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 <linux/rtnetlink.h>
  14 #include <linux/sockios.h>
  15 #include <linux/ethtool.h>
  16 #include <fcntl.h>
  17
  18 #include <rte_malloc.h>
  19 #include <rte_ethdev_driver.h>
  20 #include <rte_ethdev_pci.h>
  21 #include <rte_pci.h>
  22 #include <rte_bus_pci.h>
  23 #include <rte_common.h>
  24 #include <rte_kvargs.h>
  25 #include <rte_rwlock.h>
  26 #include <rte_spinlock.h>
  27 #include <rte_string_fns.h>
  28 #include <rte_alarm.h>
  29 #include <rte_eal_paging.h>
  30
  31 #include <mlx5_glue.h>
  32 #include <mlx5_devx_cmds.h>
  33 #include <mlx5_common.h>
  34 #include <mlx5_common_mp.h>
  35 #include <mlx5_common_mr.h>
  36 #include <mlx5_malloc.h>
  37
  38 #include "mlx5_defs.h"
  39 #include "mlx5.h"
  40 #include "mlx5_common_os.h"
  41 #include "mlx5_utils.h"
  42 #include "mlx5_rxtx.h"
  43 #include "mlx5_autoconf.h"
  44 #include "mlx5_mr.h"
  45 #include "mlx5_flow.h"
  46 #include "rte_pmd_mlx5.h"
  47 #include "mlx5_verbs.h"
  48
  49 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  50
  51 #ifndef HAVE_IBV_MLX5_MOD_MPW
  52 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  53 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  54 #endif
  55
  56 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  57 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  58 #endif
  59
  60 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
  61
  62 /* Spinlock for mlx5_shared_data allocation. */
  63 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
  64
  65 /* Process local data for secondary processes. */
  66 static struct mlx5_local_data mlx5_local_data;
  67
  68 /**
  69  * Get mlx5 device attributes. The glue function query_device_ex() is called
  70  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  71  * device attributes from the glue out parameter.
  72  *
  73  * @param dev
  74  *   Pointer to ibv context.
  75  *
  76  * @param device_attr
  77  *   Pointer to mlx5 device attributes.
  78  *
  79  * @return
  80  *   0 on success, non zero error number otherwise
  81  */
  82 int
  83 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
  84 {
  85         int err;
  86         struct ibv_device_attr_ex attr_ex;
  87         memset(device_attr, 0, sizeof(*device_attr));
  88         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
  89         if (err)
  90                 return err;
  91
  92         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
  93         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
  94         device_attr->max_sge = attr_ex.orig_attr.max_sge;
  95         device_attr->max_cq = attr_ex.orig_attr.max_cq;
  96         device_attr->max_qp = attr_ex.orig_attr.max_qp;
  97         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
  98         device_attr->max_rwq_indirection_table_size =
  99                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 100         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 101         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 102
 103         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 104         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 105         if (err)
 106                 return err;
 107
 108         device_attr->flags = dv_attr.flags;
 109         device_attr->comp_mask = dv_attr.comp_mask;
 110 #ifdef HAVE_IBV_MLX5_MOD_SWP
 111         device_attr->sw_parsing_offloads =
 112                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 113 #endif
 114         device_attr->min_single_stride_log_num_of_bytes =
 115                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 116         device_attr->max_single_stride_log_num_of_bytes =
 117                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 118         device_attr->min_single_wqe_log_num_of_strides =
 119                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 120         device_attr->max_single_wqe_log_num_of_strides =
 121                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 122         device_attr->stride_supported_qpts =
 123                 dv_attr.striding_rq_caps.supported_qpts;
 124 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 125         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 126 #endif
 127
 128         return err;
 129 }
 130
 131 /**
 132  * Verbs callback to allocate a memory. This function should allocate the space
 133  * according to the size provided residing inside a huge page.
 134  * Please note that all allocation must respect the alignment from libmlx5
 135  * (i.e. currently rte_mem_page_size()).
 136  *
 137  * @param[in] size
 138  *   The size in bytes of the memory to allocate.
 139  * @param[in] data
 140  *   A pointer to the callback data.
 141  *
 142  * @return
 143  *   Allocated buffer, NULL otherwise and rte_errno is set.
 144  */
 145 static void *
 146 mlx5_alloc_verbs_buf(size_t size, void *data)
 147 {
 148         struct mlx5_priv *priv = data;
 149         void *ret;
 150         unsigned int socket = SOCKET_ID_ANY;
 151         size_t alignment = rte_mem_page_size();
 152         if (alignment == (size_t)-1) {
 153                 DRV_LOG(ERR, "Failed to get mem page size");
 154                 rte_errno = ENOMEM;
 155                 return NULL;
 156         }
 157
 158         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 159                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 160
 161                 socket = ctrl->socket;
 162         } else if (priv->verbs_alloc_ctx.type ==
 163                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 164                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 165
 166                 socket = ctrl->socket;
 167         }
 168         MLX5_ASSERT(data != NULL);
 169         ret = mlx5_malloc(0, size, alignment, socket);
 170         if (!ret && size)
 171                 rte_errno = ENOMEM;
 172         return ret;
 173 }
 174
 175 /**
 176  * Verbs callback to free a memory.
 177  *
 178  * @param[in] ptr
 179  *   A pointer to the memory to free.
 180  * @param[in] data
 181  *   A pointer to the callback data.
 182  */
 183 static void
 184 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 185 {
 186         MLX5_ASSERT(data != NULL);
 187         mlx5_free(ptr);
 188 }
 189
 190 /**
 191  * Initialize DR related data within private structure.
 192  * Routine checks the reference counter and does actual
 193  * resources creation/initialization only if counter is zero.
 194  *
 195  * @param[in] priv
 196  *   Pointer to the private device data structure.
 197  *
 198  * @return
 199  *   Zero on success, positive error code otherwise.
 200  */
 201 static int
 202 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 203 {
 204         struct mlx5_dev_ctx_shared *sh = priv->sh;
 205         char s[MLX5_HLIST_NAMESIZE];
 206         int err = 0;
 207
 208         if (!sh->flow_tbls)
 209                 err = mlx5_alloc_table_hash_list(priv);
 210         else
 211                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 212                         (void *)sh->flow_tbls);
 213         if (err)
 214                 return err;
 215         /* Create tags hash list table. */
 216         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 217         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 218         if (!sh->tag_table) {
 219                 DRV_LOG(ERR, "tags with hash creation failed.");
 220                 err = ENOMEM;
 221                 goto error;
 222         }
 223 #ifdef HAVE_MLX5DV_DR
 224         void *domain;
 225
 226         if (sh->dv_refcnt) {
 227                 /* Shared DV/DR structures is already initialized. */
 228                 sh->dv_refcnt++;
 229                 priv->dr_shared = 1;
 230                 return 0;
 231         }
 232         /* Reference counter is zero, we should initialize structures. */
 233         domain = mlx5_glue->dr_create_domain(sh->ctx,
 234                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 235         if (!domain) {
 236                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 237                 err = errno;
 238                 goto error;
 239         }
 240         sh->rx_domain = domain;
 241         domain = mlx5_glue->dr_create_domain(sh->ctx,
 242                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 243         if (!domain) {
 244                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 245                 err = errno;
 246                 goto error;
 247         }
 248         pthread_mutex_init(&sh->dv_mutex, NULL);
 249         sh->tx_domain = domain;
 250 #ifdef HAVE_MLX5DV_DR_ESWITCH
 251         if (priv->config.dv_esw_en) {
 252                 domain  = mlx5_glue->dr_create_domain
 253                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 254                 if (!domain) {
 255                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 256                         err = errno;
 257                         goto error;
 258                 }
 259                 sh->fdb_domain = domain;
 260                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 261         }
 262 #endif
 263         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 264                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 265                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 266                 if (sh->fdb_domain)
 267                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 268         }
 269         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 270 #endif /* HAVE_MLX5DV_DR */
 271         sh->dv_refcnt++;
 272         priv->dr_shared = 1;
 273         return 0;
 274 error:
 275         /* Rollback the created objects. */
 276         if (sh->rx_domain) {
 277                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 278                 sh->rx_domain = NULL;
 279         }
 280         if (sh->tx_domain) {
 281                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 282                 sh->tx_domain = NULL;
 283         }
 284         if (sh->fdb_domain) {
 285                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 286                 sh->fdb_domain = NULL;
 287         }
 288         if (sh->esw_drop_action) {
 289                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 290                 sh->esw_drop_action = NULL;
 291         }
 292         if (sh->pop_vlan_action) {
 293                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 294                 sh->pop_vlan_action = NULL;
 295         }
 296         if (sh->tag_table) {
 297                 /* tags should be destroyed with flow before. */
 298                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 299                 sh->tag_table = NULL;
 300         }
 301         mlx5_free_table_hash_list(priv);
 302         return err;
 303 }
 304
 305 /**
 306  * Destroy DR related data within private structure.
 307  *
 308  * @param[in] priv
 309  *   Pointer to the private device data structure.
 310  */
 311 void
 312 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 313 {
 314         struct mlx5_dev_ctx_shared *sh;
 315
 316         if (!priv->dr_shared)
 317                 return;
 318         priv->dr_shared = 0;
 319         sh = priv->sh;
 320         MLX5_ASSERT(sh);
 321 #ifdef HAVE_MLX5DV_DR
 322         MLX5_ASSERT(sh->dv_refcnt);
 323         if (sh->dv_refcnt && --sh->dv_refcnt)
 324                 return;
 325         if (sh->rx_domain) {
 326                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 327                 sh->rx_domain = NULL;
 328         }
 329         if (sh->tx_domain) {
 330                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 331                 sh->tx_domain = NULL;
 332         }
 333 #ifdef HAVE_MLX5DV_DR_ESWITCH
 334         if (sh->fdb_domain) {
 335                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 336                 sh->fdb_domain = NULL;
 337         }
 338         if (sh->esw_drop_action) {
 339                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 340                 sh->esw_drop_action = NULL;
 341         }
 342 #endif
 343         if (sh->pop_vlan_action) {
 344                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 345                 sh->pop_vlan_action = NULL;
 346         }
 347         pthread_mutex_destroy(&sh->dv_mutex);
 348 #endif /* HAVE_MLX5DV_DR */
 349         if (sh->tag_table) {
 350                 /* tags should be destroyed with flow before. */
 351                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 352                 sh->tag_table = NULL;
 353         }
 354         mlx5_free_table_hash_list(priv);
 355 }
 356
 357 /**
 358  * Initialize shared data between primary and secondary process.
 359  *
 360  * A memzone is reserved by primary process and secondary processes attach to
 361  * the memzone.
 362  *
 363  * @return
 364  *   0 on success, a negative errno value otherwise and rte_errno is set.
 365  */
 366 static int
 367 mlx5_init_shared_data(void)
 368 {
 369         const struct rte_memzone *mz;
 370         int ret = 0;
 371
 372         rte_spinlock_lock(&mlx5_shared_data_lock);
 373         if (mlx5_shared_data == NULL) {
 374                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 375                         /* Allocate shared memory. */
 376                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 377                                                  sizeof(*mlx5_shared_data),
 378                                                  SOCKET_ID_ANY, 0);
 379                         if (mz == NULL) {
 380                                 DRV_LOG(ERR,
 381                                         "Cannot allocate mlx5 shared data");
 382                                 ret = -rte_errno;
 383                                 goto error;
 384                         }
 385                         mlx5_shared_data = mz->addr;
 386                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 387                         rte_spinlock_init(&mlx5_shared_data->lock);
 388                 } else {
 389                         /* Lookup allocated shared memory. */
 390                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 391                         if (mz == NULL) {
 392                                 DRV_LOG(ERR,
 393                                         "Cannot attach mlx5 shared data");
 394                                 ret = -rte_errno;
 395                                 goto error;
 396                         }
 397                         mlx5_shared_data = mz->addr;
 398                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 399                 }
 400         }
 401 error:
 402         rte_spinlock_unlock(&mlx5_shared_data_lock);
 403         return ret;
 404 }
 405
 406 /**
 407  * PMD global initialization.
 408  *
 409  * Independent from individual device, this function initializes global
 410  * per-PMD data structures distinguishing primary and secondary processes.
 411  * Hence, each initialization is called once per a process.
 412  *
 413  * @return
 414  *   0 on success, a negative errno value otherwise and rte_errno is set.
 415  */
 416 static int
 417 mlx5_init_once(void)
 418 {
 419         struct mlx5_shared_data *sd;
 420         struct mlx5_local_data *ld = &mlx5_local_data;
 421         int ret = 0;
 422
 423         if (mlx5_init_shared_data())
 424                 return -rte_errno;
 425         sd = mlx5_shared_data;
 426         MLX5_ASSERT(sd);
 427         rte_spinlock_lock(&sd->lock);
 428         switch (rte_eal_process_type()) {
 429         case RTE_PROC_PRIMARY:
 430                 if (sd->init_done)
 431                         break;
 432                 LIST_INIT(&sd->mem_event_cb_list);
 433                 rte_rwlock_init(&sd->mem_event_rwlock);
 434                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
 435                                                 mlx5_mr_mem_event_cb, NULL);
 436                 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
 437                                            mlx5_mp_os_primary_handle);
 438                 if (ret)
 439                         goto out;
 440                 sd->init_done = true;
 441                 break;
 442         case RTE_PROC_SECONDARY:
 443                 if (ld->init_done)
 444                         break;
 445                 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
 446                                              mlx5_mp_os_secondary_handle);
 447                 if (ret)
 448                         goto out;
 449                 ++sd->secondary_cnt;
 450                 ld->init_done = true;
 451                 break;
 452         default:
 453                 break;
 454         }
 455 out:
 456         rte_spinlock_unlock(&sd->lock);
 457         return ret;
 458 }
 459
 460 /**
 461  * Spawn an Ethernet device from Verbs information.
 462  *
 463  * @param dpdk_dev
 464  *   Backing DPDK device.
 465  * @param spawn
 466  *   Verbs device parameters (name, port, switch_info) to spawn.
 467  * @param config
 468  *   Device configuration parameters.
 469  *
 470  * @return
 471  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 472  *   is set. The following errors are defined:
 473  *
 474  *   EBUSY: device is not supposed to be spawned.
 475  *   EEXIST: device is already spawned
 476  */
 477 static struct rte_eth_dev *
 478 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 479                struct mlx5_dev_spawn_data *spawn,
 480                struct mlx5_dev_config config)
 481 {
 482         const struct mlx5_switch_info *switch_info = &spawn->info;
 483         struct mlx5_dev_ctx_shared *sh = NULL;
 484         struct ibv_port_attr port_attr;
 485         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 486         struct rte_eth_dev *eth_dev = NULL;
 487         struct mlx5_priv *priv = NULL;
 488         int err = 0;
 489         unsigned int hw_padding = 0;
 490         unsigned int mps;
 491         unsigned int cqe_comp;
 492         unsigned int cqe_pad = 0;
 493         unsigned int tunnel_en = 0;
 494         unsigned int mpls_en = 0;
 495         unsigned int swp = 0;
 496         unsigned int mprq = 0;
 497         unsigned int mprq_min_stride_size_n = 0;
 498         unsigned int mprq_max_stride_size_n = 0;
 499         unsigned int mprq_min_stride_num_n = 0;
 500         unsigned int mprq_max_stride_num_n = 0;
 501         struct rte_ether_addr mac;
 502         char name[RTE_ETH_NAME_MAX_LEN];
 503         int own_domain_id = 0;
 504         uint16_t port_id;
 505         unsigned int i;
 506 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 507         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 508 #endif
 509
 510         /* Determine if this port representor is supposed to be spawned. */
 511         if (switch_info->representor && dpdk_dev->devargs) {
 512                 struct rte_eth_devargs eth_da;
 513
 514                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 515                 if (err) {
 516                         rte_errno = -err;
 517                         DRV_LOG(ERR, "failed to process device arguments: %s",
 518                                 strerror(rte_errno));
 519                         return NULL;
 520                 }
 521                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 522                         if (eth_da.representor_ports[i] ==
 523                             (uint16_t)switch_info->port_name)
 524                                 break;
 525                 if (i == eth_da.nb_representor_ports) {
 526                         rte_errno = EBUSY;
 527                         return NULL;
 528                 }
 529         }
 530         /* Build device name. */
 531         if (spawn->pf_bond <  0) {
 532                 /* Single device. */
 533                 if (!switch_info->representor)
 534                         strlcpy(name, dpdk_dev->name, sizeof(name));
 535                 else
 536                         snprintf(name, sizeof(name), "%s_representor_%u",
 537                                  dpdk_dev->name, switch_info->port_name);
 538         } else {
 539                 /* Bonding device. */
 540                 if (!switch_info->representor)
 541                         snprintf(name, sizeof(name), "%s_%s",
 542                                  dpdk_dev->name,
 543                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 544                 else
 545                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 546                                  dpdk_dev->name,
 547                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 548                                  switch_info->port_name);
 549         }
 550         /* check if the device is already spawned */
 551         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 552                 rte_errno = EEXIST;
 553                 return NULL;
 554         }
 555         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 556         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 557                 struct mlx5_mp_id mp_id;
 558
 559                 eth_dev = rte_eth_dev_attach_secondary(name);
 560                 if (eth_dev == NULL) {
 561                         DRV_LOG(ERR, "can not attach rte ethdev");
 562                         rte_errno = ENOMEM;
 563                         return NULL;
 564                 }
 565                 eth_dev->device = dpdk_dev;
 566                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 567                 err = mlx5_proc_priv_init(eth_dev);
 568                 if (err)
 569                         return NULL;
 570                 mp_id.port_id = eth_dev->data->port_id;
 571                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 572                 /* Receive command fd from primary process */
 573                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 574                 if (err < 0)
 575                         goto err_secondary;
 576                 /* Remap UAR for Tx queues. */
 577                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 578                 if (err)
 579                         goto err_secondary;
 580                 /*
 581                  * Ethdev pointer is still required as input since
 582                  * the primary device is not accessible from the
 583                  * secondary process.
 584                  */
 585                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 586                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 587                 return eth_dev;
 588 err_secondary:
 589                 mlx5_dev_close(eth_dev);
 590                 return NULL;
 591         }
 592         /*
 593          * Some parameters ("tx_db_nc" in particularly) are needed in
 594          * advance to create dv/verbs device context. We proceed the
 595          * devargs here to get ones, and later proceed devargs again
 596          * to override some hardware settings.
 597          */
 598         err = mlx5_args(&config, dpdk_dev->devargs);
 599         if (err) {
 600                 err = rte_errno;
 601                 DRV_LOG(ERR, "failed to process device arguments: %s",
 602                         strerror(rte_errno));
 603                 goto error;
 604         }
 605         mlx5_malloc_mem_select(config.sys_mem_en);
 606         sh = mlx5_alloc_shared_dev_ctx(spawn, &config);
 607         if (!sh)
 608                 return NULL;
 609         config.devx = sh->devx;
 610 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 611         config.dest_tir = 1;
 612 #endif
 613 #ifdef HAVE_IBV_MLX5_MOD_SWP
 614         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 615 #endif
 616         /*
 617          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 618          * as all ConnectX-5 devices.
 619          */
 620 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 621         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 622 #endif
 623 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 624         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 625 #endif
 626         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 627         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 628                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 629                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 630                         mps = MLX5_MPW_ENHANCED;
 631                 } else {
 632                         DRV_LOG(DEBUG, "MPW is supported");
 633                         mps = MLX5_MPW;
 634                 }
 635         } else {
 636                 DRV_LOG(DEBUG, "MPW isn't supported");
 637                 mps = MLX5_MPW_DISABLED;
 638         }
 639 #ifdef HAVE_IBV_MLX5_MOD_SWP
 640         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 641                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 642         DRV_LOG(DEBUG, "SWP support: %u", swp);
 643 #endif
 644         config.swp = !!swp;
 645 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 646         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 647                 struct mlx5dv_striding_rq_caps mprq_caps =
 648                         dv_attr.striding_rq_caps;
 649
 650                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 651                         mprq_caps.min_single_stride_log_num_of_bytes);
 652                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 653                         mprq_caps.max_single_stride_log_num_of_bytes);
 654                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 655                         mprq_caps.min_single_wqe_log_num_of_strides);
 656                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 657                         mprq_caps.max_single_wqe_log_num_of_strides);
 658                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 659                         mprq_caps.supported_qpts);
 660                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 661                 mprq = 1;
 662                 mprq_min_stride_size_n =
 663                         mprq_caps.min_single_stride_log_num_of_bytes;
 664                 mprq_max_stride_size_n =
 665                         mprq_caps.max_single_stride_log_num_of_bytes;
 666                 mprq_min_stride_num_n =
 667                         mprq_caps.min_single_wqe_log_num_of_strides;
 668                 mprq_max_stride_num_n =
 669                         mprq_caps.max_single_wqe_log_num_of_strides;
 670         }
 671 #endif
 672         if (RTE_CACHE_LINE_SIZE == 128 &&
 673             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 674                 cqe_comp = 0;
 675         else
 676                 cqe_comp = 1;
 677         config.cqe_comp = cqe_comp;
 678 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 679         /* Whether device supports 128B Rx CQE padding. */
 680         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 681                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 682 #endif
 683 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 684         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 685                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 686                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 687                              (dv_attr.tunnel_offloads_caps &
 688                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 689                              (dv_attr.tunnel_offloads_caps &
 690                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 691         }
 692         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 693                 tunnel_en ? "" : "not ");
 694 #else
 695         DRV_LOG(WARNING,
 696                 "tunnel offloading disabled due to old OFED/rdma-core version");
 697 #endif
 698         config.tunnel_en = tunnel_en;
 699 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 700         mpls_en = ((dv_attr.tunnel_offloads_caps &
 701                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 702                    (dv_attr.tunnel_offloads_caps &
 703                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 704         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 705                 mpls_en ? "" : "not ");
 706 #else
 707         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 708                 " old OFED/rdma-core version or firmware configuration");
 709 #endif
 710         config.mpls_en = mpls_en;
 711         /* Check port status. */
 712         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 713         if (err) {
 714                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 715                 goto error;
 716         }
 717         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 718                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 719                 err = EINVAL;
 720                 goto error;
 721         }
 722         if (port_attr.state != IBV_PORT_ACTIVE)
 723                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 724                         mlx5_glue->port_state_str(port_attr.state),
 725                         port_attr.state);
 726         /* Allocate private eth device data. */
 727         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
 728                            sizeof(*priv),
 729                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 730         if (priv == NULL) {
 731                 DRV_LOG(ERR, "priv allocation failure");
 732                 err = ENOMEM;
 733                 goto error;
 734         }
 735         priv->sh = sh;
 736         priv->dev_port = spawn->phys_port;
 737         priv->pci_dev = spawn->pci_dev;
 738         priv->mtu = RTE_ETHER_MTU;
 739         priv->mp_id.port_id = port_id;
 740         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 741         /* Some internal functions rely on Netlink sockets, open them now. */
 742         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 743         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 744         priv->representor = !!switch_info->representor;
 745         priv->master = !!switch_info->master;
 746         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 747         priv->vport_meta_tag = 0;
 748         priv->vport_meta_mask = 0;
 749         priv->pf_bond = spawn->pf_bond;
 750 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 751         /*
 752          * The DevX port query API is implemented. E-Switch may use
 753          * either vport or reg_c[0] metadata register to match on
 754          * vport index. The engaged part of metadata register is
 755          * defined by mask.
 756          */
 757         if (switch_info->representor || switch_info->master) {
 758                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 759                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 760                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 761                                                  &devx_port);
 762                 if (err) {
 763                         DRV_LOG(WARNING,
 764                                 "can't query devx port %d on device %s",
 765                                 spawn->phys_port,
 766                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 767                         devx_port.comp_mask = 0;
 768                 }
 769         }
 770         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 771                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 772                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 773                 if (!priv->vport_meta_mask) {
 774                         DRV_LOG(ERR, "vport zero mask for port %d"
 775                                      " on bonding device %s",
 776                                      spawn->phys_port,
 777                                      mlx5_os_get_dev_device_name
 778                                                         (spawn->phys_dev));
 779                         err = ENOTSUP;
 780                         goto error;
 781                 }
 782                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 783                         DRV_LOG(ERR, "invalid vport tag for port %d"
 784                                      " on bonding device %s",
 785                                      spawn->phys_port,
 786                                      mlx5_os_get_dev_device_name
 787                                                         (spawn->phys_dev));
 788                         err = ENOTSUP;
 789                         goto error;
 790                 }
 791         }
 792         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 793                 priv->vport_id = devx_port.vport_num;
 794         } else if (spawn->pf_bond >= 0) {
 795                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 796                              " on bonding device %s",
 797                              spawn->phys_port,
 798                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 799                 err = ENOTSUP;
 800                 goto error;
 801         } else {
 802                 /* Suppose vport index in compatible way. */
 803                 priv->vport_id = switch_info->representor ?
 804                                  switch_info->port_name + 1 : -1;
 805         }
 806 #else
 807         /*
 808          * Kernel/rdma_core support single E-Switch per PF configurations
 809          * only and vport_id field contains the vport index for
 810          * associated VF, which is deduced from representor port name.
 811          * For example, let's have the IB device port 10, it has
 812          * attached network device eth0, which has port name attribute
 813          * pf0vf2, we can deduce the VF number as 2, and set vport index
 814          * as 3 (2+1). This assigning schema should be changed if the
 815          * multiple E-Switch instances per PF configurations or/and PCI
 816          * subfunctions are added.
 817          */
 818         priv->vport_id = switch_info->representor ?
 819                          switch_info->port_name + 1 : -1;
 820 #endif
 821         /* representor_id field keeps the unmodified VF index. */
 822         priv->representor_id = switch_info->representor ?
 823                                switch_info->port_name : -1;
 824         /*
 825          * Look for sibling devices in order to reuse their switch domain
 826          * if any, otherwise allocate one.
 827          */
 828         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 829                 const struct mlx5_priv *opriv =
 830                         rte_eth_devices[port_id].data->dev_private;
 831
 832                 if (!opriv ||
 833                     opriv->sh != priv->sh ||
 834                         opriv->domain_id ==
 835                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 836                         continue;
 837                 priv->domain_id = opriv->domain_id;
 838                 break;
 839         }
 840         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 841                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 842                 if (err) {
 843                         err = rte_errno;
 844                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 845                                 strerror(rte_errno));
 846                         goto error;
 847                 }
 848                 own_domain_id = 1;
 849         }
 850         /* Override some values set by hardware configuration. */
 851         mlx5_args(&config, dpdk_dev->devargs);
 852         err = mlx5_dev_check_sibling_config(priv, &config);
 853         if (err)
 854                 goto error;
 855         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 856                             IBV_DEVICE_RAW_IP_CSUM);
 857         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 858                 (config.hw_csum ? "" : "not "));
 859 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 860         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 861         DRV_LOG(DEBUG, "counters are not supported");
 862 #endif
 863 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 864         if (config.dv_flow_en) {
 865                 DRV_LOG(WARNING, "DV flow is not supported");
 866                 config.dv_flow_en = 0;
 867         }
 868 #endif
 869         config.ind_table_max_size =
 870                 sh->device_attr.max_rwq_indirection_table_size;
 871         /*
 872          * Remove this check once DPDK supports larger/variable
 873          * indirection tables.
 874          */
 875         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 876                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 877         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 878                 config.ind_table_max_size);
 879         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 880                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 881         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 882                 (config.hw_vlan_strip ? "" : "not "));
 883         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 884                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 885 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 886         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 887 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 888         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 889                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 890 #endif
 891         if (config.hw_padding && !hw_padding) {
 892                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 893                 config.hw_padding = 0;
 894         } else if (config.hw_padding) {
 895                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 896         }
 897         config.tso = (sh->device_attr.max_tso > 0 &&
 898                       (sh->device_attr.tso_supported_qpts &
 899                        (1 << IBV_QPT_RAW_PACKET)));
 900         if (config.tso)
 901                 config.tso_max_payload_sz = sh->device_attr.max_tso;
 902         /*
 903          * MPW is disabled by default, while the Enhanced MPW is enabled
 904          * by default.
 905          */
 906         if (config.mps == MLX5_ARG_UNSET)
 907                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 908                                                           MLX5_MPW_DISABLED;
 909         else
 910                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
 911         DRV_LOG(INFO, "%sMPS is %s",
 912                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " :
 913                 config.mps == MLX5_MPW ? "legacy " : "",
 914                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 915         if (config.cqe_comp && !cqe_comp) {
 916                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 917                 config.cqe_comp = 0;
 918         }
 919         if (config.cqe_pad && !cqe_pad) {
 920                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 921                 config.cqe_pad = 0;
 922         } else if (config.cqe_pad) {
 923                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 924         }
 925         if (config.devx) {
 926                 priv->counter_fallback = 0;
 927                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
 928                 if (err) {
 929                         err = -err;
 930                         goto error;
 931                 }
 932                 if (!config.hca_attr.flow_counters_dump)
 933                         priv->counter_fallback = 1;
 934 #ifndef HAVE_IBV_DEVX_ASYNC
 935                 priv->counter_fallback = 1;
 936 #endif
 937                 if (priv->counter_fallback)
 938                         DRV_LOG(INFO, "Use fall-back DV counter management");
 939                 /* Check for LRO support. */
 940                 if (config.dest_tir && config.hca_attr.lro_cap &&
 941                     config.dv_flow_en) {
 942                         /* TBD check tunnel lro caps. */
 943                         config.lro.supported = config.hca_attr.lro_cap;
 944                         DRV_LOG(DEBUG, "Device supports LRO");
 945                         /*
 946                          * If LRO timeout is not configured by application,
 947                          * use the minimal supported value.
 948                          */
 949                         if (!config.lro.timeout)
 950                                 config.lro.timeout =
 951                                 config.hca_attr.lro_timer_supported_periods[0];
 952                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 953                                 config.lro.timeout);
 954                 }
 955 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 956                 if (config.hca_attr.qos.sup && config.hca_attr.qos.srtcm_sup &&
 957                     config.dv_flow_en) {
 958                         uint8_t reg_c_mask =
 959                                 config.hca_attr.qos.flow_meter_reg_c_ids;
 960                         /*
 961                          * Meter needs two REG_C's for color match and pre-sfx
 962                          * flow match. Here get the REG_C for color match.
 963                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
 964                          */
 965                         reg_c_mask &= 0xfc;
 966                         if (__builtin_popcount(reg_c_mask) < 1) {
 967                                 priv->mtr_en = 0;
 968                                 DRV_LOG(WARNING, "No available register for"
 969                                         " meter.");
 970                         } else {
 971                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
 972                                                       REG_C_0;
 973                                 priv->mtr_en = 1;
 974                                 priv->mtr_reg_share =
 975                                       config.hca_attr.qos.flow_meter_reg_share;
 976                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
 977                                         priv->mtr_color_reg);
 978                         }
 979                 }
 980 #endif
 981         }
 982         if (config.tx_pp) {
 983                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
 984                         config.hca_attr.dev_freq_khz);
 985                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
 986                         config.hca_attr.qos.packet_pacing ? "" : "not ");
 987                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
 988                         config.hca_attr.cross_channel ? "" : "not ");
 989                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
 990                         config.hca_attr.wqe_index_ignore ? "" : "not ");
 991                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
 992                         config.hca_attr.non_wire_sq ? "" : "not ");
 993                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
 994                         config.hca_attr.log_max_static_sq_wq ? "" : "not ",
 995                         config.hca_attr.log_max_static_sq_wq);
 996                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
 997                         config.hca_attr.qos.wqe_rate_pp ? "" : "not ");
 998                 if (!config.devx) {
 999                         DRV_LOG(ERR, "DevX is required for packet pacing");
1000                         err = ENODEV;
1001                         goto error;
1002                 }
1003                 if (!config.hca_attr.qos.packet_pacing) {
1004                         DRV_LOG(ERR, "Packet pacing is not supported");
1005                         err = ENODEV;
1006                         goto error;
1007                 }
1008                 if (!config.hca_attr.cross_channel) {
1009                         DRV_LOG(ERR, "Cross channel operations are"
1010                                      " required for packet pacing");
1011                         err = ENODEV;
1012                         goto error;
1013                 }
1014                 if (!config.hca_attr.wqe_index_ignore) {
1015                         DRV_LOG(ERR, "WQE index ignore feature is"
1016                                      " required for packet pacing");
1017                         err = ENODEV;
1018                         goto error;
1019                 }
1020                 if (!config.hca_attr.non_wire_sq) {
1021                         DRV_LOG(ERR, "Non-wire SQ feature is"
1022                                      " required for packet pacing");
1023                         err = ENODEV;
1024                         goto error;
1025                 }
1026                 if (!config.hca_attr.log_max_static_sq_wq) {
1027                         DRV_LOG(ERR, "Static WQE SQ feature is"
1028                                      " required for packet pacing");
1029                         err = ENODEV;
1030                         goto error;
1031                 }
1032                 if (!config.hca_attr.qos.wqe_rate_pp) {
1033                         DRV_LOG(ERR, "WQE rate mode is required"
1034                                      " for packet pacing");
1035                         err = ENODEV;
1036                         goto error;
1037                 }
1038 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1039                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1040                              " can't create queues for packet pacing");
1041                 err = ENODEV;
1042                 goto error;
1043 #endif
1044         }
1045         if (config.devx) {
1046                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1047
1048                 err = mlx5_devx_cmd_register_read
1049                         (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1050                         reg, MLX5_ST_SZ_DW(register_mtutc));
1051                 if (!err) {
1052                         uint32_t ts_mode;
1053
1054                         /* MTUTC register is read successfully. */
1055                         ts_mode = MLX5_GET(register_mtutc, reg,
1056                                            time_stamp_mode);
1057                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1058                                 config.rt_timestamp = 1;
1059                 } else {
1060                         /* Kernel does not support register reading. */
1061                         if (config.hca_attr.dev_freq_khz ==
1062                                                  (NS_PER_S / MS_PER_S))
1063                                 config.rt_timestamp = 1;
1064                 }
1065         }
1066         /*
1067          * If HW has bug working with tunnel packet decapsulation and
1068          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1069          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1070          */
1071         if (config.hca_attr.scatter_fcs_w_decap_disable && config.decap_en)
1072                 config.hw_fcs_strip = 0;
1073         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1074                 (config.hw_fcs_strip ? "" : "not "));
1075         if (config.mprq.enabled && mprq) {
1076                 if (config.mprq.stride_num_n &&
1077                     (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1078                      config.mprq.stride_num_n < mprq_min_stride_num_n)) {
1079                         config.mprq.stride_num_n =
1080                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1081                                                 mprq_min_stride_num_n),
1082                                         mprq_max_stride_num_n);
1083                         DRV_LOG(WARNING,
1084                                 "the number of strides"
1085                                 " for Multi-Packet RQ is out of range,"
1086                                 " setting default value (%u)",
1087                                 1 << config.mprq.stride_num_n);
1088                 }
1089                 if (config.mprq.stride_size_n &&
1090                     (config.mprq.stride_size_n > mprq_max_stride_size_n ||
1091                      config.mprq.stride_size_n < mprq_min_stride_size_n)) {
1092                         config.mprq.stride_size_n =
1093                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1094                                                 mprq_min_stride_size_n),
1095                                         mprq_max_stride_size_n);
1096                         DRV_LOG(WARNING,
1097                                 "the size of a stride"
1098                                 " for Multi-Packet RQ is out of range,"
1099                                 " setting default value (%u)",
1100                                 1 << config.mprq.stride_size_n);
1101                 }
1102                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1103                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1104         } else if (config.mprq.enabled && !mprq) {
1105                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1106                 config.mprq.enabled = 0;
1107         }
1108         if (config.max_dump_files_num == 0)
1109                 config.max_dump_files_num = 128;
1110         eth_dev = rte_eth_dev_allocate(name);
1111         if (eth_dev == NULL) {
1112                 DRV_LOG(ERR, "can not allocate rte ethdev");
1113                 err = ENOMEM;
1114                 goto error;
1115         }
1116         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1117         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1118         if (priv->representor) {
1119                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1120                 eth_dev->data->representor_id = priv->representor_id;
1121         }
1122         /*
1123          * Store associated network device interface index. This index
1124          * is permanent throughout the lifetime of device. So, we may store
1125          * the ifindex here and use the cached value further.
1126          */
1127         MLX5_ASSERT(spawn->ifindex);
1128         priv->if_index = spawn->ifindex;
1129         eth_dev->data->dev_private = priv;
1130         priv->dev_data = eth_dev->data;
1131         eth_dev->data->mac_addrs = priv->mac;
1132         eth_dev->device = dpdk_dev;
1133         /* Configure the first MAC address by default. */
1134         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1135                 DRV_LOG(ERR,
1136                         "port %u cannot get MAC address, is mlx5_en"
1137                         " loaded? (errno: %s)",
1138                         eth_dev->data->port_id, strerror(rte_errno));
1139                 err = ENODEV;
1140                 goto error;
1141         }
1142         DRV_LOG(INFO,
1143                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1144                 eth_dev->data->port_id,
1145                 mac.addr_bytes[0], mac.addr_bytes[1],
1146                 mac.addr_bytes[2], mac.addr_bytes[3],
1147                 mac.addr_bytes[4], mac.addr_bytes[5]);
1148 #ifdef RTE_LIBRTE_MLX5_DEBUG
1149         {
1150                 char ifname[IF_NAMESIZE];
1151
1152                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1153                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1154                                 eth_dev->data->port_id, ifname);
1155                 else
1156                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1157                                 eth_dev->data->port_id);
1158         }
1159 #endif
1160         /* Get actual MTU if possible. */
1161         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1162         if (err) {
1163                 err = rte_errno;
1164                 goto error;
1165         }
1166         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1167                 priv->mtu);
1168         /* Initialize burst functions to prevent crashes before link-up. */
1169         eth_dev->rx_pkt_burst = removed_rx_burst;
1170         eth_dev->tx_pkt_burst = removed_tx_burst;
1171         eth_dev->dev_ops = &mlx5_os_dev_ops;
1172         /* Register MAC address. */
1173         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1174         if (config.vf && config.vf_nl_en)
1175                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1176                                       mlx5_ifindex(eth_dev),
1177                                       eth_dev->data->mac_addrs,
1178                                       MLX5_MAX_MAC_ADDRESSES);
1179         priv->flows = 0;
1180         priv->ctrl_flows = 0;
1181         TAILQ_INIT(&priv->flow_meters);
1182         TAILQ_INIT(&priv->flow_meter_profiles);
1183         /* Hint libmlx5 to use PMD allocator for data plane resources */
1184         mlx5_glue->dv_set_context_attr(sh->ctx,
1185                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1186                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1187                                 .alloc = &mlx5_alloc_verbs_buf,
1188                                 .free = &mlx5_free_verbs_buf,
1189                                 .data = priv,
1190                         }));
1191         /* Bring Ethernet device up. */
1192         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1193                 eth_dev->data->port_id);
1194         mlx5_set_link_up(eth_dev);
1195         /*
1196          * Even though the interrupt handler is not installed yet,
1197          * interrupts will still trigger on the async_fd from
1198          * Verbs context returned by ibv_open_device().
1199          */
1200         mlx5_link_update(eth_dev, 0);
1201 #ifdef HAVE_MLX5DV_DR_ESWITCH
1202         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
1203               (switch_info->representor || switch_info->master)))
1204                 config.dv_esw_en = 0;
1205 #else
1206         config.dv_esw_en = 0;
1207 #endif
1208         /* Detect minimal data bytes to inline. */
1209         mlx5_set_min_inline(spawn, &config);
1210         /* Store device configuration on private structure. */
1211         priv->config = config;
1212         /* Create context for virtual machine VLAN workaround. */
1213         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1214         if (config.dv_flow_en) {
1215                 err = mlx5_alloc_shared_dr(priv);
1216                 if (err)
1217                         goto error;
1218                 /*
1219                  * RSS id is shared with meter flow id. Meter flow id can only
1220                  * use the 24 MSB of the register.
1221                  */
1222                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1223                                      MLX5_MTR_COLOR_BITS);
1224                 if (!priv->qrss_id_pool) {
1225                         DRV_LOG(ERR, "can't create flow id pool");
1226                         err = ENOMEM;
1227                         goto error;
1228                 }
1229         }
1230         /* Supported Verbs flow priority number detection. */
1231         err = mlx5_flow_discover_priorities(eth_dev);
1232         if (err < 0) {
1233                 err = -err;
1234                 goto error;
1235         }
1236         priv->config.flow_prio = err;
1237         if (!priv->config.dv_esw_en &&
1238             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1239                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1240                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1241                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1242         }
1243         mlx5_set_metadata_mask(eth_dev);
1244         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1245             !priv->sh->dv_regc0_mask) {
1246                 DRV_LOG(ERR, "metadata mode %u is not supported "
1247                              "(no metadata reg_c[0] is available)",
1248                              priv->config.dv_xmeta_en);
1249                         err = ENOTSUP;
1250                         goto error;
1251         }
1252         /*
1253          * Allocate the buffer for flow creating, just once.
1254          * The allocation must be done before any flow creating.
1255          */
1256         mlx5_flow_alloc_intermediate(eth_dev);
1257         /* Query availability of metadata reg_c's. */
1258         err = mlx5_flow_discover_mreg_c(eth_dev);
1259         if (err < 0) {
1260                 err = -err;
1261                 goto error;
1262         }
1263         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1264                 DRV_LOG(DEBUG,
1265                         "port %u extensive metadata register is not supported",
1266                         eth_dev->data->port_id);
1267                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1268                         DRV_LOG(ERR, "metadata mode %u is not supported "
1269                                      "(no metadata registers available)",
1270                                      priv->config.dv_xmeta_en);
1271                         err = ENOTSUP;
1272                         goto error;
1273                 }
1274         }
1275         if (priv->config.dv_flow_en &&
1276             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1277             mlx5_flow_ext_mreg_supported(eth_dev) &&
1278             priv->sh->dv_regc0_mask) {
1279                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1280                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1281                 if (!priv->mreg_cp_tbl) {
1282                         err = ENOMEM;
1283                         goto error;
1284                 }
1285         }
1286         return eth_dev;
1287 error:
1288         if (priv) {
1289                 if (priv->mreg_cp_tbl)
1290                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1291                 if (priv->sh)
1292                         mlx5_os_free_shared_dr(priv);
1293                 if (priv->nl_socket_route >= 0)
1294                         close(priv->nl_socket_route);
1295                 if (priv->nl_socket_rdma >= 0)
1296                         close(priv->nl_socket_rdma);
1297                 if (priv->vmwa_context)
1298                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1299                 if (priv->qrss_id_pool)
1300                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1301                 if (own_domain_id)
1302                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1303                 mlx5_free(priv);
1304                 if (eth_dev != NULL)
1305                         eth_dev->data->dev_private = NULL;
1306         }
1307         if (eth_dev != NULL) {
1308                 /* mac_addrs must not be freed alone because part of
1309                  * dev_private
1310                  **/
1311                 eth_dev->data->mac_addrs = NULL;
1312                 rte_eth_dev_release_port(eth_dev);
1313         }
1314         if (sh)
1315                 mlx5_free_shared_dev_ctx(sh);
1316         MLX5_ASSERT(err > 0);
1317         rte_errno = err;
1318         return NULL;
1319 }
1320
1321 /**
1322  * Comparison callback to sort device data.
1323  *
1324  * This is meant to be used with qsort().
1325  *
1326  * @param a[in]
1327  *   Pointer to pointer to first data object.
1328  * @param b[in]
1329  *   Pointer to pointer to second data object.
1330  *
1331  * @return
1332  *   0 if both objects are equal, less than 0 if the first argument is less
1333  *   than the second, greater than 0 otherwise.
1334  */
1335 static int
1336 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1337 {
1338         const struct mlx5_switch_info *si_a =
1339                 &((const struct mlx5_dev_spawn_data *)a)->info;
1340         const struct mlx5_switch_info *si_b =
1341                 &((const struct mlx5_dev_spawn_data *)b)->info;
1342         int ret;
1343
1344         /* Master device first. */
1345         ret = si_b->master - si_a->master;
1346         if (ret)
1347                 return ret;
1348         /* Then representor devices. */
1349         ret = si_b->representor - si_a->representor;
1350         if (ret)
1351                 return ret;
1352         /* Unidentified devices come last in no specific order. */
1353         if (!si_a->representor)
1354                 return 0;
1355         /* Order representors by name. */
1356         return si_a->port_name - si_b->port_name;
1357 }
1358
1359 /**
1360  * Match PCI information for possible slaves of bonding device.
1361  *
1362  * @param[in] ibv_dev
1363  *   Pointer to Infiniband device structure.
1364  * @param[in] pci_dev
1365  *   Pointer to PCI device structure to match PCI address.
1366  * @param[in] nl_rdma
1367  *   Netlink RDMA group socket handle.
1368  *
1369  * @return
1370  *   negative value if no bonding device found, otherwise
1371  *   positive index of slave PF in bonding.
1372  */
1373 static int
1374 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1375                            const struct rte_pci_device *pci_dev,
1376                            int nl_rdma)
1377 {
1378         char ifname[IF_NAMESIZE + 1];
1379         unsigned int ifindex;
1380         unsigned int np, i;
1381         FILE *file = NULL;
1382         int pf = -1;
1383
1384         /*
1385          * Try to get master device name. If something goes
1386          * wrong suppose the lack of kernel support and no
1387          * bonding devices.
1388          */
1389         if (nl_rdma < 0)
1390                 return -1;
1391         if (!strstr(ibv_dev->name, "bond"))
1392                 return -1;
1393         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1394         if (!np)
1395                 return -1;
1396         /*
1397          * The Master device might not be on the predefined
1398          * port (not on port index 1, it is not garanted),
1399          * we have to scan all Infiniband device port and
1400          * find master.
1401          */
1402         for (i = 1; i <= np; ++i) {
1403                 /* Check whether Infiniband port is populated. */
1404                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1405                 if (!ifindex)
1406                         continue;
1407                 if (!if_indextoname(ifindex, ifname))
1408                         continue;
1409                 /* Try to read bonding slave names from sysfs. */
1410                 MKSTR(slaves,
1411                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1412                 file = fopen(slaves, "r");
1413                 if (file)
1414                         break;
1415         }
1416         if (!file)
1417                 return -1;
1418         /* Use safe format to check maximal buffer length. */
1419         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1420         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1421                 char tmp_str[IF_NAMESIZE + 32];
1422                 struct rte_pci_addr pci_addr;
1423                 struct mlx5_switch_info info;
1424
1425                 /* Process slave interface names in the loop. */
1426                 snprintf(tmp_str, sizeof(tmp_str),
1427                          "/sys/class/net/%s", ifname);
1428                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1429                         DRV_LOG(WARNING, "can not get PCI address"
1430                                          " for netdev \"%s\"", ifname);
1431                         continue;
1432                 }
1433                 if (pci_dev->addr.domain != pci_addr.domain ||
1434                     pci_dev->addr.bus != pci_addr.bus ||
1435                     pci_dev->addr.devid != pci_addr.devid ||
1436                     pci_dev->addr.function != pci_addr.function)
1437                         continue;
1438                 /* Slave interface PCI address match found. */
1439                 fclose(file);
1440                 snprintf(tmp_str, sizeof(tmp_str),
1441                          "/sys/class/net/%s/phys_port_name", ifname);
1442                 file = fopen(tmp_str, "rb");
1443                 if (!file)
1444                         break;
1445                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1446                 if (fscanf(file, "%32s", tmp_str) == 1)
1447                         mlx5_translate_port_name(tmp_str, &info);
1448                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1449                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1450                         pf = info.port_name;
1451                 break;
1452         }
1453         if (file)
1454                 fclose(file);
1455         return pf;
1456 }
1457
1458 /**
1459  * DPDK callback to register a PCI device.
1460  *
1461  * This function spawns Ethernet devices out of a given PCI device.
1462  *
1463  * @param[in] pci_drv
1464  *   PCI driver structure (mlx5_driver).
1465  * @param[in] pci_dev
1466  *   PCI device information.
1467  *
1468  * @return
1469  *   0 on success, a negative errno value otherwise and rte_errno is set.
1470  */
1471 int
1472 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1473                   struct rte_pci_device *pci_dev)
1474 {
1475         struct ibv_device **ibv_list;
1476         /*
1477          * Number of found IB Devices matching with requested PCI BDF.
1478          * nd != 1 means there are multiple IB devices over the same
1479          * PCI device and we have representors and master.
1480          */
1481         unsigned int nd = 0;
1482         /*
1483          * Number of found IB device Ports. nd = 1 and np = 1..n means
1484          * we have the single multiport IB device, and there may be
1485          * representors attached to some of found ports.
1486          */
1487         unsigned int np = 0;
1488         /*
1489          * Number of DPDK ethernet devices to Spawn - either over
1490          * multiple IB devices or multiple ports of single IB device.
1491          * Actually this is the number of iterations to spawn.
1492          */
1493         unsigned int ns = 0;
1494         /*
1495          * Bonding device
1496          *   < 0 - no bonding device (single one)
1497          *  >= 0 - bonding device (value is slave PF index)
1498          */
1499         int bd = -1;
1500         struct mlx5_dev_spawn_data *list = NULL;
1501         struct mlx5_dev_config dev_config;
1502         int ret;
1503
1504         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_NET) {
1505                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
1506                         " driver.");
1507                 return 1;
1508         }
1509         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1510                 mlx5_pmd_socket_init();
1511         ret = mlx5_init_once();
1512         if (ret) {
1513                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1514                         strerror(rte_errno));
1515                 return -rte_errno;
1516         }
1517         MLX5_ASSERT(pci_drv == &mlx5_driver);
1518         errno = 0;
1519         ibv_list = mlx5_glue->get_device_list(&ret);
1520         if (!ibv_list) {
1521                 rte_errno = errno ? errno : ENOSYS;
1522                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1523                 return -rte_errno;
1524         }
1525         /*
1526          * First scan the list of all Infiniband devices to find
1527          * matching ones, gathering into the list.
1528          */
1529         struct ibv_device *ibv_match[ret + 1];
1530         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1531         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1532         unsigned int i;
1533
1534         while (ret-- > 0) {
1535                 struct rte_pci_addr pci_addr;
1536
1537                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1538                 bd = mlx5_device_bond_pci_match
1539                                 (ibv_list[ret], pci_dev, nl_rdma);
1540                 if (bd >= 0) {
1541                         /*
1542                          * Bonding device detected. Only one match is allowed,
1543                          * the bonding is supported over multi-port IB device,
1544                          * there should be no matches on representor PCI
1545                          * functions or non VF LAG bonding devices with
1546                          * specified address.
1547                          */
1548                         if (nd) {
1549                                 DRV_LOG(ERR,
1550                                         "multiple PCI match on bonding device"
1551                                         "\"%s\" found", ibv_list[ret]->name);
1552                                 rte_errno = ENOENT;
1553                                 ret = -rte_errno;
1554                                 goto exit;
1555                         }
1556                         DRV_LOG(INFO, "PCI information matches for"
1557                                       " slave %d bonding device \"%s\"",
1558                                       bd, ibv_list[ret]->name);
1559                         ibv_match[nd++] = ibv_list[ret];
1560                         break;
1561                 }
1562                 if (mlx5_dev_to_pci_addr
1563                         (ibv_list[ret]->ibdev_path, &pci_addr))
1564                         continue;
1565                 if (pci_dev->addr.domain != pci_addr.domain ||
1566                     pci_dev->addr.bus != pci_addr.bus ||
1567                     pci_dev->addr.devid != pci_addr.devid ||
1568                     pci_dev->addr.function != pci_addr.function)
1569                         continue;
1570                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1571                         ibv_list[ret]->name);
1572                 ibv_match[nd++] = ibv_list[ret];
1573         }
1574         ibv_match[nd] = NULL;
1575         if (!nd) {
1576                 /* No device matches, just complain and bail out. */
1577                 DRV_LOG(WARNING,
1578                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1579                         " are kernel drivers loaded?",
1580                         pci_dev->addr.domain, pci_dev->addr.bus,
1581                         pci_dev->addr.devid, pci_dev->addr.function);
1582                 rte_errno = ENOENT;
1583                 ret = -rte_errno;
1584                 goto exit;
1585         }
1586         if (nd == 1) {
1587                 /*
1588                  * Found single matching device may have multiple ports.
1589                  * Each port may be representor, we have to check the port
1590                  * number and check the representors existence.
1591                  */
1592                 if (nl_rdma >= 0)
1593                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1594                 if (!np)
1595                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1596                                          " ports number", ibv_match[0]->name);
1597                 if (bd >= 0 && !np) {
1598                         DRV_LOG(ERR, "can not get ports"
1599                                      " for bonding device");
1600                         rte_errno = ENOENT;
1601                         ret = -rte_errno;
1602                         goto exit;
1603                 }
1604         }
1605 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1606         if (bd >= 0) {
1607                 /*
1608                  * This may happen if there is VF LAG kernel support and
1609                  * application is compiled with older rdma_core library.
1610                  */
1611                 DRV_LOG(ERR,
1612                         "No kernel/verbs support for VF LAG bonding found.");
1613                 rte_errno = ENOTSUP;
1614                 ret = -rte_errno;
1615                 goto exit;
1616         }
1617 #endif
1618         /*
1619          * Now we can determine the maximal
1620          * amount of devices to be spawned.
1621          */
1622         list = mlx5_malloc(MLX5_MEM_ZERO,
1623                            sizeof(struct mlx5_dev_spawn_data) *
1624                            (np ? np : nd),
1625                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1626         if (!list) {
1627                 DRV_LOG(ERR, "spawn data array allocation failure");
1628                 rte_errno = ENOMEM;
1629                 ret = -rte_errno;
1630                 goto exit;
1631         }
1632         if (bd >= 0 || np > 1) {
1633                 /*
1634                  * Single IB device with multiple ports found,
1635                  * it may be E-Switch master device and representors.
1636                  * We have to perform identification through the ports.
1637                  */
1638                 MLX5_ASSERT(nl_rdma >= 0);
1639                 MLX5_ASSERT(ns == 0);
1640                 MLX5_ASSERT(nd == 1);
1641                 MLX5_ASSERT(np);
1642                 for (i = 1; i <= np; ++i) {
1643                         list[ns].max_port = np;
1644                         list[ns].phys_port = i;
1645                         list[ns].phys_dev = ibv_match[0];
1646                         list[ns].eth_dev = NULL;
1647                         list[ns].pci_dev = pci_dev;
1648                         list[ns].pf_bond = bd;
1649                         list[ns].ifindex = mlx5_nl_ifindex
1650                                 (nl_rdma,
1651                                 mlx5_os_get_dev_device_name
1652                                                 (list[ns].phys_dev), i);
1653                         if (!list[ns].ifindex) {
1654                                 /*
1655                                  * No network interface index found for the
1656                                  * specified port, it means there is no
1657                                  * representor on this port. It's OK,
1658                                  * there can be disabled ports, for example
1659                                  * if sriov_numvfs < sriov_totalvfs.
1660                                  */
1661                                 continue;
1662                         }
1663                         ret = -1;
1664                         if (nl_route >= 0)
1665                                 ret = mlx5_nl_switch_info
1666                                                (nl_route,
1667                                                 list[ns].ifindex,
1668                                                 &list[ns].info);
1669                         if (ret || (!list[ns].info.representor &&
1670                                     !list[ns].info.master)) {
1671                                 /*
1672                                  * We failed to recognize representors with
1673                                  * Netlink, let's try to perform the task
1674                                  * with sysfs.
1675                                  */
1676                                 ret =  mlx5_sysfs_switch_info
1677                                                 (list[ns].ifindex,
1678                                                  &list[ns].info);
1679                         }
1680                         if (!ret && bd >= 0) {
1681                                 switch (list[ns].info.name_type) {
1682                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1683                                         if (list[ns].info.port_name == bd)
1684                                                 ns++;
1685                                         break;
1686                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1687                                         /* Fallthrough */
1688                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1689                                         if (list[ns].info.pf_num == bd)
1690                                                 ns++;
1691                                         break;
1692                                 default:
1693                                         break;
1694                                 }
1695                                 continue;
1696                         }
1697                         if (!ret && (list[ns].info.representor ^
1698                                      list[ns].info.master))
1699                                 ns++;
1700                 }
1701                 if (!ns) {
1702                         DRV_LOG(ERR,
1703                                 "unable to recognize master/representors"
1704                                 " on the IB device with multiple ports");
1705                         rte_errno = ENOENT;
1706                         ret = -rte_errno;
1707                         goto exit;
1708                 }
1709         } else {
1710                 /*
1711                  * The existence of several matching entries (nd > 1) means
1712                  * port representors have been instantiated. No existing Verbs
1713                  * call nor sysfs entries can tell them apart, this can only
1714                  * be done through Netlink calls assuming kernel drivers are
1715                  * recent enough to support them.
1716                  *
1717                  * In the event of identification failure through Netlink,
1718                  * try again through sysfs, then:
1719                  *
1720                  * 1. A single IB device matches (nd == 1) with single
1721                  *    port (np=0/1) and is not a representor, assume
1722                  *    no switch support.
1723                  *
1724                  * 2. Otherwise no safe assumptions can be made;
1725                  *    complain louder and bail out.
1726                  */
1727                 for (i = 0; i != nd; ++i) {
1728                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1729                         list[ns].max_port = 1;
1730                         list[ns].phys_port = 1;
1731                         list[ns].phys_dev = ibv_match[i];
1732                         list[ns].eth_dev = NULL;
1733                         list[ns].pci_dev = pci_dev;
1734                         list[ns].pf_bond = -1;
1735                         list[ns].ifindex = 0;
1736                         if (nl_rdma >= 0)
1737                                 list[ns].ifindex = mlx5_nl_ifindex
1738                                 (nl_rdma,
1739                                 mlx5_os_get_dev_device_name
1740                                                 (list[ns].phys_dev), 1);
1741                         if (!list[ns].ifindex) {
1742                                 char ifname[IF_NAMESIZE];
1743
1744                                 /*
1745                                  * Netlink failed, it may happen with old
1746                                  * ib_core kernel driver (before 4.16).
1747                                  * We can assume there is old driver because
1748                                  * here we are processing single ports IB
1749                                  * devices. Let's try sysfs to retrieve
1750                                  * the ifindex. The method works for
1751                                  * master device only.
1752                                  */
1753                                 if (nd > 1) {
1754                                         /*
1755                                          * Multiple devices found, assume
1756                                          * representors, can not distinguish
1757                                          * master/representor and retrieve
1758                                          * ifindex via sysfs.
1759                                          */
1760                                         continue;
1761                                 }
1762                                 ret = mlx5_get_ifname_sysfs
1763                                         (ibv_match[i]->ibdev_path, ifname);
1764                                 if (!ret)
1765                                         list[ns].ifindex =
1766                                                 if_nametoindex(ifname);
1767                                 if (!list[ns].ifindex) {
1768                                         /*
1769                                          * No network interface index found
1770                                          * for the specified device, it means
1771                                          * there it is neither representor
1772                                          * nor master.
1773                                          */
1774                                         continue;
1775                                 }
1776                         }
1777                         ret = -1;
1778                         if (nl_route >= 0)
1779                                 ret = mlx5_nl_switch_info
1780                                                (nl_route,
1781                                                 list[ns].ifindex,
1782                                                 &list[ns].info);
1783                         if (ret || (!list[ns].info.representor &&
1784                                     !list[ns].info.master)) {
1785                                 /*
1786                                  * We failed to recognize representors with
1787                                  * Netlink, let's try to perform the task
1788                                  * with sysfs.
1789                                  */
1790                                 ret =  mlx5_sysfs_switch_info
1791                                                 (list[ns].ifindex,
1792                                                  &list[ns].info);
1793                         }
1794                         if (!ret && (list[ns].info.representor ^
1795                                      list[ns].info.master)) {
1796                                 ns++;
1797                         } else if ((nd == 1) &&
1798                                    !list[ns].info.representor &&
1799                                    !list[ns].info.master) {
1800                                 /*
1801                                  * Single IB device with
1802                                  * one physical port and
1803                                  * attached network device.
1804                                  * May be SRIOV is not enabled
1805                                  * or there is no representors.
1806                                  */
1807                                 DRV_LOG(INFO, "no E-Switch support detected");
1808                                 ns++;
1809                                 break;
1810                         }
1811                 }
1812                 if (!ns) {
1813                         DRV_LOG(ERR,
1814                                 "unable to recognize master/representors"
1815                                 " on the multiple IB devices");
1816                         rte_errno = ENOENT;
1817                         ret = -rte_errno;
1818                         goto exit;
1819                 }
1820         }
1821         MLX5_ASSERT(ns);
1822         /*
1823          * Sort list to probe devices in natural order for users convenience
1824          * (i.e. master first, then representors from lowest to highest ID).
1825          */
1826         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1827         /* Default configuration. */
1828         dev_config = (struct mlx5_dev_config){
1829                 .hw_padding = 0,
1830                 .mps = MLX5_ARG_UNSET,
1831                 .dbnc = MLX5_ARG_UNSET,
1832                 .rx_vec_en = 1,
1833                 .txq_inline_max = MLX5_ARG_UNSET,
1834                 .txq_inline_min = MLX5_ARG_UNSET,
1835                 .txq_inline_mpw = MLX5_ARG_UNSET,
1836                 .txqs_inline = MLX5_ARG_UNSET,
1837                 .vf_nl_en = 1,
1838                 .mr_ext_memseg_en = 1,
1839                 .mprq = {
1840                         .enabled = 0, /* Disabled by default. */
1841                         .stride_num_n = 0,
1842                         .stride_size_n = 0,
1843                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
1844                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
1845                 },
1846                 .dv_esw_en = 1,
1847                 .dv_flow_en = 1,
1848                 .decap_en = 1,
1849                 .log_hp_size = MLX5_ARG_UNSET,
1850         };
1851         /* Device specific configuration. */
1852         switch (pci_dev->id.device_id) {
1853         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1854         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1855         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1856         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1857         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1858         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1859         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1860                 dev_config.vf = 1;
1861                 break;
1862         default:
1863                 break;
1864         }
1865         for (i = 0; i != ns; ++i) {
1866                 uint32_t restore;
1867
1868                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1869                                                  &list[i],
1870                                                  dev_config);
1871                 if (!list[i].eth_dev) {
1872                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1873                                 break;
1874                         /* Device is disabled or already spawned. Ignore it. */
1875                         continue;
1876                 }
1877                 restore = list[i].eth_dev->data->dev_flags;
1878                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1879                 /* Restore non-PCI flags cleared by the above call. */
1880                 list[i].eth_dev->data->dev_flags |= restore;
1881                 rte_eth_dev_probing_finish(list[i].eth_dev);
1882         }
1883         if (i != ns) {
1884                 DRV_LOG(ERR,
1885                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1886                         " encountering an error: %s",
1887                         pci_dev->addr.domain, pci_dev->addr.bus,
1888                         pci_dev->addr.devid, pci_dev->addr.function,
1889                         strerror(rte_errno));
1890                 ret = -rte_errno;
1891                 /* Roll back. */
1892                 while (i--) {
1893                         if (!list[i].eth_dev)
1894                                 continue;
1895                         mlx5_dev_close(list[i].eth_dev);
1896                         /* mac_addrs must not be freed because in dev_private */
1897                         list[i].eth_dev->data->mac_addrs = NULL;
1898                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1899                 }
1900                 /* Restore original error. */
1901                 rte_errno = -ret;
1902         } else {
1903                 ret = 0;
1904         }
1905 exit:
1906         /*
1907          * Do the routine cleanup:
1908          * - close opened Netlink sockets
1909          * - free allocated spawn data array
1910          * - free the Infiniband device list
1911          */
1912         if (nl_rdma >= 0)
1913                 close(nl_rdma);
1914         if (nl_route >= 0)
1915                 close(nl_route);
1916         if (list)
1917                 mlx5_free(list);
1918         MLX5_ASSERT(ibv_list);
1919         mlx5_glue->free_device_list(ibv_list);
1920         return ret;
1921 }
1922
1923 static int
1924 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1925 {
1926         char *env;
1927         int value;
1928
1929         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1930         /* Get environment variable to store. */
1931         env = getenv(MLX5_SHUT_UP_BF);
1932         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1933         if (config->dbnc == MLX5_ARG_UNSET)
1934                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1935         else
1936                 setenv(MLX5_SHUT_UP_BF,
1937                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1938         return value;
1939 }
1940
1941 static void
1942 mlx5_restore_doorbell_mapping_env(int value)
1943 {
1944         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1945         /* Restore the original environment variable state. */
1946         if (value == MLX5_ARG_UNSET)
1947                 unsetenv(MLX5_SHUT_UP_BF);
1948         else
1949                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1950 }
1951
1952 /**
1953  * Extract pdn of PD object using DV API.
1954  *
1955  * @param[in] pd
1956  *   Pointer to the verbs PD object.
1957  * @param[out] pdn
1958  *   Pointer to the PD object number variable.
1959  *
1960  * @return
1961  *   0 on success, error value otherwise.
1962  */
1963 int
1964 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1965 {
1966 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1967         struct mlx5dv_obj obj;
1968         struct mlx5dv_pd pd_info;
1969         int ret = 0;
1970
1971         obj.pd.in = pd;
1972         obj.pd.out = &pd_info;
1973         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
1974         if (ret) {
1975                 DRV_LOG(DEBUG, "Fail to get PD object info");
1976                 return ret;
1977         }
1978         *pdn = pd_info.pdn;
1979         return 0;
1980 #else
1981         (void)pd;
1982         (void)pdn;
1983         return -ENOTSUP;
1984 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
1985 }
1986
1987 /**
1988  * Function API to open IB device.
1989  *
1990  * This function calls the Linux glue APIs to open a device.
1991  *
1992  * @param[in] spawn
1993  *   Pointer to the IB device attributes (name, port, etc).
1994  * @param[out] config
1995  *   Pointer to device configuration structure.
1996  * @param[out] sh
1997  *   Pointer to shared context structure.
1998  *
1999  * @return
2000  *   0 on success, a positive error value otherwise.
2001  */
2002 int
2003 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2004                      const struct mlx5_dev_config *config,
2005                      struct mlx5_dev_ctx_shared *sh)
2006 {
2007         int dbmap_env;
2008         int err = 0;
2009
2010         sh->numa_node = spawn->pci_dev->device.numa_node;
2011         pthread_mutex_init(&sh->txpp.mutex, NULL);
2012         /*
2013          * Configure environment variable "MLX5_BF_SHUT_UP"
2014          * before the device creation. The rdma_core library
2015          * checks the variable at device creation and
2016          * stores the result internally.
2017          */
2018         dbmap_env = mlx5_config_doorbell_mapping_env(config);
2019         /* Try to open IB device with DV first, then usual Verbs. */
2020         errno = 0;
2021         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2022         if (sh->ctx) {
2023                 sh->devx = 1;
2024                 DRV_LOG(DEBUG, "DevX is supported");
2025                 /* The device is created, no need for environment. */
2026                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2027         } else {
2028                 /* The environment variable is still configured. */
2029                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2030                 err = errno ? errno : ENODEV;
2031                 /*
2032                  * The environment variable is not needed anymore,
2033                  * all device creation attempts are completed.
2034                  */
2035                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2036                 if (!sh->ctx)
2037                         return err;
2038                 DRV_LOG(DEBUG, "DevX is NOT supported");
2039                 err = 0;
2040         }
2041         return err;
2042 }
2043
2044 /**
2045  * Install shared asynchronous device events handler.
2046  * This function is implemented to support event sharing
2047  * between multiple ports of single IB device.
2048  *
2049  * @param sh
2050  *   Pointer to mlx5_dev_ctx_shared object.
2051  */
2052 void
2053 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2054 {
2055         int ret;
2056         int flags;
2057
2058         sh->intr_handle.fd = -1;
2059         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2060         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2061                     F_SETFL, flags | O_NONBLOCK);
2062         if (ret) {
2063                 DRV_LOG(INFO, "failed to change file descriptor async event"
2064                         " queue");
2065         } else {
2066                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2067                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2068                 if (rte_intr_callback_register(&sh->intr_handle,
2069                                         mlx5_dev_interrupt_handler, sh)) {
2070                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
2071                         sh->intr_handle.fd = -1;
2072                 }
2073         }
2074         if (sh->devx) {
2075 #ifdef HAVE_IBV_DEVX_ASYNC
2076                 sh->intr_handle_devx.fd = -1;
2077                 sh->devx_comp =
2078                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2079                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2080                 if (!devx_comp) {
2081                         DRV_LOG(INFO, "failed to allocate devx_comp.");
2082                         return;
2083                 }
2084                 flags = fcntl(devx_comp->fd, F_GETFL);
2085                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2086                 if (ret) {
2087                         DRV_LOG(INFO, "failed to change file descriptor"
2088                                 " devx comp");
2089                         return;
2090                 }
2091                 sh->intr_handle_devx.fd = devx_comp->fd;
2092                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2093                 if (rte_intr_callback_register(&sh->intr_handle_devx,
2094                                         mlx5_dev_interrupt_handler_devx, sh)) {
2095                         DRV_LOG(INFO, "Fail to install the devx shared"
2096                                 " interrupt.");
2097                         sh->intr_handle_devx.fd = -1;
2098                 }
2099 #endif /* HAVE_IBV_DEVX_ASYNC */
2100         }
2101 }
2102
2103 /**
2104  * Uninstall shared asynchronous device events handler.
2105  * This function is implemented to support event sharing
2106  * between multiple ports of single IB device.
2107  *
2108  * @param dev
2109  *   Pointer to mlx5_dev_ctx_shared object.
2110  */
2111 void
2112 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2113 {
2114         if (sh->intr_handle.fd >= 0)
2115                 mlx5_intr_callback_unregister(&sh->intr_handle,
2116                                               mlx5_dev_interrupt_handler, sh);
2117 #ifdef HAVE_IBV_DEVX_ASYNC
2118         if (sh->intr_handle_devx.fd >= 0)
2119                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2120                                   mlx5_dev_interrupt_handler_devx, sh);
2121         if (sh->devx_comp)
2122                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2123 #endif
2124 }
2125
2126 /**
2127  * Read statistics by a named counter.
2128  *
2129  * @param[in] priv
2130  *   Pointer to the private device data structure.
2131  * @param[in] ctr_name
2132  *   Pointer to the name of the statistic counter to read
2133  * @param[out] stat
2134  *   Pointer to read statistic value.
2135  * @return
2136  *   0 on success and stat is valud, 1 if failed to read the value
2137  *   rte_errno is set.
2138  *
2139  */
2140 int
2141 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2142                       uint64_t *stat)
2143 {
2144         int fd;
2145
2146         if (priv->sh) {
2147                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2148                           priv->sh->ibdev_path,
2149                           priv->dev_port,
2150                           ctr_name);
2151                 fd = open(path, O_RDONLY);
2152                 if (fd != -1) {
2153                         char buf[21] = {'\0'};
2154                         ssize_t n = read(fd, buf, sizeof(buf));
2155
2156                         close(fd);
2157                         if (n != -1) {
2158                                 *stat = strtoull(buf, NULL, 10);
2159                                 return 0;
2160                         }
2161                 }
2162         }
2163         *stat = 0;
2164         return 1;
2165 }
2166
2167 /**
2168  * Set the reg_mr and dereg_mr call backs
2169  *
2170  * @param reg_mr_cb[out]
2171  *   Pointer to reg_mr func
2172  * @param dereg_mr_cb[out]
2173  *   Pointer to dereg_mr func
2174  *
2175  */
2176 void
2177 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2178                       mlx5_dereg_mr_t *dereg_mr_cb)
2179 {
2180         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2181         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2182 }
2183
2184 /**
2185  * Remove a MAC address from device
2186  *
2187  * @param dev
2188  *   Pointer to Ethernet device structure.
2189  * @param index
2190  *   MAC address index.
2191  */
2192 void
2193 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2194 {
2195         struct mlx5_priv *priv = dev->data->dev_private;
2196         const int vf = priv->config.vf;
2197
2198         if (vf)
2199                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2200                                         mlx5_ifindex(dev), priv->mac_own,
2201                                         &dev->data->mac_addrs[index], index);
2202 }
2203
2204 /**
2205  * Adds a MAC address to the device
2206  *
2207  * @param dev
2208  *   Pointer to Ethernet device structure.
2209  * @param mac_addr
2210  *   MAC address to register.
2211  * @param index
2212  *   MAC address index.
2213  *
2214  * @return
2215  *   0 on success, a negative errno value otherwise
2216  */
2217 int
2218 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2219                      uint32_t index)
2220 {
2221         struct mlx5_priv *priv = dev->data->dev_private;
2222         const int vf = priv->config.vf;
2223         int ret = 0;
2224
2225         if (vf)
2226                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2227                                            mlx5_ifindex(dev), priv->mac_own,
2228                                            mac, index);
2229         return ret;
2230 }
2231
2232 /**
2233  * Modify a VF MAC address
2234  *
2235  * @param priv
2236  *   Pointer to device private data.
2237  * @param mac_addr
2238  *   MAC address to modify into.
2239  * @param iface_idx
2240  *   Net device interface index
2241  * @param vf_index
2242  *   VF index
2243  *
2244  * @return
2245  *   0 on success, a negative errno value otherwise
2246  */
2247 int
2248 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2249                            unsigned int iface_idx,
2250                            struct rte_ether_addr *mac_addr,
2251                            int vf_index)
2252 {
2253         return mlx5_nl_vf_mac_addr_modify
2254                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2255 }
2256
2257 /**
2258  * Set device promiscuous mode
2259  *
2260  * @param dev
2261  *   Pointer to Ethernet device structure.
2262  * @param enable
2263  *   0 - promiscuous is disabled, otherwise - enabled
2264  *
2265  * @return
2266  *   0 on success, a negative error value otherwise
2267  */
2268 int
2269 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2270 {
2271         struct mlx5_priv *priv = dev->data->dev_private;
2272
2273         return mlx5_nl_promisc(priv->nl_socket_route,
2274                                mlx5_ifindex(dev), !!enable);
2275 }
2276
2277 /**
2278  * Set device promiscuous mode
2279  *
2280  * @param dev
2281  *   Pointer to Ethernet device structure.
2282  * @param enable
2283  *   0 - all multicase is disabled, otherwise - enabled
2284  *
2285  * @return
2286  *   0 on success, a negative error value otherwise
2287  */
2288 int
2289 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2290 {
2291         struct mlx5_priv *priv = dev->data->dev_private;
2292
2293         return mlx5_nl_allmulti(priv->nl_socket_route,
2294                                 mlx5_ifindex(dev), !!enable);
2295 }
2296
2297 const struct eth_dev_ops mlx5_os_dev_ops = {
2298         .dev_configure = mlx5_dev_configure,
2299         .dev_start = mlx5_dev_start,
2300         .dev_stop = mlx5_dev_stop,
2301         .dev_set_link_down = mlx5_set_link_down,
2302         .dev_set_link_up = mlx5_set_link_up,
2303         .dev_close = mlx5_dev_close,
2304         .promiscuous_enable = mlx5_promiscuous_enable,
2305         .promiscuous_disable = mlx5_promiscuous_disable,
2306         .allmulticast_enable = mlx5_allmulticast_enable,
2307         .allmulticast_disable = mlx5_allmulticast_disable,
2308         .link_update = mlx5_link_update,
2309         .stats_get = mlx5_stats_get,
2310         .stats_reset = mlx5_stats_reset,
2311         .xstats_get = mlx5_xstats_get,
2312         .xstats_reset = mlx5_xstats_reset,
2313         .xstats_get_names = mlx5_xstats_get_names,
2314         .fw_version_get = mlx5_fw_version_get,
2315         .dev_infos_get = mlx5_dev_infos_get,
2316         .read_clock = mlx5_txpp_read_clock,
2317         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2318         .vlan_filter_set = mlx5_vlan_filter_set,
2319         .rx_queue_setup = mlx5_rx_queue_setup,
2320         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2321         .tx_queue_setup = mlx5_tx_queue_setup,
2322         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2323         .rx_queue_release = mlx5_rx_queue_release,
2324         .tx_queue_release = mlx5_tx_queue_release,
2325         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2326         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2327         .mac_addr_remove = mlx5_mac_addr_remove,
2328         .mac_addr_add = mlx5_mac_addr_add,
2329         .mac_addr_set = mlx5_mac_addr_set,
2330         .set_mc_addr_list = mlx5_set_mc_addr_list,
2331         .mtu_set = mlx5_dev_set_mtu,
2332         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2333         .vlan_offload_set = mlx5_vlan_offload_set,
2334         .reta_update = mlx5_dev_rss_reta_update,
2335         .reta_query = mlx5_dev_rss_reta_query,
2336         .rss_hash_update = mlx5_rss_hash_update,
2337         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2338         .filter_ctrl = mlx5_dev_filter_ctrl,
2339         .rx_descriptor_status = mlx5_rx_descriptor_status,
2340         .tx_descriptor_status = mlx5_tx_descriptor_status,
2341         .rxq_info_get = mlx5_rxq_info_get,
2342         .txq_info_get = mlx5_txq_info_get,
2343         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2344         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2345         .rx_queue_count = mlx5_rx_queue_count,
2346         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2347         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2348         .is_removed = mlx5_is_removed,
2349         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2350         .get_module_info = mlx5_get_module_info,
2351         .get_module_eeprom = mlx5_get_module_eeprom,
2352         .hairpin_cap_get = mlx5_hairpin_cap_get,
2353         .mtr_ops_get = mlx5_flow_meter_ops_get,
2354 };
2355
2356 /* Available operations from secondary process. */
2357 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2358         .stats_get = mlx5_stats_get,
2359         .stats_reset = mlx5_stats_reset,
2360         .xstats_get = mlx5_xstats_get,
2361         .xstats_reset = mlx5_xstats_reset,
2362         .xstats_get_names = mlx5_xstats_get_names,
2363         .fw_version_get = mlx5_fw_version_get,
2364         .dev_infos_get = mlx5_dev_infos_get,
2365         .read_clock = mlx5_txpp_read_clock,
2366         .rx_descriptor_status = mlx5_rx_descriptor_status,
2367         .tx_descriptor_status = mlx5_tx_descriptor_status,
2368         .rxq_info_get = mlx5_rxq_info_get,
2369         .txq_info_get = mlx5_txq_info_get,
2370         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2371         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2372         .get_module_info = mlx5_get_module_info,
2373         .get_module_eeprom = mlx5_get_module_eeprom,
2374 };
2375
2376 /* Available operations in flow isolated mode. */
2377 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2378         .dev_configure = mlx5_dev_configure,
2379         .dev_start = mlx5_dev_start,
2380         .dev_stop = mlx5_dev_stop,
2381         .dev_set_link_down = mlx5_set_link_down,
2382         .dev_set_link_up = mlx5_set_link_up,
2383         .dev_close = mlx5_dev_close,
2384         .promiscuous_enable = mlx5_promiscuous_enable,
2385         .promiscuous_disable = mlx5_promiscuous_disable,
2386         .allmulticast_enable = mlx5_allmulticast_enable,
2387         .allmulticast_disable = mlx5_allmulticast_disable,
2388         .link_update = mlx5_link_update,
2389         .stats_get = mlx5_stats_get,
2390         .stats_reset = mlx5_stats_reset,
2391         .xstats_get = mlx5_xstats_get,
2392         .xstats_reset = mlx5_xstats_reset,
2393         .xstats_get_names = mlx5_xstats_get_names,
2394         .fw_version_get = mlx5_fw_version_get,
2395         .dev_infos_get = mlx5_dev_infos_get,
2396         .read_clock = mlx5_txpp_read_clock,
2397         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2398         .vlan_filter_set = mlx5_vlan_filter_set,
2399         .rx_queue_setup = mlx5_rx_queue_setup,
2400         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2401         .tx_queue_setup = mlx5_tx_queue_setup,
2402         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2403         .rx_queue_release = mlx5_rx_queue_release,
2404         .tx_queue_release = mlx5_tx_queue_release,
2405         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2406         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2407         .mac_addr_remove = mlx5_mac_addr_remove,
2408         .mac_addr_add = mlx5_mac_addr_add,
2409         .mac_addr_set = mlx5_mac_addr_set,
2410         .set_mc_addr_list = mlx5_set_mc_addr_list,
2411         .mtu_set = mlx5_dev_set_mtu,
2412         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2413         .vlan_offload_set = mlx5_vlan_offload_set,
2414         .filter_ctrl = mlx5_dev_filter_ctrl,
2415         .rx_descriptor_status = mlx5_rx_descriptor_status,
2416         .tx_descriptor_status = mlx5_tx_descriptor_status,
2417         .rxq_info_get = mlx5_rxq_info_get,
2418         .txq_info_get = mlx5_txq_info_get,
2419         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2420         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2421         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2422         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2423         .is_removed = mlx5_is_removed,
2424         .get_module_info = mlx5_get_module_info,
2425         .get_module_eeprom = mlx5_get_module_eeprom,
2426         .hairpin_cap_get = mlx5_hairpin_cap_get,
2427         .mtr_ops_get = mlx5_flow_meter_ops_get,
2428 };