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 int
  27 mlx5_get_pci_addr(const char *dev_path, struct rte_pci_addr *pci_addr)
  28 {
  29         FILE *file;
  30         char line[32];
  31         int rc = -ENOENT;
  32         MKSTR(path, "%s/device/uevent", dev_path);
  33
  34         file = fopen(path, "rb");
  35         if (file == NULL) {
  36                 rte_errno = errno;
  37                 return -rte_errno;
  38         }
  39         while (fgets(line, sizeof(line), file) == line) {
  40                 size_t len = strlen(line);
  41
  42                 /* Truncate long lines. */
  43                 if (len == (sizeof(line) - 1)) {
  44                         while (line[(len - 1)] != '\n') {
  45                                 int ret = fgetc(file);
  46
  47                                 if (ret == EOF)
  48                                         goto exit;
  49                                 line[(len - 1)] = ret;
  50                         }
  51                         /* No match for long lines. */
  52                         continue;
  53                 }
  54                 /* Extract information. */
  55                 if (sscanf(line,
  56                            "PCI_SLOT_NAME="
  57                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
  58                            &pci_addr->domain,
  59                            &pci_addr->bus,
  60                            &pci_addr->devid,
  61                            &pci_addr->function) == 4) {
  62                         rc = 0;
  63                         break;
  64                 }
  65         }
  66 exit:
  67         fclose(file);
  68         if (rc)
  69                 rte_errno = -rc;
  70         return rc;
  71 }
  72
  73 /**
  74  * Extract port name, as a number, from sysfs or netlink information.
  75  *
  76  * @param[in] port_name_in
  77  *   String representing the port name.
  78  * @param[out] port_info_out
  79  *   Port information, including port name as a number and port name
  80  *   type if recognized
  81  *
  82  * @return
  83  *   port_name field set according to recognized name format.
  84  */
  85 void
  86 mlx5_translate_port_name(const char *port_name_in,
  87                          struct mlx5_switch_info *port_info_out)
  88 {
  89         char ctrl = 0, pf_c1, pf_c2, vf_c1, vf_c2, eol;
  90         char *end;
  91         int sc_items;
  92
  93         sc_items = sscanf(port_name_in, "%c%d",
  94                           &ctrl, &port_info_out->ctrl_num);
  95         if (sc_items == 2 && ctrl == 'c') {
  96                 port_name_in++; /* 'c' */
  97                 port_name_in += snprintf(NULL, 0, "%d",
  98                                           port_info_out->ctrl_num);
  99         }
 100         /* Check for port-name as a string of the form pf0vf0 or pf0sf0 */
 101         sc_items = sscanf(port_name_in, "%c%c%d%c%c%d%c",
 102                           &pf_c1, &pf_c2, &port_info_out->pf_num,
 103                           &vf_c1, &vf_c2, &port_info_out->port_name, &eol);
 104         if (sc_items == 6 && pf_c1 == 'p' && pf_c2 == 'f') {
 105                 if (vf_c1 == 'v' && vf_c2 == 'f') {
 106                         /* Kernel ver >= 5.0 or OFED ver >= 4.6 */
 107                         port_info_out->name_type =
 108                                         MLX5_PHYS_PORT_NAME_TYPE_PFVF;
 109                         return;
 110                 }
 111                 if (vf_c1 == 's' && vf_c2 == 'f') {
 112                         /* Kernel ver >= 5.11 or OFED ver >= 5.1 */
 113                         port_info_out->name_type =
 114                                         MLX5_PHYS_PORT_NAME_TYPE_PFSF;
 115                         return;
 116                 }
 117         }
 118         /*
 119          * Check for port-name as a string of the form p0
 120          * (support kernel ver >= 5.0, or OFED ver >= 4.6).
 121          */
 122         sc_items = sscanf(port_name_in, "%c%d%c",
 123                           &pf_c1, &port_info_out->port_name, &eol);
 124         if (sc_items == 2 && pf_c1 == 'p') {
 125                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UPLINK;
 126                 return;
 127         }
 128         /*
 129          * Check for port-name as a string of the form pf0
 130          * (support kernel ver >= 5.7 for HPF representor on BF).
 131          */
 132         sc_items = sscanf(port_name_in, "%c%c%d%c",
 133                           &pf_c1, &pf_c2, &port_info_out->pf_num, &eol);
 134         if (sc_items == 3 && pf_c1 == 'p' && pf_c2 == 'f') {
 135                 port_info_out->port_name = -1;
 136                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_PFHPF;
 137                 return;
 138         }
 139         /* Check for port-name as a number (support kernel ver < 5.0 */
 140         errno = 0;
 141         port_info_out->port_name = strtol(port_name_in, &end, 0);
 142         if (!errno &&
 143             (size_t)(end - port_name_in) == strlen(port_name_in)) {
 144                 port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_LEGACY;
 145                 return;
 146         }
 147         port_info_out->name_type = MLX5_PHYS_PORT_NAME_TYPE_UNKNOWN;
 148 }
 149
 150 int
 151 mlx5_get_ifname_sysfs(const char *ibdev_path, char *ifname)
 152 {
 153         DIR *dir;
 154         struct dirent *dent;
 155         unsigned int dev_type = 0;
 156         unsigned int dev_port_prev = ~0u;
 157         char match[IF_NAMESIZE] = "";
 158
 159         MLX5_ASSERT(ibdev_path);
 160         {
 161                 MKSTR(path, "%s/device/net", ibdev_path);
 162
 163                 dir = opendir(path);
 164                 if (dir == NULL) {
 165                         rte_errno = errno;
 166                         return -rte_errno;
 167                 }
 168         }
 169         while ((dent = readdir(dir)) != NULL) {
 170                 char *name = dent->d_name;
 171                 FILE *file;
 172                 unsigned int dev_port;
 173                 int r;
 174
 175                 if ((name[0] == '.') &&
 176                     ((name[1] == '\0') ||
 177                      ((name[1] == '.') && (name[2] == '\0'))))
 178                         continue;
 179
 180                 MKSTR(path, "%s/device/net/%s/%s",
 181                       ibdev_path, name,
 182                       (dev_type ? "dev_id" : "dev_port"));
 183
 184                 file = fopen(path, "rb");
 185                 if (file == NULL) {
 186                         if (errno != ENOENT)
 187                                 continue;
 188                         /*
 189                          * Switch to dev_id when dev_port does not exist as
 190                          * is the case with Linux kernel versions < 3.15.
 191                          */
 192 try_dev_id:
 193                         match[0] = '\0';
 194                         if (dev_type)
 195                                 break;
 196                         dev_type = 1;
 197                         dev_port_prev = ~0u;
 198                         rewinddir(dir);
 199                         continue;
 200                 }
 201                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
 202                 fclose(file);
 203                 if (r != 1)
 204                         continue;
 205                 /*
 206                  * Switch to dev_id when dev_port returns the same value for
 207                  * all ports. May happen when using a MOFED release older than
 208                  * 3.0 with a Linux kernel >= 3.15.
 209                  */
 210                 if (dev_port == dev_port_prev)
 211                         goto try_dev_id;
 212                 dev_port_prev = dev_port;
 213                 if (dev_port == 0)
 214                         strlcpy(match, name, IF_NAMESIZE);
 215         }
 216         closedir(dir);
 217         if (match[0] == '\0') {
 218                 rte_errno = ENOENT;
 219                 return -rte_errno;
 220         }
 221         strncpy(ifname, match, IF_NAMESIZE);
 222         return 0;
 223 }
 224
 225 #ifdef MLX5_GLUE
 226
 227 /**
 228  * Suffix RTE_EAL_PMD_PATH with "-glue".
 229  *
 230  * This function performs a sanity check on RTE_EAL_PMD_PATH before
 231  * suffixing its last component.
 232  *
 233  * @param buf[out]
 234  *   Output buffer, should be large enough otherwise NULL is returned.
 235  * @param size
 236  *   Size of @p out.
 237  *
 238  * @return
 239  *   Pointer to @p buf or @p NULL in case suffix cannot be appended.
 240  */
 241 static char *
 242 mlx5_glue_path(char *buf, size_t size)
 243 {
 244         static const char *const bad[] = { "/", ".", "..", NULL };
 245         const char *path = RTE_EAL_PMD_PATH;
 246         size_t len = strlen(path);
 247         size_t off;
 248         int i;
 249
 250         while (len && path[len - 1] == '/')
 251                 --len;
 252         for (off = len; off && path[off - 1] != '/'; --off)
 253                 ;
 254         for (i = 0; bad[i]; ++i)
 255                 if (!strncmp(path + off, bad[i], (int)(len - off)))
 256                         goto error;
 257         i = snprintf(buf, size, "%.*s-glue", (int)len, path);
 258         if (i == -1 || (size_t)i >= size)
 259                 goto error;
 260         return buf;
 261 error:
 262         RTE_LOG(ERR, PMD, "unable to append \"-glue\" to last component of"
 263                 " RTE_EAL_PMD_PATH (\"" RTE_EAL_PMD_PATH "\"), please"
 264                 " re-configure DPDK");
 265         return NULL;
 266 }
 267
 268 static int
 269 mlx5_glue_dlopen(void)
 270 {
 271         char glue_path[sizeof(RTE_EAL_PMD_PATH) - 1 + sizeof("-glue")];
 272         void *handle = NULL;
 273
 274         char const *path[] = {
 275                 /*
 276                  * A basic security check is necessary before trusting
 277                  * MLX5_GLUE_PATH, which may override RTE_EAL_PMD_PATH.
 278                  */
 279                 (geteuid() == getuid() && getegid() == getgid() ?
 280                  getenv("MLX5_GLUE_PATH") : NULL),
 281                 /*
 282                  * When RTE_EAL_PMD_PATH is set, use its glue-suffixed
 283                  * variant, otherwise let dlopen() look up libraries on its
 284                  * own.
 285                  */
 286                 (*RTE_EAL_PMD_PATH ?
 287                  mlx5_glue_path(glue_path, sizeof(glue_path)) : ""),
 288         };
 289         unsigned int i = 0;
 290         void **sym;
 291         const char *dlmsg;
 292
 293         while (!handle && i != RTE_DIM(path)) {
 294                 const char *end;
 295                 size_t len;
 296                 int ret;
 297
 298                 if (!path[i]) {
 299                         ++i;
 300                         continue;
 301                 }
 302                 end = strpbrk(path[i], ":;");
 303                 if (!end)
 304                         end = path[i] + strlen(path[i]);
 305                 len = end - path[i];
 306                 ret = 0;
 307                 do {
 308                         char name[ret + 1];
 309
 310                         ret = snprintf(name, sizeof(name), "%.*s%s" MLX5_GLUE,
 311                                        (int)len, path[i],
 312                                        (!len || *(end - 1) == '/') ? "" : "/");
 313                         if (ret == -1)
 314                                 break;
 315                         if (sizeof(name) != (size_t)ret + 1)
 316                                 continue;
 317                         DRV_LOG(DEBUG, "Looking for rdma-core glue as "
 318                                 "\"%s\"", name);
 319                         handle = dlopen(name, RTLD_LAZY);
 320                         break;
 321                 } while (1);
 322                 path[i] = end + 1;
 323                 if (!*end)
 324                         ++i;
 325         }
 326         if (!handle) {
 327                 rte_errno = EINVAL;
 328                 dlmsg = dlerror();
 329                 if (dlmsg)
 330                         DRV_LOG(WARNING, "Cannot load glue library: %s", dlmsg);
 331                 goto glue_error;
 332         }
 333         sym = dlsym(handle, "mlx5_glue");
 334         if (!sym || !*sym) {
 335                 rte_errno = EINVAL;
 336                 dlmsg = dlerror();
 337                 if (dlmsg)
 338                         DRV_LOG(ERR, "Cannot resolve glue symbol: %s", dlmsg);
 339                 goto glue_error;
 340         }
 341         mlx5_glue = *sym;
 342         return 0;
 343
 344 glue_error:
 345         if (handle)
 346                 dlclose(handle);
 347         return -1;
 348 }
 349
 350 #endif
 351
 352 /**
 353  * Initialization routine for run-time dependency on rdma-core.
 354  */
 355 void
 356 mlx5_glue_constructor(void)
 357 {
 358         /*
 359          * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
 360          * huge pages. Calling ibv_fork_init() during init allows
 361          * applications to use fork() safely for purposes other than
 362          * using this PMD, which is not supported in forked processes.
 363          */
 364         setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
 365         /* Match the size of Rx completion entry to the size of a cacheline. */
 366         if (RTE_CACHE_LINE_SIZE == 128)
 367                 setenv("MLX5_CQE_SIZE", "128", 0);
 368         /*
 369          * MLX5_DEVICE_FATAL_CLEANUP tells ibv_destroy functions to
 370          * cleanup all the Verbs resources even when the device was removed.
 371          */
 372         setenv("MLX5_DEVICE_FATAL_CLEANUP", "1", 1);
 373
 374 #ifdef MLX5_GLUE
 375         if (mlx5_glue_dlopen() != 0)
 376                 goto glue_error;
 377 #endif
 378
 379 #ifdef RTE_LIBRTE_MLX5_DEBUG
 380         /* Glue structure must not contain any NULL pointers. */
 381         {
 382                 unsigned int i;
 383
 384                 for (i = 0; i != sizeof(*mlx5_glue) / sizeof(void *); ++i)
 385                         MLX5_ASSERT(((const void *const *)mlx5_glue)[i]);
 386         }
 387 #endif
 388         if (strcmp(mlx5_glue->version, MLX5_GLUE_VERSION)) {
 389                 rte_errno = EINVAL;
 390                 DRV_LOG(ERR, "rdma-core glue \"%s\" mismatch: \"%s\" is "
 391                         "required", mlx5_glue->version, MLX5_GLUE_VERSION);
 392                 goto glue_error;
 393         }
 394         mlx5_glue->fork_init();
 395         return;
 396
 397 glue_error:
 398         DRV_LOG(WARNING, "Cannot initialize MLX5 common due to missing"
 399                 " run-time dependency on rdma-core libraries (libibverbs,"
 400                 " libmlx5)");
 401         mlx5_glue = NULL;
 402 }
 403
 404 struct ibv_device *
 405 mlx5_os_get_ibv_device(const struct rte_pci_addr *addr)
 406 {
 407         int n;
 408         struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
 409         struct ibv_device *ibv_match = NULL;
 410
 411         if (ibv_list == NULL) {
 412                 rte_errno = ENOSYS;
 413                 return NULL;
 414         }
 415         while (n-- > 0) {
 416                 struct rte_pci_addr paddr;
 417
 418                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
 419                 if (mlx5_get_pci_addr(ibv_list[n]->ibdev_path, &paddr) != 0)
 420                         continue;
 421                 if (rte_pci_addr_cmp(addr, &paddr) != 0)
 422                         continue;
 423                 ibv_match = ibv_list[n];
 424                 break;
 425         }
 426         if (ibv_match == NULL)
 427                 rte_errno = ENOENT;
 428         mlx5_glue->free_device_list(ibv_list);
 429         return ibv_match;
 430 }