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                         ret = 0;
  70                         break;
  71                 }
  72         }
  73         fclose(file);
  74         return 0;
  75 }
  76
  77 /**
  78  * Extract port name, as a number, from sysfs or netlink information.
  79  *
  80  * @param[in] port_name_in
  81  *   String representing the port name.
  82  * @param[out] port_info_out
  83  *   Port information, including port name as a number and port name
  84  *   type if recognized
  85  *
  86  * @return
  87  *   port_name field set according to recognized name format.
  88  */
  89 void
  90 mlx5_translate_port_name(const char *port_name_in,
  91                          struct mlx5_switch_info *port_info_out)
  92 {
  93         char pf_c1, pf_c2, vf_c1, vf_c2;
  94         char *end;
  95         int sc_items;
  96
  97         /*
  98          * Check for port-name as a string of the form pf0vf0
  99          * (support kernel ver >= 5.0 or OFED ver >= 4.6).
 100          */
 101         sc_items = sscanf(port_name_in, "%c%c%d%c%c%d",
 102                           &pf_c1, &pf_c2, &port_info_out->pf_num,
 103                           &vf_c1, &vf_c2, &port_info_out->port_name);
 104         if (sc_items == 6 &&
 105             pf_c1 == 'p' && pf_c2 == 'f' &&
 106             vf_c1 == 'v' && vf_c2 == 'f') {
 107                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFVF;
 108                 return;
 109         }
 110         /*
 111          * Check for port-name as a string of the form p0
 112          * (support kernel ver >= 5.0, or OFED ver >= 4.6).
 113          */
 114         sc_items = sscanf(port_name_in, "%c%d",
 115                           &pf_c1, &port_info_out->port_name);
 116         if (sc_items == 2 && pf_c1 == 'p') {
 117                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
 118                 return;
 119         }
 120         /*
 121          * Check for port-name as a string of the form pf0
 122          * (support kernel ver >= 5.7 for HPF representor on BF).
 123          */
 124         sc_items = sscanf(port_name_in, "%c%c%d",
 125                           &pf_c1, &pf_c2, &port_info_out->pf_num);
 126         if (sc_items == 3 && pf_c1 == 'p' && pf_c2 == 'f') {
 127                 port_info_out->port_name = -1;
 128                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFHPF;
 129                 return;
 130         }
 131         /* Check for port-name as a number (support kernel ver < 5.0 */
 132         errno = 0;
 133         port_info_out->port_name = strtol(port_name_in, &end, 0);
 134         if (!errno &&
 135             (size_t)(end - port_name_in) == strlen(port_name_in)) {
 136                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
 137                 return;
 138         }
 139         port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
 140 }
 141
 142 /**
 143  * Get kernel interface name from IB device path.
 144  *
 145  * @param[in] ibdev_path
 146  *   Pointer to IB device path.
 147  * @param[out] ifname
 148  *   Interface name output buffer.
 149  *
 150  * @return
 151  *   0 on success, a negative errno value otherwise and rte_errno is set.
 152  */
 153 int
 154 mlx5_get_ifname_sysfs(const char *ibdev_path, char *ifname)
 155 {
 156         DIR *dir;
 157         struct dirent *dent;
 158         unsigned int dev_type = 0;
 159         unsigned int dev_port_prev = ~0u;
 160         char match[IF_NAMESIZE] = "";
 161
 162         MLX5_ASSERT(ibdev_path);
 163         {
 164                 MKSTR(path, "%s/device/net", ibdev_path);
 165
 166                 dir = opendir(path);
 167                 if (dir == NULL) {
 168                         rte_errno = errno;
 169                         return -rte_errno;
 170                 }
 171         }
 172         while ((dent = readdir(dir)) != NULL) {
 173                 char *name = dent->d_name;
 174                 FILE *file;
 175                 unsigned int dev_port;
 176                 int r;
 177
 178                 if ((name[0] == '.') &&
 179                     ((name[1] == '\0') ||
 180                      ((name[1] == '.') && (name[2] == '\0'))))
 181                         continue;
 182
 183                 MKSTR(path, "%s/device/net/%s/%s",
 184                       ibdev_path, name,
 185                       (dev_type ? "dev_id" : "dev_port"));
 186
 187                 file = fopen(path, "rb");
 188                 if (file == NULL) {
 189                         if (errno != ENOENT)
 190                                 continue;
 191                         /*
 192                          * Switch to dev_id when dev_port does not exist as
 193                          * is the case with Linux kernel versions < 3.15.
 194                          */
 195 try_dev_id:
 196                         match[0] = '\0';
 197                         if (dev_type)
 198                                 break;
 199                         dev_type = 1;
 200                         dev_port_prev = ~0u;
 201                         rewinddir(dir);
 202                         continue;
 203                 }
 204                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 205                 fclose(file);
 206                 if (r != 1)
 207                         continue;
 208                 /*
 209                  * Switch to dev_id when dev_port returns the same value for
 210                  * all ports. May happen when using a MOFED release older than
 211                  * 3.0 with a Linux kernel >= 3.15.
 212                  */
 213                 if (dev_port == dev_port_prev)
 214                         goto try_dev_id;
 215                 dev_port_prev = dev_port;
 216                 if (dev_port == 0)
 217                         strlcpy(match, name, IF_NAMESIZE);
 218         }
 219         closedir(dir);
 220         if (match[0] == '\0') {
 221                 rte_errno = ENOENT;
 222                 return -rte_errno;
 223         }
 224         strncpy(ifname, match, IF_NAMESIZE);
 225         return 0;
 226 }
 227
 228 #ifdef MLX5_GLUE
 229
 230 /**
 231  * Suffix RTE_EAL_PMD_PATH with "-glue".
 232  *
 233  * This function performs a sanity check on RTE_EAL_PMD_PATH before
 234  * suffixing its last component.
 235  *
 236  * @param buf[out]
 237  *   Output buffer, should be large enough otherwise NULL is returned.
 238  * @param size
 239  *   Size of @p out.
 240  *
 241  * @return
 242  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
 243  */
 244 static char *
 245 mlx5_glue_path(char *buf, size_t size)
 246 {
 247         static const char *const bad[] = { "/", ".", "..", NULL };
 248         const char *path = RTE_EAL_PMD_PATH;
 249         size_t len = strlen(path);
 250         size_t off;
 251         int i;
 252
 253         while (len && path[len - 1] == '/')
 254                 --len;
 255         for (off = len; off && path[off - 1] != '/'; --off)
 256                 ;
 257         for (i = 0; bad[i]; ++i)
 258                 if (!strncmp(path + off, bad[i], (int)(len - off)))
 259                         goto error;
 260         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
 261         if (i == -1 || (size_t)i >= size)
 262                 goto error;
 263         return buf;
 264 error:
 265         RTE_LOG(ERR, PMD, "unable to append \"-glue\" to last component of"
 266                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"), please"
 267                 " re-configure DPDK");
 268         return NULL;
 269 }
 270
 271 static int
 272 mlx5_glue_dlopen(void)
 273 {
 274         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
 275         void *handle = NULL;
 276
 277         char const *path[] = {
 278                 /*
 279                  * A basic security check is necessary before trusting
 280                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
 281                  */
 282                 (geteuid() == getuid() && getegid() == getgid() ?
 283                  getenv("MLX5_GLUE_PATH") : NULL),
 284                 /*
 285                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
 286                  * variant, otherwise let dlopen() look up libraries on its
 287                  * own.
 288                  */
 289                 (*RTE_EAL_PMD_PATH ?
 290                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
 291         };
 292         unsigned int i = 0;
 293         void **sym;
 294         const char *dlmsg;
 295
 296         while (!handle && i != RTE_DIM(path)) {
 297                 const char *end;
 298                 size_t len;
 299                 int ret;
 300
 301                 if (!path[i]) {
 302                         ++i;
 303                         continue;
 304                 }
 305                 end = strpbrk(path[i], ":;");
 306                 if (!end)
 307                         end = path[i] + strlen(path[i]);
 308                 len = end - path[i];
 309                 ret = 0;
 310                 do {
 311                         char name[ret + 1];
 312
 313                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
 314                                        (int)len, path[i],
 315                                        (!len || *(end - 1) == '/') ? "" : "/");
 316                         if (ret == -1)
 317                                 break;
 318                         if (sizeof(name) != (size_t)ret + 1)
 319                                 continue;
 320                         DRV_LOG(DEBUG, "Looking for rdma-core glue as "
 321                                 "\"%s\"", name);
 322                         handle = dlopen(name, RTLD_LAZY);
 323                         break;
 324                 } while (1);
 325                 path[i] = end + 1;
 326                 if (!*end)
 327                         ++i;
 328         }
 329         if (!handle) {
 330                 rte_errno = EINVAL;
 331                 dlmsg = dlerror();
 332                 if (dlmsg)
 333                         DRV_LOG(WARNING, "Cannot load glue library: %s", dlmsg);
 334                 goto glue_error;
 335         }
 336         sym = dlsym(handle, "mlx5_glue");
 337         if (!sym || !*sym) {
 338                 rte_errno = EINVAL;
 339                 dlmsg = dlerror();
 340                 if (dlmsg)
 341                         DRV_LOG(ERR, "Cannot resolve glue symbol: %s", dlmsg);
 342                 goto glue_error;
 343         }
 344         mlx5_glue = *sym;
 345         return 0;
 346
 347 glue_error:
 348         if (handle)
 349                 dlclose(handle);
 350         return -1;
 351 }
 352
 353 #endif
 354
 355 /**
 356  * Initialization routine for run-time dependency on rdma-core.
 357  */
 358 void
 359 mlx5_glue_constructor(void)
 360 {
 361         /*
 362          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 363          * huge pages. Calling ibv_fork_init() during init allows
 364          * applications to use fork() safely for purposes other than
 365          * using this PMD, which is not supported in forked processes.
 366          */
 367         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 368         /* Match the size of Rx completion entry to the size of a cacheline. */
 369         if (RTE_CACHE_LINE_SIZE == 128)
 370                 setenv("MLX5_CQE_SIZE", "128", 0);
 371         /*
 372          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
 373          * cleanup all the Verbs resources even when the device was removed.
 374          */
 375         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
 376
 377 #ifdef MLX5_GLUE
 378         if (mlx5_glue_dlopen() != 0)
 379                 goto glue_error;
 380 #endif
 381
 382 #ifdef RTE_LIBRTE_MLX5_DEBUG
 383         /* Glue structure must not contain any NULL pointers. */
 384         {
 385                 unsigned int i;
 386
 387                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
 388                         MLX5_ASSERT(((const void *const *)mlx5_glue)[i]);
 389         }
 390 #endif
 391         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
 392                 rte_errno = EINVAL;
 393                 DRV_LOG(ERR, "rdma-core glue \"%s\" mismatch: \"%s\" is "
 394                         "required", mlx5_glue->version, MLX5_GLUE_VERSION);
 395                 goto glue_error;
 396         }
 397         mlx5_glue->fork_init();
 398         return;
 399
 400 glue_error:
 401         DRV_LOG(WARNING, "Cannot initialize MLX5 common due to missing"
 402                 " run-time dependency on rdma-core libraries (libibverbs,"
 403                 " libmlx5)");
 404         mlx5_glue = NULL;
 405 }
 406