lib/librte_eal/bsdapp/eal/eal.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   Copyright(c) 2014 6WIND S.A.
   6  *   All rights reserved.
   7  *
   8  *   Redistribution and use in source and binary forms, with or without
   9  *   modification, are permitted provided that the following conditions
  10  *   are met:
  11  *
  12  *     * Redistributions of source code must retain the above copyright
  13  *       notice, this list of conditions and the following disclaimer.
  14  *     * Redistributions in binary form must reproduce the above copyright
  15  *       notice, this list of conditions and the following disclaimer in
  16  *       the documentation and/or other materials provided with the
  17  *       distribution.
  18  *     * Neither the name of Intel Corporation nor the names of its
  19  *       contributors may be used to endorse or promote products derived
  20  *       from this software without specific prior written permission.
  21  *
  22  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  25  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  26  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  27  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  28  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  29  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  30  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  31  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  32  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  33  */
  34
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <stdint.h>
  38 #include <string.h>
  39 #include <stdarg.h>
  40 #include <unistd.h>
  41 #include <pthread.h>
  42 #include <syslog.h>
  43 #include <getopt.h>
  44 #include <sys/file.h>
  45 #include <stddef.h>
  46 #include <errno.h>
  47 #include <limits.h>
  48 #include <errno.h>
  49 #include <sys/mman.h>
  50 #include <sys/queue.h>
  51
  52 #include <rte_common.h>
  53 #include <rte_debug.h>
  54 #include <rte_memory.h>
  55 #include <rte_memzone.h>
  56 #include <rte_launch.h>
  57 #include <rte_tailq.h>
  58 #include <rte_eal.h>
  59 #include <rte_eal_memconfig.h>
  60 #include <rte_per_lcore.h>
  61 #include <rte_lcore.h>
  62 #include <rte_log.h>
  63 #include <rte_random.h>
  64 #include <rte_cycles.h>
  65 #include <rte_string_fns.h>
  66 #include <rte_cpuflags.h>
  67 #include <rte_interrupts.h>
  68 #include <rte_pci.h>
  69 #include <rte_dev.h>
  70 #include <rte_devargs.h>
  71 #include <rte_common.h>
  72 #include <rte_version.h>
  73 #include <rte_atomic.h>
  74 #include <malloc_heap.h>
  75 #include <rte_eth_ring.h>
  76
  77 #include "eal_private.h"
  78 #include "eal_thread.h"
  79 #include "eal_internal_cfg.h"
  80 #include "eal_filesystem.h"
  81 #include "eal_hugepages.h"
  82
  83 #define OPT_HUGE_DIR    "huge-dir"
  84 #define OPT_PROC_TYPE   "proc-type"
  85 #define OPT_NO_SHCONF   "no-shconf"
  86 #define OPT_NO_HPET     "no-hpet"
  87 #define OPT_VMWARE_TSC_MAP   "vmware-tsc-map"
  88 #define OPT_NO_PCI      "no-pci"
  89 #define OPT_NO_HUGE     "no-huge"
  90 #define OPT_FILE_PREFIX "file-prefix"
  91 #define OPT_SOCKET_MEM  "socket-mem"
  92 #define OPT_PCI_WHITELIST "pci-whitelist"
  93 #define OPT_PCI_BLACKLIST "pci-blacklist"
  94 #define OPT_VDEV        "vdev"
  95 #define OPT_SYSLOG      "syslog"
  96 #define OPT_LOG_LEVEL   "log-level"
  97
  98 #define MEMSIZE_IF_NO_HUGE_PAGE (64ULL * 1024ULL * 1024ULL)
  99
 100 #define BITS_PER_HEX 4
 101
 102 /* Allow the application to print its usage message too if set */
 103 static rte_usage_hook_t rte_application_usage_hook = NULL;
 104 /* early configuration structure, when memory config is not mmapped */
 105 static struct rte_mem_config early_mem_config;
 106
 107 /* define fd variable here, because file needs to be kept open for the
 108  * duration of the program, as we hold a write lock on it in the primary proc */
 109 static int mem_cfg_fd = -1;
 110
 111 static struct flock wr_lock = {
 112                 .l_type = F_WRLCK,
 113                 .l_whence = SEEK_SET,
 114                 .l_start = offsetof(struct rte_mem_config, memseg),
 115                 .l_len = sizeof(early_mem_config.memseg),
 116 };
 117
 118 /* Address of global and public configuration */
 119 static struct rte_config rte_config = {
 120                 .mem_config = &early_mem_config,
 121 };
 122
 123 /* internal configuration (per-core) */
 124 struct lcore_config lcore_config[RTE_MAX_LCORE];
 125
 126 /* internal configuration */
 127 struct internal_config internal_config;
 128
 129 /* used by rte_rdtsc() */
 130 int rte_cycles_vmware_tsc_map;
 131
 132 /* Return a pointer to the configuration structure */
 133 struct rte_config *
 134 rte_eal_get_configuration(void)
 135 {
 136         return &rte_config;
 137 }
 138
 139 /* parse a sysfs (or other) file containing one integer value */
 140 int
 141 eal_parse_sysfs_value(const char *filename, unsigned long *val)
 142 {
 143         FILE *f;
 144         char buf[BUFSIZ];
 145         char *end = NULL;
 146
 147         if ((f = fopen(filename, "r")) == NULL) {
 148                 RTE_LOG(ERR, EAL, "%s(): cannot open sysfs value %s\n",
 149                         __func__, filename);
 150                 return -1;
 151         }
 152
 153         if (fgets(buf, sizeof(buf), f) == NULL) {
 154                 RTE_LOG(ERR, EAL, "%s(): cannot read sysfs value %s\n",
 155                         __func__, filename);
 156                 fclose(f);
 157                 return -1;
 158         }
 159         *val = strtoul(buf, &end, 0);
 160         if ((buf[0] == '\0') || (end == NULL) || (*end != '\n')) {
 161                 RTE_LOG(ERR, EAL, "%s(): cannot parse sysfs value %s\n",
 162                                 __func__, filename);
 163                 fclose(f);
 164                 return -1;
 165         }
 166         fclose(f);
 167         return 0;
 168 }
 169
 170
 171 /* create memory configuration in shared/mmap memory. Take out
 172  * a write lock on the memsegs, so we can auto-detect primary/secondary.
 173  * This means we never close the file while running (auto-close on exit).
 174  * We also don't lock the whole file, so that in future we can use read-locks
 175  * on other parts, e.g. memzones, to detect if there are running secondary
 176  * processes. */
 177 static void
 178 rte_eal_config_create(void)
 179 {
 180         void *rte_mem_cfg_addr;
 181         int retval;
 182
 183         const char *pathname = eal_runtime_config_path();
 184
 185         if (internal_config.no_shconf)
 186                 return;
 187
 188         if (mem_cfg_fd < 0){
 189                 mem_cfg_fd = open(pathname, O_RDWR | O_CREAT, 0660);
 190                 if (mem_cfg_fd < 0)
 191                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 192         }
 193
 194         retval = ftruncate(mem_cfg_fd, sizeof(*rte_config.mem_config));
 195         if (retval < 0){
 196                 close(mem_cfg_fd);
 197                 rte_panic("Cannot resize '%s' for rte_mem_config\n", pathname);
 198         }
 199
 200         retval = fcntl(mem_cfg_fd, F_SETLK, &wr_lock);
 201         if (retval < 0){
 202                 close(mem_cfg_fd);
 203                 rte_exit(EXIT_FAILURE, "Cannot create lock on '%s'. Is another primary "
 204                                 "process running?\n", pathname);
 205         }
 206
 207         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 208                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 209
 210         if (rte_mem_cfg_addr == MAP_FAILED){
 211                 rte_panic("Cannot mmap memory for rte_config\n");
 212         }
 213         memcpy(rte_mem_cfg_addr, &early_mem_config, sizeof(early_mem_config));
 214         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 215 }
 216
 217 /* attach to an existing shared memory config */
 218 static void
 219 rte_eal_config_attach(void)
 220 {
 221         void *rte_mem_cfg_addr;
 222         const char *pathname = eal_runtime_config_path();
 223
 224         if (internal_config.no_shconf)
 225                 return;
 226
 227         if (mem_cfg_fd < 0){
 228                 mem_cfg_fd = open(pathname, O_RDWR);
 229                 if (mem_cfg_fd < 0)
 230                         rte_panic("Cannot open '%s' for rte_mem_config\n", pathname);
 231         }
 232
 233         rte_mem_cfg_addr = mmap(NULL, sizeof(*rte_config.mem_config),
 234                                 PROT_READ | PROT_WRITE, MAP_SHARED, mem_cfg_fd, 0);
 235         close(mem_cfg_fd);
 236         if (rte_mem_cfg_addr == MAP_FAILED)
 237                 rte_panic("Cannot mmap memory for rte_config\n");
 238
 239         rte_config.mem_config = (struct rte_mem_config *) rte_mem_cfg_addr;
 240 }
 241
 242 /* Detect if we are a primary or a secondary process */
 243 static enum rte_proc_type_t
 244 eal_proc_type_detect(void)
 245 {
 246         enum rte_proc_type_t ptype = RTE_PROC_PRIMARY;
 247         const char *pathname = eal_runtime_config_path();
 248
 249         /* if we can open the file but not get a write-lock we are a secondary
 250          * process. NOTE: if we get a file handle back, we keep that open
 251          * and don't close it to prevent a race condition between multiple opens */
 252         if (((mem_cfg_fd = open(pathname, O_RDWR)) >= 0) &&
 253                         (fcntl(mem_cfg_fd, F_SETLK, &wr_lock) < 0))
 254                 ptype = RTE_PROC_SECONDARY;
 255
 256         RTE_LOG(INFO, EAL, "Auto-detected process type: %s\n",
 257                         ptype == RTE_PROC_PRIMARY ? "PRIMARY" : "SECONDARY");
 258
 259         return ptype;
 260 }
 261
 262 /* Sets up rte_config structure with the pointer to shared memory config.*/
 263 static void
 264 rte_config_init(void)
 265 {
 266         rte_config.process_type = (internal_config.process_type == RTE_PROC_AUTO) ?
 267                         eal_proc_type_detect() : /* for auto, detect the type */
 268                         internal_config.process_type; /* otherwise use what's already set */
 269
 270         switch (rte_config.process_type){
 271         case RTE_PROC_PRIMARY:
 272                 rte_eal_config_create();
 273                 break;
 274         case RTE_PROC_SECONDARY:
 275                 rte_eal_config_attach();
 276                 rte_eal_mcfg_wait_complete(rte_config.mem_config);
 277                 break;
 278         case RTE_PROC_AUTO:
 279         case RTE_PROC_INVALID:
 280                 rte_panic("Invalid process type\n");
 281         }
 282 }
 283
 284 /* display usage */
 285 static void
 286 eal_usage(const char *prgname)
 287 {
 288         printf("\nUsage: %s -c COREMASK -n NUM [-m NB] [-r NUM] [-b <domain:bus:devid.func>]"
 289                "[--proc-type primary|secondary|auto] \n\n"
 290                "EAL options:\n"
 291                "  -c COREMASK  : A hexadecimal bitmask of cores to run on\n"
 292                "  -n NUM       : Number of memory channels\n"
 293                "  -v           : Display version information on startup\n"
 294                "  -m MB        : memory to allocate\n"
 295                "  -r NUM       : force number of memory ranks (don't detect)\n"
 296                "  --"OPT_LOG_LEVEL"  : set default log level\n"
 297                "  --"OPT_PROC_TYPE"  : type of this process\n"
 298                "  --"OPT_PCI_BLACKLIST", -b: add a PCI device in black list.\n"
 299                "               Prevent EAL from using this PCI device. The argument\n"
 300                "               format is <domain:bus:devid.func>.\n"
 301                "  --"OPT_PCI_WHITELIST", -w: add a PCI device in white list.\n"
 302                "               Only use the specified PCI devices. The argument format\n"
 303                "               is <[domain:]bus:devid.func>. This option can be present\n"
 304                "               several times (once per device).\n"
 305                "               [NOTE: PCI whitelist cannot be used with -b option]\n"
 306                "  --"OPT_VDEV": add a virtual device.\n"
 307                "               The argument format is <driver><id>[,key=val,...]\n"
 308                "               (ex: --vdev=eth_pcap0,iface=eth2).\n"
 309                "  --"OPT_VMWARE_TSC_MAP": use VMware TSC map instead of native RDTSC\n"
 310                "\nEAL options for DEBUG use only:\n"
 311                "  --"OPT_NO_HUGE"  : use malloc instead of hugetlbfs\n"
 312                "  --"OPT_NO_PCI"   : disable pci\n"
 313                "  --"OPT_NO_SHCONF": no shared config (mmap'd files)\n"
 314                "\n",
 315                prgname);
 316         /* Allow the application to print its usage message too if hook is set */
 317         if ( rte_application_usage_hook ) {
 318                 printf("===== Application Usage =====\n\n");
 319                 rte_application_usage_hook(prgname);
 320         }
 321 }
 322
 323 /* Set a per-application usage message */
 324 rte_usage_hook_t
 325 rte_set_application_usage_hook( rte_usage_hook_t usage_func )
 326 {
 327         rte_usage_hook_t        old_func;
 328
 329         /* Will be NULL on the first call to denote the last usage routine. */
 330         old_func                                        = rte_application_usage_hook;
 331         rte_application_usage_hook      = usage_func;
 332
 333         return old_func;
 334 }
 335
 336 /*
 337  * Parse the coremask given as argument (hexadecimal string) and fill
 338  * the global configuration (core role and core count) with the parsed
 339  * value.
 340  */
 341 static int xdigit2val(unsigned char c)
 342 {
 343         int val;
 344         if(isdigit(c))
 345                 val = c - '0';
 346         else if(isupper(c))
 347                 val = c - 'A' + 10;
 348         else
 349                 val = c - 'a' + 10;
 350         return val;
 351 }
 352 static int
 353 eal_parse_coremask(const char *coremask)
 354 {
 355         struct rte_config *cfg = rte_eal_get_configuration();
 356         int i, j, idx = 0 ;
 357         unsigned count = 0;
 358         char c;
 359         int val;
 360
 361         if (coremask == NULL)
 362                 return -1;
 363         /* Remove all blank characters ahead and after .
 364          * Remove 0x/0X if exists.
 365          */
 366         while (isblank(*coremask))
 367                 coremask++;
 368         if (coremask[0] == '0' && ((coremask[1] == 'x')
 369                 ||  (coremask[1] == 'X')) )
 370                 coremask += 2;
 371         i = strnlen(coremask, sysconf(_SC_ARG_MAX));
 372         while ((i > 0) && isblank(coremask[i - 1]))
 373                 i--;
 374         if (i == 0)
 375                 return -1;
 376
 377         for (i = i - 1; i >= 0 && idx < RTE_MAX_LCORE; i--) {
 378                 c = coremask[i];
 379                 if (isxdigit(c) == 0) {
 380                         /* invalid characters */
 381                         return (-1);
 382                 }
 383                 val = xdigit2val(c);
 384                 for(j = 0; j < BITS_PER_HEX && idx < RTE_MAX_LCORE; j++, idx++) {
 385                         if((1 << j) & val) {
 386                                 cfg->lcore_role[idx] = ROLE_RTE;
 387                                 if(count == 0)
 388                                         cfg->master_lcore = idx;
 389                                 count++;
 390                         } else  {
 391                                 cfg->lcore_role[idx] = ROLE_OFF;
 392                         }
 393                 }
 394         }
 395         for(; i >= 0; i--)
 396                 if(coremask[i] != '0')
 397                         return -1;
 398         for(; idx < RTE_MAX_LCORE; idx++)
 399                 cfg->lcore_role[idx] = ROLE_OFF;
 400         if(count == 0)
 401                 return -1;
 402         return 0;
 403 }
 404
 405 static int
 406 eal_parse_syslog(const char *facility)
 407 {
 408         int i;
 409         static struct {
 410                 const char *name;
 411                 int value;
 412         } map[] = {
 413                 { "auth", LOG_AUTH },
 414                 { "cron", LOG_CRON },
 415                 { "daemon", LOG_DAEMON },
 416                 { "ftp", LOG_FTP },
 417                 { "kern", LOG_KERN },
 418                 { "lpr", LOG_LPR },
 419                 { "mail", LOG_MAIL },
 420                 { "news", LOG_NEWS },
 421                 { "syslog", LOG_SYSLOG },
 422                 { "user", LOG_USER },
 423                 { "uucp", LOG_UUCP },
 424                 { "local0", LOG_LOCAL0 },
 425                 { "local1", LOG_LOCAL1 },
 426                 { "local2", LOG_LOCAL2 },
 427                 { "local3", LOG_LOCAL3 },
 428                 { "local4", LOG_LOCAL4 },
 429                 { "local5", LOG_LOCAL5 },
 430                 { "local6", LOG_LOCAL6 },
 431                 { "local7", LOG_LOCAL7 },
 432                 { NULL, 0 }
 433         };
 434
 435         for (i = 0; map[i].name; i++) {
 436                 if (!strcmp(facility, map[i].name)) {
 437                         internal_config.syslog_facility = map[i].value;
 438                         return 0;
 439                 }
 440         }
 441         return -1;
 442 }
 443
 444 static int
 445 eal_parse_log_level(const char *level, uint32_t *log_level)
 446 {
 447         char *end;
 448         unsigned long tmp;
 449
 450         errno = 0;
 451         tmp = strtoul(level, &end, 0);
 452
 453         /* check for errors */
 454         if ((errno != 0) || (level[0] == '\0') ||
 455             end == NULL || (*end != '\0'))
 456                 return -1;
 457
 458         /* log_level is a uint32_t */
 459         if (tmp >= UINT32_MAX)
 460                 return -1;
 461
 462         *log_level = tmp;
 463         return 0;
 464 }
 465
 466 static inline size_t
 467 eal_get_hugepage_mem_size(void)
 468 {
 469         uint64_t size = 0;
 470         unsigned i, j;
 471
 472         for (i = 0; i < internal_config.num_hugepage_sizes; i++) {
 473                 struct hugepage_info *hpi = &internal_config.hugepage_info[i];
 474                 if (hpi->hugedir != NULL) {
 475                         for (j = 0; j < RTE_MAX_NUMA_NODES; j++) {
 476                                 size += hpi->hugepage_sz * hpi->num_pages[j];
 477                         }
 478                 }
 479         }
 480
 481         return (size < SIZE_MAX) ? (size_t)(size) : SIZE_MAX;
 482 }
 483
 484 static enum rte_proc_type_t
 485 eal_parse_proc_type(const char *arg)
 486 {
 487         if (strncasecmp(arg, "primary", sizeof("primary")) == 0)
 488                 return RTE_PROC_PRIMARY;
 489         if (strncasecmp(arg, "secondary", sizeof("secondary")) == 0)
 490                 return RTE_PROC_SECONDARY;
 491         if (strncasecmp(arg, "auto", sizeof("auto")) == 0)
 492                 return RTE_PROC_AUTO;
 493
 494         return RTE_PROC_INVALID;
 495 }
 496
 497 /* Parse the argument given in the command line of the application */
 498 static int
 499 eal_parse_args(int argc, char **argv)
 500 {
 501         int opt, ret, i;
 502         char **argvopt;
 503         int option_index;
 504         int coremask_ok = 0;
 505         char *prgname = argv[0];
 506         static struct option lgopts[] = {
 507                 {OPT_NO_HUGE, 0, 0, 0},
 508                 {OPT_NO_PCI, 0, 0, 0},
 509                 {OPT_NO_HPET, 0, 0, 0},
 510                 {OPT_VMWARE_TSC_MAP, 0, 0, 0},
 511                 {OPT_HUGE_DIR, 1, 0, 0},
 512                 {OPT_NO_SHCONF, 0, 0, 0},
 513                 {OPT_PROC_TYPE, 1, 0, 0},
 514                 {OPT_FILE_PREFIX, 1, 0, 0},
 515                 {OPT_SOCKET_MEM, 1, 0, 0},
 516                 {OPT_PCI_WHITELIST, 1, 0, 'w'},
 517                 {OPT_PCI_BLACKLIST, 1, 0, 'b'},
 518                 {OPT_VDEV, 1, 0, 0},
 519                 {OPT_SYSLOG, 1, NULL, 0},
 520                 {OPT_LOG_LEVEL, 1, NULL, 0},
 521                 {0, 0, 0, 0}
 522         };
 523
 524         argvopt = argv;
 525
 526         internal_config.memory = 0;
 527         internal_config.force_nrank = 0;
 528         internal_config.force_nchannel = 0;
 529         internal_config.hugefile_prefix = HUGEFILE_PREFIX_DEFAULT;
 530         internal_config.hugepage_dir = NULL;
 531         internal_config.force_sockets = 0;
 532         internal_config.syslog_facility = LOG_DAEMON;
 533         /* default value from build option */
 534         internal_config.log_level = RTE_LOG_LEVEL;
 535 #ifdef RTE_LIBEAL_USE_HPET
 536         internal_config.no_hpet = 0;
 537 #else
 538         internal_config.no_hpet = 1;
 539 #endif
 540         /* zero out the NUMA config */
 541         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 542                 internal_config.socket_mem[i] = 0;
 543
 544         /* zero out hugedir descriptors */
 545         for (i = 0; i < MAX_HUGEPAGE_SIZES; i++)
 546                 internal_config.hugepage_info[i].lock_descriptor = 0;
 547
 548         internal_config.vmware_tsc_map = 0;
 549
 550         while ((opt = getopt_long(argc, argvopt, "b:w:c:m:n:r:v",
 551                                   lgopts, &option_index)) != EOF) {
 552
 553                 switch (opt) {
 554                 /* blacklist */
 555                 case 'b':
 556                         if (rte_eal_devargs_add(RTE_DEVTYPE_BLACKLISTED_PCI,
 557                                         optarg) < 0) {
 558                                 eal_usage(prgname);
 559                                 return (-1);
 560                         }
 561                         break;
 562                 /* whitelist */
 563                 case 'w':
 564                         if (rte_eal_devargs_add(RTE_DEVTYPE_WHITELISTED_PCI,
 565                                         optarg) < 0) {
 566                                 eal_usage(prgname);
 567                                 return -1;
 568                         }
 569                         break;
 570                 /* coremask */
 571                 case 'c':
 572                         if (eal_parse_coremask(optarg) < 0) {
 573                                 RTE_LOG(ERR, EAL, "invalid coremask\n");
 574                                 eal_usage(prgname);
 575                                 return -1;
 576                         }
 577                         coremask_ok = 1;
 578                         break;
 579                 /* size of memory */
 580                 case 'm':
 581                         internal_config.memory = atoi(optarg);
 582                         internal_config.memory *= 1024ULL;
 583                         internal_config.memory *= 1024ULL;
 584                         break;
 585                 /* force number of channels */
 586                 case 'n':
 587                         internal_config.force_nchannel = atoi(optarg);
 588                         if (internal_config.force_nchannel == 0 ||
 589                             internal_config.force_nchannel > 4) {
 590                                 RTE_LOG(ERR, EAL, "invalid channel number\n");
 591                                 eal_usage(prgname);
 592                                 return -1;
 593                         }
 594                         break;
 595                 /* force number of ranks */
 596                 case 'r':
 597                         internal_config.force_nrank = atoi(optarg);
 598                         if (internal_config.force_nrank == 0 ||
 599                             internal_config.force_nrank > 16) {
 600                                 RTE_LOG(ERR, EAL, "invalid rank number\n");
 601                                 eal_usage(prgname);
 602                                 return -1;
 603                         }
 604                         break;
 605                 case 'v':
 606                         /* since message is explicitly requested by user, we
 607                          * write message at highest log level so it can always be seen
 608                          * even if info or warning messages are disabled */
 609                         RTE_LOG(CRIT, EAL, "RTE Version: '%s'\n", rte_version());
 610                         break;
 611
 612                 /* long options */
 613                 case 0:
 614                         if (!strcmp(lgopts[option_index].name, OPT_NO_HUGE)) {
 615                                 internal_config.no_hugetlbfs = 1;
 616                         } else if (!strcmp(lgopts[option_index].name, OPT_NO_PCI)) {
 617                                 internal_config.no_pci = 1;
 618                         } else if (!strcmp(lgopts[option_index].name, OPT_NO_HPET)) {
 619                                 internal_config.no_hpet = 1;
 620                         } else if (!strcmp(lgopts[option_index].name, OPT_VMWARE_TSC_MAP)) {
 621                                 internal_config.vmware_tsc_map = 1;
 622                         } else if (!strcmp(lgopts[option_index].name, OPT_NO_SHCONF)) {
 623                                 internal_config.no_shconf = 1;
 624                         } else if (!strcmp(lgopts[option_index].name, OPT_PROC_TYPE)) {
 625                                 internal_config.process_type = eal_parse_proc_type(optarg);
 626                         } else if (!strcmp(lgopts[option_index].name, OPT_VDEV)) {
 627                                 if (rte_eal_devargs_add(RTE_DEVTYPE_VIRTUAL,
 628                                                 optarg) < 0) {
 629                                         eal_usage(prgname);
 630                                         return -1;
 631                                 }
 632                         } else if (!strcmp(lgopts[option_index].name, OPT_SYSLOG)) {
 633                                 if (eal_parse_syslog(optarg) < 0) {
 634                                         RTE_LOG(ERR, EAL, "invalid parameters for --"
 635                                                         OPT_SYSLOG "\n");
 636                                         eal_usage(prgname);
 637                                         return -1;
 638                                 }
 639                         } else if (!strcmp(lgopts[option_index].name,
 640                                          OPT_LOG_LEVEL)) {
 641                                 uint32_t log;
 642
 643                                 if (eal_parse_log_level(optarg, &log) < 0) {
 644                                         RTE_LOG(ERR, EAL,
 645                                                 "invalid parameters for --"
 646                                                 OPT_LOG_LEVEL "\n");
 647                                         eal_usage(prgname);
 648                                         return -1;
 649                                 }
 650                                 internal_config.log_level = log;
 651                         } else {
 652                                 RTE_LOG(ERR, EAL, "Option %s is not supported "
 653                                         "on FreeBSD\n",
 654                                         lgopts[option_index].name);
 655                                 eal_usage(prgname);
 656                                 return -1;
 657                         }
 658                         break;
 659
 660                 default:
 661                         eal_usage(prgname);
 662                         return -1;
 663                 }
 664         }
 665
 666         /* sanity checks */
 667         if (!coremask_ok) {
 668                 RTE_LOG(ERR, EAL, "coremask not specified\n");
 669                 eal_usage(prgname);
 670                 return -1;
 671         }
 672         if (internal_config.process_type == RTE_PROC_AUTO){
 673                 internal_config.process_type = eal_proc_type_detect();
 674         }
 675         if (internal_config.process_type == RTE_PROC_INVALID){
 676                 RTE_LOG(ERR, EAL, "Invalid process type specified\n");
 677                 eal_usage(prgname);
 678                 return -1;
 679         }
 680         if (internal_config.process_type == RTE_PROC_PRIMARY &&
 681                         internal_config.force_nchannel == 0) {
 682                 RTE_LOG(ERR, EAL, "Number of memory channels (-n) not specified\n");
 683                 eal_usage(prgname);
 684                 return -1;
 685         }
 686         if (index(internal_config.hugefile_prefix,'%') != NULL){
 687                 RTE_LOG(ERR, EAL, "Invalid char, '%%', in '"OPT_FILE_PREFIX"' option\n");
 688                 eal_usage(prgname);
 689                 return -1;
 690         }
 691         if (internal_config.memory > 0 && internal_config.force_sockets == 1) {
 692                 RTE_LOG(ERR, EAL, "Options -m and --socket-mem cannot be specified "
 693                                 "at the same time\n");
 694                 eal_usage(prgname);
 695                 return -1;
 696         }
 697         /* --no-huge doesn't make sense with either -m or --socket-mem */
 698         if (internal_config.no_hugetlbfs &&
 699                         (internal_config.memory > 0 ||
 700                                         internal_config.force_sockets == 1)) {
 701                 RTE_LOG(ERR, EAL, "Options -m or --socket-mem cannot be specified "
 702                                 "together with --no-huge!\n");
 703                 eal_usage(prgname);
 704                 return -1;
 705         }
 706
 707         if (rte_eal_devargs_type_count(RTE_DEVTYPE_WHITELISTED_PCI) != 0 &&
 708                 rte_eal_devargs_type_count(RTE_DEVTYPE_BLACKLISTED_PCI) != 0) {
 709                 RTE_LOG(ERR, EAL, "Error: blacklist [-b] and whitelist "
 710                         "[-w] options cannot be used at the same time\n");
 711                 eal_usage(prgname);
 712                 return -1;
 713         }
 714
 715         if (optind >= 0)
 716                 argv[optind-1] = prgname;
 717
 718         /* if no memory amounts were requested, this will result in 0 and
 719          * will be overriden later, right after eal_hugepage_info_init() */
 720         for (i = 0; i < RTE_MAX_NUMA_NODES; i++)
 721                 internal_config.memory += internal_config.socket_mem[i];
 722
 723         ret = optind-1;
 724         optind = 0; /* reset getopt lib */
 725         return ret;
 726 }
 727
 728 static void
 729 eal_check_mem_on_local_socket(void)
 730 {
 731         const struct rte_memseg *ms;
 732         int i, socket_id;
 733
 734         socket_id = rte_lcore_to_socket_id(rte_config.master_lcore);
 735
 736         ms = rte_eal_get_physmem_layout();
 737
 738         for (i = 0; i < RTE_MAX_MEMSEG; i++)
 739                 if (ms[i].socket_id == socket_id &&
 740                                 ms[i].len > 0)
 741                         return;
 742
 743         RTE_LOG(WARNING, EAL, "WARNING: Master core has no "
 744                         "memory on local socket!\n");
 745 }
 746
 747 static int
 748 sync_func(__attribute__((unused)) void *arg)
 749 {
 750         return 0;
 751 }
 752
 753 inline static void
 754 rte_eal_mcfg_complete(void)
 755 {
 756         /* ALL shared mem_config related INIT DONE */
 757         if (rte_config.process_type == RTE_PROC_PRIMARY)
 758                 rte_config.mem_config->magic = RTE_MAGIC;
 759 }
 760
 761 /* return non-zero if hugepages are enabled. */
 762 int rte_eal_has_hugepages(void)
 763 {
 764         return !internal_config.no_hugetlbfs;
 765 }
 766
 767 /* Abstraction for port I/0 privilege */
 768 static int
 769 rte_eal_iopl_init(void)
 770 {
 771         int fd = -1;
 772         fd = open("/dev/io", O_RDWR);
 773         if (fd < 0)
 774                 return -1;
 775         return 0;
 776 }
 777
 778 /* Launch threads, called at application init(). */
 779 int
 780 rte_eal_init(int argc, char **argv)
 781 {
 782         int i, fctret, ret;
 783         pthread_t thread_id;
 784         static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
 785
 786         if (!rte_atomic32_test_and_set(&run_once))
 787                 return -1;
 788
 789         thread_id = pthread_self();
 790
 791         if (rte_eal_log_early_init() < 0)
 792                 rte_panic("Cannot init early logs\n");
 793
 794         fctret = eal_parse_args(argc, argv);
 795         if (fctret < 0)
 796                 exit(1);
 797
 798         /* set log level as early as possible */
 799         rte_set_log_level(internal_config.log_level);
 800
 801         if (internal_config.no_hugetlbfs == 0 &&
 802                         internal_config.process_type != RTE_PROC_SECONDARY &&
 803                         eal_hugepage_info_init() < 0)
 804                 rte_panic("Cannot get hugepage information\n");
 805
 806         if (internal_config.memory == 0 && internal_config.force_sockets == 0) {
 807                 if (internal_config.no_hugetlbfs)
 808                         internal_config.memory = MEMSIZE_IF_NO_HUGE_PAGE;
 809                 else
 810                         internal_config.memory = eal_get_hugepage_mem_size();
 811         }
 812
 813         if (internal_config.vmware_tsc_map == 1) {
 814 #ifdef RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT
 815                 rte_cycles_vmware_tsc_map = 1;
 816                 RTE_LOG (DEBUG, EAL, "Using VMWARE TSC MAP, "
 817                                 "you must have monitor_control.pseudo_perfctr = TRUE\n");
 818 #else
 819                 RTE_LOG (WARNING, EAL, "Ignoring --vmware-tsc-map because "
 820                                 "RTE_LIBRTE_EAL_VMWARE_TSC_MAP_SUPPORT is not set\n");
 821 #endif
 822         }
 823
 824         rte_srand(rte_rdtsc());
 825
 826         rte_config_init();
 827
 828         if (rte_eal_iopl_init() == 0)
 829                 rte_config.flags |= EAL_FLG_HIGH_IOPL;
 830
 831         if (rte_eal_cpu_init() < 0)
 832                 rte_panic("Cannot detect lcores\n");
 833
 834         if (rte_eal_memory_init() < 0)
 835                 rte_panic("Cannot init memory\n");
 836
 837         if (rte_eal_memzone_init() < 0)
 838                 rte_panic("Cannot init memzone\n");
 839
 840         if (rte_eal_tailqs_init() < 0)
 841                 rte_panic("Cannot init tail queues for objects\n");
 842
 843 /*      if (rte_eal_log_init(argv[0], internal_config.syslog_facility) < 0)
 844                 rte_panic("Cannot init logs\n");*/
 845
 846         if (rte_eal_alarm_init() < 0)
 847                 rte_panic("Cannot init interrupt-handling thread\n");
 848
 849         if (rte_eal_intr_init() < 0)
 850                 rte_panic("Cannot init interrupt-handling thread\n");
 851
 852         if (rte_eal_timer_init() < 0)
 853                 rte_panic("Cannot init HPET or TSC timers\n");
 854
 855         if (rte_eal_pci_init() < 0)
 856                 rte_panic("Cannot init PCI\n");
 857
 858         RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
 859                 rte_config.master_lcore, thread_id);
 860
 861         eal_check_mem_on_local_socket();
 862
 863         rte_eal_mcfg_complete();
 864
 865         if (rte_eal_dev_init(PMD_INIT_PRE_PCI_PROBE) < 0)
 866                 rte_panic("Cannot init pmd devices\n");
 867
 868         RTE_LCORE_FOREACH_SLAVE(i) {
 869
 870                 /*
 871                  * create communication pipes between master thread
 872                  * and children
 873                  */
 874                 if (pipe(lcore_config[i].pipe_master2slave) < 0)
 875                         rte_panic("Cannot create pipe\n");
 876                 if (pipe(lcore_config[i].pipe_slave2master) < 0)
 877                         rte_panic("Cannot create pipe\n");
 878
 879                 lcore_config[i].state = WAIT;
 880
 881                 /* create a thread for each lcore */
 882                 ret = pthread_create(&lcore_config[i].thread_id, NULL,
 883                                      eal_thread_loop, NULL);
 884                 if (ret != 0)
 885                         rte_panic("Cannot create thread\n");
 886         }
 887
 888         eal_thread_init_master(rte_config.master_lcore);
 889
 890         /*
 891          * Launch a dummy function on all slave lcores, so that master lcore
 892          * knows they are all ready when this function returns.
 893          */
 894         rte_eal_mp_remote_launch(sync_func, NULL, SKIP_MASTER);
 895         rte_eal_mp_wait_lcore();
 896
 897         /* Probe & Initialize PCI devices */
 898         if (rte_eal_pci_probe())
 899                         rte_panic("Cannot probe PCI\n");
 900
 901         /* Initialize any outstanding devices */
 902         if (rte_eal_dev_init(PMD_INIT_POST_PCI_PROBE) < 0)
 903                 rte_panic("Cannot init pmd devices\n");
 904
 905         return fctret;
 906 }
 907
 908 /* get core role */
 909 enum rte_lcore_role_t
 910 rte_eal_lcore_role(unsigned lcore_id)
 911 {
 912         return (rte_config.lcore_role[lcore_id]);
 913 }
 914
 915 enum rte_proc_type_t
 916 rte_eal_process_type(void)
 917 {
 918         return (rte_config.process_type);
 919 }
 920