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