drivers/net/mlx5/mlx5.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2015 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stddef.h>
   7 #include <unistd.h>
   8 #include <string.h>
   9 #include <assert.h>
  10 #include <dlfcn.h>
  11 #include <stdint.h>
  12 #include <stdlib.h>
  13 #include <errno.h>
  14 #include <net/if.h>
  15 #include <sys/mman.h>
  16 #include <linux/rtnetlink.h>
  17
  18 /* Verbs header. */
  19 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  20 #ifdef PEDANTIC
  21 #pragma GCC diagnostic ignored "-Wpedantic"
  22 #endif
  23 #include <infiniband/verbs.h>
  24 #ifdef PEDANTIC
  25 #pragma GCC diagnostic error "-Wpedantic"
  26 #endif
  27
  28 #include <rte_malloc.h>
  29 #include <rte_ethdev_driver.h>
  30 #include <rte_ethdev_pci.h>
  31 #include <rte_pci.h>
  32 #include <rte_bus_pci.h>
  33 #include <rte_common.h>
  34 #include <rte_config.h>
  35 #include <rte_kvargs.h>
  36 #include <rte_rwlock.h>
  37 #include <rte_spinlock.h>
  38 #include <rte_string_fns.h>
  39 #include <rte_alarm.h>
  40
  41 #include "mlx5.h"
  42 #include "mlx5_utils.h"
  43 #include "mlx5_rxtx.h"
  44 #include "mlx5_autoconf.h"
  45 #include "mlx5_defs.h"
  46 #include "mlx5_glue.h"
  47 #include "mlx5_mr.h"
  48 #include "mlx5_flow.h"
  49
  50 /* Device parameter to enable RX completion queue compression. */
  51 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
  52
  53 /* Device parameter to enable RX completion entry padding to 128B. */
  54 #define MLX5_RXQ_CQE_PAD_EN "rxq_cqe_pad_en"
  55
  56 /* Device parameter to enable padding Rx packet to cacheline size. */
  57 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
  58
  59 /* Device parameter to enable Multi-Packet Rx queue. */
  60 #define MLX5_RX_MPRQ_EN "mprq_en"
  61
  62 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
  63 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
  64
  65 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
  66 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
  67
  68 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
  69 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
  70
  71 /* Device parameter to configure inline send. Deprecated, ignored.*/
  72 #define MLX5_TXQ_INLINE "txq_inline"
  73
  74 /* Device parameter to limit packet size to inline with ordinary SEND. */
  75 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
  76
  77 /* Device parameter to configure minimal data size to inline. */
  78 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
  79
  80 /* Device parameter to limit packet size to inline with Enhanced MPW. */
  81 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
  82
  83 /*
  84  * Device parameter to configure the number of TX queues threshold for
  85  * enabling inline send.
  86  */
  87 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
  88
  89 /*
  90  * Device parameter to configure the number of TX queues threshold for
  91  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
  92  */
  93 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
  94
  95 /* Device parameter to enable multi-packet send WQEs. */
  96 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
  97
  98 /*
  99  * Device parameter to include 2 dsegs in the title WQEBB.
 100  * Deprecated, ignored.
 101  */
 102 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
 103
 104 /*
 105  * Device parameter to limit the size of inlining packet.
 106  * Deprecated, ignored.
 107  */
 108 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
 109
 110 /*
 111  * Device parameter to enable hardware Tx vector.
 112  * Deprecated, ignored (no vectorized Tx routines anymore).
 113  */
 114 #define MLX5_TX_VEC_EN "tx_vec_en"
 115
 116 /* Device parameter to enable hardware Rx vector. */
 117 #define MLX5_RX_VEC_EN "rx_vec_en"
 118
 119 /* Allow L3 VXLAN flow creation. */
 120 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
 121
 122 /* Activate DV E-Switch flow steering. */
 123 #define MLX5_DV_ESW_EN "dv_esw_en"
 124
 125 /* Activate DV flow steering. */
 126 #define MLX5_DV_FLOW_EN "dv_flow_en"
 127
 128 /* Activate Netlink support in VF mode. */
 129 #define MLX5_VF_NL_EN "vf_nl_en"
 130
 131 /* Enable extending memsegs when creating a MR. */
 132 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
 133
 134 /* Select port representors to instantiate. */
 135 #define MLX5_REPRESENTOR "representor"
 136
 137 /* Device parameter to configure the maximum number of dump files per queue. */
 138 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
 139
 140 /* Configure timeout of LRO session (in microseconds). */
 141 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
 142
 143 #ifndef HAVE_IBV_MLX5_MOD_MPW
 144 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
 145 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
 146 #endif
 147
 148 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
 149 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
 150 #endif
 151
 152 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
 153
 154 /* Shared memory between primary and secondary processes. */
 155 struct mlx5_shared_data *mlx5_shared_data;
 156
 157 /* Spinlock for mlx5_shared_data allocation. */
 158 static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
 159
 160 /* Process local data for secondary processes. */
 161 static struct mlx5_local_data mlx5_local_data;
 162
 163 /** Driver-specific log messages type. */
 164 int mlx5_logtype;
 165
 166 /** Data associated with devices to spawn. */
 167 struct mlx5_dev_spawn_data {
 168         uint32_t ifindex; /**< Network interface index. */
 169         uint32_t max_port; /**< IB device maximal port index. */
 170         uint32_t ibv_port; /**< IB device physical port index. */
 171         int pf_bond; /**< bonding device PF index. < 0 - no bonding */
 172         struct mlx5_switch_info info; /**< Switch information. */
 173         struct ibv_device *ibv_dev; /**< Associated IB device. */
 174         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
 175         struct rte_pci_device *pci_dev; /**< Backend PCI device. */
 176 };
 177
 178 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
 179 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 180
 181 /**
 182  * Initialize the counters management structure.
 183  *
 184  * @param[in] sh
 185  *   Pointer to mlx5_ibv_shared object to free
 186  */
 187 static void
 188 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
 189 {
 190         uint8_t i;
 191
 192         TAILQ_INIT(&sh->cmng.flow_counters);
 193         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
 194                 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
 195 }
 196
 197 /**
 198  * Destroy all the resources allocated for a counter memory management.
 199  *
 200  * @param[in] mng
 201  *   Pointer to the memory management structure.
 202  */
 203 static void
 204 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
 205 {
 206         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
 207
 208         LIST_REMOVE(mng, next);
 209         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
 210         claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
 211         rte_free(mem);
 212 }
 213
 214 /**
 215  * Close and release all the resources of the counters management.
 216  *
 217  * @param[in] sh
 218  *   Pointer to mlx5_ibv_shared object to free.
 219  */
 220 static void
 221 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
 222 {
 223         struct mlx5_counter_stats_mem_mng *mng;
 224         uint8_t i;
 225         int j;
 226         int retries = 1024;
 227
 228         rte_errno = 0;
 229         while (--retries) {
 230                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
 231                 if (rte_errno != EINPROGRESS)
 232                         break;
 233                 rte_pause();
 234         }
 235         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
 236                 struct mlx5_flow_counter_pool *pool;
 237                 uint32_t batch = !!(i % 2);
 238
 239                 if (!sh->cmng.ccont[i].pools)
 240                         continue;
 241                 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 242                 while (pool) {
 243                         if (batch) {
 244                                 if (pool->min_dcs)
 245                                         claim_zero
 246                                         (mlx5_devx_cmd_destroy(pool->min_dcs));
 247                         }
 248                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
 249                                 if (pool->counters_raw[j].action)
 250                                         claim_zero
 251                                         (mlx5_glue->destroy_flow_action
 252                                                (pool->counters_raw[j].action));
 253                                 if (!batch && pool->counters_raw[j].dcs)
 254                                         claim_zero(mlx5_devx_cmd_destroy
 255                                                   (pool->counters_raw[j].dcs));
 256                         }
 257                         TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
 258                                      next);
 259                         rte_free(pool);
 260                         pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 261                 }
 262                 rte_free(sh->cmng.ccont[i].pools);
 263         }
 264         mng = LIST_FIRST(&sh->cmng.mem_mngs);
 265         while (mng) {
 266                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
 267                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
 268         }
 269         memset(&sh->cmng, 0, sizeof(sh->cmng));
 270 }
 271
 272 /**
 273  * Extract pdn of PD object using DV API.
 274  *
 275  * @param[in] pd
 276  *   Pointer to the verbs PD object.
 277  * @param[out] pdn
 278  *   Pointer to the PD object number variable.
 279  *
 280  * @return
 281  *   0 on success, error value otherwise.
 282  */
 283 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 284 static int
 285 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
 286 {
 287         struct mlx5dv_obj obj;
 288         struct mlx5dv_pd pd_info;
 289         int ret = 0;
 290
 291         obj.pd.in = pd;
 292         obj.pd.out = &pd_info;
 293         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 294         if (ret) {
 295                 DRV_LOG(DEBUG, "Fail to get PD object info");
 296                 return ret;
 297         }
 298         *pdn = pd_info.pdn;
 299         return 0;
 300 }
 301 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 302
 303 /**
 304  * Allocate shared IB device context. If there is multiport device the
 305  * master and representors will share this context, if there is single
 306  * port dedicated IB device, the context will be used by only given
 307  * port due to unification.
 308  *
 309  * Routine first searches the context for the specified IB device name,
 310  * if found the shared context assumed and reference counter is incremented.
 311  * If no context found the new one is created and initialized with specified
 312  * IB device context and parameters.
 313  *
 314  * @param[in] spawn
 315  *   Pointer to the IB device attributes (name, port, etc).
 316  *
 317  * @return
 318  *   Pointer to mlx5_ibv_shared object on success,
 319  *   otherwise NULL and rte_errno is set.
 320  */
 321 static struct mlx5_ibv_shared *
 322 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
 323 {
 324         struct mlx5_ibv_shared *sh;
 325         int err = 0;
 326         uint32_t i;
 327
 328         assert(spawn);
 329         /* Secondary process should not create the shared context. */
 330         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 331         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 332         /* Search for IB context by device name. */
 333         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
 334                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
 335                         sh->refcnt++;
 336                         goto exit;
 337                 }
 338         }
 339         /* No device found, we have to create new shared context. */
 340         assert(spawn->max_port);
 341         sh = rte_zmalloc("ethdev shared ib context",
 342                          sizeof(struct mlx5_ibv_shared) +
 343                          spawn->max_port *
 344                          sizeof(struct mlx5_ibv_shared_port),
 345                          RTE_CACHE_LINE_SIZE);
 346         if (!sh) {
 347                 DRV_LOG(ERR, "shared context allocation failure");
 348                 rte_errno  = ENOMEM;
 349                 goto exit;
 350         }
 351         /* Try to open IB device with DV first, then usual Verbs. */
 352         errno = 0;
 353         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
 354         if (sh->ctx) {
 355                 sh->devx = 1;
 356                 DRV_LOG(DEBUG, "DevX is supported");
 357         } else {
 358                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
 359                 if (!sh->ctx) {
 360                         err = errno ? errno : ENODEV;
 361                         goto error;
 362                 }
 363                 DRV_LOG(DEBUG, "DevX is NOT supported");
 364         }
 365         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
 366         if (err) {
 367                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
 368                 goto error;
 369         }
 370         sh->refcnt = 1;
 371         sh->max_port = spawn->max_port;
 372         strncpy(sh->ibdev_name, sh->ctx->device->name,
 373                 sizeof(sh->ibdev_name));
 374         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
 375                 sizeof(sh->ibdev_path));
 376         pthread_mutex_init(&sh->intr_mutex, NULL);
 377         /*
 378          * Setting port_id to max unallowed value means
 379          * there is no interrupt subhandler installed for
 380          * the given port index i.
 381          */
 382         for (i = 0; i < sh->max_port; i++)
 383                 sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
 384         sh->pd = mlx5_glue->alloc_pd(sh->ctx);
 385         if (sh->pd == NULL) {
 386                 DRV_LOG(ERR, "PD allocation failure");
 387                 err = ENOMEM;
 388                 goto error;
 389         }
 390 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 391         err = mlx5_get_pdn(sh->pd, &sh->pdn);
 392         if (err) {
 393                 DRV_LOG(ERR, "Fail to extract pdn from PD");
 394                 goto error;
 395         }
 396 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 397         /*
 398          * Once the device is added to the list of memory event
 399          * callback, its global MR cache table cannot be expanded
 400          * on the fly because of deadlock. If it overflows, lookup
 401          * should be done by searching MR list linearly, which is slow.
 402          *
 403          * At this point the device is not added to the memory
 404          * event list yet, context is just being created.
 405          */
 406         err = mlx5_mr_btree_init(&sh->mr.cache,
 407                                  MLX5_MR_BTREE_CACHE_N * 2,
 408                                  spawn->pci_dev->device.numa_node);
 409         if (err) {
 410                 err = rte_errno;
 411                 goto error;
 412         }
 413         mlx5_flow_counters_mng_init(sh);
 414         /* Add device to memory callback list. */
 415         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 416         LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
 417                          sh, mem_event_cb);
 418         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 419         /* Add context to the global device list. */
 420         LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
 421 exit:
 422         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 423         return sh;
 424 error:
 425         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 426         assert(sh);
 427         if (sh->pd)
 428                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 429         if (sh->ctx)
 430                 claim_zero(mlx5_glue->close_device(sh->ctx));
 431         rte_free(sh);
 432         assert(err > 0);
 433         rte_errno = err;
 434         return NULL;
 435 }
 436
 437 /**
 438  * Free shared IB device context. Decrement counter and if zero free
 439  * all allocated resources and close handles.
 440  *
 441  * @param[in] sh
 442  *   Pointer to mlx5_ibv_shared object to free
 443  */
 444 static void
 445 mlx5_free_shared_ibctx(struct mlx5_ibv_shared *sh)
 446 {
 447         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 448 #ifndef NDEBUG
 449         /* Check the object presence in the list. */
 450         struct mlx5_ibv_shared *lctx;
 451
 452         LIST_FOREACH(lctx, &mlx5_ibv_list, next)
 453                 if (lctx == sh)
 454                         break;
 455         assert(lctx);
 456         if (lctx != sh) {
 457                 DRV_LOG(ERR, "Freeing non-existing shared IB context");
 458                 goto exit;
 459         }
 460 #endif
 461         assert(sh);
 462         assert(sh->refcnt);
 463         /* Secondary process should not free the shared context. */
 464         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 465         if (--sh->refcnt)
 466                 goto exit;
 467         /* Release created Memory Regions. */
 468         mlx5_mr_release(sh);
 469         /* Remove from memory callback device list. */
 470         rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 471         LIST_REMOVE(sh, mem_event_cb);
 472         rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 473         /* Remove context from the global device list. */
 474         LIST_REMOVE(sh, next);
 475         /*
 476          *  Ensure there is no async event handler installed.
 477          *  Only primary process handles async device events.
 478          **/
 479         mlx5_flow_counters_mng_close(sh);
 480         assert(!sh->intr_cnt);
 481         if (sh->intr_cnt)
 482                 mlx5_intr_callback_unregister
 483                         (&sh->intr_handle, mlx5_dev_interrupt_handler, sh);
 484         pthread_mutex_destroy(&sh->intr_mutex);
 485         if (sh->pd)
 486                 claim_zero(mlx5_glue->dealloc_pd(sh->pd));
 487         if (sh->ctx)
 488                 claim_zero(mlx5_glue->close_device(sh->ctx));
 489         rte_free(sh);
 490 exit:
 491         pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 492 }
 493
 494 /**
 495  * Initialize DR related data within private structure.
 496  * Routine checks the reference counter and does actual
 497  * resources creation/initialization only if counter is zero.
 498  *
 499  * @param[in] priv
 500  *   Pointer to the private device data structure.
 501  *
 502  * @return
 503  *   Zero on success, positive error code otherwise.
 504  */
 505 static int
 506 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 507 {
 508 #ifdef HAVE_MLX5DV_DR
 509         struct mlx5_ibv_shared *sh = priv->sh;
 510         int err = 0;
 511         void *domain;
 512
 513         assert(sh);
 514         if (sh->dv_refcnt) {
 515                 /* Shared DV/DR structures is already initialized. */
 516                 sh->dv_refcnt++;
 517                 priv->dr_shared = 1;
 518                 return 0;
 519         }
 520         /* Reference counter is zero, we should initialize structures. */
 521         domain = mlx5_glue->dr_create_domain(sh->ctx,
 522                                              MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 523         if (!domain) {
 524                 DRV_LOG(ERR, "ingress mlx5dv_dr_create_domain failed");
 525                 err = errno;
 526                 goto error;
 527         }
 528         sh->rx_domain = domain;
 529         domain = mlx5_glue->dr_create_domain(sh->ctx,
 530                                              MLX5DV_DR_DOMAIN_TYPE_NIC_TX);
 531         if (!domain) {
 532                 DRV_LOG(ERR, "egress mlx5dv_dr_create_domain failed");
 533                 err = errno;
 534                 goto error;
 535         }
 536         pthread_mutex_init(&sh->dv_mutex, NULL);
 537         sh->tx_domain = domain;
 538 #ifdef HAVE_MLX5DV_DR_ESWITCH
 539         if (priv->config.dv_esw_en) {
 540                 domain  = mlx5_glue->dr_create_domain
 541                         (sh->ctx, MLX5DV_DR_DOMAIN_TYPE_FDB);
 542                 if (!domain) {
 543                         DRV_LOG(ERR, "FDB mlx5dv_dr_create_domain failed");
 544                         err = errno;
 545                         goto error;
 546                 }
 547                 sh->fdb_domain = domain;
 548                 sh->esw_drop_action = mlx5_glue->dr_create_flow_action_drop();
 549         }
 550 #endif
 551         sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
 552         sh->dv_refcnt++;
 553         priv->dr_shared = 1;
 554         return 0;
 555
 556 error:
 557        /* Rollback the created objects. */
 558         if (sh->rx_domain) {
 559                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 560                 sh->rx_domain = NULL;
 561         }
 562         if (sh->tx_domain) {
 563                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 564                 sh->tx_domain = NULL;
 565         }
 566         if (sh->fdb_domain) {
 567                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 568                 sh->fdb_domain = NULL;
 569         }
 570         if (sh->esw_drop_action) {
 571                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 572                 sh->esw_drop_action = NULL;
 573         }
 574         if (sh->pop_vlan_action) {
 575                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 576                 sh->pop_vlan_action = NULL;
 577         }
 578         return err;
 579 #else
 580         (void)priv;
 581         return 0;
 582 #endif
 583 }
 584
 585 /**
 586  * Destroy DR related data within private structure.
 587  *
 588  * @param[in] priv
 589  *   Pointer to the private device data structure.
 590  */
 591 static void
 592 mlx5_free_shared_dr(struct mlx5_priv *priv)
 593 {
 594 #ifdef HAVE_MLX5DV_DR
 595         struct mlx5_ibv_shared *sh;
 596
 597         if (!priv->dr_shared)
 598                 return;
 599         priv->dr_shared = 0;
 600         sh = priv->sh;
 601         assert(sh);
 602         assert(sh->dv_refcnt);
 603         if (sh->dv_refcnt && --sh->dv_refcnt)
 604                 return;
 605         if (sh->rx_domain) {
 606                 mlx5_glue->dr_destroy_domain(sh->rx_domain);
 607                 sh->rx_domain = NULL;
 608         }
 609         if (sh->tx_domain) {
 610                 mlx5_glue->dr_destroy_domain(sh->tx_domain);
 611                 sh->tx_domain = NULL;
 612         }
 613 #ifdef HAVE_MLX5DV_DR_ESWITCH
 614         if (sh->fdb_domain) {
 615                 mlx5_glue->dr_destroy_domain(sh->fdb_domain);
 616                 sh->fdb_domain = NULL;
 617         }
 618         if (sh->esw_drop_action) {
 619                 mlx5_glue->destroy_flow_action(sh->esw_drop_action);
 620                 sh->esw_drop_action = NULL;
 621         }
 622 #endif
 623         if (sh->pop_vlan_action) {
 624                 mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
 625                 sh->pop_vlan_action = NULL;
 626         }
 627         pthread_mutex_destroy(&sh->dv_mutex);
 628 #else
 629         (void)priv;
 630 #endif
 631 }
 632
 633 /**
 634  * Initialize shared data between primary and secondary process.
 635  *
 636  * A memzone is reserved by primary process and secondary processes attach to
 637  * the memzone.
 638  *
 639  * @return
 640  *   0 on success, a negative errno value otherwise and rte_errno is set.
 641  */
 642 static int
 643 mlx5_init_shared_data(void)
 644 {
 645         const struct rte_memzone *mz;
 646         int ret = 0;
 647
 648         rte_spinlock_lock(&mlx5_shared_data_lock);
 649         if (mlx5_shared_data == NULL) {
 650                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 651                         /* Allocate shared memory. */
 652                         mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
 653                                                  sizeof(*mlx5_shared_data),
 654                                                  SOCKET_ID_ANY, 0);
 655                         if (mz == NULL) {
 656                                 DRV_LOG(ERR,
 657                                         "Cannot allocate mlx5 shared data\n");
 658                                 ret = -rte_errno;
 659                                 goto error;
 660                         }
 661                         mlx5_shared_data = mz->addr;
 662                         memset(mlx5_shared_data, 0, sizeof(*mlx5_shared_data));
 663                         rte_spinlock_init(&mlx5_shared_data->lock);
 664                 } else {
 665                         /* Lookup allocated shared memory. */
 666                         mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
 667                         if (mz == NULL) {
 668                                 DRV_LOG(ERR,
 669                                         "Cannot attach mlx5 shared data\n");
 670                                 ret = -rte_errno;
 671                                 goto error;
 672                         }
 673                         mlx5_shared_data = mz->addr;
 674                         memset(&mlx5_local_data, 0, sizeof(mlx5_local_data));
 675                 }
 676         }
 677 error:
 678         rte_spinlock_unlock(&mlx5_shared_data_lock);
 679         return ret;
 680 }
 681
 682 /**
 683  * Retrieve integer value from environment variable.
 684  *
 685  * @param[in] name
 686  *   Environment variable name.
 687  *
 688  * @return
 689  *   Integer value, 0 if the variable is not set.
 690  */
 691 int
 692 mlx5_getenv_int(const char *name)
 693 {
 694         const char *val = getenv(name);
 695
 696         if (val == NULL)
 697                 return 0;
 698         return atoi(val);
 699 }
 700
 701 /**
 702  * Verbs callback to allocate a memory. This function should allocate the space
 703  * according to the size provided residing inside a huge page.
 704  * Please note that all allocation must respect the alignment from libmlx5
 705  * (i.e. currently sysconf(_SC_PAGESIZE)).
 706  *
 707  * @param[in] size
 708  *   The size in bytes of the memory to allocate.
 709  * @param[in] data
 710  *   A pointer to the callback data.
 711  *
 712  * @return
 713  *   Allocated buffer, NULL otherwise and rte_errno is set.
 714  */
 715 static void *
 716 mlx5_alloc_verbs_buf(size_t size, void *data)
 717 {
 718         struct mlx5_priv *priv = data;
 719         void *ret;
 720         size_t alignment = sysconf(_SC_PAGESIZE);
 721         unsigned int socket = SOCKET_ID_ANY;
 722
 723         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 724                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 725
 726                 socket = ctrl->socket;
 727         } else if (priv->verbs_alloc_ctx.type ==
 728                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 729                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 730
 731                 socket = ctrl->socket;
 732         }
 733         assert(data != NULL);
 734         ret = rte_malloc_socket(__func__, size, alignment, socket);
 735         if (!ret && size)
 736                 rte_errno = ENOMEM;
 737         return ret;
 738 }
 739
 740 /**
 741  * Verbs callback to free a memory.
 742  *
 743  * @param[in] ptr
 744  *   A pointer to the memory to free.
 745  * @param[in] data
 746  *   A pointer to the callback data.
 747  */
 748 static void
 749 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 750 {
 751         assert(data != NULL);
 752         rte_free(ptr);
 753 }
 754
 755 /**
 756  * DPDK callback to add udp tunnel port
 757  *
 758  * @param[in] dev
 759  *   A pointer to eth_dev
 760  * @param[in] udp_tunnel
 761  *   A pointer to udp tunnel
 762  *
 763  * @return
 764  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
 765  */
 766 int
 767 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
 768                          struct rte_eth_udp_tunnel *udp_tunnel)
 769 {
 770         assert(udp_tunnel != NULL);
 771         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
 772             udp_tunnel->udp_port == 4789)
 773                 return 0;
 774         if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
 775             udp_tunnel->udp_port == 4790)
 776                 return 0;
 777         return -ENOTSUP;
 778 }
 779
 780 /**
 781  * Initialize process private data structure.
 782  *
 783  * @param dev
 784  *   Pointer to Ethernet device structure.
 785  *
 786  * @return
 787  *   0 on success, a negative errno value otherwise and rte_errno is set.
 788  */
 789 int
 790 mlx5_proc_priv_init(struct rte_eth_dev *dev)
 791 {
 792         struct mlx5_priv *priv = dev->data->dev_private;
 793         struct mlx5_proc_priv *ppriv;
 794         size_t ppriv_size;
 795
 796         /*
 797          * UAR register table follows the process private structure. BlueFlame
 798          * registers for Tx queues are stored in the table.
 799          */
 800         ppriv_size =
 801                 sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
 802         ppriv = rte_malloc_socket("mlx5_proc_priv", ppriv_size,
 803                                   RTE_CACHE_LINE_SIZE, dev->device->numa_node);
 804         if (!ppriv) {
 805                 rte_errno = ENOMEM;
 806                 return -rte_errno;
 807         }
 808         ppriv->uar_table_sz = ppriv_size;
 809         dev->process_private = ppriv;
 810         return 0;
 811 }
 812
 813 /**
 814  * Un-initialize process private data structure.
 815  *
 816  * @param dev
 817  *   Pointer to Ethernet device structure.
 818  */
 819 static void
 820 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
 821 {
 822         if (!dev->process_private)
 823                 return;
 824         rte_free(dev->process_private);
 825         dev->process_private = NULL;
 826 }
 827
 828 /**
 829  * DPDK callback to close the device.
 830  *
 831  * Destroy all queues and objects, free memory.
 832  *
 833  * @param dev
 834  *   Pointer to Ethernet device structure.
 835  */
 836 static void
 837 mlx5_dev_close(struct rte_eth_dev *dev)
 838 {
 839         struct mlx5_priv *priv = dev->data->dev_private;
 840         unsigned int i;
 841         int ret;
 842
 843         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
 844                 dev->data->port_id,
 845                 ((priv->sh->ctx != NULL) ? priv->sh->ctx->device->name : ""));
 846         /* In case mlx5_dev_stop() has not been called. */
 847         mlx5_dev_interrupt_handler_uninstall(dev);
 848         mlx5_traffic_disable(dev);
 849         mlx5_flow_flush(dev, NULL);
 850         /* Prevent crashes when queues are still in use. */
 851         dev->rx_pkt_burst = removed_rx_burst;
 852         dev->tx_pkt_burst = removed_tx_burst;
 853         rte_wmb();
 854         /* Disable datapath on secondary process. */
 855         mlx5_mp_req_stop_rxtx(dev);
 856         if (priv->rxqs != NULL) {
 857                 /* XXX race condition if mlx5_rx_burst() is still running. */
 858                 usleep(1000);
 859                 for (i = 0; (i != priv->rxqs_n); ++i)
 860                         mlx5_rxq_release(dev, i);
 861                 priv->rxqs_n = 0;
 862                 priv->rxqs = NULL;
 863         }
 864         if (priv->txqs != NULL) {
 865                 /* XXX race condition if mlx5_tx_burst() is still running. */
 866                 usleep(1000);
 867                 for (i = 0; (i != priv->txqs_n); ++i)
 868                         mlx5_txq_release(dev, i);
 869                 priv->txqs_n = 0;
 870                 priv->txqs = NULL;
 871         }
 872         mlx5_proc_priv_uninit(dev);
 873         mlx5_mprq_free_mp(dev);
 874         mlx5_free_shared_dr(priv);
 875         if (priv->rss_conf.rss_key != NULL)
 876                 rte_free(priv->rss_conf.rss_key);
 877         if (priv->reta_idx != NULL)
 878                 rte_free(priv->reta_idx);
 879         if (priv->config.vf)
 880                 mlx5_nl_mac_addr_flush(dev);
 881         if (priv->nl_socket_route >= 0)
 882                 close(priv->nl_socket_route);
 883         if (priv->nl_socket_rdma >= 0)
 884                 close(priv->nl_socket_rdma);
 885         if (priv->vmwa_context)
 886                 mlx5_vlan_vmwa_exit(priv->vmwa_context);
 887         if (priv->sh) {
 888                 /*
 889                  * Free the shared context in last turn, because the cleanup
 890                  * routines above may use some shared fields, like
 891                  * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
 892                  * ifindex if Netlink fails.
 893                  */
 894                 mlx5_free_shared_ibctx(priv->sh);
 895                 priv->sh = NULL;
 896         }
 897         ret = mlx5_hrxq_verify(dev);
 898         if (ret)
 899                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
 900                         dev->data->port_id);
 901         ret = mlx5_ind_table_obj_verify(dev);
 902         if (ret)
 903                 DRV_LOG(WARNING, "port %u some indirection table still remain",
 904                         dev->data->port_id);
 905         ret = mlx5_rxq_obj_verify(dev);
 906         if (ret)
 907                 DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
 908                         dev->data->port_id);
 909         ret = mlx5_rxq_verify(dev);
 910         if (ret)
 911                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
 912                         dev->data->port_id);
 913         ret = mlx5_txq_ibv_verify(dev);
 914         if (ret)
 915                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
 916                         dev->data->port_id);
 917         ret = mlx5_txq_verify(dev);
 918         if (ret)
 919                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
 920                         dev->data->port_id);
 921         ret = mlx5_flow_verify(dev);
 922         if (ret)
 923                 DRV_LOG(WARNING, "port %u some flows still remain",
 924                         dev->data->port_id);
 925         if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
 926                 unsigned int c = 0;
 927                 uint16_t port_id;
 928
 929                 RTE_ETH_FOREACH_DEV_OF(port_id, dev->device) {
 930                         struct mlx5_priv *opriv =
 931                                 rte_eth_devices[port_id].data->dev_private;
 932
 933                         if (!opriv ||
 934                             opriv->domain_id != priv->domain_id ||
 935                             &rte_eth_devices[port_id] == dev)
 936                                 continue;
 937                         ++c;
 938                 }
 939                 if (!c)
 940                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
 941         }
 942         memset(priv, 0, sizeof(*priv));
 943         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
 944         /*
 945          * Reset mac_addrs to NULL such that it is not freed as part of
 946          * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
 947          * it is freed when dev_private is freed.
 948          */
 949         dev->data->mac_addrs = NULL;
 950 }
 951
 952 const struct eth_dev_ops mlx5_dev_ops = {
 953         .dev_configure = mlx5_dev_configure,
 954         .dev_start = mlx5_dev_start,
 955         .dev_stop = mlx5_dev_stop,
 956         .dev_set_link_down = mlx5_set_link_down,
 957         .dev_set_link_up = mlx5_set_link_up,
 958         .dev_close = mlx5_dev_close,
 959         .promiscuous_enable = mlx5_promiscuous_enable,
 960         .promiscuous_disable = mlx5_promiscuous_disable,
 961         .allmulticast_enable = mlx5_allmulticast_enable,
 962         .allmulticast_disable = mlx5_allmulticast_disable,
 963         .link_update = mlx5_link_update,
 964         .stats_get = mlx5_stats_get,
 965         .stats_reset = mlx5_stats_reset,
 966         .xstats_get = mlx5_xstats_get,
 967         .xstats_reset = mlx5_xstats_reset,
 968         .xstats_get_names = mlx5_xstats_get_names,
 969         .fw_version_get = mlx5_fw_version_get,
 970         .dev_infos_get = mlx5_dev_infos_get,
 971         .read_clock = mlx5_read_clock,
 972         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 973         .vlan_filter_set = mlx5_vlan_filter_set,
 974         .rx_queue_setup = mlx5_rx_queue_setup,
 975         .tx_queue_setup = mlx5_tx_queue_setup,
 976         .rx_queue_release = mlx5_rx_queue_release,
 977         .tx_queue_release = mlx5_tx_queue_release,
 978         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 979         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 980         .mac_addr_remove = mlx5_mac_addr_remove,
 981         .mac_addr_add = mlx5_mac_addr_add,
 982         .mac_addr_set = mlx5_mac_addr_set,
 983         .set_mc_addr_list = mlx5_set_mc_addr_list,
 984         .mtu_set = mlx5_dev_set_mtu,
 985         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 986         .vlan_offload_set = mlx5_vlan_offload_set,
 987         .reta_update = mlx5_dev_rss_reta_update,
 988         .reta_query = mlx5_dev_rss_reta_query,
 989         .rss_hash_update = mlx5_rss_hash_update,
 990         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 991         .filter_ctrl = mlx5_dev_filter_ctrl,
 992         .rx_descriptor_status = mlx5_rx_descriptor_status,
 993         .tx_descriptor_status = mlx5_tx_descriptor_status,
 994         .rx_queue_count = mlx5_rx_queue_count,
 995         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 996         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 997         .is_removed = mlx5_is_removed,
 998         .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
 999         .get_module_info = mlx5_get_module_info,
1000         .get_module_eeprom = mlx5_get_module_eeprom,
1001 };
1002
1003 /* Available operations from secondary process. */
1004 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1005         .stats_get = mlx5_stats_get,
1006         .stats_reset = mlx5_stats_reset,
1007         .xstats_get = mlx5_xstats_get,
1008         .xstats_reset = mlx5_xstats_reset,
1009         .xstats_get_names = mlx5_xstats_get_names,
1010         .fw_version_get = mlx5_fw_version_get,
1011         .dev_infos_get = mlx5_dev_infos_get,
1012         .rx_descriptor_status = mlx5_rx_descriptor_status,
1013         .tx_descriptor_status = mlx5_tx_descriptor_status,
1014         .get_module_info = mlx5_get_module_info,
1015         .get_module_eeprom = mlx5_get_module_eeprom,
1016 };
1017
1018 /* Available operations in flow isolated mode. */
1019 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1020         .dev_configure = mlx5_dev_configure,
1021         .dev_start = mlx5_dev_start,
1022         .dev_stop = mlx5_dev_stop,
1023         .dev_set_link_down = mlx5_set_link_down,
1024         .dev_set_link_up = mlx5_set_link_up,
1025         .dev_close = mlx5_dev_close,
1026         .promiscuous_enable = mlx5_promiscuous_enable,
1027         .promiscuous_disable = mlx5_promiscuous_disable,
1028         .allmulticast_enable = mlx5_allmulticast_enable,
1029         .allmulticast_disable = mlx5_allmulticast_disable,
1030         .link_update = mlx5_link_update,
1031         .stats_get = mlx5_stats_get,
1032         .stats_reset = mlx5_stats_reset,
1033         .xstats_get = mlx5_xstats_get,
1034         .xstats_reset = mlx5_xstats_reset,
1035         .xstats_get_names = mlx5_xstats_get_names,
1036         .fw_version_get = mlx5_fw_version_get,
1037         .dev_infos_get = mlx5_dev_infos_get,
1038         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1039         .vlan_filter_set = mlx5_vlan_filter_set,
1040         .rx_queue_setup = mlx5_rx_queue_setup,
1041         .tx_queue_setup = mlx5_tx_queue_setup,
1042         .rx_queue_release = mlx5_rx_queue_release,
1043         .tx_queue_release = mlx5_tx_queue_release,
1044         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1045         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1046         .mac_addr_remove = mlx5_mac_addr_remove,
1047         .mac_addr_add = mlx5_mac_addr_add,
1048         .mac_addr_set = mlx5_mac_addr_set,
1049         .set_mc_addr_list = mlx5_set_mc_addr_list,
1050         .mtu_set = mlx5_dev_set_mtu,
1051         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1052         .vlan_offload_set = mlx5_vlan_offload_set,
1053         .filter_ctrl = mlx5_dev_filter_ctrl,
1054         .rx_descriptor_status = mlx5_rx_descriptor_status,
1055         .tx_descriptor_status = mlx5_tx_descriptor_status,
1056         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1057         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1058         .is_removed = mlx5_is_removed,
1059         .get_module_info = mlx5_get_module_info,
1060         .get_module_eeprom = mlx5_get_module_eeprom,
1061 };
1062
1063 /**
1064  * Verify and store value for device argument.
1065  *
1066  * @param[in] key
1067  *   Key argument to verify.
1068  * @param[in] val
1069  *   Value associated with key.
1070  * @param opaque
1071  *   User data.
1072  *
1073  * @return
1074  *   0 on success, a negative errno value otherwise and rte_errno is set.
1075  */
1076 static int
1077 mlx5_args_check(const char *key, const char *val, void *opaque)
1078 {
1079         struct mlx5_dev_config *config = opaque;
1080         unsigned long tmp;
1081
1082         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1083         if (!strcmp(MLX5_REPRESENTOR, key))
1084                 return 0;
1085         errno = 0;
1086         tmp = strtoul(val, NULL, 0);
1087         if (errno) {
1088                 rte_errno = errno;
1089                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1090                 return -rte_errno;
1091         }
1092         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1093                 config->cqe_comp = !!tmp;
1094         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1095                 config->cqe_pad = !!tmp;
1096         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1097                 config->hw_padding = !!tmp;
1098         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1099                 config->mprq.enabled = !!tmp;
1100         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1101                 config->mprq.stride_num_n = tmp;
1102         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1103                 config->mprq.max_memcpy_len = tmp;
1104         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1105                 config->mprq.min_rxqs_num = tmp;
1106         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1107                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1108                                  " converted to txq_inline_max", key);
1109                 config->txq_inline_max = tmp;
1110         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1111                 config->txq_inline_max = tmp;
1112         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1113                 config->txq_inline_min = tmp;
1114         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1115                 config->txq_inline_mpw = tmp;
1116         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1117                 config->txqs_inline = tmp;
1118         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1119                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1120         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1121                 config->mps = !!tmp;
1122         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1123                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1124         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1125                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1126                                  " converted to txq_inline_mpw", key);
1127                 config->txq_inline_mpw = tmp;
1128         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1129                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1130         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1131                 config->rx_vec_en = !!tmp;
1132         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1133                 config->l3_vxlan_en = !!tmp;
1134         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1135                 config->vf_nl_en = !!tmp;
1136         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1137                 config->dv_esw_en = !!tmp;
1138         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1139                 config->dv_flow_en = !!tmp;
1140         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1141                 config->mr_ext_memseg_en = !!tmp;
1142         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1143                 config->max_dump_files_num = tmp;
1144         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1145                 config->lro.timeout = tmp;
1146         } else {
1147                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1148                 rte_errno = EINVAL;
1149                 return -rte_errno;
1150         }
1151         return 0;
1152 }
1153
1154 /**
1155  * Parse device parameters.
1156  *
1157  * @param config
1158  *   Pointer to device configuration structure.
1159  * @param devargs
1160  *   Device arguments structure.
1161  *
1162  * @return
1163  *   0 on success, a negative errno value otherwise and rte_errno is set.
1164  */
1165 static int
1166 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1167 {
1168         const char **params = (const char *[]){
1169                 MLX5_RXQ_CQE_COMP_EN,
1170                 MLX5_RXQ_CQE_PAD_EN,
1171                 MLX5_RXQ_PKT_PAD_EN,
1172                 MLX5_RX_MPRQ_EN,
1173                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1174                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1175                 MLX5_RXQS_MIN_MPRQ,
1176                 MLX5_TXQ_INLINE,
1177                 MLX5_TXQ_INLINE_MIN,
1178                 MLX5_TXQ_INLINE_MAX,
1179                 MLX5_TXQ_INLINE_MPW,
1180                 MLX5_TXQS_MIN_INLINE,
1181                 MLX5_TXQS_MAX_VEC,
1182                 MLX5_TXQ_MPW_EN,
1183                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1184                 MLX5_TXQ_MAX_INLINE_LEN,
1185                 MLX5_TX_VEC_EN,
1186                 MLX5_RX_VEC_EN,
1187                 MLX5_L3_VXLAN_EN,
1188                 MLX5_VF_NL_EN,
1189                 MLX5_DV_ESW_EN,
1190                 MLX5_DV_FLOW_EN,
1191                 MLX5_MR_EXT_MEMSEG_EN,
1192                 MLX5_REPRESENTOR,
1193                 MLX5_MAX_DUMP_FILES_NUM,
1194                 MLX5_LRO_TIMEOUT_USEC,
1195                 NULL,
1196         };
1197         struct rte_kvargs *kvlist;
1198         int ret = 0;
1199         int i;
1200
1201         if (devargs == NULL)
1202                 return 0;
1203         /* Following UGLY cast is done to pass checkpatch. */
1204         kvlist = rte_kvargs_parse(devargs->args, params);
1205         if (kvlist == NULL) {
1206                 rte_errno = EINVAL;
1207                 return -rte_errno;
1208         }
1209         /* Process parameters. */
1210         for (i = 0; (params[i] != NULL); ++i) {
1211                 if (rte_kvargs_count(kvlist, params[i])) {
1212                         ret = rte_kvargs_process(kvlist, params[i],
1213                                                  mlx5_args_check, config);
1214                         if (ret) {
1215                                 rte_errno = EINVAL;
1216                                 rte_kvargs_free(kvlist);
1217                                 return -rte_errno;
1218                         }
1219                 }
1220         }
1221         rte_kvargs_free(kvlist);
1222         return 0;
1223 }
1224
1225 static struct rte_pci_driver mlx5_driver;
1226
1227 /**
1228  * PMD global initialization.
1229  *
1230  * Independent from individual device, this function initializes global
1231  * per-PMD data structures distinguishing primary and secondary processes.
1232  * Hence, each initialization is called once per a process.
1233  *
1234  * @return
1235  *   0 on success, a negative errno value otherwise and rte_errno is set.
1236  */
1237 static int
1238 mlx5_init_once(void)
1239 {
1240         struct mlx5_shared_data *sd;
1241         struct mlx5_local_data *ld = &mlx5_local_data;
1242         int ret = 0;
1243
1244         if (mlx5_init_shared_data())
1245                 return -rte_errno;
1246         sd = mlx5_shared_data;
1247         assert(sd);
1248         rte_spinlock_lock(&sd->lock);
1249         switch (rte_eal_process_type()) {
1250         case RTE_PROC_PRIMARY:
1251                 if (sd->init_done)
1252                         break;
1253                 LIST_INIT(&sd->mem_event_cb_list);
1254                 rte_rwlock_init(&sd->mem_event_rwlock);
1255                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1256                                                 mlx5_mr_mem_event_cb, NULL);
1257                 ret = mlx5_mp_init_primary();
1258                 if (ret)
1259                         goto out;
1260                 sd->init_done = true;
1261                 break;
1262         case RTE_PROC_SECONDARY:
1263                 if (ld->init_done)
1264                         break;
1265                 ret = mlx5_mp_init_secondary();
1266                 if (ret)
1267                         goto out;
1268                 ++sd->secondary_cnt;
1269                 ld->init_done = true;
1270                 break;
1271         default:
1272                 break;
1273         }
1274 out:
1275         rte_spinlock_unlock(&sd->lock);
1276         return ret;
1277 }
1278
1279 /**
1280  * Configures the minimal amount of data to inline into WQE
1281  * while sending packets.
1282  *
1283  * - the txq_inline_min has the maximal priority, if this
1284  *   key is specified in devargs
1285  * - if DevX is enabled the inline mode is queried from the
1286  *   device (HCA attributes and NIC vport context if needed).
1287  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1288  *   and none (0 bytes) for other NICs
1289  *
1290  * @param spawn
1291  *   Verbs device parameters (name, port, switch_info) to spawn.
1292  * @param config
1293  *   Device configuration parameters.
1294  */
1295 static void
1296 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1297                     struct mlx5_dev_config *config)
1298 {
1299         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1300                 /* Application defines size of inlined data explicitly. */
1301                 switch (spawn->pci_dev->id.device_id) {
1302                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1303                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1304                         if (config->txq_inline_min <
1305                                        (int)MLX5_INLINE_HSIZE_L2) {
1306                                 DRV_LOG(DEBUG,
1307                                         "txq_inline_mix aligned to minimal"
1308                                         " ConnectX-4 required value %d",
1309                                         (int)MLX5_INLINE_HSIZE_L2);
1310                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1311                         }
1312                         break;
1313                 }
1314                 goto exit;
1315         }
1316         if (config->hca_attr.eth_net_offloads) {
1317                 /* We have DevX enabled, inline mode queried successfully. */
1318                 switch (config->hca_attr.wqe_inline_mode) {
1319                 case MLX5_CAP_INLINE_MODE_L2:
1320                         /* outer L2 header must be inlined. */
1321                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1322                         goto exit;
1323                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1324                         /* No inline data are required by NIC. */
1325                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1326                         config->hw_vlan_insert =
1327                                 config->hca_attr.wqe_vlan_insert;
1328                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1329                         goto exit;
1330                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1331                         /* inline mode is defined by NIC vport context. */
1332                         if (!config->hca_attr.eth_virt)
1333                                 break;
1334                         switch (config->hca_attr.vport_inline_mode) {
1335                         case MLX5_INLINE_MODE_NONE:
1336                                 config->txq_inline_min =
1337                                         MLX5_INLINE_HSIZE_NONE;
1338                                 goto exit;
1339                         case MLX5_INLINE_MODE_L2:
1340                                 config->txq_inline_min =
1341                                         MLX5_INLINE_HSIZE_L2;
1342                                 goto exit;
1343                         case MLX5_INLINE_MODE_IP:
1344                                 config->txq_inline_min =
1345                                         MLX5_INLINE_HSIZE_L3;
1346                                 goto exit;
1347                         case MLX5_INLINE_MODE_TCP_UDP:
1348                                 config->txq_inline_min =
1349                                         MLX5_INLINE_HSIZE_L4;
1350                                 goto exit;
1351                         case MLX5_INLINE_MODE_INNER_L2:
1352                                 config->txq_inline_min =
1353                                         MLX5_INLINE_HSIZE_INNER_L2;
1354                                 goto exit;
1355                         case MLX5_INLINE_MODE_INNER_IP:
1356                                 config->txq_inline_min =
1357                                         MLX5_INLINE_HSIZE_INNER_L3;
1358                                 goto exit;
1359                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1360                                 config->txq_inline_min =
1361                                         MLX5_INLINE_HSIZE_INNER_L4;
1362                                 goto exit;
1363                         }
1364                 }
1365         }
1366         /*
1367          * We get here if we are unable to deduce
1368          * inline data size with DevX. Try PCI ID
1369          * to determine old NICs.
1370          */
1371         switch (spawn->pci_dev->id.device_id) {
1372         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1373         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1374         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1375         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1376                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1377                 config->hw_vlan_insert = 0;
1378                 break;
1379         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1380         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1381         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1382         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1383                 /*
1384                  * These NICs support VLAN insertion from WQE and
1385                  * report the wqe_vlan_insert flag. But there is the bug
1386                  * and PFC control may be broken, so disable feature.
1387                  */
1388                 config->hw_vlan_insert = 0;
1389                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1390                 break;
1391         default:
1392                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1393                 break;
1394         }
1395 exit:
1396         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1397 }
1398
1399 /**
1400  * Allocate page of door-bells and register it using DevX API.
1401  *
1402  * @param [in] dev
1403  *   Pointer to Ethernet device.
1404  *
1405  * @return
1406  *   Pointer to new page on success, NULL otherwise.
1407  */
1408 static struct mlx5_devx_dbr_page *
1409 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1410 {
1411         struct mlx5_priv *priv = dev->data->dev_private;
1412         struct mlx5_devx_dbr_page *page;
1413
1414         /* Allocate space for door-bell page and management data. */
1415         page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1416                                  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1417         if (!page) {
1418                 DRV_LOG(ERR, "port %u cannot allocate dbr page",
1419                         dev->data->port_id);
1420                 return NULL;
1421         }
1422         /* Register allocated memory. */
1423         page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1424                                               MLX5_DBR_PAGE_SIZE, 0);
1425         if (!page->umem) {
1426                 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1427                         dev->data->port_id);
1428                 rte_free(page);
1429                 return NULL;
1430         }
1431         return page;
1432 }
1433
1434 /**
1435  * Find the next available door-bell, allocate new page if needed.
1436  *
1437  * @param [in] dev
1438  *   Pointer to Ethernet device.
1439  * @param [out] dbr_page
1440  *   Door-bell page containing the page data.
1441  *
1442  * @return
1443  *   Door-bell address offset on success, a negative error value otherwise.
1444  */
1445 int64_t
1446 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1447 {
1448         struct mlx5_priv *priv = dev->data->dev_private;
1449         struct mlx5_devx_dbr_page *page = NULL;
1450         uint32_t i, j;
1451
1452         LIST_FOREACH(page, &priv->dbrpgs, next)
1453                 if (page->dbr_count < MLX5_DBR_PER_PAGE)
1454                         break;
1455         if (!page) { /* No page with free door-bell exists. */
1456                 page = mlx5_alloc_dbr_page(dev);
1457                 if (!page) /* Failed to allocate new page. */
1458                         return (-1);
1459                 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1460         }
1461         /* Loop to find bitmap part with clear bit. */
1462         for (i = 0;
1463              i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1464              i++)
1465                 ; /* Empty. */
1466         /* Find the first clear bit. */
1467         j = rte_bsf64(~page->dbr_bitmap[i]);
1468         assert(i < (MLX5_DBR_PER_PAGE / 64));
1469         page->dbr_bitmap[i] |= (1 << j);
1470         page->dbr_count++;
1471         *dbr_page = page;
1472         return (((i * 64) + j) * sizeof(uint64_t));
1473 }
1474
1475 /**
1476  * Release a door-bell record.
1477  *
1478  * @param [in] dev
1479  *   Pointer to Ethernet device.
1480  * @param [in] umem_id
1481  *   UMEM ID of page containing the door-bell record to release.
1482  * @param [in] offset
1483  *   Offset of door-bell record in page.
1484  *
1485  * @return
1486  *   0 on success, a negative error value otherwise.
1487  */
1488 int32_t
1489 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1490 {
1491         struct mlx5_priv *priv = dev->data->dev_private;
1492         struct mlx5_devx_dbr_page *page = NULL;
1493         int ret = 0;
1494
1495         LIST_FOREACH(page, &priv->dbrpgs, next)
1496                 /* Find the page this address belongs to. */
1497                 if (page->umem->umem_id == umem_id)
1498                         break;
1499         if (!page)
1500                 return -EINVAL;
1501         page->dbr_count--;
1502         if (!page->dbr_count) {
1503                 /* Page not used, free it and remove from list. */
1504                 LIST_REMOVE(page, next);
1505                 if (page->umem)
1506                         ret = -mlx5_glue->devx_umem_dereg(page->umem);
1507                 rte_free(page);
1508         } else {
1509                 /* Mark in bitmap that this door-bell is not in use. */
1510                 offset /= MLX5_DBR_SIZE;
1511                 int i = offset / 64;
1512                 int j = offset % 64;
1513
1514                 page->dbr_bitmap[i] &= ~(1 << j);
1515         }
1516         return ret;
1517 }
1518
1519 /**
1520  * Spawn an Ethernet device from Verbs information.
1521  *
1522  * @param dpdk_dev
1523  *   Backing DPDK device.
1524  * @param spawn
1525  *   Verbs device parameters (name, port, switch_info) to spawn.
1526  * @param config
1527  *   Device configuration parameters.
1528  *
1529  * @return
1530  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1531  *   is set. The following errors are defined:
1532  *
1533  *   EBUSY: device is not supposed to be spawned.
1534  *   EEXIST: device is already spawned
1535  */
1536 static struct rte_eth_dev *
1537 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1538                struct mlx5_dev_spawn_data *spawn,
1539                struct mlx5_dev_config config)
1540 {
1541         const struct mlx5_switch_info *switch_info = &spawn->info;
1542         struct mlx5_ibv_shared *sh = NULL;
1543         struct ibv_port_attr port_attr;
1544         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1545         struct rte_eth_dev *eth_dev = NULL;
1546         struct mlx5_priv *priv = NULL;
1547         int err = 0;
1548         unsigned int hw_padding = 0;
1549         unsigned int mps;
1550         unsigned int cqe_comp;
1551         unsigned int cqe_pad = 0;
1552         unsigned int tunnel_en = 0;
1553         unsigned int mpls_en = 0;
1554         unsigned int swp = 0;
1555         unsigned int mprq = 0;
1556         unsigned int mprq_min_stride_size_n = 0;
1557         unsigned int mprq_max_stride_size_n = 0;
1558         unsigned int mprq_min_stride_num_n = 0;
1559         unsigned int mprq_max_stride_num_n = 0;
1560         struct rte_ether_addr mac;
1561         char name[RTE_ETH_NAME_MAX_LEN];
1562         int own_domain_id = 0;
1563         uint16_t port_id;
1564         unsigned int i;
1565
1566         /* Determine if this port representor is supposed to be spawned. */
1567         if (switch_info->representor && dpdk_dev->devargs) {
1568                 struct rte_eth_devargs eth_da;
1569
1570                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1571                 if (err) {
1572                         rte_errno = -err;
1573                         DRV_LOG(ERR, "failed to process device arguments: %s",
1574                                 strerror(rte_errno));
1575                         return NULL;
1576                 }
1577                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1578                         if (eth_da.representor_ports[i] ==
1579                             (uint16_t)switch_info->port_name)
1580                                 break;
1581                 if (i == eth_da.nb_representor_ports) {
1582                         rte_errno = EBUSY;
1583                         return NULL;
1584                 }
1585         }
1586         /* Build device name. */
1587         if (spawn->pf_bond <  0) {
1588                 /* Single device. */
1589                 if (!switch_info->representor)
1590                         strlcpy(name, dpdk_dev->name, sizeof(name));
1591                 else
1592                         snprintf(name, sizeof(name), "%s_representor_%u",
1593                                  dpdk_dev->name, switch_info->port_name);
1594         } else {
1595                 /* Bonding device. */
1596                 if (!switch_info->representor)
1597                         snprintf(name, sizeof(name), "%s_%s",
1598                                  dpdk_dev->name, spawn->ibv_dev->name);
1599                 else
1600                         snprintf(name, sizeof(name), "%s_%s_representor_%u",
1601                                  dpdk_dev->name, spawn->ibv_dev->name,
1602                                  switch_info->port_name);
1603         }
1604         /* check if the device is already spawned */
1605         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1606                 rte_errno = EEXIST;
1607                 return NULL;
1608         }
1609         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1610         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1611                 eth_dev = rte_eth_dev_attach_secondary(name);
1612                 if (eth_dev == NULL) {
1613                         DRV_LOG(ERR, "can not attach rte ethdev");
1614                         rte_errno = ENOMEM;
1615                         return NULL;
1616                 }
1617                 eth_dev->device = dpdk_dev;
1618                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1619                 err = mlx5_proc_priv_init(eth_dev);
1620                 if (err)
1621                         return NULL;
1622                 /* Receive command fd from primary process */
1623                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1624                 if (err < 0)
1625                         return NULL;
1626                 /* Remap UAR for Tx queues. */
1627                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1628                 if (err)
1629                         return NULL;
1630                 /*
1631                  * Ethdev pointer is still required as input since
1632                  * the primary device is not accessible from the
1633                  * secondary process.
1634                  */
1635                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1636                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1637                 return eth_dev;
1638         }
1639         sh = mlx5_alloc_shared_ibctx(spawn);
1640         if (!sh)
1641                 return NULL;
1642         config.devx = sh->devx;
1643 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1644         config.dest_tir = 1;
1645 #endif
1646 #ifdef HAVE_IBV_MLX5_MOD_SWP
1647         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1648 #endif
1649         /*
1650          * Multi-packet send is supported by ConnectX-4 Lx PF as well
1651          * as all ConnectX-5 devices.
1652          */
1653 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1654         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1655 #endif
1656 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1657         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1658 #endif
1659         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1660         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1661                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1662                         DRV_LOG(DEBUG, "enhanced MPW is supported");
1663                         mps = MLX5_MPW_ENHANCED;
1664                 } else {
1665                         DRV_LOG(DEBUG, "MPW is supported");
1666                         mps = MLX5_MPW;
1667                 }
1668         } else {
1669                 DRV_LOG(DEBUG, "MPW isn't supported");
1670                 mps = MLX5_MPW_DISABLED;
1671         }
1672 #ifdef HAVE_IBV_MLX5_MOD_SWP
1673         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1674                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1675         DRV_LOG(DEBUG, "SWP support: %u", swp);
1676 #endif
1677         config.swp = !!swp;
1678 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1679         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1680                 struct mlx5dv_striding_rq_caps mprq_caps =
1681                         dv_attr.striding_rq_caps;
1682
1683                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1684                         mprq_caps.min_single_stride_log_num_of_bytes);
1685                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1686                         mprq_caps.max_single_stride_log_num_of_bytes);
1687                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1688                         mprq_caps.min_single_wqe_log_num_of_strides);
1689                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1690                         mprq_caps.max_single_wqe_log_num_of_strides);
1691                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1692                         mprq_caps.supported_qpts);
1693                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1694                 mprq = 1;
1695                 mprq_min_stride_size_n =
1696                         mprq_caps.min_single_stride_log_num_of_bytes;
1697                 mprq_max_stride_size_n =
1698                         mprq_caps.max_single_stride_log_num_of_bytes;
1699                 mprq_min_stride_num_n =
1700                         mprq_caps.min_single_wqe_log_num_of_strides;
1701                 mprq_max_stride_num_n =
1702                         mprq_caps.max_single_wqe_log_num_of_strides;
1703                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1704                                                    mprq_min_stride_num_n);
1705         }
1706 #endif
1707         if (RTE_CACHE_LINE_SIZE == 128 &&
1708             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1709                 cqe_comp = 0;
1710         else
1711                 cqe_comp = 1;
1712         config.cqe_comp = cqe_comp;
1713 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1714         /* Whether device supports 128B Rx CQE padding. */
1715         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1716                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1717 #endif
1718 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1719         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1720                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1721                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1722                              (dv_attr.tunnel_offloads_caps &
1723                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1724         }
1725         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1726                 tunnel_en ? "" : "not ");
1727 #else
1728         DRV_LOG(WARNING,
1729                 "tunnel offloading disabled due to old OFED/rdma-core version");
1730 #endif
1731         config.tunnel_en = tunnel_en;
1732 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1733         mpls_en = ((dv_attr.tunnel_offloads_caps &
1734                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1735                    (dv_attr.tunnel_offloads_caps &
1736                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1737         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1738                 mpls_en ? "" : "not ");
1739 #else
1740         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1741                 " old OFED/rdma-core version or firmware configuration");
1742 #endif
1743         config.mpls_en = mpls_en;
1744         /* Check port status. */
1745         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1746         if (err) {
1747                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1748                 goto error;
1749         }
1750         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1751                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1752                 err = EINVAL;
1753                 goto error;
1754         }
1755         if (port_attr.state != IBV_PORT_ACTIVE)
1756                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1757                         mlx5_glue->port_state_str(port_attr.state),
1758                         port_attr.state);
1759         /* Allocate private eth device data. */
1760         priv = rte_zmalloc("ethdev private structure",
1761                            sizeof(*priv),
1762                            RTE_CACHE_LINE_SIZE);
1763         if (priv == NULL) {
1764                 DRV_LOG(ERR, "priv allocation failure");
1765                 err = ENOMEM;
1766                 goto error;
1767         }
1768         priv->sh = sh;
1769         priv->ibv_port = spawn->ibv_port;
1770         priv->pci_dev = spawn->pci_dev;
1771         priv->mtu = RTE_ETHER_MTU;
1772 #ifndef RTE_ARCH_64
1773         /* Initialize UAR access locks for 32bit implementations. */
1774         rte_spinlock_init(&priv->uar_lock_cq);
1775         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1776                 rte_spinlock_init(&priv->uar_lock[i]);
1777 #endif
1778         /* Some internal functions rely on Netlink sockets, open them now. */
1779         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1780         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1781         priv->nl_sn = 0;
1782         priv->representor = !!switch_info->representor;
1783         priv->master = !!switch_info->master;
1784         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1785         /*
1786          * Currently we support single E-Switch per PF configurations
1787          * only and vport_id field contains the vport index for
1788          * associated VF, which is deduced from representor port name.
1789          * For example, let's have the IB device port 10, it has
1790          * attached network device eth0, which has port name attribute
1791          * pf0vf2, we can deduce the VF number as 2, and set vport index
1792          * as 3 (2+1). This assigning schema should be changed if the
1793          * multiple E-Switch instances per PF configurations or/and PCI
1794          * subfunctions are added.
1795          */
1796         priv->vport_id = switch_info->representor ?
1797                          switch_info->port_name + 1 : -1;
1798         /* representor_id field keeps the unmodified port/VF index. */
1799         priv->representor_id = switch_info->representor ?
1800                                switch_info->port_name : -1;
1801         /*
1802          * Look for sibling devices in order to reuse their switch domain
1803          * if any, otherwise allocate one.
1804          */
1805         RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1806                 const struct mlx5_priv *opriv =
1807                         rte_eth_devices[port_id].data->dev_private;
1808
1809                 if (!opriv ||
1810                         opriv->domain_id ==
1811                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1812                         continue;
1813                 priv->domain_id = opriv->domain_id;
1814                 break;
1815         }
1816         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1817                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1818                 if (err) {
1819                         err = rte_errno;
1820                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1821                                 strerror(rte_errno));
1822                         goto error;
1823                 }
1824                 own_domain_id = 1;
1825         }
1826         err = mlx5_args(&config, dpdk_dev->devargs);
1827         if (err) {
1828                 err = rte_errno;
1829                 DRV_LOG(ERR, "failed to process device arguments: %s",
1830                         strerror(rte_errno));
1831                 goto error;
1832         }
1833         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1834                             IBV_DEVICE_RAW_IP_CSUM);
1835         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1836                 (config.hw_csum ? "" : "not "));
1837 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1838         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1839         DRV_LOG(DEBUG, "counters are not supported");
1840 #endif
1841 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1842         if (config.dv_flow_en) {
1843                 DRV_LOG(WARNING, "DV flow is not supported");
1844                 config.dv_flow_en = 0;
1845         }
1846 #endif
1847         config.ind_table_max_size =
1848                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1849         /*
1850          * Remove this check once DPDK supports larger/variable
1851          * indirection tables.
1852          */
1853         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1854                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1855         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1856                 config.ind_table_max_size);
1857         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1858                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1859         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1860                 (config.hw_vlan_strip ? "" : "not "));
1861         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1862                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1863         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1864                 (config.hw_fcs_strip ? "" : "not "));
1865 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1866         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1867 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1868         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1869                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1870 #endif
1871         if (config.hw_padding && !hw_padding) {
1872                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1873                 config.hw_padding = 0;
1874         } else if (config.hw_padding) {
1875                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1876         }
1877         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1878                       (sh->device_attr.tso_caps.supported_qpts &
1879                        (1 << IBV_QPT_RAW_PACKET)));
1880         if (config.tso)
1881                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1882         /*
1883          * MPW is disabled by default, while the Enhanced MPW is enabled
1884          * by default.
1885          */
1886         if (config.mps == MLX5_ARG_UNSET)
1887                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1888                                                           MLX5_MPW_DISABLED;
1889         else
1890                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1891         DRV_LOG(INFO, "%sMPS is %s",
1892                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1893                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1894         if (config.cqe_comp && !cqe_comp) {
1895                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1896                 config.cqe_comp = 0;
1897         }
1898         if (config.cqe_pad && !cqe_pad) {
1899                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1900                 config.cqe_pad = 0;
1901         } else if (config.cqe_pad) {
1902                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1903         }
1904         if (config.devx) {
1905                 priv->counter_fallback = 0;
1906                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
1907                 if (err) {
1908                         err = -err;
1909                         goto error;
1910                 }
1911                 if (!config.hca_attr.flow_counters_dump)
1912                         priv->counter_fallback = 1;
1913 #ifndef HAVE_IBV_DEVX_ASYNC
1914                 priv->counter_fallback = 1;
1915 #endif
1916                 if (priv->counter_fallback)
1917                         DRV_LOG(INFO, "Use fall-back DV counter management\n");
1918                 /* Check for LRO support. */
1919                 if (config.dest_tir && config.hca_attr.lro_cap) {
1920                         /* TBD check tunnel lro caps. */
1921                         config.lro.supported = config.hca_attr.lro_cap;
1922                         DRV_LOG(DEBUG, "Device supports LRO");
1923                         /*
1924                          * If LRO timeout is not configured by application,
1925                          * use the minimal supported value.
1926                          */
1927                         if (!config.lro.timeout)
1928                                 config.lro.timeout =
1929                                 config.hca_attr.lro_timer_supported_periods[0];
1930                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1931                                 config.lro.timeout);
1932                 }
1933         }
1934         if (config.mprq.enabled && mprq) {
1935                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1936                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1937                         config.mprq.stride_num_n =
1938                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1939                                         mprq_min_stride_num_n);
1940                         DRV_LOG(WARNING,
1941                                 "the number of strides"
1942                                 " for Multi-Packet RQ is out of range,"
1943                                 " setting default value (%u)",
1944                                 1 << config.mprq.stride_num_n);
1945                 }
1946                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1947                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1948         } else if (config.mprq.enabled && !mprq) {
1949                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1950                 config.mprq.enabled = 0;
1951         }
1952         if (config.max_dump_files_num == 0)
1953                 config.max_dump_files_num = 128;
1954         eth_dev = rte_eth_dev_allocate(name);
1955         if (eth_dev == NULL) {
1956                 DRV_LOG(ERR, "can not allocate rte ethdev");
1957                 err = ENOMEM;
1958                 goto error;
1959         }
1960         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1961         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1962         if (priv->representor) {
1963                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1964                 eth_dev->data->representor_id = priv->representor_id;
1965         }
1966         /*
1967          * Store associated network device interface index. This index
1968          * is permanent throughout the lifetime of device. So, we may store
1969          * the ifindex here and use the cached value further.
1970          */
1971         assert(spawn->ifindex);
1972         priv->if_index = spawn->ifindex;
1973         eth_dev->data->dev_private = priv;
1974         priv->dev_data = eth_dev->data;
1975         eth_dev->data->mac_addrs = priv->mac;
1976         eth_dev->device = dpdk_dev;
1977         /* Configure the first MAC address by default. */
1978         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1979                 DRV_LOG(ERR,
1980                         "port %u cannot get MAC address, is mlx5_en"
1981                         " loaded? (errno: %s)",
1982                         eth_dev->data->port_id, strerror(rte_errno));
1983                 err = ENODEV;
1984                 goto error;
1985         }
1986         DRV_LOG(INFO,
1987                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1988                 eth_dev->data->port_id,
1989                 mac.addr_bytes[0], mac.addr_bytes[1],
1990                 mac.addr_bytes[2], mac.addr_bytes[3],
1991                 mac.addr_bytes[4], mac.addr_bytes[5]);
1992 #ifndef NDEBUG
1993         {
1994                 char ifname[IF_NAMESIZE];
1995
1996                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1997                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1998                                 eth_dev->data->port_id, ifname);
1999                 else
2000                         DRV_LOG(DEBUG, "port %u ifname is unknown",
2001                                 eth_dev->data->port_id);
2002         }
2003 #endif
2004         /* Get actual MTU if possible. */
2005         err = mlx5_get_mtu(eth_dev, &priv->mtu);
2006         if (err) {
2007                 err = rte_errno;
2008                 goto error;
2009         }
2010         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
2011                 priv->mtu);
2012         /* Initialize burst functions to prevent crashes before link-up. */
2013         eth_dev->rx_pkt_burst = removed_rx_burst;
2014         eth_dev->tx_pkt_burst = removed_tx_burst;
2015         eth_dev->dev_ops = &mlx5_dev_ops;
2016         /* Register MAC address. */
2017         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
2018         if (config.vf && config.vf_nl_en)
2019                 mlx5_nl_mac_addr_sync(eth_dev);
2020         TAILQ_INIT(&priv->flows);
2021         TAILQ_INIT(&priv->ctrl_flows);
2022         /* Hint libmlx5 to use PMD allocator for data plane resources */
2023         struct mlx5dv_ctx_allocators alctr = {
2024                 .alloc = &mlx5_alloc_verbs_buf,
2025                 .free = &mlx5_free_verbs_buf,
2026                 .data = priv,
2027         };
2028         mlx5_glue->dv_set_context_attr(sh->ctx,
2029                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2030                                        (void *)((uintptr_t)&alctr));
2031         /* Bring Ethernet device up. */
2032         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2033                 eth_dev->data->port_id);
2034         mlx5_set_link_up(eth_dev);
2035         /*
2036          * Even though the interrupt handler is not installed yet,
2037          * interrupts will still trigger on the async_fd from
2038          * Verbs context returned by ibv_open_device().
2039          */
2040         mlx5_link_update(eth_dev, 0);
2041 #ifdef HAVE_MLX5DV_DR_ESWITCH
2042         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2043               (switch_info->representor || switch_info->master)))
2044                 config.dv_esw_en = 0;
2045 #else
2046         config.dv_esw_en = 0;
2047 #endif
2048         /* Detect minimal data bytes to inline. */
2049         mlx5_set_min_inline(spawn, &config);
2050         /* Store device configuration on private structure. */
2051         priv->config = config;
2052         /* Create context for virtual machine VLAN workaround. */
2053         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2054         if (config.dv_flow_en) {
2055                 err = mlx5_alloc_shared_dr(priv);
2056                 if (err)
2057                         goto error;
2058         }
2059         /* Supported Verbs flow priority number detection. */
2060         err = mlx5_flow_discover_priorities(eth_dev);
2061         if (err < 0) {
2062                 err = -err;
2063                 goto error;
2064         }
2065         priv->config.flow_prio = err;
2066         return eth_dev;
2067 error:
2068         if (priv) {
2069                 if (priv->sh)
2070                         mlx5_free_shared_dr(priv);
2071                 if (priv->nl_socket_route >= 0)
2072                         close(priv->nl_socket_route);
2073                 if (priv->nl_socket_rdma >= 0)
2074                         close(priv->nl_socket_rdma);
2075                 if (priv->vmwa_context)
2076                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
2077                 if (own_domain_id)
2078                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2079                 rte_free(priv);
2080                 if (eth_dev != NULL)
2081                         eth_dev->data->dev_private = NULL;
2082         }
2083         if (eth_dev != NULL) {
2084                 /* mac_addrs must not be freed alone because part of dev_private */
2085                 eth_dev->data->mac_addrs = NULL;
2086                 rte_eth_dev_release_port(eth_dev);
2087         }
2088         if (sh)
2089                 mlx5_free_shared_ibctx(sh);
2090         assert(err > 0);
2091         rte_errno = err;
2092         return NULL;
2093 }
2094
2095 /**
2096  * Comparison callback to sort device data.
2097  *
2098  * This is meant to be used with qsort().
2099  *
2100  * @param a[in]
2101  *   Pointer to pointer to first data object.
2102  * @param b[in]
2103  *   Pointer to pointer to second data object.
2104  *
2105  * @return
2106  *   0 if both objects are equal, less than 0 if the first argument is less
2107  *   than the second, greater than 0 otherwise.
2108  */
2109 static int
2110 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2111 {
2112         const struct mlx5_switch_info *si_a =
2113                 &((const struct mlx5_dev_spawn_data *)a)->info;
2114         const struct mlx5_switch_info *si_b =
2115                 &((const struct mlx5_dev_spawn_data *)b)->info;
2116         int ret;
2117
2118         /* Master device first. */
2119         ret = si_b->master - si_a->master;
2120         if (ret)
2121                 return ret;
2122         /* Then representor devices. */
2123         ret = si_b->representor - si_a->representor;
2124         if (ret)
2125                 return ret;
2126         /* Unidentified devices come last in no specific order. */
2127         if (!si_a->representor)
2128                 return 0;
2129         /* Order representors by name. */
2130         return si_a->port_name - si_b->port_name;
2131 }
2132
2133 /**
2134  * Match PCI information for possible slaves of bonding device.
2135  *
2136  * @param[in] ibv_dev
2137  *   Pointer to Infiniband device structure.
2138  * @param[in] pci_dev
2139  *   Pointer to PCI device structure to match PCI address.
2140  * @param[in] nl_rdma
2141  *   Netlink RDMA group socket handle.
2142  *
2143  * @return
2144  *   negative value if no bonding device found, otherwise
2145  *   positive index of slave PF in bonding.
2146  */
2147 static int
2148 mlx5_device_bond_pci_match(const struct ibv_device *ibv_dev,
2149                            const struct rte_pci_device *pci_dev,
2150                            int nl_rdma)
2151 {
2152         char ifname[IF_NAMESIZE + 1];
2153         unsigned int ifindex;
2154         unsigned int np, i;
2155         FILE *file = NULL;
2156         int pf = -1;
2157
2158         /*
2159          * Try to get master device name. If something goes
2160          * wrong suppose the lack of kernel support and no
2161          * bonding devices.
2162          */
2163         if (nl_rdma < 0)
2164                 return -1;
2165         if (!strstr(ibv_dev->name, "bond"))
2166                 return -1;
2167         np = mlx5_nl_portnum(nl_rdma, ibv_dev->name);
2168         if (!np)
2169                 return -1;
2170         /*
2171          * The Master device might not be on the predefined
2172          * port (not on port index 1, it is not garanted),
2173          * we have to scan all Infiniband device port and
2174          * find master.
2175          */
2176         for (i = 1; i <= np; ++i) {
2177                 /* Check whether Infiniband port is populated. */
2178                 ifindex = mlx5_nl_ifindex(nl_rdma, ibv_dev->name, i);
2179                 if (!ifindex)
2180                         continue;
2181                 if (!if_indextoname(ifindex, ifname))
2182                         continue;
2183                 /* Try to read bonding slave names from sysfs. */
2184                 MKSTR(slaves,
2185                       "/sys/class/net/%s/master/bonding/slaves", ifname);
2186                 file = fopen(slaves, "r");
2187                 if (file)
2188                         break;
2189         }
2190         if (!file)
2191                 return -1;
2192         /* Use safe format to check maximal buffer length. */
2193         assert(atol(RTE_STR(IF_NAMESIZE)) == IF_NAMESIZE);
2194         while (fscanf(file, "%" RTE_STR(IF_NAMESIZE) "s", ifname) == 1) {
2195                 char tmp_str[IF_NAMESIZE + 32];
2196                 struct rte_pci_addr pci_addr;
2197                 struct mlx5_switch_info info;
2198
2199                 /* Process slave interface names in the loop. */
2200                 snprintf(tmp_str, sizeof(tmp_str),
2201                          "/sys/class/net/%s", ifname);
2202                 if (mlx5_dev_to_pci_addr(tmp_str, &pci_addr)) {
2203                         DRV_LOG(WARNING, "can not get PCI address"
2204                                          " for netdev \"%s\"", ifname);
2205                         continue;
2206                 }
2207                 if (pci_dev->addr.domain != pci_addr.domain ||
2208                     pci_dev->addr.bus != pci_addr.bus ||
2209                     pci_dev->addr.devid != pci_addr.devid ||
2210                     pci_dev->addr.function != pci_addr.function)
2211                         continue;
2212                 /* Slave interface PCI address match found. */
2213                 fclose(file);
2214                 snprintf(tmp_str, sizeof(tmp_str),
2215                          "/sys/class/net/%s/phys_port_name", ifname);
2216                 file = fopen(tmp_str, "rb");
2217                 if (!file)
2218                         break;
2219                 info.name_type = MLX5_PHYS_PORT_NAME_TYPE_NOTSET;
2220                 if (fscanf(file, "%32s", tmp_str) == 1)
2221                         mlx5_translate_port_name(tmp_str, &info);
2222                 if (info.name_type == MLX5_PHYS_PORT_NAME_TYPE_LEGACY ||
2223                     info.name_type == MLX5_PHYS_PORT_NAME_TYPE_UPLINK)
2224                         pf = info.port_name;
2225                 break;
2226         }
2227         if (file)
2228                 fclose(file);
2229         return pf;
2230 }
2231
2232 /**
2233  * DPDK callback to register a PCI device.
2234  *
2235  * This function spawns Ethernet devices out of a given PCI device.
2236  *
2237  * @param[in] pci_drv
2238  *   PCI driver structure (mlx5_driver).
2239  * @param[in] pci_dev
2240  *   PCI device information.
2241  *
2242  * @return
2243  *   0 on success, a negative errno value otherwise and rte_errno is set.
2244  */
2245 static int
2246 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2247                struct rte_pci_device *pci_dev)
2248 {
2249         struct ibv_device **ibv_list;
2250         /*
2251          * Number of found IB Devices matching with requested PCI BDF.
2252          * nd != 1 means there are multiple IB devices over the same
2253          * PCI device and we have representors and master.
2254          */
2255         unsigned int nd = 0;
2256         /*
2257          * Number of found IB device Ports. nd = 1 and np = 1..n means
2258          * we have the single multiport IB device, and there may be
2259          * representors attached to some of found ports.
2260          */
2261         unsigned int np = 0;
2262         /*
2263          * Number of DPDK ethernet devices to Spawn - either over
2264          * multiple IB devices or multiple ports of single IB device.
2265          * Actually this is the number of iterations to spawn.
2266          */
2267         unsigned int ns = 0;
2268         /*
2269          * Bonding device
2270          *   < 0 - no bonding device (single one)
2271          *  >= 0 - bonding device (value is slave PF index)
2272          */
2273         int bd = -1;
2274         struct mlx5_dev_spawn_data *list = NULL;
2275         struct mlx5_dev_config dev_config;
2276         int ret;
2277
2278         ret = mlx5_init_once();
2279         if (ret) {
2280                 DRV_LOG(ERR, "unable to init PMD global data: %s",
2281                         strerror(rte_errno));
2282                 return -rte_errno;
2283         }
2284         assert(pci_drv == &mlx5_driver);
2285         errno = 0;
2286         ibv_list = mlx5_glue->get_device_list(&ret);
2287         if (!ibv_list) {
2288                 rte_errno = errno ? errno : ENOSYS;
2289                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2290                 return -rte_errno;
2291         }
2292         /*
2293          * First scan the list of all Infiniband devices to find
2294          * matching ones, gathering into the list.
2295          */
2296         struct ibv_device *ibv_match[ret + 1];
2297         int nl_route = mlx5_nl_init(NETLINK_ROUTE);
2298         int nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2299         unsigned int i;
2300
2301         while (ret-- > 0) {
2302                 struct rte_pci_addr pci_addr;
2303
2304                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2305                 bd = mlx5_device_bond_pci_match
2306                                 (ibv_list[ret], pci_dev, nl_rdma);
2307                 if (bd >= 0) {
2308                         /*
2309                          * Bonding device detected. Only one match is allowed,
2310                          * the bonding is supported over multi-port IB device,
2311                          * there should be no matches on representor PCI
2312                          * functions or non VF LAG bonding devices with
2313                          * specified address.
2314                          */
2315                         if (nd) {
2316                                 DRV_LOG(ERR,
2317                                         "multiple PCI match on bonding device"
2318                                         "\"%s\" found", ibv_list[ret]->name);
2319                                 rte_errno = ENOENT;
2320                                 ret = -rte_errno;
2321                                 goto exit;
2322                         }
2323                         DRV_LOG(INFO, "PCI information matches for"
2324                                       " slave %d bonding device \"%s\"",
2325                                       bd, ibv_list[ret]->name);
2326                         ibv_match[nd++] = ibv_list[ret];
2327                         break;
2328                 }
2329                 if (mlx5_dev_to_pci_addr
2330                         (ibv_list[ret]->ibdev_path, &pci_addr))
2331                         continue;
2332                 if (pci_dev->addr.domain != pci_addr.domain ||
2333                     pci_dev->addr.bus != pci_addr.bus ||
2334                     pci_dev->addr.devid != pci_addr.devid ||
2335                     pci_dev->addr.function != pci_addr.function)
2336                         continue;
2337                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2338                         ibv_list[ret]->name);
2339                 ibv_match[nd++] = ibv_list[ret];
2340         }
2341         ibv_match[nd] = NULL;
2342         if (!nd) {
2343                 /* No device matches, just complain and bail out. */
2344                 DRV_LOG(WARNING,
2345                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
2346                         " are kernel drivers loaded?",
2347                         pci_dev->addr.domain, pci_dev->addr.bus,
2348                         pci_dev->addr.devid, pci_dev->addr.function);
2349                 rte_errno = ENOENT;
2350                 ret = -rte_errno;
2351                 goto exit;
2352         }
2353         if (nd == 1) {
2354                 /*
2355                  * Found single matching device may have multiple ports.
2356                  * Each port may be representor, we have to check the port
2357                  * number and check the representors existence.
2358                  */
2359                 if (nl_rdma >= 0)
2360                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2361                 if (!np)
2362                         DRV_LOG(WARNING, "can not get IB device \"%s\""
2363                                          " ports number", ibv_match[0]->name);
2364                 if (bd >= 0 && !np) {
2365                         DRV_LOG(ERR, "can not get ports"
2366                                      " for bonding device");
2367                         rte_errno = ENOENT;
2368                         ret = -rte_errno;
2369                         goto exit;
2370                 }
2371         }
2372 #ifndef HAVE_MLX5DV_DR_DEVX_PORT
2373         if (bd >= 0) {
2374                 /*
2375                  * This may happen if there is VF LAG kernel support and
2376                  * application is compiled with older rdma_core library.
2377                  */
2378                 DRV_LOG(ERR,
2379                         "No kernel/verbs support for VF LAG bonding found.");
2380                 rte_errno = ENOTSUP;
2381                 ret = -rte_errno;
2382                 goto exit;
2383         }
2384 #endif
2385         /*
2386          * Now we can determine the maximal
2387          * amount of devices to be spawned.
2388          */
2389         list = rte_zmalloc("device spawn data",
2390                          sizeof(struct mlx5_dev_spawn_data) *
2391                          (np ? np : nd),
2392                          RTE_CACHE_LINE_SIZE);
2393         if (!list) {
2394                 DRV_LOG(ERR, "spawn data array allocation failure");
2395                 rte_errno = ENOMEM;
2396                 ret = -rte_errno;
2397                 goto exit;
2398         }
2399         if (bd >= 0 || np > 1) {
2400                 /*
2401                  * Single IB device with multiple ports found,
2402                  * it may be E-Switch master device and representors.
2403                  * We have to perform identification trough the ports.
2404                  */
2405                 assert(nl_rdma >= 0);
2406                 assert(ns == 0);
2407                 assert(nd == 1);
2408                 assert(np);
2409                 for (i = 1; i <= np; ++i) {
2410                         list[ns].max_port = np;
2411                         list[ns].ibv_port = i;
2412                         list[ns].ibv_dev = ibv_match[0];
2413                         list[ns].eth_dev = NULL;
2414                         list[ns].pci_dev = pci_dev;
2415                         list[ns].pf_bond = bd;
2416                         list[ns].ifindex = mlx5_nl_ifindex
2417                                         (nl_rdma, list[ns].ibv_dev->name, i);
2418                         if (!list[ns].ifindex) {
2419                                 /*
2420                                  * No network interface index found for the
2421                                  * specified port, it means there is no
2422                                  * representor on this port. It's OK,
2423                                  * there can be disabled ports, for example
2424                                  * if sriov_numvfs < sriov_totalvfs.
2425                                  */
2426                                 continue;
2427                         }
2428                         ret = -1;
2429                         if (nl_route >= 0)
2430                                 ret = mlx5_nl_switch_info
2431                                                (nl_route,
2432                                                 list[ns].ifindex,
2433                                                 &list[ns].info);
2434                         if (ret || (!list[ns].info.representor &&
2435                                     !list[ns].info.master)) {
2436                                 /*
2437                                  * We failed to recognize representors with
2438                                  * Netlink, let's try to perform the task
2439                                  * with sysfs.
2440                                  */
2441                                 ret =  mlx5_sysfs_switch_info
2442                                                 (list[ns].ifindex,
2443                                                  &list[ns].info);
2444                         }
2445                         if (!ret && bd >= 0) {
2446                                 switch (list[ns].info.name_type) {
2447                                 case MLX5_PHYS_PORT_NAME_TYPE_UPLINK:
2448                                         if (list[ns].info.port_name == bd)
2449                                                 ns++;
2450                                         break;
2451                                 case MLX5_PHYS_PORT_NAME_TYPE_PFVF:
2452                                         if (list[ns].info.pf_num == bd)
2453                                                 ns++;
2454                                         break;
2455                                 default:
2456                                         break;
2457                                 }
2458                                 continue;
2459                         }
2460                         if (!ret && (list[ns].info.representor ^
2461                                      list[ns].info.master))
2462                                 ns++;
2463                 }
2464                 if (!ns) {
2465                         DRV_LOG(ERR,
2466                                 "unable to recognize master/representors"
2467                                 " on the IB device with multiple ports");
2468                         rte_errno = ENOENT;
2469                         ret = -rte_errno;
2470                         goto exit;
2471                 }
2472         } else {
2473                 /*
2474                  * The existence of several matching entries (nd > 1) means
2475                  * port representors have been instantiated. No existing Verbs
2476                  * call nor sysfs entries can tell them apart, this can only
2477                  * be done through Netlink calls assuming kernel drivers are
2478                  * recent enough to support them.
2479                  *
2480                  * In the event of identification failure through Netlink,
2481                  * try again through sysfs, then:
2482                  *
2483                  * 1. A single IB device matches (nd == 1) with single
2484                  *    port (np=0/1) and is not a representor, assume
2485                  *    no switch support.
2486                  *
2487                  * 2. Otherwise no safe assumptions can be made;
2488                  *    complain louder and bail out.
2489                  */
2490                 np = 1;
2491                 for (i = 0; i != nd; ++i) {
2492                         memset(&list[ns].info, 0, sizeof(list[ns].info));
2493                         list[ns].max_port = 1;
2494                         list[ns].ibv_port = 1;
2495                         list[ns].ibv_dev = ibv_match[i];
2496                         list[ns].eth_dev = NULL;
2497                         list[ns].pci_dev = pci_dev;
2498                         list[ns].pf_bond = -1;
2499                         list[ns].ifindex = 0;
2500                         if (nl_rdma >= 0)
2501                                 list[ns].ifindex = mlx5_nl_ifindex
2502                                         (nl_rdma, list[ns].ibv_dev->name, 1);
2503                         if (!list[ns].ifindex) {
2504                                 char ifname[IF_NAMESIZE];
2505
2506                                 /*
2507                                  * Netlink failed, it may happen with old
2508                                  * ib_core kernel driver (before 4.16).
2509                                  * We can assume there is old driver because
2510                                  * here we are processing single ports IB
2511                                  * devices. Let's try sysfs to retrieve
2512                                  * the ifindex. The method works for
2513                                  * master device only.
2514                                  */
2515                                 if (nd > 1) {
2516                                         /*
2517                                          * Multiple devices found, assume
2518                                          * representors, can not distinguish
2519                                          * master/representor and retrieve
2520                                          * ifindex via sysfs.
2521                                          */
2522                                         continue;
2523                                 }
2524                                 ret = mlx5_get_master_ifname
2525                                         (ibv_match[i]->ibdev_path, &ifname);
2526                                 if (!ret)
2527                                         list[ns].ifindex =
2528                                                 if_nametoindex(ifname);
2529                                 if (!list[ns].ifindex) {
2530                                         /*
2531                                          * No network interface index found
2532                                          * for the specified device, it means
2533                                          * there it is neither representor
2534                                          * nor master.
2535                                          */
2536                                         continue;
2537                                 }
2538                         }
2539                         ret = -1;
2540                         if (nl_route >= 0)
2541                                 ret = mlx5_nl_switch_info
2542                                                (nl_route,
2543                                                 list[ns].ifindex,
2544                                                 &list[ns].info);
2545                         if (ret || (!list[ns].info.representor &&
2546                                     !list[ns].info.master)) {
2547                                 /*
2548                                  * We failed to recognize representors with
2549                                  * Netlink, let's try to perform the task
2550                                  * with sysfs.
2551                                  */
2552                                 ret =  mlx5_sysfs_switch_info
2553                                                 (list[ns].ifindex,
2554                                                  &list[ns].info);
2555                         }
2556                         if (!ret && (list[ns].info.representor ^
2557                                      list[ns].info.master)) {
2558                                 ns++;
2559                         } else if ((nd == 1) &&
2560                                    !list[ns].info.representor &&
2561                                    !list[ns].info.master) {
2562                                 /*
2563                                  * Single IB device with
2564                                  * one physical port and
2565                                  * attached network device.
2566                                  * May be SRIOV is not enabled
2567                                  * or there is no representors.
2568                                  */
2569                                 DRV_LOG(INFO, "no E-Switch support detected");
2570                                 ns++;
2571                                 break;
2572                         }
2573                 }
2574                 if (!ns) {
2575                         DRV_LOG(ERR,
2576                                 "unable to recognize master/representors"
2577                                 " on the multiple IB devices");
2578                         rte_errno = ENOENT;
2579                         ret = -rte_errno;
2580                         goto exit;
2581                 }
2582         }
2583         assert(ns);
2584         /*
2585          * Sort list to probe devices in natural order for users convenience
2586          * (i.e. master first, then representors from lowest to highest ID).
2587          */
2588         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2589         /* Default configuration. */
2590         dev_config = (struct mlx5_dev_config){
2591                 .hw_padding = 0,
2592                 .mps = MLX5_ARG_UNSET,
2593                 .rx_vec_en = 1,
2594                 .txq_inline_max = MLX5_ARG_UNSET,
2595                 .txq_inline_min = MLX5_ARG_UNSET,
2596                 .txq_inline_mpw = MLX5_ARG_UNSET,
2597                 .txqs_inline = MLX5_ARG_UNSET,
2598                 .vf_nl_en = 1,
2599                 .mr_ext_memseg_en = 1,
2600                 .mprq = {
2601                         .enabled = 0, /* Disabled by default. */
2602                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
2603                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
2604                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
2605                 },
2606                 .dv_esw_en = 1,
2607         };
2608         /* Device specific configuration. */
2609         switch (pci_dev->id.device_id) {
2610         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2611         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2612         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2613         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2614         case PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF:
2615         case PCI_DEVICE_ID_MELLANOX_CONNECTX6VF:
2616                 dev_config.vf = 1;
2617                 break;
2618         default:
2619                 break;
2620         }
2621         for (i = 0; i != ns; ++i) {
2622                 uint32_t restore;
2623
2624                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2625                                                  &list[i],
2626                                                  dev_config);
2627                 if (!list[i].eth_dev) {
2628                         if (rte_errno != EBUSY && rte_errno != EEXIST)
2629                                 break;
2630                         /* Device is disabled or already spawned. Ignore it. */
2631                         continue;
2632                 }
2633                 restore = list[i].eth_dev->data->dev_flags;
2634                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2635                 /* Restore non-PCI flags cleared by the above call. */
2636                 list[i].eth_dev->data->dev_flags |= restore;
2637                 rte_eth_dev_probing_finish(list[i].eth_dev);
2638         }
2639         if (i != ns) {
2640                 DRV_LOG(ERR,
2641                         "probe of PCI device " PCI_PRI_FMT " aborted after"
2642                         " encountering an error: %s",
2643                         pci_dev->addr.domain, pci_dev->addr.bus,
2644                         pci_dev->addr.devid, pci_dev->addr.function,
2645                         strerror(rte_errno));
2646                 ret = -rte_errno;
2647                 /* Roll back. */
2648                 while (i--) {
2649                         if (!list[i].eth_dev)
2650                                 continue;
2651                         mlx5_dev_close(list[i].eth_dev);
2652                         /* mac_addrs must not be freed because in dev_private */
2653                         list[i].eth_dev->data->mac_addrs = NULL;
2654                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2655                 }
2656                 /* Restore original error. */
2657                 rte_errno = -ret;
2658         } else {
2659                 ret = 0;
2660         }
2661 exit:
2662         /*
2663          * Do the routine cleanup:
2664          * - close opened Netlink sockets
2665          * - free allocated spawn data array
2666          * - free the Infiniband device list
2667          */
2668         if (nl_rdma >= 0)
2669                 close(nl_rdma);
2670         if (nl_route >= 0)
2671                 close(nl_route);
2672         if (list)
2673                 rte_free(list);
2674         assert(ibv_list);
2675         mlx5_glue->free_device_list(ibv_list);
2676         return ret;
2677 }
2678
2679 /**
2680  * DPDK callback to remove a PCI device.
2681  *
2682  * This function removes all Ethernet devices belong to a given PCI device.
2683  *
2684  * @param[in] pci_dev
2685  *   Pointer to the PCI device.
2686  *
2687  * @return
2688  *   0 on success, the function cannot fail.
2689  */
2690 static int
2691 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2692 {
2693         uint16_t port_id;
2694
2695         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2696                 rte_eth_dev_close(port_id);
2697         return 0;
2698 }
2699
2700 static const struct rte_pci_id mlx5_pci_id_map[] = {
2701         {
2702                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2703                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2704         },
2705         {
2706                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2707                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2708         },
2709         {
2710                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2711                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2712         },
2713         {
2714                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2715                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2716         },
2717         {
2718                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2719                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2720         },
2721         {
2722                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2723                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2724         },
2725         {
2726                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2727                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2728         },
2729         {
2730                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2731                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2732         },
2733         {
2734                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2735                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2736         },
2737         {
2738                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2739                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2740         },
2741         {
2742                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2743                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2744         },
2745         {
2746                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2747                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2748         },
2749         {
2750                 .vendor_id = 0
2751         }
2752 };
2753
2754 static struct rte_pci_driver mlx5_driver = {
2755         .driver = {
2756                 .name = MLX5_DRIVER_NAME
2757         },
2758         .id_table = mlx5_pci_id_map,
2759         .probe = mlx5_pci_probe,
2760         .remove = mlx5_pci_remove,
2761         .dma_map = mlx5_dma_map,
2762         .dma_unmap = mlx5_dma_unmap,
2763         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2764                      RTE_PCI_DRV_PROBE_AGAIN,
2765 };
2766
2767 #ifdef RTE_IBVERBS_LINK_DLOPEN
2768
2769 /**
2770  * Suffix RTE_EAL_PMD_PATH with "-glue".
2771  *
2772  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2773  * suffixing its last component.
2774  *
2775  * @param buf[out]
2776  *   Output buffer, should be large enough otherwise NULL is returned.
2777  * @param size
2778  *   Size of @p out.
2779  *
2780  * @return
2781  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2782  */
2783 static char *
2784 mlx5_glue_path(char *buf, size_t size)
2785 {
2786         static const char *const bad[] = { "/", ".", "..", NULL };
2787         const char *path = RTE_EAL_PMD_PATH;
2788         size_t len = strlen(path);
2789         size_t off;
2790         int i;
2791
2792         while (len && path[len - 1] == '/')
2793                 --len;
2794         for (off = len; off && path[off - 1] != '/'; --off)
2795                 ;
2796         for (i = 0; bad[i]; ++i)
2797                 if (!strncmp(path + off, bad[i], (int)(len - off)))
2798                         goto error;
2799         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2800         if (i == -1 || (size_t)i >= size)
2801                 goto error;
2802         return buf;
2803 error:
2804         DRV_LOG(ERR,
2805                 "unable to append \"-glue\" to last component of"
2806                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2807                 " please re-configure DPDK");
2808         return NULL;
2809 }
2810
2811 /**
2812  * Initialization routine for run-time dependency on rdma-core.
2813  */
2814 static int
2815 mlx5_glue_init(void)
2816 {
2817         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2818         const char *path[] = {
2819                 /*
2820                  * A basic security check is necessary before trusting
2821                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2822                  */
2823                 (geteuid() == getuid() && getegid() == getgid() ?
2824                  getenv("MLX5_GLUE_PATH") : NULL),
2825                 /*
2826                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2827                  * variant, otherwise let dlopen() look up libraries on its
2828                  * own.
2829                  */
2830                 (*RTE_EAL_PMD_PATH ?
2831                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2832         };
2833         unsigned int i = 0;
2834         void *handle = NULL;
2835         void **sym;
2836         const char *dlmsg;
2837
2838         while (!handle && i != RTE_DIM(path)) {
2839                 const char *end;
2840                 size_t len;
2841                 int ret;
2842
2843                 if (!path[i]) {
2844                         ++i;
2845                         continue;
2846                 }
2847                 end = strpbrk(path[i], ":;");
2848                 if (!end)
2849                         end = path[i] + strlen(path[i]);
2850                 len = end - path[i];
2851                 ret = 0;
2852                 do {
2853                         char name[ret + 1];
2854
2855                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2856                                        (int)len, path[i],
2857                                        (!len || *(end - 1) == '/') ? "" : "/");
2858                         if (ret == -1)
2859                                 break;
2860                         if (sizeof(name) != (size_t)ret + 1)
2861                                 continue;
2862                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2863                                 name);
2864                         handle = dlopen(name, RTLD_LAZY);
2865                         break;
2866                 } while (1);
2867                 path[i] = end + 1;
2868                 if (!*end)
2869                         ++i;
2870         }
2871         if (!handle) {
2872                 rte_errno = EINVAL;
2873                 dlmsg = dlerror();
2874                 if (dlmsg)
2875                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2876                 goto glue_error;
2877         }
2878         sym = dlsym(handle, "mlx5_glue");
2879         if (!sym || !*sym) {
2880                 rte_errno = EINVAL;
2881                 dlmsg = dlerror();
2882                 if (dlmsg)
2883                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2884                 goto glue_error;
2885         }
2886         mlx5_glue = *sym;
2887         return 0;
2888 glue_error:
2889         if (handle)
2890                 dlclose(handle);
2891         DRV_LOG(WARNING,
2892                 "cannot initialize PMD due to missing run-time dependency on"
2893                 " rdma-core libraries (libibverbs, libmlx5)");
2894         return -rte_errno;
2895 }
2896
2897 #endif
2898
2899 /**
2900  * Driver initialization routine.
2901  */
2902 RTE_INIT(rte_mlx5_pmd_init)
2903 {
2904         /* Initialize driver log type. */
2905         mlx5_logtype = rte_log_register("pmd.net.mlx5");
2906         if (mlx5_logtype >= 0)
2907                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2908
2909         /* Build the static tables for Verbs conversion. */
2910         mlx5_set_ptype_table();
2911         mlx5_set_cksum_table();
2912         mlx5_set_swp_types_table();
2913         /*
2914          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2915          * huge pages. Calling ibv_fork_init() during init allows
2916          * applications to use fork() safely for purposes other than
2917          * using this PMD, which is not supported in forked processes.
2918          */
2919         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2920         /* Match the size of Rx completion entry to the size of a cacheline. */
2921         if (RTE_CACHE_LINE_SIZE == 128)
2922                 setenv("MLX5_CQE_SIZE", "128", 0);
2923         /*
2924          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2925          * cleanup all the Verbs resources even when the device was removed.
2926          */
2927         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2928 #ifdef RTE_IBVERBS_LINK_DLOPEN
2929         if (mlx5_glue_init())
2930                 return;
2931         assert(mlx5_glue);
2932 #endif
2933 #ifndef NDEBUG
2934         /* Glue structure must not contain any NULL pointers. */
2935         {
2936                 unsigned int i;
2937
2938                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2939                         assert(((const void *const *)mlx5_glue)[i]);
2940         }
2941 #endif
2942         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2943                 DRV_LOG(ERR,
2944                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2945                         mlx5_glue->version, MLX5_GLUE_VERSION);
2946                 return;
2947         }
2948         mlx5_glue->fork_init();
2949         rte_pci_register(&mlx5_driver);
2950 }
2951
2952 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2953 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2954 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");