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
  50 #define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
  51
  52 #ifndef HAVE_IBV_MLX5_MOD_MPW
  53 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  54 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  55 #endif
  56
  57 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  58 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  59 #endif
  60
  61 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
  62
  63 /* Spinlock for mlx5_shared_data allocation. */
  64 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
  65
  66 /* Process local data for secondary processes. */
  67 static struct mlx5_local_data mlx5_local_data;
  68
  69 /**
  70  * Set the completion channel file descriptor interrupt as non-blocking.
  71  *
  72  * @param[in] rxq_obj
  73  *   Pointer to RQ channel object, which includes the channel fd
  74  *
  75  * @param[out] fd
  76  *   The file descriptor (representing the intetrrupt) used in this channel.
  77  *
  78  * @return
  79  *   0 on successfully setting the fd to non-blocking, non-zero otherwise.
  80  */
  81 int
  82 mlx5_os_set_nonblock_channel_fd(int fd)
  83 {
  84         int flags;
  85
  86         flags = fcntl(fd, F_GETFL);
  87         return fcntl(fd, F_SETFL, flags | O_NONBLOCK);
  88 }
  89
  90 /**
  91  * Get mlx5 device attributes. The glue function query_device_ex() is called
  92  * with out parameter of type 'struct ibv_device_attr_ex *'. Then fill in mlx5
  93  * device attributes from the glue out parameter.
  94  *
  95  * @param dev
  96  *   Pointer to ibv context.
  97  *
  98  * @param device_attr
  99  *   Pointer to mlx5 device attributes.
 100  *
 101  * @return
 102  *   0 on success, non zero error number otherwise
 103  */
 104 int
 105 mlx5_os_get_dev_attr(void *ctx, struct mlx5_dev_attr *device_attr)
 106 {
 107         int err;
 108         struct ibv_device_attr_ex attr_ex;
 109         memset(device_attr, 0, sizeof(*device_attr));
 110         err = mlx5_glue->query_device_ex(ctx, NULL, &attr_ex);
 111         if (err)
 112                 return err;
 113
 114         device_attr->device_cap_flags_ex = attr_ex.device_cap_flags_ex;
 115         device_attr->max_qp_wr = attr_ex.orig_attr.max_qp_wr;
 116         device_attr->max_sge = attr_ex.orig_attr.max_sge;
 117         device_attr->max_cq = attr_ex.orig_attr.max_cq;
 118         device_attr->max_qp = attr_ex.orig_attr.max_qp;
 119         device_attr->raw_packet_caps = attr_ex.raw_packet_caps;
 120         device_attr->max_rwq_indirection_table_size =
 121                 attr_ex.rss_caps.max_rwq_indirection_table_size;
 122         device_attr->max_tso = attr_ex.tso_caps.max_tso;
 123         device_attr->tso_supported_qpts = attr_ex.tso_caps.supported_qpts;
 124
 125         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 126         err = mlx5_glue->dv_query_device(ctx, &dv_attr);
 127         if (err)
 128                 return err;
 129
 130         device_attr->flags = dv_attr.flags;
 131         device_attr->comp_mask = dv_attr.comp_mask;
 132 #ifdef HAVE_IBV_MLX5_MOD_SWP
 133         device_attr->sw_parsing_offloads =
 134                 dv_attr.sw_parsing_caps.sw_parsing_offloads;
 135 #endif
 136         device_attr->min_single_stride_log_num_of_bytes =
 137                 dv_attr.striding_rq_caps.min_single_stride_log_num_of_bytes;
 138         device_attr->max_single_stride_log_num_of_bytes =
 139                 dv_attr.striding_rq_caps.max_single_stride_log_num_of_bytes;
 140         device_attr->min_single_wqe_log_num_of_strides =
 141                 dv_attr.striding_rq_caps.min_single_wqe_log_num_of_strides;
 142         device_attr->max_single_wqe_log_num_of_strides =
 143                 dv_attr.striding_rq_caps.max_single_wqe_log_num_of_strides;
 144         device_attr->stride_supported_qpts =
 145                 dv_attr.striding_rq_caps.supported_qpts;
 146 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 147         device_attr->tunnel_offloads_caps = dv_attr.tunnel_offloads_caps;
 148 #endif
 149
 150         return err;
 151 }
 152
 153 /**
 154  * Verbs callback to allocate a memory. This function should allocate the space
 155  * according to the size provided residing inside a huge page.
 156  * Please note that all allocation must respect the alignment from libmlx5
 157  * (i.e. currently rte_mem_page_size()).
 158  *
 159  * @param[in] size
 160  *   The size in bytes of the memory to allocate.
 161  * @param[in] data
 162  *   A pointer to the callback data.
 163  *
 164  * @return
 165  *   Allocated buffer, NULL otherwise and rte_errno is set.
 166  */
 167 static void *
 168 mlx5_alloc_verbs_buf(size_t size, void *data)
 169 {
 170         struct mlx5_priv *priv = data;
 171         void *ret;
 172         unsigned int socket = SOCKET_ID_ANY;
 173         size_t alignment = rte_mem_page_size();
 174         if (alignment == (size_t)-1) {
 175                 DRV_LOG(ERR, "Failed to get mem page size");
 176                 rte_errno = ENOMEM;
 177                 return NULL;
 178         }
 179
 180         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 181                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 182
 183                 socket = ctrl->socket;
 184         } else if (priv->verbs_alloc_ctx.type ==
 185                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 186                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 187
 188                 socket = ctrl->socket;
 189         }
 190         MLX5_ASSERT(data != NULL);
 191         ret = mlx5_malloc(0, size, alignment, socket);
 192         if (!ret && size)
 193                 rte_errno = ENOMEM;
 194         return ret;
 195 }
 196
 197 /**
 198  * Verbs callback to free a memory.
 199  *
 200  * @param[in] ptr
 201  *   A pointer to the memory to free.
 202  * @param[in] data
 203  *   A pointer to the callback data.
 204  */
 205 static void
 206 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 207 {
 208         MLX5_ASSERT(data != NULL);
 209         mlx5_free(ptr);
 210 }
 211
 212 /**
 213  * Initialize DR related data within private structure.
 214  * Routine checks the reference counter and does actual
 215  * resources creation/initialization only if counter is zero.
 216  *
 217  * @param[in] priv
 218  *   Pointer to the private device data structure.
 219  *
 220  * @return
 221  *   Zero on success, positive error code otherwise.
 222  */
 223 static int
 224 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 225 {
 226         struct mlx5_dev_ctx_shared *sh = priv->sh;
 227         char s[MLX5_HLIST_NAMESIZE];
 228         int err = 0;
 229
 230         if (!sh->flow_tbls)
 231                 err = mlx5_alloc_table_hash_list(priv);
 232         else
 233                 DRV_LOG(DEBUG, "sh->flow_tbls[%p] already created, reuse\n",
 234                         (void *)sh->flow_tbls);
 235         if (err)
 236                 return err;
 237         /* Create tags hash list table. */
 238         snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
 239         sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
 240         if (!sh->tag_table) {
 241                 DRV_LOG(ERR, "tags with hash creation failed.");
 242                 err = ENOMEM;
 243                 goto error;
 244         }
 245         snprintf(s, sizeof(s), "%s_hdr_modify", sh->ibdev_name);
 246         sh->modify_cmds = mlx5_hlist_create(s, MLX5_FLOW_HDR_MODIFY_HTABLE_SZ);
 247         if (!sh->modify_cmds) {
 248                 DRV_LOG(ERR, "hdr modify hash creation failed");
 249                 err = ENOMEM;
 250                 goto error;
 251         }
 252 #ifdef HAVE_MLX5DV_DR
 253         void *domain;
 254
 255         if (sh->dv_refcnt) {
 256                 /* Shared DV/DR structures is already initialized. */
 257                 sh->dv_refcnt++;
 258                 priv->dr_shared = 1;
 259                 return 0;
 260         }
 261         /* Reference counter is zero, we should initialize structures. */
 262         domain = mlx5_glue->dr_create_domain(sh->ctx,
 263                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 264         if (!domain) {
 265                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 266                 err = errno;
 267                 goto error;
 268         }
 269         sh->rx_domain = domain;
 270         domain = mlx5_glue->dr_create_domain(sh->ctx,
 271                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 272         if (!domain) {
 273                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 274                 err = errno;
 275                 goto error;
 276         }
 277         pthread_mutex_init(&sh->dv_mutex, NULL);
 278         sh->tx_domain = domain;
 279 #ifdef HAVE_MLX5DV_DR_ESWITCH
 280         if (priv->config.dv_esw_en) {
 281                 domain  = mlx5_glue->dr_create_domain
 282                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 283                 if (!domain) {
 284                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 285                         err = errno;
 286                         goto error;
 287                 }
 288                 sh->fdb_domain = domain;
 289                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 290         }
 291 #endif
 292         if (priv->config.reclaim_mode == MLX5_RCM_AGGR) {
 293                 mlx5_glue->dr_reclaim_domain_memory(sh->rx_domain, 1);
 294                 mlx5_glue->dr_reclaim_domain_memory(sh->tx_domain, 1);
 295                 if (sh->fdb_domain)
 296                         mlx5_glue->dr_reclaim_domain_memory(sh->fdb_domain, 1);
 297         }
 298         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 299 #endif /* HAVE_MLX5DV_DR */
 300         sh->dv_refcnt++;
 301         priv->dr_shared = 1;
 302         return 0;
 303 error:
 304         /* Rollback the created objects. */
 305         if (sh->rx_domain) {
 306                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 307                 sh->rx_domain = NULL;
 308         }
 309         if (sh->tx_domain) {
 310                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 311                 sh->tx_domain = NULL;
 312         }
 313         if (sh->fdb_domain) {
 314                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 315                 sh->fdb_domain = NULL;
 316         }
 317         if (sh->esw_drop_action) {
 318                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 319                 sh->esw_drop_action = NULL;
 320         }
 321         if (sh->pop_vlan_action) {
 322                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 323                 sh->pop_vlan_action = NULL;
 324         }
 325         if (sh->modify_cmds) {
 326                 mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
 327                 sh->modify_cmds = NULL;
 328         }
 329         if (sh->tag_table) {
 330                 /* tags should be destroyed with flow before. */
 331                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 332                 sh->tag_table = NULL;
 333         }
 334         mlx5_free_table_hash_list(priv);
 335         return err;
 336 }
 337
 338 /**
 339  * Destroy DR related data within private structure.
 340  *
 341  * @param[in] priv
 342  *   Pointer to the private device data structure.
 343  */
 344 void
 345 mlx5_os_free_shared_dr(struct mlx5_priv *priv)
 346 {
 347         struct mlx5_dev_ctx_shared *sh;
 348
 349         if (!priv->dr_shared)
 350                 return;
 351         priv->dr_shared = 0;
 352         sh = priv->sh;
 353         MLX5_ASSERT(sh);
 354 #ifdef HAVE_MLX5DV_DR
 355         MLX5_ASSERT(sh->dv_refcnt);
 356         if (sh->dv_refcnt && --sh->dv_refcnt)
 357                 return;
 358         if (sh->rx_domain) {
 359                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 360                 sh->rx_domain = NULL;
 361         }
 362         if (sh->tx_domain) {
 363                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 364                 sh->tx_domain = NULL;
 365         }
 366 #ifdef HAVE_MLX5DV_DR_ESWITCH
 367         if (sh->fdb_domain) {
 368                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 369                 sh->fdb_domain = NULL;
 370         }
 371         if (sh->esw_drop_action) {
 372                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 373                 sh->esw_drop_action = NULL;
 374         }
 375 #endif
 376         if (sh->pop_vlan_action) {
 377                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 378                 sh->pop_vlan_action = NULL;
 379         }
 380         pthread_mutex_destroy(&sh->dv_mutex);
 381 #endif /* HAVE_MLX5DV_DR */
 382         if (sh->modify_cmds) {
 383                 mlx5_hlist_destroy(sh->modify_cmds, NULL, NULL);
 384                 sh->modify_cmds = NULL;
 385         }
 386         if (sh->tag_table) {
 387                 /* tags should be destroyed with flow before. */
 388                 mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
 389                 sh->tag_table = NULL;
 390         }
 391         mlx5_free_table_hash_list(priv);
 392 }
 393
 394 /**
 395  * Initialize shared data between primary and secondary process.
 396  *
 397  * A memzone is reserved by primary process and secondary processes attach to
 398  * the memzone.
 399  *
 400  * @return
 401  *   0 on success, a negative errno value otherwise and rte_errno is set.
 402  */
 403 static int
 404 mlx5_init_shared_data(void)
 405 {
 406         const struct rte_memzone *mz;
 407         int ret = 0;
 408
 409         rte_spinlock_lock(&mlx5_shared_data_lock);
 410         if (mlx5_shared_data == NULL) {
 411                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 412                         /* Allocate shared memory. */
 413                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 414                                                  sizeof(*mlx5_shared_data),
 415                                                  SOCKET_ID_ANY, 0);
 416                         if (mz == NULL) {
 417                                 DRV_LOG(ERR,
 418                                         "Cannot allocate mlx5 shared data");
 419                                 ret = -rte_errno;
 420                                 goto error;
 421                         }
 422                         mlx5_shared_data = mz->addr;
 423                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 424                         rte_spinlock_init(&mlx5_shared_data->lock);
 425                 } else {
 426                         /* Lookup allocated shared memory. */
 427                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 428                         if (mz == NULL) {
 429                                 DRV_LOG(ERR,
 430                                         "Cannot attach mlx5 shared data");
 431                                 ret = -rte_errno;
 432                                 goto error;
 433                         }
 434                         mlx5_shared_data = mz->addr;
 435                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 436                 }
 437         }
 438 error:
 439         rte_spinlock_unlock(&mlx5_shared_data_lock);
 440         return ret;
 441 }
 442
 443 /**
 444  * PMD global initialization.
 445  *
 446  * Independent from individual device, this function initializes global
 447  * per-PMD data structures distinguishing primary and secondary processes.
 448  * Hence, each initialization is called once per a process.
 449  *
 450  * @return
 451  *   0 on success, a negative errno value otherwise and rte_errno is set.
 452  */
 453 static int
 454 mlx5_init_once(void)
 455 {
 456         struct mlx5_shared_data *sd;
 457         struct mlx5_local_data *ld = &mlx5_local_data;
 458         int ret = 0;
 459
 460         if (mlx5_init_shared_data())
 461                 return -rte_errno;
 462         sd = mlx5_shared_data;
 463         MLX5_ASSERT(sd);
 464         rte_spinlock_lock(&sd->lock);
 465         switch (rte_eal_process_type()) {
 466         case RTE_PROC_PRIMARY:
 467                 if (sd->init_done)
 468                         break;
 469                 LIST_INIT(&sd->mem_event_cb_list);
 470                 rte_rwlock_init(&sd->mem_event_rwlock);
 471                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
 472                                                 mlx5_mr_mem_event_cb, NULL);
 473                 ret = mlx5_mp_init_primary(MLX5_MP_NAME,
 474                                            mlx5_mp_os_primary_handle);
 475                 if (ret)
 476                         goto out;
 477                 sd->init_done = true;
 478                 break;
 479         case RTE_PROC_SECONDARY:
 480                 if (ld->init_done)
 481                         break;
 482                 ret = mlx5_mp_init_secondary(MLX5_MP_NAME,
 483                                              mlx5_mp_os_secondary_handle);
 484                 if (ret)
 485                         goto out;
 486                 ++sd->secondary_cnt;
 487                 ld->init_done = true;
 488                 break;
 489         default:
 490                 break;
 491         }
 492 out:
 493         rte_spinlock_unlock(&sd->lock);
 494         return ret;
 495 }
 496
 497 /**
 498  * Spawn an Ethernet device from Verbs information.
 499  *
 500  * @param dpdk_dev
 501  *   Backing DPDK device.
 502  * @param spawn
 503  *   Verbs device parameters (name, port, switch_info) to spawn.
 504  * @param config
 505  *   Device configuration parameters.
 506  *
 507  * @return
 508  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
 509  *   is set. The following errors are defined:
 510  *
 511  *   EBUSY: device is not supposed to be spawned.
 512  *   EEXIST: device is already spawned
 513  */
 514 static struct rte_eth_dev *
 515 mlx5_dev_spawn(struct rte_device *dpdk_dev,
 516                struct mlx5_dev_spawn_data *spawn,
 517                struct mlx5_dev_config *config)
 518 {
 519         const struct mlx5_switch_info *switch_info = &spawn->info;
 520         struct mlx5_dev_ctx_shared *sh = NULL;
 521         struct ibv_port_attr port_attr;
 522         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
 523         struct rte_eth_dev *eth_dev = NULL;
 524         struct mlx5_priv *priv = NULL;
 525         int err = 0;
 526         unsigned int hw_padding = 0;
 527         unsigned int mps;
 528         unsigned int cqe_comp;
 529         unsigned int cqe_pad = 0;
 530         unsigned int tunnel_en = 0;
 531         unsigned int mpls_en = 0;
 532         unsigned int swp = 0;
 533         unsigned int mprq = 0;
 534         unsigned int mprq_min_stride_size_n = 0;
 535         unsigned int mprq_max_stride_size_n = 0;
 536         unsigned int mprq_min_stride_num_n = 0;
 537         unsigned int mprq_max_stride_num_n = 0;
 538         struct rte_ether_addr mac;
 539         char name[RTE_ETH_NAME_MAX_LEN];
 540         int own_domain_id = 0;
 541         uint16_t port_id;
 542         unsigned int i;
 543 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 544         struct mlx5dv_devx_port devx_port = { .comp_mask = 0 };
 545 #endif
 546
 547         /* Determine if this port representor is supposed to be spawned. */
 548         if (switch_info->representor && dpdk_dev->devargs) {
 549                 struct rte_eth_devargs eth_da;
 550
 551                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
 552                 if (err) {
 553                         rte_errno = -err;
 554                         DRV_LOG(ERR, "failed to process device arguments: %s",
 555                                 strerror(rte_errno));
 556                         return NULL;
 557                 }
 558                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
 559                         if (eth_da.representor_ports[i] ==
 560                             (uint16_t)switch_info->port_name)
 561                                 break;
 562                 if (i == eth_da.nb_representor_ports) {
 563                         rte_errno = EBUSY;
 564                         return NULL;
 565                 }
 566         }
 567         /* Build device name. */
 568         if (spawn->pf_bond <  0) {
 569                 /* Single device. */
 570                 if (!switch_info->representor)
 571                         strlcpy(name, dpdk_dev->name, sizeof(name));
 572                 else
 573                         snprintf(name, sizeof(name), "%s_representor_%u",
 574                                  dpdk_dev->name, switch_info->port_name);
 575         } else {
 576                 /* Bonding device. */
 577                 if (!switch_info->representor)
 578                         snprintf(name, sizeof(name), "%s_%s",
 579                                  dpdk_dev->name,
 580                                  mlx5_os_get_dev_device_name(spawn->phys_dev));
 581                 else
 582                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
 583                                  dpdk_dev->name,
 584                                  mlx5_os_get_dev_device_name(spawn->phys_dev),
 585                                  switch_info->port_name);
 586         }
 587         /* check if the device is already spawned */
 588         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
 589                 rte_errno = EEXIST;
 590                 return NULL;
 591         }
 592         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
 593         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 594                 struct mlx5_mp_id mp_id;
 595
 596                 eth_dev = rte_eth_dev_attach_secondary(name);
 597                 if (eth_dev == NULL) {
 598                         DRV_LOG(ERR, "can not attach rte ethdev");
 599                         rte_errno = ENOMEM;
 600                         return NULL;
 601                 }
 602                 eth_dev->device = dpdk_dev;
 603                 eth_dev->dev_ops = &mlx5_os_dev_sec_ops;
 604                 err = mlx5_proc_priv_init(eth_dev);
 605                 if (err)
 606                         return NULL;
 607                 mp_id.port_id = eth_dev->data->port_id;
 608                 strlcpy(mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 609                 /* Receive command fd from primary process */
 610                 err = mlx5_mp_req_verbs_cmd_fd(&mp_id);
 611                 if (err < 0)
 612                         goto err_secondary;
 613                 /* Remap UAR for Tx queues. */
 614                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
 615                 if (err)
 616                         goto err_secondary;
 617                 /*
 618                  * Ethdev pointer is still required as input since
 619                  * the primary device is not accessible from the
 620                  * secondary process.
 621                  */
 622                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
 623                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
 624                 return eth_dev;
 625 err_secondary:
 626                 mlx5_dev_close(eth_dev);
 627                 return NULL;
 628         }
 629         /*
 630          * Some parameters ("tx_db_nc" in particularly) are needed in
 631          * advance to create dv/verbs device context. We proceed the
 632          * devargs here to get ones, and later proceed devargs again
 633          * to override some hardware settings.
 634          */
 635         err = mlx5_args(config, dpdk_dev->devargs);
 636         if (err) {
 637                 err = rte_errno;
 638                 DRV_LOG(ERR, "failed to process device arguments: %s",
 639                         strerror(rte_errno));
 640                 goto error;
 641         }
 642         mlx5_malloc_mem_select(config->sys_mem_en);
 643         sh = mlx5_alloc_shared_dev_ctx(spawn, config);
 644         if (!sh)
 645                 return NULL;
 646         config->devx = sh->devx;
 647 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
 648         config->dest_tir = 1;
 649 #endif
 650 #ifdef HAVE_IBV_MLX5_MOD_SWP
 651         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 652 #endif
 653         /*
 654          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 655          * as all ConnectX-5 devices.
 656          */
 657 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 658         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 659 #endif
 660 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 661         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
 662 #endif
 663         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
 664         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 665                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 666                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 667                         mps = MLX5_MPW_ENHANCED;
 668                 } else {
 669                         DRV_LOG(DEBUG, "MPW is supported");
 670                         mps = MLX5_MPW;
 671                 }
 672         } else {
 673                 DRV_LOG(DEBUG, "MPW isn't supported");
 674                 mps = MLX5_MPW_DISABLED;
 675         }
 676 #ifdef HAVE_IBV_MLX5_MOD_SWP
 677         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
 678                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
 679         DRV_LOG(DEBUG, "SWP support: %u", swp);
 680 #endif
 681         config->swp = !!swp;
 682 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 683         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
 684                 struct mlx5dv_striding_rq_caps mprq_caps =
 685                         dv_attr.striding_rq_caps;
 686
 687                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
 688                         mprq_caps.min_single_stride_log_num_of_bytes);
 689                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
 690                         mprq_caps.max_single_stride_log_num_of_bytes);
 691                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
 692                         mprq_caps.min_single_wqe_log_num_of_strides);
 693                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
 694                         mprq_caps.max_single_wqe_log_num_of_strides);
 695                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
 696                         mprq_caps.supported_qpts);
 697                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
 698                 mprq = 1;
 699                 mprq_min_stride_size_n =
 700                         mprq_caps.min_single_stride_log_num_of_bytes;
 701                 mprq_max_stride_size_n =
 702                         mprq_caps.max_single_stride_log_num_of_bytes;
 703                 mprq_min_stride_num_n =
 704                         mprq_caps.min_single_wqe_log_num_of_strides;
 705                 mprq_max_stride_num_n =
 706                         mprq_caps.max_single_wqe_log_num_of_strides;
 707         }
 708 #endif
 709         if (RTE_CACHE_LINE_SIZE == 128 &&
 710             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 711                 cqe_comp = 0;
 712         else
 713                 cqe_comp = 1;
 714         config->cqe_comp = cqe_comp;
 715 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 716         /* Whether device supports 128B Rx CQE padding. */
 717         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
 718                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
 719 #endif
 720 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 721         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 722                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
 723                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 724                              (dv_attr.tunnel_offloads_caps &
 725                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE) &&
 726                              (dv_attr.tunnel_offloads_caps &
 727                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GENEVE));
 728         }
 729         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 730                 tunnel_en ? "" : "not ");
 731 #else
 732         DRV_LOG(WARNING,
 733                 "tunnel offloading disabled due to old OFED/rdma-core version");
 734 #endif
 735         config->tunnel_en = tunnel_en;
 736 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 737         mpls_en = ((dv_attr.tunnel_offloads_caps &
 738                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
 739                    (dv_attr.tunnel_offloads_caps &
 740                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
 741         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
 742                 mpls_en ? "" : "not ");
 743 #else
 744         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
 745                 " old OFED/rdma-core version or firmware configuration");
 746 #endif
 747         config->mpls_en = mpls_en;
 748         /* Check port status. */
 749         err = mlx5_glue->query_port(sh->ctx, spawn->phys_port, &port_attr);
 750         if (err) {
 751                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
 752                 goto error;
 753         }
 754         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 755                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
 756                 err = EINVAL;
 757                 goto error;
 758         }
 759         if (port_attr.state != IBV_PORT_ACTIVE)
 760                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
 761                         mlx5_glue->port_state_str(port_attr.state),
 762                         port_attr.state);
 763         /* Allocate private eth device data. */
 764         priv = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
 765                            sizeof(*priv),
 766                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
 767         if (priv == NULL) {
 768                 DRV_LOG(ERR, "priv allocation failure");
 769                 err = ENOMEM;
 770                 goto error;
 771         }
 772         priv->sh = sh;
 773         priv->dev_port = spawn->phys_port;
 774         priv->pci_dev = spawn->pci_dev;
 775         priv->mtu = RTE_ETHER_MTU;
 776         priv->mp_id.port_id = port_id;
 777         strlcpy(priv->mp_id.name, MLX5_MP_NAME, RTE_MP_MAX_NAME_LEN);
 778         /* Some internal functions rely on Netlink sockets, open them now. */
 779         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
 780         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
 781         priv->representor = !!switch_info->representor;
 782         priv->master = !!switch_info->master;
 783         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 784         priv->vport_meta_tag = 0;
 785         priv->vport_meta_mask = 0;
 786         priv->pf_bond = spawn->pf_bond;
 787 #ifdef HAVE_MLX5DV_DR_DEVX_PORT
 788         /*
 789          * The DevX port query API is implemented. E-Switch may use
 790          * either vport or reg_c[0] metadata register to match on
 791          * vport index. The engaged part of metadata register is
 792          * defined by mask.
 793          */
 794         if (switch_info->representor || switch_info->master) {
 795                 devx_port.comp_mask = MLX5DV_DEVX_PORT_VPORT |
 796                                       MLX5DV_DEVX_PORT_MATCH_REG_C_0;
 797                 err = mlx5_glue->devx_port_query(sh->ctx, spawn->phys_port,
 798                                                  &devx_port);
 799                 if (err) {
 800                         DRV_LOG(WARNING,
 801                                 "can't query devx port %d on device %s",
 802                                 spawn->phys_port,
 803                                 mlx5_os_get_dev_device_name(spawn->phys_dev));
 804                         devx_port.comp_mask = 0;
 805                 }
 806         }
 807         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_MATCH_REG_C_0) {
 808                 priv->vport_meta_tag = devx_port.reg_c_0.value;
 809                 priv->vport_meta_mask = devx_port.reg_c_0.mask;
 810                 if (!priv->vport_meta_mask) {
 811                         DRV_LOG(ERR, "vport zero mask for port %d"
 812                                      " on bonding device %s",
 813                                      spawn->phys_port,
 814                                      mlx5_os_get_dev_device_name
 815                                                         (spawn->phys_dev));
 816                         err = ENOTSUP;
 817                         goto error;
 818                 }
 819                 if (priv->vport_meta_tag & ~priv->vport_meta_mask) {
 820                         DRV_LOG(ERR, "invalid vport tag for port %d"
 821                                      " on bonding device %s",
 822                                      spawn->phys_port,
 823                                      mlx5_os_get_dev_device_name
 824                                                         (spawn->phys_dev));
 825                         err = ENOTSUP;
 826                         goto error;
 827                 }
 828         }
 829         if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
 830                 priv->vport_id = devx_port.vport_num;
 831         } else if (spawn->pf_bond >= 0) {
 832                 DRV_LOG(ERR, "can't deduce vport index for port %d"
 833                              " on bonding device %s",
 834                              spawn->phys_port,
 835                              mlx5_os_get_dev_device_name(spawn->phys_dev));
 836                 err = ENOTSUP;
 837                 goto error;
 838         } else {
 839                 /* Suppose vport index in compatible way. */
 840                 priv->vport_id = switch_info->representor ?
 841                                  switch_info->port_name + 1 : -1;
 842         }
 843 #else
 844         /*
 845          * Kernel/rdma_core support single E-Switch per PF configurations
 846          * only and vport_id field contains the vport index for
 847          * associated VF, which is deduced from representor port name.
 848          * For example, let's have the IB device port 10, it has
 849          * attached network device eth0, which has port name attribute
 850          * pf0vf2, we can deduce the VF number as 2, and set vport index
 851          * as 3 (2+1). This assigning schema should be changed if the
 852          * multiple E-Switch instances per PF configurations or/and PCI
 853          * subfunctions are added.
 854          */
 855         priv->vport_id = switch_info->representor ?
 856                          switch_info->port_name + 1 : -1;
 857 #endif
 858         /* representor_id field keeps the unmodified VF index. */
 859         priv->representor_id = switch_info->representor ?
 860                                switch_info->port_name : -1;
 861         /*
 862          * Look for sibling devices in order to reuse their switch domain
 863          * if any, otherwise allocate one.
 864          */
 865         MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
 866                 const struct mlx5_priv *opriv =
 867                         rte_eth_devices[port_id].data->dev_private;
 868
 869                 if (!opriv ||
 870                     opriv->sh != priv->sh ||
 871                         opriv->domain_id ==
 872                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
 873                         continue;
 874                 priv->domain_id = opriv->domain_id;
 875                 break;
 876         }
 877         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 878                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
 879                 if (err) {
 880                         err = rte_errno;
 881                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
 882                                 strerror(rte_errno));
 883                         goto error;
 884                 }
 885                 own_domain_id = 1;
 886         }
 887         /* Override some values set by hardware configuration. */
 888         mlx5_args(config, dpdk_dev->devargs);
 889         err = mlx5_dev_check_sibling_config(priv, config);
 890         if (err)
 891                 goto error;
 892         config->hw_csum = !!(sh->device_attr.device_cap_flags_ex &
 893                             IBV_DEVICE_RAW_IP_CSUM);
 894         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 895                 (config->hw_csum ? "" : "not "));
 896 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
 897         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 898         DRV_LOG(DEBUG, "counters are not supported");
 899 #endif
 900 #if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
 901         if (config->dv_flow_en) {
 902                 DRV_LOG(WARNING, "DV flow is not supported");
 903                 config->dv_flow_en = 0;
 904         }
 905 #endif
 906         config->ind_table_max_size =
 907                 sh->device_attr.max_rwq_indirection_table_size;
 908         /*
 909          * Remove this check once DPDK supports larger/variable
 910          * indirection tables.
 911          */
 912         if (config->ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
 913                 config->ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 914         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 915                 config->ind_table_max_size);
 916         config->hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
 917                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 918         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 919                 (config->hw_vlan_strip ? "" : "not "));
 920         config->hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
 921                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
 922 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 923         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
 924 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 925         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
 926                         IBV_DEVICE_PCI_WRITE_END_PADDING);
 927 #endif
 928         if (config->hw_padding && !hw_padding) {
 929                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
 930                 config->hw_padding = 0;
 931         } else if (config->hw_padding) {
 932                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
 933         }
 934         config->tso = (sh->device_attr.max_tso > 0 &&
 935                       (sh->device_attr.tso_supported_qpts &
 936                        (1 << IBV_QPT_RAW_PACKET)));
 937         if (config->tso)
 938                 config->tso_max_payload_sz = sh->device_attr.max_tso;
 939         /*
 940          * MPW is disabled by default, while the Enhanced MPW is enabled
 941          * by default.
 942          */
 943         if (config->mps == MLX5_ARG_UNSET)
 944                 config->mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
 945                                                           MLX5_MPW_DISABLED;
 946         else
 947                 config->mps = config->mps ? mps : MLX5_MPW_DISABLED;
 948         DRV_LOG(INFO, "%sMPS is %s",
 949                 config->mps == MLX5_MPW_ENHANCED ? "enhanced " :
 950                 config->mps == MLX5_MPW ? "legacy " : "",
 951                 config->mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
 952         if (config->cqe_comp && !cqe_comp) {
 953                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 954                 config->cqe_comp = 0;
 955         }
 956         if (config->cqe_pad && !cqe_pad) {
 957                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
 958                 config->cqe_pad = 0;
 959         } else if (config->cqe_pad) {
 960                 DRV_LOG(INFO, "Rx CQE padding is enabled");
 961         }
 962         if (config->devx) {
 963                 priv->counter_fallback = 0;
 964                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config->hca_attr);
 965                 if (err) {
 966                         err = -err;
 967                         goto error;
 968                 }
 969                 if (!config->hca_attr.flow_counters_dump)
 970                         priv->counter_fallback = 1;
 971 #ifndef HAVE_IBV_DEVX_ASYNC
 972                 priv->counter_fallback = 1;
 973 #endif
 974                 if (priv->counter_fallback)
 975                         DRV_LOG(INFO, "Use fall-back DV counter management");
 976                 /* Check for LRO support. */
 977                 if (config->dest_tir && config->hca_attr.lro_cap &&
 978                     config->dv_flow_en) {
 979                         /* TBD check tunnel lro caps. */
 980                         config->lro.supported = config->hca_attr.lro_cap;
 981                         DRV_LOG(DEBUG, "Device supports LRO");
 982                         /*
 983                          * If LRO timeout is not configured by application,
 984                          * use the minimal supported value.
 985                          */
 986                         if (!config->lro.timeout)
 987                                 config->lro.timeout =
 988                                 config->hca_attr.lro_timer_supported_periods[0];
 989                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
 990                                 config->lro.timeout);
 991                 }
 992 #if defined(HAVE_MLX5DV_DR) && defined(HAVE_MLX5_DR_CREATE_ACTION_FLOW_METER)
 993                 if (config->hca_attr.qos.sup &&
 994                     config->hca_attr.qos.srtcm_sup &&
 995                     config->dv_flow_en) {
 996                         uint8_t reg_c_mask =
 997                                 config->hca_attr.qos.flow_meter_reg_c_ids;
 998                         /*
 999                          * Meter needs two REG_C's for color match and pre-sfx
1000                          * flow match. Here get the REG_C for color match.
1001                          * REG_C_0 and REG_C_1 is reserved for metadata feature.
1002                          */
1003                         reg_c_mask &= 0xfc;
1004                         if (__builtin_popcount(reg_c_mask) < 1) {
1005                                 priv->mtr_en = 0;
1006                                 DRV_LOG(WARNING, "No available register for"
1007                                         " meter.");
1008                         } else {
1009                                 priv->mtr_color_reg = ffs(reg_c_mask) - 1 +
1010                                                       REG_C_0;
1011                                 priv->mtr_en = 1;
1012                                 priv->mtr_reg_share =
1013                                       config->hca_attr.qos.flow_meter_reg_share;
1014                                 DRV_LOG(DEBUG, "The REG_C meter uses is %d",
1015                                         priv->mtr_color_reg);
1016                         }
1017                 }
1018 #endif
1019         }
1020         if (config->tx_pp) {
1021                 DRV_LOG(DEBUG, "Timestamp counter frequency %u kHz",
1022                         config->hca_attr.dev_freq_khz);
1023                 DRV_LOG(DEBUG, "Packet pacing is %ssupported",
1024                         config->hca_attr.qos.packet_pacing ? "" : "not ");
1025                 DRV_LOG(DEBUG, "Cross channel ops are %ssupported",
1026                         config->hca_attr.cross_channel ? "" : "not ");
1027                 DRV_LOG(DEBUG, "WQE index ignore is %ssupported",
1028                         config->hca_attr.wqe_index_ignore ? "" : "not ");
1029                 DRV_LOG(DEBUG, "Non-wire SQ feature is %ssupported",
1030                         config->hca_attr.non_wire_sq ? "" : "not ");
1031                 DRV_LOG(DEBUG, "Static WQE SQ feature is %ssupported (%d)",
1032                         config->hca_attr.log_max_static_sq_wq ? "" : "not ",
1033                         config->hca_attr.log_max_static_sq_wq);
1034                 DRV_LOG(DEBUG, "WQE rate PP mode is %ssupported",
1035                         config->hca_attr.qos.wqe_rate_pp ? "" : "not ");
1036                 if (!config->devx) {
1037                         DRV_LOG(ERR, "DevX is required for packet pacing");
1038                         err = ENODEV;
1039                         goto error;
1040                 }
1041                 if (!config->hca_attr.qos.packet_pacing) {
1042                         DRV_LOG(ERR, "Packet pacing is not supported");
1043                         err = ENODEV;
1044                         goto error;
1045                 }
1046                 if (!config->hca_attr.cross_channel) {
1047                         DRV_LOG(ERR, "Cross channel operations are"
1048                                      " required for packet pacing");
1049                         err = ENODEV;
1050                         goto error;
1051                 }
1052                 if (!config->hca_attr.wqe_index_ignore) {
1053                         DRV_LOG(ERR, "WQE index ignore feature is"
1054                                      " required for packet pacing");
1055                         err = ENODEV;
1056                         goto error;
1057                 }
1058                 if (!config->hca_attr.non_wire_sq) {
1059                         DRV_LOG(ERR, "Non-wire SQ feature is"
1060                                      " required for packet pacing");
1061                         err = ENODEV;
1062                         goto error;
1063                 }
1064                 if (!config->hca_attr.log_max_static_sq_wq) {
1065                         DRV_LOG(ERR, "Static WQE SQ feature is"
1066                                      " required for packet pacing");
1067                         err = ENODEV;
1068                         goto error;
1069                 }
1070                 if (!config->hca_attr.qos.wqe_rate_pp) {
1071                         DRV_LOG(ERR, "WQE rate mode is required"
1072                                      " for packet pacing");
1073                         err = ENODEV;
1074                         goto error;
1075                 }
1076 #ifndef HAVE_MLX5DV_DEVX_UAR_OFFSET
1077                 DRV_LOG(ERR, "DevX does not provide UAR offset,"
1078                              " can't create queues for packet pacing");
1079                 err = ENODEV;
1080                 goto error;
1081 #endif
1082         }
1083         if (config->devx) {
1084                 uint32_t reg[MLX5_ST_SZ_DW(register_mtutc)];
1085
1086                 err = config->hca_attr.access_register_user ?
1087                         mlx5_devx_cmd_register_read
1088                                 (sh->ctx, MLX5_REGISTER_ID_MTUTC, 0,
1089                                 reg, MLX5_ST_SZ_DW(register_mtutc)) : ENOTSUP;
1090                 if (!err) {
1091                         uint32_t ts_mode;
1092
1093                         /* MTUTC register is read successfully. */
1094                         ts_mode = MLX5_GET(register_mtutc, reg,
1095                                            time_stamp_mode);
1096                         if (ts_mode == MLX5_MTUTC_TIMESTAMP_MODE_REAL_TIME)
1097                                 config->rt_timestamp = 1;
1098                 } else {
1099                         /* Kernel does not support register reading. */
1100                         if (config->hca_attr.dev_freq_khz ==
1101                                                  (NS_PER_S / MS_PER_S))
1102                                 config->rt_timestamp = 1;
1103                 }
1104         }
1105         /*
1106          * If HW has bug working with tunnel packet decapsulation and
1107          * scatter FCS, and decapsulation is needed, clear the hw_fcs_strip
1108          * bit. Then DEV_RX_OFFLOAD_KEEP_CRC bit will not be set anymore.
1109          */
1110         if (config->hca_attr.scatter_fcs_w_decap_disable && config->decap_en)
1111                 config->hw_fcs_strip = 0;
1112         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1113                 (config->hw_fcs_strip ? "" : "not "));
1114         if (config->mprq.enabled && mprq) {
1115                 if (config->mprq.stride_num_n &&
1116                     (config->mprq.stride_num_n > mprq_max_stride_num_n ||
1117                      config->mprq.stride_num_n < mprq_min_stride_num_n)) {
1118                         config->mprq.stride_num_n =
1119                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1120                                                 mprq_min_stride_num_n),
1121                                         mprq_max_stride_num_n);
1122                         DRV_LOG(WARNING,
1123                                 "the number of strides"
1124                                 " for Multi-Packet RQ is out of range,"
1125                                 " setting default value (%u)",
1126                                 1 << config->mprq.stride_num_n);
1127                 }
1128                 if (config->mprq.stride_size_n &&
1129                     (config->mprq.stride_size_n > mprq_max_stride_size_n ||
1130                      config->mprq.stride_size_n < mprq_min_stride_size_n)) {
1131                         config->mprq.stride_size_n =
1132                                 RTE_MIN(RTE_MAX(MLX5_MPRQ_STRIDE_SIZE_N,
1133                                                 mprq_min_stride_size_n),
1134                                         mprq_max_stride_size_n);
1135                         DRV_LOG(WARNING,
1136                                 "the size of a stride"
1137                                 " for Multi-Packet RQ is out of range,"
1138                                 " setting default value (%u)",
1139                                 1 << config->mprq.stride_size_n);
1140                 }
1141                 config->mprq.min_stride_size_n = mprq_min_stride_size_n;
1142                 config->mprq.max_stride_size_n = mprq_max_stride_size_n;
1143         } else if (config->mprq.enabled && !mprq) {
1144                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1145                 config->mprq.enabled = 0;
1146         }
1147         if (config->max_dump_files_num == 0)
1148                 config->max_dump_files_num = 128;
1149         eth_dev = rte_eth_dev_allocate(name);
1150         if (eth_dev == NULL) {
1151                 DRV_LOG(ERR, "can not allocate rte ethdev");
1152                 err = ENOMEM;
1153                 goto error;
1154         }
1155         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1156         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1157         if (priv->representor) {
1158                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1159                 eth_dev->data->representor_id = priv->representor_id;
1160         }
1161         /*
1162          * Store associated network device interface index. This index
1163          * is permanent throughout the lifetime of device. So, we may store
1164          * the ifindex here and use the cached value further.
1165          */
1166         MLX5_ASSERT(spawn->ifindex);
1167         priv->if_index = spawn->ifindex;
1168         eth_dev->data->dev_private = priv;
1169         priv->dev_data = eth_dev->data;
1170         eth_dev->data->mac_addrs = priv->mac;
1171         eth_dev->device = dpdk_dev;
1172         /* Configure the first MAC address by default. */
1173         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1174                 DRV_LOG(ERR,
1175                         "port %u cannot get MAC address, is mlx5_en"
1176                         " loaded? (errno: %s)",
1177                         eth_dev->data->port_id, strerror(rte_errno));
1178                 err = ENODEV;
1179                 goto error;
1180         }
1181         DRV_LOG(INFO,
1182                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1183                 eth_dev->data->port_id,
1184                 mac.addr_bytes[0], mac.addr_bytes[1],
1185                 mac.addr_bytes[2], mac.addr_bytes[3],
1186                 mac.addr_bytes[4], mac.addr_bytes[5]);
1187 #ifdef RTE_LIBRTE_MLX5_DEBUG
1188         {
1189                 char ifname[IF_NAMESIZE];
1190
1191                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1192                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1193                                 eth_dev->data->port_id, ifname);
1194                 else
1195                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1196                                 eth_dev->data->port_id);
1197         }
1198 #endif
1199         /* Get actual MTU if possible. */
1200         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1201         if (err) {
1202                 err = rte_errno;
1203                 goto error;
1204         }
1205         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1206                 priv->mtu);
1207         /* Initialize burst functions to prevent crashes before link-up. */
1208         eth_dev->rx_pkt_burst = removed_rx_burst;
1209         eth_dev->tx_pkt_burst = removed_tx_burst;
1210         eth_dev->dev_ops = &mlx5_os_dev_ops;
1211         /* Register MAC address. */
1212         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1213         if (config->vf && config->vf_nl_en)
1214                 mlx5_nl_mac_addr_sync(priv->nl_socket_route,
1215                                       mlx5_ifindex(eth_dev),
1216                                       eth_dev->data->mac_addrs,
1217                                       MLX5_MAX_MAC_ADDRESSES);
1218         priv->flows = 0;
1219         priv->ctrl_flows = 0;
1220         TAILQ_INIT(&priv->flow_meters);
1221         TAILQ_INIT(&priv->flow_meter_profiles);
1222         /* Hint libmlx5 to use PMD allocator for data plane resources */
1223         mlx5_glue->dv_set_context_attr(sh->ctx,
1224                         MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1225                         (void *)((uintptr_t)&(struct mlx5dv_ctx_allocators){
1226                                 .alloc = &mlx5_alloc_verbs_buf,
1227                                 .free = &mlx5_free_verbs_buf,
1228                                 .data = priv,
1229                         }));
1230         /* Bring Ethernet device up. */
1231         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1232                 eth_dev->data->port_id);
1233         mlx5_set_link_up(eth_dev);
1234         /*
1235          * Even though the interrupt handler is not installed yet,
1236          * interrupts will still trigger on the async_fd from
1237          * Verbs context returned by ibv_open_device().
1238          */
1239         mlx5_link_update(eth_dev, 0);
1240 #ifdef HAVE_MLX5DV_DR_ESWITCH
1241         if (!(config->hca_attr.eswitch_manager && config->dv_flow_en &&
1242               (switch_info->representor || switch_info->master)))
1243                 config->dv_esw_en = 0;
1244 #else
1245         config->dv_esw_en = 0;
1246 #endif
1247         /* Detect minimal data bytes to inline. */
1248         mlx5_set_min_inline(spawn, config);
1249         /* Store device configuration on private structure. */
1250         priv->config = *config;
1251         /* Create context for virtual machine VLAN workaround. */
1252         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
1253         if (config->dv_flow_en) {
1254                 err = mlx5_alloc_shared_dr(priv);
1255                 if (err)
1256                         goto error;
1257                 /*
1258                  * RSS id is shared with meter flow id. Meter flow id can only
1259                  * use the 24 MSB of the register.
1260                  */
1261                 priv->qrss_id_pool = mlx5_flow_id_pool_alloc(UINT32_MAX >>
1262                                      MLX5_MTR_COLOR_BITS);
1263                 if (!priv->qrss_id_pool) {
1264                         DRV_LOG(ERR, "can't create flow id pool");
1265                         err = ENOMEM;
1266                         goto error;
1267                 }
1268         }
1269         /* Supported Verbs flow priority number detection. */
1270         err = mlx5_flow_discover_priorities(eth_dev);
1271         if (err < 0) {
1272                 err = -err;
1273                 goto error;
1274         }
1275         priv->config.flow_prio = err;
1276         if (!priv->config.dv_esw_en &&
1277             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1278                 DRV_LOG(WARNING, "metadata mode %u is not supported "
1279                                  "(no E-Switch)", priv->config.dv_xmeta_en);
1280                 priv->config.dv_xmeta_en = MLX5_XMETA_MODE_LEGACY;
1281         }
1282         mlx5_set_metadata_mask(eth_dev);
1283         if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1284             !priv->sh->dv_regc0_mask) {
1285                 DRV_LOG(ERR, "metadata mode %u is not supported "
1286                              "(no metadata reg_c[0] is available)",
1287                              priv->config.dv_xmeta_en);
1288                         err = ENOTSUP;
1289                         goto error;
1290         }
1291         /*
1292          * Allocate the buffer for flow creating, just once.
1293          * The allocation must be done before any flow creating.
1294          */
1295         mlx5_flow_alloc_intermediate(eth_dev);
1296         /* Query availability of metadata reg_c's. */
1297         err = mlx5_flow_discover_mreg_c(eth_dev);
1298         if (err < 0) {
1299                 err = -err;
1300                 goto error;
1301         }
1302         if (!mlx5_flow_ext_mreg_supported(eth_dev)) {
1303                 DRV_LOG(DEBUG,
1304                         "port %u extensive metadata register is not supported",
1305                         eth_dev->data->port_id);
1306                 if (priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY) {
1307                         DRV_LOG(ERR, "metadata mode %u is not supported "
1308                                      "(no metadata registers available)",
1309                                      priv->config.dv_xmeta_en);
1310                         err = ENOTSUP;
1311                         goto error;
1312                 }
1313         }
1314         if (priv->config.dv_flow_en &&
1315             priv->config.dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
1316             mlx5_flow_ext_mreg_supported(eth_dev) &&
1317             priv->sh->dv_regc0_mask) {
1318                 priv->mreg_cp_tbl = mlx5_hlist_create(MLX5_FLOW_MREG_HNAME,
1319                                                       MLX5_FLOW_MREG_HTABLE_SZ);
1320                 if (!priv->mreg_cp_tbl) {
1321                         err = ENOMEM;
1322                         goto error;
1323                 }
1324         }
1325         return eth_dev;
1326 error:
1327         if (priv) {
1328                 if (priv->mreg_cp_tbl)
1329                         mlx5_hlist_destroy(priv->mreg_cp_tbl, NULL, NULL);
1330                 if (priv->sh)
1331                         mlx5_os_free_shared_dr(priv);
1332                 if (priv->nl_socket_route >= 0)
1333                         close(priv->nl_socket_route);
1334                 if (priv->nl_socket_rdma >= 0)
1335                         close(priv->nl_socket_rdma);
1336                 if (priv->vmwa_context)
1337                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
1338                 if (priv->qrss_id_pool)
1339                         mlx5_flow_id_pool_release(priv->qrss_id_pool);
1340                 if (own_domain_id)
1341                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1342                 mlx5_free(priv);
1343                 if (eth_dev != NULL)
1344                         eth_dev->data->dev_private = NULL;
1345         }
1346         if (eth_dev != NULL) {
1347                 /* mac_addrs must not be freed alone because part of
1348                  * dev_private
1349                  **/
1350                 eth_dev->data->mac_addrs = NULL;
1351                 rte_eth_dev_release_port(eth_dev);
1352         }
1353         if (sh)
1354                 mlx5_free_shared_dev_ctx(sh);
1355         MLX5_ASSERT(err > 0);
1356         rte_errno = err;
1357         return NULL;
1358 }
1359
1360 /**
1361  * Comparison callback to sort device data.
1362  *
1363  * This is meant to be used with qsort().
1364  *
1365  * @param a[in]
1366  *   Pointer to pointer to first data object.
1367  * @param b[in]
1368  *   Pointer to pointer to second data object.
1369  *
1370  * @return
1371  *   0 if both objects are equal, less than 0 if the first argument is less
1372  *   than the second, greater than 0 otherwise.
1373  */
1374 static int
1375 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
1376 {
1377         const struct mlx5_switch_info *si_a =
1378                 &((const struct mlx5_dev_spawn_data *)a)->info;
1379         const struct mlx5_switch_info *si_b =
1380                 &((const struct mlx5_dev_spawn_data *)b)->info;
1381         int ret;
1382
1383         /* Master device first. */
1384         ret = si_b->master - si_a->master;
1385         if (ret)
1386                 return ret;
1387         /* Then representor devices. */
1388         ret = si_b->representor - si_a->representor;
1389         if (ret)
1390                 return ret;
1391         /* Unidentified devices come last in no specific order. */
1392         if (!si_a->representor)
1393                 return 0;
1394         /* Order representors by name. */
1395         return si_a->port_name - si_b->port_name;
1396 }
1397
1398 /**
1399  * Match PCI information for possible slaves of bonding device.
1400  *
1401  * @param[in] ibv_dev
1402  *   Pointer to Infiniband device structure.
1403  * @param[in] pci_dev
1404  *   Pointer to PCI device structure to match PCI address.
1405  * @param[in] nl_rdma
1406  *   Netlink RDMA group socket handle.
1407  *
1408  * @return
1409  *   negative value if no bonding device found, otherwise
1410  *   positive index of slave PF in bonding.
1411  */
1412 static int
1413 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
1414                            const struct rte_pci_device *pci_dev,
1415                            int nl_rdma)
1416 {
1417         char ifname[IF_NAMESIZE + 1];
1418         unsigned int ifindex;
1419         unsigned int np, i;
1420         FILE *file = NULL;
1421         int pf = -1;
1422
1423         /*
1424          * Try to get master device name. If something goes
1425          * wrong suppose the lack of kernel support and no
1426          * bonding devices.
1427          */
1428         if (nl_rdma < 0)
1429                 return -1;
1430         if (!strstr(ibv_dev->name, "bond"))
1431                 return -1;
1432         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
1433         if (!np)
1434                 return -1;
1435         /*
1436          * The Master device might not be on the predefined
1437          * port (not on port index 1, it is not garanted),
1438          * we have to scan all Infiniband device port and
1439          * find master.
1440          */
1441         for (i = 1; i <= np; ++i) {
1442                 /* Check whether Infiniband port is populated. */
1443                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
1444                 if (!ifindex)
1445                         continue;
1446                 if (!if_indextoname(ifindex, ifname))
1447                         continue;
1448                 /* Try to read bonding slave names from sysfs. */
1449                 MKSTR(slaves,
1450                       "/sys/class/net/%s/master/bonding/slaves", ifname);
1451                 file = fopen(slaves, "r");
1452                 if (file)
1453                         break;
1454         }
1455         if (!file)
1456                 return -1;
1457         /* Use safe format to check maximal buffer length. */
1458         MLX5_ASSERT(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
1459         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
1460                 char tmp_str[IF_NAMESIZE + 32];
1461                 struct rte_pci_addr pci_addr;
1462                 struct mlx5_switch_info info;
1463
1464                 /* Process slave interface names in the loop. */
1465                 snprintf(tmp_str, sizeof(tmp_str),
1466                          "/sys/class/net/%s", ifname);
1467                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
1468                         DRV_LOG(WARNING, "can not get PCI address"
1469                                          " for netdev \"%s\"", ifname);
1470                         continue;
1471                 }
1472                 if (pci_dev->addr.domain != pci_addr.domain ||
1473                     pci_dev->addr.bus != pci_addr.bus ||
1474                     pci_dev->addr.devid != pci_addr.devid ||
1475                     pci_dev->addr.function != pci_addr.function)
1476                         continue;
1477                 /* Slave interface PCI address match found. */
1478                 fclose(file);
1479                 snprintf(tmp_str, sizeof(tmp_str),
1480                          "/sys/class/net/%s/phys_port_name", ifname);
1481                 file = fopen(tmp_str, "rb");
1482                 if (!file)
1483                         break;
1484                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
1485                 if (fscanf(file, "%32s", tmp_str) == 1)
1486                         mlx5_translate_port_name(tmp_str, &info);
1487                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
1488                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
1489                         pf = info.port_name;
1490                 break;
1491         }
1492         if (file)
1493                 fclose(file);
1494         return pf;
1495 }
1496
1497 /**
1498  * DPDK callback to register a PCI device.
1499  *
1500  * This function spawns Ethernet devices out of a given PCI device.
1501  *
1502  * @param[in] pci_drv
1503  *   PCI driver structure (mlx5_driver).
1504  * @param[in] pci_dev
1505  *   PCI device information.
1506  *
1507  * @return
1508  *   0 on success, a negative errno value otherwise and rte_errno is set.
1509  */
1510 int
1511 mlx5_os_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1512                   struct rte_pci_device *pci_dev)
1513 {
1514         struct ibv_device **ibv_list;
1515         /*
1516          * Number of found IB Devices matching with requested PCI BDF.
1517          * nd != 1 means there are multiple IB devices over the same
1518          * PCI device and we have representors and master.
1519          */
1520         unsigned int nd = 0;
1521         /*
1522          * Number of found IB device Ports. nd = 1 and np = 1..n means
1523          * we have the single multiport IB device, and there may be
1524          * representors attached to some of found ports.
1525          */
1526         unsigned int np = 0;
1527         /*
1528          * Number of DPDK ethernet devices to Spawn - either over
1529          * multiple IB devices or multiple ports of single IB device.
1530          * Actually this is the number of iterations to spawn.
1531          */
1532         unsigned int ns = 0;
1533         /*
1534          * Bonding device
1535          *   < 0 - no bonding device (single one)
1536          *  >= 0 - bonding device (value is slave PF index)
1537          */
1538         int bd = -1;
1539         struct mlx5_dev_spawn_data *list = NULL;
1540         struct mlx5_dev_config dev_config;
1541         unsigned int dev_config_vf;
1542         int ret;
1543
1544         if (rte_eal_process_type() == RTE_PROC_PRIMARY)
1545                 mlx5_pmd_socket_init();
1546         ret = mlx5_init_once();
1547         if (ret) {
1548                 DRV_LOG(ERR, "unable to init PMD global data: %s",
1549                         strerror(rte_errno));
1550                 return -rte_errno;
1551         }
1552         errno = 0;
1553         ibv_list = mlx5_glue->get_device_list(&ret);
1554         if (!ibv_list) {
1555                 rte_errno = errno ? errno : ENOSYS;
1556                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
1557                 return -rte_errno;
1558         }
1559         /*
1560          * First scan the list of all Infiniband devices to find
1561          * matching ones, gathering into the list.
1562          */
1563         struct ibv_device *ibv_match[ret + 1];
1564         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
1565         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
1566         unsigned int i;
1567
1568         while (ret-- > 0) {
1569                 struct rte_pci_addr pci_addr;
1570
1571                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
1572                 bd = mlx5_device_bond_pci_match
1573                                 (ibv_list[ret], pci_dev, nl_rdma);
1574                 if (bd >= 0) {
1575                         /*
1576                          * Bonding device detected. Only one match is allowed,
1577                          * the bonding is supported over multi-port IB device,
1578                          * there should be no matches on representor PCI
1579                          * functions or non VF LAG bonding devices with
1580                          * specified address.
1581                          */
1582                         if (nd) {
1583                                 DRV_LOG(ERR,
1584                                         "multiple PCI match on bonding device"
1585                                         "\"%s\" found", ibv_list[ret]->name);
1586                                 rte_errno = ENOENT;
1587                                 ret = -rte_errno;
1588                                 goto exit;
1589                         }
1590                         DRV_LOG(INFO, "PCI information matches for"
1591                                       " slave %d bonding device \"%s\"",
1592                                       bd, ibv_list[ret]->name);
1593                         ibv_match[nd++] = ibv_list[ret];
1594                         break;
1595                 }
1596                 if (mlx5_dev_to_pci_addr
1597                         (ibv_list[ret]->ibdev_path, &pci_addr))
1598                         continue;
1599                 if (pci_dev->addr.domain != pci_addr.domain ||
1600                     pci_dev->addr.bus != pci_addr.bus ||
1601                     pci_dev->addr.devid != pci_addr.devid ||
1602                     pci_dev->addr.function != pci_addr.function)
1603                         continue;
1604                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
1605                         ibv_list[ret]->name);
1606                 ibv_match[nd++] = ibv_list[ret];
1607         }
1608         ibv_match[nd] = NULL;
1609         if (!nd) {
1610                 /* No device matches, just complain and bail out. */
1611                 DRV_LOG(WARNING,
1612                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
1613                         " are kernel drivers loaded?",
1614                         pci_dev->addr.domain, pci_dev->addr.bus,
1615                         pci_dev->addr.devid, pci_dev->addr.function);
1616                 rte_errno = ENOENT;
1617                 ret = -rte_errno;
1618                 goto exit;
1619         }
1620         if (nd == 1) {
1621                 /*
1622                  * Found single matching device may have multiple ports.
1623                  * Each port may be representor, we have to check the port
1624                  * number and check the representors existence.
1625                  */
1626                 if (nl_rdma >= 0)
1627                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
1628                 if (!np)
1629                         DRV_LOG(WARNING, "can not get IB device \"%s\""
1630                                          " ports number", ibv_match[0]->name);
1631                 if (bd >= 0 && !np) {
1632                         DRV_LOG(ERR, "can not get ports"
1633                                      " for bonding device");
1634                         rte_errno = ENOENT;
1635                         ret = -rte_errno;
1636                         goto exit;
1637                 }
1638         }
1639 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
1640         if (bd >= 0) {
1641                 /*
1642                  * This may happen if there is VF LAG kernel support and
1643                  * application is compiled with older rdma_core library.
1644                  */
1645                 DRV_LOG(ERR,
1646                         "No kernel/verbs support for VF LAG bonding found.");
1647                 rte_errno = ENOTSUP;
1648                 ret = -rte_errno;
1649                 goto exit;
1650         }
1651 #endif
1652         /*
1653          * Now we can determine the maximal
1654          * amount of devices to be spawned.
1655          */
1656         list = mlx5_malloc(MLX5_MEM_ZERO,
1657                            sizeof(struct mlx5_dev_spawn_data) *
1658                            (np ? np : nd),
1659                            RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
1660         if (!list) {
1661                 DRV_LOG(ERR, "spawn data array allocation failure");
1662                 rte_errno = ENOMEM;
1663                 ret = -rte_errno;
1664                 goto exit;
1665         }
1666         if (bd >= 0 || np > 1) {
1667                 /*
1668                  * Single IB device with multiple ports found,
1669                  * it may be E-Switch master device and representors.
1670                  * We have to perform identification through the ports.
1671                  */
1672                 MLX5_ASSERT(nl_rdma >= 0);
1673                 MLX5_ASSERT(ns == 0);
1674                 MLX5_ASSERT(nd == 1);
1675                 MLX5_ASSERT(np);
1676                 for (i = 1; i <= np; ++i) {
1677                         list[ns].max_port = np;
1678                         list[ns].phys_port = i;
1679                         list[ns].phys_dev = ibv_match[0];
1680                         list[ns].eth_dev = NULL;
1681                         list[ns].pci_dev = pci_dev;
1682                         list[ns].pf_bond = bd;
1683                         list[ns].ifindex = mlx5_nl_ifindex
1684                                 (nl_rdma,
1685                                 mlx5_os_get_dev_device_name
1686                                                 (list[ns].phys_dev), i);
1687                         if (!list[ns].ifindex) {
1688                                 /*
1689                                  * No network interface index found for the
1690                                  * specified port, it means there is no
1691                                  * representor on this port. It's OK,
1692                                  * there can be disabled ports, for example
1693                                  * if sriov_numvfs < sriov_totalvfs.
1694                                  */
1695                                 continue;
1696                         }
1697                         ret = -1;
1698                         if (nl_route >= 0)
1699                                 ret = mlx5_nl_switch_info
1700                                                (nl_route,
1701                                                 list[ns].ifindex,
1702                                                 &list[ns].info);
1703                         if (ret || (!list[ns].info.representor &&
1704                                     !list[ns].info.master)) {
1705                                 /*
1706                                  * We failed to recognize representors with
1707                                  * Netlink, let's try to perform the task
1708                                  * with sysfs.
1709                                  */
1710                                 ret =  mlx5_sysfs_switch_info
1711                                                 (list[ns].ifindex,
1712                                                  &list[ns].info);
1713                         }
1714                         if (!ret && bd >= 0) {
1715                                 switch (list[ns].info.name_type) {
1716                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
1717                                         if (list[ns].info.port_name == bd)
1718                                                 ns++;
1719                                         break;
1720                                 case MLX5_PHYS_PORT_NAME_TYPE_PFHPF:
1721                                         /* Fallthrough */
1722                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
1723                                         if (list[ns].info.pf_num == bd)
1724                                                 ns++;
1725                                         break;
1726                                 default:
1727                                         break;
1728                                 }
1729                                 continue;
1730                         }
1731                         if (!ret && (list[ns].info.representor ^
1732                                      list[ns].info.master))
1733                                 ns++;
1734                 }
1735                 if (!ns) {
1736                         DRV_LOG(ERR,
1737                                 "unable to recognize master/representors"
1738                                 " on the IB device with multiple ports");
1739                         rte_errno = ENOENT;
1740                         ret = -rte_errno;
1741                         goto exit;
1742                 }
1743         } else {
1744                 /*
1745                  * The existence of several matching entries (nd > 1) means
1746                  * port representors have been instantiated. No existing Verbs
1747                  * call nor sysfs entries can tell them apart, this can only
1748                  * be done through Netlink calls assuming kernel drivers are
1749                  * recent enough to support them.
1750                  *
1751                  * In the event of identification failure through Netlink,
1752                  * try again through sysfs, then:
1753                  *
1754                  * 1. A single IB device matches (nd == 1) with single
1755                  *    port (np=0/1) and is not a representor, assume
1756                  *    no switch support.
1757                  *
1758                  * 2. Otherwise no safe assumptions can be made;
1759                  *    complain louder and bail out.
1760                  */
1761                 for (i = 0; i != nd; ++i) {
1762                         memset(&list[ns].info, 0, sizeof(list[ns].info));
1763                         list[ns].max_port = 1;
1764                         list[ns].phys_port = 1;
1765                         list[ns].phys_dev = ibv_match[i];
1766                         list[ns].eth_dev = NULL;
1767                         list[ns].pci_dev = pci_dev;
1768                         list[ns].pf_bond = -1;
1769                         list[ns].ifindex = 0;
1770                         if (nl_rdma >= 0)
1771                                 list[ns].ifindex = mlx5_nl_ifindex
1772                                 (nl_rdma,
1773                                 mlx5_os_get_dev_device_name
1774                                                 (list[ns].phys_dev), 1);
1775                         if (!list[ns].ifindex) {
1776                                 char ifname[IF_NAMESIZE];
1777
1778                                 /*
1779                                  * Netlink failed, it may happen with old
1780                                  * ib_core kernel driver (before 4.16).
1781                                  * We can assume there is old driver because
1782                                  * here we are processing single ports IB
1783                                  * devices. Let's try sysfs to retrieve
1784                                  * the ifindex. The method works for
1785                                  * master device only.
1786                                  */
1787                                 if (nd > 1) {
1788                                         /*
1789                                          * Multiple devices found, assume
1790                                          * representors, can not distinguish
1791                                          * master/representor and retrieve
1792                                          * ifindex via sysfs.
1793                                          */
1794                                         continue;
1795                                 }
1796                                 ret = mlx5_get_ifname_sysfs
1797                                         (ibv_match[i]->ibdev_path, ifname);
1798                                 if (!ret)
1799                                         list[ns].ifindex =
1800                                                 if_nametoindex(ifname);
1801                                 if (!list[ns].ifindex) {
1802                                         /*
1803                                          * No network interface index found
1804                                          * for the specified device, it means
1805                                          * there it is neither representor
1806                                          * nor master.
1807                                          */
1808                                         continue;
1809                                 }
1810                         }
1811                         ret = -1;
1812                         if (nl_route >= 0)
1813                                 ret = mlx5_nl_switch_info
1814                                                (nl_route,
1815                                                 list[ns].ifindex,
1816                                                 &list[ns].info);
1817                         if (ret || (!list[ns].info.representor &&
1818                                     !list[ns].info.master)) {
1819                                 /*
1820                                  * We failed to recognize representors with
1821                                  * Netlink, let's try to perform the task
1822                                  * with sysfs.
1823                                  */
1824                                 ret =  mlx5_sysfs_switch_info
1825                                                 (list[ns].ifindex,
1826                                                  &list[ns].info);
1827                         }
1828                         if (!ret && (list[ns].info.representor ^
1829                                      list[ns].info.master)) {
1830                                 ns++;
1831                         } else if ((nd == 1) &&
1832                                    !list[ns].info.representor &&
1833                                    !list[ns].info.master) {
1834                                 /*
1835                                  * Single IB device with
1836                                  * one physical port and
1837                                  * attached network device.
1838                                  * May be SRIOV is not enabled
1839                                  * or there is no representors.
1840                                  */
1841                                 DRV_LOG(INFO, "no E-Switch support detected");
1842                                 ns++;
1843                                 break;
1844                         }
1845                 }
1846                 if (!ns) {
1847                         DRV_LOG(ERR,
1848                                 "unable to recognize master/representors"
1849                                 " on the multiple IB devices");
1850                         rte_errno = ENOENT;
1851                         ret = -rte_errno;
1852                         goto exit;
1853                 }
1854         }
1855         MLX5_ASSERT(ns);
1856         /*
1857          * Sort list to probe devices in natural order for users convenience
1858          * (i.e. master first, then representors from lowest to highest ID).
1859          */
1860         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
1861         /* Device specific configuration. */
1862         switch (pci_dev->id.device_id) {
1863         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1864         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1865         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1866         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1867         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
1868         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
1869         case PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF:
1870                 dev_config_vf = 1;
1871                 break;
1872         default:
1873                 dev_config_vf = 0;
1874                 break;
1875         }
1876         for (i = 0; i != ns; ++i) {
1877                 uint32_t restore;
1878
1879                 /* Default configuration. */
1880                 memset(&dev_config, 0, sizeof(struct mlx5_dev_config));
1881                 dev_config.vf = dev_config_vf;
1882                 dev_config.mps = MLX5_ARG_UNSET;
1883                 dev_config.dbnc = MLX5_ARG_UNSET;
1884                 dev_config.rx_vec_en = 1;
1885                 dev_config.txq_inline_max = MLX5_ARG_UNSET;
1886                 dev_config.txq_inline_min = MLX5_ARG_UNSET;
1887                 dev_config.txq_inline_mpw = MLX5_ARG_UNSET;
1888                 dev_config.txqs_inline = MLX5_ARG_UNSET;
1889                 dev_config.vf_nl_en = 1;
1890                 dev_config.mr_ext_memseg_en = 1;
1891                 dev_config.mprq.max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN;
1892                 dev_config.mprq.min_rxqs_num = MLX5_MPRQ_MIN_RXQS;
1893                 dev_config.dv_esw_en = 1;
1894                 dev_config.dv_flow_en = 1;
1895                 dev_config.decap_en = 1;
1896                 dev_config.log_hp_size = MLX5_ARG_UNSET;
1897                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
1898                                                  &list[i],
1899                                                  &dev_config);
1900                 if (!list[i].eth_dev) {
1901                         if (rte_errno != EBUSY && rte_errno != EEXIST)
1902                                 break;
1903                         /* Device is disabled or already spawned. Ignore it. */
1904                         continue;
1905                 }
1906                 restore = list[i].eth_dev->data->dev_flags;
1907                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
1908                 /* Restore non-PCI flags cleared by the above call. */
1909                 list[i].eth_dev->data->dev_flags |= restore;
1910                 rte_eth_dev_probing_finish(list[i].eth_dev);
1911         }
1912         if (i != ns) {
1913                 DRV_LOG(ERR,
1914                         "probe of PCI device " PCI_PRI_FMT " aborted after"
1915                         " encountering an error: %s",
1916                         pci_dev->addr.domain, pci_dev->addr.bus,
1917                         pci_dev->addr.devid, pci_dev->addr.function,
1918                         strerror(rte_errno));
1919                 ret = -rte_errno;
1920                 /* Roll back. */
1921                 while (i--) {
1922                         if (!list[i].eth_dev)
1923                                 continue;
1924                         mlx5_dev_close(list[i].eth_dev);
1925                         /* mac_addrs must not be freed because in dev_private */
1926                         list[i].eth_dev->data->mac_addrs = NULL;
1927                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
1928                 }
1929                 /* Restore original error. */
1930                 rte_errno = -ret;
1931         } else {
1932                 ret = 0;
1933         }
1934 exit:
1935         /*
1936          * Do the routine cleanup:
1937          * - close opened Netlink sockets
1938          * - free allocated spawn data array
1939          * - free the Infiniband device list
1940          */
1941         if (nl_rdma >= 0)
1942                 close(nl_rdma);
1943         if (nl_route >= 0)
1944                 close(nl_route);
1945         if (list)
1946                 mlx5_free(list);
1947         MLX5_ASSERT(ibv_list);
1948         mlx5_glue->free_device_list(ibv_list);
1949         return ret;
1950 }
1951
1952 static int
1953 mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
1954 {
1955         char *env;
1956         int value;
1957
1958         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1959         /* Get environment variable to store. */
1960         env = getenv(MLX5_SHUT_UP_BF);
1961         value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
1962         if (config->dbnc == MLX5_ARG_UNSET)
1963                 setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
1964         else
1965                 setenv(MLX5_SHUT_UP_BF,
1966                        config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
1967         return value;
1968 }
1969
1970 static void
1971 mlx5_restore_doorbell_mapping_env(int value)
1972 {
1973         MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1974         /* Restore the original environment variable state. */
1975         if (value == MLX5_ARG_UNSET)
1976                 unsetenv(MLX5_SHUT_UP_BF);
1977         else
1978                 setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
1979 }
1980
1981 /**
1982  * Extract pdn of PD object using DV API.
1983  *
1984  * @param[in] pd
1985  *   Pointer to the verbs PD object.
1986  * @param[out] pdn
1987  *   Pointer to the PD object number variable.
1988  *
1989  * @return
1990  *   0 on success, error value otherwise.
1991  */
1992 int
1993 mlx5_os_get_pdn(void *pd, uint32_t *pdn)
1994 {
1995 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1996         struct mlx5dv_obj obj;
1997         struct mlx5dv_pd pd_info;
1998         int ret = 0;
1999
2000         obj.pd.in = pd;
2001         obj.pd.out = &pd_info;
2002         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
2003         if (ret) {
2004                 DRV_LOG(DEBUG, "Fail to get PD object info");
2005                 return ret;
2006         }
2007         *pdn = pd_info.pdn;
2008         return 0;
2009 #else
2010         (void)pd;
2011         (void)pdn;
2012         return -ENOTSUP;
2013 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
2014 }
2015
2016 /**
2017  * Function API to open IB device.
2018  *
2019  * This function calls the Linux glue APIs to open a device.
2020  *
2021  * @param[in] spawn
2022  *   Pointer to the IB device attributes (name, port, etc).
2023  * @param[out] config
2024  *   Pointer to device configuration structure.
2025  * @param[out] sh
2026  *   Pointer to shared context structure.
2027  *
2028  * @return
2029  *   0 on success, a positive error value otherwise.
2030  */
2031 int
2032 mlx5_os_open_device(const struct mlx5_dev_spawn_data *spawn,
2033                      const struct mlx5_dev_config *config,
2034                      struct mlx5_dev_ctx_shared *sh)
2035 {
2036         int dbmap_env;
2037         int err = 0;
2038
2039         sh->numa_node = spawn->pci_dev->device.numa_node;
2040         pthread_mutex_init(&sh->txpp.mutex, NULL);
2041         /*
2042          * Configure environment variable "MLX5_BF_SHUT_UP"
2043          * before the device creation. The rdma_core library
2044          * checks the variable at device creation and
2045          * stores the result internally.
2046          */
2047         dbmap_env = mlx5_config_doorbell_mapping_env(config);
2048         /* Try to open IB device with DV first, then usual Verbs. */
2049         errno = 0;
2050         sh->ctx = mlx5_glue->dv_open_device(spawn->phys_dev);
2051         if (sh->ctx) {
2052                 sh->devx = 1;
2053                 DRV_LOG(DEBUG, "DevX is supported");
2054                 /* The device is created, no need for environment. */
2055                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2056         } else {
2057                 /* The environment variable is still configured. */
2058                 sh->ctx = mlx5_glue->open_device(spawn->phys_dev);
2059                 err = errno ? errno : ENODEV;
2060                 /*
2061                  * The environment variable is not needed anymore,
2062                  * all device creation attempts are completed.
2063                  */
2064                 mlx5_restore_doorbell_mapping_env(dbmap_env);
2065                 if (!sh->ctx)
2066                         return err;
2067                 DRV_LOG(DEBUG, "DevX is NOT supported");
2068                 err = 0;
2069         }
2070         return err;
2071 }
2072
2073 /**
2074  * Install shared asynchronous device events handler.
2075  * This function is implemented to support event sharing
2076  * between multiple ports of single IB device.
2077  *
2078  * @param sh
2079  *   Pointer to mlx5_dev_ctx_shared object.
2080  */
2081 void
2082 mlx5_os_dev_shared_handler_install(struct mlx5_dev_ctx_shared *sh)
2083 {
2084         int ret;
2085         int flags;
2086
2087         sh->intr_handle.fd = -1;
2088         flags = fcntl(((struct ibv_context *)sh->ctx)->async_fd, F_GETFL);
2089         ret = fcntl(((struct ibv_context *)sh->ctx)->async_fd,
2090                     F_SETFL, flags | O_NONBLOCK);
2091         if (ret) {
2092                 DRV_LOG(INFO, "failed to change file descriptor async event"
2093                         " queue");
2094         } else {
2095                 sh->intr_handle.fd = ((struct ibv_context *)sh->ctx)->async_fd;
2096                 sh->intr_handle.type = RTE_INTR_HANDLE_EXT;
2097                 if (rte_intr_callback_register(&sh->intr_handle,
2098                                         mlx5_dev_interrupt_handler, sh)) {
2099                         DRV_LOG(INFO, "Fail to install the shared interrupt.");
2100                         sh->intr_handle.fd = -1;
2101                 }
2102         }
2103         if (sh->devx) {
2104 #ifdef HAVE_IBV_DEVX_ASYNC
2105                 sh->intr_handle_devx.fd = -1;
2106                 sh->devx_comp =
2107                         (void *)mlx5_glue->devx_create_cmd_comp(sh->ctx);
2108                 struct mlx5dv_devx_cmd_comp *devx_comp = sh->devx_comp;
2109                 if (!devx_comp) {
2110                         DRV_LOG(INFO, "failed to allocate devx_comp.");
2111                         return;
2112                 }
2113                 flags = fcntl(devx_comp->fd, F_GETFL);
2114                 ret = fcntl(devx_comp->fd, F_SETFL, flags | O_NONBLOCK);
2115                 if (ret) {
2116                         DRV_LOG(INFO, "failed to change file descriptor"
2117                                 " devx comp");
2118                         return;
2119                 }
2120                 sh->intr_handle_devx.fd = devx_comp->fd;
2121                 sh->intr_handle_devx.type = RTE_INTR_HANDLE_EXT;
2122                 if (rte_intr_callback_register(&sh->intr_handle_devx,
2123                                         mlx5_dev_interrupt_handler_devx, sh)) {
2124                         DRV_LOG(INFO, "Fail to install the devx shared"
2125                                 " interrupt.");
2126                         sh->intr_handle_devx.fd = -1;
2127                 }
2128 #endif /* HAVE_IBV_DEVX_ASYNC */
2129         }
2130 }
2131
2132 /**
2133  * Uninstall shared asynchronous device events handler.
2134  * This function is implemented to support event sharing
2135  * between multiple ports of single IB device.
2136  *
2137  * @param dev
2138  *   Pointer to mlx5_dev_ctx_shared object.
2139  */
2140 void
2141 mlx5_os_dev_shared_handler_uninstall(struct mlx5_dev_ctx_shared *sh)
2142 {
2143         if (sh->intr_handle.fd >= 0)
2144                 mlx5_intr_callback_unregister(&sh->intr_handle,
2145                                               mlx5_dev_interrupt_handler, sh);
2146 #ifdef HAVE_IBV_DEVX_ASYNC
2147         if (sh->intr_handle_devx.fd >= 0)
2148                 rte_intr_callback_unregister(&sh->intr_handle_devx,
2149                                   mlx5_dev_interrupt_handler_devx, sh);
2150         if (sh->devx_comp)
2151                 mlx5_glue->devx_destroy_cmd_comp(sh->devx_comp);
2152 #endif
2153 }
2154
2155 /**
2156  * Read statistics by a named counter.
2157  *
2158  * @param[in] priv
2159  *   Pointer to the private device data structure.
2160  * @param[in] ctr_name
2161  *   Pointer to the name of the statistic counter to read
2162  * @param[out] stat
2163  *   Pointer to read statistic value.
2164  * @return
2165  *   0 on success and stat is valud, 1 if failed to read the value
2166  *   rte_errno is set.
2167  *
2168  */
2169 int
2170 mlx5_os_read_dev_stat(struct mlx5_priv *priv, const char *ctr_name,
2171                       uint64_t *stat)
2172 {
2173         int fd;
2174
2175         if (priv->sh) {
2176                 MKSTR(path, "%s/ports/%d/hw_counters/%s",
2177                       priv->sh->ibdev_path,
2178                       priv->dev_port,
2179                       ctr_name);
2180                 fd = open(path, O_RDONLY);
2181                 /*
2182                  * in switchdev the file location is not per port
2183                  * but rather in <ibdev_path>/hw_counters/<file_name>.
2184                  */
2185                 if (fd == -1) {
2186                         MKSTR(path1, "%s/hw_counters/%s",
2187                               priv->sh->ibdev_path,
2188                               ctr_name);
2189                         fd = open(path1, O_RDONLY);
2190                 }
2191                 if (fd != -1) {
2192                         char buf[21] = {'\0'};
2193                         ssize_t n = read(fd, buf, sizeof(buf));
2194
2195                         close(fd);
2196                         if (n != -1) {
2197                                 *stat = strtoull(buf, NULL, 10);
2198                                 return 0;
2199                         }
2200                 }
2201         }
2202         *stat = 0;
2203         return 1;
2204 }
2205
2206 /**
2207  * Set the reg_mr and dereg_mr call backs
2208  *
2209  * @param reg_mr_cb[out]
2210  *   Pointer to reg_mr func
2211  * @param dereg_mr_cb[out]
2212  *   Pointer to dereg_mr func
2213  *
2214  */
2215 void
2216 mlx5_os_set_reg_mr_cb(mlx5_reg_mr_t *reg_mr_cb,
2217                       mlx5_dereg_mr_t *dereg_mr_cb)
2218 {
2219         *reg_mr_cb = mlx5_verbs_ops.reg_mr;
2220         *dereg_mr_cb = mlx5_verbs_ops.dereg_mr;
2221 }
2222
2223 /**
2224  * Remove a MAC address from device
2225  *
2226  * @param dev
2227  *   Pointer to Ethernet device structure.
2228  * @param index
2229  *   MAC address index.
2230  */
2231 void
2232 mlx5_os_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
2233 {
2234         struct mlx5_priv *priv = dev->data->dev_private;
2235         const int vf = priv->config.vf;
2236
2237         if (vf)
2238                 mlx5_nl_mac_addr_remove(priv->nl_socket_route,
2239                                         mlx5_ifindex(dev), priv->mac_own,
2240                                         &dev->data->mac_addrs[index], index);
2241 }
2242
2243 /**
2244  * Adds a MAC address to the device
2245  *
2246  * @param dev
2247  *   Pointer to Ethernet device structure.
2248  * @param mac_addr
2249  *   MAC address to register.
2250  * @param index
2251  *   MAC address index.
2252  *
2253  * @return
2254  *   0 on success, a negative errno value otherwise
2255  */
2256 int
2257 mlx5_os_mac_addr_add(struct rte_eth_dev *dev, struct rte_ether_addr *mac,
2258                      uint32_t index)
2259 {
2260         struct mlx5_priv *priv = dev->data->dev_private;
2261         const int vf = priv->config.vf;
2262         int ret = 0;
2263
2264         if (vf)
2265                 ret = mlx5_nl_mac_addr_add(priv->nl_socket_route,
2266                                            mlx5_ifindex(dev), priv->mac_own,
2267                                            mac, index);
2268         return ret;
2269 }
2270
2271 /**
2272  * Modify a VF MAC address
2273  *
2274  * @param priv
2275  *   Pointer to device private data.
2276  * @param mac_addr
2277  *   MAC address to modify into.
2278  * @param iface_idx
2279  *   Net device interface index
2280  * @param vf_index
2281  *   VF index
2282  *
2283  * @return
2284  *   0 on success, a negative errno value otherwise
2285  */
2286 int
2287 mlx5_os_vf_mac_addr_modify(struct mlx5_priv *priv,
2288                            unsigned int iface_idx,
2289                            struct rte_ether_addr *mac_addr,
2290                            int vf_index)
2291 {
2292         return mlx5_nl_vf_mac_addr_modify
2293                 (priv->nl_socket_route, iface_idx, mac_addr, vf_index);
2294 }
2295
2296 /**
2297  * Set device promiscuous mode
2298  *
2299  * @param dev
2300  *   Pointer to Ethernet device structure.
2301  * @param enable
2302  *   0 - promiscuous is disabled, otherwise - enabled
2303  *
2304  * @return
2305  *   0 on success, a negative error value otherwise
2306  */
2307 int
2308 mlx5_os_set_promisc(struct rte_eth_dev *dev, int enable)
2309 {
2310         struct mlx5_priv *priv = dev->data->dev_private;
2311
2312         return mlx5_nl_promisc(priv->nl_socket_route,
2313                                mlx5_ifindex(dev), !!enable);
2314 }
2315
2316 /**
2317  * Set device promiscuous mode
2318  *
2319  * @param dev
2320  *   Pointer to Ethernet device structure.
2321  * @param enable
2322  *   0 - all multicase is disabled, otherwise - enabled
2323  *
2324  * @return
2325  *   0 on success, a negative error value otherwise
2326  */
2327 int
2328 mlx5_os_set_allmulti(struct rte_eth_dev *dev, int enable)
2329 {
2330         struct mlx5_priv *priv = dev->data->dev_private;
2331
2332         return mlx5_nl_allmulti(priv->nl_socket_route,
2333                                 mlx5_ifindex(dev), !!enable);
2334 }
2335
2336 /**
2337  * Flush device MAC addresses
2338  *
2339  * @param dev
2340  *   Pointer to Ethernet device structure.
2341  *
2342  */
2343 void
2344 mlx5_os_mac_addr_flush(struct rte_eth_dev *dev)
2345 {
2346         struct mlx5_priv *priv = dev->data->dev_private;
2347
2348         mlx5_nl_mac_addr_flush(priv->nl_socket_route, mlx5_ifindex(dev),
2349                                dev->data->mac_addrs,
2350                                MLX5_MAX_MAC_ADDRESSES, priv->mac_own);
2351 }
2352
2353 const struct eth_dev_ops mlx5_os_dev_ops = {
2354         .dev_configure = mlx5_dev_configure,
2355         .dev_start = mlx5_dev_start,
2356         .dev_stop = mlx5_dev_stop,
2357         .dev_set_link_down = mlx5_set_link_down,
2358         .dev_set_link_up = mlx5_set_link_up,
2359         .dev_close = mlx5_dev_close,
2360         .promiscuous_enable = mlx5_promiscuous_enable,
2361         .promiscuous_disable = mlx5_promiscuous_disable,
2362         .allmulticast_enable = mlx5_allmulticast_enable,
2363         .allmulticast_disable = mlx5_allmulticast_disable,
2364         .link_update = mlx5_link_update,
2365         .stats_get = mlx5_stats_get,
2366         .stats_reset = mlx5_stats_reset,
2367         .xstats_get = mlx5_xstats_get,
2368         .xstats_reset = mlx5_xstats_reset,
2369         .xstats_get_names = mlx5_xstats_get_names,
2370         .fw_version_get = mlx5_fw_version_get,
2371         .dev_infos_get = mlx5_dev_infos_get,
2372         .read_clock = mlx5_txpp_read_clock,
2373         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2374         .vlan_filter_set = mlx5_vlan_filter_set,
2375         .rx_queue_setup = mlx5_rx_queue_setup,
2376         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2377         .tx_queue_setup = mlx5_tx_queue_setup,
2378         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2379         .rx_queue_release = mlx5_rx_queue_release,
2380         .tx_queue_release = mlx5_tx_queue_release,
2381         .rx_queue_start = mlx5_rx_queue_start,
2382         .rx_queue_stop = mlx5_rx_queue_stop,
2383         .tx_queue_start = mlx5_tx_queue_start,
2384         .tx_queue_stop = mlx5_tx_queue_stop,
2385         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2386         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2387         .mac_addr_remove = mlx5_mac_addr_remove,
2388         .mac_addr_add = mlx5_mac_addr_add,
2389         .mac_addr_set = mlx5_mac_addr_set,
2390         .set_mc_addr_list = mlx5_set_mc_addr_list,
2391         .mtu_set = mlx5_dev_set_mtu,
2392         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2393         .vlan_offload_set = mlx5_vlan_offload_set,
2394         .reta_update = mlx5_dev_rss_reta_update,
2395         .reta_query = mlx5_dev_rss_reta_query,
2396         .rss_hash_update = mlx5_rss_hash_update,
2397         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
2398         .filter_ctrl = mlx5_dev_filter_ctrl,
2399         .rx_descriptor_status = mlx5_rx_descriptor_status,
2400         .tx_descriptor_status = mlx5_tx_descriptor_status,
2401         .rxq_info_get = mlx5_rxq_info_get,
2402         .txq_info_get = mlx5_txq_info_get,
2403         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2404         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2405         .rx_queue_count = mlx5_rx_queue_count,
2406         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2407         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2408         .is_removed = mlx5_is_removed,
2409         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
2410         .get_module_info = mlx5_get_module_info,
2411         .get_module_eeprom = mlx5_get_module_eeprom,
2412         .hairpin_cap_get = mlx5_hairpin_cap_get,
2413         .mtr_ops_get = mlx5_flow_meter_ops_get,
2414 };
2415
2416 /* Available operations from secondary process. */
2417 const struct eth_dev_ops mlx5_os_dev_sec_ops = {
2418         .stats_get = mlx5_stats_get,
2419         .stats_reset = mlx5_stats_reset,
2420         .xstats_get = mlx5_xstats_get,
2421         .xstats_reset = mlx5_xstats_reset,
2422         .xstats_get_names = mlx5_xstats_get_names,
2423         .fw_version_get = mlx5_fw_version_get,
2424         .dev_infos_get = mlx5_dev_infos_get,
2425         .read_clock = mlx5_txpp_read_clock,
2426         .rx_queue_start = mlx5_rx_queue_start,
2427         .rx_queue_stop = mlx5_rx_queue_stop,
2428         .tx_queue_start = mlx5_tx_queue_start,
2429         .tx_queue_stop = mlx5_tx_queue_stop,
2430         .rx_descriptor_status = mlx5_rx_descriptor_status,
2431         .tx_descriptor_status = mlx5_tx_descriptor_status,
2432         .rxq_info_get = mlx5_rxq_info_get,
2433         .txq_info_get = mlx5_txq_info_get,
2434         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2435         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2436         .get_module_info = mlx5_get_module_info,
2437         .get_module_eeprom = mlx5_get_module_eeprom,
2438 };
2439
2440 /* Available operations in flow isolated mode. */
2441 const struct eth_dev_ops mlx5_os_dev_ops_isolate = {
2442         .dev_configure = mlx5_dev_configure,
2443         .dev_start = mlx5_dev_start,
2444         .dev_stop = mlx5_dev_stop,
2445         .dev_set_link_down = mlx5_set_link_down,
2446         .dev_set_link_up = mlx5_set_link_up,
2447         .dev_close = mlx5_dev_close,
2448         .promiscuous_enable = mlx5_promiscuous_enable,
2449         .promiscuous_disable = mlx5_promiscuous_disable,
2450         .allmulticast_enable = mlx5_allmulticast_enable,
2451         .allmulticast_disable = mlx5_allmulticast_disable,
2452         .link_update = mlx5_link_update,
2453         .stats_get = mlx5_stats_get,
2454         .stats_reset = mlx5_stats_reset,
2455         .xstats_get = mlx5_xstats_get,
2456         .xstats_reset = mlx5_xstats_reset,
2457         .xstats_get_names = mlx5_xstats_get_names,
2458         .fw_version_get = mlx5_fw_version_get,
2459         .dev_infos_get = mlx5_dev_infos_get,
2460         .read_clock = mlx5_txpp_read_clock,
2461         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
2462         .vlan_filter_set = mlx5_vlan_filter_set,
2463         .rx_queue_setup = mlx5_rx_queue_setup,
2464         .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
2465         .tx_queue_setup = mlx5_tx_queue_setup,
2466         .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
2467         .rx_queue_release = mlx5_rx_queue_release,
2468         .tx_queue_release = mlx5_tx_queue_release,
2469         .rx_queue_start = mlx5_rx_queue_start,
2470         .rx_queue_stop = mlx5_rx_queue_stop,
2471         .tx_queue_start = mlx5_tx_queue_start,
2472         .tx_queue_stop = mlx5_tx_queue_stop,
2473         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
2474         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
2475         .mac_addr_remove = mlx5_mac_addr_remove,
2476         .mac_addr_add = mlx5_mac_addr_add,
2477         .mac_addr_set = mlx5_mac_addr_set,
2478         .set_mc_addr_list = mlx5_set_mc_addr_list,
2479         .mtu_set = mlx5_dev_set_mtu,
2480         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
2481         .vlan_offload_set = mlx5_vlan_offload_set,
2482         .filter_ctrl = mlx5_dev_filter_ctrl,
2483         .rx_descriptor_status = mlx5_rx_descriptor_status,
2484         .tx_descriptor_status = mlx5_tx_descriptor_status,
2485         .rxq_info_get = mlx5_rxq_info_get,
2486         .txq_info_get = mlx5_txq_info_get,
2487         .rx_burst_mode_get = mlx5_rx_burst_mode_get,
2488         .tx_burst_mode_get = mlx5_tx_burst_mode_get,
2489         .rx_queue_intr_enable = mlx5_rx_intr_enable,
2490         .rx_queue_intr_disable = mlx5_rx_intr_disable,
2491         .is_removed = mlx5_is_removed,
2492         .get_module_info = mlx5_get_module_info,
2493         .get_module_eeprom = mlx5_get_module_eeprom,
2494         .hairpin_cap_get = mlx5_hairpin_cap_get,
2495         .mtr_ops_get = mlx5_flow_meter_ops_get,
2496 };