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