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 #include "mlx5_nl.h"
  49 #include "mlx5_devx.h"
  50
  51 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  52
  53 #ifndef HAVE_IBV_MLX5_MOD_MPW
  54 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  55 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  56 #endif
  57
  58 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  59 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  60 #endif
  61
  62 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
  63
  64 /* Spinlock for mlx5_shared_data allocation. */
  65 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
  66
  67 /* Process local data for secondary processes. */
  68 static struct mlx5_local_data mlx5_local_data;
  69
  70 /**
  71  * Set the completion channel file descriptor interrupt as non-blocking.
  72  *
  73  * @param[in] rxq_obj
  74  *   Pointer to RQ channel object, which includes the channel fd
  75  *
  76  * @param[out] fd
  77  *   The file descriptor (representing the intetrrupt) used in this channel.
  78  *
  79  * @return
  80  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
  81  */
  82 int
  83 mlx5_os_set_nonblock_channel_fd(int fd)
  84 {
  85         int flags;
  86
  87         flags = fcntl(fd, F_GETFL);
  88         return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
  89 }
  90
  91 /**
  92  * Get mlx5 device attributes. The glue function query_device_ex() is called
  93  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  94  * device attributes from the glue out parameter.
  95  *
  96  * @param dev
  97  *   Pointer to ibv context.
  98  *
  99  * @param device_attr
 100  *   Pointer to mlx5 device attributes.
 101  *
 102  * @return
 103  *   0 on success, non zero error number otherwise
 104  */
 105 int
 106 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
 107 {
 108         int err;
 109         struct ibv_device_attr_ex attr_ex;
 110         memset(device_attr, 0, sizeof(*device_attr));
 111         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
 112         if (err)
 113                 return err;
 114
 115         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
 116         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
 117         device_attr->max_sge = attr_ex.orig_attr.max_sge;
 118         device_attr->max_cq = attr_ex.orig_attr.max_cq;
 119         device_attr->max_cqe = attr_ex.orig_attr.max_cqe;
 120         device_attr->max_mr = attr_ex.orig_attr.max_mr;
 121         device_attr->max_pd = attr_ex.orig_attr.max_pd;
 122         device_attr->max_qp = attr_ex.orig_attr.max_qp;
 123         device_attr->max_srq = attr_ex.orig_attr.max_srq;
 124         device_attr->max_srq_wr = attr_ex.orig_attr.max_srq_wr;
 125         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 126         device_attr->max_rwq_indirection_table_size =
 127                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 128         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 129         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 130
 131         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 132         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 133         if (err)
 134                 return err;
 135
 136         device_attr->flags = dv_attr.flags;
 137         device_attr->comp_mask = dv_attr.comp_mask;
 138 #ifdef HAVE_IBV_MLX5_MOD_SWP
 139         device_attr->sw_parsing_offloads =
 140                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 141 #endif
 142         device_attr->min_single_stride_log_num_of_bytes =
 143                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 144         device_attr->max_single_stride_log_num_of_bytes =
 145                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 146         device_attr->min_single_wqe_log_num_of_strides =
 147                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 148         device_attr->max_single_wqe_log_num_of_strides =
 149                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 150         device_attr->stride_supported_qpts =
 151                 dv_attr.striding_rq_caps.supported_qpts;
 152 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 153         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 154 #endif
 155
 156         return err;
 157 }
 158
 159 /**
 160  * Verbs callback to allocate a memory. This function should allocate the space
 161  * according to the size provided residing inside a huge page.
 162  * Please note that all allocation must respect the alignment from libmlx5
 163  * (i.e. currently rte_mem_page_size()).
 164  *
 165  * @param[in] size
 166  *   The size in bytes of the memory to allocate.
 167  * @param[in] data
 168  *   A pointer to the callback data.
 169  *
 170  * @return
 171  *   Allocated buffer, NULL otherwise and rte_errno is set.
 172  */
 173 static void *
 174 mlx5_alloc_verbs_buf(size_t size, void *data)
 175 {
 176         struct mlx5_dev_ctx_shared *sh = data;
 177         void *ret;
 178         size_t alignment = rte_mem_page_size();
 179         if (alignment == (size_t)-1) {
 180                 DRV_LOG(ERR, "Failed to get mem page size");
 181                 rte_errno = ENOMEM;
 182                 return NULL;
 183         }
 184
 185         MLX5_ASSERT(data != NULL);
 186         ret = mlx5_malloc(0, size, alignment, sh->numa_node);
 187         if (!ret && size)
 188                 rte_errno = ENOMEM;
 189         return ret;
 190 }
 191
 192 /**
 193  * Verbs callback to free a memory.
 194  *
 195  * @param[in] ptr
 196  *   A pointer to the memory to free.
 197  * @param[in] data
 198  *   A pointer to the callback data.
 199  */
 200 static void
 201 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 202 {
 203         MLX5_ASSERT(data != NULL);
 204         mlx5_free(ptr);
 205 }
 206
 207 /**
 208  * Initialize DR related data within private structure.
 209  * Routine checks the reference counter and does actual
 210  * resources creation/initialization only if counter is zero.
 211  *
 212  * @param[in] priv
 213  *   Pointer to the private device data structure.
 214  *
 215  * @return
 216  *   Zero on success, positive error code otherwise.
 217  */
 218 static int
 219 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 220 {
 221         struct mlx5_dev_ctx_shared *sh = priv->sh;
 222         char s[MLX5_HLIST_NAMESIZE] __rte_unused;
 223         int err;
 224
 225         MLX5_ASSERT(sh && sh->refcnt);
 226         if (sh->refcnt > 1)
 227                 return 0;
 228         err = mlx5_alloc_table_hash_list(priv);
 229         if (err)
 230                 goto error;
 231         /* The resources below are only valid with DV support. */
 232 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 233         /* Init port id action cache list. */
 234         snprintf(s, sizeof(s), "%s_port_id_action_cache", sh->ibdev_name);
 235         mlx5_cache_list_init(&sh->port_id_action_list, s, 0, sh,
 236                              flow_dv_port_id_create_cb,
 237                              flow_dv_port_id_match_cb,
 238                              flow_dv_port_id_remove_cb);
 239         /* Init push vlan action cache list. */
 240         snprintf(s, sizeof(s), "%s_push_vlan_action_cache", sh->ibdev_name);
 241         mlx5_cache_list_init(&sh->push_vlan_action_list, s, 0, sh,
 242                              flow_dv_push_vlan_create_cb,
 243                              flow_dv_push_vlan_match_cb,
 244                              flow_dv_push_vlan_remove_cb);
 245         /* Init sample action cache list. */
 246         snprintf(s, sizeof(s), "%s_sample_action_cache", sh->ibdev_name);
 247         mlx5_cache_list_init(&sh->sample_action_list, s, 0, sh,
 248                              flow_dv_sample_create_cb,
 249                              flow_dv_sample_match_cb,
 250                              flow_dv_sample_remove_cb);
 251         /* Init dest array action cache list. */
 252         snprintf(s, sizeof(s), "%s_dest_array_cache", sh->ibdev_name);
 253         mlx5_cache_list_init(&sh->dest_array_list, s, 0, sh,
 254                              flow_dv_dest_array_create_cb,
 255                              flow_dv_dest_array_match_cb,
 256                              flow_dv_dest_array_remove_cb);
 257         /* Create tags hash list table. */
 258         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 259         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE, 0,
 260                                           MLX5_HLIST_WRITE_MOST,
 261                                           flow_dv_tag_create_cb,
 262                                           flow_dv_tag_match_cb,
 263                                           flow_dv_tag_remove_cb);
 264         if (!sh->tag_table) {
 265                 DRV_LOG(ERR, "tags with hash creation failed.");
 266                 err = ENOMEM;
 267                 goto error;
 268         }
 269         sh->tag_table->ctx = sh;
 270         snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
 271         sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ,
 272                                             0, MLX5_HLIST_WRITE_MOST |
 273                                             MLX5_HLIST_DIRECT_KEY,
 274                                             flow_dv_modify_create_cb,
 275                                             flow_dv_modify_match_cb,
 276                                             flow_dv_modify_remove_cb);
 277         if (!sh->modify_cmds) {
 278                 DRV_LOG(ERR, "hdr modify hash creation failed");
 279                 err = ENOMEM;
 280                 goto error;
 281         }
 282         sh->modify_cmds->ctx = sh;
 283         snprintf(s, sizeof(s), "%s_encaps_decaps", sh->ibdev_name);
 284         sh->encaps_decaps = mlx5_hlist_create(s,
 285                                               MLX5_FLOW_ENCAP_DECAP_HTABLE_SZ,
 286                                               0, MLX5_HLIST_DIRECT_KEY |
 287                                               MLX5_HLIST_WRITE_MOST,
 288                                               flow_dv_encap_decap_create_cb,
 289                                               flow_dv_encap_decap_match_cb,
 290                                               flow_dv_encap_decap_remove_cb);
 291         if (!sh->encaps_decaps) {
 292                 DRV_LOG(ERR, "encap decap hash creation failed");
 293                 err = ENOMEM;
 294                 goto error;
 295         }
 296         sh->encaps_decaps->ctx = sh;
 297 #endif
 298 #ifdef HAVE_MLX5DV_DR
 299         void *domain;
 300
 301         /* Reference counter is zero, we should initialize structures. */
 302         domain = mlx5_glue->dr_create_domain(sh->ctx,
 303                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 304         if (!domain) {
 305                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 306                 err = errno;
 307                 goto error;
 308         }
 309         sh->rx_domain = domain;
 310         domain = mlx5_glue->dr_create_domain(sh->ctx,
 311                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 312         if (!domain) {
 313                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 314                 err = errno;
 315                 goto error;
 316         }
 317         sh->tx_domain = domain;
 318 #ifdef HAVE_MLX5DV_DR_ESWITCH
 319         if (priv->config.dv_esw_en) {
 320                 domain  = mlx5_glue->dr_create_domain
 321                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 322                 if (!domain) {
 323                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 324                         err = errno;
 325                         goto error;
 326                 }
 327                 sh->fdb_domain = domain;
 328                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 329         }
 330 #endif
 331         if (!sh->tunnel_hub)
 332                 err = mlx5_alloc_tunnel_hub(sh);
 333         if (err) {
 334                 DRV_LOG(ERR, "mlx5_alloc_tunnel_hub failed err=%d", err);
 335                 goto error;
 336         }
 337         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 338                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 339                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 340                 if (sh->fdb_domain)
 341                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 342         }
 343         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 344 #endif /* HAVE_MLX5DV_DR */
 345         sh->default_miss_action =
 346                         mlx5_glue->dr_create_flow_action_default_miss();
 347         if (!sh->default_miss_action)
 348                 DRV_LOG(WARNING, "Default miss action is not supported.");
 349         return 0;
 350 error:
 351         /* Rollback the created objects. */
 352         if (sh->rx_domain) {
 353                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 354                 sh->rx_domain = NULL;
 355         }
 356         if (sh->tx_domain) {
 357                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 358                 sh->tx_domain = NULL;
 359         }
 360         if (sh->fdb_domain) {
 361                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 362                 sh->fdb_domain = NULL;
 363         }
 364         if (sh->esw_drop_action) {
 365                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 366                 sh->esw_drop_action = NULL;
 367         }
 368         if (sh->pop_vlan_action) {
 369                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 370                 sh->pop_vlan_action = NULL;
 371         }
 372         if (sh->encaps_decaps) {
 373                 mlx5_hlist_destroy(sh->encaps_decaps);
 374                 sh->encaps_decaps = NULL;
 375         }
 376         if (sh->modify_cmds) {
 377                 mlx5_hlist_destroy(sh->modify_cmds);
 378                 sh->modify_cmds = NULL;
 379         }
 380         if (sh->tag_table) {
 381                 /* tags should be destroyed with flow before. */
 382                 mlx5_hlist_destroy(sh->tag_table);
 383                 sh->tag_table = NULL;
 384         }
 385         if (sh->tunnel_hub) {
 386                 mlx5_release_tunnel_hub(sh, priv->dev_port);
 387                 sh->tunnel_hub = NULL;
 388         }
 389         mlx5_free_table_hash_list(priv);
 390         return err;
 391 }
 392
 393 /**
 394  * Destroy DR related data within private structure.
 395  *
 396  * @param[in] priv
 397  *   Pointer to the private device data structure.
 398  */
 399 void
 400 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 401 {
 402         struct mlx5_dev_ctx_shared *sh = priv->sh;
 403
 404         MLX5_ASSERT(sh && sh->refcnt);
 405         if (sh->refcnt > 1)
 406                 return;
 407 #ifdef HAVE_MLX5DV_DR
 408         if (sh->rx_domain) {
 409                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 410                 sh->rx_domain = NULL;
 411         }
 412         if (sh->tx_domain) {
 413                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 414                 sh->tx_domain = NULL;
 415         }
 416 #ifdef HAVE_MLX5DV_DR_ESWITCH
 417         if (sh->fdb_domain) {
 418                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 419                 sh->fdb_domain = NULL;
 420         }
 421         if (sh->esw_drop_action) {
 422                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 423                 sh->esw_drop_action = NULL;
 424         }
 425 #endif
 426         if (sh->pop_vlan_action) {
 427                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 428                 sh->pop_vlan_action = NULL;
 429         }
 430 #endif /* HAVE_MLX5DV_DR */
 431         if (sh->default_miss_action)
 432                 mlx5_glue->destroy_flow_action
 433                                 (sh->default_miss_action);
 434         if (sh->encaps_decaps) {
 435                 mlx5_hlist_destroy(sh->encaps_decaps);
 436                 sh->encaps_decaps = NULL;
 437         }
 438         if (sh->modify_cmds) {
 439                 mlx5_hlist_destroy(sh->modify_cmds);
 440                 sh->modify_cmds = NULL;
 441         }
 442         if (sh->tag_table) {
 443                 /* tags should be destroyed with flow before. */
 444                 mlx5_hlist_destroy(sh->tag_table);
 445                 sh->tag_table = NULL;
 446         }
 447         if (sh->tunnel_hub) {
 448                 mlx5_release_tunnel_hub(sh, priv->dev_port);
 449                 sh->tunnel_hub = NULL;
 450         }
 451         mlx5_cache_list_destroy(&sh->port_id_action_list);
 452         mlx5_cache_list_destroy(&sh->push_vlan_action_list);
 453         mlx5_free_table_hash_list(priv);
 454 }
 455
 456 /**
 457  * Initialize shared data between primary and secondary process.
 458  *
 459  * A memzone is reserved by primary process and secondary processes attach to
 460  * the memzone.
 461  *
 462  * @return
 463  *   0 on success, a negative errno value otherwise and rte_errno is set.
 464  */
 465 static int
 466 mlx5_init_shared_data(void)
 467 {
 468         const struct rte_memzone *mz;
 469         int ret = 0;
 470
 471         rte_spinlock_lock(&mlx5_shared_data_lock);
 472         if (mlx5_shared_data == NULL) {
 473                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 474                         /* Allocate shared memory. */
 475                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 476                                                  sizeof(*mlx5_shared_data),
 477                                                  SOCKET_ID_ANY, 0);
 478                         if (mz == NULL) {
 479                                 DRV_LOG(ERR,
 480                                         "Cannot allocate mlx5 shared data");
 481                                 ret = -rte_errno;
 482                                 goto error;
 483                         }
 484                         mlx5_shared_data = mz->addr;
 485                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 486                         rte_spinlock_init(&mlx5_shared_data->lock);
 487                 } else {
 488                         /* Lookup allocated shared memory. */
 489                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 490                         if (mz == NULL) {
 491                                 DRV_LOG(ERR,
 492                                         "Cannot attach mlx5 shared data");
 493                                 ret = -rte_errno;
 494                                 goto error;
 495                         }
 496                         mlx5_shared_data = mz->addr;
 497                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 498                 }
 499         }
 500 error:
 501         rte_spinlock_unlock(&mlx5_shared_data_lock);
 502         return ret;
 503 }
 504
 505 /**
 506  * PMD global initialization.
 507  *
 508  * Independent from individual device, this function initializes global
 509  * per-PMD data structures distinguishing primary and secondary processes.
 510  * Hence, each initialization is called once per a process.
 511  *
 512  * @return
 513  *   0 on success, a negative errno value otherwise and rte_errno is set.
 514  */
 515 static int
 516 mlx5_init_once(void)
 517 {
 518         struct mlx5_shared_data *sd;
 519         struct mlx5_local_data *ld = &mlx5_local_data;
 520         int ret = 0;
 521
 522         if (mlx5_init_shared_data())
 523                 return -rte_errno;
 524         sd = mlx5_shared_data;
 525         MLX5_ASSERT(sd);
 526         rte_spinlock_lock(&sd->lock);
 527         switch (rte_eal_process_type()) {
 528         case RTE_PROC_PRIMARY:
 529                 if (sd->init_done)
 530                         break;
 531                 LIST_INIT(&sd->mem_event_cb_list);
 532                 rte_rwlock_init(&sd->mem_event_rwlock);
 533                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
 534                                                 mlx5_mr_mem_event_cb, NULL);
 535                 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
 536                                            mlx5_mp_os_primary_handle);
 537                 if (ret)
 538                         goto out;
 539                 sd->init_done = true;
 540                 break;
 541         case RTE_PROC_SECONDARY:
 542                 if (ld->init_done)
 543                         break;
 544                 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
 545                                              mlx5_mp_os_secondary_handle);
 546                 if (ret)
 547                         goto out;
 548                 ++sd->secondary_cnt;
 549                 ld->init_done = true;
 550                 break;
 551         default:
 552                 break;
 553         }
 554 out:
 555         rte_spinlock_unlock(&sd->lock);
 556         return ret;
 557 }
 558
 559 /**
 560  * Create the Tx queue DevX/Verbs object.
 561  *
 562  * @param dev
 563  *   Pointer to Ethernet device.
 564  * @param idx
 565  *   Queue index in DPDK Tx queue array.
 566  *
 567  * @return
 568  *   0 on success, a negative errno value otherwise and rte_errno is set.
 569  */
 570 static int
 571 mlx5_os_txq_obj_new(struct rte_eth_dev *dev, uint16_t idx)
 572 {
 573         struct mlx5_priv *priv = dev->data->dev_private;
 574         struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
 575         struct mlx5_txq_ctrl *txq_ctrl =
 576                         container_of(txq_data, struct mlx5_txq_ctrl, txq);
 577
 578         if (txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN)
 579                 return mlx5_txq_devx_obj_new(dev, idx);
 580 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
 581         if (!priv->config.dv_esw_en)
 582                 return mlx5_txq_devx_obj_new(dev, idx);
 583 #endif
 584         return mlx5_txq_ibv_obj_new(dev, idx);
 585 }
 586
 587 /**
 588  * Release an Tx DevX/verbs queue object.
 589  *
 590  * @param txq_obj
 591  *   DevX/Verbs Tx queue object.
 592  */
 593 static void
 594 mlx5_os_txq_obj_release(struct mlx5_txq_obj *txq_obj)
 595 {
 596         if (txq_obj->txq_ctrl->type == MLX5_TXQ_TYPE_HAIRPIN) {
 597                 mlx5_txq_devx_obj_release(txq_obj);
 598                 return;
 599         }
 600 #ifdef HAVE_MLX5DV_DEVX_UAR_OFFSET
 601         if (!txq_obj->txq_ctrl->priv->config.dv_esw_en) {
 602                 mlx5_txq_devx_obj_release(txq_obj);
 603                 return;
 604         }
 605 #endif
 606         mlx5_txq_ibv_obj_release(txq_obj);
 607 }
 608
 609 /**
 610  * DV flow counter mode detect and config.
 611  *
 612  * @param dev
 613  *   Pointer to rte_eth_dev structure.
 614  *
 615  */
 616 static void
 617 mlx5_flow_counter_mode_config(struct rte_eth_dev *dev __rte_unused)
 618 {
 619 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 620         struct mlx5_priv *priv = dev->data->dev_private;
 621         struct mlx5_dev_ctx_shared *sh = priv->sh;
 622         bool fallback;
 623
 624 #ifndef HAVE_IBV_DEVX_ASYNC
 625         fallback = true;
 626 #else
 627         fallback = false;
 628         if (!priv->config.devx || !priv->config.dv_flow_en ||
 629             !priv->config.hca_attr.flow_counters_dump ||
 630             !(priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4) ||
 631             (mlx5_flow_dv_discover_counter_offset_support(dev) == -ENOTSUP))
 632                 fallback = true;
 633 #endif
 634         if (fallback)
 635                 DRV_LOG(INFO, "Use fall-back DV counter management. Flow "
 636                         "counter dump:%d, bulk_alloc_bitmap:0x%hhx.",
 637                         priv->config.hca_attr.flow_counters_dump,
 638                         priv->config.hca_attr.flow_counter_bulk_alloc_bitmap);
 639         /* Initialize fallback mode only on the port initializes sh. */
 640         if (sh->refcnt == 1)
 641                 sh->cmng.counter_fallback = fallback;
 642         else if (fallback != sh->cmng.counter_fallback)
 643                 DRV_LOG(WARNING, "Port %d in sh has different fallback mode "
 644                         "with others:%d.", PORT_ID(priv), fallback);
 645 #endif
 646 }
 647
 648 /**
 649  * Spawn an Ethernet device from Verbs information.
 650  *
 651  * @param dpdk_dev
 652  *   Backing DPDK device.
 653  * @param spawn
 654  *   Verbs device parameters (name, port, switch_info) to spawn.
 655  * @param config
 656  *   Device configuration parameters.
 657  *
 658  * @return
 659  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 660  *   is set. The following errors are defined:
 661  *
 662  *   EBUSY: device is not supposed to be spawned.
 663  *   EEXIST: device is already spawned
 664  */
 665 static struct rte_eth_dev *
 666 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 667                struct mlx5_dev_spawn_data *spawn,
 668                struct mlx5_dev_config *config)
 669 {
 670         const struct mlx5_switch_info *switch_info = &spawn->info;
 671         struct mlx5_dev_ctx_shared *sh = NULL;
 672         struct ibv_port_attr port_attr;
 673         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 674         struct rte_eth_dev *eth_dev = NULL;
 675         struct mlx5_priv *priv = NULL;
 676         int err = 0;
 677         unsigned int hw_padding = 0;
 678         unsigned int mps;
 679         unsigned int cqe_comp;
 680         unsigned int cqe_pad = 0;
 681         unsigned int tunnel_en = 0;
 682         unsigned int mpls_en = 0;
 683         unsigned int swp = 0;
 684         unsigned int mprq = 0;
 685         unsigned int mprq_min_stride_size_n = 0;
 686         unsigned int mprq_max_stride_size_n = 0;
 687         unsigned int mprq_min_stride_num_n = 0;
 688         unsigned int mprq_max_stride_num_n = 0;
 689         struct rte_ether_addr mac;
 690         char name[RTE_ETH_NAME_MAX_LEN];
 691         int own_domain_id = 0;
 692         uint16_t port_id;
 693         unsigned int i;
 694 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 695         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 696 #endif
 697
 698         /* Determine if this port representor is supposed to be spawned. */
 699         if (switch_info->representor && dpdk_dev->devargs) {
 700                 struct rte_eth_devargs eth_da;
 701
 702                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 703                 if (err) {
 704                         rte_errno = -err;
 705                         DRV_LOG(ERR, "failed to process device arguments: %s",
 706                                 strerror(rte_errno));
 707                         return NULL;
 708                 }
 709                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 710                         if (eth_da.representor_ports[i] ==
 711                             (uint16_t)switch_info->port_name)
 712                                 break;
 713                 if (i == eth_da.nb_representor_ports) {
 714                         rte_errno = EBUSY;
 715                         return NULL;
 716                 }
 717         }
 718         /* Build device name. */
 719         if (spawn->pf_bond <  0) {
 720                 /* Single device. */
 721                 if (!switch_info->representor)
 722                         strlcpy(name, dpdk_dev->name, sizeof(name));
 723                 else
 724                         snprintf(name, sizeof(name), "%s_representor_%u",
 725                                  dpdk_dev->name, switch_info->port_name);
 726         } else {
 727                 /* Bonding device. */
 728                 if (!switch_info->representor)
 729                         snprintf(name, sizeof(name), "%s_%s",
 730                                  dpdk_dev->name,
 731                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 732                 else
 733                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 734                                  dpdk_dev->name,
 735                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 736                                  switch_info->port_name);
 737         }
 738         /* check if the device is already spawned */
 739         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 740                 rte_errno = EEXIST;
 741                 return NULL;
 742         }
 743         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 744         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 745                 struct mlx5_mp_id mp_id;
 746
 747                 eth_dev = rte_eth_dev_attach_secondary(name);
 748                 if (eth_dev == NULL) {
 749                         DRV_LOG(ERR, "can not attach rte ethdev");
 750                         rte_errno = ENOMEM;
 751                         return NULL;
 752                 }
 753                 priv = eth_dev->data->dev_private;
 754                 if (priv->sh->bond_dev != UINT16_MAX)
 755                         /* For bonding port, use primary PCI device. */
 756                         eth_dev->device =
 757                                 rte_eth_devices[priv->sh->bond_dev].device;
 758                 else
 759                         eth_dev->device = dpdk_dev;
 760                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
 761                 eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
 762                 eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
 763                 err = mlx5_proc_priv_init(eth_dev);
 764                 if (err)
 765                         return NULL;
 766                 mp_id.port_id = eth_dev->data->port_id;
 767                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 768                 /* Receive command fd from primary process */
 769                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 770                 if (err < 0)
 771                         goto err_secondary;
 772                 /* Remap UAR for Tx queues. */
 773                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 774                 if (err)
 775                         goto err_secondary;
 776                 /*
 777                  * Ethdev pointer is still required as input since
 778                  * the primary device is not accessible from the
 779                  * secondary process.
 780                  */
 781                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 782                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 783                 return eth_dev;
 784 err_secondary:
 785                 mlx5_dev_close(eth_dev);
 786                 return NULL;
 787         }
 788         /*
 789          * Some parameters ("tx_db_nc" in particularly) are needed in
 790          * advance to create dv/verbs device context. We proceed the
 791          * devargs here to get ones, and later proceed devargs again
 792          * to override some hardware settings.
 793          */
 794         err = mlx5_args(config, dpdk_dev->devargs);
 795         if (err) {
 796                 err = rte_errno;
 797                 DRV_LOG(ERR, "failed to process device arguments: %s",
 798                         strerror(rte_errno));
 799                 goto error;
 800         }
 801         if (config->dv_miss_info) {
 802                 if (switch_info->master || switch_info->representor)
 803                         config->dv_xmeta_en = MLX5_XMETA_MODE_META16;
 804         }
 805         mlx5_malloc_mem_select(config->sys_mem_en);
 806         sh = mlx5_alloc_shared_dev_ctx(spawn, config);
 807         if (!sh)
 808                 return NULL;
 809         config->devx = sh->devx;
 810 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 811         config->dest_tir = 1;
 812 #endif
 813 #ifdef HAVE_IBV_MLX5_MOD_SWP
 814         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 815 #endif
 816         /*
 817          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 818          * as all ConnectX-5 devices.
 819          */
 820 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 821         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 822 #endif
 823 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 824         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 825 #endif
 826         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 827         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 828                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 829                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 830                         mps = MLX5_MPW_ENHANCED;
 831                 } else {
 832                         DRV_LOG(DEBUG, "MPW is supported");
 833                         mps = MLX5_MPW;
 834                 }
 835         } else {
 836                 DRV_LOG(DEBUG, "MPW isn't supported");
 837                 mps = MLX5_MPW_DISABLED;
 838         }
 839 #ifdef HAVE_IBV_MLX5_MOD_SWP
 840         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 841                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 842         DRV_LOG(DEBUG, "SWP support: %u", swp);
 843 #endif
 844         config->swp = !!swp;
 845 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 846         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 847                 struct mlx5dv_striding_rq_caps mprq_caps =
 848                         dv_attr.striding_rq_caps;
 849
 850                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 851                         mprq_caps.min_single_stride_log_num_of_bytes);
 852                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 853                         mprq_caps.max_single_stride_log_num_of_bytes);
 854                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 855                         mprq_caps.min_single_wqe_log_num_of_strides);
 856                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 857                         mprq_caps.max_single_wqe_log_num_of_strides);
 858                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 859                         mprq_caps.supported_qpts);
 860                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 861                 mprq = 1;
 862                 mprq_min_stride_size_n =
 863                         mprq_caps.min_single_stride_log_num_of_bytes;
 864                 mprq_max_stride_size_n =
 865                         mprq_caps.max_single_stride_log_num_of_bytes;
 866                 mprq_min_stride_num_n =
 867                         mprq_caps.min_single_wqe_log_num_of_strides;
 868                 mprq_max_stride_num_n =
 869                         mprq_caps.max_single_wqe_log_num_of_strides;
 870         }
 871 #endif
 872         if (RTE_CACHE_LINE_SIZE == 128 &&
 873             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 874                 cqe_comp = 0;
 875         else
 876                 cqe_comp = 1;
 877         config->cqe_comp = cqe_comp;
 878 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 879         /* Whether device supports 128B Rx CQE padding. */
 880         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 881                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 882 #endif
 883 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 884         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 885                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 886                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 887                              (dv_attr.tunnel_offloads_caps &
 888                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 889                              (dv_attr.tunnel_offloads_caps &
 890                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 891         }
 892         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 893                 tunnel_en ? "" : "not ");
 894 #else
 895         DRV_LOG(WARNING,
 896                 "tunnel offloading disabled due to old OFED/rdma-core version");
 897 #endif
 898         config->tunnel_en = tunnel_en;
 899 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 900         mpls_en = ((dv_attr.tunnel_offloads_caps &
 901                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 902                    (dv_attr.tunnel_offloads_caps &
 903                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 904         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 905                 mpls_en ? "" : "not ");
 906 #else
 907         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 908                 " old OFED/rdma-core version or firmware configuration");
 909 #endif
 910         config->mpls_en = mpls_en;
 911         /* Check port status. */
 912         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 913         if (err) {
 914                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 915                 goto error;
 916         }
 917         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 918                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 919                 err = EINVAL;
 920                 goto error;
 921         }
 922         if (port_attr.state != IBV_PORT_ACTIVE)
 923                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 924                         mlx5_glue->port_state_str(port_attr.state),
 925                         port_attr.state);
 926         /* Allocate private eth device data. */
 927         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
 928                            sizeof(*priv),
 929                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 930         if (priv == NULL) {
 931                 DRV_LOG(ERR, "priv allocation failure");
 932                 err = ENOMEM;
 933                 goto error;
 934         }
 935         priv->sh = sh;
 936         priv->dev_port = spawn->phys_port;
 937         priv->pci_dev = spawn->pci_dev;
 938         priv->mtu = RTE_ETHER_MTU;
 939         priv->mp_id.port_id = port_id;
 940         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 941         /* Some internal functions rely on Netlink sockets, open them now. */
 942         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 943         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 944         priv->representor = !!switch_info->representor;
 945         priv->master = !!switch_info->master;
 946         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 947         priv->vport_meta_tag = 0;
 948         priv->vport_meta_mask = 0;
 949         priv->pf_bond = spawn->pf_bond;
 950 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 951         /*
 952          * The DevX port query API is implemented. E-Switch may use
 953          * either vport or reg_c[0] metadata register to match on
 954          * vport index. The engaged part of metadata register is
 955          * defined by mask.
 956          */
 957         if (switch_info->representor || switch_info->master) {
 958                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 959                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 960                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 961                                                  &devx_port);
 962                 if (err) {
 963                         DRV_LOG(WARNING,
 964                                 "can't query devx port %d on device %s",
 965                                 spawn->phys_port,
 966                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 967                         devx_port.comp_mask = 0;
 968                 }
 969         }
 970         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 971                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 972                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 973                 if (!priv->vport_meta_mask) {
 974                         DRV_LOG(ERR, "vport zero mask for port %d"
 975                                      " on bonding device %s",
 976                                      spawn->phys_port,
 977                                      mlx5_os_get_dev_device_name
 978                                                         (spawn->phys_dev));
 979                         err = ENOTSUP;
 980                         goto error;
 981                 }
 982                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 983                         DRV_LOG(ERR, "invalid vport tag for port %d"
 984                                      " on bonding device %s",
 985                                      spawn->phys_port,
 986                                      mlx5_os_get_dev_device_name
 987                                                         (spawn->phys_dev));
 988                         err = ENOTSUP;
 989                         goto error;
 990                 }
 991         }
 992         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 993                 priv->vport_id = devx_port.vport_num;
 994         } else if (spawn->pf_bond >= 0) {
 995                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 996                              " on bonding device %s",
 997                              spawn->phys_port,
 998                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 999                 err = ENOTSUP;
1000                 goto error;
1001         } else {
1002                 /* Suppose vport index in compatible way. */
1003                 priv->vport_id = switch_info->representor ?
1004                                  switch_info->port_name + 1 : -1;
1005         }
1006 #else
1007         /*
1008          * Kernel/rdma_core support single E-Switch per PF configurations
1009          * only and vport_id field contains the vport index for
1010          * associated VF, which is deduced from representor port name.
1011          * For example, let's have the IB device port 10, it has
1012          * attached network device eth0, which has port name attribute
1013          * pf0vf2, we can deduce the VF number as 2, and set vport index
1014          * as 3 (2+1). This assigning schema should be changed if the
1015          * multiple E-Switch instances per PF configurations or/and PCI
1016          * subfunctions are added.
1017          */
1018         priv->vport_id = switch_info->representor ?
1019                          switch_info->port_name + 1 : -1;
1020 #endif
1021         /* representor_id field keeps the unmodified VF index. */
1022         priv->representor_id = switch_info->representor ?
1023                                switch_info->port_name : -1;
1024         /*
1025          * Look for sibling devices in order to reuse their switch domain
1026          * if any, otherwise allocate one.
1027          */
1028         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1029                 const struct mlx5_priv *opriv =
1030                         rte_eth_devices[port_id].data->dev_private;
1031
1032                 if (!opriv ||
1033                     opriv->sh != priv->sh ||
1034                         opriv->domain_id ==
1035                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1036                         continue;
1037                 priv->domain_id = opriv->domain_id;
1038                 break;
1039         }
1040         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1041                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1042                 if (err) {
1043                         err = rte_errno;
1044                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1045                                 strerror(rte_errno));
1046                         goto error;
1047                 }
1048                 own_domain_id = 1;
1049         }
1050         /* Override some values set by hardware configuration. */
1051         mlx5_args(config, dpdk_dev->devargs);
1052         err = mlx5_dev_check_sibling_config(priv, config);
1053         if (err)
1054                 goto error;
1055         config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1056                             IBV_DEVICE_RAW_IP_CSUM);
1057         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1058                 (config->hw_csum ? "" : "not "));
1059 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1060         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1061         DRV_LOG(DEBUG, "counters are not supported");
1062 #endif
1063 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
1064         if (config->dv_flow_en) {
1065                 DRV_LOG(WARNING, "DV flow is not supported");
1066                 config->dv_flow_en = 0;
1067         }
1068 #endif
1069         config->ind_table_max_size =
1070                 sh->device_attr.max_rwq_indirection_table_size;
1071         /*
1072          * Remove this check once DPDK supports larger/variable
1073          * indirection tables.
1074          */
1075         if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1076                 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1077         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1078                 config->ind_table_max_size);
1079         config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1080                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1081         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1082                 (config->hw_vlan_strip ? "" : "not "));
1083         config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1084                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1085 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1086         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1087 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1088         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1089                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1090 #endif
1091         if (config->hw_padding && !hw_padding) {
1092                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1093                 config->hw_padding = 0;
1094         } else if (config->hw_padding) {
1095                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1096         }
1097         config->tso = (sh->device_attr.max_tso > 0 &&
1098                       (sh->device_attr.tso_supported_qpts &
1099                        (1 << IBV_QPT_RAW_PACKET)));
1100         if (config->tso)
1101                 config->tso_max_payload_sz = sh->device_attr.max_tso;
1102         /*
1103          * MPW is disabled by default, while the Enhanced MPW is enabled
1104          * by default.
1105          */
1106         if (config->mps == MLX5_ARG_UNSET)
1107                 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1108                                                           MLX5_MPW_DISABLED;
1109         else
1110                 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
1111         DRV_LOG(INFO, "%sMPS is %s",
1112                 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
1113                 config->mps == MLX5_MPW ? "legacy " : "",
1114                 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1115         if (config->cqe_comp && !cqe_comp) {
1116                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1117                 config->cqe_comp = 0;
1118         }
1119         if (config->cqe_pad && !cqe_pad) {
1120                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1121                 config->cqe_pad = 0;
1122         } else if (config->cqe_pad) {
1123                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1124         }
1125         if (config->devx) {
1126                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
1127                 if (err) {
1128                         err = -err;
1129                         goto error;
1130                 }
1131                 /* Check relax ordering support. */
1132                 if (!haswell_broadwell_cpu) {
1133                         sh->cmng.relaxed_ordering_write =
1134                                 config->hca_attr.relaxed_ordering_write;
1135                         sh->cmng.relaxed_ordering_read =
1136                                 config->hca_attr.relaxed_ordering_read;
1137                 } else {
1138                         sh->cmng.relaxed_ordering_read = 0;
1139                         sh->cmng.relaxed_ordering_write = 0;
1140                 }
1141                 /* Check for LRO support. */
1142                 if (config->dest_tir && config->hca_attr.lro_cap &&
1143                     config->dv_flow_en) {
1144                         /* TBD check tunnel lro caps. */
1145                         config->lro.supported = config->hca_attr.lro_cap;
1146                         DRV_LOG(DEBUG, "Device supports LRO");
1147                         /*
1148                          * If LRO timeout is not configured by application,
1149                          * use the minimal supported value.
1150                          */
1151                         if (!config->lro.timeout)
1152                                 config->lro.timeout =
1153                                 config->hca_attr.lro_timer_supported_periods[0];
1154                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1155                                 config->lro.timeout);
1156                         DRV_LOG(DEBUG, "LRO minimal size of TCP segment "
1157                                 "required for coalescing is %d bytes",
1158                                 config->hca_attr.lro_min_mss_size);
1159                 }
1160 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
1161                 if (config->hca_attr.qos.sup &&
1162                     config->hca_attr.qos.srtcm_sup &&
1163                     config->dv_flow_en) {
1164                         uint8_t reg_c_mask =
1165                                 config->hca_attr.qos.flow_meter_reg_c_ids;
1166                         /*
1167                          * Meter needs two REG_C's for color match and pre-sfx
1168                          * flow match. Here get the REG_C for color match.
1169                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
1170                          */
1171                         reg_c_mask &= 0xfc;
1172                         if (__builtin_popcount(reg_c_mask) < 1) {
1173                                 priv->mtr_en = 0;
1174                                 DRV_LOG(WARNING, "No available register for"
1175                                         " meter.");
1176                         } else {
1177                                 /*
1178                                  * The meter color register is used by the
1179                                  * flow-hit feature as well.
1180                                  * The flow-hit feature must use REG_C_3
1181                                  * Prefer REG_C_3 if it is available.
1182                                  */
1183                                 if (reg_c_mask & (1 << (REG_C_3 - REG_C_0)))
1184                                         priv->mtr_color_reg = REG_C_3;
1185                                 else
1186                                         priv->mtr_color_reg = ffs(reg_c_mask)
1187                                                               - 1 + REG_C_0;
1188                                 priv->mtr_en = 1;
1189                                 priv->mtr_reg_share =
1190                                       config->hca_attr.qos.flow_meter_reg_share;
1191                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1192                                         priv->mtr_color_reg);
1193                         }
1194                 }
1195 #endif
1196 #ifdef HAVE_MLX5_DR_CREATE_ACTION_ASO
1197                 if (config->hca_attr.flow_hit_aso &&
1198                     priv->mtr_color_reg == REG_C_3) {
1199                         sh->flow_hit_aso_en = 1;
1200                         err = mlx5_flow_aso_age_mng_init(sh);
1201                         if (err) {
1202                                 err = -err;
1203                                 goto error;
1204                         }
1205                         DRV_LOG(DEBUG, "Flow Hit ASO is supported.");
1206                 }
1207 #endif /* HAVE_MLX5_DR_CREATE_ACTION_ASO */
1208 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_SAMPLE)
1209                 if (config->hca_attr.log_max_ft_sampler_num > 0  &&
1210                     config->dv_flow_en) {
1211                         priv->sampler_en = 1;
1212                         DRV_LOG(DEBUG, "The Sampler enabled!\n");
1213                 } else {
1214                         priv->sampler_en = 0;
1215                         if (!config->hca_attr.log_max_ft_sampler_num)
1216                                 DRV_LOG(WARNING, "No available register for"
1217                                                 " Sampler.");
1218                         else
1219                                 DRV_LOG(DEBUG, "DV flow is not supported!\n");
1220                 }
1221 #endif
1222         }
1223         if (config->tx_pp) {
1224                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1225                         config->hca_attr.dev_freq_khz);
1226                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1227                         config->hca_attr.qos.packet_pacing ? "" : "not ");
1228                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1229                         config->hca_attr.cross_channel ? "" : "not ");
1230                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1231                         config->hca_attr.wqe_index_ignore ? "" : "not ");
1232                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1233                         config->hca_attr.non_wire_sq ? "" : "not ");
1234                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1235                         config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1236                         config->hca_attr.log_max_static_sq_wq);
1237                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1238                         config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1239                 if (!config->devx) {
1240                         DRV_LOG(ERR, "DevX is required for packet pacing");
1241                         err = ENODEV;
1242                         goto error;
1243                 }
1244                 if (!config->hca_attr.qos.packet_pacing) {
1245                         DRV_LOG(ERR, "Packet pacing is not supported");
1246                         err = ENODEV;
1247                         goto error;
1248                 }
1249                 if (!config->hca_attr.cross_channel) {
1250                         DRV_LOG(ERR, "Cross channel operations are"
1251                                      " required for packet pacing");
1252                         err = ENODEV;
1253                         goto error;
1254                 }
1255                 if (!config->hca_attr.wqe_index_ignore) {
1256                         DRV_LOG(ERR, "WQE index ignore feature is"
1257                                      " required for packet pacing");
1258                         err = ENODEV;
1259                         goto error;
1260                 }
1261                 if (!config->hca_attr.non_wire_sq) {
1262                         DRV_LOG(ERR, "Non-wire SQ feature is"
1263                                      " required for packet pacing");
1264                         err = ENODEV;
1265                         goto error;
1266                 }
1267                 if (!config->hca_attr.log_max_static_sq_wq) {
1268                         DRV_LOG(ERR, "Static WQE SQ feature is"
1269                                      " required for packet pacing");
1270                         err = ENODEV;
1271                         goto error;
1272                 }
1273                 if (!config->hca_attr.qos.wqe_rate_pp) {
1274                         DRV_LOG(ERR, "WQE rate mode is required"
1275                                      " for packet pacing");
1276                         err = ENODEV;
1277                         goto error;
1278                 }
1279 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1280                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1281                              " can't create queues for packet pacing");
1282                 err = ENODEV;
1283                 goto error;
1284 #endif
1285         }
1286         if (config->devx) {
1287                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1288
1289                 err = config->hca_attr.access_register_user ?
1290                         mlx5_devx_cmd_register_read
1291                                 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1292                                 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1293                 if (!err) {
1294                         uint32_t ts_mode;
1295
1296                         /* MTUTC register is read successfully. */
1297                         ts_mode = MLX5_GET(register_mtutc, reg,
1298                                            time_stamp_mode);
1299                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1300                                 config->rt_timestamp = 1;
1301                 } else {
1302                         /* Kernel does not support register reading. */
1303                         if (config->hca_attr.dev_freq_khz ==
1304                                                  (NS_PER_S / MS_PER_S))
1305                                 config->rt_timestamp = 1;
1306                 }
1307         }
1308         /*
1309          * If HW has bug working with tunnel packet decapsulation and
1310          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1311          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1312          */
1313         if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1314                 config->hw_fcs_strip = 0;
1315         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1316                 (config->hw_fcs_strip ? "" : "not "));
1317         if (config->mprq.enabled && mprq) {
1318                 if (config->mprq.stride_num_n &&
1319                     (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1320                      config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1321                         config->mprq.stride_num_n =
1322                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1323                                                 mprq_min_stride_num_n),
1324                                         mprq_max_stride_num_n);
1325                         DRV_LOG(WARNING,
1326                                 "the number of strides"
1327                                 " for Multi-Packet RQ is out of range,"
1328                                 " setting default value (%u)",
1329                                 1 << config->mprq.stride_num_n);
1330                 }
1331                 if (config->mprq.stride_size_n &&
1332                     (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1333                      config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1334                         config->mprq.stride_size_n =
1335                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1336                                                 mprq_min_stride_size_n),
1337                                         mprq_max_stride_size_n);
1338                         DRV_LOG(WARNING,
1339                                 "the size of a stride"
1340                                 " for Multi-Packet RQ is out of range,"
1341                                 " setting default value (%u)",
1342                                 1 << config->mprq.stride_size_n);
1343                 }
1344                 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1345                 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1346         } else if (config->mprq.enabled && !mprq) {
1347                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1348                 config->mprq.enabled = 0;
1349         }
1350         if (config->max_dump_files_num == 0)
1351                 config->max_dump_files_num = 128;
1352         eth_dev = rte_eth_dev_allocate(name);
1353         if (eth_dev == NULL) {
1354                 DRV_LOG(ERR, "can not allocate rte ethdev");
1355                 err = ENOMEM;
1356                 goto error;
1357         }
1358         if (priv->representor) {
1359                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1360                 eth_dev->data->representor_id = priv->representor_id;
1361         }
1362         /*
1363          * Store associated network device interface index. This index
1364          * is permanent throughout the lifetime of device. So, we may store
1365          * the ifindex here and use the cached value further.
1366          */
1367         MLX5_ASSERT(spawn->ifindex);
1368         priv->if_index = spawn->ifindex;
1369         if (priv->pf_bond >= 0 && priv->master) {
1370                 /* Get bond interface info */
1371                 err = mlx5_sysfs_bond_info(priv->if_index,
1372                                      &priv->bond_ifindex,
1373                                      priv->bond_name);
1374                 if (err)
1375                         DRV_LOG(ERR, "unable to get bond info: %s",
1376                                 strerror(rte_errno));
1377                 else
1378                         DRV_LOG(INFO, "PF device %u, bond device %u(%s)",
1379                                 priv->if_index, priv->bond_ifindex,
1380                                 priv->bond_name);
1381         }
1382         eth_dev->data->dev_private = priv;
1383         priv->dev_data = eth_dev->data;
1384         eth_dev->data->mac_addrs = priv->mac;
1385         if (spawn->pf_bond < 0) {
1386                 eth_dev->device = dpdk_dev;
1387         } else {
1388                 /* Use primary bond PCI as device. */
1389                 if (sh->bond_dev == UINT16_MAX) {
1390                         sh->bond_dev = eth_dev->data->port_id;
1391                         eth_dev->device = dpdk_dev;
1392                 } else {
1393                         eth_dev->device = rte_eth_devices[sh->bond_dev].device;
1394                 }
1395         }
1396         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1397         /* Configure the first MAC address by default. */
1398         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1399                 DRV_LOG(ERR,
1400                         "port %u cannot get MAC address, is mlx5_en"
1401                         " loaded? (errno: %s)",
1402                         eth_dev->data->port_id, strerror(rte_errno));
1403                 err = ENODEV;
1404                 goto error;
1405         }
1406         DRV_LOG(INFO,
1407                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1408                 eth_dev->data->port_id,
1409                 mac.addr_bytes[0], mac.addr_bytes[1],
1410                 mac.addr_bytes[2], mac.addr_bytes[3],
1411                 mac.addr_bytes[4], mac.addr_bytes[5]);
1412 #ifdef RTE_LIBRTE_MLX5_DEBUG
1413         {
1414                 char ifname[IF_NAMESIZE];
1415
1416                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1417                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1418                                 eth_dev->data->port_id, ifname);
1419                 else
1420                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1421                                 eth_dev->data->port_id);
1422         }
1423 #endif
1424         /* Get actual MTU if possible. */
1425         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1426         if (err) {
1427                 err = rte_errno;
1428                 goto error;
1429         }
1430         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1431                 priv->mtu);
1432         /* Initialize burst functions to prevent crashes before link-up. */
1433         eth_dev->rx_pkt_burst = removed_rx_burst;
1434         eth_dev->tx_pkt_burst = removed_tx_burst;
1435         eth_dev->dev_ops = &mlx5_dev_ops;
1436         eth_dev->rx_descriptor_status = mlx5_rx_descriptor_status;
1437         eth_dev->tx_descriptor_status = mlx5_tx_descriptor_status;
1438         eth_dev->rx_queue_count = mlx5_rx_queue_count;
1439         /* Register MAC address. */
1440         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1441         if (config->vf && config->vf_nl_en)
1442                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1443                                       mlx5_ifindex(eth_dev),
1444                                       eth_dev->data->mac_addrs,
1445                                       MLX5_MAX_MAC_ADDRESSES);
1446         priv->flows = 0;
1447         priv->ctrl_flows = 0;
1448         rte_spinlock_init(&priv->flow_list_lock);
1449         TAILQ_INIT(&priv->flow_meters);
1450         TAILQ_INIT(&priv->flow_meter_profiles);
1451         /* Hint libmlx5 to use PMD allocator for data plane resources */
1452         mlx5_glue->dv_set_context_attr(sh->ctx,
1453                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1454                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1455                                 .alloc = &mlx5_alloc_verbs_buf,
1456                                 .free = &mlx5_free_verbs_buf,
1457                                 .data = sh,
1458                         }));
1459         /* Bring Ethernet device up. */
1460         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1461                 eth_dev->data->port_id);
1462         mlx5_set_link_up(eth_dev);
1463         /*
1464          * Even though the interrupt handler is not installed yet,
1465          * interrupts will still trigger on the async_fd from
1466          * Verbs context returned by ibv_open_device().
1467          */
1468         mlx5_link_update(eth_dev, 0);
1469 #ifdef HAVE_MLX5DV_DR_ESWITCH
1470         if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1471               (switch_info->representor || switch_info->master)))
1472                 config->dv_esw_en = 0;
1473 #else
1474         config->dv_esw_en = 0;
1475 #endif
1476         /* Detect minimal data bytes to inline. */
1477         mlx5_set_min_inline(spawn, config);
1478         /* Store device configuration on private structure. */
1479         priv->config = *config;
1480         /* Create context for virtual machine VLAN workaround. */
1481         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1482         if (config->dv_flow_en) {
1483                 err = mlx5_alloc_shared_dr(priv);
1484                 if (err)
1485                         goto error;
1486         }
1487         if (config->devx && config->dv_flow_en && config->dest_tir) {
1488                 priv->obj_ops = devx_obj_ops;
1489                 priv->obj_ops.drop_action_create =
1490                                                 ibv_obj_ops.drop_action_create;
1491                 priv->obj_ops.drop_action_destroy =
1492                                                 ibv_obj_ops.drop_action_destroy;
1493 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1494                 priv->obj_ops.txq_obj_modify = ibv_obj_ops.txq_obj_modify;
1495 #else
1496                 if (config->dv_esw_en)
1497                         priv->obj_ops.txq_obj_modify =
1498                                                 ibv_obj_ops.txq_obj_modify;
1499 #endif
1500                 /* Use specific wrappers for Tx object. */
1501                 priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
1502                 priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
1503
1504         } else {
1505                 priv->obj_ops = ibv_obj_ops;
1506         }
1507         priv->drop_queue.hrxq = mlx5_drop_action_create(eth_dev);
1508         if (!priv->drop_queue.hrxq)
1509                 goto error;
1510         /* Supported Verbs flow priority number detection. */
1511         err = mlx5_flow_discover_priorities(eth_dev);
1512         if (err < 0) {
1513                 err = -err;
1514                 goto error;
1515         }
1516         priv->config.flow_prio = err;
1517         if (!priv->config.dv_esw_en &&
1518             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1519                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1520                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1521                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1522         }
1523         mlx5_set_metadata_mask(eth_dev);
1524         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1525             !priv->sh->dv_regc0_mask) {
1526                 DRV_LOG(ERR, "metadata mode %u is not supported "
1527                              "(no metadata reg_c[0] is available)",
1528                              priv->config.dv_xmeta_en);
1529                         err = ENOTSUP;
1530                         goto error;
1531         }
1532         mlx5_cache_list_init(&priv->hrxqs, "hrxq", 0, eth_dev,
1533                              mlx5_hrxq_create_cb,
1534                              mlx5_hrxq_match_cb,
1535                              mlx5_hrxq_remove_cb);
1536         /* Query availability of metadata reg_c's. */
1537         err = mlx5_flow_discover_mreg_c(eth_dev);
1538         if (err < 0) {
1539                 err = -err;
1540                 goto error;
1541         }
1542         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1543                 DRV_LOG(DEBUG,
1544                         "port %u extensive metadata register is not supported",
1545                         eth_dev->data->port_id);
1546                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1547                         DRV_LOG(ERR, "metadata mode %u is not supported "
1548                                      "(no metadata registers available)",
1549                                      priv->config.dv_xmeta_en);
1550                         err = ENOTSUP;
1551                         goto error;
1552                 }
1553         }
1554         if (priv->config.dv_flow_en &&
1555             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1556             mlx5_flow_ext_mreg_supported(eth_dev) &&
1557             priv->sh->dv_regc0_mask) {
1558                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1559                                                       MLX5_FLOW_MREG_HTABLE_SZ,
1560                                                       0, 0,
1561                                                       flow_dv_mreg_create_cb,
1562                                                       flow_dv_mreg_match_cb,
1563                                                       flow_dv_mreg_remove_cb);
1564                 if (!priv->mreg_cp_tbl) {
1565                         err = ENOMEM;
1566                         goto error;
1567                 }
1568                 priv->mreg_cp_tbl->ctx = eth_dev;
1569         }
1570         rte_spinlock_init(&priv->shared_act_sl);
1571         mlx5_flow_counter_mode_config(eth_dev);
1572         if (priv->config.dv_flow_en)
1573                 eth_dev->data->dev_flags |= RTE_ETH_DEV_FLOW_OPS_THREAD_SAFE;
1574         return eth_dev;
1575 error:
1576         if (priv) {
1577                 if (priv->mreg_cp_tbl)
1578                         mlx5_hlist_destroy(priv->mreg_cp_tbl);
1579                 if (priv->sh)
1580                         mlx5_os_free_shared_dr(priv);
1581                 if (priv->nl_socket_route >= 0)
1582                         close(priv->nl_socket_route);
1583                 if (priv->nl_socket_rdma >= 0)
1584                         close(priv->nl_socket_rdma);
1585                 if (priv->vmwa_context)
1586                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1587                 if (eth_dev && priv->drop_queue.hrxq)
1588                         mlx5_drop_action_destroy(eth_dev);
1589                 if (own_domain_id)
1590                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1591                 mlx5_cache_list_destroy(&priv->hrxqs);
1592                 mlx5_free(priv);
1593                 if (eth_dev != NULL)
1594                         eth_dev->data->dev_private = NULL;
1595         }
1596         if (eth_dev != NULL) {
1597                 /* mac_addrs must not be freed alone because part of
1598                  * dev_private
1599                  **/
1600                 eth_dev->data->mac_addrs = NULL;
1601                 rte_eth_dev_release_port(eth_dev);
1602         }
1603         if (sh)
1604                 mlx5_free_shared_dev_ctx(sh);
1605         MLX5_ASSERT(err > 0);
1606         rte_errno = err;
1607         return NULL;
1608 }
1609
1610 /**
1611  * Comparison callback to sort device data.
1612  *
1613  * This is meant to be used with qsort().
1614  *
1615  * @param a[in]
1616  *   Pointer to pointer to first data object.
1617  * @param b[in]
1618  *   Pointer to pointer to second data object.
1619  *
1620  * @return
1621  *   0 if both objects are equal, less than 0 if the first argument is less
1622  *   than the second, greater than 0 otherwise.
1623  */
1624 static int
1625 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1626 {
1627         const struct mlx5_switch_info *si_a =
1628                 &((const struct mlx5_dev_spawn_data *)a)->info;
1629         const struct mlx5_switch_info *si_b =
1630                 &((const struct mlx5_dev_spawn_data *)b)->info;
1631         int ret;
1632
1633         /* Master device first. */
1634         ret = si_b->master - si_a->master;
1635         if (ret)
1636                 return ret;
1637         /* Then representor devices. */
1638         ret = si_b->representor - si_a->representor;
1639         if (ret)
1640                 return ret;
1641         /* Unidentified devices come last in no specific order. */
1642         if (!si_a->representor)
1643                 return 0;
1644         /* Order representors by name. */
1645         return si_a->port_name - si_b->port_name;
1646 }
1647
1648 /**
1649  * Match PCI information for possible slaves of bonding device.
1650  *
1651  * @param[in] ibv_dev
1652  *   Pointer to Infiniband device structure.
1653  * @param[in] pci_dev
1654  *   Pointer to PCI device structure to match PCI address.
1655  * @param[in] nl_rdma
1656  *   Netlink RDMA group socket handle.
1657  *
1658  * @return
1659  *   negative value if no bonding device found, otherwise
1660  *   positive index of slave PF in bonding.
1661  */
1662 static int
1663 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1664                            const struct rte_pci_device *pci_dev,
1665                            int nl_rdma)
1666 {
1667         char ifname[IF_NAMESIZE + 1];
1668         unsigned int ifindex;
1669         unsigned int np, i;
1670         FILE *file = NULL;
1671         int pf = -1;
1672
1673         /*
1674          * Try to get master device name. If something goes
1675          * wrong suppose the lack of kernel support and no
1676          * bonding devices.
1677          */
1678         if (nl_rdma < 0)
1679                 return -1;
1680         if (!strstr(ibv_dev->name, "bond"))
1681                 return -1;
1682         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1683         if (!np)
1684                 return -1;
1685         /*
1686          * The Master device might not be on the predefined
1687          * port (not on port index 1, it is not garanted),
1688          * we have to scan all Infiniband device port and
1689          * find master.
1690          */
1691         for (i = 1; i <= np; ++i) {
1692                 /* Check whether Infiniband port is populated. */
1693                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1694                 if (!ifindex)
1695                         continue;
1696                 if (!if_indextoname(ifindex, ifname))
1697                         continue;
1698                 /* Try to read bonding slave names from sysfs. */
1699                 MKSTR(slaves,
1700                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1701                 file = fopen(slaves, "r");
1702                 if (file)
1703                         break;
1704         }
1705         if (!file)
1706                 return -1;
1707         /* Use safe format to check maximal buffer length. */
1708         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1709         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1710                 char tmp_str[IF_NAMESIZE + 32];
1711                 struct rte_pci_addr pci_addr;
1712                 struct mlx5_switch_info info;
1713
1714                 /* Process slave interface names in the loop. */
1715                 snprintf(tmp_str, sizeof(tmp_str),
1716                          "/sys/class/net/%s", ifname);
1717                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1718                         DRV_LOG(WARNING, "can not get PCI address"
1719                                          " for netdev \"%s\"", ifname);
1720                         continue;
1721                 }
1722                 if (pci_dev->addr.domain != pci_addr.domain ||
1723                     pci_dev->addr.bus != pci_addr.bus ||
1724                     pci_dev->addr.devid != pci_addr.devid ||
1725                     pci_dev->addr.function != pci_addr.function)
1726                         continue;
1727                 /* Slave interface PCI address match found. */
1728                 fclose(file);
1729                 snprintf(tmp_str, sizeof(tmp_str),
1730                          "/sys/class/net/%s/phys_port_name", ifname);
1731                 file = fopen(tmp_str, "rb");
1732                 if (!file)
1733                         break;
1734                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1735                 if (fscanf(file, "%32s", tmp_str) == 1)
1736                         mlx5_translate_port_name(tmp_str, &info);
1737                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1738                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1739                         pf = info.port_name;
1740                 break;
1741         }
1742         if (file)
1743                 fclose(file);
1744         return pf;
1745 }
1746
1747 /**
1748  * DPDK callback to register a PCI device.
1749  *
1750  * This function spawns Ethernet devices out of a given PCI device.
1751  *
1752  * @param[in] pci_drv
1753  *   PCI driver structure (mlx5_driver).
1754  * @param[in] pci_dev
1755  *   PCI device information.
1756  *
1757  * @return
1758  *   0 on success, a negative errno value otherwise and rte_errno is set.
1759  */
1760 int
1761 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1762                   struct rte_pci_device *pci_dev)
1763 {
1764         struct ibv_device **ibv_list;
1765         /*
1766          * Number of found IB Devices matching with requested PCI BDF.
1767          * nd != 1 means there are multiple IB devices over the same
1768          * PCI device and we have representors and master.
1769          */
1770         unsigned int nd = 0;
1771         /*
1772          * Number of found IB device Ports. nd = 1 and np = 1..n means
1773          * we have the single multiport IB device, and there may be
1774          * representors attached to some of found ports.
1775          */
1776         unsigned int np = 0;
1777         /*
1778          * Number of DPDK ethernet devices to Spawn - either over
1779          * multiple IB devices or multiple ports of single IB device.
1780          * Actually this is the number of iterations to spawn.
1781          */
1782         unsigned int ns = 0;
1783         /*
1784          * Bonding device
1785          *   < 0 - no bonding device (single one)
1786          *  >= 0 - bonding device (value is slave PF index)
1787          */
1788         int bd = -1;
1789         struct mlx5_dev_spawn_data *list = NULL;
1790         struct mlx5_dev_config dev_config;
1791         unsigned int dev_config_vf;
1792         int ret;
1793
1794         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1795                 mlx5_pmd_socket_init();
1796         ret = mlx5_init_once();
1797         if (ret) {
1798                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1799                         strerror(rte_errno));
1800                 return -rte_errno;
1801         }
1802         errno = 0;
1803         ibv_list = mlx5_glue->get_device_list(&ret);
1804         if (!ibv_list) {
1805                 rte_errno = errno ? errno : ENOSYS;
1806                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1807                 return -rte_errno;
1808         }
1809         /*
1810          * First scan the list of all Infiniband devices to find
1811          * matching ones, gathering into the list.
1812          */
1813         struct ibv_device *ibv_match[ret + 1];
1814         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1815         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1816         unsigned int i;
1817
1818         while (ret-- > 0) {
1819                 struct rte_pci_addr pci_addr;
1820
1821                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1822                 bd = mlx5_device_bond_pci_match
1823                                 (ibv_list[ret], pci_dev, nl_rdma);
1824                 if (bd >= 0) {
1825                         /*
1826                          * Bonding device detected. Only one match is allowed,
1827                          * the bonding is supported over multi-port IB device,
1828                          * there should be no matches on representor PCI
1829                          * functions or non VF LAG bonding devices with
1830                          * specified address.
1831                          */
1832                         if (nd) {
1833                                 DRV_LOG(ERR,
1834                                         "multiple PCI match on bonding device"
1835                                         "\"%s\" found", ibv_list[ret]->name);
1836                                 rte_errno = ENOENT;
1837                                 ret = -rte_errno;
1838                                 goto exit;
1839                         }
1840                         DRV_LOG(INFO, "PCI information matches for"
1841                                       " slave %d bonding device \"%s\"",
1842                                       bd, ibv_list[ret]->name);
1843                         ibv_match[nd++] = ibv_list[ret];
1844                         break;
1845                 }
1846                 if (mlx5_dev_to_pci_addr
1847                         (ibv_list[ret]->ibdev_path, &pci_addr))
1848                         continue;
1849                 if (pci_dev->addr.domain != pci_addr.domain ||
1850                     pci_dev->addr.bus != pci_addr.bus ||
1851                     pci_dev->addr.devid != pci_addr.devid ||
1852                     pci_dev->addr.function != pci_addr.function)
1853                         continue;
1854                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1855                         ibv_list[ret]->name);
1856                 ibv_match[nd++] = ibv_list[ret];
1857         }
1858         ibv_match[nd] = NULL;
1859         if (!nd) {
1860                 /* No device matches, just complain and bail out. */
1861                 DRV_LOG(WARNING,
1862                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1863                         " are kernel drivers loaded?",
1864                         pci_dev->addr.domain, pci_dev->addr.bus,
1865                         pci_dev->addr.devid, pci_dev->addr.function);
1866                 rte_errno = ENOENT;
1867                 ret = -rte_errno;
1868                 goto exit;
1869         }
1870         if (nd == 1) {
1871                 /*
1872                  * Found single matching device may have multiple ports.
1873                  * Each port may be representor, we have to check the port
1874                  * number and check the representors existence.
1875                  */
1876                 if (nl_rdma >= 0)
1877                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1878                 if (!np)
1879                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1880                                          " ports number", ibv_match[0]->name);
1881                 if (bd >= 0 && !np) {
1882                         DRV_LOG(ERR, "can not get ports"
1883                                      " for bonding device");
1884                         rte_errno = ENOENT;
1885                         ret = -rte_errno;
1886                         goto exit;
1887                 }
1888         }
1889 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1890         if (bd >= 0) {
1891                 /*
1892                  * This may happen if there is VF LAG kernel support and
1893                  * application is compiled with older rdma_core library.
1894                  */
1895                 DRV_LOG(ERR,
1896                         "No kernel/verbs support for VF LAG bonding found.");
1897                 rte_errno = ENOTSUP;
1898                 ret = -rte_errno;
1899                 goto exit;
1900         }
1901 #endif
1902         /*
1903          * Now we can determine the maximal
1904          * amount of devices to be spawned.
1905          */
1906         list = mlx5_malloc(MLX5_MEM_ZERO,
1907                            sizeof(struct mlx5_dev_spawn_data) *
1908                            (np ? np : nd),
1909                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1910         if (!list) {
1911                 DRV_LOG(ERR, "spawn data array allocation failure");
1912                 rte_errno = ENOMEM;
1913                 ret = -rte_errno;
1914                 goto exit;
1915         }
1916         if (bd >= 0 || np > 1) {
1917                 /*
1918                  * Single IB device with multiple ports found,
1919                  * it may be E-Switch master device and representors.
1920                  * We have to perform identification through the ports.
1921                  */
1922                 MLX5_ASSERT(nl_rdma >= 0);
1923                 MLX5_ASSERT(ns == 0);
1924                 MLX5_ASSERT(nd == 1);
1925                 MLX5_ASSERT(np);
1926                 for (i = 1; i <= np; ++i) {
1927                         list[ns].max_port = np;
1928                         list[ns].phys_port = i;
1929                         list[ns].phys_dev = ibv_match[0];
1930                         list[ns].eth_dev = NULL;
1931                         list[ns].pci_dev = pci_dev;
1932                         list[ns].pf_bond = bd;
1933                         list[ns].ifindex = mlx5_nl_ifindex
1934                                 (nl_rdma,
1935                                 mlx5_os_get_dev_device_name
1936                                                 (list[ns].phys_dev), i);
1937                         if (!list[ns].ifindex) {
1938                                 /*
1939                                  * No network interface index found for the
1940                                  * specified port, it means there is no
1941                                  * representor on this port. It's OK,
1942                                  * there can be disabled ports, for example
1943                                  * if sriov_numvfs < sriov_totalvfs.
1944                                  */
1945                                 continue;
1946                         }
1947                         ret = -1;
1948                         if (nl_route >= 0)
1949                                 ret = mlx5_nl_switch_info
1950                                                (nl_route,
1951                                                 list[ns].ifindex,
1952                                                 &list[ns].info);
1953                         if (ret || (!list[ns].info.representor &&
1954                                     !list[ns].info.master)) {
1955                                 /*
1956                                  * We failed to recognize representors with
1957                                  * Netlink, let's try to perform the task
1958                                  * with sysfs.
1959                                  */
1960                                 ret =  mlx5_sysfs_switch_info
1961                                                 (list[ns].ifindex,
1962                                                  &list[ns].info);
1963                         }
1964 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
1965                         if (!ret && bd >= 0) {
1966                                 switch (list[ns].info.name_type) {
1967                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1968                                         if (list[ns].info.port_name == bd)
1969                                                 ns++;
1970                                         break;
1971                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1972                                         /* Fallthrough */
1973                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1974                                         if (list[ns].info.pf_num == bd)
1975                                                 ns++;
1976                                         break;
1977                                 default:
1978                                         break;
1979                                 }
1980                                 continue;
1981                         }
1982 #endif
1983                         if (!ret && (list[ns].info.representor ^
1984                                      list[ns].info.master))
1985                                 ns++;
1986                 }
1987                 if (!ns) {
1988                         DRV_LOG(ERR,
1989                                 "unable to recognize master/representors"
1990                                 " on the IB device with multiple ports");
1991                         rte_errno = ENOENT;
1992                         ret = -rte_errno;
1993                         goto exit;
1994                 }
1995         } else {
1996                 /*
1997                  * The existence of several matching entries (nd > 1) means
1998                  * port representors have been instantiated. No existing Verbs
1999                  * call nor sysfs entries can tell them apart, this can only
2000                  * be done through Netlink calls assuming kernel drivers are
2001                  * recent enough to support them.
2002                  *
2003                  * In the event of identification failure through Netlink,
2004                  * try again through sysfs, then:
2005                  *
2006                  * 1. A single IB device matches (nd == 1) with single
2007                  *    port (np=0/1) and is not a representor, assume
2008                  *    no switch support.
2009                  *
2010                  * 2. Otherwise no safe assumptions can be made;
2011                  *    complain louder and bail out.
2012                  */
2013                 for (i = 0; i != nd; ++i) {
2014                         memset(&list[ns].info, 0, sizeof(list[ns].info));
2015                         list[ns].max_port = 1;
2016                         list[ns].phys_port = 1;
2017                         list[ns].phys_dev = ibv_match[i];
2018                         list[ns].eth_dev = NULL;
2019                         list[ns].pci_dev = pci_dev;
2020                         list[ns].pf_bond = -1;
2021                         list[ns].ifindex = 0;
2022                         if (nl_rdma >= 0)
2023                                 list[ns].ifindex = mlx5_nl_ifindex
2024                                 (nl_rdma,
2025                                 mlx5_os_get_dev_device_name
2026                                                 (list[ns].phys_dev), 1);
2027                         if (!list[ns].ifindex) {
2028                                 char ifname[IF_NAMESIZE];
2029
2030                                 /*
2031                                  * Netlink failed, it may happen with old
2032                                  * ib_core kernel driver (before 4.16).
2033                                  * We can assume there is old driver because
2034                                  * here we are processing single ports IB
2035                                  * devices. Let's try sysfs to retrieve
2036                                  * the ifindex. The method works for
2037                                  * master device only.
2038                                  */
2039                                 if (nd > 1) {
2040                                         /*
2041                                          * Multiple devices found, assume
2042                                          * representors, can not distinguish
2043                                          * master/representor and retrieve
2044                                          * ifindex via sysfs.
2045                                          */
2046                                         continue;
2047                                 }
2048                                 ret = mlx5_get_ifname_sysfs
2049                                         (ibv_match[i]->ibdev_path, ifname);
2050                                 if (!ret)
2051                                         list[ns].ifindex =
2052                                                 if_nametoindex(ifname);
2053                                 if (!list[ns].ifindex) {
2054                                         /*
2055                                          * No network interface index found
2056                                          * for the specified device, it means
2057                                          * there it is neither representor
2058                                          * nor master.
2059                                          */
2060                                         continue;
2061                                 }
2062                         }
2063                         ret = -1;
2064                         if (nl_route >= 0)
2065                                 ret = mlx5_nl_switch_info
2066                                                (nl_route,
2067                                                 list[ns].ifindex,
2068                                                 &list[ns].info);
2069                         if (ret || (!list[ns].info.representor &&
2070                                     !list[ns].info.master)) {
2071                                 /*
2072                                  * We failed to recognize representors with
2073                                  * Netlink, let's try to perform the task
2074                                  * with sysfs.
2075                                  */
2076                                 ret =  mlx5_sysfs_switch_info
2077                                                 (list[ns].ifindex,
2078                                                  &list[ns].info);
2079                         }
2080                         if (!ret && (list[ns].info.representor ^
2081                                      list[ns].info.master)) {
2082                                 ns++;
2083                         } else if ((nd == 1) &&
2084                                    !list[ns].info.representor &&
2085                                    !list[ns].info.master) {
2086                                 /*
2087                                  * Single IB device with
2088                                  * one physical port and
2089                                  * attached network device.
2090                                  * May be SRIOV is not enabled
2091                                  * or there is no representors.
2092                                  */
2093                                 DRV_LOG(INFO, "no E-Switch support detected");
2094                                 ns++;
2095                                 break;
2096                         }
2097                 }
2098                 if (!ns) {
2099                         DRV_LOG(ERR,
2100                                 "unable to recognize master/representors"
2101                                 " on the multiple IB devices");
2102                         rte_errno = ENOENT;
2103                         ret = -rte_errno;
2104                         goto exit;
2105                 }
2106         }
2107         MLX5_ASSERT(ns);
2108         /*
2109          * Sort list to probe devices in natural order for users convenience
2110          * (i.e. master first, then representors from lowest to highest ID).
2111          */
2112         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2113         /* Device specific configuration. */
2114         switch (pci_dev->id.device_id) {
2115         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2116         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2117         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2118         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2119         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2120         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2121         case PCI_DEVICE_ID_MELLANOX_CONNECTXVF:
2122                 dev_config_vf = 1;
2123                 break;
2124         default:
2125                 dev_config_vf = 0;
2126                 break;
2127         }
2128         for (i = 0; i != ns; ++i) {
2129                 uint32_t restore;
2130
2131                 /* Default configuration. */
2132                 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
2133                 dev_config.vf = dev_config_vf;
2134                 dev_config.mps = MLX5_ARG_UNSET;
2135                 dev_config.dbnc = MLX5_ARG_UNSET;
2136                 dev_config.rx_vec_en = 1;
2137                 dev_config.txq_inline_max = MLX5_ARG_UNSET;
2138                 dev_config.txq_inline_min = MLX5_ARG_UNSET;
2139                 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
2140                 dev_config.txqs_inline = MLX5_ARG_UNSET;
2141                 dev_config.vf_nl_en = 1;
2142                 dev_config.mr_ext_memseg_en = 1;
2143                 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
2144                 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
2145                 dev_config.dv_esw_en = 1;
2146                 dev_config.dv_flow_en = 1;
2147                 dev_config.decap_en = 1;
2148                 dev_config.log_hp_size = MLX5_ARG_UNSET;
2149                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2150                                                  &list[i],
2151                                                  &dev_config);
2152                 if (!list[i].eth_dev) {
2153                         if (rte_errno != EBUSY && rte_errno != EEXIST)
2154                                 break;
2155                         /* Device is disabled or already spawned. Ignore it. */
2156                         continue;
2157                 }
2158                 restore = list[i].eth_dev->data->dev_flags;
2159                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2160                 /* Restore non-PCI flags cleared by the above call. */
2161                 list[i].eth_dev->data->dev_flags |= restore;
2162                 rte_eth_dev_probing_finish(list[i].eth_dev);
2163         }
2164         if (i != ns) {
2165                 DRV_LOG(ERR,
2166                         "probe of PCI device " PCI_PRI_FMT " aborted after"
2167                         " encountering an error: %s",
2168                         pci_dev->addr.domain, pci_dev->addr.bus,
2169                         pci_dev->addr.devid, pci_dev->addr.function,
2170                         strerror(rte_errno));
2171                 ret = -rte_errno;
2172                 /* Roll back. */
2173                 while (i--) {
2174                         if (!list[i].eth_dev)
2175                                 continue;
2176                         mlx5_dev_close(list[i].eth_dev);
2177                         /* mac_addrs must not be freed because in dev_private */
2178                         list[i].eth_dev->data->mac_addrs = NULL;
2179                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2180                 }
2181                 /* Restore original error. */
2182                 rte_errno = -ret;
2183         } else {
2184                 ret = 0;
2185         }
2186 exit:
2187         /*
2188          * Do the routine cleanup:
2189          * - close opened Netlink sockets
2190          * - free allocated spawn data array
2191          * - free the Infiniband device list
2192          */
2193         if (nl_rdma >= 0)
2194                 close(nl_rdma);
2195         if (nl_route >= 0)
2196                 close(nl_route);
2197         if (list)
2198                 mlx5_free(list);
2199         MLX5_ASSERT(ibv_list);
2200         mlx5_glue->free_device_list(ibv_list);
2201         return ret;
2202 }
2203
2204 static int
2205 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
2206 {
2207         char *env;
2208         int value;
2209
2210         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2211         /* Get environment variable to store. */
2212         env = getenv(MLX5_SHUT_UP_BF);
2213         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
2214         if (config->dbnc == MLX5_ARG_UNSET)
2215                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
2216         else
2217                 setenv(MLX5_SHUT_UP_BF,
2218                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
2219         return value;
2220 }
2221
2222 static void
2223 mlx5_restore_doorbell_mapping_env(int value)
2224 {
2225         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
2226         /* Restore the original environment variable state. */
2227         if (value == MLX5_ARG_UNSET)
2228                 unsetenv(MLX5_SHUT_UP_BF);
2229         else
2230                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
2231 }
2232
2233 /**
2234  * Extract pdn of PD object using DV API.
2235  *
2236  * @param[in] pd
2237  *   Pointer to the verbs PD object.
2238  * @param[out] pdn
2239  *   Pointer to the PD object number variable.
2240  *
2241  * @return
2242  *   0 on success, error value otherwise.
2243  */
2244 int
2245 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
2246 {
2247 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2248         struct mlx5dv_obj obj;
2249         struct mlx5dv_pd pd_info;
2250         int ret = 0;
2251
2252         obj.pd.in = pd;
2253         obj.pd.out = &pd_info;
2254         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2255         if (ret) {
2256                 DRV_LOG(DEBUG, "Fail to get PD object info");
2257                 return ret;
2258         }
2259         *pdn = pd_info.pdn;
2260         return 0;
2261 #else
2262         (void)pd;
2263         (void)pdn;
2264         return -ENOTSUP;
2265 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2266 }
2267
2268 /**
2269  * Function API to open IB device.
2270  *
2271  * This function calls the Linux glue APIs to open a device.
2272  *
2273  * @param[in] spawn
2274  *   Pointer to the IB device attributes (name, port, etc).
2275  * @param[out] config
2276  *   Pointer to device configuration structure.
2277  * @param[out] sh
2278  *   Pointer to shared context structure.
2279  *
2280  * @return
2281  *   0 on success, a positive error value otherwise.
2282  */
2283 int
2284 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2285                      const struct mlx5_dev_config *config,
2286                      struct mlx5_dev_ctx_shared *sh)
2287 {
2288         int dbmap_env;
2289         int err = 0;
2290
2291         sh->numa_node = spawn->pci_dev->device.numa_node;
2292         pthread_mutex_init(&sh->txpp.mutex, NULL);
2293         /*
2294          * Configure environment variable "MLX5_BF_SHUT_UP"
2295          * before the device creation. The rdma_core library
2296          * checks the variable at device creation and
2297          * stores the result internally.
2298          */
2299         dbmap_env = mlx5_config_doorbell_mapping_env(config);
2300         /* Try to open IB device with DV first, then usual Verbs. */
2301         errno = 0;
2302         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2303         if (sh->ctx) {
2304                 sh->devx = 1;
2305                 DRV_LOG(DEBUG, "DevX is supported");
2306                 /* The device is created, no need for environment. */
2307                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2308         } else {
2309                 /* The environment variable is still configured. */
2310                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2311                 err = errno ? errno : ENODEV;
2312                 /*
2313                  * The environment variable is not needed anymore,
2314                  * all device creation attempts are completed.
2315                  */
2316                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2317                 if (!sh->ctx)
2318                         return err;
2319                 DRV_LOG(DEBUG, "DevX is NOT supported");
2320                 err = 0;
2321         }
2322         if (!err && sh->ctx) {
2323                 /* Hint libmlx5 to use PMD allocator for data plane resources */
2324                 mlx5_glue->dv_set_context_attr(sh->ctx,
2325                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2326                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
2327                                 .alloc = &mlx5_alloc_verbs_buf,
2328                                 .free = &mlx5_free_verbs_buf,
2329                                 .data = sh,
2330                         }));
2331         }
2332         return err;
2333 }
2334
2335 /**
2336  * Install shared asynchronous device events handler.
2337  * This function is implemented to support event sharing
2338  * between multiple ports of single IB device.
2339  *
2340  * @param sh
2341  *   Pointer to mlx5_dev_ctx_shared object.
2342  */
2343 void
2344 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2345 {
2346         int ret;
2347         int flags;
2348
2349         sh->intr_handle.fd = -1;
2350         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2351         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2352                     F_SETFL, flags | O_NONBLOCK);
2353         if (ret) {
2354                 DRV_LOG(INFO, "failed to change file descriptor async event"
2355                         " queue");
2356         } else {
2357                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2358                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2359                 if (rte_intr_callback_register(&sh->intr_handle,
2360                                         mlx5_dev_interrupt_handler, sh)) {
2361                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
2362                         sh->intr_handle.fd = -1;
2363                 }
2364         }
2365         if (sh->devx) {
2366 #ifdef HAVE_IBV_DEVX_ASYNC
2367                 sh->intr_handle_devx.fd = -1;
2368                 sh->devx_comp =
2369                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2370                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2371                 if (!devx_comp) {
2372                         DRV_LOG(INFO, "failed to allocate devx_comp.");
2373                         return;
2374                 }
2375                 flags = fcntl(devx_comp->fd, F_GETFL);
2376                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2377                 if (ret) {
2378                         DRV_LOG(INFO, "failed to change file descriptor"
2379                                 " devx comp");
2380                         return;
2381                 }
2382                 sh->intr_handle_devx.fd = devx_comp->fd;
2383                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2384                 if (rte_intr_callback_register(&sh->intr_handle_devx,
2385                                         mlx5_dev_interrupt_handler_devx, sh)) {
2386                         DRV_LOG(INFO, "Fail to install the devx shared"
2387                                 " interrupt.");
2388                         sh->intr_handle_devx.fd = -1;
2389                 }
2390 #endif /* HAVE_IBV_DEVX_ASYNC */
2391         }
2392 }
2393
2394 /**
2395  * Uninstall shared asynchronous device events handler.
2396  * This function is implemented to support event sharing
2397  * between multiple ports of single IB device.
2398  *
2399  * @param dev
2400  *   Pointer to mlx5_dev_ctx_shared object.
2401  */
2402 void
2403 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2404 {
2405         if (sh->intr_handle.fd >= 0)
2406                 mlx5_intr_callback_unregister(&sh->intr_handle,
2407                                               mlx5_dev_interrupt_handler, sh);
2408 #ifdef HAVE_IBV_DEVX_ASYNC
2409         if (sh->intr_handle_devx.fd >= 0)
2410                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2411                                   mlx5_dev_interrupt_handler_devx, sh);
2412         if (sh->devx_comp)
2413                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2414 #endif
2415 }
2416
2417 /**
2418  * Read statistics by a named counter.
2419  *
2420  * @param[in] priv
2421  *   Pointer to the private device data structure.
2422  * @param[in] ctr_name
2423  *   Pointer to the name of the statistic counter to read
2424  * @param[out] stat
2425  *   Pointer to read statistic value.
2426  * @return
2427  *   0 on success and stat is valud, 1 if failed to read the value
2428  *   rte_errno is set.
2429  *
2430  */
2431 int
2432 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2433                       uint64_t *stat)
2434 {
2435         int fd;
2436
2437         if (priv->sh) {
2438                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2439                       priv->sh->ibdev_path,
2440                       priv->dev_port,
2441                       ctr_name);
2442                 fd = open(path, O_RDONLY);
2443                 /*
2444                  * in switchdev the file location is not per port
2445                  * but rather in <ibdev_path>/hw_counters/<file_name>.
2446                  */
2447                 if (fd == -1) {
2448                         MKSTR(path1, "%s/hw_counters/%s",
2449                               priv->sh->ibdev_path,
2450                               ctr_name);
2451                         fd = open(path1, O_RDONLY);
2452                 }
2453                 if (fd != -1) {
2454                         char buf[21] = {'\0'};
2455                         ssize_t n = read(fd, buf, sizeof(buf));
2456
2457                         close(fd);
2458                         if (n != -1) {
2459                                 *stat = strtoull(buf, NULL, 10);
2460                                 return 0;
2461                         }
2462                 }
2463         }
2464         *stat = 0;
2465         return 1;
2466 }
2467
2468 /**
2469  * Set the reg_mr and dereg_mr call backs
2470  *
2471  * @param reg_mr_cb[out]
2472  *   Pointer to reg_mr func
2473  * @param dereg_mr_cb[out]
2474  *   Pointer to dereg_mr func
2475  *
2476  */
2477 void
2478 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2479                       mlx5_dereg_mr_t *dereg_mr_cb)
2480 {
2481         *reg_mr_cb = mlx5_mr_verbs_ops.reg_mr;
2482         *dereg_mr_cb = mlx5_mr_verbs_ops.dereg_mr;
2483 }
2484
2485 /**
2486  * Remove a MAC address from device
2487  *
2488  * @param dev
2489  *   Pointer to Ethernet device structure.
2490  * @param index
2491  *   MAC address index.
2492  */
2493 void
2494 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2495 {
2496         struct mlx5_priv *priv = dev->data->dev_private;
2497         const int vf = priv->config.vf;
2498
2499         if (vf)
2500                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2501                                         mlx5_ifindex(dev), priv->mac_own,
2502                                         &dev->data->mac_addrs[index], index);
2503 }
2504
2505 /**
2506  * Adds a MAC address to the device
2507  *
2508  * @param dev
2509  *   Pointer to Ethernet device structure.
2510  * @param mac_addr
2511  *   MAC address to register.
2512  * @param index
2513  *   MAC address index.
2514  *
2515  * @return
2516  *   0 on success, a negative errno value otherwise
2517  */
2518 int
2519 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2520                      uint32_t index)
2521 {
2522         struct mlx5_priv *priv = dev->data->dev_private;
2523         const int vf = priv->config.vf;
2524         int ret = 0;
2525
2526         if (vf)
2527                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2528                                            mlx5_ifindex(dev), priv->mac_own,
2529                                            mac, index);
2530         return ret;
2531 }
2532
2533 /**
2534  * Modify a VF MAC address
2535  *
2536  * @param priv
2537  *   Pointer to device private data.
2538  * @param mac_addr
2539  *   MAC address to modify into.
2540  * @param iface_idx
2541  *   Net device interface index
2542  * @param vf_index
2543  *   VF index
2544  *
2545  * @return
2546  *   0 on success, a negative errno value otherwise
2547  */
2548 int
2549 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2550                            unsigned int iface_idx,
2551                            struct rte_ether_addr *mac_addr,
2552                            int vf_index)
2553 {
2554         return mlx5_nl_vf_mac_addr_modify
2555                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2556 }
2557
2558 /**
2559  * Set device promiscuous mode
2560  *
2561  * @param dev
2562  *   Pointer to Ethernet device structure.
2563  * @param enable
2564  *   0 - promiscuous is disabled, otherwise - enabled
2565  *
2566  * @return
2567  *   0 on success, a negative error value otherwise
2568  */
2569 int
2570 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2571 {
2572         struct mlx5_priv *priv = dev->data->dev_private;
2573
2574         return mlx5_nl_promisc(priv->nl_socket_route,
2575                                mlx5_ifindex(dev), !!enable);
2576 }
2577
2578 /**
2579  * Set device promiscuous mode
2580  *
2581  * @param dev
2582  *   Pointer to Ethernet device structure.
2583  * @param enable
2584  *   0 - all multicase is disabled, otherwise - enabled
2585  *
2586  * @return
2587  *   0 on success, a negative error value otherwise
2588  */
2589 int
2590 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2591 {
2592         struct mlx5_priv *priv = dev->data->dev_private;
2593
2594         return mlx5_nl_allmulti(priv->nl_socket_route,
2595                                 mlx5_ifindex(dev), !!enable);
2596 }
2597
2598 /**
2599  * Flush device MAC addresses
2600  *
2601  * @param dev
2602  *   Pointer to Ethernet device structure.
2603  *
2604  */
2605 void
2606 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2607 {
2608         struct mlx5_priv *priv = dev->data->dev_private;
2609
2610         mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2611                                dev->data->mac_addrs,
2612                                MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2613 }