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         struct mlx5_switch_info info; /**< Switch information. */
 172         struct ibv_device *ibv_dev; /**< Associated IB device. */
 173         struct rte_eth_dev *eth_dev; /**< Associated Ethernet device. */
 174         struct rte_pci_device *pci_dev; /**< Backend PCI device. */
 175 };
 176
 177 static LIST_HEAD(, mlx5_ibv_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
 178 static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 179
 180 /**
 181  * Initialize the counters management structure.
 182  *
 183  * @param[in] sh
 184  *   Pointer to mlx5_ibv_shared object to free
 185  */
 186 static void
 187 mlx5_flow_counters_mng_init(struct mlx5_ibv_shared *sh)
 188 {
 189         uint8_t i;
 190
 191         TAILQ_INIT(&sh->cmng.flow_counters);
 192         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i)
 193                 TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
 194 }
 195
 196 /**
 197  * Destroy all the resources allocated for a counter memory management.
 198  *
 199  * @param[in] mng
 200  *   Pointer to the memory management structure.
 201  */
 202 static void
 203 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
 204 {
 205         uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
 206
 207         LIST_REMOVE(mng, next);
 208         claim_zero(mlx5_devx_cmd_destroy(mng->dm));
 209         claim_zero(mlx5_glue->devx_umem_dereg(mng->umem));
 210         rte_free(mem);
 211 }
 212
 213 /**
 214  * Close and release all the resources of the counters management.
 215  *
 216  * @param[in] sh
 217  *   Pointer to mlx5_ibv_shared object to free.
 218  */
 219 static void
 220 mlx5_flow_counters_mng_close(struct mlx5_ibv_shared *sh)
 221 {
 222         struct mlx5_counter_stats_mem_mng *mng;
 223         uint8_t i;
 224         int j;
 225         int retries = 1024;
 226
 227         rte_errno = 0;
 228         while (--retries) {
 229                 rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
 230                 if (rte_errno != EINPROGRESS)
 231                         break;
 232                 rte_pause();
 233         }
 234         for (i = 0; i < RTE_DIM(sh->cmng.ccont); ++i) {
 235                 struct mlx5_flow_counter_pool *pool;
 236                 uint32_t batch = !!(i % 2);
 237
 238                 if (!sh->cmng.ccont[i].pools)
 239                         continue;
 240                 pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 241                 while (pool) {
 242                         if (batch) {
 243                                 if (pool->min_dcs)
 244                                         claim_zero
 245                                         (mlx5_devx_cmd_destroy(pool->min_dcs));
 246                         }
 247                         for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
 248                                 if (pool->counters_raw[j].action)
 249                                         claim_zero
 250                                         (mlx5_glue->destroy_flow_action
 251                                                (pool->counters_raw[j].action));
 252                                 if (!batch && pool->counters_raw[j].dcs)
 253                                         claim_zero(mlx5_devx_cmd_destroy
 254                                                   (pool->counters_raw[j].dcs));
 255                         }
 256                         TAILQ_REMOVE(&sh->cmng.ccont[i].pool_list, pool,
 257                                      next);
 258                         rte_free(pool);
 259                         pool = TAILQ_FIRST(&sh->cmng.ccont[i].pool_list);
 260                 }
 261                 rte_free(sh->cmng.ccont[i].pools);
 262         }
 263         mng = LIST_FIRST(&sh->cmng.mem_mngs);
 264         while (mng) {
 265                 mlx5_flow_destroy_counter_stat_mem_mng(mng);
 266                 mng = LIST_FIRST(&sh->cmng.mem_mngs);
 267         }
 268         memset(&sh->cmng, 0, sizeof(sh->cmng));
 269 }
 270
 271 /**
 272  * Extract pdn of PD object using DV API.
 273  *
 274  * @param[in] pd
 275  *   Pointer to the verbs PD object.
 276  * @param[out] pdn
 277  *   Pointer to the PD object number variable.
 278  *
 279  * @return
 280  *   0 on success, error value otherwise.
 281  */
 282 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 283 static int
 284 mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
 285 {
 286         struct mlx5dv_obj obj;
 287         struct mlx5dv_pd pd_info;
 288         int ret = 0;
 289
 290         obj.pd.in = pd;
 291         obj.pd.out = &pd_info;
 292         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 293         if (ret) {
 294                 DRV_LOG(DEBUG, "Fail to get PD object info");
 295                 return ret;
 296         }
 297         *pdn = pd_info.pdn;
 298         return 0;
 299 }
 300 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 301
 302 /**
 303  * Allocate shared IB device context. If there is multiport device the
 304  * master and representors will share this context, if there is single
 305  * port dedicated IB device, the context will be used by only given
 306  * port due to unification.
 307  *
 308  * Routine first searches the context for the specified IB device name,
 309  * if found the shared context assumed and reference counter is incremented.
 310  * If no context found the new one is created and initialized with specified
 311  * IB device context and parameters.
 312  *
 313  * @param[in] spawn
 314  *   Pointer to the IB device attributes (name, port, etc).
 315  *
 316  * @return
 317  *   Pointer to mlx5_ibv_shared object on success,
 318  *   otherwise NULL and rte_errno is set.
 319  */
 320 static struct mlx5_ibv_shared *
 321 mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
 322 {
 323         struct mlx5_ibv_shared *sh;
 324         int err = 0;
 325         uint32_t i;
 326
 327         assert(spawn);
 328         /* Secondary process should not create the shared context. */
 329         assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
 330         pthread_mutex_lock(&mlx5_ibv_list_mutex);
 331         /* Search for IB context by device name. */
 332         LIST_FOREACH(sh, &mlx5_ibv_list, next) {
 333                 if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
 334                         sh->refcnt++;
 335                         goto exit;
 336                 }
 337         }
 338         /* No device found, we have to create new shared context. */
 339         assert(spawn->max_port);
 340         sh = rte_zmalloc("ethdev shared ib context",
 341                          sizeof(struct mlx5_ibv_shared) +
 342                          spawn->max_port *
 343                          sizeof(struct mlx5_ibv_shared_port),
 344                          RTE_CACHE_LINE_SIZE);
 345         if (!sh) {
 346                 DRV_LOG(ERR, "shared context allocation failure");
 347                 rte_errno  = ENOMEM;
 348                 goto exit;
 349         }
 350         /* Try to open IB device with DV first, then usual Verbs. */
 351         errno = 0;
 352         sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
 353         if (sh->ctx) {
 354                 sh->devx = 1;
 355                 DRV_LOG(DEBUG, "DevX is supported");
 356         } else {
 357                 sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
 358                 if (!sh->ctx) {
 359                         err = errno ? errno : ENODEV;
 360                         goto error;
 361                 }
 362                 DRV_LOG(DEBUG, "DevX is NOT supported");
 363         }
 364         err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
 365         if (err) {
 366                 DRV_LOG(DEBUG, "ibv_query_device_ex() failed");
 367                 goto error;
 368         }
 369         sh->refcnt = 1;
 370         sh->max_port = spawn->max_port;
 371         strncpy(sh->ibdev_name, sh->ctx->device->name,
 372                 sizeof(sh->ibdev_name));
 373         strncpy(sh->ibdev_path, sh->ctx->device->ibdev_path,
 374                 sizeof(sh->ibdev_path));
 375         sh->pci_dev = spawn->pci_dev;
 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                                  sh->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 };
1000
1001 /* Available operations from secondary process. */
1002 static const struct eth_dev_ops mlx5_dev_sec_ops = {
1003         .stats_get = mlx5_stats_get,
1004         .stats_reset = mlx5_stats_reset,
1005         .xstats_get = mlx5_xstats_get,
1006         .xstats_reset = mlx5_xstats_reset,
1007         .xstats_get_names = mlx5_xstats_get_names,
1008         .fw_version_get = mlx5_fw_version_get,
1009         .dev_infos_get = mlx5_dev_infos_get,
1010         .rx_descriptor_status = mlx5_rx_descriptor_status,
1011         .tx_descriptor_status = mlx5_tx_descriptor_status,
1012 };
1013
1014 /* Available operations in flow isolated mode. */
1015 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1016         .dev_configure = mlx5_dev_configure,
1017         .dev_start = mlx5_dev_start,
1018         .dev_stop = mlx5_dev_stop,
1019         .dev_set_link_down = mlx5_set_link_down,
1020         .dev_set_link_up = mlx5_set_link_up,
1021         .dev_close = mlx5_dev_close,
1022         .promiscuous_enable = mlx5_promiscuous_enable,
1023         .promiscuous_disable = mlx5_promiscuous_disable,
1024         .allmulticast_enable = mlx5_allmulticast_enable,
1025         .allmulticast_disable = mlx5_allmulticast_disable,
1026         .link_update = mlx5_link_update,
1027         .stats_get = mlx5_stats_get,
1028         .stats_reset = mlx5_stats_reset,
1029         .xstats_get = mlx5_xstats_get,
1030         .xstats_reset = mlx5_xstats_reset,
1031         .xstats_get_names = mlx5_xstats_get_names,
1032         .fw_version_get = mlx5_fw_version_get,
1033         .dev_infos_get = mlx5_dev_infos_get,
1034         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1035         .vlan_filter_set = mlx5_vlan_filter_set,
1036         .rx_queue_setup = mlx5_rx_queue_setup,
1037         .tx_queue_setup = mlx5_tx_queue_setup,
1038         .rx_queue_release = mlx5_rx_queue_release,
1039         .tx_queue_release = mlx5_tx_queue_release,
1040         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1041         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1042         .mac_addr_remove = mlx5_mac_addr_remove,
1043         .mac_addr_add = mlx5_mac_addr_add,
1044         .mac_addr_set = mlx5_mac_addr_set,
1045         .set_mc_addr_list = mlx5_set_mc_addr_list,
1046         .mtu_set = mlx5_dev_set_mtu,
1047         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1048         .vlan_offload_set = mlx5_vlan_offload_set,
1049         .filter_ctrl = mlx5_dev_filter_ctrl,
1050         .rx_descriptor_status = mlx5_rx_descriptor_status,
1051         .tx_descriptor_status = mlx5_tx_descriptor_status,
1052         .rx_queue_intr_enable = mlx5_rx_intr_enable,
1053         .rx_queue_intr_disable = mlx5_rx_intr_disable,
1054         .is_removed = mlx5_is_removed,
1055 };
1056
1057 /**
1058  * Verify and store value for device argument.
1059  *
1060  * @param[in] key
1061  *   Key argument to verify.
1062  * @param[in] val
1063  *   Value associated with key.
1064  * @param opaque
1065  *   User data.
1066  *
1067  * @return
1068  *   0 on success, a negative errno value otherwise and rte_errno is set.
1069  */
1070 static int
1071 mlx5_args_check(const char *key, const char *val, void *opaque)
1072 {
1073         struct mlx5_dev_config *config = opaque;
1074         unsigned long tmp;
1075
1076         /* No-op, port representors are processed in mlx5_dev_spawn(). */
1077         if (!strcmp(MLX5_REPRESENTOR, key))
1078                 return 0;
1079         errno = 0;
1080         tmp = strtoul(val, NULL, 0);
1081         if (errno) {
1082                 rte_errno = errno;
1083                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1084                 return -rte_errno;
1085         }
1086         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1087                 config->cqe_comp = !!tmp;
1088         } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
1089                 config->cqe_pad = !!tmp;
1090         } else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1091                 config->hw_padding = !!tmp;
1092         } else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1093                 config->mprq.enabled = !!tmp;
1094         } else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1095                 config->mprq.stride_num_n = tmp;
1096         } else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1097                 config->mprq.max_memcpy_len = tmp;
1098         } else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1099                 config->mprq.min_rxqs_num = tmp;
1100         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1101                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1102                                  " converted to txq_inline_max", key);
1103                 config->txq_inline_max = tmp;
1104         } else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1105                 config->txq_inline_max = tmp;
1106         } else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1107                 config->txq_inline_min = tmp;
1108         } else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1109                 config->txq_inline_mpw = tmp;
1110         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1111                 config->txqs_inline = tmp;
1112         } else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1113                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1114         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1115                 config->mps = !!tmp;
1116         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1117                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1118         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1119                 DRV_LOG(WARNING, "%s: deprecated parameter,"
1120                                  " converted to txq_inline_mpw", key);
1121                 config->txq_inline_mpw = tmp;
1122         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1123                 DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1124         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1125                 config->rx_vec_en = !!tmp;
1126         } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1127                 config->l3_vxlan_en = !!tmp;
1128         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1129                 config->vf_nl_en = !!tmp;
1130         } else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1131                 config->dv_esw_en = !!tmp;
1132         } else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1133                 config->dv_flow_en = !!tmp;
1134         } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1135                 config->mr_ext_memseg_en = !!tmp;
1136         } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1137                 config->max_dump_files_num = tmp;
1138         } else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1139                 config->lro.timeout = tmp;
1140         } else {
1141                 DRV_LOG(WARNING, "%s: unknown parameter", key);
1142                 rte_errno = EINVAL;
1143                 return -rte_errno;
1144         }
1145         return 0;
1146 }
1147
1148 /**
1149  * Parse device parameters.
1150  *
1151  * @param config
1152  *   Pointer to device configuration structure.
1153  * @param devargs
1154  *   Device arguments structure.
1155  *
1156  * @return
1157  *   0 on success, a negative errno value otherwise and rte_errno is set.
1158  */
1159 static int
1160 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1161 {
1162         const char **params = (const char *[]){
1163                 MLX5_RXQ_CQE_COMP_EN,
1164                 MLX5_RXQ_CQE_PAD_EN,
1165                 MLX5_RXQ_PKT_PAD_EN,
1166                 MLX5_RX_MPRQ_EN,
1167                 MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1168                 MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1169                 MLX5_RXQS_MIN_MPRQ,
1170                 MLX5_TXQ_INLINE,
1171                 MLX5_TXQ_INLINE_MIN,
1172                 MLX5_TXQ_INLINE_MAX,
1173                 MLX5_TXQ_INLINE_MPW,
1174                 MLX5_TXQS_MIN_INLINE,
1175                 MLX5_TXQS_MAX_VEC,
1176                 MLX5_TXQ_MPW_EN,
1177                 MLX5_TXQ_MPW_HDR_DSEG_EN,
1178                 MLX5_TXQ_MAX_INLINE_LEN,
1179                 MLX5_TX_VEC_EN,
1180                 MLX5_RX_VEC_EN,
1181                 MLX5_L3_VXLAN_EN,
1182                 MLX5_VF_NL_EN,
1183                 MLX5_DV_ESW_EN,
1184                 MLX5_DV_FLOW_EN,
1185                 MLX5_MR_EXT_MEMSEG_EN,
1186                 MLX5_REPRESENTOR,
1187                 MLX5_MAX_DUMP_FILES_NUM,
1188                 MLX5_LRO_TIMEOUT_USEC,
1189                 NULL,
1190         };
1191         struct rte_kvargs *kvlist;
1192         int ret = 0;
1193         int i;
1194
1195         if (devargs == NULL)
1196                 return 0;
1197         /* Following UGLY cast is done to pass checkpatch. */
1198         kvlist = rte_kvargs_parse(devargs->args, params);
1199         if (kvlist == NULL) {
1200                 rte_errno = EINVAL;
1201                 return -rte_errno;
1202         }
1203         /* Process parameters. */
1204         for (i = 0; (params[i] != NULL); ++i) {
1205                 if (rte_kvargs_count(kvlist, params[i])) {
1206                         ret = rte_kvargs_process(kvlist, params[i],
1207                                                  mlx5_args_check, config);
1208                         if (ret) {
1209                                 rte_errno = EINVAL;
1210                                 rte_kvargs_free(kvlist);
1211                                 return -rte_errno;
1212                         }
1213                 }
1214         }
1215         rte_kvargs_free(kvlist);
1216         return 0;
1217 }
1218
1219 static struct rte_pci_driver mlx5_driver;
1220
1221 /**
1222  * PMD global initialization.
1223  *
1224  * Independent from individual device, this function initializes global
1225  * per-PMD data structures distinguishing primary and secondary processes.
1226  * Hence, each initialization is called once per a process.
1227  *
1228  * @return
1229  *   0 on success, a negative errno value otherwise and rte_errno is set.
1230  */
1231 static int
1232 mlx5_init_once(void)
1233 {
1234         struct mlx5_shared_data *sd;
1235         struct mlx5_local_data *ld = &mlx5_local_data;
1236         int ret = 0;
1237
1238         if (mlx5_init_shared_data())
1239                 return -rte_errno;
1240         sd = mlx5_shared_data;
1241         assert(sd);
1242         rte_spinlock_lock(&sd->lock);
1243         switch (rte_eal_process_type()) {
1244         case RTE_PROC_PRIMARY:
1245                 if (sd->init_done)
1246                         break;
1247                 LIST_INIT(&sd->mem_event_cb_list);
1248                 rte_rwlock_init(&sd->mem_event_rwlock);
1249                 rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
1250                                                 mlx5_mr_mem_event_cb, NULL);
1251                 ret = mlx5_mp_init_primary();
1252                 if (ret)
1253                         goto out;
1254                 sd->init_done = true;
1255                 break;
1256         case RTE_PROC_SECONDARY:
1257                 if (ld->init_done)
1258                         break;
1259                 ret = mlx5_mp_init_secondary();
1260                 if (ret)
1261                         goto out;
1262                 ++sd->secondary_cnt;
1263                 ld->init_done = true;
1264                 break;
1265         default:
1266                 break;
1267         }
1268 out:
1269         rte_spinlock_unlock(&sd->lock);
1270         return ret;
1271 }
1272
1273 /**
1274  * Configures the minimal amount of data to inline into WQE
1275  * while sending packets.
1276  *
1277  * - the txq_inline_min has the maximal priority, if this
1278  *   key is specified in devargs
1279  * - if DevX is enabled the inline mode is queried from the
1280  *   device (HCA attributes and NIC vport context if needed).
1281  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4LX
1282  *   and none (0 bytes) for other NICs
1283  *
1284  * @param spawn
1285  *   Verbs device parameters (name, port, switch_info) to spawn.
1286  * @param config
1287  *   Device configuration parameters.
1288  */
1289 static void
1290 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1291                     struct mlx5_dev_config *config)
1292 {
1293         if (config->txq_inline_min != MLX5_ARG_UNSET) {
1294                 /* Application defines size of inlined data explicitly. */
1295                 switch (spawn->pci_dev->id.device_id) {
1296                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1297                 case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1298                         if (config->txq_inline_min <
1299                                        (int)MLX5_INLINE_HSIZE_L2) {
1300                                 DRV_LOG(DEBUG,
1301                                         "txq_inline_mix aligned to minimal"
1302                                         " ConnectX-4 required value %d",
1303                                         (int)MLX5_INLINE_HSIZE_L2);
1304                                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1305                         }
1306                         break;
1307                 }
1308                 goto exit;
1309         }
1310         if (config->hca_attr.eth_net_offloads) {
1311                 /* We have DevX enabled, inline mode queried successfully. */
1312                 switch (config->hca_attr.wqe_inline_mode) {
1313                 case MLX5_CAP_INLINE_MODE_L2:
1314                         /* outer L2 header must be inlined. */
1315                         config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1316                         goto exit;
1317                 case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1318                         /* No inline data are required by NIC. */
1319                         config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1320                         config->hw_vlan_insert =
1321                                 config->hca_attr.wqe_vlan_insert;
1322                         DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1323                         goto exit;
1324                 case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1325                         /* inline mode is defined by NIC vport context. */
1326                         if (!config->hca_attr.eth_virt)
1327                                 break;
1328                         switch (config->hca_attr.vport_inline_mode) {
1329                         case MLX5_INLINE_MODE_NONE:
1330                                 config->txq_inline_min =
1331                                         MLX5_INLINE_HSIZE_NONE;
1332                                 goto exit;
1333                         case MLX5_INLINE_MODE_L2:
1334                                 config->txq_inline_min =
1335                                         MLX5_INLINE_HSIZE_L2;
1336                                 goto exit;
1337                         case MLX5_INLINE_MODE_IP:
1338                                 config->txq_inline_min =
1339                                         MLX5_INLINE_HSIZE_L3;
1340                                 goto exit;
1341                         case MLX5_INLINE_MODE_TCP_UDP:
1342                                 config->txq_inline_min =
1343                                         MLX5_INLINE_HSIZE_L4;
1344                                 goto exit;
1345                         case MLX5_INLINE_MODE_INNER_L2:
1346                                 config->txq_inline_min =
1347                                         MLX5_INLINE_HSIZE_INNER_L2;
1348                                 goto exit;
1349                         case MLX5_INLINE_MODE_INNER_IP:
1350                                 config->txq_inline_min =
1351                                         MLX5_INLINE_HSIZE_INNER_L3;
1352                                 goto exit;
1353                         case MLX5_INLINE_MODE_INNER_TCP_UDP:
1354                                 config->txq_inline_min =
1355                                         MLX5_INLINE_HSIZE_INNER_L4;
1356                                 goto exit;
1357                         }
1358                 }
1359         }
1360         /*
1361          * We get here if we are unable to deduce
1362          * inline data size with DevX. Try PCI ID
1363          * to determine old NICs.
1364          */
1365         switch (spawn->pci_dev->id.device_id) {
1366         case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1367         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1368         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1369         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1370                 config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1371                 config->hw_vlan_insert = 0;
1372                 break;
1373         case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1374         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1375         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1376         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1377                 /*
1378                  * These NICs support VLAN insertion from WQE and
1379                  * report the wqe_vlan_insert flag. But there is the bug
1380                  * and PFC control may be broken, so disable feature.
1381                  */
1382                 config->hw_vlan_insert = 0;
1383                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1384                 break;
1385         default:
1386                 config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1387                 break;
1388         }
1389 exit:
1390         DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1391 }
1392
1393 /**
1394  * Allocate page of door-bells and register it using DevX API.
1395  *
1396  * @param [in] dev
1397  *   Pointer to Ethernet device.
1398  *
1399  * @return
1400  *   Pointer to new page on success, NULL otherwise.
1401  */
1402 static struct mlx5_devx_dbr_page *
1403 mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
1404 {
1405         struct mlx5_priv *priv = dev->data->dev_private;
1406         struct mlx5_devx_dbr_page *page;
1407
1408         /* Allocate space for door-bell page and management data. */
1409         page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
1410                                  RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1411         if (!page) {
1412                 DRV_LOG(ERR, "port %u cannot allocate dbr page",
1413                         dev->data->port_id);
1414                 return NULL;
1415         }
1416         /* Register allocated memory. */
1417         page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
1418                                               MLX5_DBR_PAGE_SIZE, 0);
1419         if (!page->umem) {
1420                 DRV_LOG(ERR, "port %u cannot umem reg dbr page",
1421                         dev->data->port_id);
1422                 rte_free(page);
1423                 return NULL;
1424         }
1425         return page;
1426 }
1427
1428 /**
1429  * Find the next available door-bell, allocate new page if needed.
1430  *
1431  * @param [in] dev
1432  *   Pointer to Ethernet device.
1433  * @param [out] dbr_page
1434  *   Door-bell page containing the page data.
1435  *
1436  * @return
1437  *   Door-bell address offset on success, a negative error value otherwise.
1438  */
1439 int64_t
1440 mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
1441 {
1442         struct mlx5_priv *priv = dev->data->dev_private;
1443         struct mlx5_devx_dbr_page *page = NULL;
1444         uint32_t i, j;
1445
1446         LIST_FOREACH(page, &priv->dbrpgs, next)
1447                 if (page->dbr_count < MLX5_DBR_PER_PAGE)
1448                         break;
1449         if (!page) { /* No page with free door-bell exists. */
1450                 page = mlx5_alloc_dbr_page(dev);
1451                 if (!page) /* Failed to allocate new page. */
1452                         return (-1);
1453                 LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
1454         }
1455         /* Loop to find bitmap part with clear bit. */
1456         for (i = 0;
1457              i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
1458              i++)
1459                 ; /* Empty. */
1460         /* Find the first clear bit. */
1461         j = rte_bsf64(~page->dbr_bitmap[i]);
1462         assert(i < (MLX5_DBR_PER_PAGE / 64));
1463         page->dbr_bitmap[i] |= (1 << j);
1464         page->dbr_count++;
1465         *dbr_page = page;
1466         return (((i * 64) + j) * sizeof(uint64_t));
1467 }
1468
1469 /**
1470  * Release a door-bell record.
1471  *
1472  * @param [in] dev
1473  *   Pointer to Ethernet device.
1474  * @param [in] umem_id
1475  *   UMEM ID of page containing the door-bell record to release.
1476  * @param [in] offset
1477  *   Offset of door-bell record in page.
1478  *
1479  * @return
1480  *   0 on success, a negative error value otherwise.
1481  */
1482 int32_t
1483 mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
1484 {
1485         struct mlx5_priv *priv = dev->data->dev_private;
1486         struct mlx5_devx_dbr_page *page = NULL;
1487         int ret = 0;
1488
1489         LIST_FOREACH(page, &priv->dbrpgs, next)
1490                 /* Find the page this address belongs to. */
1491                 if (page->umem->umem_id == umem_id)
1492                         break;
1493         if (!page)
1494                 return -EINVAL;
1495         page->dbr_count--;
1496         if (!page->dbr_count) {
1497                 /* Page not used, free it and remove from list. */
1498                 LIST_REMOVE(page, next);
1499                 if (page->umem)
1500                         ret = -mlx5_glue->devx_umem_dereg(page->umem);
1501                 rte_free(page);
1502         } else {
1503                 /* Mark in bitmap that this door-bell is not in use. */
1504                 offset /= MLX5_DBR_SIZE;
1505                 int i = offset / 64;
1506                 int j = offset % 64;
1507
1508                 page->dbr_bitmap[i] &= ~(1 << j);
1509         }
1510         return ret;
1511 }
1512
1513 /**
1514  * Spawn an Ethernet device from Verbs information.
1515  *
1516  * @param dpdk_dev
1517  *   Backing DPDK device.
1518  * @param spawn
1519  *   Verbs device parameters (name, port, switch_info) to spawn.
1520  * @param config
1521  *   Device configuration parameters.
1522  *
1523  * @return
1524  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
1525  *   is set. The following errors are defined:
1526  *
1527  *   EBUSY: device is not supposed to be spawned.
1528  *   EEXIST: device is already spawned
1529  */
1530 static struct rte_eth_dev *
1531 mlx5_dev_spawn(struct rte_device *dpdk_dev,
1532                struct mlx5_dev_spawn_data *spawn,
1533                struct mlx5_dev_config config)
1534 {
1535         const struct mlx5_switch_info *switch_info = &spawn->info;
1536         struct mlx5_ibv_shared *sh = NULL;
1537         struct ibv_port_attr port_attr;
1538         struct mlx5dv_context dv_attr = { .comp_mask = 0 };
1539         struct rte_eth_dev *eth_dev = NULL;
1540         struct mlx5_priv *priv = NULL;
1541         int err = 0;
1542         unsigned int hw_padding = 0;
1543         unsigned int mps;
1544         unsigned int cqe_comp;
1545         unsigned int cqe_pad = 0;
1546         unsigned int tunnel_en = 0;
1547         unsigned int mpls_en = 0;
1548         unsigned int swp = 0;
1549         unsigned int mprq = 0;
1550         unsigned int mprq_min_stride_size_n = 0;
1551         unsigned int mprq_max_stride_size_n = 0;
1552         unsigned int mprq_min_stride_num_n = 0;
1553         unsigned int mprq_max_stride_num_n = 0;
1554         struct rte_ether_addr mac;
1555         char name[RTE_ETH_NAME_MAX_LEN];
1556         int own_domain_id = 0;
1557         uint16_t port_id;
1558         unsigned int i;
1559
1560         /* Determine if this port representor is supposed to be spawned. */
1561         if (switch_info->representor && dpdk_dev->devargs) {
1562                 struct rte_eth_devargs eth_da;
1563
1564                 err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
1565                 if (err) {
1566                         rte_errno = -err;
1567                         DRV_LOG(ERR, "failed to process device arguments: %s",
1568                                 strerror(rte_errno));
1569                         return NULL;
1570                 }
1571                 for (i = 0; i < eth_da.nb_representor_ports; ++i)
1572                         if (eth_da.representor_ports[i] ==
1573                             (uint16_t)switch_info->port_name)
1574                                 break;
1575                 if (i == eth_da.nb_representor_ports) {
1576                         rte_errno = EBUSY;
1577                         return NULL;
1578                 }
1579         }
1580         /* Build device name. */
1581         if (!switch_info->representor)
1582                 strlcpy(name, dpdk_dev->name, sizeof(name));
1583         else
1584                 snprintf(name, sizeof(name), "%s_representor_%u",
1585                          dpdk_dev->name, switch_info->port_name);
1586         /* check if the device is already spawned */
1587         if (rte_eth_dev_get_port_by_name(name, &port_id) == 0) {
1588                 rte_errno = EEXIST;
1589                 return NULL;
1590         }
1591         DRV_LOG(DEBUG, "naming Ethernet device \"%s\"", name);
1592         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1593                 eth_dev = rte_eth_dev_attach_secondary(name);
1594                 if (eth_dev == NULL) {
1595                         DRV_LOG(ERR, "can not attach rte ethdev");
1596                         rte_errno = ENOMEM;
1597                         return NULL;
1598                 }
1599                 eth_dev->device = dpdk_dev;
1600                 eth_dev->dev_ops = &mlx5_dev_sec_ops;
1601                 err = mlx5_proc_priv_init(eth_dev);
1602                 if (err)
1603                         return NULL;
1604                 /* Receive command fd from primary process */
1605                 err = mlx5_mp_req_verbs_cmd_fd(eth_dev);
1606                 if (err < 0)
1607                         return NULL;
1608                 /* Remap UAR for Tx queues. */
1609                 err = mlx5_tx_uar_init_secondary(eth_dev, err);
1610                 if (err)
1611                         return NULL;
1612                 /*
1613                  * Ethdev pointer is still required as input since
1614                  * the primary device is not accessible from the
1615                  * secondary process.
1616                  */
1617                 eth_dev->rx_pkt_burst = mlx5_select_rx_function(eth_dev);
1618                 eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
1619                 return eth_dev;
1620         }
1621         sh = mlx5_alloc_shared_ibctx(spawn);
1622         if (!sh)
1623                 return NULL;
1624         config.devx = sh->devx;
1625 #ifdef HAVE_MLX5DV_DR_ACTION_DEST_DEVX_TIR
1626         config.dest_tir = 1;
1627 #endif
1628 #ifdef HAVE_IBV_MLX5_MOD_SWP
1629         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
1630 #endif
1631         /*
1632          * Multi-packet send is supported by ConnectX-4 Lx PF as well
1633          * as all ConnectX-5 devices.
1634          */
1635 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1636         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
1637 #endif
1638 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1639         dv_attr.comp_mask |= MLX5DV_CONTEXT_MASK_STRIDING_RQ;
1640 #endif
1641         mlx5_glue->dv_query_device(sh->ctx, &dv_attr);
1642         if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
1643                 if (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
1644                         DRV_LOG(DEBUG, "enhanced MPW is supported");
1645                         mps = MLX5_MPW_ENHANCED;
1646                 } else {
1647                         DRV_LOG(DEBUG, "MPW is supported");
1648                         mps = MLX5_MPW;
1649                 }
1650         } else {
1651                 DRV_LOG(DEBUG, "MPW isn't supported");
1652                 mps = MLX5_MPW_DISABLED;
1653         }
1654 #ifdef HAVE_IBV_MLX5_MOD_SWP
1655         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
1656                 swp = dv_attr.sw_parsing_caps.sw_parsing_offloads;
1657         DRV_LOG(DEBUG, "SWP support: %u", swp);
1658 #endif
1659         config.swp = !!swp;
1660 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
1661         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_STRIDING_RQ) {
1662                 struct mlx5dv_striding_rq_caps mprq_caps =
1663                         dv_attr.striding_rq_caps;
1664
1665                 DRV_LOG(DEBUG, "\tmin_single_stride_log_num_of_bytes: %d",
1666                         mprq_caps.min_single_stride_log_num_of_bytes);
1667                 DRV_LOG(DEBUG, "\tmax_single_stride_log_num_of_bytes: %d",
1668                         mprq_caps.max_single_stride_log_num_of_bytes);
1669                 DRV_LOG(DEBUG, "\tmin_single_wqe_log_num_of_strides: %d",
1670                         mprq_caps.min_single_wqe_log_num_of_strides);
1671                 DRV_LOG(DEBUG, "\tmax_single_wqe_log_num_of_strides: %d",
1672                         mprq_caps.max_single_wqe_log_num_of_strides);
1673                 DRV_LOG(DEBUG, "\tsupported_qpts: %d",
1674                         mprq_caps.supported_qpts);
1675                 DRV_LOG(DEBUG, "device supports Multi-Packet RQ");
1676                 mprq = 1;
1677                 mprq_min_stride_size_n =
1678                         mprq_caps.min_single_stride_log_num_of_bytes;
1679                 mprq_max_stride_size_n =
1680                         mprq_caps.max_single_stride_log_num_of_bytes;
1681                 mprq_min_stride_num_n =
1682                         mprq_caps.min_single_wqe_log_num_of_strides;
1683                 mprq_max_stride_num_n =
1684                         mprq_caps.max_single_wqe_log_num_of_strides;
1685                 config.mprq.stride_num_n = RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1686                                                    mprq_min_stride_num_n);
1687         }
1688 #endif
1689         if (RTE_CACHE_LINE_SIZE == 128 &&
1690             !(dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
1691                 cqe_comp = 0;
1692         else
1693                 cqe_comp = 1;
1694         config.cqe_comp = cqe_comp;
1695 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
1696         /* Whether device supports 128B Rx CQE padding. */
1697         cqe_pad = RTE_CACHE_LINE_SIZE == 128 &&
1698                   (dv_attr.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_PAD);
1699 #endif
1700 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1701         if (dv_attr.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
1702                 tunnel_en = ((dv_attr.tunnel_offloads_caps &
1703                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
1704                              (dv_attr.tunnel_offloads_caps &
1705                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
1706         }
1707         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
1708                 tunnel_en ? "" : "not ");
1709 #else
1710         DRV_LOG(WARNING,
1711                 "tunnel offloading disabled due to old OFED/rdma-core version");
1712 #endif
1713         config.tunnel_en = tunnel_en;
1714 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1715         mpls_en = ((dv_attr.tunnel_offloads_caps &
1716                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_GRE) &&
1717                    (dv_attr.tunnel_offloads_caps &
1718                     MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_CW_MPLS_OVER_UDP));
1719         DRV_LOG(DEBUG, "MPLS over GRE/UDP tunnel offloading is %ssupported",
1720                 mpls_en ? "" : "not ");
1721 #else
1722         DRV_LOG(WARNING, "MPLS over GRE/UDP tunnel offloading disabled due to"
1723                 " old OFED/rdma-core version or firmware configuration");
1724 #endif
1725         config.mpls_en = mpls_en;
1726         /* Check port status. */
1727         err = mlx5_glue->query_port(sh->ctx, spawn->ibv_port, &port_attr);
1728         if (err) {
1729                 DRV_LOG(ERR, "port query failed: %s", strerror(err));
1730                 goto error;
1731         }
1732         if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
1733                 DRV_LOG(ERR, "port is not configured in Ethernet mode");
1734                 err = EINVAL;
1735                 goto error;
1736         }
1737         if (port_attr.state != IBV_PORT_ACTIVE)
1738                 DRV_LOG(DEBUG, "port is not active: \"%s\" (%d)",
1739                         mlx5_glue->port_state_str(port_attr.state),
1740                         port_attr.state);
1741         /* Allocate private eth device data. */
1742         priv = rte_zmalloc("ethdev private structure",
1743                            sizeof(*priv),
1744                            RTE_CACHE_LINE_SIZE);
1745         if (priv == NULL) {
1746                 DRV_LOG(ERR, "priv allocation failure");
1747                 err = ENOMEM;
1748                 goto error;
1749         }
1750         priv->sh = sh;
1751         priv->ibv_port = spawn->ibv_port;
1752         priv->mtu = RTE_ETHER_MTU;
1753 #ifndef RTE_ARCH_64
1754         /* Initialize UAR access locks for 32bit implementations. */
1755         rte_spinlock_init(&priv->uar_lock_cq);
1756         for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
1757                 rte_spinlock_init(&priv->uar_lock[i]);
1758 #endif
1759         /* Some internal functions rely on Netlink sockets, open them now. */
1760         priv->nl_socket_rdma = mlx5_nl_init(NETLINK_RDMA);
1761         priv->nl_socket_route = mlx5_nl_init(NETLINK_ROUTE);
1762         priv->nl_sn = 0;
1763         priv->representor = !!switch_info->representor;
1764         priv->master = !!switch_info->master;
1765         priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1766         /*
1767          * Currently we support single E-Switch per PF configurations
1768          * only and vport_id field contains the vport index for
1769          * associated VF, which is deduced from representor port name.
1770          * For example, let's have the IB device port 10, it has
1771          * attached network device eth0, which has port name attribute
1772          * pf0vf2, we can deduce the VF number as 2, and set vport index
1773          * as 3 (2+1). This assigning schema should be changed if the
1774          * multiple E-Switch instances per PF configurations or/and PCI
1775          * subfunctions are added.
1776          */
1777         priv->vport_id = switch_info->representor ?
1778                          switch_info->port_name + 1 : -1;
1779         /* representor_id field keeps the unmodified port/VF index. */
1780         priv->representor_id = switch_info->representor ?
1781                                switch_info->port_name : -1;
1782         /*
1783          * Look for sibling devices in order to reuse their switch domain
1784          * if any, otherwise allocate one.
1785          */
1786         RTE_ETH_FOREACH_DEV_OF(port_id, dpdk_dev) {
1787                 const struct mlx5_priv *opriv =
1788                         rte_eth_devices[port_id].data->dev_private;
1789
1790                 if (!opriv ||
1791                         opriv->domain_id ==
1792                         RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID)
1793                         continue;
1794                 priv->domain_id = opriv->domain_id;
1795                 break;
1796         }
1797         if (priv->domain_id == RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1798                 err = rte_eth_switch_domain_alloc(&priv->domain_id);
1799                 if (err) {
1800                         err = rte_errno;
1801                         DRV_LOG(ERR, "unable to allocate switch domain: %s",
1802                                 strerror(rte_errno));
1803                         goto error;
1804                 }
1805                 own_domain_id = 1;
1806         }
1807         err = mlx5_args(&config, dpdk_dev->devargs);
1808         if (err) {
1809                 err = rte_errno;
1810                 DRV_LOG(ERR, "failed to process device arguments: %s",
1811                         strerror(rte_errno));
1812                 goto error;
1813         }
1814         config.hw_csum = !!(sh->device_attr.device_cap_flags_ex &
1815                             IBV_DEVICE_RAW_IP_CSUM);
1816         DRV_LOG(DEBUG, "checksum offloading is %ssupported",
1817                 (config.hw_csum ? "" : "not "));
1818 #if !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) && \
1819         !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1820         DRV_LOG(DEBUG, "counters are not supported");
1821 #endif
1822 #ifndef HAVE_IBV_FLOW_DV_SUPPORT
1823         if (config.dv_flow_en) {
1824                 DRV_LOG(WARNING, "DV flow is not supported");
1825                 config.dv_flow_en = 0;
1826         }
1827 #endif
1828         config.ind_table_max_size =
1829                 sh->device_attr.rss_caps.max_rwq_indirection_table_size;
1830         /*
1831          * Remove this check once DPDK supports larger/variable
1832          * indirection tables.
1833          */
1834         if (config.ind_table_max_size > (unsigned int)ETH_RSS_RETA_SIZE_512)
1835                 config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
1836         DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
1837                 config.ind_table_max_size);
1838         config.hw_vlan_strip = !!(sh->device_attr.raw_packet_caps &
1839                                   IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
1840         DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
1841                 (config.hw_vlan_strip ? "" : "not "));
1842         config.hw_fcs_strip = !!(sh->device_attr.raw_packet_caps &
1843                                  IBV_RAW_PACKET_CAP_SCATTER_FCS);
1844         DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
1845                 (config.hw_fcs_strip ? "" : "not "));
1846 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
1847         hw_padding = !!sh->device_attr.rx_pad_end_addr_align;
1848 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
1849         hw_padding = !!(sh->device_attr.device_cap_flags_ex &
1850                         IBV_DEVICE_PCI_WRITE_END_PADDING);
1851 #endif
1852         if (config.hw_padding && !hw_padding) {
1853                 DRV_LOG(DEBUG, "Rx end alignment padding isn't supported");
1854                 config.hw_padding = 0;
1855         } else if (config.hw_padding) {
1856                 DRV_LOG(DEBUG, "Rx end alignment padding is enabled");
1857         }
1858         config.tso = (sh->device_attr.tso_caps.max_tso > 0 &&
1859                       (sh->device_attr.tso_caps.supported_qpts &
1860                        (1 << IBV_QPT_RAW_PACKET)));
1861         if (config.tso)
1862                 config.tso_max_payload_sz = sh->device_attr.tso_caps.max_tso;
1863         /*
1864          * MPW is disabled by default, while the Enhanced MPW is enabled
1865          * by default.
1866          */
1867         if (config.mps == MLX5_ARG_UNSET)
1868                 config.mps = (mps == MLX5_MPW_ENHANCED) ? MLX5_MPW_ENHANCED :
1869                                                           MLX5_MPW_DISABLED;
1870         else
1871                 config.mps = config.mps ? mps : MLX5_MPW_DISABLED;
1872         DRV_LOG(INFO, "%sMPS is %s",
1873                 config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
1874                 config.mps != MLX5_MPW_DISABLED ? "enabled" : "disabled");
1875         if (config.cqe_comp && !cqe_comp) {
1876                 DRV_LOG(WARNING, "Rx CQE compression isn't supported");
1877                 config.cqe_comp = 0;
1878         }
1879         if (config.cqe_pad && !cqe_pad) {
1880                 DRV_LOG(WARNING, "Rx CQE padding isn't supported");
1881                 config.cqe_pad = 0;
1882         } else if (config.cqe_pad) {
1883                 DRV_LOG(INFO, "Rx CQE padding is enabled");
1884         }
1885         if (config.devx) {
1886                 priv->counter_fallback = 0;
1887                 err = mlx5_devx_cmd_query_hca_attr(sh->ctx, &config.hca_attr);
1888                 if (err) {
1889                         err = -err;
1890                         goto error;
1891                 }
1892                 if (!config.hca_attr.flow_counters_dump)
1893                         priv->counter_fallback = 1;
1894 #ifndef HAVE_IBV_DEVX_ASYNC
1895                 priv->counter_fallback = 1;
1896 #endif
1897                 if (priv->counter_fallback)
1898                         DRV_LOG(INFO, "Use fall-back DV counter management\n");
1899                 /* Check for LRO support. */
1900                 if (config.dest_tir && config.hca_attr.lro_cap) {
1901                         /* TBD check tunnel lro caps. */
1902                         config.lro.supported = config.hca_attr.lro_cap;
1903                         DRV_LOG(DEBUG, "Device supports LRO");
1904                         /*
1905                          * If LRO timeout is not configured by application,
1906                          * use the minimal supported value.
1907                          */
1908                         if (!config.lro.timeout)
1909                                 config.lro.timeout =
1910                                 config.hca_attr.lro_timer_supported_periods[0];
1911                         DRV_LOG(DEBUG, "LRO session timeout set to %d usec",
1912                                 config.lro.timeout);
1913                 }
1914         }
1915         if (config.mprq.enabled && mprq) {
1916                 if (config.mprq.stride_num_n > mprq_max_stride_num_n ||
1917                     config.mprq.stride_num_n < mprq_min_stride_num_n) {
1918                         config.mprq.stride_num_n =
1919                                 RTE_MAX(MLX5_MPRQ_STRIDE_NUM_N,
1920                                         mprq_min_stride_num_n);
1921                         DRV_LOG(WARNING,
1922                                 "the number of strides"
1923                                 " for Multi-Packet RQ is out of range,"
1924                                 " setting default value (%u)",
1925                                 1 << config.mprq.stride_num_n);
1926                 }
1927                 config.mprq.min_stride_size_n = mprq_min_stride_size_n;
1928                 config.mprq.max_stride_size_n = mprq_max_stride_size_n;
1929         } else if (config.mprq.enabled && !mprq) {
1930                 DRV_LOG(WARNING, "Multi-Packet RQ isn't supported");
1931                 config.mprq.enabled = 0;
1932         }
1933         if (config.max_dump_files_num == 0)
1934                 config.max_dump_files_num = 128;
1935         eth_dev = rte_eth_dev_allocate(name);
1936         if (eth_dev == NULL) {
1937                 DRV_LOG(ERR, "can not allocate rte ethdev");
1938                 err = ENOMEM;
1939                 goto error;
1940         }
1941         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
1942         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1943         if (priv->representor) {
1944                 eth_dev->data->dev_flags |= RTE_ETH_DEV_REPRESENTOR;
1945                 eth_dev->data->representor_id = priv->representor_id;
1946         }
1947         /*
1948          * Store associated network device interface index. This index
1949          * is permanent throughout the lifetime of device. So, we may store
1950          * the ifindex here and use the cached value further.
1951          */
1952         assert(spawn->ifindex);
1953         priv->if_index = spawn->ifindex;
1954         eth_dev->data->dev_private = priv;
1955         priv->dev_data = eth_dev->data;
1956         eth_dev->data->mac_addrs = priv->mac;
1957         eth_dev->device = dpdk_dev;
1958         /* Configure the first MAC address by default. */
1959         if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
1960                 DRV_LOG(ERR,
1961                         "port %u cannot get MAC address, is mlx5_en"
1962                         " loaded? (errno: %s)",
1963                         eth_dev->data->port_id, strerror(rte_errno));
1964                 err = ENODEV;
1965                 goto error;
1966         }
1967         DRV_LOG(INFO,
1968                 "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
1969                 eth_dev->data->port_id,
1970                 mac.addr_bytes[0], mac.addr_bytes[1],
1971                 mac.addr_bytes[2], mac.addr_bytes[3],
1972                 mac.addr_bytes[4], mac.addr_bytes[5]);
1973 #ifndef NDEBUG
1974         {
1975                 char ifname[IF_NAMESIZE];
1976
1977                 if (mlx5_get_ifname(eth_dev, &ifname) == 0)
1978                         DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
1979                                 eth_dev->data->port_id, ifname);
1980                 else
1981                         DRV_LOG(DEBUG, "port %u ifname is unknown",
1982                                 eth_dev->data->port_id);
1983         }
1984 #endif
1985         /* Get actual MTU if possible. */
1986         err = mlx5_get_mtu(eth_dev, &priv->mtu);
1987         if (err) {
1988                 err = rte_errno;
1989                 goto error;
1990         }
1991         DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
1992                 priv->mtu);
1993         /* Initialize burst functions to prevent crashes before link-up. */
1994         eth_dev->rx_pkt_burst = removed_rx_burst;
1995         eth_dev->tx_pkt_burst = removed_tx_burst;
1996         eth_dev->dev_ops = &mlx5_dev_ops;
1997         /* Register MAC address. */
1998         claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
1999         if (config.vf && config.vf_nl_en)
2000                 mlx5_nl_mac_addr_sync(eth_dev);
2001         TAILQ_INIT(&priv->flows);
2002         TAILQ_INIT(&priv->ctrl_flows);
2003         /* Hint libmlx5 to use PMD allocator for data plane resources */
2004         struct mlx5dv_ctx_allocators alctr = {
2005                 .alloc = &mlx5_alloc_verbs_buf,
2006                 .free = &mlx5_free_verbs_buf,
2007                 .data = priv,
2008         };
2009         mlx5_glue->dv_set_context_attr(sh->ctx,
2010                                        MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
2011                                        (void *)((uintptr_t)&alctr));
2012         /* Bring Ethernet device up. */
2013         DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
2014                 eth_dev->data->port_id);
2015         mlx5_set_link_up(eth_dev);
2016         /*
2017          * Even though the interrupt handler is not installed yet,
2018          * interrupts will still trigger on the async_fd from
2019          * Verbs context returned by ibv_open_device().
2020          */
2021         mlx5_link_update(eth_dev, 0);
2022 #ifdef HAVE_MLX5DV_DR_ESWITCH
2023         if (!(config.hca_attr.eswitch_manager && config.dv_flow_en &&
2024               (switch_info->representor || switch_info->master)))
2025                 config.dv_esw_en = 0;
2026 #else
2027         config.dv_esw_en = 0;
2028 #endif
2029         /* Detect minimal data bytes to inline. */
2030         mlx5_set_min_inline(spawn, &config);
2031         /* Store device configuration on private structure. */
2032         priv->config = config;
2033         /* Create context for virtual machine VLAN workaround. */
2034         priv->vmwa_context = mlx5_vlan_vmwa_init(eth_dev, spawn->ifindex);
2035         if (config.dv_flow_en) {
2036                 err = mlx5_alloc_shared_dr(priv);
2037                 if (err)
2038                         goto error;
2039         }
2040         /* Supported Verbs flow priority number detection. */
2041         err = mlx5_flow_discover_priorities(eth_dev);
2042         if (err < 0) {
2043                 err = -err;
2044                 goto error;
2045         }
2046         priv->config.flow_prio = err;
2047         return eth_dev;
2048 error:
2049         if (priv) {
2050                 if (priv->sh)
2051                         mlx5_free_shared_dr(priv);
2052                 if (priv->nl_socket_route >= 0)
2053                         close(priv->nl_socket_route);
2054                 if (priv->nl_socket_rdma >= 0)
2055                         close(priv->nl_socket_rdma);
2056                 if (priv->vmwa_context)
2057                         mlx5_vlan_vmwa_exit(priv->vmwa_context);
2058                 if (own_domain_id)
2059                         claim_zero(rte_eth_switch_domain_free(priv->domain_id));
2060                 rte_free(priv);
2061                 if (eth_dev != NULL)
2062                         eth_dev->data->dev_private = NULL;
2063         }
2064         if (eth_dev != NULL) {
2065                 /* mac_addrs must not be freed alone because part of dev_private */
2066                 eth_dev->data->mac_addrs = NULL;
2067                 rte_eth_dev_release_port(eth_dev);
2068         }
2069         if (sh)
2070                 mlx5_free_shared_ibctx(sh);
2071         assert(err > 0);
2072         rte_errno = err;
2073         return NULL;
2074 }
2075
2076 /**
2077  * Comparison callback to sort device data.
2078  *
2079  * This is meant to be used with qsort().
2080  *
2081  * @param a[in]
2082  *   Pointer to pointer to first data object.
2083  * @param b[in]
2084  *   Pointer to pointer to second data object.
2085  *
2086  * @return
2087  *   0 if both objects are equal, less than 0 if the first argument is less
2088  *   than the second, greater than 0 otherwise.
2089  */
2090 static int
2091 mlx5_dev_spawn_data_cmp(const void *a, const void *b)
2092 {
2093         const struct mlx5_switch_info *si_a =
2094                 &((const struct mlx5_dev_spawn_data *)a)->info;
2095         const struct mlx5_switch_info *si_b =
2096                 &((const struct mlx5_dev_spawn_data *)b)->info;
2097         int ret;
2098
2099         /* Master device first. */
2100         ret = si_b->master - si_a->master;
2101         if (ret)
2102                 return ret;
2103         /* Then representor devices. */
2104         ret = si_b->representor - si_a->representor;
2105         if (ret)
2106                 return ret;
2107         /* Unidentified devices come last in no specific order. */
2108         if (!si_a->representor)
2109                 return 0;
2110         /* Order representors by name. */
2111         return si_a->port_name - si_b->port_name;
2112 }
2113
2114 /**
2115  * DPDK callback to register a PCI device.
2116  *
2117  * This function spawns Ethernet devices out of a given PCI device.
2118  *
2119  * @param[in] pci_drv
2120  *   PCI driver structure (mlx5_driver).
2121  * @param[in] pci_dev
2122  *   PCI device information.
2123  *
2124  * @return
2125  *   0 on success, a negative errno value otherwise and rte_errno is set.
2126  */
2127 static int
2128 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2129                struct rte_pci_device *pci_dev)
2130 {
2131         struct ibv_device **ibv_list;
2132         /*
2133          * Number of found IB Devices matching with requested PCI BDF.
2134          * nd != 1 means there are multiple IB devices over the same
2135          * PCI device and we have representors and master.
2136          */
2137         unsigned int nd = 0;
2138         /*
2139          * Number of found IB device Ports. nd = 1 and np = 1..n means
2140          * we have the single multiport IB device, and there may be
2141          * representors attached to some of found ports.
2142          */
2143         unsigned int np = 0;
2144         /*
2145          * Number of DPDK ethernet devices to Spawn - either over
2146          * multiple IB devices or multiple ports of single IB device.
2147          * Actually this is the number of iterations to spawn.
2148          */
2149         unsigned int ns = 0;
2150         struct mlx5_dev_config dev_config;
2151         int ret;
2152
2153         ret = mlx5_init_once();
2154         if (ret) {
2155                 DRV_LOG(ERR, "unable to init PMD global data: %s",
2156                         strerror(rte_errno));
2157                 return -rte_errno;
2158         }
2159         assert(pci_drv == &mlx5_driver);
2160         errno = 0;
2161         ibv_list = mlx5_glue->get_device_list(&ret);
2162         if (!ibv_list) {
2163                 rte_errno = errno ? errno : ENOSYS;
2164                 DRV_LOG(ERR, "cannot list devices, is ib_uverbs loaded?");
2165                 return -rte_errno;
2166         }
2167         /*
2168          * First scan the list of all Infiniband devices to find
2169          * matching ones, gathering into the list.
2170          */
2171         struct ibv_device *ibv_match[ret + 1];
2172         int nl_route = -1;
2173         int nl_rdma = -1;
2174         unsigned int i;
2175
2176         while (ret-- > 0) {
2177                 struct rte_pci_addr pci_addr;
2178
2179                 DRV_LOG(DEBUG, "checking device \"%s\"", ibv_list[ret]->name);
2180                 if (mlx5_ibv_device_to_pci_addr(ibv_list[ret], &pci_addr))
2181                         continue;
2182                 if (pci_dev->addr.domain != pci_addr.domain ||
2183                     pci_dev->addr.bus != pci_addr.bus ||
2184                     pci_dev->addr.devid != pci_addr.devid ||
2185                     pci_dev->addr.function != pci_addr.function)
2186                         continue;
2187                 DRV_LOG(INFO, "PCI information matches for device \"%s\"",
2188                         ibv_list[ret]->name);
2189                 ibv_match[nd++] = ibv_list[ret];
2190         }
2191         ibv_match[nd] = NULL;
2192         if (!nd) {
2193                 /* No device matches, just complain and bail out. */
2194                 mlx5_glue->free_device_list(ibv_list);
2195                 DRV_LOG(WARNING,
2196                         "no Verbs device matches PCI device " PCI_PRI_FMT ","
2197                         " are kernel drivers loaded?",
2198                         pci_dev->addr.domain, pci_dev->addr.bus,
2199                         pci_dev->addr.devid, pci_dev->addr.function);
2200                 rte_errno = ENOENT;
2201                 ret = -rte_errno;
2202                 return ret;
2203         }
2204         nl_route = mlx5_nl_init(NETLINK_ROUTE);
2205         nl_rdma = mlx5_nl_init(NETLINK_RDMA);
2206         if (nd == 1) {
2207                 /*
2208                  * Found single matching device may have multiple ports.
2209                  * Each port may be representor, we have to check the port
2210                  * number and check the representors existence.
2211                  */
2212                 if (nl_rdma >= 0)
2213                         np = mlx5_nl_portnum(nl_rdma, ibv_match[0]->name);
2214                 if (!np)
2215                         DRV_LOG(WARNING, "can not get IB device \"%s\""
2216                                          " ports number", ibv_match[0]->name);
2217         }
2218         /*
2219          * Now we can determine the maximal
2220          * amount of devices to be spawned.
2221          */
2222         struct mlx5_dev_spawn_data list[np ? np : nd];
2223
2224         if (np > 1) {
2225                 /*
2226                  * Single IB device with multiple ports found,
2227                  * it may be E-Switch master device and representors.
2228                  * We have to perform identification trough the ports.
2229                  */
2230                 assert(nl_rdma >= 0);
2231                 assert(ns == 0);
2232                 assert(nd == 1);
2233                 for (i = 1; i <= np; ++i) {
2234                         list[ns].max_port = np;
2235                         list[ns].ibv_port = i;
2236                         list[ns].ibv_dev = ibv_match[0];
2237                         list[ns].eth_dev = NULL;
2238                         list[ns].pci_dev = pci_dev;
2239                         list[ns].ifindex = mlx5_nl_ifindex
2240                                         (nl_rdma, list[ns].ibv_dev->name, i);
2241                         if (!list[ns].ifindex) {
2242                                 /*
2243                                  * No network interface index found for the
2244                                  * specified port, it means there is no
2245                                  * representor on this port. It's OK,
2246                                  * there can be disabled ports, for example
2247                                  * if sriov_numvfs < sriov_totalvfs.
2248                                  */
2249                                 continue;
2250                         }
2251                         ret = -1;
2252                         if (nl_route >= 0)
2253                                 ret = mlx5_nl_switch_info
2254                                                (nl_route,
2255                                                 list[ns].ifindex,
2256                                                 &list[ns].info);
2257                         if (ret || (!list[ns].info.representor &&
2258                                     !list[ns].info.master)) {
2259                                 /*
2260                                  * We failed to recognize representors with
2261                                  * Netlink, let's try to perform the task
2262                                  * with sysfs.
2263                                  */
2264                                 ret =  mlx5_sysfs_switch_info
2265                                                 (list[ns].ifindex,
2266                                                  &list[ns].info);
2267                         }
2268                         if (!ret && (list[ns].info.representor ^
2269                                      list[ns].info.master))
2270                                 ns++;
2271                 }
2272                 if (!ns) {
2273                         DRV_LOG(ERR,
2274                                 "unable to recognize master/representors"
2275                                 " on the IB device with multiple ports");
2276                         rte_errno = ENOENT;
2277                         ret = -rte_errno;
2278                         goto exit;
2279                 }
2280         } else {
2281                 /*
2282                  * The existence of several matching entries (nd > 1) means
2283                  * port representors have been instantiated. No existing Verbs
2284                  * call nor sysfs entries can tell them apart, this can only
2285                  * be done through Netlink calls assuming kernel drivers are
2286                  * recent enough to support them.
2287                  *
2288                  * In the event of identification failure through Netlink,
2289                  * try again through sysfs, then:
2290                  *
2291                  * 1. A single IB device matches (nd == 1) with single
2292                  *    port (np=0/1) and is not a representor, assume
2293                  *    no switch support.
2294                  *
2295                  * 2. Otherwise no safe assumptions can be made;
2296                  *    complain louder and bail out.
2297                  */
2298                 np = 1;
2299                 for (i = 0; i != nd; ++i) {
2300                         memset(&list[ns].info, 0, sizeof(list[ns].info));
2301                         list[ns].max_port = 1;
2302                         list[ns].ibv_port = 1;
2303                         list[ns].ibv_dev = ibv_match[i];
2304                         list[ns].eth_dev = NULL;
2305                         list[ns].pci_dev = pci_dev;
2306                         list[ns].ifindex = 0;
2307                         if (nl_rdma >= 0)
2308                                 list[ns].ifindex = mlx5_nl_ifindex
2309                                         (nl_rdma, list[ns].ibv_dev->name, 1);
2310                         if (!list[ns].ifindex) {
2311                                 char ifname[IF_NAMESIZE];
2312
2313                                 /*
2314                                  * Netlink failed, it may happen with old
2315                                  * ib_core kernel driver (before 4.16).
2316                                  * We can assume there is old driver because
2317                                  * here we are processing single ports IB
2318                                  * devices. Let's try sysfs to retrieve
2319                                  * the ifindex. The method works for
2320                                  * master device only.
2321                                  */
2322                                 if (nd > 1) {
2323                                         /*
2324                                          * Multiple devices found, assume
2325                                          * representors, can not distinguish
2326                                          * master/representor and retrieve
2327                                          * ifindex via sysfs.
2328                                          */
2329                                         continue;
2330                                 }
2331                                 ret = mlx5_get_master_ifname
2332                                         (ibv_match[i]->ibdev_path, &ifname);
2333                                 if (!ret)
2334                                         list[ns].ifindex =
2335                                                 if_nametoindex(ifname);
2336                                 if (!list[ns].ifindex) {
2337                                         /*
2338                                          * No network interface index found
2339                                          * for the specified device, it means
2340                                          * there it is neither representor
2341                                          * nor master.
2342                                          */
2343                                         continue;
2344                                 }
2345                         }
2346                         ret = -1;
2347                         if (nl_route >= 0)
2348                                 ret = mlx5_nl_switch_info
2349                                                (nl_route,
2350                                                 list[ns].ifindex,
2351                                                 &list[ns].info);
2352                         if (ret || (!list[ns].info.representor &&
2353                                     !list[ns].info.master)) {
2354                                 /*
2355                                  * We failed to recognize representors with
2356                                  * Netlink, let's try to perform the task
2357                                  * with sysfs.
2358                                  */
2359                                 ret =  mlx5_sysfs_switch_info
2360                                                 (list[ns].ifindex,
2361                                                  &list[ns].info);
2362                         }
2363                         if (!ret && (list[ns].info.representor ^
2364                                      list[ns].info.master)) {
2365                                 ns++;
2366                         } else if ((nd == 1) &&
2367                                    !list[ns].info.representor &&
2368                                    !list[ns].info.master) {
2369                                 /*
2370                                  * Single IB device with
2371                                  * one physical port and
2372                                  * attached network device.
2373                                  * May be SRIOV is not enabled
2374                                  * or there is no representors.
2375                                  */
2376                                 DRV_LOG(INFO, "no E-Switch support detected");
2377                                 ns++;
2378                                 break;
2379                         }
2380                 }
2381                 if (!ns) {
2382                         DRV_LOG(ERR,
2383                                 "unable to recognize master/representors"
2384                                 " on the multiple IB devices");
2385                         rte_errno = ENOENT;
2386                         ret = -rte_errno;
2387                         goto exit;
2388                 }
2389         }
2390         assert(ns);
2391         /*
2392          * Sort list to probe devices in natural order for users convenience
2393          * (i.e. master first, then representors from lowest to highest ID).
2394          */
2395         qsort(list, ns, sizeof(*list), mlx5_dev_spawn_data_cmp);
2396         /* Default configuration. */
2397         dev_config = (struct mlx5_dev_config){
2398                 .hw_padding = 0,
2399                 .mps = MLX5_ARG_UNSET,
2400                 .rx_vec_en = 1,
2401                 .txq_inline_max = MLX5_ARG_UNSET,
2402                 .txq_inline_min = MLX5_ARG_UNSET,
2403                 .txq_inline_mpw = MLX5_ARG_UNSET,
2404                 .txqs_inline = MLX5_ARG_UNSET,
2405                 .vf_nl_en = 1,
2406                 .mr_ext_memseg_en = 1,
2407                 .mprq = {
2408                         .enabled = 0, /* Disabled by default. */
2409                         .stride_num_n = MLX5_MPRQ_STRIDE_NUM_N,
2410                         .max_memcpy_len = MLX5_MPRQ_MEMCPY_DEFAULT_LEN,
2411                         .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
2412                 },
2413                 .dv_esw_en = 1,
2414         };
2415         /* Device specific configuration. */
2416         switch (pci_dev->id.device_id) {
2417         case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
2418         case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
2419         case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
2420         case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
2421                 dev_config.vf = 1;
2422                 break;
2423         default:
2424                 break;
2425         }
2426         for (i = 0; i != ns; ++i) {
2427                 uint32_t restore;
2428
2429                 list[i].eth_dev = mlx5_dev_spawn(&pci_dev->device,
2430                                                  &list[i],
2431                                                  dev_config);
2432                 if (!list[i].eth_dev) {
2433                         if (rte_errno != EBUSY && rte_errno != EEXIST)
2434                                 break;
2435                         /* Device is disabled or already spawned. Ignore it. */
2436                         continue;
2437                 }
2438                 restore = list[i].eth_dev->data->dev_flags;
2439                 rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
2440                 /* Restore non-PCI flags cleared by the above call. */
2441                 list[i].eth_dev->data->dev_flags |= restore;
2442                 rte_eth_dev_probing_finish(list[i].eth_dev);
2443         }
2444         if (i != ns) {
2445                 DRV_LOG(ERR,
2446                         "probe of PCI device " PCI_PRI_FMT " aborted after"
2447                         " encountering an error: %s",
2448                         pci_dev->addr.domain, pci_dev->addr.bus,
2449                         pci_dev->addr.devid, pci_dev->addr.function,
2450                         strerror(rte_errno));
2451                 ret = -rte_errno;
2452                 /* Roll back. */
2453                 while (i--) {
2454                         if (!list[i].eth_dev)
2455                                 continue;
2456                         mlx5_dev_close(list[i].eth_dev);
2457                         /* mac_addrs must not be freed because in dev_private */
2458                         list[i].eth_dev->data->mac_addrs = NULL;
2459                         claim_zero(rte_eth_dev_release_port(list[i].eth_dev));
2460                 }
2461                 /* Restore original error. */
2462                 rte_errno = -ret;
2463         } else {
2464                 ret = 0;
2465         }
2466 exit:
2467         /*
2468          * Do the routine cleanup:
2469          * - close opened Netlink sockets
2470          * - free the Infiniband device list
2471          */
2472         if (nl_rdma >= 0)
2473                 close(nl_rdma);
2474         if (nl_route >= 0)
2475                 close(nl_route);
2476         assert(ibv_list);
2477         mlx5_glue->free_device_list(ibv_list);
2478         return ret;
2479 }
2480
2481 /**
2482  * DPDK callback to remove a PCI device.
2483  *
2484  * This function removes all Ethernet devices belong to a given PCI device.
2485  *
2486  * @param[in] pci_dev
2487  *   Pointer to the PCI device.
2488  *
2489  * @return
2490  *   0 on success, the function cannot fail.
2491  */
2492 static int
2493 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2494 {
2495         uint16_t port_id;
2496
2497         RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device)
2498                 rte_eth_dev_close(port_id);
2499         return 0;
2500 }
2501
2502 static const struct rte_pci_id mlx5_pci_id_map[] = {
2503         {
2504                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2505                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2506         },
2507         {
2508                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2509                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2510         },
2511         {
2512                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2513                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2514         },
2515         {
2516                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2517                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2518         },
2519         {
2520                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2521                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2522         },
2523         {
2524                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2525                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2526         },
2527         {
2528                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2529                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2530         },
2531         {
2532                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2533                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2534         },
2535         {
2536                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2537                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2538         },
2539         {
2540                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2541                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2542         },
2543         {
2544                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2545                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2546         },
2547         {
2548                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2549                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2550         },
2551         {
2552                 .vendor_id = 0
2553         }
2554 };
2555
2556 static struct rte_pci_driver mlx5_driver = {
2557         .driver = {
2558                 .name = MLX5_DRIVER_NAME
2559         },
2560         .id_table = mlx5_pci_id_map,
2561         .probe = mlx5_pci_probe,
2562         .remove = mlx5_pci_remove,
2563         .dma_map = mlx5_dma_map,
2564         .dma_unmap = mlx5_dma_unmap,
2565         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
2566                      RTE_PCI_DRV_PROBE_AGAIN,
2567 };
2568
2569 #ifdef RTE_IBVERBS_LINK_DLOPEN
2570
2571 /**
2572  * Suffix RTE_EAL_PMD_PATH with "-glue".
2573  *
2574  * This function performs a sanity check on RTE_EAL_PMD_PATH before
2575  * suffixing its last component.
2576  *
2577  * @param buf[out]
2578  *   Output buffer, should be large enough otherwise NULL is returned.
2579  * @param size
2580  *   Size of @p out.
2581  *
2582  * @return
2583  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
2584  */
2585 static char *
2586 mlx5_glue_path(char *buf, size_t size)
2587 {
2588         static const char *const bad[] = { "/", ".", "..", NULL };
2589         const char *path = RTE_EAL_PMD_PATH;
2590         size_t len = strlen(path);
2591         size_t off;
2592         int i;
2593
2594         while (len && path[len - 1] == '/')
2595                 --len;
2596         for (off = len; off && path[off - 1] != '/'; --off)
2597                 ;
2598         for (i = 0; bad[i]; ++i)
2599                 if (!strncmp(path + off, bad[i], (int)(len - off)))
2600                         goto error;
2601         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
2602         if (i == -1 || (size_t)i >= size)
2603                 goto error;
2604         return buf;
2605 error:
2606         DRV_LOG(ERR,
2607                 "unable to append \"-glue\" to last component of"
2608                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
2609                 " please re-configure DPDK");
2610         return NULL;
2611 }
2612
2613 /**
2614  * Initialization routine for run-time dependency on rdma-core.
2615  */
2616 static int
2617 mlx5_glue_init(void)
2618 {
2619         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
2620         const char *path[] = {
2621                 /*
2622                  * A basic security check is necessary before trusting
2623                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
2624                  */
2625                 (geteuid() == getuid() && getegid() == getgid() ?
2626                  getenv("MLX5_GLUE_PATH") : NULL),
2627                 /*
2628                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
2629                  * variant, otherwise let dlopen() look up libraries on its
2630                  * own.
2631                  */
2632                 (*RTE_EAL_PMD_PATH ?
2633                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
2634         };
2635         unsigned int i = 0;
2636         void *handle = NULL;
2637         void **sym;
2638         const char *dlmsg;
2639
2640         while (!handle && i != RTE_DIM(path)) {
2641                 const char *end;
2642                 size_t len;
2643                 int ret;
2644
2645                 if (!path[i]) {
2646                         ++i;
2647                         continue;
2648                 }
2649                 end = strpbrk(path[i], ":;");
2650                 if (!end)
2651                         end = path[i] + strlen(path[i]);
2652                 len = end - path[i];
2653                 ret = 0;
2654                 do {
2655                         char name[ret + 1];
2656
2657                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
2658                                        (int)len, path[i],
2659                                        (!len || *(end - 1) == '/') ? "" : "/");
2660                         if (ret == -1)
2661                                 break;
2662                         if (sizeof(name) != (size_t)ret + 1)
2663                                 continue;
2664                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
2665                                 name);
2666                         handle = dlopen(name, RTLD_LAZY);
2667                         break;
2668                 } while (1);
2669                 path[i] = end + 1;
2670                 if (!*end)
2671                         ++i;
2672         }
2673         if (!handle) {
2674                 rte_errno = EINVAL;
2675                 dlmsg = dlerror();
2676                 if (dlmsg)
2677                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
2678                 goto glue_error;
2679         }
2680         sym = dlsym(handle, "mlx5_glue");
2681         if (!sym || !*sym) {
2682                 rte_errno = EINVAL;
2683                 dlmsg = dlerror();
2684                 if (dlmsg)
2685                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
2686                 goto glue_error;
2687         }
2688         mlx5_glue = *sym;
2689         return 0;
2690 glue_error:
2691         if (handle)
2692                 dlclose(handle);
2693         DRV_LOG(WARNING,
2694                 "cannot initialize PMD due to missing run-time dependency on"
2695                 " rdma-core libraries (libibverbs, libmlx5)");
2696         return -rte_errno;
2697 }
2698
2699 #endif
2700
2701 /**
2702  * Driver initialization routine.
2703  */
2704 RTE_INIT(rte_mlx5_pmd_init)
2705 {
2706         /* Initialize driver log type. */
2707         mlx5_logtype = rte_log_register("pmd.net.mlx5");
2708         if (mlx5_logtype >= 0)
2709                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
2710
2711         /* Build the static tables for Verbs conversion. */
2712         mlx5_set_ptype_table();
2713         mlx5_set_cksum_table();
2714         mlx5_set_swp_types_table();
2715         /*
2716          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
2717          * huge pages. Calling ibv_fork_init() during init allows
2718          * applications to use fork() safely for purposes other than
2719          * using this PMD, which is not supported in forked processes.
2720          */
2721         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
2722         /* Match the size of Rx completion entry to the size of a cacheline. */
2723         if (RTE_CACHE_LINE_SIZE == 128)
2724                 setenv("MLX5_CQE_SIZE", "128", 0);
2725         /*
2726          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
2727          * cleanup all the Verbs resources even when the device was removed.
2728          */
2729         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
2730 #ifdef RTE_IBVERBS_LINK_DLOPEN
2731         if (mlx5_glue_init())
2732                 return;
2733         assert(mlx5_glue);
2734 #endif
2735 #ifndef NDEBUG
2736         /* Glue structure must not contain any NULL pointers. */
2737         {
2738                 unsigned int i;
2739
2740                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
2741                         assert(((const void *const *)mlx5_glue)[i]);
2742         }
2743 #endif
2744         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
2745                 DRV_LOG(ERR,
2746                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
2747                         mlx5_glue->version, MLX5_GLUE_VERSION);
2748                 return;
2749         }
2750         mlx5_glue->fork_init();
2751         rte_pci_register(&mlx5_driver);
2752 }
2753
2754 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2755 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2756 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");