lib/eal/freebsd/eal.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 Intel Corporation.
   3  * Copyright(c) 2014 6WIND S.A.
   4  */
   5
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <stdint.h>
   9 #include <string.h>
  10 #include <stdarg.h>
  11 #include <unistd.h>
  12 #include <pthread.h>
  13 #include <syslog.h>
  14 #include <getopt.h>
  15 #include <sys/file.h>
  16 #include <stddef.h>
  17 #include <errno.h>
  18 #include <limits.h>
  19 #include <sys/mman.h>
  20 #include <sys/queue.h>
  21 #include <sys/stat.h>
  22
  23 #include <rte_compat.h>
  24 #include <rte_common.h>
  25 #include <rte_debug.h>
  26 #include <rte_memory.h>
  27 #include <rte_launch.h>
  28 #include <rte_eal.h>
  29 #include <rte_errno.h>
  30 #include <rte_per_lcore.h>
  31 #include <rte_lcore.h>
  32 #include <rte_service_component.h>
  33 #include <rte_log.h>
  34 #include <rte_random.h>
  35 #include <rte_cycles.h>
  36 #include <rte_string_fns.h>
  37 #include <rte_cpuflags.h>
  38 #include <rte_interrupts.h>
  39 #include <rte_bus.h>
  40 #include <rte_dev.h>
  41 #include <rte_devargs.h>
  42 #include <rte_version.h>
  43 #include <rte_vfio.h>
  44 #include <malloc_heap.h>
  45 #include <telemetry_internal.h>
  46
  47 #include "eal_private.h"
  48 #include "eal_thread.h"
  49 #include "eal_internal_cfg.h"
  50 #include "eal_filesystem.h"
  51 #include "eal_hugepages.h"
  52 #include "eal_options.h"
  53 #include "eal_memcfg.h"
  54 #include "eal_trace.h"
  55
  56 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  57
  58 /* define fd variable here, because file needs to be kept open for the
  59  * duration of the program, as we hold a write lock on it in the primary proc */
  60 static int mem_cfg_fd = -1;
  61
  62 static struct flock wr_lock = {
  63                 .l_type = F_WRLCK,
  64                 .l_whence = SEEK_SET,
  65                 .l_start = offsetof(struct rte_mem_config, memsegs),
  66                 .l_len = RTE_SIZEOF_FIELD(struct rte_mem_config, memsegs),
  67 };
  68
  69 /* internal configuration (per-core) */
  70 struct lcore_config lcore_config[RTE_MAX_LCORE];
  71
  72 /* used by rte_rdtsc() */
  73 int rte_cycles_vmware_tsc_map;
  74
  75 static const char *default_runtime_dir = "/var/run";
  76
  77 int
  78 eal_create_runtime_dir(void)
  79 {
  80         const char *directory = default_runtime_dir;
  81         const char *xdg_runtime_dir = getenv("XDG_RUNTIME_DIR");
  82         const char *fallback = "/tmp";
  83         char run_dir[PATH_MAX];
  84         char tmp[PATH_MAX];
  85         int ret;
  86
  87         if (getuid() != 0) {
  88                 /* try XDG path first, fall back to /tmp */
  89                 if (xdg_runtime_dir != NULL)
  90                         directory = xdg_runtime_dir;
  91                 else
  92                         directory = fallback;
  93         }
  94         /* create DPDK subdirectory under runtime dir */
  95         ret = snprintf(tmp, sizeof(tmp), "%s/dpdk", directory);
  96         if (ret < 0 || ret == sizeof(tmp)) {
  97                 RTE_LOG(ERR, EAL, "Error creating DPDK runtime path name\n");
  98                 return -1;
  99         }
 100
 101         /* create prefix-specific subdirectory under DPDK runtime dir */
 102         ret = snprintf(run_dir, sizeof(run_dir), "%s/%s",
 103                         tmp, eal_get_hugefile_prefix());
 104         if (ret < 0 || ret == sizeof(run_dir)) {
 105                 RTE_LOG(ERR, EAL, "Error creating prefix-specific runtime path name\n");
 106                 return -1;
 107         }
 108
 109         /* create the path if it doesn't exist. no "mkdir -p" here, so do it
 110          * step by step.
 111          */
 112         ret = mkdir(tmp, 0700);
 113         if (ret < 0 && errno != EEXIST) {
 114                 RTE_LOG(ERR, EAL, "Error creating '%s': %s\n",
 115                         tmp, strerror(errno));
 116                 return -1;
 117         }
 118
 119         ret = mkdir(run_dir, 0700);
 120         if (ret < 0 && errno != EEXIST) {
 121                 RTE_LOG(ERR, EAL, "Error creating '%s': %s\n",
 122                         run_dir, strerror(errno));
 123                 return -1;
 124         }
 125
 126         if (eal_set_runtime_dir(run_dir, sizeof(run_dir)))
 127                 return -1;
 128
 129         return 0;
 130 }
 131
 132 int
 133 eal_clean_runtime_dir(void)
 134 {
 135         /* FreeBSD doesn't need this implemented for now, because, unlike Linux,
 136          * FreeBSD doesn't create per-process files, so no need to clean up.
 137          */
 138         return 0;
 139 }
 140
 141 /* parse a sysfs (or other) file containing one integer value */
 142 int
 143 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 144 {
 145         FILE *f;
 146         char buf[BUFSIZ];
 147         char *end = NULL;
 148
 149         if ((f = fopen(filename, "r")) == NULL) {
 150                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 151                         __func__, filename);
 152                 return -1;
 153         }
 154
 155         if (fgets(buf, sizeof(buf), f) == NULL) {
 156                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 157                         __func__, filename);
 158                 fclose(f);
 159                 return -1;
 160         }
 161         *val = strtoul(buf, &end, 0);
 162         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 163                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 164                                 __func__, filename);
 165                 fclose(f);
 166                 return -1;
 167         }
 168         fclose(f);
 169         return 0;
 170 }
 171
 172
 173 /* create memory configuration in shared/mmap memory. Take out
 174  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 175  * This means we never close the file while running (auto-close on exit).
 176  * We also don't lock the whole file, so that in future we can use read-locks
 177  * on other parts, e.g. memzones, to detect if there are running secondary
 178  * processes. */
 179 static int
 180 rte_eal_config_create(void)
 181 {
 182         struct rte_config *config = rte_eal_get_configuration();
 183         const struct internal_config *internal_conf =
 184                 eal_get_internal_configuration();
 185         size_t page_sz = sysconf(_SC_PAGE_SIZE);
 186         size_t cfg_len = sizeof(struct rte_mem_config);
 187         size_t cfg_len_aligned = RTE_ALIGN(cfg_len, page_sz);
 188         void *rte_mem_cfg_addr, *mapped_mem_cfg_addr;
 189         int retval;
 190
 191         const char *pathname = eal_runtime_config_path();
 192
 193         if (internal_conf->no_shconf)
 194                 return 0;
 195
 196         /* map the config before base address so that we don't waste a page */
 197         if (internal_conf->base_virtaddr != 0)
 198                 rte_mem_cfg_addr = (void *)
 199                         RTE_ALIGN_FLOOR(internal_conf->base_virtaddr -
 200                         sizeof(struct rte_mem_config), page_sz);
 201         else
 202                 rte_mem_cfg_addr = NULL;
 203
 204         if (mem_cfg_fd < 0){
 205                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0600);
 206                 if (mem_cfg_fd < 0) {
 207                         RTE_LOG(ERR, EAL, "Cannot open '%s' for rte_mem_config\n",
 208                                 pathname);
 209                         return -1;
 210                 }
 211         }
 212
 213         retval = ftruncate(mem_cfg_fd, cfg_len);
 214         if (retval < 0){
 215                 close(mem_cfg_fd);
 216                 mem_cfg_fd = -1;
 217                 RTE_LOG(ERR, EAL, "Cannot resize '%s' for rte_mem_config\n",
 218                         pathname);
 219                 return -1;
 220         }
 221
 222         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 223         if (retval < 0){
 224                 close(mem_cfg_fd);
 225                 mem_cfg_fd = -1;
 226                 RTE_LOG(ERR, EAL, "Cannot create lock on '%s'. Is another primary "
 227                         "process running?\n", pathname);
 228                 return -1;
 229         }
 230
 231         /* reserve space for config */
 232         rte_mem_cfg_addr = eal_get_virtual_area(rte_mem_cfg_addr,
 233                         &cfg_len_aligned, page_sz, 0, 0);
 234         if (rte_mem_cfg_addr == NULL) {
 235                 RTE_LOG(ERR, EAL, "Cannot mmap memory for rte_config\n");
 236                 close(mem_cfg_fd);
 237                 mem_cfg_fd = -1;
 238                 return -1;
 239         }
 240
 241         /* remap the actual file into the space we've just reserved */
 242         mapped_mem_cfg_addr = mmap(rte_mem_cfg_addr,
 243                         cfg_len_aligned, PROT_READ | PROT_WRITE,
 244                         MAP_SHARED | MAP_FIXED, mem_cfg_fd, 0);
 245         if (mapped_mem_cfg_addr == MAP_FAILED) {
 246                 RTE_LOG(ERR, EAL, "Cannot remap memory for rte_config\n");
 247                 munmap(rte_mem_cfg_addr, cfg_len);
 248                 close(mem_cfg_fd);
 249                 mem_cfg_fd = -1;
 250                 return -1;
 251         }
 252
 253         memcpy(rte_mem_cfg_addr, config->mem_config, sizeof(struct rte_mem_config));
 254         config->mem_config = rte_mem_cfg_addr;
 255
 256         /* store address of the config in the config itself so that secondary
 257          * processes could later map the config into this exact location
 258          */
 259         config->mem_config->mem_cfg_addr = (uintptr_t) rte_mem_cfg_addr;
 260         return 0;
 261 }
 262
 263 /* attach to an existing shared memory config */
 264 static int
 265 rte_eal_config_attach(void)
 266 {
 267         void *rte_mem_cfg_addr;
 268         const char *pathname = eal_runtime_config_path();
 269         struct rte_config *config = rte_eal_get_configuration();
 270         const struct internal_config *internal_conf =
 271                 eal_get_internal_configuration();
 272
 273
 274         if (internal_conf->no_shconf)
 275                 return 0;
 276
 277         if (mem_cfg_fd < 0){
 278                 mem_cfg_fd = open(pathname, O_RDWR);
 279                 if (mem_cfg_fd < 0) {
 280                         RTE_LOG(ERR, EAL, "Cannot open '%s' for rte_mem_config\n",
 281                                 pathname);
 282                         return -1;
 283                 }
 284         }
 285
 286         rte_mem_cfg_addr = mmap(NULL, sizeof(*config->mem_config),
 287                                 PROT_READ, MAP_SHARED, mem_cfg_fd, 0);
 288         /* don't close the fd here, it will be closed on reattach */
 289         if (rte_mem_cfg_addr == MAP_FAILED) {
 290                 close(mem_cfg_fd);
 291                 mem_cfg_fd = -1;
 292                 RTE_LOG(ERR, EAL, "Cannot mmap memory for rte_config! error %i (%s)\n",
 293                         errno, strerror(errno));
 294                 return -1;
 295         }
 296
 297         config->mem_config = rte_mem_cfg_addr;
 298
 299         return 0;
 300 }
 301
 302 /* reattach the shared config at exact memory location primary process has it */
 303 static int
 304 rte_eal_config_reattach(void)
 305 {
 306         struct rte_mem_config *mem_config;
 307         void *rte_mem_cfg_addr;
 308         struct rte_config *config = rte_eal_get_configuration();
 309         const struct internal_config *internal_conf =
 310                 eal_get_internal_configuration();
 311
 312         if (internal_conf->no_shconf)
 313                 return 0;
 314
 315         /* save the address primary process has mapped shared config to */
 316         rte_mem_cfg_addr =
 317                         (void *)(uintptr_t)config->mem_config->mem_cfg_addr;
 318
 319         /* unmap original config */
 320         munmap(config->mem_config, sizeof(struct rte_mem_config));
 321
 322         /* remap the config at proper address */
 323         mem_config = (struct rte_mem_config *) mmap(rte_mem_cfg_addr,
 324                         sizeof(*mem_config), PROT_READ | PROT_WRITE, MAP_SHARED,
 325                         mem_cfg_fd, 0);
 326         close(mem_cfg_fd);
 327         mem_cfg_fd = -1;
 328
 329         if (mem_config == MAP_FAILED || mem_config != rte_mem_cfg_addr) {
 330                 if (mem_config != MAP_FAILED) {
 331                         /* errno is stale, don't use */
 332                         RTE_LOG(ERR, EAL, "Cannot mmap memory for rte_config at [%p], got [%p]"
 333                                           " - please use '--" OPT_BASE_VIRTADDR
 334                                           "' option\n",
 335                                 rte_mem_cfg_addr, mem_config);
 336                         munmap(mem_config, sizeof(struct rte_mem_config));
 337                         return -1;
 338                 }
 339                 RTE_LOG(ERR, EAL, "Cannot mmap memory for rte_config! error %i (%s)\n",
 340                         errno, strerror(errno));
 341                 return -1;
 342         }
 343
 344         config->mem_config = mem_config;
 345
 346         return 0;
 347 }
 348
 349 /* Detect if we are a primary or a secondary process */
 350 enum rte_proc_type_t
 351 eal_proc_type_detect(void)
 352 {
 353         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 354         const char *pathname = eal_runtime_config_path();
 355         const struct internal_config *internal_conf =
 356                 eal_get_internal_configuration();
 357
 358         /* if there no shared config, there can be no secondary processes */
 359         if (!internal_conf->no_shconf) {
 360                 /* if we can open the file but not get a write-lock we are a
 361                  * secondary process. NOTE: if we get a file handle back, we
 362                  * keep that open and don't close it to prevent a race condition
 363                  * between multiple opens.
 364                  */
 365                 if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 366                                 (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 367                         ptype = RTE_PROC_SECONDARY;
 368         }
 369
 370         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 371                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 372
 373         return ptype;
 374 }
 375
 376 /* Sets up rte_config structure with the pointer to shared memory config.*/
 377 static int
 378 rte_config_init(void)
 379 {
 380         struct rte_config *config = rte_eal_get_configuration();
 381         const struct internal_config *internal_conf =
 382                 eal_get_internal_configuration();
 383
 384         config->process_type = internal_conf->process_type;
 385
 386         switch (config->process_type) {
 387         case RTE_PROC_PRIMARY:
 388                 if (rte_eal_config_create() < 0)
 389                         return -1;
 390                 eal_mcfg_update_from_internal();
 391                 break;
 392         case RTE_PROC_SECONDARY:
 393                 if (rte_eal_config_attach() < 0)
 394                         return -1;
 395                 eal_mcfg_wait_complete();
 396                 if (eal_mcfg_check_version() < 0) {
 397                         RTE_LOG(ERR, EAL, "Primary and secondary process DPDK version mismatch\n");
 398                         return -1;
 399                 }
 400                 if (rte_eal_config_reattach() < 0)
 401                         return -1;
 402                 if (!__rte_mp_enable()) {
 403                         RTE_LOG(ERR, EAL, "Primary process refused secondary attachment\n");
 404                         return -1;
 405                 }
 406                 eal_mcfg_update_internal();
 407                 break;
 408         case RTE_PROC_AUTO:
 409         case RTE_PROC_INVALID:
 410                 RTE_LOG(ERR, EAL, "Invalid process type %d\n",
 411                         config->process_type);
 412                 return -1;
 413         }
 414
 415         return 0;
 416 }
 417
 418 /* display usage */
 419 static void
 420 eal_usage(const char *prgname)
 421 {
 422         rte_usage_hook_t hook = eal_get_application_usage_hook();
 423
 424         printf("\nUsage: %s ", prgname);
 425         eal_common_usage();
 426         /* Allow the application to print its usage message too if hook is set */
 427         if (hook) {
 428                 printf("===== Application Usage =====\n\n");
 429                 (hook)(prgname);
 430         }
 431 }
 432
 433 static inline size_t
 434 eal_get_hugepage_mem_size(void)
 435 {
 436         uint64_t size = 0;
 437         unsigned i, j;
 438         struct internal_config *internal_conf =
 439                 eal_get_internal_configuration();
 440
 441         for (i = 0; i < internal_conf->num_hugepage_sizes; i++) {
 442                 struct hugepage_info *hpi = &internal_conf->hugepage_info[i];
 443                 if (strnlen(hpi->hugedir, sizeof(hpi->hugedir)) != 0) {
 444                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 445                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 446                         }
 447                 }
 448         }
 449
 450         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 451 }
 452
 453 /* Parse the arguments for --log-level only */
 454 static void
 455 eal_log_level_parse(int argc, char **argv)
 456 {
 457         int opt;
 458         char **argvopt;
 459         int option_index;
 460         const int old_optind = optind;
 461         const int old_optopt = optopt;
 462         const int old_optreset = optreset;
 463         char * const old_optarg = optarg;
 464         struct internal_config *internal_conf =
 465                 eal_get_internal_configuration();
 466
 467         argvopt = argv;
 468         optind = 1;
 469         optreset = 1;
 470
 471         while ((opt = getopt_long(argc, argvopt, eal_short_options,
 472                                   eal_long_options, &option_index)) != EOF) {
 473
 474                 int ret;
 475
 476                 /* getopt is not happy, stop right now */
 477                 if (opt == '?')
 478                         break;
 479
 480                 ret = (opt == OPT_LOG_LEVEL_NUM) ?
 481                     eal_parse_common_option(opt, optarg, internal_conf) : 0;
 482
 483                 /* common parser is not happy */
 484                 if (ret < 0)
 485                         break;
 486         }
 487
 488         /* restore getopt lib */
 489         optind = old_optind;
 490         optopt = old_optopt;
 491         optreset = old_optreset;
 492         optarg = old_optarg;
 493 }
 494
 495 /* Parse the argument given in the command line of the application */
 496 static int
 497 eal_parse_args(int argc, char **argv)
 498 {
 499         int opt, ret;
 500         char **argvopt;
 501         int option_index;
 502         char *prgname = argv[0];
 503         const int old_optind = optind;
 504         const int old_optopt = optopt;
 505         const int old_optreset = optreset;
 506         char * const old_optarg = optarg;
 507         struct internal_config *internal_conf =
 508                 eal_get_internal_configuration();
 509
 510         argvopt = argv;
 511         optind = 1;
 512         optreset = 1;
 513
 514         while ((opt = getopt_long(argc, argvopt, eal_short_options,
 515                                   eal_long_options, &option_index)) != EOF) {
 516
 517                 /* getopt didn't recognise the option */
 518                 if (opt == '?') {
 519                         eal_usage(prgname);
 520                         ret = -1;
 521                         goto out;
 522                 }
 523
 524                 /* eal_log_level_parse() already handled this option */
 525                 if (opt == OPT_LOG_LEVEL_NUM)
 526                         continue;
 527
 528                 ret = eal_parse_common_option(opt, optarg, internal_conf);
 529                 /* common parser is not happy */
 530                 if (ret < 0) {
 531                         eal_usage(prgname);
 532                         ret = -1;
 533                         goto out;
 534                 }
 535                 /* common parser handled this option */
 536                 if (ret == 0)
 537                         continue;
 538
 539                 switch (opt) {
 540                 case OPT_MBUF_POOL_OPS_NAME_NUM:
 541                 {
 542                         char *ops_name = strdup(optarg);
 543                         if (ops_name == NULL)
 544                                 RTE_LOG(ERR, EAL, "Could not store mbuf pool ops name\n");
 545                         else {
 546                                 /* free old ops name */
 547                                 if (internal_conf->user_mbuf_pool_ops_name !=
 548                                                 NULL)
 549                                         free(internal_conf->user_mbuf_pool_ops_name);
 550
 551                                 internal_conf->user_mbuf_pool_ops_name =
 552                                                 ops_name;
 553                         }
 554                         break;
 555                 }
 556                 case 'h':
 557                         eal_usage(prgname);
 558                         exit(EXIT_SUCCESS);
 559                 default:
 560                         if (opt < OPT_LONG_MIN_NUM && isprint(opt)) {
 561                                 RTE_LOG(ERR, EAL, "Option %c is not supported "
 562                                         "on FreeBSD\n", opt);
 563                         } else if (opt >= OPT_LONG_MIN_NUM &&
 564                                    opt < OPT_LONG_MAX_NUM) {
 565                                 RTE_LOG(ERR, EAL, "Option %s is not supported "
 566                                         "on FreeBSD\n",
 567                                         eal_long_options[option_index].name);
 568                         } else {
 569                                 RTE_LOG(ERR, EAL, "Option %d is not supported "
 570                                         "on FreeBSD\n", opt);
 571                         }
 572                         eal_usage(prgname);
 573                         ret = -1;
 574                         goto out;
 575                 }
 576         }
 577
 578         /* create runtime data directory. In no_shconf mode, skip any errors */
 579         if (eal_create_runtime_dir() < 0) {
 580                 if (internal_conf->no_shconf == 0) {
 581                         RTE_LOG(ERR, EAL, "Cannot create runtime directory\n");
 582                         ret = -1;
 583                         goto out;
 584                 } else
 585                         RTE_LOG(WARNING, EAL, "No DPDK runtime directory created\n");
 586         }
 587
 588         if (eal_adjust_config(internal_conf) != 0) {
 589                 ret = -1;
 590                 goto out;
 591         }
 592
 593         /* sanity checks */
 594         if (eal_check_common_options(internal_conf) != 0) {
 595                 eal_usage(prgname);
 596                 ret = -1;
 597                 goto out;
 598         }
 599
 600         if (optind >= 0)
 601                 argv[optind-1] = prgname;
 602         ret = optind-1;
 603
 604 out:
 605         /* restore getopt lib */
 606         optind = old_optind;
 607         optopt = old_optopt;
 608         optreset = old_optreset;
 609         optarg = old_optarg;
 610
 611         return ret;
 612 }
 613
 614 static int
 615 check_socket(const struct rte_memseg_list *msl, void *arg)
 616 {
 617         int *socket_id = arg;
 618
 619         if (msl->external)
 620                 return 0;
 621
 622         if (msl->socket_id == *socket_id && msl->memseg_arr.count != 0)
 623                 return 1;
 624
 625         return 0;
 626 }
 627
 628 static void
 629 eal_check_mem_on_local_socket(void)
 630 {
 631         int socket_id;
 632         const struct rte_config *config = rte_eal_get_configuration();
 633
 634         socket_id = rte_lcore_to_socket_id(config->main_lcore);
 635
 636         if (rte_memseg_list_walk(check_socket, &socket_id) == 0)
 637                 RTE_LOG(WARNING, EAL, "WARNING: Main core has no memory on local socket!\n");
 638 }
 639
 640
 641 static int
 642 sync_func(__rte_unused void *arg)
 643 {
 644         return 0;
 645 }
 646 /* Abstraction for port I/0 privilege */
 647 int
 648 rte_eal_iopl_init(void)
 649 {
 650         static int fd = -1;
 651
 652         if (fd < 0)
 653                 fd = open("/dev/io", O_RDWR);
 654
 655         if (fd < 0)
 656                 return -1;
 657         /* keep fd open for iopl */
 658         return 0;
 659 }
 660
 661 static void rte_eal_init_alert(const char *msg)
 662 {
 663         fprintf(stderr, "EAL: FATAL: %s\n", msg);
 664         RTE_LOG(ERR, EAL, "%s\n", msg);
 665 }
 666
 667 /* Launch threads, called at application init(). */
 668 int
 669 rte_eal_init(int argc, char **argv)
 670 {
 671         int i, fctret, ret;
 672         pthread_t thread_id;
 673         static uint32_t run_once;
 674         uint32_t has_run = 0;
 675         char cpuset[RTE_CPU_AFFINITY_STR_LEN];
 676         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 677         const struct rte_config *config = rte_eal_get_configuration();
 678         struct internal_config *internal_conf =
 679                 eal_get_internal_configuration();
 680         bool has_phys_addr;
 681         enum rte_iova_mode iova_mode;
 682
 683         /* checks if the machine is adequate */
 684         if (!rte_cpu_is_supported()) {
 685                 rte_eal_init_alert("unsupported cpu type.");
 686                 rte_errno = ENOTSUP;
 687                 return -1;
 688         }
 689
 690         if (!__atomic_compare_exchange_n(&run_once, &has_run, 1, 0,
 691                                         __ATOMIC_RELAXED, __ATOMIC_RELAXED)) {
 692                 rte_eal_init_alert("already called initialization.");
 693                 rte_errno = EALREADY;
 694                 return -1;
 695         }
 696
 697         thread_id = pthread_self();
 698
 699         eal_reset_internal_config(internal_conf);
 700
 701         /* clone argv to report out later in telemetry */
 702         eal_save_args(argc, argv);
 703
 704         /* set log level as early as possible */
 705         eal_log_level_parse(argc, argv);
 706
 707         if (rte_eal_cpu_init() < 0) {
 708                 rte_eal_init_alert("Cannot detect lcores.");
 709                 rte_errno = ENOTSUP;
 710                 return -1;
 711         }
 712
 713         fctret = eal_parse_args(argc, argv);
 714         if (fctret < 0) {
 715                 rte_eal_init_alert("Invalid 'command line' arguments.");
 716                 rte_errno = EINVAL;
 717                 __atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
 718                 return -1;
 719         }
 720
 721         /* FreeBSD always uses legacy memory model */
 722         internal_conf->legacy_mem = true;
 723         if (internal_conf->in_memory) {
 724                 RTE_LOG(WARNING, EAL, "Warning: ignoring unsupported flag, '%s'\n", OPT_IN_MEMORY);
 725                 internal_conf->in_memory = false;
 726         }
 727
 728         if (eal_plugins_init() < 0) {
 729                 rte_eal_init_alert("Cannot init plugins");
 730                 rte_errno = EINVAL;
 731                 __atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
 732                 return -1;
 733         }
 734
 735         if (eal_trace_init() < 0) {
 736                 rte_eal_init_alert("Cannot init trace");
 737                 rte_errno = EFAULT;
 738                 __atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
 739                 return -1;
 740         }
 741
 742         if (eal_option_device_parse()) {
 743                 rte_errno = ENODEV;
 744                 __atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
 745                 return -1;
 746         }
 747
 748         if (rte_config_init() < 0) {
 749                 rte_eal_init_alert("Cannot init config");
 750                 return -1;
 751         }
 752
 753         if (rte_eal_intr_init() < 0) {
 754                 rte_eal_init_alert("Cannot init interrupt-handling thread");
 755                 return -1;
 756         }
 757
 758         if (rte_eal_alarm_init() < 0) {
 759                 rte_eal_init_alert("Cannot init alarm");
 760                 /* rte_eal_alarm_init sets rte_errno on failure. */
 761                 return -1;
 762         }
 763
 764         /* Put mp channel init before bus scan so that we can init the vdev
 765          * bus through mp channel in the secondary process before the bus scan.
 766          */
 767         if (rte_mp_channel_init() < 0 && rte_errno != ENOTSUP) {
 768                 rte_eal_init_alert("failed to init mp channel");
 769                 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
 770                         rte_errno = EFAULT;
 771                         return -1;
 772                 }
 773         }
 774
 775         if (rte_bus_scan()) {
 776                 rte_eal_init_alert("Cannot scan the buses for devices");
 777                 rte_errno = ENODEV;
 778                 __atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
 779                 return -1;
 780         }
 781
 782         /*
 783          * PA are only available for hugepages via contigmem.
 784          * If contigmem is inaccessible, rte_eal_hugepage_init() will fail
 785          * with a message describing the cause.
 786          */
 787         has_phys_addr = internal_conf->no_hugetlbfs == 0;
 788         iova_mode = internal_conf->iova_mode;
 789         if (iova_mode == RTE_IOVA_PA && !has_phys_addr) {
 790                 rte_eal_init_alert("Cannot use IOVA as 'PA' since physical addresses are not available");
 791                 rte_errno = EINVAL;
 792                 return -1;
 793         }
 794         if (iova_mode == RTE_IOVA_DC) {
 795                 RTE_LOG(DEBUG, EAL, "Specific IOVA mode is not requested, autodetecting\n");
 796                 if (has_phys_addr) {
 797                         RTE_LOG(DEBUG, EAL, "Selecting IOVA mode according to bus requests\n");
 798                         iova_mode = rte_bus_get_iommu_class();
 799                         if (iova_mode == RTE_IOVA_DC)
 800                                 iova_mode = RTE_IOVA_PA;
 801                 } else {
 802                         iova_mode = RTE_IOVA_VA;
 803                 }
 804         }
 805         rte_eal_get_configuration()->iova_mode = iova_mode;
 806         RTE_LOG(INFO, EAL, "Selected IOVA mode '%s'\n",
 807                 rte_eal_iova_mode() == RTE_IOVA_PA ? "PA" : "VA");
 808
 809         if (internal_conf->no_hugetlbfs == 0) {
 810                 /* rte_config isn't initialized yet */
 811                 ret = internal_conf->process_type == RTE_PROC_PRIMARY ?
 812                         eal_hugepage_info_init() :
 813                         eal_hugepage_info_read();
 814                 if (ret < 0) {
 815                         rte_eal_init_alert("Cannot get hugepage information.");
 816                         rte_errno = EACCES;
 817                         __atomic_store_n(&run_once, 0, __ATOMIC_RELAXED);
 818                         return -1;
 819                 }
 820         }
 821
 822         if (internal_conf->memory == 0 && internal_conf->force_sockets == 0) {
 823                 if (internal_conf->no_hugetlbfs)
 824                         internal_conf->memory = MEMSIZE_IF_NO_HUGE_PAGE;
 825                 else
 826                         internal_conf->memory = eal_get_hugepage_mem_size();
 827         }
 828
 829         if (internal_conf->vmware_tsc_map == 1) {
 830 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 831                 rte_cycles_vmware_tsc_map = 1;
 832                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 833                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 834 #else
 835                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 836                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 837 #endif
 838         }
 839
 840         /* in secondary processes, memory init may allocate additional fbarrays
 841          * not present in primary processes, so to avoid any potential issues,
 842          * initialize memzones first.
 843          */
 844         if (rte_eal_memzone_init() < 0) {
 845                 rte_eal_init_alert("Cannot init memzone");
 846                 rte_errno = ENODEV;
 847                 return -1;
 848         }
 849
 850         if (rte_eal_memory_init() < 0) {
 851                 rte_eal_init_alert("Cannot init memory");
 852                 rte_errno = ENOMEM;
 853                 return -1;
 854         }
 855
 856         if (rte_eal_malloc_heap_init() < 0) {
 857                 rte_eal_init_alert("Cannot init malloc heap");
 858                 rte_errno = ENODEV;
 859                 return -1;
 860         }
 861
 862         if (rte_eal_tailqs_init() < 0) {
 863                 rte_eal_init_alert("Cannot init tail queues for objects");
 864                 rte_errno = EFAULT;
 865                 return -1;
 866         }
 867
 868         if (rte_eal_timer_init() < 0) {
 869                 rte_eal_init_alert("Cannot init HPET or TSC timers");
 870                 rte_errno = ENOTSUP;
 871                 return -1;
 872         }
 873
 874         eal_check_mem_on_local_socket();
 875
 876         if (pthread_setaffinity_np(pthread_self(), sizeof(rte_cpuset_t),
 877                         &lcore_config[config->main_lcore].cpuset) != 0) {
 878                 rte_eal_init_alert("Cannot set affinity");
 879                 rte_errno = EINVAL;
 880                 return -1;
 881         }
 882         __rte_thread_init(config->main_lcore,
 883                 &lcore_config[config->main_lcore].cpuset);
 884
 885         ret = eal_thread_dump_current_affinity(cpuset, sizeof(cpuset));
 886
 887         RTE_LOG(DEBUG, EAL, "Main lcore %u is ready (tid=%p;cpuset=[%s%s])\n",
 888                 config->main_lcore, thread_id, cpuset,
 889                 ret == 0 ? "" : "...");
 890
 891         RTE_LCORE_FOREACH_WORKER(i) {
 892
 893                 /*
 894                  * create communication pipes between main thread
 895                  * and children
 896                  */
 897                 if (pipe(lcore_config[i].pipe_main2worker) < 0)
 898                         rte_panic("Cannot create pipe\n");
 899                 if (pipe(lcore_config[i].pipe_worker2main) < 0)
 900                         rte_panic("Cannot create pipe\n");
 901
 902                 lcore_config[i].state = WAIT;
 903
 904                 /* create a thread for each lcore */
 905                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 906                                      eal_thread_loop, NULL);
 907                 if (ret != 0)
 908                         rte_panic("Cannot create thread\n");
 909
 910                 /* Set thread_name for aid in debugging. */
 911                 snprintf(thread_name, sizeof(thread_name),
 912                                 "lcore-worker-%d", i);
 913                 rte_thread_setname(lcore_config[i].thread_id, thread_name);
 914
 915                 ret = pthread_setaffinity_np(lcore_config[i].thread_id,
 916                         sizeof(rte_cpuset_t), &lcore_config[i].cpuset);
 917                 if (ret != 0)
 918                         rte_panic("Cannot set affinity\n");
 919         }
 920
 921         /*
 922          * Launch a dummy function on all worker lcores, so that main lcore
 923          * knows they are all ready when this function returns.
 924          */
 925         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MAIN);
 926         rte_eal_mp_wait_lcore();
 927
 928         /* initialize services so vdevs register service during bus_probe. */
 929         ret = rte_service_init();
 930         if (ret) {
 931                 rte_eal_init_alert("rte_service_init() failed");
 932                 rte_errno = -ret;
 933                 return -1;
 934         }
 935
 936         /* Probe all the buses and devices/drivers on them */
 937         if (rte_bus_probe()) {
 938                 rte_eal_init_alert("Cannot probe devices");
 939                 rte_errno = ENOTSUP;
 940                 return -1;
 941         }
 942
 943         /* initialize default service/lcore mappings and start running. Ignore
 944          * -ENOTSUP, as it indicates no service coremask passed to EAL.
 945          */
 946         ret = rte_service_start_with_defaults();
 947         if (ret < 0 && ret != -ENOTSUP) {
 948                 rte_errno = -ret;
 949                 return -1;
 950         }
 951
 952         /*
 953          * Clean up unused files in runtime directory. We do this at the end of
 954          * init and not at the beginning because we want to clean stuff up
 955          * whether we are primary or secondary process, but we cannot remove
 956          * primary process' files because secondary should be able to run even
 957          * if primary process is dead.
 958          *
 959          * In no_shconf mode, no runtime directory is created in the first
 960          * place, so no cleanup needed.
 961          */
 962         if (!internal_conf->no_shconf && eal_clean_runtime_dir() < 0) {
 963                 rte_eal_init_alert("Cannot clear runtime directory");
 964                 return -1;
 965         }
 966         if (rte_eal_process_type() == RTE_PROC_PRIMARY && !internal_conf->no_telemetry) {
 967                 int tlog = rte_log_register_type_and_pick_level(
 968                                 "lib.telemetry", RTE_LOG_WARNING);
 969                 if (tlog < 0)
 970                         tlog = RTE_LOGTYPE_EAL;
 971                 if (rte_telemetry_init(rte_eal_get_runtime_dir(),
 972                                 rte_version(),
 973                                 &internal_conf->ctrl_cpuset, rte_log, tlog) != 0)
 974                         return -1;
 975         }
 976
 977         eal_mcfg_complete();
 978
 979         return fctret;
 980 }
 981
 982 int
 983 rte_eal_cleanup(void)
 984 {
 985         struct internal_config *internal_conf =
 986                 eal_get_internal_configuration();
 987         rte_service_finalize();
 988         rte_mp_channel_cleanup();
 989         /* after this point, any DPDK pointers will become dangling */
 990         rte_eal_memory_detach();
 991         rte_eal_alarm_cleanup();
 992         rte_trace_save();
 993         eal_trace_fini();
 994         eal_cleanup_config(internal_conf);
 995         return 0;
 996 }
 997
 998 int rte_eal_create_uio_dev(void)
 999 {
1000         const struct internal_config *internal_conf =
1001                 eal_get_internal_configuration();
1002         return internal_conf->create_uio_dev;
1003 }
1004
1005 enum rte_intr_mode
1006 rte_eal_vfio_intr_mode(void)
1007 {
1008         return RTE_INTR_MODE_NONE;
1009 }
1010
1011 void
1012 rte_eal_vfio_get_vf_token(__rte_unused rte_uuid_t vf_token)
1013 {
1014 }
1015
1016 int rte_vfio_setup_device(__rte_unused const char *sysfs_base,
1017                       __rte_unused const char *dev_addr,
1018                       __rte_unused int *vfio_dev_fd,
1019                       __rte_unused struct vfio_device_info *device_info)
1020 {
1021         return -1;
1022 }
1023
1024 int rte_vfio_release_device(__rte_unused const char *sysfs_base,
1025                         __rte_unused const char *dev_addr,
1026                         __rte_unused int fd)
1027 {
1028         return -1;
1029 }
1030
1031 int rte_vfio_enable(__rte_unused const char *modname)
1032 {
1033         return -1;
1034 }
1035
1036 int rte_vfio_is_enabled(__rte_unused const char *modname)
1037 {
1038         return 0;
1039 }
1040
1041 int rte_vfio_noiommu_is_enabled(void)
1042 {
1043         return 0;
1044 }
1045
1046 int rte_vfio_clear_group(__rte_unused int vfio_group_fd)
1047 {
1048         return -1;
1049 }
1050
1051 int
1052 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
1053                        __rte_unused const char *dev_addr,
1054                        __rte_unused int *iommu_group_num)
1055 {
1056         return -1;
1057 }
1058
1059 int
1060 rte_vfio_get_container_fd(void)
1061 {
1062         return -1;
1063 }
1064
1065 int
1066 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
1067 {
1068         return -1;
1069 }
1070
1071 int
1072 rte_vfio_container_create(void)
1073 {
1074         return -1;
1075 }
1076
1077 int
1078 rte_vfio_container_destroy(__rte_unused int container_fd)
1079 {
1080         return -1;
1081 }
1082
1083 int
1084 rte_vfio_container_group_bind(__rte_unused int container_fd,
1085                 __rte_unused int iommu_group_num)
1086 {
1087         return -1;
1088 }
1089
1090 int
1091 rte_vfio_container_group_unbind(__rte_unused int container_fd,
1092                 __rte_unused int iommu_group_num)
1093 {
1094         return -1;
1095 }
1096
1097 int
1098 rte_vfio_container_dma_map(__rte_unused int container_fd,
1099                         __rte_unused uint64_t vaddr,
1100                         __rte_unused uint64_t iova,
1101                         __rte_unused uint64_t len)
1102 {
1103         return -1;
1104 }
1105
1106 int
1107 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
1108                         __rte_unused uint64_t vaddr,
1109                         __rte_unused uint64_t iova,
1110                         __rte_unused uint64_t len)
1111 {
1112         return -1;
1113 }