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_eal_memconfig.h>
  36 #include <rte_kvargs.h>
  37
  38 #include "mlx5.h"
  39 #include "mlx5_utils.h"
  40 #include "mlx5_rxtx.h"
  41 #include "mlx5_autoconf.h"
  42 #include "mlx5_defs.h"
  43 #include "mlx5_glue.h"
  44
  45 /* Device parameter to enable RX completion queue compression. */
  46 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
  47
  48 /* Device parameter to configure inline send. */
  49 #define MLX5_TXQ_INLINE "txq_inline"
  50
  51 /*
  52  * Device parameter to configure the number of TX queues threshold for
  53  * enabling inline send.
  54  */
  55 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
  56
  57 /* Device parameter to enable multi-packet send WQEs. */
  58 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
  59
  60 /* Device parameter to include 2 dsegs in the title WQEBB. */
  61 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
  62
  63 /* Device parameter to limit the size of inlining packet. */
  64 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
  65
  66 /* Device parameter to enable hardware Tx vector. */
  67 #define MLX5_TX_VEC_EN "tx_vec_en"
  68
  69 /* Device parameter to enable hardware Rx vector. */
  70 #define MLX5_RX_VEC_EN "rx_vec_en"
  71
  72 /* Activate Netlink support in VF mode. */
  73 #define MLX5_VF_NL_EN "vf_nl_en"
  74
  75 #ifndef HAVE_IBV_MLX5_MOD_MPW
  76 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
  77 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
  78 #endif
  79
  80 #ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
  81 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
  82 #endif
  83
  84 /** Driver-specific log messages type. */
  85 int mlx5_logtype;
  86
  87 /**
  88  * Retrieve integer value from environment variable.
  89  *
  90  * @param[in] name
  91  *   Environment variable name.
  92  *
  93  * @return
  94  *   Integer value, 0 if the variable is not set.
  95  */
  96 int
  97 mlx5_getenv_int(const char *name)
  98 {
  99         const char *val = getenv(name);
 100
 101         if (val == NULL)
 102                 return 0;
 103         return atoi(val);
 104 }
 105
 106 /**
 107  * Verbs callback to allocate a memory. This function should allocate the space
 108  * according to the size provided residing inside a huge page.
 109  * Please note that all allocation must respect the alignment from libmlx5
 110  * (i.e. currently sysconf(_SC_PAGESIZE)).
 111  *
 112  * @param[in] size
 113  *   The size in bytes of the memory to allocate.
 114  * @param[in] data
 115  *   A pointer to the callback data.
 116  *
 117  * @return
 118  *   Allocated buffer, NULL otherwise and rte_errno is set.
 119  */
 120 static void *
 121 mlx5_alloc_verbs_buf(size_t size, void *data)
 122 {
 123         struct priv *priv = data;
 124         void *ret;
 125         size_t alignment = sysconf(_SC_PAGESIZE);
 126         unsigned int socket = SOCKET_ID_ANY;
 127
 128         if (priv->verbs_alloc_ctx.type == MLX5_VERBS_ALLOC_TYPE_TX_QUEUE) {
 129                 const struct mlx5_txq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 130
 131                 socket = ctrl->socket;
 132         } else if (priv->verbs_alloc_ctx.type ==
 133                    MLX5_VERBS_ALLOC_TYPE_RX_QUEUE) {
 134                 const struct mlx5_rxq_ctrl *ctrl = priv->verbs_alloc_ctx.obj;
 135
 136                 socket = ctrl->socket;
 137         }
 138         assert(data != NULL);
 139         ret = rte_malloc_socket(__func__, size, alignment, socket);
 140         if (!ret && size)
 141                 rte_errno = ENOMEM;
 142         return ret;
 143 }
 144
 145 /**
 146  * Verbs callback to free a memory.
 147  *
 148  * @param[in] ptr
 149  *   A pointer to the memory to free.
 150  * @param[in] data
 151  *   A pointer to the callback data.
 152  */
 153 static void
 154 mlx5_free_verbs_buf(void *ptr, void *data __rte_unused)
 155 {
 156         assert(data != NULL);
 157         rte_free(ptr);
 158 }
 159
 160 /**
 161  * DPDK callback to close the device.
 162  *
 163  * Destroy all queues and objects, free memory.
 164  *
 165  * @param dev
 166  *   Pointer to Ethernet device structure.
 167  */
 168 static void
 169 mlx5_dev_close(struct rte_eth_dev *dev)
 170 {
 171         struct priv *priv = dev->data->dev_private;
 172         unsigned int i;
 173         int ret;
 174
 175         DRV_LOG(DEBUG, "port %u closing device \"%s\"",
 176                 dev->data->port_id,
 177                 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
 178         /* In case mlx5_dev_stop() has not been called. */
 179         mlx5_dev_interrupt_handler_uninstall(dev);
 180         mlx5_traffic_disable(dev);
 181         /* Prevent crashes when queues are still in use. */
 182         dev->rx_pkt_burst = removed_rx_burst;
 183         dev->tx_pkt_burst = removed_tx_burst;
 184         if (priv->rxqs != NULL) {
 185                 /* XXX race condition if mlx5_rx_burst() is still running. */
 186                 usleep(1000);
 187                 for (i = 0; (i != priv->rxqs_n); ++i)
 188                         mlx5_rxq_release(dev, i);
 189                 priv->rxqs_n = 0;
 190                 priv->rxqs = NULL;
 191         }
 192         if (priv->txqs != NULL) {
 193                 /* XXX race condition if mlx5_tx_burst() is still running. */
 194                 usleep(1000);
 195                 for (i = 0; (i != priv->txqs_n); ++i)
 196                         mlx5_txq_release(dev, i);
 197                 priv->txqs_n = 0;
 198                 priv->txqs = NULL;
 199         }
 200         if (priv->pd != NULL) {
 201                 assert(priv->ctx != NULL);
 202                 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
 203                 claim_zero(mlx5_glue->close_device(priv->ctx));
 204         } else
 205                 assert(priv->ctx == NULL);
 206         if (priv->rss_conf.rss_key != NULL)
 207                 rte_free(priv->rss_conf.rss_key);
 208         if (priv->reta_idx != NULL)
 209                 rte_free(priv->reta_idx);
 210         if (priv->primary_socket)
 211                 mlx5_socket_uninit(dev);
 212         if (priv->config.vf)
 213                 mlx5_nl_mac_addr_flush(dev);
 214         if (priv->nl_socket >= 0)
 215                 close(priv->nl_socket);
 216         ret = mlx5_hrxq_ibv_verify(dev);
 217         if (ret)
 218                 DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
 219                         dev->data->port_id);
 220         ret = mlx5_ind_table_ibv_verify(dev);
 221         if (ret)
 222                 DRV_LOG(WARNING, "port %u some indirection table still remain",
 223                         dev->data->port_id);
 224         ret = mlx5_rxq_ibv_verify(dev);
 225         if (ret)
 226                 DRV_LOG(WARNING, "port %u some Verbs Rx queue still remain",
 227                         dev->data->port_id);
 228         ret = mlx5_rxq_verify(dev);
 229         if (ret)
 230                 DRV_LOG(WARNING, "port %u some Rx queues still remain",
 231                         dev->data->port_id);
 232         ret = mlx5_txq_ibv_verify(dev);
 233         if (ret)
 234                 DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
 235                         dev->data->port_id);
 236         ret = mlx5_txq_verify(dev);
 237         if (ret)
 238                 DRV_LOG(WARNING, "port %u some Tx queues still remain",
 239                         dev->data->port_id);
 240         ret = mlx5_flow_verify(dev);
 241         if (ret)
 242                 DRV_LOG(WARNING, "port %u some flows still remain",
 243                         dev->data->port_id);
 244         ret = mlx5_mr_verify(dev);
 245         if (ret)
 246                 DRV_LOG(WARNING, "port %u some memory region still remain",
 247                         dev->data->port_id);
 248         memset(priv, 0, sizeof(*priv));
 249 }
 250
 251 const struct eth_dev_ops mlx5_dev_ops = {
 252         .dev_configure = mlx5_dev_configure,
 253         .dev_start = mlx5_dev_start,
 254         .dev_stop = mlx5_dev_stop,
 255         .dev_set_link_down = mlx5_set_link_down,
 256         .dev_set_link_up = mlx5_set_link_up,
 257         .dev_close = mlx5_dev_close,
 258         .promiscuous_enable = mlx5_promiscuous_enable,
 259         .promiscuous_disable = mlx5_promiscuous_disable,
 260         .allmulticast_enable = mlx5_allmulticast_enable,
 261         .allmulticast_disable = mlx5_allmulticast_disable,
 262         .link_update = mlx5_link_update,
 263         .stats_get = mlx5_stats_get,
 264         .stats_reset = mlx5_stats_reset,
 265         .xstats_get = mlx5_xstats_get,
 266         .xstats_reset = mlx5_xstats_reset,
 267         .xstats_get_names = mlx5_xstats_get_names,
 268         .dev_infos_get = mlx5_dev_infos_get,
 269         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 270         .vlan_filter_set = mlx5_vlan_filter_set,
 271         .rx_queue_setup = mlx5_rx_queue_setup,
 272         .tx_queue_setup = mlx5_tx_queue_setup,
 273         .rx_queue_release = mlx5_rx_queue_release,
 274         .tx_queue_release = mlx5_tx_queue_release,
 275         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 276         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 277         .mac_addr_remove = mlx5_mac_addr_remove,
 278         .mac_addr_add = mlx5_mac_addr_add,
 279         .mac_addr_set = mlx5_mac_addr_set,
 280         .set_mc_addr_list = mlx5_set_mc_addr_list,
 281         .mtu_set = mlx5_dev_set_mtu,
 282         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 283         .vlan_offload_set = mlx5_vlan_offload_set,
 284         .reta_update = mlx5_dev_rss_reta_update,
 285         .reta_query = mlx5_dev_rss_reta_query,
 286         .rss_hash_update = mlx5_rss_hash_update,
 287         .rss_hash_conf_get = mlx5_rss_hash_conf_get,
 288         .filter_ctrl = mlx5_dev_filter_ctrl,
 289         .rx_descriptor_status = mlx5_rx_descriptor_status,
 290         .tx_descriptor_status = mlx5_tx_descriptor_status,
 291         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 292         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 293         .is_removed = mlx5_is_removed,
 294 };
 295
 296 static const struct eth_dev_ops mlx5_dev_sec_ops = {
 297         .stats_get = mlx5_stats_get,
 298         .stats_reset = mlx5_stats_reset,
 299         .xstats_get = mlx5_xstats_get,
 300         .xstats_reset = mlx5_xstats_reset,
 301         .xstats_get_names = mlx5_xstats_get_names,
 302         .dev_infos_get = mlx5_dev_infos_get,
 303         .rx_descriptor_status = mlx5_rx_descriptor_status,
 304         .tx_descriptor_status = mlx5_tx_descriptor_status,
 305 };
 306
 307 /* Available operators in flow isolated mode. */
 308 const struct eth_dev_ops mlx5_dev_ops_isolate = {
 309         .dev_configure = mlx5_dev_configure,
 310         .dev_start = mlx5_dev_start,
 311         .dev_stop = mlx5_dev_stop,
 312         .dev_set_link_down = mlx5_set_link_down,
 313         .dev_set_link_up = mlx5_set_link_up,
 314         .dev_close = mlx5_dev_close,
 315         .link_update = mlx5_link_update,
 316         .stats_get = mlx5_stats_get,
 317         .stats_reset = mlx5_stats_reset,
 318         .xstats_get = mlx5_xstats_get,
 319         .xstats_reset = mlx5_xstats_reset,
 320         .xstats_get_names = mlx5_xstats_get_names,
 321         .dev_infos_get = mlx5_dev_infos_get,
 322         .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
 323         .vlan_filter_set = mlx5_vlan_filter_set,
 324         .rx_queue_setup = mlx5_rx_queue_setup,
 325         .tx_queue_setup = mlx5_tx_queue_setup,
 326         .rx_queue_release = mlx5_rx_queue_release,
 327         .tx_queue_release = mlx5_tx_queue_release,
 328         .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
 329         .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
 330         .mac_addr_remove = mlx5_mac_addr_remove,
 331         .mac_addr_add = mlx5_mac_addr_add,
 332         .mac_addr_set = mlx5_mac_addr_set,
 333         .set_mc_addr_list = mlx5_set_mc_addr_list,
 334         .mtu_set = mlx5_dev_set_mtu,
 335         .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
 336         .vlan_offload_set = mlx5_vlan_offload_set,
 337         .filter_ctrl = mlx5_dev_filter_ctrl,
 338         .rx_descriptor_status = mlx5_rx_descriptor_status,
 339         .tx_descriptor_status = mlx5_tx_descriptor_status,
 340         .rx_queue_intr_enable = mlx5_rx_intr_enable,
 341         .rx_queue_intr_disable = mlx5_rx_intr_disable,
 342         .is_removed = mlx5_is_removed,
 343 };
 344
 345 static struct {
 346         struct rte_pci_addr pci_addr; /* associated PCI address */
 347         uint32_t ports; /* physical ports bitfield. */
 348 } mlx5_dev[32];
 349
 350 /**
 351  * Get device index in mlx5_dev[] from PCI bus address.
 352  *
 353  * @param[in] pci_addr
 354  *   PCI bus address to look for.
 355  *
 356  * @return
 357  *   mlx5_dev[] index on success, -1 on failure.
 358  */
 359 static int
 360 mlx5_dev_idx(struct rte_pci_addr *pci_addr)
 361 {
 362         unsigned int i;
 363         int ret = -1;
 364
 365         assert(pci_addr != NULL);
 366         for (i = 0; (i != RTE_DIM(mlx5_dev)); ++i) {
 367                 if ((mlx5_dev[i].pci_addr.domain == pci_addr->domain) &&
 368                     (mlx5_dev[i].pci_addr.bus == pci_addr->bus) &&
 369                     (mlx5_dev[i].pci_addr.devid == pci_addr->devid) &&
 370                     (mlx5_dev[i].pci_addr.function == pci_addr->function))
 371                         return i;
 372                 if ((mlx5_dev[i].ports == 0) && (ret == -1))
 373                         ret = i;
 374         }
 375         return ret;
 376 }
 377
 378 /**
 379  * Verify and store value for device argument.
 380  *
 381  * @param[in] key
 382  *   Key argument to verify.
 383  * @param[in] val
 384  *   Value associated with key.
 385  * @param opaque
 386  *   User data.
 387  *
 388  * @return
 389  *   0 on success, a negative errno value otherwise and rte_errno is set.
 390  */
 391 static int
 392 mlx5_args_check(const char *key, const char *val, void *opaque)
 393 {
 394         struct mlx5_dev_config *config = opaque;
 395         unsigned long tmp;
 396
 397         errno = 0;
 398         tmp = strtoul(val, NULL, 0);
 399         if (errno) {
 400                 rte_errno = errno;
 401                 DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
 402                 return -rte_errno;
 403         }
 404         if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
 405                 config->cqe_comp = !!tmp;
 406         } else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
 407                 config->txq_inline = tmp;
 408         } else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
 409                 config->txqs_inline = tmp;
 410         } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
 411                 config->mps = !!tmp ? config->mps : 0;
 412         } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
 413                 config->mpw_hdr_dseg = !!tmp;
 414         } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
 415                 config->inline_max_packet_sz = tmp;
 416         } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
 417                 config->tx_vec_en = !!tmp;
 418         } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
 419                 config->rx_vec_en = !!tmp;
 420         } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
 421                 config->vf_nl_en = !!tmp;
 422         } else {
 423                 DRV_LOG(WARNING, "%s: unknown parameter", key);
 424                 rte_errno = EINVAL;
 425                 return -rte_errno;
 426         }
 427         return 0;
 428 }
 429
 430 /**
 431  * Parse device parameters.
 432  *
 433  * @param config
 434  *   Pointer to device configuration structure.
 435  * @param devargs
 436  *   Device arguments structure.
 437  *
 438  * @return
 439  *   0 on success, a negative errno value otherwise and rte_errno is set.
 440  */
 441 static int
 442 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
 443 {
 444         const char **params = (const char *[]){
 445                 MLX5_RXQ_CQE_COMP_EN,
 446                 MLX5_TXQ_INLINE,
 447                 MLX5_TXQS_MIN_INLINE,
 448                 MLX5_TXQ_MPW_EN,
 449                 MLX5_TXQ_MPW_HDR_DSEG_EN,
 450                 MLX5_TXQ_MAX_INLINE_LEN,
 451                 MLX5_TX_VEC_EN,
 452                 MLX5_RX_VEC_EN,
 453                 MLX5_VF_NL_EN,
 454                 NULL,
 455         };
 456         struct rte_kvargs *kvlist;
 457         int ret = 0;
 458         int i;
 459
 460         if (devargs == NULL)
 461                 return 0;
 462         /* Following UGLY cast is done to pass checkpatch. */
 463         kvlist = rte_kvargs_parse(devargs->args, params);
 464         if (kvlist == NULL)
 465                 return 0;
 466         /* Process parameters. */
 467         for (i = 0; (params[i] != NULL); ++i) {
 468                 if (rte_kvargs_count(kvlist, params[i])) {
 469                         ret = rte_kvargs_process(kvlist, params[i],
 470                                                  mlx5_args_check, config);
 471                         if (ret) {
 472                                 rte_errno = EINVAL;
 473                                 rte_kvargs_free(kvlist);
 474                                 return -rte_errno;
 475                         }
 476                 }
 477         }
 478         rte_kvargs_free(kvlist);
 479         return 0;
 480 }
 481
 482 static struct rte_pci_driver mlx5_driver;
 483
 484 /*
 485  * Reserved UAR address space for TXQ UAR(hw doorbell) mapping, process
 486  * local resource used by both primary and secondary to avoid duplicate
 487  * reservation.
 488  * The space has to be available on both primary and secondary process,
 489  * TXQ UAR maps to this area using fixed mmap w/o double check.
 490  */
 491 static void *uar_base;
 492
 493 static int
 494 find_lower_va_bound(const struct rte_memseg_list *msl __rte_unused,
 495                 const struct rte_memseg *ms, void *arg)
 496 {
 497         void **addr = arg;
 498
 499         if (*addr == NULL)
 500                 *addr = ms->addr;
 501         else
 502                 *addr = RTE_MIN(*addr, ms->addr);
 503
 504         return 0;
 505 }
 506
 507 /**
 508  * Reserve UAR address space for primary process.
 509  *
 510  * @param[in] dev
 511  *   Pointer to Ethernet device.
 512  *
 513  * @return
 514  *   0 on success, a negative errno value otherwise and rte_errno is set.
 515  */
 516 static int
 517 mlx5_uar_init_primary(struct rte_eth_dev *dev)
 518 {
 519         struct priv *priv = dev->data->dev_private;
 520         void *addr = (void *)0;
 521
 522         if (uar_base) { /* UAR address space mapped. */
 523                 priv->uar_base = uar_base;
 524                 return 0;
 525         }
 526         /* find out lower bound of hugepage segments */
 527         rte_memseg_walk(find_lower_va_bound, &addr);
 528
 529         /* keep distance to hugepages to minimize potential conflicts. */
 530         addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
 531         /* anonymous mmap, no real memory consumption. */
 532         addr = mmap(addr, MLX5_UAR_SIZE,
 533                     PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 534         if (addr == MAP_FAILED) {
 535                 DRV_LOG(ERR,
 536                         "port %u failed to reserve UAR address space, please"
 537                         " adjust MLX5_UAR_SIZE or try --base-virtaddr",
 538                         dev->data->port_id);
 539                 rte_errno = ENOMEM;
 540                 return -rte_errno;
 541         }
 542         /* Accept either same addr or a new addr returned from mmap if target
 543          * range occupied.
 544          */
 545         DRV_LOG(INFO, "port %u reserved UAR address space: %p",
 546                 dev->data->port_id, addr);
 547         priv->uar_base = addr; /* for primary and secondary UAR re-mmap. */
 548         uar_base = addr; /* process local, don't reserve again. */
 549         return 0;
 550 }
 551
 552 /**
 553  * Reserve UAR address space for secondary process, align with
 554  * primary process.
 555  *
 556  * @param[in] dev
 557  *   Pointer to Ethernet device.
 558  *
 559  * @return
 560  *   0 on success, a negative errno value otherwise and rte_errno is set.
 561  */
 562 static int
 563 mlx5_uar_init_secondary(struct rte_eth_dev *dev)
 564 {
 565         struct priv *priv = dev->data->dev_private;
 566         void *addr;
 567
 568         assert(priv->uar_base);
 569         if (uar_base) { /* already reserved. */
 570                 assert(uar_base == priv->uar_base);
 571                 return 0;
 572         }
 573         /* anonymous mmap, no real memory consumption. */
 574         addr = mmap(priv->uar_base, MLX5_UAR_SIZE,
 575                     PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 576         if (addr == MAP_FAILED) {
 577                 DRV_LOG(ERR, "port %u UAR mmap failed: %p size: %llu",
 578                         dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
 579                 rte_errno = ENXIO;
 580                 return -rte_errno;
 581         }
 582         if (priv->uar_base != addr) {
 583                 DRV_LOG(ERR,
 584                         "port %u UAR address %p size %llu occupied, please"
 585                         " adjust MLX5_UAR_OFFSET or try EAL parameter"
 586                         " --base-virtaddr",
 587                         dev->data->port_id, priv->uar_base, MLX5_UAR_SIZE);
 588                 rte_errno = ENXIO;
 589                 return -rte_errno;
 590         }
 591         uar_base = addr; /* process local, don't reserve again */
 592         DRV_LOG(INFO, "port %u reserved UAR address space: %p",
 593                 dev->data->port_id, addr);
 594         return 0;
 595 }
 596
 597 /**
 598  * DPDK callback to register a PCI device.
 599  *
 600  * This function creates an Ethernet device for each port of a given
 601  * PCI device.
 602  *
 603  * @param[in] pci_drv
 604  *   PCI driver structure (mlx5_driver).
 605  * @param[in] pci_dev
 606  *   PCI device information.
 607  *
 608  * @return
 609  *   0 on success, a negative errno value otherwise and rte_errno is set.
 610  */
 611 static int
 612 mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 613                struct rte_pci_device *pci_dev)
 614 {
 615         struct ibv_device **list = NULL;
 616         struct ibv_device *ibv_dev;
 617         int err = 0;
 618         struct ibv_context *attr_ctx = NULL;
 619         struct ibv_device_attr_ex device_attr;
 620         unsigned int vf;
 621         unsigned int mps;
 622         unsigned int cqe_comp;
 623         unsigned int tunnel_en = 0;
 624         unsigned int swp = 0;
 625         int idx;
 626         int i;
 627         struct mlx5dv_context attrs_out = {0};
 628 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 629         struct ibv_counter_set_description cs_desc;
 630 #endif
 631
 632         assert(pci_drv == &mlx5_driver);
 633         /* Get mlx5_dev[] index. */
 634         idx = mlx5_dev_idx(&pci_dev->addr);
 635         if (idx == -1) {
 636                 DRV_LOG(ERR, "this driver cannot support any more adapters");
 637                 err = ENOMEM;
 638                 goto error;
 639         }
 640         DRV_LOG(DEBUG, "using driver device index %d", idx);
 641         /* Save PCI address. */
 642         mlx5_dev[idx].pci_addr = pci_dev->addr;
 643         list = mlx5_glue->get_device_list(&i);
 644         if (list == NULL) {
 645                 assert(errno);
 646                 err = errno;
 647                 if (errno == ENOSYS)
 648                         DRV_LOG(ERR,
 649                                 "cannot list devices, is ib_uverbs loaded?");
 650                 goto error;
 651         }
 652         assert(i >= 0);
 653         /*
 654          * For each listed device, check related sysfs entry against
 655          * the provided PCI ID.
 656          */
 657         while (i != 0) {
 658                 struct rte_pci_addr pci_addr;
 659
 660                 --i;
 661                 DRV_LOG(DEBUG, "checking device \"%s\"", list[i]->name);
 662                 if (mlx5_ibv_device_to_pci_addr(list[i], &pci_addr))
 663                         continue;
 664                 if ((pci_dev->addr.domain != pci_addr.domain) ||
 665                     (pci_dev->addr.bus != pci_addr.bus) ||
 666                     (pci_dev->addr.devid != pci_addr.devid) ||
 667                     (pci_dev->addr.function != pci_addr.function))
 668                         continue;
 669                 DRV_LOG(INFO, "PCI information matches, using device \"%s\"",
 670                         list[i]->name);
 671                 vf = ((pci_dev->id.device_id ==
 672                        PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
 673                       (pci_dev->id.device_id ==
 674                        PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF) ||
 675                       (pci_dev->id.device_id ==
 676                        PCI_DEVICE_ID_MELLANOX_CONNECTX5VF) ||
 677                       (pci_dev->id.device_id ==
 678                        PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF));
 679                 attr_ctx = mlx5_glue->open_device(list[i]);
 680                 rte_errno = errno;
 681                 err = rte_errno;
 682                 break;
 683         }
 684         if (attr_ctx == NULL) {
 685                 mlx5_glue->free_device_list(list);
 686                 switch (err) {
 687                 case 0:
 688                         DRV_LOG(ERR,
 689                                 "cannot access device, is mlx5_ib loaded?");
 690                         err = ENODEV;
 691                         goto error;
 692                 case EINVAL:
 693                         DRV_LOG(ERR,
 694                                 "cannot use device, are drivers up to date?");
 695                         goto error;
 696                 }
 697         }
 698         ibv_dev = list[i];
 699         DRV_LOG(DEBUG, "device opened");
 700 #ifdef HAVE_IBV_MLX5_MOD_SWP
 701         attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
 702 #endif
 703         /*
 704          * Multi-packet send is supported by ConnectX-4 Lx PF as well
 705          * as all ConnectX-5 devices.
 706          */
 707 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 708         attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS;
 709 #endif
 710         mlx5_glue->dv_query_device(attr_ctx, &attrs_out);
 711         if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
 712                 if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
 713                         DRV_LOG(DEBUG, "enhanced MPW is supported");
 714                         mps = MLX5_MPW_ENHANCED;
 715                 } else {
 716                         DRV_LOG(DEBUG, "MPW is supported");
 717                         mps = MLX5_MPW;
 718                 }
 719         } else {
 720                 DRV_LOG(DEBUG, "MPW isn't supported");
 721                 mps = MLX5_MPW_DISABLED;
 722         }
 723 #ifdef HAVE_IBV_MLX5_MOD_SWP
 724         if (attrs_out.comp_mask | MLX5DV_CONTEXT_MASK_SWP)
 725                 swp = attrs_out.sw_parsing_caps.sw_parsing_offloads;
 726         DRV_LOG(DEBUG, "SWP support: %u", swp);
 727 #endif
 728         if (RTE_CACHE_LINE_SIZE == 128 &&
 729             !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
 730                 cqe_comp = 0;
 731         else
 732                 cqe_comp = 1;
 733 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 734         if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_TUNNEL_OFFLOADS) {
 735                 tunnel_en = ((attrs_out.tunnel_offloads_caps &
 736                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_VXLAN) &&
 737                              (attrs_out.tunnel_offloads_caps &
 738                               MLX5DV_RAW_PACKET_CAP_TUNNELED_OFFLOAD_GRE));
 739         }
 740         DRV_LOG(DEBUG, "tunnel offloading is %ssupported",
 741                 tunnel_en ? "" : "not ");
 742 #else
 743         DRV_LOG(WARNING,
 744                 "tunnel offloading disabled due to old OFED/rdma-core version");
 745 #endif
 746         if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr)) {
 747                 err = errno;
 748                 goto error;
 749         }
 750         DRV_LOG(INFO, "%u port(s) detected",
 751                 device_attr.orig_attr.phys_port_cnt);
 752         for (i = 0; i < device_attr.orig_attr.phys_port_cnt; i++) {
 753                 char name[RTE_ETH_NAME_MAX_LEN];
 754                 int len;
 755                 uint32_t port = i + 1; /* ports are indexed from one */
 756                 uint32_t test = (1 << i);
 757                 struct ibv_context *ctx = NULL;
 758                 struct ibv_port_attr port_attr;
 759                 struct ibv_pd *pd = NULL;
 760                 struct priv *priv = NULL;
 761                 struct rte_eth_dev *eth_dev = NULL;
 762                 struct ibv_device_attr_ex device_attr_ex;
 763                 struct ether_addr mac;
 764                 struct mlx5_dev_config config = {
 765                         .cqe_comp = cqe_comp,
 766                         .mps = mps,
 767                         .tunnel_en = tunnel_en,
 768                         .tx_vec_en = 1,
 769                         .rx_vec_en = 1,
 770                         .mpw_hdr_dseg = 0,
 771                         .txq_inline = MLX5_ARG_UNSET,
 772                         .txqs_inline = MLX5_ARG_UNSET,
 773                         .inline_max_packet_sz = MLX5_ARG_UNSET,
 774                         .vf_nl_en = 1,
 775                         .swp = !!swp,
 776                 };
 777
 778                 len = snprintf(name, sizeof(name), PCI_PRI_FMT,
 779                          pci_dev->addr.domain, pci_dev->addr.bus,
 780                          pci_dev->addr.devid, pci_dev->addr.function);
 781                 if (device_attr.orig_attr.phys_port_cnt > 1)
 782                         snprintf(name + len, sizeof(name), " port %u", i);
 783                 mlx5_dev[idx].ports |= test;
 784                 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
 785                         eth_dev = rte_eth_dev_attach_secondary(name);
 786                         if (eth_dev == NULL) {
 787                                 DRV_LOG(ERR, "can not attach rte ethdev");
 788                                 rte_errno = ENOMEM;
 789                                 err = rte_errno;
 790                                 goto error;
 791                         }
 792                         eth_dev->device = &pci_dev->device;
 793                         eth_dev->dev_ops = &mlx5_dev_sec_ops;
 794                         err = mlx5_uar_init_secondary(eth_dev);
 795                         if (err)
 796                                 goto error;
 797                         /* Receive command fd from primary process */
 798                         err = mlx5_socket_connect(eth_dev);
 799                         if (err)
 800                                 goto error;
 801                         /* Remap UAR for Tx queues. */
 802                         err = mlx5_tx_uar_remap(eth_dev, err);
 803                         if (err)
 804                                 goto error;
 805                         /*
 806                          * Ethdev pointer is still required as input since
 807                          * the primary device is not accessible from the
 808                          * secondary process.
 809                          */
 810                         eth_dev->rx_pkt_burst =
 811                                 mlx5_select_rx_function(eth_dev);
 812                         eth_dev->tx_pkt_burst =
 813                                 mlx5_select_tx_function(eth_dev);
 814                         continue;
 815                 }
 816                 DRV_LOG(DEBUG, "using port %u (%08" PRIx32 ")", port, test);
 817                 ctx = mlx5_glue->open_device(ibv_dev);
 818                 if (ctx == NULL) {
 819                         err = ENODEV;
 820                         goto port_error;
 821                 }
 822                 /* Check port status. */
 823                 err = mlx5_glue->query_port(ctx, port, &port_attr);
 824                 if (err) {
 825                         DRV_LOG(ERR, "port query failed: %s", strerror(err));
 826                         goto port_error;
 827                 }
 828                 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
 829                         DRV_LOG(ERR,
 830                                 "port %d is not configured in Ethernet mode",
 831                                 port);
 832                         err = EINVAL;
 833                         goto port_error;
 834                 }
 835                 if (port_attr.state != IBV_PORT_ACTIVE)
 836                         DRV_LOG(DEBUG, "port %d is not active: \"%s\" (%d)",
 837                                 port,
 838                                 mlx5_glue->port_state_str(port_attr.state),
 839                                 port_attr.state);
 840                 /* Allocate protection domain. */
 841                 pd = mlx5_glue->alloc_pd(ctx);
 842                 if (pd == NULL) {
 843                         DRV_LOG(ERR, "PD allocation failure");
 844                         err = ENOMEM;
 845                         goto port_error;
 846                 }
 847                 mlx5_dev[idx].ports |= test;
 848                 /* from rte_ethdev.c */
 849                 priv = rte_zmalloc("ethdev private structure",
 850                                    sizeof(*priv),
 851                                    RTE_CACHE_LINE_SIZE);
 852                 if (priv == NULL) {
 853                         DRV_LOG(ERR, "priv allocation failure");
 854                         err = ENOMEM;
 855                         goto port_error;
 856                 }
 857                 priv->ctx = ctx;
 858                 strncpy(priv->ibdev_path, priv->ctx->device->ibdev_path,
 859                         sizeof(priv->ibdev_path));
 860                 priv->device_attr = device_attr;
 861                 priv->port = port;
 862                 priv->pd = pd;
 863                 priv->mtu = ETHER_MTU;
 864                 err = mlx5_args(&config, pci_dev->device.devargs);
 865                 if (err) {
 866                         DRV_LOG(ERR, "failed to process device arguments: %s",
 867                                 strerror(err));
 868                         goto port_error;
 869                 }
 870                 if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
 871                         DRV_LOG(ERR, "ibv_query_device_ex() failed");
 872                         err = errno;
 873                         goto port_error;
 874                 }
 875                 config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
 876                                     IBV_DEVICE_RAW_IP_CSUM);
 877                 DRV_LOG(DEBUG, "checksum offloading is %ssupported",
 878                         (config.hw_csum ? "" : "not "));
 879 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 880                 config.flow_counter_en = !!(device_attr.max_counter_sets);
 881                 mlx5_glue->describe_counter_set(ctx, 0, &cs_desc);
 882                 DRV_LOG(DEBUG,
 883                         "counter type = %d, num of cs = %ld, attributes = %d",
 884                         cs_desc.counter_type, cs_desc.num_of_cs,
 885                         cs_desc.attributes);
 886 #endif
 887                 config.ind_table_max_size =
 888                         device_attr_ex.rss_caps.max_rwq_indirection_table_size;
 889                 /* Remove this check once DPDK supports larger/variable
 890                  * indirection tables. */
 891                 if (config.ind_table_max_size >
 892                                 (unsigned int)ETH_RSS_RETA_SIZE_512)
 893                         config.ind_table_max_size = ETH_RSS_RETA_SIZE_512;
 894                 DRV_LOG(DEBUG, "maximum Rx indirection table size is %u",
 895                         config.ind_table_max_size);
 896                 config.hw_vlan_strip = !!(device_attr_ex.raw_packet_caps &
 897                                          IBV_RAW_PACKET_CAP_CVLAN_STRIPPING);
 898                 DRV_LOG(DEBUG, "VLAN stripping is %ssupported",
 899                         (config.hw_vlan_strip ? "" : "not "));
 900
 901                 config.hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
 902                                          IBV_RAW_PACKET_CAP_SCATTER_FCS);
 903                 DRV_LOG(DEBUG, "FCS stripping configuration is %ssupported",
 904                         (config.hw_fcs_strip ? "" : "not "));
 905
 906 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
 907                 config.hw_padding = !!device_attr_ex.rx_pad_end_addr_align;
 908 #endif
 909                 DRV_LOG(DEBUG,
 910                         "hardware Rx end alignment padding is %ssupported",
 911                         (config.hw_padding ? "" : "not "));
 912                 config.vf = vf;
 913                 config.tso = ((device_attr_ex.tso_caps.max_tso > 0) &&
 914                               (device_attr_ex.tso_caps.supported_qpts &
 915                               (1 << IBV_QPT_RAW_PACKET)));
 916                 if (config.tso)
 917                         config.tso_max_payload_sz =
 918                                         device_attr_ex.tso_caps.max_tso;
 919                 if (config.mps && !mps) {
 920                         DRV_LOG(ERR,
 921                                 "multi-packet send not supported on this device"
 922                                 " (" MLX5_TXQ_MPW_EN ")");
 923                         err = ENOTSUP;
 924                         goto port_error;
 925                 }
 926                 DRV_LOG(INFO, "%s MPS is %s",
 927                         config.mps == MLX5_MPW_ENHANCED ? "enhanced " : "",
 928                         config.mps != MLX5_MPW_DISABLED ? "enabled" :
 929                         "disabled");
 930                 if (config.cqe_comp && !cqe_comp) {
 931                         DRV_LOG(WARNING, "Rx CQE compression isn't supported");
 932                         config.cqe_comp = 0;
 933                 }
 934                 eth_dev = rte_eth_dev_allocate(name);
 935                 if (eth_dev == NULL) {
 936                         DRV_LOG(ERR, "can not allocate rte ethdev");
 937                         err = ENOMEM;
 938                         goto port_error;
 939                 }
 940                 eth_dev->data->dev_private = priv;
 941                 priv->dev = eth_dev;
 942                 eth_dev->data->mac_addrs = priv->mac;
 943                 eth_dev->device = &pci_dev->device;
 944                 rte_eth_copy_pci_info(eth_dev, pci_dev);
 945                 eth_dev->device->driver = &mlx5_driver.driver;
 946                 err = mlx5_uar_init_primary(eth_dev);
 947                 if (err)
 948                         goto port_error;
 949                 /* Configure the first MAC address by default. */
 950                 if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
 951                         DRV_LOG(ERR,
 952                                 "port %u cannot get MAC address, is mlx5_en"
 953                                 " loaded? (errno: %s)",
 954                                 eth_dev->data->port_id, strerror(errno));
 955                         err = ENODEV;
 956                         goto port_error;
 957                 }
 958                 DRV_LOG(INFO,
 959                         "port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
 960                         eth_dev->data->port_id,
 961                         mac.addr_bytes[0], mac.addr_bytes[1],
 962                         mac.addr_bytes[2], mac.addr_bytes[3],
 963                         mac.addr_bytes[4], mac.addr_bytes[5]);
 964 #ifndef NDEBUG
 965                 {
 966                         char ifname[IF_NAMESIZE];
 967
 968                         if (mlx5_get_ifname(eth_dev, &ifname) == 0)
 969                                 DRV_LOG(DEBUG, "port %u ifname is \"%s\"",
 970                                         eth_dev->data->port_id, ifname);
 971                         else
 972                                 DRV_LOG(DEBUG, "port %u ifname is unknown",
 973                                         eth_dev->data->port_id);
 974                 }
 975 #endif
 976                 /* Get actual MTU if possible. */
 977                 err = mlx5_get_mtu(eth_dev, &priv->mtu);
 978                 if (err)
 979                         goto port_error;
 980                 DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
 981                         priv->mtu);
 982                 /*
 983                  * Initialize burst functions to prevent crashes before link-up.
 984                  */
 985                 eth_dev->rx_pkt_burst = removed_rx_burst;
 986                 eth_dev->tx_pkt_burst = removed_tx_burst;
 987                 eth_dev->dev_ops = &mlx5_dev_ops;
 988                 /* Register MAC address. */
 989                 claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
 990                 priv->nl_socket = -1;
 991                 priv->nl_sn = 0;
 992                 if (vf && config.vf_nl_en) {
 993                         priv->nl_socket = mlx5_nl_init(RTMGRP_LINK);
 994                         if (priv->nl_socket < 0)
 995                                 priv->nl_socket = -1;
 996                         mlx5_nl_mac_addr_sync(eth_dev);
 997                 }
 998                 TAILQ_INIT(&priv->flows);
 999                 TAILQ_INIT(&priv->ctrl_flows);
1000                 /* Hint libmlx5 to use PMD allocator for data plane resources */
1001                 struct mlx5dv_ctx_allocators alctr = {
1002                         .alloc = &mlx5_alloc_verbs_buf,
1003                         .free = &mlx5_free_verbs_buf,
1004                         .data = priv,
1005                 };
1006                 mlx5_glue->dv_set_context_attr(ctx,
1007                                                MLX5DV_CTX_ATTR_BUF_ALLOCATORS,
1008                                                (void *)((uintptr_t)&alctr));
1009                 /* Bring Ethernet device up. */
1010                 DRV_LOG(DEBUG, "port %u forcing Ethernet interface up",
1011                         eth_dev->data->port_id);
1012                 mlx5_set_link_up(eth_dev);
1013                 /*
1014                  * Even though the interrupt handler is not installed yet,
1015                  * interrupts will still trigger on the asyn_fd from
1016                  * Verbs context returned by ibv_open_device().
1017                  */
1018                 mlx5_link_update(eth_dev, 0);
1019                 /* Store device configuration on private structure. */
1020                 priv->config = config;
1021                 continue;
1022 port_error:
1023                 if (priv)
1024                         rte_free(priv);
1025                 if (pd)
1026                         claim_zero(mlx5_glue->dealloc_pd(pd));
1027                 if (ctx)
1028                         claim_zero(mlx5_glue->close_device(ctx));
1029                 break;
1030         }
1031         /*
1032          * XXX if something went wrong in the loop above, there is a resource
1033          * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
1034          * long as the dpdk does not provide a way to deallocate a ethdev and a
1035          * way to enumerate the registered ethdevs to free the previous ones.
1036          */
1037         /* no port found, complain */
1038         if (!mlx5_dev[idx].ports) {
1039                 rte_errno = ENODEV;
1040                 err = rte_errno;
1041         }
1042 error:
1043         if (attr_ctx)
1044                 claim_zero(mlx5_glue->close_device(attr_ctx));
1045         if (list)
1046                 mlx5_glue->free_device_list(list);
1047         if (err) {
1048                 rte_errno = err;
1049                 return -rte_errno;
1050         }
1051         return 0;
1052 }
1053
1054 static const struct rte_pci_id mlx5_pci_id_map[] = {
1055         {
1056                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1057                                PCI_DEVICE_ID_MELLANOX_CONNECTX4)
1058         },
1059         {
1060                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1061                                PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
1062         },
1063         {
1064                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1065                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
1066         },
1067         {
1068                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1069                                PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
1070         },
1071         {
1072                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1073                                PCI_DEVICE_ID_MELLANOX_CONNECTX5)
1074         },
1075         {
1076                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1077                                PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
1078         },
1079         {
1080                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1081                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
1082         },
1083         {
1084                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
1085                                PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
1086         },
1087         {
1088                 .vendor_id = 0
1089         }
1090 };
1091
1092 static struct rte_pci_driver mlx5_driver = {
1093         .driver = {
1094                 .name = MLX5_DRIVER_NAME
1095         },
1096         .id_table = mlx5_pci_id_map,
1097         .probe = mlx5_pci_probe,
1098         .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV,
1099 };
1100
1101 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1102
1103 /**
1104  * Suffix RTE_EAL_PMD_PATH with "-glue".
1105  *
1106  * This function performs a sanity check on RTE_EAL_PMD_PATH before
1107  * suffixing its last component.
1108  *
1109  * @param buf[out]
1110  *   Output buffer, should be large enough otherwise NULL is returned.
1111  * @param size
1112  *   Size of @p out.
1113  *
1114  * @return
1115  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
1116  */
1117 static char *
1118 mlx5_glue_path(char *buf, size_t size)
1119 {
1120         static const char *const bad[] = { "/", ".", "..", NULL };
1121         const char *path = RTE_EAL_PMD_PATH;
1122         size_t len = strlen(path);
1123         size_t off;
1124         int i;
1125
1126         while (len && path[len - 1] == '/')
1127                 --len;
1128         for (off = len; off && path[off - 1] != '/'; --off)
1129                 ;
1130         for (i = 0; bad[i]; ++i)
1131                 if (!strncmp(path + off, bad[i], (int)(len - off)))
1132                         goto error;
1133         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
1134         if (i == -1 || (size_t)i >= size)
1135                 goto error;
1136         return buf;
1137 error:
1138         DRV_LOG(ERR,
1139                 "unable to append \"-glue\" to last component of"
1140                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"),"
1141                 " please re-configure DPDK");
1142         return NULL;
1143 }
1144
1145 /**
1146  * Initialization routine for run-time dependency on rdma-core.
1147  */
1148 static int
1149 mlx5_glue_init(void)
1150 {
1151         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
1152         const char *path[] = {
1153                 /*
1154                  * A basic security check is necessary before trusting
1155                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
1156                  */
1157                 (geteuid() == getuid() && getegid() == getgid() ?
1158                  getenv("MLX5_GLUE_PATH") : NULL),
1159                 /*
1160                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
1161                  * variant, otherwise let dlopen() look up libraries on its
1162                  * own.
1163                  */
1164                 (*RTE_EAL_PMD_PATH ?
1165                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
1166         };
1167         unsigned int i = 0;
1168         void *handle = NULL;
1169         void **sym;
1170         const char *dlmsg;
1171
1172         while (!handle && i != RTE_DIM(path)) {
1173                 const char *end;
1174                 size_t len;
1175                 int ret;
1176
1177                 if (!path[i]) {
1178                         ++i;
1179                         continue;
1180                 }
1181                 end = strpbrk(path[i], ":;");
1182                 if (!end)
1183                         end = path[i] + strlen(path[i]);
1184                 len = end - path[i];
1185                 ret = 0;
1186                 do {
1187                         char name[ret + 1];
1188
1189                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
1190                                        (int)len, path[i],
1191                                        (!len || *(end - 1) == '/') ? "" : "/");
1192                         if (ret == -1)
1193                                 break;
1194                         if (sizeof(name) != (size_t)ret + 1)
1195                                 continue;
1196                         DRV_LOG(DEBUG, "looking for rdma-core glue as \"%s\"",
1197                                 name);
1198                         handle = dlopen(name, RTLD_LAZY);
1199                         break;
1200                 } while (1);
1201                 path[i] = end + 1;
1202                 if (!*end)
1203                         ++i;
1204         }
1205         if (!handle) {
1206                 rte_errno = EINVAL;
1207                 dlmsg = dlerror();
1208                 if (dlmsg)
1209                         DRV_LOG(WARNING, "cannot load glue library: %s", dlmsg);
1210                 goto glue_error;
1211         }
1212         sym = dlsym(handle, "mlx5_glue");
1213         if (!sym || !*sym) {
1214                 rte_errno = EINVAL;
1215                 dlmsg = dlerror();
1216                 if (dlmsg)
1217                         DRV_LOG(ERR, "cannot resolve glue symbol: %s", dlmsg);
1218                 goto glue_error;
1219         }
1220         mlx5_glue = *sym;
1221         return 0;
1222 glue_error:
1223         if (handle)
1224                 dlclose(handle);
1225         DRV_LOG(WARNING,
1226                 "cannot initialize PMD due to missing run-time dependency on"
1227                 " rdma-core libraries (libibverbs, libmlx5)");
1228         return -rte_errno;
1229 }
1230
1231 #endif
1232
1233 /**
1234  * Driver initialization routine.
1235  */
1236 RTE_INIT(rte_mlx5_pmd_init);
1237 static void
1238 rte_mlx5_pmd_init(void)
1239 {
1240         /* Build the static tables for Verbs conversion. */
1241         mlx5_set_ptype_table();
1242         mlx5_set_cksum_table();
1243         mlx5_set_swp_types_table();
1244         /*
1245          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
1246          * huge pages. Calling ibv_fork_init() during init allows
1247          * applications to use fork() safely for purposes other than
1248          * using this PMD, which is not supported in forked processes.
1249          */
1250         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
1251         /* Match the size of Rx completion entry to the size of a cacheline. */
1252         if (RTE_CACHE_LINE_SIZE == 128)
1253                 setenv("MLX5_CQE_SIZE", "128", 0);
1254 #ifdef RTE_LIBRTE_MLX5_DLOPEN_DEPS
1255         if (mlx5_glue_init())
1256                 return;
1257         assert(mlx5_glue);
1258 #endif
1259 #ifndef NDEBUG
1260         /* Glue structure must not contain any NULL pointers. */
1261         {
1262                 unsigned int i;
1263
1264                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
1265                         assert(((const void *const *)mlx5_glue)[i]);
1266         }
1267 #endif
1268         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
1269                 DRV_LOG(ERR,
1270                         "rdma-core glue \"%s\" mismatch: \"%s\" is required",
1271                         mlx5_glue->version, MLX5_GLUE_VERSION);
1272                 return;
1273         }
1274         mlx5_glue->fork_init();
1275         rte_pci_register(&mlx5_driver);
1276 }
1277
1278 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
1279 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
1280 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
1281
1282 /** Initialize driver log type. */
1283 RTE_INIT(vdev_netvsc_init_log)
1284 {
1285         mlx5_logtype = rte_log_register("pmd.net.mlx5");
1286         if (mlx5_logtype >= 0)
1287                 rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
1288 }