drivers/common/mlx5/linux/mlx5_common_os.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2020 Mellanox Technologies, Ltd
   3  */
   4
   5 #include <unistd.h>
   6 #include <string.h>
   7 #include <stdio.h>
   8 #ifdef RTE_IBVERBS_LINK_DLOPEN
   9 #include <dlfcn.h>
  10 #endif
  11 #include <dirent.h>
  12 #include <net/if.h>
  13
  14 #include <rte_errno.h>
  15 #include <rte_string_fns.h>
  16
  17 #include "mlx5_common.h"
  18 #include "mlx5_common_log.h"
  19 #include "mlx5_common_os.h"
  20 #include "mlx5_glue.h"
  21
  22 #ifdef MLX5_GLUE
  23 const struct mlx5_glue *mlx5_glue;
  24 #endif
  25
  26 /**
  27  * Get PCI information by sysfs device path.
  28  *
  29  * @param dev_path
  30  *   Pointer to device sysfs folder name.
  31  * @param[out] pci_addr
  32  *   PCI bus address output buffer.
  33  *
  34  * @return
  35  *   0 on success, a negative errno value otherwise and rte_errno is set.
  36  */
  37 int
  38 mlx5_dev_to_pci_addr(const char *dev_path,
  39                      struct rte_pci_addr *pci_addr)
  40 {
  41         FILE *file;
  42         char line[32];
  43         int rc = -ENOENT;
  44         MKSTR(path, "%s/device/uevent", dev_path);
  45
  46         file = fopen(path, "rb");
  47         if (file == NULL) {
  48                 rte_errno = errno;
  49                 return -rte_errno;
  50         }
  51         while (fgets(line, sizeof(line), file) == line) {
  52                 size_t len = strlen(line);
  53
  54                 /* Truncate long lines. */
  55                 if (len == (sizeof(line) - 1)) {
  56                         while (line[(len - 1)] != '\n') {
  57                                 int ret = fgetc(file);
  58
  59                                 if (ret == EOF)
  60                                         goto exit;
  61                                 line[(len - 1)] = ret;
  62                         }
  63                         /* No match for long lines. */
  64                         continue;
  65                 }
  66                 /* Extract information. */
  67                 if (sscanf(line,
  68                            "PCI_SLOT_NAME="
  69                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
  70                            &pci_addr->domain,
  71                            &pci_addr->bus,
  72                            &pci_addr->devid,
  73                            &pci_addr->function) == 4) {
  74                         rc = 0;
  75                         break;
  76                 }
  77         }
  78 exit:
  79         fclose(file);
  80         if (rc)
  81                 rte_errno = -rc;
  82         return rc;
  83 }
  84
  85 /**
  86  * Extract port name, as a number, from sysfs or netlink information.
  87  *
  88  * @param[in] port_name_in
  89  *   String representing the port name.
  90  * @param[out] port_info_out
  91  *   Port information, including port name as a number and port name
  92  *   type if recognized
  93  *
  94  * @return
  95  *   port_name field set according to recognized name format.
  96  */
  97 void
  98 mlx5_translate_port_name(const char *port_name_in,
  99                          struct mlx5_switch_info *port_info_out)
 100 {
 101         char ctrl = 0, pf_c1, pf_c2, vf_c1, vf_c2, eol;
 102         char *end;
 103         int sc_items;
 104
 105         sc_items = sscanf(port_name_in, "%c%d",
 106                           &ctrl, &port_info_out->ctrl_num);
 107         if (sc_items == 2 && ctrl == 'c') {
 108                 port_name_in++; /* 'c' */
 109                 port_name_in += snprintf(NULL, 0, "%d",
 110                                           port_info_out->ctrl_num);
 111         }
 112         /* Check for port-name as a string of the form pf0vf0 or pf0sf0 */
 113         sc_items = sscanf(port_name_in, "%c%c%d%c%c%d%c",
 114                           &pf_c1, &pf_c2, &port_info_out->pf_num,
 115                           &vf_c1, &vf_c2, &port_info_out->port_name, &eol);
 116         if (sc_items == 6 && pf_c1 == 'p' && pf_c2 == 'f') {
 117                 if (vf_c1 == 'v' && vf_c2 == 'f') {
 118                         /* Kernel ver >= 5.0 or OFED ver >= 4.6 */
 119                         port_info_out->name_type =
 120                                         MLX5_PHYS_PORT_NAME_TYPE_PFVF;
 121                         return;
 122                 }
 123                 if (vf_c1 == 's' && vf_c2 == 'f') {
 124                         /* Kernel ver >= 5.11 or OFED ver >= 5.1 */
 125                         port_info_out->name_type =
 126                                         MLX5_PHYS_PORT_NAME_TYPE_PFSF;
 127                         return;
 128                 }
 129         }
 130         /*
 131          * Check for port-name as a string of the form p0
 132          * (support kernel ver >= 5.0, or OFED ver >= 4.6).
 133          */
 134         sc_items = sscanf(port_name_in, "%c%d%c",
 135                           &pf_c1, &port_info_out->port_name, &eol);
 136         if (sc_items == 2 && pf_c1 == 'p') {
 137                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
 138                 return;
 139         }
 140         /*
 141          * Check for port-name as a string of the form pf0
 142          * (support kernel ver >= 5.7 for HPF representor on BF).
 143          */
 144         sc_items = sscanf(port_name_in, "%c%c%d%c",
 145                           &pf_c1, &pf_c2, &port_info_out->pf_num, &eol);
 146         if (sc_items == 3 && pf_c1 == 'p' && pf_c2 == 'f') {
 147                 port_info_out->port_name = -1;
 148                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFHPF;
 149                 return;
 150         }
 151         /* Check for port-name as a number (support kernel ver < 5.0 */
 152         errno = 0;
 153         port_info_out->port_name = strtol(port_name_in, &end, 0);
 154         if (!errno &&
 155             (size_t)(end - port_name_in) == strlen(port_name_in)) {
 156                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
 157                 return;
 158         }
 159         port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
 160 }
 161
 162 /**
 163  * Get kernel interface name from IB device path.
 164  *
 165  * @param[in] ibdev_path
 166  *   Pointer to IB device path.
 167  * @param[out] ifname
 168  *   Interface name output buffer.
 169  *
 170  * @return
 171  *   0 on success, a negative errno value otherwise and rte_errno is set.
 172  */
 173 int
 174 mlx5_get_ifname_sysfs(const char *ibdev_path, char *ifname)
 175 {
 176         DIR *dir;
 177         struct dirent *dent;
 178         unsigned int dev_type = 0;
 179         unsigned int dev_port_prev = ~0u;
 180         char match[IF_NAMESIZE] = "";
 181
 182         MLX5_ASSERT(ibdev_path);
 183         {
 184                 MKSTR(path, "%s/device/net", ibdev_path);
 185
 186                 dir = opendir(path);
 187                 if (dir == NULL) {
 188                         rte_errno = errno;
 189                         return -rte_errno;
 190                 }
 191         }
 192         while ((dent = readdir(dir)) != NULL) {
 193                 char *name = dent->d_name;
 194                 FILE *file;
 195                 unsigned int dev_port;
 196                 int r;
 197
 198                 if ((name[0] == '.') &&
 199                     ((name[1] == '\0') ||
 200                      ((name[1] == '.') && (name[2] == '\0'))))
 201                         continue;
 202
 203                 MKSTR(path, "%s/device/net/%s/%s",
 204                       ibdev_path, name,
 205                       (dev_type ? "dev_id" : "dev_port"));
 206
 207                 file = fopen(path, "rb");
 208                 if (file == NULL) {
 209                         if (errno != ENOENT)
 210                                 continue;
 211                         /*
 212                          * Switch to dev_id when dev_port does not exist as
 213                          * is the case with Linux kernel versions < 3.15.
 214                          */
 215 try_dev_id:
 216                         match[0] = '\0';
 217                         if (dev_type)
 218                                 break;
 219                         dev_type = 1;
 220                         dev_port_prev = ~0u;
 221                         rewinddir(dir);
 222                         continue;
 223                 }
 224                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 225                 fclose(file);
 226                 if (r != 1)
 227                         continue;
 228                 /*
 229                  * Switch to dev_id when dev_port returns the same value for
 230                  * all ports. May happen when using a MOFED release older than
 231                  * 3.0 with a Linux kernel >= 3.15.
 232                  */
 233                 if (dev_port == dev_port_prev)
 234                         goto try_dev_id;
 235                 dev_port_prev = dev_port;
 236                 if (dev_port == 0)
 237                         strlcpy(match, name, IF_NAMESIZE);
 238         }
 239         closedir(dir);
 240         if (match[0] == '\0') {
 241                 rte_errno = ENOENT;
 242                 return -rte_errno;
 243         }
 244         strncpy(ifname, match, IF_NAMESIZE);
 245         return 0;
 246 }
 247
 248 #ifdef MLX5_GLUE
 249
 250 /**
 251  * Suffix RTE_EAL_PMD_PATH with "-glue".
 252  *
 253  * This function performs a sanity check on RTE_EAL_PMD_PATH before
 254  * suffixing its last component.
 255  *
 256  * @param buf[out]
 257  *   Output buffer, should be large enough otherwise NULL is returned.
 258  * @param size
 259  *   Size of @p out.
 260  *
 261  * @return
 262  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
 263  */
 264 static char *
 265 mlx5_glue_path(char *buf, size_t size)
 266 {
 267         static const char *const bad[] = { "/", ".", "..", NULL };
 268         const char *path = RTE_EAL_PMD_PATH;
 269         size_t len = strlen(path);
 270         size_t off;
 271         int i;
 272
 273         while (len && path[len - 1] == '/')
 274                 --len;
 275         for (off = len; off && path[off - 1] != '/'; --off)
 276                 ;
 277         for (i = 0; bad[i]; ++i)
 278                 if (!strncmp(path + off, bad[i], (int)(len - off)))
 279                         goto error;
 280         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
 281         if (i == -1 || (size_t)i >= size)
 282                 goto error;
 283         return buf;
 284 error:
 285         RTE_LOG(ERR, PMD, "unable to append \"-glue\" to last component of"
 286                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"), please"
 287                 " re-configure DPDK");
 288         return NULL;
 289 }
 290
 291 static int
 292 mlx5_glue_dlopen(void)
 293 {
 294         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
 295         void *handle = NULL;
 296
 297         char const *path[] = {
 298                 /*
 299                  * A basic security check is necessary before trusting
 300                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
 301                  */
 302                 (geteuid() == getuid() && getegid() == getgid() ?
 303                  getenv("MLX5_GLUE_PATH") : NULL),
 304                 /*
 305                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
 306                  * variant, otherwise let dlopen() look up libraries on its
 307                  * own.
 308                  */
 309                 (*RTE_EAL_PMD_PATH ?
 310                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
 311         };
 312         unsigned int i = 0;
 313         void **sym;
 314         const char *dlmsg;
 315
 316         while (!handle && i != RTE_DIM(path)) {
 317                 const char *end;
 318                 size_t len;
 319                 int ret;
 320
 321                 if (!path[i]) {
 322                         ++i;
 323                         continue;
 324                 }
 325                 end = strpbrk(path[i], ":;");
 326                 if (!end)
 327                         end = path[i] + strlen(path[i]);
 328                 len = end - path[i];
 329                 ret = 0;
 330                 do {
 331                         char name[ret + 1];
 332
 333                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
 334                                        (int)len, path[i],
 335                                        (!len || *(end - 1) == '/') ? "" : "/");
 336                         if (ret == -1)
 337                                 break;
 338                         if (sizeof(name) != (size_t)ret + 1)
 339                                 continue;
 340                         DRV_LOG(DEBUG, "Looking for rdma-core glue as "
 341                                 "\"%s\"", name);
 342                         handle = dlopen(name, RTLD_LAZY);
 343                         break;
 344                 } while (1);
 345                 path[i] = end + 1;
 346                 if (!*end)
 347                         ++i;
 348         }
 349         if (!handle) {
 350                 rte_errno = EINVAL;
 351                 dlmsg = dlerror();
 352                 if (dlmsg)
 353                         DRV_LOG(WARNING, "Cannot load glue library: %s", dlmsg);
 354                 goto glue_error;
 355         }
 356         sym = dlsym(handle, "mlx5_glue");
 357         if (!sym || !*sym) {
 358                 rte_errno = EINVAL;
 359                 dlmsg = dlerror();
 360                 if (dlmsg)
 361                         DRV_LOG(ERR, "Cannot resolve glue symbol: %s", dlmsg);
 362                 goto glue_error;
 363         }
 364         mlx5_glue = *sym;
 365         return 0;
 366
 367 glue_error:
 368         if (handle)
 369                 dlclose(handle);
 370         return -1;
 371 }
 372
 373 #endif
 374
 375 /**
 376  * Initialization routine for run-time dependency on rdma-core.
 377  */
 378 void
 379 mlx5_glue_constructor(void)
 380 {
 381         /*
 382          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 383          * huge pages. Calling ibv_fork_init() during init allows
 384          * applications to use fork() safely for purposes other than
 385          * using this PMD, which is not supported in forked processes.
 386          */
 387         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 388         /* Match the size of Rx completion entry to the size of a cacheline. */
 389         if (RTE_CACHE_LINE_SIZE == 128)
 390                 setenv("MLX5_CQE_SIZE", "128", 0);
 391         /*
 392          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
 393          * cleanup all the Verbs resources even when the device was removed.
 394          */
 395         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
 396
 397 #ifdef MLX5_GLUE
 398         if (mlx5_glue_dlopen() != 0)
 399                 goto glue_error;
 400 #endif
 401
 402 #ifdef RTE_LIBRTE_MLX5_DEBUG
 403         /* Glue structure must not contain any NULL pointers. */
 404         {
 405                 unsigned int i;
 406
 407                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
 408                         MLX5_ASSERT(((const void *const *)mlx5_glue)[i]);
 409         }
 410 #endif
 411         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
 412                 rte_errno = EINVAL;
 413                 DRV_LOG(ERR, "rdma-core glue \"%s\" mismatch: \"%s\" is "
 414                         "required", mlx5_glue->version, MLX5_GLUE_VERSION);
 415                 goto glue_error;
 416         }
 417         mlx5_glue->fork_init();
 418         return;
 419
 420 glue_error:
 421         DRV_LOG(WARNING, "Cannot initialize MLX5 common due to missing"
 422                 " run-time dependency on rdma-core libraries (libibverbs,"
 423                 " libmlx5)");
 424         mlx5_glue = NULL;
 425 }
 426
 427 struct ibv_device *
 428 mlx5_os_get_ibv_device(struct rte_pci_addr *addr)
 429 {
 430         int n;
 431         struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
 432         struct ibv_device *ibv_match = NULL;
 433
 434         if (ibv_list == NULL) {
 435                 rte_errno = ENOSYS;
 436                 return NULL;
 437         }
 438         while (n-- > 0) {
 439                 struct rte_pci_addr paddr;
 440
 441                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
 442                 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &paddr) != 0)
 443                         continue;
 444                 if (rte_pci_addr_cmp(addr, &paddr) != 0)
 445                         continue;
 446                 ibv_match = ibv_list[n];
 447                 break;
 448         }
 449         if (ibv_match == NULL)
 450                 rte_errno = ENOENT;
 451         mlx5_glue->free_device_list(ibv_list);
 452         return ibv_match;
 453 }