLX5_ASSERT(ns);
1686         /*
1687          * Sort list to probe devices in natural order for users convenience
1688          * (i.e. master first, then representors from lowest to highest ID).
1689          */
1690         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1691         /* Default configuration. */
1692         dev_config = (struct mlx5_dev_config){
1693                 .hw_padding = 0,
1694                 .mps = MLX5_ARG_UNSET,
1695                 .dbnc = MLX5_ARG_UNSET,
1696                 .rx_vec_en = 1,
1697                 .txq_inline_max = MLX5_ARG_UNSET,
1698                 .txq_inline_min = MLX5_ARG_UNSET,
1699                 .txq_inline_mpw = MLX5_ARG_UNSET,
1700                 .txqs_inline = MLX5_ARG_UNSET,
1701                 .vf_nl_en = 1,
1702                 .mr_ext_memseg_en = 1,
1703                 .mprq = {
1704                         .enabled = 0, /* Disabled by default. */
1705                         .stride_num_n = 0,
1706                         .stride_size_n = 0,
1707                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1708                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1709                 },
1710                 .dv_esw_en = 1,
1711                 .dv_flow_en = 1,
1712                 .log_hp_size = MLX5_ARG_UNSET,
1713         };
1714         /* Device specific configuration. */
1715         switch (pci_dev->id.device_id) {
1716         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1717         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1718         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1719         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1720         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1721         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1722         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1723                 dev_config.vf = 1;
1724                 break;
1725         default:
1726                 break;
1727         }
1728         for (i = 0; i != ns; ++i) {
1729                 uint32_t restore;
1730
1731                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1732                                                  &list[i],
1733                                                  dev_config);
1734                 if (!list[i].eth_dev) {
1735                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1736                                 break;
1737                         /* Device is disabled or already spawned. Ignore it. */
1738                         continue;
1739                 }
1740                 restore = list[i].eth_dev->data->dev_flags;
1741                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1742                 /* Restore non-PCI flags cleared by the above call. */
1743                 list[i].eth_dev->data->dev_flags |= restore;
1744                 rte_eth_dev_probing_finish(list[i].eth_dev);
1745         }
1746         if (i != ns) {
1747                 DRV_LOG(ERR,
1748                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1749                         " encountering an error: %s",
1750                         pci_dev->addr.domain, pci_dev->addr.bus,
1751                         pci_dev->addr.devid, pci_dev->addr.function,
1752                         strerror(rte_errno));
1753                 ret = -rte_errno;
1754                 /* Roll back. */
1755                 while (i--) {
1756                         if (!list[i].eth_dev)
1757                                 continue;
1758                         mlx5_dev_close(list[i].eth_dev);
1759                         /* mac_addrs must not be freed because in dev_private */
1760                         list[i].eth_dev->data->mac_addrs = NULL;
1761                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1762                 }
1763                 /* Restore original error. */
1764                 rte_errno = -ret;
1765         } else {
1766                 ret = 0;
1767         }
1768 exit:
1769         /*
1770          * Do the routine cleanup:
1771          * - close opened Netlink sockets
1772          * - free allocated spawn data array
1773          * - free the Infiniband device list
1774          */
1775         if (nl_rdma >= 0)
1776                 close(nl_rdma);
1777         if (nl_route >= 0)
1778                 close(nl_route);
1779         if (list)
1780                 rte_free(list);
1781         MLX5_ASSERT(ibv_list);
1782         mlx5_glue->free_device_list(ibv_list);
1783         return ret;
1784 }
1785
1786 static int
1787 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1788 {
1789         char *env;
1790         int value;
1791
1792         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1793         /* Get environment variable to store. */
1794         env = getenv(MLX5_SHUT_UP_BF);
1795         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1796         if (config->dbnc == MLX5_ARG_UNSET)
1797                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1798         else
1799                 setenv(MLX5_SHUT_UP_BF,
1800                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1801         return value;
1802 }
1803
1804 static void
1805 mlx5_restore_doorbell_mapping_env(int value)
1806 {
1807         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1808         /* Restore the original environment variable state. */
1809         if (value == MLX5_ARG_UNSET)
1810                 unsetenv(MLX5_SHUT_UP_BF);
1811         else
1812                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1813 }
1814
1815 /**
1816  * Extract pdn of PD object using DV API.
1817  *
1818  * @param[in] pd
1819  *   Pointer to the verbs PD object.
1820  * @param[out] pdn
1821  *   Pointer to the PD object number variable.
1822  *
1823  * @return
1824  *   0 on success, error value otherwise.
1825  */
1826 int
1827 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1828 {
1829 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1830         struct mlx5dv_obj obj;
1831         struct mlx5dv_pd pd_info;
1832         int ret = 0;
1833
1834         obj.pd.in = pd;
1835         obj.pd.out = &pd_info;
1836         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1837         if (ret) {
1838                 DRV_LOG(DEBUG, "Fail to get PD object info");
1839                 return ret;
1840         }
1841         *pdn = pd_info.pdn;
1842         return 0;
1843 #else
1844         (void)pd;
1845         (void)pdn;
1846         return -ENOTSUP;
1847 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1848 }
1849
1850 /**
1851  * Function API to open IB device.
1852  *
1853  * This function calls the Linux glue APIs to open a device.
1854  *
1855  * @param[in] spawn
1856  *   Pointer to the IB device attributes (name, port, etc).
1857  * @param[out] config
1858  *   Pointer to device configuration structure.
1859  * @param[out] sh
1860  *   Pointer to shared context structure.
1861  *
1862  * @return
1863  *   0 on success, a positive error value otherwise.
1864  */
1865 int
1866 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1867                      const struct mlx5_dev_config *config,
1868                      struct mlx5_dev_ctx_shared *sh)
1869 {
1870         int dbmap_env;
1871         int err = 0;
1872
1873         sh->numa_node = spawn->pci_dev->device.numa_node;
1874         pthread_mutex_init(&sh->txpp.mutex, NULL);
1875         /*
1876          * Configure environment variable "MLX5_BF_SHUT_UP"
1877          * before the device creation. The rdma_core library
1878          * checks the variable at device creation and
1879          * stores the result internally.
1880          */
1881         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1882         /* Try to open IB device with DV first, then usual Verbs. */
1883         errno = 0;
1884         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1885         if (sh->ctx) {
1886                 sh->devx = 1;
1887                 DRV_LOG(DEBUG, "DevX is supported");
1888                 /* The device is created, no need for environment. */
1889                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1890         } else {
1891                 /* The environment variable is still configured. */
1892                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1893                 err = errno ? errno : ENODEV;
1894                 /*
1895                  * The environment variable is not needed anymore,
1896                  * all device creation attempts are completed.
1897                  */
1898                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1899                 if (!sh->ctx)
1900                         return err;
1901                 DRV_LOG(DEBUG, "DevX is NOT supported");
1902                 err = 0;
1903         }
1904         return err;
1905 }
1906
1907 /**
1908  * Install shared asynchronous device events handler.
1909  * This function is implemented to support event sharing
1910  * between multiple ports of single IB device.
1911  *
1912  * @param sh
1913  *   Pointer to mlx5_dev_ctx_shared object.
1914  */
1915 void
1916 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1917 {
1918         int ret;
1919         int flags;
1920
1921         sh->intr_handle.fd = -1;
1922         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1923         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1924                     F_SETFL, flags | O_NONBLOCK);
1925         if (ret) {
1926                 DRV_LOG(INFO, "failed to change file descriptor async event"
1927                         " queue");
1928         } else {
1929                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1930                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1931                 if (rte_intr_callback_register(&sh->intr_handle,
1932                                         mlx5_dev_interrupt_handler, sh)) {
1933                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1934                         sh->intr_handle.fd = -1;
1935                 }
1936         }
1937         if (sh->devx) {
1938 #ifdef HAVE_IBV_DEVX_ASYNC
1939                 sh->intr_handle_devx.fd = -1;
1940                 sh->devx_comp =
1941                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1942                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1943                 if (!devx_comp) {
1944                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1945                         return;
1946                 }
1947                 flags = fcntl(devx_comp->fd, F_GETFL);
1948                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1949                 if (ret) {
1950                         DRV_LOG(INFO, "failed to change file descriptor"
1951                                 " devx comp");
1952                         return;
1953                 }
1954                 sh->intr_handle_devx.fd = devx_comp->fd;
1955                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1956                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1957                                         mlx5_dev_interrupt_handler_devx, sh)) {
1958                         DRV_LOG(INFO, "Fail to install the devx shared"
1959                                 " interrupt.");
1960                         sh->intr_handle_devx.fd = -1;
1961                 }
1962 #endif /* HAVE_IBV_DEVX_ASYNC */
1963         }
1964 }
1965
1966 /**
1967  * Uninstall shared asynchronous device events handler.
1968  * This function is implemented to support event sharing
1969  * between multiple ports of single IB device.
1970  *
1971  * @param dev
1972  *   Pointer to mlx5_dev_ctx_shared object.
1973  */
1974 void
1975 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
1976 {
1977         if (sh->intr_handle.fd >= 0)
1978                 mlx5_intr_callback_unregister(&sh->intr_handle,
1979                                               mlx5_dev_interrupt_handler, sh);
1980 #ifdef HAVE_IBV_DEVX_ASYNC
1981         if (sh->intr_handle_devx.fd >= 0)
1982                 rte_intr_callback_unregister(&sh->intr_handle_devx,
1983                                   mlx5_dev_interrupt_handler_devx, sh);
1984         if (sh->devx_comp)
1985                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1986 #endif
1987 }
1988
1989 /**
1990  * Read statistics by a named counter.
1991  *
1992  * @param[in] priv
1993  *   Pointer to the private device data structure.
1994  * @param[in] ctr_name
1995  *   Pointer to the name of the statistic counter to read
1996  * @param[out] stat
1997  *   Pointer to read statistic value.
1998  * @return
1999  *   0 on success and stat is valud, 1 if failed to read the value
2000  *   rte_errno is set.
2001  *
2002  */
2003 int
2004 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2005                       uint64_t *stat)
2006 {
2007         int fd;
2008
2009         if (priv->sh) {
2010                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2011                           priv->sh->ibdev_path,
2012                           priv->dev_port,
2013                           ctr_name);
2014                 fd = open(path, O_RDONLY);
2015                 if (fd != -1) {
2016                         char buf[21] = {'\0'};
2017                         ssize_t n = read(fd, buf, sizeof(buf));
2018
2019                         close(fd);
2020                         if (n != -1) {
2021                                 *stat = strtoull(buf, NULL, 10);
2022                                 return 0;
2023                         }
2024                 }
2025         }
2026         *stat = 0;
2027         return 1;
2028 }
2029
2030 /**
2031  * Read device counters table.
2032  *
2033  * @param dev
2034  *   Pointer to Ethernet device.
2035  * @param[out] stats
2036  *   Counters table output buffer.
2037  *
2038  * @return
2039  *   0 on success and stats is filled, negative errno value otherwise and
2040  *   rte_errno is set.
2041  */
2042 int
2043 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
2044 {
2045         struct mlx5_priv *priv = dev->data->dev_private;
2046         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2047         unsigned int i;
2048         struct ifreq ifr;
2049         unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
2050         unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
2051         struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
2052         int ret;
2053
2054         et_stats->cmd = ETHTOOL_GSTATS;
2055         et_stats->n_stats = xstats_ctrl->stats_n;
2056         ifr.ifr_data = (caddr_t)et_stats;
2057         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2058         if (ret) {
2059                 DRV_LOG(WARNING,
2060                         "port %u unable to read statistic values from device",
2061                         dev->data->port_id);
2062                 return ret;
2063         }
2064         for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2065                 if (xstats_ctrl->info[i].dev) {
2066                         ret = mlx5_os_read_dev_stat(priv,
2067                                             xstats_ctrl->info[i].ctr_name,
2068                                             &stats[i]);
2069                         /* return last xstats counter if fail to read. */
2070                         if (ret == 0)
2071                                 xstats_ctrl->xstats[i] = stats[i];
2072                         else
2073                                 stats[i] = xstats_ctrl->xstats[i];
2074                 } else {
2075                         stats[i] = (uint64_t)
2076                                 et_stats->data[xstats_ctrl->dev_table_idx[i]];
2077                 }
2078         }
2079         return 0;
2080 }
2081
2082 /**
2083  * Query the number of statistics provided by ETHTOOL.
2084  *
2085  * @param dev
2086  *   Pointer to Ethernet device.
2087  *
2088  * @return
2089  *   Number of statistics on success, negative errno value otherwise and
2090  *   rte_errno is set.
2091  */
2092 int
2093 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2094 {
2095         struct ethtool_drvinfo drvinfo;
2096         struct ifreq ifr;
2097         int ret;
2098
2099         drvinfo.cmd = ETHTOOL_GDRVINFO;
2100         ifr.ifr_data = (caddr_t)&drvinfo;
2101         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2102         if (ret) {
2103                 DRV_LOG(WARNING, "port %u unable to query number of statistics",
2104                         dev->data->port_id);
2105                 return ret;
2106         }
2107         return drvinfo.n_stats;
2108 }
2109
2110 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2111         {
2112                 .dpdk_name = "rx_port_unicast_bytes",
2113                 .ctr_name = "rx_vport_unicast_bytes",
2114         },
2115         {
2116                 .dpdk_name = "rx_port_multicast_bytes",
2117                 .ctr_name = "rx_vport_multicast_bytes",
2118         },
2119         {
2120                 .dpdk_name = "rx_port_broadcast_bytes",
2121                 .ctr_name = "rx_vport_broadcast_bytes",
2122         },
2123         {
2124                 .dpdk_name = "rx_port_unicast_packets",
2125                 .ctr_name = "rx_vport_unicast_packets",
2126         },
2127         {
2128                 .dpdk_name = "rx_port_multicast_packets",
2129                 .ctr_name = "rx_vport_multicast_packets",
2130         },
2131         {
2132                 .dpdk_name = "rx_port_broadcast_packets",
2133                 .ctr_name = "rx_vport_broadcast_packets",
2134         },
2135         {
2136                 .dpdk_name = "tx_port_unicast_bytes",
2137                 .ctr_name = "tx_vport_unicast_bytes",
2138         },
2139         {
2140                 .dpdk_name = "tx_port_multicast_bytes",
2141                 .ctr_name = "tx_vport_multicast_bytes",
2142         },
2143         {
2144                 .dpdk_name = "tx_port_broadcast_bytes",
2145                 .ctr_name = "tx_vport_broadcast_bytes",
2146         },
2147         {
2148                 .dpdk_name = "tx_port_unicast_packets",
2149                 .ctr_name = "tx_vport_unicast_packets",
2150         },
2151         {
2152                 .dpdk_name = "tx_port_multicast_packets",
2153                 .ctr_name = "tx_vport_multicast_packets",
2154         },
2155         {
2156                 .dpdk_name = "tx_port_broadcast_packets",
2157                 .ctr_name = "tx_vport_broadcast_packets",
2158         },
2159         {
2160                 .dpdk_name = "rx_wqe_err",
2161                 .ctr_name = "rx_wqe_err",
2162         },
2163         {
2164                 .dpdk_name = "rx_crc_errors_phy",
2165                 .ctr_name = "rx_crc_errors_phy",
2166         },
2167         {
2168                 .dpdk_name = "rx_in_range_len_errors_phy",
2169                 .ctr_name = "rx_in_range_len_errors_phy",
2170         },
2171         {
2172                 .dpdk_name = "rx_symbol_err_phy",
2173                 .ctr_name = "rx_symbol_err_phy",
2174         },
2175         {
2176                 .dpdk_name = "tx_errors_phy",
2177                 .ctr_name = "tx_errors_phy",
2178         },
2179         {
2180                 .dpdk_name = "rx_out_of_buffer",
2181                 .ctr_name = "out_of_buffer",
2182                 .dev = 1,
2183         },
2184         {
2185                 .dpdk_name = "tx_packets_phy",
2186                 .ctr_name = "tx_packets_phy",
2187         },
2188         {
2189                 .dpdk_name = "rx_packets_phy",
2190                 .ctr_name = "rx_packets_phy",
2191         },
2192         {
2193                 .dpdk_name = "tx_discards_phy",
2194                 .ctr_name = "tx_discards_phy",
2195         },
2196         {
2197                 .dpdk_name = "rx_discards_phy",
2198                 .ctr_name = "rx_discards_phy",
2199         },
2200         {
2201                 .dpdk_name = "tx_bytes_phy",
2202                 .ctr_name = "tx_bytes_phy",
2203         },
2204         {
2205                 .dpdk_name = "rx_bytes_phy",
2206                 .ctr_name = "rx_bytes_phy",
2207         },
2208         /* Representor only */
2209         {
2210                 .dpdk_name = "rx_packets",
2211                 .ctr_name = "vport_rx_packets",
2212         },
2213         {
2214                 .dpdk_name = "rx_bytes",
2215                 .ctr_name = "vport_rx_bytes",
2216         },
2217         {
2218                 .dpdk_name = "tx_packets",
2219                 .ctr_name = "vport_tx_packets",
2220         },
2221         {
2222                 .dpdk_name = "tx_bytes",
2223                 .ctr_name = "vport_tx_bytes",
2224         },
2225 };
2226
2227 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2228
2229 /**
2230  * Init the structures to read device counters.
2231  *
2232  * @param dev
2233  *   Pointer to Ethernet device.
2234  */
2235 void
2236 mlx5_os_stats_init(struct rte_eth_dev *dev)
2237 {
2238         struct mlx5_priv *priv = dev->data->dev_private;
2239         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2240         struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2241         unsigned int i;
2242         unsigned int j;
2243         struct ifreq ifr;
2244         struct ethtool_gstrings *strings = NULL;
2245         unsigned int dev_stats_n;
2246         unsigned int str_sz;
2247         int ret;
2248
2249         /* So that it won't aggregate for each init. */
2250         xstats_ctrl->mlx5_stats_n = 0;
2251         ret = mlx5_os_get_stats_n(dev);
2252         if (ret < 0) {
2253                 DRV_LOG(WARNING, "port %u no extended statistics available",
2254                         dev->data->port_id);
2255                 return;
2256         }
2257         dev_stats_n = ret;
2258         /* Allocate memory to grab stat names and values. */
2259         str_sz = dev_stats_n * ETH_GSTRING_LEN;
2260         strings = (struct ethtool_gstrings *)
2261                   rte_malloc("xstats_strings",
2262                              str_sz + sizeof(struct ethtool_gstrings), 0);
2263         if (!strings) {
2264                 DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2265                      dev->data->port_id);
2266                 return;
2267         }
2268         strings->cmd = ETHTOOL_GSTRINGS;
2269         strings->string_set = ETH_SS_STATS;
2270         strings->len = dev_stats_n;
2271         ifr.ifr_data = (caddr_t)strings;
2272         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2273         if (ret) {
2274                 DRV_LOG(WARNING, "port %u unable to get statistic names",
2275                         dev->data->port_id);
2276                 goto free;
2277         }
2278         for (i = 0; i != dev_stats_n; ++i) {
2279                 const char *curr_string = (const char *)
2280                         &strings->data[i * ETH_GSTRING_LEN];
2281
2282                 for (j = 0; j != xstats_n; ++j) {
2283                         if (!strcmp(mlx5_counters_init[j].ctr_name,
2284                                     curr_string)) {
2285                                 unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2286
2287                                 xstats_ctrl->dev_table_idx[idx] = i;
2288                                 xstats_ctrl->info[idx] = mlx5_counters_init[j];
2289                                 break;
2290                         }
2291                 }
2292         }
2293         /* Add dev counters. */
2294         for (i = 0; i != xstats_n; ++i) {
2295                 if (mlx5_counters_init[i].dev) {
2296                         unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2297
2298                         xstats_ctrl->info[idx] = mlx5_counters_init[i];
2299                         xstats_ctrl->hw_stats[idx] = 0;
2300                 }
2301         }
2302         MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2303         xstats_ctrl->stats_n = dev_stats_n;
2304         /* Copy to base at first time. */
2305         ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2306         if (ret)
2307                 DRV_LOG(ERR, "port %u cannot read device counters: %s",
2308                         dev->data->port_id, strerror(rte_errno));
2309         mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2310         stats_ctrl->imissed = 0;
2311 free:
2312         rte_free(strings);
2313 }
2314
2315 /**
2316  * Set the reg_mr and dereg_mr call backs
2317  *
2318  * @param reg_mr_cb[out]
2319  *   Pointer to reg_mr func
2320  * @param dereg_mr_cb[out]
2321  *   Pointer to dereg_mr func
2322  *
2323  */
2324 void
2325 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2326                       mlx5_dereg_mr_t *dereg_mr_cb)
2327 {
2328         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2329         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2330 }
2331
2332 const struct eth_dev_ops mlx5_os_dev_ops = {
2333         .dev_configure = mlx5_dev_configure,
2334         .dev_start = mlx5_dev_start,
2335         .dev_stop = mlx5_dev_stop,
2336         .dev_set_link_down = mlx5_set_link_down,
2337         .dev_set_link_up = mlx5_set_link_up,
2338         .dev_close = mlx5_dev_close,
2339         .promiscuous_enable = mlx5_promiscuous_enable,
2340         .promiscuous_disable = mlx5_promiscuous_disable,
2341         .allmulticast_enable = mlx5_allmulticast_enable,
2342         .allmulticast_disable = mlx5_allmulticast_disable,
2343         .link_update = mlx5_link_update,
2344         .stats_get = mlx5_stats_get,
2345         .stats_reset = mlx5_stats_reset,
2346         .xstats_get = mlx5_xstats_get,
2347         .xstats_reset = mlx5_xstats_reset,
2348         .xstats_get_names = mlx5_xstats_get_names,
2349         .fw_version_get = mlx5_fw_version_get,
2350         .dev_infos_get = mlx5_dev_infos_get,
2351         .read_clock = mlx5_txpp_read_clock,
2352         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2353         .vlan_filter_set = mlx5_vlan_filter_set,
2354         .rx_queue_setup = mlx5_rx_queue_setup,
2355         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2356         .tx_queue_setup = mlx5_tx_queue_setup,
2357         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2358         .rx_queue_release = mlx5_rx_queue_release,
2359         .tx_queue_release = mlx5_tx_queue_release,
2360         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2361         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2362         .mac_addr_remove = mlx5_mac_addr_remove,
2363         .mac_addr_add = mlx5_mac_addr_add,
2364         .mac_addr_set = mlx5_mac_addr_set,
2365         .set_mc_addr_list = mlx5_set_mc_addr_list,
2366         .mtu_set = mlx5_dev_set_mtu,
2367         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2368         .vlan_offload_set = mlx5_vlan_offload_set,
2369         .reta_update = mlx5_dev_rss_reta_update,
2370         .reta_query = mlx5_dev_rss_reta_query,
2371         .rss_hash_update = mlx5_rss_hash_update,
2372         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2373         .filter_ctrl = mlx5_dev_filter_ctrl,
2374         .rx_descriptor_status = mlx5_rx_descriptor_status,
2375         .tx_descriptor_status = mlx5_tx_descriptor_status,
2376         .rxq_info_get = mlx5_rxq_info_get,
2377         .txq_info_get = mlx5_txq_info_get,
2378         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2379         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2380         .rx_queue_count = mlx5_rx_queue_count,
2381         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2382         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2383         .is_removed = mlx5_is_removed,
2384         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2385         .get_module_info = mlx5_get_module_info,
2386         .get_module_eeprom = mlx5_get_module_eeprom,
2387         .hairpin_cap_get = mlx5_hairpin_cap_get,
2388         .mtr_ops_get = mlx5_flow_meter_ops_get,
2389 };
2390
2391 /* Available operations from secondary process. */
2392 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2393         .stats_get = mlx5_stats_get,
2394         .stats_reset = mlx5_stats_reset,
2395         .xstats_get = mlx5_xstats_get,
2396         .xstats_reset = mlx5_xstats_reset,
2397         .xstats_get_names = mlx5_xstats_get_names,
2398         .fw_version_get = mlx5_fw_version_get,
2399         .dev_infos_get = mlx5_dev_infos_get,
2400         .read_clock = mlx5_txpp_read_clock,
2401         .rx_descriptor_status = mlx5_rx_descriptor_status,
2402         .tx_descriptor_status = mlx5_tx_descriptor_status,
2403         .rxq_info_get = mlx5_rxq_info_get,
2404         .txq_info_get = mlx5_txq_info_get,
2405         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2406         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2407         .get_module_info = mlx5_get_module_info,
2408         .get_module_eeprom = mlx5_get_module_eeprom,
2409 };
2410
2411 /* Available operations in flow isolated mode. */
2412 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2413         .dev_configure = mlx5_dev_configure,
2414         .dev_start = mlx5_dev_start,
2415         .dev_stop = mlx5_dev_stop,
2416         .dev_set_link_down = mlx5_set_link_down,
2417         .dev_set_link_up = mlx5_set_link_up,
2418         .dev_close = mlx5_dev_close,
2419         .promiscuous_enable = mlx5_promiscuous_enable,
2420         .promiscuous_disable = mlx5_promiscuous_disable,
2421         .allmulticast_enable = mlx5_allmulticast_enable,
2422         .allmulticast_disable = mlx5_allmulticast_disable,
2423         .link_update = mlx5_link_update,
2424         .stats_get = mlx5_stats_get,
2425         .stats_reset = mlx5_stats_reset,
2426         .xstats_get = mlx5_xstats_get,
2427         .xstats_reset = mlx5_xstats_reset,
2428         .xstats_get_names = mlx5_xstats_get_names,
2429         .fw_version_get = mlx5_fw_version_get,
2430         .dev_infos_get = mlx5_dev_infos_get,
2431         .read_clock = mlx5_txpp_read_clock,
2432         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2433         .vlan_filter_set = mlx5_vlan_filter_set,
2434         .rx_queue_setup = mlx5_rx_queue_setup,
2435         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2436         .tx_queue_setup = mlx5_tx_queue_setup,
2437         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2438         .rx_queue_release = mlx5_rx_queue_release,
2439         .tx_queue_release = mlx5_tx_queue_release,
2440         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2441         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2442         .mac_addr_remove = mlx5_mac_addr_remove,
2443         .mac_addr_add = mlx5_mac_addr_add,
2444         .mac_addr_set = mlx5_mac_addr_set,
2445         .set_mc_addr_list = mlx5_set_mc_addr_list,
2446         .mtu_set = mlx5_dev_set_mtu,
2447         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2448         .vlan_offload_set = mlx5_vlan_offload_set,
2449         .filter_ctrl = mlx5_dev_filter_ctrl,
2450         .rx_descriptor_status = mlx5_rx_descriptor_status,
2451         .tx_descriptor_status = mlx5_tx_descriptor_status,
2452         .rxq_info_get = mlx5_rxq_info_get,
2453         .txq_info_get = mlx5_txq_info_get,
2454         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2455         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2456         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2457         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2458         .is_removed = mlx5_is_removed,
2459         .get_module_info = mlx5_get_module_info,
2460         .get_module_eeprom = mlx5_get_module_eeprom,
2461         .hairpin_cap_get = mlx5_hairpin_cap_get,
2462         .mtr_ops_get = mlx5_flow_meter_ops_get,
2463 };