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_utils.h"
  50 #include "mlx5_rxtx.h"
  51 #include "mlx5_autoconf.h"
  52 #include "mlx5_mr.h"
  53 #include "mlx5_flow.h"
  54 #include "rte_pmd_mlx5.h"
  55 #include "mlx5_verbs.h"
  56
  57 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  58
  59 #ifndef HAVE_IBV_MLX5_MOD_MPW
  60 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  61 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  62 #endif
  63
  64 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  65 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  66 #endif
  67
  68 /**
  69  * Get device name. Given an ibv_device pointer - return a
  70  * pointer to the corresponding device name.
  71  *
  72  * @param[in] dev
  73  *   Pointer to ibv device.
  74  *
  75  * @return
  76  *   Pointer to device name if dev is valid, NULL otherwise.
  77  */
  78 const char *
  79 mlx5_os_get_dev_device_name(void *dev)
  80 {
  81         if (!dev)
  82                 return NULL;
  83         return ((struct ibv_device *)dev)->name;
  84 }
  85
  86 /**
  87  * Get ibv device name. Given an ibv_context pointer - return a
  88  * pointer to the corresponding device name.
  89  *
  90  * @param[in] ctx
  91  *   Pointer to ibv context.
  92  *
  93  * @return
  94  *   Pointer to device name if ctx is valid, NULL otherwise.
  95  */
  96 const char *
  97 mlx5_os_get_ctx_device_name(void *ctx)
  98 {
  99         if (!ctx)
 100                 return NULL;
 101         return ((struct ibv_context *)ctx)->device->name;
 102 }
 103
 104 /**
 105  * Get ibv device path name. Given an ibv_context pointer - return a
 106  * pointer to the corresponding device path name.
 107  *
 108  * @param[in] ctx
 109  *   Pointer to ibv context.
 110  *
 111  * @return
 112  *   Pointer to device path name if ctx is valid, NULL otherwise.
 113  */
 114 const char *
 115 mlx5_os_get_ctx_device_path(void *ctx)
 116 {
 117         if (!ctx)
 118                 return NULL;
 119
 120         return ((struct ibv_context *)ctx)->device->ibdev_path;
 121 }
 122
 123 /**
 124  * Get umem id. Given a pointer to umem object of type
 125  * 'struct mlx5dv_devx_umem *' - return its id.
 126  *
 127  * @param[in] umem
 128  *   Pointer to umem object.
 129  *
 130  * @return
 131  *   The umem id if umem is valid, 0 otherwise.
 132  */
 133 uint32_t
 134 mlx5_os_get_umem_id(void *umem)
 135 {
 136         if (!umem)
 137                 return 0;
 138         return ((struct mlx5dv_devx_umem *)umem)->umem_id;
 139 }
 140
 141 /**
 142  * Get mlx5 device attributes. The glue function query_device_ex() is called
 143  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
 144  * device attributes from the glue out parameter.
 145  *
 146  * @param dev
 147  *   Pointer to ibv context.
 148  *
 149  * @param device_attr
 150  *   Pointer to mlx5 device attributes.
 151  *
 152  * @return
 153  *   0 on success, non zero error number otherwise
 154  */
 155 int
 156 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
 157 {
 158         int err;
 159         struct ibv_device_attr_ex attr_ex;
 160         memset(device_attr, 0, sizeof(*device_attr));
 161         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
 162         if (err)
 163                 return err;
 164
 165         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
 166         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
 167         device_attr->max_sge = attr_ex.orig_attr.max_sge;
 168         device_attr->max_cq = attr_ex.orig_attr.max_cq;
 169         device_attr->max_qp = attr_ex.orig_attr.max_qp;
 170         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 171         device_attr->max_rwq_indirection_table_size =
 172                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 173         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 174         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 175
 176         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 177         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 178         if (err)
 179                 return err;
 180
 181         device_attr->flags = dv_attr.flags;
 182         device_attr->comp_mask = dv_attr.comp_mask;
 183 #ifdef HAVE_IBV_MLX5_MOD_SWP
 184         device_attr->sw_parsing_offloads =
 185                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 186 #endif
 187         device_attr->min_single_stride_log_num_of_bytes =
 188                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 189         device_attr->max_single_stride_log_num_of_bytes =
 190                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 191         device_attr->min_single_wqe_log_num_of_strides =
 192                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 193         device_attr->max_single_wqe_log_num_of_strides =
 194                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 195         device_attr->stride_supported_qpts =
 196                 dv_attr.striding_rq_caps.supported_qpts;
 197 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 198         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 199 #endif
 200
 201         return err;
 202 }
 203
 204 /**
 205  * Verbs callback to allocate a memory. This function should allocate the space
 206  * according to the size provided residing inside a huge page.
 207  * Please note that all allocation must respect the alignment from libmlx5
 208  * (i.e. currently sysconf(_SC_PAGESIZE)).
 209  *
 210  * @param[in] size
 211  *   The size in bytes of the memory to allocate.
 212  * @param[in] data
 213  *   A pointer to the callback data.
 214  *
 215  * @return
 216  *   Allocated buffer, NULL otherwise and rte_errno is set.
 217  */
 218 static void *
 219 mlx5_alloc_verbs_buf(size_t size, void *data)
 220 {
 221         struct mlx5_priv *priv = data;
 222         void *ret;
 223         size_t alignment = sysconf(_SC_PAGESIZE);
 224         unsigned int socket = SOCKET_ID_ANY;
 225
 226         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 227                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 228
 229                 socket = ctrl->socket;
 230         } else if (priv->verbs_alloc_ctx.type ==
 231                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 232                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 233
 234                 socket = ctrl->socket;
 235         }
 236         MLX5_ASSERT(data != NULL);
 237         ret = rte_malloc_socket(__func__, size, alignment, socket);
 238         if (!ret && size)
 239                 rte_errno = ENOMEM;
 240         return ret;
 241 }
 242
 243 /**
 244  * Verbs callback to free a memory.
 245  *
 246  * @param[in] ptr
 247  *   A pointer to the memory to free.
 248  * @param[in] data
 249  *   A pointer to the callback data.
 250  */
 251 static void
 252 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 253 {
 254         MLX5_ASSERT(data != NULL);
 255         rte_free(ptr);
 256 }
 257
 258 /**
 259  * Initialize DR related data within private structure.
 260  * Routine checks the reference counter and does actual
 261  * resources creation/initialization only if counter is zero.
 262  *
 263  * @param[in] priv
 264  *   Pointer to the private device data structure.
 265  *
 266  * @return
 267  *   Zero on success, positive error code otherwise.
 268  */
 269 static int
 270 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 271 {
 272         struct mlx5_dev_ctx_shared *sh = priv->sh;
 273         char s[MLX5_HLIST_NAMESIZE];
 274         int err = 0;
 275
 276         if (!sh->flow_tbls)
 277                 err = mlx5_alloc_table_hash_list(priv);
 278         else
 279                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 280                         (void *)sh->flow_tbls);
 281         if (err)
 282                 return err;
 283         /* Create tags hash list table. */
 284         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 285         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 286         if (!sh->tag_table) {
 287                 DRV_LOG(ERR, "tags with hash creation failed.\n");
 288                 err = ENOMEM;
 289                 goto error;
 290         }
 291 #ifdef HAVE_MLX5DV_DR
 292         void *domain;
 293
 294         if (sh->dv_refcnt) {
 295                 /* Shared DV/DR structures is already initialized. */
 296                 sh->dv_refcnt++;
 297                 priv->dr_shared = 1;
 298                 return 0;
 299         }
 300         /* Reference counter is zero, we should initialize structures. */
 301         domain = mlx5_glue->dr_create_domain(sh->ctx,
 302                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 303         if (!domain) {
 304                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 305                 err = errno;
 306                 goto error;
 307         }
 308         sh->rx_domain = domain;
 309         domain = mlx5_glue->dr_create_domain(sh->ctx,
 310                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 311         if (!domain) {
 312                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 313                 err = errno;
 314                 goto error;
 315         }
 316         pthread_mutex_init(&sh->dv_mutex, NULL);
 317         sh->tx_domain = domain;
 318 #ifdef HAVE_MLX5DV_DR_ESWITCH
 319         if (priv->config.dv_esw_en) {
 320                 domain  = mlx5_glue->dr_create_domain
 321                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 322                 if (!domain) {
 323                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 324                         err = errno;
 325                         goto error;
 326                 }
 327                 sh->fdb_domain = domain;
 328                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 329         }
 330 #endif
 331         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 332                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 333                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 334                 if (sh->fdb_domain)
 335                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 336         }
 337         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 338 #endif /* HAVE_MLX5DV_DR */
 339         sh->dv_refcnt++;
 340         priv->dr_shared = 1;
 341         return 0;
 342 error:
 343         /* Rollback the created objects. */
 344         if (sh->rx_domain) {
 345                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 346                 sh->rx_domain = NULL;
 347         }
 348         if (sh->tx_domain) {
 349                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 350                 sh->tx_domain = NULL;
 351         }
 352         if (sh->fdb_domain) {
 353                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 354                 sh->fdb_domain = NULL;
 355         }
 356         if (sh->esw_drop_action) {
 357                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 358                 sh->esw_drop_action = NULL;
 359         }
 360         if (sh->pop_vlan_action) {
 361                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 362                 sh->pop_vlan_action = NULL;
 363         }
 364         if (sh->tag_table) {
 365                 /* tags should be destroyed with flow before. */
 366                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 367                 sh->tag_table = NULL;
 368         }
 369         mlx5_free_table_hash_list(priv);
 370         return err;
 371 }
 372
 373 /**
 374  * Destroy DR related data within private structure.
 375  *
 376  * @param[in] priv
 377  *   Pointer to the private device data structure.
 378  */
 379 void
 380 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 381 {
 382         struct mlx5_dev_ctx_shared *sh;
 383
 384         if (!priv->dr_shared)
 385                 return;
 386         priv->dr_shared = 0;
 387         sh = priv->sh;
 388         MLX5_ASSERT(sh);
 389 #ifdef HAVE_MLX5DV_DR
 390         MLX5_ASSERT(sh->dv_refcnt);
 391         if (sh->dv_refcnt && --sh->dv_refcnt)
 392                 return;
 393         if (sh->rx_domain) {
 394                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 395                 sh->rx_domain = NULL;
 396         }
 397         if (sh->tx_domain) {
 398                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 399                 sh->tx_domain = NULL;
 400         }
 401 #ifdef HAVE_MLX5DV_DR_ESWITCH
 402         if (sh->fdb_domain) {
 403                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 404                 sh->fdb_domain = NULL;
 405         }
 406         if (sh->esw_drop_action) {
 407                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 408                 sh->esw_drop_action = NULL;
 409         }
 410 #endif
 411         if (sh->pop_vlan_action) {
 412                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 413                 sh->pop_vlan_action = NULL;
 414         }
 415         pthread_mutex_destroy(&sh->dv_mutex);
 416 #endif /* HAVE_MLX5DV_DR */
 417         if (sh->tag_table) {
 418                 /* tags should be destroyed with flow before. */
 419                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 420                 sh->tag_table = NULL;
 421         }
 422         mlx5_free_table_hash_list(priv);
 423 }
 424
 425 /**
 426  * Spawn an Ethernet device from Verbs information.
 427  *
 428  * @param dpdk_dev
 429  *   Backing DPDK device.
 430  * @param spawn
 431  *   Verbs device parameters (name, port, switch_info) to spawn.
 432  * @param config
 433  *   Device configuration parameters.
 434  *
 435  * @return
 436  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 437  *   is set. The following errors are defined:
 438  *
 439  *   EBUSY: device is not supposed to be spawned.
 440  *   EEXIST: device is already spawned
 441  */
 442 static struct rte_eth_dev *
 443 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 444                struct mlx5_dev_spawn_data *spawn,
 445                struct mlx5_dev_config config)
 446 {
 447         const struct mlx5_switch_info *switch_info = &spawn->info;
 448         struct mlx5_dev_ctx_shared *sh = NULL;
 449         struct ibv_port_attr port_attr;
 450         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 451         struct rte_eth_dev *eth_dev = NULL;
 452         struct mlx5_priv *priv = NULL;
 453         int err = 0;
 454         unsigned int hw_padding = 0;
 455         unsigned int mps;
 456         unsigned int cqe_comp;
 457         unsigned int cqe_pad = 0;
 458         unsigned int tunnel_en = 0;
 459         unsigned int mpls_en = 0;
 460         unsigned int swp = 0;
 461         unsigned int mprq = 0;
 462         unsigned int mprq_min_stride_size_n = 0;
 463         unsigned int mprq_max_stride_size_n = 0;
 464         unsigned int mprq_min_stride_num_n = 0;
 465         unsigned int mprq_max_stride_num_n = 0;
 466         struct rte_ether_addr mac;
 467         char name[RTE_ETH_NAME_MAX_LEN];
 468         int own_domain_id = 0;
 469         uint16_t port_id;
 470         unsigned int i;
 471 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 472         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 473 #endif
 474
 475         /* Determine if this port representor is supposed to be spawned. */
 476         if (switch_info->representor && dpdk_dev->devargs) {
 477                 struct rte_eth_devargs eth_da;
 478
 479                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 480                 if (err) {
 481                         rte_errno = -err;
 482                         DRV_LOG(ERR, "failed to process device arguments: %s",
 483                                 strerror(rte_errno));
 484                         return NULL;
 485                 }
 486                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 487                         if (eth_da.representor_ports[i] ==
 488                             (uint16_t)switch_info->port_name)
 489                                 break;
 490                 if (i == eth_da.nb_representor_ports) {
 491                         rte_errno = EBUSY;
 492                         return NULL;
 493                 }
 494         }
 495         /* Build device name. */
 496         if (spawn->pf_bond <  0) {
 497                 /* Single device. */
 498                 if (!switch_info->representor)
 499                         strlcpy(name, dpdk_dev->name, sizeof(name));
 500                 else
 501                         snprintf(name, sizeof(name), "%s_representor_%u",
 502                                  dpdk_dev->name, switch_info->port_name);
 503         } else {
 504                 /* Bonding device. */
 505                 if (!switch_info->representor)
 506                         snprintf(name, sizeof(name), "%s_%s",
 507                                  dpdk_dev->name,
 508                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 509                 else
 510                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 511                                  dpdk_dev->name,
 512                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 513                                  switch_info->port_name);
 514         }
 515         /* check if the device is already spawned */
 516         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 517                 rte_errno = EEXIST;
 518                 return NULL;
 519         }
 520         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 521         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 522                 struct mlx5_mp_id mp_id;
 523
 524                 eth_dev = rte_eth_dev_attach_secondary(name);
 525                 if (eth_dev == NULL) {
 526                         DRV_LOG(ERR, "can not attach rte ethdev");
 527                         rte_errno = ENOMEM;
 528                         return NULL;
 529                 }
 530                 eth_dev->device = dpdk_dev;
 531                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 532                 err = mlx5_proc_priv_init(eth_dev);
 533                 if (err)
 534                         return NULL;
 535                 mp_id.port_id = eth_dev->data->port_id;
 536                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 537                 /* Receive command fd from primary process */
 538                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 539                 if (err < 0)
 540                         goto err_secondary;
 541                 /* Remap UAR for Tx queues. */
 542                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 543                 if (err)
 544                         goto err_secondary;
 545                 /*
 546                  * Ethdev pointer is still required as input since
 547                  * the primary device is not accessible from the
 548                  * secondary process.
 549                  */
 550                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 551                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 552                 return eth_dev;
 553 err_secondary:
 554                 mlx5_dev_close(eth_dev);
 555                 return NULL;
 556         }
 557         /*
 558          * Some parameters ("tx_db_nc" in particularly) are needed in
 559          * advance to create dv/verbs device context. We proceed the
 560          * devargs here to get ones, and later proceed devargs again
 561          * to override some hardware settings.
 562          */
 563         err = mlx5_args(&config, dpdk_dev->devargs);
 564         if (err) {
 565                 err = rte_errno;
 566                 DRV_LOG(ERR, "failed to process device arguments: %s",
 567                         strerror(rte_errno));
 568                 goto error;
 569         }
 570         sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
 571         if (!sh)
 572                 return NULL;
 573         config.devx = sh->devx;
 574 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 575         config.dest_tir = 1;
 576 #endif
 577 #ifdef HAVE_IBV_MLX5_MOD_SWP
 578         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 579 #endif
 580         /*
 581          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 582          * as all ConnectX-5 devices.
 583          */
 584 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 585         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 586 #endif
 587 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 588         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 589 #endif
 590         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 591         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 592                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 593                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 594                         mps = MLX5_MPW_ENHANCED;
 595                 } else {
 596                         DRV_LOG(DEBUG, "MPW is supported");
 597                         mps = MLX5_MPW;
 598                 }
 599         } else {
 600                 DRV_LOG(DEBUG, "MPW isn't supported");
 601                 mps = MLX5_MPW_DISABLED;
 602         }
 603 #ifdef HAVE_IBV_MLX5_MOD_SWP
 604         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 605                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 606         DRV_LOG(DEBUG, "SWP support: %u", swp);
 607 #endif
 608         config.swp = !!swp;
 609 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 610         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 611                 struct mlx5dv_striding_rq_caps mprq_caps =
 612                         dv_attr.striding_rq_caps;
 613
 614                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 615                         mprq_caps.min_single_stride_log_num_of_bytes);
 616                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 617                         mprq_caps.max_single_stride_log_num_of_bytes);
 618                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 619                         mprq_caps.min_single_wqe_log_num_of_strides);
 620                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 621                         mprq_caps.max_single_wqe_log_num_of_strides);
 622                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 623                         mprq_caps.supported_qpts);
 624                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 625                 mprq = 1;
 626                 mprq_min_stride_size_n =
 627                         mprq_caps.min_single_stride_log_num_of_bytes;
 628                 mprq_max_stride_size_n =
 629                         mprq_caps.max_single_stride_log_num_of_bytes;
 630                 mprq_min_stride_num_n =
 631                         mprq_caps.min_single_wqe_log_num_of_strides;
 632                 mprq_max_stride_num_n =
 633                         mprq_caps.max_single_wqe_log_num_of_strides;
 634         }
 635 #endif
 636         if (RTE_CACHE_LINE_SIZE == 128 &&
 637             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 638                 cqe_comp = 0;
 639         else
 640                 cqe_comp = 1;
 641         config.cqe_comp = cqe_comp;
 642 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 643         /* Whether device supports 128B Rx CQE padding. */
 644         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 645                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 646 #endif
 647 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 648         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 649                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 650                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 651                              (dv_attr.tunnel_offloads_caps &
 652                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 653                              (dv_attr.tunnel_offloads_caps &
 654                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 655         }
 656         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 657                 tunnel_en ? "" : "not ");
 658 #else
 659         DRV_LOG(WARNING,
 660                 "tunnel offloading disabled due to old OFED/rdma-core version");
 661 #endif
 662         config.tunnel_en = tunnel_en;
 663 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 664         mpls_en = ((dv_attr.tunnel_offloads_caps &
 665                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 666                    (dv_attr.tunnel_offloads_caps &
 667                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 668         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 669                 mpls_en ? "" : "not ");
 670 #else
 671         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 672                 " old OFED/rdma-core version or firmware configuration");
 673 #endif
 674         config.mpls_en = mpls_en;
 675         /* Check port status. */
 676         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 677         if (err) {
 678                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 679                 goto error;
 680         }
 681         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 682                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 683                 err = EINVAL;
 684                 goto error;
 685         }
 686         if (port_attr.state != IBV_PORT_ACTIVE)
 687                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 688                         mlx5_glue->port_state_str(port_attr.state),
 689                         port_attr.state);
 690         /* Allocate private eth device data. */
 691         priv = rte_zmalloc("ethdev private structure",
 692                            sizeof(*priv),
 693                            RTE_CACHE_LINE_SIZE);
 694         if (priv == NULL) {
 695                 DRV_LOG(ERR, "priv allocation failure");
 696                 err = ENOMEM;
 697                 goto error;
 698         }
 699         priv->sh = sh;
 700         priv->dev_port = spawn->phys_port;
 701         priv->pci_dev = spawn->pci_dev;
 702         priv->mtu = RTE_ETHER_MTU;
 703         priv->mp_id.port_id = port_id;
 704         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 705 #ifndef RTE_ARCH_64
 706         /* Initialize UAR access locks for 32bit implementations. */
 707         rte_spinlock_init(&priv->uar_lock_cq);
 708         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
 709                 rte_spinlock_init(&priv->uar_lock[i]);
 710 #endif
 711         /* Some internal functions rely on Netlink sockets, open them now. */
 712         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 713         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 714         priv->representor = !!switch_info->representor;
 715         priv->master = !!switch_info->master;
 716         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 717         priv->vport_meta_tag = 0;
 718         priv->vport_meta_mask = 0;
 719         priv->pf_bond = spawn->pf_bond;
 720 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 721         /*
 722          * The DevX port query API is implemented. E-Switch may use
 723          * either vport or reg_c[0] metadata register to match on
 724          * vport index. The engaged part of metadata register is
 725          * defined by mask.
 726          */
 727         if (switch_info->representor || switch_info->master) {
 728                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 729                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 730                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 731                                                  &devx_port);
 732                 if (err) {
 733                         DRV_LOG(WARNING,
 734                                 "can't query devx port %d on device %s",
 735                                 spawn->phys_port,
 736                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 737                         devx_port.comp_mask = 0;
 738                 }
 739         }
 740         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 741                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 742                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 743                 if (!priv->vport_meta_mask) {
 744                         DRV_LOG(ERR, "vport zero mask for port %d"
 745                                      " on bonding device %s",
 746                                      spawn->phys_port,
 747                                      mlx5_os_get_dev_device_name
 748                                                         (spawn->phys_dev));
 749                         err = ENOTSUP;
 750                         goto error;
 751                 }
 752                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 753                         DRV_LOG(ERR, "invalid vport tag for port %d"
 754                                      " on bonding device %s",
 755                                      spawn->phys_port,
 756                                      mlx5_os_get_dev_device_name
 757                                                         (spawn->phys_dev));
 758                         err = ENOTSUP;
 759                         goto error;
 760                 }
 761         }
 762         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 763                 priv->vport_id = devx_port.vport_num;
 764         } else if (spawn->pf_bond >= 0) {
 765                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 766                              " on bonding device %s",
 767                              spawn->phys_port,
 768                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 769                 err = ENOTSUP;
 770                 goto error;
 771         } else {
 772                 /* Suppose vport index in compatible way. */
 773                 priv->vport_id = switch_info->representor ?
 774                                  switch_info->port_name + 1 : -1;
 775         }
 776 #else
 777         /*
 778          * Kernel/rdma_core support single E-Switch per PF configurations
 779          * only and vport_id field contains the vport index for
 780          * associated VF, which is deduced from representor port name.
 781          * For example, let's have the IB device port 10, it has
 782          * attached network device eth0, which has port name attribute
 783          * pf0vf2, we can deduce the VF number as 2, and set vport index
 784          * as 3 (2+1). This assigning schema should be changed if the
 785          * multiple E-Switch instances per PF configurations or/and PCI
 786          * subfunctions are added.
 787          */
 788         priv->vport_id = switch_info->representor ?
 789                          switch_info->port_name + 1 : -1;
 790 #endif
 791         /* representor_id field keeps the unmodified VF index. */
 792         priv->representor_id = switch_info->representor ?
 793                                switch_info->port_name : -1;
 794         /*
 795          * Look for sibling devices in order to reuse their switch domain
 796          * if any, otherwise allocate one.
 797          */
 798         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 799                 const struct mlx5_priv *opriv =
 800                         rte_eth_devices[port_id].data->dev_private;
 801
 802                 if (!opriv ||
 803                     opriv->sh != priv->sh ||
 804                         opriv->domain_id ==
 805                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 806                         continue;
 807                 priv->domain_id = opriv->domain_id;
 808                 break;
 809         }
 810         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 811                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 812                 if (err) {
 813                         err = rte_errno;
 814                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 815                                 strerror(rte_errno));
 816                         goto error;
 817                 }
 818                 own_domain_id = 1;
 819         }
 820         /* Override some values set by hardware configuration. */
 821         mlx5_args(&config, dpdk_dev->devargs);
 822         err = mlx5_dev_check_sibling_config(priv, &config);
 823         if (err)
 824                 goto error;
 825         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 826                             IBV_DEVICE_RAW_IP_CSUM);
 827         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 828                 (config.hw_csum ? "" : "not "));
 829 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 830         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 831         DRV_LOG(DEBUG, "counters are not supported");
 832 #endif
 833 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 834         if (config.dv_flow_en) {
 835                 DRV_LOG(WARNING, "DV flow is not supported");
 836                 config.dv_flow_en = 0;
 837         }
 838 #endif
 839         config.ind_table_max_size =
 840                 sh->device_attr.max_rwq_indirection_table_size;
 841         /*
 842          * Remove this check once DPDK supports larger/variable
 843          * indirection tables.
 844          */
 845         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 846                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 847         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 848                 config.ind_table_max_size);
 849         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 850                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 851         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 852                 (config.hw_vlan_strip ? "" : "not "));
 853         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 854                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 855         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 856                 (config.hw_fcs_strip ? "" : "not "));
 857 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 858         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 859 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 860         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 861                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 862 #endif
 863         if (config.hw_padding && !hw_padding) {
 864                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 865                 config.hw_padding = 0;
 866         } else if (config.hw_padding) {
 867                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 868         }
 869         config.tso = (sh->device_attr.max_tso > 0 &&
 870                       (sh->device_attr.tso_supported_qpts &
 871                        (1 << IBV_QPT_RAW_PACKET)));
 872         if (config.tso)
 873                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 874         /*
 875          * MPW is disabled by default, while the Enhanced MPW is enabled
 876          * by default.
 877          */
 878         if (config.mps == MLX5_ARG_UNSET)
 879                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 880                                                           MLX5_MPW_DISABLED;
 881         else
 882                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 883         DRV_LOG(INFO, "%sMPS is %s",
 884                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 885                 config.mps == MLX5_MPW ? "legacy " : "",
 886                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 887         if (config.cqe_comp && !cqe_comp) {
 888                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 889                 config.cqe_comp = 0;
 890         }
 891         if (config.cqe_pad && !cqe_pad) {
 892                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 893                 config.cqe_pad = 0;
 894         } else if (config.cqe_pad) {
 895                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 896         }
 897         if (config.devx) {
 898                 priv->counter_fallback = 0;
 899                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 900                 if (err) {
 901                         err = -err;
 902                         goto error;
 903                 }
 904                 if (!config.hca_attr.flow_counters_dump)
 905                         priv->counter_fallback = 1;
 906 #ifndef HAVE_IBV_DEVX_ASYNC
 907                 priv->counter_fallback = 1;
 908 #endif
 909                 if (priv->counter_fallback)
 910                         DRV_LOG(INFO, "Use fall-back DV counter management");
 911                 /* Check for LRO support. */
 912                 if (config.dest_tir && config.hca_attr.lro_cap &&
 913                     config.dv_flow_en) {
 914                         /* TBD check tunnel lro caps. */
 915                         config.lro.supported = config.hca_attr.lro_cap;
 916                         DRV_LOG(DEBUG, "Device supports LRO");
 917                         /*
 918                          * If LRO timeout is not configured by application,
 919                          * use the minimal supported value.
 920                          */
 921                         if (!config.lro.timeout)
 922                                 config.lro.timeout =
 923                                 config.hca_attr.lro_timer_supported_periods[0];
 924                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 925                                 config.lro.timeout);
 926                 }
 927 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 928                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 929                     config.dv_flow_en) {
 930                         uint8_t reg_c_mask =
 931                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 932                         /*
 933                          * Meter needs two REG_C's for color match and pre-sfx
 934                          * flow match. Here get the REG_C for color match.
 935                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 936                          */
 937                         reg_c_mask &= 0xfc;
 938                         if (__builtin_popcount(reg_c_mask) < 1) {
 939                                 priv->mtr_en = 0;
 940                                 DRV_LOG(WARNING, "No available register for"
 941                                         " meter.");
 942                         } else {
 943                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 944                                                       REG_C_0;
 945                                 priv->mtr_en = 1;
 946                                 priv->mtr_reg_share =
 947                                       config.hca_attr.qos.flow_meter_reg_share;
 948                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 949                                         priv->mtr_color_reg);
 950                         }
 951                 }
 952 #endif
 953         }
 954         if (config.mprq.enabled && mprq) {
 955                 if (config.mprq.stride_num_n &&
 956                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
 957                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
 958                         config.mprq.stride_num_n =
 959                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
 960                                                 mprq_min_stride_num_n),
 961                                         mprq_max_stride_num_n);
 962                         DRV_LOG(WARNING,
 963                                 "the number of strides"
 964                                 " for Multi-Packet RQ is out of range,"
 965                                 " setting default value (%u)",
 966                                 1 << config.mprq.stride_num_n);
 967                 }
 968                 if (config.mprq.stride_size_n &&
 969                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
 970                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
 971                         config.mprq.stride_size_n =
 972                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
 973                                                 mprq_min_stride_size_n),
 974                                         mprq_max_stride_size_n);
 975                         DRV_LOG(WARNING,
 976                                 "the size of a stride"
 977                                 " for Multi-Packet RQ is out of range,"
 978                                 " setting default value (%u)",
 979                                 1 << config.mprq.stride_size_n);
 980                 }
 981                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
 982                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
 983         } else if (config.mprq.enabled && !mprq) {
 984                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
 985                 config.mprq.enabled = 0;
 986         }
 987         if (config.max_dump_files_num == 0)
 988                 config.max_dump_files_num = 128;
 989         eth_dev = rte_eth_dev_allocate(name);
 990         if (eth_dev == NULL) {
 991                 DRV_LOG(ERR, "can not allocate rte ethdev");
 992                 err = ENOMEM;
 993                 goto error;
 994         }
 995         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
 996         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 997         if (priv->representor) {
 998                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
 999                 eth_dev->data->representor_id = priv->representor_id;
1000         }
1001         /*
1002          * Store associated network device interface index. This index
1003          * is permanent throughout the lifetime of device. So, we may store
1004          * the ifindex here and use the cached value further.
1005          */
1006         MLX5_ASSERT(spawn->ifindex);
1007         priv->if_index = spawn->ifindex;
1008         eth_dev->data->dev_private = priv;
1009         priv->dev_data = eth_dev->data;
1010         eth_dev->data->mac_addrs = priv->mac;
1011         eth_dev->device = dpdk_dev;
1012         /* Configure the first MAC address by default. */
1013         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1014                 DRV_LOG(ERR,
1015                         "port %u cannot get MAC address, is mlx5_en"
1016                         " loaded? (errno: %s)",
1017                         eth_dev->data->port_id, strerror(rte_errno));
1018                 err = ENODEV;
1019                 goto error;
1020         }
1021         DRV_LOG(INFO,
1022                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1023                 eth_dev->data->port_id,
1024                 mac.addr_bytes[0], mac.addr_bytes[1],
1025                 mac.addr_bytes[2], mac.addr_bytes[3],
1026                 mac.addr_bytes[4], mac.addr_bytes[5]);
1027 #ifdef RTE_LIBRTE_MLX5_DEBUG
1028         {
1029                 char ifname[IF_NAMESIZE];
1030
1031                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1032                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1033                                 eth_dev->data->port_id, ifname);
1034                 else
1035                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1036                                 eth_dev->data->port_id);
1037         }
1038 #endif
1039         /* Get actual MTU if possible. */
1040         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1041         if (err) {
1042                 err = rte_errno;
1043                 goto error;
1044         }
1045         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1046                 priv->mtu);
1047         /* Initialize burst functions to prevent crashes before link-up. */
1048         eth_dev->rx_pkt_burst = removed_rx_burst;
1049         eth_dev->tx_pkt_burst = removed_tx_burst;
1050         eth_dev->dev_ops = &mlx5_os_dev_ops;
1051         /* Register MAC address. */
1052         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1053         if (config.vf && config.vf_nl_en)
1054                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1055                                       mlx5_ifindex(eth_dev),
1056                                       eth_dev->data->mac_addrs,
1057                                       MLX5_MAX_MAC_ADDRESSES);
1058         priv->flows = 0;
1059         priv->ctrl_flows = 0;
1060         TAILQ_INIT(&priv->flow_meters);
1061         TAILQ_INIT(&priv->flow_meter_profiles);
1062         /* Hint libmlx5 to use PMD allocator for data plane resources */
1063         struct mlx5dv_ctx_allocators alctr = {
1064                 .alloc = &mlx5_alloc_verbs_buf,
1065                 .free = &mlx5_free_verbs_buf,
1066                 .data = priv,
1067         };
1068         mlx5_glue->dv_set_context_attr(sh->ctx,
1069                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1070                                        (void *)((uintptr_t)&alctr));
1071         /* Bring Ethernet device up. */
1072         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1073                 eth_dev->data->port_id);
1074         mlx5_set_link_up(eth_dev);
1075         /*
1076          * Even though the interrupt handler is not installed yet,
1077          * interrupts will still trigger on the async_fd from
1078          * Verbs context returned by ibv_open_device().
1079          */
1080         mlx5_link_update(eth_dev, 0);
1081 #ifdef HAVE_MLX5DV_DR_ESWITCH
1082         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1083               (switch_info->representor || switch_info->master)))
1084                 config.dv_esw_en = 0;
1085 #else
1086         config.dv_esw_en = 0;
1087 #endif
1088         /* Detect minimal data bytes to inline. */
1089         mlx5_set_min_inline(spawn, &config);
1090         /* Store device configuration on private structure. */
1091         priv->config = config;
1092         /* Create context for virtual machine VLAN workaround. */
1093         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1094         if (config.dv_flow_en) {
1095                 err = mlx5_alloc_shared_dr(priv);
1096                 if (err)
1097                         goto error;
1098                 /*
1099                  * RSS id is shared with meter flow id. Meter flow id can only
1100                  * use the 24 MSB of the register.
1101                  */
1102                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1103                                      MLX5_MTR_COLOR_BITS);
1104                 if (!priv->qrss_id_pool) {
1105                         DRV_LOG(ERR, "can't create flow id pool");
1106                         err = ENOMEM;
1107                         goto error;
1108                 }
1109         }
1110         /* Supported Verbs flow priority number detection. */
1111         err = mlx5_flow_discover_priorities(eth_dev);
1112         if (err < 0) {
1113                 err = -err;
1114                 goto error;
1115         }
1116         priv->config.flow_prio = err;
1117         if (!priv->config.dv_esw_en &&
1118             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1119                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1120                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1121                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1122         }
1123         mlx5_set_metadata_mask(eth_dev);
1124         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1125             !priv->sh->dv_regc0_mask) {
1126                 DRV_LOG(ERR, "metadata mode %u is not supported "
1127                              "(no metadata reg_c[0] is available)",
1128                              priv->config.dv_xmeta_en);
1129                         err = ENOTSUP;
1130                         goto error;
1131         }
1132         /*
1133          * Allocate the buffer for flow creating, just once.
1134          * The allocation must be done before any flow creating.
1135          */
1136         mlx5_flow_alloc_intermediate(eth_dev);
1137         /* Query availability of metadata reg_c's. */
1138         err = mlx5_flow_discover_mreg_c(eth_dev);
1139         if (err < 0) {
1140                 err = -err;
1141                 goto error;
1142         }
1143         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1144                 DRV_LOG(DEBUG,
1145                         "port %u extensive metadata register is not supported",
1146                         eth_dev->data->port_id);
1147                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1148                         DRV_LOG(ERR, "metadata mode %u is not supported "
1149                                      "(no metadata registers available)",
1150                                      priv->config.dv_xmeta_en);
1151                         err = ENOTSUP;
1152                         goto error;
1153                 }
1154         }
1155         if (priv->config.dv_flow_en &&
1156             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1157             mlx5_flow_ext_mreg_supported(eth_dev) &&
1158             priv->sh->dv_regc0_mask) {
1159                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1160                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1161                 if (!priv->mreg_cp_tbl) {
1162                         err = ENOMEM;
1163                         goto error;
1164                 }
1165         }
1166         return eth_dev;
1167 error:
1168         if (priv) {
1169                 if (priv->mreg_cp_tbl)
1170                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1171                 if (priv->sh)
1172                         mlx5_os_free_shared_dr(priv);
1173                 if (priv->nl_socket_route >= 0)
1174                         close(priv->nl_socket_route);
1175                 if (priv->nl_socket_rdma >= 0)
1176                         close(priv->nl_socket_rdma);
1177                 if (priv->vmwa_context)
1178                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1179                 if (priv->qrss_id_pool)
1180                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1181                 if (own_domain_id)
1182                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1183                 rte_free(priv);
1184                 if (eth_dev != NULL)
1185                         eth_dev->data->dev_private = NULL;
1186         }
1187         if (eth_dev != NULL) {
1188                 /* mac_addrs must not be freed alone because part of
1189                  * dev_private
1190                  **/
1191                 eth_dev->data->mac_addrs = NULL;
1192                 rte_eth_dev_release_port(eth_dev);
1193         }
1194         if (sh)
1195                 mlx5_free_shared_dev_ctx(sh);
1196         MLX5_ASSERT(err > 0);
1197         rte_errno = err;
1198         return NULL;
1199 }
1200
1201 /**
1202  * Comparison callback to sort device data.
1203  *
1204  * This is meant to be used with qsort().
1205  *
1206  * @param a[in]
1207  *   Pointer to pointer to first data object.
1208  * @param b[in]
1209  *   Pointer to pointer to second data object.
1210  *
1211  * @return
1212  *   0 if both objects are equal, less than 0 if the first argument is less
1213  *   than the second, greater than 0 otherwise.
1214  */
1215 static int
1216 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1217 {
1218         const struct mlx5_switch_info *si_a =
1219                 &((const struct mlx5_dev_spawn_data *)a)->info;
1220         const struct mlx5_switch_info *si_b =
1221                 &((const struct mlx5_dev_spawn_data *)b)->info;
1222         int ret;
1223
1224         /* Master device first. */
1225         ret = si_b->master - si_a->master;
1226         if (ret)
1227                 return ret;
1228         /* Then representor devices. */
1229         ret = si_b->representor - si_a->representor;
1230         if (ret)
1231                 return ret;
1232         /* Unidentified devices come last in no specific order. */
1233         if (!si_a->representor)
1234                 return 0;
1235         /* Order representors by name. */
1236         return si_a->port_name - si_b->port_name;
1237 }
1238
1239 /**
1240  * Match PCI information for possible slaves of bonding device.
1241  *
1242  * @param[in] ibv_dev
1243  *   Pointer to Infiniband device structure.
1244  * @param[in] pci_dev
1245  *   Pointer to PCI device structure to match PCI address.
1246  * @param[in] nl_rdma
1247  *   Netlink RDMA group socket handle.
1248  *
1249  * @return
1250  *   negative value if no bonding device found, otherwise
1251  *   positive index of slave PF in bonding.
1252  */
1253 static int
1254 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1255                            const struct rte_pci_device *pci_dev,
1256                            int nl_rdma)
1257 {
1258         char ifname[IF_NAMESIZE + 1];
1259         unsigned int ifindex;
1260         unsigned int np, i;
1261         FILE *file = NULL;
1262         int pf = -1;
1263
1264         /*
1265          * Try to get master device name. If something goes
1266          * wrong suppose the lack of kernel support and no
1267          * bonding devices.
1268          */
1269         if (nl_rdma < 0)
1270                 return -1;
1271         if (!strstr(ibv_dev->name, "bond"))
1272                 return -1;
1273         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1274         if (!np)
1275                 return -1;
1276         /*
1277          * The Master device might not be on the predefined
1278          * port (not on port index 1, it is not garanted),
1279          * we have to scan all Infiniband device port and
1280          * find master.
1281          */
1282         for (i = 1; i <= np; ++i) {
1283                 /* Check whether Infiniband port is populated. */
1284                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1285                 if (!ifindex)
1286                         continue;
1287                 if (!if_indextoname(ifindex, ifname))
1288                         continue;
1289                 /* Try to read bonding slave names from sysfs. */
1290                 MKSTR(slaves,
1291                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1292                 file = fopen(slaves, "r");
1293                 if (file)
1294                         break;
1295         }
1296         if (!file)
1297                 return -1;
1298         /* Use safe format to check maximal buffer length. */
1299         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1300         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1301                 char tmp_str[IF_NAMESIZE + 32];
1302                 struct rte_pci_addr pci_addr;
1303                 struct mlx5_switch_info info;
1304
1305                 /* Process slave interface names in the loop. */
1306                 snprintf(tmp_str, sizeof(tmp_str),
1307                          "/sys/class/net/%s", ifname);
1308                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1309                         DRV_LOG(WARNING, "can not get PCI address"
1310                                          " for netdev \"%s\"", ifname);
1311                         continue;
1312                 }
1313                 if (pci_dev->addr.domain != pci_addr.domain ||
1314                     pci_dev->addr.bus != pci_addr.bus ||
1315                     pci_dev->addr.devid != pci_addr.devid ||
1316                     pci_dev->addr.function != pci_addr.function)
1317                         continue;
1318                 /* Slave interface PCI address match found. */
1319                 fclose(file);
1320                 snprintf(tmp_str, sizeof(tmp_str),
1321                          "/sys/class/net/%s/phys_port_name", ifname);
1322                 file = fopen(tmp_str, "rb");
1323                 if (!file)
1324                         break;
1325                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1326                 if (fscanf(file, "%32s", tmp_str) == 1)
1327                         mlx5_translate_port_name(tmp_str, &info);
1328                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1329                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1330                         pf = info.port_name;
1331                 break;
1332         }
1333         if (file)
1334                 fclose(file);
1335         return pf;
1336 }
1337
1338 /**
1339  * DPDK callback to register a PCI device.
1340  *
1341  * This function spawns Ethernet devices out of a given PCI device.
1342  *
1343  * @param[in] pci_drv
1344  *   PCI driver structure (mlx5_driver).
1345  * @param[in] pci_dev
1346  *   PCI device information.
1347  *
1348  * @return
1349  *   0 on success, a negative errno value otherwise and rte_errno is set.
1350  */
1351 int
1352 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1353                   struct rte_pci_device *pci_dev)
1354 {
1355         struct ibv_device **ibv_list;
1356         /*
1357          * Number of found IB Devices matching with requested PCI BDF.
1358          * nd != 1 means there are multiple IB devices over the same
1359          * PCI device and we have representors and master.
1360          */
1361         unsigned int nd = 0;
1362         /*
1363          * Number of found IB device Ports. nd = 1 and np = 1..n means
1364          * we have the single multiport IB device, and there may be
1365          * representors attached to some of found ports.
1366          */
1367         unsigned int np = 0;
1368         /*
1369          * Number of DPDK ethernet devices to Spawn - either over
1370          * multiple IB devices or multiple ports of single IB device.
1371          * Actually this is the number of iterations to spawn.
1372          */
1373         unsigned int ns = 0;
1374         /*
1375          * Bonding device
1376          *   < 0 - no bonding device (single one)
1377          *  >= 0 - bonding device (value is slave PF index)
1378          */
1379         int bd = -1;
1380         struct mlx5_dev_spawn_data *list = NULL;
1381         struct mlx5_dev_config dev_config;
1382         int ret;
1383
1384         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1385                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1386                         " driver.");
1387                 return 1;
1388         }
1389         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1390                 mlx5_pmd_socket_init();
1391         ret = mlx5_init_once();
1392         if (ret) {
1393                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1394                         strerror(rte_errno));
1395                 return -rte_errno;
1396         }
1397         MLX5_ASSERT(pci_drv == &mlx5_driver);
1398         errno = 0;
1399         ibv_list = mlx5_glue->get_device_list(&ret);
1400         if (!ibv_list) {
1401                 rte_errno = errno ? errno : ENOSYS;
1402                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1403                 return -rte_errno;
1404         }
1405         /*
1406          * First scan the list of all Infiniband devices to find
1407          * matching ones, gathering into the list.
1408          */
1409         struct ibv_device *ibv_match[ret + 1];
1410         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1411         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1412         unsigned int i;
1413
1414         while (ret-- > 0) {
1415                 struct rte_pci_addr pci_addr;
1416
1417                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1418                 bd = mlx5_device_bond_pci_match
1419                                 (ibv_list[ret], pci_dev, nl_rdma);
1420                 if (bd >= 0) {
1421                         /*
1422                          * Bonding device detected. Only one match is allowed,
1423                          * the bonding is supported over multi-port IB device,
1424                          * there should be no matches on representor PCI
1425                          * functions or non VF LAG bonding devices with
1426                          * specified address.
1427                          */
1428                         if (nd) {
1429                                 DRV_LOG(ERR,
1430                                         "multiple PCI match on bonding device"
1431                                         "\"%s\" found", ibv_list[ret]->name);
1432                                 rte_errno = ENOENT;
1433                                 ret = -rte_errno;
1434                                 goto exit;
1435                         }
1436                         DRV_LOG(INFO, "PCI information matches for"
1437                                       " slave %d bonding device \"%s\"",
1438                                       bd, ibv_list[ret]->name);
1439                         ibv_match[nd++] = ibv_list[ret];
1440                         break;
1441                 }
1442                 if (mlx5_dev_to_pci_addr
1443                         (ibv_list[ret]->ibdev_path, &pci_addr))
1444                         continue;
1445                 if (pci_dev->addr.domain != pci_addr.domain ||
1446                     pci_dev->addr.bus != pci_addr.bus ||
1447                     pci_dev->addr.devid != pci_addr.devid ||
1448                     pci_dev->addr.function != pci_addr.function)
1449                         continue;
1450                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1451                         ibv_list[ret]->name);
1452                 ibv_match[nd++] = ibv_list[ret];
1453         }
1454         ibv_match[nd] = NULL;
1455         if (!nd) {
1456                 /* No device matches, just complain and bail out. */
1457                 DRV_LOG(WARNING,
1458                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1459                         " are kernel drivers loaded?",
1460                         pci_dev->addr.domain, pci_dev->addr.bus,
1461                         pci_dev->addr.devid, pci_dev->addr.function);
1462                 rte_errno = ENOENT;
1463                 ret = -rte_errno;
1464                 goto exit;
1465         }
1466         if (nd == 1) {
1467                 /*
1468                  * Found single matching device may have multiple ports.
1469                  * Each port may be representor, we have to check the port
1470                  * number and check the representors existence.
1471                  */
1472                 if (nl_rdma >= 0)
1473                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1474                 if (!np)
1475                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1476                                          " ports number", ibv_match[0]->name);
1477                 if (bd >= 0 && !np) {
1478                         DRV_LOG(ERR, "can not get ports"
1479                                      " for bonding device");
1480                         rte_errno = ENOENT;
1481                         ret = -rte_errno;
1482                         goto exit;
1483                 }
1484         }
1485 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1486         if (bd >= 0) {
1487                 /*
1488                  * This may happen if there is VF LAG kernel support and
1489                  * application is compiled with older rdma_core library.
1490                  */
1491                 DRV_LOG(ERR,
1492                         "No kernel/verbs support for VF LAG bonding found.");
1493                 rte_errno = ENOTSUP;
1494                 ret = -rte_errno;
1495                 goto exit;
1496         }
1497 #endif
1498         /*
1499          * Now we can determine the maximal
1500          * amount of devices to be spawned.
1501          */
1502         list = rte_zmalloc("device spawn data",
1503                          sizeof(struct mlx5_dev_spawn_data) *
1504                          (np ? np : nd),
1505                          RTE_CACHE_LINE_SIZE);
1506         if (!list) {
1507                 DRV_LOG(ERR, "spawn data array allocation failure");
1508                 rte_errno = ENOMEM;
1509                 ret = -rte_errno;
1510                 goto exit;
1511         }
1512         if (bd >= 0 || np > 1) {
1513                 /*
1514                  * Single IB device with multiple ports found,
1515                  * it may be E-Switch master device and representors.
1516                  * We have to perform identification through the ports.
1517                  */
1518                 MLX5_ASSERT(nl_rdma >= 0);
1519                 MLX5_ASSERT(ns == 0);
1520                 MLX5_ASSERT(nd == 1);
1521                 MLX5_ASSERT(np);
1522                 for (i = 1; i <= np; ++i) {
1523                         list[ns].max_port = np;
1524                         list[ns].phys_port = i;
1525                         list[ns].phys_dev = ibv_match[0];
1526                         list[ns].eth_dev = NULL;
1527                         list[ns].pci_dev = pci_dev;
1528                         list[ns].pf_bond = bd;
1529                         list[ns].ifindex = mlx5_nl_ifindex
1530                                 (nl_rdma,
1531                                 mlx5_os_get_dev_device_name
1532                                                 (list[ns].phys_dev), i);
1533                         if (!list[ns].ifindex) {
1534                                 /*
1535                                  * No network interface index found for the
1536                                  * specified port, it means there is no
1537                                  * representor on this port. It's OK,
1538                                  * there can be disabled ports, for example
1539                                  * if sriov_numvfs < sriov_totalvfs.
1540                                  */
1541                                 continue;
1542                         }
1543                         ret = -1;
1544                         if (nl_route >= 0)
1545                                 ret = mlx5_nl_switch_info
1546                                                (nl_route,
1547                                                 list[ns].ifindex,
1548                                                 &list[ns].info);
1549                         if (ret || (!list[ns].info.representor &&
1550                                     !list[ns].info.master)) {
1551                                 /*
1552                                  * We failed to recognize representors with
1553                                  * Netlink, let's try to perform the task
1554                                  * with sysfs.
1555                                  */
1556                                 ret =  mlx5_sysfs_switch_info
1557                                                 (list[ns].ifindex,
1558                                                  &list[ns].info);
1559                         }
1560                         if (!ret && bd >= 0) {
1561                                 switch (list[ns].info.name_type) {
1562                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1563                                         if (list[ns].info.port_name == bd)
1564                                                 ns++;
1565                                         break;
1566                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1567                                         if (list[ns].info.pf_num == bd)
1568                                                 ns++;
1569                                         break;
1570                                 default:
1571                                         break;
1572                                 }
1573                                 continue;
1574                         }
1575                         if (!ret && (list[ns].info.representor ^
1576                                      list[ns].info.master))
1577                                 ns++;
1578                 }
1579                 if (!ns) {
1580                         DRV_LOG(ERR,
1581                                 "unable to recognize master/representors"
1582                                 " on the IB device with multiple ports");
1583                         rte_errno = ENOENT;
1584                         ret = -rte_errno;
1585                         goto exit;
1586                 }
1587         } else {
1588                 /*
1589                  * The existence of several matching entries (nd > 1) means
1590                  * port representors have been instantiated. No existing Verbs
1591                  * call nor sysfs entries can tell them apart, this can only
1592                  * be done through Netlink calls assuming kernel drivers are
1593                  * recent enough to support them.
1594                  *
1595                  * In the event of identification failure through Netlink,
1596                  * try again through sysfs, then:
1597                  *
1598                  * 1. A single IB device matches (nd == 1) with single
1599                  *    port (np=0/1) and is not a representor, assume
1600                  *    no switch support.
1601                  *
1602                  * 2. Otherwise no safe assumptions can be made;
1603                  *    complain louder and bail out.
1604                  */
1605                 for (i = 0; i != nd; ++i) {
1606                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1607                         list[ns].max_port = 1;
1608                         list[ns].phys_port = 1;
1609                         list[ns].phys_dev = ibv_match[i];
1610                         list[ns].eth_dev = NULL;
1611                         list[ns].pci_dev = pci_dev;
1612                         list[ns].pf_bond = -1;
1613                         list[ns].ifindex = 0;
1614                         if (nl_rdma >= 0)
1615                                 list[ns].ifindex = mlx5_nl_ifindex
1616                                 (nl_rdma,
1617                                 mlx5_os_get_dev_device_name
1618                                                 (list[ns].phys_dev), 1);
1619                         if (!list[ns].ifindex) {
1620                                 char ifname[IF_NAMESIZE];
1621
1622                                 /*
1623                                  * Netlink failed, it may happen with old
1624                                  * ib_core kernel driver (before 4.16).
1625                                  * We can assume there is old driver because
1626                                  * here we are processing single ports IB
1627                                  * devices. Let's try sysfs to retrieve
1628                                  * the ifindex. The method works for
1629                                  * master device only.
1630                                  */
1631                                 if (nd > 1) {
1632                                         /*
1633                                          * Multiple devices found, assume
1634                                          * representors, can not distinguish
1635                                          * master/representor and retrieve
1636                                          * ifindex via sysfs.
1637                                          */
1638                                         continue;
1639                                 }
1640                                 ret = mlx5_get_master_ifname
1641                                         (ibv_match[i]->ibdev_path, &ifname);
1642                                 if (!ret)
1643                                         list[ns].ifindex =
1644                                                 if_nametoindex(ifname);
1645                                 if (!list[ns].ifindex) {
1646                                         /*
1647                                          * No network interface index found
1648                                          * for the specified device, it means
1649                                          * there it is neither representor
1650                                          * nor master.
1651                                          */
1652                                         continue;
1653                                 }
1654                         }
1655                         ret = -1;
1656                         if (nl_route >= 0)
1657                                 ret = mlx5_nl_switch_info
1658                                                (nl_route,
1659                                                 list[ns].ifindex,
1660                                                 &list[ns].info);
1661                         if (ret || (!list[ns].info.representor &&
1662                                     !list[ns].info.master)) {
1663                                 /*
1664                                  * We failed to recognize representors with
1665                                  * Netlink, let's try to perform the task
1666                                  * with sysfs.
1667                                  */
1668                                 ret =  mlx5_sysfs_switch_info
1669                                                 (list[ns].ifindex,
1670                                                  &list[ns].info);
1671                         }
1672                         if (!ret && (list[ns].info.representor ^
1673                                      list[ns].info.master)) {
1674                                 ns++;
1675                         } else if ((nd == 1) &&
1676                                    !list[ns].info.representor &&
1677                                    !list[ns].info.master) {
1678                                 /*
1679                                  * Single IB device with
1680                                  * one physical port and
1681                                  * attached network device.
1682                                  * May be SRIOV is not enabled
1683                                  * or there is no representors.
1684                                  */
1685                                 DRV_LOG(INFO, "no E-Switch support detected");
1686                                 ns++;
1687                                 break;
1688                         }
1689                 }
1690                 if (!ns) {
1691                         DRV_LOG(ERR,
1692                                 "unable to recognize master/representors"
1693                                 " on the multiple IB devices");
1694                         rte_errno = ENOENT;
1695                         ret = -rte_errno;
1696                         goto exit;
1697                 }
1698         }
1699         MLX5_ASSERT(ns);
1700         /*
1701          * Sort list to probe devices in natural order for users convenience
1702          * (i.e. master first, then representors from lowest to highest ID).
1703          */
1704         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1705         /* Default configuration. */
1706         dev_config = (struct mlx5_dev_config){
1707                 .hw_padding = 0,
1708                 .mps = MLX5_ARG_UNSET,
1709                 .dbnc = MLX5_ARG_UNSET,
1710                 .rx_vec_en = 1,
1711                 .txq_inline_max = MLX5_ARG_UNSET,
1712                 .txq_inline_min = MLX5_ARG_UNSET,
1713                 .txq_inline_mpw = MLX5_ARG_UNSET,
1714                 .txqs_inline = MLX5_ARG_UNSET,
1715                 .vf_nl_en = 1,
1716                 .mr_ext_memseg_en = 1,
1717                 .mprq = {
1718                         .enabled = 0, /* Disabled by default. */
1719                         .stride_num_n = 0,
1720                         .stride_size_n = 0,
1721                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1722                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1723                 },
1724                 .dv_esw_en = 1,
1725                 .dv_flow_en = 1,
1726                 .log_hp_size = MLX5_ARG_UNSET,
1727         };
1728         /* Device specific configuration. */
1729         switch (pci_dev->id.device_id) {
1730         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1731         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1732         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1733         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1734         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1735         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1736         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1737                 dev_config.vf = 1;
1738                 break;
1739         default:
1740                 break;
1741         }
1742         for (i = 0; i != ns; ++i) {
1743                 uint32_t restore;
1744
1745                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1746                                                  &list[i],
1747                                                  dev_config);
1748                 if (!list[i].eth_dev) {
1749                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1750                                 break;
1751                         /* Device is disabled or already spawned. Ignore it. */
1752                         continue;
1753                 }
1754                 restore = list[i].eth_dev->data->dev_flags;
1755                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1756                 /* Restore non-PCI flags cleared by the above call. */
1757                 list[i].eth_dev->data->dev_flags |= restore;
1758                 rte_eth_dev_probing_finish(list[i].eth_dev);
1759         }
1760         if (i != ns) {
1761                 DRV_LOG(ERR,
1762                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1763                         " encountering an error: %s",
1764                         pci_dev->addr.domain, pci_dev->addr.bus,
1765                         pci_dev->addr.devid, pci_dev->addr.function,
1766                         strerror(rte_errno));
1767                 ret = -rte_errno;
1768                 /* Roll back. */
1769                 while (i--) {
1770                         if (!list[i].eth_dev)
1771                                 continue;
1772                         mlx5_dev_close(list[i].eth_dev);
1773                         /* mac_addrs must not be freed because in dev_private */
1774                         list[i].eth_dev->data->mac_addrs = NULL;
1775                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1776                 }
1777                 /* Restore original error. */
1778                 rte_errno = -ret;
1779         } else {
1780                 ret = 0;
1781         }
1782 exit:
1783         /*
1784          * Do the routine cleanup:
1785          * - close opened Netlink sockets
1786          * - free allocated spawn data array
1787          * - free the Infiniband device list
1788          */
1789         if (nl_rdma >= 0)
1790                 close(nl_rdma);
1791         if (nl_route >= 0)
1792                 close(nl_route);
1793         if (list)
1794                 rte_free(list);
1795         MLX5_ASSERT(ibv_list);
1796         mlx5_glue->free_device_list(ibv_list);
1797         return ret;
1798 }
1799
1800 static int
1801 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1802 {
1803         char *env;
1804         int value;
1805
1806         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1807         /* Get environment variable to store. */
1808         env = getenv(MLX5_SHUT_UP_BF);
1809         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1810         if (config->dbnc == MLX5_ARG_UNSET)
1811                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1812         else
1813                 setenv(MLX5_SHUT_UP_BF,
1814                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1815         return value;
1816 }
1817
1818 static void
1819 mlx5_restore_doorbell_mapping_env(int value)
1820 {
1821         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1822         /* Restore the original environment variable state. */
1823         if (value == MLX5_ARG_UNSET)
1824                 unsetenv(MLX5_SHUT_UP_BF);
1825         else
1826                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1827 }
1828
1829 /**
1830  * Extract pdn of PD object using DV API.
1831  *
1832  * @param[in] pd
1833  *   Pointer to the verbs PD object.
1834  * @param[out] pdn
1835  *   Pointer to the PD object number variable.
1836  *
1837  * @return
1838  *   0 on success, error value otherwise.
1839  */
1840 int
1841 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1842 {
1843 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1844         struct mlx5dv_obj obj;
1845         struct mlx5dv_pd pd_info;
1846         int ret = 0;
1847
1848         obj.pd.in = pd;
1849         obj.pd.out = &pd_info;
1850         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1851         if (ret) {
1852                 DRV_LOG(DEBUG, "Fail to get PD object info");
1853                 return ret;
1854         }
1855         *pdn = pd_info.pdn;
1856         return 0;
1857 #else
1858         (void)pd;
1859         (void)pdn;
1860         return -ENOTSUP;
1861 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1862 }
1863
1864 /**
1865  * Function API to open IB device.
1866  *
1867  * This function calls the Linux glue APIs to open a device.
1868  *
1869  * @param[in] spawn
1870  *   Pointer to the IB device attributes (name, port, etc).
1871  * @param[out] config
1872  *   Pointer to device configuration structure.
1873  * @param[out] sh
1874  *   Pointer to shared context structure.
1875  *
1876  * @return
1877  *   0 on success, a positive error value otherwise.
1878  */
1879 int
1880 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
1881                      const struct mlx5_dev_config *config,
1882                      struct mlx5_dev_ctx_shared *sh)
1883 {
1884         int dbmap_env;
1885         int err = 0;
1886         /*
1887          * Configure environment variable "MLX5_BF_SHUT_UP"
1888          * before the device creation. The rdma_core library
1889          * checks the variable at device creation and
1890          * stores the result internally.
1891          */
1892         dbmap_env = mlx5_config_doorbell_mapping_env(config);
1893         /* Try to open IB device with DV first, then usual Verbs. */
1894         errno = 0;
1895         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
1896         if (sh->ctx) {
1897                 sh->devx = 1;
1898                 DRV_LOG(DEBUG, "DevX is supported");
1899                 /* The device is created, no need for environment. */
1900                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1901         } else {
1902                 /* The environment variable is still configured. */
1903                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
1904                 err = errno ? errno : ENODEV;
1905                 /*
1906                  * The environment variable is not needed anymore,
1907                  * all device creation attempts are completed.
1908                  */
1909                 mlx5_restore_doorbell_mapping_env(dbmap_env);
1910                 if (!sh->ctx)
1911                         return err;
1912                 DRV_LOG(DEBUG, "DevX is NOT supported");
1913                 err = 0;
1914         }
1915         return err;
1916 }
1917
1918 /**
1919  * Install shared asynchronous device events handler.
1920  * This function is implemented to support event sharing
1921  * between multiple ports of single IB device.
1922  *
1923  * @param sh
1924  *   Pointer to mlx5_dev_ctx_shared object.
1925  */
1926 void
1927 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
1928 {
1929         int ret;
1930         int flags;
1931
1932         sh->intr_handle.fd = -1;
1933         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
1934         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
1935                     F_SETFL, flags | O_NONBLOCK);
1936         if (ret) {
1937                 DRV_LOG(INFO, "failed to change file descriptor async event"
1938                         " queue");
1939         } else {
1940                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
1941                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
1942                 if (rte_intr_callback_register(&sh->intr_handle,
1943                                         mlx5_dev_interrupt_handler, sh)) {
1944                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
1945                         sh->intr_handle.fd = -1;
1946                 }
1947         }
1948         if (sh->devx) {
1949 #ifdef HAVE_IBV_DEVX_ASYNC
1950                 sh->intr_handle_devx.fd = -1;
1951                 sh->devx_comp =
1952                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
1953                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
1954                 if (!devx_comp) {
1955                         DRV_LOG(INFO, "failed to allocate devx_comp.");
1956                         return;
1957                 }
1958                 flags = fcntl(devx_comp->fd, F_GETFL);
1959                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
1960                 if (ret) {
1961                         DRV_LOG(INFO, "failed to change file descriptor"
1962                                 " devx comp");
1963                         return;
1964                 }
1965                 sh->intr_handle_devx.fd = devx_comp->fd;
1966                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
1967                 if (rte_intr_callback_register(&sh->intr_handle_devx,
1968                                         mlx5_dev_interrupt_handler_devx, sh)) {
1969                         DRV_LOG(INFO, "Fail to install the devx shared"
1970                                 " interrupt.");
1971                         sh->intr_handle_devx.fd = -1;
1972                 }
1973 #endif /* HAVE_IBV_DEVX_ASYNC */
1974         }
1975 }
1976
1977 /**
1978  * Uninstall shared asynchronous device events handler.
1979  * This function is implemented to support event sharing
1980  * between multiple ports of single IB device.
1981  *
1982  * @param dev
1983  *   Pointer to mlx5_dev_ctx_shared object.
1984  */
1985 void
1986 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
1987 {
1988         if (sh->intr_handle.fd >= 0)
1989                 mlx5_intr_callback_unregister(&sh->intr_handle,
1990                                               mlx5_dev_interrupt_handler, sh);
1991 #ifdef HAVE_IBV_DEVX_ASYNC
1992         if (sh->intr_handle_devx.fd >= 0)
1993                 rte_intr_callback_unregister(&sh->intr_handle_devx,
1994                                   mlx5_dev_interrupt_handler_devx, sh);
1995         if (sh->devx_comp)
1996                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
1997 #endif
1998 }
1999
2000 /**
2001  * Read statistics by a named counter.
2002  *
2003  * @param[in] priv
2004  *   Pointer to the private device data structure.
2005  * @param[in] ctr_name
2006  *   Pointer to the name of the statistic counter to read
2007  * @param[out] stat
2008  *   Pointer to read statistic value.
2009  * @return
2010  *   0 on success and stat is valud, 1 if failed to read the value
2011  *   rte_errno is set.
2012  *
2013  */
2014 int
2015 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2016                       uint64_t *stat)
2017 {
2018         int fd;
2019
2020         if (priv->sh) {
2021                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2022                           priv->sh->ibdev_path,
2023                           priv->dev_port,
2024                           ctr_name);
2025                 fd = open(path, O_RDONLY);
2026                 if (fd != -1) {
2027                         char buf[21] = {'\0'};
2028                         ssize_t n = read(fd, buf, sizeof(buf));
2029
2030                         close(fd);
2031                         if (n != -1) {
2032                                 *stat = strtoull(buf, NULL, 10);
2033                                 return 0;
2034                         }
2035                 }
2036         }
2037         *stat = 0;
2038         return 1;
2039 }
2040
2041 /**
2042  * Read device counters table.
2043  *
2044  * @param dev
2045  *   Pointer to Ethernet device.
2046  * @param[out] stats
2047  *   Counters table output buffer.
2048  *
2049  * @return
2050  *   0 on success and stats is filled, negative errno value otherwise and
2051  *   rte_errno is set.
2052  */
2053 int
2054 mlx5_os_read_dev_counters(struct rte_eth_dev *dev, uint64_t *stats)
2055 {
2056         struct mlx5_priv *priv = dev->data->dev_private;
2057         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2058         unsigned int i;
2059         struct ifreq ifr;
2060         unsigned int stats_sz = xstats_ctrl->stats_n * sizeof(uint64_t);
2061         unsigned char et_stat_buf[sizeof(struct ethtool_stats) + stats_sz];
2062         struct ethtool_stats *et_stats = (struct ethtool_stats *)et_stat_buf;
2063         int ret;
2064
2065         et_stats->cmd = ETHTOOL_GSTATS;
2066         et_stats->n_stats = xstats_ctrl->stats_n;
2067         ifr.ifr_data = (caddr_t)et_stats;
2068         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2069         if (ret) {
2070                 DRV_LOG(WARNING,
2071                         "port %u unable to read statistic values from device",
2072                         dev->data->port_id);
2073                 return ret;
2074         }
2075         for (i = 0; i != xstats_ctrl->mlx5_stats_n; ++i) {
2076                 if (xstats_ctrl->info[i].dev) {
2077                         ret = mlx5_os_read_dev_stat(priv,
2078                                             xstats_ctrl->info[i].ctr_name,
2079                                             &stats[i]);
2080                         /* return last xstats counter if fail to read. */
2081                         if (ret == 0)
2082                                 xstats_ctrl->xstats[i] = stats[i];
2083                         else
2084                                 stats[i] = xstats_ctrl->xstats[i];
2085                 } else {
2086                         stats[i] = (uint64_t)
2087                                 et_stats->data[xstats_ctrl->dev_table_idx[i]];
2088                 }
2089         }
2090         return 0;
2091 }
2092
2093 /**
2094  * Query the number of statistics provided by ETHTOOL.
2095  *
2096  * @param dev
2097  *   Pointer to Ethernet device.
2098  *
2099  * @return
2100  *   Number of statistics on success, negative errno value otherwise and
2101  *   rte_errno is set.
2102  */
2103 int
2104 mlx5_os_get_stats_n(struct rte_eth_dev *dev)
2105 {
2106         struct ethtool_drvinfo drvinfo;
2107         struct ifreq ifr;
2108         int ret;
2109
2110         drvinfo.cmd = ETHTOOL_GDRVINFO;
2111         ifr.ifr_data = (caddr_t)&drvinfo;
2112         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2113         if (ret) {
2114                 DRV_LOG(WARNING, "port %u unable to query number of statistics",
2115                         dev->data->port_id);
2116                 return ret;
2117         }
2118         return drvinfo.n_stats;
2119 }
2120
2121 static const struct mlx5_counter_ctrl mlx5_counters_init[] = {
2122         {
2123                 .dpdk_name = "rx_port_unicast_bytes",
2124                 .ctr_name = "rx_vport_unicast_bytes",
2125         },
2126         {
2127                 .dpdk_name = "rx_port_multicast_bytes",
2128                 .ctr_name = "rx_vport_multicast_bytes",
2129         },
2130         {
2131                 .dpdk_name = "rx_port_broadcast_bytes",
2132                 .ctr_name = "rx_vport_broadcast_bytes",
2133         },
2134         {
2135                 .dpdk_name = "rx_port_unicast_packets",
2136                 .ctr_name = "rx_vport_unicast_packets",
2137         },
2138         {
2139                 .dpdk_name = "rx_port_multicast_packets",
2140                 .ctr_name = "rx_vport_multicast_packets",
2141         },
2142         {
2143                 .dpdk_name = "rx_port_broadcast_packets",
2144                 .ctr_name = "rx_vport_broadcast_packets",
2145         },
2146         {
2147                 .dpdk_name = "tx_port_unicast_bytes",
2148                 .ctr_name = "tx_vport_unicast_bytes",
2149         },
2150         {
2151                 .dpdk_name = "tx_port_multicast_bytes",
2152                 .ctr_name = "tx_vport_multicast_bytes",
2153         },
2154         {
2155                 .dpdk_name = "tx_port_broadcast_bytes",
2156                 .ctr_name = "tx_vport_broadcast_bytes",
2157         },
2158         {
2159                 .dpdk_name = "tx_port_unicast_packets",
2160                 .ctr_name = "tx_vport_unicast_packets",
2161         },
2162         {
2163                 .dpdk_name = "tx_port_multicast_packets",
2164                 .ctr_name = "tx_vport_multicast_packets",
2165         },
2166         {
2167                 .dpdk_name = "tx_port_broadcast_packets",
2168                 .ctr_name = "tx_vport_broadcast_packets",
2169         },
2170         {
2171                 .dpdk_name = "rx_wqe_err",
2172                 .ctr_name = "rx_wqe_err",
2173         },
2174         {
2175                 .dpdk_name = "rx_crc_errors_phy",
2176                 .ctr_name = "rx_crc_errors_phy",
2177         },
2178         {
2179                 .dpdk_name = "rx_in_range_len_errors_phy",
2180                 .ctr_name = "rx_in_range_len_errors_phy",
2181         },
2182         {
2183                 .dpdk_name = "rx_symbol_err_phy",
2184                 .ctr_name = "rx_symbol_err_phy",
2185         },
2186         {
2187                 .dpdk_name = "tx_errors_phy",
2188                 .ctr_name = "tx_errors_phy",
2189         },
2190         {
2191                 .dpdk_name = "rx_out_of_buffer",
2192                 .ctr_name = "out_of_buffer",
2193                 .dev = 1,
2194         },
2195         {
2196                 .dpdk_name = "tx_packets_phy",
2197                 .ctr_name = "tx_packets_phy",
2198         },
2199         {
2200                 .dpdk_name = "rx_packets_phy",
2201                 .ctr_name = "rx_packets_phy",
2202         },
2203         {
2204                 .dpdk_name = "tx_discards_phy",
2205                 .ctr_name = "tx_discards_phy",
2206         },
2207         {
2208                 .dpdk_name = "rx_discards_phy",
2209                 .ctr_name = "rx_discards_phy",
2210         },
2211         {
2212                 .dpdk_name = "tx_bytes_phy",
2213                 .ctr_name = "tx_bytes_phy",
2214         },
2215         {
2216                 .dpdk_name = "rx_bytes_phy",
2217                 .ctr_name = "rx_bytes_phy",
2218         },
2219         /* Representor only */
2220         {
2221                 .dpdk_name = "rx_packets",
2222                 .ctr_name = "vport_rx_packets",
2223         },
2224         {
2225                 .dpdk_name = "rx_bytes",
2226                 .ctr_name = "vport_rx_bytes",
2227         },
2228         {
2229                 .dpdk_name = "tx_packets",
2230                 .ctr_name = "vport_tx_packets",
2231         },
2232         {
2233                 .dpdk_name = "tx_bytes",
2234                 .ctr_name = "vport_tx_bytes",
2235         },
2236 };
2237
2238 static const unsigned int xstats_n = RTE_DIM(mlx5_counters_init);
2239
2240 /**
2241  * Init the structures to read device counters.
2242  *
2243  * @param dev
2244  *   Pointer to Ethernet device.
2245  */
2246 void
2247 mlx5_os_stats_init(struct rte_eth_dev *dev)
2248 {
2249         struct mlx5_priv *priv = dev->data->dev_private;
2250         struct mlx5_xstats_ctrl *xstats_ctrl = &priv->xstats_ctrl;
2251         struct mlx5_stats_ctrl *stats_ctrl = &priv->stats_ctrl;
2252         unsigned int i;
2253         unsigned int j;
2254         struct ifreq ifr;
2255         struct ethtool_gstrings *strings = NULL;
2256         unsigned int dev_stats_n;
2257         unsigned int str_sz;
2258         int ret;
2259
2260         /* So that it won't aggregate for each init. */
2261         xstats_ctrl->mlx5_stats_n = 0;
2262         ret = mlx5_os_get_stats_n(dev);
2263         if (ret < 0) {
2264                 DRV_LOG(WARNING, "port %u no extended statistics available",
2265                         dev->data->port_id);
2266                 return;
2267         }
2268         dev_stats_n = ret;
2269         /* Allocate memory to grab stat names and values. */
2270         str_sz = dev_stats_n * ETH_GSTRING_LEN;
2271         strings = (struct ethtool_gstrings *)
2272                   rte_malloc("xstats_strings",
2273                              str_sz + sizeof(struct ethtool_gstrings), 0);
2274         if (!strings) {
2275                 DRV_LOG(WARNING, "port %u unable to allocate memory for xstats",
2276                      dev->data->port_id);
2277                 return;
2278         }
2279         strings->cmd = ETHTOOL_GSTRINGS;
2280         strings->string_set = ETH_SS_STATS;
2281         strings->len = dev_stats_n;
2282         ifr.ifr_data = (caddr_t)strings;
2283         ret = mlx5_ifreq(dev, SIOCETHTOOL, &ifr);
2284         if (ret) {
2285                 DRV_LOG(WARNING, "port %u unable to get statistic names",
2286                         dev->data->port_id);
2287                 goto free;
2288         }
2289         for (i = 0; i != dev_stats_n; ++i) {
2290                 const char *curr_string = (const char *)
2291                         &strings->data[i * ETH_GSTRING_LEN];
2292
2293                 for (j = 0; j != xstats_n; ++j) {
2294                         if (!strcmp(mlx5_counters_init[j].ctr_name,
2295                                     curr_string)) {
2296                                 unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2297
2298                                 xstats_ctrl->dev_table_idx[idx] = i;
2299                                 xstats_ctrl->info[idx] = mlx5_counters_init[j];
2300                                 break;
2301                         }
2302                 }
2303         }
2304         /* Add dev counters. */
2305         for (i = 0; i != xstats_n; ++i) {
2306                 if (mlx5_counters_init[i].dev) {
2307                         unsigned int idx = xstats_ctrl->mlx5_stats_n++;
2308
2309                         xstats_ctrl->info[idx] = mlx5_counters_init[i];
2310                         xstats_ctrl->hw_stats[idx] = 0;
2311                 }
2312         }
2313         MLX5_ASSERT(xstats_ctrl->mlx5_stats_n <= MLX5_MAX_XSTATS);
2314         xstats_ctrl->stats_n = dev_stats_n;
2315         /* Copy to base at first time. */
2316         ret = mlx5_os_read_dev_counters(dev, xstats_ctrl->base);
2317         if (ret)
2318                 DRV_LOG(ERR, "port %u cannot read device counters: %s",
2319                         dev->data->port_id, strerror(rte_errno));
2320         mlx5_os_read_dev_stat(priv, "out_of_buffer", &stats_ctrl->imissed_base);
2321         stats_ctrl->imissed = 0;
2322 free:
2323         rte_free(strings);
2324 }
2325
2326 /**
2327  * Set the reg_mr and dereg_mr call backs
2328  *
2329  * @param reg_mr_cb[out]
2330  *   Pointer to reg_mr func
2331  * @param dereg_mr_cb[out]
2332  *   Pointer to dereg_mr func
2333  *
2334  */
2335 void
2336 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2337                       mlx5_dereg_mr_t *dereg_mr_cb)
2338 {
2339         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2340         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2341 }
2342
2343 const struct eth_dev_ops mlx5_os_dev_ops = {
2344         .dev_configure = mlx5_dev_configure,
2345         .dev_start = mlx5_dev_start,
2346         .dev_stop = mlx5_dev_stop,
2347         .dev_set_link_down = mlx5_set_link_down,
2348         .dev_set_link_up = mlx5_set_link_up,
2349         .dev_close = mlx5_dev_close,
2350         .promiscuous_enable = mlx5_promiscuous_enable,
2351         .promiscuous_disable = mlx5_promiscuous_disable,
2352         .allmulticast_enable = mlx5_allmulticast_enable,
2353         .allmulticast_disable = mlx5_allmulticast_disable,
2354         .link_update = mlx5_link_update,
2355         .stats_get = mlx5_stats_get,
2356         .stats_reset = mlx5_stats_reset,
2357         .xstats_get = mlx5_xstats_get,
2358         .xstats_reset = mlx5_xstats_reset,
2359         .xstats_get_names = mlx5_xstats_get_names,
2360         .fw_version_get = mlx5_fw_version_get,
2361         .dev_infos_get = mlx5_dev_infos_get,
2362         .read_clock = mlx5_read_clock,
2363         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2364         .vlan_filter_set = mlx5_vlan_filter_set,
2365         .rx_queue_setup = mlx5_rx_queue_setup,
2366         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2367         .tx_queue_setup = mlx5_tx_queue_setup,
2368         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2369         .rx_queue_release = mlx5_rx_queue_release,
2370         .tx_queue_release = mlx5_tx_queue_release,
2371         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2372         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2373         .mac_addr_remove = mlx5_mac_addr_remove,
2374         .mac_addr_add = mlx5_mac_addr_add,
2375         .mac_addr_set = mlx5_mac_addr_set,
2376         .set_mc_addr_list = mlx5_set_mc_addr_list,
2377         .mtu_set = mlx5_dev_set_mtu,
2378         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2379         .vlan_offload_set = mlx5_vlan_offload_set,
2380         .reta_update = mlx5_dev_rss_reta_update,
2381         .reta_query = mlx5_dev_rss_reta_query,
2382         .rss_hash_update = mlx5_rss_hash_update,
2383         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2384         .filter_ctrl = mlx5_dev_filter_ctrl,
2385         .rx_descriptor_status = mlx5_rx_descriptor_status,
2386         .tx_descriptor_status = mlx5_tx_descriptor_status,
2387         .rxq_info_get = mlx5_rxq_info_get,
2388         .txq_info_get = mlx5_txq_info_get,
2389         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2390         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2391         .rx_queue_count = mlx5_rx_queue_count,
2392         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2393         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2394         .is_removed = mlx5_is_removed,
2395         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2396         .get_module_info = mlx5_get_module_info,
2397         .get_module_eeprom = mlx5_get_module_eeprom,
2398         .hairpin_cap_get = mlx5_hairpin_cap_get,
2399         .mtr_ops_get = mlx5_flow_meter_ops_get,
2400 };
2401
2402 /* Available operations from secondary process. */
2403 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2404         .stats_get = mlx5_stats_get,
2405         .stats_reset = mlx5_stats_reset,
2406         .xstats_get = mlx5_xstats_get,
2407         .xstats_reset = mlx5_xstats_reset,
2408         .xstats_get_names = mlx5_xstats_get_names,
2409         .fw_version_get = mlx5_fw_version_get,
2410         .dev_infos_get = mlx5_dev_infos_get,
2411         .rx_descriptor_status = mlx5_rx_descriptor_status,
2412         .tx_descriptor_status = mlx5_tx_descriptor_status,
2413         .rxq_info_get = mlx5_rxq_info_get,
2414         .txq_info_get = mlx5_txq_info_get,
2415         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2416         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2417         .get_module_info = mlx5_get_module_info,
2418         .get_module_eeprom = mlx5_get_module_eeprom,
2419 };
2420
2421 /* Available operations in flow isolated mode. */
2422 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2423         .dev_configure = mlx5_dev_configure,
2424         .dev_start = mlx5_dev_start,
2425         .dev_stop = mlx5_dev_stop,
2426         .dev_set_link_down = mlx5_set_link_down,
2427         .dev_set_link_up = mlx5_set_link_up,
2428         .dev_close = mlx5_dev_close,
2429         .promiscuous_enable = mlx5_promiscuous_enable,
2430         .promiscuous_disable = mlx5_promiscuous_disable,
2431         .allmulticast_enable = mlx5_allmulticast_enable,
2432         .allmulticast_disable = mlx5_allmulticast_disable,
2433         .link_update = mlx5_link_update,
2434         .stats_get = mlx5_stats_get,
2435         .stats_reset = mlx5_stats_reset,
2436         .xstats_get = mlx5_xstats_get,
2437         .xstats_reset = mlx5_xstats_reset,
2438         .xstats_get_names = mlx5_xstats_get_names,
2439         .fw_version_get = mlx5_fw_version_get,
2440         .dev_infos_get = mlx5_dev_infos_get,
2441         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2442         .vlan_filter_set = mlx5_vlan_filter_set,
2443         .rx_queue_setup = mlx5_rx_queue_setup,
2444         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2445         .tx_queue_setup = mlx5_tx_queue_setup,
2446         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2447         .rx_queue_release = mlx5_rx_queue_release,
2448         .tx_queue_release = mlx5_tx_queue_release,
2449         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2450         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2451         .mac_addr_remove = mlx5_mac_addr_remove,
2452         .mac_addr_add = mlx5_mac_addr_add,
2453         .mac_addr_set = mlx5_mac_addr_set,
2454         .set_mc_addr_list = mlx5_set_mc_addr_list,
2455         .mtu_set = mlx5_dev_set_mtu,
2456         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2457         .vlan_offload_set = mlx5_vlan_offload_set,
2458         .filter_ctrl = mlx5_dev_filter_ctrl,
2459         .rx_descriptor_status = mlx5_rx_descriptor_status,
2460         .tx_descriptor_status = mlx5_tx_descriptor_status,
2461         .rxq_info_get = mlx5_rxq_info_get,
2462         .txq_info_get = mlx5_txq_info_get,
2463         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2464         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2465         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2466         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2467         .is_removed = mlx5_is_removed,
2468         .get_module_info = mlx5_get_module_info,
2469         .get_module_eeprom = mlx5_get_module_eeprom,
2470         .hairpin_cap_get = mlx5_hairpin_cap_get,
2471         .mtr_ops_get = mlx5_flow_meter_ops_get,
2472 };